[llvm] [AMDGPU][GlobalIsel] Add register bank legalization rules for buffer atomic min max (PR #180975)
via llvm-commits
llvm-commits at lists.llvm.org
Wed Feb 11 08:58:14 PST 2026
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
Author: Syadus Sefat (mssefat)
<details>
<summary>Changes</summary>
This patch adds register bank legalization rules for buffer atomic min max operations in the AMDGPU GlobalISel pipeline.
---
Full diff: https://github.com/llvm/llvm-project/pull/180975.diff
2 Files Affected:
- (modified) llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp (+4-1)
- (added) llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.buffer.atomic.integer-minmax.ll (+282)
``````````diff
diff --git a/llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp b/llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp
index 61db779ae0b49..602100f3b9b83 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp
@@ -817,7 +817,10 @@ RegBankLegalizeRules::RegBankLegalizeRules(const GCNSubtarget &_ST,
.Div(S64, {{Vgpr64},
{Vgpr64, Vgpr64, SgprV4S32_WF, Vgpr32, Vgpr32, Sgpr32_WF}});
- addRulesForGOpcs({G_AMDGPU_BUFFER_ATOMIC_SWAP}, Standard)
+ addRulesForGOpcs({G_AMDGPU_BUFFER_ATOMIC_SWAP, G_AMDGPU_BUFFER_ATOMIC_UMAX,
+ G_AMDGPU_BUFFER_ATOMIC_UMIN, G_AMDGPU_BUFFER_ATOMIC_SMAX,
+ G_AMDGPU_BUFFER_ATOMIC_SMIN},
+ Standard)
.Div(S32, {{Vgpr32}, {Vgpr32, SgprV4S32_WF, Vgpr32, Vgpr32, Sgpr32_WF}})
.Div(S64, {{Vgpr64}, {Vgpr64, SgprV4S32_WF, Vgpr32, Vgpr32, Sgpr32_WF}});
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.buffer.atomic.integer-minmax.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.buffer.atomic.integer-minmax.ll
new file mode 100644
index 0000000000000..d3cd8839d127b
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.raw.buffer.atomic.integer-minmax.ll
@@ -0,0 +1,282 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+; RUN: llc -debug -global-isel -mtriple=amdgcn -mcpu=gfx1010 -new-reg-bank-select < %s | FileCheck %s
+
+define amdgpu_ps float @test1(i32 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+; CHECK-LABEL: test1:
+; CHECK: ; %bb.0:
+; CHECK-NEXT: buffer_atomic_smin v0, v1, s[0:3], s4 offen glc
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: ; return to shader part epilog
+ %ret = call i32 @llvm.amdgcn.raw.buffer.atomic.smin.i32(i32 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+ %cast = bitcast i32 %ret to float
+ ret float %cast
+}
+
+define amdgpu_ps float @test2(i32 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+; CHECK-LABEL: test2:
+; CHECK: ; %bb.0:
+; CHECK-NEXT: buffer_atomic_smax v0, v1, s[0:3], s4 offen glc
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: ; return to shader part epilog
+ %ret = call i32 @llvm.amdgcn.raw.buffer.atomic.smax.i32(i32 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+ %cast = bitcast i32 %ret to float
+ ret float %cast
+}
+
+define amdgpu_ps float @test3(i32 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+; CHECK-LABEL: test3:
+; CHECK: ; %bb.0:
+; CHECK-NEXT: buffer_atomic_umin v0, v1, s[0:3], s4 offen glc
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: ; return to shader part epilog
+ %ret = call i32 @llvm.amdgcn.raw.buffer.atomic.umin.i32(i32 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+ %cast = bitcast i32 %ret to float
+ ret float %cast
+}
+
+define amdgpu_ps float @test4(i32 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+; CHECK-LABEL: test4:
+; CHECK: ; %bb.0:
+; CHECK-NEXT: buffer_atomic_umax v0, v1, s[0:3], s4 offen glc
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: ; return to shader part epilog
+ %ret = call i32 @llvm.amdgcn.raw.buffer.atomic.umax.i32(i32 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+ %cast = bitcast i32 %ret to float
+ ret float %cast
+}
+
+
+define amdgpu_ps <2 x float> @test5(i64 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+; CHECK-LABEL: test5:
+; CHECK: ; %bb.0:
+; CHECK-NEXT: buffer_atomic_smin_x2 v[0:1], v2, s[0:3], s4 offen glc
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: ; return to shader part epilog
+ %ret = call i64 @llvm.amdgcn.raw.buffer.atomic.smin.i64(i64 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+ %cast = bitcast i64 %ret to <2 x float>
+ ret <2 x float> %cast
+}
+
+define amdgpu_ps <2 x float> @test6(i64 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+; CHECK-LABEL: test6:
+; CHECK: ; %bb.0:
+; CHECK-NEXT: buffer_atomic_smax_x2 v[0:1], v2, s[0:3], s4 offen glc
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: ; return to shader part epilog
+ %ret = call i64 @llvm.amdgcn.raw.buffer.atomic.smax.i64(i64 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+ %cast = bitcast i64 %ret to <2 x float>
+ ret <2 x float> %cast
+}
+
+define amdgpu_ps <2 x float> @test7(i64 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+; CHECK-LABEL: test7:
+; CHECK: ; %bb.0:
+; CHECK-NEXT: buffer_atomic_umin_x2 v[0:1], v2, s[0:3], s4 offen glc
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: ; return to shader part epilog
+ %ret = call i64 @llvm.amdgcn.raw.buffer.atomic.umin.i64(i64 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+ %cast = bitcast i64 %ret to <2 x float>
+ ret <2 x float> %cast
+}
+
+define amdgpu_ps <2 x float> @test8(i64 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+; CHECK-LABEL: test8:
+; CHECK: ; %bb.0:
+; CHECK-NEXT: buffer_atomic_umax_x2 v[0:1], v2, s[0:3], s4 offen glc
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: ; return to shader part epilog
+ %ret = call i64 @llvm.amdgcn.raw.buffer.atomic.umax.i64(i64 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+ %cast = bitcast i64 %ret to <2 x float>
+ ret <2 x float> %cast
+}
+
+define amdgpu_ps float @wf_test1(i32 %val, <4 x i32> %rsrc, i32 %voffset, i32 inreg %soffset) {
+; CHECK-LABEL: wf_test1:
+; CHECK: ; %bb.0:
+; CHECK-NEXT: s_mov_b32 s2, exec_lo
+; CHECK-NEXT: .LBB8_1: ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT: v_readfirstlane_b32 s4, v1
+; CHECK-NEXT: v_readfirstlane_b32 s5, v2
+; CHECK-NEXT: v_readfirstlane_b32 s6, v3
+; CHECK-NEXT: v_readfirstlane_b32 s7, v4
+; CHECK-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[1:2]
+; CHECK-NEXT: v_cmp_eq_u64_e64 s1, s[6:7], v[3:4]
+; CHECK-NEXT: s_and_b32 s1, vcc_lo, s1
+; CHECK-NEXT: s_and_saveexec_b32 s1, s1
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: buffer_atomic_smin v0, v5, s[4:7], s0 offen glc
+; CHECK-NEXT: ; implicit-def: $vgpr1
+; CHECK-NEXT: ; implicit-def: $vgpr5
+; CHECK-NEXT: ; implicit-def: $vgpr3_vgpr4
+; CHECK-NEXT: s_waitcnt_depctr depctr_vm_vsrc(0)
+; CHECK-NEXT: s_xor_b32 exec_lo, exec_lo, s1
+; CHECK-NEXT: s_cbranch_execnz .LBB8_1
+; CHECK-NEXT: ; %bb.2:
+; CHECK-NEXT: s_mov_b32 exec_lo, s2
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: ; return to shader part epilog
+ %ret = call i32 @llvm.amdgcn.raw.buffer.atomic.smin.i32(i32 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+ %cast = bitcast i32 %ret to float
+ ret float %cast
+}
+
+define amdgpu_ps <2 x float> @wf_test2(i64 %val, <4 x i32> %rsrc, i32 %voffset, i32 inreg %soffset) {
+; CHECK-LABEL: wf_test2:
+; CHECK: ; %bb.0:
+; CHECK-NEXT: s_mov_b32 s2, exec_lo
+; CHECK-NEXT: .LBB9_1: ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT: v_readfirstlane_b32 s4, v2
+; CHECK-NEXT: v_readfirstlane_b32 s5, v3
+; CHECK-NEXT: v_readfirstlane_b32 s6, v4
+; CHECK-NEXT: v_readfirstlane_b32 s7, v5
+; CHECK-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[2:3]
+; CHECK-NEXT: v_cmp_eq_u64_e64 s1, s[6:7], v[4:5]
+; CHECK-NEXT: s_and_b32 s1, vcc_lo, s1
+; CHECK-NEXT: s_and_saveexec_b32 s1, s1
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: buffer_atomic_smax_x2 v[0:1], v6, s[4:7], s0 offen glc
+; CHECK-NEXT: ; implicit-def: $vgpr2
+; CHECK-NEXT: ; implicit-def: $vgpr6
+; CHECK-NEXT: ; implicit-def: $vgpr4_vgpr5
+; CHECK-NEXT: s_waitcnt_depctr depctr_vm_vsrc(0)
+; CHECK-NEXT: s_xor_b32 exec_lo, exec_lo, s1
+; CHECK-NEXT: s_cbranch_execnz .LBB9_1
+; CHECK-NEXT: ; %bb.2:
+; CHECK-NEXT: s_mov_b32 exec_lo, s2
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: ; return to shader part epilog
+ %ret = call i64 @llvm.amdgcn.raw.buffer.atomic.smax.i64(i64 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+ %cast = bitcast i64 %ret to <2 x float>
+ ret <2 x float> %cast
+}
+
+; Waterfall tests - divergent soffset (VGPR soffset)
+define amdgpu_ps float @wf_test3(i32 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 %soffset) {
+; CHECK-LABEL: wf_test3:
+; CHECK: ; %bb.0:
+; CHECK-NEXT: s_mov_b32 s4, exec_lo
+; CHECK-NEXT: .LBB10_1: ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT: v_readfirstlane_b32 s6, v2
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc_lo, s6, v2
+; CHECK-NEXT: s_and_saveexec_b32 s5, vcc_lo
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: buffer_atomic_umin v0, v1, s[0:3], s6 offen glc
+; CHECK-NEXT: ; implicit-def: $vgpr2
+; CHECK-NEXT: ; implicit-def: $vgpr1
+; CHECK-NEXT: s_waitcnt_depctr depctr_vm_vsrc(0)
+; CHECK-NEXT: s_xor_b32 exec_lo, exec_lo, s5
+; CHECK-NEXT: s_cbranch_execnz .LBB10_1
+; CHECK-NEXT: ; %bb.2:
+; CHECK-NEXT: s_mov_b32 exec_lo, s4
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: ; return to shader part epilog
+ %ret = call i32 @llvm.amdgcn.raw.buffer.atomic.umin.i32(i32 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+ %cast = bitcast i32 %ret to float
+ ret float %cast
+}
+
+define amdgpu_ps <2 x float> @wf_test5(i64 %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 %soffset) {
+; CHECK-LABEL: wf_test5:
+; CHECK: ; %bb.0:
+; CHECK-NEXT: s_mov_b32 s4, exec_lo
+; CHECK-NEXT: .LBB11_1: ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT: v_readfirstlane_b32 s6, v3
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc_lo, s6, v3
+; CHECK-NEXT: s_and_saveexec_b32 s5, vcc_lo
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: buffer_atomic_umax_x2 v[0:1], v2, s[0:3], s6 offen glc
+; CHECK-NEXT: ; implicit-def: $vgpr3
+; CHECK-NEXT: ; implicit-def: $vgpr2
+; CHECK-NEXT: s_waitcnt_depctr depctr_vm_vsrc(0)
+; CHECK-NEXT: s_xor_b32 exec_lo, exec_lo, s5
+; CHECK-NEXT: s_cbranch_execnz .LBB11_1
+; CHECK-NEXT: ; %bb.2:
+; CHECK-NEXT: s_mov_b32 exec_lo, s4
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: ; return to shader part epilog
+ %ret = call i64 @llvm.amdgcn.raw.buffer.atomic.umax.i64(i64 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+ %cast = bitcast i64 %ret to <2 x float>
+ ret <2 x float> %cast
+}
+
+; Waterfall tests - both divergent (VGPR rsrc and soffset)
+define amdgpu_ps float @wf_test6(i32 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset) {
+; CHECK-LABEL: wf_test6:
+; CHECK: ; %bb.0:
+; CHECK-NEXT: s_mov_b32 s2, exec_lo
+; CHECK-NEXT: .LBB12_1: ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT: v_readfirstlane_b32 s4, v1
+; CHECK-NEXT: v_readfirstlane_b32 s5, v2
+; CHECK-NEXT: v_readfirstlane_b32 s6, v3
+; CHECK-NEXT: v_readfirstlane_b32 s7, v4
+; CHECK-NEXT: v_readfirstlane_b32 s3, v6
+; CHECK-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[1:2]
+; CHECK-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[3:4]
+; CHECK-NEXT: v_cmp_eq_u32_e64 s1, s3, v6
+; CHECK-NEXT: s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT: s_and_b32 s0, s0, s1
+; CHECK-NEXT: s_and_saveexec_b32 s0, s0
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: buffer_atomic_smin v0, v5, s[4:7], s3 offen glc
+; CHECK-NEXT: ; implicit-def: $vgpr1
+; CHECK-NEXT: ; implicit-def: $vgpr6
+; CHECK-NEXT: ; implicit-def: $vgpr5
+; CHECK-NEXT: ; implicit-def: $vgpr3_vgpr4
+; CHECK-NEXT: s_waitcnt_depctr depctr_vm_vsrc(0)
+; CHECK-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; CHECK-NEXT: s_cbranch_execnz .LBB12_1
+; CHECK-NEXT: ; %bb.2:
+; CHECK-NEXT: s_mov_b32 exec_lo, s2
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: ; return to shader part epilog
+ %ret = call i32 @llvm.amdgcn.raw.buffer.atomic.smin.i32(i32 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+ %cast = bitcast i32 %ret to float
+ ret float %cast
+}
+
+define amdgpu_ps <2 x float> @wf_test7(i64 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset) {
+; CHECK-LABEL: wf_test7:
+; CHECK: ; %bb.0:
+; CHECK-NEXT: s_mov_b32 s2, exec_lo
+; CHECK-NEXT: .LBB13_1: ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT: v_readfirstlane_b32 s4, v2
+; CHECK-NEXT: v_readfirstlane_b32 s5, v3
+; CHECK-NEXT: v_readfirstlane_b32 s6, v4
+; CHECK-NEXT: v_readfirstlane_b32 s7, v5
+; CHECK-NEXT: v_readfirstlane_b32 s3, v7
+; CHECK-NEXT: v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[2:3]
+; CHECK-NEXT: v_cmp_eq_u64_e64 s0, s[6:7], v[4:5]
+; CHECK-NEXT: v_cmp_eq_u32_e64 s1, s3, v7
+; CHECK-NEXT: s_and_b32 s0, vcc_lo, s0
+; CHECK-NEXT: s_and_b32 s0, s0, s1
+; CHECK-NEXT: s_and_saveexec_b32 s0, s0
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: buffer_atomic_umin_x2 v[0:1], v6, s[4:7], s3 offen glc
+; CHECK-NEXT: ; implicit-def: $vgpr2
+; CHECK-NEXT: ; implicit-def: $vgpr7
+; CHECK-NEXT: ; implicit-def: $vgpr6
+; CHECK-NEXT: ; implicit-def: $vgpr4_vgpr5
+; CHECK-NEXT: s_waitcnt_depctr depctr_vm_vsrc(0)
+; CHECK-NEXT: s_xor_b32 exec_lo, exec_lo, s0
+; CHECK-NEXT: s_cbranch_execnz .LBB13_1
+; CHECK-NEXT: ; %bb.2:
+; CHECK-NEXT: s_mov_b32 exec_lo, s2
+; CHECK-NEXT: s_waitcnt vmcnt(0)
+; CHECK-NEXT: ; return to shader part epilog
+ %ret = call i64 @llvm.amdgcn.raw.buffer.atomic.umin.i64(i64 %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+ %cast = bitcast i64 %ret to <2 x float>
+ ret <2 x float> %cast
+}
+
+
+declare i32 @llvm.amdgcn.raw.buffer.atomic.smin.i32(i32, <4 x i32>, i32, i32, i32 immarg) #0
+declare i32 @llvm.amdgcn.raw.buffer.atomic.smax.i32(i32, <4 x i32>, i32, i32, i32 immarg) #0
+declare i32 @llvm.amdgcn.raw.buffer.atomic.umin.i32(i32, <4 x i32>, i32, i32, i32 immarg) #0
+declare i32 @llvm.amdgcn.raw.buffer.atomic.umax.i32(i32, <4 x i32>, i32, i32, i32 immarg) #0
+
+declare i64 @llvm.amdgcn.raw.buffer.atomic.smin.i64(i64, <4 x i32>, i32, i32, i32 immarg) #0
+declare i64 @llvm.amdgcn.raw.buffer.atomic.smax.i64(i64, <4 x i32>, i32, i32, i32 immarg) #0
+declare i64 @llvm.amdgcn.raw.buffer.atomic.umin.i64(i64, <4 x i32>, i32, i32, i32 immarg) #0
+declare i64 @llvm.amdgcn.raw.buffer.atomic.umax.i64(i64, <4 x i32>, i32, i32, i32 immarg) #0
+
+attributes #0 = { nounwind }
``````````
</details>
https://github.com/llvm/llvm-project/pull/180975
More information about the llvm-commits
mailing list