[llvm] 7c8b127 - [SPIR-V] Add pass to remove spv_ptrcast intrinsics (#128896)
via llvm-commits
llvm-commits at lists.llvm.org
Tue Mar 4 01:30:51 PST 2025
Author: Nathan Gauër
Date: 2025-03-04T10:30:46+01:00
New Revision: 7c8b1275bc87f6b5983788a6e603504725b490c9
URL: https://github.com/llvm/llvm-project/commit/7c8b1275bc87f6b5983788a6e603504725b490c9
DIFF: https://github.com/llvm/llvm-project/commit/7c8b1275bc87f6b5983788a6e603504725b490c9.diff
LOG: [SPIR-V] Add pass to remove spv_ptrcast intrinsics (#128896)
OpenCL is allowed to cast pointers, meaning they can resolve some type
mismatches this way. In logical SPIR-V, those are restricted. This new
pass legalizes such pointer cast when targeting logical SPIR-V.
For now, this pass supports 3 cases we witnessed:
- loading a vec3 from a vec4*.
- loading a scalar from a vec*.
- loading the 1st element of an array.
---------
Co-authored-by: Steven Perron <stevenperron at google.com>
Added:
llvm/lib/Target/SPIRV/SPIRVLegalizePointerCast.cpp
llvm/test/CodeGen/SPIRV/pointers/array-skips-gep.ll
llvm/test/CodeGen/SPIRV/pointers/getelementptr-downcast-vector.ll
Modified:
llvm/lib/Target/SPIRV/CMakeLists.txt
llvm/lib/Target/SPIRV/SPIRV.h
llvm/lib/Target/SPIRV/SPIRVEmitIntrinsics.cpp
llvm/lib/Target/SPIRV/SPIRVGlobalRegistry.cpp
llvm/lib/Target/SPIRV/SPIRVGlobalRegistry.h
llvm/lib/Target/SPIRV/SPIRVTargetMachine.cpp
llvm/lib/Target/SPIRV/SPIRVUtils.cpp
llvm/lib/Target/SPIRV/SPIRVUtils.h
Removed:
################################################################################
diff --git a/llvm/lib/Target/SPIRV/CMakeLists.txt b/llvm/lib/Target/SPIRV/CMakeLists.txt
index efdd8c8d24fbd..4a2b534b948d6 100644
--- a/llvm/lib/Target/SPIRV/CMakeLists.txt
+++ b/llvm/lib/Target/SPIRV/CMakeLists.txt
@@ -27,6 +27,7 @@ add_llvm_target(SPIRVCodeGen
SPIRVInstrInfo.cpp
SPIRVInstructionSelector.cpp
SPIRVStripConvergentIntrinsics.cpp
+ SPIRVLegalizePointerCast.cpp
SPIRVMergeRegionExitTargets.cpp
SPIRVISelLowering.cpp
SPIRVLegalizerInfo.cpp
diff --git a/llvm/lib/Target/SPIRV/SPIRV.h b/llvm/lib/Target/SPIRV/SPIRV.h
index 6d00a046ff7ca..d765dfe370be2 100644
--- a/llvm/lib/Target/SPIRV/SPIRV.h
+++ b/llvm/lib/Target/SPIRV/SPIRV.h
@@ -23,6 +23,7 @@ ModulePass *createSPIRVPrepareFunctionsPass(const SPIRVTargetMachine &TM);
FunctionPass *createSPIRVStructurizerPass();
FunctionPass *createSPIRVMergeRegionExitTargetsPass();
FunctionPass *createSPIRVStripConvergenceIntrinsicsPass();
+FunctionPass *createSPIRVLegalizePointerCastPass(SPIRVTargetMachine *TM);
FunctionPass *createSPIRVRegularizerPass();
FunctionPass *createSPIRVPreLegalizerCombiner();
FunctionPass *createSPIRVPreLegalizerPass();
diff --git a/llvm/lib/Target/SPIRV/SPIRVEmitIntrinsics.cpp b/llvm/lib/Target/SPIRV/SPIRVEmitIntrinsics.cpp
index d6177058231d9..3e032f5ed4ba4 100644
--- a/llvm/lib/Target/SPIRV/SPIRVEmitIntrinsics.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVEmitIntrinsics.cpp
@@ -57,12 +57,6 @@ void initializeSPIRVEmitIntrinsicsPass(PassRegistry &);
namespace {
-inline MetadataAsValue *buildMD(Value *Arg) {
- LLVMContext &Ctx = Arg->getContext();
- return MetadataAsValue::get(
- Ctx, MDNode::get(Ctx, ValueAsMetadata::getConstant(Arg)));
-}
-
class SPIRVEmitIntrinsics
: public ModulePass,
public InstVisitor<SPIRVEmitIntrinsics, Instruction *> {
@@ -142,24 +136,9 @@ class SPIRVEmitIntrinsics
void preprocessCompositeConstants(IRBuilder<> &B);
void preprocessUndefs(IRBuilder<> &B);
- CallInst *buildIntrWithMD(Intrinsic::ID IntrID, ArrayRef<Type *> Types,
- Value *Arg, Value *Arg2, ArrayRef<Constant *> Imms,
- IRBuilder<> &B) {
- SmallVector<Value *, 4> Args;
- Args.push_back(Arg2);
- Args.push_back(buildMD(Arg));
- for (auto *Imm : Imms)
- Args.push_back(Imm);
- return B.CreateIntrinsic(IntrID, {Types}, Args);
- }
-
Type *reconstructType(Value *Op, bool UnknownElemTypeI8,
bool IsPostprocessing);
- void buildAssignType(IRBuilder<> &B, Type *ElemTy, Value *Arg);
- void buildAssignPtr(IRBuilder<> &B, Type *ElemTy, Value *Arg);
- void updateAssignType(CallInst *AssignCI, Value *Arg, Value *OfType);
-
void replaceMemInstrUses(Instruction *Old, Instruction *New, IRBuilder<> &B);
void processInstrAfterVisit(Instruction *I, IRBuilder<> &B);
bool insertAssignPtrTypeIntrs(Instruction *I, IRBuilder<> &B,
@@ -273,18 +252,6 @@ bool expectIgnoredInIRTranslation(const Instruction *I) {
}
}
-bool allowEmitFakeUse(const Value *Arg) {
- if (isSpvIntrinsic(Arg))
- return false;
- if (dyn_cast<AtomicCmpXchgInst>(Arg) || dyn_cast<InsertValueInst>(Arg) ||
- dyn_cast<UndefValue>(Arg))
- return false;
- if (const auto *LI = dyn_cast<LoadInst>(Arg))
- if (LI->getType()->isAggregateType())
- return false;
- return true;
-}
-
} // namespace
char SPIRVEmitIntrinsics::ID = 0;
@@ -355,10 +322,7 @@ static void emitAssignName(Instruction *I, IRBuilder<> &B) {
void SPIRVEmitIntrinsics::replaceAllUsesWith(Value *Src, Value *Dest,
bool DeleteOld) {
- Src->replaceAllUsesWith(Dest);
- // Update deduced type records
- GR->updateIfExistDeducedElementType(Src, Dest, DeleteOld);
- GR->updateIfExistAssignPtrTypeInstr(Src, Dest, DeleteOld);
+ GR->replaceAllUsesWith(Src, Dest, DeleteOld);
// Update uncomplete type records if any
if (isTodoType(Src)) {
if (DeleteOld)
@@ -425,57 +389,6 @@ Type *SPIRVEmitIntrinsics::reconstructType(Value *Op, bool UnknownElemTypeI8,
return nullptr;
}
-void SPIRVEmitIntrinsics::buildAssignType(IRBuilder<> &B, Type *Ty,
- Value *Arg) {
- Value *OfType = getNormalizedPoisonValue(Ty);
- CallInst *AssignCI = nullptr;
- if (Arg->getType()->isAggregateType() && Ty->isAggregateType() &&
- allowEmitFakeUse(Arg)) {
- LLVMContext &Ctx = Arg->getContext();
- SmallVector<Metadata *, 2> ArgMDs{
- MDNode::get(Ctx, ValueAsMetadata::getConstant(OfType)),
- MDString::get(Ctx, Arg->getName())};
- B.CreateIntrinsic(Intrinsic::spv_value_md, {},
- {MetadataAsValue::get(Ctx, MDTuple::get(Ctx, ArgMDs))});
- AssignCI = B.CreateIntrinsic(Intrinsic::fake_use, {}, {Arg});
- } else {
- AssignCI = buildIntrWithMD(Intrinsic::spv_assign_type, {Arg->getType()},
- OfType, Arg, {}, B);
- }
- GR->addAssignPtrTypeInstr(Arg, AssignCI);
-}
-
-void SPIRVEmitIntrinsics::buildAssignPtr(IRBuilder<> &B, Type *ElemTy,
- Value *Arg) {
- ElemTy = normalizeType(ElemTy);
- Value *OfType = PoisonValue::get(ElemTy);
- CallInst *AssignPtrTyCI = GR->findAssignPtrTypeInstr(Arg);
- if (AssignPtrTyCI == nullptr ||
- AssignPtrTyCI->getParent()->getParent() != CurrF) {
- AssignPtrTyCI = buildIntrWithMD(
- Intrinsic::spv_assign_ptr_type, {Arg->getType()}, OfType, Arg,
- {B.getInt32(getPointerAddressSpace(Arg->getType()))}, B);
- GR->addDeducedElementType(AssignPtrTyCI, ElemTy);
- GR->addDeducedElementType(Arg, ElemTy);
- GR->addAssignPtrTypeInstr(Arg, AssignPtrTyCI);
- } else {
- updateAssignType(AssignPtrTyCI, Arg, OfType);
- }
-}
-
-void SPIRVEmitIntrinsics::updateAssignType(CallInst *AssignCI, Value *Arg,
- Value *OfType) {
- AssignCI->setArgOperand(1, buildMD(OfType));
- if (cast<IntrinsicInst>(AssignCI)->getIntrinsicID() !=
- Intrinsic::spv_assign_ptr_type)
- return;
-
- // update association with the pointee type
- Type *ElemTy = normalizeType(OfType->getType());
- GR->addDeducedElementType(AssignCI, ElemTy);
- GR->addDeducedElementType(Arg, ElemTy);
-}
-
CallInst *SPIRVEmitIntrinsics::buildSpvPtrcast(Function *F, Value *Op,
Type *ElemTy) {
IRBuilder<> B(Op->getContext());
@@ -495,7 +408,7 @@ CallInst *SPIRVEmitIntrinsics::buildSpvPtrcast(Function *F, Value *Op,
B.getInt32(getPointerAddressSpace(OpTy))};
CallInst *PtrCasted =
B.CreateIntrinsic(Intrinsic::spv_ptrcast, {Types}, Args);
- buildAssignPtr(B, ElemTy, PtrCasted);
+ GR->buildAssignPtr(B, ElemTy, PtrCasted);
return PtrCasted;
}
@@ -1032,7 +945,8 @@ bool SPIRVEmitIntrinsics::deduceOperandElementTypeFunctionRet(
continue;
if (CallInst *AssignCI = GR->findAssignPtrTypeInstr(CI)) {
if (Type *PrevElemTy = GR->findDeducedElementType(CI)) {
- updateAssignType(AssignCI, CI, getNormalizedPoisonValue(OpElemTy));
+ GR->updateAssignType(AssignCI, CI,
+ getNormalizedPoisonValue(OpElemTy));
propagateElemType(CI, PrevElemTy, VisitedSubst);
}
}
@@ -1218,7 +1132,7 @@ void SPIRVEmitIntrinsics::deduceOperandElementType(
{B.getInt32(getPointerAddressSpace(OpTy))}, B);
GR->addAssignPtrTypeInstr(Op, CI);
} else {
- updateAssignType(AssignCI, Op, OpTyVal);
+ GR->updateAssignType(AssignCI, Op, OpTyVal);
DenseSet<std::pair<Value *, Value *>> VisitedSubst{
std::make_pair(I, Op)};
propagateElemTypeRec(Op, KnownElemTy, PrevElemTy, VisitedSubst);
@@ -1508,7 +1422,7 @@ void SPIRVEmitIntrinsics::insertAssignPtrTypeTargetExt(
CallInst *AssignCI = GR->findAssignPtrTypeInstr(V);
if (!AssignCI) {
- buildAssignType(B, AssignedType, V);
+ GR->buildAssignType(B, AssignedType, V);
return;
}
@@ -1528,7 +1442,7 @@ void SPIRVEmitIntrinsics::insertAssignPtrTypeTargetExt(
// Our previous guess about the type seems to be wrong, let's update
// inferred type according to a new, more precise type information.
- updateAssignType(AssignCI, V, getNormalizedPoisonValue(AssignedType));
+ GR->updateAssignType(AssignCI, V, getNormalizedPoisonValue(AssignedType));
}
void SPIRVEmitIntrinsics::replacePointerOperandWithPtrCast(
@@ -1585,7 +1499,7 @@ void SPIRVEmitIntrinsics::replacePointerOperandWithPtrCast(
if (FirstPtrCastOrAssignPtrType) {
// If this would be the first spv_ptrcast, do not emit spv_ptrcast and
// emit spv_assign_ptr_type instead.
- buildAssignPtr(B, ExpectedElementType, Pointer);
+ GR->buildAssignPtr(B, ExpectedElementType, Pointer);
return;
} else if (isTodoType(Pointer)) {
eraseTodoType(Pointer);
@@ -1597,10 +1511,10 @@ void SPIRVEmitIntrinsics::replacePointerOperandWithPtrCast(
assert(PrevElemTy);
DenseSet<std::pair<Value *, Value *>> VisitedSubst{
std::make_pair(I, Pointer)};
- updateAssignType(AssignCI, Pointer, ExpectedElementVal);
+ GR->updateAssignType(AssignCI, Pointer, ExpectedElementVal);
propagateElemType(Pointer, PrevElemTy, VisitedSubst);
} else {
- buildAssignPtr(B, ExpectedElementType, Pointer);
+ GR->buildAssignPtr(B, ExpectedElementType, Pointer);
}
return;
}
@@ -1613,7 +1527,7 @@ void SPIRVEmitIntrinsics::replacePointerOperandWithPtrCast(
auto *PtrCastI = B.CreateIntrinsic(Intrinsic::spv_ptrcast, {Types}, Args);
I->setOperand(OperandToReplace, PtrCastI);
// We need to set up a pointee type for the newly created spv_ptrcast.
- buildAssignPtr(B, ExpectedElementType, PtrCastI);
+ GR->buildAssignPtr(B, ExpectedElementType, PtrCastI);
}
void SPIRVEmitIntrinsics::insertPtrCastOrAssignTypeInstr(Instruction *I,
@@ -1929,7 +1843,7 @@ bool SPIRVEmitIntrinsics::insertAssignPtrTypeIntrs(Instruction *I,
setInsertPointAfterDef(B, I);
if (Type *ElemTy = deduceElementType(I, UnknownElemTypeI8)) {
- buildAssignPtr(B, ElemTy, I);
+ GR->buildAssignPtr(B, ElemTy, I);
return false;
}
return true;
@@ -1962,8 +1876,8 @@ void SPIRVEmitIntrinsics::insertAssignTypeIntrs(Instruction *I,
setInsertPointAfterDef(B, I);
switch (ResIt->second) {
case WellKnownTypes::Event:
- buildAssignType(B, TargetExtType::get(I->getContext(), "spirv.Event"),
- I);
+ GR->buildAssignType(
+ B, TargetExtType::get(I->getContext(), "spirv.Event"), I);
break;
}
}
@@ -2008,7 +1922,7 @@ void SPIRVEmitIntrinsics::insertAssignTypeIntrs(Instruction *I,
}
}
TypeToAssign = restoreMutatedType(GR, I, TypeToAssign);
- buildAssignType(B, TypeToAssign, I);
+ GR->buildAssignType(B, TypeToAssign, I);
}
for (const auto &Op : I->operands()) {
if (isa<ConstantPointerNull>(Op) || isa<UndefValue>(Op) ||
@@ -2025,10 +1939,11 @@ void SPIRVEmitIntrinsics::insertAssignTypeIntrs(Instruction *I,
Type *OpTy = Op->getType();
Type *OpTyElem = getPointeeType(OpTy);
if (OpTyElem) {
- buildAssignPtr(B, OpTyElem, Op);
+ GR->buildAssignPtr(B, OpTyElem, Op);
} else if (isPointerTy(OpTy)) {
Type *ElemTy = GR->findDeducedElementType(Op);
- buildAssignPtr(B, ElemTy ? ElemTy : deduceElementType(Op, true), Op);
+ GR->buildAssignPtr(B, ElemTy ? ElemTy : deduceElementType(Op, true),
+ Op);
} else {
CallInst *AssignCI =
buildIntrWithMD(Intrinsic::spv_assign_type, {OpTy},
@@ -2089,14 +2004,14 @@ void SPIRVEmitIntrinsics::processInstrAfterVisit(Instruction *I,
if (!IsConstComposite && isPointerTy(OpTy) &&
(OpElemTy = GR->findDeducedElementType(Op)) != nullptr &&
OpElemTy != IntegerType::getInt8Ty(I->getContext())) {
- buildAssignPtr(B, IntegerType::getInt8Ty(I->getContext()), NewOp);
+ GR->buildAssignPtr(B, IntegerType::getInt8Ty(I->getContext()), NewOp);
SmallVector<Type *, 2> Types = {OpTy, OpTy};
SmallVector<Value *, 2> Args = {
NewOp, buildMD(getNormalizedPoisonValue(OpElemTy)),
B.getInt32(getPointerAddressSpace(OpTy))};
CallInst *PtrCasted =
B.CreateIntrinsic(Intrinsic::spv_ptrcast, {Types}, Args);
- buildAssignPtr(B, OpElemTy, PtrCasted);
+ GR->buildAssignPtr(B, OpElemTy, PtrCasted);
NewOp = PtrCasted;
}
I->setOperand(OpNo, NewOp);
@@ -2178,7 +2093,7 @@ void SPIRVEmitIntrinsics::processParamTypesByFunHeader(Function *F,
continue;
if (hasPointeeTypeAttr(Arg) &&
(ElemTy = getPointeeTypeByAttr(Arg)) != nullptr) {
- buildAssignPtr(B, ElemTy, Arg);
+ GR->buildAssignPtr(B, ElemTy, Arg);
continue;
}
// search in function's call sites
@@ -2194,7 +2109,7 @@ void SPIRVEmitIntrinsics::processParamTypesByFunHeader(Function *F,
break;
}
if (ElemTy) {
- buildAssignPtr(B, ElemTy, Arg);
+ GR->buildAssignPtr(B, ElemTy, Arg);
continue;
}
if (HaveFunPtrs) {
@@ -2206,7 +2121,7 @@ void SPIRVEmitIntrinsics::processParamTypesByFunHeader(Function *F,
SmallVector<std::pair<Value *, unsigned>> Ops;
deduceOperandElementTypeFunctionPointer(CI, Ops, ElemTy, false);
if (ElemTy) {
- buildAssignPtr(B, ElemTy, Arg);
+ GR->buildAssignPtr(B, ElemTy, Arg);
break;
}
}
@@ -2225,11 +2140,11 @@ void SPIRVEmitIntrinsics::processParamTypes(Function *F, IRBuilder<> &B) {
if (!ElemTy && (ElemTy = deduceFunParamElementType(F, OpIdx)) != nullptr) {
if (CallInst *AssignCI = GR->findAssignPtrTypeInstr(Arg)) {
DenseSet<std::pair<Value *, Value *>> VisitedSubst;
- updateAssignType(AssignCI, Arg, getNormalizedPoisonValue(ElemTy));
+ GR->updateAssignType(AssignCI, Arg, getNormalizedPoisonValue(ElemTy));
propagateElemType(Arg, IntegerType::getInt8Ty(F->getContext()),
VisitedSubst);
} else {
- buildAssignPtr(B, ElemTy, Arg);
+ GR->buildAssignPtr(B, ElemTy, Arg);
}
}
}
@@ -2279,7 +2194,7 @@ bool SPIRVEmitIntrinsics::processFunctionPointers(Module &M) {
continue;
if (II->getIntrinsicID() == Intrinsic::spv_assign_ptr_type ||
II->getIntrinsicID() == Intrinsic::spv_ptrcast) {
- updateAssignType(II, &F, getNormalizedPoisonValue(FPElemTy));
+ GR->updateAssignType(II, &F, getNormalizedPoisonValue(FPElemTy));
break;
}
}
@@ -2331,7 +2246,7 @@ void SPIRVEmitIntrinsics::applyDemangledPtrArgTypes(IRBuilder<> &B) {
if (!hasPointeeTypeAttr(Arg)) {
B.SetInsertPointPastAllocas(Arg->getParent());
B.SetCurrentDebugLocation(DebugLoc());
- buildAssignPtr(B, ElemTy, Arg);
+ GR->buildAssignPtr(B, ElemTy, Arg);
}
} else if (isa<GetElementPtrInst>(Param)) {
replaceUsesOfWithSpvPtrcast(Param, normalizeType(ElemTy), CI,
@@ -2344,7 +2259,7 @@ void SPIRVEmitIntrinsics::applyDemangledPtrArgTypes(IRBuilder<> &B) {
->getParent()
->getEntryBlock()
.getFirstNonPHIOrDbgOrAlloca());
- buildAssignPtr(B, ElemTy, Param);
+ GR->buildAssignPtr(B, ElemTy, Param);
}
CallInst *Ref = dyn_cast<CallInst>(Param);
if (!Ref)
diff --git a/llvm/lib/Target/SPIRV/SPIRVGlobalRegistry.cpp b/llvm/lib/Target/SPIRV/SPIRVGlobalRegistry.cpp
index 0ed414ebc8bbe..e9b517d3eeaaa 100644
--- a/llvm/lib/Target/SPIRV/SPIRVGlobalRegistry.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVGlobalRegistry.cpp
@@ -22,6 +22,9 @@
#include "SPIRVUtils.h"
#include "llvm/ADT/APInt.h"
#include "llvm/IR/Constants.h"
+#include "llvm/IR/IntrinsicInst.h"
+#include "llvm/IR/Intrinsics.h"
+#include "llvm/IR/IntrinsicsSPIRV.h"
#include "llvm/IR/Type.h"
#include "llvm/Support/Casting.h"
#include <cassert>
@@ -29,6 +32,20 @@
using namespace llvm;
+namespace {
+
+bool allowEmitFakeUse(const Value *Arg) {
+ if (isSpvIntrinsic(Arg))
+ return false;
+ if (dyn_cast<AtomicCmpXchgInst>(Arg) || dyn_cast<InsertValueInst>(Arg) ||
+ dyn_cast<UndefValue>(Arg))
+ return false;
+ if (const auto *LI = dyn_cast<LoadInst>(Arg))
+ if (LI->getType()->isAggregateType())
+ return false;
+ return true;
+}
+
inline unsigned typeToAddressSpace(const Type *Ty) {
if (auto PType = dyn_cast<TypedPointerType>(Ty))
return PType->getAddressSpace();
@@ -40,6 +57,8 @@ inline unsigned typeToAddressSpace(const Type *Ty) {
report_fatal_error("Unable to convert LLVM type to SPIRVType", true);
}
+} // anonymous namespace
+
SPIRVGlobalRegistry::SPIRVGlobalRegistry(unsigned PointerSize)
: PointerSize(PointerSize), Bound(0) {}
@@ -1739,3 +1758,62 @@ LLT SPIRVGlobalRegistry::getRegType(SPIRVType *SpvType) const {
}
return LLT::scalar(64);
}
+
+void SPIRVGlobalRegistry::replaceAllUsesWith(Value *Old, Value *New,
+ bool DeleteOld) {
+ Old->replaceAllUsesWith(New);
+ updateIfExistDeducedElementType(Old, New, DeleteOld);
+ updateIfExistAssignPtrTypeInstr(Old, New, DeleteOld);
+}
+
+void SPIRVGlobalRegistry::buildAssignType(IRBuilder<> &B, Type *Ty,
+ Value *Arg) {
+ Value *OfType = getNormalizedPoisonValue(Ty);
+ CallInst *AssignCI = nullptr;
+ if (Arg->getType()->isAggregateType() && Ty->isAggregateType() &&
+ allowEmitFakeUse(Arg)) {
+ LLVMContext &Ctx = Arg->getContext();
+ SmallVector<Metadata *, 2> ArgMDs{
+ MDNode::get(Ctx, ValueAsMetadata::getConstant(OfType)),
+ MDString::get(Ctx, Arg->getName())};
+ B.CreateIntrinsic(Intrinsic::spv_value_md, {},
+ {MetadataAsValue::get(Ctx, MDTuple::get(Ctx, ArgMDs))});
+ AssignCI = B.CreateIntrinsic(Intrinsic::fake_use, {}, {Arg});
+ } else {
+ AssignCI = buildIntrWithMD(Intrinsic::spv_assign_type, {Arg->getType()},
+ OfType, Arg, {}, B);
+ }
+ addAssignPtrTypeInstr(Arg, AssignCI);
+}
+
+void SPIRVGlobalRegistry::buildAssignPtr(IRBuilder<> &B, Type *ElemTy,
+ Value *Arg) {
+ Value *OfType = PoisonValue::get(ElemTy);
+ CallInst *AssignPtrTyCI = findAssignPtrTypeInstr(Arg);
+ Function *CurrF =
+ B.GetInsertBlock() ? B.GetInsertBlock()->getParent() : nullptr;
+ if (AssignPtrTyCI == nullptr ||
+ AssignPtrTyCI->getParent()->getParent() != CurrF) {
+ AssignPtrTyCI = buildIntrWithMD(
+ Intrinsic::spv_assign_ptr_type, {Arg->getType()}, OfType, Arg,
+ {B.getInt32(getPointerAddressSpace(Arg->getType()))}, B);
+ addDeducedElementType(AssignPtrTyCI, ElemTy);
+ addDeducedElementType(Arg, ElemTy);
+ addAssignPtrTypeInstr(Arg, AssignPtrTyCI);
+ } else {
+ updateAssignType(AssignPtrTyCI, Arg, OfType);
+ }
+}
+
+void SPIRVGlobalRegistry::updateAssignType(CallInst *AssignCI, Value *Arg,
+ Value *OfType) {
+ AssignCI->setArgOperand(1, buildMD(OfType));
+ if (cast<IntrinsicInst>(AssignCI)->getIntrinsicID() !=
+ Intrinsic::spv_assign_ptr_type)
+ return;
+
+ // update association with the pointee type
+ Type *ElemTy = OfType->getType();
+ addDeducedElementType(AssignCI, ElemTy);
+ addDeducedElementType(Arg, ElemTy);
+}
diff --git a/llvm/lib/Target/SPIRV/SPIRVGlobalRegistry.h b/llvm/lib/Target/SPIRV/SPIRVGlobalRegistry.h
index 2c24ba79ea8e6..778232e7013d4 100644
--- a/llvm/lib/Target/SPIRV/SPIRVGlobalRegistry.h
+++ b/llvm/lib/Target/SPIRV/SPIRVGlobalRegistry.h
@@ -620,6 +620,14 @@ class SPIRVGlobalRegistry {
const TargetRegisterClass *getRegClass(SPIRVType *SpvType) const;
LLT getRegType(SPIRVType *SpvType) const;
+
+ // Replace all uses of a |Old| with |New| updates the global registry type
+ // mappings.
+ void replaceAllUsesWith(Value *Old, Value *New, bool DeleteOld = true);
+
+ void buildAssignType(IRBuilder<> &B, Type *Ty, Value *Arg);
+ void buildAssignPtr(IRBuilder<> &B, Type *ElemTy, Value *Arg);
+ void updateAssignType(CallInst *AssignCI, Value *Arg, Value *OfType);
};
} // end namespace llvm
#endif // LLLVM_LIB_TARGET_SPIRV_SPIRVTYPEMANAGER_H
diff --git a/llvm/lib/Target/SPIRV/SPIRVLegalizePointerCast.cpp b/llvm/lib/Target/SPIRV/SPIRVLegalizePointerCast.cpp
new file mode 100644
index 0000000000000..ec419d25cd317
--- /dev/null
+++ b/llvm/lib/Target/SPIRV/SPIRVLegalizePointerCast.cpp
@@ -0,0 +1,224 @@
+//===-- SPIRVLegalizePointerCast.cpp ----------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// The LLVM IR has multiple legal patterns we cannot lower to Logical SPIR-V.
+// This pass modifies such loads to have an IR we can directly lower to valid
+// logical SPIR-V.
+// OpenCL can avoid this because they rely on ptrcast, which is not supported
+// by logical SPIR-V.
+//
+// This pass relies on the assign_ptr_type intrinsic to deduce the type of the
+// pointed values, must replace all occurences of `ptrcast`. This is why
+// unhandled cases are reported as unreachable: we MUST cover all cases.
+//
+// 1. Loading the first element of an array
+//
+// %array = [10 x i32]
+// %value = load i32, ptr %array
+//
+// LLVM can skip the GEP instruction, and only request loading the first 4
+// bytes. In logical SPIR-V, we need an OpAccessChain to access the first
+// element. This pass will add a getelementptr instruction before the load.
+//
+//
+// 2. Implicit downcast from load
+//
+// %1 = getelementptr <4 x i32>, ptr %vec4, i64 0
+// %2 = load <3 x i32>, ptr %1
+//
+// The pointer in the GEP instruction is only used for offset computations,
+// but it doesn't NEED to match the pointed type. OpAccessChain however
+// requires this. Also, LLVM loads define the bitwidth of the load, not the
+// pointer. In this example, we can guess %vec4 is a vec4 thanks to the GEP
+// instruction basetype, but we only want to load the first 3 elements, hence
+// do a partial load. In logical SPIR-V, this is not legal. What we must do
+// is load the full vector (basetype), extract 3 elements, and recombine them
+// to form a 3-element vector.
+//
+//===----------------------------------------------------------------------===//
+
+#include "SPIRV.h"
+#include "SPIRVSubtarget.h"
+#include "SPIRVTargetMachine.h"
+#include "SPIRVUtils.h"
+#include "llvm/CodeGen/IntrinsicLowering.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/IntrinsicInst.h"
+#include "llvm/IR/Intrinsics.h"
+#include "llvm/IR/IntrinsicsSPIRV.h"
+#include "llvm/Transforms/Utils/Cloning.h"
+#include "llvm/Transforms/Utils/LowerMemIntrinsics.h"
+
+using namespace llvm;
+
+namespace llvm {
+void initializeSPIRVLegalizePointerCastPass(PassRegistry &);
+}
+
+class SPIRVLegalizePointerCast : public FunctionPass {
+
+ // Builds the `spv_assign_type` assigning |Ty| to |Value| at the current
+ // builder position.
+ void buildAssignType(IRBuilder<> &B, Type *Ty, Value *Arg) {
+ Value *OfType = PoisonValue::get(Ty);
+ CallInst *AssignCI = buildIntrWithMD(Intrinsic::spv_assign_type,
+ {Arg->getType()}, OfType, Arg, {}, B);
+ GR->addAssignPtrTypeInstr(Arg, AssignCI);
+ }
+
+ // Loads parts of the vector of type |SourceType| from the pointer |Source|
+ // and create a new vector of type |TargetType|. |TargetType| must be a vector
+ // type, and element types of |TargetType| and |SourceType| must match.
+ // Returns the loaded value.
+ Value *loadVectorFromVector(IRBuilder<> &B, FixedVectorType *SourceType,
+ FixedVectorType *TargetType, Value *Source) {
+ // We expect the codegen to avoid doing implicit bitcast from a load.
+ assert(TargetType->getElementType() == SourceType->getElementType());
+ assert(TargetType->getNumElements() < SourceType->getNumElements());
+
+ LoadInst *NewLoad = B.CreateLoad(SourceType, Source);
+ buildAssignType(B, SourceType, NewLoad);
+
+ SmallVector<int> Mask(/* Size= */ TargetType->getNumElements(),
+ /* Value= */ 0);
+ Value *Output = B.CreateShuffleVector(NewLoad, NewLoad, Mask);
+ buildAssignType(B, TargetType, Output);
+ return Output;
+ }
+
+ // Loads the first value in an aggregate pointed by |Source| of containing
+ // elements of type |ElementType|. Load flags will be copied from |BadLoad|,
+ // which should be the load being legalized. Returns the loaded value.
+ Value *loadFirstValueFromAggregate(IRBuilder<> &B, Type *ElementType,
+ Value *Source, LoadInst *BadLoad) {
+ SmallVector<Type *, 2> Types = {BadLoad->getPointerOperandType(),
+ BadLoad->getPointerOperandType()};
+ SmallVector<Value *, 3> Args{/* isInBounds= */ B.getInt1(false), Source,
+ B.getInt32(0), B.getInt32(0)};
+ auto *GEP = B.CreateIntrinsic(Intrinsic::spv_gep, {Types}, {Args});
+ GR->buildAssignPtr(B, ElementType, GEP);
+
+ const auto *TLI = TM->getSubtargetImpl()->getTargetLowering();
+ MachineMemOperand::Flags Flags = TLI->getLoadMemOperandFlags(
+ *BadLoad, BadLoad->getFunction()->getDataLayout());
+ Instruction *LI = B.CreateIntrinsic(
+ Intrinsic::spv_load, {BadLoad->getOperand(0)->getType()},
+ {GEP, B.getInt16(Flags), B.getInt8(BadLoad->getAlign().value())});
+ buildAssignType(B, ElementType, LI);
+ return LI;
+ }
+
+ // Replaces the load instruction to get rid of the ptrcast used as source
+ // operand.
+ void transformLoad(IRBuilder<> &B, LoadInst *LI, Value *CastedOperand,
+ Value *OriginalOperand) {
+ Type *FromTy = GR->findDeducedElementType(OriginalOperand);
+ Type *ToTy = GR->findDeducedElementType(CastedOperand);
+ Value *Output = nullptr;
+
+ auto *SAT = dyn_cast<ArrayType>(FromTy);
+ auto *SVT = dyn_cast<FixedVectorType>(FromTy);
+ auto *DVT = dyn_cast<FixedVectorType>(ToTy);
+
+ B.SetInsertPoint(LI);
+
+ // Destination is the element type of Source, and source is an array ->
+ // Loading 1st element.
+ // - float a = array[0];
+ if (SAT && SAT->getElementType() == ToTy)
+ Output = loadFirstValueFromAggregate(B, SAT->getElementType(),
+ OriginalOperand, LI);
+ // Destination is the element type of Source, and source is a vector ->
+ // Vector to scalar.
+ // - float a = vector.x;
+ else if (!DVT && SVT && SVT->getElementType() == ToTy) {
+ Output = loadFirstValueFromAggregate(B, SVT->getElementType(),
+ OriginalOperand, LI);
+ }
+ // Destination is a smaller vector than source.
+ // - float3 v3 = vector4;
+ else if (SVT && DVT)
+ Output = loadVectorFromVector(B, SVT, DVT, OriginalOperand);
+ else
+ llvm_unreachable("Unimplemented implicit down-cast from load.");
+
+ GR->replaceAllUsesWith(LI, Output, /* DeleteOld= */ true);
+ DeadInstructions.push_back(LI);
+ }
+
+ void legalizePointerCast(IntrinsicInst *II) {
+ Value *CastedOperand = II;
+ Value *OriginalOperand = II->getOperand(0);
+
+ IRBuilder<> B(II->getContext());
+ std::vector<Value *> Users;
+ for (Use &U : II->uses())
+ Users.push_back(U.getUser());
+
+ for (Value *User : Users) {
+ if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
+ transformLoad(B, LI, CastedOperand, OriginalOperand);
+ continue;
+ }
+
+ IntrinsicInst *Intrin = dyn_cast<IntrinsicInst>(User);
+ if (Intrin->getIntrinsicID() == Intrinsic::spv_assign_ptr_type) {
+ DeadInstructions.push_back(Intrin);
+ continue;
+ }
+
+ llvm_unreachable("Unsupported ptrcast user. Please fix.");
+ }
+
+ DeadInstructions.push_back(II);
+ }
+
+public:
+ SPIRVLegalizePointerCast(SPIRVTargetMachine *TM) : FunctionPass(ID), TM(TM) {
+ initializeSPIRVLegalizePointerCastPass(*PassRegistry::getPassRegistry());
+ };
+
+ virtual bool runOnFunction(Function &F) override {
+ const SPIRVSubtarget &ST = TM->getSubtarget<SPIRVSubtarget>(F);
+ GR = ST.getSPIRVGlobalRegistry();
+ DeadInstructions.clear();
+
+ std::vector<IntrinsicInst *> WorkList;
+ for (auto &BB : F) {
+ for (auto &I : BB) {
+ auto *II = dyn_cast<IntrinsicInst>(&I);
+ if (II && II->getIntrinsicID() == Intrinsic::spv_ptrcast)
+ WorkList.push_back(II);
+ }
+ }
+
+ for (IntrinsicInst *II : WorkList)
+ legalizePointerCast(II);
+
+ for (Instruction *I : DeadInstructions)
+ I->eraseFromParent();
+
+ return DeadInstructions.size() != 0;
+ }
+
+private:
+ SPIRVTargetMachine *TM = nullptr;
+ SPIRVGlobalRegistry *GR = nullptr;
+ std::vector<Instruction *> DeadInstructions;
+
+public:
+ static char ID;
+};
+
+char SPIRVLegalizePointerCast::ID = 0;
+INITIALIZE_PASS(SPIRVLegalizePointerCast, "spirv-legalize-bitcast",
+ "SPIRV legalize bitcast pass", false, false)
+
+FunctionPass *llvm::createSPIRVLegalizePointerCastPass(SPIRVTargetMachine *TM) {
+ return new SPIRVLegalizePointerCast(TM);
+}
diff --git a/llvm/lib/Target/SPIRV/SPIRVTargetMachine.cpp b/llvm/lib/Target/SPIRV/SPIRVTargetMachine.cpp
index 098c7a6fba50e..0aa214dd354ee 100644
--- a/llvm/lib/Target/SPIRV/SPIRVTargetMachine.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVTargetMachine.cpp
@@ -209,6 +209,8 @@ void SPIRVPassConfig::addIRPasses() {
void SPIRVPassConfig::addISelPrepare() {
addPass(createSPIRVEmitIntrinsicsPass(&getTM<SPIRVTargetMachine>()));
+ if (TM.getSubtargetImpl()->isVulkanEnv())
+ addPass(createSPIRVLegalizePointerCastPass(&getTM<SPIRVTargetMachine>()));
TargetPassConfig::addISelPrepare();
}
diff --git a/llvm/lib/Target/SPIRV/SPIRVUtils.cpp b/llvm/lib/Target/SPIRV/SPIRVUtils.cpp
index 815646dd7d3d4..ce4f6d6c9288f 100644
--- a/llvm/lib/Target/SPIRV/SPIRVUtils.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVUtils.cpp
@@ -775,6 +775,17 @@ Register createVirtualRegister(
MIRBuilder);
}
+CallInst *buildIntrWithMD(Intrinsic::ID IntrID, ArrayRef<Type *> Types,
+ Value *Arg, Value *Arg2, ArrayRef<Constant *> Imms,
+ IRBuilder<> &B) {
+ SmallVector<Value *, 4> Args;
+ Args.push_back(Arg2);
+ Args.push_back(buildMD(Arg));
+ for (auto *Imm : Imms)
+ Args.push_back(Imm);
+ return B.CreateIntrinsic(IntrID, {Types}, Args);
+}
+
// Return true if there is an opaque pointer type nested in the argument.
bool isNestedPointer(const Type *Ty) {
if (Ty->isPtrOrPtrVectorTy())
diff --git a/llvm/lib/Target/SPIRV/SPIRVUtils.h b/llvm/lib/Target/SPIRV/SPIRVUtils.h
index 1346ea80a98fc..180d3a29d170f 100644
--- a/llvm/lib/Target/SPIRV/SPIRVUtils.h
+++ b/llvm/lib/Target/SPIRV/SPIRVUtils.h
@@ -405,6 +405,16 @@ inline PoisonValue *getNormalizedPoisonValue(Type *Ty) {
return PoisonValue::get(normalizeType(Ty));
}
+inline MetadataAsValue *buildMD(Value *Arg) {
+ LLVMContext &Ctx = Arg->getContext();
+ return MetadataAsValue::get(
+ Ctx, MDNode::get(Ctx, ValueAsMetadata::getConstant(Arg)));
+}
+
+CallInst *buildIntrWithMD(Intrinsic::ID IntrID, ArrayRef<Type *> Types,
+ Value *Arg, Value *Arg2, ArrayRef<Constant *> Imms,
+ IRBuilder<> &B);
+
MachineInstr *getVRegDef(MachineRegisterInfo &MRI, Register Reg);
#define SPIRV_BACKEND_SERVICE_FUN_NAME "__spirv_backend_service_fun"
diff --git a/llvm/test/CodeGen/SPIRV/pointers/array-skips-gep.ll b/llvm/test/CodeGen/SPIRV/pointers/array-skips-gep.ll
new file mode 100644
index 0000000000000..85663d6c3eeda
--- /dev/null
+++ b/llvm/test/CodeGen/SPIRV/pointers/array-skips-gep.ll
@@ -0,0 +1,37 @@
+; RUN: llc -verify-machineinstrs -O0 -mtriple=spirv-unknown-vulkan-compute %s -o - | FileCheck %s
+; RUN: %if spirv-tools %{ llc -O0 -mtriple=spirv-unknown-vulkan %s -o - -filetype=obj | spirv-val %}
+
+; CHECK-DAG: %[[#uint_ty:]] = OpTypeInt 32 0
+; CHECK-DAG: %[[#uint_0:]] = OpConstant %[[#uint_ty]] 0
+; CHECK-DAG: %[[#int_10:]] = OpConstant %[[#uint_ty]] 10
+; CHECK-DAG: %[[#array_ty:]] = OpTypeArray %[[#uint_ty]] %[[#int_10]]
+; CHECK-DAG: %[[#array_fp:]] = OpTypePointer Function %[[#array_ty]]
+; CHECK-DAG: %[[#array_pp:]] = OpTypePointer Private %[[#array_ty]]
+; CHECK-DAG: %[[#int_fp:]] = OpTypePointer Function %[[#uint_ty]]
+; CHECK-DAG: %[[#int_pp:]] = OpTypePointer Private %[[#uint_ty]]
+
+ at gv = external addrspace(10) global [10 x i32]
+; CHECK: %[[#gv:]] = OpVariable %[[#array_pp]] Private
+
+define internal spir_func i32 @foo() {
+ %array = alloca [10 x i32], align 4
+; CHECK: %[[#array:]] = OpVariable %[[#array_fp:]] Function
+
+ ; Direct load from the pointer index. This requires an OpAccessChain
+ %1 = load i32, ptr %array, align 4
+; CHECK: %[[#ptr:]] = OpAccessChain %[[#int_fp]] %[[#array]] %[[#uint_0]]
+; CHECK: %[[#val:]] = OpLoad %[[#uint_ty]] %[[#ptr]] Aligned 4
+
+ ret i32 %1
+; CHECK: OpReturnValue %[[#val]]
+}
+
+define internal spir_func i32 @bar() {
+ ; Direct load from the pointer index. This requires an OpAccessChain
+ %1 = load i32, ptr addrspace(10) @gv
+; CHECK: %[[#ptr:]] = OpAccessChain %[[#int_pp]] %[[#gv]] %[[#uint_0]]
+; CHECK: %[[#val:]] = OpLoad %[[#uint_ty]] %[[#ptr]] Aligned 4
+
+ ret i32 %1
+; CHECK: OpReturnValue %[[#val]]
+}
diff --git a/llvm/test/CodeGen/SPIRV/pointers/getelementptr-downcast-vector.ll b/llvm/test/CodeGen/SPIRV/pointers/getelementptr-downcast-vector.ll
new file mode 100644
index 0000000000000..d4131fa8a2658
--- /dev/null
+++ b/llvm/test/CodeGen/SPIRV/pointers/getelementptr-downcast-vector.ll
@@ -0,0 +1,110 @@
+; RUN: llc -verify-machineinstrs -O0 -mtriple=spirv-unknown-vulkan-compute %s -o - | FileCheck %s
+; RUN: %if spirv-tools %{ llc -O0 -mtriple=spirv-unknown-vulkan %s -o - -filetype=obj | spirv-val %}
+
+; CHECK-DAG: %[[#uint:]] = OpTypeInt 32 0
+; CHECK-DAG: %[[#uint_pp:]] = OpTypePointer Private %[[#uint]]
+; CHECK-DAG: %[[#uint_fp:]] = OpTypePointer Function %[[#uint]]
+; CHECK-DAG: %[[#uint_0:]] = OpConstant %[[#uint]] 0
+; CHECK-DAG: %[[#v2:]] = OpTypeVector %[[#uint]] 2
+; CHECK-DAG: %[[#v3:]] = OpTypeVector %[[#uint]] 3
+; CHECK-DAG: %[[#v4:]] = OpTypeVector %[[#uint]] 4
+; CHECK-DAG: %[[#v4_pp:]] = OpTypePointer Private %[[#v4]]
+; CHECK-DAG: %[[#v4_fp:]] = OpTypePointer Function %[[#v4]]
+
+define internal spir_func <3 x i32> @foo(ptr addrspace(10) %a) {
+
+ %1 = getelementptr inbounds <4 x i32>, ptr addrspace(10) %a, i64 0
+; CHECK: %[[#tmp:]] = OpInBoundsAccessChain %[[#v4_pp]] %[[#]]
+
+ ; partial loading of a vector: v4 -> v3.
+ %2 = load <3 x i32>, ptr addrspace(10) %1, align 16
+; CHECK: %[[#load:]] = OpLoad %[[#v4]] %[[#tmp]] Aligned 16
+; CHECK: %[[#val:]] = OpVectorShuffle %[[#v3]] %[[#load]] %[[#load]] 0 0 0
+
+ ret <3 x i32> %2
+; CHECK: OpReturnValue %[[#val]]
+}
+
+define internal spir_func <3 x i32> @fooDefault(ptr %a) {
+
+ %1 = getelementptr inbounds <4 x i32>, ptr %a, i64 0
+; CHECK: %[[#tmp:]] = OpInBoundsAccessChain %[[#v4_fp]] %[[#]]
+
+ ; partial loading of a vector: v4 -> v3.
+ %2 = load <3 x i32>, ptr %1, align 16
+; CHECK: %[[#load:]] = OpLoad %[[#v4]] %[[#tmp]] Aligned 16
+; CHECK: %[[#val:]] = OpVectorShuffle %[[#v3]] %[[#load]] %[[#load]] 0 0 0
+
+ ret <3 x i32> %2
+; CHECK: OpReturnValue %[[#val]]
+}
+
+define internal spir_func <3 x i32> @fooBounds(ptr %a) {
+
+ %1 = getelementptr <4 x i32>, ptr %a, i64 0
+; CHECK: %[[#tmp:]] = OpAccessChain %[[#v4_fp]] %[[#]]
+
+ ; partial loading of a vector: v4 -> v3.
+ %2 = load <3 x i32>, ptr %1, align 16
+; CHECK: %[[#load:]] = OpLoad %[[#v4]] %[[#tmp]] Aligned 16
+; CHECK: %[[#val:]] = OpVectorShuffle %[[#v3]] %[[#load]] %[[#load]] 0 0 0
+
+ ret <3 x i32> %2
+; CHECK: OpReturnValue %[[#val]]
+}
+
+define internal spir_func <2 x i32> @bar(ptr addrspace(10) %a) {
+
+ %1 = getelementptr inbounds <4 x i32>, ptr addrspace(10) %a, i64 0
+; CHECK: %[[#tmp:]] = OpInBoundsAccessChain %[[#v4_pp]] %[[#]]
+
+ ; partial loading of a vector: v4 -> v2.
+ %2 = load <2 x i32>, ptr addrspace(10) %1, align 16
+; CHECK: %[[#load:]] = OpLoad %[[#v4]] %[[#tmp]] Aligned 16
+; CHECK: %[[#val:]] = OpVectorShuffle %[[#v2]] %[[#load]] %[[#load]] 0 0
+
+ ret <2 x i32> %2
+; CHECK: OpReturnValue %[[#val]]
+}
+
+define internal spir_func i32 @baz(ptr addrspace(10) %a) {
+
+ %1 = getelementptr inbounds <4 x i32>, ptr addrspace(10) %a, i64 0
+; CHECK: %[[#tmp:]] = OpInBoundsAccessChain %[[#v4_pp]] %[[#]]
+
+ ; Loading of the first scalar of a vector: v4 -> int.
+ %2 = load i32, ptr addrspace(10) %1, align 16
+; CHECK: %[[#ptr:]] = OpAccessChain %[[#uint_pp]] %[[#tmp]] %[[#uint_0]]
+; CHECK: %[[#val:]] = OpLoad %[[#uint]] %[[#ptr]] Aligned 16
+
+ ret i32 %2
+; CHECK: OpReturnValue %[[#val]]
+}
+
+define internal spir_func i32 @bazDefault(ptr %a) {
+
+ %1 = getelementptr inbounds <4 x i32>, ptr %a, i64 0
+; CHECK: %[[#tmp:]] = OpInBoundsAccessChain %[[#v4_fp]] %[[#]]
+
+ ; Loading of the first scalar of a vector: v4 -> int.
+ %2 = load i32, ptr %1, align 16
+; CHECK: %[[#ptr:]] = OpAccessChain %[[#uint_fp]] %[[#tmp]] %[[#uint_0]]
+; CHECK: %[[#val:]] = OpLoad %[[#uint]] %[[#ptr]] Aligned 16
+
+ ret i32 %2
+; CHECK: OpReturnValue %[[#val]]
+}
+
+define internal spir_func i32 @bazBounds(ptr %a) {
+
+ %1 = getelementptr <4 x i32>, ptr %a, i64 0
+; CHECK: %[[#tmp:]] = OpAccessChain %[[#v4_fp]] %[[#]]
+
+ ; Loading of the first scalar of a vector: v4 -> int.
+ %2 = load i32, ptr %1, align 16
+; CHECK: %[[#ptr:]] = OpAccessChain %[[#uint_fp]] %[[#tmp]] %[[#uint_0]]
+; CHECK: %[[#val:]] = OpLoad %[[#uint]] %[[#ptr]] Aligned 16
+
+ ret i32 %2
+; CHECK: OpReturnValue %[[#val]]
+}
More information about the llvm-commits
mailing list