[llvm] [AMDGPU] Avoid setting GLC for image atomics when result is unused (PR #150742)
Harrison Hao via llvm-commits
llvm-commits at lists.llvm.org
Sat Jul 26 01:39:07 PDT 2025
https://github.com/harrisonGPU created https://github.com/llvm/llvm-project/pull/150742
Select glc=0 (no-return) for image atomics when result is unused
When the result of an image atomic instruction isn't used, we can safely set glc=0 to avoid the return value.
This avoids unnecessary waits on vmcnt and can improve performance.
>From 83fb88cf94d7d0d5dc18a39dfe8033c3afed5699 Mon Sep 17 00:00:00 2001
From: Harrison Hao <tsworld1314 at gmail.com>
Date: Sat, 26 Jul 2025 05:33:34 +0000
Subject: [PATCH] [AMDGPU] Avoid setting GLC for image atomics when result is
unused
---
.../AMDGPU/AMDGPUInstructionSelector.cpp | 12 +++++--
llvm/lib/Target/AMDGPU/SIISelLowering.cpp | 6 ++--
.../llvm.amdgcn.image.atomic.dim.ll | 24 ++++++-------
.../llvm.amdgcn.image.atomic.dim.mir | 36 +++++++++++++------
.../AMDGPU/llvm.amdgcn.image.atomic.pk.add.ll | 20 +++++------
5 files changed, 61 insertions(+), 37 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
index 8975486caa770..eb41f9806d860 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
@@ -2123,8 +2123,10 @@ bool AMDGPUInstructionSelector::selectImageIntrinsic(
assert((!IsTexFail || DMaskLanes >= 1) && "should have legalized this");
unsigned CPol = MI.getOperand(ArgOffset + Intr->CachePolicyIndex).getImm();
- if (BaseOpcode->Atomic)
- CPol |= AMDGPU::CPol::GLC; // TODO no-return optimization
+ // Keep GLC only when the atomic's result is actually used.
+ if (BaseOpcode->Atomic && VDataOut && !MRI->use_empty(VDataOut))
+ CPol |= AMDGPU::CPol::GLC;
+
if (CPol & ~((IsGFX12Plus ? AMDGPU::CPol::ALL : AMDGPU::CPol::ALL_pregfx12) |
AMDGPU::CPol::VOLATILE))
return false;
@@ -2214,7 +2216,11 @@ bool AMDGPUInstructionSelector::selectImageIntrinsic(
}
} else {
- MIB.addDef(VDataOut); // vdata output
+ // If VDataOut is unused, mark it as dead to avoid unnecessary VGPR usage.
+ if (BaseOpcode->Atomic && MRI->use_empty(VDataOut))
+ MIB.addDef(VDataOut, RegState::Dead);
+ else
+ MIB.addDef(VDataOut);
}
}
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index bc0fd8d4e814b..ab9f941378ded 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -8780,8 +8780,10 @@ SDValue SITargetLowering::lowerImage(SDValue Op,
}
unsigned CPol = Op.getConstantOperandVal(ArgOffset + Intr->CachePolicyIndex);
- if (BaseOpcode->Atomic)
- CPol |= AMDGPU::CPol::GLC; // TODO no-return optimization
+ // Keep GLC only when the atomic's result is actually used.
+ if (BaseOpcode->Atomic && !Op.getValue(0).use_empty())
+ CPol |= AMDGPU::CPol::GLC;
+
if (CPol & ~((IsGFX12Plus ? AMDGPU::CPol::ALL : AMDGPU::CPol::ALL_pregfx12) |
AMDGPU::CPol::VOLATILE))
return Op;
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.atomic.dim.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.atomic.dim.ll
index 221e2fd4f00f7..09e1fca3f2677 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.atomic.dim.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.atomic.dim.ll
@@ -1200,7 +1200,7 @@ define amdgpu_ps void @atomic_cmpswap_i32_1d_no_return(<8 x i32> inreg %rsrc, i3
; GFX6-NEXT: s_mov_b32 s5, s7
; GFX6-NEXT: s_mov_b32 s6, s8
; GFX6-NEXT: s_mov_b32 s7, s9
-; GFX6-NEXT: image_atomic_cmpswap v[0:1], v2, s[0:7] dmask:0x3 unorm glc
+; GFX6-NEXT: image_atomic_cmpswap v[0:1], v2, s[0:7] dmask:0x3 unorm
; GFX6-NEXT: s_endpgm
;
; GFX8-LABEL: atomic_cmpswap_i32_1d_no_return:
@@ -1213,7 +1213,7 @@ define amdgpu_ps void @atomic_cmpswap_i32_1d_no_return(<8 x i32> inreg %rsrc, i3
; GFX8-NEXT: s_mov_b32 s5, s7
; GFX8-NEXT: s_mov_b32 s6, s8
; GFX8-NEXT: s_mov_b32 s7, s9
-; GFX8-NEXT: image_atomic_cmpswap v[0:1], v2, s[0:7] dmask:0x3 unorm glc
+; GFX8-NEXT: image_atomic_cmpswap v[0:1], v2, s[0:7] dmask:0x3 unorm
; GFX8-NEXT: s_endpgm
;
; GFX900-LABEL: atomic_cmpswap_i32_1d_no_return:
@@ -1226,7 +1226,7 @@ define amdgpu_ps void @atomic_cmpswap_i32_1d_no_return(<8 x i32> inreg %rsrc, i3
; GFX900-NEXT: s_mov_b32 s5, s7
; GFX900-NEXT: s_mov_b32 s6, s8
; GFX900-NEXT: s_mov_b32 s7, s9
-; GFX900-NEXT: image_atomic_cmpswap v[0:1], v2, s[0:7] dmask:0x3 unorm glc
+; GFX900-NEXT: image_atomic_cmpswap v[0:1], v2, s[0:7] dmask:0x3 unorm
; GFX900-NEXT: s_endpgm
;
; GFX90A-LABEL: atomic_cmpswap_i32_1d_no_return:
@@ -1239,7 +1239,7 @@ define amdgpu_ps void @atomic_cmpswap_i32_1d_no_return(<8 x i32> inreg %rsrc, i3
; GFX90A-NEXT: s_mov_b32 s5, s7
; GFX90A-NEXT: s_mov_b32 s6, s8
; GFX90A-NEXT: s_mov_b32 s7, s9
-; GFX90A-NEXT: image_atomic_cmpswap v[0:1], v2, s[0:7] dmask:0x3 unorm glc
+; GFX90A-NEXT: image_atomic_cmpswap v[0:1], v2, s[0:7] dmask:0x3 unorm
; GFX90A-NEXT: s_endpgm
;
; GFX10PLUS-LABEL: atomic_cmpswap_i32_1d_no_return:
@@ -1252,7 +1252,7 @@ define amdgpu_ps void @atomic_cmpswap_i32_1d_no_return(<8 x i32> inreg %rsrc, i3
; GFX10PLUS-NEXT: s_mov_b32 s5, s7
; GFX10PLUS-NEXT: s_mov_b32 s6, s8
; GFX10PLUS-NEXT: s_mov_b32 s7, s9
-; GFX10PLUS-NEXT: image_atomic_cmpswap v[0:1], v2, s[0:7] dmask:0x3 dim:SQ_RSRC_IMG_1D unorm glc
+; GFX10PLUS-NEXT: image_atomic_cmpswap v[0:1], v2, s[0:7] dmask:0x3 dim:SQ_RSRC_IMG_1D unorm
; GFX10PLUS-NEXT: s_endpgm
;
; GFX12-LABEL: atomic_cmpswap_i32_1d_no_return:
@@ -1265,7 +1265,7 @@ define amdgpu_ps void @atomic_cmpswap_i32_1d_no_return(<8 x i32> inreg %rsrc, i3
; GFX12-NEXT: s_mov_b32 s5, s7
; GFX12-NEXT: s_mov_b32 s6, s8
; GFX12-NEXT: s_mov_b32 s7, s9
-; GFX12-NEXT: image_atomic_cmpswap v[0:1], v2, s[0:7] dmask:0x3 dim:SQ_RSRC_IMG_1D th:TH_ATOMIC_RETURN
+; GFX12-NEXT: image_atomic_cmpswap v[0:1], v2, s[0:7] dmask:0x3 dim:SQ_RSRC_IMG_1D
; GFX12-NEXT: s_endpgm
main_body:
%v = call i32 @llvm.amdgcn.image.atomic.cmpswap.1d.i32.i32(i32 %cmp, i32 %swap, i32 %s, <8 x i32> %rsrc, i32 0, i32 0)
@@ -3194,7 +3194,7 @@ define amdgpu_ps void @atomic_cmpswap_i64_1d_no_return(<8 x i32> inreg %rsrc, i6
; GFX6-NEXT: s_mov_b32 s5, s7
; GFX6-NEXT: s_mov_b32 s6, s8
; GFX6-NEXT: s_mov_b32 s7, s9
-; GFX6-NEXT: image_atomic_cmpswap v[0:3], v4, s[0:7] dmask:0xf unorm glc
+; GFX6-NEXT: image_atomic_cmpswap v[0:3], v4, s[0:7] dmask:0xf unorm
; GFX6-NEXT: s_endpgm
;
; GFX8-LABEL: atomic_cmpswap_i64_1d_no_return:
@@ -3207,7 +3207,7 @@ define amdgpu_ps void @atomic_cmpswap_i64_1d_no_return(<8 x i32> inreg %rsrc, i6
; GFX8-NEXT: s_mov_b32 s5, s7
; GFX8-NEXT: s_mov_b32 s6, s8
; GFX8-NEXT: s_mov_b32 s7, s9
-; GFX8-NEXT: image_atomic_cmpswap v[0:3], v4, s[0:7] dmask:0xf unorm glc
+; GFX8-NEXT: image_atomic_cmpswap v[0:3], v4, s[0:7] dmask:0xf unorm
; GFX8-NEXT: s_endpgm
;
; GFX900-LABEL: atomic_cmpswap_i64_1d_no_return:
@@ -3220,7 +3220,7 @@ define amdgpu_ps void @atomic_cmpswap_i64_1d_no_return(<8 x i32> inreg %rsrc, i6
; GFX900-NEXT: s_mov_b32 s5, s7
; GFX900-NEXT: s_mov_b32 s6, s8
; GFX900-NEXT: s_mov_b32 s7, s9
-; GFX900-NEXT: image_atomic_cmpswap v[0:3], v4, s[0:7] dmask:0xf unorm glc
+; GFX900-NEXT: image_atomic_cmpswap v[0:3], v4, s[0:7] dmask:0xf unorm
; GFX900-NEXT: s_endpgm
;
; GFX90A-LABEL: atomic_cmpswap_i64_1d_no_return:
@@ -3233,7 +3233,7 @@ define amdgpu_ps void @atomic_cmpswap_i64_1d_no_return(<8 x i32> inreg %rsrc, i6
; GFX90A-NEXT: s_mov_b32 s5, s7
; GFX90A-NEXT: s_mov_b32 s6, s8
; GFX90A-NEXT: s_mov_b32 s7, s9
-; GFX90A-NEXT: image_atomic_cmpswap v[0:3], v4, s[0:7] dmask:0xf unorm glc
+; GFX90A-NEXT: image_atomic_cmpswap v[0:3], v4, s[0:7] dmask:0xf unorm
; GFX90A-NEXT: s_endpgm
;
; GFX10PLUS-LABEL: atomic_cmpswap_i64_1d_no_return:
@@ -3246,7 +3246,7 @@ define amdgpu_ps void @atomic_cmpswap_i64_1d_no_return(<8 x i32> inreg %rsrc, i6
; GFX10PLUS-NEXT: s_mov_b32 s5, s7
; GFX10PLUS-NEXT: s_mov_b32 s6, s8
; GFX10PLUS-NEXT: s_mov_b32 s7, s9
-; GFX10PLUS-NEXT: image_atomic_cmpswap v[0:3], v4, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm glc
+; GFX10PLUS-NEXT: image_atomic_cmpswap v[0:3], v4, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm
; GFX10PLUS-NEXT: s_endpgm
;
; GFX12-LABEL: atomic_cmpswap_i64_1d_no_return:
@@ -3259,7 +3259,7 @@ define amdgpu_ps void @atomic_cmpswap_i64_1d_no_return(<8 x i32> inreg %rsrc, i6
; GFX12-NEXT: s_mov_b32 s5, s7
; GFX12-NEXT: s_mov_b32 s6, s8
; GFX12-NEXT: s_mov_b32 s7, s9
-; GFX12-NEXT: image_atomic_cmpswap v[0:3], v4, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D th:TH_ATOMIC_RETURN
+; GFX12-NEXT: image_atomic_cmpswap v[0:3], v4, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D
; GFX12-NEXT: s_endpgm
main_body:
%v = call i64 @llvm.amdgcn.image.atomic.cmpswap.1d.i64.i32(i64 %cmp, i64 %swap, i32 %s, <8 x i32> %rsrc, i32 0, i32 0)
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.atomic.dim.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.atomic.dim.mir
index 292fa4be1ca1d..a95b0ee434956 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.atomic.dim.mir
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.atomic.dim.mir
@@ -25,6 +25,7 @@ body: |
; GFX6-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY killed [[IMAGE_ATOMIC_CMPSWAP_V2_V1_si]].sub0
; GFX6-NEXT: $vgpr0 = COPY [[COPY3]]
; GFX6-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0
+ ;
; GFX8-LABEL: name: atomic_cmpswap_i32_1d
; GFX8: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $vgpr0_vgpr1, $vgpr2
; GFX8-NEXT: {{ $}}
@@ -35,6 +36,7 @@ body: |
; GFX8-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY killed [[IMAGE_ATOMIC_CMPSWAP_V2_V1_vi]].sub0
; GFX8-NEXT: $vgpr0 = COPY [[COPY3]]
; GFX8-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0
+ ;
; GFX10-LABEL: name: atomic_cmpswap_i32_1d
; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $vgpr0_vgpr1, $vgpr2
; GFX10-NEXT: {{ $}}
@@ -45,6 +47,7 @@ body: |
; GFX10-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY killed [[IMAGE_ATOMIC_CMPSWAP_V2_V1_gfx10_]].sub0
; GFX10-NEXT: $vgpr0 = COPY [[COPY3]]
; GFX10-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0
+ ;
; GFX11-LABEL: name: atomic_cmpswap_i32_1d
; GFX11: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $vgpr0_vgpr1, $vgpr2
; GFX11-NEXT: {{ $}}
@@ -55,6 +58,7 @@ body: |
; GFX11-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY killed [[IMAGE_ATOMIC_CMPSWAP_V2_V1_gfx11_]].sub0
; GFX11-NEXT: $vgpr0 = COPY [[COPY3]]
; GFX11-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0
+ ;
; GFX12-LABEL: name: atomic_cmpswap_i32_1d
; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $vgpr0_vgpr1, $vgpr2
; GFX12-NEXT: {{ $}}
@@ -89,39 +93,43 @@ body: |
; GFX6-NEXT: [[COPY:%[0-9]+]]:sgpr_256 = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7
; GFX6-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
; GFX6-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
- ; GFX6-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V2_V1_si:%[0-9]+]]:vreg_64 = IMAGE_ATOMIC_CMPSWAP_V2_V1_si [[COPY1]], [[COPY2]], [[COPY]], 3, 1, 1, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), addrspace 8)
+ ; GFX6-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V2_V1_si:%[0-9]+]]:vreg_64 = IMAGE_ATOMIC_CMPSWAP_V2_V1_si [[COPY1]], [[COPY2]], [[COPY]], 3, 1, 0, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), addrspace 8)
; GFX6-NEXT: S_ENDPGM 0
+ ;
; GFX8-LABEL: name: atomic_cmpswap_i32_1d_no_return
; GFX8: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $vgpr0_vgpr1, $vgpr2
; GFX8-NEXT: {{ $}}
; GFX8-NEXT: [[COPY:%[0-9]+]]:sgpr_256 = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7
; GFX8-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
; GFX8-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
- ; GFX8-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V1_V1_vi:%[0-9]+]]:vreg_64 = IMAGE_ATOMIC_CMPSWAP_V2_V1_vi [[COPY1]], [[COPY2]], [[COPY]], 3, 1, 1, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), addrspace 8)
+ ; GFX8-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V2_V1_vi:%[0-9]+]]:vreg_64 = IMAGE_ATOMIC_CMPSWAP_V2_V1_vi [[COPY1]], [[COPY2]], [[COPY]], 3, 1, 0, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), addrspace 8)
; GFX8-NEXT: S_ENDPGM 0
+ ;
; GFX10-LABEL: name: atomic_cmpswap_i32_1d_no_return
; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $vgpr0_vgpr1, $vgpr2
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_256 = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7
; GFX10-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
; GFX10-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
- ; GFX10-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V2_V1_gfx10_:%[0-9]+]]:vreg_64 = IMAGE_ATOMIC_CMPSWAP_V2_V1_gfx10 [[COPY1]], [[COPY2]], [[COPY]], 3, 0, 1, 1, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), addrspace 8)
+ ; GFX10-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V2_V1_gfx10_:%[0-9]+]]:vreg_64 = IMAGE_ATOMIC_CMPSWAP_V2_V1_gfx10 [[COPY1]], [[COPY2]], [[COPY]], 3, 0, 1, 0, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), addrspace 8)
; GFX10-NEXT: S_ENDPGM 0
+ ;
; GFX11-LABEL: name: atomic_cmpswap_i32_1d_no_return
; GFX11: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $vgpr0_vgpr1, $vgpr2
; GFX11-NEXT: {{ $}}
; GFX11-NEXT: [[COPY:%[0-9]+]]:sgpr_256 = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7
; GFX11-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
; GFX11-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
- ; GFX11-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V2_V1_gfx11_:%[0-9]+]]:vreg_64 = IMAGE_ATOMIC_CMPSWAP_V2_V1_gfx11 [[COPY1]], [[COPY2]], [[COPY]], 3, 0, 1, 1, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), addrspace 8)
+ ; GFX11-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V2_V1_gfx11_:%[0-9]+]]:vreg_64 = IMAGE_ATOMIC_CMPSWAP_V2_V1_gfx11 [[COPY1]], [[COPY2]], [[COPY]], 3, 0, 1, 0, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), addrspace 8)
; GFX11-NEXT: S_ENDPGM 0
+ ;
; GFX12-LABEL: name: atomic_cmpswap_i32_1d_no_return
; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $vgpr0_vgpr1, $vgpr2
; GFX12-NEXT: {{ $}}
; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_256 = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7
; GFX12-NEXT: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
; GFX12-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
- ; GFX12-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V2_V1_gfx12_:%[0-9]+]]:vreg_64 = IMAGE_ATOMIC_CMPSWAP_V2_V1_gfx12 [[COPY1]], [[COPY2]], [[COPY]], 3, 0, 1, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), addrspace 8)
+ ; GFX12-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V2_V1_gfx12_:%[0-9]+]]:vreg_64 = IMAGE_ATOMIC_CMPSWAP_V2_V1_gfx12 [[COPY1]], [[COPY2]], [[COPY]], 3, 0, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), addrspace 8)
; GFX12-NEXT: S_ENDPGM 0
%0:sgpr(<8 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7
%1:vgpr(<2 x s32>) = COPY $vgpr0_vgpr1
@@ -150,6 +158,7 @@ body: |
; GFX6-NEXT: [[COPY3:%[0-9]+]]:vreg_64 = COPY killed [[IMAGE_ATOMIC_CMPSWAP_V4_V1_si]].sub0_sub1
; GFX6-NEXT: $vgpr0_vgpr1 = COPY [[COPY3]]
; GFX6-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0_vgpr1
+ ;
; GFX8-LABEL: name: atomic_cmpswap_i64_1d
; GFX8: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4
; GFX8-NEXT: {{ $}}
@@ -160,6 +169,7 @@ body: |
; GFX8-NEXT: [[COPY3:%[0-9]+]]:vreg_64 = COPY killed [[IMAGE_ATOMIC_CMPSWAP_V4_V1_vi]].sub0_sub1
; GFX8-NEXT: $vgpr0_vgpr1 = COPY [[COPY3]]
; GFX8-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0_vgpr1
+ ;
; GFX10-LABEL: name: atomic_cmpswap_i64_1d
; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4
; GFX10-NEXT: {{ $}}
@@ -170,6 +180,7 @@ body: |
; GFX10-NEXT: [[COPY3:%[0-9]+]]:vreg_64 = COPY killed [[IMAGE_ATOMIC_CMPSWAP_V4_V1_gfx10_]].sub0_sub1
; GFX10-NEXT: $vgpr0_vgpr1 = COPY [[COPY3]]
; GFX10-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0_vgpr1
+ ;
; GFX11-LABEL: name: atomic_cmpswap_i64_1d
; GFX11: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4
; GFX11-NEXT: {{ $}}
@@ -180,6 +191,7 @@ body: |
; GFX11-NEXT: [[COPY3:%[0-9]+]]:vreg_64 = COPY killed [[IMAGE_ATOMIC_CMPSWAP_V4_V1_gfx11_]].sub0_sub1
; GFX11-NEXT: $vgpr0_vgpr1 = COPY [[COPY3]]
; GFX11-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0_vgpr1
+ ;
; GFX12-LABEL: name: atomic_cmpswap_i64_1d
; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4
; GFX12-NEXT: {{ $}}
@@ -214,39 +226,43 @@ body: |
; GFX6-NEXT: [[COPY:%[0-9]+]]:sgpr_256 = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7
; GFX6-NEXT: [[COPY1:%[0-9]+]]:vreg_128 = COPY $vgpr0_vgpr1_vgpr2_vgpr3
; GFX6-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr4
- ; GFX6-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V4_V1_si:%[0-9]+]]:vreg_128 = IMAGE_ATOMIC_CMPSWAP_V4_V1_si [[COPY1]], [[COPY2]], [[COPY]], 15, 1, 1, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), addrspace 8)
+ ; GFX6-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V4_V1_si:%[0-9]+]]:vreg_128 = IMAGE_ATOMIC_CMPSWAP_V4_V1_si [[COPY1]], [[COPY2]], [[COPY]], 15, 1, 0, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), addrspace 8)
; GFX6-NEXT: S_ENDPGM 0
+ ;
; GFX8-LABEL: name: atomic_cmpswap_i64_1d_no_return
; GFX8: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4
; GFX8-NEXT: {{ $}}
; GFX8-NEXT: [[COPY:%[0-9]+]]:sgpr_256 = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7
; GFX8-NEXT: [[COPY1:%[0-9]+]]:vreg_128 = COPY $vgpr0_vgpr1_vgpr2_vgpr3
; GFX8-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr4
- ; GFX8-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V4_V1_vi:%[0-9]+]]:vreg_128 = IMAGE_ATOMIC_CMPSWAP_V4_V1_vi [[COPY1]], [[COPY2]], [[COPY]], 15, 1, 1, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), addrspace 8)
+ ; GFX8-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V4_V1_vi:%[0-9]+]]:vreg_128 = IMAGE_ATOMIC_CMPSWAP_V4_V1_vi [[COPY1]], [[COPY2]], [[COPY]], 15, 1, 0, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), addrspace 8)
; GFX8-NEXT: S_ENDPGM 0
+ ;
; GFX10-LABEL: name: atomic_cmpswap_i64_1d_no_return
; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4
; GFX10-NEXT: {{ $}}
; GFX10-NEXT: [[COPY:%[0-9]+]]:sgpr_256 = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7
; GFX10-NEXT: [[COPY1:%[0-9]+]]:vreg_128 = COPY $vgpr0_vgpr1_vgpr2_vgpr3
; GFX10-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr4
- ; GFX10-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V4_V1_gfx10_:%[0-9]+]]:vreg_128 = IMAGE_ATOMIC_CMPSWAP_V4_V1_gfx10 [[COPY1]], [[COPY2]], [[COPY]], 15, 0, 1, 1, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), addrspace 8)
+ ; GFX10-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V4_V1_gfx10_:%[0-9]+]]:vreg_128 = IMAGE_ATOMIC_CMPSWAP_V4_V1_gfx10 [[COPY1]], [[COPY2]], [[COPY]], 15, 0, 1, 0, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), addrspace 8)
; GFX10-NEXT: S_ENDPGM 0
+ ;
; GFX11-LABEL: name: atomic_cmpswap_i64_1d_no_return
; GFX11: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4
; GFX11-NEXT: {{ $}}
; GFX11-NEXT: [[COPY:%[0-9]+]]:sgpr_256 = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7
; GFX11-NEXT: [[COPY1:%[0-9]+]]:vreg_128 = COPY $vgpr0_vgpr1_vgpr2_vgpr3
; GFX11-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr4
- ; GFX11-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V4_V1_gfx11_:%[0-9]+]]:vreg_128 = IMAGE_ATOMIC_CMPSWAP_V4_V1_gfx11 [[COPY1]], [[COPY2]], [[COPY]], 15, 0, 1, 1, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), addrspace 8)
+ ; GFX11-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V4_V1_gfx11_:%[0-9]+]]:vreg_128 = IMAGE_ATOMIC_CMPSWAP_V4_V1_gfx11 [[COPY1]], [[COPY2]], [[COPY]], 15, 0, 1, 0, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), addrspace 8)
; GFX11-NEXT: S_ENDPGM 0
+ ;
; GFX12-LABEL: name: atomic_cmpswap_i64_1d_no_return
; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4
; GFX12-NEXT: {{ $}}
; GFX12-NEXT: [[COPY:%[0-9]+]]:sgpr_256 = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7
; GFX12-NEXT: [[COPY1:%[0-9]+]]:vreg_128 = COPY $vgpr0_vgpr1_vgpr2_vgpr3
; GFX12-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr4
- ; GFX12-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V4_V1_gfx12_:%[0-9]+]]:vreg_128 = IMAGE_ATOMIC_CMPSWAP_V4_V1_gfx12 [[COPY1]], [[COPY2]], [[COPY]], 15, 0, 1, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), addrspace 8)
+ ; GFX12-NEXT: [[IMAGE_ATOMIC_CMPSWAP_V4_V1_gfx12_:%[0-9]+]]:vreg_128 = IMAGE_ATOMIC_CMPSWAP_V4_V1_gfx12 [[COPY1]], [[COPY2]], [[COPY]], 15, 0, 0, 0, 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), addrspace 8)
; GFX12-NEXT: S_ENDPGM 0
%0:sgpr(<8 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7
%1:vgpr(<2 x s64>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.atomic.pk.add.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.atomic.pk.add.ll
index 7be0d9ca329aa..382518377ef25 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.atomic.pk.add.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.atomic.pk.add.ll
@@ -41,14 +41,14 @@ main_body:
define amdgpu_ps float @atomic_pk_add_f16_1d_v2_noret(<8 x i32> inreg %rsrc, <2 x half> %data, i32 %s) {
; GFX12-SDAG-LABEL: atomic_pk_add_f16_1d_v2_noret:
; GFX12-SDAG: ; %bb.0: ; %main_body
-; GFX12-SDAG-NEXT: image_atomic_pk_add_f16 v0, v1, s[0:7] dmask:0x1 dim:SQ_RSRC_IMG_1D th:TH_ATOMIC_RETURN
+; GFX12-SDAG-NEXT: image_atomic_pk_add_f16 v0, v1, s[0:7] dmask:0x1 dim:SQ_RSRC_IMG_1D
; GFX12-SDAG-NEXT: s_wait_loadcnt 0x0
; GFX12-SDAG-NEXT: v_mov_b32_e32 v0, 1.0
; GFX12-SDAG-NEXT: ; return to shader part epilog
;
; GFX12-GISEL-LABEL: atomic_pk_add_f16_1d_v2_noret:
; GFX12-GISEL: ; %bb.0: ; %main_body
-; GFX12-GISEL-NEXT: image_atomic_pk_add_f16 v0, v1, s[0:7] dmask:0x1 dim:SQ_RSRC_IMG_1D th:TH_ATOMIC_RETURN
+; GFX12-GISEL-NEXT: image_atomic_pk_add_f16 v0, v1, s[0:7] dmask:0x1 dim:SQ_RSRC_IMG_1D
; GFX12-GISEL-NEXT: s_wait_loadcnt 0x0
; GFX12-GISEL-NEXT: v_mov_b32_e32 v0, 1.0
; GFX12-GISEL-NEXT: ; return to shader part epilog
@@ -79,14 +79,14 @@ main_body:
define amdgpu_ps float @atomic_pk_add_f16_1d_v4_noret(<8 x i32> inreg %rsrc, <4 x half> %data, i32 %s) {
; GFX12-SDAG-LABEL: atomic_pk_add_f16_1d_v4_noret:
; GFX12-SDAG: ; %bb.0: ; %main_body
-; GFX12-SDAG-NEXT: image_atomic_pk_add_f16 v[0:1], v2, s[0:7] dmask:0x3 dim:SQ_RSRC_IMG_1D th:TH_ATOMIC_RETURN
+; GFX12-SDAG-NEXT: image_atomic_pk_add_f16 v[0:1], v2, s[0:7] dmask:0x3 dim:SQ_RSRC_IMG_1D
; GFX12-SDAG-NEXT: s_wait_loadcnt 0x0
; GFX12-SDAG-NEXT: v_mov_b32_e32 v0, 1.0
; GFX12-SDAG-NEXT: ; return to shader part epilog
;
; GFX12-GISEL-LABEL: atomic_pk_add_f16_1d_v4_noret:
; GFX12-GISEL: ; %bb.0: ; %main_body
-; GFX12-GISEL-NEXT: image_atomic_pk_add_f16 v[0:1], v2, s[0:7] dmask:0x3 dim:SQ_RSRC_IMG_1D th:TH_ATOMIC_RETURN
+; GFX12-GISEL-NEXT: image_atomic_pk_add_f16 v[0:1], v2, s[0:7] dmask:0x3 dim:SQ_RSRC_IMG_1D
; GFX12-GISEL-NEXT: s_wait_loadcnt 0x0
; GFX12-GISEL-NEXT: v_mov_b32_e32 v0, 1.0
; GFX12-GISEL-NEXT: ; return to shader part epilog
@@ -126,14 +126,14 @@ main_body:
define amdgpu_ps float @atomic_pk_add_bf16_1d_v2_noret(<8 x i32> inreg %rsrc, <2 x bfloat> %data, i32 %s) {
; GFX12-SDAG-LABEL: atomic_pk_add_bf16_1d_v2_noret:
; GFX12-SDAG: ; %bb.0: ; %main_body
-; GFX12-SDAG-NEXT: image_atomic_pk_add_bf16 v0, v1, s[0:7] dmask:0x1 dim:SQ_RSRC_IMG_1D th:TH_ATOMIC_RETURN
+; GFX12-SDAG-NEXT: image_atomic_pk_add_bf16 v0, v1, s[0:7] dmask:0x1 dim:SQ_RSRC_IMG_1D
; GFX12-SDAG-NEXT: s_wait_loadcnt 0x0
; GFX12-SDAG-NEXT: v_mov_b32_e32 v0, 1.0
; GFX12-SDAG-NEXT: ; return to shader part epilog
;
; GFX12-GISEL-LABEL: atomic_pk_add_bf16_1d_v2_noret:
; GFX12-GISEL: ; %bb.0: ; %main_body
-; GFX12-GISEL-NEXT: image_atomic_pk_add_bf16 v0, v1, s[0:7] dmask:0x1 dim:SQ_RSRC_IMG_1D th:TH_ATOMIC_RETURN
+; GFX12-GISEL-NEXT: image_atomic_pk_add_bf16 v0, v1, s[0:7] dmask:0x1 dim:SQ_RSRC_IMG_1D
; GFX12-GISEL-NEXT: s_wait_loadcnt 0x0
; GFX12-GISEL-NEXT: v_mov_b32_e32 v0, 1.0
; GFX12-GISEL-NEXT: ; return to shader part epilog
@@ -173,14 +173,14 @@ main_body:
define amdgpu_ps float @atomic_pk_add_bf16_1d_v4_noret(<8 x i32> inreg %rsrc, <4 x bfloat> %data, i32 %s) {
; GFX12-SDAG-LABEL: atomic_pk_add_bf16_1d_v4_noret:
; GFX12-SDAG: ; %bb.0: ; %main_body
-; GFX12-SDAG-NEXT: image_atomic_pk_add_bf16 v[0:1], v2, s[0:7] dmask:0x3 dim:SQ_RSRC_IMG_1D th:TH_ATOMIC_RETURN
+; GFX12-SDAG-NEXT: image_atomic_pk_add_bf16 v[0:1], v2, s[0:7] dmask:0x3 dim:SQ_RSRC_IMG_1D
; GFX12-SDAG-NEXT: s_wait_loadcnt 0x0
; GFX12-SDAG-NEXT: v_mov_b32_e32 v0, 1.0
; GFX12-SDAG-NEXT: ; return to shader part epilog
;
; GFX12-GISEL-LABEL: atomic_pk_add_bf16_1d_v4_noret:
; GFX12-GISEL: ; %bb.0: ; %main_body
-; GFX12-GISEL-NEXT: image_atomic_pk_add_bf16 v[0:1], v2, s[0:7] dmask:0x3 dim:SQ_RSRC_IMG_1D th:TH_ATOMIC_RETURN
+; GFX12-GISEL-NEXT: image_atomic_pk_add_bf16 v[0:1], v2, s[0:7] dmask:0x3 dim:SQ_RSRC_IMG_1D
; GFX12-GISEL-NEXT: s_wait_loadcnt 0x0
; GFX12-GISEL-NEXT: v_mov_b32_e32 v0, 1.0
; GFX12-GISEL-NEXT: ; return to shader part epilog
@@ -192,14 +192,14 @@ main_body:
define amdgpu_ps float @atomic_pk_add_bf16_1d_v4_nt(<8 x i32> inreg %rsrc, <4 x bfloat> %data, i32 %s) {
; GFX12-SDAG-LABEL: atomic_pk_add_bf16_1d_v4_nt:
; GFX12-SDAG: ; %bb.0: ; %main_body
-; GFX12-SDAG-NEXT: image_atomic_pk_add_bf16 v[0:1], v2, s[0:7] dmask:0x3 dim:SQ_RSRC_IMG_1D th:TH_ATOMIC_NT_RETURN
+; GFX12-SDAG-NEXT: image_atomic_pk_add_bf16 v[0:1], v2, s[0:7] dmask:0x3 dim:SQ_RSRC_IMG_1D th:TH_ATOMIC_NT
; GFX12-SDAG-NEXT: s_wait_loadcnt 0x0
; GFX12-SDAG-NEXT: v_mov_b32_e32 v0, 1.0
; GFX12-SDAG-NEXT: ; return to shader part epilog
;
; GFX12-GISEL-LABEL: atomic_pk_add_bf16_1d_v4_nt:
; GFX12-GISEL: ; %bb.0: ; %main_body
-; GFX12-GISEL-NEXT: image_atomic_pk_add_bf16 v[0:1], v2, s[0:7] dmask:0x3 dim:SQ_RSRC_IMG_1D th:TH_ATOMIC_NT_RETURN
+; GFX12-GISEL-NEXT: image_atomic_pk_add_bf16 v[0:1], v2, s[0:7] dmask:0x3 dim:SQ_RSRC_IMG_1D th:TH_ATOMIC_NT
; GFX12-GISEL-NEXT: s_wait_loadcnt 0x0
; GFX12-GISEL-NEXT: v_mov_b32_e32 v0, 1.0
; GFX12-GISEL-NEXT: ; return to shader part epilog
More information about the llvm-commits
mailing list