[llvm] [AMDGPU] Introduce orderign parameter to atomic intrinsics and introduce new llvm.amdgcn.image.atomic.load intrinsic. (PR #73613)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Dec 22 05:38:56 PST 2023
https://github.com/sstipanovic updated https://github.com/llvm/llvm-project/pull/73613
>From 33960a42cad7f7156017fa05fab5a6b2455fb3f9 Mon Sep 17 00:00:00 2001
From: Stefan Stipanovic <Stefan.Stipanovic at amd.com>
Date: Fri, 22 Dec 2023 14:38:14 +0100
Subject: [PATCH] [AMDGPU] Introduce llvm.amdgcn.image.atomic.load intrinsic.
This intrinsic should behave mostly identically to an llvm.amdgcn.image.load,
except that: - It is not marked as IntrReadMem. This is to ensure that
the implied memory semantics are preserved. - When lowering, it's
MachineMemOperand is to get the "acquire" memory semantics.
---
llvm/include/llvm/CodeGen/SelectionDAG.h | 9 ++++--
llvm/include/llvm/CodeGen/TargetLowering.h | 2 ++
llvm/include/llvm/IR/IntrinsicsAMDGPU.td | 16 ++++++----
llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp | 6 ++--
.../lib/CodeGen/SelectionDAG/SelectionDAG.cpp | 10 +++++--
.../SelectionDAG/SelectionDAGBuilder.cpp | 6 ++--
llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.h | 1 +
.../AMDGPU/AMDGPUInstructionSelector.cpp | 2 ++
llvm/lib/Target/AMDGPU/MIMGInstructions.td | 30 +++++++++++++++++++
llvm/lib/Target/AMDGPU/SIDefines.h | 2 ++
llvm/lib/Target/AMDGPU/SIISelLowering.cpp | 14 +++++++++
llvm/test/CodeGen/AMDGPU/atomic-image-load.ll | 29 ++++++++++++++++++
12 files changed, 111 insertions(+), 16 deletions(-)
create mode 100644 llvm/test/CodeGen/AMDGPU/atomic-image-load.ll
diff --git a/llvm/include/llvm/CodeGen/SelectionDAG.h b/llvm/include/llvm/CodeGen/SelectionDAG.h
index e867448b9d5512..c7a6d4862969b1 100644
--- a/llvm/include/llvm/CodeGen/SelectionDAG.h
+++ b/llvm/include/llvm/CodeGen/SelectionDAG.h
@@ -36,6 +36,7 @@
#include "llvm/IR/Metadata.h"
#include "llvm/Support/Allocator.h"
#include "llvm/Support/ArrayRecycler.h"
+#include "llvm/Support/AtomicOrdering.h"
#include "llvm/Support/CodeGen.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/RecyclingAllocator.h"
@@ -1297,7 +1298,8 @@ class SelectionDAG {
EVT MemVT, MachinePointerInfo PtrInfo, Align Alignment,
MachineMemOperand::Flags Flags = MachineMemOperand::MOLoad |
MachineMemOperand::MOStore,
- uint64_t Size = 0, const AAMDNodes &AAInfo = AAMDNodes());
+ uint64_t Size = 0, const AAMDNodes &AAInfo = AAMDNodes(),
+ AtomicOrdering Ordering = AtomicOrdering::NotAtomic);
inline SDValue getMemIntrinsicNode(
unsigned Opcode, const SDLoc &dl, SDVTList VTList, ArrayRef<SDValue> Ops,
@@ -1305,11 +1307,12 @@ class SelectionDAG {
MaybeAlign Alignment = std::nullopt,
MachineMemOperand::Flags Flags = MachineMemOperand::MOLoad |
MachineMemOperand::MOStore,
- uint64_t Size = 0, const AAMDNodes &AAInfo = AAMDNodes()) {
+ uint64_t Size = 0, const AAMDNodes &AAInfo = AAMDNodes(),
+ AtomicOrdering Ordering = AtomicOrdering::NotAtomic) {
// Ensure that codegen never sees alignment 0
return getMemIntrinsicNode(Opcode, dl, VTList, Ops, MemVT, PtrInfo,
Alignment.value_or(getEVTAlign(MemVT)), Flags,
- Size, AAInfo);
+ Size, AAInfo, Ordering);
}
SDValue getMemIntrinsicNode(unsigned Opcode, const SDLoc &dl, SDVTList VTList,
diff --git a/llvm/include/llvm/CodeGen/TargetLowering.h b/llvm/include/llvm/CodeGen/TargetLowering.h
index 490125164ab341..78b423177c233f 100644
--- a/llvm/include/llvm/CodeGen/TargetLowering.h
+++ b/llvm/include/llvm/CodeGen/TargetLowering.h
@@ -1154,6 +1154,8 @@ class TargetLoweringBase {
MaybeAlign align = Align(1); // alignment
MachineMemOperand::Flags flags = MachineMemOperand::MONone;
+
+ AtomicOrdering ordering = AtomicOrdering::NotAtomic;
IntrinsicInfo() = default;
};
diff --git a/llvm/include/llvm/IR/IntrinsicsAMDGPU.td b/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
index 06f9c0445bceac..03c04605ba8137 100644
--- a/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
+++ b/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
@@ -796,7 +796,8 @@ class AMDGPUImageDimIntrinsicEval<AMDGPUDimProfile P_> {
// All dimension-aware intrinsics are derived from this class.
class AMDGPUImageDimIntrinsic<AMDGPUDimProfile P_,
list<IntrinsicProperty> props,
- list<SDNodeProperty> sdnodeprops> : DefaultAttrsIntrinsic<
+ list<SDNodeProperty> sdnodeprops,
+ string name = ""> : Intrinsic<
P_.RetTypes, // vdata(VGPR) -- for load/atomic-with-return
!listconcat(
!foreach(arg, P_.DataArgs, arg.Type), // vdata(VGPR) -- for store/atomic
@@ -806,13 +807,14 @@ class AMDGPUImageDimIntrinsic<AMDGPUDimProfile P_,
!if(P_.IsSample, [llvm_v4i32_ty, // samp(SGPR)
llvm_i1_ty], []), // unorm(imm)
[llvm_i32_ty, // texfailctrl(imm; bit 0 = tfe, bit 1 = lwe)
- llvm_i32_ty]), // cachepolicy(imm; bit 0 = glc, bit 1 = slc, bit 2 = dlc)
+ llvm_i32_ty]), // cachepolicy(imm; bit 0 = glc, bit 1 = slc, bit 2 = dlc)
- !listconcat(props,
+ !listconcat(props, [IntrNoCallback, IntrNoFree, IntrWillReturn],
!if(P_.IsAtomic, [], [ImmArg<ArgIndex<AMDGPUImageDimIntrinsicEval<P_>.DmaskArgIndex>>]),
!if(P_.IsSample, [ImmArg<ArgIndex<AMDGPUImageDimIntrinsicEval<P_>.UnormArgIndex>>], []),
[ImmArg<ArgIndex<AMDGPUImageDimIntrinsicEval<P_>.TexFailCtrlArgIndex>>,
- ImmArg<ArgIndex<AMDGPUImageDimIntrinsicEval<P_>.CachePolicyArgIndex>>]),
+ ImmArg<ArgIndex<AMDGPUImageDimIntrinsicEval<P_>.CachePolicyArgIndex>>],
+ !if(!or(P_.IsAtomic, !eq(name, "int_amdgcn_image_atomic_load")), [], [IntrNoSync])),
"", sdnodeprops>,
@@ -858,7 +860,7 @@ defset list<AMDGPUImageDimIntrinsic> AMDGPUImageDimIntrinsics = {
def !strconcat(NAME, "_", dim.Name)
: AMDGPUImageDimIntrinsic<
AMDGPUDimNoSampleProfile<opmod, dim, retty, dataargs, Mip>,
- props, sdnodeprops>;
+ props, sdnodeprops, NAME>;
}
}
@@ -866,6 +868,10 @@ defset list<AMDGPUImageDimIntrinsic> AMDGPUImageDimIntrinsics = {
: AMDGPUImageDimIntrinsicsAll<"LOAD", [llvm_any_ty], [], [IntrReadMem],
[SDNPMemOperand]>,
AMDGPUImageDMaskIntrinsic;
+ defm int_amdgcn_image_atomic_load
+ : AMDGPUImageDimIntrinsicsAll<"ATOMIC_LOAD", [llvm_any_ty], [], [],
+ [SDNPMemOperand]>,
+ AMDGPUImageDMaskIntrinsic;
defm int_amdgcn_image_load_mip
: AMDGPUImageDimIntrinsicsNoMsaa<"LOAD_MIP", [llvm_any_ty], [],
[IntrReadMem, IntrWillReturn], [SDNPMemOperand], 1>,
diff --git a/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp b/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp
index 14a4e72152e7c4..93127600e0d125 100644
--- a/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/IRTranslator.cpp
@@ -2607,8 +2607,10 @@ bool IRTranslator::translateCall(const User &U, MachineIRBuilder &MIRBuilder) {
MPI = MachinePointerInfo(Info.ptrVal, Info.offset);
else if (Info.fallbackAddressSpace)
MPI = MachinePointerInfo(*Info.fallbackAddressSpace);
- MIB.addMemOperand(
- MF->getMachineMemOperand(MPI, Info.flags, MemTy, Alignment, CI.getAAMetadata()));
+ MIB.addMemOperand(MF->getMachineMemOperand(
+ MPI, Info.flags, MemTy, Alignment, CI.getAAMetadata(),
+ /*Ranges*/ nullptr, /*SSID*/ SyncScope::System, Info.ordering,
+ Info.ordering));
}
return true;
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 51ae8b703e50f2..9e94d6016b5cb1 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -55,8 +55,10 @@
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/GlobalValue.h"
+#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/Metadata.h"
#include "llvm/IR/Type.h"
+#include "llvm/Support/AtomicOrdering.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CodeGen.h"
#include "llvm/Support/Compiler.h"
@@ -8335,15 +8337,17 @@ SDValue SelectionDAG::getMergeValues(ArrayRef<SDValue> Ops, const SDLoc &dl) {
SDValue SelectionDAG::getMemIntrinsicNode(
unsigned Opcode, const SDLoc &dl, SDVTList VTList, ArrayRef<SDValue> Ops,
EVT MemVT, MachinePointerInfo PtrInfo, Align Alignment,
- MachineMemOperand::Flags Flags, uint64_t Size, const AAMDNodes &AAInfo) {
+ MachineMemOperand::Flags Flags, uint64_t Size, const AAMDNodes &AAInfo,
+ AtomicOrdering Ordering) {
if (!Size && MemVT.isScalableVector())
Size = MemoryLocation::UnknownSize;
else if (!Size)
Size = MemVT.getStoreSize();
MachineFunction &MF = getMachineFunction();
- MachineMemOperand *MMO =
- MF.getMachineMemOperand(PtrInfo, Flags, Size, Alignment, AAInfo);
+ MachineMemOperand *MMO = MF.getMachineMemOperand(
+ PtrInfo, Flags, Size, Alignment, AAInfo, /*Ranges*/ nullptr,
+ /*SSID*/ SyncScope::System, Ordering, Ordering);
return getMemIntrinsicNode(Opcode, dl, VTList, Ops, MemVT, MMO);
}
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
index ed1c96a873748f..f0ff6b16756f8b 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -4986,9 +4986,9 @@ void SelectionDAGBuilder::visitTargetIntrinsic(const CallInst &I,
MPI = MachinePointerInfo(Info.ptrVal, Info.offset);
else if (Info.fallbackAddressSpace)
MPI = MachinePointerInfo(*Info.fallbackAddressSpace);
- Result = DAG.getMemIntrinsicNode(Info.opc, getCurSDLoc(), VTs, Ops,
- Info.memVT, MPI, Info.align, Info.flags,
- Info.size, I.getAAMetadata());
+ Result = DAG.getMemIntrinsicNode(
+ Info.opc, getCurSDLoc(), VTs, Ops, Info.memVT, MPI, Info.align,
+ Info.flags, Info.size, I.getAAMetadata(), Info.ordering);
} else if (!HasChain) {
Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurSDLoc(), VTs, Ops);
} else if (!I.getType()->isVoidTy()) {
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.h b/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.h
index 515decea3921ed..585d99ffc862ff 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.h
@@ -81,6 +81,7 @@ struct ImageDimIntrinsicInfo {
uint8_t UnormIndex;
uint8_t TexFailCtrlIndex;
uint8_t CachePolicyIndex;
+ // uint8_t AtomicOrderingIndex;
uint8_t BiasTyArg;
uint8_t GradientTyArg;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
index d24c7da964ce85..902f7874067c49 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
@@ -17,6 +17,7 @@
#include "AMDGPUInstrInfo.h"
#include "AMDGPURegisterBankInfo.h"
#include "AMDGPUTargetMachine.h"
+#include "SIDefines.h"
#include "SIMachineFunctionInfo.h"
#include "Utils/AMDGPUBaseInfo.h"
#include "llvm/CodeGen/GlobalISel/GIMatchTableExecutorImpl.h"
@@ -1901,6 +1902,7 @@ bool AMDGPUInstructionSelector::selectImageIntrinsic(
assert((!IsTexFail || DMaskLanes >= 1) && "should have legalized this");
unsigned CPol = MI.getOperand(ArgOffset + Intr->CachePolicyIndex).getImm();
+ CPol &= ~AMDGPU::CPol::ATOMIC_ORDERING;
if (BaseOpcode->Atomic)
CPol |= AMDGPU::CPol::GLC; // TODO no-return optimization
if (CPol & ~AMDGPU::CPol::ALL)
diff --git a/llvm/lib/Target/AMDGPU/MIMGInstructions.td b/llvm/lib/Target/AMDGPU/MIMGInstructions.td
index 240366c8e7daae..25aabcc5b0b405 100644
--- a/llvm/lib/Target/AMDGPU/MIMGInstructions.td
+++ b/llvm/lib/Target/AMDGPU/MIMGInstructions.td
@@ -653,6 +653,31 @@ multiclass MIMG_NoSampler <mimgopc op, string asm, bit has_d16, bit mip = 0,
}
}
+multiclass MIMG_NoSampler_Pseudo <mimgopc op, string asm, bit has_d16, bit mip = 0,
+ bit isResInfo = 0,
+ bit msaa = 0> {
+ def "" : MIMGBaseOpcode {
+ let Coordinates = !not(isResInfo);
+ let LodOrClampOrMip = mip;
+ let HasD16 = has_d16;
+ let MSAA = msaa;
+ }
+
+ let BaseOpcode = !cast<MIMGBaseOpcode>(NAME),
+ mayLoad = !not(isResInfo), isCodeGenOnly = 1 in {
+ let VDataDwords = 1 in
+ defm _V1 : MIMG_NoSampler_Src_Helper <op, asm, VGPR_32, 1, msaa>;
+ let VDataDwords = 2 in
+ defm _V2 : MIMG_NoSampler_Src_Helper <op, asm, VReg_64, 0, msaa>;
+ let VDataDwords = 3 in
+ defm _V3 : MIMG_NoSampler_Src_Helper <op, asm, VReg_96, 0, msaa>;
+ let VDataDwords = 4 in
+ defm _V4 : MIMG_NoSampler_Src_Helper <op, asm, VReg_128, 0, msaa>;
+ let VDataDwords = 5 in
+ defm _V5 : MIMG_NoSampler_Src_Helper <op, asm, VReg_160, 0, msaa>;
+ }
+}
+
class MIMG_Store_Helper <mimgopc op, string asm,
RegisterClass data_rc,
RegisterClass addr_rc,
@@ -1554,7 +1579,11 @@ defm IMAGE_ATOMIC_FCMPSWAP : MIMG_Atomic <mimgopc<MIMG.NOP, MIMG.NOP, 0x1d,
defm IMAGE_ATOMIC_FMIN : MIMG_Atomic <mimgopc<MIMG.NOP, MIMG.NOP, 0x1e, MIMG.NOP>, "image_atomic_fmin", 0, 1>;
defm IMAGE_ATOMIC_FMAX : MIMG_Atomic <mimgopc<MIMG.NOP, MIMG.NOP, 0x1f, MIMG.NOP>, "image_atomic_fmax", 0, 1>;
+
+defm IMAGE_ATOMIC_LOAD : MIMG_NoSampler_Pseudo <mimgopc<0x00, 0x00, 0x00>, "image_load", 1>;
+
defm IMAGE_SAMPLE : MIMG_Sampler_WQM <mimgopc<0x1b, 0x1b, 0x20>, AMDGPUSample>;
+
let OtherPredicates = [HasExtendedImageInsts] in {
defm IMAGE_SAMPLE_CL : MIMG_Sampler_WQM <mimgopc<0x40, 0x40, 0x21>, AMDGPUSample_cl>;
defm IMAGE_SAMPLE_D : MIMG_Sampler <mimgopc<0x1c, 0x1c, 0x22>, AMDGPUSample_d>;
@@ -1698,6 +1727,7 @@ class ImageDimIntrinsicInfo<AMDGPUImageDimIntrinsic I> {
bits<8> UnormIndex = DimEval.UnormArgIndex;
bits<8> TexFailCtrlIndex = DimEval.TexFailCtrlArgIndex;
bits<8> CachePolicyIndex = DimEval.CachePolicyArgIndex;
+ // bits<8> AtomicOrderingIndex = DimEval.AtomicOrderingIndex;
bits<8> BiasTyArg = !add(I.P.NumRetAndDataAnyTypes,
!if(!eq(NumOffsetArgs, 0), 0, I.P.ExtraAddrArgs[0].Type.isAny));
diff --git a/llvm/lib/Target/AMDGPU/SIDefines.h b/llvm/lib/Target/AMDGPU/SIDefines.h
index 659ff75e13d01f..cf6eedf5f2b72d 100644
--- a/llvm/lib/Target/AMDGPU/SIDefines.h
+++ b/llvm/lib/Target/AMDGPU/SIDefines.h
@@ -353,6 +353,7 @@ enum CPol {
SLC = 2,
DLC = 4,
SCC = 16,
+ // ATOMIC_ORDERING = 32,
SC0 = GLC,
SC1 = SCC,
NT = SLC,
@@ -397,6 +398,7 @@ enum CPol {
TH_TYPE_STORE = 1 << 8, // TH_STORE policy
TH_TYPE_ATOMIC = 1 << 9, // TH_ATOMIC policy
TH_REAL_BYPASS = 1 << 10, // is TH=3 bypass policy or not
+ ATOMIC_ORDERING = 0x70, // Atomic ordering bits mask
};
} // namespace CPol
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index a7f4d63229b7ef..e3a03e88a95545 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -17,8 +17,10 @@
#include "AMDGPUTargetMachine.h"
#include "GCNSubtarget.h"
#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
+#include "SIDefines.h"
#include "SIMachineFunctionInfo.h"
#include "SIRegisterInfo.h"
+#include "Utils/AMDGPUBaseInfo.h"
#include "llvm/ADT/APInt.h"
#include "llvm/ADT/FloatingPointMode.h"
#include "llvm/ADT/Statistic.h"
@@ -39,6 +41,7 @@
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/IntrinsicsAMDGPU.h"
#include "llvm/IR/IntrinsicsR600.h"
+#include "llvm/Support/AtomicOrdering.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Support/ModRef.h"
@@ -1131,6 +1134,17 @@ bool SITargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
// XXX - Should this be volatile without known ordering?
Info.flags |= MachineMemOperand::MOVolatile;
+ if (RsrcIntr->IsImage) {
+ auto Idx = CI.arg_size() - 1;
+ unsigned OrderingArg =
+ cast<ConstantInt>(CI.getArgOperand(Idx))->getZExtValue();
+ auto Ordering = (OrderingArg & AMDGPU::CPol::ATOMIC_ORDERING) >> 4;
+ unsigned ClearedCPol = OrderingArg & ~AMDGPU::CPol::ATOMIC_ORDERING;
+ ConstantInt *CPol = ConstantInt::get(IntegerType::getInt32Ty(CI.getContext()), ClearedCPol);
+ const_cast<CallInst &>(CI).setArgOperand(Idx, CPol);
+ Info.ordering = static_cast<AtomicOrdering>(Ordering);
+ }
+
switch (IntrID) {
default:
break;
diff --git a/llvm/test/CodeGen/AMDGPU/atomic-image-load.ll b/llvm/test/CodeGen/AMDGPU/atomic-image-load.ll
new file mode 100644
index 00000000000000..95e2695ee97ffc
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/atomic-image-load.ll
@@ -0,0 +1,29 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=gfx1010 < %s | FileCheck %s -check-prefix=GFX10
+; RUN: llc -march=amdgcn -mcpu=gfx900 < %s | FileCheck %s -check-prefix=GFX9
+; RUN: llc -march=amdgcn -mcpu=gfx1100 < %s | FileCheck %s -check-prefix=GFX11
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx1010 < %s | FileCheck %s -check-prefix=GFX10
+
+define amdgpu_ps void @test(<8 x i32> inreg %load, <8 x i32> inreg %store) {
+; GFX10-LABEL: test:
+; GFX10: ; %bb.0:
+; GFX10-NEXT: v_mov_b32_e32 v0, 0
+; GFX10-NEXT: image_load v0, v0, s[0:7] dmask:0x1 dim:SQ_RSRC_IMG_1D unorm
+; GFX10-NEXT: s_endpgm
+;
+; GFX9-LABEL: test:
+; GFX9: ; %bb.0:
+; GFX9-NEXT: v_mov_b32_e32 v0, 0
+; GFX9-NEXT: image_load v0, v0, s[0:7] dmask:0x1 unorm
+; GFX9-NEXT: s_endpgm
+;
+; GFX11-LABEL: test:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: v_mov_b32_e32 v0, 0
+; GFX11-NEXT: image_load v0, v0, s[0:7] dmask:0x1 dim:SQ_RSRC_IMG_1D unorm
+; GFX11-NEXT: s_endpgm
+ %data0 = call float @llvm.amdgcn.image.atomic.load.1d.f32.i32(i32 1, i32 0, <8 x i32> %load, i32 0, i32 112)
+ ret void
+}
+
+declare float @llvm.amdgcn.image.atomic.load.1d.f32.i32(i32 immarg, i32, <8 x i32>, i32 immarg, i32 immarg)
More information about the llvm-commits
mailing list