diff options
Diffstat (limited to 'lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp')
| -rw-r--r-- | lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp | 16 |
1 files changed, 9 insertions, 7 deletions
diff --git a/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp b/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp index c147830e12ed..743dc7a0d00b 100644 --- a/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp +++ b/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp @@ -16,7 +16,6 @@ #include "AMDGPUSubtarget.h" #include "AMDGPUTargetMachine.h" #include "llvm/ADT/StringRef.h" -#include "llvm/Analysis/DivergenceAnalysis.h" #include "llvm/Analysis/Loads.h" #include "llvm/CodeGen/Passes.h" #include "llvm/CodeGen/TargetPassConfig.h" @@ -84,8 +83,8 @@ bool AMDGPULowerKernelArguments::runOnFunction(Function &F) { return false; CallInst *KernArgSegment = - Builder.CreateIntrinsic(Intrinsic::amdgcn_kernarg_segment_ptr, nullptr, - F.getName() + ".kernarg.segment"); + Builder.CreateIntrinsic(Intrinsic::amdgcn_kernarg_segment_ptr, {}, {}, + nullptr, F.getName() + ".kernarg.segment"); KernArgSegment->addAttribute(AttributeList::ReturnIndex, Attribute::NonNull); KernArgSegment->addAttribute(AttributeList::ReturnIndex, @@ -123,14 +122,17 @@ bool AMDGPULowerKernelArguments::runOnFunction(Function &F) { VectorType *VT = dyn_cast<VectorType>(ArgTy); bool IsV3 = VT && VT->getNumElements() == 3; + bool DoShiftOpt = Size < 32 && !ArgTy->isAggregateType(); + VectorType *V4Ty = nullptr; int64_t AlignDownOffset = alignDown(EltOffset, 4); int64_t OffsetDiff = EltOffset - AlignDownOffset; - unsigned AdjustedAlign = MinAlign(KernArgBaseAlign, AlignDownOffset); + unsigned AdjustedAlign = MinAlign(DoShiftOpt ? AlignDownOffset : EltOffset, + KernArgBaseAlign); Value *ArgPtr; - if (Size < 32 && !ArgTy->isAggregateType()) { // FIXME: Handle aggregate types + if (DoShiftOpt) { // FIXME: Handle aggregate types // Since we don't have sub-dword scalar loads, avoid doing an extload by // loading earlier than the argument address, and extracting the relevant // bits. @@ -148,7 +150,7 @@ bool AMDGPULowerKernelArguments::runOnFunction(Function &F) { } else { ArgPtr = Builder.CreateConstInBoundsGEP1_64( KernArgSegment, - AlignDownOffset, + EltOffset, Arg.getName() + ".kernarg.offset"); ArgPtr = Builder.CreateBitCast(ArgPtr, ArgTy->getPointerTo(AS), ArgPtr->getName() + ".cast"); @@ -199,7 +201,7 @@ bool AMDGPULowerKernelArguments::runOnFunction(Function &F) { // TODO: Convert noalias arg to !noalias - if (Size < 32 && !ArgTy->isAggregateType()) { + if (DoShiftOpt) { Value *ExtractBits = OffsetDiff == 0 ? Load : Builder.CreateLShr(Load, OffsetDiff * 8); |