AMDGPUInstCombineIntrinsic.cpp File Reference

#include "AMDGPUInstrInfo.h"
#include "AMDGPUTargetTransformInfo.h"
#include "GCNSubtarget.h"
#include "llvm/ADT/FloatingPointMode.h"
#include "llvm/IR/IntrinsicsAMDGPU.h"
#include "llvm/Transforms/InstCombine/InstCombiner.h"
#include <optional>
#include "InstCombineTables.inc"

Go to the source code of this file.

Macros
#define	DEBUG_TYPE   "AMDGPUtti"
#define	GET_AMDGPUImageDMaskIntrinsicTable_IMPL

Functions
static APFloat	fmed3AMDGCN (const APFloat &Src0, const APFloat &Src1, const APFloat &Src2)
static bool	canSafelyConvertTo16Bit (Value &V, bool IsFloat)
static Value *	convertTo16Bit (Value &V, InstCombiner::BuilderTy &Builder)
static std::optional< Instruction * >	modifyIntrinsicCall (IntrinsicInst &OldIntr, Instruction &InstToReplace, unsigned NewIntr, InstCombiner &IC, std::function< void(SmallVectorImpl< Value * > &, SmallVectorImpl< Type * > &)> Func)
	Applies Func(OldIntr.Args, OldIntr.ArgTys), creates intrinsic call with modified arguments (based on OldIntr) and replaces InstToReplace with this newly created intrinsic call.
static std::optional< Instruction * >	simplifyAMDGCNImageIntrinsic (const GCNSubtarget *ST, const AMDGPU::ImageDimIntrinsicInfo *ImageDimIntr, IntrinsicInst &II, InstCombiner &IC)
static bool	matchFPExtFromF16 (Value *Arg, Value *&FPExtSrc)
	Match an fpext from half to float, or a constant we can convert.
static Value *	simplifyAMDGCNMemoryIntrinsicDemanded (InstCombiner &IC, IntrinsicInst &II, APInt DemandedElts, int DMaskIdx=-1)
	Implement SimplifyDemandedVectorElts for amdgcn buffer and image intrinsics.

Macro Definition Documentation

DEBUG_TYPE

#define DEBUG_TYPE   "AMDGPUtti"

Definition at line 28 of file AMDGPUInstCombineIntrinsic.cpp.

GET_AMDGPUImageDMaskIntrinsicTable_IMPL

#define GET_AMDGPUImageDMaskIntrinsicTable_IMPL

Definition at line 36 of file AMDGPUInstCombineIntrinsic.cpp.

Function Documentation

canSafelyConvertTo16Bit()

static bool canSafelyConvertTo16Bit ( Value &  V, bool  IsFloat  )

static

Definition at line 66 of file AMDGPUInstCombineIntrinsic.cpp.

References llvm::APFloat::convert(), llvm::APInt::getActiveBits(), llvm::Value::getType(), llvm::Type::isHalfTy(), llvm::Type::isIntegerTy(), llvm::PatternMatch::m_FPExt(), llvm::PatternMatch::m_Value(), llvm::PatternMatch::m_ZExt(), and llvm::PatternMatch::match().

Referenced by simplifyAMDGCNImageIntrinsic().

convertTo16Bit()

static Value * convertTo16Bit ( Value &  V, InstCombiner::BuilderTy &  Builder  )

static

Definition at line 104 of file AMDGPUInstCombineIntrinsic.cpp.

References Builder, llvm::Type::getHalfTy(), llvm::Type::getInt16Ty(), llvm::Type::isFloatingPointTy(), llvm::Type::isIntegerTy(), and llvm_unreachable.

fmed3AMDGCN()

static APFloat fmed3AMDGCN ( const APFloat &  Src0, const APFloat &  Src1, const APFloat &  Src2  )

static

Definition at line 45 of file AMDGPUInstCombineIntrinsic.cpp.

References assert(), llvm::APFloat::compare(), and llvm::maxnum().

Referenced by llvm::GCNTTIImpl::instCombineIntrinsic().

matchFPExtFromF16()

static bool matchFPExtFromF16 ( Value *  Arg, Value *&  FPExtSrc  )

static

Match an fpext from half to float, or a constant we can convert.

Definition at line 360 of file AMDGPUInstCombineIntrinsic.cpp.

References Arg, llvm::APFloat::convert(), llvm::ConstantFP::get(), llvm::Type::getHalfTy(), llvm::Value::getType(), llvm::ConstantFP::getValueAPF(), llvm::APFloatBase::IEEEhalf(), llvm::Type::isHalfTy(), llvm::PatternMatch::m_ConstantFP(), llvm::PatternMatch::m_FPExt(), llvm::PatternMatch::m_OneUse(), llvm::PatternMatch::m_Value(), llvm::PatternMatch::match(), and llvm::APFloatBase::rmNearestTiesToEven.

Referenced by llvm::GCNTTIImpl::instCombineIntrinsic().

modifyIntrinsicCall()

static std::optional< Instruction * > modifyIntrinsicCall ( IntrinsicInst &  OldIntr, Instruction &  InstToReplace, unsigned  NewIntr, InstCombiner &  IC, std::function< void(SmallVectorImpl< Value * > &, SmallVectorImpl< Type * > &)>  Func  )

static

Applies Func(OldIntr.Args, OldIntr.ArgTys), creates intrinsic call with modified arguments (based on OldIntr) and replaces InstToReplace with this newly created intrinsic call.

Definition at line 119 of file AMDGPUInstCombineIntrinsic.cpp.

References llvm::CallBase::args(), llvm::InstCombiner::Builder, llvm::Instruction::copyFastMathFlags(), llvm::Instruction::copyMetadata(), llvm::IRBuilderBase::CreateCall(), llvm::InstCombiner::eraseInstFromFunction(), llvm::CallBase::getCalledFunction(), llvm::Intrinsic::getDeclaration(), llvm::Intrinsic::getIntrinsicSignature(), llvm::Instruction::getModule(), llvm::Value::getType(), I, llvm::Type::isVoidTy(), llvm::InstCombiner::replaceInstUsesWith(), and llvm::Value::takeName().

Referenced by simplifyAMDGCNImageIntrinsic().

simplifyAMDGCNImageIntrinsic()

static std::optional< Instruction * > simplifyAMDGCNImageIntrinsic ( const GCNSubtarget *  ST, const AMDGPU::ImageDimIntrinsicInfo *  ImageDimIntr, IntrinsicInst &  II, InstCombiner &  IC  )

static

Definition at line 155 of file AMDGPUInstCombineIntrinsic.cpp.

References assert(), llvm::AMDGPU::ImageDimIntrinsicInfo::BaseOpcode, llvm::AMDGPU::ImageDimIntrinsicInfo::BiasIndex, canSafelyConvertTo16Bit(), llvm::AMDGPU::ImageDimIntrinsicInfo::CoordStart, llvm::AMDGPU::ImageDimIntrinsicInfo::Dim, for(), llvm::Value::getContext(), llvm::Type::getHalfTy(), llvm::AMDGPU::getImageDimIntrinsicByBaseOpcode(), llvm::Type::getInt16Ty(), llvm::IntrinsicInst::getIntrinsicID(), llvm::AMDGPU::getMIMGBaseOpcodeInfo(), llvm::AMDGPU::getMIMGBiasMappingInfo(), llvm::AMDGPU::getMIMGLZMappingInfo(), llvm::AMDGPU::getMIMGMIPMappingInfo(), llvm::AMDGPU::getMIMGOffsetMappingInfo(), llvm::User::getOperand(), llvm::Type::getScalarType(), llvm::Value::getType(), llvm::AMDGPU::ImageDimIntrinsicInfo::GradientStart, llvm::AMDGPU::MIMGBaseOpcodeInfo::HasD16, llvm::Value::hasOneUse(), if(), llvm::AMDGPU::ImageDimIntrinsicInfo::Intr, llvm::Type::isFloatingPointTy(), llvm::Type::isHalfTy(), llvm::AMDGPU::ImageDimIntrinsicInfo::LodIndex, llvm::AMDGPU::ImageDimIntrinsicInfo::MipIndex, modifyIntrinsicCall(), llvm::AMDGPU::ImageDimIntrinsicInfo::NumBiasArgs, llvm::AMDGPU::ImageDimIntrinsicInfo::OffsetIndex, llvm::AMDGPU::MIMGBaseOpcodeInfo::Sampler, llvm::Instruction::user_back(), and llvm::AMDGPU::ImageDimIntrinsicInfo::VAddrEnd.

Referenced by llvm::GCNTTIImpl::instCombineIntrinsic().

simplifyAMDGCNMemoryIntrinsicDemanded()

static Value * simplifyAMDGCNMemoryIntrinsicDemanded ( InstCombiner &  IC, IntrinsicInst &  II, APInt  DemandedElts, int  DMaskIdx = -1  )

static

Implement SimplifyDemandedVectorElts for amdgcn buffer and image intrinsics.

Note: This only supports non-TFE/LWE image intrinsic calls; those have struct returns.

Definition at line 1052 of file AMDGPUInstCombineIntrinsic.cpp.

References llvm::CallBase::args(), llvm::InstCombiner::Builder, llvm::Instruction::copyMetadata(), llvm::APInt::countr_zero(), llvm::IRBuilderBase::CreateAdd(), llvm::IRBuilderBase::CreateCall(), llvm::IRBuilderBase::CreateInsertElement(), llvm::IRBuilderBase::CreateShuffleVector(), llvm::FixedVectorType::get(), llvm::UndefValue::get(), llvm::ConstantInt::get(), llvm::APInt::getActiveBits(), llvm::CallBase::getCalledFunction(), llvm::InstCombiner::getDataLayout(), llvm::Intrinsic::getDeclaration(), llvm::IntrinsicInst::getIntrinsicID(), llvm::Intrinsic::getIntrinsicSignature(), llvm::Instruction::getModule(), llvm::ConstantInt::getType(), llvm::Value::getType(), llvm::DataLayout::getTypeSizeInBits(), llvm::ConstantInt::getZExtValue(), llvm::APInt::isMask(), llvm::Offset, llvm::APInt::popcount(), llvm::popcount(), llvm::SmallVectorTemplateBase< T, bool >::push_back(), llvm::CallBase::setArgOperand(), llvm::IRBuilderBase::SetInsertPoint(), and llvm::Value::takeName().

Referenced by llvm::GCNTTIImpl::simplifyDemandedVectorEltsIntrinsic().