#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"

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 APInt	trimTrailingZerosInVector (InstCombiner &IC, Value UseV, Instruction I)

static Value *	simplifyAMDGCNMemoryIntrinsicDemanded (InstCombiner &IC, IntrinsicInst &II, APInt DemandedElts, int DMaskIdx, bool IsLoad)
	Implement SimplifyDemandedVectorElts for amdgcn buffer and image intrinsics.

static bool	canContractSqrtToRsq (const FPMathOperator *SqrtOp)
	Return true if it's legal to contract llvm.amdgcn.rcp(llvm.sqrt)

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

◆ canContractSqrtToRsq()

static bool canContractSqrtToRsq ( const FPMathOperator * SqrtOp )

static

Return true if it's legal to contract llvm.amdgcn.rcp(llvm.sqrt)

Definition at line 412 of file AMDGPUInstCombineIntrinsic.cpp.

References llvm::FPMathOperator::getFPAccuracy(), llvm::Value::getType(), llvm::FPMathOperator::hasApproxFunc(), llvm::Type::isFloatTy(), and llvm::Type::isHalfTy().

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

◆ 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 llvm::APFloat::convert(), llvm::ConstantFP::get(), llvm::Value::getContext(), 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.

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

◆ simplifyAMDGCNMemoryIntrinsicDemanded()

static Value * simplifyAMDGCNMemoryIntrinsicDemanded	(	InstCombiner &	IC,
		IntrinsicInst &	II,
		APInt	DemandedElts,
		int	DMaskIdx,
		bool	IsLoad
	)

static

Implement SimplifyDemandedVectorElts for amdgcn buffer and image intrinsics.

The result of simplifying amdgcn image and buffer store intrinsics is updating definitions of the intrinsics vector argument, not Uses of the result like image and buffer loads. Note: This only supports non-TFE/LWE image intrinsic calls; those have struct returns.

Definition at line 1156 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::CreateExtractElement(), 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::User::getOperand(), 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().