1 //===-- NVPTXTargetTransformInfo.h - NVPTX specific TTI ---------*- C++ -*-===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 /// \file 9 /// This file a TargetTransformInfo::Concept conforming object specific to the 10 /// NVPTX target machine. It uses the target's detailed information to 11 /// provide more precise answers to certain TTI queries, while letting the 12 /// target independent and default TTI implementations handle the rest. 13 /// 14 //===----------------------------------------------------------------------===// 15 16 #ifndef LLVM_LIB_TARGET_NVPTX_NVPTXTARGETTRANSFORMINFO_H 17 #define LLVM_LIB_TARGET_NVPTX_NVPTXTARGETTRANSFORMINFO_H 18 19 #include "NVPTXTargetMachine.h" 20 #include "MCTargetDesc/NVPTXBaseInfo.h" 21 #include "llvm/Analysis/TargetTransformInfo.h" 22 #include "llvm/CodeGen/BasicTTIImpl.h" 23 #include "llvm/CodeGen/TargetLowering.h" 24 25 namespace llvm { 26 27 class NVPTXTTIImpl : public BasicTTIImplBase<NVPTXTTIImpl> { 28 typedef BasicTTIImplBase<NVPTXTTIImpl> BaseT; 29 typedef TargetTransformInfo TTI; 30 friend BaseT; 31 32 const NVPTXSubtarget *ST; 33 const NVPTXTargetLowering *TLI; 34 getST()35 const NVPTXSubtarget *getST() const { return ST; }; getTLI()36 const NVPTXTargetLowering *getTLI() const { return TLI; }; 37 38 public: NVPTXTTIImpl(const NVPTXTargetMachine * TM,const Function & F)39 explicit NVPTXTTIImpl(const NVPTXTargetMachine *TM, const Function &F) 40 : BaseT(TM, F.getParent()->getDataLayout()), ST(TM->getSubtargetImpl()), 41 TLI(ST->getTargetLowering()) {} 42 hasBranchDivergence()43 bool hasBranchDivergence() { return true; } 44 45 bool isSourceOfDivergence(const Value *V); 46 getFlatAddressSpace()47 unsigned getFlatAddressSpace() const { 48 return AddressSpace::ADDRESS_SPACE_GENERIC; 49 } 50 51 // Loads and stores can be vectorized if the alignment is at least as big as 52 // the load/store we want to vectorize. isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,unsigned Alignment,unsigned AddrSpace)53 bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes, 54 unsigned Alignment, 55 unsigned AddrSpace) const { 56 return Alignment >= ChainSizeInBytes; 57 } isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,unsigned Alignment,unsigned AddrSpace)58 bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes, 59 unsigned Alignment, 60 unsigned AddrSpace) const { 61 return isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment, AddrSpace); 62 } 63 64 // NVPTX has infinite registers of all kinds, but the actual machine doesn't. 65 // We conservatively return 1 here which is just enough to enable the 66 // vectorizers but disables heuristics based on the number of registers. 67 // FIXME: Return a more reasonable number, while keeping an eye on 68 // LoopVectorizer's unrolling heuristics. getNumberOfRegisters(bool Vector)69 unsigned getNumberOfRegisters(bool Vector) const { return 1; } 70 71 // Only <2 x half> should be vectorized, so always return 32 for the vector 72 // register size. getRegisterBitWidth(bool Vector)73 unsigned getRegisterBitWidth(bool Vector) const { return 32; } getMinVectorRegisterBitWidth()74 unsigned getMinVectorRegisterBitWidth() const { return 32; } 75 76 // We don't want to prevent inlining because of target-cpu and -features 77 // attributes that were added to newer versions of LLVM/Clang: There are 78 // no incompatible functions in PTX, ptxas will throw errors in such cases. areInlineCompatible(const Function * Caller,const Function * Callee)79 bool areInlineCompatible(const Function *Caller, 80 const Function *Callee) const { 81 return true; 82 } 83 84 // Increase the inlining cost threshold by a factor of 5, reflecting that 85 // calls are particularly expensive in NVPTX. getInliningThresholdMultiplier()86 unsigned getInliningThresholdMultiplier() { return 5; } 87 88 int getArithmeticInstrCost( 89 unsigned Opcode, Type *Ty, 90 TTI::OperandValueKind Opd1Info = TTI::OK_AnyValue, 91 TTI::OperandValueKind Opd2Info = TTI::OK_AnyValue, 92 TTI::OperandValueProperties Opd1PropInfo = TTI::OP_None, 93 TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None, 94 ArrayRef<const Value *> Args = ArrayRef<const Value *>(), 95 const Instruction *CxtI = nullptr); 96 97 void getUnrollingPreferences(Loop *L, ScalarEvolution &SE, 98 TTI::UnrollingPreferences &UP); hasVolatileVariant(Instruction * I,unsigned AddrSpace)99 bool hasVolatileVariant(Instruction *I, unsigned AddrSpace) { 100 // Volatile loads/stores are only supported for shared and global address 101 // spaces, or for generic AS that maps to them. 102 if (!(AddrSpace == llvm::ADDRESS_SPACE_GENERIC || 103 AddrSpace == llvm::ADDRESS_SPACE_GLOBAL || 104 AddrSpace == llvm::ADDRESS_SPACE_SHARED)) 105 return false; 106 107 switch(I->getOpcode()){ 108 default: 109 return false; 110 case Instruction::Load: 111 case Instruction::Store: 112 return true; 113 } 114 } 115 }; 116 117 } // end namespace llvm 118 119 #endif 120