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 35 const NVPTXSubtarget *getST() const { return ST; }; 36 const NVPTXTargetLowering *getTLI() const { return TLI; }; 37 38 public: 39 explicit NVPTXTTIImpl(const NVPTXTargetMachine *TM, const Function &F) 40 : BaseT(TM, F.getParent()->getDataLayout()), ST(TM->getSubtargetImpl()), 41 TLI(ST->getTargetLowering()) {} 42 43 bool hasBranchDivergence() { return true; } 44 45 bool isSourceOfDivergence(const Value *V); 46 47 unsigned getFlatAddressSpace() const { 48 return AddressSpace::ADDRESS_SPACE_GENERIC; 49 } 50 51 Optional<Instruction *> instCombineIntrinsic(InstCombiner &IC, 52 IntrinsicInst &II) const; 53 54 // Loads and stores can be vectorized if the alignment is at least as big as 55 // the load/store we want to vectorize. 56 bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes, Align Alignment, 57 unsigned AddrSpace) const { 58 return Alignment >= ChainSizeInBytes; 59 } 60 bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes, Align Alignment, 61 unsigned AddrSpace) const { 62 return isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment, AddrSpace); 63 } 64 65 // NVPTX has infinite registers of all kinds, but the actual machine doesn't. 66 // We conservatively return 1 here which is just enough to enable the 67 // vectorizers but disables heuristics based on the number of registers. 68 // FIXME: Return a more reasonable number, while keeping an eye on 69 // LoopVectorizer's unrolling heuristics. 70 unsigned getNumberOfRegisters(bool Vector) const { return 1; } 71 72 // Only <2 x half> should be vectorized, so always return 32 for the vector 73 // register size. 74 TypeSize getRegisterBitWidth(TargetTransformInfo::RegisterKind K) const { 75 return TypeSize::getFixed(32); 76 } 77 unsigned getMinVectorRegisterBitWidth() const { return 32; } 78 79 // We don't want to prevent inlining because of target-cpu and -features 80 // attributes that were added to newer versions of LLVM/Clang: There are 81 // no incompatible functions in PTX, ptxas will throw errors in such cases. 82 bool areInlineCompatible(const Function *Caller, 83 const Function *Callee) const { 84 return true; 85 } 86 87 // Increase the inlining cost threshold by a factor of 5, reflecting that 88 // calls are particularly expensive in NVPTX. 89 unsigned getInliningThresholdMultiplier() { return 5; } 90 91 InstructionCost getArithmeticInstrCost( 92 unsigned Opcode, Type *Ty, 93 TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput, 94 TTI::OperandValueKind Opd1Info = TTI::OK_AnyValue, 95 TTI::OperandValueKind Opd2Info = TTI::OK_AnyValue, 96 TTI::OperandValueProperties Opd1PropInfo = TTI::OP_None, 97 TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None, 98 ArrayRef<const Value *> Args = ArrayRef<const Value *>(), 99 const Instruction *CxtI = nullptr); 100 101 void getUnrollingPreferences(Loop *L, ScalarEvolution &SE, 102 TTI::UnrollingPreferences &UP); 103 104 void getPeelingPreferences(Loop *L, ScalarEvolution &SE, 105 TTI::PeelingPreferences &PP); 106 107 bool hasVolatileVariant(Instruction *I, unsigned AddrSpace) { 108 // Volatile loads/stores are only supported for shared and global address 109 // spaces, or for generic AS that maps to them. 110 if (!(AddrSpace == llvm::ADDRESS_SPACE_GENERIC || 111 AddrSpace == llvm::ADDRESS_SPACE_GLOBAL || 112 AddrSpace == llvm::ADDRESS_SPACE_SHARED)) 113 return false; 114 115 switch(I->getOpcode()){ 116 default: 117 return false; 118 case Instruction::Load: 119 case Instruction::Store: 120 return true; 121 } 122 } 123 }; 124 125 } // end namespace llvm 126 127 #endif 128