1 //===-- WebAssemblyTargetTransformInfo.cpp - WebAssembly-specific TTI -----===// 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 /// 9 /// \file 10 /// This file defines the WebAssembly-specific TargetTransformInfo 11 /// implementation. 12 /// 13 //===----------------------------------------------------------------------===// 14 15 #include "WebAssemblyTargetTransformInfo.h" 16 #include "llvm/CodeGen/CostTable.h" 17 #include "llvm/Support/Debug.h" 18 using namespace llvm; 19 20 #define DEBUG_TYPE "wasmtti" 21 22 TargetTransformInfo::PopcntSupportKind 23 WebAssemblyTTIImpl::getPopcntSupport(unsigned TyWidth) const { 24 assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2"); 25 return TargetTransformInfo::PSK_FastHardware; 26 } 27 28 unsigned WebAssemblyTTIImpl::getNumberOfRegisters(unsigned ClassID) const { 29 unsigned Result = BaseT::getNumberOfRegisters(ClassID); 30 31 // For SIMD, use at least 16 registers, as a rough guess. 32 bool Vector = (ClassID == 1); 33 if (Vector) 34 Result = std::max(Result, 16u); 35 36 return Result; 37 } 38 39 unsigned WebAssemblyTTIImpl::getRegisterBitWidth(bool Vector) const { 40 if (Vector && getST()->hasSIMD128()) 41 return 128; 42 43 return 64; 44 } 45 46 unsigned WebAssemblyTTIImpl::getArithmeticInstrCost( 47 unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind, 48 TTI::OperandValueKind Opd1Info, 49 TTI::OperandValueKind Opd2Info, TTI::OperandValueProperties Opd1PropInfo, 50 TTI::OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args, 51 const Instruction *CxtI) { 52 53 unsigned Cost = BasicTTIImplBase<WebAssemblyTTIImpl>::getArithmeticInstrCost( 54 Opcode, Ty, CostKind, Opd1Info, Opd2Info, Opd1PropInfo, Opd2PropInfo); 55 56 if (auto *VTy = dyn_cast<VectorType>(Ty)) { 57 switch (Opcode) { 58 case Instruction::LShr: 59 case Instruction::AShr: 60 case Instruction::Shl: 61 // SIMD128's shifts currently only accept a scalar shift count. For each 62 // element, we'll need to extract, op, insert. The following is a rough 63 // approxmation. 64 if (Opd2Info != TTI::OK_UniformValue && 65 Opd2Info != TTI::OK_UniformConstantValue) 66 Cost = 67 cast<FixedVectorType>(VTy)->getNumElements() * 68 (TargetTransformInfo::TCC_Basic + 69 getArithmeticInstrCost(Opcode, VTy->getElementType(), CostKind) + 70 TargetTransformInfo::TCC_Basic); 71 break; 72 } 73 } 74 return Cost; 75 } 76 77 unsigned WebAssemblyTTIImpl::getVectorInstrCost(unsigned Opcode, Type *Val, 78 unsigned Index) { 79 unsigned Cost = BasicTTIImplBase::getVectorInstrCost(Opcode, Val, Index); 80 81 // SIMD128's insert/extract currently only take constant indices. 82 if (Index == -1u) 83 return Cost + 25 * TargetTransformInfo::TCC_Expensive; 84 85 return Cost; 86 } 87 88 bool WebAssemblyTTIImpl::areInlineCompatible(const Function *Caller, 89 const Function *Callee) const { 90 // Allow inlining only when the Callee has a subset of the Caller's 91 // features. In principle, we should be able to inline regardless of any 92 // features because WebAssembly supports features at module granularity, not 93 // function granularity, but without this restriction it would be possible for 94 // a module to "forget" about features if all the functions that used them 95 // were inlined. 96 const TargetMachine &TM = getTLI()->getTargetMachine(); 97 98 const FeatureBitset &CallerBits = 99 TM.getSubtargetImpl(*Caller)->getFeatureBits(); 100 const FeatureBitset &CalleeBits = 101 TM.getSubtargetImpl(*Callee)->getFeatureBits(); 102 103 return (CallerBits & CalleeBits) == CalleeBits; 104 } 105 106 void WebAssemblyTTIImpl::getUnrollingPreferences( 107 Loop *L, ScalarEvolution &SE, TTI::UnrollingPreferences &UP) const { 108 // Scan the loop: don't unroll loops with calls. This is a standard approach 109 // for most (all?) targets. 110 for (BasicBlock *BB : L->blocks()) 111 for (Instruction &I : *BB) 112 if (isa<CallInst>(I) || isa<InvokeInst>(I)) 113 if (const Function *F = cast<CallBase>(I).getCalledFunction()) 114 if (isLoweredToCall(F)) 115 return; 116 117 // The chosen threshold is within the range of 'LoopMicroOpBufferSize' of 118 // the various microarchitectures that use the BasicTTI implementation and 119 // has been selected through heuristics across multiple cores and runtimes. 120 UP.Partial = UP.Runtime = UP.UpperBound = true; 121 UP.PartialThreshold = 30; 122 123 // Avoid unrolling when optimizing for size. 124 UP.OptSizeThreshold = 0; 125 UP.PartialOptSizeThreshold = 0; 126 127 // Set number of instructions optimized when "back edge" 128 // becomes "fall through" to default value of 2. 129 UP.BEInsns = 2; 130 } 131