[llvm] 9d9d68f - AMDGPU: Fix using underaligned atomicrmw in atomic scan tests
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Mon Jul 7 05:44:15 PDT 2025
Author: Matt Arsenault
Date: 2025-07-07T21:43:48+09:00
New Revision: 9d9d68fd3b3e65398b6b106bfaa803e86befe44b
URL: https://github.com/llvm/llvm-project/commit/9d9d68fd3b3e65398b6b106bfaa803e86befe44b
DIFF: https://github.com/llvm/llvm-project/commit/9d9d68fd3b3e65398b6b106bfaa803e86befe44b.diff
LOG: AMDGPU: Fix using underaligned atomicrmw in atomic scan tests
These do not work correctly and emit a broken libcall. The actual
lowering of the atomic is unimportant for the test anyway.
Added:
Modified:
llvm/test/CodeGen/AMDGPU/global_atomics_scan_fadd.ll
llvm/test/CodeGen/AMDGPU/global_atomics_scan_fmax.ll
llvm/test/CodeGen/AMDGPU/global_atomics_scan_fmin.ll
llvm/test/CodeGen/AMDGPU/global_atomics_scan_fsub.ll
Removed:
################################################################################
diff --git a/llvm/test/CodeGen/AMDGPU/global_atomics_scan_fadd.ll b/llvm/test/CodeGen/AMDGPU/global_atomics_scan_fadd.ll
index 82c58394c03bb..37756d15861be 100644
--- a/llvm/test/CodeGen/AMDGPU/global_atomics_scan_fadd.ll
+++ b/llvm/test/CodeGen/AMDGPU/global_atomics_scan_fadd.ll
@@ -5736,966 +5736,469 @@ define amdgpu_kernel void @global_atomic_fadd_uni_address_div_value_default_scop
define amdgpu_kernel void @global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe(ptr addrspace(1) %ptr) #0 {
; GFX7LESS-LABEL: global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe:
; GFX7LESS: ; %bb.0:
-; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], exec
-; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v3, s0, 0
-; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v3, s1, v3
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX7LESS-NEXT: s_and_saveexec_b64 s[2:3], vcc
+; GFX7LESS-NEXT: s_mov_b64 s[2:3], exec
+; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s2, 0
+; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s3, v0
+; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX7LESS-NEXT: s_cbranch_execz .LBB9_3
; GFX7LESS-NEXT: ; %bb.1:
-; GFX7LESS-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
-; GFX7LESS-NEXT: s_bcnt1_i32_b64 s2, s[0:1]
-; GFX7LESS-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-NEXT: v_lshlrev_b32_e32 v3, 20, v2
+; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
+; GFX7LESS-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX7LESS-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX7LESS-NEXT: v_cvt_f64_u32_e32 v[1:2], s2
-; GFX7LESS-NEXT: v_or_b32_e32 v4, v0, v4
-; GFX7LESS-NEXT: v_mul_f64 v[41:42], v[1:2], 4.0
+; GFX7LESS-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
+; GFX7LESS-NEXT: v_cvt_f64_u32_e32 v[0:1], s2
+; GFX7LESS-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, s0
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, s1
-; GFX7LESS-NEXT: v_or_b32_e32 v40, v4, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, s6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, s7
+; GFX7LESS-NEXT: s_mov_b32 s2, -1
; GFX7LESS-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-NEXT: v_add_f64 v[2:3], v[0:1], v[41:42]
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:12
-; GFX7LESS-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:8
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX7LESS-NEXT: s_waitcnt expcnt(2)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[0:3], 0 glc
+; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-NEXT: s_cbranch_execnz .LBB9_2
; GFX7LESS-NEXT: .LBB9_3:
; GFX7LESS-NEXT: s_endpgm
;
; GFX9-LABEL: global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-NEXT: s_mov_b32 s66, -1
-; GFX9-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-NEXT: s_mov_b64 s[0:1], exec
-; GFX9-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-NEXT: v_mbcnt_lo_u32_b32 v3, s0, 0
-; GFX9-NEXT: s_add_u32 s64, s64, s11
-; GFX9-NEXT: v_mbcnt_hi_u32_b32 v3, s1, v3
-; GFX9-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX9-NEXT: s_movk_i32 s32, 0x800
-; GFX9-NEXT: s_and_saveexec_b64 s[2:3], vcc
+; GFX9-NEXT: s_mov_b64 s[2:3], exec
+; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
+; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
+; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX9-NEXT: s_cbranch_execz .LBB9_3
; GFX9-NEXT: ; %bb.1:
-; GFX9-NEXT: s_bcnt1_i32_b64 s0, s[0:1]
-; GFX9-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX9-NEXT: v_cvt_f64_u32_e32 v[3:4], s0
-; GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-NEXT: s_mov_b32 s33, s10
-; GFX9-NEXT: v_mul_f64 v[41:42], v[3:4], 4.0
+; GFX9-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
+; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9-NEXT: v_cvt_f64_u32_e32 v[0:1], s2
+; GFX9-NEXT: s_mov_b64 s[2:3], 0
+; GFX9-NEXT: v_mov_b32_e32 v6, 0
+; GFX9-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX9-NEXT: s_mov_b32 s50, s9
-; GFX9-NEXT: s_mov_b32 s51, s8
-; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
+; GFX9-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v4, s1
-; GFX9-NEXT: s_mov_b64 s[54:55], 0
-; GFX9-NEXT: v_mov_b32_e32 v3, s0
-; GFX9-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-NEXT: v_mov_b32_e32 v3, s5
; GFX9-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f64 v[0:1], v[3:4], v[41:42]
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:4
-; GFX9-NEXT: buffer_store_dword v3, off, s[64:67], 0
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX9-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: buffer_load_dword v3, off, s[64:67], 0
-; GFX9-NEXT: buffer_load_dword v4, off, s[64:67], 0 offset:4
-; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-NEXT: s_cbranch_execnz .LBB9_2
; GFX9-NEXT: .LBB9_3:
; GFX9-NEXT: s_endpgm
;
; GFX1064-LABEL: global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1064: ; %bb.0:
-; GFX1064-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-NEXT: s_mov_b32 s66, -1
-; GFX1064-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-NEXT: s_add_u32 s64, s64, s11
-; GFX1064-NEXT: s_mov_b32 s33, s10
-; GFX1064-NEXT: s_mov_b64 s[10:11], exec
-; GFX1064-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v3, s10, 0
-; GFX1064-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-NEXT: s_movk_i32 s32, 0x800
-; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v3, s11, v3
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX1064-NEXT: s_mov_b64 s[2:3], exec
+; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
+; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
+; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX1064-NEXT: s_cbranch_execz .LBB9_3
; GFX1064-NEXT: ; %bb.1:
-; GFX1064-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1064-NEXT: s_bcnt1_i32_b64 s0, s[10:11]
-; GFX1064-NEXT: s_mov_b32 s50, s9
-; GFX1064-NEXT: v_cvt_f64_u32_e32 v[3:4], s0
-; GFX1064-NEXT: s_mov_b32 s51, s8
-; GFX1064-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1064-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1064-NEXT: s_mov_b64 s[54:55], 0
+; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1064-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
+; GFX1064-NEXT: v_mov_b32_e32 v6, 0
+; GFX1064-NEXT: v_cvt_f64_u32_e32 v[0:1], s2
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1064-NEXT: v_mul_f64 v[41:42], v[3:4], 4.0
-; GFX1064-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1064-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1064-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX1064-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX1064-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: v_mov_b32_e32 v2, s1
-; GFX1064-NEXT: v_mov_b32_e32 v1, s0
+; GFX1064-NEXT: v_mov_b32_e32 v2, s2
+; GFX1064-NEXT: v_mov_b32_e32 v3, s3
+; GFX1064-NEXT: s_mov_b64 s[2:3], 0
; GFX1064-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1064-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1064-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX1064-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1064-NEXT: s_add_u32 s8, s36, 44
-; GFX1064-NEXT: s_addc_u32 s9, s37, 0
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: s_clause 0x1
-; GFX1064-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX1064-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1064-NEXT: v_mov_b32_e32 v3, v1
+; GFX1064-NEXT: v_mov_b32_e32 v2, v0
+; GFX1064-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1064-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX1064-NEXT: s_cbranch_execnz .LBB9_2
; GFX1064-NEXT: .LBB9_3:
; GFX1064-NEXT: s_endpgm
;
; GFX1032-LABEL: global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1032: ; %bb.0:
-; GFX1032-NEXT: s_mov_b32 s50, s9
-; GFX1032-NEXT: s_mov_b32 s9, exec_lo
-; GFX1032-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v3, s9, 0
-; GFX1032-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-NEXT: s_mov_b32 s66, -1
-; GFX1032-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-NEXT: s_add_u32 s64, s64, s11
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
-; GFX1032-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-NEXT: s_mov_b32 s54, 0
-; GFX1032-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-NEXT: s_mov_b32 s3, exec_lo
+; GFX1032-NEXT: s_mov_b32 s2, 0
+; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, s3, 0
+; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX1032-NEXT: s_cbranch_execz .LBB9_3
; GFX1032-NEXT: ; %bb.1:
-; GFX1032-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1032-NEXT: s_bcnt1_i32_b32 s0, s9
-; GFX1032-NEXT: s_mov_b32 s33, s10
-; GFX1032-NEXT: v_cvt_f64_u32_e32 v[3:4], s0
-; GFX1032-NEXT: s_mov_b32 s51, s8
-; GFX1032-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1032-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1032-NEXT: s_bcnt1_i32_b32 s3, s3
+; GFX1032-NEXT: v_mov_b32_e32 v6, 0
+; GFX1032-NEXT: v_cvt_f64_u32_e32 v[0:1], s3
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1032-NEXT: v_mul_f64 v[41:42], v[3:4], 4.0
-; GFX1032-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1032-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1032-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX1032-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX1032-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: v_mov_b32_e32 v2, s1
-; GFX1032-NEXT: v_mov_b32_e32 v1, s0
+; GFX1032-NEXT: v_mov_b32_e32 v2, s4
+; GFX1032-NEXT: v_mov_b32_e32 v3, s5
; GFX1032-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1032-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1032-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX1032-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1032-NEXT: s_add_u32 s8, s36, 44
-; GFX1032-NEXT: s_addc_u32 s9, s37, 0
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: s_clause 0x1
-; GFX1032-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
+; GFX1032-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1032-NEXT: v_mov_b32_e32 v3, v1
+; GFX1032-NEXT: v_mov_b32_e32 v2, v0
+; GFX1032-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
; GFX1032-NEXT: s_cbranch_execnz .LBB9_2
; GFX1032-NEXT: .LBB9_3:
; GFX1032-NEXT: s_endpgm
;
; GFX1164-LABEL: global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1164: ; %bb.0:
-; GFX1164-NEXT: s_mov_b32 s33, s10
-; GFX1164-NEXT: s_mov_b64 s[10:11], exec
-; GFX1164-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, s10, 0
-; GFX1164-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-NEXT: s_mov_b32 s32, 32
; GFX1164-NEXT: s_mov_b64 s[0:1], exec
+; GFX1164-NEXT: s_mov_b64 s[2:3], exec
+; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, s11, v0
+; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, s1, v0
; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1164-NEXT: s_cbranch_execz .LBB9_3
; GFX1164-NEXT: ; %bb.1:
-; GFX1164-NEXT: s_bcnt1_i32_b64 s0, s[10:11]
-; GFX1164-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
+; GFX1164-NEXT: s_bcnt1_i32_b64 s0, s[0:1]
+; GFX1164-NEXT: v_mov_b32_e32 v6, 0
; GFX1164-NEXT: v_cvt_f64_u32_e32 v[0:1], s0
-; GFX1164-NEXT: s_mov_b32 s50, s9
-; GFX1164-NEXT: s_mov_b32 s51, s8
-; GFX1164-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1164-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1164-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1164-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-NEXT: v_mul_f64 v[41:42], v[0:1], 4.0
+; GFX1164-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: v_mov_b32_e32 v2, s1
-; GFX1164-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-NEXT: .p2align 6
+; GFX1164-NEXT: v_mov_b32_e32 v2, s2
+; GFX1164-NEXT: v_mov_b32_e32 v3, s3
+; GFX1164-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1164-NEXT: s_waitcnt vmcnt(0)
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1164-NEXT: s_add_u32 s8, s36, 44
-; GFX1164-NEXT: s_addc_u32 s9, s37, 0
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1164-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-NEXT: s_cbranch_execnz .LBB9_2
; GFX1164-NEXT: .LBB9_3:
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-NEXT: s_endpgm
;
; GFX1132-LABEL: global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1132: ; %bb.0:
-; GFX1132-NEXT: s_mov_b32 s8, exec_lo
-; GFX1132-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, s8, 0
-; GFX1132-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1132-NEXT: s_mov_b32 s54, 0
-; GFX1132-NEXT: s_mov_b32 s32, 32
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
+; GFX1132-NEXT: s_mov_b32 s2, 0
+; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
+; GFX1132-NEXT: s_mov_b32 s1, exec_lo
+; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-NEXT: s_cbranch_execz .LBB9_3
; GFX1132-NEXT: ; %bb.1:
-; GFX1132-NEXT: s_bcnt1_i32_b32 s0, s8
-; GFX1132-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
+; GFX1132-NEXT: s_bcnt1_i32_b32 s0, s0
+; GFX1132-NEXT: v_mov_b32_e32 v6, 0
; GFX1132-NEXT: v_cvt_f64_u32_e32 v[0:1], s0
-; GFX1132-NEXT: s_mov_b32 s33, s15
-; GFX1132-NEXT: s_mov_b32 s50, s14
-; GFX1132-NEXT: s_mov_b32 s51, s13
-; GFX1132-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1132-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1132-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-NEXT: v_mul_f64 v[41:42], v[0:1], 4.0
+; GFX1132-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-NEXT: .p2align 6
+; GFX1132-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX1132-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1132-NEXT: s_waitcnt vmcnt(0)
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1132-NEXT: s_add_u32 s8, s36, 44
-; GFX1132-NEXT: s_addc_u32 s9, s37, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1132-NEXT: s_mov_b32 s12, s51
-; GFX1132-NEXT: s_mov_b32 s13, s50
-; GFX1132-NEXT: s_mov_b32 s14, s33
-; GFX1132-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1132-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-NEXT: s_cbranch_execnz .LBB9_2
; GFX1132-NEXT: .LBB9_3:
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-NEXT: s_endpgm
;
; GFX7LESS-DPP-LABEL: global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe:
; GFX7LESS-DPP: ; %bb.0:
-; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], exec
-; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v3, s0, 0
-; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v3, s1, v3
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[2:3], vcc
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], exec
+; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s2, 0
+; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s3, v0
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX7LESS-DPP-NEXT: s_cbranch_execz .LBB9_3
; GFX7LESS-DPP-NEXT: ; %bb.1:
-; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-DPP-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
-; GFX7LESS-DPP-NEXT: s_bcnt1_i32_b64 s2, s[0:1]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX7LESS-DPP-NEXT: v_cvt_f64_u32_e32 v[1:2], s2
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v4, v0, v4
-; GFX7LESS-DPP-NEXT: v_mul_f64 v[41:42], v[1:2], 4.0
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s3, 0xf000
+; GFX7LESS-DPP-NEXT: v_cvt_f64_u32_e32 v[0:1], s2
+; GFX7LESS-DPP-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, s0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, s1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v40, v4, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s7
+; GFX7LESS-DPP-NEXT: s_mov_b32 s2, -1
; GFX7LESS-DPP-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-DPP-NEXT: v_add_f64 v[2:3], v[0:1], v[41:42]
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:8
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(2)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-DPP-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[0:3], 0 glc
+; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB9_2
; GFX7LESS-DPP-NEXT: .LBB9_3:
; GFX7LESS-DPP-NEXT: s_endpgm
;
; GFX9-DPP-LABEL: global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe:
; GFX9-DPP: ; %bb.0:
-; GFX9-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-DPP-NEXT: s_mov_b32 s66, -1
-; GFX9-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], exec
-; GFX9-DPP-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, s0, 0
-; GFX9-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, s1, v3
-; GFX9-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX9-DPP-NEXT: s_and_saveexec_b64 s[2:3], vcc
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], exec
+; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
+; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
+; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX9-DPP-NEXT: s_cbranch_execz .LBB9_3
; GFX9-DPP-NEXT: ; %bb.1:
-; GFX9-DPP-NEXT: s_bcnt1_i32_b64 s0, s[0:1]
-; GFX9-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX9-DPP-NEXT: v_cvt_f64_u32_e32 v[3:4], s0
-; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-DPP-NEXT: s_mov_b32 s33, s10
-; GFX9-DPP-NEXT: v_mul_f64 v[41:42], v[3:4], 4.0
+; GFX9-DPP-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
+; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9-DPP-NEXT: v_cvt_f64_u32_e32 v[0:1], s2
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
+; GFX9-DPP-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX9-DPP-NEXT: s_mov_b32 s50, s9
-; GFX9-DPP-NEXT: s_mov_b32 s51, s8
-; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
+; GFX9-DPP-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, s1
-; GFX9-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s0
-; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s5
; GFX9-DPP-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX9-DPP-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX9-DPP-NEXT: v_add_f64 v[0:1], v[3:4], v[41:42]
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:4
-; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX9-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-DPP-NEXT: buffer_load_dword v3, off, s[64:67], 0
-; GFX9-DPP-NEXT: buffer_load_dword v4, off, s[64:67], 0 offset:4
-; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-DPP-NEXT: s_cbranch_execnz .LBB9_2
; GFX9-DPP-NEXT: .LBB9_3:
; GFX9-DPP-NEXT: s_endpgm
;
; GFX1064-DPP-LABEL: global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1064-DPP: ; %bb.0:
-; GFX1064-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1064-DPP-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], exec
-; GFX1064-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, s10, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, s11, v3
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], exec
+; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
+; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
+; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX1064-DPP-NEXT: s_cbranch_execz .LBB9_3
; GFX1064-DPP-NEXT: ; %bb.1:
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1064-DPP-NEXT: s_bcnt1_i32_b64 s0, s[10:11]
-; GFX1064-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1064-DPP-NEXT: v_cvt_f64_u32_e32 v[3:4], s0
-; GFX1064-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1064-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1064-DPP-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
+; GFX1064-DPP-NEXT: v_cvt_f64_u32_e32 v[0:1], s2
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1064-DPP-NEXT: v_mul_f64 v[41:42], v[3:4], 4.0
-; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX1064-DPP-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, s0
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s2
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s3
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1064-DPP-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1064-DPP-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1064-DPP-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1064-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1064-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-DPP-NEXT: s_clause 0x1
-; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX1064-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX1064-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB9_2
; GFX1064-DPP-NEXT: .LBB9_3:
; GFX1064-DPP-NEXT: s_endpgm
;
; GFX1032-DPP-LABEL: global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1032-DPP: ; %bb.0:
-; GFX1032-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1032-DPP-NEXT: s_mov_b32 s9, exec_lo
-; GFX1032-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, s9, 0
-; GFX1032-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1032-DPP-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
-; GFX1032-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-DPP-NEXT: s_mov_b32 s3, exec_lo
+; GFX1032-DPP-NEXT: s_mov_b32 s2, 0
+; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s3, 0
+; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX1032-DPP-NEXT: s_cbranch_execz .LBB9_3
; GFX1032-DPP-NEXT: ; %bb.1:
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1032-DPP-NEXT: s_bcnt1_i32_b32 s0, s9
-; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1032-DPP-NEXT: v_cvt_f64_u32_e32 v[3:4], s0
-; GFX1032-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1032-DPP-NEXT: s_bcnt1_i32_b32 s3, s3
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
+; GFX1032-DPP-NEXT: v_cvt_f64_u32_e32 v[0:1], s3
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1032-DPP-NEXT: v_mul_f64 v[41:42], v[3:4], 4.0
-; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX1032-DPP-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, s0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s4
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s5
; GFX1032-DPP-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1032-DPP-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1032-DPP-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1032-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1032-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-DPP-NEXT: s_clause 0x1
-; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
+; GFX1032-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX1032-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB9_2
; GFX1032-DPP-NEXT: .LBB9_3:
; GFX1032-DPP-NEXT: s_endpgm
;
; GFX1164-DPP-LABEL: global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1164-DPP: ; %bb.0:
-; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], exec
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s10, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], exec
+; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, s11, v0
+; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, s1, v0
; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1164-DPP-NEXT: s_cbranch_execz .LBB9_3
; GFX1164-DPP-NEXT: ; %bb.1:
-; GFX1164-DPP-NEXT: s_bcnt1_i32_b64 s0, s[10:11]
-; GFX1164-DPP-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
+; GFX1164-DPP-NEXT: s_bcnt1_i32_b64 s0, s[0:1]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
; GFX1164-DPP-NEXT: v_cvt_f64_u32_e32 v[0:1], s0
-; GFX1164-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1164-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1164-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1164-DPP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_mul_f64 v[41:42], v[0:1], 4.0
+; GFX1164-DPP-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-DPP-NEXT: .p2align 6
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s2
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s3
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-DPP-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1164-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1164-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[54:55]
-; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB9_2
+; GFX1164-DPP-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
+; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB9_2
; GFX1164-DPP-NEXT: .LBB9_3:
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-DPP-NEXT: s_endpgm
;
; GFX1132-DPP-LABEL: global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1132-DPP: ; %bb.0:
-; GFX1132-DPP-NEXT: s_mov_b32 s8, exec_lo
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s8, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1132-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
+; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
+; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
+; GFX1132-DPP-NEXT: s_mov_b32 s1, exec_lo
+; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-DPP-NEXT: s_cbranch_execz .LBB9_3
; GFX1132-DPP-NEXT: ; %bb.1:
-; GFX1132-DPP-NEXT: s_bcnt1_i32_b32 s0, s8
-; GFX1132-DPP-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
+; GFX1132-DPP-NEXT: s_bcnt1_i32_b32 s0, s0
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v6, 0
; GFX1132-DPP-NEXT: v_cvt_f64_u32_e32 v[0:1], s0
-; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
-; GFX1132-DPP-NEXT: s_mov_b32 s50, s14
-; GFX1132-DPP-NEXT: s_mov_b32 s51, s13
-; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1132-DPP-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_mul_f64 v[41:42], v[0:1], 4.0
+; GFX1132-DPP-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-DPP-NEXT: .p2align 6
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX1132-DPP-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1132-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1132-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1132-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1132-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-DPP-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB9_2
; GFX1132-DPP-NEXT: .LBB9_3:
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-DPP-NEXT: s_endpgm
- %result = atomicrmw fadd ptr addrspace(1) %ptr, double 4.0 syncscope("agent") monotonic, align 4, !amdgpu.no.fine.grained.memory !1, !amdgpu.ignore.denormal.mode !1
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, double 4.0 syncscope("agent") monotonic, align 8, !amdgpu.no.fine.grained.memory !1, !amdgpu.ignore.denormal.mode !1
ret void
}
define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_agent_scope_align4_unsafe(ptr addrspace(1) %ptr) #0 {
; GFX7LESS-LABEL: global_atomic_fadd_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX7LESS: ; %bb.0:
-; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-NEXT: s_mov_b32 s32, 0
+; GFX7LESS-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX7LESS-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX7LESS-NEXT: s_mov_b32 s38, -1
+; GFX7LESS-NEXT: s_mov_b32 s39, 0xe8f000
+; GFX7LESS-NEXT: s_add_u32 s36, s36, s11
+; GFX7LESS-NEXT: s_addc_u32 s37, s37, 0
+; GFX7LESS-NEXT: s_mov_b32 s14, s10
+; GFX7LESS-NEXT: s_mov_b32 s13, s9
+; GFX7LESS-NEXT: s_mov_b32 s12, s8
+; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX7LESS-NEXT: s_add_u32 s8, s34, 44
+; GFX7LESS-NEXT: s_addc_u32 s9, s35, 0
+; GFX7LESS-NEXT: s_getpc_b64 s[4:5]
+; GFX7LESS-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX7LESS-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
+; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GFX7LESS-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX7LESS-NEXT: v_or_b32_e32 v40, v0, v2
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX7LESS-NEXT: v_or_b32_e32 v31, v0, v2
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[38:39]
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX7LESS-NEXT: s_mov_b64 s[0:1], exec
-; GFX7LESS-NEXT: v_mov_b32_e32 v41, 0
-; GFX7LESS-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
+; GFX7LESS-NEXT: v_bfrev_b32_e32 v5, 1
; GFX7LESS-NEXT: .LBB10_1: ; %ComputeLoop
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-NEXT: s_ff1_i32_b64 s4, s[0:1]
@@ -6705,7 +6208,7 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_agent
; GFX7LESS-NEXT: s_andn2_b64 s[0:1], s[0:1], s[4:5]
; GFX7LESS-NEXT: v_cmp_ne_u64_e64 s[4:5], s[0:1], 0
; GFX7LESS-NEXT: s_and_b64 vcc, exec, s[4:5]
-; GFX7LESS-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX7LESS-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX7LESS-NEXT: s_cbranch_vccnz .LBB10_1
; GFX7LESS-NEXT: ; %bb.2: ; %ComputeEnd
; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
@@ -6715,102 +6218,70 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_agent
; GFX7LESS-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX7LESS-NEXT: s_cbranch_execz .LBB10_5
; GFX7LESS-NEXT: ; %bb.3:
-; GFX7LESS-NEXT: s_load_dwordx2 s[52:53], s[36:37], 0x9
-; GFX7LESS-NEXT: s_mov_b32 s55, 0xf000
-; GFX7LESS-NEXT: s_mov_b32 s54, -1
+; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x9
+; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
+; GFX7LESS-NEXT: s_mov_b32 s2, -1
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: buffer_load_dwordx2 v[0:1], off, s[52:55], 0
-; GFX7LESS-NEXT: s_mov_b64 s[54:55], 0
+; GFX7LESS-NEXT: buffer_load_dwordx2 v[2:3], off, s[0:3], 0
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
; GFX7LESS-NEXT: .LBB10_4: ; %atomicrmw.start
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-NEXT: v_add_f64 v[2:3], v[0:1], v[41:42]
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:12
-; GFX7LESS-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:8
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX7LESS-NEXT: s_waitcnt expcnt(2)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[0:3], 0 glc
+; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-NEXT: s_cbranch_execnz .LBB10_4
; GFX7LESS-NEXT: .LBB10_5:
; GFX7LESS-NEXT: s_endpgm
;
; GFX9-LABEL: global_atomic_fadd_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-NEXT: s_mov_b32 s66, -1
-; GFX9-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-NEXT: s_add_u32 s64, s64, s11
-; GFX9-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-NEXT: s_mov_b32 s51, s8
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_mov_b32 s50, s9
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX9-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX9-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX9-NEXT: s_mov_b32 s38, -1
+; GFX9-NEXT: s_mov_b32 s39, 0xe00000
+; GFX9-NEXT: s_add_u32 s36, s36, s11
+; GFX9-NEXT: s_addc_u32 s37, s37, 0
+; GFX9-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX9-NEXT: s_mov_b32 s12, s8
+; GFX9-NEXT: s_add_u32 s8, s34, 44
+; GFX9-NEXT: s_mov_b32 s13, s9
+; GFX9-NEXT: s_addc_u32 s9, s35, 0
+; GFX9-NEXT: s_getpc_b64 s[4:5]
+; GFX9-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX9-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
+; GFX9-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
+; GFX9-NEXT: s_mov_b32 s14, s10
+; GFX9-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-NEXT: s_mov_b32 s33, s10
-; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX9-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: s_movk_i32 s32, 0x800
+; GFX9-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX9-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX9-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX9-NEXT: s_mov_b32 s32, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: v_mov_b32_e32 v41, 0
+; GFX9-NEXT: v_mov_b32_e32 v4, 0
; GFX9-NEXT: s_mov_b64 s[0:1], exec
-; GFX9-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX9-NEXT: v_bfrev_b32_e32 v5, 1
; GFX9-NEXT: .LBB10_1: ; %ComputeLoop
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-NEXT: s_ff1_i32_b64 s4, s[0:1]
; GFX9-NEXT: v_readlane_b32 s3, v1, s4
; GFX9-NEXT: v_readlane_b32 s2, v0, s4
-; GFX9-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX9-NEXT: s_lshl_b64 s[2:3], 1, s4
; GFX9-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
; GFX9-NEXT: s_cmp_lg_u64 s[0:1], 0
@@ -6823,99 +6294,64 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_agent
; GFX9-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX9-NEXT: s_cbranch_execz .LBB10_5
; GFX9-NEXT: ; %bb.3:
-; GFX9-NEXT: s_load_dwordx2 s[52:53], s[36:37], 0x24
-; GFX9-NEXT: v_mov_b32_e32 v0, 0
-; GFX9-NEXT: s_mov_b64 s[54:55], 0
+; GFX9-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX9-NEXT: v_mov_b32_e32 v6, 0
+; GFX9-NEXT: s_mov_b64 s[2:3], 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX9-NEXT: global_load_dwordx2 v[2:3], v6, s[0:1]
; GFX9-NEXT: .LBB10_4: ; %atomicrmw.start
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX9-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-NEXT: s_cbranch_execnz .LBB10_4
; GFX9-NEXT: .LBB10_5:
; GFX9-NEXT: s_endpgm
;
; GFX1064-LABEL: global_atomic_fadd_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX1064: ; %bb.0:
-; GFX1064-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-NEXT: s_mov_b32 s66, -1
-; GFX1064-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-NEXT: s_add_u32 s64, s64, s11
+; GFX1064-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1064-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1064-NEXT: s_mov_b32 s38, -1
+; GFX1064-NEXT: s_mov_b32 s39, 0x31e16000
+; GFX1064-NEXT: s_add_u32 s36, s36, s11
; GFX1064-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1064-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-NEXT: s_mov_b32 s51, s8
+; GFX1064-NEXT: s_addc_u32 s37, s37, 0
+; GFX1064-NEXT: s_mov_b32 s12, s8
; GFX1064-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-NEXT: s_mov_b32 s50, s9
+; GFX1064-NEXT: s_mov_b32 s13, s9
; GFX1064-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1064-NEXT: s_getpc_b64 s[4:5]
+; GFX1064-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1064-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1064-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1064-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1064-NEXT: s_mov_b32 s33, s10
-; GFX1064-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1064-NEXT: s_mov_b32 s14, s10
+; GFX1064-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1064-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1064-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1064-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: s_movk_i32 s32, 0x800
+; GFX1064-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1064-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1064-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1064-NEXT: s_mov_b32 s32, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: v_mov_b32_e32 v41, 0
-; GFX1064-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX1064-NEXT: v_mov_b32_e32 v4, 0
+; GFX1064-NEXT: v_bfrev_b32_e32 v5, 1
; GFX1064-NEXT: s_mov_b64 s[0:1], exec
; GFX1064-NEXT: .LBB10_1: ; %ComputeLoop
; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1064-NEXT: s_ff1_i32_b64 s4, s[0:1]
; GFX1064-NEXT: v_readlane_b32 s3, v1, s4
; GFX1064-NEXT: v_readlane_b32 s2, v0, s4
-; GFX1064-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX1064-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX1064-NEXT: s_lshl_b64 s[2:3], 1, s4
; GFX1064-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
; GFX1064-NEXT: s_cmp_lg_u64 s[0:1], 0
@@ -6928,93 +6364,57 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_agent
; GFX1064-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX1064-NEXT: s_cbranch_execz .LBB10_5
; GFX1064-NEXT: ; %bb.3:
-; GFX1064-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1064-NEXT: v_mov_b32_e32 v0, 0
-; GFX1064-NEXT: s_mov_b64 s[54:55], 0
+; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX1064-NEXT: v_mov_b32_e32 v6, 0
+; GFX1064-NEXT: s_mov_b64 s[2:3], 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX1064-NEXT: global_load_dwordx2 v[2:3], v6, s[0:1]
; GFX1064-NEXT: .LBB10_4: ; %atomicrmw.start
; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1064-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1064-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: s_clause 0x1
-; GFX1064-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX1064-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1064-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1064-NEXT: s_waitcnt vmcnt(0)
+; GFX1064-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1064-NEXT: v_mov_b32_e32 v3, v1
+; GFX1064-NEXT: v_mov_b32_e32 v2, v0
+; GFX1064-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1064-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX1064-NEXT: s_cbranch_execnz .LBB10_4
; GFX1064-NEXT: .LBB10_5:
; GFX1064-NEXT: s_endpgm
;
; GFX1032-LABEL: global_atomic_fadd_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX1032: ; %bb.0:
-; GFX1032-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-NEXT: s_mov_b32 s66, -1
-; GFX1032-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-NEXT: s_add_u32 s64, s64, s11
+; GFX1032-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1032-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1032-NEXT: s_mov_b32 s38, -1
+; GFX1032-NEXT: s_mov_b32 s39, 0x31c16000
+; GFX1032-NEXT: s_add_u32 s36, s36, s11
; GFX1032-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1032-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-NEXT: s_mov_b32 s51, s8
+; GFX1032-NEXT: s_addc_u32 s37, s37, 0
+; GFX1032-NEXT: s_mov_b32 s12, s8
; GFX1032-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-NEXT: s_mov_b32 s50, s9
+; GFX1032-NEXT: s_mov_b32 s13, s9
; GFX1032-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1032-NEXT: s_getpc_b64 s[4:5]
+; GFX1032-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1032-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1032-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1032-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1032-NEXT: s_mov_b32 s33, s10
-; GFX1032-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-NEXT: s_mov_b32 s14, s10
+; GFX1032-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1032-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1032-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1032-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1032-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1032-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1032-NEXT: s_mov_b32 s32, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: v_mov_b32_e32 v41, 0
-; GFX1032-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX1032-NEXT: v_mov_b32_e32 v4, 0
+; GFX1032-NEXT: v_bfrev_b32_e32 v5, 1
; GFX1032-NEXT: s_mov_b32 s0, exec_lo
; GFX1032-NEXT: .LBB10_1: ; %ComputeLoop
; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
@@ -7023,61 +6423,32 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_agent
; GFX1032-NEXT: v_readlane_b32 s2, v0, s1
; GFX1032-NEXT: s_lshl_b32 s1, 1, s1
; GFX1032-NEXT: s_andn2_b32 s0, s0, s1
-; GFX1032-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX1032-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX1032-NEXT: s_cmp_lg_u32 s0, 0
; GFX1032-NEXT: s_cbranch_scc1 .LBB10_1
; GFX1032-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1032-NEXT: s_mov_b32 s54, 0
+; GFX1032-NEXT: s_mov_b32 s2, 0
; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX1032-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX1032-NEXT: s_cbranch_execz .LBB10_5
; GFX1032-NEXT: ; %bb.3:
-; GFX1032-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1032-NEXT: v_mov_b32_e32 v0, 0
+; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX1032-NEXT: v_mov_b32_e32 v6, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX1032-NEXT: global_load_dwordx2 v[2:3], v6, s[0:1]
; GFX1032-NEXT: .LBB10_4: ; %atomicrmw.start
; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1032-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1032-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: s_clause 0x1
-; GFX1032-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
+; GFX1032-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1032-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1032-NEXT: s_waitcnt vmcnt(0)
+; GFX1032-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1032-NEXT: v_mov_b32_e32 v3, v1
+; GFX1032-NEXT: v_mov_b32_e32 v2, v0
+; GFX1032-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
; GFX1032-NEXT: s_cbranch_execnz .LBB10_4
; GFX1032-NEXT: .LBB10_5:
; GFX1032-NEXT: s_endpgm
@@ -7085,31 +6456,24 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_agent
; GFX1164-LABEL: global_atomic_fadd_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX1164: ; %bb.0:
; GFX1164-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1164-NEXT: s_mov_b32 s51, s8
+; GFX1164-NEXT: s_mov_b32 s12, s8
; GFX1164-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-NEXT: s_mov_b32 s50, s9
+; GFX1164-NEXT: s_mov_b32 s13, s9
; GFX1164-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1164-NEXT: s_getpc_b64 s[4:5]
+; GFX1164-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1164-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1164-NEXT: v_mov_b32_e32 v31, v0
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: s_mov_b32 s33, s10
-; GFX1164-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX1164-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1164-NEXT: s_mov_b32 s14, s10
+; GFX1164-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1164-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1164-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: s_mov_b32 s32, 32
-; GFX1164-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1164-NEXT: s_mov_b32 s32, 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: v_mov_b32_e32 v41, 0
-; GFX1164-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX1164-NEXT: s_swappc_b64 s[30:31], s[16:17]
+; GFX1164-NEXT: v_mov_b32_e32 v4, 0
+; GFX1164-NEXT: v_bfrev_b32_e32 v5, 1
; GFX1164-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-NEXT: .LBB10_1: ; %ComputeLoop
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
@@ -7118,7 +6482,7 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_agent
; GFX1164-NEXT: v_readlane_b32 s3, v1, s4
; GFX1164-NEXT: v_readlane_b32 s2, v0, s4
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX1164-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX1164-NEXT: s_lshl_b64 s[2:3], 1, s4
; GFX1164-NEXT: s_and_not1_b64 s[0:1], s[0:1], s[2:3]
; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
@@ -7133,81 +6497,48 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_agent
; GFX1164-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX1164-NEXT: s_cbranch_execz .LBB10_5
; GFX1164-NEXT: ; %bb.3:
-; GFX1164-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1164-NEXT: v_mov_b32_e32 v0, 0
-; GFX1164-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1164-NEXT: v_mov_b32_e32 v6, 0
+; GFX1164-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-NEXT: .p2align 6
+; GFX1164-NEXT: global_load_b64 v[2:3], v6, s[0:1]
; GFX1164-NEXT: .LBB10_4: ; %atomicrmw.start
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1164-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1164-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-NEXT: s_cbranch_execnz .LBB10_4
; GFX1164-NEXT: .LBB10_5:
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-NEXT: s_endpgm
;
; GFX1132-LABEL: global_atomic_fadd_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX1132: ; %bb.0:
; GFX1132-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1132-NEXT: s_mov_b64 s[48:49], s[0:1]
+; GFX1132-NEXT: v_mov_b32_e32 v31, v0
; GFX1132-NEXT: s_add_u32 s8, s34, 44
; GFX1132-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX1132-NEXT: v_mov_b32_e32 v31, v0
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1132-NEXT: s_mov_b32 s50, s14
-; GFX1132-NEXT: s_mov_b32 s51, s13
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
+; GFX1132-NEXT: s_getpc_b64 s[4:5]
+; GFX1132-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1132-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1132-NEXT: s_mov_b32 s12, s13
+; GFX1132-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1132-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1132-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX1132-NEXT: s_mov_b32 s13, s14
; GFX1132-NEXT: s_mov_b32 s14, s15
-; GFX1132-NEXT: s_mov_b32 s32, 32
-; GFX1132-NEXT: s_mov_b32 s33, s15
-; GFX1132-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-NEXT: s_mov_b32 s32, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: v_mov_b32_e32 v41, 0
-; GFX1132-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX1132-NEXT: s_swappc_b64 s[30:31], s[16:17]
+; GFX1132-NEXT: v_mov_b32_e32 v4, 0
+; GFX1132-NEXT: v_bfrev_b32_e32 v5, 1
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-NEXT: .LBB10_1: ; %ComputeLoop
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
@@ -7218,522 +6549,358 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_agent
; GFX1132-NEXT: s_lshl_b32 s1, 1, s1
; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX1132-NEXT: s_and_not1_b32 s0, s0, s1
-; GFX1132-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX1132-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX1132-NEXT: s_cmp_lg_u32 s0, 0
; GFX1132-NEXT: s_cbranch_scc1 .LBB10_1
; GFX1132-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-NEXT: s_mov_b32 s54, 0
+; GFX1132-NEXT: s_mov_b32 s2, 0
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX1132-NEXT: s_cbranch_execz .LBB10_5
; GFX1132-NEXT: ; %bb.3:
-; GFX1132-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1132-NEXT: v_mov_b32_e32 v0, 0
+; GFX1132-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1132-NEXT: v_mov_b32_e32 v6, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-NEXT: .p2align 6
+; GFX1132-NEXT: global_load_b64 v[2:3], v6, s[0:1]
; GFX1132-NEXT: .LBB10_4: ; %atomicrmw.start
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1132-NEXT: s_add_u32 s8, s34, 44
-; GFX1132-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1132-NEXT: s_mov_b32 s12, s51
-; GFX1132-NEXT: s_mov_b32 s13, s50
-; GFX1132-NEXT: s_mov_b32 s14, s33
-; GFX1132-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
-; GFX1132-NEXT: s_cbranch_execnz .LBB10_4
-; GFX1132-NEXT: .LBB10_5:
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
+; GFX1132-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1132-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
+; GFX1132-NEXT: s_cbranch_execnz .LBB10_4
+; GFX1132-NEXT: .LBB10_5:
; GFX1132-NEXT: s_endpgm
;
; GFX7LESS-DPP-LABEL: global_atomic_fadd_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX7LESS-DPP: ; %bb.0:
-; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-DPP-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s55, 0xf000
-; GFX7LESS-DPP-NEXT: s_mov_b32 s54, -1
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s32, 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s50, -1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s51, 0xe8f000
+; GFX7LESS-DPP-NEXT: s_add_u32 s48, s48, s11
+; GFX7LESS-DPP-NEXT: s_addc_u32 s49, s49, 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s10
+; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s8
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[36:37], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s39, 0xf000
+; GFX7LESS-DPP-NEXT: s_mov_b32 s38, -1
+; GFX7LESS-DPP-NEXT: s_add_u32 s8, s4, 44
+; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s5, 0
+; GFX7LESS-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX7LESS-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX7LESS-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GFX7LESS-DPP-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v42, v0, v2
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX7LESS-DPP-NEXT: v_or_b32_e32 v31, v0, v2
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v42
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v41, v1
-; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[0:1], off, s[52:55], 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[4:5], off, s[36:39], 0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], 0
; GFX7LESS-DPP-NEXT: .LBB10_1: ; %atomicrmw.start
; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-DPP-NEXT: v_add_f64 v[2:3], v[0:1], v[40:41]
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:8
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(2)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v42
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-DPP-NEXT: v_add_f64 v[2:3], v[4:5], v[0:1]
; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v9, v5
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v8, v4
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v2
+; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[36:39], 0 glc
+; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[4:5]
+; GFX7LESS-DPP-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, v6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, v7
+; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[0:1]
; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB10_1
; GFX7LESS-DPP-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX7LESS-DPP-NEXT: s_endpgm
;
; GFX9-DPP-LABEL: global_atomic_fadd_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX9-DPP: ; %bb.0:
-; GFX9-DPP-NEXT: s_mov_b32 s68, SCRATCH_RSRC_DWORD0
-; GFX9-DPP-NEXT: s_mov_b32 s69, SCRATCH_RSRC_DWORD1
-; GFX9-DPP-NEXT: s_mov_b32 s70, -1
-; GFX9-DPP-NEXT: s_mov_b32 s71, 0xe00000
-; GFX9-DPP-NEXT: s_add_u32 s68, s68, s11
-; GFX9-DPP-NEXT: s_addc_u32 s69, s69, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-DPP-NEXT: s_mov_b32 s51, s8
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_mov_b32 s50, s9
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX9-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX9-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX9-DPP-NEXT: s_mov_b32 s38, -1
+; GFX9-DPP-NEXT: s_mov_b32 s39, 0xe00000
+; GFX9-DPP-NEXT: s_add_u32 s36, s36, s11
+; GFX9-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX9-DPP-NEXT: s_mov_b32 s12, s8
+; GFX9-DPP-NEXT: s_add_u32 s8, s34, 44
+; GFX9-DPP-NEXT: s_mov_b32 s13, s9
+; GFX9-DPP-NEXT: s_addc_u32 s9, s35, 0
+; GFX9-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX9-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX9-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
+; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
+; GFX9-DPP-NEXT: s_mov_b32 s14, s10
+; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-DPP-NEXT: s_mov_b32 s33, s10
-; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[68:69]
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[70:71]
-; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX9-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX9-DPP-NEXT: s_mov_b32 s32, 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX9-DPP-NEXT: v_cndmask_b32_e64 v11, v9, v1, s[0:1]
-; GFX9-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:1 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:1 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v4, 1
+; GFX9-DPP-NEXT: v_cndmask_b32_e64 v6, v4, v1, s[0:1]
+; GFX9-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:1 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:1 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:2 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:2 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:2 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:2 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:4 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:4 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:4 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:4 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:8 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:8 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:8 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:8 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_bcast:15 row_mask:0xa bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_bcast:15 row_mask:0xa bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_bcast:15 row_mask:0xa bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_bcast:15 row_mask:0xa bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
; GFX9-DPP-NEXT: s_nop 1
-; GFX9-DPP-NEXT: v_mov_b32_dpp v9, v11 row_bcast:31 row_mask:0xc bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v10 row_bcast:31 row_mask:0xc bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
+; GFX9-DPP-NEXT: v_mov_b32_dpp v4, v6 row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v3, v5 row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[3:4], v[5:6], v[3:4]
; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 0
; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v1
; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX9-DPP-NEXT: v_readlane_b32 s53, v9, 63
-; GFX9-DPP-NEXT: v_readlane_b32 s52, v8, 63
+; GFX9-DPP-NEXT: v_readlane_b32 s3, v4, 63
+; GFX9-DPP-NEXT: v_readlane_b32 s2, v3, 63
; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[2:3]
; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
+; GFX9-DPP-NEXT: s_and_saveexec_b64 s[2:3], vcc
; GFX9-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX9-DPP-NEXT: ; %bb.1:
-; GFX9-DPP-NEXT: s_load_dwordx2 s[54:55], s[36:37], 0x24
-; GFX9-DPP-NEXT: s_mov_b64 s[64:65], 0
+; GFX9-DPP-NEXT: s_load_dwordx2 s[2:3], s[34:35], 0x24
+; GFX9-DPP-NEXT: s_mov_b64 s[4:5], 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[54:55]
+; GFX9-DPP-NEXT: global_load_dwordx2 v[11:12], v0, s[2:3]
; GFX9-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX9-DPP-NEXT: v_add_f64 v[3:4], v[1:2], s[52:53]
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[68:69]
-; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[68:71], 0 offset:4
-; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[68:71], 0
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[68:71], 0 offset:12
-; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[68:71], 0 offset:8
-; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[70:71]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s54
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s55
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[68:71], 0
-; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[68:71], 0 offset:4
-; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-DPP-NEXT: s_or_b64 s[64:65], vcc, s[64:65]
-; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[64:65]
+; GFX9-DPP-NEXT: v_add_f64 v[9:10], v[11:12], s[0:1]
+; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[1:2], v0, v[9:12], s[2:3] glc
+; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[1:2], v[11:12]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v12, v2
+; GFX9-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v11, v1
+; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX9-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX9-DPP-NEXT: .LBB10_3:
; GFX9-DPP-NEXT: s_endpgm
;
; GFX1064-DPP-LABEL: global_atomic_fadd_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX1064-DPP: ; %bb.0:
-; GFX1064-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1064-DPP-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-DPP-NEXT: s_add_u32 s64, s64, s11
+; GFX1064-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1064-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1064-DPP-NEXT: s_mov_b32 s38, -1
+; GFX1064-DPP-NEXT: s_mov_b32 s39, 0x31e16000
+; GFX1064-DPP-NEXT: s_add_u32 s36, s36, s11
; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1064-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1064-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX1064-DPP-NEXT: s_mov_b32 s12, s8
; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1064-DPP-NEXT: s_mov_b32 s13, s9
; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1064-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1064-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1064-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1064-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX1064-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1064-DPP-NEXT: s_mov_b32 s32, 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1064-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v11, 0x80000000, v1, s[0:1]
-; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
-; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1064-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_readlane_b32 s3, v9, 0
-; GFX1064-DPP-NEXT: v_readlane_b32 s2, v8, 0
-; GFX1064-DPP-NEXT: v_readlane_b32 s5, v9, 32
-; GFX1064-DPP-NEXT: v_readlane_b32 s4, v8, 32
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], s[2:3], s[4:5]
+; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v4, 1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v6, 0x80000000, v1, s[0:1]
+; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[5:6], v[3:4]
+; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
+; GFX1064-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_readlane_b32 s3, v4, 0
+; GFX1064-DPP-NEXT: v_readlane_b32 s2, v3, 0
+; GFX1064-DPP-NEXT: v_readlane_b32 s5, v4, 32
+; GFX1064-DPP-NEXT: v_readlane_b32 s4, v3, 32
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], s[2:3], s[4:5]
; GFX1064-DPP-NEXT: s_mov_b64 exec, s[0:1]
; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v41, v8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v42, v9
-; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v7, exec_hi, v0
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, v3
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, v4
+; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v7
; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX1064-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX1064-DPP-NEXT: ; %bb.1:
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1064-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX1064-DPP-NEXT: global_load_dwordx2 v[9:10], v2, s[0:1]
; GFX1064-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-DPP-NEXT: s_clause 0x1
-; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX1064-DPP-NEXT: v_add_f64 v[7:8], v[9:10], v[0:1]
+; GFX1064-DPP-NEXT: global_atomic_cmpswap_x2 v[7:8], v2, v[7:10], s[0:1] glc
+; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1064-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[7:8], v[9:10]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, v8
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v9, v7
+; GFX1064-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX1064-DPP-NEXT: .LBB10_3:
; GFX1064-DPP-NEXT: s_endpgm
;
; GFX1032-DPP-LABEL: global_atomic_fadd_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX1032-DPP: ; %bb.0:
-; GFX1032-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1032-DPP-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-DPP-NEXT: s_add_u32 s64, s64, s11
+; GFX1032-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1032-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1032-DPP-NEXT: s_mov_b32 s38, -1
+; GFX1032-DPP-NEXT: s_mov_b32 s39, 0x31c16000
+; GFX1032-DPP-NEXT: s_add_u32 s36, s36, s11
; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1032-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1032-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX1032-DPP-NEXT: s_mov_b32 s12, s8
; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1032-DPP-NEXT: s_mov_b32 s13, s9
; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1032-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1032-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1032-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1032-DPP-NEXT: s_mov_b32 s32, 0
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1032-DPP-NEXT: s_or_saveexec_b32 s0, -1
-; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v11, 0x80000000, v1, s0
-; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
-; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1032-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v4, 1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v6, 0x80000000, v1, s0
+; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[5:6], v[3:4]
+; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
+; GFX1032-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
; GFX1032-DPP-NEXT: s_mov_b32 exec_lo, s0
-; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v41, v8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v42, v9
-; GFX1032-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
+; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v7, exec_lo, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, v3
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, v4
+; GFX1032-DPP-NEXT: s_mov_b32 s2, 0
+; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v7
; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX1032-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX1032-DPP-NEXT: ; %bb.1:
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX1032-DPP-NEXT: global_load_dwordx2 v[9:10], v2, s[0:1]
; GFX1032-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-DPP-NEXT: s_clause 0x1
-; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
+; GFX1032-DPP-NEXT: v_add_f64 v[7:8], v[9:10], v[0:1]
+; GFX1032-DPP-NEXT: global_atomic_cmpswap_x2 v[7:8], v2, v[7:10], s[0:1] glc
+; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1032-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[7:8], v[9:10]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, v8
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v9, v7
+; GFX1032-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX1032-DPP-NEXT: .LBB10_3:
; GFX1032-DPP-NEXT: s_endpgm
@@ -7741,238 +6908,175 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_agent
; GFX1164-DPP-LABEL: global_atomic_fadd_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX1164-DPP: ; %bb.0:
; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1164-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1164-DPP-NEXT: s_mov_b32 s12, s8
; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1164-DPP-NEXT: s_mov_b32 s13, s9
; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1164-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1164-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1164-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v0
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX1164-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1164-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1164-DPP-NEXT: s_mov_b32 s32, 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
+; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1164-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v11, 0x80000000, v1, s[0:1]
-; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
+; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v3, 1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v5, 0x80000000, v1, s[0:1]
+; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s[0:1]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
-; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[4:5], v[2:3]
+; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1164-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
+; GFX1164-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1164-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_permlane64_b32 v11, v9
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_permlane64_b32 v5, v3
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_permlane64_b32 v10, v8
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_permlane64_b32 v4, v2
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_mov_b64 exec, s[0:1]
-; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v41, v8
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v42, v9
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-DPP-NEXT: s_waitcnt_depctr 0xfffe
-; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
+; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v6, exec_hi, v0
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, v2
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, v3
+; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
; GFX1164-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX1164-DPP-NEXT: ; %bb.1:
-; GFX1164-DPP-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1164-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-DPP-NEXT: .p2align 6
+; GFX1164-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
; GFX1164-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-DPP-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-DPP-NEXT: v_add_f64 v[6:7], v[8:9], v[0:1]
+; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
+; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[8:9]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v9, v7
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, v6
+; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX1164-DPP-NEXT: .LBB10_3:
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-DPP-NEXT: s_endpgm
;
; GFX1132-DPP-LABEL: global_atomic_fadd_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX1132-DPP: ; %bb.0:
; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1132-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1132-DPP-NEXT: s_mov_b32 s50, s14
-; GFX1132-DPP-NEXT: s_mov_b32 s51, s13
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
+; GFX1132-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1132-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1132-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1132-DPP-NEXT: s_mov_b32 s12, s13
+; GFX1132-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX1132-DPP-NEXT: s_mov_b32 s13, s14
; GFX1132-DPP-NEXT: s_mov_b32 s14, s15
-; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-DPP-NEXT: s_mov_b32 s32, 0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
+; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1132-DPP-NEXT: s_or_saveexec_b32 s0, -1
-; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v11, 0x80000000, v1, s0
-; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
+; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v3, 1
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v5, 0x80000000, v1, s0
+; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
-; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[4:5], v[2:3]
+; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1132-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
+; GFX1132-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1132-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
; GFX1132-DPP-NEXT: s_mov_b32 exec_lo, s0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v41, v8
-; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v0, v2
+; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v6, exec_lo, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v42, v9
-; GFX1132-DPP-NEXT: s_mov_b32 s54, 0
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v10, 0 :: v_dual_mov_b32 v1, v3
+; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
-; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
+; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
; GFX1132-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX1132-DPP-NEXT: ; %bb.1:
-; GFX1132-DPP-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
+; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-DPP-NEXT: .p2align 6
+; GFX1132-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
; GFX1132-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-DPP-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1132-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1132-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-DPP-NEXT: v_add_f64 v[6:7], v[8:9], v[0:1]
+; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
+; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[6:7], v[8:9]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6
+; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX1132-DPP-NEXT: .LBB10_3:
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-DPP-NEXT: s_endpgm
%divValue = call double @div.float.value()
- %result = atomicrmw fadd ptr addrspace(1) %ptr, double %divValue syncscope("agent") monotonic, align 4, !amdgpu.no.fine.grained.memory !1, !amdgpu.ignore.denormal.mode !1
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, double %divValue syncscope("agent") monotonic, align 8, !amdgpu.no.fine.grained.memory !1, !amdgpu.ignore.denormal.mode !1
ret void
}
@@ -11760,998 +10864,555 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_agent
define amdgpu_kernel void @global_atomic_fadd_double_uni_address_uni_value_default_scope_strictfp(ptr addrspace(1) %ptr) #2 {
; GFX7LESS-LABEL: global_atomic_fadd_double_uni_address_uni_value_default_scope_strictfp:
; GFX7LESS: ; %bb.0:
-; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v5, exec_lo, 0
-; GFX7LESS-NEXT: s_bcnt1_i32_b64 s0, exec
-; GFX7LESS-NEXT: s_mov_b32 s1, 0x43300000
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0xc3300000
-; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v5, exec_hi, v5
-; GFX7LESS-NEXT: v_add_f64 v[3:4], s[0:1], v[3:4]
-; GFX7LESS-NEXT: v_mul_f64 v[41:42], 4.0, v[3:4]
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v5
+; GFX7LESS-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GFX7LESS-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GFX7LESS-NEXT: s_mov_b32 s14, -1
+; GFX7LESS-NEXT: s_mov_b32 s15, 0xe8f000
+; GFX7LESS-NEXT: s_add_u32 s12, s12, s11
+; GFX7LESS-NEXT: s_addc_u32 s13, s13, 0
+; GFX7LESS-NEXT: s_mov_b64 s[2:3], exec
+; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s2, 0
+; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s3, v0
+; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX7LESS-NEXT: s_cbranch_execz .LBB16_3
; GFX7LESS-NEXT: ; %bb.1:
-; GFX7LESS-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
+; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
+; GFX7LESS-NEXT: s_bcnt1_i32_b64 s6, s[2:3]
+; GFX7LESS-NEXT: s_mov_b32 s7, 0x43300000
+; GFX7LESS-NEXT: v_mov_b32_e32 v0, 0
+; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0xc3300000
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX7LESS-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX7LESS-NEXT: v_or_b32_e32 v3, v0, v1
+; GFX7LESS-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0x0
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
+; GFX7LESS-NEXT: v_add_f64 v[0:1], s[6:7], v[0:1]
+; GFX7LESS-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, s0
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, s1
-; GFX7LESS-NEXT: v_or_b32_e32 v40, v3, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, s8
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, s9
+; GFX7LESS-NEXT: s_mov_b32 s2, -1
; GFX7LESS-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-NEXT: v_add_f64 v[2:3], v[0:1], v[41:42]
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:12
-; GFX7LESS-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:8
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX7LESS-NEXT: s_waitcnt expcnt(2)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[0:3], 0 glc
+; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-NEXT: s_cbranch_execnz .LBB16_2
; GFX7LESS-NEXT: .LBB16_3:
; GFX7LESS-NEXT: s_endpgm
;
; GFX9-LABEL: global_atomic_fadd_double_uni_address_uni_value_default_scope_strictfp:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-NEXT: s_mov_b32 s66, -1
-; GFX9-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-NEXT: s_add_u32 s64, s64, s11
-; GFX9-NEXT: v_mov_b32_e32 v3, 0
-; GFX9-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-NEXT: s_bcnt1_i32_b64 s0, exec
-; GFX9-NEXT: v_mov_b32_e32 v4, 0xc3300000
-; GFX9-NEXT: s_mov_b32 s1, 0x43300000
-; GFX9-NEXT: v_add_f64 v[3:4], s[0:1], v[3:4]
-; GFX9-NEXT: s_movk_i32 s32, 0x800
-; GFX9-NEXT: v_mul_f64 v[41:42], 4.0, v[3:4]
-; GFX9-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX9-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
+; GFX9-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GFX9-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GFX9-NEXT: s_mov_b32 s14, -1
+; GFX9-NEXT: s_mov_b64 s[0:1], exec
+; GFX9-NEXT: s_mov_b32 s15, 0xe00000
+; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
+; GFX9-NEXT: s_add_u32 s12, s12, s11
+; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, s1, v0
+; GFX9-NEXT: s_addc_u32 s13, s13, 0
+; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX9-NEXT: s_and_saveexec_b64 s[2:3], vcc
; GFX9-NEXT: s_cbranch_execz .LBB16_3
; GFX9-NEXT: ; %bb.1:
-; GFX9-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX9-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX9-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX9-NEXT: s_mov_b32 s33, s10
-; GFX9-NEXT: s_mov_b32 s50, s9
+; GFX9-NEXT: v_mov_b32_e32 v0, 0
+; GFX9-NEXT: s_bcnt1_i32_b64 s0, s[0:1]
+; GFX9-NEXT: v_mov_b32_e32 v1, 0xc3300000
+; GFX9-NEXT: s_mov_b32 s1, 0x43300000
+; GFX9-NEXT: v_add_f64 v[0:1], s[0:1], v[0:1]
+; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9-NEXT: s_mov_b64 s[2:3], 0
+; GFX9-NEXT: v_mov_b32_e32 v6, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX9-NEXT: s_mov_b32 s51, s8
-; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX9-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX9-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v2, s1
-; GFX9-NEXT: s_mov_b64 s[54:55], 0
-; GFX9-NEXT: v_mov_b32_e32 v1, s0
-; GFX9-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX9-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-NEXT: v_mov_b32_e32 v3, s5
; GFX9-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX9-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-NEXT: s_cbranch_execnz .LBB16_2
; GFX9-NEXT: .LBB16_3:
; GFX9-NEXT: s_endpgm
;
; GFX1064-LABEL: global_atomic_fadd_double_uni_address_uni_value_default_scope_strictfp:
; GFX1064: ; %bb.0:
-; GFX1064-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-NEXT: s_mov_b32 s66, -1
-; GFX1064-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-NEXT: s_add_u32 s64, s64, s11
-; GFX1064-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-NEXT: s_bcnt1_i32_b64 s0, exec
-; GFX1064-NEXT: s_mov_b32 s1, 0x43300000
-; GFX1064-NEXT: s_movk_i32 s32, 0x800
-; GFX1064-NEXT: v_add_f64 v[3:4], 0xc3300000, s[0:1]
-; GFX1064-NEXT: v_mul_f64 v[41:42], 4.0, v[3:4]
-; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
+; GFX1064-NEXT: s_mov_b64 s[0:1], exec
+; GFX1064-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
+; GFX1064-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GFX1064-NEXT: s_mov_b32 s14, -1
+; GFX1064-NEXT: s_mov_b32 s15, 0x31e16000
+; GFX1064-NEXT: s_add_u32 s12, s12, s11
+; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, s1, v0
+; GFX1064-NEXT: s_addc_u32 s13, s13, 0
+; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX1064-NEXT: s_and_saveexec_b64 s[2:3], vcc
; GFX1064-NEXT: s_cbranch_execz .LBB16_3
; GFX1064-NEXT: ; %bb.1:
-; GFX1064-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1064-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1064-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1064-NEXT: s_mov_b32 s33, s10
-; GFX1064-NEXT: s_mov_b32 s50, s9
-; GFX1064-NEXT: s_mov_b32 s51, s8
-; GFX1064-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1064-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1064-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1064-NEXT: s_mov_b64 s[54:55], 0
+; GFX1064-NEXT: s_bcnt1_i32_b64 s2, s[0:1]
+; GFX1064-NEXT: s_mov_b32 s3, 0x43300000
+; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1064-NEXT: v_add_f64 v[0:1], 0xc3300000, s[2:3]
+; GFX1064-NEXT: v_mov_b32_e32 v6, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
+; GFX1064-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX1064-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: v_mov_b32_e32 v2, s1
-; GFX1064-NEXT: v_mov_b32_e32 v1, s0
+; GFX1064-NEXT: v_mov_b32_e32 v2, s2
+; GFX1064-NEXT: v_mov_b32_e32 v3, s3
+; GFX1064-NEXT: s_mov_b64 s[2:3], 0
; GFX1064-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1064-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1064-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX1064-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1064-NEXT: s_add_u32 s8, s36, 44
-; GFX1064-NEXT: s_addc_u32 s9, s37, 0
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: s_clause 0x1
-; GFX1064-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX1064-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1064-NEXT: v_mov_b32_e32 v3, v1
+; GFX1064-NEXT: v_mov_b32_e32 v2, v0
+; GFX1064-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1064-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX1064-NEXT: s_cbranch_execnz .LBB16_2
; GFX1064-NEXT: .LBB16_3:
; GFX1064-NEXT: s_endpgm
;
; GFX1032-LABEL: global_atomic_fadd_double_uni_address_uni_value_default_scope_strictfp:
; GFX1032: ; %bb.0:
-; GFX1032-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-NEXT: s_mov_b32 s66, -1
-; GFX1032-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-NEXT: s_add_u32 s64, s64, s11
-; GFX1032-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-NEXT: s_bcnt1_i32_b32 s0, exec_lo
-; GFX1032-NEXT: s_mov_b32 s1, 0x43300000
-; GFX1032-NEXT: s_mov_b32 s54, 0
-; GFX1032-NEXT: v_add_f64 v[3:4], 0xc3300000, s[0:1]
-; GFX1032-NEXT: s_movk_i32 s32, 0x400
-; GFX1032-NEXT: v_mul_f64 v[41:42], 4.0, v[3:4]
-; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
-; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
+; GFX1032-NEXT: s_mov_b32 s0, exec_lo
+; GFX1032-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
+; GFX1032-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GFX1032-NEXT: s_mov_b32 s14, -1
+; GFX1032-NEXT: s_mov_b32 s15, 0x31c16000
+; GFX1032-NEXT: s_add_u32 s12, s12, s11
+; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
+; GFX1032-NEXT: s_addc_u32 s13, s13, 0
+; GFX1032-NEXT: s_mov_b32 s2, 0
+; GFX1032-NEXT: s_and_saveexec_b32 s1, vcc_lo
; GFX1032-NEXT: s_cbranch_execz .LBB16_3
; GFX1032-NEXT: ; %bb.1:
-; GFX1032-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1032-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1032-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1032-NEXT: s_mov_b32 s33, s10
-; GFX1032-NEXT: s_mov_b32 s50, s9
-; GFX1032-NEXT: s_mov_b32 s51, s8
-; GFX1032-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1032-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1032-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-NEXT: s_bcnt1_i32_b32 s6, s0
+; GFX1032-NEXT: s_mov_b32 s7, 0x43300000
+; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1032-NEXT: v_add_f64 v[0:1], 0xc3300000, s[6:7]
+; GFX1032-NEXT: v_mov_b32_e32 v6, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
+; GFX1032-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX1032-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: v_mov_b32_e32 v2, s1
-; GFX1032-NEXT: v_mov_b32_e32 v1, s0
+; GFX1032-NEXT: v_mov_b32_e32 v2, s4
+; GFX1032-NEXT: v_mov_b32_e32 v3, s5
; GFX1032-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1032-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1032-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX1032-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1032-NEXT: s_add_u32 s8, s36, 44
-; GFX1032-NEXT: s_addc_u32 s9, s37, 0
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: s_clause 0x1
-; GFX1032-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
+; GFX1032-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1032-NEXT: v_mov_b32_e32 v3, v1
+; GFX1032-NEXT: v_mov_b32_e32 v2, v0
+; GFX1032-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
; GFX1032-NEXT: s_cbranch_execnz .LBB16_2
; GFX1032-NEXT: .LBB16_3:
; GFX1032-NEXT: s_endpgm
;
; GFX1164-LABEL: global_atomic_fadd_double_uni_address_uni_value_default_scope_strictfp:
; GFX1164: ; %bb.0:
-; GFX1164-NEXT: s_mov_b64 s[48:49], s[0:1]
; GFX1164-NEXT: s_bcnt1_i32_b64 s0, exec
-; GFX1164-NEXT: v_mov_b32_e32 v40, v0
; GFX1164-NEXT: v_mov_b32_e32 v0, 0x43300000
; GFX1164-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-NEXT: s_mov_b32 s32, 32
-; GFX1164-NEXT: s_clause 0x1
-; GFX1164-NEXT: scratch_store_b32 off, v0, off offset:20
-; GFX1164-NEXT: scratch_store_b32 off, v1, off offset:16
-; GFX1164-NEXT: scratch_load_b64 v[0:1], off, off offset:16
+; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
; GFX1164-NEXT: s_mov_b64 s[0:1], exec
-; GFX1164-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-NEXT: v_add_f64 v[0:1], 0xc3300000, v[0:1]
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX1164-NEXT: v_mul_f64 v[41:42], 4.0, v[0:1]
-; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX1164-NEXT: s_clause 0x1
+; GFX1164-NEXT: scratch_store_b32 off, v0, off offset:4
+; GFX1164-NEXT: scratch_store_b32 off, v1, off
+; GFX1164-NEXT: scratch_load_b64 v[0:1], off, off
+; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
+; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v2
; GFX1164-NEXT: s_cbranch_execz .LBB16_3
; GFX1164-NEXT: ; %bb.1:
-; GFX1164-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1164-NEXT: s_mov_b32 s33, s10
-; GFX1164-NEXT: s_mov_b32 s50, s9
-; GFX1164-NEXT: s_mov_b32 s51, s8
-; GFX1164-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1164-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1164-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-NEXT: v_add_f64 v[0:1], 0xc3300000, v[0:1]
+; GFX1164-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1164-NEXT: v_mov_b32_e32 v6, 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1164-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
+; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX1164-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: v_mov_b32_e32 v2, s1
-; GFX1164-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-NEXT: .p2align 6
+; GFX1164-NEXT: v_mov_b32_e32 v2, s2
+; GFX1164-NEXT: v_mov_b32_e32 v3, s3
+; GFX1164-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1164-NEXT: s_waitcnt vmcnt(0)
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1164-NEXT: s_add_u32 s8, s36, 44
-; GFX1164-NEXT: s_addc_u32 s9, s37, 0
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1164-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-NEXT: s_cbranch_execnz .LBB16_2
; GFX1164-NEXT: .LBB16_3:
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-NEXT: s_endpgm
;
; GFX1132-LABEL: global_atomic_fadd_double_uni_address_uni_value_default_scope_strictfp:
; GFX1132: ; %bb.0:
-; GFX1132-NEXT: s_mov_b64 s[48:49], s[0:1]
; GFX1132-NEXT: s_bcnt1_i32_b32 s0, exec_lo
; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX1132-NEXT: v_dual_mov_b32 v40, v0 :: v_dual_mov_b32 v1, s0
-; GFX1132-NEXT: v_mov_b32_e32 v0, 0x43300000
-; GFX1132-NEXT: s_mov_b32 s54, 0
-; GFX1132-NEXT: s_clause 0x1
-; GFX1132-NEXT: scratch_store_b32 off, v0, off offset:20
-; GFX1132-NEXT: scratch_store_b32 off, v1, off offset:16
-; GFX1132-NEXT: scratch_load_b64 v[0:1], off, off offset:16
-; GFX1132-NEXT: s_mov_b32 s32, 32
+; GFX1132-NEXT: v_dual_mov_b32 v0, 0x43300000 :: v_dual_mov_b32 v1, s0
+; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
+; GFX1132-NEXT: s_mov_b32 s2, 0
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
-; GFX1132-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-NEXT: v_add_f64 v[0:1], 0xc3300000, v[0:1]
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX1132-NEXT: v_mul_f64 v[41:42], 4.0, v[0:1]
-; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
+; GFX1132-NEXT: s_clause 0x1
+; GFX1132-NEXT: scratch_store_b32 off, v0, off offset:4
+; GFX1132-NEXT: scratch_store_b32 off, v1, off
+; GFX1132-NEXT: scratch_load_b64 v[0:1], off, off
+; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v2
; GFX1132-NEXT: s_cbranch_execz .LBB16_3
; GFX1132-NEXT: ; %bb.1:
-; GFX1132-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1132-NEXT: s_mov_b32 s33, s15
-; GFX1132-NEXT: s_mov_b32 s50, s14
-; GFX1132-NEXT: s_mov_b32 s51, s13
-; GFX1132-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1132-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-NEXT: v_add_f64 v[0:1], 0xc3300000, v[0:1]
+; GFX1132-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1132-NEXT: v_mov_b32_e32 v6, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1132-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
+; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX1132-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-NEXT: .p2align 6
+; GFX1132-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX1132-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1132-NEXT: s_waitcnt vmcnt(0)
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1132-NEXT: s_add_u32 s8, s36, 44
-; GFX1132-NEXT: s_addc_u32 s9, s37, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1132-NEXT: s_mov_b32 s12, s51
-; GFX1132-NEXT: s_mov_b32 s13, s50
-; GFX1132-NEXT: s_mov_b32 s14, s33
-; GFX1132-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1132-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-NEXT: s_cbranch_execnz .LBB16_2
; GFX1132-NEXT: .LBB16_3:
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-NEXT: s_endpgm
;
; GFX7LESS-DPP-LABEL: global_atomic_fadd_double_uni_address_uni_value_default_scope_strictfp:
; GFX7LESS-DPP: ; %bb.0:
-; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v5, exec_lo, 0
-; GFX7LESS-DPP-NEXT: s_bcnt1_i32_b64 s0, exec
-; GFX7LESS-DPP-NEXT: s_mov_b32 s1, 0x43300000
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0xc3300000
-; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v5, exec_hi, v5
-; GFX7LESS-DPP-NEXT: v_add_f64 v[3:4], s[0:1], v[3:4]
-; GFX7LESS-DPP-NEXT: v_mul_f64 v[41:42], 4.0, v[3:4]
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v5
+; GFX7LESS-DPP-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s14, -1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s15, 0xe8f000
+; GFX7LESS-DPP-NEXT: s_add_u32 s12, s12, s11
+; GFX7LESS-DPP-NEXT: s_addc_u32 s13, s13, 0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], exec
+; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s2, 0
+; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s3, v0
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX7LESS-DPP-NEXT: s_cbranch_execz .LBB16_3
; GFX7LESS-DPP-NEXT: ; %bb.1:
-; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-DPP-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_bcnt1_i32_b64 s6, s[2:3]
+; GFX7LESS-DPP-NEXT: s_mov_b32 s7, 0x43300000
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 0
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0xc3300000
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v3, v0, v1
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s3, 0xf000
+; GFX7LESS-DPP-NEXT: v_add_f64 v[0:1], s[6:7], v[0:1]
+; GFX7LESS-DPP-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, s0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, s1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v40, v3, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s8
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s2, -1
; GFX7LESS-DPP-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-DPP-NEXT: v_add_f64 v[2:3], v[0:1], v[41:42]
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:8
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(2)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-DPP-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[0:3], 0 glc
+; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB16_2
; GFX7LESS-DPP-NEXT: .LBB16_3:
; GFX7LESS-DPP-NEXT: s_endpgm
;
; GFX9-DPP-LABEL: global_atomic_fadd_double_uni_address_uni_value_default_scope_strictfp:
; GFX9-DPP: ; %bb.0:
-; GFX9-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-DPP-NEXT: s_mov_b32 s66, -1
-; GFX9-DPP-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, 0
-; GFX9-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-DPP-NEXT: s_bcnt1_i32_b64 s0, exec
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0xc3300000
-; GFX9-DPP-NEXT: s_mov_b32 s1, 0x43300000
-; GFX9-DPP-NEXT: v_add_f64 v[3:4], s[0:1], v[3:4]
-; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX9-DPP-NEXT: v_mul_f64 v[41:42], 4.0, v[3:4]
-; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
+; GFX9-DPP-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GFX9-DPP-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GFX9-DPP-NEXT: s_mov_b32 s14, -1
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], exec
+; GFX9-DPP-NEXT: s_mov_b32 s15, 0xe00000
+; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
+; GFX9-DPP-NEXT: s_add_u32 s12, s12, s11
+; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, s1, v0
+; GFX9-DPP-NEXT: s_addc_u32 s13, s13, 0
+; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX9-DPP-NEXT: s_and_saveexec_b64 s[2:3], vcc
; GFX9-DPP-NEXT: s_cbranch_execz .LBB16_3
; GFX9-DPP-NEXT: ; %bb.1:
-; GFX9-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX9-DPP-NEXT: s_mov_b32 s33, s10
-; GFX9-DPP-NEXT: s_mov_b32 s50, s9
+; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 0
+; GFX9-DPP-NEXT: s_bcnt1_i32_b64 s0, s[0:1]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0xc3300000
+; GFX9-DPP-NEXT: s_mov_b32 s1, 0x43300000
+; GFX9-DPP-NEXT: v_add_f64 v[0:1], s[0:1], v[0:1]
+; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX9-DPP-NEXT: s_mov_b32 s51, s8
-; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX9-DPP-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX9-DPP-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX9-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s5
; GFX9-DPP-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX9-DPP-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX9-DPP-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-DPP-NEXT: s_cbranch_execnz .LBB16_2
; GFX9-DPP-NEXT: .LBB16_3:
; GFX9-DPP-NEXT: s_endpgm
;
; GFX1064-DPP-LABEL: global_atomic_fadd_double_uni_address_uni_value_default_scope_strictfp:
; GFX1064-DPP: ; %bb.0:
-; GFX1064-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1064-DPP-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX1064-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-DPP-NEXT: s_bcnt1_i32_b64 s0, exec
-; GFX1064-DPP-NEXT: s_mov_b32 s1, 0x43300000
-; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX1064-DPP-NEXT: v_add_f64 v[3:4], 0xc3300000, s[0:1]
-; GFX1064-DPP-NEXT: v_mul_f64 v[41:42], 4.0, v[3:4]
-; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
+; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], exec
+; GFX1064-DPP-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
+; GFX1064-DPP-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GFX1064-DPP-NEXT: s_mov_b32 s14, -1
+; GFX1064-DPP-NEXT: s_mov_b32 s15, 0x31e16000
+; GFX1064-DPP-NEXT: s_add_u32 s12, s12, s11
+; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, s1, v0
+; GFX1064-DPP-NEXT: s_addc_u32 s13, s13, 0
+; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[2:3], vcc
; GFX1064-DPP-NEXT: s_cbranch_execz .LBB16_3
; GFX1064-DPP-NEXT: ; %bb.1:
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1064-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1064-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1064-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1064-DPP-NEXT: s_bcnt1_i32_b64 s2, s[0:1]
+; GFX1064-DPP-NEXT: s_mov_b32 s3, 0x43300000
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1064-DPP-NEXT: v_add_f64 v[0:1], 0xc3300000, s[2:3]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX1064-DPP-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, s0
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s2
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s3
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1064-DPP-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1064-DPP-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1064-DPP-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1064-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1064-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-DPP-NEXT: s_clause 0x1
-; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX1064-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX1064-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB16_2
; GFX1064-DPP-NEXT: .LBB16_3:
; GFX1064-DPP-NEXT: s_endpgm
;
; GFX1032-DPP-LABEL: global_atomic_fadd_double_uni_address_uni_value_default_scope_strictfp:
; GFX1032-DPP: ; %bb.0:
-; GFX1032-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1032-DPP-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX1032-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-DPP-NEXT: s_bcnt1_i32_b32 s0, exec_lo
-; GFX1032-DPP-NEXT: s_mov_b32 s1, 0x43300000
-; GFX1032-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1032-DPP-NEXT: v_add_f64 v[3:4], 0xc3300000, s[0:1]
-; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
-; GFX1032-DPP-NEXT: v_mul_f64 v[41:42], 4.0, v[3:4]
-; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
-; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
+; GFX1032-DPP-NEXT: s_mov_b32 s0, exec_lo
+; GFX1032-DPP-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
+; GFX1032-DPP-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GFX1032-DPP-NEXT: s_mov_b32 s14, -1
+; GFX1032-DPP-NEXT: s_mov_b32 s15, 0x31c16000
+; GFX1032-DPP-NEXT: s_add_u32 s12, s12, s11
+; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
+; GFX1032-DPP-NEXT: s_addc_u32 s13, s13, 0
+; GFX1032-DPP-NEXT: s_mov_b32 s2, 0
+; GFX1032-DPP-NEXT: s_and_saveexec_b32 s1, vcc_lo
; GFX1032-DPP-NEXT: s_cbranch_execz .LBB16_3
; GFX1032-DPP-NEXT: ; %bb.1:
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1032-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1032-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-DPP-NEXT: s_bcnt1_i32_b32 s6, s0
+; GFX1032-DPP-NEXT: s_mov_b32 s7, 0x43300000
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1032-DPP-NEXT: v_add_f64 v[0:1], 0xc3300000, s[6:7]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX1032-DPP-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, s0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s4
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s5
; GFX1032-DPP-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1032-DPP-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1032-DPP-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1032-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1032-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-DPP-NEXT: s_clause 0x1
-; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
+; GFX1032-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX1032-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB16_2
; GFX1032-DPP-NEXT: .LBB16_3:
; GFX1032-DPP-NEXT: s_endpgm
;
; GFX1164-DPP-LABEL: global_atomic_fadd_double_uni_address_uni_value_default_scope_strictfp:
; GFX1164-DPP: ; %bb.0:
-; GFX1164-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
; GFX1164-DPP-NEXT: s_bcnt1_i32_b64 s0, exec
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 0x43300000
; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1164-DPP-NEXT: s_clause 0x1
-; GFX1164-DPP-NEXT: scratch_store_b32 off, v0, off offset:20
-; GFX1164-DPP-NEXT: scratch_store_b32 off, v1, off offset:16
-; GFX1164-DPP-NEXT: scratch_load_b64 v[0:1], off, off offset:16
+; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
-; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-DPP-NEXT: v_add_f64 v[0:1], 0xc3300000, v[0:1]
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_mul_f64 v[41:42], 4.0, v[0:1]
-; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX1164-DPP-NEXT: s_clause 0x1
+; GFX1164-DPP-NEXT: scratch_store_b32 off, v0, off offset:4
+; GFX1164-DPP-NEXT: scratch_store_b32 off, v1, off
+; GFX1164-DPP-NEXT: scratch_load_b64 v[0:1], off, off
+; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
+; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v2
; GFX1164-DPP-NEXT: s_cbranch_execz .LBB16_3
; GFX1164-DPP-NEXT: ; %bb.1:
-; GFX1164-DPP-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1164-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1164-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1164-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-DPP-NEXT: v_add_f64 v[0:1], 0xc3300000, v[0:1]
+; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1164-DPP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX1164-DPP-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-DPP-NEXT: .p2align 6
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s2
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s3
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-DPP-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1164-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1164-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-DPP-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB16_2
; GFX1164-DPP-NEXT: .LBB16_3:
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-DPP-NEXT: s_endpgm
;
; GFX1132-DPP-LABEL: global_atomic_fadd_double_uni_address_uni_value_default_scope_strictfp:
; GFX1132-DPP: ; %bb.0:
-; GFX1132-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
; GFX1132-DPP-NEXT: s_bcnt1_i32_b32 s0, exec_lo
; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v40, v0 :: v_dual_mov_b32 v1, s0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v0, 0x43300000
-; GFX1132-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1132-DPP-NEXT: s_clause 0x1
-; GFX1132-DPP-NEXT: scratch_store_b32 off, v0, off offset:20
-; GFX1132-DPP-NEXT: scratch_store_b32 off, v1, off offset:16
-; GFX1132-DPP-NEXT: scratch_load_b64 v[0:1], off, off offset:16
-; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v0, 0x43300000 :: v_dual_mov_b32 v1, s0
+; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
+; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
-; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-DPP-NEXT: v_add_f64 v[0:1], 0xc3300000, v[0:1]
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_mul_f64 v[41:42], 4.0, v[0:1]
-; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
+; GFX1132-DPP-NEXT: s_clause 0x1
+; GFX1132-DPP-NEXT: scratch_store_b32 off, v0, off offset:4
+; GFX1132-DPP-NEXT: scratch_store_b32 off, v1, off
+; GFX1132-DPP-NEXT: scratch_load_b64 v[0:1], off, off
+; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v2
; GFX1132-DPP-NEXT: s_cbranch_execz .LBB16_3
; GFX1132-DPP-NEXT: ; %bb.1:
-; GFX1132-DPP-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
-; GFX1132-DPP-NEXT: s_mov_b32 s50, s14
-; GFX1132-DPP-NEXT: s_mov_b32 s51, s13
-; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-DPP-NEXT: v_add_f64 v[0:1], 0xc3300000, v[0:1]
+; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v6, 0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1132-DPP-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
+; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX1132-DPP-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-DPP-NEXT: .p2align 6
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX1132-DPP-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1132-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1132-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1132-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1132-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-DPP-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB16_2
; GFX1132-DPP-NEXT: .LBB16_3:
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-DPP-NEXT: s_endpgm
- %result = atomicrmw fadd ptr addrspace(1) %ptr, double 4.0 monotonic, align 4, !amdgpu.no.fine.grained.memory !1, !amdgpu.ignore.denormal.mode !1
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, double 4.0 monotonic, align 8, !amdgpu.no.fine.grained.memory !1, !amdgpu.ignore.denormal.mode !1
ret void
}
define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_default_scope_strictfp(ptr addrspace(1) %ptr) #2 {
; GFX7LESS-LABEL: global_atomic_fadd_double_uni_address_div_value_default_scope_strictfp:
; GFX7LESS: ; %bb.0:
-; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-NEXT: s_mov_b32 s32, 0
+; GFX7LESS-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX7LESS-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX7LESS-NEXT: s_mov_b32 s38, -1
+; GFX7LESS-NEXT: s_mov_b32 s39, 0xe8f000
+; GFX7LESS-NEXT: s_add_u32 s36, s36, s11
+; GFX7LESS-NEXT: s_addc_u32 s37, s37, 0
+; GFX7LESS-NEXT: s_mov_b32 s14, s10
+; GFX7LESS-NEXT: s_mov_b32 s13, s9
+; GFX7LESS-NEXT: s_mov_b32 s12, s8
+; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX7LESS-NEXT: s_add_u32 s8, s34, 44
+; GFX7LESS-NEXT: s_addc_u32 s9, s35, 0
+; GFX7LESS-NEXT: s_getpc_b64 s[4:5]
+; GFX7LESS-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX7LESS-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
+; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GFX7LESS-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX7LESS-NEXT: v_or_b32_e32 v40, v0, v2
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX7LESS-NEXT: v_or_b32_e32 v31, v0, v2
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[38:39]
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX7LESS-NEXT: s_mov_b64 s[0:1], exec
-; GFX7LESS-NEXT: v_mov_b32_e32 v41, 0
-; GFX7LESS-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
+; GFX7LESS-NEXT: v_bfrev_b32_e32 v5, 1
; GFX7LESS-NEXT: .LBB17_1: ; %ComputeLoop
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-NEXT: s_ff1_i32_b64 s4, s[0:1]
@@ -12761,7 +11422,7 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_defau
; GFX7LESS-NEXT: s_andn2_b64 s[0:1], s[0:1], s[4:5]
; GFX7LESS-NEXT: v_cmp_ne_u64_e64 s[4:5], s[0:1], 0
; GFX7LESS-NEXT: s_and_b64 vcc, exec, s[4:5]
-; GFX7LESS-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX7LESS-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX7LESS-NEXT: s_cbranch_vccnz .LBB17_1
; GFX7LESS-NEXT: ; %bb.2: ; %ComputeEnd
; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
@@ -12771,102 +11432,70 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_defau
; GFX7LESS-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX7LESS-NEXT: s_cbranch_execz .LBB17_5
; GFX7LESS-NEXT: ; %bb.3:
-; GFX7LESS-NEXT: s_load_dwordx2 s[52:53], s[36:37], 0x9
-; GFX7LESS-NEXT: s_mov_b32 s55, 0xf000
-; GFX7LESS-NEXT: s_mov_b32 s54, -1
+; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x9
+; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
+; GFX7LESS-NEXT: s_mov_b32 s2, -1
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: buffer_load_dwordx2 v[0:1], off, s[52:55], 0
-; GFX7LESS-NEXT: s_mov_b64 s[54:55], 0
+; GFX7LESS-NEXT: buffer_load_dwordx2 v[2:3], off, s[0:3], 0
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
; GFX7LESS-NEXT: .LBB17_4: ; %atomicrmw.start
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-NEXT: v_add_f64 v[2:3], v[0:1], v[41:42]
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:12
-; GFX7LESS-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:8
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX7LESS-NEXT: s_waitcnt expcnt(2)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[0:3], 0 glc
+; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-NEXT: s_cbranch_execnz .LBB17_4
; GFX7LESS-NEXT: .LBB17_5:
; GFX7LESS-NEXT: s_endpgm
;
; GFX9-LABEL: global_atomic_fadd_double_uni_address_div_value_default_scope_strictfp:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-NEXT: s_mov_b32 s66, -1
-; GFX9-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-NEXT: s_add_u32 s64, s64, s11
-; GFX9-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-NEXT: s_mov_b32 s51, s8
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_mov_b32 s50, s9
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX9-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX9-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX9-NEXT: s_mov_b32 s38, -1
+; GFX9-NEXT: s_mov_b32 s39, 0xe00000
+; GFX9-NEXT: s_add_u32 s36, s36, s11
+; GFX9-NEXT: s_addc_u32 s37, s37, 0
+; GFX9-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX9-NEXT: s_mov_b32 s12, s8
+; GFX9-NEXT: s_add_u32 s8, s34, 44
+; GFX9-NEXT: s_mov_b32 s13, s9
+; GFX9-NEXT: s_addc_u32 s9, s35, 0
+; GFX9-NEXT: s_getpc_b64 s[4:5]
+; GFX9-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX9-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
+; GFX9-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
+; GFX9-NEXT: s_mov_b32 s14, s10
+; GFX9-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-NEXT: s_mov_b32 s33, s10
-; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX9-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: s_movk_i32 s32, 0x800
+; GFX9-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX9-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX9-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX9-NEXT: s_mov_b32 s32, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: v_mov_b32_e32 v41, 0
+; GFX9-NEXT: v_mov_b32_e32 v4, 0
; GFX9-NEXT: s_mov_b64 s[0:1], exec
-; GFX9-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX9-NEXT: v_bfrev_b32_e32 v5, 1
; GFX9-NEXT: .LBB17_1: ; %ComputeLoop
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-NEXT: s_ff1_i32_b64 s4, s[0:1]
; GFX9-NEXT: v_readlane_b32 s3, v1, s4
; GFX9-NEXT: v_readlane_b32 s2, v0, s4
-; GFX9-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX9-NEXT: s_lshl_b64 s[2:3], 1, s4
; GFX9-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
; GFX9-NEXT: s_cmp_lg_u64 s[0:1], 0
@@ -12879,99 +11508,64 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_defau
; GFX9-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX9-NEXT: s_cbranch_execz .LBB17_5
; GFX9-NEXT: ; %bb.3:
-; GFX9-NEXT: s_load_dwordx2 s[52:53], s[36:37], 0x24
-; GFX9-NEXT: v_mov_b32_e32 v0, 0
-; GFX9-NEXT: s_mov_b64 s[54:55], 0
+; GFX9-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX9-NEXT: v_mov_b32_e32 v6, 0
+; GFX9-NEXT: s_mov_b64 s[2:3], 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX9-NEXT: global_load_dwordx2 v[2:3], v6, s[0:1]
; GFX9-NEXT: .LBB17_4: ; %atomicrmw.start
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX9-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-NEXT: s_cbranch_execnz .LBB17_4
; GFX9-NEXT: .LBB17_5:
; GFX9-NEXT: s_endpgm
;
; GFX1064-LABEL: global_atomic_fadd_double_uni_address_div_value_default_scope_strictfp:
; GFX1064: ; %bb.0:
-; GFX1064-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-NEXT: s_mov_b32 s66, -1
-; GFX1064-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-NEXT: s_add_u32 s64, s64, s11
+; GFX1064-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1064-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1064-NEXT: s_mov_b32 s38, -1
+; GFX1064-NEXT: s_mov_b32 s39, 0x31e16000
+; GFX1064-NEXT: s_add_u32 s36, s36, s11
; GFX1064-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1064-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-NEXT: s_mov_b32 s51, s8
+; GFX1064-NEXT: s_addc_u32 s37, s37, 0
+; GFX1064-NEXT: s_mov_b32 s12, s8
; GFX1064-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-NEXT: s_mov_b32 s50, s9
+; GFX1064-NEXT: s_mov_b32 s13, s9
; GFX1064-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1064-NEXT: s_getpc_b64 s[4:5]
+; GFX1064-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1064-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1064-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1064-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1064-NEXT: s_mov_b32 s33, s10
-; GFX1064-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1064-NEXT: s_mov_b32 s14, s10
+; GFX1064-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1064-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1064-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1064-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: s_movk_i32 s32, 0x800
+; GFX1064-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1064-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1064-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1064-NEXT: s_mov_b32 s32, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: v_mov_b32_e32 v41, 0
-; GFX1064-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX1064-NEXT: v_mov_b32_e32 v4, 0
+; GFX1064-NEXT: v_bfrev_b32_e32 v5, 1
; GFX1064-NEXT: s_mov_b64 s[0:1], exec
; GFX1064-NEXT: .LBB17_1: ; %ComputeLoop
; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1064-NEXT: s_ff1_i32_b64 s4, s[0:1]
; GFX1064-NEXT: v_readlane_b32 s3, v1, s4
; GFX1064-NEXT: v_readlane_b32 s2, v0, s4
-; GFX1064-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX1064-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX1064-NEXT: s_lshl_b64 s[2:3], 1, s4
; GFX1064-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
; GFX1064-NEXT: s_cmp_lg_u64 s[0:1], 0
@@ -12984,93 +11578,57 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_defau
; GFX1064-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX1064-NEXT: s_cbranch_execz .LBB17_5
; GFX1064-NEXT: ; %bb.3:
-; GFX1064-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1064-NEXT: v_mov_b32_e32 v0, 0
-; GFX1064-NEXT: s_mov_b64 s[54:55], 0
+; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX1064-NEXT: v_mov_b32_e32 v6, 0
+; GFX1064-NEXT: s_mov_b64 s[2:3], 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX1064-NEXT: global_load_dwordx2 v[2:3], v6, s[0:1]
; GFX1064-NEXT: .LBB17_4: ; %atomicrmw.start
; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1064-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1064-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: s_clause 0x1
-; GFX1064-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX1064-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1064-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1064-NEXT: s_waitcnt vmcnt(0)
+; GFX1064-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1064-NEXT: v_mov_b32_e32 v3, v1
+; GFX1064-NEXT: v_mov_b32_e32 v2, v0
+; GFX1064-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1064-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX1064-NEXT: s_cbranch_execnz .LBB17_4
; GFX1064-NEXT: .LBB17_5:
; GFX1064-NEXT: s_endpgm
;
; GFX1032-LABEL: global_atomic_fadd_double_uni_address_div_value_default_scope_strictfp:
; GFX1032: ; %bb.0:
-; GFX1032-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-NEXT: s_mov_b32 s66, -1
-; GFX1032-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-NEXT: s_add_u32 s64, s64, s11
+; GFX1032-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1032-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1032-NEXT: s_mov_b32 s38, -1
+; GFX1032-NEXT: s_mov_b32 s39, 0x31c16000
+; GFX1032-NEXT: s_add_u32 s36, s36, s11
; GFX1032-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1032-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-NEXT: s_mov_b32 s51, s8
+; GFX1032-NEXT: s_addc_u32 s37, s37, 0
+; GFX1032-NEXT: s_mov_b32 s12, s8
; GFX1032-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-NEXT: s_mov_b32 s50, s9
+; GFX1032-NEXT: s_mov_b32 s13, s9
; GFX1032-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1032-NEXT: s_getpc_b64 s[4:5]
+; GFX1032-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1032-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1032-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1032-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1032-NEXT: s_mov_b32 s33, s10
-; GFX1032-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-NEXT: s_mov_b32 s14, s10
+; GFX1032-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1032-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1032-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1032-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1032-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1032-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1032-NEXT: s_mov_b32 s32, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: v_mov_b32_e32 v41, 0
-; GFX1032-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX1032-NEXT: v_mov_b32_e32 v4, 0
+; GFX1032-NEXT: v_bfrev_b32_e32 v5, 1
; GFX1032-NEXT: s_mov_b32 s0, exec_lo
; GFX1032-NEXT: .LBB17_1: ; %ComputeLoop
; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
@@ -13079,61 +11637,32 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_defau
; GFX1032-NEXT: v_readlane_b32 s2, v0, s1
; GFX1032-NEXT: s_lshl_b32 s1, 1, s1
; GFX1032-NEXT: s_andn2_b32 s0, s0, s1
-; GFX1032-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX1032-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX1032-NEXT: s_cmp_lg_u32 s0, 0
; GFX1032-NEXT: s_cbranch_scc1 .LBB17_1
; GFX1032-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1032-NEXT: s_mov_b32 s54, 0
+; GFX1032-NEXT: s_mov_b32 s2, 0
; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX1032-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX1032-NEXT: s_cbranch_execz .LBB17_5
; GFX1032-NEXT: ; %bb.3:
-; GFX1032-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1032-NEXT: v_mov_b32_e32 v0, 0
+; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX1032-NEXT: v_mov_b32_e32 v6, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX1032-NEXT: global_load_dwordx2 v[2:3], v6, s[0:1]
; GFX1032-NEXT: .LBB17_4: ; %atomicrmw.start
; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1032-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1032-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: s_clause 0x1
-; GFX1032-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
+; GFX1032-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1032-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1032-NEXT: s_waitcnt vmcnt(0)
+; GFX1032-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1032-NEXT: v_mov_b32_e32 v3, v1
+; GFX1032-NEXT: v_mov_b32_e32 v2, v0
+; GFX1032-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
; GFX1032-NEXT: s_cbranch_execnz .LBB17_4
; GFX1032-NEXT: .LBB17_5:
; GFX1032-NEXT: s_endpgm
@@ -13141,31 +11670,24 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_defau
; GFX1164-LABEL: global_atomic_fadd_double_uni_address_div_value_default_scope_strictfp:
; GFX1164: ; %bb.0:
; GFX1164-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1164-NEXT: s_mov_b32 s51, s8
+; GFX1164-NEXT: s_mov_b32 s12, s8
; GFX1164-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-NEXT: s_mov_b32 s50, s9
+; GFX1164-NEXT: s_mov_b32 s13, s9
; GFX1164-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1164-NEXT: s_getpc_b64 s[4:5]
+; GFX1164-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1164-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1164-NEXT: v_mov_b32_e32 v31, v0
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: s_mov_b32 s33, s10
-; GFX1164-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX1164-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1164-NEXT: s_mov_b32 s14, s10
+; GFX1164-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1164-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1164-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: s_mov_b32 s32, 32
-; GFX1164-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1164-NEXT: s_mov_b32 s32, 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: v_mov_b32_e32 v41, 0
-; GFX1164-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX1164-NEXT: s_swappc_b64 s[30:31], s[16:17]
+; GFX1164-NEXT: v_mov_b32_e32 v4, 0
+; GFX1164-NEXT: v_bfrev_b32_e32 v5, 1
; GFX1164-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-NEXT: .LBB17_1: ; %ComputeLoop
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
@@ -13174,7 +11696,7 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_defau
; GFX1164-NEXT: v_readlane_b32 s3, v1, s4
; GFX1164-NEXT: v_readlane_b32 s2, v0, s4
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX1164-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX1164-NEXT: s_lshl_b64 s[2:3], 1, s4
; GFX1164-NEXT: s_and_not1_b64 s[0:1], s[0:1], s[2:3]
; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
@@ -13189,81 +11711,48 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_defau
; GFX1164-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX1164-NEXT: s_cbranch_execz .LBB17_5
; GFX1164-NEXT: ; %bb.3:
-; GFX1164-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1164-NEXT: v_mov_b32_e32 v0, 0
-; GFX1164-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1164-NEXT: v_mov_b32_e32 v6, 0
+; GFX1164-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-NEXT: .p2align 6
+; GFX1164-NEXT: global_load_b64 v[2:3], v6, s[0:1]
; GFX1164-NEXT: .LBB17_4: ; %atomicrmw.start
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1164-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1164-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-NEXT: s_cbranch_execnz .LBB17_4
; GFX1164-NEXT: .LBB17_5:
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-NEXT: s_endpgm
;
; GFX1132-LABEL: global_atomic_fadd_double_uni_address_div_value_default_scope_strictfp:
; GFX1132: ; %bb.0:
; GFX1132-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1132-NEXT: s_mov_b64 s[48:49], s[0:1]
+; GFX1132-NEXT: v_mov_b32_e32 v31, v0
; GFX1132-NEXT: s_add_u32 s8, s34, 44
; GFX1132-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX1132-NEXT: v_mov_b32_e32 v31, v0
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1132-NEXT: s_mov_b32 s50, s14
-; GFX1132-NEXT: s_mov_b32 s51, s13
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
+; GFX1132-NEXT: s_getpc_b64 s[4:5]
+; GFX1132-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1132-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1132-NEXT: s_mov_b32 s12, s13
+; GFX1132-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1132-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1132-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX1132-NEXT: s_mov_b32 s13, s14
; GFX1132-NEXT: s_mov_b32 s14, s15
-; GFX1132-NEXT: s_mov_b32 s32, 32
-; GFX1132-NEXT: s_mov_b32 s33, s15
-; GFX1132-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-NEXT: s_mov_b32 s32, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: v_mov_b32_e32 v41, 0
-; GFX1132-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX1132-NEXT: s_swappc_b64 s[30:31], s[16:17]
+; GFX1132-NEXT: v_mov_b32_e32 v4, 0
+; GFX1132-NEXT: v_bfrev_b32_e32 v5, 1
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-NEXT: .LBB17_1: ; %ComputeLoop
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
@@ -13274,522 +11763,358 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_defau
; GFX1132-NEXT: s_lshl_b32 s1, 1, s1
; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX1132-NEXT: s_and_not1_b32 s0, s0, s1
-; GFX1132-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX1132-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX1132-NEXT: s_cmp_lg_u32 s0, 0
; GFX1132-NEXT: s_cbranch_scc1 .LBB17_1
; GFX1132-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-NEXT: s_mov_b32 s54, 0
+; GFX1132-NEXT: s_mov_b32 s2, 0
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX1132-NEXT: s_cbranch_execz .LBB17_5
; GFX1132-NEXT: ; %bb.3:
-; GFX1132-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1132-NEXT: v_mov_b32_e32 v0, 0
+; GFX1132-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1132-NEXT: v_mov_b32_e32 v6, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-NEXT: .p2align 6
+; GFX1132-NEXT: global_load_b64 v[2:3], v6, s[0:1]
; GFX1132-NEXT: .LBB17_4: ; %atomicrmw.start
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1132-NEXT: s_add_u32 s8, s34, 44
-; GFX1132-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1132-NEXT: s_mov_b32 s12, s51
-; GFX1132-NEXT: s_mov_b32 s13, s50
-; GFX1132-NEXT: s_mov_b32 s14, s33
-; GFX1132-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-NEXT: v_add_f64 v[0:1], v[2:3], v[4:5]
+; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1132-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-NEXT: s_cbranch_execnz .LBB17_4
; GFX1132-NEXT: .LBB17_5:
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-NEXT: s_endpgm
;
; GFX7LESS-DPP-LABEL: global_atomic_fadd_double_uni_address_div_value_default_scope_strictfp:
; GFX7LESS-DPP: ; %bb.0:
-; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-DPP-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s55, 0xf000
-; GFX7LESS-DPP-NEXT: s_mov_b32 s54, -1
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s32, 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s50, -1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s51, 0xe8f000
+; GFX7LESS-DPP-NEXT: s_add_u32 s48, s48, s11
+; GFX7LESS-DPP-NEXT: s_addc_u32 s49, s49, 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s10
+; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s8
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[36:37], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s39, 0xf000
+; GFX7LESS-DPP-NEXT: s_mov_b32 s38, -1
+; GFX7LESS-DPP-NEXT: s_add_u32 s8, s4, 44
+; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s5, 0
+; GFX7LESS-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX7LESS-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX7LESS-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GFX7LESS-DPP-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v42, v0, v2
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX7LESS-DPP-NEXT: v_or_b32_e32 v31, v0, v2
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v42
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v41, v1
-; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[0:1], off, s[52:55], 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[4:5], off, s[36:39], 0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], 0
; GFX7LESS-DPP-NEXT: .LBB17_1: ; %atomicrmw.start
; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-DPP-NEXT: v_add_f64 v[2:3], v[0:1], v[40:41]
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:8
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(2)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v42
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-DPP-NEXT: v_add_f64 v[2:3], v[4:5], v[0:1]
; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v9, v5
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v8, v4
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v2
+; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[36:39], 0 glc
+; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[4:5]
+; GFX7LESS-DPP-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, v6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, v7
+; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[0:1]
; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB17_1
; GFX7LESS-DPP-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX7LESS-DPP-NEXT: s_endpgm
;
; GFX9-DPP-LABEL: global_atomic_fadd_double_uni_address_div_value_default_scope_strictfp:
; GFX9-DPP: ; %bb.0:
-; GFX9-DPP-NEXT: s_mov_b32 s68, SCRATCH_RSRC_DWORD0
-; GFX9-DPP-NEXT: s_mov_b32 s69, SCRATCH_RSRC_DWORD1
-; GFX9-DPP-NEXT: s_mov_b32 s70, -1
-; GFX9-DPP-NEXT: s_mov_b32 s71, 0xe00000
-; GFX9-DPP-NEXT: s_add_u32 s68, s68, s11
-; GFX9-DPP-NEXT: s_addc_u32 s69, s69, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-DPP-NEXT: s_mov_b32 s51, s8
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_mov_b32 s50, s9
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX9-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX9-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX9-DPP-NEXT: s_mov_b32 s38, -1
+; GFX9-DPP-NEXT: s_mov_b32 s39, 0xe00000
+; GFX9-DPP-NEXT: s_add_u32 s36, s36, s11
+; GFX9-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX9-DPP-NEXT: s_mov_b32 s12, s8
+; GFX9-DPP-NEXT: s_add_u32 s8, s34, 44
+; GFX9-DPP-NEXT: s_mov_b32 s13, s9
+; GFX9-DPP-NEXT: s_addc_u32 s9, s35, 0
+; GFX9-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX9-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX9-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
+; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
+; GFX9-DPP-NEXT: s_mov_b32 s14, s10
+; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-DPP-NEXT: s_mov_b32 s33, s10
-; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[68:69]
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[70:71]
-; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX9-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX9-DPP-NEXT: s_mov_b32 s32, 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX9-DPP-NEXT: v_cndmask_b32_e64 v11, v9, v1, s[0:1]
-; GFX9-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:1 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:1 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v4, 1
+; GFX9-DPP-NEXT: v_cndmask_b32_e64 v6, v4, v1, s[0:1]
+; GFX9-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:1 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:1 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:2 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:2 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:2 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:2 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:4 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:4 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:4 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:4 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:8 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:8 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:8 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:8 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_bcast:15 row_mask:0xa bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_bcast:15 row_mask:0xa bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_bcast:15 row_mask:0xa bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_bcast:15 row_mask:0xa bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
; GFX9-DPP-NEXT: s_nop 1
-; GFX9-DPP-NEXT: v_mov_b32_dpp v9, v11 row_bcast:31 row_mask:0xc bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v10 row_bcast:31 row_mask:0xc bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
+; GFX9-DPP-NEXT: v_mov_b32_dpp v4, v6 row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v3, v5 row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[3:4], v[5:6], v[3:4]
; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 0
; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v1
; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX9-DPP-NEXT: v_readlane_b32 s53, v9, 63
-; GFX9-DPP-NEXT: v_readlane_b32 s52, v8, 63
+; GFX9-DPP-NEXT: v_readlane_b32 s3, v4, 63
+; GFX9-DPP-NEXT: v_readlane_b32 s2, v3, 63
; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[2:3]
; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
+; GFX9-DPP-NEXT: s_and_saveexec_b64 s[2:3], vcc
; GFX9-DPP-NEXT: s_cbranch_execz .LBB17_3
; GFX9-DPP-NEXT: ; %bb.1:
-; GFX9-DPP-NEXT: s_load_dwordx2 s[54:55], s[36:37], 0x24
-; GFX9-DPP-NEXT: s_mov_b64 s[64:65], 0
+; GFX9-DPP-NEXT: s_load_dwordx2 s[2:3], s[34:35], 0x24
+; GFX9-DPP-NEXT: s_mov_b64 s[4:5], 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[54:55]
+; GFX9-DPP-NEXT: global_load_dwordx2 v[11:12], v0, s[2:3]
; GFX9-DPP-NEXT: .LBB17_2: ; %atomicrmw.start
; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX9-DPP-NEXT: v_add_f64 v[3:4], v[1:2], s[52:53]
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[68:69]
-; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[68:71], 0 offset:4
-; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[68:71], 0
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[68:71], 0 offset:12
-; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[68:71], 0 offset:8
-; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[70:71]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s54
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s55
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[68:71], 0
-; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[68:71], 0 offset:4
-; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-DPP-NEXT: s_or_b64 s[64:65], vcc, s[64:65]
-; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[64:65]
+; GFX9-DPP-NEXT: v_add_f64 v[9:10], v[11:12], s[0:1]
+; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[1:2], v0, v[9:12], s[2:3] glc
+; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[1:2], v[11:12]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v12, v2
+; GFX9-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v11, v1
+; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX9-DPP-NEXT: s_cbranch_execnz .LBB17_2
; GFX9-DPP-NEXT: .LBB17_3:
; GFX9-DPP-NEXT: s_endpgm
;
; GFX1064-DPP-LABEL: global_atomic_fadd_double_uni_address_div_value_default_scope_strictfp:
; GFX1064-DPP: ; %bb.0:
-; GFX1064-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1064-DPP-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-DPP-NEXT: s_add_u32 s64, s64, s11
+; GFX1064-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1064-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1064-DPP-NEXT: s_mov_b32 s38, -1
+; GFX1064-DPP-NEXT: s_mov_b32 s39, 0x31e16000
+; GFX1064-DPP-NEXT: s_add_u32 s36, s36, s11
; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1064-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1064-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX1064-DPP-NEXT: s_mov_b32 s12, s8
; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1064-DPP-NEXT: s_mov_b32 s13, s9
; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1064-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1064-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1064-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1064-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX1064-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1064-DPP-NEXT: s_mov_b32 s32, 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1064-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v11, 0x80000000, v1, s[0:1]
-; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
-; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1064-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_readlane_b32 s3, v9, 0
-; GFX1064-DPP-NEXT: v_readlane_b32 s2, v8, 0
-; GFX1064-DPP-NEXT: v_readlane_b32 s5, v9, 32
-; GFX1064-DPP-NEXT: v_readlane_b32 s4, v8, 32
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], s[2:3], s[4:5]
+; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v4, 1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v6, 0x80000000, v1, s[0:1]
+; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[5:6], v[3:4]
+; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
+; GFX1064-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_readlane_b32 s3, v4, 0
+; GFX1064-DPP-NEXT: v_readlane_b32 s2, v3, 0
+; GFX1064-DPP-NEXT: v_readlane_b32 s5, v4, 32
+; GFX1064-DPP-NEXT: v_readlane_b32 s4, v3, 32
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], s[2:3], s[4:5]
; GFX1064-DPP-NEXT: s_mov_b64 exec, s[0:1]
; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v41, v8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v42, v9
-; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v7, exec_hi, v0
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, v3
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, v4
+; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v7
; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX1064-DPP-NEXT: s_cbranch_execz .LBB17_3
; GFX1064-DPP-NEXT: ; %bb.1:
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1064-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX1064-DPP-NEXT: global_load_dwordx2 v[9:10], v2, s[0:1]
; GFX1064-DPP-NEXT: .LBB17_2: ; %atomicrmw.start
; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-DPP-NEXT: s_clause 0x1
-; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX1064-DPP-NEXT: v_add_f64 v[7:8], v[9:10], v[0:1]
+; GFX1064-DPP-NEXT: global_atomic_cmpswap_x2 v[7:8], v2, v[7:10], s[0:1] glc
+; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1064-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[7:8], v[9:10]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, v8
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v9, v7
+; GFX1064-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB17_2
; GFX1064-DPP-NEXT: .LBB17_3:
; GFX1064-DPP-NEXT: s_endpgm
;
; GFX1032-DPP-LABEL: global_atomic_fadd_double_uni_address_div_value_default_scope_strictfp:
; GFX1032-DPP: ; %bb.0:
-; GFX1032-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1032-DPP-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-DPP-NEXT: s_add_u32 s64, s64, s11
+; GFX1032-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1032-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1032-DPP-NEXT: s_mov_b32 s38, -1
+; GFX1032-DPP-NEXT: s_mov_b32 s39, 0x31c16000
+; GFX1032-DPP-NEXT: s_add_u32 s36, s36, s11
; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1032-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1032-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX1032-DPP-NEXT: s_mov_b32 s12, s8
; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1032-DPP-NEXT: s_mov_b32 s13, s9
; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1032-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1032-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1032-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1032-DPP-NEXT: s_mov_b32 s32, 0
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1032-DPP-NEXT: s_or_saveexec_b32 s0, -1
-; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v11, 0x80000000, v1, s0
-; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
-; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1032-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v4, 1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v6, 0x80000000, v1, s0
+; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[5:6], v[3:4]
+; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
+; GFX1032-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
; GFX1032-DPP-NEXT: s_mov_b32 exec_lo, s0
-; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v41, v8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v42, v9
-; GFX1032-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
+; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v7, exec_lo, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, v3
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, v4
+; GFX1032-DPP-NEXT: s_mov_b32 s2, 0
+; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v7
; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX1032-DPP-NEXT: s_cbranch_execz .LBB17_3
; GFX1032-DPP-NEXT: ; %bb.1:
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX1032-DPP-NEXT: global_load_dwordx2 v[9:10], v2, s[0:1]
; GFX1032-DPP-NEXT: .LBB17_2: ; %atomicrmw.start
; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-DPP-NEXT: s_clause 0x1
-; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
+; GFX1032-DPP-NEXT: v_add_f64 v[7:8], v[9:10], v[0:1]
+; GFX1032-DPP-NEXT: global_atomic_cmpswap_x2 v[7:8], v2, v[7:10], s[0:1] glc
+; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1032-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[7:8], v[9:10]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, v8
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v9, v7
+; GFX1032-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB17_2
; GFX1032-DPP-NEXT: .LBB17_3:
; GFX1032-DPP-NEXT: s_endpgm
@@ -13797,238 +12122,175 @@ define amdgpu_kernel void @global_atomic_fadd_double_uni_address_div_value_defau
; GFX1164-DPP-LABEL: global_atomic_fadd_double_uni_address_div_value_default_scope_strictfp:
; GFX1164-DPP: ; %bb.0:
; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1164-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1164-DPP-NEXT: s_mov_b32 s12, s8
; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1164-DPP-NEXT: s_mov_b32 s13, s9
; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1164-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1164-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1164-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v0
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX1164-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1164-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1164-DPP-NEXT: s_mov_b32 s32, 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
+; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1164-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v11, 0x80000000, v1, s[0:1]
-; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
+; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v3, 1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v5, 0x80000000, v1, s[0:1]
+; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s[0:1]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
-; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[4:5], v[2:3]
+; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1164-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
+; GFX1164-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1164-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_permlane64_b32 v11, v9
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_permlane64_b32 v5, v3
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_permlane64_b32 v10, v8
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_permlane64_b32 v4, v2
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_mov_b64 exec, s[0:1]
-; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v41, v8
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v42, v9
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-DPP-NEXT: s_waitcnt_depctr 0xfffe
-; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
+; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v6, exec_hi, v0
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, v2
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, v3
+; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
; GFX1164-DPP-NEXT: s_cbranch_execz .LBB17_3
; GFX1164-DPP-NEXT: ; %bb.1:
-; GFX1164-DPP-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1164-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-DPP-NEXT: .p2align 6
+; GFX1164-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
; GFX1164-DPP-NEXT: .LBB17_2: ; %atomicrmw.start
; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-DPP-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-DPP-NEXT: v_add_f64 v[6:7], v[8:9], v[0:1]
+; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
+; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[8:9]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v9, v7
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, v6
+; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB17_2
; GFX1164-DPP-NEXT: .LBB17_3:
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-DPP-NEXT: s_endpgm
;
; GFX1132-DPP-LABEL: global_atomic_fadd_double_uni_address_div_value_default_scope_strictfp:
; GFX1132-DPP: ; %bb.0:
; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1132-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1132-DPP-NEXT: s_mov_b32 s50, s14
-; GFX1132-DPP-NEXT: s_mov_b32 s51, s13
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
+; GFX1132-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1132-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1132-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1132-DPP-NEXT: s_mov_b32 s12, s13
+; GFX1132-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX1132-DPP-NEXT: s_mov_b32 s13, s14
; GFX1132-DPP-NEXT: s_mov_b32 s14, s15
-; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-DPP-NEXT: s_mov_b32 s32, 0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
+; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1132-DPP-NEXT: s_or_saveexec_b32 s0, -1
-; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v11, 0x80000000, v1, s0
-; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
+; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v3, 1
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v5, 0x80000000, v1, s0
+; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
-; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[4:5], v[2:3]
+; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1132-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
+; GFX1132-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1132-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
; GFX1132-DPP-NEXT: s_mov_b32 exec_lo, s0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v41, v8
-; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v0, v2
+; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v6, exec_lo, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v42, v9
-; GFX1132-DPP-NEXT: s_mov_b32 s54, 0
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v10, 0 :: v_dual_mov_b32 v1, v3
+; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
-; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
+; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
; GFX1132-DPP-NEXT: s_cbranch_execz .LBB17_3
; GFX1132-DPP-NEXT: ; %bb.1:
-; GFX1132-DPP-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
+; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-DPP-NEXT: .p2align 6
+; GFX1132-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
; GFX1132-DPP-NEXT: .LBB17_2: ; %atomicrmw.start
; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-DPP-NEXT: v_add_f64 v[3:4], v[1:2], v[41:42]
-; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1132-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1132-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-DPP-NEXT: v_add_f64 v[6:7], v[8:9], v[0:1]
+; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
+; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[6:7], v[8:9]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6
+; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB17_2
; GFX1132-DPP-NEXT: .LBB17_3:
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-DPP-NEXT: s_endpgm
%divValue = call double @div.float.value() strictfp
- %result = atomicrmw fadd ptr addrspace(1) %ptr, double %divValue monotonic, align 4, !amdgpu.no.fine.grained.memory !1, !amdgpu.ignore.denormal.mode !1
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, double %divValue monotonic, align 8, !amdgpu.no.fine.grained.memory !1, !amdgpu.ignore.denormal.mode !1
ret void
}
@@ -14374,7 +12636,7 @@ define amdgpu_kernel void @global_atomic_fadd_uni_address_uni_value_system_scope
; GFX1132-DPP-NEXT: global_atomic_add_f32 v1, v0, s[2:3]
; GFX1132-DPP-NEXT: .LBB18_2:
; GFX1132-DPP-NEXT: s_endpgm
- %result = atomicrmw fadd ptr addrspace(1) %ptr, float 4.0 monotonic, align 4, !amdgpu.no.fine.grained.memory !1, !amdgpu.no.remote.memory !1, !amdgpu.ignore.denormal.mode !1
+ %result = atomicrmw fadd ptr addrspace(1) %ptr, float 4.0 monotonic, align 8, !amdgpu.no.fine.grained.memory !1, !amdgpu.no.remote.memory !1, !amdgpu.ignore.denormal.mode !1
ret void
}
diff --git a/llvm/test/CodeGen/AMDGPU/global_atomics_scan_fmax.ll b/llvm/test/CodeGen/AMDGPU/global_atomics_scan_fmax.ll
index f8f911b693e09..c8b24f71d5735 100644
--- a/llvm/test/CodeGen/AMDGPU/global_atomics_scan_fmax.ll
+++ b/llvm/test/CodeGen/AMDGPU/global_atomics_scan_fmax.ll
@@ -3134,932 +3134,377 @@ define amdgpu_kernel void @global_atomic_fmax_uni_address_div_value_default_scop
define amdgpu_kernel void @global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe(ptr addrspace(1) %ptr) #0 {
; GFX7LESS-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
; GFX7LESS: ; %bb.0:
-; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v3, exec_lo, 0
-; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v3, exec_hi, v3
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
+; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
+; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX7LESS-NEXT: s_cbranch_execz .LBB6_3
; GFX7LESS-NEXT: ; %bb.1:
-; GFX7LESS-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
+; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX7LESS-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX7LESS-NEXT: v_or_b32_e32 v3, v0, v1
+; GFX7LESS-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, s0
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, s1
-; GFX7LESS-NEXT: v_or_b32_e32 v40, v3, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, s6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, s7
+; GFX7LESS-NEXT: s_mov_b32 s2, -1
; GFX7LESS-NEXT: .LBB6_2: ; %atomicrmw.start
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], 4.0
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7LESS-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-NEXT: v_mov_b32_e32 v7, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v6, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v5, v1
+; GFX7LESS-NEXT: v_mov_b32_e32 v4, v0
+; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, v4
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, v5
+; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-NEXT: s_cbranch_execnz .LBB6_2
; GFX7LESS-NEXT: .LBB6_3:
; GFX7LESS-NEXT: s_endpgm
;
; GFX9-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-NEXT: s_mov_b32 s66, -1
-; GFX9-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX9-NEXT: s_add_u32 s64, s64, s11
-; GFX9-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX9-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX9-NEXT: s_movk_i32 s32, 0x800
+; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX9-NEXT: s_cbranch_execz .LBB6_3
; GFX9-NEXT: ; %bb.1:
-; GFX9-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX9-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX9-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX9-NEXT: s_mov_b32 s33, s10
-; GFX9-NEXT: s_mov_b32 s50, s9
+; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9-NEXT: s_mov_b64 s[2:3], 0
+; GFX9-NEXT: v_mov_b32_e32 v4, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX9-NEXT: s_mov_b32 s51, s8
-; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX9-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v2, s1
-; GFX9-NEXT: s_mov_b64 s[54:55], 0
-; GFX9-NEXT: v_mov_b32_e32 v1, s0
-; GFX9-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX9-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-NEXT: v_mov_b32_e32 v3, s5
; GFX9-NEXT: .LBB6_2: ; %atomicrmw.start
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX9-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX9-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-NEXT: s_cbranch_execnz .LBB6_2
; GFX9-NEXT: .LBB6_3:
; GFX9-NEXT: s_endpgm
;
; GFX1064-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1064: ; %bb.0:
-; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1064-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-NEXT: s_mov_b32 s66, -1
-; GFX1064-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX1064-NEXT: s_add_u32 s64, s64, s11
-; GFX1064-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-NEXT: s_movk_i32 s32, 0x800
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
-; GFX1064-NEXT: s_cbranch_execz .LBB6_3
+; GFX1064-NEXT: s_cbranch_execz .LBB6_2
; GFX1064-NEXT: ; %bb.1:
-; GFX1064-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1064-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1064-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1064-NEXT: s_mov_b32 s33, s10
-; GFX1064-NEXT: s_mov_b32 s50, s9
-; GFX1064-NEXT: s_mov_b32 s51, s8
-; GFX1064-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1064-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1064-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1064-NEXT: s_mov_b64 s[54:55], 0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: v_mov_b32_e32 v2, s1
-; GFX1064-NEXT: v_mov_b32_e32 v1, s0
-; GFX1064-NEXT: .LBB6_2: ; %atomicrmw.start
-; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1064-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1064-NEXT: s_add_u32 s8, s36, 44
-; GFX1064-NEXT: s_addc_u32 s9, s37, 0
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX1064-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-NEXT: v_mov_b32_e32 v4, 0
+; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1064-NEXT: v_mov_b32_e32 v0, 0
+; GFX1064-NEXT: v_mov_b32_e32 v1, 0x40100000
+; GFX1064-NEXT: v_mov_b32_e32 v2, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: s_clause 0x1
-; GFX1064-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-NEXT: s_andn2_b64 exec, exec, s[54:55]
-; GFX1064-NEXT: s_cbranch_execnz .LBB6_2
-; GFX1064-NEXT: .LBB6_3:
+; GFX1064-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
+; GFX1064-NEXT: .LBB6_2:
; GFX1064-NEXT: s_endpgm
;
; GFX1032-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1032: ; %bb.0:
-; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1032-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-NEXT: s_mov_b32 s66, -1
-; GFX1032-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
-; GFX1032-NEXT: s_add_u32 s64, s64, s11
-; GFX1032-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-NEXT: s_mov_b32 s54, 0
-; GFX1032-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX1032-NEXT: s_cbranch_execz .LBB6_3
+; GFX1032-NEXT: s_cbranch_execz .LBB6_2
; GFX1032-NEXT: ; %bb.1:
-; GFX1032-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1032-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1032-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1032-NEXT: s_mov_b32 s33, s10
-; GFX1032-NEXT: s_mov_b32 s50, s9
-; GFX1032-NEXT: s_mov_b32 s51, s8
-; GFX1032-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1032-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1032-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: v_mov_b32_e32 v2, s1
-; GFX1032-NEXT: v_mov_b32_e32 v1, s0
-; GFX1032-NEXT: .LBB6_2: ; %atomicrmw.start
-; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1032-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1032-NEXT: s_add_u32 s8, s36, 44
-; GFX1032-NEXT: s_addc_u32 s9, s37, 0
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX1032-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-NEXT: v_mov_b32_e32 v4, 0
+; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1032-NEXT: v_mov_b32_e32 v0, 0
+; GFX1032-NEXT: v_mov_b32_e32 v1, 0x40100000
+; GFX1032-NEXT: v_mov_b32_e32 v2, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: s_clause 0x1
-; GFX1032-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
-; GFX1032-NEXT: s_cbranch_execnz .LBB6_2
-; GFX1032-NEXT: .LBB6_3:
+; GFX1032-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
+; GFX1032-NEXT: .LBB6_2:
; GFX1032-NEXT: s_endpgm
;
; GFX1164-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1164: ; %bb.0:
-; GFX1164-NEXT: v_mov_b32_e32 v40, v0
; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-NEXT: s_mov_b32 s32, 32
; GFX1164-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1164-NEXT: s_cbranch_execz .LBB6_3
; GFX1164-NEXT: ; %bb.1:
-; GFX1164-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1164-NEXT: s_mov_b32 s33, s10
-; GFX1164-NEXT: s_mov_b32 s50, s9
-; GFX1164-NEXT: s_mov_b32 s51, s8
-; GFX1164-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1164-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1164-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1164-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1164-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: v_mov_b32_e32 v2, s1
-; GFX1164-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-NEXT: .p2align 6
+; GFX1164-NEXT: v_mov_b32_e32 v2, s2
+; GFX1164-NEXT: v_mov_b32_e32 v3, s3
+; GFX1164-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-NEXT: .LBB6_2: ; %atomicrmw.start
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1164-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1164-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
; GFX1164-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1164-NEXT: s_add_u32 s8, s36, 44
-; GFX1164-NEXT: s_addc_u32 s9, s37, 0
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX1164-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-NEXT: s_cbranch_execnz .LBB6_2
; GFX1164-NEXT: .LBB6_3:
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-NEXT: s_endpgm
;
; GFX1132-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1132: ; %bb.0:
-; GFX1132-NEXT: v_mov_b32_e32 v40, v0
; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1132-NEXT: s_mov_b32 s54, 0
-; GFX1132-NEXT: s_mov_b32 s32, 32
+; GFX1132-NEXT: s_mov_b32 s2, 0
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
+; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-NEXT: s_cbranch_execz .LBB6_3
; GFX1132-NEXT: ; %bb.1:
-; GFX1132-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1132-NEXT: s_mov_b32 s33, s15
-; GFX1132-NEXT: s_mov_b32 s50, s14
-; GFX1132-NEXT: s_mov_b32 s51, s13
-; GFX1132-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1132-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1132-NEXT: v_mov_b32_e32 v4, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1132-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-NEXT: .p2align 6
+; GFX1132-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX1132-NEXT: .LBB6_2: ; %atomicrmw.start
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1132-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1132-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
; GFX1132-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1132-NEXT: s_add_u32 s8, s36, 44
-; GFX1132-NEXT: s_addc_u32 s9, s37, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1132-NEXT: s_mov_b32 s12, s51
-; GFX1132-NEXT: s_mov_b32 s13, s50
-; GFX1132-NEXT: s_mov_b32 s14, s33
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX1132-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX1132-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-NEXT: s_cbranch_execnz .LBB6_2
; GFX1132-NEXT: .LBB6_3:
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-NEXT: s_endpgm
;
; GFX7LESS-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
; GFX7LESS-DPP: ; %bb.0:
-; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v3, exec_lo, 0
-; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v3, exec_hi, v3
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
+; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX7LESS-DPP-NEXT: s_cbranch_execz .LBB6_3
; GFX7LESS-DPP-NEXT: ; %bb.1:
-; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-DPP-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v3, v0, v1
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s3, 0xf000
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, s0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, s1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v40, v3, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s7
+; GFX7LESS-DPP-NEXT: s_mov_b32 s2, -1
; GFX7LESS-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-DPP-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[2:3], 4.0
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, v1
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, v0
+; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX7LESS-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v4
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v5
+; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB6_2
; GFX7LESS-DPP-NEXT: .LBB6_3:
; GFX7LESS-DPP-NEXT: s_endpgm
;
; GFX9-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
; GFX9-DPP: ; %bb.0:
-; GFX9-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-DPP-NEXT: s_mov_b32 s66, -1
-; GFX9-DPP-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX9-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX9-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX9-DPP-NEXT: s_cbranch_execz .LBB6_3
; GFX9-DPP-NEXT: ; %bb.1:
-; GFX9-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX9-DPP-NEXT: s_mov_b32 s33, s10
-; GFX9-DPP-NEXT: s_mov_b32 s50, s9
+; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX9-DPP-NEXT: s_mov_b32 s51, s8
-; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX9-DPP-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX9-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s5
; GFX9-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX9-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX9-DPP-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-DPP-NEXT: s_cbranch_execnz .LBB6_2
; GFX9-DPP-NEXT: .LBB6_3:
; GFX9-DPP-NEXT: s_endpgm
;
; GFX1064-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1064-DPP: ; %bb.0:
-; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1064-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1064-DPP-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX1064-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX1064-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
-; GFX1064-DPP-NEXT: s_cbranch_execz .LBB6_3
+; GFX1064-DPP-NEXT: s_cbranch_execz .LBB6_2
; GFX1064-DPP-NEXT: ; %bb.1:
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1064-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1064-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1064-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1064-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
-; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1064-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1064-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0x40100000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-DPP-NEXT: s_clause 0x1
-; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
-; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB6_2
-; GFX1064-DPP-NEXT: .LBB6_3:
+; GFX1064-DPP-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
+; GFX1064-DPP-NEXT: .LBB6_2:
; GFX1064-DPP-NEXT: s_endpgm
;
; GFX1032-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1032-DPP: ; %bb.0:
-; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1032-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1032-DPP-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
-; GFX1032-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX1032-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX1032-DPP-NEXT: s_cbranch_execz .LBB6_3
+; GFX1032-DPP-NEXT: s_cbranch_execz .LBB6_2
; GFX1032-DPP-NEXT: ; %bb.1:
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1032-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1032-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1032-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
-; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1032-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1032-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0x40100000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, 0
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-DPP-NEXT: s_clause 0x1
-; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
-; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB6_2
-; GFX1032-DPP-NEXT: .LBB6_3:
+; GFX1032-DPP-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
+; GFX1032-DPP-NEXT: .LBB6_2:
; GFX1032-DPP-NEXT: s_endpgm
;
; GFX1164-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1164-DPP: ; %bb.0:
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1164-DPP-NEXT: s_cbranch_execz .LBB6_3
; GFX1164-DPP-NEXT: ; %bb.1:
-; GFX1164-DPP-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1164-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1164-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1164-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1164-DPP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-DPP-NEXT: .p2align 6
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s2
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s3
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1164-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1164-DPP-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1164-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1164-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB6_2
; GFX1164-DPP-NEXT: .LBB6_3:
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-DPP-NEXT: s_endpgm
;
; GFX1132-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1132-DPP: ; %bb.0:
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1132-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
+; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
+; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-DPP-NEXT: s_cbranch_execz .LBB6_3
; GFX1132-DPP-NEXT: ; %bb.1:
-; GFX1132-DPP-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
-; GFX1132-DPP-NEXT: s_mov_b32 s50, s14
-; GFX1132-DPP-NEXT: s_mov_b32 s51, s13
-; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1132-DPP-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-DPP-NEXT: .p2align 6
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX1132-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1132-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1132-DPP-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1132-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1132-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1132-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1132-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB6_2
; GFX1132-DPP-NEXT: .LBB6_3:
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-DPP-NEXT: s_endpgm
- %result = atomicrmw fmax ptr addrspace(1) %ptr, double 4.0 syncscope("agent") monotonic, align 4, !amdgpu.no.fine.grained.memory !1
+ %result = atomicrmw fmax ptr addrspace(1) %ptr, double 4.0 syncscope("agent") monotonic, align 8, !amdgpu.no.fine.grained.memory !1
ret void
}
define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe(ptr addrspace(1) %ptr) #0 {
; GFX7LESS-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
; GFX7LESS: ; %bb.0:
-; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-NEXT: s_mov_b32 s32, 0
+; GFX7LESS-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX7LESS-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX7LESS-NEXT: s_mov_b32 s38, -1
+; GFX7LESS-NEXT: s_mov_b32 s39, 0xe8f000
+; GFX7LESS-NEXT: s_add_u32 s36, s36, s11
+; GFX7LESS-NEXT: s_addc_u32 s37, s37, 0
+; GFX7LESS-NEXT: s_mov_b32 s14, s10
+; GFX7LESS-NEXT: s_mov_b32 s13, s9
+; GFX7LESS-NEXT: s_mov_b32 s12, s8
+; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX7LESS-NEXT: s_add_u32 s8, s34, 44
+; GFX7LESS-NEXT: s_addc_u32 s9, s35, 0
+; GFX7LESS-NEXT: s_getpc_b64 s[4:5]
+; GFX7LESS-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX7LESS-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
+; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GFX7LESS-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX7LESS-NEXT: v_or_b32_e32 v40, v0, v2
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX7LESS-NEXT: v_or_b32_e32 v31, v0, v2
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[38:39]
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX7LESS-NEXT: s_mov_b64 s[0:1], exec
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
+; GFX7LESS-NEXT: v_mov_b32_e32 v5, 0x7ff80000
; GFX7LESS-NEXT: .LBB7_1: ; %ComputeLoop
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-NEXT: s_ff1_i32_b64 s4, s[0:1]
-; GFX7LESS-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX7LESS-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
; GFX7LESS-NEXT: v_readlane_b32 s3, v1, s4
; GFX7LESS-NEXT: v_readlane_b32 s2, v0, s4
; GFX7LESS-NEXT: s_lshl_b64 s[4:5], 1, s4
@@ -4067,7 +3512,7 @@ define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_agent
; GFX7LESS-NEXT: s_andn2_b64 s[0:1], s[0:1], s[4:5]
; GFX7LESS-NEXT: v_cmp_ne_u64_e64 s[2:3], s[0:1], 0
; GFX7LESS-NEXT: s_and_b64 vcc, exec, s[2:3]
-; GFX7LESS-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
+; GFX7LESS-NEXT: v_max_f64 v[4:5], v[2:3], v[4:5]
; GFX7LESS-NEXT: s_cbranch_vccnz .LBB7_1
; GFX7LESS-NEXT: ; %bb.2: ; %ComputeEnd
; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
@@ -4077,109 +3522,77 @@ define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_agent
; GFX7LESS-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX7LESS-NEXT: s_cbranch_execz .LBB7_5
; GFX7LESS-NEXT: ; %bb.3:
-; GFX7LESS-NEXT: s_load_dwordx2 s[52:53], s[36:37], 0x9
-; GFX7LESS-NEXT: s_mov_b32 s55, 0xf000
-; GFX7LESS-NEXT: s_mov_b32 s54, -1
+; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x9
+; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
+; GFX7LESS-NEXT: s_mov_b32 s2, -1
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: buffer_load_dwordx2 v[0:1], off, s[52:55], 0
-; GFX7LESS-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
+; GFX7LESS-NEXT: buffer_load_dwordx2 v[2:3], off, s[0:3], 0
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
; GFX7LESS-NEXT: .LBB7_4: ; %atomicrmw.start
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], v[41:42]
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7LESS-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[0:3], 0 glc
+; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-NEXT: s_cbranch_execnz .LBB7_4
; GFX7LESS-NEXT: .LBB7_5:
; GFX7LESS-NEXT: s_endpgm
;
; GFX9-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-NEXT: s_mov_b32 s66, -1
-; GFX9-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-NEXT: s_add_u32 s64, s64, s11
-; GFX9-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-NEXT: s_mov_b32 s51, s8
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_mov_b32 s50, s9
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX9-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX9-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX9-NEXT: s_mov_b32 s38, -1
+; GFX9-NEXT: s_mov_b32 s39, 0xe00000
+; GFX9-NEXT: s_add_u32 s36, s36, s11
+; GFX9-NEXT: s_addc_u32 s37, s37, 0
+; GFX9-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX9-NEXT: s_mov_b32 s12, s8
+; GFX9-NEXT: s_add_u32 s8, s34, 44
+; GFX9-NEXT: s_mov_b32 s13, s9
+; GFX9-NEXT: s_addc_u32 s9, s35, 0
+; GFX9-NEXT: s_getpc_b64 s[4:5]
+; GFX9-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX9-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
+; GFX9-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
+; GFX9-NEXT: s_mov_b32 s14, s10
+; GFX9-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-NEXT: s_mov_b32 s33, s10
-; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX9-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: s_movk_i32 s32, 0x800
+; GFX9-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX9-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX9-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX9-NEXT: s_mov_b32 s32, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: v_mov_b32_e32 v2, 0
+; GFX9-NEXT: v_mov_b32_e32 v4, 0
; GFX9-NEXT: s_mov_b64 s[0:1], exec
-; GFX9-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX9-NEXT: v_mov_b32_e32 v5, 0x7ff80000
; GFX9-NEXT: .LBB7_1: ; %ComputeLoop
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-NEXT: s_ff1_i32_b64 s4, s[0:1]
; GFX9-NEXT: v_readlane_b32 s3, v1, s4
; GFX9-NEXT: v_readlane_b32 s2, v0, s4
-; GFX9-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX9-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
; GFX9-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
; GFX9-NEXT: s_lshl_b64 s[2:3], 1, s4
; GFX9-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
; GFX9-NEXT: s_cmp_lg_u64 s[0:1], 0
-; GFX9-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
+; GFX9-NEXT: v_max_f64 v[4:5], v[2:3], v[4:5]
; GFX9-NEXT: s_cbranch_scc1 .LBB7_1
; GFX9-NEXT: ; %bb.2: ; %ComputeEnd
; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
@@ -4189,90 +3602,55 @@ define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_agent
; GFX9-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX9-NEXT: s_cbranch_execz .LBB7_5
; GFX9-NEXT: ; %bb.3:
-; GFX9-NEXT: s_load_dwordx2 s[52:53], s[36:37], 0x24
-; GFX9-NEXT: v_mov_b32_e32 v0, 0
-; GFX9-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
-; GFX9-NEXT: s_mov_b64 s[54:55], 0
+; GFX9-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX9-NEXT: v_mov_b32_e32 v6, 0
+; GFX9-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX9-NEXT: s_mov_b64 s[2:3], 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: global_load_dwordx2 v[4:5], v0, s[52:53]
+; GFX9-NEXT: global_load_dwordx2 v[2:3], v6, s[0:1]
; GFX9-NEXT: .LBB7_4: ; %atomicrmw.start
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: buffer_store_dword v5, off, s[64:67], 0 offset:4
-; GFX9-NEXT: buffer_store_dword v4, off, s[64:67], 0
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
-; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX9-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: buffer_load_dword v4, off, s[64:67], 0
-; GFX9-NEXT: buffer_load_dword v5, off, s[64:67], 0 offset:4
-; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX9-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
+; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-NEXT: s_cbranch_execnz .LBB7_4
; GFX9-NEXT: .LBB7_5:
; GFX9-NEXT: s_endpgm
;
; GFX1064-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
; GFX1064: ; %bb.0:
-; GFX1064-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-NEXT: s_mov_b32 s66, -1
-; GFX1064-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-NEXT: s_add_u32 s64, s64, s11
+; GFX1064-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1064-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1064-NEXT: s_mov_b32 s38, -1
+; GFX1064-NEXT: s_mov_b32 s39, 0x31e16000
+; GFX1064-NEXT: s_add_u32 s36, s36, s11
; GFX1064-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1064-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-NEXT: s_mov_b32 s51, s8
+; GFX1064-NEXT: s_addc_u32 s37, s37, 0
+; GFX1064-NEXT: s_mov_b32 s12, s8
; GFX1064-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-NEXT: s_mov_b32 s50, s9
+; GFX1064-NEXT: s_mov_b32 s13, s9
; GFX1064-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1064-NEXT: s_getpc_b64 s[4:5]
+; GFX1064-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1064-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1064-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1064-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1064-NEXT: s_mov_b32 s33, s10
-; GFX1064-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1064-NEXT: s_mov_b32 s14, s10
+; GFX1064-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1064-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1064-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1064-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: s_movk_i32 s32, 0x800
+; GFX1064-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1064-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1064-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1064-NEXT: s_mov_b32 s32, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1064-NEXT: v_mov_b32_e32 v2, 0
@@ -4296,93 +3674,42 @@ define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_agent
; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX1064-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
-; GFX1064-NEXT: s_cbranch_execz .LBB7_5
+; GFX1064-NEXT: s_cbranch_execz .LBB7_4
; GFX1064-NEXT: ; %bb.3:
-; GFX1064-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
+; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1064-NEXT: v_mov_b32_e32 v0, 0
-; GFX1064-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
-; GFX1064-NEXT: s_mov_b64 s[54:55], 0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: global_load_dwordx2 v[4:5], v0, s[52:53]
-; GFX1064-NEXT: .LBB7_4: ; %atomicrmw.start
-; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1064-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX1064-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-NEXT: buffer_store_dword v5, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: buffer_store_dword v4, off, s[64:67], 0
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
-; GFX1064-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX1064-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX1064-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-NEXT: v_mov_b32_e32 v1, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: s_clause 0x1
-; GFX1064-NEXT: buffer_load_dword v4, off, s[64:67], 0
-; GFX1064-NEXT: buffer_load_dword v5, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-NEXT: s_andn2_b64 exec, exec, s[54:55]
-; GFX1064-NEXT: s_cbranch_execnz .LBB7_4
-; GFX1064-NEXT: .LBB7_5:
+; GFX1064-NEXT: global_atomic_fmax_x2 v0, v[2:3], s[0:1]
+; GFX1064-NEXT: .LBB7_4:
; GFX1064-NEXT: s_endpgm
;
; GFX1032-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
; GFX1032: ; %bb.0:
-; GFX1032-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-NEXT: s_mov_b32 s66, -1
-; GFX1032-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-NEXT: s_add_u32 s64, s64, s11
+; GFX1032-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1032-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1032-NEXT: s_mov_b32 s38, -1
+; GFX1032-NEXT: s_mov_b32 s39, 0x31c16000
+; GFX1032-NEXT: s_add_u32 s36, s36, s11
; GFX1032-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1032-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-NEXT: s_mov_b32 s51, s8
+; GFX1032-NEXT: s_addc_u32 s37, s37, 0
+; GFX1032-NEXT: s_mov_b32 s12, s8
; GFX1032-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-NEXT: s_mov_b32 s50, s9
+; GFX1032-NEXT: s_mov_b32 s13, s9
; GFX1032-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1032-NEXT: s_getpc_b64 s[4:5]
+; GFX1032-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1032-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1032-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1032-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1032-NEXT: s_mov_b32 s33, s10
-; GFX1032-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-NEXT: s_mov_b32 s14, s10
+; GFX1032-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1032-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1032-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1032-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1032-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1032-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1032-NEXT: s_mov_b32 s32, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1032-NEXT: v_mov_b32_e32 v2, 0
@@ -4402,96 +3729,45 @@ define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_agent
; GFX1032-NEXT: s_cbranch_scc1 .LBB7_1
; GFX1032-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1032-NEXT: s_mov_b32 s54, 0
; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX1032-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX1032-NEXT: s_cbranch_execz .LBB7_5
+; GFX1032-NEXT: s_cbranch_execz .LBB7_4
; GFX1032-NEXT: ; %bb.3:
-; GFX1032-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
+; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1032-NEXT: v_mov_b32_e32 v0, 0
-; GFX1032-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: global_load_dwordx2 v[4:5], v0, s[52:53]
-; GFX1032-NEXT: .LBB7_4: ; %atomicrmw.start
-; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1032-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX1032-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-NEXT: buffer_store_dword v5, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: buffer_store_dword v4, off, s[64:67], 0
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
-; GFX1032-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX1032-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX1032-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-NEXT: v_mov_b32_e32 v1, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: s_clause 0x1
-; GFX1032-NEXT: buffer_load_dword v4, off, s[64:67], 0
-; GFX1032-NEXT: buffer_load_dword v5, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
-; GFX1032-NEXT: s_cbranch_execnz .LBB7_4
-; GFX1032-NEXT: .LBB7_5:
+; GFX1032-NEXT: global_atomic_fmax_x2 v0, v[2:3], s[0:1]
+; GFX1032-NEXT: .LBB7_4:
; GFX1032-NEXT: s_endpgm
;
; GFX1164-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
; GFX1164: ; %bb.0:
; GFX1164-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1164-NEXT: s_mov_b32 s51, s8
+; GFX1164-NEXT: s_mov_b32 s12, s8
; GFX1164-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-NEXT: s_mov_b32 s50, s9
+; GFX1164-NEXT: s_mov_b32 s13, s9
; GFX1164-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1164-NEXT: s_getpc_b64 s[4:5]
+; GFX1164-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1164-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1164-NEXT: v_mov_b32_e32 v31, v0
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: s_mov_b32 s33, s10
-; GFX1164-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX1164-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1164-NEXT: s_mov_b32 s14, s10
+; GFX1164-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1164-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1164-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: s_mov_b32 s32, 32
-; GFX1164-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1164-NEXT: s_mov_b32 s32, 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: v_mov_b32_e32 v2, 0
-; GFX1164-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX1164-NEXT: s_swappc_b64 s[30:31], s[16:17]
+; GFX1164-NEXT: v_mov_b32_e32 v4, 0
+; GFX1164-NEXT: v_mov_b32_e32 v5, 0x7ff80000
; GFX1164-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-NEXT: .LBB7_1: ; %ComputeLoop
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1164-NEXT: s_ctz_i32_b64 s4, s[0:1]
-; GFX1164-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX1164-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
; GFX1164-NEXT: v_readlane_b32 s3, v1, s4
; GFX1164-NEXT: v_readlane_b32 s2, v0, s4
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
@@ -4500,7 +3776,7 @@ define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_agent
; GFX1164-NEXT: s_and_not1_b64 s[0:1], s[0:1], s[2:3]
; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1164-NEXT: s_cmp_lg_u64 s[0:1], 0
-; GFX1164-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-NEXT: v_max_f64 v[4:5], v[2:3], v[4:5]
; GFX1164-NEXT: s_cbranch_scc1 .LBB7_1
; GFX1164-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
@@ -4511,89 +3787,57 @@ define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_agent
; GFX1164-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX1164-NEXT: s_cbranch_execz .LBB7_5
; GFX1164-NEXT: ; %bb.3:
-; GFX1164-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1164-NEXT: v_mov_b32_e32 v0, 0
-; GFX1164-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
-; GFX1164-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1164-NEXT: v_mov_b32_e32 v6, 0
+; GFX1164-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: global_load_b64 v[4:5], v0, s[52:53]
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-NEXT: .p2align 6
+; GFX1164-NEXT: global_load_b64 v[2:3], v6, s[0:1]
; GFX1164-NEXT: .LBB7_4: ; %atomicrmw.start
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX1164-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
-; GFX1164-NEXT: scratch_store_b64 off, v[4:5], off
-; GFX1164-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-NEXT: scratch_store_b64 off, v[0:1], off offset:8
-; GFX1164-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: scratch_load_b64 v[4:5], off, off
-; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1164-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
+; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1164-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-NEXT: s_cbranch_execnz .LBB7_4
; GFX1164-NEXT: .LBB7_5:
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-NEXT: s_endpgm
;
; GFX1132-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
; GFX1132: ; %bb.0:
; GFX1132-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1132-NEXT: s_mov_b64 s[48:49], s[0:1]
+; GFX1132-NEXT: v_mov_b32_e32 v31, v0
; GFX1132-NEXT: s_add_u32 s8, s34, 44
; GFX1132-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX1132-NEXT: v_mov_b32_e32 v31, v0
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1132-NEXT: s_mov_b32 s50, s14
-; GFX1132-NEXT: s_mov_b32 s51, s13
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
+; GFX1132-NEXT: s_getpc_b64 s[4:5]
+; GFX1132-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1132-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1132-NEXT: s_mov_b32 s12, s13
+; GFX1132-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1132-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1132-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX1132-NEXT: s_mov_b32 s13, s14
; GFX1132-NEXT: s_mov_b32 s14, s15
-; GFX1132-NEXT: s_mov_b32 s32, 32
-; GFX1132-NEXT: s_mov_b32 s33, s15
-; GFX1132-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-NEXT: s_mov_b32 s32, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: v_mov_b32_e32 v2, 0
-; GFX1132-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX1132-NEXT: s_swappc_b64 s[30:31], s[16:17]
+; GFX1132-NEXT: v_mov_b32_e32 v4, 0
+; GFX1132-NEXT: v_mov_b32_e32 v5, 0x7ff80000
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-NEXT: .LBB7_1: ; %ComputeLoop
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1132-NEXT: s_ctz_i32_b32 s1, s0
-; GFX1132-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX1132-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
; GFX1132-NEXT: v_readlane_b32 s3, v1, s1
; GFX1132-NEXT: v_readlane_b32 s2, v0, s1
; GFX1132-NEXT: s_lshl_b32 s1, 1, s1
@@ -4601,811 +3845,557 @@ define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_agent
; GFX1132-NEXT: s_and_not1_b32 s0, s0, s1
; GFX1132-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
; GFX1132-NEXT: s_cmp_lg_u32 s0, 0
-; GFX1132-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-NEXT: v_max_f64 v[4:5], v[2:3], v[4:5]
; GFX1132-NEXT: s_cbranch_scc1 .LBB7_1
; GFX1132-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-NEXT: s_mov_b32 s54, 0
+; GFX1132-NEXT: s_mov_b32 s2, 0
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX1132-NEXT: s_cbranch_execz .LBB7_5
; GFX1132-NEXT: ; %bb.3:
-; GFX1132-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1132-NEXT: v_mov_b32_e32 v0, 0
+; GFX1132-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1132-NEXT: v_mov_b32_e32 v6, 0
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX1132-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
+; GFX1132-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: global_load_b64 v[4:5], v0, s[52:53]
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-NEXT: .p2align 6
+; GFX1132-NEXT: global_load_b64 v[2:3], v6, s[0:1]
; GFX1132-NEXT: .LBB7_4: ; %atomicrmw.start
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX1132-NEXT: s_add_u32 s8, s34, 44
-; GFX1132-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-NEXT: v_mov_b32_e32 v31, v40
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: v_mov_b32_e32 v3, s53
-; GFX1132-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1132-NEXT: s_mov_b32 s12, s51
-; GFX1132-NEXT: s_mov_b32 s13, s50
-; GFX1132-NEXT: s_mov_b32 s14, s33
-; GFX1132-NEXT: v_mov_b32_e32 v6, 0
-; GFX1132-NEXT: v_mov_b32_e32 v2, s52
-; GFX1132-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
-; GFX1132-NEXT: scratch_store_b64 off, v[4:5], off
-; GFX1132-NEXT: v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v5, 8
-; GFX1132-NEXT: scratch_store_b64 off, v[0:1], off offset:8
-; GFX1132-NEXT: v_dual_mov_b32 v0, 8 :: v_dual_mov_b32 v1, 0
-; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: scratch_load_b64 v[4:5], off, off
-; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1132-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
+; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1132-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-NEXT: s_cbranch_execnz .LBB7_4
; GFX1132-NEXT: .LBB7_5:
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-NEXT: s_endpgm
;
; GFX7LESS-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
; GFX7LESS-DPP: ; %bb.0:
-; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-DPP-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s55, 0xf000
-; GFX7LESS-DPP-NEXT: s_mov_b32 s54, -1
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s32, 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s50, -1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s51, 0xe8f000
+; GFX7LESS-DPP-NEXT: s_add_u32 s48, s48, s11
+; GFX7LESS-DPP-NEXT: s_addc_u32 s49, s49, 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s10
+; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s8
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[36:37], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s39, 0xf000
+; GFX7LESS-DPP-NEXT: s_mov_b32 s38, -1
+; GFX7LESS-DPP-NEXT: s_add_u32 s8, s4, 44
+; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s5, 0
+; GFX7LESS-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX7LESS-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX7LESS-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GFX7LESS-DPP-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v40, v0, v2
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX7LESS-DPP-NEXT: v_or_b32_e32 v31, v0, v2
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[2:3], off, s[52:55], 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-DPP-NEXT: v_max_f64 v[41:42], v[0:1], v[0:1]
+; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[2:3], off, s[36:39], 0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], 0
+; GFX7LESS-DPP-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
; GFX7LESS-DPP-NEXT: .LBB7_1: ; %atomicrmw.start
; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
-; GFX7LESS-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v3, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX7LESS-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[36:39], 0 glc
+; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-DPP-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[0:1]
; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB7_1
; GFX7LESS-DPP-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX7LESS-DPP-NEXT: s_endpgm
;
; GFX9-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
; GFX9-DPP: ; %bb.0:
-; GFX9-DPP-NEXT: s_mov_b32 s68, SCRATCH_RSRC_DWORD0
-; GFX9-DPP-NEXT: s_mov_b32 s69, SCRATCH_RSRC_DWORD1
-; GFX9-DPP-NEXT: s_mov_b32 s70, -1
-; GFX9-DPP-NEXT: s_mov_b32 s71, 0xe00000
-; GFX9-DPP-NEXT: s_add_u32 s68, s68, s11
-; GFX9-DPP-NEXT: s_addc_u32 s69, s69, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-DPP-NEXT: s_mov_b32 s51, s8
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_mov_b32 s50, s9
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX9-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX9-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX9-DPP-NEXT: s_mov_b32 s38, -1
+; GFX9-DPP-NEXT: s_mov_b32 s39, 0xe00000
+; GFX9-DPP-NEXT: s_add_u32 s36, s36, s11
+; GFX9-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX9-DPP-NEXT: s_mov_b32 s12, s8
+; GFX9-DPP-NEXT: s_add_u32 s8, s34, 44
+; GFX9-DPP-NEXT: s_mov_b32 s13, s9
+; GFX9-DPP-NEXT: s_addc_u32 s9, s35, 0
+; GFX9-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX9-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX9-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
+; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
+; GFX9-DPP-NEXT: s_mov_b32 s14, s10
+; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-DPP-NEXT: s_mov_b32 s33, s10
-; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[68:69]
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[70:71]
-; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX9-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX9-DPP-NEXT: s_mov_b32 s32, 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
-; GFX9-DPP-NEXT: v_cndmask_b32_e64 v11, v9, v1, s[0:1]
-; GFX9-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:1 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:1 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
-; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:2 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:2 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
-; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:4 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:4 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
-; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:8 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:8 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
-; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_bcast:15 row_mask:0xa bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_bcast:15 row_mask:0xa bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: s_nop 1
-; GFX9-DPP-NEXT: v_mov_b32_dpp v9, v11 row_bcast:31 row_mask:0xc bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v10 row_bcast:31 row_mask:0xc bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX9-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
-; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0x7ff80000
+; GFX9-DPP-NEXT: v_cndmask_b32_e64 v6, v4, v1, s[0:1]
+; GFX9-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:1 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:1 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
+; GFX9-DPP-NEXT: s_nop 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:2 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:2 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
+; GFX9-DPP-NEXT: s_nop 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:4 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:4 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
+; GFX9-DPP-NEXT: s_nop 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:8 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:8 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
+; GFX9-DPP-NEXT: s_nop 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_bcast:15 row_mask:0xa bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_bcast:15 row_mask:0xa bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: s_nop 1
+; GFX9-DPP-NEXT: v_mov_b32_dpp v4, v6 row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v3, v5 row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[3:4]
+; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
+; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 0
; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v1
; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX9-DPP-NEXT: v_readlane_b32 s53, v9, 63
-; GFX9-DPP-NEXT: v_readlane_b32 s52, v8, 63
+; GFX9-DPP-NEXT: v_readlane_b32 s3, v4, 63
+; GFX9-DPP-NEXT: v_readlane_b32 s2, v3, 63
; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[2:3]
; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
+; GFX9-DPP-NEXT: s_and_saveexec_b64 s[2:3], vcc
; GFX9-DPP-NEXT: s_cbranch_execz .LBB7_3
; GFX9-DPP-NEXT: ; %bb.1:
-; GFX9-DPP-NEXT: s_load_dwordx2 s[54:55], s[36:37], 0x24
-; GFX9-DPP-NEXT: s_mov_b64 s[64:65], 0
+; GFX9-DPP-NEXT: s_load_dwordx2 s[2:3], s[34:35], 0x24
+; GFX9-DPP-NEXT: s_mov_b64 s[4:5], 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[54:55]
+; GFX9-DPP-NEXT: global_load_dwordx2 v[11:12], v0, s[2:3]
; GFX9-DPP-NEXT: .LBB7_2: ; %atomicrmw.start
; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX9-DPP-NEXT: v_max_f64 v[3:4], s[52:53], s[52:53]
+; GFX9-DPP-NEXT: v_max_f64 v[1:2], s[0:1], s[0:1]
; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[1:2], v[1:2]
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[68:69]
-; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[68:71], 0 offset:4
-; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[68:71], 0
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
-; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[68:71], 0 offset:12
-; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[68:71], 0 offset:8
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[70:71]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s54
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s55
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[68:71], 0
-; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[68:71], 0 offset:4
-; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-DPP-NEXT: s_or_b64 s[64:65], vcc, s[64:65]
-; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[64:65]
+; GFX9-DPP-NEXT: v_max_f64 v[9:10], v[11:12], v[11:12]
+; GFX9-DPP-NEXT: v_max_f64 v[9:10], v[9:10], v[1:2]
+; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[1:2], v0, v[9:12], s[2:3] glc
+; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[1:2], v[11:12]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v12, v2
+; GFX9-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v11, v1
+; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX9-DPP-NEXT: s_cbranch_execnz .LBB7_2
; GFX9-DPP-NEXT: .LBB7_3:
; GFX9-DPP-NEXT: s_endpgm
;
; GFX1064-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
; GFX1064-DPP: ; %bb.0:
-; GFX1064-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1064-DPP-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-DPP-NEXT: s_add_u32 s64, s64, s11
+; GFX1064-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1064-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1064-DPP-NEXT: s_mov_b32 s38, -1
+; GFX1064-DPP-NEXT: s_mov_b32 s39, 0x31e16000
+; GFX1064-DPP-NEXT: s_add_u32 s36, s36, s11
; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1064-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1064-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX1064-DPP-NEXT: s_mov_b32 s12, s8
; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1064-DPP-NEXT: s_mov_b32 s13, s9
; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1064-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1064-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1064-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1064-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX1064-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1064-DPP-NEXT: s_mov_b32 s32, 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1064-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s[0:1]
-; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1064-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_readlane_b32 s3, v9, 0
-; GFX1064-DPP-NEXT: v_readlane_b32 s5, v9, 32
-; GFX1064-DPP-NEXT: v_readlane_b32 s4, v8, 32
-; GFX1064-DPP-NEXT: v_readlane_b32 s2, v8, 0
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], s[4:5], s[4:5]
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], s[2:3], s[2:3]
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0x7ff80000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v6, 0x7ff80000, v1, s[0:1]
+; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[3:4]
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
+; GFX1064-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_readlane_b32 s3, v4, 0
+; GFX1064-DPP-NEXT: v_readlane_b32 s5, v4, 32
+; GFX1064-DPP-NEXT: v_readlane_b32 s4, v3, 32
+; GFX1064-DPP-NEXT: v_readlane_b32 s2, v3, 0
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], s[4:5], s[4:5]
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], s[2:3], s[2:3]
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
; GFX1064-DPP-NEXT: s_mov_b64 exec, s[0:1]
; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v41, v8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v42, v9
-; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v7, exec_hi, v0
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, v3
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, v4
+; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v7
; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
-; GFX1064-DPP-NEXT: s_cbranch_execz .LBB7_3
+; GFX1064-DPP-NEXT: s_cbranch_execz .LBB7_2
; GFX1064-DPP-NEXT: ; %bb.1:
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1064-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
-; GFX1064-DPP-NEXT: .LBB7_2: ; %atomicrmw.start
-; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[41:42], v[41:42]
-; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[1:2], v[1:2]
-; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-DPP-NEXT: s_clause 0x1
-; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
-; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB7_2
-; GFX1064-DPP-NEXT: .LBB7_3:
+; GFX1064-DPP-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
+; GFX1064-DPP-NEXT: .LBB7_2:
; GFX1064-DPP-NEXT: s_endpgm
;
; GFX1032-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
; GFX1032-DPP: ; %bb.0:
-; GFX1032-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1032-DPP-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-DPP-NEXT: s_add_u32 s64, s64, s11
+; GFX1032-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1032-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1032-DPP-NEXT: s_mov_b32 s38, -1
+; GFX1032-DPP-NEXT: s_mov_b32 s39, 0x31c16000
+; GFX1032-DPP-NEXT: s_add_u32 s36, s36, s11
; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1032-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1032-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX1032-DPP-NEXT: s_mov_b32 s12, s8
; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1032-DPP-NEXT: s_mov_b32 s13, s9
; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1032-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1032-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1032-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1032-DPP-NEXT: s_mov_b32 s32, 0
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1032-DPP-NEXT: s_or_saveexec_b32 s0, -1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s0
-; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1032-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0x7ff80000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v6, 0x7ff80000, v1, s0
+; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[3:4]
+; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
+; GFX1032-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
; GFX1032-DPP-NEXT: s_mov_b32 exec_lo, s0
-; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, v8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, v9
-; GFX1032-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
+; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v7, exec_lo, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, v3
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, v4
+; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v7
; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX1032-DPP-NEXT: s_cbranch_execz .LBB7_3
+; GFX1032-DPP-NEXT: s_cbranch_execz .LBB7_2
; GFX1032-DPP-NEXT: ; %bb.1:
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1032-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
-; GFX1032-DPP-NEXT: .LBB7_2: ; %atomicrmw.start
-; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[41:42]
-; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-DPP-NEXT: s_clause 0x1
-; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
-; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB7_2
-; GFX1032-DPP-NEXT: .LBB7_3:
+; GFX1032-DPP-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
+; GFX1032-DPP-NEXT: .LBB7_2:
; GFX1032-DPP-NEXT: s_endpgm
;
; GFX1164-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
; GFX1164-DPP: ; %bb.0:
; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1164-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1164-DPP-NEXT: s_mov_b32 s12, s8
; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1164-DPP-NEXT: s_mov_b32 s13, s9
; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1164-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1164-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1164-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v0
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX1164-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1164-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1164-DPP-NEXT: s_mov_b32 s32, 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
+; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1164-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s[0:1]
-; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v5, 0x7ff80000, v1, s[0:1]
+; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s[0:1]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[4:5], v[2:3]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 0x7ff80000
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 0x7ff80000
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 0x7ff80000
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1164-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
+; GFX1164-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1164-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_permlane64_b32 v11, v9
-; GFX1164-DPP-NEXT: v_permlane64_b32 v10, v8
+; GFX1164-DPP-NEXT: v_permlane64_b32 v5, v3
+; GFX1164-DPP-NEXT: v_permlane64_b32 v4, v2
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_mov_b64 exec, s[0:1]
-; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, v8
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, v9
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
-; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
+; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v6, exec_hi, v0
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, v2
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, v3
+; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
; GFX1164-DPP-NEXT: s_cbranch_execz .LBB7_3
; GFX1164-DPP-NEXT: ; %bb.1:
-; GFX1164-DPP-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1164-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
-; GFX1164-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX1164-DPP-NEXT: v_max_f64 v[0:1], v[0:1], v[0:1]
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-DPP-NEXT: .p2align 6
+; GFX1164-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
; GFX1164-DPP-NEXT: .LBB7_2: ; %atomicrmw.start
; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[41:42]
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-DPP-NEXT: v_max_f64 v[6:7], v[8:9], v[8:9]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1164-DPP-NEXT: v_max_f64 v[6:7], v[6:7], v[0:1]
+; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
+; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[8:9]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v9, v7
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, v6
+; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB7_2
; GFX1164-DPP-NEXT: .LBB7_3:
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-DPP-NEXT: s_endpgm
;
; GFX1132-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_agent_scope_unsafe:
; GFX1132-DPP: ; %bb.0:
; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1132-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1132-DPP-NEXT: s_mov_b32 s50, s14
-; GFX1132-DPP-NEXT: s_mov_b32 s51, s13
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
+; GFX1132-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1132-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1132-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1132-DPP-NEXT: s_mov_b32 s12, s13
+; GFX1132-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX1132-DPP-NEXT: s_mov_b32 s13, s14
; GFX1132-DPP-NEXT: s_mov_b32 s14, s15
-; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-DPP-NEXT: s_mov_b32 s32, 0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
+; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1132-DPP-NEXT: s_or_saveexec_b32 s0, -1
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v9, 0x7ff80000 :: v_dual_mov_b32 v8, 0
-; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s0
-; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, 0x7ff80000 :: v_dual_mov_b32 v2, 0
+; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v5, 0x7ff80000, v1, s0
+; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
+; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[4:5], v[2:3]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 0x7ff80000 :: v_dual_mov_b32 v4, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 0x7ff80000 :: v_dual_mov_b32 v4, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 0x7ff80000 :: v_dual_mov_b32 v4, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1132-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
+; GFX1132-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1132-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
; GFX1132-DPP-NEXT: s_mov_b32 exec_lo, s0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v3, v8
-; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, v9
-; GFX1132-DPP-NEXT: s_mov_b32 s54, 0
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v0, v2
+; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v6, exec_lo, 0
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v10, 0 :: v_dual_mov_b32 v1, v3
+; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
+; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
; GFX1132-DPP-NEXT: s_cbranch_execz .LBB7_3
; GFX1132-DPP-NEXT: ; %bb.1:
-; GFX1132-DPP-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1132-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
+; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1132-DPP-NEXT: v_max_f64 v[0:1], v[0:1], v[0:1]
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-DPP-NEXT: .p2align 6
+; GFX1132-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
; GFX1132-DPP-NEXT: .LBB7_2: ; %atomicrmw.start
; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1132-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1132-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[41:42]
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-DPP-NEXT: v_max_f64 v[6:7], v[8:9], v[8:9]
+; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1132-DPP-NEXT: v_max_f64 v[6:7], v[6:7], v[0:1]
+; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
+; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[6:7], v[8:9]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6
+; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB7_2
; GFX1132-DPP-NEXT: .LBB7_3:
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-DPP-NEXT: s_endpgm
%divValue = call double @div.double.value()
- %result = atomicrmw fmax ptr addrspace(1) %ptr, double %divValue syncscope("agent") monotonic, align 4, !amdgpu.no.fine.grained.memory !1
+ %result = atomicrmw fmax ptr addrspace(1) %ptr, double %divValue syncscope("agent") monotonic, align 8, !amdgpu.no.fine.grained.memory !1
ret void
}
@@ -6680,932 +5670,377 @@ define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_one_a
define amdgpu_kernel void @global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe(ptr addrspace(1) %ptr) #0 {
; GFX7LESS-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
; GFX7LESS: ; %bb.0:
-; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v3, exec_lo, 0
-; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v3, exec_hi, v3
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
+; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
+; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX7LESS-NEXT: s_cbranch_execz .LBB10_3
; GFX7LESS-NEXT: ; %bb.1:
-; GFX7LESS-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
+; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX7LESS-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX7LESS-NEXT: v_or_b32_e32 v3, v0, v1
+; GFX7LESS-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, s0
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, s1
-; GFX7LESS-NEXT: v_or_b32_e32 v40, v3, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, s6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, s7
+; GFX7LESS-NEXT: s_mov_b32 s2, -1
; GFX7LESS-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], 4.0
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7LESS-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-NEXT: v_mov_b32_e32 v7, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v6, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v5, v1
+; GFX7LESS-NEXT: v_mov_b32_e32 v4, v0
+; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, v4
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, v5
+; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-NEXT: s_cbranch_execnz .LBB10_2
; GFX7LESS-NEXT: .LBB10_3:
; GFX7LESS-NEXT: s_endpgm
;
; GFX9-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-NEXT: s_mov_b32 s66, -1
-; GFX9-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX9-NEXT: s_add_u32 s64, s64, s11
-; GFX9-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX9-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX9-NEXT: s_movk_i32 s32, 0x800
+; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX9-NEXT: s_cbranch_execz .LBB10_3
; GFX9-NEXT: ; %bb.1:
-; GFX9-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX9-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX9-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX9-NEXT: s_mov_b32 s33, s10
-; GFX9-NEXT: s_mov_b32 s50, s9
+; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9-NEXT: s_mov_b64 s[2:3], 0
+; GFX9-NEXT: v_mov_b32_e32 v4, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX9-NEXT: s_mov_b32 s51, s8
-; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX9-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v2, s1
-; GFX9-NEXT: s_mov_b64 s[54:55], 0
-; GFX9-NEXT: v_mov_b32_e32 v1, s0
-; GFX9-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX9-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-NEXT: v_mov_b32_e32 v3, s5
; GFX9-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX9-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX9-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-NEXT: s_cbranch_execnz .LBB10_2
; GFX9-NEXT: .LBB10_3:
; GFX9-NEXT: s_endpgm
;
; GFX1064-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
; GFX1064: ; %bb.0:
-; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1064-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-NEXT: s_mov_b32 s66, -1
-; GFX1064-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX1064-NEXT: s_add_u32 s64, s64, s11
-; GFX1064-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-NEXT: s_movk_i32 s32, 0x800
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
-; GFX1064-NEXT: s_cbranch_execz .LBB10_3
+; GFX1064-NEXT: s_cbranch_execz .LBB10_2
; GFX1064-NEXT: ; %bb.1:
-; GFX1064-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1064-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1064-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1064-NEXT: s_mov_b32 s33, s10
-; GFX1064-NEXT: s_mov_b32 s50, s9
-; GFX1064-NEXT: s_mov_b32 s51, s8
-; GFX1064-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1064-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1064-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1064-NEXT: s_mov_b64 s[54:55], 0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: v_mov_b32_e32 v2, s1
-; GFX1064-NEXT: v_mov_b32_e32 v1, s0
-; GFX1064-NEXT: .LBB10_2: ; %atomicrmw.start
-; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1064-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1064-NEXT: s_add_u32 s8, s36, 44
-; GFX1064-NEXT: s_addc_u32 s9, s37, 0
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX1064-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-NEXT: v_mov_b32_e32 v4, 0
+; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1064-NEXT: v_mov_b32_e32 v0, 0
+; GFX1064-NEXT: v_mov_b32_e32 v1, 0x40100000
+; GFX1064-NEXT: v_mov_b32_e32 v2, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: s_clause 0x1
-; GFX1064-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-NEXT: s_andn2_b64 exec, exec, s[54:55]
-; GFX1064-NEXT: s_cbranch_execnz .LBB10_2
-; GFX1064-NEXT: .LBB10_3:
+; GFX1064-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
+; GFX1064-NEXT: .LBB10_2:
; GFX1064-NEXT: s_endpgm
;
; GFX1032-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
; GFX1032: ; %bb.0:
-; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1032-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-NEXT: s_mov_b32 s66, -1
-; GFX1032-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
-; GFX1032-NEXT: s_add_u32 s64, s64, s11
-; GFX1032-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-NEXT: s_mov_b32 s54, 0
-; GFX1032-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX1032-NEXT: s_cbranch_execz .LBB10_3
+; GFX1032-NEXT: s_cbranch_execz .LBB10_2
; GFX1032-NEXT: ; %bb.1:
-; GFX1032-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1032-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1032-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1032-NEXT: s_mov_b32 s33, s10
-; GFX1032-NEXT: s_mov_b32 s50, s9
-; GFX1032-NEXT: s_mov_b32 s51, s8
-; GFX1032-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1032-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1032-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: v_mov_b32_e32 v2, s1
-; GFX1032-NEXT: v_mov_b32_e32 v1, s0
-; GFX1032-NEXT: .LBB10_2: ; %atomicrmw.start
-; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1032-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1032-NEXT: s_add_u32 s8, s36, 44
-; GFX1032-NEXT: s_addc_u32 s9, s37, 0
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX1032-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-NEXT: v_mov_b32_e32 v4, 0
+; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1032-NEXT: v_mov_b32_e32 v0, 0
+; GFX1032-NEXT: v_mov_b32_e32 v1, 0x40100000
+; GFX1032-NEXT: v_mov_b32_e32 v2, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: s_clause 0x1
-; GFX1032-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
-; GFX1032-NEXT: s_cbranch_execnz .LBB10_2
-; GFX1032-NEXT: .LBB10_3:
+; GFX1032-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
+; GFX1032-NEXT: .LBB10_2:
; GFX1032-NEXT: s_endpgm
;
; GFX1164-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
; GFX1164: ; %bb.0:
-; GFX1164-NEXT: v_mov_b32_e32 v40, v0
; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-NEXT: s_mov_b32 s32, 32
; GFX1164-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1164-NEXT: s_cbranch_execz .LBB10_3
; GFX1164-NEXT: ; %bb.1:
-; GFX1164-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1164-NEXT: s_mov_b32 s33, s10
-; GFX1164-NEXT: s_mov_b32 s50, s9
-; GFX1164-NEXT: s_mov_b32 s51, s8
-; GFX1164-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1164-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1164-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1164-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1164-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: v_mov_b32_e32 v2, s1
-; GFX1164-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-NEXT: .p2align 6
+; GFX1164-NEXT: v_mov_b32_e32 v2, s2
+; GFX1164-NEXT: v_mov_b32_e32 v3, s3
+; GFX1164-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1164-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1164-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
; GFX1164-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1164-NEXT: s_add_u32 s8, s36, 44
-; GFX1164-NEXT: s_addc_u32 s9, s37, 0
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX1164-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-NEXT: s_cbranch_execnz .LBB10_2
; GFX1164-NEXT: .LBB10_3:
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-NEXT: s_endpgm
;
; GFX1132-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
; GFX1132: ; %bb.0:
-; GFX1132-NEXT: v_mov_b32_e32 v40, v0
; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1132-NEXT: s_mov_b32 s54, 0
-; GFX1132-NEXT: s_mov_b32 s32, 32
+; GFX1132-NEXT: s_mov_b32 s2, 0
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
+; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-NEXT: s_cbranch_execz .LBB10_3
; GFX1132-NEXT: ; %bb.1:
-; GFX1132-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1132-NEXT: s_mov_b32 s33, s15
-; GFX1132-NEXT: s_mov_b32 s50, s14
-; GFX1132-NEXT: s_mov_b32 s51, s13
-; GFX1132-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1132-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1132-NEXT: v_mov_b32_e32 v4, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1132-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-NEXT: .p2align 6
+; GFX1132-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX1132-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1132-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1132-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
; GFX1132-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1132-NEXT: s_add_u32 s8, s36, 44
-; GFX1132-NEXT: s_addc_u32 s9, s37, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1132-NEXT: s_mov_b32 s12, s51
-; GFX1132-NEXT: s_mov_b32 s13, s50
-; GFX1132-NEXT: s_mov_b32 s14, s33
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX1132-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX1132-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-NEXT: s_cbranch_execnz .LBB10_2
; GFX1132-NEXT: .LBB10_3:
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-NEXT: s_endpgm
;
; GFX7LESS-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
; GFX7LESS-DPP: ; %bb.0:
-; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v3, exec_lo, 0
-; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v3, exec_hi, v3
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
+; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX7LESS-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX7LESS-DPP-NEXT: ; %bb.1:
-; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-DPP-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v3, v0, v1
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s3, 0xf000
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, s0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, s1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v40, v3, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s7
+; GFX7LESS-DPP-NEXT: s_mov_b32 s2, -1
; GFX7LESS-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-DPP-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[2:3], 4.0
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, v1
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, v0
+; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX7LESS-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v4
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v5
+; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX7LESS-DPP-NEXT: .LBB10_3:
; GFX7LESS-DPP-NEXT: s_endpgm
;
; GFX9-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
; GFX9-DPP: ; %bb.0:
-; GFX9-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-DPP-NEXT: s_mov_b32 s66, -1
-; GFX9-DPP-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX9-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX9-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX9-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX9-DPP-NEXT: ; %bb.1:
-; GFX9-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX9-DPP-NEXT: s_mov_b32 s33, s10
-; GFX9-DPP-NEXT: s_mov_b32 s50, s9
+; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX9-DPP-NEXT: s_mov_b32 s51, s8
-; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX9-DPP-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX9-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s5
; GFX9-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX9-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX9-DPP-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX9-DPP-NEXT: .LBB10_3:
; GFX9-DPP-NEXT: s_endpgm
;
; GFX1064-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
; GFX1064-DPP: ; %bb.0:
-; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1064-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1064-DPP-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX1064-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX1064-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
-; GFX1064-DPP-NEXT: s_cbranch_execz .LBB10_3
+; GFX1064-DPP-NEXT: s_cbranch_execz .LBB10_2
; GFX1064-DPP-NEXT: ; %bb.1:
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1064-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1064-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1064-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1064-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
-; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1064-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1064-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0x40100000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-DPP-NEXT: s_clause 0x1
-; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
-; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB10_2
-; GFX1064-DPP-NEXT: .LBB10_3:
+; GFX1064-DPP-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
+; GFX1064-DPP-NEXT: .LBB10_2:
; GFX1064-DPP-NEXT: s_endpgm
;
; GFX1032-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
; GFX1032-DPP: ; %bb.0:
-; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1032-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1032-DPP-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
-; GFX1032-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX1032-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX1032-DPP-NEXT: s_cbranch_execz .LBB10_3
+; GFX1032-DPP-NEXT: s_cbranch_execz .LBB10_2
; GFX1032-DPP-NEXT: ; %bb.1:
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1032-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1032-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1032-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
-; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1032-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1032-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0x40100000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, 0
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-DPP-NEXT: s_clause 0x1
-; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
-; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB10_2
-; GFX1032-DPP-NEXT: .LBB10_3:
+; GFX1032-DPP-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
+; GFX1032-DPP-NEXT: .LBB10_2:
; GFX1032-DPP-NEXT: s_endpgm
;
; GFX1164-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
; GFX1164-DPP: ; %bb.0:
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1164-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX1164-DPP-NEXT: ; %bb.1:
-; GFX1164-DPP-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1164-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1164-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1164-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1164-DPP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-DPP-NEXT: .p2align 6
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s2
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s3
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1164-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1164-DPP-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1164-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1164-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX1164-DPP-NEXT: .LBB10_3:
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-DPP-NEXT: s_endpgm
;
; GFX1132-DPP-LABEL: global_atomic_fmax_double_uni_address_uni_value_default_scope_unsafe:
; GFX1132-DPP: ; %bb.0:
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1132-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
+; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
+; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX1132-DPP-NEXT: ; %bb.1:
-; GFX1132-DPP-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
-; GFX1132-DPP-NEXT: s_mov_b32 s50, s14
-; GFX1132-DPP-NEXT: s_mov_b32 s51, s13
-; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1132-DPP-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-DPP-NEXT: .p2align 6
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX1132-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1132-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1132-DPP-NEXT: v_max_f64 v[0:1], v[0:1], 4.0
+; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1132-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1132-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1132-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1132-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[3:4], 4.0
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX1132-DPP-NEXT: .LBB10_3:
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-DPP-NEXT: s_endpgm
- %result = atomicrmw fmax ptr addrspace(1) %ptr, double 4.0 monotonic, align 4, !amdgpu.no.fine.grained.memory !1
+ %result = atomicrmw fmax ptr addrspace(1) %ptr, double 4.0 monotonic, align 8, !amdgpu.no.fine.grained.memory !1
ret void
}
define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe(ptr addrspace(1) %ptr) #0 {
; GFX7LESS-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
; GFX7LESS: ; %bb.0:
-; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-NEXT: s_mov_b32 s32, 0
+; GFX7LESS-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX7LESS-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX7LESS-NEXT: s_mov_b32 s38, -1
+; GFX7LESS-NEXT: s_mov_b32 s39, 0xe8f000
+; GFX7LESS-NEXT: s_add_u32 s36, s36, s11
+; GFX7LESS-NEXT: s_addc_u32 s37, s37, 0
+; GFX7LESS-NEXT: s_mov_b32 s14, s10
+; GFX7LESS-NEXT: s_mov_b32 s13, s9
+; GFX7LESS-NEXT: s_mov_b32 s12, s8
+; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX7LESS-NEXT: s_add_u32 s8, s34, 44
+; GFX7LESS-NEXT: s_addc_u32 s9, s35, 0
+; GFX7LESS-NEXT: s_getpc_b64 s[4:5]
+; GFX7LESS-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX7LESS-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
+; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GFX7LESS-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX7LESS-NEXT: v_or_b32_e32 v40, v0, v2
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX7LESS-NEXT: v_or_b32_e32 v31, v0, v2
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[38:39]
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX7LESS-NEXT: s_mov_b64 s[0:1], exec
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
+; GFX7LESS-NEXT: v_mov_b32_e32 v5, 0x7ff80000
; GFX7LESS-NEXT: .LBB11_1: ; %ComputeLoop
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-NEXT: s_ff1_i32_b64 s4, s[0:1]
-; GFX7LESS-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX7LESS-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
; GFX7LESS-NEXT: v_readlane_b32 s3, v1, s4
; GFX7LESS-NEXT: v_readlane_b32 s2, v0, s4
; GFX7LESS-NEXT: s_lshl_b64 s[4:5], 1, s4
@@ -7613,7 +6048,7 @@ define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_defau
; GFX7LESS-NEXT: s_andn2_b64 s[0:1], s[0:1], s[4:5]
; GFX7LESS-NEXT: v_cmp_ne_u64_e64 s[2:3], s[0:1], 0
; GFX7LESS-NEXT: s_and_b64 vcc, exec, s[2:3]
-; GFX7LESS-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
+; GFX7LESS-NEXT: v_max_f64 v[4:5], v[2:3], v[4:5]
; GFX7LESS-NEXT: s_cbranch_vccnz .LBB11_1
; GFX7LESS-NEXT: ; %bb.2: ; %ComputeEnd
; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
@@ -7623,109 +6058,77 @@ define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_defau
; GFX7LESS-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX7LESS-NEXT: s_cbranch_execz .LBB11_5
; GFX7LESS-NEXT: ; %bb.3:
-; GFX7LESS-NEXT: s_load_dwordx2 s[52:53], s[36:37], 0x9
-; GFX7LESS-NEXT: s_mov_b32 s55, 0xf000
-; GFX7LESS-NEXT: s_mov_b32 s54, -1
+; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x9
+; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
+; GFX7LESS-NEXT: s_mov_b32 s2, -1
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: buffer_load_dwordx2 v[0:1], off, s[52:55], 0
-; GFX7LESS-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
+; GFX7LESS-NEXT: buffer_load_dwordx2 v[2:3], off, s[0:3], 0
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
; GFX7LESS-NEXT: .LBB11_4: ; %atomicrmw.start
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], v[41:42]
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7LESS-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[0:3], 0 glc
+; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-NEXT: s_cbranch_execnz .LBB11_4
; GFX7LESS-NEXT: .LBB11_5:
; GFX7LESS-NEXT: s_endpgm
;
; GFX9-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-NEXT: s_mov_b32 s66, -1
-; GFX9-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-NEXT: s_add_u32 s64, s64, s11
-; GFX9-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-NEXT: s_mov_b32 s51, s8
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_mov_b32 s50, s9
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX9-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX9-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX9-NEXT: s_mov_b32 s38, -1
+; GFX9-NEXT: s_mov_b32 s39, 0xe00000
+; GFX9-NEXT: s_add_u32 s36, s36, s11
+; GFX9-NEXT: s_addc_u32 s37, s37, 0
+; GFX9-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX9-NEXT: s_mov_b32 s12, s8
+; GFX9-NEXT: s_add_u32 s8, s34, 44
+; GFX9-NEXT: s_mov_b32 s13, s9
+; GFX9-NEXT: s_addc_u32 s9, s35, 0
+; GFX9-NEXT: s_getpc_b64 s[4:5]
+; GFX9-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX9-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
+; GFX9-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
+; GFX9-NEXT: s_mov_b32 s14, s10
+; GFX9-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-NEXT: s_mov_b32 s33, s10
-; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX9-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: s_movk_i32 s32, 0x800
+; GFX9-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX9-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX9-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX9-NEXT: s_mov_b32 s32, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: v_mov_b32_e32 v2, 0
+; GFX9-NEXT: v_mov_b32_e32 v4, 0
; GFX9-NEXT: s_mov_b64 s[0:1], exec
-; GFX9-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX9-NEXT: v_mov_b32_e32 v5, 0x7ff80000
; GFX9-NEXT: .LBB11_1: ; %ComputeLoop
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-NEXT: s_ff1_i32_b64 s4, s[0:1]
; GFX9-NEXT: v_readlane_b32 s3, v1, s4
; GFX9-NEXT: v_readlane_b32 s2, v0, s4
-; GFX9-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX9-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
; GFX9-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
; GFX9-NEXT: s_lshl_b64 s[2:3], 1, s4
; GFX9-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
; GFX9-NEXT: s_cmp_lg_u64 s[0:1], 0
-; GFX9-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
+; GFX9-NEXT: v_max_f64 v[4:5], v[2:3], v[4:5]
; GFX9-NEXT: s_cbranch_scc1 .LBB11_1
; GFX9-NEXT: ; %bb.2: ; %ComputeEnd
; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
@@ -7735,90 +6138,55 @@ define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_defau
; GFX9-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX9-NEXT: s_cbranch_execz .LBB11_5
; GFX9-NEXT: ; %bb.3:
-; GFX9-NEXT: s_load_dwordx2 s[52:53], s[36:37], 0x24
-; GFX9-NEXT: v_mov_b32_e32 v0, 0
-; GFX9-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
-; GFX9-NEXT: s_mov_b64 s[54:55], 0
+; GFX9-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX9-NEXT: v_mov_b32_e32 v6, 0
+; GFX9-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX9-NEXT: s_mov_b64 s[2:3], 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: global_load_dwordx2 v[4:5], v0, s[52:53]
+; GFX9-NEXT: global_load_dwordx2 v[2:3], v6, s[0:1]
; GFX9-NEXT: .LBB11_4: ; %atomicrmw.start
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: buffer_store_dword v5, off, s[64:67], 0 offset:4
-; GFX9-NEXT: buffer_store_dword v4, off, s[64:67], 0
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
-; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX9-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: buffer_load_dword v4, off, s[64:67], 0
-; GFX9-NEXT: buffer_load_dword v5, off, s[64:67], 0 offset:4
-; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX9-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
+; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-NEXT: s_cbranch_execnz .LBB11_4
; GFX9-NEXT: .LBB11_5:
; GFX9-NEXT: s_endpgm
;
; GFX1064-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
; GFX1064: ; %bb.0:
-; GFX1064-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-NEXT: s_mov_b32 s66, -1
-; GFX1064-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-NEXT: s_add_u32 s64, s64, s11
+; GFX1064-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1064-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1064-NEXT: s_mov_b32 s38, -1
+; GFX1064-NEXT: s_mov_b32 s39, 0x31e16000
+; GFX1064-NEXT: s_add_u32 s36, s36, s11
; GFX1064-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1064-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-NEXT: s_mov_b32 s51, s8
+; GFX1064-NEXT: s_addc_u32 s37, s37, 0
+; GFX1064-NEXT: s_mov_b32 s12, s8
; GFX1064-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-NEXT: s_mov_b32 s50, s9
+; GFX1064-NEXT: s_mov_b32 s13, s9
; GFX1064-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1064-NEXT: s_getpc_b64 s[4:5]
+; GFX1064-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1064-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1064-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1064-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1064-NEXT: s_mov_b32 s33, s10
-; GFX1064-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1064-NEXT: s_mov_b32 s14, s10
+; GFX1064-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1064-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1064-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1064-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: s_movk_i32 s32, 0x800
+; GFX1064-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1064-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1064-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1064-NEXT: s_mov_b32 s32, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1064-NEXT: v_mov_b32_e32 v2, 0
@@ -7842,93 +6210,42 @@ define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_defau
; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX1064-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
-; GFX1064-NEXT: s_cbranch_execz .LBB11_5
+; GFX1064-NEXT: s_cbranch_execz .LBB11_4
; GFX1064-NEXT: ; %bb.3:
-; GFX1064-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
+; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1064-NEXT: v_mov_b32_e32 v0, 0
-; GFX1064-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
-; GFX1064-NEXT: s_mov_b64 s[54:55], 0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: global_load_dwordx2 v[4:5], v0, s[52:53]
-; GFX1064-NEXT: .LBB11_4: ; %atomicrmw.start
-; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1064-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX1064-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-NEXT: buffer_store_dword v5, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: buffer_store_dword v4, off, s[64:67], 0
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
-; GFX1064-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX1064-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX1064-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-NEXT: v_mov_b32_e32 v1, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: s_clause 0x1
-; GFX1064-NEXT: buffer_load_dword v4, off, s[64:67], 0
-; GFX1064-NEXT: buffer_load_dword v5, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-NEXT: s_andn2_b64 exec, exec, s[54:55]
-; GFX1064-NEXT: s_cbranch_execnz .LBB11_4
-; GFX1064-NEXT: .LBB11_5:
+; GFX1064-NEXT: global_atomic_fmax_x2 v0, v[2:3], s[0:1]
+; GFX1064-NEXT: .LBB11_4:
; GFX1064-NEXT: s_endpgm
;
; GFX1032-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
; GFX1032: ; %bb.0:
-; GFX1032-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-NEXT: s_mov_b32 s66, -1
-; GFX1032-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-NEXT: s_add_u32 s64, s64, s11
+; GFX1032-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1032-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1032-NEXT: s_mov_b32 s38, -1
+; GFX1032-NEXT: s_mov_b32 s39, 0x31c16000
+; GFX1032-NEXT: s_add_u32 s36, s36, s11
; GFX1032-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1032-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-NEXT: s_mov_b32 s51, s8
+; GFX1032-NEXT: s_addc_u32 s37, s37, 0
+; GFX1032-NEXT: s_mov_b32 s12, s8
; GFX1032-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-NEXT: s_mov_b32 s50, s9
+; GFX1032-NEXT: s_mov_b32 s13, s9
; GFX1032-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1032-NEXT: s_getpc_b64 s[4:5]
+; GFX1032-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1032-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1032-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1032-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1032-NEXT: s_mov_b32 s33, s10
-; GFX1032-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-NEXT: s_mov_b32 s14, s10
+; GFX1032-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1032-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1032-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1032-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1032-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1032-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1032-NEXT: s_mov_b32 s32, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1032-NEXT: v_mov_b32_e32 v2, 0
@@ -7948,96 +6265,45 @@ define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_defau
; GFX1032-NEXT: s_cbranch_scc1 .LBB11_1
; GFX1032-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1032-NEXT: s_mov_b32 s54, 0
; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX1032-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX1032-NEXT: s_cbranch_execz .LBB11_5
+; GFX1032-NEXT: s_cbranch_execz .LBB11_4
; GFX1032-NEXT: ; %bb.3:
-; GFX1032-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
+; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1032-NEXT: v_mov_b32_e32 v0, 0
-; GFX1032-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: global_load_dwordx2 v[4:5], v0, s[52:53]
-; GFX1032-NEXT: .LBB11_4: ; %atomicrmw.start
-; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1032-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX1032-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-NEXT: buffer_store_dword v5, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: buffer_store_dword v4, off, s[64:67], 0
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
-; GFX1032-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX1032-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX1032-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-NEXT: v_mov_b32_e32 v1, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: s_clause 0x1
-; GFX1032-NEXT: buffer_load_dword v4, off, s[64:67], 0
-; GFX1032-NEXT: buffer_load_dword v5, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
-; GFX1032-NEXT: s_cbranch_execnz .LBB11_4
-; GFX1032-NEXT: .LBB11_5:
+; GFX1032-NEXT: global_atomic_fmax_x2 v0, v[2:3], s[0:1]
+; GFX1032-NEXT: .LBB11_4:
; GFX1032-NEXT: s_endpgm
;
; GFX1164-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
; GFX1164: ; %bb.0:
; GFX1164-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1164-NEXT: s_mov_b32 s51, s8
+; GFX1164-NEXT: s_mov_b32 s12, s8
; GFX1164-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-NEXT: s_mov_b32 s50, s9
+; GFX1164-NEXT: s_mov_b32 s13, s9
; GFX1164-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1164-NEXT: s_getpc_b64 s[4:5]
+; GFX1164-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1164-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1164-NEXT: v_mov_b32_e32 v31, v0
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: s_mov_b32 s33, s10
-; GFX1164-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX1164-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1164-NEXT: s_mov_b32 s14, s10
+; GFX1164-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1164-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1164-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: s_mov_b32 s32, 32
-; GFX1164-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1164-NEXT: s_mov_b32 s32, 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: v_mov_b32_e32 v2, 0
-; GFX1164-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX1164-NEXT: s_swappc_b64 s[30:31], s[16:17]
+; GFX1164-NEXT: v_mov_b32_e32 v4, 0
+; GFX1164-NEXT: v_mov_b32_e32 v5, 0x7ff80000
; GFX1164-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-NEXT: .LBB11_1: ; %ComputeLoop
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1164-NEXT: s_ctz_i32_b64 s4, s[0:1]
-; GFX1164-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX1164-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
; GFX1164-NEXT: v_readlane_b32 s3, v1, s4
; GFX1164-NEXT: v_readlane_b32 s2, v0, s4
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
@@ -8046,7 +6312,7 @@ define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_defau
; GFX1164-NEXT: s_and_not1_b64 s[0:1], s[0:1], s[2:3]
; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1164-NEXT: s_cmp_lg_u64 s[0:1], 0
-; GFX1164-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-NEXT: v_max_f64 v[4:5], v[2:3], v[4:5]
; GFX1164-NEXT: s_cbranch_scc1 .LBB11_1
; GFX1164-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
@@ -8057,89 +6323,57 @@ define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_defau
; GFX1164-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX1164-NEXT: s_cbranch_execz .LBB11_5
; GFX1164-NEXT: ; %bb.3:
-; GFX1164-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1164-NEXT: v_mov_b32_e32 v0, 0
-; GFX1164-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
-; GFX1164-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1164-NEXT: v_mov_b32_e32 v6, 0
+; GFX1164-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: global_load_b64 v[4:5], v0, s[52:53]
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-NEXT: .p2align 6
+; GFX1164-NEXT: global_load_b64 v[2:3], v6, s[0:1]
; GFX1164-NEXT: .LBB11_4: ; %atomicrmw.start
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX1164-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
-; GFX1164-NEXT: scratch_store_b64 off, v[4:5], off
-; GFX1164-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-NEXT: scratch_store_b64 off, v[0:1], off offset:8
-; GFX1164-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: scratch_load_b64 v[4:5], off, off
-; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1164-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
+; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1164-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-NEXT: s_cbranch_execnz .LBB11_4
; GFX1164-NEXT: .LBB11_5:
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-NEXT: s_endpgm
;
; GFX1132-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
; GFX1132: ; %bb.0:
; GFX1132-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1132-NEXT: s_mov_b64 s[48:49], s[0:1]
+; GFX1132-NEXT: v_mov_b32_e32 v31, v0
; GFX1132-NEXT: s_add_u32 s8, s34, 44
; GFX1132-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX1132-NEXT: v_mov_b32_e32 v31, v0
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1132-NEXT: s_mov_b32 s50, s14
-; GFX1132-NEXT: s_mov_b32 s51, s13
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
+; GFX1132-NEXT: s_getpc_b64 s[4:5]
+; GFX1132-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1132-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1132-NEXT: s_mov_b32 s12, s13
+; GFX1132-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1132-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1132-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX1132-NEXT: s_mov_b32 s13, s14
; GFX1132-NEXT: s_mov_b32 s14, s15
-; GFX1132-NEXT: s_mov_b32 s32, 32
-; GFX1132-NEXT: s_mov_b32 s33, s15
-; GFX1132-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-NEXT: s_mov_b32 s32, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: v_mov_b32_e32 v2, 0
-; GFX1132-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX1132-NEXT: s_swappc_b64 s[30:31], s[16:17]
+; GFX1132-NEXT: v_mov_b32_e32 v4, 0
+; GFX1132-NEXT: v_mov_b32_e32 v5, 0x7ff80000
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-NEXT: .LBB11_1: ; %ComputeLoop
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1132-NEXT: s_ctz_i32_b32 s1, s0
-; GFX1132-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX1132-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
; GFX1132-NEXT: v_readlane_b32 s3, v1, s1
; GFX1132-NEXT: v_readlane_b32 s2, v0, s1
; GFX1132-NEXT: s_lshl_b32 s1, 1, s1
@@ -8147,811 +6381,557 @@ define amdgpu_kernel void @global_atomic_fmax_double_uni_address_div_value_defau
; GFX1132-NEXT: s_and_not1_b32 s0, s0, s1
; GFX1132-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
; GFX1132-NEXT: s_cmp_lg_u32 s0, 0
-; GFX1132-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-NEXT: v_max_f64 v[4:5], v[2:3], v[4:5]
; GFX1132-NEXT: s_cbranch_scc1 .LBB11_1
; GFX1132-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-NEXT: s_mov_b32 s54, 0
+; GFX1132-NEXT: s_mov_b32 s2, 0
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX1132-NEXT: s_cbranch_execz .LBB11_5
; GFX1132-NEXT: ; %bb.3:
-; GFX1132-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1132-NEXT: v_mov_b32_e32 v0, 0
+; GFX1132-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1132-NEXT: v_mov_b32_e32 v6, 0
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX1132-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
+; GFX1132-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: global_load_b64 v[4:5], v0, s[52:53]
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-NEXT: .p2align 6
+; GFX1132-NEXT: global_load_b64 v[2:3], v6, s[0:1]
; GFX1132-NEXT: .LBB11_4: ; %atomicrmw.start
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX1132-NEXT: s_add_u32 s8, s34, 44
-; GFX1132-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-NEXT: v_mov_b32_e32 v31, v40
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: v_mov_b32_e32 v3, s53
-; GFX1132-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1132-NEXT: s_mov_b32 s12, s51
-; GFX1132-NEXT: s_mov_b32 s13, s50
-; GFX1132-NEXT: s_mov_b32 s14, s33
-; GFX1132-NEXT: v_mov_b32_e32 v6, 0
-; GFX1132-NEXT: v_mov_b32_e32 v2, s52
-; GFX1132-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
-; GFX1132-NEXT: scratch_store_b64 off, v[4:5], off
-; GFX1132-NEXT: v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v5, 8
-; GFX1132-NEXT: scratch_store_b64 off, v[0:1], off offset:8
-; GFX1132-NEXT: v_dual_mov_b32 v0, 8 :: v_dual_mov_b32 v1, 0
-; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: scratch_load_b64 v[4:5], off, off
-; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1132-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
+; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1132-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-NEXT: s_cbranch_execnz .LBB11_4
; GFX1132-NEXT: .LBB11_5:
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-NEXT: s_endpgm
;
; GFX7LESS-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
; GFX7LESS-DPP: ; %bb.0:
-; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-DPP-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s55, 0xf000
-; GFX7LESS-DPP-NEXT: s_mov_b32 s54, -1
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s32, 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s50, -1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s51, 0xe8f000
+; GFX7LESS-DPP-NEXT: s_add_u32 s48, s48, s11
+; GFX7LESS-DPP-NEXT: s_addc_u32 s49, s49, 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s10
+; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s8
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[36:37], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s39, 0xf000
+; GFX7LESS-DPP-NEXT: s_mov_b32 s38, -1
+; GFX7LESS-DPP-NEXT: s_add_u32 s8, s4, 44
+; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s5, 0
+; GFX7LESS-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX7LESS-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX7LESS-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GFX7LESS-DPP-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v40, v0, v2
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX7LESS-DPP-NEXT: v_or_b32_e32 v31, v0, v2
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[2:3], off, s[52:55], 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-DPP-NEXT: v_max_f64 v[41:42], v[0:1], v[0:1]
+; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[2:3], off, s[36:39], 0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], 0
+; GFX7LESS-DPP-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
; GFX7LESS-DPP-NEXT: .LBB11_1: ; %atomicrmw.start
; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
-; GFX7LESS-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[0:1], v[41:42]
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[0:1], v[4:5]
; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v3, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX7LESS-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[36:39], 0 glc
+; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-DPP-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[0:1]
; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB11_1
; GFX7LESS-DPP-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX7LESS-DPP-NEXT: s_endpgm
;
; GFX9-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
; GFX9-DPP: ; %bb.0:
-; GFX9-DPP-NEXT: s_mov_b32 s68, SCRATCH_RSRC_DWORD0
-; GFX9-DPP-NEXT: s_mov_b32 s69, SCRATCH_RSRC_DWORD1
-; GFX9-DPP-NEXT: s_mov_b32 s70, -1
-; GFX9-DPP-NEXT: s_mov_b32 s71, 0xe00000
-; GFX9-DPP-NEXT: s_add_u32 s68, s68, s11
-; GFX9-DPP-NEXT: s_addc_u32 s69, s69, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-DPP-NEXT: s_mov_b32 s51, s8
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_mov_b32 s50, s9
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX9-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX9-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX9-DPP-NEXT: s_mov_b32 s38, -1
+; GFX9-DPP-NEXT: s_mov_b32 s39, 0xe00000
+; GFX9-DPP-NEXT: s_add_u32 s36, s36, s11
+; GFX9-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX9-DPP-NEXT: s_mov_b32 s12, s8
+; GFX9-DPP-NEXT: s_add_u32 s8, s34, 44
+; GFX9-DPP-NEXT: s_mov_b32 s13, s9
+; GFX9-DPP-NEXT: s_addc_u32 s9, s35, 0
+; GFX9-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX9-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX9-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
+; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
+; GFX9-DPP-NEXT: s_mov_b32 s14, s10
+; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-DPP-NEXT: s_mov_b32 s33, s10
-; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[68:69]
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[70:71]
-; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX9-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX9-DPP-NEXT: s_mov_b32 s32, 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
-; GFX9-DPP-NEXT: v_cndmask_b32_e64 v11, v9, v1, s[0:1]
-; GFX9-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:1 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:1 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0x7ff80000
+; GFX9-DPP-NEXT: v_cndmask_b32_e64 v6, v4, v1, s[0:1]
+; GFX9-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:1 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:1 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:2 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:2 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:2 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:2 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:4 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:4 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:4 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:4 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:8 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:8 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:8 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:8 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_bcast:15 row_mask:0xa bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_bcast:15 row_mask:0xa bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[12:13]
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_bcast:15 row_mask:0xa bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_bcast:15 row_mask:0xa bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[7:8]
; GFX9-DPP-NEXT: s_nop 1
-; GFX9-DPP-NEXT: v_mov_b32_dpp v9, v11 row_bcast:31 row_mask:0xc bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v10 row_bcast:31 row_mask:0xc bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX9-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
+; GFX9-DPP-NEXT: v_mov_b32_dpp v4, v6 row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v3, v5 row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[3:4]
+; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 0
; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v1
; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX9-DPP-NEXT: v_readlane_b32 s53, v9, 63
-; GFX9-DPP-NEXT: v_readlane_b32 s52, v8, 63
+; GFX9-DPP-NEXT: v_readlane_b32 s3, v4, 63
+; GFX9-DPP-NEXT: v_readlane_b32 s2, v3, 63
; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[2:3]
; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
+; GFX9-DPP-NEXT: s_and_saveexec_b64 s[2:3], vcc
; GFX9-DPP-NEXT: s_cbranch_execz .LBB11_3
; GFX9-DPP-NEXT: ; %bb.1:
-; GFX9-DPP-NEXT: s_load_dwordx2 s[54:55], s[36:37], 0x24
-; GFX9-DPP-NEXT: s_mov_b64 s[64:65], 0
+; GFX9-DPP-NEXT: s_load_dwordx2 s[2:3], s[34:35], 0x24
+; GFX9-DPP-NEXT: s_mov_b64 s[4:5], 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[54:55]
+; GFX9-DPP-NEXT: global_load_dwordx2 v[11:12], v0, s[2:3]
; GFX9-DPP-NEXT: .LBB11_2: ; %atomicrmw.start
; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX9-DPP-NEXT: v_max_f64 v[3:4], s[52:53], s[52:53]
+; GFX9-DPP-NEXT: v_max_f64 v[1:2], s[0:1], s[0:1]
; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[1:2], v[1:2]
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[68:69]
-; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[68:71], 0 offset:4
-; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[68:71], 0
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
-; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[68:71], 0 offset:12
-; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[68:71], 0 offset:8
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[70:71]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s54
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s55
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[68:71], 0
-; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[68:71], 0 offset:4
-; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-DPP-NEXT: s_or_b64 s[64:65], vcc, s[64:65]
-; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[64:65]
+; GFX9-DPP-NEXT: v_max_f64 v[9:10], v[11:12], v[11:12]
+; GFX9-DPP-NEXT: v_max_f64 v[9:10], v[9:10], v[1:2]
+; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[1:2], v0, v[9:12], s[2:3] glc
+; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[1:2], v[11:12]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v12, v2
+; GFX9-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v11, v1
+; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX9-DPP-NEXT: s_cbranch_execnz .LBB11_2
; GFX9-DPP-NEXT: .LBB11_3:
; GFX9-DPP-NEXT: s_endpgm
;
; GFX1064-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
; GFX1064-DPP: ; %bb.0:
-; GFX1064-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1064-DPP-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-DPP-NEXT: s_add_u32 s64, s64, s11
+; GFX1064-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1064-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1064-DPP-NEXT: s_mov_b32 s38, -1
+; GFX1064-DPP-NEXT: s_mov_b32 s39, 0x31e16000
+; GFX1064-DPP-NEXT: s_add_u32 s36, s36, s11
; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1064-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1064-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX1064-DPP-NEXT: s_mov_b32 s12, s8
; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1064-DPP-NEXT: s_mov_b32 s13, s9
; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1064-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1064-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1064-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1064-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX1064-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1064-DPP-NEXT: s_mov_b32 s32, 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1064-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s[0:1]
-; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1064-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_readlane_b32 s3, v9, 0
-; GFX1064-DPP-NEXT: v_readlane_b32 s5, v9, 32
-; GFX1064-DPP-NEXT: v_readlane_b32 s4, v8, 32
-; GFX1064-DPP-NEXT: v_readlane_b32 s2, v8, 0
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], s[4:5], s[4:5]
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], s[2:3], s[2:3]
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0x7ff80000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v6, 0x7ff80000, v1, s[0:1]
+; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[3:4]
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
+; GFX1064-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_readlane_b32 s3, v4, 0
+; GFX1064-DPP-NEXT: v_readlane_b32 s5, v4, 32
+; GFX1064-DPP-NEXT: v_readlane_b32 s4, v3, 32
+; GFX1064-DPP-NEXT: v_readlane_b32 s2, v3, 0
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], s[4:5], s[4:5]
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], s[2:3], s[2:3]
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
; GFX1064-DPP-NEXT: s_mov_b64 exec, s[0:1]
; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v41, v8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v42, v9
-; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v7, exec_hi, v0
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, v3
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, v4
+; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v7
; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
-; GFX1064-DPP-NEXT: s_cbranch_execz .LBB11_3
+; GFX1064-DPP-NEXT: s_cbranch_execz .LBB11_2
; GFX1064-DPP-NEXT: ; %bb.1:
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1064-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
-; GFX1064-DPP-NEXT: .LBB11_2: ; %atomicrmw.start
-; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[41:42], v[41:42]
-; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[1:2], v[1:2]
-; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-DPP-NEXT: s_clause 0x1
-; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
-; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB11_2
-; GFX1064-DPP-NEXT: .LBB11_3:
+; GFX1064-DPP-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
+; GFX1064-DPP-NEXT: .LBB11_2:
; GFX1064-DPP-NEXT: s_endpgm
;
; GFX1032-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
; GFX1032-DPP: ; %bb.0:
-; GFX1032-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1032-DPP-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-DPP-NEXT: s_add_u32 s64, s64, s11
+; GFX1032-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1032-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1032-DPP-NEXT: s_mov_b32 s38, -1
+; GFX1032-DPP-NEXT: s_mov_b32 s39, 0x31c16000
+; GFX1032-DPP-NEXT: s_add_u32 s36, s36, s11
; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1032-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1032-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX1032-DPP-NEXT: s_mov_b32 s12, s8
; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1032-DPP-NEXT: s_mov_b32 s13, s9
; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1032-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1032-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1032-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1032-DPP-NEXT: s_mov_b32 s32, 0
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1032-DPP-NEXT: s_or_saveexec_b32 s0, -1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s0
-; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1032-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0x7ff80000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v6, 0x7ff80000, v1, s0
+; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[3:4]
+; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[5:6], v[3:4]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
+; GFX1032-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[5:6]
; GFX1032-DPP-NEXT: s_mov_b32 exec_lo, s0
-; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, v8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, v9
-; GFX1032-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
+; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v7, exec_lo, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, v3
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, v4
+; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v7
; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX1032-DPP-NEXT: s_cbranch_execz .LBB11_3
+; GFX1032-DPP-NEXT: s_cbranch_execz .LBB11_2
; GFX1032-DPP-NEXT: ; %bb.1:
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1032-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
-; GFX1032-DPP-NEXT: .LBB11_2: ; %atomicrmw.start
-; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[41:42]
-; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-DPP-NEXT: s_clause 0x1
-; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
-; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB11_2
-; GFX1032-DPP-NEXT: .LBB11_3:
+; GFX1032-DPP-NEXT: global_atomic_fmax_x2 v2, v[0:1], s[0:1]
+; GFX1032-DPP-NEXT: .LBB11_2:
; GFX1032-DPP-NEXT: s_endpgm
;
; GFX1164-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
; GFX1164-DPP: ; %bb.0:
; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1164-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1164-DPP-NEXT: s_mov_b32 s12, s8
; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1164-DPP-NEXT: s_mov_b32 s13, s9
; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1164-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1164-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1164-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v0
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX1164-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1164-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1164-DPP-NEXT: s_mov_b32 s32, 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
+; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1164-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s[0:1]
-; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v5, 0x7ff80000, v1, s[0:1]
+; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s[0:1]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[4:5], v[2:3]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 0x7ff80000
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 0x7ff80000
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 0x7ff80000
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1164-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
+; GFX1164-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1164-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_permlane64_b32 v11, v9
-; GFX1164-DPP-NEXT: v_permlane64_b32 v10, v8
+; GFX1164-DPP-NEXT: v_permlane64_b32 v5, v3
+; GFX1164-DPP-NEXT: v_permlane64_b32 v4, v2
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_mov_b64 exec, s[0:1]
-; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, v8
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, v9
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
-; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
+; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v6, exec_hi, v0
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, v2
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, v3
+; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
; GFX1164-DPP-NEXT: s_cbranch_execz .LBB11_3
; GFX1164-DPP-NEXT: ; %bb.1:
-; GFX1164-DPP-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1164-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
-; GFX1164-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX1164-DPP-NEXT: v_max_f64 v[0:1], v[0:1], v[0:1]
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-DPP-NEXT: .p2align 6
+; GFX1164-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
; GFX1164-DPP-NEXT: .LBB11_2: ; %atomicrmw.start
; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[41:42]
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-DPP-NEXT: v_max_f64 v[6:7], v[8:9], v[8:9]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1164-DPP-NEXT: v_max_f64 v[6:7], v[6:7], v[0:1]
+; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
+; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[8:9]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v9, v7
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, v6
+; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB11_2
; GFX1164-DPP-NEXT: .LBB11_3:
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-DPP-NEXT: s_endpgm
;
; GFX1132-DPP-LABEL: global_atomic_fmax_double_uni_address_div_value_default_scope_unsafe:
; GFX1132-DPP: ; %bb.0:
; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1132-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1132-DPP-NEXT: s_mov_b32 s50, s14
-; GFX1132-DPP-NEXT: s_mov_b32 s51, s13
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
+; GFX1132-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1132-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1132-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1132-DPP-NEXT: s_mov_b32 s12, s13
+; GFX1132-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX1132-DPP-NEXT: s_mov_b32 s13, s14
; GFX1132-DPP-NEXT: s_mov_b32 s14, s15
-; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-DPP-NEXT: s_mov_b32 s32, 0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
+; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1132-DPP-NEXT: s_or_saveexec_b32 s0, -1
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v9, 0x7ff80000 :: v_dual_mov_b32 v8, 0
-; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s0
-; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, 0x7ff80000 :: v_dual_mov_b32 v2, 0
+; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v5, 0x7ff80000, v1, s0
+; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[10:11], v[8:9]
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
+; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[4:5], v[2:3]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 0x7ff80000 :: v_dual_mov_b32 v4, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 0x7ff80000 :: v_dual_mov_b32 v4, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 0x7ff80000 :: v_dual_mov_b32 v4, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1132-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
+; GFX1132-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1132-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[10:11]
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[4:5]
; GFX1132-DPP-NEXT: s_mov_b32 exec_lo, s0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v3, v8
-; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, v9
-; GFX1132-DPP-NEXT: s_mov_b32 s54, 0
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v0, v2
+; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v6, exec_lo, 0
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v10, 0 :: v_dual_mov_b32 v1, v3
+; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
+; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
; GFX1132-DPP-NEXT: s_cbranch_execz .LBB11_3
; GFX1132-DPP-NEXT: ; %bb.1:
-; GFX1132-DPP-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1132-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
+; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1132-DPP-NEXT: v_max_f64 v[0:1], v[0:1], v[0:1]
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-DPP-NEXT: .p2align 6
+; GFX1132-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
; GFX1132-DPP-NEXT: .LBB11_2: ; %atomicrmw.start
; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1132-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1132-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[41:42]
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-DPP-NEXT: v_max_f64 v[6:7], v[8:9], v[8:9]
+; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1132-DPP-NEXT: v_max_f64 v[6:7], v[6:7], v[0:1]
+; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
+; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[6:7], v[8:9]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6
+; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB11_2
; GFX1132-DPP-NEXT: .LBB11_3:
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-DPP-NEXT: s_endpgm
%divValue = call double @div.double.value()
- %result = atomicrmw fmax ptr addrspace(1) %ptr, double %divValue monotonic, align 4, !amdgpu.no.fine.grained.memory !1
+ %result = atomicrmw fmax ptr addrspace(1) %ptr, double %divValue monotonic, align 8, !amdgpu.no.fine.grained.memory !1
ret void
}
diff --git a/llvm/test/CodeGen/AMDGPU/global_atomics_scan_fmin.ll b/llvm/test/CodeGen/AMDGPU/global_atomics_scan_fmin.ll
index 1f76a476107a3..4fccfc051d56d 100644
--- a/llvm/test/CodeGen/AMDGPU/global_atomics_scan_fmin.ll
+++ b/llvm/test/CodeGen/AMDGPU/global_atomics_scan_fmin.ll
@@ -3134,932 +3134,377 @@ define amdgpu_kernel void @global_atomic_fmin_uni_address_div_value_default_scop
define amdgpu_kernel void @global_atomic_fmin_double_uni_address_uni_value_agent_scope_unsafe(ptr addrspace(1) %ptr) #0 {
; GFX7LESS-LABEL: global_atomic_fmin_double_uni_address_uni_value_agent_scope_unsafe:
; GFX7LESS: ; %bb.0:
-; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v3, exec_lo, 0
-; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v3, exec_hi, v3
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
+; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
+; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX7LESS-NEXT: s_cbranch_execz .LBB6_3
; GFX7LESS-NEXT: ; %bb.1:
-; GFX7LESS-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
+; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX7LESS-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX7LESS-NEXT: v_or_b32_e32 v3, v0, v1
+; GFX7LESS-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, s0
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, s1
-; GFX7LESS-NEXT: v_or_b32_e32 v40, v3, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, s6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, s7
+; GFX7LESS-NEXT: s_mov_b32 s2, -1
; GFX7LESS-NEXT: .LBB6_2: ; %atomicrmw.start
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_min_f64 v[0:1], v[2:3], 4.0
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7LESS-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-NEXT: v_mov_b32_e32 v7, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v6, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v5, v1
+; GFX7LESS-NEXT: v_mov_b32_e32 v4, v0
+; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, v4
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, v5
+; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-NEXT: s_cbranch_execnz .LBB6_2
; GFX7LESS-NEXT: .LBB6_3:
; GFX7LESS-NEXT: s_endpgm
;
; GFX9-LABEL: global_atomic_fmin_double_uni_address_uni_value_agent_scope_unsafe:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-NEXT: s_mov_b32 s66, -1
-; GFX9-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX9-NEXT: s_add_u32 s64, s64, s11
-; GFX9-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX9-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX9-NEXT: s_movk_i32 s32, 0x800
+; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX9-NEXT: s_cbranch_execz .LBB6_3
; GFX9-NEXT: ; %bb.1:
-; GFX9-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX9-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX9-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX9-NEXT: s_mov_b32 s33, s10
-; GFX9-NEXT: s_mov_b32 s50, s9
+; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9-NEXT: s_mov_b64 s[2:3], 0
+; GFX9-NEXT: v_mov_b32_e32 v4, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX9-NEXT: s_mov_b32 s51, s8
-; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX9-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v2, s1
-; GFX9-NEXT: s_mov_b64 s[54:55], 0
-; GFX9-NEXT: v_mov_b32_e32 v1, s0
-; GFX9-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX9-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-NEXT: v_mov_b32_e32 v3, s5
; GFX9-NEXT: .LBB6_2: ; %atomicrmw.start
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX9-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX9-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-NEXT: s_cbranch_execnz .LBB6_2
; GFX9-NEXT: .LBB6_3:
; GFX9-NEXT: s_endpgm
;
; GFX1064-LABEL: global_atomic_fmin_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1064: ; %bb.0:
-; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1064-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-NEXT: s_mov_b32 s66, -1
-; GFX1064-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX1064-NEXT: s_add_u32 s64, s64, s11
-; GFX1064-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-NEXT: s_movk_i32 s32, 0x800
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
-; GFX1064-NEXT: s_cbranch_execz .LBB6_3
+; GFX1064-NEXT: s_cbranch_execz .LBB6_2
; GFX1064-NEXT: ; %bb.1:
-; GFX1064-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1064-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1064-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1064-NEXT: s_mov_b32 s33, s10
-; GFX1064-NEXT: s_mov_b32 s50, s9
-; GFX1064-NEXT: s_mov_b32 s51, s8
-; GFX1064-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1064-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1064-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1064-NEXT: s_mov_b64 s[54:55], 0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: v_mov_b32_e32 v2, s1
-; GFX1064-NEXT: v_mov_b32_e32 v1, s0
-; GFX1064-NEXT: .LBB6_2: ; %atomicrmw.start
-; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1064-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1064-NEXT: s_add_u32 s8, s36, 44
-; GFX1064-NEXT: s_addc_u32 s9, s37, 0
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX1064-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-NEXT: v_mov_b32_e32 v4, 0
+; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1064-NEXT: v_mov_b32_e32 v0, 0
+; GFX1064-NEXT: v_mov_b32_e32 v1, 0x40100000
+; GFX1064-NEXT: v_mov_b32_e32 v2, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: s_clause 0x1
-; GFX1064-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-NEXT: s_andn2_b64 exec, exec, s[54:55]
-; GFX1064-NEXT: s_cbranch_execnz .LBB6_2
-; GFX1064-NEXT: .LBB6_3:
+; GFX1064-NEXT: global_atomic_fmin_x2 v2, v[0:1], s[0:1]
+; GFX1064-NEXT: .LBB6_2:
; GFX1064-NEXT: s_endpgm
;
; GFX1032-LABEL: global_atomic_fmin_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1032: ; %bb.0:
-; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1032-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-NEXT: s_mov_b32 s66, -1
-; GFX1032-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
-; GFX1032-NEXT: s_add_u32 s64, s64, s11
-; GFX1032-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-NEXT: s_mov_b32 s54, 0
-; GFX1032-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX1032-NEXT: s_cbranch_execz .LBB6_3
+; GFX1032-NEXT: s_cbranch_execz .LBB6_2
; GFX1032-NEXT: ; %bb.1:
-; GFX1032-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1032-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1032-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1032-NEXT: s_mov_b32 s33, s10
-; GFX1032-NEXT: s_mov_b32 s50, s9
-; GFX1032-NEXT: s_mov_b32 s51, s8
-; GFX1032-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1032-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1032-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: v_mov_b32_e32 v2, s1
-; GFX1032-NEXT: v_mov_b32_e32 v1, s0
-; GFX1032-NEXT: .LBB6_2: ; %atomicrmw.start
-; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1032-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1032-NEXT: s_add_u32 s8, s36, 44
-; GFX1032-NEXT: s_addc_u32 s9, s37, 0
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX1032-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-NEXT: v_mov_b32_e32 v4, 0
+; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1032-NEXT: v_mov_b32_e32 v0, 0
+; GFX1032-NEXT: v_mov_b32_e32 v1, 0x40100000
+; GFX1032-NEXT: v_mov_b32_e32 v2, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: s_clause 0x1
-; GFX1032-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
-; GFX1032-NEXT: s_cbranch_execnz .LBB6_2
-; GFX1032-NEXT: .LBB6_3:
+; GFX1032-NEXT: global_atomic_fmin_x2 v2, v[0:1], s[0:1]
+; GFX1032-NEXT: .LBB6_2:
; GFX1032-NEXT: s_endpgm
;
; GFX1164-LABEL: global_atomic_fmin_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1164: ; %bb.0:
-; GFX1164-NEXT: v_mov_b32_e32 v40, v0
; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-NEXT: s_mov_b32 s32, 32
; GFX1164-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1164-NEXT: s_cbranch_execz .LBB6_3
; GFX1164-NEXT: ; %bb.1:
-; GFX1164-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1164-NEXT: s_mov_b32 s33, s10
-; GFX1164-NEXT: s_mov_b32 s50, s9
-; GFX1164-NEXT: s_mov_b32 s51, s8
-; GFX1164-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1164-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1164-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1164-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1164-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: v_mov_b32_e32 v2, s1
-; GFX1164-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-NEXT: .p2align 6
+; GFX1164-NEXT: v_mov_b32_e32 v2, s2
+; GFX1164-NEXT: v_mov_b32_e32 v3, s3
+; GFX1164-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-NEXT: .LBB6_2: ; %atomicrmw.start
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1164-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1164-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
; GFX1164-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1164-NEXT: s_add_u32 s8, s36, 44
-; GFX1164-NEXT: s_addc_u32 s9, s37, 0
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX1164-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-NEXT: s_cbranch_execnz .LBB6_2
; GFX1164-NEXT: .LBB6_3:
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-NEXT: s_endpgm
;
; GFX1132-LABEL: global_atomic_fmin_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1132: ; %bb.0:
-; GFX1132-NEXT: v_mov_b32_e32 v40, v0
; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1132-NEXT: s_mov_b32 s54, 0
-; GFX1132-NEXT: s_mov_b32 s32, 32
+; GFX1132-NEXT: s_mov_b32 s2, 0
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
+; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-NEXT: s_cbranch_execz .LBB6_3
; GFX1132-NEXT: ; %bb.1:
-; GFX1132-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1132-NEXT: s_mov_b32 s33, s15
-; GFX1132-NEXT: s_mov_b32 s50, s14
-; GFX1132-NEXT: s_mov_b32 s51, s13
-; GFX1132-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1132-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1132-NEXT: v_mov_b32_e32 v4, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1132-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-NEXT: .p2align 6
+; GFX1132-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX1132-NEXT: .LBB6_2: ; %atomicrmw.start
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1132-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1132-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
; GFX1132-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1132-NEXT: s_add_u32 s8, s36, 44
-; GFX1132-NEXT: s_addc_u32 s9, s37, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1132-NEXT: s_mov_b32 s12, s51
-; GFX1132-NEXT: s_mov_b32 s13, s50
-; GFX1132-NEXT: s_mov_b32 s14, s33
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX1132-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX1132-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-NEXT: s_cbranch_execnz .LBB6_2
; GFX1132-NEXT: .LBB6_3:
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-NEXT: s_endpgm
;
; GFX7LESS-DPP-LABEL: global_atomic_fmin_double_uni_address_uni_value_agent_scope_unsafe:
; GFX7LESS-DPP: ; %bb.0:
-; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v3, exec_lo, 0
-; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v3, exec_hi, v3
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
+; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX7LESS-DPP-NEXT: s_cbranch_execz .LBB6_3
; GFX7LESS-DPP-NEXT: ; %bb.1:
-; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-DPP-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v3, v0, v1
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s3, 0xf000
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, s0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, s1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v40, v3, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s7
+; GFX7LESS-DPP-NEXT: s_mov_b32 s2, -1
; GFX7LESS-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-DPP-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_min_f64 v[0:1], v[2:3], 4.0
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7LESS-DPP-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, v1
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, v0
+; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX7LESS-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v4
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v5
+; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB6_2
; GFX7LESS-DPP-NEXT: .LBB6_3:
; GFX7LESS-DPP-NEXT: s_endpgm
;
; GFX9-DPP-LABEL: global_atomic_fmin_double_uni_address_uni_value_agent_scope_unsafe:
; GFX9-DPP: ; %bb.0:
-; GFX9-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-DPP-NEXT: s_mov_b32 s66, -1
-; GFX9-DPP-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX9-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX9-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX9-DPP-NEXT: s_cbranch_execz .LBB6_3
; GFX9-DPP-NEXT: ; %bb.1:
-; GFX9-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX9-DPP-NEXT: s_mov_b32 s33, s10
-; GFX9-DPP-NEXT: s_mov_b32 s50, s9
+; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX9-DPP-NEXT: s_mov_b32 s51, s8
-; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX9-DPP-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX9-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s5
; GFX9-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX9-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX9-DPP-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-DPP-NEXT: s_cbranch_execnz .LBB6_2
; GFX9-DPP-NEXT: .LBB6_3:
; GFX9-DPP-NEXT: s_endpgm
;
; GFX1064-DPP-LABEL: global_atomic_fmin_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1064-DPP: ; %bb.0:
-; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1064-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1064-DPP-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX1064-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX1064-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
-; GFX1064-DPP-NEXT: s_cbranch_execz .LBB6_3
+; GFX1064-DPP-NEXT: s_cbranch_execz .LBB6_2
; GFX1064-DPP-NEXT: ; %bb.1:
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1064-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1064-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1064-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1064-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
-; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1064-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1064-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0x40100000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-DPP-NEXT: s_clause 0x1
-; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
-; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB6_2
-; GFX1064-DPP-NEXT: .LBB6_3:
+; GFX1064-DPP-NEXT: global_atomic_fmin_x2 v2, v[0:1], s[0:1]
+; GFX1064-DPP-NEXT: .LBB6_2:
; GFX1064-DPP-NEXT: s_endpgm
;
; GFX1032-DPP-LABEL: global_atomic_fmin_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1032-DPP: ; %bb.0:
-; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1032-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1032-DPP-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
-; GFX1032-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX1032-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX1032-DPP-NEXT: s_cbranch_execz .LBB6_3
+; GFX1032-DPP-NEXT: s_cbranch_execz .LBB6_2
; GFX1032-DPP-NEXT: ; %bb.1:
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1032-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1032-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1032-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
-; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1032-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1032-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0x40100000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, 0
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-DPP-NEXT: s_clause 0x1
-; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
-; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB6_2
-; GFX1032-DPP-NEXT: .LBB6_3:
+; GFX1032-DPP-NEXT: global_atomic_fmin_x2 v2, v[0:1], s[0:1]
+; GFX1032-DPP-NEXT: .LBB6_2:
; GFX1032-DPP-NEXT: s_endpgm
;
; GFX1164-DPP-LABEL: global_atomic_fmin_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1164-DPP: ; %bb.0:
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1164-DPP-NEXT: s_cbranch_execz .LBB6_3
; GFX1164-DPP-NEXT: ; %bb.1:
-; GFX1164-DPP-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1164-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1164-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1164-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1164-DPP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-DPP-NEXT: .p2align 6
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s2
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s3
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1164-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1164-DPP-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1164-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1164-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB6_2
; GFX1164-DPP-NEXT: .LBB6_3:
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-DPP-NEXT: s_endpgm
;
; GFX1132-DPP-LABEL: global_atomic_fmin_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1132-DPP: ; %bb.0:
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1132-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
+; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
+; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-DPP-NEXT: s_cbranch_execz .LBB6_3
; GFX1132-DPP-NEXT: ; %bb.1:
-; GFX1132-DPP-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
-; GFX1132-DPP-NEXT: s_mov_b32 s50, s14
-; GFX1132-DPP-NEXT: s_mov_b32 s51, s13
-; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1132-DPP-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-DPP-NEXT: .p2align 6
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX1132-DPP-NEXT: .LBB6_2: ; %atomicrmw.start
; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1132-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1132-DPP-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1132-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1132-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1132-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1132-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX1132-DPP-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB6_2
; GFX1132-DPP-NEXT: .LBB6_3:
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-DPP-NEXT: s_endpgm
- %result = atomicrmw fmin ptr addrspace(1) %ptr, double 4.0 syncscope("agent") monotonic, align 4, !amdgpu.no.fine.grained.memory !1
+ %result = atomicrmw fmin ptr addrspace(1) %ptr, double 4.0 syncscope("agent") monotonic, align 8, !amdgpu.no.fine.grained.memory !1
ret void
}
define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_agent_scope_unsafe(ptr addrspace(1) %ptr) #0 {
; GFX7LESS-LABEL: global_atomic_fmin_double_uni_address_div_value_agent_scope_unsafe:
; GFX7LESS: ; %bb.0:
-; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-NEXT: s_mov_b32 s32, 0
+; GFX7LESS-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX7LESS-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX7LESS-NEXT: s_mov_b32 s38, -1
+; GFX7LESS-NEXT: s_mov_b32 s39, 0xe8f000
+; GFX7LESS-NEXT: s_add_u32 s36, s36, s11
+; GFX7LESS-NEXT: s_addc_u32 s37, s37, 0
+; GFX7LESS-NEXT: s_mov_b32 s14, s10
+; GFX7LESS-NEXT: s_mov_b32 s13, s9
+; GFX7LESS-NEXT: s_mov_b32 s12, s8
+; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX7LESS-NEXT: s_add_u32 s8, s34, 44
+; GFX7LESS-NEXT: s_addc_u32 s9, s35, 0
+; GFX7LESS-NEXT: s_getpc_b64 s[4:5]
+; GFX7LESS-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX7LESS-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
+; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GFX7LESS-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX7LESS-NEXT: v_or_b32_e32 v40, v0, v2
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX7LESS-NEXT: v_or_b32_e32 v31, v0, v2
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[38:39]
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX7LESS-NEXT: s_mov_b64 s[0:1], exec
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
+; GFX7LESS-NEXT: v_mov_b32_e32 v5, 0x7ff80000
; GFX7LESS-NEXT: .LBB7_1: ; %ComputeLoop
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-NEXT: s_ff1_i32_b64 s4, s[0:1]
-; GFX7LESS-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX7LESS-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
; GFX7LESS-NEXT: v_readlane_b32 s3, v1, s4
; GFX7LESS-NEXT: v_readlane_b32 s2, v0, s4
; GFX7LESS-NEXT: s_lshl_b64 s[4:5], 1, s4
@@ -4067,7 +3512,7 @@ define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_agent
; GFX7LESS-NEXT: s_andn2_b64 s[0:1], s[0:1], s[4:5]
; GFX7LESS-NEXT: v_cmp_ne_u64_e64 s[2:3], s[0:1], 0
; GFX7LESS-NEXT: s_and_b64 vcc, exec, s[2:3]
-; GFX7LESS-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
+; GFX7LESS-NEXT: v_min_f64 v[4:5], v[2:3], v[4:5]
; GFX7LESS-NEXT: s_cbranch_vccnz .LBB7_1
; GFX7LESS-NEXT: ; %bb.2: ; %ComputeEnd
; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
@@ -4077,109 +3522,77 @@ define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_agent
; GFX7LESS-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX7LESS-NEXT: s_cbranch_execz .LBB7_5
; GFX7LESS-NEXT: ; %bb.3:
-; GFX7LESS-NEXT: s_load_dwordx2 s[52:53], s[36:37], 0x9
-; GFX7LESS-NEXT: s_mov_b32 s55, 0xf000
-; GFX7LESS-NEXT: s_mov_b32 s54, -1
+; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x9
+; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
+; GFX7LESS-NEXT: s_mov_b32 s2, -1
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: buffer_load_dwordx2 v[0:1], off, s[52:55], 0
-; GFX7LESS-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
+; GFX7LESS-NEXT: buffer_load_dwordx2 v[2:3], off, s[0:3], 0
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
; GFX7LESS-NEXT: .LBB7_4: ; %atomicrmw.start
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_min_f64 v[0:1], v[2:3], v[41:42]
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7LESS-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[0:3], 0 glc
+; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-NEXT: s_cbranch_execnz .LBB7_4
; GFX7LESS-NEXT: .LBB7_5:
; GFX7LESS-NEXT: s_endpgm
;
; GFX9-LABEL: global_atomic_fmin_double_uni_address_div_value_agent_scope_unsafe:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-NEXT: s_mov_b32 s66, -1
-; GFX9-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-NEXT: s_add_u32 s64, s64, s11
-; GFX9-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-NEXT: s_mov_b32 s51, s8
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_mov_b32 s50, s9
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX9-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX9-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX9-NEXT: s_mov_b32 s38, -1
+; GFX9-NEXT: s_mov_b32 s39, 0xe00000
+; GFX9-NEXT: s_add_u32 s36, s36, s11
+; GFX9-NEXT: s_addc_u32 s37, s37, 0
+; GFX9-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX9-NEXT: s_mov_b32 s12, s8
+; GFX9-NEXT: s_add_u32 s8, s34, 44
+; GFX9-NEXT: s_mov_b32 s13, s9
+; GFX9-NEXT: s_addc_u32 s9, s35, 0
+; GFX9-NEXT: s_getpc_b64 s[4:5]
+; GFX9-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX9-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
+; GFX9-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
+; GFX9-NEXT: s_mov_b32 s14, s10
+; GFX9-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-NEXT: s_mov_b32 s33, s10
-; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX9-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: s_movk_i32 s32, 0x800
+; GFX9-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX9-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX9-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX9-NEXT: s_mov_b32 s32, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: v_mov_b32_e32 v2, 0
+; GFX9-NEXT: v_mov_b32_e32 v4, 0
; GFX9-NEXT: s_mov_b64 s[0:1], exec
-; GFX9-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX9-NEXT: v_mov_b32_e32 v5, 0x7ff80000
; GFX9-NEXT: .LBB7_1: ; %ComputeLoop
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-NEXT: s_ff1_i32_b64 s4, s[0:1]
; GFX9-NEXT: v_readlane_b32 s3, v1, s4
; GFX9-NEXT: v_readlane_b32 s2, v0, s4
-; GFX9-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX9-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
; GFX9-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
; GFX9-NEXT: s_lshl_b64 s[2:3], 1, s4
; GFX9-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
; GFX9-NEXT: s_cmp_lg_u64 s[0:1], 0
-; GFX9-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
+; GFX9-NEXT: v_min_f64 v[4:5], v[2:3], v[4:5]
; GFX9-NEXT: s_cbranch_scc1 .LBB7_1
; GFX9-NEXT: ; %bb.2: ; %ComputeEnd
; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
@@ -4189,90 +3602,55 @@ define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_agent
; GFX9-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX9-NEXT: s_cbranch_execz .LBB7_5
; GFX9-NEXT: ; %bb.3:
-; GFX9-NEXT: s_load_dwordx2 s[52:53], s[36:37], 0x24
-; GFX9-NEXT: v_mov_b32_e32 v0, 0
-; GFX9-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
-; GFX9-NEXT: s_mov_b64 s[54:55], 0
+; GFX9-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX9-NEXT: v_mov_b32_e32 v6, 0
+; GFX9-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX9-NEXT: s_mov_b64 s[2:3], 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: global_load_dwordx2 v[4:5], v0, s[52:53]
+; GFX9-NEXT: global_load_dwordx2 v[2:3], v6, s[0:1]
; GFX9-NEXT: .LBB7_4: ; %atomicrmw.start
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: buffer_store_dword v5, off, s[64:67], 0 offset:4
-; GFX9-NEXT: buffer_store_dword v4, off, s[64:67], 0
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: v_min_f64 v[0:1], v[0:1], v[41:42]
-; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX9-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: buffer_load_dword v4, off, s[64:67], 0
-; GFX9-NEXT: buffer_load_dword v5, off, s[64:67], 0 offset:4
-; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX9-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
+; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-NEXT: s_cbranch_execnz .LBB7_4
; GFX9-NEXT: .LBB7_5:
; GFX9-NEXT: s_endpgm
;
; GFX1064-LABEL: global_atomic_fmin_double_uni_address_div_value_agent_scope_unsafe:
; GFX1064: ; %bb.0:
-; GFX1064-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-NEXT: s_mov_b32 s66, -1
-; GFX1064-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-NEXT: s_add_u32 s64, s64, s11
+; GFX1064-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1064-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1064-NEXT: s_mov_b32 s38, -1
+; GFX1064-NEXT: s_mov_b32 s39, 0x31e16000
+; GFX1064-NEXT: s_add_u32 s36, s36, s11
; GFX1064-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1064-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-NEXT: s_mov_b32 s51, s8
+; GFX1064-NEXT: s_addc_u32 s37, s37, 0
+; GFX1064-NEXT: s_mov_b32 s12, s8
; GFX1064-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-NEXT: s_mov_b32 s50, s9
+; GFX1064-NEXT: s_mov_b32 s13, s9
; GFX1064-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1064-NEXT: s_getpc_b64 s[4:5]
+; GFX1064-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1064-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1064-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1064-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1064-NEXT: s_mov_b32 s33, s10
-; GFX1064-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1064-NEXT: s_mov_b32 s14, s10
+; GFX1064-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1064-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1064-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1064-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: s_movk_i32 s32, 0x800
+; GFX1064-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1064-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1064-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1064-NEXT: s_mov_b32 s32, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1064-NEXT: v_mov_b32_e32 v2, 0
@@ -4296,93 +3674,42 @@ define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_agent
; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX1064-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
-; GFX1064-NEXT: s_cbranch_execz .LBB7_5
+; GFX1064-NEXT: s_cbranch_execz .LBB7_4
; GFX1064-NEXT: ; %bb.3:
-; GFX1064-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
+; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1064-NEXT: v_mov_b32_e32 v0, 0
-; GFX1064-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
-; GFX1064-NEXT: s_mov_b64 s[54:55], 0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: global_load_dwordx2 v[4:5], v0, s[52:53]
-; GFX1064-NEXT: .LBB7_4: ; %atomicrmw.start
-; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1064-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX1064-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-NEXT: buffer_store_dword v5, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: buffer_store_dword v4, off, s[64:67], 0
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: v_min_f64 v[0:1], v[0:1], v[41:42]
-; GFX1064-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX1064-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX1064-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-NEXT: v_mov_b32_e32 v1, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: s_clause 0x1
-; GFX1064-NEXT: buffer_load_dword v4, off, s[64:67], 0
-; GFX1064-NEXT: buffer_load_dword v5, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-NEXT: s_andn2_b64 exec, exec, s[54:55]
-; GFX1064-NEXT: s_cbranch_execnz .LBB7_4
-; GFX1064-NEXT: .LBB7_5:
+; GFX1064-NEXT: global_atomic_fmin_x2 v0, v[2:3], s[0:1]
+; GFX1064-NEXT: .LBB7_4:
; GFX1064-NEXT: s_endpgm
;
; GFX1032-LABEL: global_atomic_fmin_double_uni_address_div_value_agent_scope_unsafe:
; GFX1032: ; %bb.0:
-; GFX1032-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-NEXT: s_mov_b32 s66, -1
-; GFX1032-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-NEXT: s_add_u32 s64, s64, s11
+; GFX1032-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1032-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1032-NEXT: s_mov_b32 s38, -1
+; GFX1032-NEXT: s_mov_b32 s39, 0x31c16000
+; GFX1032-NEXT: s_add_u32 s36, s36, s11
; GFX1032-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1032-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-NEXT: s_mov_b32 s51, s8
+; GFX1032-NEXT: s_addc_u32 s37, s37, 0
+; GFX1032-NEXT: s_mov_b32 s12, s8
; GFX1032-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-NEXT: s_mov_b32 s50, s9
+; GFX1032-NEXT: s_mov_b32 s13, s9
; GFX1032-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1032-NEXT: s_getpc_b64 s[4:5]
+; GFX1032-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1032-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1032-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1032-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1032-NEXT: s_mov_b32 s33, s10
-; GFX1032-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-NEXT: s_mov_b32 s14, s10
+; GFX1032-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1032-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1032-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1032-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1032-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1032-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1032-NEXT: s_mov_b32 s32, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1032-NEXT: v_mov_b32_e32 v2, 0
@@ -4402,96 +3729,45 @@ define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_agent
; GFX1032-NEXT: s_cbranch_scc1 .LBB7_1
; GFX1032-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1032-NEXT: s_mov_b32 s54, 0
; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX1032-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX1032-NEXT: s_cbranch_execz .LBB7_5
+; GFX1032-NEXT: s_cbranch_execz .LBB7_4
; GFX1032-NEXT: ; %bb.3:
-; GFX1032-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
+; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1032-NEXT: v_mov_b32_e32 v0, 0
-; GFX1032-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: global_load_dwordx2 v[4:5], v0, s[52:53]
-; GFX1032-NEXT: .LBB7_4: ; %atomicrmw.start
-; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1032-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX1032-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-NEXT: buffer_store_dword v5, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: buffer_store_dword v4, off, s[64:67], 0
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: v_min_f64 v[0:1], v[0:1], v[41:42]
-; GFX1032-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX1032-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX1032-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-NEXT: v_mov_b32_e32 v1, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: s_clause 0x1
-; GFX1032-NEXT: buffer_load_dword v4, off, s[64:67], 0
-; GFX1032-NEXT: buffer_load_dword v5, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
-; GFX1032-NEXT: s_cbranch_execnz .LBB7_4
-; GFX1032-NEXT: .LBB7_5:
+; GFX1032-NEXT: global_atomic_fmin_x2 v0, v[2:3], s[0:1]
+; GFX1032-NEXT: .LBB7_4:
; GFX1032-NEXT: s_endpgm
;
; GFX1164-LABEL: global_atomic_fmin_double_uni_address_div_value_agent_scope_unsafe:
; GFX1164: ; %bb.0:
; GFX1164-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1164-NEXT: s_mov_b32 s51, s8
+; GFX1164-NEXT: s_mov_b32 s12, s8
; GFX1164-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-NEXT: s_mov_b32 s50, s9
+; GFX1164-NEXT: s_mov_b32 s13, s9
; GFX1164-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1164-NEXT: s_getpc_b64 s[4:5]
+; GFX1164-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1164-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1164-NEXT: v_mov_b32_e32 v31, v0
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: s_mov_b32 s33, s10
-; GFX1164-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX1164-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1164-NEXT: s_mov_b32 s14, s10
+; GFX1164-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1164-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1164-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: s_mov_b32 s32, 32
-; GFX1164-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1164-NEXT: s_mov_b32 s32, 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: v_mov_b32_e32 v2, 0
-; GFX1164-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX1164-NEXT: s_swappc_b64 s[30:31], s[16:17]
+; GFX1164-NEXT: v_mov_b32_e32 v4, 0
+; GFX1164-NEXT: v_mov_b32_e32 v5, 0x7ff80000
; GFX1164-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-NEXT: .LBB7_1: ; %ComputeLoop
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1164-NEXT: s_ctz_i32_b64 s4, s[0:1]
-; GFX1164-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX1164-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
; GFX1164-NEXT: v_readlane_b32 s3, v1, s4
; GFX1164-NEXT: v_readlane_b32 s2, v0, s4
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
@@ -4500,7 +3776,7 @@ define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_agent
; GFX1164-NEXT: s_and_not1_b64 s[0:1], s[0:1], s[2:3]
; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1164-NEXT: s_cmp_lg_u64 s[0:1], 0
-; GFX1164-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-NEXT: v_min_f64 v[4:5], v[2:3], v[4:5]
; GFX1164-NEXT: s_cbranch_scc1 .LBB7_1
; GFX1164-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
@@ -4511,89 +3787,57 @@ define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_agent
; GFX1164-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX1164-NEXT: s_cbranch_execz .LBB7_5
; GFX1164-NEXT: ; %bb.3:
-; GFX1164-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1164-NEXT: v_mov_b32_e32 v0, 0
-; GFX1164-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
-; GFX1164-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1164-NEXT: v_mov_b32_e32 v6, 0
+; GFX1164-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: global_load_b64 v[4:5], v0, s[52:53]
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-NEXT: .p2align 6
+; GFX1164-NEXT: global_load_b64 v[2:3], v6, s[0:1]
; GFX1164-NEXT: .LBB7_4: ; %atomicrmw.start
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX1164-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: v_min_f64 v[0:1], v[0:1], v[41:42]
-; GFX1164-NEXT: scratch_store_b64 off, v[4:5], off
-; GFX1164-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-NEXT: scratch_store_b64 off, v[0:1], off offset:8
-; GFX1164-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: scratch_load_b64 v[4:5], off, off
-; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1164-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
+; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1164-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-NEXT: s_cbranch_execnz .LBB7_4
; GFX1164-NEXT: .LBB7_5:
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-NEXT: s_endpgm
;
; GFX1132-LABEL: global_atomic_fmin_double_uni_address_div_value_agent_scope_unsafe:
; GFX1132: ; %bb.0:
; GFX1132-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1132-NEXT: s_mov_b64 s[48:49], s[0:1]
+; GFX1132-NEXT: v_mov_b32_e32 v31, v0
; GFX1132-NEXT: s_add_u32 s8, s34, 44
; GFX1132-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX1132-NEXT: v_mov_b32_e32 v31, v0
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1132-NEXT: s_mov_b32 s50, s14
-; GFX1132-NEXT: s_mov_b32 s51, s13
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
+; GFX1132-NEXT: s_getpc_b64 s[4:5]
+; GFX1132-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1132-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1132-NEXT: s_mov_b32 s12, s13
+; GFX1132-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1132-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1132-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX1132-NEXT: s_mov_b32 s13, s14
; GFX1132-NEXT: s_mov_b32 s14, s15
-; GFX1132-NEXT: s_mov_b32 s32, 32
-; GFX1132-NEXT: s_mov_b32 s33, s15
-; GFX1132-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-NEXT: s_mov_b32 s32, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: v_mov_b32_e32 v2, 0
-; GFX1132-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX1132-NEXT: s_swappc_b64 s[30:31], s[16:17]
+; GFX1132-NEXT: v_mov_b32_e32 v4, 0
+; GFX1132-NEXT: v_mov_b32_e32 v5, 0x7ff80000
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-NEXT: .LBB7_1: ; %ComputeLoop
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1132-NEXT: s_ctz_i32_b32 s1, s0
-; GFX1132-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX1132-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
; GFX1132-NEXT: v_readlane_b32 s3, v1, s1
; GFX1132-NEXT: v_readlane_b32 s2, v0, s1
; GFX1132-NEXT: s_lshl_b32 s1, 1, s1
@@ -4601,811 +3845,557 @@ define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_agent
; GFX1132-NEXT: s_and_not1_b32 s0, s0, s1
; GFX1132-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
; GFX1132-NEXT: s_cmp_lg_u32 s0, 0
-; GFX1132-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-NEXT: v_min_f64 v[4:5], v[2:3], v[4:5]
; GFX1132-NEXT: s_cbranch_scc1 .LBB7_1
; GFX1132-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-NEXT: s_mov_b32 s54, 0
+; GFX1132-NEXT: s_mov_b32 s2, 0
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX1132-NEXT: s_cbranch_execz .LBB7_5
; GFX1132-NEXT: ; %bb.3:
-; GFX1132-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1132-NEXT: v_mov_b32_e32 v0, 0
+; GFX1132-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1132-NEXT: v_mov_b32_e32 v6, 0
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX1132-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
+; GFX1132-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: global_load_b64 v[4:5], v0, s[52:53]
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-NEXT: .p2align 6
+; GFX1132-NEXT: global_load_b64 v[2:3], v6, s[0:1]
; GFX1132-NEXT: .LBB7_4: ; %atomicrmw.start
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX1132-NEXT: s_add_u32 s8, s34, 44
-; GFX1132-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-NEXT: v_mov_b32_e32 v31, v40
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: v_mov_b32_e32 v3, s53
-; GFX1132-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1132-NEXT: s_mov_b32 s12, s51
-; GFX1132-NEXT: s_mov_b32 s13, s50
-; GFX1132-NEXT: s_mov_b32 s14, s33
-; GFX1132-NEXT: v_mov_b32_e32 v6, 0
-; GFX1132-NEXT: v_mov_b32_e32 v2, s52
-; GFX1132-NEXT: v_min_f64 v[0:1], v[0:1], v[41:42]
-; GFX1132-NEXT: scratch_store_b64 off, v[4:5], off
-; GFX1132-NEXT: v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v5, 8
-; GFX1132-NEXT: scratch_store_b64 off, v[0:1], off offset:8
-; GFX1132-NEXT: v_dual_mov_b32 v0, 8 :: v_dual_mov_b32 v1, 0
-; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: scratch_load_b64 v[4:5], off, off
-; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1132-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
+; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1132-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-NEXT: s_cbranch_execnz .LBB7_4
; GFX1132-NEXT: .LBB7_5:
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-NEXT: s_endpgm
;
; GFX7LESS-DPP-LABEL: global_atomic_fmin_double_uni_address_div_value_agent_scope_unsafe:
; GFX7LESS-DPP: ; %bb.0:
-; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-DPP-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s55, 0xf000
-; GFX7LESS-DPP-NEXT: s_mov_b32 s54, -1
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s32, 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s50, -1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s51, 0xe8f000
+; GFX7LESS-DPP-NEXT: s_add_u32 s48, s48, s11
+; GFX7LESS-DPP-NEXT: s_addc_u32 s49, s49, 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s10
+; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s8
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[36:37], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s39, 0xf000
+; GFX7LESS-DPP-NEXT: s_mov_b32 s38, -1
+; GFX7LESS-DPP-NEXT: s_add_u32 s8, s4, 44
+; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s5, 0
+; GFX7LESS-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX7LESS-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX7LESS-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GFX7LESS-DPP-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v40, v0, v2
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX7LESS-DPP-NEXT: v_or_b32_e32 v31, v0, v2
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[2:3], off, s[52:55], 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-DPP-NEXT: v_max_f64 v[41:42], v[0:1], v[0:1]
+; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[2:3], off, s[36:39], 0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], 0
+; GFX7LESS-DPP-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
; GFX7LESS-DPP-NEXT: .LBB7_1: ; %atomicrmw.start
; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
-; GFX7LESS-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: v_min_f64 v[0:1], v[0:1], v[41:42]
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v3, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX7LESS-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[36:39], 0 glc
+; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-DPP-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[0:1]
; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB7_1
; GFX7LESS-DPP-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX7LESS-DPP-NEXT: s_endpgm
;
; GFX9-DPP-LABEL: global_atomic_fmin_double_uni_address_div_value_agent_scope_unsafe:
; GFX9-DPP: ; %bb.0:
-; GFX9-DPP-NEXT: s_mov_b32 s68, SCRATCH_RSRC_DWORD0
-; GFX9-DPP-NEXT: s_mov_b32 s69, SCRATCH_RSRC_DWORD1
-; GFX9-DPP-NEXT: s_mov_b32 s70, -1
-; GFX9-DPP-NEXT: s_mov_b32 s71, 0xe00000
-; GFX9-DPP-NEXT: s_add_u32 s68, s68, s11
-; GFX9-DPP-NEXT: s_addc_u32 s69, s69, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-DPP-NEXT: s_mov_b32 s51, s8
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_mov_b32 s50, s9
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX9-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX9-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX9-DPP-NEXT: s_mov_b32 s38, -1
+; GFX9-DPP-NEXT: s_mov_b32 s39, 0xe00000
+; GFX9-DPP-NEXT: s_add_u32 s36, s36, s11
+; GFX9-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX9-DPP-NEXT: s_mov_b32 s12, s8
+; GFX9-DPP-NEXT: s_add_u32 s8, s34, 44
+; GFX9-DPP-NEXT: s_mov_b32 s13, s9
+; GFX9-DPP-NEXT: s_addc_u32 s9, s35, 0
+; GFX9-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX9-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX9-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
+; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
+; GFX9-DPP-NEXT: s_mov_b32 s14, s10
+; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-DPP-NEXT: s_mov_b32 s33, s10
-; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[68:69]
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[70:71]
-; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX9-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX9-DPP-NEXT: s_mov_b32 s32, 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
-; GFX9-DPP-NEXT: v_cndmask_b32_e64 v11, v9, v1, s[0:1]
-; GFX9-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:1 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:1 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_min_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
-; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:2 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:2 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_min_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
-; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:4 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:4 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_min_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
-; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:8 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:8 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_min_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
-; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_bcast:15 row_mask:0xa bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_bcast:15 row_mask:0xa bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_min_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: s_nop 1
-; GFX9-DPP-NEXT: v_mov_b32_dpp v9, v11 row_bcast:31 row_mask:0xc bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v10 row_bcast:31 row_mask:0xc bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX9-DPP-NEXT: v_min_f64 v[8:9], v[10:11], v[8:9]
-; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0x7ff80000
+; GFX9-DPP-NEXT: v_cndmask_b32_e64 v6, v4, v1, s[0:1]
+; GFX9-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:1 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:1 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_min_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
+; GFX9-DPP-NEXT: s_nop 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:2 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:2 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_min_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
+; GFX9-DPP-NEXT: s_nop 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:4 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:4 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_min_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
+; GFX9-DPP-NEXT: s_nop 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:8 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:8 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_min_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
+; GFX9-DPP-NEXT: s_nop 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_bcast:15 row_mask:0xa bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_bcast:15 row_mask:0xa bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_min_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: s_nop 1
+; GFX9-DPP-NEXT: v_mov_b32_dpp v4, v6 row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v3, v5 row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[3:4]
+; GFX9-DPP-NEXT: v_min_f64 v[3:4], v[5:6], v[3:4]
+; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 0
; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v1
; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX9-DPP-NEXT: v_readlane_b32 s53, v9, 63
-; GFX9-DPP-NEXT: v_readlane_b32 s52, v8, 63
+; GFX9-DPP-NEXT: v_readlane_b32 s3, v4, 63
+; GFX9-DPP-NEXT: v_readlane_b32 s2, v3, 63
; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[2:3]
; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
+; GFX9-DPP-NEXT: s_and_saveexec_b64 s[2:3], vcc
; GFX9-DPP-NEXT: s_cbranch_execz .LBB7_3
; GFX9-DPP-NEXT: ; %bb.1:
-; GFX9-DPP-NEXT: s_load_dwordx2 s[54:55], s[36:37], 0x24
-; GFX9-DPP-NEXT: s_mov_b64 s[64:65], 0
+; GFX9-DPP-NEXT: s_load_dwordx2 s[2:3], s[34:35], 0x24
+; GFX9-DPP-NEXT: s_mov_b64 s[4:5], 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[54:55]
+; GFX9-DPP-NEXT: global_load_dwordx2 v[11:12], v0, s[2:3]
; GFX9-DPP-NEXT: .LBB7_2: ; %atomicrmw.start
; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX9-DPP-NEXT: v_max_f64 v[3:4], s[52:53], s[52:53]
+; GFX9-DPP-NEXT: v_max_f64 v[1:2], s[0:1], s[0:1]
; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[1:2], v[1:2]
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[68:69]
-; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[68:71], 0 offset:4
-; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[68:71], 0
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: v_min_f64 v[3:4], v[5:6], v[3:4]
-; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[68:71], 0 offset:12
-; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[68:71], 0 offset:8
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[70:71]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s54
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s55
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[68:71], 0
-; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[68:71], 0 offset:4
-; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-DPP-NEXT: s_or_b64 s[64:65], vcc, s[64:65]
-; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[64:65]
+; GFX9-DPP-NEXT: v_max_f64 v[9:10], v[11:12], v[11:12]
+; GFX9-DPP-NEXT: v_min_f64 v[9:10], v[9:10], v[1:2]
+; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[1:2], v0, v[9:12], s[2:3] glc
+; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[1:2], v[11:12]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v12, v2
+; GFX9-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v11, v1
+; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX9-DPP-NEXT: s_cbranch_execnz .LBB7_2
; GFX9-DPP-NEXT: .LBB7_3:
; GFX9-DPP-NEXT: s_endpgm
;
; GFX1064-DPP-LABEL: global_atomic_fmin_double_uni_address_div_value_agent_scope_unsafe:
; GFX1064-DPP: ; %bb.0:
-; GFX1064-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1064-DPP-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-DPP-NEXT: s_add_u32 s64, s64, s11
+; GFX1064-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1064-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1064-DPP-NEXT: s_mov_b32 s38, -1
+; GFX1064-DPP-NEXT: s_mov_b32 s39, 0x31e16000
+; GFX1064-DPP-NEXT: s_add_u32 s36, s36, s11
; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1064-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1064-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX1064-DPP-NEXT: s_mov_b32 s12, s8
; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1064-DPP-NEXT: s_mov_b32 s13, s9
; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1064-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1064-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1064-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1064-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX1064-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1064-DPP-NEXT: s_mov_b32 s32, 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1064-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s[0:1]
-; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX1064-DPP-NEXT: v_min_f64 v[8:9], v[10:11], v[8:9]
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1064-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_readlane_b32 s3, v9, 0
-; GFX1064-DPP-NEXT: v_readlane_b32 s5, v9, 32
-; GFX1064-DPP-NEXT: v_readlane_b32 s4, v8, 32
-; GFX1064-DPP-NEXT: v_readlane_b32 s2, v8, 0
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], s[4:5], s[4:5]
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], s[2:3], s[2:3]
-; GFX1064-DPP-NEXT: v_min_f64 v[8:9], v[10:11], v[8:9]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0x7ff80000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v6, 0x7ff80000, v1, s[0:1]
+; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[3:4]
+; GFX1064-DPP-NEXT: v_min_f64 v[3:4], v[5:6], v[3:4]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
+; GFX1064-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_readlane_b32 s3, v4, 0
+; GFX1064-DPP-NEXT: v_readlane_b32 s5, v4, 32
+; GFX1064-DPP-NEXT: v_readlane_b32 s4, v3, 32
+; GFX1064-DPP-NEXT: v_readlane_b32 s2, v3, 0
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], s[4:5], s[4:5]
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], s[2:3], s[2:3]
+; GFX1064-DPP-NEXT: v_min_f64 v[3:4], v[5:6], v[3:4]
; GFX1064-DPP-NEXT: s_mov_b64 exec, s[0:1]
; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v41, v8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v42, v9
-; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v7, exec_hi, v0
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, v3
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, v4
+; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v7
; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
-; GFX1064-DPP-NEXT: s_cbranch_execz .LBB7_3
+; GFX1064-DPP-NEXT: s_cbranch_execz .LBB7_2
; GFX1064-DPP-NEXT: ; %bb.1:
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1064-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
-; GFX1064-DPP-NEXT: .LBB7_2: ; %atomicrmw.start
-; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[41:42], v[41:42]
-; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[1:2], v[1:2]
-; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: v_min_f64 v[3:4], v[5:6], v[3:4]
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-DPP-NEXT: s_clause 0x1
-; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
-; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB7_2
-; GFX1064-DPP-NEXT: .LBB7_3:
+; GFX1064-DPP-NEXT: global_atomic_fmin_x2 v2, v[0:1], s[0:1]
+; GFX1064-DPP-NEXT: .LBB7_2:
; GFX1064-DPP-NEXT: s_endpgm
;
; GFX1032-DPP-LABEL: global_atomic_fmin_double_uni_address_div_value_agent_scope_unsafe:
; GFX1032-DPP: ; %bb.0:
-; GFX1032-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1032-DPP-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-DPP-NEXT: s_add_u32 s64, s64, s11
+; GFX1032-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1032-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1032-DPP-NEXT: s_mov_b32 s38, -1
+; GFX1032-DPP-NEXT: s_mov_b32 s39, 0x31c16000
+; GFX1032-DPP-NEXT: s_add_u32 s36, s36, s11
; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1032-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1032-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX1032-DPP-NEXT: s_mov_b32 s12, s8
; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1032-DPP-NEXT: s_mov_b32 s13, s9
; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1032-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1032-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1032-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1032-DPP-NEXT: s_mov_b32 s32, 0
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1032-DPP-NEXT: s_or_saveexec_b32 s0, -1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s0
-; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX1032-DPP-NEXT: v_min_f64 v[8:9], v[10:11], v[8:9]
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1032-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0x7ff80000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v6, 0x7ff80000, v1, s0
+; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[3:4]
+; GFX1032-DPP-NEXT: v_min_f64 v[3:4], v[5:6], v[3:4]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
+; GFX1032-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[5:6]
; GFX1032-DPP-NEXT: s_mov_b32 exec_lo, s0
-; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, v8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, v9
-; GFX1032-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
+; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v7, exec_lo, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, v3
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, v4
+; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v7
; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX1032-DPP-NEXT: s_cbranch_execz .LBB7_3
+; GFX1032-DPP-NEXT: s_cbranch_execz .LBB7_2
; GFX1032-DPP-NEXT: ; %bb.1:
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1032-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
-; GFX1032-DPP-NEXT: .LBB7_2: ; %atomicrmw.start
-; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[41:42]
-; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-DPP-NEXT: s_clause 0x1
-; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
-; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB7_2
-; GFX1032-DPP-NEXT: .LBB7_3:
+; GFX1032-DPP-NEXT: global_atomic_fmin_x2 v2, v[0:1], s[0:1]
+; GFX1032-DPP-NEXT: .LBB7_2:
; GFX1032-DPP-NEXT: s_endpgm
;
; GFX1164-DPP-LABEL: global_atomic_fmin_double_uni_address_div_value_agent_scope_unsafe:
; GFX1164-DPP: ; %bb.0:
; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1164-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1164-DPP-NEXT: s_mov_b32 s12, s8
; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1164-DPP-NEXT: s_mov_b32 s13, s9
; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1164-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1164-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1164-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v0
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX1164-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1164-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1164-DPP-NEXT: s_mov_b32 s32, 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
+; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1164-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s[0:1]
-; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v5, 0x7ff80000, v1, s[0:1]
+; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s[0:1]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX1164-DPP-NEXT: v_min_f64 v[8:9], v[10:11], v[8:9]
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX1164-DPP-NEXT: v_min_f64 v[2:3], v[4:5], v[2:3]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 0x7ff80000
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 0x7ff80000
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 0x7ff80000
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1164-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
+; GFX1164-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1164-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_permlane64_b32 v11, v9
-; GFX1164-DPP-NEXT: v_permlane64_b32 v10, v8
+; GFX1164-DPP-NEXT: v_permlane64_b32 v5, v3
+; GFX1164-DPP-NEXT: v_permlane64_b32 v4, v2
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_mov_b64 exec, s[0:1]
-; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, v8
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, v9
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
-; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
+; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v6, exec_hi, v0
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, v2
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, v3
+; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
; GFX1164-DPP-NEXT: s_cbranch_execz .LBB7_3
; GFX1164-DPP-NEXT: ; %bb.1:
-; GFX1164-DPP-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1164-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
-; GFX1164-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX1164-DPP-NEXT: v_max_f64 v[0:1], v[0:1], v[0:1]
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-DPP-NEXT: .p2align 6
+; GFX1164-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
; GFX1164-DPP-NEXT: .LBB7_2: ; %atomicrmw.start
; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[41:42]
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-DPP-NEXT: v_max_f64 v[6:7], v[8:9], v[8:9]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1164-DPP-NEXT: v_min_f64 v[6:7], v[6:7], v[0:1]
+; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
+; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[8:9]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v9, v7
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, v6
+; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB7_2
; GFX1164-DPP-NEXT: .LBB7_3:
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-DPP-NEXT: s_endpgm
;
; GFX1132-DPP-LABEL: global_atomic_fmin_double_uni_address_div_value_agent_scope_unsafe:
; GFX1132-DPP: ; %bb.0:
; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1132-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1132-DPP-NEXT: s_mov_b32 s50, s14
-; GFX1132-DPP-NEXT: s_mov_b32 s51, s13
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
+; GFX1132-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1132-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1132-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1132-DPP-NEXT: s_mov_b32 s12, s13
+; GFX1132-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX1132-DPP-NEXT: s_mov_b32 s13, s14
; GFX1132-DPP-NEXT: s_mov_b32 s14, s15
-; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-DPP-NEXT: s_mov_b32 s32, 0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
+; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1132-DPP-NEXT: s_or_saveexec_b32 s0, -1
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v9, 0x7ff80000 :: v_dual_mov_b32 v8, 0
-; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s0
-; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, 0x7ff80000 :: v_dual_mov_b32 v2, 0
+; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v5, 0x7ff80000, v1, s0
+; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX1132-DPP-NEXT: v_min_f64 v[8:9], v[10:11], v[8:9]
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
+; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX1132-DPP-NEXT: v_min_f64 v[2:3], v[4:5], v[2:3]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 0x7ff80000 :: v_dual_mov_b32 v4, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1132-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1132-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 0x7ff80000 :: v_dual_mov_b32 v4, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1132-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1132-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 0x7ff80000 :: v_dual_mov_b32 v4, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1132-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1132-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1132-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
+; GFX1132-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1132-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1132-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1132-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
; GFX1132-DPP-NEXT: s_mov_b32 exec_lo, s0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v3, v8
-; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, v9
-; GFX1132-DPP-NEXT: s_mov_b32 s54, 0
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v0, v2
+; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v6, exec_lo, 0
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v10, 0 :: v_dual_mov_b32 v1, v3
+; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
+; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
; GFX1132-DPP-NEXT: s_cbranch_execz .LBB7_3
; GFX1132-DPP-NEXT: ; %bb.1:
-; GFX1132-DPP-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1132-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
+; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1132-DPP-NEXT: v_max_f64 v[0:1], v[0:1], v[0:1]
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-DPP-NEXT: .p2align 6
+; GFX1132-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
; GFX1132-DPP-NEXT: .LBB7_2: ; %atomicrmw.start
; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1132-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1132-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX1132-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[41:42]
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-DPP-NEXT: v_max_f64 v[6:7], v[8:9], v[8:9]
+; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1132-DPP-NEXT: v_min_f64 v[6:7], v[6:7], v[0:1]
+; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
+; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[6:7], v[8:9]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6
+; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB7_2
; GFX1132-DPP-NEXT: .LBB7_3:
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-DPP-NEXT: s_endpgm
%divValue = call double @div.double.value()
- %result = atomicrmw fmin ptr addrspace(1) %ptr, double %divValue syncscope("agent") monotonic, align 4, !amdgpu.no.fine.grained.memory !1
+ %result = atomicrmw fmin ptr addrspace(1) %ptr, double %divValue syncscope("agent") monotonic, align 8, !amdgpu.no.fine.grained.memory !1
ret void
}
@@ -6680,932 +5670,377 @@ define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_one_a
define amdgpu_kernel void @global_atomic_fmin_double_uni_address_uni_value_default_scope_unsafe(ptr addrspace(1) %ptr) #0 {
; GFX7LESS-LABEL: global_atomic_fmin_double_uni_address_uni_value_default_scope_unsafe:
; GFX7LESS: ; %bb.0:
-; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v3, exec_lo, 0
-; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v3, exec_hi, v3
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
+; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
+; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX7LESS-NEXT: s_cbranch_execz .LBB10_3
; GFX7LESS-NEXT: ; %bb.1:
-; GFX7LESS-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
+; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX7LESS-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX7LESS-NEXT: v_or_b32_e32 v3, v0, v1
+; GFX7LESS-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, s0
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, s1
-; GFX7LESS-NEXT: v_or_b32_e32 v40, v3, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, s6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, s7
+; GFX7LESS-NEXT: s_mov_b32 s2, -1
; GFX7LESS-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_min_f64 v[0:1], v[2:3], 4.0
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7LESS-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-NEXT: v_mov_b32_e32 v7, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v6, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v5, v1
+; GFX7LESS-NEXT: v_mov_b32_e32 v4, v0
+; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, v4
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, v5
+; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-NEXT: s_cbranch_execnz .LBB10_2
; GFX7LESS-NEXT: .LBB10_3:
; GFX7LESS-NEXT: s_endpgm
;
; GFX9-LABEL: global_atomic_fmin_double_uni_address_uni_value_default_scope_unsafe:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-NEXT: s_mov_b32 s66, -1
-; GFX9-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX9-NEXT: s_add_u32 s64, s64, s11
-; GFX9-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX9-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX9-NEXT: s_movk_i32 s32, 0x800
+; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX9-NEXT: s_cbranch_execz .LBB10_3
; GFX9-NEXT: ; %bb.1:
-; GFX9-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX9-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX9-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX9-NEXT: s_mov_b32 s33, s10
-; GFX9-NEXT: s_mov_b32 s50, s9
+; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9-NEXT: s_mov_b64 s[2:3], 0
+; GFX9-NEXT: v_mov_b32_e32 v4, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX9-NEXT: s_mov_b32 s51, s8
-; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX9-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v2, s1
-; GFX9-NEXT: s_mov_b64 s[54:55], 0
-; GFX9-NEXT: v_mov_b32_e32 v1, s0
-; GFX9-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX9-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-NEXT: v_mov_b32_e32 v3, s5
; GFX9-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX9-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX9-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-NEXT: s_cbranch_execnz .LBB10_2
; GFX9-NEXT: .LBB10_3:
; GFX9-NEXT: s_endpgm
;
; GFX1064-LABEL: global_atomic_fmin_double_uni_address_uni_value_default_scope_unsafe:
; GFX1064: ; %bb.0:
-; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1064-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-NEXT: s_mov_b32 s66, -1
-; GFX1064-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX1064-NEXT: s_add_u32 s64, s64, s11
-; GFX1064-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-NEXT: s_movk_i32 s32, 0x800
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
-; GFX1064-NEXT: s_cbranch_execz .LBB10_3
+; GFX1064-NEXT: s_cbranch_execz .LBB10_2
; GFX1064-NEXT: ; %bb.1:
-; GFX1064-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1064-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1064-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1064-NEXT: s_mov_b32 s33, s10
-; GFX1064-NEXT: s_mov_b32 s50, s9
-; GFX1064-NEXT: s_mov_b32 s51, s8
-; GFX1064-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1064-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1064-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1064-NEXT: s_mov_b64 s[54:55], 0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: v_mov_b32_e32 v2, s1
-; GFX1064-NEXT: v_mov_b32_e32 v1, s0
-; GFX1064-NEXT: .LBB10_2: ; %atomicrmw.start
-; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1064-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1064-NEXT: s_add_u32 s8, s36, 44
-; GFX1064-NEXT: s_addc_u32 s9, s37, 0
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX1064-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-NEXT: v_mov_b32_e32 v4, 0
+; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1064-NEXT: v_mov_b32_e32 v0, 0
+; GFX1064-NEXT: v_mov_b32_e32 v1, 0x40100000
+; GFX1064-NEXT: v_mov_b32_e32 v2, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: s_clause 0x1
-; GFX1064-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-NEXT: s_andn2_b64 exec, exec, s[54:55]
-; GFX1064-NEXT: s_cbranch_execnz .LBB10_2
-; GFX1064-NEXT: .LBB10_3:
+; GFX1064-NEXT: global_atomic_fmin_x2 v2, v[0:1], s[0:1]
+; GFX1064-NEXT: .LBB10_2:
; GFX1064-NEXT: s_endpgm
;
; GFX1032-LABEL: global_atomic_fmin_double_uni_address_uni_value_default_scope_unsafe:
; GFX1032: ; %bb.0:
-; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1032-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-NEXT: s_mov_b32 s66, -1
-; GFX1032-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
-; GFX1032-NEXT: s_add_u32 s64, s64, s11
-; GFX1032-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-NEXT: s_mov_b32 s54, 0
-; GFX1032-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX1032-NEXT: s_cbranch_execz .LBB10_3
+; GFX1032-NEXT: s_cbranch_execz .LBB10_2
; GFX1032-NEXT: ; %bb.1:
-; GFX1032-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1032-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1032-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1032-NEXT: s_mov_b32 s33, s10
-; GFX1032-NEXT: s_mov_b32 s50, s9
-; GFX1032-NEXT: s_mov_b32 s51, s8
-; GFX1032-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1032-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1032-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: v_mov_b32_e32 v2, s1
-; GFX1032-NEXT: v_mov_b32_e32 v1, s0
-; GFX1032-NEXT: .LBB10_2: ; %atomicrmw.start
-; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1032-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1032-NEXT: s_add_u32 s8, s36, 44
-; GFX1032-NEXT: s_addc_u32 s9, s37, 0
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX1032-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-NEXT: v_mov_b32_e32 v4, 0
+; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1032-NEXT: v_mov_b32_e32 v0, 0
+; GFX1032-NEXT: v_mov_b32_e32 v1, 0x40100000
+; GFX1032-NEXT: v_mov_b32_e32 v2, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: s_clause 0x1
-; GFX1032-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
-; GFX1032-NEXT: s_cbranch_execnz .LBB10_2
-; GFX1032-NEXT: .LBB10_3:
+; GFX1032-NEXT: global_atomic_fmin_x2 v2, v[0:1], s[0:1]
+; GFX1032-NEXT: .LBB10_2:
; GFX1032-NEXT: s_endpgm
;
; GFX1164-LABEL: global_atomic_fmin_double_uni_address_uni_value_default_scope_unsafe:
; GFX1164: ; %bb.0:
-; GFX1164-NEXT: v_mov_b32_e32 v40, v0
; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-NEXT: s_mov_b32 s32, 32
; GFX1164-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1164-NEXT: s_cbranch_execz .LBB10_3
; GFX1164-NEXT: ; %bb.1:
-; GFX1164-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1164-NEXT: s_mov_b32 s33, s10
-; GFX1164-NEXT: s_mov_b32 s50, s9
-; GFX1164-NEXT: s_mov_b32 s51, s8
-; GFX1164-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1164-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1164-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1164-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1164-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: v_mov_b32_e32 v2, s1
-; GFX1164-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-NEXT: .p2align 6
+; GFX1164-NEXT: v_mov_b32_e32 v2, s2
+; GFX1164-NEXT: v_mov_b32_e32 v3, s3
+; GFX1164-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1164-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1164-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
; GFX1164-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1164-NEXT: s_add_u32 s8, s36, 44
-; GFX1164-NEXT: s_addc_u32 s9, s37, 0
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX1164-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-NEXT: s_cbranch_execnz .LBB10_2
; GFX1164-NEXT: .LBB10_3:
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-NEXT: s_endpgm
;
; GFX1132-LABEL: global_atomic_fmin_double_uni_address_uni_value_default_scope_unsafe:
; GFX1132: ; %bb.0:
-; GFX1132-NEXT: v_mov_b32_e32 v40, v0
; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1132-NEXT: s_mov_b32 s54, 0
-; GFX1132-NEXT: s_mov_b32 s32, 32
+; GFX1132-NEXT: s_mov_b32 s2, 0
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
+; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-NEXT: s_cbranch_execz .LBB10_3
; GFX1132-NEXT: ; %bb.1:
-; GFX1132-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1132-NEXT: s_mov_b32 s33, s15
-; GFX1132-NEXT: s_mov_b32 s50, s14
-; GFX1132-NEXT: s_mov_b32 s51, s13
-; GFX1132-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1132-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1132-NEXT: v_mov_b32_e32 v4, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1132-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-NEXT: .p2align 6
+; GFX1132-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX1132-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1132-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1132-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
; GFX1132-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1132-NEXT: s_add_u32 s8, s36, 44
-; GFX1132-NEXT: s_addc_u32 s9, s37, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1132-NEXT: s_mov_b32 s12, s51
-; GFX1132-NEXT: s_mov_b32 s13, s50
-; GFX1132-NEXT: s_mov_b32 s14, s33
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX1132-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX1132-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-NEXT: s_cbranch_execnz .LBB10_2
; GFX1132-NEXT: .LBB10_3:
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-NEXT: s_endpgm
;
; GFX7LESS-DPP-LABEL: global_atomic_fmin_double_uni_address_uni_value_default_scope_unsafe:
; GFX7LESS-DPP: ; %bb.0:
-; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v3, exec_lo, 0
-; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v3, exec_hi, v3
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
+; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v0, exec_hi, v0
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX7LESS-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX7LESS-DPP-NEXT: ; %bb.1:
-; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-DPP-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v3, v0, v1
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s3, 0xf000
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, s0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, s1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v40, v3, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s7
+; GFX7LESS-DPP-NEXT: s_mov_b32 s2, -1
; GFX7LESS-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-DPP-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_min_f64 v[0:1], v[2:3], 4.0
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7LESS-DPP-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, v1
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, v0
+; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[4:7], off, s[0:3], 0 glc
+; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
+; GFX7LESS-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v4
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v5
+; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX7LESS-DPP-NEXT: .LBB10_3:
; GFX7LESS-DPP-NEXT: s_endpgm
;
; GFX9-DPP-LABEL: global_atomic_fmin_double_uni_address_uni_value_default_scope_unsafe:
; GFX9-DPP: ; %bb.0:
-; GFX9-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-DPP-NEXT: s_mov_b32 s66, -1
-; GFX9-DPP-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX9-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX9-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX9-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX9-DPP-NEXT: ; %bb.1:
-; GFX9-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX9-DPP-NEXT: s_mov_b32 s33, s10
-; GFX9-DPP-NEXT: s_mov_b32 s50, s9
+; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX9-DPP-NEXT: s_mov_b32 s51, s8
-; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX9-DPP-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX9-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s5
; GFX9-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX9-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX9-DPP-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[0:1], v4, v[0:3], s[0:1] glc
; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX9-DPP-NEXT: .LBB10_3:
; GFX9-DPP-NEXT: s_endpgm
;
; GFX1064-DPP-LABEL: global_atomic_fmin_double_uni_address_uni_value_default_scope_unsafe:
; GFX1064-DPP: ; %bb.0:
-; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1064-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1064-DPP-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX1064-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX1064-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
-; GFX1064-DPP-NEXT: s_cbranch_execz .LBB10_3
+; GFX1064-DPP-NEXT: s_cbranch_execz .LBB10_2
; GFX1064-DPP-NEXT: ; %bb.1:
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1064-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1064-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1064-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1064-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
-; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1064-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1064-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0x40100000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-DPP-NEXT: s_clause 0x1
-; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
-; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB10_2
-; GFX1064-DPP-NEXT: .LBB10_3:
+; GFX1064-DPP-NEXT: global_atomic_fmin_x2 v2, v[0:1], s[0:1]
+; GFX1064-DPP-NEXT: .LBB10_2:
; GFX1064-DPP-NEXT: s_endpgm
;
; GFX1032-DPP-LABEL: global_atomic_fmin_double_uni_address_uni_value_default_scope_unsafe:
; GFX1032-DPP: ; %bb.0:
-; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1032-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1032-DPP-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
-; GFX1032-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX1032-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
+; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX1032-DPP-NEXT: s_cbranch_execz .LBB10_3
+; GFX1032-DPP-NEXT: s_cbranch_execz .LBB10_2
; GFX1032-DPP-NEXT: ; %bb.1:
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1032-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1032-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1032-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
-; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1032-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1032-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0x40100000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, 0
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-DPP-NEXT: s_clause 0x1
-; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
-; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB10_2
-; GFX1032-DPP-NEXT: .LBB10_3:
+; GFX1032-DPP-NEXT: global_atomic_fmin_x2 v2, v[0:1], s[0:1]
+; GFX1032-DPP-NEXT: .LBB10_2:
; GFX1032-DPP-NEXT: s_endpgm
;
; GFX1164-DPP-LABEL: global_atomic_fmin_double_uni_address_uni_value_default_scope_unsafe:
; GFX1164-DPP: ; %bb.0:
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1164-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX1164-DPP-NEXT: ; %bb.1:
-; GFX1164-DPP-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1164-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1164-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1164-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1164-DPP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-DPP-NEXT: .p2align 6
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s2
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s3
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1164-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1164-DPP-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1164-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1164-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX1164-DPP-NEXT: .LBB10_3:
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-DPP-NEXT: s_endpgm
;
; GFX1132-DPP-LABEL: global_atomic_fmin_double_uni_address_uni_value_default_scope_unsafe:
; GFX1132-DPP: ; %bb.0:
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1132-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
+; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
+; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX1132-DPP-NEXT: ; %bb.1:
-; GFX1132-DPP-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
-; GFX1132-DPP-NEXT: s_mov_b32 s50, s14
-; GFX1132-DPP-NEXT: s_mov_b32 s51, s13
-; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1132-DPP-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-DPP-NEXT: .p2align 6
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX1132-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1132-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1132-DPP-NEXT: v_min_f64 v[0:1], v[0:1], 4.0
+; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v4, v[0:3], s[0:1] glc
; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1132-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1132-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1132-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1132-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX1132-DPP-NEXT: v_min_f64 v[3:4], v[3:4], 4.0
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX1132-DPP-NEXT: .LBB10_3:
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-DPP-NEXT: s_endpgm
- %result = atomicrmw fmin ptr addrspace(1) %ptr, double 4.0 monotonic, align 4, !amdgpu.no.fine.grained.memory !1
+ %result = atomicrmw fmin ptr addrspace(1) %ptr, double 4.0 monotonic, align 8, !amdgpu.no.fine.grained.memory !1
ret void
}
define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_default_scope_unsafe(ptr addrspace(1) %ptr) #0 {
; GFX7LESS-LABEL: global_atomic_fmin_double_uni_address_div_value_default_scope_unsafe:
; GFX7LESS: ; %bb.0:
-; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-NEXT: s_mov_b32 s32, 0
+; GFX7LESS-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX7LESS-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX7LESS-NEXT: s_mov_b32 s38, -1
+; GFX7LESS-NEXT: s_mov_b32 s39, 0xe8f000
+; GFX7LESS-NEXT: s_add_u32 s36, s36, s11
+; GFX7LESS-NEXT: s_addc_u32 s37, s37, 0
+; GFX7LESS-NEXT: s_mov_b32 s14, s10
+; GFX7LESS-NEXT: s_mov_b32 s13, s9
+; GFX7LESS-NEXT: s_mov_b32 s12, s8
+; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX7LESS-NEXT: s_add_u32 s8, s34, 44
+; GFX7LESS-NEXT: s_addc_u32 s9, s35, 0
+; GFX7LESS-NEXT: s_getpc_b64 s[4:5]
+; GFX7LESS-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX7LESS-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
+; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GFX7LESS-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX7LESS-NEXT: v_or_b32_e32 v40, v0, v2
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX7LESS-NEXT: v_or_b32_e32 v31, v0, v2
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[38:39]
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX7LESS-NEXT: s_mov_b64 s[0:1], exec
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
+; GFX7LESS-NEXT: v_mov_b32_e32 v5, 0x7ff80000
; GFX7LESS-NEXT: .LBB11_1: ; %ComputeLoop
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-NEXT: s_ff1_i32_b64 s4, s[0:1]
-; GFX7LESS-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX7LESS-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
; GFX7LESS-NEXT: v_readlane_b32 s3, v1, s4
; GFX7LESS-NEXT: v_readlane_b32 s2, v0, s4
; GFX7LESS-NEXT: s_lshl_b64 s[4:5], 1, s4
@@ -7613,7 +6048,7 @@ define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_defau
; GFX7LESS-NEXT: s_andn2_b64 s[0:1], s[0:1], s[4:5]
; GFX7LESS-NEXT: v_cmp_ne_u64_e64 s[2:3], s[0:1], 0
; GFX7LESS-NEXT: s_and_b64 vcc, exec, s[2:3]
-; GFX7LESS-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
+; GFX7LESS-NEXT: v_min_f64 v[4:5], v[2:3], v[4:5]
; GFX7LESS-NEXT: s_cbranch_vccnz .LBB11_1
; GFX7LESS-NEXT: ; %bb.2: ; %ComputeEnd
; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
@@ -7623,109 +6058,77 @@ define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_defau
; GFX7LESS-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX7LESS-NEXT: s_cbranch_execz .LBB11_5
; GFX7LESS-NEXT: ; %bb.3:
-; GFX7LESS-NEXT: s_load_dwordx2 s[52:53], s[36:37], 0x9
-; GFX7LESS-NEXT: s_mov_b32 s55, 0xf000
-; GFX7LESS-NEXT: s_mov_b32 s54, -1
+; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x9
+; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
+; GFX7LESS-NEXT: s_mov_b32 s2, -1
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: buffer_load_dwordx2 v[0:1], off, s[52:55], 0
-; GFX7LESS-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
+; GFX7LESS-NEXT: buffer_load_dwordx2 v[2:3], off, s[0:3], 0
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
; GFX7LESS-NEXT: .LBB11_4: ; %atomicrmw.start
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-NEXT: v_max_f64 v[2:3], v[0:1], v[0:1]
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_min_f64 v[0:1], v[2:3], v[41:42]
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX7LESS-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[0:3], 0 glc
+; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-NEXT: s_cbranch_execnz .LBB11_4
; GFX7LESS-NEXT: .LBB11_5:
; GFX7LESS-NEXT: s_endpgm
;
; GFX9-LABEL: global_atomic_fmin_double_uni_address_div_value_default_scope_unsafe:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-NEXT: s_mov_b32 s66, -1
-; GFX9-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-NEXT: s_add_u32 s64, s64, s11
-; GFX9-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-NEXT: s_mov_b32 s51, s8
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_mov_b32 s50, s9
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX9-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX9-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX9-NEXT: s_mov_b32 s38, -1
+; GFX9-NEXT: s_mov_b32 s39, 0xe00000
+; GFX9-NEXT: s_add_u32 s36, s36, s11
+; GFX9-NEXT: s_addc_u32 s37, s37, 0
+; GFX9-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX9-NEXT: s_mov_b32 s12, s8
+; GFX9-NEXT: s_add_u32 s8, s34, 44
+; GFX9-NEXT: s_mov_b32 s13, s9
+; GFX9-NEXT: s_addc_u32 s9, s35, 0
+; GFX9-NEXT: s_getpc_b64 s[4:5]
+; GFX9-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX9-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
+; GFX9-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
+; GFX9-NEXT: s_mov_b32 s14, s10
+; GFX9-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-NEXT: s_mov_b32 s33, s10
-; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX9-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: s_movk_i32 s32, 0x800
+; GFX9-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX9-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX9-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX9-NEXT: s_mov_b32 s32, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: v_mov_b32_e32 v2, 0
+; GFX9-NEXT: v_mov_b32_e32 v4, 0
; GFX9-NEXT: s_mov_b64 s[0:1], exec
-; GFX9-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX9-NEXT: v_mov_b32_e32 v5, 0x7ff80000
; GFX9-NEXT: .LBB11_1: ; %ComputeLoop
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-NEXT: s_ff1_i32_b64 s4, s[0:1]
; GFX9-NEXT: v_readlane_b32 s3, v1, s4
; GFX9-NEXT: v_readlane_b32 s2, v0, s4
-; GFX9-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX9-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
; GFX9-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
; GFX9-NEXT: s_lshl_b64 s[2:3], 1, s4
; GFX9-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
; GFX9-NEXT: s_cmp_lg_u64 s[0:1], 0
-; GFX9-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
+; GFX9-NEXT: v_min_f64 v[4:5], v[2:3], v[4:5]
; GFX9-NEXT: s_cbranch_scc1 .LBB11_1
; GFX9-NEXT: ; %bb.2: ; %ComputeEnd
; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
@@ -7735,90 +6138,55 @@ define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_defau
; GFX9-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX9-NEXT: s_cbranch_execz .LBB11_5
; GFX9-NEXT: ; %bb.3:
-; GFX9-NEXT: s_load_dwordx2 s[52:53], s[36:37], 0x24
-; GFX9-NEXT: v_mov_b32_e32 v0, 0
-; GFX9-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
-; GFX9-NEXT: s_mov_b64 s[54:55], 0
+; GFX9-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX9-NEXT: v_mov_b32_e32 v6, 0
+; GFX9-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX9-NEXT: s_mov_b64 s[2:3], 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: global_load_dwordx2 v[4:5], v0, s[52:53]
+; GFX9-NEXT: global_load_dwordx2 v[2:3], v6, s[0:1]
; GFX9-NEXT: .LBB11_4: ; %atomicrmw.start
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: buffer_store_dword v5, off, s[64:67], 0 offset:4
-; GFX9-NEXT: buffer_store_dword v4, off, s[64:67], 0
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: v_min_f64 v[0:1], v[0:1], v[41:42]
-; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX9-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: buffer_load_dword v4, off, s[64:67], 0
-; GFX9-NEXT: buffer_load_dword v5, off, s[64:67], 0 offset:4
-; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX9-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
+; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-NEXT: s_cbranch_execnz .LBB11_4
; GFX9-NEXT: .LBB11_5:
; GFX9-NEXT: s_endpgm
;
; GFX1064-LABEL: global_atomic_fmin_double_uni_address_div_value_default_scope_unsafe:
; GFX1064: ; %bb.0:
-; GFX1064-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-NEXT: s_mov_b32 s66, -1
-; GFX1064-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-NEXT: s_add_u32 s64, s64, s11
+; GFX1064-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1064-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1064-NEXT: s_mov_b32 s38, -1
+; GFX1064-NEXT: s_mov_b32 s39, 0x31e16000
+; GFX1064-NEXT: s_add_u32 s36, s36, s11
; GFX1064-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1064-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-NEXT: s_mov_b32 s51, s8
+; GFX1064-NEXT: s_addc_u32 s37, s37, 0
+; GFX1064-NEXT: s_mov_b32 s12, s8
; GFX1064-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-NEXT: s_mov_b32 s50, s9
+; GFX1064-NEXT: s_mov_b32 s13, s9
; GFX1064-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1064-NEXT: s_getpc_b64 s[4:5]
+; GFX1064-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1064-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1064-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1064-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1064-NEXT: s_mov_b32 s33, s10
-; GFX1064-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1064-NEXT: s_mov_b32 s14, s10
+; GFX1064-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1064-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1064-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1064-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: s_movk_i32 s32, 0x800
+; GFX1064-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1064-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1064-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1064-NEXT: s_mov_b32 s32, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1064-NEXT: v_mov_b32_e32 v2, 0
@@ -7842,93 +6210,42 @@ define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_defau
; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX1064-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
-; GFX1064-NEXT: s_cbranch_execz .LBB11_5
+; GFX1064-NEXT: s_cbranch_execz .LBB11_4
; GFX1064-NEXT: ; %bb.3:
-; GFX1064-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
+; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1064-NEXT: v_mov_b32_e32 v0, 0
-; GFX1064-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
-; GFX1064-NEXT: s_mov_b64 s[54:55], 0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: global_load_dwordx2 v[4:5], v0, s[52:53]
-; GFX1064-NEXT: .LBB11_4: ; %atomicrmw.start
-; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1064-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX1064-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-NEXT: buffer_store_dword v5, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: buffer_store_dword v4, off, s[64:67], 0
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: v_min_f64 v[0:1], v[0:1], v[41:42]
-; GFX1064-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX1064-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX1064-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-NEXT: v_mov_b32_e32 v1, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: s_clause 0x1
-; GFX1064-NEXT: buffer_load_dword v4, off, s[64:67], 0
-; GFX1064-NEXT: buffer_load_dword v5, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-NEXT: s_andn2_b64 exec, exec, s[54:55]
-; GFX1064-NEXT: s_cbranch_execnz .LBB11_4
-; GFX1064-NEXT: .LBB11_5:
+; GFX1064-NEXT: global_atomic_fmin_x2 v0, v[2:3], s[0:1]
+; GFX1064-NEXT: .LBB11_4:
; GFX1064-NEXT: s_endpgm
;
; GFX1032-LABEL: global_atomic_fmin_double_uni_address_div_value_default_scope_unsafe:
; GFX1032: ; %bb.0:
-; GFX1032-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-NEXT: s_mov_b32 s66, -1
-; GFX1032-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-NEXT: s_add_u32 s64, s64, s11
+; GFX1032-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1032-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1032-NEXT: s_mov_b32 s38, -1
+; GFX1032-NEXT: s_mov_b32 s39, 0x31c16000
+; GFX1032-NEXT: s_add_u32 s36, s36, s11
; GFX1032-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1032-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-NEXT: s_mov_b32 s51, s8
+; GFX1032-NEXT: s_addc_u32 s37, s37, 0
+; GFX1032-NEXT: s_mov_b32 s12, s8
; GFX1032-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-NEXT: s_mov_b32 s50, s9
+; GFX1032-NEXT: s_mov_b32 s13, s9
; GFX1032-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1032-NEXT: s_getpc_b64 s[4:5]
+; GFX1032-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1032-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1032-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1032-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1032-NEXT: s_mov_b32 s33, s10
-; GFX1032-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-NEXT: s_mov_b32 s14, s10
+; GFX1032-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1032-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1032-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1032-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1032-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1032-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1032-NEXT: s_mov_b32 s32, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1032-NEXT: v_mov_b32_e32 v2, 0
@@ -7948,96 +6265,45 @@ define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_defau
; GFX1032-NEXT: s_cbranch_scc1 .LBB11_1
; GFX1032-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1032-NEXT: s_mov_b32 s54, 0
; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX1032-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX1032-NEXT: s_cbranch_execz .LBB11_5
+; GFX1032-NEXT: s_cbranch_execz .LBB11_4
; GFX1032-NEXT: ; %bb.3:
-; GFX1032-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
+; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1032-NEXT: v_mov_b32_e32 v0, 0
-; GFX1032-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: global_load_dwordx2 v[4:5], v0, s[52:53]
-; GFX1032-NEXT: .LBB11_4: ; %atomicrmw.start
-; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1032-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX1032-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-NEXT: buffer_store_dword v5, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: buffer_store_dword v4, off, s[64:67], 0
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: v_min_f64 v[0:1], v[0:1], v[41:42]
-; GFX1032-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX1032-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX1032-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-NEXT: v_mov_b32_e32 v1, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: s_clause 0x1
-; GFX1032-NEXT: buffer_load_dword v4, off, s[64:67], 0
-; GFX1032-NEXT: buffer_load_dword v5, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
-; GFX1032-NEXT: s_cbranch_execnz .LBB11_4
-; GFX1032-NEXT: .LBB11_5:
+; GFX1032-NEXT: global_atomic_fmin_x2 v0, v[2:3], s[0:1]
+; GFX1032-NEXT: .LBB11_4:
; GFX1032-NEXT: s_endpgm
;
; GFX1164-LABEL: global_atomic_fmin_double_uni_address_div_value_default_scope_unsafe:
; GFX1164: ; %bb.0:
; GFX1164-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1164-NEXT: s_mov_b32 s51, s8
+; GFX1164-NEXT: s_mov_b32 s12, s8
; GFX1164-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-NEXT: s_mov_b32 s50, s9
+; GFX1164-NEXT: s_mov_b32 s13, s9
; GFX1164-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1164-NEXT: s_getpc_b64 s[4:5]
+; GFX1164-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1164-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1164-NEXT: v_mov_b32_e32 v31, v0
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: s_mov_b32 s33, s10
-; GFX1164-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX1164-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1164-NEXT: s_mov_b32 s14, s10
+; GFX1164-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1164-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1164-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: s_mov_b32 s32, 32
-; GFX1164-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1164-NEXT: s_mov_b32 s32, 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: v_mov_b32_e32 v2, 0
-; GFX1164-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX1164-NEXT: s_swappc_b64 s[30:31], s[16:17]
+; GFX1164-NEXT: v_mov_b32_e32 v4, 0
+; GFX1164-NEXT: v_mov_b32_e32 v5, 0x7ff80000
; GFX1164-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-NEXT: .LBB11_1: ; %ComputeLoop
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1164-NEXT: s_ctz_i32_b64 s4, s[0:1]
-; GFX1164-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX1164-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
; GFX1164-NEXT: v_readlane_b32 s3, v1, s4
; GFX1164-NEXT: v_readlane_b32 s2, v0, s4
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
@@ -8046,7 +6312,7 @@ define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_defau
; GFX1164-NEXT: s_and_not1_b64 s[0:1], s[0:1], s[2:3]
; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1164-NEXT: s_cmp_lg_u64 s[0:1], 0
-; GFX1164-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-NEXT: v_min_f64 v[4:5], v[2:3], v[4:5]
; GFX1164-NEXT: s_cbranch_scc1 .LBB11_1
; GFX1164-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
@@ -8057,89 +6323,57 @@ define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_defau
; GFX1164-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX1164-NEXT: s_cbranch_execz .LBB11_5
; GFX1164-NEXT: ; %bb.3:
-; GFX1164-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1164-NEXT: v_mov_b32_e32 v0, 0
-; GFX1164-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
-; GFX1164-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1164-NEXT: v_mov_b32_e32 v6, 0
+; GFX1164-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: global_load_b64 v[4:5], v0, s[52:53]
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-NEXT: .p2align 6
+; GFX1164-NEXT: global_load_b64 v[2:3], v6, s[0:1]
; GFX1164-NEXT: .LBB11_4: ; %atomicrmw.start
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX1164-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: v_min_f64 v[0:1], v[0:1], v[41:42]
-; GFX1164-NEXT: scratch_store_b64 off, v[4:5], off
-; GFX1164-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-NEXT: scratch_store_b64 off, v[0:1], off offset:8
-; GFX1164-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: scratch_load_b64 v[4:5], off, off
-; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1164-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
+; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1164-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-NEXT: s_cbranch_execnz .LBB11_4
; GFX1164-NEXT: .LBB11_5:
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-NEXT: s_endpgm
;
; GFX1132-LABEL: global_atomic_fmin_double_uni_address_div_value_default_scope_unsafe:
; GFX1132: ; %bb.0:
; GFX1132-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1132-NEXT: s_mov_b64 s[48:49], s[0:1]
+; GFX1132-NEXT: v_mov_b32_e32 v31, v0
; GFX1132-NEXT: s_add_u32 s8, s34, 44
; GFX1132-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX1132-NEXT: v_mov_b32_e32 v31, v0
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1132-NEXT: s_mov_b32 s50, s14
-; GFX1132-NEXT: s_mov_b32 s51, s13
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
+; GFX1132-NEXT: s_getpc_b64 s[4:5]
+; GFX1132-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1132-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1132-NEXT: s_mov_b32 s12, s13
+; GFX1132-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1132-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1132-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX1132-NEXT: s_mov_b32 s13, s14
; GFX1132-NEXT: s_mov_b32 s14, s15
-; GFX1132-NEXT: s_mov_b32 s32, 32
-; GFX1132-NEXT: s_mov_b32 s33, s15
-; GFX1132-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-NEXT: s_mov_b32 s32, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: v_mov_b32_e32 v2, 0
-; GFX1132-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX1132-NEXT: s_swappc_b64 s[30:31], s[16:17]
+; GFX1132-NEXT: v_mov_b32_e32 v4, 0
+; GFX1132-NEXT: v_mov_b32_e32 v5, 0x7ff80000
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-NEXT: .LBB11_1: ; %ComputeLoop
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1132-NEXT: s_ctz_i32_b32 s1, s0
-; GFX1132-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX1132-NEXT: v_max_f64 v[2:3], v[4:5], v[4:5]
; GFX1132-NEXT: v_readlane_b32 s3, v1, s1
; GFX1132-NEXT: v_readlane_b32 s2, v0, s1
; GFX1132-NEXT: s_lshl_b32 s1, 1, s1
@@ -8147,811 +6381,557 @@ define amdgpu_kernel void @global_atomic_fmin_double_uni_address_div_value_defau
; GFX1132-NEXT: s_and_not1_b32 s0, s0, s1
; GFX1132-NEXT: v_max_f64 v[4:5], s[2:3], s[2:3]
; GFX1132-NEXT: s_cmp_lg_u32 s0, 0
-; GFX1132-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-NEXT: v_min_f64 v[4:5], v[2:3], v[4:5]
; GFX1132-NEXT: s_cbranch_scc1 .LBB11_1
; GFX1132-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-NEXT: s_mov_b32 s54, 0
+; GFX1132-NEXT: s_mov_b32 s2, 0
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX1132-NEXT: s_cbranch_execz .LBB11_5
; GFX1132-NEXT: ; %bb.3:
-; GFX1132-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1132-NEXT: v_mov_b32_e32 v0, 0
+; GFX1132-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1132-NEXT: v_mov_b32_e32 v6, 0
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX1132-NEXT: v_max_f64 v[41:42], v[2:3], v[2:3]
+; GFX1132-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: global_load_b64 v[4:5], v0, s[52:53]
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-NEXT: .p2align 6
+; GFX1132-NEXT: global_load_b64 v[2:3], v6, s[0:1]
; GFX1132-NEXT: .LBB11_4: ; %atomicrmw.start
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-NEXT: v_max_f64 v[0:1], v[4:5], v[4:5]
-; GFX1132-NEXT: s_add_u32 s8, s34, 44
-; GFX1132-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-NEXT: v_mov_b32_e32 v31, v40
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: v_mov_b32_e32 v3, s53
-; GFX1132-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1132-NEXT: s_mov_b32 s12, s51
-; GFX1132-NEXT: s_mov_b32 s13, s50
-; GFX1132-NEXT: s_mov_b32 s14, s33
-; GFX1132-NEXT: v_mov_b32_e32 v6, 0
-; GFX1132-NEXT: v_mov_b32_e32 v2, s52
-; GFX1132-NEXT: v_min_f64 v[0:1], v[0:1], v[41:42]
-; GFX1132-NEXT: scratch_store_b64 off, v[4:5], off
-; GFX1132-NEXT: v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v5, 8
-; GFX1132-NEXT: scratch_store_b64 off, v[0:1], off offset:8
-; GFX1132-NEXT: v_dual_mov_b32 v0, 8 :: v_dual_mov_b32 v1, 0
-; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: scratch_load_b64 v[4:5], off, off
-; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
+; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1132-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
+; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1132-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-NEXT: s_cbranch_execnz .LBB11_4
; GFX1132-NEXT: .LBB11_5:
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-NEXT: s_endpgm
;
; GFX7LESS-DPP-LABEL: global_atomic_fmin_double_uni_address_div_value_default_scope_unsafe:
; GFX7LESS-DPP: ; %bb.0:
-; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-DPP-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s55, 0xf000
-; GFX7LESS-DPP-NEXT: s_mov_b32 s54, -1
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s32, 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s50, -1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s51, 0xe8f000
+; GFX7LESS-DPP-NEXT: s_add_u32 s48, s48, s11
+; GFX7LESS-DPP-NEXT: s_addc_u32 s49, s49, 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s10
+; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s8
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[36:37], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s39, 0xf000
+; GFX7LESS-DPP-NEXT: s_mov_b32 s38, -1
+; GFX7LESS-DPP-NEXT: s_add_u32 s8, s4, 44
+; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s5, 0
+; GFX7LESS-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX7LESS-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX7LESS-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GFX7LESS-DPP-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v40, v0, v2
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX7LESS-DPP-NEXT: v_or_b32_e32 v31, v0, v2
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[2:3], off, s[52:55], 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-DPP-NEXT: v_max_f64 v[41:42], v[0:1], v[0:1]
+; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[2:3], off, s[36:39], 0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], 0
+; GFX7LESS-DPP-NEXT: v_max_f64 v[4:5], v[0:1], v[0:1]
; GFX7LESS-DPP-NEXT: .LBB11_1: ; %atomicrmw.start
; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
; GFX7LESS-DPP-NEXT: v_max_f64 v[0:1], v[2:3], v[2:3]
-; GFX7LESS-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: v_min_f64 v[0:1], v[0:1], v[41:42]
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: v_min_f64 v[0:1], v[0:1], v[4:5]
; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v3, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX7LESS-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[36:39], 0 glc
+; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-DPP-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[0:1]
; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB11_1
; GFX7LESS-DPP-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX7LESS-DPP-NEXT: s_endpgm
;
; GFX9-DPP-LABEL: global_atomic_fmin_double_uni_address_div_value_default_scope_unsafe:
; GFX9-DPP: ; %bb.0:
-; GFX9-DPP-NEXT: s_mov_b32 s68, SCRATCH_RSRC_DWORD0
-; GFX9-DPP-NEXT: s_mov_b32 s69, SCRATCH_RSRC_DWORD1
-; GFX9-DPP-NEXT: s_mov_b32 s70, -1
-; GFX9-DPP-NEXT: s_mov_b32 s71, 0xe00000
-; GFX9-DPP-NEXT: s_add_u32 s68, s68, s11
-; GFX9-DPP-NEXT: s_addc_u32 s69, s69, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-DPP-NEXT: s_mov_b32 s51, s8
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_mov_b32 s50, s9
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX9-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX9-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX9-DPP-NEXT: s_mov_b32 s38, -1
+; GFX9-DPP-NEXT: s_mov_b32 s39, 0xe00000
+; GFX9-DPP-NEXT: s_add_u32 s36, s36, s11
+; GFX9-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX9-DPP-NEXT: s_mov_b32 s12, s8
+; GFX9-DPP-NEXT: s_add_u32 s8, s34, 44
+; GFX9-DPP-NEXT: s_mov_b32 s13, s9
+; GFX9-DPP-NEXT: s_addc_u32 s9, s35, 0
+; GFX9-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX9-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX9-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
+; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
+; GFX9-DPP-NEXT: s_mov_b32 s14, s10
+; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-DPP-NEXT: s_mov_b32 s33, s10
-; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[68:69]
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[70:71]
-; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX9-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX9-DPP-NEXT: s_mov_b32 s32, 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
-; GFX9-DPP-NEXT: v_cndmask_b32_e64 v11, v9, v1, s[0:1]
-; GFX9-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:1 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:1 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_min_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0x7ff80000
+; GFX9-DPP-NEXT: v_cndmask_b32_e64 v6, v4, v1, s[0:1]
+; GFX9-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:1 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:1 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_min_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:2 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:2 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_min_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:2 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:2 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_min_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:4 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:4 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_min_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:4 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:4 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_min_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:8 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:8 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_min_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v13, 0x7ff80000
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:8 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:8 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_min_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0x7ff80000
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_bcast:15 row_mask:0xa bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_bcast:15 row_mask:0xa bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[12:13], v[12:13], v[12:13]
-; GFX9-DPP-NEXT: v_min_f64 v[10:11], v[10:11], v[12:13]
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_bcast:15 row_mask:0xa bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_bcast:15 row_mask:0xa bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[7:8], v[7:8], v[7:8]
+; GFX9-DPP-NEXT: v_min_f64 v[5:6], v[5:6], v[7:8]
; GFX9-DPP-NEXT: s_nop 1
-; GFX9-DPP-NEXT: v_mov_b32_dpp v9, v11 row_bcast:31 row_mask:0xc bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v10 row_bcast:31 row_mask:0xc bank_mask:0xf
-; GFX9-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX9-DPP-NEXT: v_min_f64 v[8:9], v[10:11], v[8:9]
+; GFX9-DPP-NEXT: v_mov_b32_dpp v4, v6 row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v3, v5 row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[3:4]
+; GFX9-DPP-NEXT: v_min_f64 v[3:4], v[5:6], v[3:4]
; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 0
; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v1
; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX9-DPP-NEXT: v_readlane_b32 s53, v9, 63
-; GFX9-DPP-NEXT: v_readlane_b32 s52, v8, 63
+; GFX9-DPP-NEXT: v_readlane_b32 s3, v4, 63
+; GFX9-DPP-NEXT: v_readlane_b32 s2, v3, 63
; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[2:3]
; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
+; GFX9-DPP-NEXT: s_and_saveexec_b64 s[2:3], vcc
; GFX9-DPP-NEXT: s_cbranch_execz .LBB11_3
; GFX9-DPP-NEXT: ; %bb.1:
-; GFX9-DPP-NEXT: s_load_dwordx2 s[54:55], s[36:37], 0x24
-; GFX9-DPP-NEXT: s_mov_b64 s[64:65], 0
+; GFX9-DPP-NEXT: s_load_dwordx2 s[2:3], s[34:35], 0x24
+; GFX9-DPP-NEXT: s_mov_b64 s[4:5], 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[54:55]
+; GFX9-DPP-NEXT: global_load_dwordx2 v[11:12], v0, s[2:3]
; GFX9-DPP-NEXT: .LBB11_2: ; %atomicrmw.start
; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX9-DPP-NEXT: v_max_f64 v[3:4], s[52:53], s[52:53]
+; GFX9-DPP-NEXT: v_max_f64 v[1:2], s[0:1], s[0:1]
; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX9-DPP-NEXT: v_max_f64 v[5:6], v[1:2], v[1:2]
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[68:69]
-; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[68:71], 0 offset:4
-; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[68:71], 0
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: v_min_f64 v[3:4], v[5:6], v[3:4]
-; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[68:71], 0 offset:12
-; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[68:71], 0 offset:8
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[70:71]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s54
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s55
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[68:71], 0
-; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[68:71], 0 offset:4
-; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-DPP-NEXT: s_or_b64 s[64:65], vcc, s[64:65]
-; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[64:65]
+; GFX9-DPP-NEXT: v_max_f64 v[9:10], v[11:12], v[11:12]
+; GFX9-DPP-NEXT: v_min_f64 v[9:10], v[9:10], v[1:2]
+; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[1:2], v0, v[9:12], s[2:3] glc
+; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[1:2], v[11:12]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v12, v2
+; GFX9-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v11, v1
+; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX9-DPP-NEXT: s_cbranch_execnz .LBB11_2
; GFX9-DPP-NEXT: .LBB11_3:
; GFX9-DPP-NEXT: s_endpgm
;
; GFX1064-DPP-LABEL: global_atomic_fmin_double_uni_address_div_value_default_scope_unsafe:
; GFX1064-DPP: ; %bb.0:
-; GFX1064-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1064-DPP-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-DPP-NEXT: s_add_u32 s64, s64, s11
+; GFX1064-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1064-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1064-DPP-NEXT: s_mov_b32 s38, -1
+; GFX1064-DPP-NEXT: s_mov_b32 s39, 0x31e16000
+; GFX1064-DPP-NEXT: s_add_u32 s36, s36, s11
; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1064-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1064-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX1064-DPP-NEXT: s_mov_b32 s12, s8
; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1064-DPP-NEXT: s_mov_b32 s13, s9
; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1064-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1064-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1064-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1064-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX1064-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1064-DPP-NEXT: s_mov_b32 s32, 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1064-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s[0:1]
-; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX1064-DPP-NEXT: v_min_f64 v[8:9], v[10:11], v[8:9]
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1064-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1064-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_readlane_b32 s3, v9, 0
-; GFX1064-DPP-NEXT: v_readlane_b32 s5, v9, 32
-; GFX1064-DPP-NEXT: v_readlane_b32 s4, v8, 32
-; GFX1064-DPP-NEXT: v_readlane_b32 s2, v8, 0
-; GFX1064-DPP-NEXT: v_max_f64 v[8:9], s[4:5], s[4:5]
-; GFX1064-DPP-NEXT: v_max_f64 v[10:11], s[2:3], s[2:3]
-; GFX1064-DPP-NEXT: v_min_f64 v[8:9], v[10:11], v[8:9]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0x7ff80000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v6, 0x7ff80000, v1, s[0:1]
+; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[3:4]
+; GFX1064-DPP-NEXT: v_min_f64 v[3:4], v[5:6], v[3:4]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
+; GFX1064-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1064-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_readlane_b32 s3, v4, 0
+; GFX1064-DPP-NEXT: v_readlane_b32 s5, v4, 32
+; GFX1064-DPP-NEXT: v_readlane_b32 s4, v3, 32
+; GFX1064-DPP-NEXT: v_readlane_b32 s2, v3, 0
+; GFX1064-DPP-NEXT: v_max_f64 v[3:4], s[4:5], s[4:5]
+; GFX1064-DPP-NEXT: v_max_f64 v[5:6], s[2:3], s[2:3]
+; GFX1064-DPP-NEXT: v_min_f64 v[3:4], v[5:6], v[3:4]
; GFX1064-DPP-NEXT: s_mov_b64 exec, s[0:1]
; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v41, v8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v42, v9
-; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v7, exec_hi, v0
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, v3
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, v4
+; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v7
; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
-; GFX1064-DPP-NEXT: s_cbranch_execz .LBB11_3
+; GFX1064-DPP-NEXT: s_cbranch_execz .LBB11_2
; GFX1064-DPP-NEXT: ; %bb.1:
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1064-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
-; GFX1064-DPP-NEXT: .LBB11_2: ; %atomicrmw.start
-; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1064-DPP-NEXT: v_max_f64 v[3:4], v[41:42], v[41:42]
-; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-DPP-NEXT: v_max_f64 v[5:6], v[1:2], v[1:2]
-; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: v_min_f64 v[3:4], v[5:6], v[3:4]
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-DPP-NEXT: s_clause 0x1
-; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
-; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB11_2
-; GFX1064-DPP-NEXT: .LBB11_3:
+; GFX1064-DPP-NEXT: global_atomic_fmin_x2 v2, v[0:1], s[0:1]
+; GFX1064-DPP-NEXT: .LBB11_2:
; GFX1064-DPP-NEXT: s_endpgm
;
; GFX1032-DPP-LABEL: global_atomic_fmin_double_uni_address_div_value_default_scope_unsafe:
; GFX1032-DPP: ; %bb.0:
-; GFX1032-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1032-DPP-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-DPP-NEXT: s_add_u32 s64, s64, s11
+; GFX1032-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1032-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1032-DPP-NEXT: s_mov_b32 s38, -1
+; GFX1032-DPP-NEXT: s_mov_b32 s39, 0x31c16000
+; GFX1032-DPP-NEXT: s_add_u32 s36, s36, s11
; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1032-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1032-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX1032-DPP-NEXT: s_mov_b32 s12, s8
; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1032-DPP-NEXT: s_mov_b32 s13, s9
; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1032-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1032-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1032-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1032-DPP-NEXT: s_mov_b32 s32, 0
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1032-DPP-NEXT: s_or_saveexec_b32 s0, -1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s0
-; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX1032-DPP-NEXT: v_min_f64 v[8:9], v[10:11], v[8:9]
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1032-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
-; GFX1032-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1032-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0x7ff80000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v6, 0x7ff80000, v1, s0
+; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[3:4], v[3:4]
+; GFX1032-DPP-NEXT: v_min_f64 v[3:4], v[5:6], v[3:4]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0x7ff80000
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
+; GFX1032-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1032-DPP-NEXT: v_max_f64 v[5:6], v[5:6], v[5:6]
+; GFX1032-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[5:6]
; GFX1032-DPP-NEXT: s_mov_b32 exec_lo, s0
-; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, v8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, v9
-; GFX1032-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
+; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v7, exec_lo, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, v3
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, v4
+; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v7
; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
-; GFX1032-DPP-NEXT: s_cbranch_execz .LBB11_3
+; GFX1032-DPP-NEXT: s_cbranch_execz .LBB11_2
; GFX1032-DPP-NEXT: ; %bb.1:
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1032-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
-; GFX1032-DPP-NEXT: .LBB11_2: ; %atomicrmw.start
-; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[41:42]
-; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-DPP-NEXT: s_clause 0x1
-; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
-; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB11_2
-; GFX1032-DPP-NEXT: .LBB11_3:
+; GFX1032-DPP-NEXT: global_atomic_fmin_x2 v2, v[0:1], s[0:1]
+; GFX1032-DPP-NEXT: .LBB11_2:
; GFX1032-DPP-NEXT: s_endpgm
;
; GFX1164-DPP-LABEL: global_atomic_fmin_double_uni_address_div_value_default_scope_unsafe:
; GFX1164-DPP: ; %bb.0:
; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1164-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1164-DPP-NEXT: s_mov_b32 s12, s8
; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1164-DPP-NEXT: s_mov_b32 s13, s9
; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
+; GFX1164-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1164-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1164-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v0
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX1164-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1164-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1164-DPP-NEXT: s_mov_b32 s32, 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
+; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1164-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v9, 0x7ff80000
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s[0:1]
-; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, 0x7ff80000
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v5, 0x7ff80000, v1, s[0:1]
+; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s[0:1]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX1164-DPP-NEXT: v_min_f64 v[8:9], v[10:11], v[8:9]
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1164-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX1164-DPP-NEXT: v_min_f64 v[2:3], v[4:5], v[2:3]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 0x7ff80000
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 0x7ff80000
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v11, 0x7ff80000
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 0x7ff80000
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1164-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
+; GFX1164-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1164-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_permlane64_b32 v11, v9
-; GFX1164-DPP-NEXT: v_permlane64_b32 v10, v8
+; GFX1164-DPP-NEXT: v_permlane64_b32 v5, v3
+; GFX1164-DPP-NEXT: v_permlane64_b32 v4, v2
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1164-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1164-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_mov_b64 exec, s[0:1]
-; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, v8
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, v9
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
-; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
+; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v6, exec_hi, v0
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, v2
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, v3
+; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
; GFX1164-DPP-NEXT: s_cbranch_execz .LBB11_3
; GFX1164-DPP-NEXT: ; %bb.1:
-; GFX1164-DPP-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1164-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
-; GFX1164-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX1164-DPP-NEXT: v_max_f64 v[0:1], v[0:1], v[0:1]
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-DPP-NEXT: .p2align 6
+; GFX1164-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
; GFX1164-DPP-NEXT: .LBB11_2: ; %atomicrmw.start
; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[41:42]
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-DPP-NEXT: v_max_f64 v[6:7], v[8:9], v[8:9]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1164-DPP-NEXT: v_min_f64 v[6:7], v[6:7], v[0:1]
+; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
+; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[8:9]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v9, v7
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, v6
+; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB11_2
; GFX1164-DPP-NEXT: .LBB11_3:
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-DPP-NEXT: s_endpgm
;
; GFX1132-DPP-LABEL: global_atomic_fmin_double_uni_address_div_value_default_scope_unsafe:
; GFX1132-DPP: ; %bb.0:
; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1132-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, div.double.value at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, div.double.value at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1132-DPP-NEXT: s_mov_b32 s50, s14
-; GFX1132-DPP-NEXT: s_mov_b32 s51, s13
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
+; GFX1132-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1132-DPP-NEXT: s_add_u32 s4, s4, div.double.value at gotpcrel32@lo+4
+; GFX1132-DPP-NEXT: s_addc_u32 s5, s5, div.double.value at gotpcrel32@hi+12
; GFX1132-DPP-NEXT: s_mov_b32 s12, s13
+; GFX1132-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX1132-DPP-NEXT: s_mov_b32 s13, s14
; GFX1132-DPP-NEXT: s_mov_b32 s14, s15
-; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-DPP-NEXT: s_mov_b32 s32, 0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
+; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1132-DPP-NEXT: s_or_saveexec_b32 s0, -1
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v9, 0x7ff80000 :: v_dual_mov_b32 v8, 0
-; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v11, 0x7ff80000, v1, s0
-; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, 0x7ff80000 :: v_dual_mov_b32 v2, 0
+; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v5, 0x7ff80000, v1, s0
+; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[8:9], v[8:9], v[8:9]
-; GFX1132-DPP-NEXT: v_min_f64 v[8:9], v[10:11], v[8:9]
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
+; GFX1132-DPP-NEXT: v_max_f64 v[2:3], v[2:3], v[2:3]
+; GFX1132-DPP-NEXT: v_min_f64 v[2:3], v[4:5], v[2:3]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 0x7ff80000 :: v_dual_mov_b32 v4, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1132-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1132-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 0x7ff80000 :: v_dual_mov_b32 v4, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1132-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v11, 0x7ff80000 :: v_dual_mov_b32 v10, 0
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1132-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 0x7ff80000 :: v_dual_mov_b32 v4, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1132-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1132-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1132-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
+; GFX1132-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1132-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_max_f64 v[10:11], v[10:11], v[10:11]
-; GFX1132-DPP-NEXT: v_min_f64 v[8:9], v[8:9], v[10:11]
+; GFX1132-DPP-NEXT: v_max_f64 v[4:5], v[4:5], v[4:5]
+; GFX1132-DPP-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5]
; GFX1132-DPP-NEXT: s_mov_b32 exec_lo, s0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v3, v8
-; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, v9
-; GFX1132-DPP-NEXT: s_mov_b32 s54, 0
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v0, v2
+; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v6, exec_lo, 0
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v10, 0 :: v_dual_mov_b32 v1, v3
+; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
+; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
; GFX1132-DPP-NEXT: s_cbranch_execz .LBB11_3
; GFX1132-DPP-NEXT: ; %bb.1:
-; GFX1132-DPP-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1132-DPP-NEXT: v_max_f64 v[41:42], v[3:4], v[3:4]
+; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1132-DPP-NEXT: v_max_f64 v[0:1], v[0:1], v[0:1]
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-DPP-NEXT: .p2align 6
+; GFX1132-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
; GFX1132-DPP-NEXT: .LBB11_2: ; %atomicrmw.start
; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-DPP-NEXT: v_max_f64 v[3:4], v[1:2], v[1:2]
-; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1132-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1132-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX1132-DPP-NEXT: v_min_f64 v[3:4], v[3:4], v[41:42]
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-DPP-NEXT: v_max_f64 v[6:7], v[8:9], v[8:9]
+; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1132-DPP-NEXT: v_min_f64 v[6:7], v[6:7], v[0:1]
+; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
+; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[6:7], v[8:9]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6
+; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB11_2
; GFX1132-DPP-NEXT: .LBB11_3:
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-DPP-NEXT: s_endpgm
%divValue = call double @div.double.value()
- %result = atomicrmw fmin ptr addrspace(1) %ptr, double %divValue monotonic, align 4, !amdgpu.no.fine.grained.memory !1
+ %result = atomicrmw fmin ptr addrspace(1) %ptr, double %divValue monotonic, align 8, !amdgpu.no.fine.grained.memory !1
ret void
}
diff --git a/llvm/test/CodeGen/AMDGPU/global_atomics_scan_fsub.ll b/llvm/test/CodeGen/AMDGPU/global_atomics_scan_fsub.ll
index 9db3c37045ccf..bb119ebd11daa 100644
--- a/llvm/test/CodeGen/AMDGPU/global_atomics_scan_fsub.ll
+++ b/llvm/test/CodeGen/AMDGPU/global_atomics_scan_fsub.ll
@@ -6064,966 +6064,469 @@ define amdgpu_kernel void @global_atomic_fsub_uni_address_div_value_default_scop
define amdgpu_kernel void @global_atomic_fsub_double_uni_address_uni_value_agent_scope_unsafe(ptr addrspace(1) %ptr) #0 {
; GFX7LESS-LABEL: global_atomic_fsub_double_uni_address_uni_value_agent_scope_unsafe:
; GFX7LESS: ; %bb.0:
-; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], exec
-; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v3, s0, 0
-; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v3, s1, v3
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX7LESS-NEXT: s_and_saveexec_b64 s[2:3], vcc
+; GFX7LESS-NEXT: s_mov_b64 s[2:3], exec
+; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s2, 0
+; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s3, v0
+; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX7LESS-NEXT: s_cbranch_execz .LBB9_3
; GFX7LESS-NEXT: ; %bb.1:
-; GFX7LESS-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
-; GFX7LESS-NEXT: s_bcnt1_i32_b64 s2, s[0:1]
-; GFX7LESS-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-NEXT: v_lshlrev_b32_e32 v3, 20, v2
+; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
+; GFX7LESS-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX7LESS-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX7LESS-NEXT: v_cvt_f64_u32_e32 v[1:2], s2
-; GFX7LESS-NEXT: v_or_b32_e32 v4, v0, v4
-; GFX7LESS-NEXT: v_mul_f64 v[41:42], v[1:2], 4.0
+; GFX7LESS-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
+; GFX7LESS-NEXT: v_cvt_f64_u32_e32 v[0:1], s2
+; GFX7LESS-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, s0
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, s1
-; GFX7LESS-NEXT: v_or_b32_e32 v40, v4, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, s6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, s7
+; GFX7LESS-NEXT: s_mov_b32 s2, -1
; GFX7LESS-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-NEXT: v_add_f64 v[2:3], v[0:1], -v[41:42]
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:12
-; GFX7LESS-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:8
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX7LESS-NEXT: s_waitcnt expcnt(2)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[0:3], 0 glc
+; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-NEXT: s_cbranch_execnz .LBB9_2
; GFX7LESS-NEXT: .LBB9_3:
; GFX7LESS-NEXT: s_endpgm
;
; GFX9-LABEL: global_atomic_fsub_double_uni_address_uni_value_agent_scope_unsafe:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-NEXT: s_mov_b32 s66, -1
-; GFX9-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-NEXT: s_mov_b64 s[0:1], exec
-; GFX9-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-NEXT: v_mbcnt_lo_u32_b32 v3, s0, 0
-; GFX9-NEXT: s_add_u32 s64, s64, s11
-; GFX9-NEXT: v_mbcnt_hi_u32_b32 v3, s1, v3
-; GFX9-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX9-NEXT: s_movk_i32 s32, 0x800
-; GFX9-NEXT: s_and_saveexec_b64 s[2:3], vcc
+; GFX9-NEXT: s_mov_b64 s[2:3], exec
+; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
+; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
+; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX9-NEXT: s_cbranch_execz .LBB9_3
; GFX9-NEXT: ; %bb.1:
-; GFX9-NEXT: s_bcnt1_i32_b64 s0, s[0:1]
-; GFX9-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX9-NEXT: v_cvt_f64_u32_e32 v[3:4], s0
-; GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-NEXT: s_mov_b32 s33, s10
-; GFX9-NEXT: v_mul_f64 v[41:42], v[3:4], 4.0
+; GFX9-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
+; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9-NEXT: v_cvt_f64_u32_e32 v[0:1], s2
+; GFX9-NEXT: s_mov_b64 s[2:3], 0
+; GFX9-NEXT: v_mov_b32_e32 v6, 0
+; GFX9-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX9-NEXT: s_mov_b32 s50, s9
-; GFX9-NEXT: s_mov_b32 s51, s8
-; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
+; GFX9-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v4, s1
-; GFX9-NEXT: s_mov_b64 s[54:55], 0
-; GFX9-NEXT: v_mov_b32_e32 v3, s0
-; GFX9-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-NEXT: v_mov_b32_e32 v3, s5
; GFX9-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f64 v[0:1], v[3:4], -v[41:42]
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:4
-; GFX9-NEXT: buffer_store_dword v3, off, s[64:67], 0
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX9-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: buffer_load_dword v3, off, s[64:67], 0
-; GFX9-NEXT: buffer_load_dword v4, off, s[64:67], 0 offset:4
-; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-NEXT: s_cbranch_execnz .LBB9_2
; GFX9-NEXT: .LBB9_3:
; GFX9-NEXT: s_endpgm
;
; GFX1064-LABEL: global_atomic_fsub_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1064: ; %bb.0:
-; GFX1064-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-NEXT: s_mov_b32 s66, -1
-; GFX1064-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-NEXT: s_add_u32 s64, s64, s11
-; GFX1064-NEXT: s_mov_b32 s33, s10
-; GFX1064-NEXT: s_mov_b64 s[10:11], exec
-; GFX1064-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v3, s10, 0
-; GFX1064-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-NEXT: s_movk_i32 s32, 0x800
-; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v3, s11, v3
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX1064-NEXT: s_mov_b64 s[2:3], exec
+; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
+; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
+; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX1064-NEXT: s_cbranch_execz .LBB9_3
; GFX1064-NEXT: ; %bb.1:
-; GFX1064-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1064-NEXT: s_bcnt1_i32_b64 s0, s[10:11]
-; GFX1064-NEXT: s_mov_b32 s50, s9
-; GFX1064-NEXT: v_cvt_f64_u32_e32 v[3:4], s0
-; GFX1064-NEXT: s_mov_b32 s51, s8
-; GFX1064-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1064-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1064-NEXT: s_mov_b64 s[54:55], 0
+; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1064-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
+; GFX1064-NEXT: v_mov_b32_e32 v6, 0
+; GFX1064-NEXT: v_cvt_f64_u32_e32 v[0:1], s2
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1064-NEXT: v_mul_f64 v[41:42], v[3:4], 4.0
-; GFX1064-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1064-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1064-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX1064-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX1064-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: v_mov_b32_e32 v2, s1
-; GFX1064-NEXT: v_mov_b32_e32 v1, s0
+; GFX1064-NEXT: v_mov_b32_e32 v2, s2
+; GFX1064-NEXT: v_mov_b32_e32 v3, s3
+; GFX1064-NEXT: s_mov_b64 s[2:3], 0
; GFX1064-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1064-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1064-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX1064-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1064-NEXT: s_add_u32 s8, s36, 44
-; GFX1064-NEXT: s_addc_u32 s9, s37, 0
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: s_clause 0x1
-; GFX1064-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX1064-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1064-NEXT: v_mov_b32_e32 v3, v1
+; GFX1064-NEXT: v_mov_b32_e32 v2, v0
+; GFX1064-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1064-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX1064-NEXT: s_cbranch_execnz .LBB9_2
; GFX1064-NEXT: .LBB9_3:
; GFX1064-NEXT: s_endpgm
;
; GFX1032-LABEL: global_atomic_fsub_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1032: ; %bb.0:
-; GFX1032-NEXT: s_mov_b32 s50, s9
-; GFX1032-NEXT: s_mov_b32 s9, exec_lo
-; GFX1032-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v3, s9, 0
-; GFX1032-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-NEXT: s_mov_b32 s66, -1
-; GFX1032-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-NEXT: s_add_u32 s64, s64, s11
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
-; GFX1032-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-NEXT: s_mov_b32 s54, 0
-; GFX1032-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-NEXT: s_mov_b32 s3, exec_lo
+; GFX1032-NEXT: s_mov_b32 s2, 0
+; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, s3, 0
+; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX1032-NEXT: s_cbranch_execz .LBB9_3
; GFX1032-NEXT: ; %bb.1:
-; GFX1032-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1032-NEXT: s_bcnt1_i32_b32 s0, s9
-; GFX1032-NEXT: s_mov_b32 s33, s10
-; GFX1032-NEXT: v_cvt_f64_u32_e32 v[3:4], s0
-; GFX1032-NEXT: s_mov_b32 s51, s8
-; GFX1032-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1032-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1032-NEXT: s_bcnt1_i32_b32 s3, s3
+; GFX1032-NEXT: v_mov_b32_e32 v6, 0
+; GFX1032-NEXT: v_cvt_f64_u32_e32 v[0:1], s3
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1032-NEXT: v_mul_f64 v[41:42], v[3:4], 4.0
-; GFX1032-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1032-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1032-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX1032-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX1032-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: v_mov_b32_e32 v2, s1
-; GFX1032-NEXT: v_mov_b32_e32 v1, s0
+; GFX1032-NEXT: v_mov_b32_e32 v2, s4
+; GFX1032-NEXT: v_mov_b32_e32 v3, s5
; GFX1032-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1032-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1032-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX1032-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1032-NEXT: s_add_u32 s8, s36, 44
-; GFX1032-NEXT: s_addc_u32 s9, s37, 0
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: s_clause 0x1
-; GFX1032-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
+; GFX1032-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1032-NEXT: v_mov_b32_e32 v3, v1
+; GFX1032-NEXT: v_mov_b32_e32 v2, v0
+; GFX1032-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
; GFX1032-NEXT: s_cbranch_execnz .LBB9_2
; GFX1032-NEXT: .LBB9_3:
; GFX1032-NEXT: s_endpgm
;
; GFX1164-LABEL: global_atomic_fsub_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1164: ; %bb.0:
-; GFX1164-NEXT: s_mov_b32 s33, s10
-; GFX1164-NEXT: s_mov_b64 s[10:11], exec
-; GFX1164-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, s10, 0
-; GFX1164-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-NEXT: s_mov_b32 s32, 32
; GFX1164-NEXT: s_mov_b64 s[0:1], exec
+; GFX1164-NEXT: s_mov_b64 s[2:3], exec
+; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, s11, v0
+; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, s1, v0
; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1164-NEXT: s_cbranch_execz .LBB9_3
; GFX1164-NEXT: ; %bb.1:
-; GFX1164-NEXT: s_bcnt1_i32_b64 s0, s[10:11]
-; GFX1164-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
+; GFX1164-NEXT: s_bcnt1_i32_b64 s0, s[0:1]
+; GFX1164-NEXT: v_mov_b32_e32 v6, 0
; GFX1164-NEXT: v_cvt_f64_u32_e32 v[0:1], s0
-; GFX1164-NEXT: s_mov_b32 s50, s9
-; GFX1164-NEXT: s_mov_b32 s51, s8
-; GFX1164-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1164-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1164-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1164-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-NEXT: v_mul_f64 v[41:42], v[0:1], 4.0
+; GFX1164-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: v_mov_b32_e32 v2, s1
-; GFX1164-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-NEXT: .p2align 6
+; GFX1164-NEXT: v_mov_b32_e32 v2, s2
+; GFX1164-NEXT: v_mov_b32_e32 v3, s3
+; GFX1164-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1164-NEXT: s_waitcnt vmcnt(0)
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1164-NEXT: s_add_u32 s8, s36, 44
-; GFX1164-NEXT: s_addc_u32 s9, s37, 0
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1164-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-NEXT: s_cbranch_execnz .LBB9_2
; GFX1164-NEXT: .LBB9_3:
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-NEXT: s_endpgm
;
; GFX1132-LABEL: global_atomic_fsub_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1132: ; %bb.0:
-; GFX1132-NEXT: s_mov_b32 s8, exec_lo
-; GFX1132-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, s8, 0
-; GFX1132-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1132-NEXT: s_mov_b32 s54, 0
-; GFX1132-NEXT: s_mov_b32 s32, 32
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
+; GFX1132-NEXT: s_mov_b32 s2, 0
+; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
+; GFX1132-NEXT: s_mov_b32 s1, exec_lo
+; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-NEXT: s_cbranch_execz .LBB9_3
; GFX1132-NEXT: ; %bb.1:
-; GFX1132-NEXT: s_bcnt1_i32_b32 s0, s8
-; GFX1132-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
+; GFX1132-NEXT: s_bcnt1_i32_b32 s0, s0
+; GFX1132-NEXT: v_mov_b32_e32 v6, 0
; GFX1132-NEXT: v_cvt_f64_u32_e32 v[0:1], s0
-; GFX1132-NEXT: s_mov_b32 s33, s15
-; GFX1132-NEXT: s_mov_b32 s50, s14
-; GFX1132-NEXT: s_mov_b32 s51, s13
-; GFX1132-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1132-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1132-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-NEXT: v_mul_f64 v[41:42], v[0:1], 4.0
+; GFX1132-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-NEXT: .p2align 6
+; GFX1132-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX1132-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1132-NEXT: s_waitcnt vmcnt(0)
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1132-NEXT: s_add_u32 s8, s36, 44
-; GFX1132-NEXT: s_addc_u32 s9, s37, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1132-NEXT: s_mov_b32 s12, s51
-; GFX1132-NEXT: s_mov_b32 s13, s50
-; GFX1132-NEXT: s_mov_b32 s14, s33
-; GFX1132-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1132-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-NEXT: s_cbranch_execnz .LBB9_2
; GFX1132-NEXT: .LBB9_3:
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-NEXT: s_endpgm
;
; GFX7LESS-DPP-LABEL: global_atomic_fsub_double_uni_address_uni_value_agent_scope_unsafe:
; GFX7LESS-DPP: ; %bb.0:
-; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], exec
-; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v3, s0, 0
-; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v3, s1, v3
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[2:3], vcc
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], exec
+; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s2, 0
+; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s3, v0
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX7LESS-DPP-NEXT: s_cbranch_execz .LBB9_3
; GFX7LESS-DPP-NEXT: ; %bb.1:
-; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-DPP-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
-; GFX7LESS-DPP-NEXT: s_bcnt1_i32_b64 s2, s[0:1]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX7LESS-DPP-NEXT: v_cvt_f64_u32_e32 v[1:2], s2
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v4, v0, v4
-; GFX7LESS-DPP-NEXT: v_mul_f64 v[41:42], v[1:2], 4.0
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[6:7], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s3, 0xf000
+; GFX7LESS-DPP-NEXT: v_cvt_f64_u32_e32 v[0:1], s2
+; GFX7LESS-DPP-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, s0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, s1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v40, v4, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s7
+; GFX7LESS-DPP-NEXT: s_mov_b32 s2, -1
; GFX7LESS-DPP-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-DPP-NEXT: v_add_f64 v[2:3], v[0:1], -v[41:42]
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:8
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(2)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-DPP-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[0:3], 0 glc
+; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB9_2
; GFX7LESS-DPP-NEXT: .LBB9_3:
; GFX7LESS-DPP-NEXT: s_endpgm
;
; GFX9-DPP-LABEL: global_atomic_fsub_double_uni_address_uni_value_agent_scope_unsafe:
; GFX9-DPP: ; %bb.0:
-; GFX9-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-DPP-NEXT: s_mov_b32 s66, -1
-; GFX9-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], exec
-; GFX9-DPP-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, s0, 0
-; GFX9-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, s1, v3
-; GFX9-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX9-DPP-NEXT: s_and_saveexec_b64 s[2:3], vcc
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], exec
+; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
+; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
+; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX9-DPP-NEXT: s_cbranch_execz .LBB9_3
; GFX9-DPP-NEXT: ; %bb.1:
-; GFX9-DPP-NEXT: s_bcnt1_i32_b64 s0, s[0:1]
-; GFX9-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX9-DPP-NEXT: v_cvt_f64_u32_e32 v[3:4], s0
-; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-DPP-NEXT: s_mov_b32 s33, s10
-; GFX9-DPP-NEXT: v_mul_f64 v[41:42], v[3:4], 4.0
+; GFX9-DPP-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
+; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9-DPP-NEXT: v_cvt_f64_u32_e32 v[0:1], s2
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
+; GFX9-DPP-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX9-DPP-NEXT: s_mov_b32 s50, s9
-; GFX9-DPP-NEXT: s_mov_b32 s51, s8
-; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
+; GFX9-DPP-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, s1
-; GFX9-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s0
-; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s5
; GFX9-DPP-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX9-DPP-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX9-DPP-NEXT: v_add_f64 v[0:1], v[3:4], -v[41:42]
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:4
-; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:12
-; GFX9-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0 offset:8
-; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-DPP-NEXT: buffer_load_dword v3, off, s[64:67], 0
-; GFX9-DPP-NEXT: buffer_load_dword v4, off, s[64:67], 0 offset:4
-; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-DPP-NEXT: s_cbranch_execnz .LBB9_2
; GFX9-DPP-NEXT: .LBB9_3:
; GFX9-DPP-NEXT: s_endpgm
;
; GFX1064-DPP-LABEL: global_atomic_fsub_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1064-DPP: ; %bb.0:
-; GFX1064-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1064-DPP-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], exec
-; GFX1064-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, s10, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, s11, v3
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], exec
+; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s2, 0
+; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, s3, v0
+; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX1064-DPP-NEXT: s_cbranch_execz .LBB9_3
; GFX1064-DPP-NEXT: ; %bb.1:
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1064-DPP-NEXT: s_bcnt1_i32_b64 s0, s[10:11]
-; GFX1064-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1064-DPP-NEXT: v_cvt_f64_u32_e32 v[3:4], s0
-; GFX1064-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1064-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1064-DPP-NEXT: s_bcnt1_i32_b64 s2, s[2:3]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
+; GFX1064-DPP-NEXT: v_cvt_f64_u32_e32 v[0:1], s2
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1064-DPP-NEXT: v_mul_f64 v[41:42], v[3:4], 4.0
-; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX1064-DPP-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, s0
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s2
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s3
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1064-DPP-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1064-DPP-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1064-DPP-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1064-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1064-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-DPP-NEXT: s_clause 0x1
-; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX1064-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX1064-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB9_2
; GFX1064-DPP-NEXT: .LBB9_3:
; GFX1064-DPP-NEXT: s_endpgm
;
; GFX1032-DPP-LABEL: global_atomic_fsub_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1032-DPP: ; %bb.0:
-; GFX1032-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1032-DPP-NEXT: s_mov_b32 s9, exec_lo
-; GFX1032-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, s9, 0
-; GFX1032-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1032-DPP-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
-; GFX1032-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-DPP-NEXT: s_mov_b32 s3, exec_lo
+; GFX1032-DPP-NEXT: s_mov_b32 s2, 0
+; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s3, 0
+; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX1032-DPP-NEXT: s_cbranch_execz .LBB9_3
; GFX1032-DPP-NEXT: ; %bb.1:
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1032-DPP-NEXT: s_bcnt1_i32_b32 s0, s9
-; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1032-DPP-NEXT: v_cvt_f64_u32_e32 v[3:4], s0
-; GFX1032-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1032-DPP-NEXT: s_bcnt1_i32_b32 s3, s3
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
+; GFX1032-DPP-NEXT: v_cvt_f64_u32_e32 v[0:1], s3
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX1032-DPP-NEXT: v_mul_f64 v[41:42], v[3:4], 4.0
-; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX1032-DPP-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, s0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s4
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s5
; GFX1032-DPP-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1032-DPP-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1032-DPP-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1032-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1032-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-DPP-NEXT: s_clause 0x1
-; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
+; GFX1032-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX1032-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB9_2
; GFX1032-DPP-NEXT: .LBB9_3:
; GFX1032-DPP-NEXT: s_endpgm
;
; GFX1164-DPP-LABEL: global_atomic_fsub_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1164-DPP: ; %bb.0:
-; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], exec
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s10, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], exec
+; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, s11, v0
+; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, s1, v0
; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1164-DPP-NEXT: s_cbranch_execz .LBB9_3
; GFX1164-DPP-NEXT: ; %bb.1:
-; GFX1164-DPP-NEXT: s_bcnt1_i32_b64 s0, s[10:11]
-; GFX1164-DPP-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
+; GFX1164-DPP-NEXT: s_bcnt1_i32_b64 s0, s[0:1]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
; GFX1164-DPP-NEXT: v_cvt_f64_u32_e32 v[0:1], s0
-; GFX1164-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1164-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1164-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1164-DPP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_mul_f64 v[41:42], v[0:1], 4.0
+; GFX1164-DPP-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-DPP-NEXT: .p2align 6
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s2
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s3
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-DPP-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1164-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1164-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-DPP-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB9_2
; GFX1164-DPP-NEXT: .LBB9_3:
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-DPP-NEXT: s_endpgm
;
; GFX1132-DPP-LABEL: global_atomic_fsub_double_uni_address_uni_value_agent_scope_unsafe:
; GFX1132-DPP: ; %bb.0:
-; GFX1132-DPP-NEXT: s_mov_b32 s8, exec_lo
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s8, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1132-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
+; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
+; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
+; GFX1132-DPP-NEXT: s_mov_b32 s1, exec_lo
+; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-DPP-NEXT: s_cbranch_execz .LBB9_3
; GFX1132-DPP-NEXT: ; %bb.1:
-; GFX1132-DPP-NEXT: s_bcnt1_i32_b32 s0, s8
-; GFX1132-DPP-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
+; GFX1132-DPP-NEXT: s_bcnt1_i32_b32 s0, s0
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v6, 0
; GFX1132-DPP-NEXT: v_cvt_f64_u32_e32 v[0:1], s0
-; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
-; GFX1132-DPP-NEXT: s_mov_b32 s50, s14
-; GFX1132-DPP-NEXT: s_mov_b32 s51, s13
-; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1132-DPP-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_mul_f64 v[41:42], v[0:1], 4.0
+; GFX1132-DPP-NEXT: v_mul_f64 v[4:5], v[0:1], 4.0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-DPP-NEXT: .p2align 6
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX1132-DPP-NEXT: .LBB9_2: ; %atomicrmw.start
; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1132-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1132-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1132-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1132-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-DPP-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB9_2
; GFX1132-DPP-NEXT: .LBB9_3:
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-DPP-NEXT: s_endpgm
- %result = atomicrmw fsub ptr addrspace(1) %ptr, double 4.0 syncscope("agent") monotonic, align 4
+ %result = atomicrmw fsub ptr addrspace(1) %ptr, double 4.0 syncscope("agent") monotonic, align 8
ret void
}
define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_agent_scope_align4_unsafe(ptr addrspace(1) %ptr) #0 {
; GFX7LESS-LABEL: global_atomic_fsub_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX7LESS: ; %bb.0:
-; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-NEXT: s_mov_b32 s32, 0
+; GFX7LESS-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX7LESS-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX7LESS-NEXT: s_mov_b32 s38, -1
+; GFX7LESS-NEXT: s_mov_b32 s39, 0xe8f000
+; GFX7LESS-NEXT: s_add_u32 s36, s36, s11
+; GFX7LESS-NEXT: s_addc_u32 s37, s37, 0
+; GFX7LESS-NEXT: s_mov_b32 s14, s10
+; GFX7LESS-NEXT: s_mov_b32 s13, s9
+; GFX7LESS-NEXT: s_mov_b32 s12, s8
+; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX7LESS-NEXT: s_add_u32 s8, s34, 44
+; GFX7LESS-NEXT: s_addc_u32 s9, s35, 0
+; GFX7LESS-NEXT: s_getpc_b64 s[4:5]
+; GFX7LESS-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX7LESS-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
+; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GFX7LESS-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX7LESS-NEXT: v_or_b32_e32 v40, v0, v2
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX7LESS-NEXT: v_or_b32_e32 v31, v0, v2
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[38:39]
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX7LESS-NEXT: s_mov_b64 s[0:1], exec
-; GFX7LESS-NEXT: v_mov_b32_e32 v41, 0
-; GFX7LESS-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
+; GFX7LESS-NEXT: v_bfrev_b32_e32 v5, 1
; GFX7LESS-NEXT: .LBB10_1: ; %ComputeLoop
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-NEXT: s_ff1_i32_b64 s4, s[0:1]
@@ -7033,7 +6536,7 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_agent
; GFX7LESS-NEXT: s_andn2_b64 s[0:1], s[0:1], s[4:5]
; GFX7LESS-NEXT: v_cmp_ne_u64_e64 s[4:5], s[0:1], 0
; GFX7LESS-NEXT: s_and_b64 vcc, exec, s[4:5]
-; GFX7LESS-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX7LESS-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX7LESS-NEXT: s_cbranch_vccnz .LBB10_1
; GFX7LESS-NEXT: ; %bb.2: ; %ComputeEnd
; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
@@ -7043,102 +6546,70 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_agent
; GFX7LESS-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX7LESS-NEXT: s_cbranch_execz .LBB10_5
; GFX7LESS-NEXT: ; %bb.3:
-; GFX7LESS-NEXT: s_load_dwordx2 s[52:53], s[36:37], 0x9
-; GFX7LESS-NEXT: s_mov_b32 s55, 0xf000
-; GFX7LESS-NEXT: s_mov_b32 s54, -1
+; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x9
+; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
+; GFX7LESS-NEXT: s_mov_b32 s2, -1
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: buffer_load_dwordx2 v[0:1], off, s[52:55], 0
-; GFX7LESS-NEXT: s_mov_b64 s[54:55], 0
+; GFX7LESS-NEXT: buffer_load_dwordx2 v[2:3], off, s[0:3], 0
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
; GFX7LESS-NEXT: .LBB10_4: ; %atomicrmw.start
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-NEXT: v_add_f64 v[2:3], v[0:1], -v[41:42]
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:12
-; GFX7LESS-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:8
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX7LESS-NEXT: s_waitcnt expcnt(2)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[0:3], 0 glc
+; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-NEXT: s_cbranch_execnz .LBB10_4
; GFX7LESS-NEXT: .LBB10_5:
; GFX7LESS-NEXT: s_endpgm
;
; GFX9-LABEL: global_atomic_fsub_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-NEXT: s_mov_b32 s66, -1
-; GFX9-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-NEXT: s_add_u32 s64, s64, s11
-; GFX9-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-NEXT: s_mov_b32 s51, s8
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_mov_b32 s50, s9
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX9-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX9-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX9-NEXT: s_mov_b32 s38, -1
+; GFX9-NEXT: s_mov_b32 s39, 0xe00000
+; GFX9-NEXT: s_add_u32 s36, s36, s11
+; GFX9-NEXT: s_addc_u32 s37, s37, 0
+; GFX9-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX9-NEXT: s_mov_b32 s12, s8
+; GFX9-NEXT: s_add_u32 s8, s34, 44
+; GFX9-NEXT: s_mov_b32 s13, s9
+; GFX9-NEXT: s_addc_u32 s9, s35, 0
+; GFX9-NEXT: s_getpc_b64 s[4:5]
+; GFX9-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX9-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
+; GFX9-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
+; GFX9-NEXT: s_mov_b32 s14, s10
+; GFX9-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-NEXT: s_mov_b32 s33, s10
-; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX9-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: s_movk_i32 s32, 0x800
+; GFX9-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX9-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX9-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX9-NEXT: s_mov_b32 s32, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: v_mov_b32_e32 v41, 0
+; GFX9-NEXT: v_mov_b32_e32 v4, 0
; GFX9-NEXT: s_mov_b64 s[0:1], exec
-; GFX9-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX9-NEXT: v_bfrev_b32_e32 v5, 1
; GFX9-NEXT: .LBB10_1: ; %ComputeLoop
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-NEXT: s_ff1_i32_b64 s4, s[0:1]
; GFX9-NEXT: v_readlane_b32 s3, v1, s4
; GFX9-NEXT: v_readlane_b32 s2, v0, s4
-; GFX9-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX9-NEXT: s_lshl_b64 s[2:3], 1, s4
; GFX9-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
; GFX9-NEXT: s_cmp_lg_u64 s[0:1], 0
@@ -7151,99 +6622,64 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_agent
; GFX9-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX9-NEXT: s_cbranch_execz .LBB10_5
; GFX9-NEXT: ; %bb.3:
-; GFX9-NEXT: s_load_dwordx2 s[52:53], s[36:37], 0x24
-; GFX9-NEXT: v_mov_b32_e32 v0, 0
-; GFX9-NEXT: s_mov_b64 s[54:55], 0
+; GFX9-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX9-NEXT: v_mov_b32_e32 v6, 0
+; GFX9-NEXT: s_mov_b64 s[2:3], 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX9-NEXT: global_load_dwordx2 v[2:3], v6, s[0:1]
; GFX9-NEXT: .LBB10_4: ; %atomicrmw.start
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX9-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-NEXT: s_cbranch_execnz .LBB10_4
; GFX9-NEXT: .LBB10_5:
; GFX9-NEXT: s_endpgm
;
; GFX1064-LABEL: global_atomic_fsub_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX1064: ; %bb.0:
-; GFX1064-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-NEXT: s_mov_b32 s66, -1
-; GFX1064-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-NEXT: s_add_u32 s64, s64, s11
+; GFX1064-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1064-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1064-NEXT: s_mov_b32 s38, -1
+; GFX1064-NEXT: s_mov_b32 s39, 0x31e16000
+; GFX1064-NEXT: s_add_u32 s36, s36, s11
; GFX1064-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1064-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-NEXT: s_mov_b32 s51, s8
+; GFX1064-NEXT: s_addc_u32 s37, s37, 0
+; GFX1064-NEXT: s_mov_b32 s12, s8
; GFX1064-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-NEXT: s_mov_b32 s50, s9
+; GFX1064-NEXT: s_mov_b32 s13, s9
; GFX1064-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1064-NEXT: s_getpc_b64 s[4:5]
+; GFX1064-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1064-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1064-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1064-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1064-NEXT: s_mov_b32 s33, s10
-; GFX1064-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1064-NEXT: s_mov_b32 s14, s10
+; GFX1064-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1064-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1064-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1064-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: s_movk_i32 s32, 0x800
+; GFX1064-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1064-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1064-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1064-NEXT: s_mov_b32 s32, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: v_mov_b32_e32 v41, 0
-; GFX1064-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX1064-NEXT: v_mov_b32_e32 v4, 0
+; GFX1064-NEXT: v_bfrev_b32_e32 v5, 1
; GFX1064-NEXT: s_mov_b64 s[0:1], exec
; GFX1064-NEXT: .LBB10_1: ; %ComputeLoop
; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1064-NEXT: s_ff1_i32_b64 s4, s[0:1]
; GFX1064-NEXT: v_readlane_b32 s3, v1, s4
; GFX1064-NEXT: v_readlane_b32 s2, v0, s4
-; GFX1064-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX1064-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX1064-NEXT: s_lshl_b64 s[2:3], 1, s4
; GFX1064-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
; GFX1064-NEXT: s_cmp_lg_u64 s[0:1], 0
@@ -7256,93 +6692,57 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_agent
; GFX1064-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX1064-NEXT: s_cbranch_execz .LBB10_5
; GFX1064-NEXT: ; %bb.3:
-; GFX1064-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1064-NEXT: v_mov_b32_e32 v0, 0
-; GFX1064-NEXT: s_mov_b64 s[54:55], 0
+; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX1064-NEXT: v_mov_b32_e32 v6, 0
+; GFX1064-NEXT: s_mov_b64 s[2:3], 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX1064-NEXT: global_load_dwordx2 v[2:3], v6, s[0:1]
; GFX1064-NEXT: .LBB10_4: ; %atomicrmw.start
; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1064-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1064-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: s_clause 0x1
-; GFX1064-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX1064-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1064-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1064-NEXT: s_waitcnt vmcnt(0)
+; GFX1064-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1064-NEXT: v_mov_b32_e32 v3, v1
+; GFX1064-NEXT: v_mov_b32_e32 v2, v0
+; GFX1064-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1064-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX1064-NEXT: s_cbranch_execnz .LBB10_4
; GFX1064-NEXT: .LBB10_5:
; GFX1064-NEXT: s_endpgm
;
; GFX1032-LABEL: global_atomic_fsub_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX1032: ; %bb.0:
-; GFX1032-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-NEXT: s_mov_b32 s66, -1
-; GFX1032-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-NEXT: s_add_u32 s64, s64, s11
+; GFX1032-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1032-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1032-NEXT: s_mov_b32 s38, -1
+; GFX1032-NEXT: s_mov_b32 s39, 0x31c16000
+; GFX1032-NEXT: s_add_u32 s36, s36, s11
; GFX1032-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1032-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-NEXT: s_mov_b32 s51, s8
+; GFX1032-NEXT: s_addc_u32 s37, s37, 0
+; GFX1032-NEXT: s_mov_b32 s12, s8
; GFX1032-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-NEXT: s_mov_b32 s50, s9
+; GFX1032-NEXT: s_mov_b32 s13, s9
; GFX1032-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1032-NEXT: s_getpc_b64 s[4:5]
+; GFX1032-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1032-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1032-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1032-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1032-NEXT: s_mov_b32 s33, s10
-; GFX1032-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-NEXT: s_mov_b32 s14, s10
+; GFX1032-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1032-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1032-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1032-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1032-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1032-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1032-NEXT: s_mov_b32 s32, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: v_mov_b32_e32 v41, 0
-; GFX1032-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX1032-NEXT: v_mov_b32_e32 v4, 0
+; GFX1032-NEXT: v_bfrev_b32_e32 v5, 1
; GFX1032-NEXT: s_mov_b32 s0, exec_lo
; GFX1032-NEXT: .LBB10_1: ; %ComputeLoop
; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
@@ -7351,61 +6751,32 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_agent
; GFX1032-NEXT: v_readlane_b32 s2, v0, s1
; GFX1032-NEXT: s_lshl_b32 s1, 1, s1
; GFX1032-NEXT: s_andn2_b32 s0, s0, s1
-; GFX1032-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX1032-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX1032-NEXT: s_cmp_lg_u32 s0, 0
; GFX1032-NEXT: s_cbranch_scc1 .LBB10_1
; GFX1032-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1032-NEXT: s_mov_b32 s54, 0
+; GFX1032-NEXT: s_mov_b32 s2, 0
; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX1032-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX1032-NEXT: s_cbranch_execz .LBB10_5
; GFX1032-NEXT: ; %bb.3:
-; GFX1032-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1032-NEXT: v_mov_b32_e32 v0, 0
+; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX1032-NEXT: v_mov_b32_e32 v6, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX1032-NEXT: global_load_dwordx2 v[2:3], v6, s[0:1]
; GFX1032-NEXT: .LBB10_4: ; %atomicrmw.start
; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1032-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1032-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: s_clause 0x1
-; GFX1032-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
+; GFX1032-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1032-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1032-NEXT: s_waitcnt vmcnt(0)
+; GFX1032-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1032-NEXT: v_mov_b32_e32 v3, v1
+; GFX1032-NEXT: v_mov_b32_e32 v2, v0
+; GFX1032-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
; GFX1032-NEXT: s_cbranch_execnz .LBB10_4
; GFX1032-NEXT: .LBB10_5:
; GFX1032-NEXT: s_endpgm
@@ -7413,31 +6784,24 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_agent
; GFX1164-LABEL: global_atomic_fsub_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX1164: ; %bb.0:
; GFX1164-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1164-NEXT: s_mov_b32 s51, s8
+; GFX1164-NEXT: s_mov_b32 s12, s8
; GFX1164-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-NEXT: s_mov_b32 s50, s9
+; GFX1164-NEXT: s_mov_b32 s13, s9
; GFX1164-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1164-NEXT: s_getpc_b64 s[4:5]
+; GFX1164-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1164-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1164-NEXT: v_mov_b32_e32 v31, v0
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: s_mov_b32 s33, s10
-; GFX1164-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX1164-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1164-NEXT: s_mov_b32 s14, s10
+; GFX1164-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1164-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1164-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: s_mov_b32 s32, 32
-; GFX1164-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1164-NEXT: s_mov_b32 s32, 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: v_mov_b32_e32 v41, 0
-; GFX1164-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX1164-NEXT: s_swappc_b64 s[30:31], s[16:17]
+; GFX1164-NEXT: v_mov_b32_e32 v4, 0
+; GFX1164-NEXT: v_bfrev_b32_e32 v5, 1
; GFX1164-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-NEXT: .LBB10_1: ; %ComputeLoop
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
@@ -7446,7 +6810,7 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_agent
; GFX1164-NEXT: v_readlane_b32 s3, v1, s4
; GFX1164-NEXT: v_readlane_b32 s2, v0, s4
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX1164-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX1164-NEXT: s_lshl_b64 s[2:3], 1, s4
; GFX1164-NEXT: s_and_not1_b64 s[0:1], s[0:1], s[2:3]
; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
@@ -7461,81 +6825,48 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_agent
; GFX1164-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX1164-NEXT: s_cbranch_execz .LBB10_5
; GFX1164-NEXT: ; %bb.3:
-; GFX1164-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1164-NEXT: v_mov_b32_e32 v0, 0
-; GFX1164-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1164-NEXT: v_mov_b32_e32 v6, 0
+; GFX1164-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-NEXT: .p2align 6
+; GFX1164-NEXT: global_load_b64 v[2:3], v6, s[0:1]
; GFX1164-NEXT: .LBB10_4: ; %atomicrmw.start
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1164-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1164-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-NEXT: s_cbranch_execnz .LBB10_4
; GFX1164-NEXT: .LBB10_5:
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-NEXT: s_endpgm
;
; GFX1132-LABEL: global_atomic_fsub_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX1132: ; %bb.0:
; GFX1132-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1132-NEXT: s_mov_b64 s[48:49], s[0:1]
+; GFX1132-NEXT: v_mov_b32_e32 v31, v0
; GFX1132-NEXT: s_add_u32 s8, s34, 44
; GFX1132-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX1132-NEXT: v_mov_b32_e32 v31, v0
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1132-NEXT: s_mov_b32 s50, s14
-; GFX1132-NEXT: s_mov_b32 s51, s13
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
+; GFX1132-NEXT: s_getpc_b64 s[4:5]
+; GFX1132-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1132-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1132-NEXT: s_mov_b32 s12, s13
+; GFX1132-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1132-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1132-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX1132-NEXT: s_mov_b32 s13, s14
; GFX1132-NEXT: s_mov_b32 s14, s15
-; GFX1132-NEXT: s_mov_b32 s32, 32
-; GFX1132-NEXT: s_mov_b32 s33, s15
-; GFX1132-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-NEXT: s_mov_b32 s32, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: v_mov_b32_e32 v41, 0
-; GFX1132-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX1132-NEXT: s_swappc_b64 s[30:31], s[16:17]
+; GFX1132-NEXT: v_mov_b32_e32 v4, 0
+; GFX1132-NEXT: v_bfrev_b32_e32 v5, 1
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-NEXT: .LBB10_1: ; %ComputeLoop
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
@@ -7546,522 +6877,358 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_agent
; GFX1132-NEXT: s_lshl_b32 s1, 1, s1
; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX1132-NEXT: s_and_not1_b32 s0, s0, s1
-; GFX1132-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX1132-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX1132-NEXT: s_cmp_lg_u32 s0, 0
; GFX1132-NEXT: s_cbranch_scc1 .LBB10_1
; GFX1132-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-NEXT: s_mov_b32 s54, 0
+; GFX1132-NEXT: s_mov_b32 s2, 0
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX1132-NEXT: s_cbranch_execz .LBB10_5
; GFX1132-NEXT: ; %bb.3:
-; GFX1132-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1132-NEXT: v_mov_b32_e32 v0, 0
+; GFX1132-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1132-NEXT: v_mov_b32_e32 v6, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-NEXT: .p2align 6
+; GFX1132-NEXT: global_load_b64 v[2:3], v6, s[0:1]
; GFX1132-NEXT: .LBB10_4: ; %atomicrmw.start
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1132-NEXT: s_add_u32 s8, s34, 44
-; GFX1132-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1132-NEXT: s_mov_b32 s12, s51
-; GFX1132-NEXT: s_mov_b32 s13, s50
-; GFX1132-NEXT: s_mov_b32 s14, s33
-; GFX1132-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1132-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-NEXT: s_cbranch_execnz .LBB10_4
; GFX1132-NEXT: .LBB10_5:
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-NEXT: s_endpgm
;
; GFX7LESS-DPP-LABEL: global_atomic_fsub_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX7LESS-DPP: ; %bb.0:
-; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-DPP-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s55, 0xf000
-; GFX7LESS-DPP-NEXT: s_mov_b32 s54, -1
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v42, v0, v2
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v42
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v41, v1
-; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[0:1], off, s[52:55], 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-DPP-NEXT: .LBB10_1: ; %atomicrmw.start
-; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-DPP-NEXT: v_add_f64 v[2:3], v[0:1], -v[40:41]
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:8
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(2)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v42
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
-; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB10_1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s32, 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s50, -1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s51, 0xe8f000
+; GFX7LESS-DPP-NEXT: s_add_u32 s48, s48, s11
+; GFX7LESS-DPP-NEXT: s_addc_u32 s49, s49, 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s10
+; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s8
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[36:37], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s39, 0xf000
+; GFX7LESS-DPP-NEXT: s_mov_b32 s38, -1
+; GFX7LESS-DPP-NEXT: s_add_u32 s8, s4, 44
+; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s5, 0
+; GFX7LESS-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX7LESS-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX7LESS-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
+; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
+; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
+; GFX7LESS-DPP-NEXT: v_or_b32_e32 v0, v0, v1
+; GFX7LESS-DPP-NEXT: v_or_b32_e32 v31, v0, v2
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
+; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
+; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
+; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[4:5], off, s[36:39], 0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], 0
+; GFX7LESS-DPP-NEXT: .LBB10_1: ; %atomicrmw.start
+; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-DPP-NEXT: v_add_f64 v[2:3], v[4:5], -v[0:1]
+; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v9, v5
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v8, v4
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v2
+; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[36:39], 0 glc
+; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[4:5]
+; GFX7LESS-DPP-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, v6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, v7
+; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[0:1]
+; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB10_1
; GFX7LESS-DPP-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX7LESS-DPP-NEXT: s_endpgm
;
; GFX9-DPP-LABEL: global_atomic_fsub_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX9-DPP: ; %bb.0:
-; GFX9-DPP-NEXT: s_mov_b32 s68, SCRATCH_RSRC_DWORD0
-; GFX9-DPP-NEXT: s_mov_b32 s69, SCRATCH_RSRC_DWORD1
-; GFX9-DPP-NEXT: s_mov_b32 s70, -1
-; GFX9-DPP-NEXT: s_mov_b32 s71, 0xe00000
-; GFX9-DPP-NEXT: s_add_u32 s68, s68, s11
-; GFX9-DPP-NEXT: s_addc_u32 s69, s69, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-DPP-NEXT: s_mov_b32 s51, s8
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_mov_b32 s50, s9
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX9-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX9-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX9-DPP-NEXT: s_mov_b32 s38, -1
+; GFX9-DPP-NEXT: s_mov_b32 s39, 0xe00000
+; GFX9-DPP-NEXT: s_add_u32 s36, s36, s11
+; GFX9-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX9-DPP-NEXT: s_mov_b32 s12, s8
+; GFX9-DPP-NEXT: s_add_u32 s8, s34, 44
+; GFX9-DPP-NEXT: s_mov_b32 s13, s9
+; GFX9-DPP-NEXT: s_addc_u32 s9, s35, 0
+; GFX9-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX9-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX9-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
+; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
+; GFX9-DPP-NEXT: s_mov_b32 s14, s10
+; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-DPP-NEXT: s_mov_b32 s33, s10
-; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[68:69]
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[70:71]
-; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX9-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX9-DPP-NEXT: s_mov_b32 s32, 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX9-DPP-NEXT: v_cndmask_b32_e64 v11, v9, v1, s[0:1]
-; GFX9-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:1 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:1 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v4, 1
+; GFX9-DPP-NEXT: v_cndmask_b32_e64 v6, v4, v1, s[0:1]
+; GFX9-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:1 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:1 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:2 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:2 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:2 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:2 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:4 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:4 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:4 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:4 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:8 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:8 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:8 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:8 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_bcast:15 row_mask:0xa bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_bcast:15 row_mask:0xa bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_bcast:15 row_mask:0xa bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_bcast:15 row_mask:0xa bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
; GFX9-DPP-NEXT: s_nop 1
-; GFX9-DPP-NEXT: v_mov_b32_dpp v9, v11 row_bcast:31 row_mask:0xc bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v10 row_bcast:31 row_mask:0xc bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
+; GFX9-DPP-NEXT: v_mov_b32_dpp v4, v6 row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v3, v5 row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[3:4], v[5:6], v[3:4]
; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 0
; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v1
; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX9-DPP-NEXT: v_readlane_b32 s53, v9, 63
-; GFX9-DPP-NEXT: v_readlane_b32 s52, v8, 63
+; GFX9-DPP-NEXT: v_readlane_b32 s3, v4, 63
+; GFX9-DPP-NEXT: v_readlane_b32 s2, v3, 63
; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[2:3]
; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
+; GFX9-DPP-NEXT: s_and_saveexec_b64 s[2:3], vcc
; GFX9-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX9-DPP-NEXT: ; %bb.1:
-; GFX9-DPP-NEXT: s_load_dwordx2 s[54:55], s[36:37], 0x24
-; GFX9-DPP-NEXT: s_mov_b64 s[64:65], 0
+; GFX9-DPP-NEXT: s_load_dwordx2 s[2:3], s[34:35], 0x24
+; GFX9-DPP-NEXT: s_mov_b64 s[4:5], 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[54:55]
+; GFX9-DPP-NEXT: global_load_dwordx2 v[11:12], v0, s[2:3]
; GFX9-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX9-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -s[52:53]
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[68:69]
-; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[68:71], 0 offset:4
-; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[68:71], 0
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[68:71], 0 offset:12
-; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[68:71], 0 offset:8
-; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[70:71]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s54
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s55
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[68:71], 0
-; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[68:71], 0 offset:4
-; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-DPP-NEXT: s_or_b64 s[64:65], vcc, s[64:65]
-; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[64:65]
+; GFX9-DPP-NEXT: v_add_f64 v[9:10], v[11:12], -s[0:1]
+; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[1:2], v0, v[9:12], s[2:3] glc
+; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[1:2], v[11:12]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v12, v2
+; GFX9-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v11, v1
+; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX9-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX9-DPP-NEXT: .LBB10_3:
; GFX9-DPP-NEXT: s_endpgm
;
; GFX1064-DPP-LABEL: global_atomic_fsub_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX1064-DPP: ; %bb.0:
-; GFX1064-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1064-DPP-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-DPP-NEXT: s_add_u32 s64, s64, s11
+; GFX1064-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1064-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1064-DPP-NEXT: s_mov_b32 s38, -1
+; GFX1064-DPP-NEXT: s_mov_b32 s39, 0x31e16000
+; GFX1064-DPP-NEXT: s_add_u32 s36, s36, s11
; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1064-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1064-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX1064-DPP-NEXT: s_mov_b32 s12, s8
; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1064-DPP-NEXT: s_mov_b32 s13, s9
; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1064-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1064-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1064-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1064-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX1064-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1064-DPP-NEXT: s_mov_b32 s32, 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1064-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v11, 0x80000000, v1, s[0:1]
-; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
-; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1064-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_readlane_b32 s3, v9, 0
-; GFX1064-DPP-NEXT: v_readlane_b32 s2, v8, 0
-; GFX1064-DPP-NEXT: v_readlane_b32 s5, v9, 32
-; GFX1064-DPP-NEXT: v_readlane_b32 s4, v8, 32
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], s[2:3], s[4:5]
+; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v4, 1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v6, 0x80000000, v1, s[0:1]
+; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[5:6], v[3:4]
+; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
+; GFX1064-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_readlane_b32 s3, v4, 0
+; GFX1064-DPP-NEXT: v_readlane_b32 s2, v3, 0
+; GFX1064-DPP-NEXT: v_readlane_b32 s5, v4, 32
+; GFX1064-DPP-NEXT: v_readlane_b32 s4, v3, 32
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], s[2:3], s[4:5]
; GFX1064-DPP-NEXT: s_mov_b64 exec, s[0:1]
; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v41, v8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v42, v9
-; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v7, exec_hi, v0
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, v3
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, v4
+; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v7
; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX1064-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX1064-DPP-NEXT: ; %bb.1:
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1064-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX1064-DPP-NEXT: global_load_dwordx2 v[9:10], v2, s[0:1]
; GFX1064-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-DPP-NEXT: s_clause 0x1
-; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX1064-DPP-NEXT: v_add_f64 v[7:8], v[9:10], -v[0:1]
+; GFX1064-DPP-NEXT: global_atomic_cmpswap_x2 v[7:8], v2, v[7:10], s[0:1] glc
+; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1064-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[7:8], v[9:10]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, v8
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v9, v7
+; GFX1064-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX1064-DPP-NEXT: .LBB10_3:
; GFX1064-DPP-NEXT: s_endpgm
;
; GFX1032-DPP-LABEL: global_atomic_fsub_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX1032-DPP: ; %bb.0:
-; GFX1032-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1032-DPP-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-DPP-NEXT: s_add_u32 s64, s64, s11
+; GFX1032-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1032-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1032-DPP-NEXT: s_mov_b32 s38, -1
+; GFX1032-DPP-NEXT: s_mov_b32 s39, 0x31c16000
+; GFX1032-DPP-NEXT: s_add_u32 s36, s36, s11
; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1032-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1032-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX1032-DPP-NEXT: s_mov_b32 s12, s8
; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1032-DPP-NEXT: s_mov_b32 s13, s9
; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1032-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1032-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1032-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1032-DPP-NEXT: s_mov_b32 s32, 0
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1032-DPP-NEXT: s_or_saveexec_b32 s0, -1
-; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v11, 0x80000000, v1, s0
-; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
-; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1032-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v4, 1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v6, 0x80000000, v1, s0
+; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[5:6], v[3:4]
+; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
+; GFX1032-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
; GFX1032-DPP-NEXT: s_mov_b32 exec_lo, s0
-; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v41, v8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v42, v9
-; GFX1032-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
+; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v7, exec_lo, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, v3
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, v4
+; GFX1032-DPP-NEXT: s_mov_b32 s2, 0
+; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v7
; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX1032-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX1032-DPP-NEXT: ; %bb.1:
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX1032-DPP-NEXT: global_load_dwordx2 v[9:10], v2, s[0:1]
; GFX1032-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-DPP-NEXT: s_clause 0x1
-; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
+; GFX1032-DPP-NEXT: v_add_f64 v[7:8], v[9:10], -v[0:1]
+; GFX1032-DPP-NEXT: global_atomic_cmpswap_x2 v[7:8], v2, v[7:10], s[0:1] glc
+; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1032-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[7:8], v[9:10]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, v8
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v9, v7
+; GFX1032-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX1032-DPP-NEXT: .LBB10_3:
; GFX1032-DPP-NEXT: s_endpgm
@@ -8069,238 +7236,175 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_agent
; GFX1164-DPP-LABEL: global_atomic_fsub_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX1164-DPP: ; %bb.0:
; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1164-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1164-DPP-NEXT: s_mov_b32 s12, s8
; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1164-DPP-NEXT: s_mov_b32 s13, s9
; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1164-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1164-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1164-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v0
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX1164-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1164-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1164-DPP-NEXT: s_mov_b32 s32, 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
+; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1164-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v11, 0x80000000, v1, s[0:1]
-; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
+; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v3, 1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v5, 0x80000000, v1, s[0:1]
+; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s[0:1]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
-; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[4:5], v[2:3]
+; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1164-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
+; GFX1164-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1164-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_permlane64_b32 v11, v9
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_permlane64_b32 v5, v3
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_permlane64_b32 v10, v8
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_permlane64_b32 v4, v2
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_mov_b64 exec, s[0:1]
-; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v41, v8
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v42, v9
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-DPP-NEXT: s_waitcnt_depctr 0xfffe
-; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
+; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v6, exec_hi, v0
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, v2
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, v3
+; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
; GFX1164-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX1164-DPP-NEXT: ; %bb.1:
-; GFX1164-DPP-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1164-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-DPP-NEXT: .p2align 6
+; GFX1164-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
; GFX1164-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-DPP-NEXT: v_add_f64 v[6:7], v[8:9], -v[0:1]
+; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
+; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[8:9]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v9, v7
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, v6
+; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX1164-DPP-NEXT: .LBB10_3:
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-DPP-NEXT: s_endpgm
;
; GFX1132-DPP-LABEL: global_atomic_fsub_double_uni_address_div_value_agent_scope_align4_unsafe:
; GFX1132-DPP: ; %bb.0:
; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1132-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1132-DPP-NEXT: s_mov_b32 s50, s14
-; GFX1132-DPP-NEXT: s_mov_b32 s51, s13
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
+; GFX1132-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1132-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1132-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1132-DPP-NEXT: s_mov_b32 s12, s13
+; GFX1132-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX1132-DPP-NEXT: s_mov_b32 s13, s14
; GFX1132-DPP-NEXT: s_mov_b32 s14, s15
-; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-DPP-NEXT: s_mov_b32 s32, 0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
+; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1132-DPP-NEXT: s_or_saveexec_b32 s0, -1
-; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v11, 0x80000000, v1, s0
-; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
+; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v3, 1
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v5, 0x80000000, v1, s0
+; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
-; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[4:5], v[2:3]
+; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1132-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
+; GFX1132-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1132-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
; GFX1132-DPP-NEXT: s_mov_b32 exec_lo, s0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v41, v8
-; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v0, v2
+; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v6, exec_lo, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v42, v9
-; GFX1132-DPP-NEXT: s_mov_b32 s54, 0
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v10, 0 :: v_dual_mov_b32 v1, v3
+; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
-; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
+; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
; GFX1132-DPP-NEXT: s_cbranch_execz .LBB10_3
; GFX1132-DPP-NEXT: ; %bb.1:
-; GFX1132-DPP-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
+; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-DPP-NEXT: .p2align 6
+; GFX1132-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
; GFX1132-DPP-NEXT: .LBB10_2: ; %atomicrmw.start
; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1132-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1132-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-DPP-NEXT: v_add_f64 v[6:7], v[8:9], -v[0:1]
+; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
+; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[6:7], v[8:9]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6
+; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB10_2
; GFX1132-DPP-NEXT: .LBB10_3:
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-DPP-NEXT: s_endpgm
%divValue = call double @div.float.value()
- %result = atomicrmw fsub ptr addrspace(1) %ptr, double %divValue syncscope("agent") monotonic, align 4
+ %result = atomicrmw fsub ptr addrspace(1) %ptr, double %divValue syncscope("agent") monotonic, align 8
ret void
}
@@ -12086,998 +11190,555 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_agent
define amdgpu_kernel void @global_atomic_fsub_double_uni_address_uni_value_default_scope_strictfp(ptr addrspace(1) %ptr) #2 {
; GFX7LESS-LABEL: global_atomic_fsub_double_uni_address_uni_value_default_scope_strictfp:
; GFX7LESS: ; %bb.0:
-; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v5, exec_lo, 0
-; GFX7LESS-NEXT: s_bcnt1_i32_b64 s0, exec
-; GFX7LESS-NEXT: s_mov_b32 s1, 0x43300000
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0xc3300000
-; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v5, exec_hi, v5
-; GFX7LESS-NEXT: v_add_f64 v[3:4], s[0:1], v[3:4]
-; GFX7LESS-NEXT: v_mul_f64 v[41:42], 4.0, v[3:4]
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v5
+; GFX7LESS-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GFX7LESS-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GFX7LESS-NEXT: s_mov_b32 s14, -1
+; GFX7LESS-NEXT: s_mov_b32 s15, 0xe8f000
+; GFX7LESS-NEXT: s_add_u32 s12, s12, s11
+; GFX7LESS-NEXT: s_addc_u32 s13, s13, 0
+; GFX7LESS-NEXT: s_mov_b64 s[2:3], exec
+; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s2, 0
+; GFX7LESS-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s3, v0
+; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7LESS-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX7LESS-NEXT: s_cbranch_execz .LBB16_3
; GFX7LESS-NEXT: ; %bb.1:
-; GFX7LESS-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
+; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
+; GFX7LESS-NEXT: s_bcnt1_i32_b64 s6, s[2:3]
+; GFX7LESS-NEXT: s_mov_b32 s7, 0x43300000
+; GFX7LESS-NEXT: v_mov_b32_e32 v0, 0
+; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0xc3300000
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX7LESS-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX7LESS-NEXT: v_or_b32_e32 v3, v0, v1
+; GFX7LESS-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0x0
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
+; GFX7LESS-NEXT: v_add_f64 v[0:1], s[6:7], v[0:1]
+; GFX7LESS-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, s0
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, s1
-; GFX7LESS-NEXT: v_or_b32_e32 v40, v3, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, s8
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, s9
+; GFX7LESS-NEXT: s_mov_b32 s2, -1
; GFX7LESS-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-NEXT: v_add_f64 v[2:3], v[0:1], -v[41:42]
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:12
-; GFX7LESS-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:8
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX7LESS-NEXT: s_waitcnt expcnt(2)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[0:3], 0 glc
+; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-NEXT: s_cbranch_execnz .LBB16_2
; GFX7LESS-NEXT: .LBB16_3:
; GFX7LESS-NEXT: s_endpgm
;
; GFX9-LABEL: global_atomic_fsub_double_uni_address_uni_value_default_scope_strictfp:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-NEXT: s_mov_b32 s66, -1
-; GFX9-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-NEXT: s_add_u32 s64, s64, s11
-; GFX9-NEXT: v_mov_b32_e32 v3, 0
-; GFX9-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-NEXT: s_bcnt1_i32_b64 s0, exec
-; GFX9-NEXT: v_mov_b32_e32 v4, 0xc3300000
-; GFX9-NEXT: s_mov_b32 s1, 0x43300000
-; GFX9-NEXT: v_add_f64 v[3:4], s[0:1], v[3:4]
-; GFX9-NEXT: s_movk_i32 s32, 0x800
-; GFX9-NEXT: v_mul_f64 v[41:42], 4.0, v[3:4]
-; GFX9-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX9-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX9-NEXT: s_and_saveexec_b64 s[0:1], vcc
+; GFX9-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GFX9-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GFX9-NEXT: s_mov_b32 s14, -1
+; GFX9-NEXT: s_mov_b64 s[0:1], exec
+; GFX9-NEXT: s_mov_b32 s15, 0xe00000
+; GFX9-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
+; GFX9-NEXT: s_add_u32 s12, s12, s11
+; GFX9-NEXT: v_mbcnt_hi_u32_b32 v0, s1, v0
+; GFX9-NEXT: s_addc_u32 s13, s13, 0
+; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX9-NEXT: s_and_saveexec_b64 s[2:3], vcc
; GFX9-NEXT: s_cbranch_execz .LBB16_3
; GFX9-NEXT: ; %bb.1:
-; GFX9-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX9-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX9-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX9-NEXT: s_mov_b32 s33, s10
-; GFX9-NEXT: s_mov_b32 s50, s9
+; GFX9-NEXT: v_mov_b32_e32 v0, 0
+; GFX9-NEXT: s_bcnt1_i32_b64 s0, s[0:1]
+; GFX9-NEXT: v_mov_b32_e32 v1, 0xc3300000
+; GFX9-NEXT: s_mov_b32 s1, 0x43300000
+; GFX9-NEXT: v_add_f64 v[0:1], s[0:1], v[0:1]
+; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9-NEXT: s_mov_b64 s[2:3], 0
+; GFX9-NEXT: v_mov_b32_e32 v6, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX9-NEXT: s_mov_b32 s51, s8
-; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX9-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX9-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: v_mov_b32_e32 v2, s1
-; GFX9-NEXT: s_mov_b64 s[54:55], 0
-; GFX9-NEXT: v_mov_b32_e32 v1, s0
-; GFX9-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX9-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-NEXT: v_mov_b32_e32 v3, s5
; GFX9-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX9-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX9-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-NEXT: s_cbranch_execnz .LBB16_2
; GFX9-NEXT: .LBB16_3:
; GFX9-NEXT: s_endpgm
;
; GFX1064-LABEL: global_atomic_fsub_double_uni_address_uni_value_default_scope_strictfp:
; GFX1064: ; %bb.0:
-; GFX1064-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-NEXT: s_mov_b32 s66, -1
-; GFX1064-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-NEXT: s_add_u32 s64, s64, s11
-; GFX1064-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-NEXT: s_bcnt1_i32_b64 s0, exec
-; GFX1064-NEXT: s_mov_b32 s1, 0x43300000
-; GFX1064-NEXT: s_movk_i32 s32, 0x800
-; GFX1064-NEXT: v_add_f64 v[3:4], 0xc3300000, s[0:1]
-; GFX1064-NEXT: v_mul_f64 v[41:42], 4.0, v[3:4]
-; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX1064-NEXT: s_and_saveexec_b64 s[0:1], vcc
+; GFX1064-NEXT: s_mov_b64 s[0:1], exec
+; GFX1064-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GFX1064-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
+; GFX1064-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GFX1064-NEXT: s_mov_b32 s14, -1
+; GFX1064-NEXT: s_mov_b32 s15, 0x31e16000
+; GFX1064-NEXT: s_add_u32 s12, s12, s11
+; GFX1064-NEXT: v_mbcnt_hi_u32_b32 v0, s1, v0
+; GFX1064-NEXT: s_addc_u32 s13, s13, 0
+; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX1064-NEXT: s_and_saveexec_b64 s[2:3], vcc
; GFX1064-NEXT: s_cbranch_execz .LBB16_3
; GFX1064-NEXT: ; %bb.1:
-; GFX1064-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1064-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1064-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1064-NEXT: s_mov_b32 s33, s10
-; GFX1064-NEXT: s_mov_b32 s50, s9
-; GFX1064-NEXT: s_mov_b32 s51, s8
-; GFX1064-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1064-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1064-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1064-NEXT: s_mov_b64 s[54:55], 0
+; GFX1064-NEXT: s_bcnt1_i32_b64 s2, s[0:1]
+; GFX1064-NEXT: s_mov_b32 s3, 0x43300000
+; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1064-NEXT: v_add_f64 v[0:1], 0xc3300000, s[2:3]
+; GFX1064-NEXT: v_mov_b32_e32 v6, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
+; GFX1064-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX1064-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: v_mov_b32_e32 v2, s1
-; GFX1064-NEXT: v_mov_b32_e32 v1, s0
+; GFX1064-NEXT: v_mov_b32_e32 v2, s2
+; GFX1064-NEXT: v_mov_b32_e32 v3, s3
+; GFX1064-NEXT: s_mov_b64 s[2:3], 0
; GFX1064-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1064-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1064-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX1064-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1064-NEXT: s_add_u32 s8, s36, 44
-; GFX1064-NEXT: s_addc_u32 s9, s37, 0
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: s_clause 0x1
-; GFX1064-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX1064-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1064-NEXT: v_mov_b32_e32 v3, v1
+; GFX1064-NEXT: v_mov_b32_e32 v2, v0
+; GFX1064-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1064-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX1064-NEXT: s_cbranch_execnz .LBB16_2
; GFX1064-NEXT: .LBB16_3:
; GFX1064-NEXT: s_endpgm
;
; GFX1032-LABEL: global_atomic_fsub_double_uni_address_uni_value_default_scope_strictfp:
; GFX1032: ; %bb.0:
-; GFX1032-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-NEXT: s_mov_b32 s66, -1
-; GFX1032-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-NEXT: s_add_u32 s64, s64, s11
-; GFX1032-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-NEXT: s_bcnt1_i32_b32 s0, exec_lo
-; GFX1032-NEXT: s_mov_b32 s1, 0x43300000
-; GFX1032-NEXT: s_mov_b32 s54, 0
-; GFX1032-NEXT: v_add_f64 v[3:4], 0xc3300000, s[0:1]
-; GFX1032-NEXT: s_movk_i32 s32, 0x400
-; GFX1032-NEXT: v_mul_f64 v[41:42], 4.0, v[3:4]
-; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
-; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
+; GFX1032-NEXT: s_mov_b32 s0, exec_lo
+; GFX1032-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
+; GFX1032-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GFX1032-NEXT: s_mov_b32 s14, -1
+; GFX1032-NEXT: s_mov_b32 s15, 0x31c16000
+; GFX1032-NEXT: s_add_u32 s12, s12, s11
+; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
+; GFX1032-NEXT: s_addc_u32 s13, s13, 0
+; GFX1032-NEXT: s_mov_b32 s2, 0
+; GFX1032-NEXT: s_and_saveexec_b32 s1, vcc_lo
; GFX1032-NEXT: s_cbranch_execz .LBB16_3
; GFX1032-NEXT: ; %bb.1:
-; GFX1032-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1032-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1032-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1032-NEXT: s_mov_b32 s33, s10
-; GFX1032-NEXT: s_mov_b32 s50, s9
-; GFX1032-NEXT: s_mov_b32 s51, s8
-; GFX1032-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1032-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1032-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-NEXT: s_bcnt1_i32_b32 s6, s0
+; GFX1032-NEXT: s_mov_b32 s7, 0x43300000
+; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1032-NEXT: v_add_f64 v[0:1], 0xc3300000, s[6:7]
+; GFX1032-NEXT: v_mov_b32_e32 v6, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
+; GFX1032-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX1032-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: v_mov_b32_e32 v2, s1
-; GFX1032-NEXT: v_mov_b32_e32 v1, s0
+; GFX1032-NEXT: v_mov_b32_e32 v2, s4
+; GFX1032-NEXT: v_mov_b32_e32 v3, s5
; GFX1032-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1032-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1032-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX1032-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1032-NEXT: s_add_u32 s8, s36, 44
-; GFX1032-NEXT: s_addc_u32 s9, s37, 0
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: s_clause 0x1
-; GFX1032-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
+; GFX1032-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1032-NEXT: v_mov_b32_e32 v3, v1
+; GFX1032-NEXT: v_mov_b32_e32 v2, v0
+; GFX1032-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
; GFX1032-NEXT: s_cbranch_execnz .LBB16_2
; GFX1032-NEXT: .LBB16_3:
; GFX1032-NEXT: s_endpgm
;
; GFX1164-LABEL: global_atomic_fsub_double_uni_address_uni_value_default_scope_strictfp:
; GFX1164: ; %bb.0:
-; GFX1164-NEXT: s_mov_b64 s[48:49], s[0:1]
; GFX1164-NEXT: s_bcnt1_i32_b64 s0, exec
-; GFX1164-NEXT: v_mov_b32_e32 v40, v0
; GFX1164-NEXT: v_mov_b32_e32 v0, 0x43300000
; GFX1164-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-NEXT: s_mov_b32 s32, 32
-; GFX1164-NEXT: s_clause 0x1
-; GFX1164-NEXT: scratch_store_b32 off, v0, off offset:20
-; GFX1164-NEXT: scratch_store_b32 off, v1, off offset:16
-; GFX1164-NEXT: scratch_load_b64 v[0:1], off, off offset:16
+; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
; GFX1164-NEXT: s_mov_b64 s[0:1], exec
-; GFX1164-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-NEXT: v_add_f64 v[0:1], 0xc3300000, v[0:1]
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX1164-NEXT: v_mul_f64 v[41:42], 4.0, v[0:1]
-; GFX1164-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX1164-NEXT: s_clause 0x1
+; GFX1164-NEXT: scratch_store_b32 off, v0, off offset:4
+; GFX1164-NEXT: scratch_store_b32 off, v1, off
+; GFX1164-NEXT: scratch_load_b64 v[0:1], off, off
+; GFX1164-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v0
+; GFX1164-NEXT: v_cmpx_eq_u32_e32 0, v2
; GFX1164-NEXT: s_cbranch_execz .LBB16_3
; GFX1164-NEXT: ; %bb.1:
-; GFX1164-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1164-NEXT: s_mov_b32 s33, s10
-; GFX1164-NEXT: s_mov_b32 s50, s9
-; GFX1164-NEXT: s_mov_b32 s51, s8
-; GFX1164-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1164-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1164-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-NEXT: v_add_f64 v[0:1], 0xc3300000, v[0:1]
+; GFX1164-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1164-NEXT: v_mov_b32_e32 v6, 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1164-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
+; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX1164-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: v_mov_b32_e32 v2, s1
-; GFX1164-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-NEXT: .p2align 6
+; GFX1164-NEXT: v_mov_b32_e32 v2, s2
+; GFX1164-NEXT: v_mov_b32_e32 v3, s3
+; GFX1164-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1164-NEXT: s_waitcnt vmcnt(0)
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1164-NEXT: s_add_u32 s8, s36, 44
-; GFX1164-NEXT: s_addc_u32 s9, s37, 0
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1164-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-NEXT: s_cbranch_execnz .LBB16_2
; GFX1164-NEXT: .LBB16_3:
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-NEXT: s_endpgm
;
; GFX1132-LABEL: global_atomic_fsub_double_uni_address_uni_value_default_scope_strictfp:
; GFX1132: ; %bb.0:
-; GFX1132-NEXT: s_mov_b64 s[48:49], s[0:1]
; GFX1132-NEXT: s_bcnt1_i32_b32 s0, exec_lo
; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX1132-NEXT: v_dual_mov_b32 v40, v0 :: v_dual_mov_b32 v1, s0
-; GFX1132-NEXT: v_mov_b32_e32 v0, 0x43300000
-; GFX1132-NEXT: s_mov_b32 s54, 0
-; GFX1132-NEXT: s_clause 0x1
-; GFX1132-NEXT: scratch_store_b32 off, v0, off offset:20
-; GFX1132-NEXT: scratch_store_b32 off, v1, off offset:16
-; GFX1132-NEXT: scratch_load_b64 v[0:1], off, off offset:16
-; GFX1132-NEXT: s_mov_b32 s32, 32
+; GFX1132-NEXT: v_dual_mov_b32 v0, 0x43300000 :: v_dual_mov_b32 v1, s0
+; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
+; GFX1132-NEXT: s_mov_b32 s2, 0
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
-; GFX1132-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-NEXT: v_add_f64 v[0:1], 0xc3300000, v[0:1]
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX1132-NEXT: v_mul_f64 v[41:42], 4.0, v[0:1]
-; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
+; GFX1132-NEXT: s_clause 0x1
+; GFX1132-NEXT: scratch_store_b32 off, v0, off offset:4
+; GFX1132-NEXT: scratch_store_b32 off, v1, off
+; GFX1132-NEXT: scratch_load_b64 v[0:1], off, off
+; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v2
; GFX1132-NEXT: s_cbranch_execz .LBB16_3
; GFX1132-NEXT: ; %bb.1:
-; GFX1132-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1132-NEXT: s_mov_b32 s33, s15
-; GFX1132-NEXT: s_mov_b32 s50, s14
-; GFX1132-NEXT: s_mov_b32 s51, s13
-; GFX1132-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1132-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-NEXT: v_add_f64 v[0:1], 0xc3300000, v[0:1]
+; GFX1132-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1132-NEXT: v_mov_b32_e32 v6, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1132-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
+; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX1132-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-NEXT: .p2align 6
+; GFX1132-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX1132-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1132-NEXT: s_waitcnt vmcnt(0)
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1132-NEXT: s_add_u32 s8, s36, 44
-; GFX1132-NEXT: s_addc_u32 s9, s37, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1132-NEXT: s_mov_b32 s12, s51
-; GFX1132-NEXT: s_mov_b32 s13, s50
-; GFX1132-NEXT: s_mov_b32 s14, s33
-; GFX1132-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1132-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-NEXT: s_cbranch_execnz .LBB16_2
; GFX1132-NEXT: .LBB16_3:
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-NEXT: s_endpgm
;
; GFX7LESS-DPP-LABEL: global_atomic_fsub_double_uni_address_uni_value_default_scope_strictfp:
; GFX7LESS-DPP: ; %bb.0:
-; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v5, exec_lo, 0
-; GFX7LESS-DPP-NEXT: s_bcnt1_i32_b64 s0, exec
-; GFX7LESS-DPP-NEXT: s_mov_b32 s1, 0x43300000
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0xc3300000
-; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v5, exec_hi, v5
-; GFX7LESS-DPP-NEXT: v_add_f64 v[3:4], s[0:1], v[3:4]
-; GFX7LESS-DPP-NEXT: v_mul_f64 v[41:42], 4.0, v[3:4]
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v5
+; GFX7LESS-DPP-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s14, -1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s15, 0xe8f000
+; GFX7LESS-DPP-NEXT: s_add_u32 s12, s12, s11
+; GFX7LESS-DPP-NEXT: s_addc_u32 s13, s13, 0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], exec
+; GFX7LESS-DPP-NEXT: v_mbcnt_lo_u32_b32_e64 v0, s2, 0
+; GFX7LESS-DPP-NEXT: v_mbcnt_hi_u32_b32_e32 v0, s3, v0
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7LESS-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX7LESS-DPP-NEXT: s_cbranch_execz .LBB16_3
; GFX7LESS-DPP-NEXT: ; %bb.1:
-; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-DPP-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_bcnt1_i32_b64 s6, s[2:3]
+; GFX7LESS-DPP-NEXT: s_mov_b32 s7, 0x43300000
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 0
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0xc3300000
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v3, v0, v1
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s3, 0xf000
+; GFX7LESS-DPP-NEXT: v_add_f64 v[0:1], s[6:7], v[0:1]
+; GFX7LESS-DPP-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, s0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, s1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v40, v3, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s8
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s2, -1
; GFX7LESS-DPP-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-DPP-NEXT: v_add_f64 v[2:3], v[0:1], -v[41:42]
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:8
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(2)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-DPP-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[0:3], 0 glc
+; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB16_2
; GFX7LESS-DPP-NEXT: .LBB16_3:
; GFX7LESS-DPP-NEXT: s_endpgm
;
; GFX9-DPP-LABEL: global_atomic_fsub_double_uni_address_uni_value_default_scope_strictfp:
; GFX9-DPP: ; %bb.0:
-; GFX9-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-DPP-NEXT: s_mov_b32 s66, -1
-; GFX9-DPP-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, 0
-; GFX9-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-DPP-NEXT: s_bcnt1_i32_b64 s0, exec
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0xc3300000
-; GFX9-DPP-NEXT: s_mov_b32 s1, 0x43300000
-; GFX9-DPP-NEXT: v_add_f64 v[3:4], s[0:1], v[3:4]
-; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX9-DPP-NEXT: v_mul_f64 v[41:42], 4.0, v[3:4]
-; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
+; GFX9-DPP-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GFX9-DPP-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GFX9-DPP-NEXT: s_mov_b32 s14, -1
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], exec
+; GFX9-DPP-NEXT: s_mov_b32 s15, 0xe00000
+; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
+; GFX9-DPP-NEXT: s_add_u32 s12, s12, s11
+; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, s1, v0
+; GFX9-DPP-NEXT: s_addc_u32 s13, s13, 0
+; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX9-DPP-NEXT: s_and_saveexec_b64 s[2:3], vcc
; GFX9-DPP-NEXT: s_cbranch_execz .LBB16_3
; GFX9-DPP-NEXT: ; %bb.1:
-; GFX9-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX9-DPP-NEXT: s_mov_b32 s33, s10
-; GFX9-DPP-NEXT: s_mov_b32 s50, s9
+; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 0
+; GFX9-DPP-NEXT: s_bcnt1_i32_b64 s0, s[0:1]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0xc3300000
+; GFX9-DPP-NEXT: s_mov_b32 s1, 0x43300000
+; GFX9-DPP-NEXT: v_add_f64 v[0:1], s[0:1], v[0:1]
+; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
-; GFX9-DPP-NEXT: s_mov_b32 s51, s8
-; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX9-DPP-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX9-DPP-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX9-DPP-NEXT: s_mov_b64 s[54:55], 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
+; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s4
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s5
; GFX9-DPP-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX9-DPP-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX9-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-DPP-NEXT: s_cbranch_execnz .LBB16_2
; GFX9-DPP-NEXT: .LBB16_3:
; GFX9-DPP-NEXT: s_endpgm
;
; GFX1064-DPP-LABEL: global_atomic_fsub_double_uni_address_uni_value_default_scope_strictfp:
; GFX1064-DPP: ; %bb.0:
-; GFX1064-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1064-DPP-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX1064-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-DPP-NEXT: s_bcnt1_i32_b64 s0, exec
-; GFX1064-DPP-NEXT: s_mov_b32 s1, 0x43300000
-; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX1064-DPP-NEXT: v_add_f64 v[3:4], 0xc3300000, s[0:1]
-; GFX1064-DPP-NEXT: v_mul_f64 v[41:42], 4.0, v[3:4]
-; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v3, exec_hi, v3
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
-; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
+; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], exec
+; GFX1064-DPP-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
+; GFX1064-DPP-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GFX1064-DPP-NEXT: s_mov_b32 s14, -1
+; GFX1064-DPP-NEXT: s_mov_b32 s15, 0x31e16000
+; GFX1064-DPP-NEXT: s_add_u32 s12, s12, s11
+; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, s1, v0
+; GFX1064-DPP-NEXT: s_addc_u32 s13, s13, 0
+; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
+; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[2:3], vcc
; GFX1064-DPP-NEXT: s_cbranch_execz .LBB16_3
; GFX1064-DPP-NEXT: ; %bb.1:
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1064-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1064-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1064-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1064-DPP-NEXT: s_bcnt1_i32_b64 s2, s[0:1]
+; GFX1064-DPP-NEXT: s_mov_b32 s3, 0x43300000
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1064-DPP-NEXT: v_add_f64 v[0:1], 0xc3300000, s[2:3]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x0
+; GFX1064-DPP-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, s0
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s2
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s3
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1064-DPP-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1064-DPP-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1064-DPP-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1064-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1064-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-DPP-NEXT: s_clause 0x1
-; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX1064-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX1064-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB16_2
; GFX1064-DPP-NEXT: .LBB16_3:
; GFX1064-DPP-NEXT: s_endpgm
;
; GFX1032-DPP-LABEL: global_atomic_fsub_double_uni_address_uni_value_default_scope_strictfp:
; GFX1032-DPP: ; %bb.0:
-; GFX1032-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1032-DPP-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX1032-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-DPP-NEXT: s_bcnt1_i32_b32 s0, exec_lo
-; GFX1032-DPP-NEXT: s_mov_b32 s1, 0x43300000
-; GFX1032-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1032-DPP-NEXT: v_add_f64 v[3:4], 0xc3300000, s[0:1]
-; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
-; GFX1032-DPP-NEXT: v_mul_f64 v[41:42], 4.0, v[3:4]
-; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v3, exec_lo, 0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
-; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
+; GFX1032-DPP-NEXT: s_mov_b32 s0, exec_lo
+; GFX1032-DPP-NEXT: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, s0, 0
+; GFX1032-DPP-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GFX1032-DPP-NEXT: s_mov_b32 s14, -1
+; GFX1032-DPP-NEXT: s_mov_b32 s15, 0x31c16000
+; GFX1032-DPP-NEXT: s_add_u32 s12, s12, s11
+; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
+; GFX1032-DPP-NEXT: s_addc_u32 s13, s13, 0
+; GFX1032-DPP-NEXT: s_mov_b32 s2, 0
+; GFX1032-DPP-NEXT: s_and_saveexec_b32 s1, vcc_lo
; GFX1032-DPP-NEXT: s_cbranch_execz .LBB16_3
; GFX1032-DPP-NEXT: ; %bb.1:
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x24
-; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v3, 20, v2
-; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v4, 10, v1
-; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1032-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1032-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v4, v3
-; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-DPP-NEXT: s_bcnt1_i32_b32 s6, s0
+; GFX1032-DPP-NEXT: s_mov_b32 s7, 0x43300000
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24
+; GFX1032-DPP-NEXT: v_add_f64 v[0:1], 0xc3300000, s[6:7]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[52:53], 0x0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[4:5], s[0:1], 0x0
+; GFX1032-DPP-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, s0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s4
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s5
; GFX1032-DPP-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX1032-DPP-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1032-DPP-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1032-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1032-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-DPP-NEXT: s_clause 0x1
-; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
+; GFX1032-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX1032-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB16_2
; GFX1032-DPP-NEXT: .LBB16_3:
; GFX1032-DPP-NEXT: s_endpgm
;
; GFX1164-DPP-LABEL: global_atomic_fsub_double_uni_address_uni_value_default_scope_strictfp:
; GFX1164-DPP: ; %bb.0:
-; GFX1164-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
; GFX1164-DPP-NEXT: s_bcnt1_i32_b64 s0, exec
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 0x43300000
; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1164-DPP-NEXT: s_clause 0x1
-; GFX1164-DPP-NEXT: scratch_store_b32 off, v0, off offset:20
-; GFX1164-DPP-NEXT: scratch_store_b32 off, v1, off offset:16
-; GFX1164-DPP-NEXT: scratch_load_b64 v[0:1], off, off offset:16
+; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
-; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-DPP-NEXT: v_add_f64 v[0:1], 0xc3300000, v[0:1]
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_mul_f64 v[41:42], 4.0, v[0:1]
-; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v0, exec_hi, v0
+; GFX1164-DPP-NEXT: s_clause 0x1
+; GFX1164-DPP-NEXT: scratch_store_b32 off, v0, off offset:4
+; GFX1164-DPP-NEXT: scratch_store_b32 off, v1, off
+; GFX1164-DPP-NEXT: scratch_load_b64 v[0:1], off, off
+; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v2, exec_hi, v2
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
+; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v2
; GFX1164-DPP-NEXT: s_cbranch_execz .LBB16_3
; GFX1164-DPP-NEXT: ; %bb.1:
-; GFX1164-DPP-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1164-DPP-NEXT: s_mov_b32 s50, s9
-; GFX1164-DPP-NEXT: s_mov_b32 s51, s8
-; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX1164-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-DPP-NEXT: v_add_f64 v[0:1], 0xc3300000, v[0:1]
+; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1164-DPP-NEXT: s_load_b64 s[2:3], s[0:1], 0x0
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX1164-DPP-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, s0
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-DPP-NEXT: .p2align 6
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s2
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s3
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-DPP-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1164-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1164-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-DPP-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB16_2
; GFX1164-DPP-NEXT: .LBB16_3:
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-DPP-NEXT: s_endpgm
;
; GFX1132-DPP-LABEL: global_atomic_fsub_double_uni_address_uni_value_default_scope_strictfp:
; GFX1132-DPP: ; %bb.0:
-; GFX1132-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
; GFX1132-DPP-NEXT: s_bcnt1_i32_b32 s0, exec_lo
; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v40, v0 :: v_dual_mov_b32 v1, s0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v0, 0x43300000
-; GFX1132-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1132-DPP-NEXT: s_clause 0x1
-; GFX1132-DPP-NEXT: scratch_store_b32 off, v0, off offset:20
-; GFX1132-DPP-NEXT: scratch_store_b32 off, v1, off offset:16
-; GFX1132-DPP-NEXT: scratch_load_b64 v[0:1], off, off offset:16
-; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v0, 0x43300000 :: v_dual_mov_b32 v1, s0
+; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v2, exec_lo, 0
+; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
-; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-DPP-NEXT: v_add_f64 v[0:1], 0xc3300000, v[0:1]
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_mul_f64 v[41:42], 4.0, v[0:1]
-; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v0
+; GFX1132-DPP-NEXT: s_clause 0x1
+; GFX1132-DPP-NEXT: scratch_store_b32 off, v0, off offset:4
+; GFX1132-DPP-NEXT: scratch_store_b32 off, v1, off
+; GFX1132-DPP-NEXT: scratch_load_b64 v[0:1], off, off
+; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v2
; GFX1132-DPP-NEXT: s_cbranch_execz .LBB16_3
; GFX1132-DPP-NEXT: ; %bb.1:
-; GFX1132-DPP-NEXT: s_load_b64 s[52:53], s[4:5], 0x24
-; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
-; GFX1132-DPP-NEXT: s_mov_b32 s50, s14
-; GFX1132-DPP-NEXT: s_mov_b32 s51, s13
-; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-DPP-NEXT: v_add_f64 v[0:1], 0xc3300000, v[0:1]
+; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v6, 0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[52:53], 0x0
+; GFX1132-DPP-NEXT: s_load_b64 s[4:5], s[0:1], 0x0
+; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX1132-DPP-NEXT: v_mul_f64 v[4:5], 4.0, v[0:1]
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-DPP-NEXT: .p2align 6
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX1132-DPP-NEXT: .LBB16_2: ; %atomicrmw.start
; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1132-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX1132-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX1132-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1132-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-DPP-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB16_2
; GFX1132-DPP-NEXT: .LBB16_3:
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-DPP-NEXT: s_endpgm
- %result = atomicrmw fsub ptr addrspace(1) %ptr, double 4.0 monotonic, align 4
+ %result = atomicrmw fsub ptr addrspace(1) %ptr, double 4.0 monotonic, align 8
ret void
}
define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_default_scope_strictfp(ptr addrspace(1) %ptr) #2 {
; GFX7LESS-LABEL: global_atomic_fsub_double_uni_address_div_value_default_scope_strictfp:
; GFX7LESS: ; %bb.0:
-; GFX7LESS-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-NEXT: s_mov_b32 s32, 0
+; GFX7LESS-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX7LESS-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX7LESS-NEXT: s_mov_b32 s38, -1
+; GFX7LESS-NEXT: s_mov_b32 s39, 0xe8f000
+; GFX7LESS-NEXT: s_add_u32 s36, s36, s11
+; GFX7LESS-NEXT: s_addc_u32 s37, s37, 0
+; GFX7LESS-NEXT: s_mov_b32 s14, s10
+; GFX7LESS-NEXT: s_mov_b32 s13, s9
+; GFX7LESS-NEXT: s_mov_b32 s12, s8
+; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX7LESS-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX7LESS-NEXT: s_add_u32 s8, s34, 44
+; GFX7LESS-NEXT: s_addc_u32 s9, s35, 0
+; GFX7LESS-NEXT: s_getpc_b64 s[4:5]
+; GFX7LESS-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX7LESS-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
+; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX7LESS-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX7LESS-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GFX7LESS-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX7LESS-NEXT: v_or_b32_e32 v40, v0, v2
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX7LESS-NEXT: v_or_b32_e32 v31, v0, v2
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[38:39]
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX7LESS-NEXT: s_mov_b64 s[0:1], exec
-; GFX7LESS-NEXT: v_mov_b32_e32 v41, 0
-; GFX7LESS-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
+; GFX7LESS-NEXT: v_bfrev_b32_e32 v5, 1
; GFX7LESS-NEXT: .LBB17_1: ; %ComputeLoop
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-NEXT: s_ff1_i32_b64 s4, s[0:1]
@@ -13087,7 +11748,7 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_defau
; GFX7LESS-NEXT: s_andn2_b64 s[0:1], s[0:1], s[4:5]
; GFX7LESS-NEXT: v_cmp_ne_u64_e64 s[4:5], s[0:1], 0
; GFX7LESS-NEXT: s_and_b64 vcc, exec, s[4:5]
-; GFX7LESS-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX7LESS-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX7LESS-NEXT: s_cbranch_vccnz .LBB17_1
; GFX7LESS-NEXT: ; %bb.2: ; %ComputeEnd
; GFX7LESS-NEXT: v_mbcnt_lo_u32_b32_e64 v0, exec_lo, 0
@@ -13097,102 +11758,70 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_defau
; GFX7LESS-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX7LESS-NEXT: s_cbranch_execz .LBB17_5
; GFX7LESS-NEXT: ; %bb.3:
-; GFX7LESS-NEXT: s_load_dwordx2 s[52:53], s[36:37], 0x9
-; GFX7LESS-NEXT: s_mov_b32 s55, 0xf000
-; GFX7LESS-NEXT: s_mov_b32 s54, -1
+; GFX7LESS-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x9
+; GFX7LESS-NEXT: s_mov_b32 s3, 0xf000
+; GFX7LESS-NEXT: s_mov_b32 s2, -1
; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: buffer_load_dwordx2 v[0:1], off, s[52:55], 0
-; GFX7LESS-NEXT: s_mov_b64 s[54:55], 0
+; GFX7LESS-NEXT: buffer_load_dwordx2 v[2:3], off, s[0:3], 0
+; GFX7LESS-NEXT: s_mov_b64 s[4:5], 0
; GFX7LESS-NEXT: .LBB17_4: ; %atomicrmw.start
; GFX7LESS-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-NEXT: v_add_f64 v[2:3], v[0:1], -v[41:42]
-; GFX7LESS-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:12
-; GFX7LESS-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:8
-; GFX7LESS-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX7LESS-NEXT: s_waitcnt expcnt(2)
-; GFX7LESS-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-NEXT: v_mov_b32_e32 v31, v40
-; GFX7LESS-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
; GFX7LESS-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-NEXT: v_mov_b32_e32 v9, v3
+; GFX7LESS-NEXT: v_mov_b32_e32 v8, v2
+; GFX7LESS-NEXT: v_mov_b32_e32 v7, v1
+; GFX7LESS-NEXT: v_mov_b32_e32 v6, v0
+; GFX7LESS-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[0:3], 0 glc
+; GFX7LESS-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[2:3]
+; GFX7LESS-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX7LESS-NEXT: v_mov_b32_e32 v2, v6
+; GFX7LESS-NEXT: v_mov_b32_e32 v3, v7
+; GFX7LESS-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX7LESS-NEXT: s_cbranch_execnz .LBB17_4
; GFX7LESS-NEXT: .LBB17_5:
; GFX7LESS-NEXT: s_endpgm
;
; GFX9-LABEL: global_atomic_fsub_double_uni_address_div_value_default_scope_strictfp:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX9-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX9-NEXT: s_mov_b32 s66, -1
-; GFX9-NEXT: s_mov_b32 s67, 0xe00000
-; GFX9-NEXT: s_add_u32 s64, s64, s11
-; GFX9-NEXT: s_addc_u32 s65, s65, 0
-; GFX9-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-NEXT: s_mov_b32 s51, s8
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_mov_b32 s50, s9
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX9-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX9-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX9-NEXT: s_mov_b32 s38, -1
+; GFX9-NEXT: s_mov_b32 s39, 0xe00000
+; GFX9-NEXT: s_add_u32 s36, s36, s11
+; GFX9-NEXT: s_addc_u32 s37, s37, 0
+; GFX9-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX9-NEXT: s_mov_b32 s12, s8
+; GFX9-NEXT: s_add_u32 s8, s34, 44
+; GFX9-NEXT: s_mov_b32 s13, s9
+; GFX9-NEXT: s_addc_u32 s9, s35, 0
+; GFX9-NEXT: s_getpc_b64 s[4:5]
+; GFX9-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX9-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
+; GFX9-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
+; GFX9-NEXT: s_mov_b32 s14, s10
+; GFX9-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-NEXT: s_mov_b32 s33, s10
-; GFX9-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX9-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: s_movk_i32 s32, 0x800
+; GFX9-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX9-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX9-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX9-NEXT: s_mov_b32 s32, 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: v_mov_b32_e32 v41, 0
+; GFX9-NEXT: v_mov_b32_e32 v4, 0
; GFX9-NEXT: s_mov_b64 s[0:1], exec
-; GFX9-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX9-NEXT: v_bfrev_b32_e32 v5, 1
; GFX9-NEXT: .LBB17_1: ; %ComputeLoop
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-NEXT: s_ff1_i32_b64 s4, s[0:1]
; GFX9-NEXT: v_readlane_b32 s3, v1, s4
; GFX9-NEXT: v_readlane_b32 s2, v0, s4
-; GFX9-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX9-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX9-NEXT: s_lshl_b64 s[2:3], 1, s4
; GFX9-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
; GFX9-NEXT: s_cmp_lg_u64 s[0:1], 0
@@ -13205,99 +11834,64 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_defau
; GFX9-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX9-NEXT: s_cbranch_execz .LBB17_5
; GFX9-NEXT: ; %bb.3:
-; GFX9-NEXT: s_load_dwordx2 s[52:53], s[36:37], 0x24
-; GFX9-NEXT: v_mov_b32_e32 v0, 0
-; GFX9-NEXT: s_mov_b64 s[54:55], 0
+; GFX9-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX9-NEXT: v_mov_b32_e32 v6, 0
+; GFX9-NEXT: s_mov_b64 s[2:3], 0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX9-NEXT: global_load_dwordx2 v[2:3], v6, s[0:1]
; GFX9-NEXT: .LBB17_4: ; %atomicrmw.start
; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX9-NEXT: s_add_u32 s8, s36, 44
-; GFX9-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-NEXT: s_getpc_b64 s[0:1]
-; GFX9-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX9-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX9-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX9-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX9-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-NEXT: s_mov_b32 s12, s51
-; GFX9-NEXT: s_mov_b32 s13, s50
-; GFX9-NEXT: s_mov_b32 s14, s33
-; GFX9-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX9-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-NEXT: v_mov_b32_e32 v2, s52
-; GFX9-NEXT: v_mov_b32_e32 v3, s53
-; GFX9-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX9-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX9-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX9-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX9-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX9-NEXT: s_waitcnt vmcnt(0)
+; GFX9-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v3, v1
+; GFX9-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX9-NEXT: v_mov_b32_e32 v2, v0
+; GFX9-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX9-NEXT: s_cbranch_execnz .LBB17_4
; GFX9-NEXT: .LBB17_5:
; GFX9-NEXT: s_endpgm
;
; GFX1064-LABEL: global_atomic_fsub_double_uni_address_div_value_default_scope_strictfp:
; GFX1064: ; %bb.0:
-; GFX1064-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-NEXT: s_mov_b32 s66, -1
-; GFX1064-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-NEXT: s_add_u32 s64, s64, s11
+; GFX1064-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1064-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1064-NEXT: s_mov_b32 s38, -1
+; GFX1064-NEXT: s_mov_b32 s39, 0x31e16000
+; GFX1064-NEXT: s_add_u32 s36, s36, s11
; GFX1064-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1064-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-NEXT: s_mov_b32 s51, s8
+; GFX1064-NEXT: s_addc_u32 s37, s37, 0
+; GFX1064-NEXT: s_mov_b32 s12, s8
; GFX1064-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-NEXT: s_mov_b32 s50, s9
+; GFX1064-NEXT: s_mov_b32 s13, s9
; GFX1064-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1064-NEXT: s_getpc_b64 s[4:5]
+; GFX1064-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1064-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1064-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1064-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1064-NEXT: s_mov_b32 s33, s10
-; GFX1064-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1064-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1064-NEXT: s_mov_b32 s14, s10
+; GFX1064-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1064-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1064-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1064-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: s_movk_i32 s32, 0x800
+; GFX1064-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1064-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1064-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1064-NEXT: s_mov_b32 s32, 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: v_mov_b32_e32 v41, 0
-; GFX1064-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX1064-NEXT: v_mov_b32_e32 v4, 0
+; GFX1064-NEXT: v_bfrev_b32_e32 v5, 1
; GFX1064-NEXT: s_mov_b64 s[0:1], exec
; GFX1064-NEXT: .LBB17_1: ; %ComputeLoop
; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1064-NEXT: s_ff1_i32_b64 s4, s[0:1]
; GFX1064-NEXT: v_readlane_b32 s3, v1, s4
; GFX1064-NEXT: v_readlane_b32 s2, v0, s4
-; GFX1064-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX1064-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX1064-NEXT: s_lshl_b64 s[2:3], 1, s4
; GFX1064-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3]
; GFX1064-NEXT: s_cmp_lg_u64 s[0:1], 0
@@ -13310,93 +11904,57 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_defau
; GFX1064-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX1064-NEXT: s_cbranch_execz .LBB17_5
; GFX1064-NEXT: ; %bb.3:
-; GFX1064-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1064-NEXT: v_mov_b32_e32 v0, 0
-; GFX1064-NEXT: s_mov_b64 s[54:55], 0
+; GFX1064-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX1064-NEXT: v_mov_b32_e32 v6, 0
+; GFX1064-NEXT: s_mov_b64 s[2:3], 0
; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX1064-NEXT: global_load_dwordx2 v[2:3], v6, s[0:1]
; GFX1064-NEXT: .LBB17_4: ; %atomicrmw.start
; GFX1064-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1064-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1064-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-NEXT: s_mov_b32 s12, s51
-; GFX1064-NEXT: s_mov_b32 s13, s50
-; GFX1064-NEXT: s_mov_b32 s14, s33
-; GFX1064-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-NEXT: s_clause 0x1
-; GFX1064-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX1064-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1064-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1064-NEXT: s_waitcnt vmcnt(0)
+; GFX1064-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1064-NEXT: v_mov_b32_e32 v3, v1
+; GFX1064-NEXT: v_mov_b32_e32 v2, v0
+; GFX1064-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1064-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX1064-NEXT: s_cbranch_execnz .LBB17_4
; GFX1064-NEXT: .LBB17_5:
; GFX1064-NEXT: s_endpgm
;
; GFX1032-LABEL: global_atomic_fsub_double_uni_address_div_value_default_scope_strictfp:
; GFX1032: ; %bb.0:
-; GFX1032-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-NEXT: s_mov_b32 s66, -1
-; GFX1032-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-NEXT: s_add_u32 s64, s64, s11
+; GFX1032-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1032-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1032-NEXT: s_mov_b32 s38, -1
+; GFX1032-NEXT: s_mov_b32 s39, 0x31c16000
+; GFX1032-NEXT: s_add_u32 s36, s36, s11
; GFX1032-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1032-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-NEXT: s_mov_b32 s51, s8
+; GFX1032-NEXT: s_addc_u32 s37, s37, 0
+; GFX1032-NEXT: s_mov_b32 s12, s8
; GFX1032-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-NEXT: s_mov_b32 s50, s9
+; GFX1032-NEXT: s_mov_b32 s13, s9
; GFX1032-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1032-NEXT: s_getpc_b64 s[4:5]
+; GFX1032-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1032-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1032-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1032-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1032-NEXT: s_mov_b32 s33, s10
-; GFX1032-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1032-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-NEXT: s_mov_b32 s14, s10
+; GFX1032-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1032-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1032-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1032-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1032-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1032-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1032-NEXT: s_mov_b32 s32, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: v_mov_b32_e32 v41, 0
-; GFX1032-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX1032-NEXT: v_mov_b32_e32 v4, 0
+; GFX1032-NEXT: v_bfrev_b32_e32 v5, 1
; GFX1032-NEXT: s_mov_b32 s0, exec_lo
; GFX1032-NEXT: .LBB17_1: ; %ComputeLoop
; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
@@ -13405,61 +11963,32 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_defau
; GFX1032-NEXT: v_readlane_b32 s2, v0, s1
; GFX1032-NEXT: s_lshl_b32 s1, 1, s1
; GFX1032-NEXT: s_andn2_b32 s0, s0, s1
-; GFX1032-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX1032-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX1032-NEXT: s_cmp_lg_u32 s0, 0
; GFX1032-NEXT: s_cbranch_scc1 .LBB17_1
; GFX1032-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1032-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1032-NEXT: s_mov_b32 s54, 0
+; GFX1032-NEXT: s_mov_b32 s2, 0
; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX1032-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX1032-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX1032-NEXT: s_cbranch_execz .LBB17_5
; GFX1032-NEXT: ; %bb.3:
-; GFX1032-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1032-NEXT: v_mov_b32_e32 v0, 0
+; GFX1032-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX1032-NEXT: v_mov_b32_e32 v6, 0
; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX1032-NEXT: global_load_dwordx2 v[2:3], v6, s[0:1]
; GFX1032-NEXT: .LBB17_4: ; %atomicrmw.start
; GFX1032-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1032-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1032-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-NEXT: s_mov_b32 s12, s51
-; GFX1032-NEXT: s_mov_b32 s13, s50
-; GFX1032-NEXT: s_mov_b32 s14, s33
-; GFX1032-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-NEXT: s_clause 0x1
-; GFX1032-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
+; GFX1032-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1032-NEXT: global_atomic_cmpswap_x2 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1032-NEXT: s_waitcnt vmcnt(0)
+; GFX1032-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1032-NEXT: v_mov_b32_e32 v3, v1
+; GFX1032-NEXT: v_mov_b32_e32 v2, v0
+; GFX1032-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1032-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
; GFX1032-NEXT: s_cbranch_execnz .LBB17_4
; GFX1032-NEXT: .LBB17_5:
; GFX1032-NEXT: s_endpgm
@@ -13467,31 +11996,24 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_defau
; GFX1164-LABEL: global_atomic_fsub_double_uni_address_div_value_default_scope_strictfp:
; GFX1164: ; %bb.0:
; GFX1164-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1164-NEXT: s_mov_b32 s51, s8
+; GFX1164-NEXT: s_mov_b32 s12, s8
; GFX1164-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-NEXT: s_mov_b32 s50, s9
+; GFX1164-NEXT: s_mov_b32 s13, s9
; GFX1164-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1164-NEXT: s_getpc_b64 s[4:5]
+; GFX1164-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1164-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1164-NEXT: v_mov_b32_e32 v31, v0
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: s_mov_b32 s33, s10
-; GFX1164-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX1164-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1164-NEXT: s_mov_b32 s14, s10
+; GFX1164-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1164-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1164-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: s_mov_b32 s32, 32
-; GFX1164-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1164-NEXT: s_mov_b32 s32, 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: v_mov_b32_e32 v41, 0
-; GFX1164-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX1164-NEXT: s_swappc_b64 s[30:31], s[16:17]
+; GFX1164-NEXT: v_mov_b32_e32 v4, 0
+; GFX1164-NEXT: v_bfrev_b32_e32 v5, 1
; GFX1164-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-NEXT: .LBB17_1: ; %ComputeLoop
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
@@ -13500,7 +12022,7 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_defau
; GFX1164-NEXT: v_readlane_b32 s3, v1, s4
; GFX1164-NEXT: v_readlane_b32 s2, v0, s4
; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX1164-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX1164-NEXT: s_lshl_b64 s[2:3], 1, s4
; GFX1164-NEXT: s_and_not1_b64 s[0:1], s[0:1], s[2:3]
; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
@@ -13515,81 +12037,48 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_defau
; GFX1164-NEXT: s_xor_b64 s[0:1], exec, s[0:1]
; GFX1164-NEXT: s_cbranch_execz .LBB17_5
; GFX1164-NEXT: ; %bb.3:
-; GFX1164-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1164-NEXT: v_mov_b32_e32 v0, 0
-; GFX1164-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1164-NEXT: v_mov_b32_e32 v6, 0
+; GFX1164-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-NEXT: .p2align 6
+; GFX1164-NEXT: global_load_b64 v[2:3], v6, s[0:1]
; GFX1164-NEXT: .LBB17_4: ; %atomicrmw.start
; GFX1164-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1164-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-NEXT: s_mov_b32 s12, s51
-; GFX1164-NEXT: s_mov_b32 s13, s50
-; GFX1164-NEXT: s_mov_b32 s14, s33
-; GFX1164-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1164-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1164-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[2:3]
+; GFX1164-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-NEXT: v_mov_b32_e32 v2, v0
+; GFX1164-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-NEXT: s_cbranch_execnz .LBB17_4
; GFX1164-NEXT: .LBB17_5:
-; GFX1164-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-NEXT: s_endpgm
;
; GFX1132-LABEL: global_atomic_fsub_double_uni_address_div_value_default_scope_strictfp:
; GFX1132: ; %bb.0:
; GFX1132-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1132-NEXT: s_mov_b64 s[48:49], s[0:1]
+; GFX1132-NEXT: v_mov_b32_e32 v31, v0
; GFX1132-NEXT: s_add_u32 s8, s34, 44
; GFX1132-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX1132-NEXT: v_mov_b32_e32 v31, v0
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1132-NEXT: s_mov_b32 s50, s14
-; GFX1132-NEXT: s_mov_b32 s51, s13
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
+; GFX1132-NEXT: s_getpc_b64 s[4:5]
+; GFX1132-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1132-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1132-NEXT: s_mov_b32 s12, s13
+; GFX1132-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1132-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1132-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GFX1132-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX1132-NEXT: s_mov_b32 s13, s14
; GFX1132-NEXT: s_mov_b32 s14, s15
-; GFX1132-NEXT: s_mov_b32 s32, 32
-; GFX1132-NEXT: s_mov_b32 s33, s15
-; GFX1132-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-NEXT: s_mov_b32 s32, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: v_mov_b32_e32 v41, 0
-; GFX1132-NEXT: v_bfrev_b32_e32 v42, 1
+; GFX1132-NEXT: s_swappc_b64 s[30:31], s[16:17]
+; GFX1132-NEXT: v_mov_b32_e32 v4, 0
+; GFX1132-NEXT: v_bfrev_b32_e32 v5, 1
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-NEXT: .LBB17_1: ; %ComputeLoop
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
@@ -13600,522 +12089,358 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_defau
; GFX1132-NEXT: s_lshl_b32 s1, 1, s1
; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX1132-NEXT: s_and_not1_b32 s0, s0, s1
-; GFX1132-NEXT: v_add_f64 v[41:42], v[41:42], s[2:3]
+; GFX1132-NEXT: v_add_f64 v[4:5], v[4:5], s[2:3]
; GFX1132-NEXT: s_cmp_lg_u32 s0, 0
; GFX1132-NEXT: s_cbranch_scc1 .LBB17_1
; GFX1132-NEXT: ; %bb.2: ; %ComputeEnd
; GFX1132-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1132-NEXT: s_mov_b32 s54, 0
+; GFX1132-NEXT: s_mov_b32 s2, 0
; GFX1132-NEXT: s_mov_b32 s0, exec_lo
; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX1132-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX1132-NEXT: s_cbranch_execz .LBB17_5
; GFX1132-NEXT: ; %bb.3:
-; GFX1132-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1132-NEXT: v_mov_b32_e32 v0, 0
+; GFX1132-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1132-NEXT: v_mov_b32_e32 v6, 0
; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-NEXT: .p2align 6
+; GFX1132-NEXT: global_load_b64 v[2:3], v6, s[0:1]
; GFX1132-NEXT: .LBB17_4: ; %atomicrmw.start
; GFX1132-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1132-NEXT: s_add_u32 s8, s34, 44
-; GFX1132-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1132-NEXT: s_mov_b32 s12, s51
-; GFX1132-NEXT: s_mov_b32 s13, s50
-; GFX1132-NEXT: s_mov_b32 s14, s33
-; GFX1132-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-NEXT: v_add_f64 v[0:1], v[2:3], -v[4:5]
+; GFX1132-NEXT: global_atomic_cmpswap_b64 v[0:1], v6, v[0:3], s[0:1] glc
+; GFX1132-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[2:3]
+; GFX1132-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
+; GFX1132-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-NEXT: s_cbranch_execnz .LBB17_4
; GFX1132-NEXT: .LBB17_5:
-; GFX1132-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-NEXT: s_endpgm
;
; GFX7LESS-DPP-LABEL: global_atomic_fsub_double_uni_address_div_value_default_scope_strictfp:
; GFX7LESS-DPP: ; %bb.0:
-; GFX7LESS-DPP-NEXT: s_movk_i32 s32, 0x800
-; GFX7LESS-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s66, -1
-; GFX7LESS-DPP-NEXT: s_mov_b32 s67, 0xe8f000
-; GFX7LESS-DPP-NEXT: s_add_u32 s64, s64, s11
-; GFX7LESS-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX7LESS-DPP-NEXT: s_mov_b32 s33, s10
-; GFX7LESS-DPP-NEXT: s_mov_b32 s50, s9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s51, s8
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[52:53], s[4:5], 0x9
-; GFX7LESS-DPP-NEXT: s_mov_b32 s55, 0xf000
-; GFX7LESS-DPP-NEXT: s_mov_b32 s54, -1
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s32, 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s48, SCRATCH_RSRC_DWORD0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s49, SCRATCH_RSRC_DWORD1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s50, -1
+; GFX7LESS-DPP-NEXT: s_mov_b32 s51, 0xe8f000
+; GFX7LESS-DPP-NEXT: s_add_u32 s48, s48, s11
+; GFX7LESS-DPP-NEXT: s_addc_u32 s49, s49, 0
+; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s10
+; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s8
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[36:37], s[4:5], 0x9
+; GFX7LESS-DPP-NEXT: s_mov_b32 s39, 0xf000
+; GFX7LESS-DPP-NEXT: s_mov_b32 s38, -1
+; GFX7LESS-DPP-NEXT: s_add_u32 s8, s4, 44
+; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s5, 0
+; GFX7LESS-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX7LESS-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX7LESS-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
+; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX7LESS-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GFX7LESS-DPP-NEXT: v_or_b32_e32 v0, v0, v1
-; GFX7LESS-DPP-NEXT: v_or_b32_e32 v42, v0, v2
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX7LESS-DPP-NEXT: v_or_b32_e32 v31, v0, v2
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v42
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[48:49]
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[50:51]
; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v41, v1
-; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[0:1], off, s[52:55], 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX7LESS-DPP-NEXT: buffer_load_dwordx2 v[4:5], off, s[36:39], 0
+; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], 0
; GFX7LESS-DPP-NEXT: .LBB17_1: ; %atomicrmw.start
; GFX7LESS-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX7LESS-DPP-NEXT: v_add_f64 v[2:3], v[0:1], -v[40:41]
-; GFX7LESS-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: buffer_store_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX7LESS-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:12
-; GFX7LESS-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:8
-; GFX7LESS-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX7LESS-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX7LESS-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX7LESS-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX7LESS-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(2)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX7LESS-DPP-NEXT: s_mov_b32 s12, s51
-; GFX7LESS-DPP-NEXT: s_mov_b32 s13, s50
-; GFX7LESS-DPP-NEXT: s_mov_b32 s14, s33
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v31, v42
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX7LESS-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
+; GFX7LESS-DPP-NEXT: v_add_f64 v[2:3], v[4:5], -v[0:1]
; GFX7LESS-DPP-NEXT: s_waitcnt expcnt(0)
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX7LESS-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX7LESS-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX7LESS-DPP-NEXT: v_and_b32_e32 v2, 1, v0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v0, off, s[64:67], 0
-; GFX7LESS-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0 offset:4
-; GFX7LESS-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
-; GFX7LESS-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v9, v5
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v8, v4
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v7, v3
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v6, v2
+; GFX7LESS-DPP-NEXT: buffer_atomic_cmpswap_x2 v[6:9], off, s[36:39], 0 glc
+; GFX7LESS-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX7LESS-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[4:5]
+; GFX7LESS-DPP-NEXT: s_or_b64 s[0:1], vcc, s[0:1]
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v4, v6
+; GFX7LESS-DPP-NEXT: v_mov_b32_e32 v5, v7
+; GFX7LESS-DPP-NEXT: s_andn2_b64 exec, exec, s[0:1]
; GFX7LESS-DPP-NEXT: s_cbranch_execnz .LBB17_1
; GFX7LESS-DPP-NEXT: ; %bb.2: ; %atomicrmw.end
; GFX7LESS-DPP-NEXT: s_endpgm
;
; GFX9-DPP-LABEL: global_atomic_fsub_double_uni_address_div_value_default_scope_strictfp:
; GFX9-DPP: ; %bb.0:
-; GFX9-DPP-NEXT: s_mov_b32 s68, SCRATCH_RSRC_DWORD0
-; GFX9-DPP-NEXT: s_mov_b32 s69, SCRATCH_RSRC_DWORD1
-; GFX9-DPP-NEXT: s_mov_b32 s70, -1
-; GFX9-DPP-NEXT: s_mov_b32 s71, 0xe00000
-; GFX9-DPP-NEXT: s_add_u32 s68, s68, s11
-; GFX9-DPP-NEXT: s_addc_u32 s69, s69, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[36:37], s[4:5]
-; GFX9-DPP-NEXT: s_mov_b32 s51, s8
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_mov_b32 s50, s9
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX9-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX9-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX9-DPP-NEXT: s_mov_b32 s38, -1
+; GFX9-DPP-NEXT: s_mov_b32 s39, 0xe00000
+; GFX9-DPP-NEXT: s_add_u32 s36, s36, s11
+; GFX9-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
+; GFX9-DPP-NEXT: s_mov_b32 s12, s8
+; GFX9-DPP-NEXT: s_add_u32 s8, s34, 44
+; GFX9-DPP-NEXT: s_mov_b32 s13, s9
+; GFX9-DPP-NEXT: s_addc_u32 s9, s35, 0
+; GFX9-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX9-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX9-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
+; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
+; GFX9-DPP-NEXT: s_mov_b32 s14, s10
+; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX9-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX9-DPP-NEXT: s_mov_b32 s33, s10
-; GFX9-DPP-NEXT: s_mov_b64 s[34:35], s[6:7]
-; GFX9-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
-; GFX9-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
+; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[68:69]
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[70:71]
-; GFX9-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX9-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX9-DPP-NEXT: s_mov_b32 s32, 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX9-DPP-NEXT: v_cndmask_b32_e64 v11, v9, v1, s[0:1]
-; GFX9-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:1 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:1 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v4, 1
+; GFX9-DPP-NEXT: v_cndmask_b32_e64 v6, v4, v1, s[0:1]
+; GFX9-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
+; GFX9-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:1 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:1 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:2 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:2 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:2 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:2 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:4 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:4 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:4 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:4 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_shr:8 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_shr:8 row_mask:0xf bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
-; GFX9-DPP-NEXT: v_bfrev_b32_e32 v13, 1
-; GFX9-DPP-NEXT: v_mov_b32_e32 v12, 0
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_shr:8 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_shr:8 row_mask:0xf bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
+; GFX9-DPP-NEXT: v_bfrev_b32_e32 v8, 1
+; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
; GFX9-DPP-NEXT: s_nop 0
-; GFX9-DPP-NEXT: v_mov_b32_dpp v13, v11 row_bcast:15 row_mask:0xa bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v12, v10 row_bcast:15 row_mask:0xa bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[10:11], v[10:11], v[12:13]
+; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v6 row_bcast:15 row_mask:0xa bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v7, v5 row_bcast:15 row_mask:0xa bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[5:6], v[5:6], v[7:8]
; GFX9-DPP-NEXT: s_nop 1
-; GFX9-DPP-NEXT: v_mov_b32_dpp v9, v11 row_bcast:31 row_mask:0xc bank_mask:0xf
-; GFX9-DPP-NEXT: v_mov_b32_dpp v8, v10 row_bcast:31 row_mask:0xc bank_mask:0xf
-; GFX9-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
+; GFX9-DPP-NEXT: v_mov_b32_dpp v4, v6 row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9-DPP-NEXT: v_mov_b32_dpp v3, v5 row_bcast:31 row_mask:0xc bank_mask:0xf
+; GFX9-DPP-NEXT: v_add_f64 v[3:4], v[5:6], v[3:4]
; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 0
; GFX9-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
; GFX9-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v1
; GFX9-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX9-DPP-NEXT: v_readlane_b32 s53, v9, 63
-; GFX9-DPP-NEXT: v_readlane_b32 s52, v8, 63
+; GFX9-DPP-NEXT: v_readlane_b32 s3, v4, 63
+; GFX9-DPP-NEXT: v_readlane_b32 s2, v3, 63
; GFX9-DPP-NEXT: s_mov_b64 exec, s[0:1]
+; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[2:3]
; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
-; GFX9-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
+; GFX9-DPP-NEXT: s_and_saveexec_b64 s[2:3], vcc
; GFX9-DPP-NEXT: s_cbranch_execz .LBB17_3
; GFX9-DPP-NEXT: ; %bb.1:
-; GFX9-DPP-NEXT: s_load_dwordx2 s[54:55], s[36:37], 0x24
-; GFX9-DPP-NEXT: s_mov_b64 s[64:65], 0
+; GFX9-DPP-NEXT: s_load_dwordx2 s[2:3], s[34:35], 0x24
+; GFX9-DPP-NEXT: s_mov_b64 s[4:5], 0
; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[54:55]
+; GFX9-DPP-NEXT: global_load_dwordx2 v[11:12], v0, s[2:3]
; GFX9-DPP-NEXT: .LBB17_2: ; %atomicrmw.start
; GFX9-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX9-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -s[52:53]
-; GFX9-DPP-NEXT: s_add_u32 s8, s36, 44
-; GFX9-DPP-NEXT: s_addc_u32 s9, s37, 0
-; GFX9-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX9-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX9-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX9-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX9-DPP-NEXT: s_mov_b64 s[0:1], s[68:69]
-; GFX9-DPP-NEXT: buffer_store_dword v2, off, s[68:71], 0 offset:4
-; GFX9-DPP-NEXT: buffer_store_dword v1, off, s[68:71], 0
-; GFX9-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX9-DPP-NEXT: buffer_store_dword v4, off, s[68:71], 0 offset:12
-; GFX9-DPP-NEXT: buffer_store_dword v3, off, s[68:71], 0 offset:8
-; GFX9-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX9-DPP-NEXT: s_mov_b64 s[10:11], s[34:35]
-; GFX9-DPP-NEXT: s_mov_b32 s12, s51
-; GFX9-DPP-NEXT: s_mov_b32 s13, s50
-; GFX9-DPP-NEXT: s_mov_b32 s14, s33
-; GFX9-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX9-DPP-NEXT: s_mov_b64 s[2:3], s[70:71]
-; GFX9-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v2, s54
-; GFX9-DPP-NEXT: v_mov_b32_e32 v3, s55
-; GFX9-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX9-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX9-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX9-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX9-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX9-DPP-NEXT: buffer_load_dword v1, off, s[68:71], 0
-; GFX9-DPP-NEXT: buffer_load_dword v2, off, s[68:71], 0 offset:4
-; GFX9-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX9-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX9-DPP-NEXT: s_or_b64 s[64:65], vcc, s[64:65]
-; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[64:65]
+; GFX9-DPP-NEXT: v_add_f64 v[9:10], v[11:12], -s[0:1]
+; GFX9-DPP-NEXT: global_atomic_cmpswap_x2 v[1:2], v0, v[9:12], s[2:3] glc
+; GFX9-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX9-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[1:2], v[11:12]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v12, v2
+; GFX9-DPP-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
+; GFX9-DPP-NEXT: v_mov_b32_e32 v11, v1
+; GFX9-DPP-NEXT: s_andn2_b64 exec, exec, s[4:5]
; GFX9-DPP-NEXT: s_cbranch_execnz .LBB17_2
; GFX9-DPP-NEXT: .LBB17_3:
; GFX9-DPP-NEXT: s_endpgm
;
; GFX1064-DPP-LABEL: global_atomic_fsub_double_uni_address_div_value_default_scope_strictfp:
; GFX1064-DPP: ; %bb.0:
-; GFX1064-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1064-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1064-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1064-DPP-NEXT: s_mov_b32 s67, 0x31e16000
-; GFX1064-DPP-NEXT: s_add_u32 s64, s64, s11
+; GFX1064-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1064-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1064-DPP-NEXT: s_mov_b32 s38, -1
+; GFX1064-DPP-NEXT: s_mov_b32 s39, 0x31e16000
+; GFX1064-DPP-NEXT: s_add_u32 s36, s36, s11
; GFX1064-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1064-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1064-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1064-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX1064-DPP-NEXT: s_mov_b32 s12, s8
; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1064-DPP-NEXT: s_mov_b32 s13, s9
; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1064-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1064-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1064-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1064-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1064-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1064-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1064-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1064-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1064-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: s_movk_i32 s32, 0x800
+; GFX1064-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1064-DPP-NEXT: s_mov_b32 s32, 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1064-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v11, 0x80000000, v1, s[0:1]
-; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
-; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1064-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1064-DPP-NEXT: v_readlane_b32 s3, v9, 0
-; GFX1064-DPP-NEXT: v_readlane_b32 s2, v8, 0
-; GFX1064-DPP-NEXT: v_readlane_b32 s5, v9, 32
-; GFX1064-DPP-NEXT: v_readlane_b32 s4, v8, 32
-; GFX1064-DPP-NEXT: v_add_f64 v[8:9], s[2:3], s[4:5]
+; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v4, 1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v6, 0x80000000, v1, s[0:1]
+; GFX1064-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s[0:1]
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[5:6], v[3:4]
+; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
+; GFX1064-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1064-DPP-NEXT: v_readlane_b32 s3, v4, 0
+; GFX1064-DPP-NEXT: v_readlane_b32 s2, v3, 0
+; GFX1064-DPP-NEXT: v_readlane_b32 s5, v4, 32
+; GFX1064-DPP-NEXT: v_readlane_b32 s4, v3, 32
+; GFX1064-DPP-NEXT: v_add_f64 v[3:4], s[2:3], s[4:5]
; GFX1064-DPP-NEXT: s_mov_b64 exec, s[0:1]
; GFX1064-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v41, v8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v42, v9
-; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1064-DPP-NEXT: v_mbcnt_hi_u32_b32 v7, exec_hi, v0
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, v3
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, v4
+; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 0, v7
; GFX1064-DPP-NEXT: s_and_saveexec_b64 s[0:1], vcc
; GFX1064-DPP-NEXT: s_cbranch_execz .LBB17_3
; GFX1064-DPP-NEXT: ; %bb.1:
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
-; GFX1064-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1064-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
+; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX1064-DPP-NEXT: global_load_dwordx2 v[9:10], v2, s[0:1]
; GFX1064-DPP-NEXT: .LBB17_2: ; %atomicrmw.start
; GFX1064-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1064-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1064-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1064-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1064-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1064-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1064-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1064-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1064-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1064-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1064-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1064-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1064-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1064-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1064-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1064-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1064-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1064-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1064-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1064-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1064-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1064-DPP-NEXT: s_clause 0x1
-; GFX1064-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1064-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1064-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1064-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1064-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[54:55]
+; GFX1064-DPP-NEXT: v_add_f64 v[7:8], v[9:10], -v[0:1]
+; GFX1064-DPP-NEXT: global_atomic_cmpswap_x2 v[7:8], v2, v[7:10], s[0:1] glc
+; GFX1064-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1064-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[7:8], v[9:10]
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v10, v8
+; GFX1064-DPP-NEXT: v_mov_b32_e32 v9, v7
+; GFX1064-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1064-DPP-NEXT: s_andn2_b64 exec, exec, s[2:3]
; GFX1064-DPP-NEXT: s_cbranch_execnz .LBB17_2
; GFX1064-DPP-NEXT: .LBB17_3:
; GFX1064-DPP-NEXT: s_endpgm
;
; GFX1032-DPP-LABEL: global_atomic_fsub_double_uni_address_div_value_default_scope_strictfp:
; GFX1032-DPP: ; %bb.0:
-; GFX1032-DPP-NEXT: s_mov_b32 s64, SCRATCH_RSRC_DWORD0
-; GFX1032-DPP-NEXT: s_mov_b32 s65, SCRATCH_RSRC_DWORD1
-; GFX1032-DPP-NEXT: s_mov_b32 s66, -1
-; GFX1032-DPP-NEXT: s_mov_b32 s67, 0x31c16000
-; GFX1032-DPP-NEXT: s_add_u32 s64, s64, s11
+; GFX1032-DPP-NEXT: s_mov_b32 s36, SCRATCH_RSRC_DWORD0
+; GFX1032-DPP-NEXT: s_mov_b32 s37, SCRATCH_RSRC_DWORD1
+; GFX1032-DPP-NEXT: s_mov_b32 s38, -1
+; GFX1032-DPP-NEXT: s_mov_b32 s39, 0x31c16000
+; GFX1032-DPP-NEXT: s_add_u32 s36, s36, s11
; GFX1032-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1032-DPP-NEXT: s_addc_u32 s65, s65, 0
-; GFX1032-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1032-DPP-NEXT: s_addc_u32 s37, s37, 0
+; GFX1032-DPP-NEXT: s_mov_b32 s12, s8
; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1032-DPP-NEXT: s_mov_b32 s13, s9
; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1032-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1032-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1032-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v2, 20, v2
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX1032-DPP-NEXT: v_lshlrev_b32_e32 v1, 10, v1
-; GFX1032-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1032-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1032-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1032-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1032-DPP-NEXT: v_or3_b32 v40, v0, v1, v2
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: s_movk_i32 s32, 0x400
+; GFX1032-DPP-NEXT: v_or3_b32 v31, v0, v1, v2
+; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[36:37]
+; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[38:39]
+; GFX1032-DPP-NEXT: s_mov_b32 s32, 0
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1032-DPP-NEXT: s_or_saveexec_b32 s0, -1
-; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v11, 0x80000000, v1, s0
-; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
-; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1032-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1032-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
-; GFX1032-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v4, 1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, 0
+; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v6, 0x80000000, v1, s0
+; GFX1032-DPP-NEXT: v_cndmask_b32_e64 v5, 0, v0, s0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v4, v6 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[5:6], v[3:4]
+; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_bfrev_b32_e32 v6, 1
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 0
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v6, v4 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
+; GFX1032-DPP-NEXT: v_permlanex16_b32 v6, v4, 0, 0
+; GFX1032-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[3:4], v[5:6]
; GFX1032-DPP-NEXT: s_mov_b32 exec_lo, s0
-; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v41, v8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v42, v9
-; GFX1032-DPP-NEXT: s_mov_b32 s54, 0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
+; GFX1032-DPP-NEXT: v_mbcnt_lo_u32_b32 v7, exec_lo, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, v3
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, v4
+; GFX1032-DPP-NEXT: s_mov_b32 s2, 0
+; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v7
; GFX1032-DPP-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX1032-DPP-NEXT: s_cbranch_execz .LBB17_3
; GFX1032-DPP-NEXT: ; %bb.1:
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[52:53], s[34:35], 0x24
+; GFX1032-DPP-NEXT: s_load_dwordx2 s[0:1], s[34:35], 0x24
; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: global_load_dwordx2 v[1:2], v0, s[52:53]
+; GFX1032-DPP-NEXT: global_load_dwordx2 v[9:10], v2, s[0:1]
; GFX1032-DPP-NEXT: .LBB17_2: ; %atomicrmw.start
; GFX1032-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1032-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1032-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1032-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1032-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1032-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1032-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1032-DPP-NEXT: buffer_store_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: buffer_store_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: s_load_dwordx2 s[16:17], s[0:1], 0x0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1032-DPP-NEXT: s_mov_b64 s[0:1], s[64:65]
-; GFX1032-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1032-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1032-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1032-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1032-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1032-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1032-DPP-NEXT: s_mov_b64 s[2:3], s[66:67]
-; GFX1032-DPP-NEXT: buffer_store_dword v4, off, s[64:67], 0 offset:12
-; GFX1032-DPP-NEXT: buffer_store_dword v3, off, s[64:67], 0 offset:8
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1032-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1032-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1032-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
-; GFX1032-DPP-NEXT: s_clause 0x1
-; GFX1032-DPP-NEXT: buffer_load_dword v1, off, s[64:67], 0
-; GFX1032-DPP-NEXT: buffer_load_dword v2, off, s[64:67], 0 offset:4
-; GFX1032-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1032-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1032-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s54
+; GFX1032-DPP-NEXT: v_add_f64 v[7:8], v[9:10], -v[0:1]
+; GFX1032-DPP-NEXT: global_atomic_cmpswap_x2 v[7:8], v2, v[7:10], s[0:1] glc
+; GFX1032-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1032-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[7:8], v[9:10]
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v10, v8
+; GFX1032-DPP-NEXT: v_mov_b32_e32 v9, v7
+; GFX1032-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1032-DPP-NEXT: s_andn2_b32 exec_lo, exec_lo, s2
; GFX1032-DPP-NEXT: s_cbranch_execnz .LBB17_2
; GFX1032-DPP-NEXT: .LBB17_3:
; GFX1032-DPP-NEXT: s_endpgm
@@ -14123,238 +12448,175 @@ define amdgpu_kernel void @global_atomic_fsub_double_uni_address_div_value_defau
; GFX1164-DPP-LABEL: global_atomic_fsub_double_uni_address_div_value_default_scope_strictfp:
; GFX1164-DPP: ; %bb.0:
; GFX1164-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1164-DPP-NEXT: s_mov_b32 s51, s8
+; GFX1164-DPP-NEXT: s_mov_b32 s12, s8
; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-DPP-NEXT: s_mov_b32 s50, s9
+; GFX1164-DPP-NEXT: s_mov_b32 s13, s9
; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
+; GFX1164-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1164-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1164-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v0
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: s_mov_b32 s33, s10
-; GFX1164-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
+; GFX1164-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1164-DPP-NEXT: s_mov_b32 s14, s10
+; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1164-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1164-DPP-NEXT: s_mov_b32 s32, 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
+; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1164-DPP-NEXT: s_or_saveexec_b64 s[0:1], -1
-; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v11, 0x80000000, v1, s[0:1]
-; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s[0:1]
+; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v3, 1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v5, 0x80000000, v1, s[0:1]
+; GFX1164-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s[0:1]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
-; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[4:5], v[2:3]
+; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1164-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
+; GFX1164-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1164-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1164-DPP-NEXT: v_permlane64_b32 v11, v9
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1164-DPP-NEXT: v_permlane64_b32 v5, v3
; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-DPP-NEXT: v_permlane64_b32 v10, v8
-; GFX1164-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1164-DPP-NEXT: v_permlane64_b32 v4, v2
+; GFX1164-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
; GFX1164-DPP-NEXT: s_mov_b64 exec, s[0:1]
-; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX1164-DPP-NEXT: v_mbcnt_lo_u32_b32 v0, exec_lo, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v41, v8
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_2)
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v42, v9
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v10, 0
; GFX1164-DPP-NEXT: s_mov_b64 s[0:1], exec
; GFX1164-DPP-NEXT: s_waitcnt_depctr 0xfffe
-; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v1, exec_hi, v0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 0
-; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
+; GFX1164-DPP-NEXT: v_mbcnt_hi_u32_b32 v6, exec_hi, v0
+; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, v2
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, v3
+; GFX1164-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
; GFX1164-DPP-NEXT: s_cbranch_execz .LBB17_3
; GFX1164-DPP-NEXT: ; %bb.1:
-; GFX1164-DPP-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
-; GFX1164-DPP-NEXT: s_mov_b64 s[54:55], 0
+; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
+; GFX1164-DPP-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1164-DPP-NEXT: .p2align 6
+; GFX1164-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
; GFX1164-DPP-NEXT: .LBB17_2: ; %atomicrmw.start
; GFX1164-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1164-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1164-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1164-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1164-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1164-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1164-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1164-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v31, v40
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v0, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v1, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v2, s52
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v5, 8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v6, 0
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1164-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1164-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1164-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1164-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1164-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1164-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1164-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v3, s53
-; GFX1164-DPP-NEXT: v_mov_b32_e32 v4, 0
-; GFX1164-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1164-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1164-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1164-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1164-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1164-DPP-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX1164-DPP-NEXT: s_or_b64 s[54:55], vcc, s[54:55]
-; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[54:55]
+; GFX1164-DPP-NEXT: v_add_f64 v[6:7], v[8:9], -v[0:1]
+; GFX1164-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
+; GFX1164-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1164-DPP-NEXT: v_cmp_eq_u64_e32 vcc, v[6:7], v[8:9]
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v9, v7
+; GFX1164-DPP-NEXT: v_mov_b32_e32 v8, v6
+; GFX1164-DPP-NEXT: s_or_b64 s[2:3], vcc, s[2:3]
+; GFX1164-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1164-DPP-NEXT: s_and_not1_b64 exec, exec, s[2:3]
; GFX1164-DPP-NEXT: s_cbranch_execnz .LBB17_2
; GFX1164-DPP-NEXT: .LBB17_3:
-; GFX1164-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1164-DPP-NEXT: s_endpgm
;
; GFX1132-DPP-LABEL: global_atomic_fsub_double_uni_address_div_value_default_scope_strictfp:
; GFX1132-DPP: ; %bb.0:
; GFX1132-DPP-NEXT: s_mov_b64 s[34:35], s[4:5]
-; GFX1132-DPP-NEXT: s_mov_b64 s[48:49], s[0:1]
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, div.float.value at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, div.float.value at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v31, v0
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: s_mov_b64 s[36:37], s[6:7]
-; GFX1132-DPP-NEXT: s_mov_b32 s50, s14
-; GFX1132-DPP-NEXT: s_mov_b32 s51, s13
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
+; GFX1132-DPP-NEXT: s_getpc_b64 s[4:5]
+; GFX1132-DPP-NEXT: s_add_u32 s4, s4, div.float.value at gotpcrel32@lo+4
+; GFX1132-DPP-NEXT: s_addc_u32 s5, s5, div.float.value at gotpcrel32@hi+12
; GFX1132-DPP-NEXT: s_mov_b32 s12, s13
+; GFX1132-DPP-NEXT: s_load_b64 s[16:17], s[4:5], 0x0
+; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[6:7]
+; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[0:1]
+; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX1132-DPP-NEXT: s_mov_b32 s13, s14
; GFX1132-DPP-NEXT: s_mov_b32 s14, s15
-; GFX1132-DPP-NEXT: s_mov_b32 s32, 32
-; GFX1132-DPP-NEXT: s_mov_b32 s33, s15
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v40, v0
-; GFX1132-DPP-NEXT: s_mov_b64 s[38:39], s[2:3]
+; GFX1132-DPP-NEXT: s_mov_b32 s32, 0
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
+; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX1132-DPP-NEXT: s_or_saveexec_b32 s0, -1
-; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v9, 1
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v8, 0
-; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v11, 0x80000000, v1, s0
-; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v10, 0, v0, s0
+; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v3, 1
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v2, 0
+; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v5, 0x80000000, v1, s0
+; GFX1132-DPP-NEXT: v_cndmask_b32_e64 v4, 0, v0, s0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v9, v11 row_xmask:1 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v8, v10 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v3, v5 row_xmask:1 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v2, v4 row_xmask:1 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[10:11], v[8:9]
-; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[4:5], v[2:3]
+; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:2 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:2 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v10, 0
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:2 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:4 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:4 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:4 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
-; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v11, 1
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v10, 0
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v11, v9 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
+; GFX1132-DPP-NEXT: v_bfrev_b32_e32 v5, 1
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v4, 0
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v5, v3 row_xmask:8 row_mask:0xf bank_mask:0xf
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_mov_b32_dpp v10, v8 row_xmask:8 row_mask:0xf bank_mask:0xf
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1132-DPP-NEXT: v_mov_b32_dpp v4, v2 row_xmask:8 row_mask:0xf bank_mask:0xf
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_permlanex16_b32 v11, v9, 0, 0
-; GFX1132-DPP-NEXT: v_permlanex16_b32 v10, v8, 0, 0
+; GFX1132-DPP-NEXT: v_permlanex16_b32 v5, v3, 0, 0
+; GFX1132-DPP-NEXT: v_permlanex16_b32 v4, v2, 0, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX1132-DPP-NEXT: v_add_f64 v[8:9], v[8:9], v[10:11]
+; GFX1132-DPP-NEXT: v_add_f64 v[2:3], v[2:3], v[4:5]
; GFX1132-DPP-NEXT: s_mov_b32 exec_lo, s0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v41, v8
-; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v1, exec_lo, 0
+; GFX1132-DPP-NEXT: v_mov_b32_e32 v0, v2
+; GFX1132-DPP-NEXT: v_mbcnt_lo_u32_b32 v6, exec_lo, 0
; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v42, v9
-; GFX1132-DPP-NEXT: s_mov_b32 s54, 0
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v10, 0 :: v_dual_mov_b32 v1, v3
+; GFX1132-DPP-NEXT: s_mov_b32 s2, 0
; GFX1132-DPP-NEXT: s_mov_b32 s0, exec_lo
-; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v1
+; GFX1132-DPP-NEXT: v_cmpx_eq_u32_e32 0, v6
; GFX1132-DPP-NEXT: s_cbranch_execz .LBB17_3
; GFX1132-DPP-NEXT: ; %bb.1:
-; GFX1132-DPP-NEXT: s_load_b64 s[52:53], s[34:35], 0x24
+; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[34:35], 0x24
; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: global_load_b64 v[1:2], v0, s[52:53]
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x1
-; GFX1132-DPP-NEXT: .p2align 6
+; GFX1132-DPP-NEXT: global_load_b64 v[8:9], v10, s[0:1]
; GFX1132-DPP-NEXT: .LBB17_2: ; %atomicrmw.start
; GFX1132-DPP-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
-; GFX1132-DPP-NEXT: v_add_f64 v[3:4], v[1:2], -v[41:42]
-; GFX1132-DPP-NEXT: s_add_u32 s8, s34, 44
-; GFX1132-DPP-NEXT: s_addc_u32 s9, s35, 0
-; GFX1132-DPP-NEXT: s_getpc_b64 s[0:1]
-; GFX1132-DPP-NEXT: s_add_u32 s0, s0, __atomic_compare_exchange at gotpcrel32@lo+4
-; GFX1132-DPP-NEXT: s_addc_u32 s1, s1, __atomic_compare_exchange at gotpcrel32@hi+12
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[1:2], off
-; GFX1132-DPP-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v31, v40 :: v_dual_mov_b32 v0, 8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s52
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v5, 8 :: v_dual_mov_b32 v6, 0
-; GFX1132-DPP-NEXT: v_mov_b32_e32 v7, 0
-; GFX1132-DPP-NEXT: s_mov_b64 s[4:5], s[48:49]
-; GFX1132-DPP-NEXT: s_mov_b64 s[6:7], s[38:39]
-; GFX1132-DPP-NEXT: s_mov_b64 s[10:11], s[36:37]
-; GFX1132-DPP-NEXT: s_mov_b32 s12, s51
-; GFX1132-DPP-NEXT: s_mov_b32 s13, s50
-; GFX1132-DPP-NEXT: s_mov_b32 s14, s33
-; GFX1132-DPP-NEXT: scratch_store_b64 off, v[3:4], off offset:8
-; GFX1132-DPP-NEXT: v_dual_mov_b32 v3, s53 :: v_dual_mov_b32 v4, 0
-; GFX1132-DPP-NEXT: s_waitcnt lgkmcnt(0)
-; GFX1132-DPP-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX1132-DPP-NEXT: scratch_load_b64 v[1:2], off, off
-; GFX1132-DPP-NEXT: v_and_b32_e32 v0, 1, v0
-; GFX1132-DPP-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; GFX1132-DPP-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
-; GFX1132-DPP-NEXT: s_or_b32 s54, vcc_lo, s54
-; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s54
+; GFX1132-DPP-NEXT: v_add_f64 v[6:7], v[8:9], -v[0:1]
+; GFX1132-DPP-NEXT: global_atomic_cmpswap_b64 v[6:7], v10, v[6:9], s[0:1] glc
+; GFX1132-DPP-NEXT: s_waitcnt vmcnt(0)
+; GFX1132-DPP-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[6:7], v[8:9]
+; GFX1132-DPP-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6
+; GFX1132-DPP-NEXT: s_or_b32 s2, vcc_lo, s2
+; GFX1132-DPP-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-DPP-NEXT: s_and_not1_b32 exec_lo, exec_lo, s2
; GFX1132-DPP-NEXT: s_cbranch_execnz .LBB17_2
; GFX1132-DPP-NEXT: .LBB17_3:
-; GFX1132-DPP-NEXT: s_set_inst_prefetch_distance 0x2
; GFX1132-DPP-NEXT: s_endpgm
%divValue = call double @div.float.value() strictfp
- %result = atomicrmw fsub ptr addrspace(1) %ptr, double %divValue monotonic, align 4
+ %result = atomicrmw fsub ptr addrspace(1) %ptr, double %divValue monotonic, align 8
ret void
}
More information about the llvm-commits
mailing list