1 //===-- AMDGPUTargetTransformInfo.h - AMDGPU specific TTI -------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This file a TargetTransformInfo::Concept conforming object specific to the
11 /// AMDGPU target machine. It uses the target's detailed information to
12 /// provide more precise answers to certain TTI queries, while letting the
13 /// target independent and default TTI implementations handle the rest.
15 //===----------------------------------------------------------------------===//
17 #ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUTARGETTRANSFORMINFO_H
18 #define LLVM_LIB_TARGET_AMDGPU_AMDGPUTARGETTRANSFORMINFO_H
21 #include "AMDGPUTargetMachine.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/CodeGen/BasicTTIImpl.h"
26 class AMDGPUTargetLowering;
28 class AMDGPUTTIImpl final : public BasicTTIImplBase<AMDGPUTTIImpl> {
29 typedef BasicTTIImplBase<AMDGPUTTIImpl> BaseT;
30 typedef TargetTransformInfo TTI;
33 const AMDGPUSubtarget *ST;
34 const AMDGPUTargetLowering *TLI;
35 bool IsGraphicsShader;
37 const AMDGPUSubtarget *getST() const { return ST; }
38 const AMDGPUTargetLowering *getTLI() const { return TLI; }
41 static inline int getFullRateInstrCost() {
42 return TargetTransformInfo::TCC_Basic;
45 static inline int getHalfRateInstrCost() {
46 return 2 * TargetTransformInfo::TCC_Basic;
49 // TODO: The size is usually 8 bytes, but takes 4x as many cycles. Maybe
51 static inline int getQuarterRateInstrCost() {
52 return 3 * TargetTransformInfo::TCC_Basic;
55 // On some parts, normal fp64 operations are half rate, and others
56 // quarter. This also applies to some integer operations.
57 inline int get64BitInstrCost() const {
58 return ST->hasHalfRate64Ops() ?
59 getHalfRateInstrCost() : getQuarterRateInstrCost();
63 explicit AMDGPUTTIImpl(const AMDGPUTargetMachine *TM, const Function &F)
64 : BaseT(TM, F.getParent()->getDataLayout()),
65 ST(TM->getSubtargetImpl(F)),
66 TLI(ST->getTargetLowering()),
67 IsGraphicsShader(AMDGPU::isShader(F.getCallingConv())) {}
69 bool hasBranchDivergence() { return true; }
71 void getUnrollingPreferences(Loop *L, TTI::UnrollingPreferences &UP);
73 TTI::PopcntSupportKind getPopcntSupport(unsigned TyWidth) {
74 assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
75 return TTI::PSK_FastHardware;
78 unsigned getNumberOfRegisters(bool Vector);
79 unsigned getRegisterBitWidth(bool Vector);
80 unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const;
82 bool isLegalToVectorizeMemChain(unsigned ChainSizeInBytes,
84 unsigned AddrSpace) const;
85 bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
87 unsigned AddrSpace) const;
88 bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
90 unsigned AddrSpace) const;
92 unsigned getMaxInterleaveFactor(unsigned VF);
94 int getArithmeticInstrCost(
95 unsigned Opcode, Type *Ty,
96 TTI::OperandValueKind Opd1Info = TTI::OK_AnyValue,
97 TTI::OperandValueKind Opd2Info = TTI::OK_AnyValue,
98 TTI::OperandValueProperties Opd1PropInfo = TTI::OP_None,
99 TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None,
100 ArrayRef<const Value *> Args = ArrayRef<const Value *>());
102 unsigned getCFInstrCost(unsigned Opcode);
104 int getVectorInstrCost(unsigned Opcode, Type *ValTy, unsigned Index);
105 bool isSourceOfDivergence(const Value *V) const;
107 unsigned getFlatAddressSpace() const {
108 // Don't bother running InferAddressSpaces pass on graphics shaders which
109 // don't use flat addressing.
110 if (IsGraphicsShader)
112 return ST->hasFlatAddressSpace() ?
113 ST->getAMDGPUAS().FLAT_ADDRESS : ST->getAMDGPUAS().UNKNOWN_ADDRESS_SPACE;
116 unsigned getVectorSplitCost() { return 0; }
119 } // end namespace llvm