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