| File: | build/source/llvm/lib/Target/ARM/MVEGatherScatterLowering.cpp |
| Warning: | line 204, column 39 The result of the left shift is undefined due to shifting by '128', which is greater or equal to the width of type 'unsigned long long' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===- MVEGatherScatterLowering.cpp - Gather/Scatter lowering -------------===// | ||||||||
| 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 | /// This pass custom lowers llvm.gather and llvm.scatter instructions to | ||||||||
| 10 | /// arm.mve.gather and arm.mve.scatter intrinsics, optimising the code to | ||||||||
| 11 | /// produce a better final result as we go. | ||||||||
| 12 | // | ||||||||
| 13 | //===----------------------------------------------------------------------===// | ||||||||
| 14 | |||||||||
| 15 | #include "ARM.h" | ||||||||
| 16 | #include "ARMBaseInstrInfo.h" | ||||||||
| 17 | #include "ARMSubtarget.h" | ||||||||
| 18 | #include "llvm/Analysis/LoopInfo.h" | ||||||||
| 19 | #include "llvm/Analysis/TargetTransformInfo.h" | ||||||||
| 20 | #include "llvm/Analysis/ValueTracking.h" | ||||||||
| 21 | #include "llvm/CodeGen/TargetLowering.h" | ||||||||
| 22 | #include "llvm/CodeGen/TargetPassConfig.h" | ||||||||
| 23 | #include "llvm/CodeGen/TargetSubtargetInfo.h" | ||||||||
| 24 | #include "llvm/InitializePasses.h" | ||||||||
| 25 | #include "llvm/IR/BasicBlock.h" | ||||||||
| 26 | #include "llvm/IR/Constant.h" | ||||||||
| 27 | #include "llvm/IR/Constants.h" | ||||||||
| 28 | #include "llvm/IR/DerivedTypes.h" | ||||||||
| 29 | #include "llvm/IR/Function.h" | ||||||||
| 30 | #include "llvm/IR/InstrTypes.h" | ||||||||
| 31 | #include "llvm/IR/Instruction.h" | ||||||||
| 32 | #include "llvm/IR/Instructions.h" | ||||||||
| 33 | #include "llvm/IR/IntrinsicInst.h" | ||||||||
| 34 | #include "llvm/IR/Intrinsics.h" | ||||||||
| 35 | #include "llvm/IR/IntrinsicsARM.h" | ||||||||
| 36 | #include "llvm/IR/IRBuilder.h" | ||||||||
| 37 | #include "llvm/IR/PatternMatch.h" | ||||||||
| 38 | #include "llvm/IR/Type.h" | ||||||||
| 39 | #include "llvm/IR/Value.h" | ||||||||
| 40 | #include "llvm/Pass.h" | ||||||||
| 41 | #include "llvm/Support/Casting.h" | ||||||||
| 42 | #include "llvm/Transforms/Utils/Local.h" | ||||||||
| 43 | #include <algorithm> | ||||||||
| 44 | #include <cassert> | ||||||||
| 45 | |||||||||
| 46 | using namespace llvm; | ||||||||
| 47 | |||||||||
| 48 | #define DEBUG_TYPE"arm-mve-gather-scatter-lowering" "arm-mve-gather-scatter-lowering" | ||||||||
| 49 | |||||||||
| 50 | cl::opt<bool> EnableMaskedGatherScatters( | ||||||||
| 51 | "enable-arm-maskedgatscat", cl::Hidden, cl::init(true), | ||||||||
| 52 | cl::desc("Enable the generation of masked gathers and scatters")); | ||||||||
| 53 | |||||||||
| 54 | namespace { | ||||||||
| 55 | |||||||||
| 56 | class MVEGatherScatterLowering : public FunctionPass { | ||||||||
| 57 | public: | ||||||||
| 58 | static char ID; // Pass identification, replacement for typeid | ||||||||
| 59 | |||||||||
| 60 | explicit MVEGatherScatterLowering() : FunctionPass(ID) { | ||||||||
| 61 | initializeMVEGatherScatterLoweringPass(*PassRegistry::getPassRegistry()); | ||||||||
| 62 | } | ||||||||
| 63 | |||||||||
| 64 | bool runOnFunction(Function &F) override; | ||||||||
| 65 | |||||||||
| 66 | StringRef getPassName() const override { | ||||||||
| 67 | return "MVE gather/scatter lowering"; | ||||||||
| 68 | } | ||||||||
| 69 | |||||||||
| 70 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||||||
| 71 | AU.setPreservesCFG(); | ||||||||
| 72 | AU.addRequired<TargetPassConfig>(); | ||||||||
| 73 | AU.addRequired<LoopInfoWrapperPass>(); | ||||||||
| 74 | FunctionPass::getAnalysisUsage(AU); | ||||||||
| 75 | } | ||||||||
| 76 | |||||||||
| 77 | private: | ||||||||
| 78 | LoopInfo *LI = nullptr; | ||||||||
| 79 | const DataLayout *DL; | ||||||||
| 80 | |||||||||
| 81 | // Check this is a valid gather with correct alignment | ||||||||
| 82 | bool isLegalTypeAndAlignment(unsigned NumElements, unsigned ElemSize, | ||||||||
| 83 | Align Alignment); | ||||||||
| 84 | // Check whether Ptr is hidden behind a bitcast and look through it | ||||||||
| 85 | void lookThroughBitcast(Value *&Ptr); | ||||||||
| 86 | // Decompose a ptr into Base and Offsets, potentially using a GEP to return a | ||||||||
| 87 | // scalar base and vector offsets, or else fallback to using a base of 0 and | ||||||||
| 88 | // offset of Ptr where possible. | ||||||||
| 89 | Value *decomposePtr(Value *Ptr, Value *&Offsets, int &Scale, | ||||||||
| 90 | FixedVectorType *Ty, Type *MemoryTy, | ||||||||
| 91 | IRBuilder<> &Builder); | ||||||||
| 92 | // Check for a getelementptr and deduce base and offsets from it, on success | ||||||||
| 93 | // returning the base directly and the offsets indirectly using the Offsets | ||||||||
| 94 | // argument | ||||||||
| 95 | Value *decomposeGEP(Value *&Offsets, FixedVectorType *Ty, | ||||||||
| 96 | GetElementPtrInst *GEP, IRBuilder<> &Builder); | ||||||||
| 97 | // Compute the scale of this gather/scatter instruction | ||||||||
| 98 | int computeScale(unsigned GEPElemSize, unsigned MemoryElemSize); | ||||||||
| 99 | // If the value is a constant, or derived from constants via additions | ||||||||
| 100 | // and multilications, return its numeric value | ||||||||
| 101 | std::optional<int64_t> getIfConst(const Value *V); | ||||||||
| 102 | // If Inst is an add instruction, check whether one summand is a | ||||||||
| 103 | // constant. If so, scale this constant and return it together with | ||||||||
| 104 | // the other summand. | ||||||||
| 105 | std::pair<Value *, int64_t> getVarAndConst(Value *Inst, int TypeScale); | ||||||||
| 106 | |||||||||
| 107 | Instruction *lowerGather(IntrinsicInst *I); | ||||||||
| 108 | // Create a gather from a base + vector of offsets | ||||||||
| 109 | Instruction *tryCreateMaskedGatherOffset(IntrinsicInst *I, Value *Ptr, | ||||||||
| 110 | Instruction *&Root, | ||||||||
| 111 | IRBuilder<> &Builder); | ||||||||
| 112 | // Create a gather from a vector of pointers | ||||||||
| 113 | Instruction *tryCreateMaskedGatherBase(IntrinsicInst *I, Value *Ptr, | ||||||||
| 114 | IRBuilder<> &Builder, | ||||||||
| 115 | int64_t Increment = 0); | ||||||||
| 116 | // Create an incrementing gather from a vector of pointers | ||||||||
| 117 | Instruction *tryCreateMaskedGatherBaseWB(IntrinsicInst *I, Value *Ptr, | ||||||||
| 118 | IRBuilder<> &Builder, | ||||||||
| 119 | int64_t Increment = 0); | ||||||||
| 120 | |||||||||
| 121 | Instruction *lowerScatter(IntrinsicInst *I); | ||||||||
| 122 | // Create a scatter to a base + vector of offsets | ||||||||
| 123 | Instruction *tryCreateMaskedScatterOffset(IntrinsicInst *I, Value *Offsets, | ||||||||
| 124 | IRBuilder<> &Builder); | ||||||||
| 125 | // Create a scatter to a vector of pointers | ||||||||
| 126 | Instruction *tryCreateMaskedScatterBase(IntrinsicInst *I, Value *Ptr, | ||||||||
| 127 | IRBuilder<> &Builder, | ||||||||
| 128 | int64_t Increment = 0); | ||||||||
| 129 | // Create an incrementing scatter from a vector of pointers | ||||||||
| 130 | Instruction *tryCreateMaskedScatterBaseWB(IntrinsicInst *I, Value *Ptr, | ||||||||
| 131 | IRBuilder<> &Builder, | ||||||||
| 132 | int64_t Increment = 0); | ||||||||
| 133 | |||||||||
| 134 | // QI gathers and scatters can increment their offsets on their own if | ||||||||
| 135 | // the increment is a constant value (digit) | ||||||||
| 136 | Instruction *tryCreateIncrementingGatScat(IntrinsicInst *I, Value *Ptr, | ||||||||
| 137 | IRBuilder<> &Builder); | ||||||||
| 138 | // QI gathers/scatters can increment their offsets on their own if the | ||||||||
| 139 | // increment is a constant value (digit) - this creates a writeback QI | ||||||||
| 140 | // gather/scatter | ||||||||
| 141 | Instruction *tryCreateIncrementingWBGatScat(IntrinsicInst *I, Value *BasePtr, | ||||||||
| 142 | Value *Ptr, unsigned TypeScale, | ||||||||
| 143 | IRBuilder<> &Builder); | ||||||||
| 144 | |||||||||
| 145 | // Optimise the base and offsets of the given address | ||||||||
| 146 | bool optimiseAddress(Value *Address, BasicBlock *BB, LoopInfo *LI); | ||||||||
| 147 | // Try to fold consecutive geps together into one | ||||||||
| 148 | Value *foldGEP(GetElementPtrInst *GEP, Value *&Offsets, unsigned &Scale, | ||||||||
| 149 | IRBuilder<> &Builder); | ||||||||
| 150 | // Check whether these offsets could be moved out of the loop they're in | ||||||||
| 151 | bool optimiseOffsets(Value *Offsets, BasicBlock *BB, LoopInfo *LI); | ||||||||
| 152 | // Pushes the given add out of the loop | ||||||||
| 153 | void pushOutAdd(PHINode *&Phi, Value *OffsSecondOperand, unsigned StartIndex); | ||||||||
| 154 | // Pushes the given mul or shl out of the loop | ||||||||
| 155 | void pushOutMulShl(unsigned Opc, PHINode *&Phi, Value *IncrementPerRound, | ||||||||
| 156 | Value *OffsSecondOperand, unsigned LoopIncrement, | ||||||||
| 157 | IRBuilder<> &Builder); | ||||||||
| 158 | }; | ||||||||
| 159 | |||||||||
| 160 | } // end anonymous namespace | ||||||||
| 161 | |||||||||
| 162 | char MVEGatherScatterLowering::ID = 0; | ||||||||
| 163 | |||||||||
| 164 | INITIALIZE_PASS(MVEGatherScatterLowering, DEBUG_TYPE,static void *initializeMVEGatherScatterLoweringPassOnce(PassRegistry &Registry) { PassInfo *PI = new PassInfo( "MVE gather/scattering lowering pass" , "arm-mve-gather-scatter-lowering", &MVEGatherScatterLowering ::ID, PassInfo::NormalCtor_t(callDefaultCtor<MVEGatherScatterLowering >), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeMVEGatherScatterLoweringPassFlag ; void llvm::initializeMVEGatherScatterLoweringPass(PassRegistry &Registry) { llvm::call_once(InitializeMVEGatherScatterLoweringPassFlag , initializeMVEGatherScatterLoweringPassOnce, std::ref(Registry )); } | ||||||||
| 165 | "MVE gather/scattering lowering pass", false, false)static void *initializeMVEGatherScatterLoweringPassOnce(PassRegistry &Registry) { PassInfo *PI = new PassInfo( "MVE gather/scattering lowering pass" , "arm-mve-gather-scatter-lowering", &MVEGatherScatterLowering ::ID, PassInfo::NormalCtor_t(callDefaultCtor<MVEGatherScatterLowering >), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeMVEGatherScatterLoweringPassFlag ; void llvm::initializeMVEGatherScatterLoweringPass(PassRegistry &Registry) { llvm::call_once(InitializeMVEGatherScatterLoweringPassFlag , initializeMVEGatherScatterLoweringPassOnce, std::ref(Registry )); } | ||||||||
| 166 | |||||||||
| 167 | Pass *llvm::createMVEGatherScatterLoweringPass() { | ||||||||
| 168 | return new MVEGatherScatterLowering(); | ||||||||
| 169 | } | ||||||||
| 170 | |||||||||
| 171 | bool MVEGatherScatterLowering::isLegalTypeAndAlignment(unsigned NumElements, | ||||||||
| 172 | unsigned ElemSize, | ||||||||
| 173 | Align Alignment) { | ||||||||
| 174 | if (((NumElements == 4 && | ||||||||
| 175 | (ElemSize == 32 || ElemSize == 16 || ElemSize == 8)) || | ||||||||
| 176 | (NumElements == 8 && (ElemSize == 16 || ElemSize == 8)) || | ||||||||
| 177 | (NumElements == 16 && ElemSize == 8)) && | ||||||||
| 178 | Alignment >= ElemSize / 8) | ||||||||
| 179 | return true; | ||||||||
| 180 | LLVM_DEBUG(dbgs() << "masked gathers/scatters: instruction does not have "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: instruction does not have " << "valid alignment or vector type \n"; } } while (false ) | ||||||||
| 181 | << "valid alignment or vector type \n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: instruction does not have " << "valid alignment or vector type \n"; } } while (false ); | ||||||||
| 182 | return false; | ||||||||
| 183 | } | ||||||||
| 184 | |||||||||
| 185 | static bool checkOffsetSize(Value *Offsets, unsigned TargetElemCount) { | ||||||||
| 186 | // Offsets that are not of type <N x i32> are sign extended by the | ||||||||
| 187 | // getelementptr instruction, and MVE gathers/scatters treat the offset as | ||||||||
| 188 | // unsigned. Thus, if the element size is smaller than 32, we can only allow | ||||||||
| 189 | // positive offsets - i.e., the offsets are not allowed to be variables we | ||||||||
| 190 | // can't look into. | ||||||||
| 191 | // Additionally, <N x i32> offsets have to either originate from a zext of a | ||||||||
| 192 | // vector with element types smaller or equal the type of the gather we're | ||||||||
| 193 | // looking at, or consist of constants that we can check are small enough | ||||||||
| 194 | // to fit into the gather type. | ||||||||
| 195 | // Thus we check that 0 < value < 2^TargetElemSize. | ||||||||
| 196 | unsigned TargetElemSize = 128 / TargetElemCount; | ||||||||
| 197 | unsigned OffsetElemSize = cast<FixedVectorType>(Offsets->getType()) | ||||||||
| 198 | ->getElementType() | ||||||||
| 199 | ->getScalarSizeInBits(); | ||||||||
| 200 | if (OffsetElemSize != TargetElemSize || OffsetElemSize != 32) { | ||||||||
| 201 | Constant *ConstOff = dyn_cast<Constant>(Offsets); | ||||||||
| 202 | if (!ConstOff
| ||||||||
| 203 | return false; | ||||||||
| 204 | int64_t TargetElemMaxSize = (1ULL << TargetElemSize); | ||||||||
| |||||||||
| 205 | auto CheckValueSize = [TargetElemMaxSize](Value *OffsetElem) { | ||||||||
| 206 | ConstantInt *OConst = dyn_cast<ConstantInt>(OffsetElem); | ||||||||
| 207 | if (!OConst) | ||||||||
| 208 | return false; | ||||||||
| 209 | int SExtValue = OConst->getSExtValue(); | ||||||||
| 210 | if (SExtValue >= TargetElemMaxSize || SExtValue < 0) | ||||||||
| 211 | return false; | ||||||||
| 212 | return true; | ||||||||
| 213 | }; | ||||||||
| 214 | if (isa<FixedVectorType>(ConstOff->getType())) { | ||||||||
| 215 | for (unsigned i = 0; i < TargetElemCount; i++) { | ||||||||
| 216 | if (!CheckValueSize(ConstOff->getAggregateElement(i))) | ||||||||
| 217 | return false; | ||||||||
| 218 | } | ||||||||
| 219 | } else { | ||||||||
| 220 | if (!CheckValueSize(ConstOff)) | ||||||||
| 221 | return false; | ||||||||
| 222 | } | ||||||||
| 223 | } | ||||||||
| 224 | return true; | ||||||||
| 225 | } | ||||||||
| 226 | |||||||||
| 227 | Value *MVEGatherScatterLowering::decomposePtr(Value *Ptr, Value *&Offsets, | ||||||||
| 228 | int &Scale, FixedVectorType *Ty, | ||||||||
| 229 | Type *MemoryTy, | ||||||||
| 230 | IRBuilder<> &Builder) { | ||||||||
| 231 | if (auto *GEP = dyn_cast<GetElementPtrInst>(Ptr)) { | ||||||||
| 232 | if (Value *V = decomposeGEP(Offsets, Ty, GEP, Builder)) { | ||||||||
| 233 | Scale = | ||||||||
| 234 | computeScale(GEP->getSourceElementType()->getPrimitiveSizeInBits(), | ||||||||
| 235 | MemoryTy->getScalarSizeInBits()); | ||||||||
| 236 | return Scale == -1 ? nullptr : V; | ||||||||
| 237 | } | ||||||||
| 238 | } | ||||||||
| 239 | |||||||||
| 240 | // If we couldn't use the GEP (or it doesn't exist), attempt to use a | ||||||||
| 241 | // BasePtr of 0 with Ptr as the Offsets, so long as there are only 4 | ||||||||
| 242 | // elements. | ||||||||
| 243 | FixedVectorType *PtrTy = cast<FixedVectorType>(Ptr->getType()); | ||||||||
| 244 | if (PtrTy->getNumElements() != 4 || MemoryTy->getScalarSizeInBits() == 32) | ||||||||
| 245 | return nullptr; | ||||||||
| 246 | Value *Zero = ConstantInt::get(Builder.getInt32Ty(), 0); | ||||||||
| 247 | Value *BasePtr = Builder.CreateIntToPtr(Zero, Builder.getInt8PtrTy()); | ||||||||
| 248 | Offsets = Builder.CreatePtrToInt( | ||||||||
| 249 | Ptr, FixedVectorType::get(Builder.getInt32Ty(), 4)); | ||||||||
| 250 | Scale = 0; | ||||||||
| 251 | return BasePtr; | ||||||||
| 252 | } | ||||||||
| 253 | |||||||||
| 254 | Value *MVEGatherScatterLowering::decomposeGEP(Value *&Offsets, | ||||||||
| 255 | FixedVectorType *Ty, | ||||||||
| 256 | GetElementPtrInst *GEP, | ||||||||
| 257 | IRBuilder<> &Builder) { | ||||||||
| 258 | if (!GEP) { | ||||||||
| 259 | LLVM_DEBUG(dbgs() << "masked gathers/scatters: no getelementpointer "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: no getelementpointer " << "found\n"; } } while (false) | ||||||||
| 260 | << "found\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: no getelementpointer " << "found\n"; } } while (false); | ||||||||
| 261 | return nullptr; | ||||||||
| 262 | } | ||||||||
| 263 | LLVM_DEBUG(dbgs() << "masked gathers/scatters: getelementpointer found."do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: getelementpointer found." << " Looking at intrinsic for base + vector of offsets\n" ; } } while (false) | ||||||||
| 264 | << " Looking at intrinsic for base + vector of offsets\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: getelementpointer found." << " Looking at intrinsic for base + vector of offsets\n" ; } } while (false); | ||||||||
| 265 | Value *GEPPtr = GEP->getPointerOperand(); | ||||||||
| 266 | Offsets = GEP->getOperand(1); | ||||||||
| 267 | if (GEPPtr->getType()->isVectorTy() || | ||||||||
| 268 | !isa<FixedVectorType>(Offsets->getType())) | ||||||||
| 269 | return nullptr; | ||||||||
| 270 | |||||||||
| 271 | if (GEP->getNumOperands() != 2) { | ||||||||
| 272 | LLVM_DEBUG(dbgs() << "masked gathers/scatters: getelementptr with too many"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: getelementptr with too many" << " operands. Expanding.\n"; } } while (false) | ||||||||
| 273 | << " operands. Expanding.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: getelementptr with too many" << " operands. Expanding.\n"; } } while (false); | ||||||||
| 274 | return nullptr; | ||||||||
| 275 | } | ||||||||
| 276 | Offsets = GEP->getOperand(1); | ||||||||
| 277 | unsigned OffsetsElemCount = | ||||||||
| 278 | cast<FixedVectorType>(Offsets->getType())->getNumElements(); | ||||||||
| 279 | // Paranoid check whether the number of parallel lanes is the same | ||||||||
| 280 | assert(Ty->getNumElements() == OffsetsElemCount)(static_cast <bool> (Ty->getNumElements() == OffsetsElemCount ) ? void (0) : __assert_fail ("Ty->getNumElements() == OffsetsElemCount" , "llvm/lib/Target/ARM/MVEGatherScatterLowering.cpp", 280, __extension__ __PRETTY_FUNCTION__)); | ||||||||
| 281 | |||||||||
| 282 | ZExtInst *ZextOffs = dyn_cast<ZExtInst>(Offsets); | ||||||||
| 283 | if (ZextOffs) | ||||||||
| 284 | Offsets = ZextOffs->getOperand(0); | ||||||||
| 285 | FixedVectorType *OffsetType = cast<FixedVectorType>(Offsets->getType()); | ||||||||
| 286 | |||||||||
| 287 | // If the offsets are already being zext-ed to <N x i32>, that relieves us of | ||||||||
| 288 | // having to make sure that they won't overflow. | ||||||||
| 289 | if (!ZextOffs || cast<FixedVectorType>(ZextOffs->getDestTy()) | ||||||||
| 290 | ->getElementType() | ||||||||
| 291 | ->getScalarSizeInBits() != 32) | ||||||||
| 292 | if (!checkOffsetSize(Offsets, OffsetsElemCount)) | ||||||||
| 293 | return nullptr; | ||||||||
| 294 | |||||||||
| 295 | // The offset sizes have been checked; if any truncating or zext-ing is | ||||||||
| 296 | // required to fix them, do that now | ||||||||
| 297 | if (Ty != Offsets->getType()) { | ||||||||
| 298 | if ((Ty->getElementType()->getScalarSizeInBits() < | ||||||||
| 299 | OffsetType->getElementType()->getScalarSizeInBits())) { | ||||||||
| 300 | Offsets = Builder.CreateTrunc(Offsets, Ty); | ||||||||
| 301 | } else { | ||||||||
| 302 | Offsets = Builder.CreateZExt(Offsets, VectorType::getInteger(Ty)); | ||||||||
| 303 | } | ||||||||
| 304 | } | ||||||||
| 305 | // If none of the checks failed, return the gep's base pointer | ||||||||
| 306 | LLVM_DEBUG(dbgs() << "masked gathers/scatters: found correct offsets\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: found correct offsets\n" ; } } while (false); | ||||||||
| 307 | return GEPPtr; | ||||||||
| 308 | } | ||||||||
| 309 | |||||||||
| 310 | void MVEGatherScatterLowering::lookThroughBitcast(Value *&Ptr) { | ||||||||
| 311 | // Look through bitcast instruction if #elements is the same | ||||||||
| 312 | if (auto *BitCast = dyn_cast<BitCastInst>(Ptr)) { | ||||||||
| 313 | auto *BCTy = cast<FixedVectorType>(BitCast->getType()); | ||||||||
| 314 | auto *BCSrcTy = cast<FixedVectorType>(BitCast->getOperand(0)->getType()); | ||||||||
| 315 | if (BCTy->getNumElements() == BCSrcTy->getNumElements()) { | ||||||||
| 316 | LLVM_DEBUG(dbgs() << "masked gathers/scatters: looking through "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: looking through " << "bitcast\n"; } } while (false) | ||||||||
| 317 | << "bitcast\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: looking through " << "bitcast\n"; } } while (false); | ||||||||
| 318 | Ptr = BitCast->getOperand(0); | ||||||||
| 319 | } | ||||||||
| 320 | } | ||||||||
| 321 | } | ||||||||
| 322 | |||||||||
| 323 | int MVEGatherScatterLowering::computeScale(unsigned GEPElemSize, | ||||||||
| 324 | unsigned MemoryElemSize) { | ||||||||
| 325 | // This can be a 32bit load/store scaled by 4, a 16bit load/store scaled by 2, | ||||||||
| 326 | // or a 8bit, 16bit or 32bit load/store scaled by 1 | ||||||||
| 327 | if (GEPElemSize == 32 && MemoryElemSize == 32) | ||||||||
| 328 | return 2; | ||||||||
| 329 | else if (GEPElemSize == 16 && MemoryElemSize == 16) | ||||||||
| 330 | return 1; | ||||||||
| 331 | else if (GEPElemSize == 8) | ||||||||
| 332 | return 0; | ||||||||
| 333 | LLVM_DEBUG(dbgs() << "masked gathers/scatters: incorrect scale. Can't "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: incorrect scale. Can't " << "create intrinsic\n"; } } while (false) | ||||||||
| 334 | << "create intrinsic\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: incorrect scale. Can't " << "create intrinsic\n"; } } while (false); | ||||||||
| 335 | return -1; | ||||||||
| 336 | } | ||||||||
| 337 | |||||||||
| 338 | std::optional<int64_t> MVEGatherScatterLowering::getIfConst(const Value *V) { | ||||||||
| 339 | const Constant *C = dyn_cast<Constant>(V); | ||||||||
| 340 | if (C && C->getSplatValue()) | ||||||||
| 341 | return std::optional<int64_t>{C->getUniqueInteger().getSExtValue()}; | ||||||||
| 342 | if (!isa<Instruction>(V)) | ||||||||
| 343 | return std::optional<int64_t>{}; | ||||||||
| 344 | |||||||||
| 345 | const Instruction *I = cast<Instruction>(V); | ||||||||
| 346 | if (I->getOpcode() == Instruction::Add || I->getOpcode() == Instruction::Or || | ||||||||
| 347 | I->getOpcode() == Instruction::Mul || | ||||||||
| 348 | I->getOpcode() == Instruction::Shl) { | ||||||||
| 349 | std::optional<int64_t> Op0 = getIfConst(I->getOperand(0)); | ||||||||
| 350 | std::optional<int64_t> Op1 = getIfConst(I->getOperand(1)); | ||||||||
| 351 | if (!Op0 || !Op1) | ||||||||
| 352 | return std::optional<int64_t>{}; | ||||||||
| 353 | if (I->getOpcode() == Instruction::Add) | ||||||||
| 354 | return std::optional<int64_t>{*Op0 + *Op1}; | ||||||||
| 355 | if (I->getOpcode() == Instruction::Mul) | ||||||||
| 356 | return std::optional<int64_t>{*Op0 * *Op1}; | ||||||||
| 357 | if (I->getOpcode() == Instruction::Shl) | ||||||||
| 358 | return std::optional<int64_t>{*Op0 << *Op1}; | ||||||||
| 359 | if (I->getOpcode() == Instruction::Or) | ||||||||
| 360 | return std::optional<int64_t>{*Op0 | *Op1}; | ||||||||
| 361 | } | ||||||||
| 362 | return std::optional<int64_t>{}; | ||||||||
| 363 | } | ||||||||
| 364 | |||||||||
| 365 | // Return true if I is an Or instruction that is equivalent to an add, due to | ||||||||
| 366 | // the operands having no common bits set. | ||||||||
| 367 | static bool isAddLikeOr(Instruction *I, const DataLayout &DL) { | ||||||||
| 368 | return I->getOpcode() == Instruction::Or && | ||||||||
| 369 | haveNoCommonBitsSet(I->getOperand(0), I->getOperand(1), DL); | ||||||||
| 370 | } | ||||||||
| 371 | |||||||||
| 372 | std::pair<Value *, int64_t> | ||||||||
| 373 | MVEGatherScatterLowering::getVarAndConst(Value *Inst, int TypeScale) { | ||||||||
| 374 | std::pair<Value *, int64_t> ReturnFalse = | ||||||||
| 375 | std::pair<Value *, int64_t>(nullptr, 0); | ||||||||
| 376 | // At this point, the instruction we're looking at must be an add or an | ||||||||
| 377 | // add-like-or. | ||||||||
| 378 | Instruction *Add = dyn_cast<Instruction>(Inst); | ||||||||
| 379 | if (Add == nullptr || | ||||||||
| 380 | (Add->getOpcode() != Instruction::Add && !isAddLikeOr(Add, *DL))) | ||||||||
| 381 | return ReturnFalse; | ||||||||
| 382 | |||||||||
| 383 | Value *Summand; | ||||||||
| 384 | std::optional<int64_t> Const; | ||||||||
| 385 | // Find out which operand the value that is increased is | ||||||||
| 386 | if ((Const = getIfConst(Add->getOperand(0)))) | ||||||||
| 387 | Summand = Add->getOperand(1); | ||||||||
| 388 | else if ((Const = getIfConst(Add->getOperand(1)))) | ||||||||
| 389 | Summand = Add->getOperand(0); | ||||||||
| 390 | else | ||||||||
| 391 | return ReturnFalse; | ||||||||
| 392 | |||||||||
| 393 | // Check that the constant is small enough for an incrementing gather | ||||||||
| 394 | int64_t Immediate = *Const << TypeScale; | ||||||||
| 395 | if (Immediate > 512 || Immediate < -512 || Immediate % 4 != 0) | ||||||||
| 396 | return ReturnFalse; | ||||||||
| 397 | |||||||||
| 398 | return std::pair<Value *, int64_t>(Summand, Immediate); | ||||||||
| 399 | } | ||||||||
| 400 | |||||||||
| 401 | Instruction *MVEGatherScatterLowering::lowerGather(IntrinsicInst *I) { | ||||||||
| 402 | using namespace PatternMatch; | ||||||||
| 403 | LLVM_DEBUG(dbgs() << "masked gathers: checking transform preconditions\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers: checking transform preconditions\n" << *I << "\n"; } } while (false) | ||||||||
| 404 | << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers: checking transform preconditions\n" << *I << "\n"; } } while (false); | ||||||||
| 405 | |||||||||
| 406 | // @llvm.masked.gather.*(Ptrs, alignment, Mask, Src0) | ||||||||
| 407 | // Attempt to turn the masked gather in I into a MVE intrinsic | ||||||||
| 408 | // Potentially optimising the addressing modes as we do so. | ||||||||
| 409 | auto *Ty = cast<FixedVectorType>(I->getType()); | ||||||||
| 410 | Value *Ptr = I->getArgOperand(0); | ||||||||
| 411 | Align Alignment = cast<ConstantInt>(I->getArgOperand(1))->getAlignValue(); | ||||||||
| 412 | Value *Mask = I->getArgOperand(2); | ||||||||
| 413 | Value *PassThru = I->getArgOperand(3); | ||||||||
| 414 | |||||||||
| 415 | if (!isLegalTypeAndAlignment(Ty->getNumElements(), Ty->getScalarSizeInBits(), | ||||||||
| 416 | Alignment)) | ||||||||
| 417 | return nullptr; | ||||||||
| 418 | lookThroughBitcast(Ptr); | ||||||||
| 419 | assert(Ptr->getType()->isVectorTy() && "Unexpected pointer type")(static_cast <bool> (Ptr->getType()->isVectorTy() && "Unexpected pointer type") ? void (0) : __assert_fail ("Ptr->getType()->isVectorTy() && \"Unexpected pointer type\"" , "llvm/lib/Target/ARM/MVEGatherScatterLowering.cpp", 419, __extension__ __PRETTY_FUNCTION__)); | ||||||||
| 420 | |||||||||
| 421 | IRBuilder<> Builder(I->getContext()); | ||||||||
| 422 | Builder.SetInsertPoint(I); | ||||||||
| 423 | Builder.SetCurrentDebugLocation(I->getDebugLoc()); | ||||||||
| 424 | |||||||||
| 425 | Instruction *Root = I; | ||||||||
| 426 | |||||||||
| 427 | Instruction *Load = tryCreateIncrementingGatScat(I, Ptr, Builder); | ||||||||
| 428 | if (!Load) | ||||||||
| 429 | Load = tryCreateMaskedGatherOffset(I, Ptr, Root, Builder); | ||||||||
| 430 | if (!Load) | ||||||||
| 431 | Load = tryCreateMaskedGatherBase(I, Ptr, Builder); | ||||||||
| 432 | if (!Load) | ||||||||
| 433 | return nullptr; | ||||||||
| 434 | |||||||||
| 435 | if (!isa<UndefValue>(PassThru) && !match(PassThru, m_Zero())) { | ||||||||
| 436 | LLVM_DEBUG(dbgs() << "masked gathers: found non-trivial passthru - "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers: found non-trivial passthru - " << "creating select\n"; } } while (false) | ||||||||
| 437 | << "creating select\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers: found non-trivial passthru - " << "creating select\n"; } } while (false); | ||||||||
| 438 | Load = SelectInst::Create(Mask, Load, PassThru); | ||||||||
| 439 | Builder.Insert(Load); | ||||||||
| 440 | } | ||||||||
| 441 | |||||||||
| 442 | Root->replaceAllUsesWith(Load); | ||||||||
| 443 | Root->eraseFromParent(); | ||||||||
| 444 | if (Root != I) | ||||||||
| 445 | // If this was an extending gather, we need to get rid of the sext/zext | ||||||||
| 446 | // sext/zext as well as of the gather itself | ||||||||
| 447 | I->eraseFromParent(); | ||||||||
| 448 | |||||||||
| 449 | LLVM_DEBUG(dbgs() << "masked gathers: successfully built masked gather\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers: successfully built masked gather\n" << *Load << "\n"; } } while (false) | ||||||||
| 450 | << *Load << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers: successfully built masked gather\n" << *Load << "\n"; } } while (false); | ||||||||
| 451 | return Load; | ||||||||
| 452 | } | ||||||||
| 453 | |||||||||
| 454 | Instruction *MVEGatherScatterLowering::tryCreateMaskedGatherBase( | ||||||||
| 455 | IntrinsicInst *I, Value *Ptr, IRBuilder<> &Builder, int64_t Increment) { | ||||||||
| 456 | using namespace PatternMatch; | ||||||||
| 457 | auto *Ty = cast<FixedVectorType>(I->getType()); | ||||||||
| 458 | LLVM_DEBUG(dbgs() << "masked gathers: loading from vector of pointers\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers: loading from vector of pointers\n" ; } } while (false); | ||||||||
| 459 | if (Ty->getNumElements() != 4 || Ty->getScalarSizeInBits() != 32) | ||||||||
| 460 | // Can't build an intrinsic for this | ||||||||
| 461 | return nullptr; | ||||||||
| 462 | Value *Mask = I->getArgOperand(2); | ||||||||
| 463 | if (match(Mask, m_One())) | ||||||||
| 464 | return Builder.CreateIntrinsic(Intrinsic::arm_mve_vldr_gather_base, | ||||||||
| 465 | {Ty, Ptr->getType()}, | ||||||||
| 466 | {Ptr, Builder.getInt32(Increment)}); | ||||||||
| 467 | else | ||||||||
| 468 | return Builder.CreateIntrinsic( | ||||||||
| 469 | Intrinsic::arm_mve_vldr_gather_base_predicated, | ||||||||
| 470 | {Ty, Ptr->getType(), Mask->getType()}, | ||||||||
| 471 | {Ptr, Builder.getInt32(Increment), Mask}); | ||||||||
| 472 | } | ||||||||
| 473 | |||||||||
| 474 | Instruction *MVEGatherScatterLowering::tryCreateMaskedGatherBaseWB( | ||||||||
| 475 | IntrinsicInst *I, Value *Ptr, IRBuilder<> &Builder, int64_t Increment) { | ||||||||
| 476 | using namespace PatternMatch; | ||||||||
| 477 | auto *Ty = cast<FixedVectorType>(I->getType()); | ||||||||
| 478 | LLVM_DEBUG(dbgs() << "masked gathers: loading from vector of pointers with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers: loading from vector of pointers with " << "writeback\n"; } } while (false) | ||||||||
| 479 | << "writeback\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers: loading from vector of pointers with " << "writeback\n"; } } while (false); | ||||||||
| 480 | if (Ty->getNumElements() != 4 || Ty->getScalarSizeInBits() != 32) | ||||||||
| 481 | // Can't build an intrinsic for this | ||||||||
| 482 | return nullptr; | ||||||||
| 483 | Value *Mask = I->getArgOperand(2); | ||||||||
| 484 | if (match(Mask, m_One())) | ||||||||
| 485 | return Builder.CreateIntrinsic(Intrinsic::arm_mve_vldr_gather_base_wb, | ||||||||
| 486 | {Ty, Ptr->getType()}, | ||||||||
| 487 | {Ptr, Builder.getInt32(Increment)}); | ||||||||
| 488 | else | ||||||||
| 489 | return Builder.CreateIntrinsic( | ||||||||
| 490 | Intrinsic::arm_mve_vldr_gather_base_wb_predicated, | ||||||||
| 491 | {Ty, Ptr->getType(), Mask->getType()}, | ||||||||
| 492 | {Ptr, Builder.getInt32(Increment), Mask}); | ||||||||
| 493 | } | ||||||||
| 494 | |||||||||
| 495 | Instruction *MVEGatherScatterLowering::tryCreateMaskedGatherOffset( | ||||||||
| 496 | IntrinsicInst *I, Value *Ptr, Instruction *&Root, IRBuilder<> &Builder) { | ||||||||
| 497 | using namespace PatternMatch; | ||||||||
| 498 | |||||||||
| 499 | Type *MemoryTy = I->getType(); | ||||||||
| 500 | Type *ResultTy = MemoryTy; | ||||||||
| 501 | |||||||||
| 502 | unsigned Unsigned = 1; | ||||||||
| 503 | // The size of the gather was already checked in isLegalTypeAndAlignment; | ||||||||
| 504 | // if it was not a full vector width an appropriate extend should follow. | ||||||||
| 505 | auto *Extend = Root; | ||||||||
| 506 | bool TruncResult = false; | ||||||||
| 507 | if (MemoryTy->getPrimitiveSizeInBits() < 128) { | ||||||||
| 508 | if (I->hasOneUse()) { | ||||||||
| 509 | // If the gather has a single extend of the correct type, use an extending | ||||||||
| 510 | // gather and replace the ext. In which case the correct root to replace | ||||||||
| 511 | // is not the CallInst itself, but the instruction which extends it. | ||||||||
| 512 | Instruction* User = cast<Instruction>(*I->users().begin()); | ||||||||
| 513 | if (isa<SExtInst>(User) && | ||||||||
| 514 | User->getType()->getPrimitiveSizeInBits() == 128) { | ||||||||
| 515 | LLVM_DEBUG(dbgs() << "masked gathers: Incorporating extend: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers: Incorporating extend: " << *User << "\n"; } } while (false) | ||||||||
| 516 | << *User << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers: Incorporating extend: " << *User << "\n"; } } while (false); | ||||||||
| 517 | Extend = User; | ||||||||
| 518 | ResultTy = User->getType(); | ||||||||
| 519 | Unsigned = 0; | ||||||||
| 520 | } else if (isa<ZExtInst>(User) && | ||||||||
| 521 | User->getType()->getPrimitiveSizeInBits() == 128) { | ||||||||
| 522 | LLVM_DEBUG(dbgs() << "masked gathers: Incorporating extend: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers: Incorporating extend: " << *ResultTy << "\n"; } } while (false) | ||||||||
| 523 | << *ResultTy << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers: Incorporating extend: " << *ResultTy << "\n"; } } while (false); | ||||||||
| 524 | Extend = User; | ||||||||
| 525 | ResultTy = User->getType(); | ||||||||
| 526 | } | ||||||||
| 527 | } | ||||||||
| 528 | |||||||||
| 529 | // If an extend hasn't been found and the type is an integer, create an | ||||||||
| 530 | // extending gather and truncate back to the original type. | ||||||||
| 531 | if (ResultTy->getPrimitiveSizeInBits() < 128 && | ||||||||
| 532 | ResultTy->isIntOrIntVectorTy()) { | ||||||||
| 533 | ResultTy = ResultTy->getWithNewBitWidth( | ||||||||
| 534 | 128 / cast<FixedVectorType>(ResultTy)->getNumElements()); | ||||||||
| 535 | TruncResult = true; | ||||||||
| 536 | LLVM_DEBUG(dbgs() << "masked gathers: Small input type, truncing to: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers: Small input type, truncing to: " << *ResultTy << "\n"; } } while (false) | ||||||||
| 537 | << *ResultTy << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers: Small input type, truncing to: " << *ResultTy << "\n"; } } while (false); | ||||||||
| 538 | } | ||||||||
| 539 | |||||||||
| 540 | // The final size of the gather must be a full vector width | ||||||||
| 541 | if (ResultTy->getPrimitiveSizeInBits() != 128) { | ||||||||
| 542 | LLVM_DEBUG(dbgs() << "masked gathers: Extend needed but not provided "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers: Extend needed but not provided " "from the correct type. Expanding\n"; } } while (false) | ||||||||
| 543 | "from the correct type. Expanding\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers: Extend needed but not provided " "from the correct type. Expanding\n"; } } while (false); | ||||||||
| 544 | return nullptr; | ||||||||
| 545 | } | ||||||||
| 546 | } | ||||||||
| 547 | |||||||||
| 548 | Value *Offsets; | ||||||||
| 549 | int Scale; | ||||||||
| 550 | Value *BasePtr = decomposePtr( | ||||||||
| 551 | Ptr, Offsets, Scale, cast<FixedVectorType>(ResultTy), MemoryTy, Builder); | ||||||||
| 552 | if (!BasePtr) | ||||||||
| 553 | return nullptr; | ||||||||
| 554 | |||||||||
| 555 | Root = Extend; | ||||||||
| 556 | Value *Mask = I->getArgOperand(2); | ||||||||
| 557 | Instruction *Load = nullptr; | ||||||||
| 558 | if (!match(Mask, m_One())) | ||||||||
| 559 | Load = Builder.CreateIntrinsic( | ||||||||
| 560 | Intrinsic::arm_mve_vldr_gather_offset_predicated, | ||||||||
| 561 | {ResultTy, BasePtr->getType(), Offsets->getType(), Mask->getType()}, | ||||||||
| 562 | {BasePtr, Offsets, Builder.getInt32(MemoryTy->getScalarSizeInBits()), | ||||||||
| 563 | Builder.getInt32(Scale), Builder.getInt32(Unsigned), Mask}); | ||||||||
| 564 | else | ||||||||
| 565 | Load = Builder.CreateIntrinsic( | ||||||||
| 566 | Intrinsic::arm_mve_vldr_gather_offset, | ||||||||
| 567 | {ResultTy, BasePtr->getType(), Offsets->getType()}, | ||||||||
| 568 | {BasePtr, Offsets, Builder.getInt32(MemoryTy->getScalarSizeInBits()), | ||||||||
| 569 | Builder.getInt32(Scale), Builder.getInt32(Unsigned)}); | ||||||||
| 570 | |||||||||
| 571 | if (TruncResult) { | ||||||||
| 572 | Load = TruncInst::Create(Instruction::Trunc, Load, MemoryTy); | ||||||||
| 573 | Builder.Insert(Load); | ||||||||
| 574 | } | ||||||||
| 575 | return Load; | ||||||||
| 576 | } | ||||||||
| 577 | |||||||||
| 578 | Instruction *MVEGatherScatterLowering::lowerScatter(IntrinsicInst *I) { | ||||||||
| 579 | using namespace PatternMatch; | ||||||||
| 580 | LLVM_DEBUG(dbgs() << "masked scatters: checking transform preconditions\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked scatters: checking transform preconditions\n" << *I << "\n"; } } while (false) | ||||||||
| 581 | << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked scatters: checking transform preconditions\n" << *I << "\n"; } } while (false); | ||||||||
| 582 | |||||||||
| 583 | // @llvm.masked.scatter.*(data, ptrs, alignment, mask) | ||||||||
| 584 | // Attempt to turn the masked scatter in I into a MVE intrinsic | ||||||||
| 585 | // Potentially optimising the addressing modes as we do so. | ||||||||
| 586 | Value *Input = I->getArgOperand(0); | ||||||||
| 587 | Value *Ptr = I->getArgOperand(1); | ||||||||
| 588 | Align Alignment = cast<ConstantInt>(I->getArgOperand(2))->getAlignValue(); | ||||||||
| 589 | auto *Ty = cast<FixedVectorType>(Input->getType()); | ||||||||
| 590 | |||||||||
| 591 | if (!isLegalTypeAndAlignment(Ty->getNumElements(), Ty->getScalarSizeInBits(), | ||||||||
| 592 | Alignment)) | ||||||||
| 593 | return nullptr; | ||||||||
| 594 | |||||||||
| 595 | lookThroughBitcast(Ptr); | ||||||||
| 596 | assert(Ptr->getType()->isVectorTy() && "Unexpected pointer type")(static_cast <bool> (Ptr->getType()->isVectorTy() && "Unexpected pointer type") ? void (0) : __assert_fail ("Ptr->getType()->isVectorTy() && \"Unexpected pointer type\"" , "llvm/lib/Target/ARM/MVEGatherScatterLowering.cpp", 596, __extension__ __PRETTY_FUNCTION__)); | ||||||||
| 597 | |||||||||
| 598 | IRBuilder<> Builder(I->getContext()); | ||||||||
| 599 | Builder.SetInsertPoint(I); | ||||||||
| 600 | Builder.SetCurrentDebugLocation(I->getDebugLoc()); | ||||||||
| 601 | |||||||||
| 602 | Instruction *Store = tryCreateIncrementingGatScat(I, Ptr, Builder); | ||||||||
| 603 | if (!Store) | ||||||||
| 604 | Store = tryCreateMaskedScatterOffset(I, Ptr, Builder); | ||||||||
| 605 | if (!Store) | ||||||||
| 606 | Store = tryCreateMaskedScatterBase(I, Ptr, Builder); | ||||||||
| 607 | if (!Store) | ||||||||
| 608 | return nullptr; | ||||||||
| 609 | |||||||||
| 610 | LLVM_DEBUG(dbgs() << "masked scatters: successfully built masked scatter\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked scatters: successfully built masked scatter\n" << *Store << "\n"; } } while (false) | ||||||||
| 611 | << *Store << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked scatters: successfully built masked scatter\n" << *Store << "\n"; } } while (false); | ||||||||
| 612 | I->eraseFromParent(); | ||||||||
| 613 | return Store; | ||||||||
| 614 | } | ||||||||
| 615 | |||||||||
| 616 | Instruction *MVEGatherScatterLowering::tryCreateMaskedScatterBase( | ||||||||
| 617 | IntrinsicInst *I, Value *Ptr, IRBuilder<> &Builder, int64_t Increment) { | ||||||||
| 618 | using namespace PatternMatch; | ||||||||
| 619 | Value *Input = I->getArgOperand(0); | ||||||||
| 620 | auto *Ty = cast<FixedVectorType>(Input->getType()); | ||||||||
| 621 | // Only QR variants allow truncating | ||||||||
| 622 | if (!(Ty->getNumElements() == 4 && Ty->getScalarSizeInBits() == 32)) { | ||||||||
| 623 | // Can't build an intrinsic for this | ||||||||
| 624 | return nullptr; | ||||||||
| 625 | } | ||||||||
| 626 | Value *Mask = I->getArgOperand(3); | ||||||||
| 627 | // int_arm_mve_vstr_scatter_base(_predicated) addr, offset, data(, mask) | ||||||||
| 628 | LLVM_DEBUG(dbgs() << "masked scatters: storing to a vector of pointers\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked scatters: storing to a vector of pointers\n" ; } } while (false); | ||||||||
| 629 | if (match(Mask, m_One())) | ||||||||
| 630 | return Builder.CreateIntrinsic(Intrinsic::arm_mve_vstr_scatter_base, | ||||||||
| 631 | {Ptr->getType(), Input->getType()}, | ||||||||
| 632 | {Ptr, Builder.getInt32(Increment), Input}); | ||||||||
| 633 | else | ||||||||
| 634 | return Builder.CreateIntrinsic( | ||||||||
| 635 | Intrinsic::arm_mve_vstr_scatter_base_predicated, | ||||||||
| 636 | {Ptr->getType(), Input->getType(), Mask->getType()}, | ||||||||
| 637 | {Ptr, Builder.getInt32(Increment), Input, Mask}); | ||||||||
| 638 | } | ||||||||
| 639 | |||||||||
| 640 | Instruction *MVEGatherScatterLowering::tryCreateMaskedScatterBaseWB( | ||||||||
| 641 | IntrinsicInst *I, Value *Ptr, IRBuilder<> &Builder, int64_t Increment) { | ||||||||
| 642 | using namespace PatternMatch; | ||||||||
| 643 | Value *Input = I->getArgOperand(0); | ||||||||
| 644 | auto *Ty = cast<FixedVectorType>(Input->getType()); | ||||||||
| 645 | LLVM_DEBUG(dbgs() << "masked scatters: storing to a vector of pointers "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked scatters: storing to a vector of pointers " << "with writeback\n"; } } while (false) | ||||||||
| 646 | << "with writeback\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked scatters: storing to a vector of pointers " << "with writeback\n"; } } while (false); | ||||||||
| 647 | if (Ty->getNumElements() != 4 || Ty->getScalarSizeInBits() != 32) | ||||||||
| 648 | // Can't build an intrinsic for this | ||||||||
| 649 | return nullptr; | ||||||||
| 650 | Value *Mask = I->getArgOperand(3); | ||||||||
| 651 | if (match(Mask, m_One())) | ||||||||
| 652 | return Builder.CreateIntrinsic(Intrinsic::arm_mve_vstr_scatter_base_wb, | ||||||||
| 653 | {Ptr->getType(), Input->getType()}, | ||||||||
| 654 | {Ptr, Builder.getInt32(Increment), Input}); | ||||||||
| 655 | else | ||||||||
| 656 | return Builder.CreateIntrinsic( | ||||||||
| 657 | Intrinsic::arm_mve_vstr_scatter_base_wb_predicated, | ||||||||
| 658 | {Ptr->getType(), Input->getType(), Mask->getType()}, | ||||||||
| 659 | {Ptr, Builder.getInt32(Increment), Input, Mask}); | ||||||||
| 660 | } | ||||||||
| 661 | |||||||||
| 662 | Instruction *MVEGatherScatterLowering::tryCreateMaskedScatterOffset( | ||||||||
| 663 | IntrinsicInst *I, Value *Ptr, IRBuilder<> &Builder) { | ||||||||
| 664 | using namespace PatternMatch; | ||||||||
| 665 | Value *Input = I->getArgOperand(0); | ||||||||
| 666 | Value *Mask = I->getArgOperand(3); | ||||||||
| 667 | Type *InputTy = Input->getType(); | ||||||||
| 668 | Type *MemoryTy = InputTy; | ||||||||
| 669 | |||||||||
| 670 | LLVM_DEBUG(dbgs() << "masked scatters: getelementpointer found. Storing"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked scatters: getelementpointer found. Storing" << " to base + vector of offsets\n"; } } while (false) | ||||||||
| 671 | << " to base + vector of offsets\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked scatters: getelementpointer found. Storing" << " to base + vector of offsets\n"; } } while (false); | ||||||||
| 672 | // If the input has been truncated, try to integrate that trunc into the | ||||||||
| 673 | // scatter instruction (we don't care about alignment here) | ||||||||
| 674 | if (TruncInst *Trunc = dyn_cast<TruncInst>(Input)) { | ||||||||
| 675 | Value *PreTrunc = Trunc->getOperand(0); | ||||||||
| 676 | Type *PreTruncTy = PreTrunc->getType(); | ||||||||
| 677 | if (PreTruncTy->getPrimitiveSizeInBits() == 128) { | ||||||||
| 678 | Input = PreTrunc; | ||||||||
| 679 | InputTy = PreTruncTy; | ||||||||
| 680 | } | ||||||||
| 681 | } | ||||||||
| 682 | bool ExtendInput = false; | ||||||||
| 683 | if (InputTy->getPrimitiveSizeInBits() < 128 && | ||||||||
| 684 | InputTy->isIntOrIntVectorTy()) { | ||||||||
| 685 | // If we can't find a trunc to incorporate into the instruction, create an | ||||||||
| 686 | // implicit one with a zext, so that we can still create a scatter. We know | ||||||||
| 687 | // that the input type is 4x/8x/16x and of type i8/i16/i32, so any type | ||||||||
| 688 | // smaller than 128 bits will divide evenly into a 128bit vector. | ||||||||
| 689 | InputTy = InputTy->getWithNewBitWidth( | ||||||||
| 690 | 128 / cast<FixedVectorType>(InputTy)->getNumElements()); | ||||||||
| 691 | ExtendInput = true; | ||||||||
| 692 | LLVM_DEBUG(dbgs() << "masked scatters: Small input type, will extend:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked scatters: Small input type, will extend:\n" << *Input << "\n"; } } while (false) | ||||||||
| 693 | << *Input << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked scatters: Small input type, will extend:\n" << *Input << "\n"; } } while (false); | ||||||||
| 694 | } | ||||||||
| 695 | if (InputTy->getPrimitiveSizeInBits() != 128) { | ||||||||
| 696 | LLVM_DEBUG(dbgs() << "masked scatters: cannot create scatters for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked scatters: cannot create scatters for " "non-standard input types. Expanding.\n"; } } while (false) | ||||||||
| 697 | "non-standard input types. Expanding.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked scatters: cannot create scatters for " "non-standard input types. Expanding.\n"; } } while (false); | ||||||||
| 698 | return nullptr; | ||||||||
| 699 | } | ||||||||
| 700 | |||||||||
| 701 | Value *Offsets; | ||||||||
| 702 | int Scale; | ||||||||
| 703 | Value *BasePtr = decomposePtr( | ||||||||
| 704 | Ptr, Offsets, Scale, cast<FixedVectorType>(InputTy), MemoryTy, Builder); | ||||||||
| 705 | if (!BasePtr) | ||||||||
| 706 | return nullptr; | ||||||||
| 707 | |||||||||
| 708 | if (ExtendInput) | ||||||||
| 709 | Input = Builder.CreateZExt(Input, InputTy); | ||||||||
| 710 | if (!match(Mask, m_One())) | ||||||||
| 711 | return Builder.CreateIntrinsic( | ||||||||
| 712 | Intrinsic::arm_mve_vstr_scatter_offset_predicated, | ||||||||
| 713 | {BasePtr->getType(), Offsets->getType(), Input->getType(), | ||||||||
| 714 | Mask->getType()}, | ||||||||
| 715 | {BasePtr, Offsets, Input, | ||||||||
| 716 | Builder.getInt32(MemoryTy->getScalarSizeInBits()), | ||||||||
| 717 | Builder.getInt32(Scale), Mask}); | ||||||||
| 718 | else | ||||||||
| 719 | return Builder.CreateIntrinsic( | ||||||||
| 720 | Intrinsic::arm_mve_vstr_scatter_offset, | ||||||||
| 721 | {BasePtr->getType(), Offsets->getType(), Input->getType()}, | ||||||||
| 722 | {BasePtr, Offsets, Input, | ||||||||
| 723 | Builder.getInt32(MemoryTy->getScalarSizeInBits()), | ||||||||
| 724 | Builder.getInt32(Scale)}); | ||||||||
| 725 | } | ||||||||
| 726 | |||||||||
| 727 | Instruction *MVEGatherScatterLowering::tryCreateIncrementingGatScat( | ||||||||
| 728 | IntrinsicInst *I, Value *Ptr, IRBuilder<> &Builder) { | ||||||||
| 729 | FixedVectorType *Ty; | ||||||||
| 730 | if (I->getIntrinsicID() == Intrinsic::masked_gather) | ||||||||
| 731 | Ty = cast<FixedVectorType>(I->getType()); | ||||||||
| 732 | else | ||||||||
| 733 | Ty = cast<FixedVectorType>(I->getArgOperand(0)->getType()); | ||||||||
| 734 | |||||||||
| 735 | // Incrementing gathers only exist for v4i32 | ||||||||
| 736 | if (Ty->getNumElements() != 4 || Ty->getScalarSizeInBits() != 32) | ||||||||
| 737 | return nullptr; | ||||||||
| 738 | // Incrementing gathers are not beneficial outside of a loop | ||||||||
| 739 | Loop *L = LI->getLoopFor(I->getParent()); | ||||||||
| 740 | if (L == nullptr) | ||||||||
| 741 | return nullptr; | ||||||||
| 742 | |||||||||
| 743 | // Decompose the GEP into Base and Offsets | ||||||||
| 744 | GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr); | ||||||||
| 745 | Value *Offsets; | ||||||||
| 746 | Value *BasePtr = decomposeGEP(Offsets, Ty, GEP, Builder); | ||||||||
| 747 | if (!BasePtr) | ||||||||
| 748 | return nullptr; | ||||||||
| 749 | |||||||||
| 750 | LLVM_DEBUG(dbgs() << "masked gathers/scatters: trying to build incrementing "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: trying to build incrementing " "wb gather/scatter\n"; } } while (false) | ||||||||
| 751 | "wb gather/scatter\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: trying to build incrementing " "wb gather/scatter\n"; } } while (false); | ||||||||
| 752 | |||||||||
| 753 | // The gep was in charge of making sure the offsets are scaled correctly | ||||||||
| 754 | // - calculate that factor so it can be applied by hand | ||||||||
| 755 | int TypeScale = | ||||||||
| 756 | computeScale(DL->getTypeSizeInBits(GEP->getOperand(0)->getType()), | ||||||||
| 757 | DL->getTypeSizeInBits(GEP->getType()) / | ||||||||
| 758 | cast<FixedVectorType>(GEP->getType())->getNumElements()); | ||||||||
| 759 | if (TypeScale == -1) | ||||||||
| 760 | return nullptr; | ||||||||
| 761 | |||||||||
| 762 | if (GEP->hasOneUse()) { | ||||||||
| 763 | // Only in this case do we want to build a wb gather, because the wb will | ||||||||
| 764 | // change the phi which does affect other users of the gep (which will still | ||||||||
| 765 | // be using the phi in the old way) | ||||||||
| 766 | if (auto *Load = tryCreateIncrementingWBGatScat(I, BasePtr, Offsets, | ||||||||
| 767 | TypeScale, Builder)) | ||||||||
| 768 | return Load; | ||||||||
| 769 | } | ||||||||
| 770 | |||||||||
| 771 | LLVM_DEBUG(dbgs() << "masked gathers/scatters: trying to build incrementing "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: trying to build incrementing " "non-wb gather/scatter\n"; } } while (false) | ||||||||
| 772 | "non-wb gather/scatter\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: trying to build incrementing " "non-wb gather/scatter\n"; } } while (false); | ||||||||
| 773 | |||||||||
| 774 | std::pair<Value *, int64_t> Add = getVarAndConst(Offsets, TypeScale); | ||||||||
| 775 | if (Add.first == nullptr) | ||||||||
| 776 | return nullptr; | ||||||||
| 777 | Value *OffsetsIncoming = Add.first; | ||||||||
| 778 | int64_t Immediate = Add.second; | ||||||||
| 779 | |||||||||
| 780 | // Make sure the offsets are scaled correctly | ||||||||
| 781 | Instruction *ScaledOffsets = BinaryOperator::Create( | ||||||||
| 782 | Instruction::Shl, OffsetsIncoming, | ||||||||
| 783 | Builder.CreateVectorSplat(Ty->getNumElements(), Builder.getInt32(TypeScale)), | ||||||||
| 784 | "ScaledIndex", I); | ||||||||
| 785 | // Add the base to the offsets | ||||||||
| 786 | OffsetsIncoming = BinaryOperator::Create( | ||||||||
| 787 | Instruction::Add, ScaledOffsets, | ||||||||
| 788 | Builder.CreateVectorSplat( | ||||||||
| 789 | Ty->getNumElements(), | ||||||||
| 790 | Builder.CreatePtrToInt( | ||||||||
| 791 | BasePtr, | ||||||||
| 792 | cast<VectorType>(ScaledOffsets->getType())->getElementType())), | ||||||||
| 793 | "StartIndex", I); | ||||||||
| 794 | |||||||||
| 795 | if (I->getIntrinsicID() == Intrinsic::masked_gather) | ||||||||
| 796 | return tryCreateMaskedGatherBase(I, OffsetsIncoming, Builder, Immediate); | ||||||||
| 797 | else | ||||||||
| 798 | return tryCreateMaskedScatterBase(I, OffsetsIncoming, Builder, Immediate); | ||||||||
| 799 | } | ||||||||
| 800 | |||||||||
| 801 | Instruction *MVEGatherScatterLowering::tryCreateIncrementingWBGatScat( | ||||||||
| 802 | IntrinsicInst *I, Value *BasePtr, Value *Offsets, unsigned TypeScale, | ||||||||
| 803 | IRBuilder<> &Builder) { | ||||||||
| 804 | // Check whether this gather's offset is incremented by a constant - if so, | ||||||||
| 805 | // and the load is of the right type, we can merge this into a QI gather | ||||||||
| 806 | Loop *L = LI->getLoopFor(I->getParent()); | ||||||||
| 807 | // Offsets that are worth merging into this instruction will be incremented | ||||||||
| 808 | // by a constant, thus we're looking for an add of a phi and a constant | ||||||||
| 809 | PHINode *Phi = dyn_cast<PHINode>(Offsets); | ||||||||
| 810 | if (Phi == nullptr || Phi->getNumIncomingValues() != 2 || | ||||||||
| 811 | Phi->getParent() != L->getHeader() || Phi->getNumUses() != 2) | ||||||||
| 812 | // No phi means no IV to write back to; if there is a phi, we expect it | ||||||||
| 813 | // to have exactly two incoming values; the only phis we are interested in | ||||||||
| 814 | // will be loop IV's and have exactly two uses, one in their increment and | ||||||||
| 815 | // one in the gather's gep | ||||||||
| 816 | return nullptr; | ||||||||
| 817 | |||||||||
| 818 | unsigned IncrementIndex = | ||||||||
| 819 | Phi->getIncomingBlock(0) == L->getLoopLatch() ? 0 : 1; | ||||||||
| 820 | // Look through the phi to the phi increment | ||||||||
| 821 | Offsets = Phi->getIncomingValue(IncrementIndex); | ||||||||
| 822 | |||||||||
| 823 | std::pair<Value *, int64_t> Add = getVarAndConst(Offsets, TypeScale); | ||||||||
| 824 | if (Add.first == nullptr) | ||||||||
| 825 | return nullptr; | ||||||||
| 826 | Value *OffsetsIncoming = Add.first; | ||||||||
| 827 | int64_t Immediate = Add.second; | ||||||||
| 828 | if (OffsetsIncoming != Phi) | ||||||||
| 829 | // Then the increment we are looking at is not an increment of the | ||||||||
| 830 | // induction variable, and we don't want to do a writeback | ||||||||
| 831 | return nullptr; | ||||||||
| 832 | |||||||||
| 833 | Builder.SetInsertPoint(&Phi->getIncomingBlock(1 - IncrementIndex)->back()); | ||||||||
| 834 | unsigned NumElems = | ||||||||
| 835 | cast<FixedVectorType>(OffsetsIncoming->getType())->getNumElements(); | ||||||||
| 836 | |||||||||
| 837 | // Make sure the offsets are scaled correctly | ||||||||
| 838 | Instruction *ScaledOffsets = BinaryOperator::Create( | ||||||||
| 839 | Instruction::Shl, Phi->getIncomingValue(1 - IncrementIndex), | ||||||||
| 840 | Builder.CreateVectorSplat(NumElems, Builder.getInt32(TypeScale)), | ||||||||
| 841 | "ScaledIndex", &Phi->getIncomingBlock(1 - IncrementIndex)->back()); | ||||||||
| 842 | // Add the base to the offsets | ||||||||
| 843 | OffsetsIncoming = BinaryOperator::Create( | ||||||||
| 844 | Instruction::Add, ScaledOffsets, | ||||||||
| 845 | Builder.CreateVectorSplat( | ||||||||
| 846 | NumElems, | ||||||||
| 847 | Builder.CreatePtrToInt( | ||||||||
| 848 | BasePtr, | ||||||||
| 849 | cast<VectorType>(ScaledOffsets->getType())->getElementType())), | ||||||||
| 850 | "StartIndex", &Phi->getIncomingBlock(1 - IncrementIndex)->back()); | ||||||||
| 851 | // The gather is pre-incrementing | ||||||||
| 852 | OffsetsIncoming = BinaryOperator::Create( | ||||||||
| 853 | Instruction::Sub, OffsetsIncoming, | ||||||||
| 854 | Builder.CreateVectorSplat(NumElems, Builder.getInt32(Immediate)), | ||||||||
| 855 | "PreIncrementStartIndex", | ||||||||
| 856 | &Phi->getIncomingBlock(1 - IncrementIndex)->back()); | ||||||||
| 857 | Phi->setIncomingValue(1 - IncrementIndex, OffsetsIncoming); | ||||||||
| 858 | |||||||||
| 859 | Builder.SetInsertPoint(I); | ||||||||
| 860 | |||||||||
| 861 | Instruction *EndResult; | ||||||||
| 862 | Instruction *NewInduction; | ||||||||
| 863 | if (I->getIntrinsicID() == Intrinsic::masked_gather) { | ||||||||
| 864 | // Build the incrementing gather | ||||||||
| 865 | Value *Load = tryCreateMaskedGatherBaseWB(I, Phi, Builder, Immediate); | ||||||||
| 866 | // One value to be handed to whoever uses the gather, one is the loop | ||||||||
| 867 | // increment | ||||||||
| 868 | EndResult = ExtractValueInst::Create(Load, 0, "Gather"); | ||||||||
| 869 | NewInduction = ExtractValueInst::Create(Load, 1, "GatherIncrement"); | ||||||||
| 870 | Builder.Insert(EndResult); | ||||||||
| 871 | Builder.Insert(NewInduction); | ||||||||
| 872 | } else { | ||||||||
| 873 | // Build the incrementing scatter | ||||||||
| 874 | EndResult = NewInduction = | ||||||||
| 875 | tryCreateMaskedScatterBaseWB(I, Phi, Builder, Immediate); | ||||||||
| 876 | } | ||||||||
| 877 | Instruction *AddInst = cast<Instruction>(Offsets); | ||||||||
| 878 | AddInst->replaceAllUsesWith(NewInduction); | ||||||||
| 879 | AddInst->eraseFromParent(); | ||||||||
| 880 | Phi->setIncomingValue(IncrementIndex, NewInduction); | ||||||||
| 881 | |||||||||
| 882 | return EndResult; | ||||||||
| 883 | } | ||||||||
| 884 | |||||||||
| 885 | void MVEGatherScatterLowering::pushOutAdd(PHINode *&Phi, | ||||||||
| 886 | Value *OffsSecondOperand, | ||||||||
| 887 | unsigned StartIndex) { | ||||||||
| 888 | LLVM_DEBUG(dbgs() << "masked gathers/scatters: optimising add instruction\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: optimising add instruction\n" ; } } while (false); | ||||||||
| 889 | Instruction *InsertionPoint = | ||||||||
| 890 | &cast<Instruction>(Phi->getIncomingBlock(StartIndex)->back()); | ||||||||
| 891 | // Initialize the phi with a vector that contains a sum of the constants | ||||||||
| 892 | Instruction *NewIndex = BinaryOperator::Create( | ||||||||
| 893 | Instruction::Add, Phi->getIncomingValue(StartIndex), OffsSecondOperand, | ||||||||
| 894 | "PushedOutAdd", InsertionPoint); | ||||||||
| 895 | unsigned IncrementIndex = StartIndex == 0 ? 1 : 0; | ||||||||
| 896 | |||||||||
| 897 | // Order such that start index comes first (this reduces mov's) | ||||||||
| 898 | Phi->addIncoming(NewIndex, Phi->getIncomingBlock(StartIndex)); | ||||||||
| 899 | Phi->addIncoming(Phi->getIncomingValue(IncrementIndex), | ||||||||
| 900 | Phi->getIncomingBlock(IncrementIndex)); | ||||||||
| 901 | Phi->removeIncomingValue(IncrementIndex); | ||||||||
| 902 | Phi->removeIncomingValue(StartIndex); | ||||||||
| 903 | } | ||||||||
| 904 | |||||||||
| 905 | void MVEGatherScatterLowering::pushOutMulShl(unsigned Opcode, PHINode *&Phi, | ||||||||
| 906 | Value *IncrementPerRound, | ||||||||
| 907 | Value *OffsSecondOperand, | ||||||||
| 908 | unsigned LoopIncrement, | ||||||||
| 909 | IRBuilder<> &Builder) { | ||||||||
| 910 | LLVM_DEBUG(dbgs() << "masked gathers/scatters: optimising mul instruction\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: optimising mul instruction\n" ; } } while (false); | ||||||||
| 911 | |||||||||
| 912 | // Create a new scalar add outside of the loop and transform it to a splat | ||||||||
| 913 | // by which loop variable can be incremented | ||||||||
| 914 | Instruction *InsertionPoint = &cast<Instruction>( | ||||||||
| 915 | Phi->getIncomingBlock(LoopIncrement == 1 ? 0 : 1)->back()); | ||||||||
| 916 | |||||||||
| 917 | // Create a new index | ||||||||
| 918 | Value *StartIndex = | ||||||||
| 919 | BinaryOperator::Create((Instruction::BinaryOps)Opcode, | ||||||||
| 920 | Phi->getIncomingValue(LoopIncrement == 1 ? 0 : 1), | ||||||||
| 921 | OffsSecondOperand, "PushedOutMul", InsertionPoint); | ||||||||
| 922 | |||||||||
| 923 | Instruction *Product = | ||||||||
| 924 | BinaryOperator::Create((Instruction::BinaryOps)Opcode, IncrementPerRound, | ||||||||
| 925 | OffsSecondOperand, "Product", InsertionPoint); | ||||||||
| 926 | // Increment NewIndex by Product instead of the multiplication | ||||||||
| 927 | Instruction *NewIncrement = BinaryOperator::Create( | ||||||||
| 928 | Instruction::Add, Phi, Product, "IncrementPushedOutMul", | ||||||||
| 929 | cast<Instruction>(Phi->getIncomingBlock(LoopIncrement)->back()) | ||||||||
| 930 | .getPrevNode()); | ||||||||
| 931 | |||||||||
| 932 | Phi->addIncoming(StartIndex, | ||||||||
| 933 | Phi->getIncomingBlock(LoopIncrement == 1 ? 0 : 1)); | ||||||||
| 934 | Phi->addIncoming(NewIncrement, Phi->getIncomingBlock(LoopIncrement)); | ||||||||
| 935 | Phi->removeIncomingValue((unsigned)0); | ||||||||
| 936 | Phi->removeIncomingValue((unsigned)0); | ||||||||
| 937 | } | ||||||||
| 938 | |||||||||
| 939 | // Check whether all usages of this instruction are as offsets of | ||||||||
| 940 | // gathers/scatters or simple arithmetics only used by gathers/scatters | ||||||||
| 941 | static bool hasAllGatScatUsers(Instruction *I, const DataLayout &DL) { | ||||||||
| 942 | if (I->hasNUses(0)) { | ||||||||
| 943 | return false; | ||||||||
| 944 | } | ||||||||
| 945 | bool Gatscat = true; | ||||||||
| 946 | for (User *U : I->users()) { | ||||||||
| 947 | if (!isa<Instruction>(U)) | ||||||||
| 948 | return false; | ||||||||
| 949 | if (isa<GetElementPtrInst>(U) || | ||||||||
| 950 | isGatherScatter(dyn_cast<IntrinsicInst>(U))) { | ||||||||
| 951 | return Gatscat; | ||||||||
| 952 | } else { | ||||||||
| 953 | unsigned OpCode = cast<Instruction>(U)->getOpcode(); | ||||||||
| 954 | if ((OpCode == Instruction::Add || OpCode == Instruction::Mul || | ||||||||
| 955 | OpCode == Instruction::Shl || | ||||||||
| 956 | isAddLikeOr(cast<Instruction>(U), DL)) && | ||||||||
| 957 | hasAllGatScatUsers(cast<Instruction>(U), DL)) { | ||||||||
| 958 | continue; | ||||||||
| 959 | } | ||||||||
| 960 | return false; | ||||||||
| 961 | } | ||||||||
| 962 | } | ||||||||
| 963 | return Gatscat; | ||||||||
| 964 | } | ||||||||
| 965 | |||||||||
| 966 | bool MVEGatherScatterLowering::optimiseOffsets(Value *Offsets, BasicBlock *BB, | ||||||||
| 967 | LoopInfo *LI) { | ||||||||
| 968 | LLVM_DEBUG(dbgs() << "masked gathers/scatters: trying to optimize: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: trying to optimize: " << *Offsets << "\n"; } } while (false) | ||||||||
| 969 | << *Offsets << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: trying to optimize: " << *Offsets << "\n"; } } while (false); | ||||||||
| 970 | // Optimise the addresses of gathers/scatters by moving invariant | ||||||||
| 971 | // calculations out of the loop | ||||||||
| 972 | if (!isa<Instruction>(Offsets)) | ||||||||
| 973 | return false; | ||||||||
| 974 | Instruction *Offs = cast<Instruction>(Offsets); | ||||||||
| 975 | if (Offs->getOpcode() != Instruction::Add && !isAddLikeOr(Offs, *DL) && | ||||||||
| 976 | Offs->getOpcode() != Instruction::Mul && | ||||||||
| 977 | Offs->getOpcode() != Instruction::Shl) | ||||||||
| 978 | return false; | ||||||||
| 979 | Loop *L = LI->getLoopFor(BB); | ||||||||
| 980 | if (L == nullptr) | ||||||||
| 981 | return false; | ||||||||
| 982 | if (!Offs->hasOneUse()) { | ||||||||
| 983 | if (!hasAllGatScatUsers(Offs, *DL)) | ||||||||
| 984 | return false; | ||||||||
| 985 | } | ||||||||
| 986 | |||||||||
| 987 | // Find out which, if any, operand of the instruction | ||||||||
| 988 | // is a phi node | ||||||||
| 989 | PHINode *Phi; | ||||||||
| 990 | int OffsSecondOp; | ||||||||
| 991 | if (isa<PHINode>(Offs->getOperand(0))) { | ||||||||
| 992 | Phi = cast<PHINode>(Offs->getOperand(0)); | ||||||||
| 993 | OffsSecondOp = 1; | ||||||||
| 994 | } else if (isa<PHINode>(Offs->getOperand(1))) { | ||||||||
| 995 | Phi = cast<PHINode>(Offs->getOperand(1)); | ||||||||
| 996 | OffsSecondOp = 0; | ||||||||
| 997 | } else { | ||||||||
| 998 | bool Changed = false; | ||||||||
| 999 | if (isa<Instruction>(Offs->getOperand(0)) && | ||||||||
| 1000 | L->contains(cast<Instruction>(Offs->getOperand(0)))) | ||||||||
| 1001 | Changed |= optimiseOffsets(Offs->getOperand(0), BB, LI); | ||||||||
| 1002 | if (isa<Instruction>(Offs->getOperand(1)) && | ||||||||
| 1003 | L->contains(cast<Instruction>(Offs->getOperand(1)))) | ||||||||
| 1004 | Changed |= optimiseOffsets(Offs->getOperand(1), BB, LI); | ||||||||
| 1005 | if (!Changed) | ||||||||
| 1006 | return false; | ||||||||
| 1007 | if (isa<PHINode>(Offs->getOperand(0))) { | ||||||||
| 1008 | Phi = cast<PHINode>(Offs->getOperand(0)); | ||||||||
| 1009 | OffsSecondOp = 1; | ||||||||
| 1010 | } else if (isa<PHINode>(Offs->getOperand(1))) { | ||||||||
| 1011 | Phi = cast<PHINode>(Offs->getOperand(1)); | ||||||||
| 1012 | OffsSecondOp = 0; | ||||||||
| 1013 | } else { | ||||||||
| 1014 | return false; | ||||||||
| 1015 | } | ||||||||
| 1016 | } | ||||||||
| 1017 | // A phi node we want to perform this function on should be from the | ||||||||
| 1018 | // loop header. | ||||||||
| 1019 | if (Phi->getParent() != L->getHeader()) | ||||||||
| 1020 | return false; | ||||||||
| 1021 | |||||||||
| 1022 | // We're looking for a simple add recurrence. | ||||||||
| 1023 | BinaryOperator *IncInstruction; | ||||||||
| 1024 | Value *Start, *IncrementPerRound; | ||||||||
| 1025 | if (!matchSimpleRecurrence(Phi, IncInstruction, Start, IncrementPerRound) || | ||||||||
| 1026 | IncInstruction->getOpcode() != Instruction::Add) | ||||||||
| 1027 | return false; | ||||||||
| 1028 | |||||||||
| 1029 | int IncrementingBlock = Phi->getIncomingValue(0) == IncInstruction ? 0 : 1; | ||||||||
| 1030 | |||||||||
| 1031 | // Get the value that is added to/multiplied with the phi | ||||||||
| 1032 | Value *OffsSecondOperand = Offs->getOperand(OffsSecondOp); | ||||||||
| 1033 | |||||||||
| 1034 | if (IncrementPerRound->getType() != OffsSecondOperand->getType() || | ||||||||
| 1035 | !L->isLoopInvariant(OffsSecondOperand)) | ||||||||
| 1036 | // Something has gone wrong, abort | ||||||||
| 1037 | return false; | ||||||||
| 1038 | |||||||||
| 1039 | // Only proceed if the increment per round is a constant or an instruction | ||||||||
| 1040 | // which does not originate from within the loop | ||||||||
| 1041 | if (!isa<Constant>(IncrementPerRound) && | ||||||||
| 1042 | !(isa<Instruction>(IncrementPerRound) && | ||||||||
| 1043 | !L->contains(cast<Instruction>(IncrementPerRound)))) | ||||||||
| 1044 | return false; | ||||||||
| 1045 | |||||||||
| 1046 | // If the phi is not used by anything else, we can just adapt it when | ||||||||
| 1047 | // replacing the instruction; if it is, we'll have to duplicate it | ||||||||
| 1048 | PHINode *NewPhi; | ||||||||
| 1049 | if (Phi->getNumUses() == 2) { | ||||||||
| 1050 | // No other users -> reuse existing phi (One user is the instruction | ||||||||
| 1051 | // we're looking at, the other is the phi increment) | ||||||||
| 1052 | if (IncInstruction->getNumUses() != 1) { | ||||||||
| 1053 | // If the incrementing instruction does have more users than | ||||||||
| 1054 | // our phi, we need to copy it | ||||||||
| 1055 | IncInstruction = BinaryOperator::Create( | ||||||||
| 1056 | Instruction::BinaryOps(IncInstruction->getOpcode()), Phi, | ||||||||
| 1057 | IncrementPerRound, "LoopIncrement", IncInstruction); | ||||||||
| 1058 | Phi->setIncomingValue(IncrementingBlock, IncInstruction); | ||||||||
| 1059 | } | ||||||||
| 1060 | NewPhi = Phi; | ||||||||
| 1061 | } else { | ||||||||
| 1062 | // There are other users -> create a new phi | ||||||||
| 1063 | NewPhi = PHINode::Create(Phi->getType(), 2, "NewPhi", Phi); | ||||||||
| 1064 | // Copy the incoming values of the old phi | ||||||||
| 1065 | NewPhi->addIncoming(Phi->getIncomingValue(IncrementingBlock == 1 ? 0 : 1), | ||||||||
| 1066 | Phi->getIncomingBlock(IncrementingBlock == 1 ? 0 : 1)); | ||||||||
| 1067 | IncInstruction = BinaryOperator::Create( | ||||||||
| 1068 | Instruction::BinaryOps(IncInstruction->getOpcode()), NewPhi, | ||||||||
| 1069 | IncrementPerRound, "LoopIncrement", IncInstruction); | ||||||||
| 1070 | NewPhi->addIncoming(IncInstruction, | ||||||||
| 1071 | Phi->getIncomingBlock(IncrementingBlock)); | ||||||||
| 1072 | IncrementingBlock = 1; | ||||||||
| 1073 | } | ||||||||
| 1074 | |||||||||
| 1075 | IRBuilder<> Builder(BB->getContext()); | ||||||||
| 1076 | Builder.SetInsertPoint(Phi); | ||||||||
| 1077 | Builder.SetCurrentDebugLocation(Offs->getDebugLoc()); | ||||||||
| 1078 | |||||||||
| 1079 | switch (Offs->getOpcode()) { | ||||||||
| 1080 | case Instruction::Add: | ||||||||
| 1081 | case Instruction::Or: | ||||||||
| 1082 | pushOutAdd(NewPhi, OffsSecondOperand, IncrementingBlock == 1 ? 0 : 1); | ||||||||
| 1083 | break; | ||||||||
| 1084 | case Instruction::Mul: | ||||||||
| 1085 | case Instruction::Shl: | ||||||||
| 1086 | pushOutMulShl(Offs->getOpcode(), NewPhi, IncrementPerRound, | ||||||||
| 1087 | OffsSecondOperand, IncrementingBlock, Builder); | ||||||||
| 1088 | break; | ||||||||
| 1089 | default: | ||||||||
| 1090 | return false; | ||||||||
| 1091 | } | ||||||||
| 1092 | LLVM_DEBUG(dbgs() << "masked gathers/scatters: simplified loop variable "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: simplified loop variable " << "add/mul\n"; } } while (false) | ||||||||
| 1093 | << "add/mul\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: simplified loop variable " << "add/mul\n"; } } while (false); | ||||||||
| 1094 | |||||||||
| 1095 | // The instruction has now been "absorbed" into the phi value | ||||||||
| 1096 | Offs->replaceAllUsesWith(NewPhi); | ||||||||
| 1097 | if (Offs->hasNUses(0)) | ||||||||
| 1098 | Offs->eraseFromParent(); | ||||||||
| 1099 | // Clean up the old increment in case it's unused because we built a new | ||||||||
| 1100 | // one | ||||||||
| 1101 | if (IncInstruction->hasNUses(0)) | ||||||||
| 1102 | IncInstruction->eraseFromParent(); | ||||||||
| 1103 | |||||||||
| 1104 | return true; | ||||||||
| 1105 | } | ||||||||
| 1106 | |||||||||
| 1107 | static Value *CheckAndCreateOffsetAdd(Value *X, unsigned ScaleX, Value *Y, | ||||||||
| 1108 | unsigned ScaleY, IRBuilder<> &Builder) { | ||||||||
| 1109 | // Splat the non-vector value to a vector of the given type - if the value is | ||||||||
| 1110 | // a constant (and its value isn't too big), we can even use this opportunity | ||||||||
| 1111 | // to scale it to the size of the vector elements | ||||||||
| 1112 | auto FixSummands = [&Builder](FixedVectorType *&VT, Value *&NonVectorVal) { | ||||||||
| 1113 | ConstantInt *Const; | ||||||||
| 1114 | if ((Const = dyn_cast<ConstantInt>(NonVectorVal)) && | ||||||||
| 1115 | VT->getElementType() != NonVectorVal->getType()) { | ||||||||
| 1116 | unsigned TargetElemSize = VT->getElementType()->getPrimitiveSizeInBits(); | ||||||||
| 1117 | uint64_t N = Const->getZExtValue(); | ||||||||
| 1118 | if (N < (unsigned)(1 << (TargetElemSize - 1))) { | ||||||||
| 1119 | NonVectorVal = Builder.CreateVectorSplat( | ||||||||
| 1120 | VT->getNumElements(), Builder.getIntN(TargetElemSize, N)); | ||||||||
| 1121 | return; | ||||||||
| 1122 | } | ||||||||
| 1123 | } | ||||||||
| 1124 | NonVectorVal = | ||||||||
| 1125 | Builder.CreateVectorSplat(VT->getNumElements(), NonVectorVal); | ||||||||
| 1126 | }; | ||||||||
| 1127 | |||||||||
| 1128 | FixedVectorType *XElType = dyn_cast<FixedVectorType>(X->getType()); | ||||||||
| 1129 | FixedVectorType *YElType = dyn_cast<FixedVectorType>(Y->getType()); | ||||||||
| 1130 | // If one of X, Y is not a vector, we have to splat it in order | ||||||||
| 1131 | // to add the two of them. | ||||||||
| 1132 | if (XElType
| ||||||||
| 1133 | FixSummands(XElType, Y); | ||||||||
| 1134 | YElType = cast<FixedVectorType>(Y->getType()); | ||||||||
| 1135 | } else if (YElType && !XElType) { | ||||||||
| 1136 | FixSummands(YElType, X); | ||||||||
| 1137 | XElType = cast<FixedVectorType>(X->getType()); | ||||||||
| 1138 | } | ||||||||
| 1139 | assert(XElType && YElType && "Unknown vector types")(static_cast <bool> (XElType && YElType && "Unknown vector types") ? void (0) : __assert_fail ("XElType && YElType && \"Unknown vector types\"" , "llvm/lib/Target/ARM/MVEGatherScatterLowering.cpp", 1139, __extension__ __PRETTY_FUNCTION__)); | ||||||||
| 1140 | // Check that the summands are of compatible types | ||||||||
| 1141 | if (XElType != YElType) { | ||||||||
| 1142 | LLVM_DEBUG(dbgs() << "masked gathers/scatters: incompatible gep offsets\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "masked gathers/scatters: incompatible gep offsets\n" ; } } while (false); | ||||||||
| 1143 | return nullptr; | ||||||||
| 1144 | } | ||||||||
| 1145 | |||||||||
| 1146 | if (XElType->getElementType()->getScalarSizeInBits() != 32) { | ||||||||
| 1147 | // Check that by adding the vectors we do not accidentally | ||||||||
| 1148 | // create an overflow | ||||||||
| 1149 | Constant *ConstX = dyn_cast<Constant>(X); | ||||||||
| 1150 | Constant *ConstY = dyn_cast<Constant>(Y); | ||||||||
| 1151 | if (!ConstX
| ||||||||
| 1152 | return nullptr; | ||||||||
| 1153 | unsigned TargetElemSize = 128 / XElType->getNumElements(); | ||||||||
| 1154 | for (unsigned i = 0; i < XElType->getNumElements(); i++) { | ||||||||
| 1155 | ConstantInt *ConstXEl = | ||||||||
| 1156 | dyn_cast<ConstantInt>(ConstX->getAggregateElement(i)); | ||||||||
| 1157 | ConstantInt *ConstYEl = | ||||||||
| 1158 | dyn_cast<ConstantInt>(ConstY->getAggregateElement(i)); | ||||||||
| 1159 | if (!ConstXEl
| ||||||||
| 1160 | ConstXEl->getZExtValue() * ScaleX + | ||||||||
| 1161 | ConstYEl->getZExtValue() * ScaleY >= | ||||||||
| 1162 | (unsigned)(1 << (TargetElemSize - 1))) | ||||||||
| 1163 | return nullptr; | ||||||||
| 1164 | } | ||||||||
| 1165 | } | ||||||||
| 1166 | |||||||||
| 1167 | Value *XScale = Builder.CreateVectorSplat( | ||||||||
| 1168 | XElType->getNumElements(), | ||||||||
| 1169 | Builder.getIntN(XElType->getScalarSizeInBits(), ScaleX)); | ||||||||
| 1170 | Value *YScale = Builder.CreateVectorSplat( | ||||||||
| 1171 | YElType->getNumElements(), | ||||||||
| 1172 | Builder.getIntN(YElType->getScalarSizeInBits(), ScaleY)); | ||||||||
| 1173 | Value *Add = Builder.CreateAdd(Builder.CreateMul(X, XScale), | ||||||||
| 1174 | Builder.CreateMul(Y, YScale)); | ||||||||
| 1175 | |||||||||
| 1176 | if (checkOffsetSize(Add, XElType->getNumElements())) | ||||||||
| 1177 | return Add; | ||||||||
| 1178 | else | ||||||||
| 1179 | return nullptr; | ||||||||
| 1180 | } | ||||||||
| 1181 | |||||||||
| 1182 | Value *MVEGatherScatterLowering::foldGEP(GetElementPtrInst *GEP, | ||||||||
| 1183 | Value *&Offsets, unsigned &Scale, | ||||||||
| 1184 | IRBuilder<> &Builder) { | ||||||||
| 1185 | Value *GEPPtr = GEP->getPointerOperand(); | ||||||||
| 1186 | Offsets = GEP->getOperand(1); | ||||||||
| 1187 | Scale = DL->getTypeAllocSize(GEP->getSourceElementType()); | ||||||||
| 1188 | // We only merge geps with constant offsets, because only for those | ||||||||
| 1189 | // we can make sure that we do not cause an overflow | ||||||||
| 1190 | if (GEP->getNumIndices() != 1 || !isa<Constant>(Offsets)) | ||||||||
| |||||||||
| 1191 | return nullptr; | ||||||||
| 1192 | if (GetElementPtrInst *BaseGEP = dyn_cast<GetElementPtrInst>(GEPPtr)) { | ||||||||
| 1193 | // Merge the two geps into one | ||||||||
| 1194 | Value *BaseBasePtr = foldGEP(BaseGEP, Offsets, Scale, Builder); | ||||||||
| 1195 | if (!BaseBasePtr
| ||||||||
| 1196 | return nullptr; | ||||||||
| 1197 | Offsets = CheckAndCreateOffsetAdd( | ||||||||
| 1198 | Offsets, Scale, GEP->getOperand(1), | ||||||||
| 1199 | DL->getTypeAllocSize(GEP->getSourceElementType()), Builder); | ||||||||
| 1200 | if (Offsets == nullptr) | ||||||||
| 1201 | return nullptr; | ||||||||
| 1202 | Scale = 1; // Scale is always an i8 at this point. | ||||||||
| 1203 | return BaseBasePtr; | ||||||||
| 1204 | } | ||||||||
| 1205 | return GEPPtr; | ||||||||
| 1206 | } | ||||||||
| 1207 | |||||||||
| 1208 | bool MVEGatherScatterLowering::optimiseAddress(Value *Address, BasicBlock *BB, | ||||||||
| 1209 | LoopInfo *LI) { | ||||||||
| 1210 | GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Address); | ||||||||
| 1211 | if (!GEP) | ||||||||
| 1212 | return false; | ||||||||
| 1213 | bool Changed = false; | ||||||||
| 1214 | if (GEP->hasOneUse() && isa<GetElementPtrInst>(GEP->getPointerOperand())) { | ||||||||
| 1215 | IRBuilder<> Builder(GEP->getContext()); | ||||||||
| 1216 | Builder.SetInsertPoint(GEP); | ||||||||
| 1217 | Builder.SetCurrentDebugLocation(GEP->getDebugLoc()); | ||||||||
| 1218 | Value *Offsets; | ||||||||
| 1219 | unsigned Scale; | ||||||||
| 1220 | Value *Base = foldGEP(GEP, Offsets, Scale, Builder); | ||||||||
| 1221 | // We only want to merge the geps if there is a real chance that they can be | ||||||||
| 1222 | // used by an MVE gather; thus the offset has to have the correct size | ||||||||
| 1223 | // (always i32 if it is not of vector type) and the base has to be a | ||||||||
| 1224 | // pointer. | ||||||||
| 1225 | if (Offsets && Base && Base != GEP) { | ||||||||
| 1226 | assert(Scale == 1 && "Expected to fold GEP to a scale of 1")(static_cast <bool> (Scale == 1 && "Expected to fold GEP to a scale of 1" ) ? void (0) : __assert_fail ("Scale == 1 && \"Expected to fold GEP to a scale of 1\"" , "llvm/lib/Target/ARM/MVEGatherScatterLowering.cpp", 1226, __extension__ __PRETTY_FUNCTION__)); | ||||||||
| 1227 | Type *BaseTy = Builder.getInt8PtrTy(); | ||||||||
| 1228 | if (auto *VecTy = dyn_cast<FixedVectorType>(Base->getType())) | ||||||||
| 1229 | BaseTy = FixedVectorType::get(BaseTy, VecTy); | ||||||||
| 1230 | GetElementPtrInst *NewAddress = GetElementPtrInst::Create( | ||||||||
| 1231 | Builder.getInt8Ty(), Builder.CreateBitCast(Base, BaseTy), Offsets, | ||||||||
| 1232 | "gep.merged", GEP); | ||||||||
| 1233 | LLVM_DEBUG(dbgs() << "Folded GEP: " << *GEPdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "Folded GEP: " << *GEP << "\n new : " << *NewAddress << "\n"; } } while (false) | ||||||||
| 1234 | << "\n new : " << *NewAddress << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("arm-mve-gather-scatter-lowering")) { dbgs() << "Folded GEP: " << *GEP << "\n new : " << *NewAddress << "\n"; } } while (false); | ||||||||
| 1235 | GEP->replaceAllUsesWith( | ||||||||
| 1236 | Builder.CreateBitCast(NewAddress, GEP->getType())); | ||||||||
| 1237 | GEP = NewAddress; | ||||||||
| 1238 | Changed = true; | ||||||||
| 1239 | } | ||||||||
| 1240 | } | ||||||||
| 1241 | Changed |= optimiseOffsets(GEP->getOperand(1), GEP->getParent(), LI); | ||||||||
| 1242 | return Changed; | ||||||||
| 1243 | } | ||||||||
| 1244 | |||||||||
| 1245 | bool MVEGatherScatterLowering::runOnFunction(Function &F) { | ||||||||
| 1246 | if (!EnableMaskedGatherScatters) | ||||||||
| 1247 | return false; | ||||||||
| 1248 | auto &TPC = getAnalysis<TargetPassConfig>(); | ||||||||
| 1249 | auto &TM = TPC.getTM<TargetMachine>(); | ||||||||
| 1250 | auto *ST = &TM.getSubtarget<ARMSubtarget>(F); | ||||||||
| 1251 | if (!ST->hasMVEIntegerOps()) | ||||||||
| 1252 | return false; | ||||||||
| 1253 | LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | ||||||||
| 1254 | DL = &F.getParent()->getDataLayout(); | ||||||||
| 1255 | SmallVector<IntrinsicInst *, 4> Gathers; | ||||||||
| 1256 | SmallVector<IntrinsicInst *, 4> Scatters; | ||||||||
| 1257 | |||||||||
| 1258 | bool Changed = false; | ||||||||
| 1259 | |||||||||
| 1260 | for (BasicBlock &BB : F) { | ||||||||
| 1261 | Changed |= SimplifyInstructionsInBlock(&BB); | ||||||||
| 1262 | |||||||||
| 1263 | for (Instruction &I : BB) { | ||||||||
| 1264 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I); | ||||||||
| 1265 | if (II && II->getIntrinsicID() == Intrinsic::masked_gather && | ||||||||
| 1266 | isa<FixedVectorType>(II->getType())) { | ||||||||
| 1267 | Gathers.push_back(II); | ||||||||
| 1268 | Changed |= optimiseAddress(II->getArgOperand(0), II->getParent(), LI); | ||||||||
| 1269 | } else if (II && II->getIntrinsicID() == Intrinsic::masked_scatter && | ||||||||
| 1270 | isa<FixedVectorType>(II->getArgOperand(0)->getType())) { | ||||||||
| 1271 | Scatters.push_back(II); | ||||||||
| 1272 | Changed |= optimiseAddress(II->getArgOperand(1), II->getParent(), LI); | ||||||||
| 1273 | } | ||||||||
| 1274 | } | ||||||||
| 1275 | } | ||||||||
| 1276 | for (unsigned i = 0; i < Gathers.size(); i++) { | ||||||||
| 1277 | IntrinsicInst *I = Gathers[i]; | ||||||||
| 1278 | Instruction *L = lowerGather(I); | ||||||||
| 1279 | if (L == nullptr) | ||||||||
| 1280 | continue; | ||||||||
| 1281 | |||||||||
| 1282 | // Get rid of any now dead instructions | ||||||||
| 1283 | SimplifyInstructionsInBlock(L->getParent()); | ||||||||
| 1284 | Changed = true; | ||||||||
| 1285 | } | ||||||||
| 1286 | |||||||||
| 1287 | for (unsigned i = 0; i < Scatters.size(); i++) { | ||||||||
| 1288 | IntrinsicInst *I = Scatters[i]; | ||||||||
| 1289 | Instruction *S = lowerScatter(I); | ||||||||
| 1290 | if (S == nullptr) | ||||||||
| 1291 | continue; | ||||||||
| 1292 | |||||||||
| 1293 | // Get rid of any now dead instructions | ||||||||
| 1294 | SimplifyInstructionsInBlock(S->getParent()); | ||||||||
| 1295 | Changed = true; | ||||||||
| 1296 | } | ||||||||
| 1297 | return Changed; | ||||||||
| 1298 | } |
| 1 | //===- llvm/IRBuilder.h - Builder for LLVM Instructions ---------*- 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 | // |
| 9 | // This file defines the IRBuilder class, which is used as a convenient way |
| 10 | // to create LLVM instructions with a consistent and simplified interface. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #ifndef LLVM_IR_IRBUILDER_H |
| 15 | #define LLVM_IR_IRBUILDER_H |
| 16 | |
| 17 | #include "llvm-c/Types.h" |
| 18 | #include "llvm/ADT/ArrayRef.h" |
| 19 | #include "llvm/ADT/STLExtras.h" |
| 20 | #include "llvm/ADT/StringRef.h" |
| 21 | #include "llvm/ADT/Twine.h" |
| 22 | #include "llvm/IR/BasicBlock.h" |
| 23 | #include "llvm/IR/Constant.h" |
| 24 | #include "llvm/IR/ConstantFolder.h" |
| 25 | #include "llvm/IR/Constants.h" |
| 26 | #include "llvm/IR/DataLayout.h" |
| 27 | #include "llvm/IR/DebugLoc.h" |
| 28 | #include "llvm/IR/DerivedTypes.h" |
| 29 | #include "llvm/IR/FPEnv.h" |
| 30 | #include "llvm/IR/Function.h" |
| 31 | #include "llvm/IR/GlobalVariable.h" |
| 32 | #include "llvm/IR/InstrTypes.h" |
| 33 | #include "llvm/IR/Instruction.h" |
| 34 | #include "llvm/IR/Instructions.h" |
| 35 | #include "llvm/IR/Intrinsics.h" |
| 36 | #include "llvm/IR/LLVMContext.h" |
| 37 | #include "llvm/IR/Module.h" |
| 38 | #include "llvm/IR/Operator.h" |
| 39 | #include "llvm/IR/Type.h" |
| 40 | #include "llvm/IR/Value.h" |
| 41 | #include "llvm/IR/ValueHandle.h" |
| 42 | #include "llvm/Support/AtomicOrdering.h" |
| 43 | #include "llvm/Support/CBindingWrapping.h" |
| 44 | #include "llvm/Support/Casting.h" |
| 45 | #include <cassert> |
| 46 | #include <cstdint> |
| 47 | #include <functional> |
| 48 | #include <optional> |
| 49 | #include <utility> |
| 50 | |
| 51 | namespace llvm { |
| 52 | |
| 53 | class APInt; |
| 54 | class Use; |
| 55 | |
| 56 | /// This provides the default implementation of the IRBuilder |
| 57 | /// 'InsertHelper' method that is called whenever an instruction is created by |
| 58 | /// IRBuilder and needs to be inserted. |
| 59 | /// |
| 60 | /// By default, this inserts the instruction at the insertion point. |
| 61 | class IRBuilderDefaultInserter { |
| 62 | public: |
| 63 | virtual ~IRBuilderDefaultInserter(); |
| 64 | |
| 65 | virtual void InsertHelper(Instruction *I, const Twine &Name, |
| 66 | BasicBlock *BB, |
| 67 | BasicBlock::iterator InsertPt) const { |
| 68 | if (BB) |
| 69 | I->insertInto(BB, InsertPt); |
| 70 | I->setName(Name); |
| 71 | } |
| 72 | }; |
| 73 | |
| 74 | /// Provides an 'InsertHelper' that calls a user-provided callback after |
| 75 | /// performing the default insertion. |
| 76 | class IRBuilderCallbackInserter : public IRBuilderDefaultInserter { |
| 77 | std::function<void(Instruction *)> Callback; |
| 78 | |
| 79 | public: |
| 80 | ~IRBuilderCallbackInserter() override; |
| 81 | |
| 82 | IRBuilderCallbackInserter(std::function<void(Instruction *)> Callback) |
| 83 | : Callback(std::move(Callback)) {} |
| 84 | |
| 85 | void InsertHelper(Instruction *I, const Twine &Name, |
| 86 | BasicBlock *BB, |
| 87 | BasicBlock::iterator InsertPt) const override { |
| 88 | IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt); |
| 89 | Callback(I); |
| 90 | } |
| 91 | }; |
| 92 | |
| 93 | /// Common base class shared among various IRBuilders. |
| 94 | class IRBuilderBase { |
| 95 | /// Pairs of (metadata kind, MDNode *) that should be added to all newly |
| 96 | /// created instructions, like !dbg metadata. |
| 97 | SmallVector<std::pair<unsigned, MDNode *>, 2> MetadataToCopy; |
| 98 | |
| 99 | /// Add or update the an entry (Kind, MD) to MetadataToCopy, if \p MD is not |
| 100 | /// null. If \p MD is null, remove the entry with \p Kind. |
| 101 | void AddOrRemoveMetadataToCopy(unsigned Kind, MDNode *MD) { |
| 102 | if (!MD) { |
| 103 | erase_if(MetadataToCopy, [Kind](const std::pair<unsigned, MDNode *> &KV) { |
| 104 | return KV.first == Kind; |
| 105 | }); |
| 106 | return; |
| 107 | } |
| 108 | |
| 109 | for (auto &KV : MetadataToCopy) |
| 110 | if (KV.first == Kind) { |
| 111 | KV.second = MD; |
| 112 | return; |
| 113 | } |
| 114 | |
| 115 | MetadataToCopy.emplace_back(Kind, MD); |
| 116 | } |
| 117 | |
| 118 | protected: |
| 119 | BasicBlock *BB; |
| 120 | BasicBlock::iterator InsertPt; |
| 121 | LLVMContext &Context; |
| 122 | const IRBuilderFolder &Folder; |
| 123 | const IRBuilderDefaultInserter &Inserter; |
| 124 | |
| 125 | MDNode *DefaultFPMathTag; |
| 126 | FastMathFlags FMF; |
| 127 | |
| 128 | bool IsFPConstrained = false; |
| 129 | fp::ExceptionBehavior DefaultConstrainedExcept = fp::ebStrict; |
| 130 | RoundingMode DefaultConstrainedRounding = RoundingMode::Dynamic; |
| 131 | |
| 132 | ArrayRef<OperandBundleDef> DefaultOperandBundles; |
| 133 | |
| 134 | public: |
| 135 | IRBuilderBase(LLVMContext &context, const IRBuilderFolder &Folder, |
| 136 | const IRBuilderDefaultInserter &Inserter, MDNode *FPMathTag, |
| 137 | ArrayRef<OperandBundleDef> OpBundles) |
| 138 | : Context(context), Folder(Folder), Inserter(Inserter), |
| 139 | DefaultFPMathTag(FPMathTag), DefaultOperandBundles(OpBundles) { |
| 140 | ClearInsertionPoint(); |
| 141 | } |
| 142 | |
| 143 | /// Insert and return the specified instruction. |
| 144 | template<typename InstTy> |
| 145 | InstTy *Insert(InstTy *I, const Twine &Name = "") const { |
| 146 | Inserter.InsertHelper(I, Name, BB, InsertPt); |
| 147 | AddMetadataToInst(I); |
| 148 | return I; |
| 149 | } |
| 150 | |
| 151 | /// No-op overload to handle constants. |
| 152 | Constant *Insert(Constant *C, const Twine& = "") const { |
| 153 | return C; |
| 154 | } |
| 155 | |
| 156 | Value *Insert(Value *V, const Twine &Name = "") const { |
| 157 | if (Instruction *I = dyn_cast<Instruction>(V)) |
| 158 | return Insert(I, Name); |
| 159 | assert(isa<Constant>(V))(static_cast <bool> (isa<Constant>(V)) ? void (0) : __assert_fail ("isa<Constant>(V)", "llvm/include/llvm/IR/IRBuilder.h" , 159, __extension__ __PRETTY_FUNCTION__)); |
| 160 | return V; |
| 161 | } |
| 162 | |
| 163 | //===--------------------------------------------------------------------===// |
| 164 | // Builder configuration methods |
| 165 | //===--------------------------------------------------------------------===// |
| 166 | |
| 167 | /// Clear the insertion point: created instructions will not be |
| 168 | /// inserted into a block. |
| 169 | void ClearInsertionPoint() { |
| 170 | BB = nullptr; |
| 171 | InsertPt = BasicBlock::iterator(); |
| 172 | } |
| 173 | |
| 174 | BasicBlock *GetInsertBlock() const { return BB; } |
| 175 | BasicBlock::iterator GetInsertPoint() const { return InsertPt; } |
| 176 | LLVMContext &getContext() const { return Context; } |
| 177 | |
| 178 | /// This specifies that created instructions should be appended to the |
| 179 | /// end of the specified block. |
| 180 | void SetInsertPoint(BasicBlock *TheBB) { |
| 181 | BB = TheBB; |
| 182 | InsertPt = BB->end(); |
| 183 | } |
| 184 | |
| 185 | /// This specifies that created instructions should be inserted before |
| 186 | /// the specified instruction. |
| 187 | void SetInsertPoint(Instruction *I) { |
| 188 | BB = I->getParent(); |
| 189 | InsertPt = I->getIterator(); |
| 190 | assert(InsertPt != BB->end() && "Can't read debug loc from end()")(static_cast <bool> (InsertPt != BB->end() && "Can't read debug loc from end()") ? void (0) : __assert_fail ("InsertPt != BB->end() && \"Can't read debug loc from end()\"" , "llvm/include/llvm/IR/IRBuilder.h", 190, __extension__ __PRETTY_FUNCTION__ )); |
| 191 | SetCurrentDebugLocation(I->getDebugLoc()); |
| 192 | } |
| 193 | |
| 194 | /// This specifies that created instructions should be inserted at the |
| 195 | /// specified point. |
| 196 | void SetInsertPoint(BasicBlock *TheBB, BasicBlock::iterator IP) { |
| 197 | BB = TheBB; |
| 198 | InsertPt = IP; |
| 199 | if (IP != TheBB->end()) |
| 200 | SetCurrentDebugLocation(IP->getDebugLoc()); |
| 201 | } |
| 202 | |
| 203 | /// This specifies that created instructions should inserted at the beginning |
| 204 | /// end of the specified function, but after already existing static alloca |
| 205 | /// instructions that are at the start. |
| 206 | void SetInsertPointPastAllocas(Function *F) { |
| 207 | BB = &F->getEntryBlock(); |
| 208 | InsertPt = BB->getFirstNonPHIOrDbgOrAlloca(); |
| 209 | } |
| 210 | |
| 211 | /// Set location information used by debugging information. |
| 212 | void SetCurrentDebugLocation(DebugLoc L) { |
| 213 | AddOrRemoveMetadataToCopy(LLVMContext::MD_dbg, L.getAsMDNode()); |
| 214 | } |
| 215 | |
| 216 | /// Collect metadata with IDs \p MetadataKinds from \p Src which should be |
| 217 | /// added to all created instructions. Entries present in MedataDataToCopy but |
| 218 | /// not on \p Src will be dropped from MetadataToCopy. |
| 219 | void CollectMetadataToCopy(Instruction *Src, |
| 220 | ArrayRef<unsigned> MetadataKinds) { |
| 221 | for (unsigned K : MetadataKinds) |
| 222 | AddOrRemoveMetadataToCopy(K, Src->getMetadata(K)); |
| 223 | } |
| 224 | |
| 225 | /// Get location information used by debugging information. |
| 226 | DebugLoc getCurrentDebugLocation() const; |
| 227 | |
| 228 | /// If this builder has a current debug location, set it on the |
| 229 | /// specified instruction. |
| 230 | void SetInstDebugLocation(Instruction *I) const; |
| 231 | |
| 232 | /// Add all entries in MetadataToCopy to \p I. |
| 233 | void AddMetadataToInst(Instruction *I) const { |
| 234 | for (const auto &KV : MetadataToCopy) |
| 235 | I->setMetadata(KV.first, KV.second); |
| 236 | } |
| 237 | |
| 238 | /// Get the return type of the current function that we're emitting |
| 239 | /// into. |
| 240 | Type *getCurrentFunctionReturnType() const; |
| 241 | |
| 242 | /// InsertPoint - A saved insertion point. |
| 243 | class InsertPoint { |
| 244 | BasicBlock *Block = nullptr; |
| 245 | BasicBlock::iterator Point; |
| 246 | |
| 247 | public: |
| 248 | /// Creates a new insertion point which doesn't point to anything. |
| 249 | InsertPoint() = default; |
| 250 | |
| 251 | /// Creates a new insertion point at the given location. |
| 252 | InsertPoint(BasicBlock *InsertBlock, BasicBlock::iterator InsertPoint) |
| 253 | : Block(InsertBlock), Point(InsertPoint) {} |
| 254 | |
| 255 | /// Returns true if this insert point is set. |
| 256 | bool isSet() const { return (Block != nullptr); } |
| 257 | |
| 258 | BasicBlock *getBlock() const { return Block; } |
| 259 | BasicBlock::iterator getPoint() const { return Point; } |
| 260 | }; |
| 261 | |
| 262 | /// Returns the current insert point. |
| 263 | InsertPoint saveIP() const { |
| 264 | return InsertPoint(GetInsertBlock(), GetInsertPoint()); |
| 265 | } |
| 266 | |
| 267 | /// Returns the current insert point, clearing it in the process. |
| 268 | InsertPoint saveAndClearIP() { |
| 269 | InsertPoint IP(GetInsertBlock(), GetInsertPoint()); |
| 270 | ClearInsertionPoint(); |
| 271 | return IP; |
| 272 | } |
| 273 | |
| 274 | /// Sets the current insert point to a previously-saved location. |
| 275 | void restoreIP(InsertPoint IP) { |
| 276 | if (IP.isSet()) |
| 277 | SetInsertPoint(IP.getBlock(), IP.getPoint()); |
| 278 | else |
| 279 | ClearInsertionPoint(); |
| 280 | } |
| 281 | |
| 282 | /// Get the floating point math metadata being used. |
| 283 | MDNode *getDefaultFPMathTag() const { return DefaultFPMathTag; } |
| 284 | |
| 285 | /// Get the flags to be applied to created floating point ops |
| 286 | FastMathFlags getFastMathFlags() const { return FMF; } |
| 287 | |
| 288 | FastMathFlags &getFastMathFlags() { return FMF; } |
| 289 | |
| 290 | /// Clear the fast-math flags. |
| 291 | void clearFastMathFlags() { FMF.clear(); } |
| 292 | |
| 293 | /// Set the floating point math metadata to be used. |
| 294 | void setDefaultFPMathTag(MDNode *FPMathTag) { DefaultFPMathTag = FPMathTag; } |
| 295 | |
| 296 | /// Set the fast-math flags to be used with generated fp-math operators |
| 297 | void setFastMathFlags(FastMathFlags NewFMF) { FMF = NewFMF; } |
| 298 | |
| 299 | /// Enable/Disable use of constrained floating point math. When |
| 300 | /// enabled the CreateF<op>() calls instead create constrained |
| 301 | /// floating point intrinsic calls. Fast math flags are unaffected |
| 302 | /// by this setting. |
| 303 | void setIsFPConstrained(bool IsCon) { IsFPConstrained = IsCon; } |
| 304 | |
| 305 | /// Query for the use of constrained floating point math |
| 306 | bool getIsFPConstrained() { return IsFPConstrained; } |
| 307 | |
| 308 | /// Set the exception handling to be used with constrained floating point |
| 309 | void setDefaultConstrainedExcept(fp::ExceptionBehavior NewExcept) { |
| 310 | #ifndef NDEBUG |
| 311 | std::optional<StringRef> ExceptStr = |
| 312 | convertExceptionBehaviorToStr(NewExcept); |
| 313 | assert(ExceptStr && "Garbage strict exception behavior!")(static_cast <bool> (ExceptStr && "Garbage strict exception behavior!" ) ? void (0) : __assert_fail ("ExceptStr && \"Garbage strict exception behavior!\"" , "llvm/include/llvm/IR/IRBuilder.h", 313, __extension__ __PRETTY_FUNCTION__ )); |
| 314 | #endif |
| 315 | DefaultConstrainedExcept = NewExcept; |
| 316 | } |
| 317 | |
| 318 | /// Set the rounding mode handling to be used with constrained floating point |
| 319 | void setDefaultConstrainedRounding(RoundingMode NewRounding) { |
| 320 | #ifndef NDEBUG |
| 321 | std::optional<StringRef> RoundingStr = |
| 322 | convertRoundingModeToStr(NewRounding); |
| 323 | assert(RoundingStr && "Garbage strict rounding mode!")(static_cast <bool> (RoundingStr && "Garbage strict rounding mode!" ) ? void (0) : __assert_fail ("RoundingStr && \"Garbage strict rounding mode!\"" , "llvm/include/llvm/IR/IRBuilder.h", 323, __extension__ __PRETTY_FUNCTION__ )); |
| 324 | #endif |
| 325 | DefaultConstrainedRounding = NewRounding; |
| 326 | } |
| 327 | |
| 328 | /// Get the exception handling used with constrained floating point |
| 329 | fp::ExceptionBehavior getDefaultConstrainedExcept() { |
| 330 | return DefaultConstrainedExcept; |
| 331 | } |
| 332 | |
| 333 | /// Get the rounding mode handling used with constrained floating point |
| 334 | RoundingMode getDefaultConstrainedRounding() { |
| 335 | return DefaultConstrainedRounding; |
| 336 | } |
| 337 | |
| 338 | void setConstrainedFPFunctionAttr() { |
| 339 | assert(BB && "Must have a basic block to set any function attributes!")(static_cast <bool> (BB && "Must have a basic block to set any function attributes!" ) ? void (0) : __assert_fail ("BB && \"Must have a basic block to set any function attributes!\"" , "llvm/include/llvm/IR/IRBuilder.h", 339, __extension__ __PRETTY_FUNCTION__ )); |
| 340 | |
| 341 | Function *F = BB->getParent(); |
| 342 | if (!F->hasFnAttribute(Attribute::StrictFP)) { |
| 343 | F->addFnAttr(Attribute::StrictFP); |
| 344 | } |
| 345 | } |
| 346 | |
| 347 | void setConstrainedFPCallAttr(CallBase *I) { |
| 348 | I->addFnAttr(Attribute::StrictFP); |
| 349 | } |
| 350 | |
| 351 | void setDefaultOperandBundles(ArrayRef<OperandBundleDef> OpBundles) { |
| 352 | DefaultOperandBundles = OpBundles; |
| 353 | } |
| 354 | |
| 355 | //===--------------------------------------------------------------------===// |
| 356 | // RAII helpers. |
| 357 | //===--------------------------------------------------------------------===// |
| 358 | |
| 359 | // RAII object that stores the current insertion point and restores it |
| 360 | // when the object is destroyed. This includes the debug location. |
| 361 | class InsertPointGuard { |
| 362 | IRBuilderBase &Builder; |
| 363 | AssertingVH<BasicBlock> Block; |
| 364 | BasicBlock::iterator Point; |
| 365 | DebugLoc DbgLoc; |
| 366 | |
| 367 | public: |
| 368 | InsertPointGuard(IRBuilderBase &B) |
| 369 | : Builder(B), Block(B.GetInsertBlock()), Point(B.GetInsertPoint()), |
| 370 | DbgLoc(B.getCurrentDebugLocation()) {} |
| 371 | |
| 372 | InsertPointGuard(const InsertPointGuard &) = delete; |
| 373 | InsertPointGuard &operator=(const InsertPointGuard &) = delete; |
| 374 | |
| 375 | ~InsertPointGuard() { |
| 376 | Builder.restoreIP(InsertPoint(Block, Point)); |
| 377 | Builder.SetCurrentDebugLocation(DbgLoc); |
| 378 | } |
| 379 | }; |
| 380 | |
| 381 | // RAII object that stores the current fast math settings and restores |
| 382 | // them when the object is destroyed. |
| 383 | class FastMathFlagGuard { |
| 384 | IRBuilderBase &Builder; |
| 385 | FastMathFlags FMF; |
| 386 | MDNode *FPMathTag; |
| 387 | bool IsFPConstrained; |
| 388 | fp::ExceptionBehavior DefaultConstrainedExcept; |
| 389 | RoundingMode DefaultConstrainedRounding; |
| 390 | |
| 391 | public: |
| 392 | FastMathFlagGuard(IRBuilderBase &B) |
| 393 | : Builder(B), FMF(B.FMF), FPMathTag(B.DefaultFPMathTag), |
| 394 | IsFPConstrained(B.IsFPConstrained), |
| 395 | DefaultConstrainedExcept(B.DefaultConstrainedExcept), |
| 396 | DefaultConstrainedRounding(B.DefaultConstrainedRounding) {} |
| 397 | |
| 398 | FastMathFlagGuard(const FastMathFlagGuard &) = delete; |
| 399 | FastMathFlagGuard &operator=(const FastMathFlagGuard &) = delete; |
| 400 | |
| 401 | ~FastMathFlagGuard() { |
| 402 | Builder.FMF = FMF; |
| 403 | Builder.DefaultFPMathTag = FPMathTag; |
| 404 | Builder.IsFPConstrained = IsFPConstrained; |
| 405 | Builder.DefaultConstrainedExcept = DefaultConstrainedExcept; |
| 406 | Builder.DefaultConstrainedRounding = DefaultConstrainedRounding; |
| 407 | } |
| 408 | }; |
| 409 | |
| 410 | // RAII object that stores the current default operand bundles and restores |
| 411 | // them when the object is destroyed. |
| 412 | class OperandBundlesGuard { |
| 413 | IRBuilderBase &Builder; |
| 414 | ArrayRef<OperandBundleDef> DefaultOperandBundles; |
| 415 | |
| 416 | public: |
| 417 | OperandBundlesGuard(IRBuilderBase &B) |
| 418 | : Builder(B), DefaultOperandBundles(B.DefaultOperandBundles) {} |
| 419 | |
| 420 | OperandBundlesGuard(const OperandBundlesGuard &) = delete; |
| 421 | OperandBundlesGuard &operator=(const OperandBundlesGuard &) = delete; |
| 422 | |
| 423 | ~OperandBundlesGuard() { |
| 424 | Builder.DefaultOperandBundles = DefaultOperandBundles; |
| 425 | } |
| 426 | }; |
| 427 | |
| 428 | |
| 429 | //===--------------------------------------------------------------------===// |
| 430 | // Miscellaneous creation methods. |
| 431 | //===--------------------------------------------------------------------===// |
| 432 | |
| 433 | /// Make a new global variable with initializer type i8* |
| 434 | /// |
| 435 | /// Make a new global variable with an initializer that has array of i8 type |
| 436 | /// filled in with the null terminated string value specified. The new global |
| 437 | /// variable will be marked mergable with any others of the same contents. If |
| 438 | /// Name is specified, it is the name of the global variable created. |
| 439 | /// |
| 440 | /// If no module is given via \p M, it is take from the insertion point basic |
| 441 | /// block. |
| 442 | GlobalVariable *CreateGlobalString(StringRef Str, const Twine &Name = "", |
| 443 | unsigned AddressSpace = 0, |
| 444 | Module *M = nullptr); |
| 445 | |
| 446 | /// Get a constant value representing either true or false. |
| 447 | ConstantInt *getInt1(bool V) { |
| 448 | return ConstantInt::get(getInt1Ty(), V); |
| 449 | } |
| 450 | |
| 451 | /// Get the constant value for i1 true. |
| 452 | ConstantInt *getTrue() { |
| 453 | return ConstantInt::getTrue(Context); |
| 454 | } |
| 455 | |
| 456 | /// Get the constant value for i1 false. |
| 457 | ConstantInt *getFalse() { |
| 458 | return ConstantInt::getFalse(Context); |
| 459 | } |
| 460 | |
| 461 | /// Get a constant 8-bit value. |
| 462 | ConstantInt *getInt8(uint8_t C) { |
| 463 | return ConstantInt::get(getInt8Ty(), C); |
| 464 | } |
| 465 | |
| 466 | /// Get a constant 16-bit value. |
| 467 | ConstantInt *getInt16(uint16_t C) { |
| 468 | return ConstantInt::get(getInt16Ty(), C); |
| 469 | } |
| 470 | |
| 471 | /// Get a constant 32-bit value. |
| 472 | ConstantInt *getInt32(uint32_t C) { |
| 473 | return ConstantInt::get(getInt32Ty(), C); |
| 474 | } |
| 475 | |
| 476 | /// Get a constant 64-bit value. |
| 477 | ConstantInt *getInt64(uint64_t C) { |
| 478 | return ConstantInt::get(getInt64Ty(), C); |
| 479 | } |
| 480 | |
| 481 | /// Get a constant N-bit value, zero extended or truncated from |
| 482 | /// a 64-bit value. |
| 483 | ConstantInt *getIntN(unsigned N, uint64_t C) { |
| 484 | return ConstantInt::get(getIntNTy(N), C); |
| 485 | } |
| 486 | |
| 487 | /// Get a constant integer value. |
| 488 | ConstantInt *getInt(const APInt &AI) { |
| 489 | return ConstantInt::get(Context, AI); |
| 490 | } |
| 491 | |
| 492 | //===--------------------------------------------------------------------===// |
| 493 | // Type creation methods |
| 494 | //===--------------------------------------------------------------------===// |
| 495 | |
| 496 | /// Fetch the type representing a single bit |
| 497 | IntegerType *getInt1Ty() { |
| 498 | return Type::getInt1Ty(Context); |
| 499 | } |
| 500 | |
| 501 | /// Fetch the type representing an 8-bit integer. |
| 502 | IntegerType *getInt8Ty() { |
| 503 | return Type::getInt8Ty(Context); |
| 504 | } |
| 505 | |
| 506 | /// Fetch the type representing a 16-bit integer. |
| 507 | IntegerType *getInt16Ty() { |
| 508 | return Type::getInt16Ty(Context); |
| 509 | } |
| 510 | |
| 511 | /// Fetch the type representing a 32-bit integer. |
| 512 | IntegerType *getInt32Ty() { |
| 513 | return Type::getInt32Ty(Context); |
| 514 | } |
| 515 | |
| 516 | /// Fetch the type representing a 64-bit integer. |
| 517 | IntegerType *getInt64Ty() { |
| 518 | return Type::getInt64Ty(Context); |
| 519 | } |
| 520 | |
| 521 | /// Fetch the type representing a 128-bit integer. |
| 522 | IntegerType *getInt128Ty() { return Type::getInt128Ty(Context); } |
| 523 | |
| 524 | /// Fetch the type representing an N-bit integer. |
| 525 | IntegerType *getIntNTy(unsigned N) { |
| 526 | return Type::getIntNTy(Context, N); |
| 527 | } |
| 528 | |
| 529 | /// Fetch the type representing a 16-bit floating point value. |
| 530 | Type *getHalfTy() { |
| 531 | return Type::getHalfTy(Context); |
| 532 | } |
| 533 | |
| 534 | /// Fetch the type representing a 16-bit brain floating point value. |
| 535 | Type *getBFloatTy() { |
| 536 | return Type::getBFloatTy(Context); |
| 537 | } |
| 538 | |
| 539 | /// Fetch the type representing a 32-bit floating point value. |
| 540 | Type *getFloatTy() { |
| 541 | return Type::getFloatTy(Context); |
| 542 | } |
| 543 | |
| 544 | /// Fetch the type representing a 64-bit floating point value. |
| 545 | Type *getDoubleTy() { |
| 546 | return Type::getDoubleTy(Context); |
| 547 | } |
| 548 | |
| 549 | /// Fetch the type representing void. |
| 550 | Type *getVoidTy() { |
| 551 | return Type::getVoidTy(Context); |
| 552 | } |
| 553 | |
| 554 | /// Fetch the type representing a pointer. |
| 555 | PointerType *getPtrTy(unsigned AddrSpace = 0) { |
| 556 | return PointerType::get(Context, AddrSpace); |
| 557 | } |
| 558 | |
| 559 | /// Fetch the type representing a pointer to an 8-bit integer value. |
| 560 | PointerType *getInt8PtrTy(unsigned AddrSpace = 0) { |
| 561 | return Type::getInt8PtrTy(Context, AddrSpace); |
| 562 | } |
| 563 | |
| 564 | /// Fetch the type of an integer with size at least as big as that of a |
| 565 | /// pointer in the given address space. |
| 566 | IntegerType *getIntPtrTy(const DataLayout &DL, unsigned AddrSpace = 0) { |
| 567 | return DL.getIntPtrType(Context, AddrSpace); |
| 568 | } |
| 569 | |
| 570 | //===--------------------------------------------------------------------===// |
| 571 | // Intrinsic creation methods |
| 572 | //===--------------------------------------------------------------------===// |
| 573 | |
| 574 | /// Create and insert a memset to the specified pointer and the |
| 575 | /// specified value. |
| 576 | /// |
| 577 | /// If the pointer isn't an i8*, it will be converted. If a TBAA tag is |
| 578 | /// specified, it will be added to the instruction. Likewise with alias.scope |
| 579 | /// and noalias tags. |
| 580 | CallInst *CreateMemSet(Value *Ptr, Value *Val, uint64_t Size, |
| 581 | MaybeAlign Align, bool isVolatile = false, |
| 582 | MDNode *TBAATag = nullptr, MDNode *ScopeTag = nullptr, |
| 583 | MDNode *NoAliasTag = nullptr) { |
| 584 | return CreateMemSet(Ptr, Val, getInt64(Size), Align, isVolatile, |
| 585 | TBAATag, ScopeTag, NoAliasTag); |
| 586 | } |
| 587 | |
| 588 | CallInst *CreateMemSet(Value *Ptr, Value *Val, Value *Size, MaybeAlign Align, |
| 589 | bool isVolatile = false, MDNode *TBAATag = nullptr, |
| 590 | MDNode *ScopeTag = nullptr, |
| 591 | MDNode *NoAliasTag = nullptr); |
| 592 | |
| 593 | CallInst *CreateMemSetInline(Value *Dst, MaybeAlign DstAlign, Value *Val, |
| 594 | Value *Size, bool IsVolatile = false, |
| 595 | MDNode *TBAATag = nullptr, |
| 596 | MDNode *ScopeTag = nullptr, |
| 597 | MDNode *NoAliasTag = nullptr); |
| 598 | |
| 599 | /// Create and insert an element unordered-atomic memset of the region of |
| 600 | /// memory starting at the given pointer to the given value. |
| 601 | /// |
| 602 | /// If the pointer isn't an i8*, it will be converted. If a TBAA tag is |
| 603 | /// specified, it will be added to the instruction. Likewise with alias.scope |
| 604 | /// and noalias tags. |
| 605 | CallInst *CreateElementUnorderedAtomicMemSet(Value *Ptr, Value *Val, |
| 606 | uint64_t Size, Align Alignment, |
| 607 | uint32_t ElementSize, |
| 608 | MDNode *TBAATag = nullptr, |
| 609 | MDNode *ScopeTag = nullptr, |
| 610 | MDNode *NoAliasTag = nullptr) { |
| 611 | return CreateElementUnorderedAtomicMemSet(Ptr, Val, getInt64(Size), |
| 612 | Align(Alignment), ElementSize, |
| 613 | TBAATag, ScopeTag, NoAliasTag); |
| 614 | } |
| 615 | |
| 616 | CallInst *CreateElementUnorderedAtomicMemSet(Value *Ptr, Value *Val, |
| 617 | Value *Size, Align Alignment, |
| 618 | uint32_t ElementSize, |
| 619 | MDNode *TBAATag = nullptr, |
| 620 | MDNode *ScopeTag = nullptr, |
| 621 | MDNode *NoAliasTag = nullptr); |
| 622 | |
| 623 | /// Create and insert a memcpy between the specified pointers. |
| 624 | /// |
| 625 | /// If the pointers aren't i8*, they will be converted. If a TBAA tag is |
| 626 | /// specified, it will be added to the instruction. Likewise with alias.scope |
| 627 | /// and noalias tags. |
| 628 | CallInst *CreateMemCpy(Value *Dst, MaybeAlign DstAlign, Value *Src, |
| 629 | MaybeAlign SrcAlign, uint64_t Size, |
| 630 | bool isVolatile = false, MDNode *TBAATag = nullptr, |
| 631 | MDNode *TBAAStructTag = nullptr, |
| 632 | MDNode *ScopeTag = nullptr, |
| 633 | MDNode *NoAliasTag = nullptr) { |
| 634 | return CreateMemCpy(Dst, DstAlign, Src, SrcAlign, getInt64(Size), |
| 635 | isVolatile, TBAATag, TBAAStructTag, ScopeTag, |
| 636 | NoAliasTag); |
| 637 | } |
| 638 | |
| 639 | CallInst *CreateMemTransferInst( |
| 640 | Intrinsic::ID IntrID, Value *Dst, MaybeAlign DstAlign, Value *Src, |
| 641 | MaybeAlign SrcAlign, Value *Size, bool isVolatile = false, |
| 642 | MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, |
| 643 | MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr); |
| 644 | |
| 645 | CallInst *CreateMemCpy(Value *Dst, MaybeAlign DstAlign, Value *Src, |
| 646 | MaybeAlign SrcAlign, Value *Size, |
| 647 | bool isVolatile = false, MDNode *TBAATag = nullptr, |
| 648 | MDNode *TBAAStructTag = nullptr, |
| 649 | MDNode *ScopeTag = nullptr, |
| 650 | MDNode *NoAliasTag = nullptr) { |
| 651 | return CreateMemTransferInst(Intrinsic::memcpy, Dst, DstAlign, Src, |
| 652 | SrcAlign, Size, isVolatile, TBAATag, |
| 653 | TBAAStructTag, ScopeTag, NoAliasTag); |
| 654 | } |
| 655 | |
| 656 | CallInst * |
| 657 | CreateMemCpyInline(Value *Dst, MaybeAlign DstAlign, Value *Src, |
| 658 | MaybeAlign SrcAlign, Value *Size, bool IsVolatile = false, |
| 659 | MDNode *TBAATag = nullptr, MDNode *TBAAStructTag = nullptr, |
| 660 | MDNode *ScopeTag = nullptr, MDNode *NoAliasTag = nullptr); |
| 661 | |
| 662 | /// Create and insert an element unordered-atomic memcpy between the |
| 663 | /// specified pointers. |
| 664 | /// |
| 665 | /// DstAlign/SrcAlign are the alignments of the Dst/Src pointers, respectively. |
| 666 | /// |
| 667 | /// If the pointers aren't i8*, they will be converted. If a TBAA tag is |
| 668 | /// specified, it will be added to the instruction. Likewise with alias.scope |
| 669 | /// and noalias tags. |
| 670 | CallInst *CreateElementUnorderedAtomicMemCpy( |
| 671 | Value *Dst, Align DstAlign, Value *Src, Align SrcAlign, Value *Size, |
| 672 | uint32_t ElementSize, MDNode *TBAATag = nullptr, |
| 673 | MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, |
| 674 | MDNode *NoAliasTag = nullptr); |
| 675 | |
| 676 | CallInst *CreateMemMove(Value *Dst, MaybeAlign DstAlign, Value *Src, |
| 677 | MaybeAlign SrcAlign, uint64_t Size, |
| 678 | bool isVolatile = false, MDNode *TBAATag = nullptr, |
| 679 | MDNode *ScopeTag = nullptr, |
| 680 | MDNode *NoAliasTag = nullptr) { |
| 681 | return CreateMemMove(Dst, DstAlign, Src, SrcAlign, getInt64(Size), |
| 682 | isVolatile, TBAATag, ScopeTag, NoAliasTag); |
| 683 | } |
| 684 | |
| 685 | CallInst *CreateMemMove(Value *Dst, MaybeAlign DstAlign, Value *Src, |
| 686 | MaybeAlign SrcAlign, Value *Size, |
| 687 | bool isVolatile = false, MDNode *TBAATag = nullptr, |
| 688 | MDNode *ScopeTag = nullptr, |
| 689 | MDNode *NoAliasTag = nullptr); |
| 690 | |
| 691 | /// \brief Create and insert an element unordered-atomic memmove between the |
| 692 | /// specified pointers. |
| 693 | /// |
| 694 | /// DstAlign/SrcAlign are the alignments of the Dst/Src pointers, |
| 695 | /// respectively. |
| 696 | /// |
| 697 | /// If the pointers aren't i8*, they will be converted. If a TBAA tag is |
| 698 | /// specified, it will be added to the instruction. Likewise with alias.scope |
| 699 | /// and noalias tags. |
| 700 | CallInst *CreateElementUnorderedAtomicMemMove( |
| 701 | Value *Dst, Align DstAlign, Value *Src, Align SrcAlign, Value *Size, |
| 702 | uint32_t ElementSize, MDNode *TBAATag = nullptr, |
| 703 | MDNode *TBAAStructTag = nullptr, MDNode *ScopeTag = nullptr, |
| 704 | MDNode *NoAliasTag = nullptr); |
| 705 | |
| 706 | private: |
| 707 | CallInst *getReductionIntrinsic(Intrinsic::ID ID, Value *Src); |
| 708 | |
| 709 | public: |
| 710 | /// Create a sequential vector fadd reduction intrinsic of the source vector. |
| 711 | /// The first parameter is a scalar accumulator value. An unordered reduction |
| 712 | /// can be created by adding the reassoc fast-math flag to the resulting |
| 713 | /// sequential reduction. |
| 714 | CallInst *CreateFAddReduce(Value *Acc, Value *Src); |
| 715 | |
| 716 | /// Create a sequential vector fmul reduction intrinsic of the source vector. |
| 717 | /// The first parameter is a scalar accumulator value. An unordered reduction |
| 718 | /// can be created by adding the reassoc fast-math flag to the resulting |
| 719 | /// sequential reduction. |
| 720 | CallInst *CreateFMulReduce(Value *Acc, Value *Src); |
| 721 | |
| 722 | /// Create a vector int add reduction intrinsic of the source vector. |
| 723 | CallInst *CreateAddReduce(Value *Src); |
| 724 | |
| 725 | /// Create a vector int mul reduction intrinsic of the source vector. |
| 726 | CallInst *CreateMulReduce(Value *Src); |
| 727 | |
| 728 | /// Create a vector int AND reduction intrinsic of the source vector. |
| 729 | CallInst *CreateAndReduce(Value *Src); |
| 730 | |
| 731 | /// Create a vector int OR reduction intrinsic of the source vector. |
| 732 | CallInst *CreateOrReduce(Value *Src); |
| 733 | |
| 734 | /// Create a vector int XOR reduction intrinsic of the source vector. |
| 735 | CallInst *CreateXorReduce(Value *Src); |
| 736 | |
| 737 | /// Create a vector integer max reduction intrinsic of the source |
| 738 | /// vector. |
| 739 | CallInst *CreateIntMaxReduce(Value *Src, bool IsSigned = false); |
| 740 | |
| 741 | /// Create a vector integer min reduction intrinsic of the source |
| 742 | /// vector. |
| 743 | CallInst *CreateIntMinReduce(Value *Src, bool IsSigned = false); |
| 744 | |
| 745 | /// Create a vector float max reduction intrinsic of the source |
| 746 | /// vector. |
| 747 | CallInst *CreateFPMaxReduce(Value *Src); |
| 748 | |
| 749 | /// Create a vector float min reduction intrinsic of the source |
| 750 | /// vector. |
| 751 | CallInst *CreateFPMinReduce(Value *Src); |
| 752 | |
| 753 | /// Create a lifetime.start intrinsic. |
| 754 | /// |
| 755 | /// If the pointer isn't i8* it will be converted. |
| 756 | CallInst *CreateLifetimeStart(Value *Ptr, ConstantInt *Size = nullptr); |
| 757 | |
| 758 | /// Create a lifetime.end intrinsic. |
| 759 | /// |
| 760 | /// If the pointer isn't i8* it will be converted. |
| 761 | CallInst *CreateLifetimeEnd(Value *Ptr, ConstantInt *Size = nullptr); |
| 762 | |
| 763 | /// Create a call to invariant.start intrinsic. |
| 764 | /// |
| 765 | /// If the pointer isn't i8* it will be converted. |
| 766 | CallInst *CreateInvariantStart(Value *Ptr, ConstantInt *Size = nullptr); |
| 767 | |
| 768 | /// Create a call to llvm.threadlocal.address intrinsic. |
| 769 | CallInst *CreateThreadLocalAddress(Value *Ptr); |
| 770 | |
| 771 | /// Create a call to Masked Load intrinsic |
| 772 | CallInst *CreateMaskedLoad(Type *Ty, Value *Ptr, Align Alignment, Value *Mask, |
| 773 | Value *PassThru = nullptr, const Twine &Name = ""); |
| 774 | |
| 775 | /// Create a call to Masked Store intrinsic |
| 776 | CallInst *CreateMaskedStore(Value *Val, Value *Ptr, Align Alignment, |
| 777 | Value *Mask); |
| 778 | |
| 779 | /// Create a call to Masked Gather intrinsic |
| 780 | CallInst *CreateMaskedGather(Type *Ty, Value *Ptrs, Align Alignment, |
| 781 | Value *Mask = nullptr, Value *PassThru = nullptr, |
| 782 | const Twine &Name = ""); |
| 783 | |
| 784 | /// Create a call to Masked Scatter intrinsic |
| 785 | CallInst *CreateMaskedScatter(Value *Val, Value *Ptrs, Align Alignment, |
| 786 | Value *Mask = nullptr); |
| 787 | |
| 788 | /// Create a call to Masked Expand Load intrinsic |
| 789 | CallInst *CreateMaskedExpandLoad(Type *Ty, Value *Ptr, Value *Mask = nullptr, |
| 790 | Value *PassThru = nullptr, |
| 791 | const Twine &Name = ""); |
| 792 | |
| 793 | /// Create a call to Masked Compress Store intrinsic |
| 794 | CallInst *CreateMaskedCompressStore(Value *Val, Value *Ptr, |
| 795 | Value *Mask = nullptr); |
| 796 | |
| 797 | /// Create an assume intrinsic call that allows the optimizer to |
| 798 | /// assume that the provided condition will be true. |
| 799 | /// |
| 800 | /// The optional argument \p OpBundles specifies operand bundles that are |
| 801 | /// added to the call instruction. |
| 802 | CallInst * |
| 803 | CreateAssumption(Value *Cond, |
| 804 | ArrayRef<OperandBundleDef> OpBundles = std::nullopt); |
| 805 | |
| 806 | /// Create a llvm.experimental.noalias.scope.decl intrinsic call. |
| 807 | Instruction *CreateNoAliasScopeDeclaration(Value *Scope); |
| 808 | Instruction *CreateNoAliasScopeDeclaration(MDNode *ScopeTag) { |
| 809 | return CreateNoAliasScopeDeclaration( |
| 810 | MetadataAsValue::get(Context, ScopeTag)); |
| 811 | } |
| 812 | |
| 813 | /// Create a call to the experimental.gc.statepoint intrinsic to |
| 814 | /// start a new statepoint sequence. |
| 815 | CallInst *CreateGCStatepointCall(uint64_t ID, uint32_t NumPatchBytes, |
| 816 | FunctionCallee ActualCallee, |
| 817 | ArrayRef<Value *> CallArgs, |
| 818 | std::optional<ArrayRef<Value *>> DeoptArgs, |
| 819 | ArrayRef<Value *> GCArgs, |
| 820 | const Twine &Name = ""); |
| 821 | |
| 822 | /// Create a call to the experimental.gc.statepoint intrinsic to |
| 823 | /// start a new statepoint sequence. |
| 824 | CallInst *CreateGCStatepointCall(uint64_t ID, uint32_t NumPatchBytes, |
| 825 | FunctionCallee ActualCallee, uint32_t Flags, |
| 826 | ArrayRef<Value *> CallArgs, |
| 827 | std::optional<ArrayRef<Use>> TransitionArgs, |
| 828 | std::optional<ArrayRef<Use>> DeoptArgs, |
| 829 | ArrayRef<Value *> GCArgs, |
| 830 | const Twine &Name = ""); |
| 831 | |
| 832 | /// Conveninence function for the common case when CallArgs are filled |
| 833 | /// in using ArrayRef(CS.arg_begin(), CS.arg_end()); Use needs to be |
| 834 | /// .get()'ed to get the Value pointer. |
| 835 | CallInst *CreateGCStatepointCall(uint64_t ID, uint32_t NumPatchBytes, |
| 836 | FunctionCallee ActualCallee, |
| 837 | ArrayRef<Use> CallArgs, |
| 838 | std::optional<ArrayRef<Value *>> DeoptArgs, |
| 839 | ArrayRef<Value *> GCArgs, |
| 840 | const Twine &Name = ""); |
| 841 | |
| 842 | /// Create an invoke to the experimental.gc.statepoint intrinsic to |
| 843 | /// start a new statepoint sequence. |
| 844 | InvokeInst * |
| 845 | CreateGCStatepointInvoke(uint64_t ID, uint32_t NumPatchBytes, |
| 846 | FunctionCallee ActualInvokee, BasicBlock *NormalDest, |
| 847 | BasicBlock *UnwindDest, ArrayRef<Value *> InvokeArgs, |
| 848 | std::optional<ArrayRef<Value *>> DeoptArgs, |
| 849 | ArrayRef<Value *> GCArgs, const Twine &Name = ""); |
| 850 | |
| 851 | /// Create an invoke to the experimental.gc.statepoint intrinsic to |
| 852 | /// start a new statepoint sequence. |
| 853 | InvokeInst *CreateGCStatepointInvoke( |
| 854 | uint64_t ID, uint32_t NumPatchBytes, FunctionCallee ActualInvokee, |
| 855 | BasicBlock *NormalDest, BasicBlock *UnwindDest, uint32_t Flags, |
| 856 | ArrayRef<Value *> InvokeArgs, std::optional<ArrayRef<Use>> TransitionArgs, |
| 857 | std::optional<ArrayRef<Use>> DeoptArgs, ArrayRef<Value *> GCArgs, |
| 858 | const Twine &Name = ""); |
| 859 | |
| 860 | // Convenience function for the common case when CallArgs are filled in using |
| 861 | // ArrayRef(CS.arg_begin(), CS.arg_end()); Use needs to be .get()'ed to |
| 862 | // get the Value *. |
| 863 | InvokeInst * |
| 864 | CreateGCStatepointInvoke(uint64_t ID, uint32_t NumPatchBytes, |
| 865 | FunctionCallee ActualInvokee, BasicBlock *NormalDest, |
| 866 | BasicBlock *UnwindDest, ArrayRef<Use> InvokeArgs, |
| 867 | std::optional<ArrayRef<Value *>> DeoptArgs, |
| 868 | ArrayRef<Value *> GCArgs, const Twine &Name = ""); |
| 869 | |
| 870 | /// Create a call to the experimental.gc.result intrinsic to extract |
| 871 | /// the result from a call wrapped in a statepoint. |
| 872 | CallInst *CreateGCResult(Instruction *Statepoint, |
| 873 | Type *ResultType, |
| 874 | const Twine &Name = ""); |
| 875 | |
| 876 | /// Create a call to the experimental.gc.relocate intrinsics to |
| 877 | /// project the relocated value of one pointer from the statepoint. |
| 878 | CallInst *CreateGCRelocate(Instruction *Statepoint, |
| 879 | int BaseOffset, |
| 880 | int DerivedOffset, |
| 881 | Type *ResultType, |
| 882 | const Twine &Name = ""); |
| 883 | |
| 884 | /// Create a call to the experimental.gc.pointer.base intrinsic to get the |
| 885 | /// base pointer for the specified derived pointer. |
| 886 | CallInst *CreateGCGetPointerBase(Value *DerivedPtr, const Twine &Name = ""); |
| 887 | |
| 888 | /// Create a call to the experimental.gc.get.pointer.offset intrinsic to get |
| 889 | /// the offset of the specified derived pointer from its base. |
| 890 | CallInst *CreateGCGetPointerOffset(Value *DerivedPtr, const Twine &Name = ""); |
| 891 | |
| 892 | /// Create a call to llvm.vscale, multiplied by \p Scaling. The type of VScale |
| 893 | /// will be the same type as that of \p Scaling. |
| 894 | Value *CreateVScale(Constant *Scaling, const Twine &Name = ""); |
| 895 | |
| 896 | /// Create an expression which evaluates to the number of elements in \p EC |
| 897 | /// at runtime. |
| 898 | Value *CreateElementCount(Type *DstType, ElementCount EC); |
| 899 | |
| 900 | /// Create an expression which evaluates to the number of units in \p Size |
| 901 | /// at runtime. This works for both units of bits and bytes. |
| 902 | Value *CreateTypeSize(Type *DstType, TypeSize Size); |
| 903 | |
| 904 | /// Creates a vector of type \p DstType with the linear sequence <0, 1, ...> |
| 905 | Value *CreateStepVector(Type *DstType, const Twine &Name = ""); |
| 906 | |
| 907 | /// Create a call to intrinsic \p ID with 1 operand which is mangled on its |
| 908 | /// type. |
| 909 | CallInst *CreateUnaryIntrinsic(Intrinsic::ID ID, Value *V, |
| 910 | Instruction *FMFSource = nullptr, |
| 911 | const Twine &Name = ""); |
| 912 | |
| 913 | /// Create a call to intrinsic \p ID with 2 operands which is mangled on the |
| 914 | /// first type. |
| 915 | CallInst *CreateBinaryIntrinsic(Intrinsic::ID ID, Value *LHS, Value *RHS, |
| 916 | Instruction *FMFSource = nullptr, |
| 917 | const Twine &Name = ""); |
| 918 | |
| 919 | /// Create a call to intrinsic \p ID with \p Args, mangled using \p Types. If |
| 920 | /// \p FMFSource is provided, copy fast-math-flags from that instruction to |
| 921 | /// the intrinsic. |
| 922 | CallInst *CreateIntrinsic(Intrinsic::ID ID, ArrayRef<Type *> Types, |
| 923 | ArrayRef<Value *> Args, |
| 924 | Instruction *FMFSource = nullptr, |
| 925 | const Twine &Name = ""); |
| 926 | |
| 927 | /// Create a call to intrinsic \p ID with \p RetTy and \p Args. If |
| 928 | /// \p FMFSource is provided, copy fast-math-flags from that instruction to |
| 929 | /// the intrinsic. |
| 930 | CallInst *CreateIntrinsic(Type *RetTy, Intrinsic::ID ID, |
| 931 | ArrayRef<Value *> Args, |
| 932 | Instruction *FMFSource = nullptr, |
| 933 | const Twine &Name = ""); |
| 934 | |
| 935 | /// Create call to the minnum intrinsic. |
| 936 | CallInst *CreateMinNum(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 937 | return CreateBinaryIntrinsic(Intrinsic::minnum, LHS, RHS, nullptr, Name); |
| 938 | } |
| 939 | |
| 940 | /// Create call to the maxnum intrinsic. |
| 941 | CallInst *CreateMaxNum(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 942 | return CreateBinaryIntrinsic(Intrinsic::maxnum, LHS, RHS, nullptr, Name); |
| 943 | } |
| 944 | |
| 945 | /// Create call to the minimum intrinsic. |
| 946 | CallInst *CreateMinimum(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 947 | return CreateBinaryIntrinsic(Intrinsic::minimum, LHS, RHS, nullptr, Name); |
| 948 | } |
| 949 | |
| 950 | /// Create call to the maximum intrinsic. |
| 951 | CallInst *CreateMaximum(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 952 | return CreateBinaryIntrinsic(Intrinsic::maximum, LHS, RHS, nullptr, Name); |
| 953 | } |
| 954 | |
| 955 | /// Create call to the copysign intrinsic. |
| 956 | CallInst *CreateCopySign(Value *LHS, Value *RHS, |
| 957 | Instruction *FMFSource = nullptr, |
| 958 | const Twine &Name = "") { |
| 959 | return CreateBinaryIntrinsic(Intrinsic::copysign, LHS, RHS, FMFSource, |
| 960 | Name); |
| 961 | } |
| 962 | |
| 963 | /// Create a call to the arithmetic_fence intrinsic. |
| 964 | CallInst *CreateArithmeticFence(Value *Val, Type *DstType, |
| 965 | const Twine &Name = "") { |
| 966 | return CreateIntrinsic(Intrinsic::arithmetic_fence, DstType, Val, nullptr, |
| 967 | Name); |
| 968 | } |
| 969 | |
| 970 | /// Create a call to the vector.extract intrinsic. |
| 971 | CallInst *CreateExtractVector(Type *DstType, Value *SrcVec, Value *Idx, |
| 972 | const Twine &Name = "") { |
| 973 | return CreateIntrinsic(Intrinsic::vector_extract, |
| 974 | {DstType, SrcVec->getType()}, {SrcVec, Idx}, nullptr, |
| 975 | Name); |
| 976 | } |
| 977 | |
| 978 | /// Create a call to the vector.insert intrinsic. |
| 979 | CallInst *CreateInsertVector(Type *DstType, Value *SrcVec, Value *SubVec, |
| 980 | Value *Idx, const Twine &Name = "") { |
| 981 | return CreateIntrinsic(Intrinsic::vector_insert, |
| 982 | {DstType, SubVec->getType()}, {SrcVec, SubVec, Idx}, |
| 983 | nullptr, Name); |
| 984 | } |
| 985 | |
| 986 | private: |
| 987 | /// Create a call to a masked intrinsic with given Id. |
| 988 | CallInst *CreateMaskedIntrinsic(Intrinsic::ID Id, ArrayRef<Value *> Ops, |
| 989 | ArrayRef<Type *> OverloadedTypes, |
| 990 | const Twine &Name = ""); |
| 991 | |
| 992 | Value *getCastedInt8PtrValue(Value *Ptr); |
| 993 | |
| 994 | //===--------------------------------------------------------------------===// |
| 995 | // Instruction creation methods: Terminators |
| 996 | //===--------------------------------------------------------------------===// |
| 997 | |
| 998 | private: |
| 999 | /// Helper to add branch weight and unpredictable metadata onto an |
| 1000 | /// instruction. |
| 1001 | /// \returns The annotated instruction. |
| 1002 | template <typename InstTy> |
| 1003 | InstTy *addBranchMetadata(InstTy *I, MDNode *Weights, MDNode *Unpredictable) { |
| 1004 | if (Weights) |
| 1005 | I->setMetadata(LLVMContext::MD_prof, Weights); |
| 1006 | if (Unpredictable) |
| 1007 | I->setMetadata(LLVMContext::MD_unpredictable, Unpredictable); |
| 1008 | return I; |
| 1009 | } |
| 1010 | |
| 1011 | public: |
| 1012 | /// Create a 'ret void' instruction. |
| 1013 | ReturnInst *CreateRetVoid() { |
| 1014 | return Insert(ReturnInst::Create(Context)); |
| 1015 | } |
| 1016 | |
| 1017 | /// Create a 'ret <val>' instruction. |
| 1018 | ReturnInst *CreateRet(Value *V) { |
| 1019 | return Insert(ReturnInst::Create(Context, V)); |
| 1020 | } |
| 1021 | |
| 1022 | /// Create a sequence of N insertvalue instructions, |
| 1023 | /// with one Value from the retVals array each, that build a aggregate |
| 1024 | /// return value one value at a time, and a ret instruction to return |
| 1025 | /// the resulting aggregate value. |
| 1026 | /// |
| 1027 | /// This is a convenience function for code that uses aggregate return values |
| 1028 | /// as a vehicle for having multiple return values. |
| 1029 | ReturnInst *CreateAggregateRet(Value *const *retVals, unsigned N) { |
| 1030 | Value *V = PoisonValue::get(getCurrentFunctionReturnType()); |
| 1031 | for (unsigned i = 0; i != N; ++i) |
| 1032 | V = CreateInsertValue(V, retVals[i], i, "mrv"); |
| 1033 | return Insert(ReturnInst::Create(Context, V)); |
| 1034 | } |
| 1035 | |
| 1036 | /// Create an unconditional 'br label X' instruction. |
| 1037 | BranchInst *CreateBr(BasicBlock *Dest) { |
| 1038 | return Insert(BranchInst::Create(Dest)); |
| 1039 | } |
| 1040 | |
| 1041 | /// Create a conditional 'br Cond, TrueDest, FalseDest' |
| 1042 | /// instruction. |
| 1043 | BranchInst *CreateCondBr(Value *Cond, BasicBlock *True, BasicBlock *False, |
| 1044 | MDNode *BranchWeights = nullptr, |
| 1045 | MDNode *Unpredictable = nullptr) { |
| 1046 | return Insert(addBranchMetadata(BranchInst::Create(True, False, Cond), |
| 1047 | BranchWeights, Unpredictable)); |
| 1048 | } |
| 1049 | |
| 1050 | /// Create a conditional 'br Cond, TrueDest, FalseDest' |
| 1051 | /// instruction. Copy branch meta data if available. |
| 1052 | BranchInst *CreateCondBr(Value *Cond, BasicBlock *True, BasicBlock *False, |
| 1053 | Instruction *MDSrc) { |
| 1054 | BranchInst *Br = BranchInst::Create(True, False, Cond); |
| 1055 | if (MDSrc) { |
| 1056 | unsigned WL[4] = {LLVMContext::MD_prof, LLVMContext::MD_unpredictable, |
| 1057 | LLVMContext::MD_make_implicit, LLVMContext::MD_dbg}; |
| 1058 | Br->copyMetadata(*MDSrc, WL); |
| 1059 | } |
| 1060 | return Insert(Br); |
| 1061 | } |
| 1062 | |
| 1063 | /// Create a switch instruction with the specified value, default dest, |
| 1064 | /// and with a hint for the number of cases that will be added (for efficient |
| 1065 | /// allocation). |
| 1066 | SwitchInst *CreateSwitch(Value *V, BasicBlock *Dest, unsigned NumCases = 10, |
| 1067 | MDNode *BranchWeights = nullptr, |
| 1068 | MDNode *Unpredictable = nullptr) { |
| 1069 | return Insert(addBranchMetadata(SwitchInst::Create(V, Dest, NumCases), |
| 1070 | BranchWeights, Unpredictable)); |
| 1071 | } |
| 1072 | |
| 1073 | /// Create an indirect branch instruction with the specified address |
| 1074 | /// operand, with an optional hint for the number of destinations that will be |
| 1075 | /// added (for efficient allocation). |
| 1076 | IndirectBrInst *CreateIndirectBr(Value *Addr, unsigned NumDests = 10) { |
| 1077 | return Insert(IndirectBrInst::Create(Addr, NumDests)); |
| 1078 | } |
| 1079 | |
| 1080 | /// Create an invoke instruction. |
| 1081 | InvokeInst *CreateInvoke(FunctionType *Ty, Value *Callee, |
| 1082 | BasicBlock *NormalDest, BasicBlock *UnwindDest, |
| 1083 | ArrayRef<Value *> Args, |
| 1084 | ArrayRef<OperandBundleDef> OpBundles, |
| 1085 | const Twine &Name = "") { |
| 1086 | InvokeInst *II = |
| 1087 | InvokeInst::Create(Ty, Callee, NormalDest, UnwindDest, Args, OpBundles); |
| 1088 | if (IsFPConstrained) |
| 1089 | setConstrainedFPCallAttr(II); |
| 1090 | return Insert(II, Name); |
| 1091 | } |
| 1092 | InvokeInst *CreateInvoke(FunctionType *Ty, Value *Callee, |
| 1093 | BasicBlock *NormalDest, BasicBlock *UnwindDest, |
| 1094 | ArrayRef<Value *> Args = std::nullopt, |
| 1095 | const Twine &Name = "") { |
| 1096 | InvokeInst *II = |
| 1097 | InvokeInst::Create(Ty, Callee, NormalDest, UnwindDest, Args); |
| 1098 | if (IsFPConstrained) |
| 1099 | setConstrainedFPCallAttr(II); |
| 1100 | return Insert(II, Name); |
| 1101 | } |
| 1102 | |
| 1103 | InvokeInst *CreateInvoke(FunctionCallee Callee, BasicBlock *NormalDest, |
| 1104 | BasicBlock *UnwindDest, ArrayRef<Value *> Args, |
| 1105 | ArrayRef<OperandBundleDef> OpBundles, |
| 1106 | const Twine &Name = "") { |
| 1107 | return CreateInvoke(Callee.getFunctionType(), Callee.getCallee(), |
| 1108 | NormalDest, UnwindDest, Args, OpBundles, Name); |
| 1109 | } |
| 1110 | |
| 1111 | InvokeInst *CreateInvoke(FunctionCallee Callee, BasicBlock *NormalDest, |
| 1112 | BasicBlock *UnwindDest, |
| 1113 | ArrayRef<Value *> Args = std::nullopt, |
| 1114 | const Twine &Name = "") { |
| 1115 | return CreateInvoke(Callee.getFunctionType(), Callee.getCallee(), |
| 1116 | NormalDest, UnwindDest, Args, Name); |
| 1117 | } |
| 1118 | |
| 1119 | /// \brief Create a callbr instruction. |
| 1120 | CallBrInst *CreateCallBr(FunctionType *Ty, Value *Callee, |
| 1121 | BasicBlock *DefaultDest, |
| 1122 | ArrayRef<BasicBlock *> IndirectDests, |
| 1123 | ArrayRef<Value *> Args = std::nullopt, |
| 1124 | const Twine &Name = "") { |
| 1125 | return Insert(CallBrInst::Create(Ty, Callee, DefaultDest, IndirectDests, |
| 1126 | Args), Name); |
| 1127 | } |
| 1128 | CallBrInst *CreateCallBr(FunctionType *Ty, Value *Callee, |
| 1129 | BasicBlock *DefaultDest, |
| 1130 | ArrayRef<BasicBlock *> IndirectDests, |
| 1131 | ArrayRef<Value *> Args, |
| 1132 | ArrayRef<OperandBundleDef> OpBundles, |
| 1133 | const Twine &Name = "") { |
| 1134 | return Insert( |
| 1135 | CallBrInst::Create(Ty, Callee, DefaultDest, IndirectDests, Args, |
| 1136 | OpBundles), Name); |
| 1137 | } |
| 1138 | |
| 1139 | CallBrInst *CreateCallBr(FunctionCallee Callee, BasicBlock *DefaultDest, |
| 1140 | ArrayRef<BasicBlock *> IndirectDests, |
| 1141 | ArrayRef<Value *> Args = std::nullopt, |
| 1142 | const Twine &Name = "") { |
| 1143 | return CreateCallBr(Callee.getFunctionType(), Callee.getCallee(), |
| 1144 | DefaultDest, IndirectDests, Args, Name); |
| 1145 | } |
| 1146 | CallBrInst *CreateCallBr(FunctionCallee Callee, BasicBlock *DefaultDest, |
| 1147 | ArrayRef<BasicBlock *> IndirectDests, |
| 1148 | ArrayRef<Value *> Args, |
| 1149 | ArrayRef<OperandBundleDef> OpBundles, |
| 1150 | const Twine &Name = "") { |
| 1151 | return CreateCallBr(Callee.getFunctionType(), Callee.getCallee(), |
| 1152 | DefaultDest, IndirectDests, Args, Name); |
| 1153 | } |
| 1154 | |
| 1155 | ResumeInst *CreateResume(Value *Exn) { |
| 1156 | return Insert(ResumeInst::Create(Exn)); |
| 1157 | } |
| 1158 | |
| 1159 | CleanupReturnInst *CreateCleanupRet(CleanupPadInst *CleanupPad, |
| 1160 | BasicBlock *UnwindBB = nullptr) { |
| 1161 | return Insert(CleanupReturnInst::Create(CleanupPad, UnwindBB)); |
| 1162 | } |
| 1163 | |
| 1164 | CatchSwitchInst *CreateCatchSwitch(Value *ParentPad, BasicBlock *UnwindBB, |
| 1165 | unsigned NumHandlers, |
| 1166 | const Twine &Name = "") { |
| 1167 | return Insert(CatchSwitchInst::Create(ParentPad, UnwindBB, NumHandlers), |
| 1168 | Name); |
| 1169 | } |
| 1170 | |
| 1171 | CatchPadInst *CreateCatchPad(Value *ParentPad, ArrayRef<Value *> Args, |
| 1172 | const Twine &Name = "") { |
| 1173 | return Insert(CatchPadInst::Create(ParentPad, Args), Name); |
| 1174 | } |
| 1175 | |
| 1176 | CleanupPadInst *CreateCleanupPad(Value *ParentPad, |
| 1177 | ArrayRef<Value *> Args = std::nullopt, |
| 1178 | const Twine &Name = "") { |
| 1179 | return Insert(CleanupPadInst::Create(ParentPad, Args), Name); |
| 1180 | } |
| 1181 | |
| 1182 | CatchReturnInst *CreateCatchRet(CatchPadInst *CatchPad, BasicBlock *BB) { |
| 1183 | return Insert(CatchReturnInst::Create(CatchPad, BB)); |
| 1184 | } |
| 1185 | |
| 1186 | UnreachableInst *CreateUnreachable() { |
| 1187 | return Insert(new UnreachableInst(Context)); |
| 1188 | } |
| 1189 | |
| 1190 | //===--------------------------------------------------------------------===// |
| 1191 | // Instruction creation methods: Binary Operators |
| 1192 | //===--------------------------------------------------------------------===// |
| 1193 | private: |
| 1194 | BinaryOperator *CreateInsertNUWNSWBinOp(BinaryOperator::BinaryOps Opc, |
| 1195 | Value *LHS, Value *RHS, |
| 1196 | const Twine &Name, |
| 1197 | bool HasNUW, bool HasNSW) { |
| 1198 | BinaryOperator *BO = Insert(BinaryOperator::Create(Opc, LHS, RHS), Name); |
| 1199 | if (HasNUW) BO->setHasNoUnsignedWrap(); |
| 1200 | if (HasNSW) BO->setHasNoSignedWrap(); |
| 1201 | return BO; |
| 1202 | } |
| 1203 | |
| 1204 | Instruction *setFPAttrs(Instruction *I, MDNode *FPMD, |
| 1205 | FastMathFlags FMF) const { |
| 1206 | if (!FPMD) |
| 1207 | FPMD = DefaultFPMathTag; |
| 1208 | if (FPMD) |
| 1209 | I->setMetadata(LLVMContext::MD_fpmath, FPMD); |
| 1210 | I->setFastMathFlags(FMF); |
| 1211 | return I; |
| 1212 | } |
| 1213 | |
| 1214 | Value *getConstrainedFPRounding(std::optional<RoundingMode> Rounding) { |
| 1215 | RoundingMode UseRounding = DefaultConstrainedRounding; |
| 1216 | |
| 1217 | if (Rounding) |
| 1218 | UseRounding = *Rounding; |
| 1219 | |
| 1220 | std::optional<StringRef> RoundingStr = |
| 1221 | convertRoundingModeToStr(UseRounding); |
| 1222 | assert(RoundingStr && "Garbage strict rounding mode!")(static_cast <bool> (RoundingStr && "Garbage strict rounding mode!" ) ? void (0) : __assert_fail ("RoundingStr && \"Garbage strict rounding mode!\"" , "llvm/include/llvm/IR/IRBuilder.h", 1222, __extension__ __PRETTY_FUNCTION__ )); |
| 1223 | auto *RoundingMDS = MDString::get(Context, *RoundingStr); |
| 1224 | |
| 1225 | return MetadataAsValue::get(Context, RoundingMDS); |
| 1226 | } |
| 1227 | |
| 1228 | Value *getConstrainedFPExcept(std::optional<fp::ExceptionBehavior> Except) { |
| 1229 | std::optional<StringRef> ExceptStr = convertExceptionBehaviorToStr( |
| 1230 | Except.value_or(DefaultConstrainedExcept)); |
| 1231 | assert(ExceptStr && "Garbage strict exception behavior!")(static_cast <bool> (ExceptStr && "Garbage strict exception behavior!" ) ? void (0) : __assert_fail ("ExceptStr && \"Garbage strict exception behavior!\"" , "llvm/include/llvm/IR/IRBuilder.h", 1231, __extension__ __PRETTY_FUNCTION__ )); |
| 1232 | auto *ExceptMDS = MDString::get(Context, *ExceptStr); |
| 1233 | |
| 1234 | return MetadataAsValue::get(Context, ExceptMDS); |
| 1235 | } |
| 1236 | |
| 1237 | Value *getConstrainedFPPredicate(CmpInst::Predicate Predicate) { |
| 1238 | assert(CmpInst::isFPPredicate(Predicate) &&(static_cast <bool> (CmpInst::isFPPredicate(Predicate) && Predicate != CmpInst::FCMP_FALSE && Predicate != CmpInst ::FCMP_TRUE && "Invalid constrained FP comparison predicate!" ) ? void (0) : __assert_fail ("CmpInst::isFPPredicate(Predicate) && Predicate != CmpInst::FCMP_FALSE && Predicate != CmpInst::FCMP_TRUE && \"Invalid constrained FP comparison predicate!\"" , "llvm/include/llvm/IR/IRBuilder.h", 1241, __extension__ __PRETTY_FUNCTION__ )) |
| 1239 | Predicate != CmpInst::FCMP_FALSE &&(static_cast <bool> (CmpInst::isFPPredicate(Predicate) && Predicate != CmpInst::FCMP_FALSE && Predicate != CmpInst ::FCMP_TRUE && "Invalid constrained FP comparison predicate!" ) ? void (0) : __assert_fail ("CmpInst::isFPPredicate(Predicate) && Predicate != CmpInst::FCMP_FALSE && Predicate != CmpInst::FCMP_TRUE && \"Invalid constrained FP comparison predicate!\"" , "llvm/include/llvm/IR/IRBuilder.h", 1241, __extension__ __PRETTY_FUNCTION__ )) |
| 1240 | Predicate != CmpInst::FCMP_TRUE &&(static_cast <bool> (CmpInst::isFPPredicate(Predicate) && Predicate != CmpInst::FCMP_FALSE && Predicate != CmpInst ::FCMP_TRUE && "Invalid constrained FP comparison predicate!" ) ? void (0) : __assert_fail ("CmpInst::isFPPredicate(Predicate) && Predicate != CmpInst::FCMP_FALSE && Predicate != CmpInst::FCMP_TRUE && \"Invalid constrained FP comparison predicate!\"" , "llvm/include/llvm/IR/IRBuilder.h", 1241, __extension__ __PRETTY_FUNCTION__ )) |
| 1241 | "Invalid constrained FP comparison predicate!")(static_cast <bool> (CmpInst::isFPPredicate(Predicate) && Predicate != CmpInst::FCMP_FALSE && Predicate != CmpInst ::FCMP_TRUE && "Invalid constrained FP comparison predicate!" ) ? void (0) : __assert_fail ("CmpInst::isFPPredicate(Predicate) && Predicate != CmpInst::FCMP_FALSE && Predicate != CmpInst::FCMP_TRUE && \"Invalid constrained FP comparison predicate!\"" , "llvm/include/llvm/IR/IRBuilder.h", 1241, __extension__ __PRETTY_FUNCTION__ )); |
| 1242 | |
| 1243 | StringRef PredicateStr = CmpInst::getPredicateName(Predicate); |
| 1244 | auto *PredicateMDS = MDString::get(Context, PredicateStr); |
| 1245 | |
| 1246 | return MetadataAsValue::get(Context, PredicateMDS); |
| 1247 | } |
| 1248 | |
| 1249 | public: |
| 1250 | Value *CreateAdd(Value *LHS, Value *RHS, const Twine &Name = "", |
| 1251 | bool HasNUW = false, bool HasNSW = false) { |
| 1252 | if (Value *V = |
| 1253 | Folder.FoldNoWrapBinOp(Instruction::Add, LHS, RHS, HasNUW, HasNSW)) |
| 1254 | return V; |
| 1255 | return CreateInsertNUWNSWBinOp(Instruction::Add, LHS, RHS, Name, HasNUW, |
| 1256 | HasNSW); |
| 1257 | } |
| 1258 | |
| 1259 | Value *CreateNSWAdd(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 1260 | return CreateAdd(LHS, RHS, Name, false, true); |
| 1261 | } |
| 1262 | |
| 1263 | Value *CreateNUWAdd(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 1264 | return CreateAdd(LHS, RHS, Name, true, false); |
| 1265 | } |
| 1266 | |
| 1267 | Value *CreateSub(Value *LHS, Value *RHS, const Twine &Name = "", |
| 1268 | bool HasNUW = false, bool HasNSW = false) { |
| 1269 | if (Value *V = |
| 1270 | Folder.FoldNoWrapBinOp(Instruction::Sub, LHS, RHS, HasNUW, HasNSW)) |
| 1271 | return V; |
| 1272 | return CreateInsertNUWNSWBinOp(Instruction::Sub, LHS, RHS, Name, HasNUW, |
| 1273 | HasNSW); |
| 1274 | } |
| 1275 | |
| 1276 | Value *CreateNSWSub(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 1277 | return CreateSub(LHS, RHS, Name, false, true); |
| 1278 | } |
| 1279 | |
| 1280 | Value *CreateNUWSub(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 1281 | return CreateSub(LHS, RHS, Name, true, false); |
| 1282 | } |
| 1283 | |
| 1284 | Value *CreateMul(Value *LHS, Value *RHS, const Twine &Name = "", |
| 1285 | bool HasNUW = false, bool HasNSW = false) { |
| 1286 | if (Value *V = |
| 1287 | Folder.FoldNoWrapBinOp(Instruction::Mul, LHS, RHS, HasNUW, HasNSW)) |
| 1288 | return V; |
| 1289 | return CreateInsertNUWNSWBinOp(Instruction::Mul, LHS, RHS, Name, HasNUW, |
| 1290 | HasNSW); |
| 1291 | } |
| 1292 | |
| 1293 | Value *CreateNSWMul(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 1294 | return CreateMul(LHS, RHS, Name, false, true); |
| 1295 | } |
| 1296 | |
| 1297 | Value *CreateNUWMul(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 1298 | return CreateMul(LHS, RHS, Name, true, false); |
| 1299 | } |
| 1300 | |
| 1301 | Value *CreateUDiv(Value *LHS, Value *RHS, const Twine &Name = "", |
| 1302 | bool isExact = false) { |
| 1303 | if (Value *V = Folder.FoldExactBinOp(Instruction::UDiv, LHS, RHS, isExact)) |
| 1304 | return V; |
| 1305 | if (!isExact) |
| 1306 | return Insert(BinaryOperator::CreateUDiv(LHS, RHS), Name); |
| 1307 | return Insert(BinaryOperator::CreateExactUDiv(LHS, RHS), Name); |
| 1308 | } |
| 1309 | |
| 1310 | Value *CreateExactUDiv(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 1311 | return CreateUDiv(LHS, RHS, Name, true); |
| 1312 | } |
| 1313 | |
| 1314 | Value *CreateSDiv(Value *LHS, Value *RHS, const Twine &Name = "", |
| 1315 | bool isExact = false) { |
| 1316 | if (Value *V = Folder.FoldExactBinOp(Instruction::SDiv, LHS, RHS, isExact)) |
| 1317 | return V; |
| 1318 | if (!isExact) |
| 1319 | return Insert(BinaryOperator::CreateSDiv(LHS, RHS), Name); |
| 1320 | return Insert(BinaryOperator::CreateExactSDiv(LHS, RHS), Name); |
| 1321 | } |
| 1322 | |
| 1323 | Value *CreateExactSDiv(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 1324 | return CreateSDiv(LHS, RHS, Name, true); |
| 1325 | } |
| 1326 | |
| 1327 | Value *CreateURem(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 1328 | if (Value *V = Folder.FoldBinOp(Instruction::URem, LHS, RHS)) |
| 1329 | return V; |
| 1330 | return Insert(BinaryOperator::CreateURem(LHS, RHS), Name); |
| 1331 | } |
| 1332 | |
| 1333 | Value *CreateSRem(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 1334 | if (Value *V = Folder.FoldBinOp(Instruction::SRem, LHS, RHS)) |
| 1335 | return V; |
| 1336 | return Insert(BinaryOperator::CreateSRem(LHS, RHS), Name); |
| 1337 | } |
| 1338 | |
| 1339 | Value *CreateShl(Value *LHS, Value *RHS, const Twine &Name = "", |
| 1340 | bool HasNUW = false, bool HasNSW = false) { |
| 1341 | if (Value *V = |
| 1342 | Folder.FoldNoWrapBinOp(Instruction::Shl, LHS, RHS, HasNUW, HasNSW)) |
| 1343 | return V; |
| 1344 | return CreateInsertNUWNSWBinOp(Instruction::Shl, LHS, RHS, Name, |
| 1345 | HasNUW, HasNSW); |
| 1346 | } |
| 1347 | |
| 1348 | Value *CreateShl(Value *LHS, const APInt &RHS, const Twine &Name = "", |
| 1349 | bool HasNUW = false, bool HasNSW = false) { |
| 1350 | return CreateShl(LHS, ConstantInt::get(LHS->getType(), RHS), Name, |
| 1351 | HasNUW, HasNSW); |
| 1352 | } |
| 1353 | |
| 1354 | Value *CreateShl(Value *LHS, uint64_t RHS, const Twine &Name = "", |
| 1355 | bool HasNUW = false, bool HasNSW = false) { |
| 1356 | return CreateShl(LHS, ConstantInt::get(LHS->getType(), RHS), Name, |
| 1357 | HasNUW, HasNSW); |
| 1358 | } |
| 1359 | |
| 1360 | Value *CreateLShr(Value *LHS, Value *RHS, const Twine &Name = "", |
| 1361 | bool isExact = false) { |
| 1362 | if (Value *V = Folder.FoldExactBinOp(Instruction::LShr, LHS, RHS, isExact)) |
| 1363 | return V; |
| 1364 | if (!isExact) |
| 1365 | return Insert(BinaryOperator::CreateLShr(LHS, RHS), Name); |
| 1366 | return Insert(BinaryOperator::CreateExactLShr(LHS, RHS), Name); |
| 1367 | } |
| 1368 | |
| 1369 | Value *CreateLShr(Value *LHS, const APInt &RHS, const Twine &Name = "", |
| 1370 | bool isExact = false) { |
| 1371 | return CreateLShr(LHS, ConstantInt::get(LHS->getType(), RHS), Name,isExact); |
| 1372 | } |
| 1373 | |
| 1374 | Value *CreateLShr(Value *LHS, uint64_t RHS, const Twine &Name = "", |
| 1375 | bool isExact = false) { |
| 1376 | return CreateLShr(LHS, ConstantInt::get(LHS->getType(), RHS), Name,isExact); |
| 1377 | } |
| 1378 | |
| 1379 | Value *CreateAShr(Value *LHS, Value *RHS, const Twine &Name = "", |
| 1380 | bool isExact = false) { |
| 1381 | if (Value *V = Folder.FoldExactBinOp(Instruction::AShr, LHS, RHS, isExact)) |
| 1382 | return V; |
| 1383 | if (!isExact) |
| 1384 | return Insert(BinaryOperator::CreateAShr(LHS, RHS), Name); |
| 1385 | return Insert(BinaryOperator::CreateExactAShr(LHS, RHS), Name); |
| 1386 | } |
| 1387 | |
| 1388 | Value *CreateAShr(Value *LHS, const APInt &RHS, const Twine &Name = "", |
| 1389 | bool isExact = false) { |
| 1390 | return CreateAShr(LHS, ConstantInt::get(LHS->getType(), RHS), Name,isExact); |
| 1391 | } |
| 1392 | |
| 1393 | Value *CreateAShr(Value *LHS, uint64_t RHS, const Twine &Name = "", |
| 1394 | bool isExact = false) { |
| 1395 | return CreateAShr(LHS, ConstantInt::get(LHS->getType(), RHS), Name,isExact); |
| 1396 | } |
| 1397 | |
| 1398 | Value *CreateAnd(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 1399 | if (auto *V = Folder.FoldBinOp(Instruction::And, LHS, RHS)) |
| 1400 | return V; |
| 1401 | return Insert(BinaryOperator::CreateAnd(LHS, RHS), Name); |
| 1402 | } |
| 1403 | |
| 1404 | Value *CreateAnd(Value *LHS, const APInt &RHS, const Twine &Name = "") { |
| 1405 | return CreateAnd(LHS, ConstantInt::get(LHS->getType(), RHS), Name); |
| 1406 | } |
| 1407 | |
| 1408 | Value *CreateAnd(Value *LHS, uint64_t RHS, const Twine &Name = "") { |
| 1409 | return CreateAnd(LHS, ConstantInt::get(LHS->getType(), RHS), Name); |
| 1410 | } |
| 1411 | |
| 1412 | Value *CreateAnd(ArrayRef<Value*> Ops) { |
| 1413 | assert(!Ops.empty())(static_cast <bool> (!Ops.empty()) ? void (0) : __assert_fail ("!Ops.empty()", "llvm/include/llvm/IR/IRBuilder.h", 1413, __extension__ __PRETTY_FUNCTION__)); |
| 1414 | Value *Accum = Ops[0]; |
| 1415 | for (unsigned i = 1; i < Ops.size(); i++) |
| 1416 | Accum = CreateAnd(Accum, Ops[i]); |
| 1417 | return Accum; |
| 1418 | } |
| 1419 | |
| 1420 | Value *CreateOr(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 1421 | if (auto *V = Folder.FoldBinOp(Instruction::Or, LHS, RHS)) |
| 1422 | return V; |
| 1423 | return Insert(BinaryOperator::CreateOr(LHS, RHS), Name); |
| 1424 | } |
| 1425 | |
| 1426 | Value *CreateOr(Value *LHS, const APInt &RHS, const Twine &Name = "") { |
| 1427 | return CreateOr(LHS, ConstantInt::get(LHS->getType(), RHS), Name); |
| 1428 | } |
| 1429 | |
| 1430 | Value *CreateOr(Value *LHS, uint64_t RHS, const Twine &Name = "") { |
| 1431 | return CreateOr(LHS, ConstantInt::get(LHS->getType(), RHS), Name); |
| 1432 | } |
| 1433 | |
| 1434 | Value *CreateOr(ArrayRef<Value*> Ops) { |
| 1435 | assert(!Ops.empty())(static_cast <bool> (!Ops.empty()) ? void (0) : __assert_fail ("!Ops.empty()", "llvm/include/llvm/IR/IRBuilder.h", 1435, __extension__ __PRETTY_FUNCTION__)); |
| 1436 | Value *Accum = Ops[0]; |
| 1437 | for (unsigned i = 1; i < Ops.size(); i++) |
| 1438 | Accum = CreateOr(Accum, Ops[i]); |
| 1439 | return Accum; |
| 1440 | } |
| 1441 | |
| 1442 | Value *CreateXor(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 1443 | if (Value *V = Folder.FoldBinOp(Instruction::Xor, LHS, RHS)) |
| 1444 | return V; |
| 1445 | return Insert(BinaryOperator::CreateXor(LHS, RHS), Name); |
| 1446 | } |
| 1447 | |
| 1448 | Value *CreateXor(Value *LHS, const APInt &RHS, const Twine &Name = "") { |
| 1449 | return CreateXor(LHS, ConstantInt::get(LHS->getType(), RHS), Name); |
| 1450 | } |
| 1451 | |
| 1452 | Value *CreateXor(Value *LHS, uint64_t RHS, const Twine &Name = "") { |
| 1453 | return CreateXor(LHS, ConstantInt::get(LHS->getType(), RHS), Name); |
| 1454 | } |
| 1455 | |
| 1456 | Value *CreateFAdd(Value *L, Value *R, const Twine &Name = "", |
| 1457 | MDNode *FPMD = nullptr) { |
| 1458 | if (IsFPConstrained) |
| 1459 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_fadd, |
| 1460 | L, R, nullptr, Name, FPMD); |
| 1461 | |
| 1462 | if (Value *V = Folder.FoldBinOpFMF(Instruction::FAdd, L, R, FMF)) |
| 1463 | return V; |
| 1464 | Instruction *I = setFPAttrs(BinaryOperator::CreateFAdd(L, R), FPMD, FMF); |
| 1465 | return Insert(I, Name); |
| 1466 | } |
| 1467 | |
| 1468 | /// Copy fast-math-flags from an instruction rather than using the builder's |
| 1469 | /// default FMF. |
| 1470 | Value *CreateFAddFMF(Value *L, Value *R, Instruction *FMFSource, |
| 1471 | const Twine &Name = "") { |
| 1472 | if (IsFPConstrained) |
| 1473 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_fadd, |
| 1474 | L, R, FMFSource, Name); |
| 1475 | |
| 1476 | FastMathFlags FMF = FMFSource->getFastMathFlags(); |
| 1477 | if (Value *V = Folder.FoldBinOpFMF(Instruction::FAdd, L, R, FMF)) |
| 1478 | return V; |
| 1479 | Instruction *I = setFPAttrs(BinaryOperator::CreateFAdd(L, R), nullptr, FMF); |
| 1480 | return Insert(I, Name); |
| 1481 | } |
| 1482 | |
| 1483 | Value *CreateFSub(Value *L, Value *R, const Twine &Name = "", |
| 1484 | MDNode *FPMD = nullptr) { |
| 1485 | if (IsFPConstrained) |
| 1486 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_fsub, |
| 1487 | L, R, nullptr, Name, FPMD); |
| 1488 | |
| 1489 | if (Value *V = Folder.FoldBinOpFMF(Instruction::FSub, L, R, FMF)) |
| 1490 | return V; |
| 1491 | Instruction *I = setFPAttrs(BinaryOperator::CreateFSub(L, R), FPMD, FMF); |
| 1492 | return Insert(I, Name); |
| 1493 | } |
| 1494 | |
| 1495 | /// Copy fast-math-flags from an instruction rather than using the builder's |
| 1496 | /// default FMF. |
| 1497 | Value *CreateFSubFMF(Value *L, Value *R, Instruction *FMFSource, |
| 1498 | const Twine &Name = "") { |
| 1499 | if (IsFPConstrained) |
| 1500 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_fsub, |
| 1501 | L, R, FMFSource, Name); |
| 1502 | |
| 1503 | FastMathFlags FMF = FMFSource->getFastMathFlags(); |
| 1504 | if (Value *V = Folder.FoldBinOpFMF(Instruction::FSub, L, R, FMF)) |
| 1505 | return V; |
| 1506 | Instruction *I = setFPAttrs(BinaryOperator::CreateFSub(L, R), nullptr, FMF); |
| 1507 | return Insert(I, Name); |
| 1508 | } |
| 1509 | |
| 1510 | Value *CreateFMul(Value *L, Value *R, const Twine &Name = "", |
| 1511 | MDNode *FPMD = nullptr) { |
| 1512 | if (IsFPConstrained) |
| 1513 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_fmul, |
| 1514 | L, R, nullptr, Name, FPMD); |
| 1515 | |
| 1516 | if (Value *V = Folder.FoldBinOpFMF(Instruction::FMul, L, R, FMF)) |
| 1517 | return V; |
| 1518 | Instruction *I = setFPAttrs(BinaryOperator::CreateFMul(L, R), FPMD, FMF); |
| 1519 | return Insert(I, Name); |
| 1520 | } |
| 1521 | |
| 1522 | /// Copy fast-math-flags from an instruction rather than using the builder's |
| 1523 | /// default FMF. |
| 1524 | Value *CreateFMulFMF(Value *L, Value *R, Instruction *FMFSource, |
| 1525 | const Twine &Name = "") { |
| 1526 | if (IsFPConstrained) |
| 1527 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_fmul, |
| 1528 | L, R, FMFSource, Name); |
| 1529 | |
| 1530 | FastMathFlags FMF = FMFSource->getFastMathFlags(); |
| 1531 | if (Value *V = Folder.FoldBinOpFMF(Instruction::FMul, L, R, FMF)) |
| 1532 | return V; |
| 1533 | Instruction *I = setFPAttrs(BinaryOperator::CreateFMul(L, R), nullptr, FMF); |
| 1534 | return Insert(I, Name); |
| 1535 | } |
| 1536 | |
| 1537 | Value *CreateFDiv(Value *L, Value *R, const Twine &Name = "", |
| 1538 | MDNode *FPMD = nullptr) { |
| 1539 | if (IsFPConstrained) |
| 1540 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_fdiv, |
| 1541 | L, R, nullptr, Name, FPMD); |
| 1542 | |
| 1543 | if (Value *V = Folder.FoldBinOpFMF(Instruction::FDiv, L, R, FMF)) |
| 1544 | return V; |
| 1545 | Instruction *I = setFPAttrs(BinaryOperator::CreateFDiv(L, R), FPMD, FMF); |
| 1546 | return Insert(I, Name); |
| 1547 | } |
| 1548 | |
| 1549 | /// Copy fast-math-flags from an instruction rather than using the builder's |
| 1550 | /// default FMF. |
| 1551 | Value *CreateFDivFMF(Value *L, Value *R, Instruction *FMFSource, |
| 1552 | const Twine &Name = "") { |
| 1553 | if (IsFPConstrained) |
| 1554 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_fdiv, |
| 1555 | L, R, FMFSource, Name); |
| 1556 | |
| 1557 | FastMathFlags FMF = FMFSource->getFastMathFlags(); |
| 1558 | if (Value *V = Folder.FoldBinOpFMF(Instruction::FDiv, L, R, FMF)) |
| 1559 | return V; |
| 1560 | Instruction *I = setFPAttrs(BinaryOperator::CreateFDiv(L, R), nullptr, FMF); |
| 1561 | return Insert(I, Name); |
| 1562 | } |
| 1563 | |
| 1564 | Value *CreateFRem(Value *L, Value *R, const Twine &Name = "", |
| 1565 | MDNode *FPMD = nullptr) { |
| 1566 | if (IsFPConstrained) |
| 1567 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_frem, |
| 1568 | L, R, nullptr, Name, FPMD); |
| 1569 | |
| 1570 | if (Value *V = Folder.FoldBinOpFMF(Instruction::FRem, L, R, FMF)) return V; |
| 1571 | Instruction *I = setFPAttrs(BinaryOperator::CreateFRem(L, R), FPMD, FMF); |
| 1572 | return Insert(I, Name); |
| 1573 | } |
| 1574 | |
| 1575 | /// Copy fast-math-flags from an instruction rather than using the builder's |
| 1576 | /// default FMF. |
| 1577 | Value *CreateFRemFMF(Value *L, Value *R, Instruction *FMFSource, |
| 1578 | const Twine &Name = "") { |
| 1579 | if (IsFPConstrained) |
| 1580 | return CreateConstrainedFPBinOp(Intrinsic::experimental_constrained_frem, |
| 1581 | L, R, FMFSource, Name); |
| 1582 | |
| 1583 | FastMathFlags FMF = FMFSource->getFastMathFlags(); |
| 1584 | if (Value *V = Folder.FoldBinOpFMF(Instruction::FRem, L, R, FMF)) return V; |
| 1585 | Instruction *I = setFPAttrs(BinaryOperator::CreateFRem(L, R), nullptr, FMF); |
| 1586 | return Insert(I, Name); |
| 1587 | } |
| 1588 | |
| 1589 | Value *CreateBinOp(Instruction::BinaryOps Opc, |
| 1590 | Value *LHS, Value *RHS, const Twine &Name = "", |
| 1591 | MDNode *FPMathTag = nullptr) { |
| 1592 | if (Value *V = Folder.FoldBinOp(Opc, LHS, RHS)) return V; |
| 1593 | Instruction *BinOp = BinaryOperator::Create(Opc, LHS, RHS); |
| 1594 | if (isa<FPMathOperator>(BinOp)) |
| 1595 | setFPAttrs(BinOp, FPMathTag, FMF); |
| 1596 | return Insert(BinOp, Name); |
| 1597 | } |
| 1598 | |
| 1599 | Value *CreateLogicalAnd(Value *Cond1, Value *Cond2, const Twine &Name = "") { |
| 1600 | assert(Cond2->getType()->isIntOrIntVectorTy(1))(static_cast <bool> (Cond2->getType()->isIntOrIntVectorTy (1)) ? void (0) : __assert_fail ("Cond2->getType()->isIntOrIntVectorTy(1)" , "llvm/include/llvm/IR/IRBuilder.h", 1600, __extension__ __PRETTY_FUNCTION__ )); |
| 1601 | return CreateSelect(Cond1, Cond2, |
| 1602 | ConstantInt::getNullValue(Cond2->getType()), Name); |
| 1603 | } |
| 1604 | |
| 1605 | Value *CreateLogicalOr(Value *Cond1, Value *Cond2, const Twine &Name = "") { |
| 1606 | assert(Cond2->getType()->isIntOrIntVectorTy(1))(static_cast <bool> (Cond2->getType()->isIntOrIntVectorTy (1)) ? void (0) : __assert_fail ("Cond2->getType()->isIntOrIntVectorTy(1)" , "llvm/include/llvm/IR/IRBuilder.h", 1606, __extension__ __PRETTY_FUNCTION__ )); |
| 1607 | return CreateSelect(Cond1, ConstantInt::getAllOnesValue(Cond2->getType()), |
| 1608 | Cond2, Name); |
| 1609 | } |
| 1610 | |
| 1611 | Value *CreateLogicalOp(Instruction::BinaryOps Opc, Value *Cond1, Value *Cond2, |
| 1612 | const Twine &Name = "") { |
| 1613 | switch (Opc) { |
| 1614 | case Instruction::And: |
| 1615 | return CreateLogicalAnd(Cond1, Cond2, Name); |
| 1616 | case Instruction::Or: |
| 1617 | return CreateLogicalOr(Cond1, Cond2, Name); |
| 1618 | default: |
| 1619 | break; |
| 1620 | } |
| 1621 | llvm_unreachable("Not a logical operation.")::llvm::llvm_unreachable_internal("Not a logical operation.", "llvm/include/llvm/IR/IRBuilder.h", 1621); |
| 1622 | } |
| 1623 | |
| 1624 | // NOTE: this is sequential, non-commutative, ordered reduction! |
| 1625 | Value *CreateLogicalOr(ArrayRef<Value *> Ops) { |
| 1626 | assert(!Ops.empty())(static_cast <bool> (!Ops.empty()) ? void (0) : __assert_fail ("!Ops.empty()", "llvm/include/llvm/IR/IRBuilder.h", 1626, __extension__ __PRETTY_FUNCTION__)); |
| 1627 | Value *Accum = Ops[0]; |
| 1628 | for (unsigned i = 1; i < Ops.size(); i++) |
| 1629 | Accum = CreateLogicalOr(Accum, Ops[i]); |
| 1630 | return Accum; |
| 1631 | } |
| 1632 | |
| 1633 | CallInst *CreateConstrainedFPBinOp( |
| 1634 | Intrinsic::ID ID, Value *L, Value *R, Instruction *FMFSource = nullptr, |
| 1635 | const Twine &Name = "", MDNode *FPMathTag = nullptr, |
| 1636 | std::optional<RoundingMode> Rounding = std::nullopt, |
| 1637 | std::optional<fp::ExceptionBehavior> Except = std::nullopt); |
| 1638 | |
| 1639 | Value *CreateNeg(Value *V, const Twine &Name = "", bool HasNUW = false, |
| 1640 | bool HasNSW = false) { |
| 1641 | return CreateSub(Constant::getNullValue(V->getType()), V, Name, HasNUW, |
| 1642 | HasNSW); |
| 1643 | } |
| 1644 | |
| 1645 | Value *CreateNSWNeg(Value *V, const Twine &Name = "") { |
| 1646 | return CreateNeg(V, Name, false, true); |
| 1647 | } |
| 1648 | |
| 1649 | Value *CreateNUWNeg(Value *V, const Twine &Name = "") { |
| 1650 | return CreateNeg(V, Name, true, false); |
| 1651 | } |
| 1652 | |
| 1653 | Value *CreateFNeg(Value *V, const Twine &Name = "", |
| 1654 | MDNode *FPMathTag = nullptr) { |
| 1655 | if (Value *Res = Folder.FoldUnOpFMF(Instruction::FNeg, V, FMF)) |
| 1656 | return Res; |
| 1657 | return Insert(setFPAttrs(UnaryOperator::CreateFNeg(V), FPMathTag, FMF), |
| 1658 | Name); |
| 1659 | } |
| 1660 | |
| 1661 | /// Copy fast-math-flags from an instruction rather than using the builder's |
| 1662 | /// default FMF. |
| 1663 | Value *CreateFNegFMF(Value *V, Instruction *FMFSource, |
| 1664 | const Twine &Name = "") { |
| 1665 | FastMathFlags FMF = FMFSource->getFastMathFlags(); |
| 1666 | if (Value *Res = Folder.FoldUnOpFMF(Instruction::FNeg, V, FMF)) |
| 1667 | return Res; |
| 1668 | return Insert(setFPAttrs(UnaryOperator::CreateFNeg(V), nullptr, FMF), |
| 1669 | Name); |
| 1670 | } |
| 1671 | |
| 1672 | Value *CreateNot(Value *V, const Twine &Name = "") { |
| 1673 | return CreateXor(V, Constant::getAllOnesValue(V->getType()), Name); |
| 1674 | } |
| 1675 | |
| 1676 | Value *CreateUnOp(Instruction::UnaryOps Opc, |
| 1677 | Value *V, const Twine &Name = "", |
| 1678 | MDNode *FPMathTag = nullptr) { |
| 1679 | if (Value *Res = Folder.FoldUnOpFMF(Opc, V, FMF)) |
| 1680 | return Res; |
| 1681 | Instruction *UnOp = UnaryOperator::Create(Opc, V); |
| 1682 | if (isa<FPMathOperator>(UnOp)) |
| 1683 | setFPAttrs(UnOp, FPMathTag, FMF); |
| 1684 | return Insert(UnOp, Name); |
| 1685 | } |
| 1686 | |
| 1687 | /// Create either a UnaryOperator or BinaryOperator depending on \p Opc. |
| 1688 | /// Correct number of operands must be passed accordingly. |
| 1689 | Value *CreateNAryOp(unsigned Opc, ArrayRef<Value *> Ops, |
| 1690 | const Twine &Name = "", MDNode *FPMathTag = nullptr); |
| 1691 | |
| 1692 | //===--------------------------------------------------------------------===// |
| 1693 | // Instruction creation methods: Memory Instructions |
| 1694 | //===--------------------------------------------------------------------===// |
| 1695 | |
| 1696 | AllocaInst *CreateAlloca(Type *Ty, unsigned AddrSpace, |
| 1697 | Value *ArraySize = nullptr, const Twine &Name = "") { |
| 1698 | const DataLayout &DL = BB->getModule()->getDataLayout(); |
| 1699 | Align AllocaAlign = DL.getPrefTypeAlign(Ty); |
| 1700 | return Insert(new AllocaInst(Ty, AddrSpace, ArraySize, AllocaAlign), Name); |
| 1701 | } |
| 1702 | |
| 1703 | AllocaInst *CreateAlloca(Type *Ty, Value *ArraySize = nullptr, |
| 1704 | const Twine &Name = "") { |
| 1705 | const DataLayout &DL = BB->getModule()->getDataLayout(); |
| 1706 | Align AllocaAlign = DL.getPrefTypeAlign(Ty); |
| 1707 | unsigned AddrSpace = DL.getAllocaAddrSpace(); |
| 1708 | return Insert(new AllocaInst(Ty, AddrSpace, ArraySize, AllocaAlign), Name); |
| 1709 | } |
| 1710 | |
| 1711 | /// Provided to resolve 'CreateLoad(Ty, Ptr, "...")' correctly, instead of |
| 1712 | /// converting the string to 'bool' for the isVolatile parameter. |
| 1713 | LoadInst *CreateLoad(Type *Ty, Value *Ptr, const char *Name) { |
| 1714 | return CreateAlignedLoad(Ty, Ptr, MaybeAlign(), Name); |
| 1715 | } |
| 1716 | |
| 1717 | LoadInst *CreateLoad(Type *Ty, Value *Ptr, const Twine &Name = "") { |
| 1718 | return CreateAlignedLoad(Ty, Ptr, MaybeAlign(), Name); |
| 1719 | } |
| 1720 | |
| 1721 | LoadInst *CreateLoad(Type *Ty, Value *Ptr, bool isVolatile, |
| 1722 | const Twine &Name = "") { |
| 1723 | return CreateAlignedLoad(Ty, Ptr, MaybeAlign(), isVolatile, Name); |
| 1724 | } |
| 1725 | |
| 1726 | StoreInst *CreateStore(Value *Val, Value *Ptr, bool isVolatile = false) { |
| 1727 | return CreateAlignedStore(Val, Ptr, MaybeAlign(), isVolatile); |
| 1728 | } |
| 1729 | |
| 1730 | LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, MaybeAlign Align, |
| 1731 | const char *Name) { |
| 1732 | return CreateAlignedLoad(Ty, Ptr, Align, /*isVolatile*/false, Name); |
| 1733 | } |
| 1734 | |
| 1735 | LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, MaybeAlign Align, |
| 1736 | const Twine &Name = "") { |
| 1737 | return CreateAlignedLoad(Ty, Ptr, Align, /*isVolatile*/false, Name); |
| 1738 | } |
| 1739 | |
| 1740 | LoadInst *CreateAlignedLoad(Type *Ty, Value *Ptr, MaybeAlign Align, |
| 1741 | bool isVolatile, const Twine &Name = "") { |
| 1742 | if (!Align) { |
| 1743 | const DataLayout &DL = BB->getModule()->getDataLayout(); |
| 1744 | Align = DL.getABITypeAlign(Ty); |
| 1745 | } |
| 1746 | return Insert(new LoadInst(Ty, Ptr, Twine(), isVolatile, *Align), Name); |
| 1747 | } |
| 1748 | |
| 1749 | StoreInst *CreateAlignedStore(Value *Val, Value *Ptr, MaybeAlign Align, |
| 1750 | bool isVolatile = false) { |
| 1751 | if (!Align) { |
| 1752 | const DataLayout &DL = BB->getModule()->getDataLayout(); |
| 1753 | Align = DL.getABITypeAlign(Val->getType()); |
| 1754 | } |
| 1755 | return Insert(new StoreInst(Val, Ptr, isVolatile, *Align)); |
| 1756 | } |
| 1757 | FenceInst *CreateFence(AtomicOrdering Ordering, |
| 1758 | SyncScope::ID SSID = SyncScope::System, |
| 1759 | const Twine &Name = "") { |
| 1760 | return Insert(new FenceInst(Context, Ordering, SSID), Name); |
| 1761 | } |
| 1762 | |
| 1763 | AtomicCmpXchgInst * |
| 1764 | CreateAtomicCmpXchg(Value *Ptr, Value *Cmp, Value *New, MaybeAlign Align, |
| 1765 | AtomicOrdering SuccessOrdering, |
| 1766 | AtomicOrdering FailureOrdering, |
| 1767 | SyncScope::ID SSID = SyncScope::System) { |
| 1768 | if (!Align) { |
| 1769 | const DataLayout &DL = BB->getModule()->getDataLayout(); |
| 1770 | Align = llvm::Align(DL.getTypeStoreSize(New->getType())); |
| 1771 | } |
| 1772 | |
| 1773 | return Insert(new AtomicCmpXchgInst(Ptr, Cmp, New, *Align, SuccessOrdering, |
| 1774 | FailureOrdering, SSID)); |
| 1775 | } |
| 1776 | |
| 1777 | AtomicRMWInst *CreateAtomicRMW(AtomicRMWInst::BinOp Op, Value *Ptr, |
| 1778 | Value *Val, MaybeAlign Align, |
| 1779 | AtomicOrdering Ordering, |
| 1780 | SyncScope::ID SSID = SyncScope::System) { |
| 1781 | if (!Align) { |
| 1782 | const DataLayout &DL = BB->getModule()->getDataLayout(); |
| 1783 | Align = llvm::Align(DL.getTypeStoreSize(Val->getType())); |
| 1784 | } |
| 1785 | |
| 1786 | return Insert(new AtomicRMWInst(Op, Ptr, Val, *Align, Ordering, SSID)); |
| 1787 | } |
| 1788 | |
| 1789 | Value *CreateGEP(Type *Ty, Value *Ptr, ArrayRef<Value *> IdxList, |
| 1790 | const Twine &Name = "", bool IsInBounds = false) { |
| 1791 | if (auto *V = Folder.FoldGEP(Ty, Ptr, IdxList, IsInBounds)) |
| 1792 | return V; |
| 1793 | return Insert(IsInBounds |
| 1794 | ? GetElementPtrInst::CreateInBounds(Ty, Ptr, IdxList) |
| 1795 | : GetElementPtrInst::Create(Ty, Ptr, IdxList), |
| 1796 | Name); |
| 1797 | } |
| 1798 | |
| 1799 | Value *CreateInBoundsGEP(Type *Ty, Value *Ptr, ArrayRef<Value *> IdxList, |
| 1800 | const Twine &Name = "") { |
| 1801 | return CreateGEP(Ty, Ptr, IdxList, Name, /* IsInBounds */ true); |
| 1802 | } |
| 1803 | |
| 1804 | Value *CreateConstGEP1_32(Type *Ty, Value *Ptr, unsigned Idx0, |
| 1805 | const Twine &Name = "") { |
| 1806 | Value *Idx = ConstantInt::get(Type::getInt32Ty(Context), Idx0); |
| 1807 | |
| 1808 | if (auto *V = Folder.FoldGEP(Ty, Ptr, Idx, /*IsInBounds=*/false)) |
| 1809 | return V; |
| 1810 | |
| 1811 | return Insert(GetElementPtrInst::Create(Ty, Ptr, Idx), Name); |
| 1812 | } |
| 1813 | |
| 1814 | Value *CreateConstInBoundsGEP1_32(Type *Ty, Value *Ptr, unsigned Idx0, |
| 1815 | const Twine &Name = "") { |
| 1816 | Value *Idx = ConstantInt::get(Type::getInt32Ty(Context), Idx0); |
| 1817 | |
| 1818 | if (auto *V = Folder.FoldGEP(Ty, Ptr, Idx, /*IsInBounds=*/true)) |
| 1819 | return V; |
| 1820 | |
| 1821 | return Insert(GetElementPtrInst::CreateInBounds(Ty, Ptr, Idx), Name); |
| 1822 | } |
| 1823 | |
| 1824 | Value *CreateConstGEP2_32(Type *Ty, Value *Ptr, unsigned Idx0, unsigned Idx1, |
| 1825 | const Twine &Name = "") { |
| 1826 | Value *Idxs[] = { |
| 1827 | ConstantInt::get(Type::getInt32Ty(Context), Idx0), |
| 1828 | ConstantInt::get(Type::getInt32Ty(Context), Idx1) |
| 1829 | }; |
| 1830 | |
| 1831 | if (auto *V = Folder.FoldGEP(Ty, Ptr, Idxs, /*IsInBounds=*/false)) |
| 1832 | return V; |
| 1833 | |
| 1834 | return Insert(GetElementPtrInst::Create(Ty, Ptr, Idxs), Name); |
| 1835 | } |
| 1836 | |
| 1837 | Value *CreateConstInBoundsGEP2_32(Type *Ty, Value *Ptr, unsigned Idx0, |
| 1838 | unsigned Idx1, const Twine &Name = "") { |
| 1839 | Value *Idxs[] = { |
| 1840 | ConstantInt::get(Type::getInt32Ty(Context), Idx0), |
| 1841 | ConstantInt::get(Type::getInt32Ty(Context), Idx1) |
| 1842 | }; |
| 1843 | |
| 1844 | if (auto *V = Folder.FoldGEP(Ty, Ptr, Idxs, /*IsInBounds=*/true)) |
| 1845 | return V; |
| 1846 | |
| 1847 | return Insert(GetElementPtrInst::CreateInBounds(Ty, Ptr, Idxs), Name); |
| 1848 | } |
| 1849 | |
| 1850 | Value *CreateConstGEP1_64(Type *Ty, Value *Ptr, uint64_t Idx0, |
| 1851 | const Twine &Name = "") { |
| 1852 | Value *Idx = ConstantInt::get(Type::getInt64Ty(Context), Idx0); |
| 1853 | |
| 1854 | if (auto *V = Folder.FoldGEP(Ty, Ptr, Idx, /*IsInBounds=*/false)) |
| 1855 | return V; |
| 1856 | |
| 1857 | return Insert(GetElementPtrInst::Create(Ty, Ptr, Idx), Name); |
| 1858 | } |
| 1859 | |
| 1860 | Value *CreateConstInBoundsGEP1_64(Type *Ty, Value *Ptr, uint64_t Idx0, |
| 1861 | const Twine &Name = "") { |
| 1862 | Value *Idx = ConstantInt::get(Type::getInt64Ty(Context), Idx0); |
| 1863 | |
| 1864 | if (auto *V = Folder.FoldGEP(Ty, Ptr, Idx, /*IsInBounds=*/true)) |
| 1865 | return V; |
| 1866 | |
| 1867 | return Insert(GetElementPtrInst::CreateInBounds(Ty, Ptr, Idx), Name); |
| 1868 | } |
| 1869 | |
| 1870 | Value *CreateConstGEP2_64(Type *Ty, Value *Ptr, uint64_t Idx0, uint64_t Idx1, |
| 1871 | const Twine &Name = "") { |
| 1872 | Value *Idxs[] = { |
| 1873 | ConstantInt::get(Type::getInt64Ty(Context), Idx0), |
| 1874 | ConstantInt::get(Type::getInt64Ty(Context), Idx1) |
| 1875 | }; |
| 1876 | |
| 1877 | if (auto *V = Folder.FoldGEP(Ty, Ptr, Idxs, /*IsInBounds=*/false)) |
| 1878 | return V; |
| 1879 | |
| 1880 | return Insert(GetElementPtrInst::Create(Ty, Ptr, Idxs), Name); |
| 1881 | } |
| 1882 | |
| 1883 | Value *CreateConstInBoundsGEP2_64(Type *Ty, Value *Ptr, uint64_t Idx0, |
| 1884 | uint64_t Idx1, const Twine &Name = "") { |
| 1885 | Value *Idxs[] = { |
| 1886 | ConstantInt::get(Type::getInt64Ty(Context), Idx0), |
| 1887 | ConstantInt::get(Type::getInt64Ty(Context), Idx1) |
| 1888 | }; |
| 1889 | |
| 1890 | if (auto *V = Folder.FoldGEP(Ty, Ptr, Idxs, /*IsInBounds=*/true)) |
| 1891 | return V; |
| 1892 | |
| 1893 | return Insert(GetElementPtrInst::CreateInBounds(Ty, Ptr, Idxs), Name); |
| 1894 | } |
| 1895 | |
| 1896 | Value *CreateStructGEP(Type *Ty, Value *Ptr, unsigned Idx, |
| 1897 | const Twine &Name = "") { |
| 1898 | return CreateConstInBoundsGEP2_32(Ty, Ptr, 0, Idx, Name); |
| 1899 | } |
| 1900 | |
| 1901 | /// Same as CreateGlobalString, but return a pointer with "i8*" type |
| 1902 | /// instead of a pointer to array of i8. |
| 1903 | /// |
| 1904 | /// If no module is given via \p M, it is take from the insertion point basic |
| 1905 | /// block. |
| 1906 | Constant *CreateGlobalStringPtr(StringRef Str, const Twine &Name = "", |
| 1907 | unsigned AddressSpace = 0, |
| 1908 | Module *M = nullptr) { |
| 1909 | GlobalVariable *GV = CreateGlobalString(Str, Name, AddressSpace, M); |
| 1910 | Constant *Zero = ConstantInt::get(Type::getInt32Ty(Context), 0); |
| 1911 | Constant *Indices[] = {Zero, Zero}; |
| 1912 | return ConstantExpr::getInBoundsGetElementPtr(GV->getValueType(), GV, |
| 1913 | Indices); |
| 1914 | } |
| 1915 | |
| 1916 | //===--------------------------------------------------------------------===// |
| 1917 | // Instruction creation methods: Cast/Conversion Operators |
| 1918 | //===--------------------------------------------------------------------===// |
| 1919 | |
| 1920 | Value *CreateTrunc(Value *V, Type *DestTy, const Twine &Name = "") { |
| 1921 | return CreateCast(Instruction::Trunc, V, DestTy, Name); |
| 1922 | } |
| 1923 | |
| 1924 | Value *CreateZExt(Value *V, Type *DestTy, const Twine &Name = "") { |
| 1925 | return CreateCast(Instruction::ZExt, V, DestTy, Name); |
| 1926 | } |
| 1927 | |
| 1928 | Value *CreateSExt(Value *V, Type *DestTy, const Twine &Name = "") { |
| 1929 | return CreateCast(Instruction::SExt, V, DestTy, Name); |
| 1930 | } |
| 1931 | |
| 1932 | /// Create a ZExt or Trunc from the integer value V to DestTy. Return |
| 1933 | /// the value untouched if the type of V is already DestTy. |
| 1934 | Value *CreateZExtOrTrunc(Value *V, Type *DestTy, |
| 1935 | const Twine &Name = "") { |
| 1936 | assert(V->getType()->isIntOrIntVectorTy() &&(static_cast <bool> (V->getType()->isIntOrIntVectorTy () && DestTy->isIntOrIntVectorTy() && "Can only zero extend/truncate integers!" ) ? void (0) : __assert_fail ("V->getType()->isIntOrIntVectorTy() && DestTy->isIntOrIntVectorTy() && \"Can only zero extend/truncate integers!\"" , "llvm/include/llvm/IR/IRBuilder.h", 1938, __extension__ __PRETTY_FUNCTION__ )) |
| 1937 | DestTy->isIntOrIntVectorTy() &&(static_cast <bool> (V->getType()->isIntOrIntVectorTy () && DestTy->isIntOrIntVectorTy() && "Can only zero extend/truncate integers!" ) ? void (0) : __assert_fail ("V->getType()->isIntOrIntVectorTy() && DestTy->isIntOrIntVectorTy() && \"Can only zero extend/truncate integers!\"" , "llvm/include/llvm/IR/IRBuilder.h", 1938, __extension__ __PRETTY_FUNCTION__ )) |
| 1938 | "Can only zero extend/truncate integers!")(static_cast <bool> (V->getType()->isIntOrIntVectorTy () && DestTy->isIntOrIntVectorTy() && "Can only zero extend/truncate integers!" ) ? void (0) : __assert_fail ("V->getType()->isIntOrIntVectorTy() && DestTy->isIntOrIntVectorTy() && \"Can only zero extend/truncate integers!\"" , "llvm/include/llvm/IR/IRBuilder.h", 1938, __extension__ __PRETTY_FUNCTION__ )); |
| 1939 | Type *VTy = V->getType(); |
| 1940 | if (VTy->getScalarSizeInBits() < DestTy->getScalarSizeInBits()) |
| 1941 | return CreateZExt(V, DestTy, Name); |
| 1942 | if (VTy->getScalarSizeInBits() > DestTy->getScalarSizeInBits()) |
| 1943 | return CreateTrunc(V, DestTy, Name); |
| 1944 | return V; |
| 1945 | } |
| 1946 | |
| 1947 | /// Create a SExt or Trunc from the integer value V to DestTy. Return |
| 1948 | /// the value untouched if the type of V is already DestTy. |
| 1949 | Value *CreateSExtOrTrunc(Value *V, Type *DestTy, |
| 1950 | const Twine &Name = "") { |
| 1951 | assert(V->getType()->isIntOrIntVectorTy() &&(static_cast <bool> (V->getType()->isIntOrIntVectorTy () && DestTy->isIntOrIntVectorTy() && "Can only sign extend/truncate integers!" ) ? void (0) : __assert_fail ("V->getType()->isIntOrIntVectorTy() && DestTy->isIntOrIntVectorTy() && \"Can only sign extend/truncate integers!\"" , "llvm/include/llvm/IR/IRBuilder.h", 1953, __extension__ __PRETTY_FUNCTION__ )) |
| 1952 | DestTy->isIntOrIntVectorTy() &&(static_cast <bool> (V->getType()->isIntOrIntVectorTy () && DestTy->isIntOrIntVectorTy() && "Can only sign extend/truncate integers!" ) ? void (0) : __assert_fail ("V->getType()->isIntOrIntVectorTy() && DestTy->isIntOrIntVectorTy() && \"Can only sign extend/truncate integers!\"" , "llvm/include/llvm/IR/IRBuilder.h", 1953, __extension__ __PRETTY_FUNCTION__ )) |
| 1953 | "Can only sign extend/truncate integers!")(static_cast <bool> (V->getType()->isIntOrIntVectorTy () && DestTy->isIntOrIntVectorTy() && "Can only sign extend/truncate integers!" ) ? void (0) : __assert_fail ("V->getType()->isIntOrIntVectorTy() && DestTy->isIntOrIntVectorTy() && \"Can only sign extend/truncate integers!\"" , "llvm/include/llvm/IR/IRBuilder.h", 1953, __extension__ __PRETTY_FUNCTION__ )); |
| 1954 | Type *VTy = V->getType(); |
| 1955 | if (VTy->getScalarSizeInBits() < DestTy->getScalarSizeInBits()) |
| 1956 | return CreateSExt(V, DestTy, Name); |
| 1957 | if (VTy->getScalarSizeInBits() > DestTy->getScalarSizeInBits()) |
| 1958 | return CreateTrunc(V, DestTy, Name); |
| 1959 | return V; |
| 1960 | } |
| 1961 | |
| 1962 | Value *CreateFPToUI(Value *V, Type *DestTy, const Twine &Name = "") { |
| 1963 | if (IsFPConstrained) |
| 1964 | return CreateConstrainedFPCast(Intrinsic::experimental_constrained_fptoui, |
| 1965 | V, DestTy, nullptr, Name); |
| 1966 | return CreateCast(Instruction::FPToUI, V, DestTy, Name); |
| 1967 | } |
| 1968 | |
| 1969 | Value *CreateFPToSI(Value *V, Type *DestTy, const Twine &Name = "") { |
| 1970 | if (IsFPConstrained) |
| 1971 | return CreateConstrainedFPCast(Intrinsic::experimental_constrained_fptosi, |
| 1972 | V, DestTy, nullptr, Name); |
| 1973 | return CreateCast(Instruction::FPToSI, V, DestTy, Name); |
| 1974 | } |
| 1975 | |
| 1976 | Value *CreateUIToFP(Value *V, Type *DestTy, const Twine &Name = ""){ |
| 1977 | if (IsFPConstrained) |
| 1978 | return CreateConstrainedFPCast(Intrinsic::experimental_constrained_uitofp, |
| 1979 | V, DestTy, nullptr, Name); |
| 1980 | return CreateCast(Instruction::UIToFP, V, DestTy, Name); |
| 1981 | } |
| 1982 | |
| 1983 | Value *CreateSIToFP(Value *V, Type *DestTy, const Twine &Name = ""){ |
| 1984 | if (IsFPConstrained) |
| 1985 | return CreateConstrainedFPCast(Intrinsic::experimental_constrained_sitofp, |
| 1986 | V, DestTy, nullptr, Name); |
| 1987 | return CreateCast(Instruction::SIToFP, V, DestTy, Name); |
| 1988 | } |
| 1989 | |
| 1990 | Value *CreateFPTrunc(Value *V, Type *DestTy, |
| 1991 | const Twine &Name = "") { |
| 1992 | if (IsFPConstrained) |
| 1993 | return CreateConstrainedFPCast( |
| 1994 | Intrinsic::experimental_constrained_fptrunc, V, DestTy, nullptr, |
| 1995 | Name); |
| 1996 | return CreateCast(Instruction::FPTrunc, V, DestTy, Name); |
| 1997 | } |
| 1998 | |
| 1999 | Value *CreateFPExt(Value *V, Type *DestTy, const Twine &Name = "") { |
| 2000 | if (IsFPConstrained) |
| 2001 | return CreateConstrainedFPCast(Intrinsic::experimental_constrained_fpext, |
| 2002 | V, DestTy, nullptr, Name); |
| 2003 | return CreateCast(Instruction::FPExt, V, DestTy, Name); |
| 2004 | } |
| 2005 | |
| 2006 | Value *CreatePtrToInt(Value *V, Type *DestTy, |
| 2007 | const Twine &Name = "") { |
| 2008 | return CreateCast(Instruction::PtrToInt, V, DestTy, Name); |
| 2009 | } |
| 2010 | |
| 2011 | Value *CreateIntToPtr(Value *V, Type *DestTy, |
| 2012 | const Twine &Name = "") { |
| 2013 | return CreateCast(Instruction::IntToPtr, V, DestTy, Name); |
| 2014 | } |
| 2015 | |
| 2016 | Value *CreateBitCast(Value *V, Type *DestTy, |
| 2017 | const Twine &Name = "") { |
| 2018 | return CreateCast(Instruction::BitCast, V, DestTy, Name); |
| 2019 | } |
| 2020 | |
| 2021 | Value *CreateAddrSpaceCast(Value *V, Type *DestTy, |
| 2022 | const Twine &Name = "") { |
| 2023 | return CreateCast(Instruction::AddrSpaceCast, V, DestTy, Name); |
| 2024 | } |
| 2025 | |
| 2026 | Value *CreateZExtOrBitCast(Value *V, Type *DestTy, |
| 2027 | const Twine &Name = "") { |
| 2028 | if (V->getType() == DestTy) |
| 2029 | return V; |
| 2030 | if (auto *VC = dyn_cast<Constant>(V)) |
| 2031 | return Insert(Folder.CreateZExtOrBitCast(VC, DestTy), Name); |
| 2032 | return Insert(CastInst::CreateZExtOrBitCast(V, DestTy), Name); |
| 2033 | } |
| 2034 | |
| 2035 | Value *CreateSExtOrBitCast(Value *V, Type *DestTy, |
| 2036 | const Twine &Name = "") { |
| 2037 | if (V->getType() == DestTy) |
| 2038 | return V; |
| 2039 | if (auto *VC = dyn_cast<Constant>(V)) |
| 2040 | return Insert(Folder.CreateSExtOrBitCast(VC, DestTy), Name); |
| 2041 | return Insert(CastInst::CreateSExtOrBitCast(V, DestTy), Name); |
| 2042 | } |
| 2043 | |
| 2044 | Value *CreateTruncOrBitCast(Value *V, Type *DestTy, |
| 2045 | const Twine &Name = "") { |
| 2046 | if (V->getType() == DestTy) |
| 2047 | return V; |
| 2048 | if (auto *VC = dyn_cast<Constant>(V)) |
| 2049 | return Insert(Folder.CreateTruncOrBitCast(VC, DestTy), Name); |
| 2050 | return Insert(CastInst::CreateTruncOrBitCast(V, DestTy), Name); |
| 2051 | } |
| 2052 | |
| 2053 | Value *CreateCast(Instruction::CastOps Op, Value *V, Type *DestTy, |
| 2054 | const Twine &Name = "") { |
| 2055 | if (V->getType() == DestTy) |
| 2056 | return V; |
| 2057 | if (auto *VC = dyn_cast<Constant>(V)) |
| 2058 | return Insert(Folder.CreateCast(Op, VC, DestTy), Name); |
| 2059 | return Insert(CastInst::Create(Op, V, DestTy), Name); |
| 2060 | } |
| 2061 | |
| 2062 | Value *CreatePointerCast(Value *V, Type *DestTy, |
| 2063 | const Twine &Name = "") { |
| 2064 | if (V->getType() == DestTy) |
| 2065 | return V; |
| 2066 | if (auto *VC = dyn_cast<Constant>(V)) |
| 2067 | return Insert(Folder.CreatePointerCast(VC, DestTy), Name); |
| 2068 | return Insert(CastInst::CreatePointerCast(V, DestTy), Name); |
| 2069 | } |
| 2070 | |
| 2071 | Value *CreatePointerBitCastOrAddrSpaceCast(Value *V, Type *DestTy, |
| 2072 | const Twine &Name = "") { |
| 2073 | if (V->getType() == DestTy) |
| 2074 | return V; |
| 2075 | |
| 2076 | if (auto *VC = dyn_cast<Constant>(V)) { |
| 2077 | return Insert(Folder.CreatePointerBitCastOrAddrSpaceCast(VC, DestTy), |
| 2078 | Name); |
| 2079 | } |
| 2080 | |
| 2081 | return Insert(CastInst::CreatePointerBitCastOrAddrSpaceCast(V, DestTy), |
| 2082 | Name); |
| 2083 | } |
| 2084 | |
| 2085 | Value *CreateIntCast(Value *V, Type *DestTy, bool isSigned, |
| 2086 | const Twine &Name = "") { |
| 2087 | if (V->getType() == DestTy) |
| 2088 | return V; |
| 2089 | if (auto *VC = dyn_cast<Constant>(V)) |
| 2090 | return Insert(Folder.CreateIntCast(VC, DestTy, isSigned), Name); |
| 2091 | return Insert(CastInst::CreateIntegerCast(V, DestTy, isSigned), Name); |
| 2092 | } |
| 2093 | |
| 2094 | Value *CreateBitOrPointerCast(Value *V, Type *DestTy, |
| 2095 | const Twine &Name = "") { |
| 2096 | if (V->getType() == DestTy) |
| 2097 | return V; |
| 2098 | if (V->getType()->isPtrOrPtrVectorTy() && DestTy->isIntOrIntVectorTy()) |
| 2099 | return CreatePtrToInt(V, DestTy, Name); |
| 2100 | if (V->getType()->isIntOrIntVectorTy() && DestTy->isPtrOrPtrVectorTy()) |
| 2101 | return CreateIntToPtr(V, DestTy, Name); |
| 2102 | |
| 2103 | return CreateBitCast(V, DestTy, Name); |
| 2104 | } |
| 2105 | |
| 2106 | Value *CreateFPCast(Value *V, Type *DestTy, const Twine &Name = "") { |
| 2107 | if (V->getType() == DestTy) |
| 2108 | return V; |
| 2109 | if (auto *VC = dyn_cast<Constant>(V)) |
| 2110 | return Insert(Folder.CreateFPCast(VC, DestTy), Name); |
| 2111 | return Insert(CastInst::CreateFPCast(V, DestTy), Name); |
| 2112 | } |
| 2113 | |
| 2114 | CallInst *CreateConstrainedFPCast( |
| 2115 | Intrinsic::ID ID, Value *V, Type *DestTy, |
| 2116 | Instruction *FMFSource = nullptr, const Twine &Name = "", |
| 2117 | MDNode *FPMathTag = nullptr, |
| 2118 | std::optional<RoundingMode> Rounding = std::nullopt, |
| 2119 | std::optional<fp::ExceptionBehavior> Except = std::nullopt); |
| 2120 | |
| 2121 | // Provided to resolve 'CreateIntCast(Ptr, Ptr, "...")', giving a |
| 2122 | // compile time error, instead of converting the string to bool for the |
| 2123 | // isSigned parameter. |
| 2124 | Value *CreateIntCast(Value *, Type *, const char *) = delete; |
| 2125 | |
| 2126 | //===--------------------------------------------------------------------===// |
| 2127 | // Instruction creation methods: Compare Instructions |
| 2128 | //===--------------------------------------------------------------------===// |
| 2129 | |
| 2130 | Value *CreateICmpEQ(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 2131 | return CreateICmp(ICmpInst::ICMP_EQ, LHS, RHS, Name); |
| 2132 | } |
| 2133 | |
| 2134 | Value *CreateICmpNE(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 2135 | return CreateICmp(ICmpInst::ICMP_NE, LHS, RHS, Name); |
| 2136 | } |
| 2137 | |
| 2138 | Value *CreateICmpUGT(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 2139 | return CreateICmp(ICmpInst::ICMP_UGT, LHS, RHS, Name); |
| 2140 | } |
| 2141 | |
| 2142 | Value *CreateICmpUGE(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 2143 | return CreateICmp(ICmpInst::ICMP_UGE, LHS, RHS, Name); |
| 2144 | } |
| 2145 | |
| 2146 | Value *CreateICmpULT(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 2147 | return CreateICmp(ICmpInst::ICMP_ULT, LHS, RHS, Name); |
| 2148 | } |
| 2149 | |
| 2150 | Value *CreateICmpULE(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 2151 | return CreateICmp(ICmpInst::ICMP_ULE, LHS, RHS, Name); |
| 2152 | } |
| 2153 | |
| 2154 | Value *CreateICmpSGT(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 2155 | return CreateICmp(ICmpInst::ICMP_SGT, LHS, RHS, Name); |
| 2156 | } |
| 2157 | |
| 2158 | Value *CreateICmpSGE(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 2159 | return CreateICmp(ICmpInst::ICMP_SGE, LHS, RHS, Name); |
| 2160 | } |
| 2161 | |
| 2162 | Value *CreateICmpSLT(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 2163 | return CreateICmp(ICmpInst::ICMP_SLT, LHS, RHS, Name); |
| 2164 | } |
| 2165 | |
| 2166 | Value *CreateICmpSLE(Value *LHS, Value *RHS, const Twine &Name = "") { |
| 2167 | return CreateICmp(ICmpInst::ICMP_SLE, LHS, RHS, Name); |
| 2168 | } |
| 2169 | |
| 2170 | Value *CreateFCmpOEQ(Value *LHS, Value *RHS, const Twine &Name = "", |
| 2171 | MDNode *FPMathTag = nullptr) { |
| 2172 | return CreateFCmp(FCmpInst::FCMP_OEQ, LHS, RHS, Name, FPMathTag); |
| 2173 | } |
| 2174 | |
| 2175 | Value *CreateFCmpOGT(Value *LHS, Value *RHS, const Twine &Name = "", |
| 2176 | MDNode *FPMathTag = nullptr) { |
| 2177 | return CreateFCmp(FCmpInst::FCMP_OGT, LHS, RHS, Name, FPMathTag); |
| 2178 | } |
| 2179 | |
| 2180 | Value *CreateFCmpOGE(Value *LHS, Value *RHS, const Twine &Name = "", |
| 2181 | MDNode *FPMathTag = nullptr) { |
| 2182 | return CreateFCmp(FCmpInst::FCMP_OGE, LHS, RHS, Name, FPMathTag); |
| 2183 | } |
| 2184 | |
| 2185 | Value *CreateFCmpOLT(Value *LHS, Value *RHS, const Twine &Name = "", |
| 2186 | MDNode *FPMathTag = nullptr) { |
| 2187 | return CreateFCmp(FCmpInst::FCMP_OLT, LHS, RHS, Name, FPMathTag); |
| 2188 | } |
| 2189 | |
| 2190 | Value *CreateFCmpOLE(Value *LHS, Value *RHS, const Twine &Name = "", |
| 2191 | MDNode *FPMathTag = nullptr) { |
| 2192 | return CreateFCmp(FCmpInst::FCMP_OLE, LHS, RHS, Name, FPMathTag); |
| 2193 | } |
| 2194 | |
| 2195 | Value *CreateFCmpONE(Value *LHS, Value *RHS, const Twine &Name = "", |
| 2196 | MDNode *FPMathTag = nullptr) { |
| 2197 | return CreateFCmp(FCmpInst::FCMP_ONE, LHS, RHS, Name, FPMathTag); |
| 2198 | } |
| 2199 | |
| 2200 | Value *CreateFCmpORD(Value *LHS, Value *RHS, const Twine &Name = "", |
| 2201 | MDNode *FPMathTag = nullptr) { |
| 2202 | return CreateFCmp(FCmpInst::FCMP_ORD, LHS, RHS, Name, FPMathTag); |
| 2203 | } |
| 2204 | |
| 2205 | Value *CreateFCmpUNO(Value *LHS, Value *RHS, const Twine &Name = "", |
| 2206 | MDNode *FPMathTag = nullptr) { |
| 2207 | return CreateFCmp(FCmpInst::FCMP_UNO, LHS, RHS, Name, FPMathTag); |
| 2208 | } |
| 2209 | |
| 2210 | Value *CreateFCmpUEQ(Value *LHS, Value *RHS, const Twine &Name = "", |
| 2211 | MDNode *FPMathTag = nullptr) { |
| 2212 | return CreateFCmp(FCmpInst::FCMP_UEQ, LHS, RHS, Name, FPMathTag); |
| 2213 | } |
| 2214 | |
| 2215 | Value *CreateFCmpUGT(Value *LHS, Value *RHS, const Twine &Name = "", |
| 2216 | MDNode *FPMathTag = nullptr) { |
| 2217 | return CreateFCmp(FCmpInst::FCMP_UGT, LHS, RHS, Name, FPMathTag); |
| 2218 | } |
| 2219 | |
| 2220 | Value *CreateFCmpUGE(Value *LHS, Value *RHS, const Twine &Name = "", |
| 2221 | MDNode *FPMathTag = nullptr) { |
| 2222 | return CreateFCmp(FCmpInst::FCMP_UGE, LHS, RHS, Name, FPMathTag); |
| 2223 | } |
| 2224 | |
| 2225 | Value *CreateFCmpULT(Value *LHS, Value *RHS, const Twine &Name = "", |
| 2226 | MDNode *FPMathTag = nullptr) { |
| 2227 | return CreateFCmp(FCmpInst::FCMP_ULT, LHS, RHS, Name, FPMathTag); |
| 2228 | } |
| 2229 | |
| 2230 | Value *CreateFCmpULE(Value *LHS, Value *RHS, const Twine &Name = "", |
| 2231 | MDNode *FPMathTag = nullptr) { |
| 2232 | return CreateFCmp(FCmpInst::FCMP_ULE, LHS, RHS, Name, FPMathTag); |
| 2233 | } |
| 2234 | |
| 2235 | Value *CreateFCmpUNE(Value *LHS, Value *RHS, const Twine &Name = "", |
| 2236 | MDNode *FPMathTag = nullptr) { |
| 2237 | return CreateFCmp(FCmpInst::FCMP_UNE, LHS, RHS, Name, FPMathTag); |
| 2238 | } |
| 2239 | |
| 2240 | Value *CreateICmp(CmpInst::Predicate P, Value *LHS, Value *RHS, |
| 2241 | const Twine &Name = "") { |
| 2242 | if (auto *V = Folder.FoldICmp(P, LHS, RHS)) |
| 2243 | return V; |
| 2244 | return Insert(new ICmpInst(P, LHS, RHS), Name); |
| 2245 | } |
| 2246 | |
| 2247 | // Create a quiet floating-point comparison (i.e. one that raises an FP |
| 2248 | // exception only in the case where an input is a signaling NaN). |
| 2249 | // Note that this differs from CreateFCmpS only if IsFPConstrained is true. |
| 2250 | Value *CreateFCmp(CmpInst::Predicate P, Value *LHS, Value *RHS, |
| 2251 | const Twine &Name = "", MDNode *FPMathTag = nullptr) { |
| 2252 | return CreateFCmpHelper(P, LHS, RHS, Name, FPMathTag, false); |
| 2253 | } |
| 2254 | |
| 2255 | Value *CreateCmp(CmpInst::Predicate Pred, Value *LHS, Value *RHS, |
| 2256 | const Twine &Name = "", MDNode *FPMathTag = nullptr) { |
| 2257 | return CmpInst::isFPPredicate(Pred) |
| 2258 | ? CreateFCmp(Pred, LHS, RHS, Name, FPMathTag) |
| 2259 | : CreateICmp(Pred, LHS, RHS, Name); |
| 2260 | } |
| 2261 | |
| 2262 | // Create a signaling floating-point comparison (i.e. one that raises an FP |
| 2263 | // exception whenever an input is any NaN, signaling or quiet). |
| 2264 | // Note that this differs from CreateFCmp only if IsFPConstrained is true. |
| 2265 | Value *CreateFCmpS(CmpInst::Predicate P, Value *LHS, Value *RHS, |
| 2266 | const Twine &Name = "", MDNode *FPMathTag = nullptr) { |
| 2267 | return CreateFCmpHelper(P, LHS, RHS, Name, FPMathTag, true); |
| 2268 | } |
| 2269 | |
| 2270 | private: |
| 2271 | // Helper routine to create either a signaling or a quiet FP comparison. |
| 2272 | Value *CreateFCmpHelper(CmpInst::Predicate P, Value *LHS, Value *RHS, |
| 2273 | const Twine &Name, MDNode *FPMathTag, |
| 2274 | bool IsSignaling); |
| 2275 | |
| 2276 | public: |
| 2277 | CallInst *CreateConstrainedFPCmp( |
| 2278 | Intrinsic::ID ID, CmpInst::Predicate P, Value *L, Value *R, |
| 2279 | const Twine &Name = "", |
| 2280 | std::optional<fp::ExceptionBehavior> Except = std::nullopt); |
| 2281 | |
| 2282 | //===--------------------------------------------------------------------===// |
| 2283 | // Instruction creation methods: Other Instructions |
| 2284 | //===--------------------------------------------------------------------===// |
| 2285 | |
| 2286 | PHINode *CreatePHI(Type *Ty, unsigned NumReservedValues, |
| 2287 | const Twine &Name = "") { |
| 2288 | PHINode *Phi = PHINode::Create(Ty, NumReservedValues); |
| 2289 | if (isa<FPMathOperator>(Phi)) |
| 2290 | setFPAttrs(Phi, nullptr /* MDNode* */, FMF); |
| 2291 | return Insert(Phi, Name); |
| 2292 | } |
| 2293 | |
| 2294 | private: |
| 2295 | CallInst *createCallHelper(Function *Callee, ArrayRef<Value *> Ops, |
| 2296 | const Twine &Name = "", |
| 2297 | Instruction *FMFSource = nullptr, |
| 2298 | ArrayRef<OperandBundleDef> OpBundles = {}); |
| 2299 | |
| 2300 | public: |
| 2301 | CallInst *CreateCall(FunctionType *FTy, Value *Callee, |
| 2302 | ArrayRef<Value *> Args = std::nullopt, |
| 2303 | const Twine &Name = "", MDNode *FPMathTag = nullptr) { |
| 2304 | CallInst *CI = CallInst::Create(FTy, Callee, Args, DefaultOperandBundles); |
| 2305 | if (IsFPConstrained) |
| 2306 | setConstrainedFPCallAttr(CI); |
| 2307 | if (isa<FPMathOperator>(CI)) |
| 2308 | setFPAttrs(CI, FPMathTag, FMF); |
| 2309 | return Insert(CI, Name); |
| 2310 | } |
| 2311 | |
| 2312 | CallInst *CreateCall(FunctionType *FTy, Value *Callee, ArrayRef<Value *> Args, |
| 2313 | ArrayRef<OperandBundleDef> OpBundles, |
| 2314 | const Twine &Name = "", MDNode *FPMathTag = nullptr) { |
| 2315 | CallInst *CI = CallInst::Create(FTy, Callee, Args, OpBundles); |
| 2316 | if (IsFPConstrained) |
| 2317 | setConstrainedFPCallAttr(CI); |
| 2318 | if (isa<FPMathOperator>(CI)) |
| 2319 | setFPAttrs(CI, FPMathTag, FMF); |
| 2320 | return Insert(CI, Name); |
| 2321 | } |
| 2322 | |
| 2323 | CallInst *CreateCall(FunctionCallee Callee, |
| 2324 | ArrayRef<Value *> Args = std::nullopt, |
| 2325 | const Twine &Name = "", MDNode *FPMathTag = nullptr) { |
| 2326 | return CreateCall(Callee.getFunctionType(), Callee.getCallee(), Args, Name, |
| 2327 | FPMathTag); |
| 2328 | } |
| 2329 | |
| 2330 | CallInst *CreateCall(FunctionCallee Callee, ArrayRef<Value *> Args, |
| 2331 | ArrayRef<OperandBundleDef> OpBundles, |
| 2332 | const Twine &Name = "", MDNode *FPMathTag = nullptr) { |
| 2333 | return CreateCall(Callee.getFunctionType(), Callee.getCallee(), Args, |
| 2334 | OpBundles, Name, FPMathTag); |
| 2335 | } |
| 2336 | |
| 2337 | CallInst *CreateConstrainedFPCall( |
| 2338 | Function *Callee, ArrayRef<Value *> Args, const Twine &Name = "", |
| 2339 | std::optional<RoundingMode> Rounding = std::nullopt, |
| 2340 | std::optional<fp::ExceptionBehavior> Except = std::nullopt); |
| 2341 | |
| 2342 | Value *CreateSelect(Value *C, Value *True, Value *False, |
| 2343 | const Twine &Name = "", Instruction *MDFrom = nullptr); |
| 2344 | |
| 2345 | VAArgInst *CreateVAArg(Value *List, Type *Ty, const Twine &Name = "") { |
| 2346 | return Insert(new VAArgInst(List, Ty), Name); |
| 2347 | } |
| 2348 | |
| 2349 | Value *CreateExtractElement(Value *Vec, Value *Idx, |
| 2350 | const Twine &Name = "") { |
| 2351 | if (Value *V = Folder.FoldExtractElement(Vec, Idx)) |
| 2352 | return V; |
| 2353 | return Insert(ExtractElementInst::Create(Vec, Idx), Name); |
| 2354 | } |
| 2355 | |
| 2356 | Value *CreateExtractElement(Value *Vec, uint64_t Idx, |
| 2357 | const Twine &Name = "") { |
| 2358 | return CreateExtractElement(Vec, getInt64(Idx), Name); |
| 2359 | } |
| 2360 | |
| 2361 | Value *CreateInsertElement(Type *VecTy, Value *NewElt, Value *Idx, |
| 2362 | const Twine &Name = "") { |
| 2363 | return CreateInsertElement(PoisonValue::get(VecTy), NewElt, Idx, Name); |
| 2364 | } |
| 2365 | |
| 2366 | Value *CreateInsertElement(Type *VecTy, Value *NewElt, uint64_t Idx, |
| 2367 | const Twine &Name = "") { |
| 2368 | return CreateInsertElement(PoisonValue::get(VecTy), NewElt, Idx, Name); |
| 2369 | } |
| 2370 | |
| 2371 | Value *CreateInsertElement(Value *Vec, Value *NewElt, Value *Idx, |
| 2372 | const Twine &Name = "") { |
| 2373 | if (Value *V = Folder.FoldInsertElement(Vec, NewElt, Idx)) |
| 2374 | return V; |
| 2375 | return Insert(InsertElementInst::Create(Vec, NewElt, Idx), Name); |
| 2376 | } |
| 2377 | |
| 2378 | Value *CreateInsertElement(Value *Vec, Value *NewElt, uint64_t Idx, |
| 2379 | const Twine &Name = "") { |
| 2380 | return CreateInsertElement(Vec, NewElt, getInt64(Idx), Name); |
| 2381 | } |
| 2382 | |
| 2383 | Value *CreateShuffleVector(Value *V1, Value *V2, Value *Mask, |
| 2384 | const Twine &Name = "") { |
| 2385 | SmallVector<int, 16> IntMask; |
| 2386 | ShuffleVectorInst::getShuffleMask(cast<Constant>(Mask), IntMask); |
| 2387 | return CreateShuffleVector(V1, V2, IntMask, Name); |
| 2388 | } |
| 2389 | |
| 2390 | /// See class ShuffleVectorInst for a description of the mask representation. |
| 2391 | Value *CreateShuffleVector(Value *V1, Value *V2, ArrayRef<int> Mask, |
| 2392 | const Twine &Name = "") { |
| 2393 | if (Value *V = Folder.FoldShuffleVector(V1, V2, Mask)) |
| 2394 | return V; |
| 2395 | return Insert(new ShuffleVectorInst(V1, V2, Mask), Name); |
| 2396 | } |
| 2397 | |
| 2398 | /// Create a unary shuffle. The second vector operand of the IR instruction |
| 2399 | /// is poison. |
| 2400 | Value *CreateShuffleVector(Value *V, ArrayRef<int> Mask, |
| 2401 | const Twine &Name = "") { |
| 2402 | return CreateShuffleVector(V, PoisonValue::get(V->getType()), Mask, Name); |
| 2403 | } |
| 2404 | |
| 2405 | Value *CreateExtractValue(Value *Agg, ArrayRef<unsigned> Idxs, |
| 2406 | const Twine &Name = "") { |
| 2407 | if (auto *V = Folder.FoldExtractValue(Agg, Idxs)) |
| 2408 | return V; |
| 2409 | return Insert(ExtractValueInst::Create(Agg, Idxs), Name); |
| 2410 | } |
| 2411 | |
| 2412 | Value *CreateInsertValue(Value *Agg, Value *Val, ArrayRef<unsigned> Idxs, |
| 2413 | const Twine &Name = "") { |
| 2414 | if (auto *V = Folder.FoldInsertValue(Agg, Val, Idxs)) |
| 2415 | return V; |
| 2416 | return Insert(InsertValueInst::Create(Agg, Val, Idxs), Name); |
| 2417 | } |
| 2418 | |
| 2419 | LandingPadInst *CreateLandingPad(Type *Ty, unsigned NumClauses, |
| 2420 | const Twine &Name = "") { |
| 2421 | return Insert(LandingPadInst::Create(Ty, NumClauses), Name); |
| 2422 | } |
| 2423 | |
| 2424 | Value *CreateFreeze(Value *V, const Twine &Name = "") { |
| 2425 | return Insert(new FreezeInst(V), Name); |
| 2426 | } |
| 2427 | |
| 2428 | //===--------------------------------------------------------------------===// |
| 2429 | // Utility creation methods |
| 2430 | //===--------------------------------------------------------------------===// |
| 2431 | |
| 2432 | /// Return a boolean value testing if \p Arg == 0. |
| 2433 | Value *CreateIsNull(Value *Arg, const Twine &Name = "") { |
| 2434 | return CreateICmpEQ(Arg, ConstantInt::getNullValue(Arg->getType()), Name); |
| 2435 | } |
| 2436 | |
| 2437 | /// Return a boolean value testing if \p Arg != 0. |
| 2438 | Value *CreateIsNotNull(Value *Arg, const Twine &Name = "") { |
| 2439 | return CreateICmpNE(Arg, ConstantInt::getNullValue(Arg->getType()), Name); |
| 2440 | } |
| 2441 | |
| 2442 | /// Return a boolean value testing if \p Arg < 0. |
| 2443 | Value *CreateIsNeg(Value *Arg, const Twine &Name = "") { |
| 2444 | return CreateICmpSLT(Arg, ConstantInt::getNullValue(Arg->getType()), Name); |
| 2445 | } |
| 2446 | |
| 2447 | /// Return a boolean value testing if \p Arg > -1. |
| 2448 | Value *CreateIsNotNeg(Value *Arg, const Twine &Name = "") { |
| 2449 | return CreateICmpSGT(Arg, ConstantInt::getAllOnesValue(Arg->getType()), |
| 2450 | Name); |
| 2451 | } |
| 2452 | |
| 2453 | /// Return the i64 difference between two pointer values, dividing out |
| 2454 | /// the size of the pointed-to objects. |
| 2455 | /// |
| 2456 | /// This is intended to implement C-style pointer subtraction. As such, the |
| 2457 | /// pointers must be appropriately aligned for their element types and |
| 2458 | /// pointing into the same object. |
| 2459 | Value *CreatePtrDiff(Type *ElemTy, Value *LHS, Value *RHS, |
| 2460 | const Twine &Name = ""); |
| 2461 | |
| 2462 | /// Create a launder.invariant.group intrinsic call. If Ptr type is |
| 2463 | /// different from pointer to i8, it's casted to pointer to i8 in the same |
| 2464 | /// address space before call and casted back to Ptr type after call. |
| 2465 | Value *CreateLaunderInvariantGroup(Value *Ptr); |
| 2466 | |
| 2467 | /// \brief Create a strip.invariant.group intrinsic call. If Ptr type is |
| 2468 | /// different from pointer to i8, it's casted to pointer to i8 in the same |
| 2469 | /// address space before call and casted back to Ptr type after call. |
| 2470 | Value *CreateStripInvariantGroup(Value *Ptr); |
| 2471 | |
| 2472 | /// Return a vector value that contains the vector V reversed |
| 2473 | Value *CreateVectorReverse(Value *V, const Twine &Name = ""); |
| 2474 | |
| 2475 | /// Return a vector splice intrinsic if using scalable vectors, otherwise |
| 2476 | /// return a shufflevector. If the immediate is positive, a vector is |
| 2477 | /// extracted from concat(V1, V2), starting at Imm. If the immediate |
| 2478 | /// is negative, we extract -Imm elements from V1 and the remaining |
| 2479 | /// elements from V2. Imm is a signed integer in the range |
| 2480 | /// -VL <= Imm < VL (where VL is the runtime vector length of the |
| 2481 | /// source/result vector) |
| 2482 | Value *CreateVectorSplice(Value *V1, Value *V2, int64_t Imm, |
| 2483 | const Twine &Name = ""); |
| 2484 | |
| 2485 | /// Return a vector value that contains \arg V broadcasted to \p |
| 2486 | /// NumElts elements. |
| 2487 | Value *CreateVectorSplat(unsigned NumElts, Value *V, const Twine &Name = ""); |
| 2488 | |
| 2489 | /// Return a vector value that contains \arg V broadcasted to \p |
| 2490 | /// EC elements. |
| 2491 | Value *CreateVectorSplat(ElementCount EC, Value *V, const Twine &Name = ""); |
| 2492 | |
| 2493 | /// Return a value that has been extracted from a larger integer type. |
| 2494 | Value *CreateExtractInteger(const DataLayout &DL, Value *From, |
| 2495 | IntegerType *ExtractedTy, uint64_t Offset, |
| 2496 | const Twine &Name); |
| 2497 | |
| 2498 | Value *CreatePreserveArrayAccessIndex(Type *ElTy, Value *Base, |
| 2499 | unsigned Dimension, unsigned LastIndex, |
| 2500 | MDNode *DbgInfo); |
| 2501 | |
| 2502 | Value *CreatePreserveUnionAccessIndex(Value *Base, unsigned FieldIndex, |
| 2503 | MDNode *DbgInfo); |
| 2504 | |
| 2505 | Value *CreatePreserveStructAccessIndex(Type *ElTy, Value *Base, |
| 2506 | unsigned Index, unsigned FieldIndex, |
| 2507 | MDNode *DbgInfo); |
| 2508 | |
| 2509 | private: |
| 2510 | /// Helper function that creates an assume intrinsic call that |
| 2511 | /// represents an alignment assumption on the provided pointer \p PtrValue |
| 2512 | /// with offset \p OffsetValue and alignment value \p AlignValue. |
| 2513 | CallInst *CreateAlignmentAssumptionHelper(const DataLayout &DL, |
| 2514 | Value *PtrValue, Value *AlignValue, |
| 2515 | Value *OffsetValue); |
| 2516 | |
| 2517 | public: |
| 2518 | /// Create an assume intrinsic call that represents an alignment |
| 2519 | /// assumption on the provided pointer. |
| 2520 | /// |
| 2521 | /// An optional offset can be provided, and if it is provided, the offset |
| 2522 | /// must be subtracted from the provided pointer to get the pointer with the |
| 2523 | /// specified alignment. |
| 2524 | CallInst *CreateAlignmentAssumption(const DataLayout &DL, Value *PtrValue, |
| 2525 | unsigned Alignment, |
| 2526 | Value *OffsetValue = nullptr); |
| 2527 | |
| 2528 | /// Create an assume intrinsic call that represents an alignment |
| 2529 | /// assumption on the provided pointer. |
| 2530 | /// |
| 2531 | /// An optional offset can be provided, and if it is provided, the offset |
| 2532 | /// must be subtracted from the provided pointer to get the pointer with the |
| 2533 | /// specified alignment. |
| 2534 | /// |
| 2535 | /// This overload handles the condition where the Alignment is dependent |
| 2536 | /// on an existing value rather than a static value. |
| 2537 | CallInst *CreateAlignmentAssumption(const DataLayout &DL, Value *PtrValue, |
| 2538 | Value *Alignment, |
| 2539 | Value *OffsetValue = nullptr); |
| 2540 | }; |
| 2541 | |
| 2542 | /// This provides a uniform API for creating instructions and inserting |
| 2543 | /// them into a basic block: either at the end of a BasicBlock, or at a specific |
| 2544 | /// iterator location in a block. |
| 2545 | /// |
| 2546 | /// Note that the builder does not expose the full generality of LLVM |
| 2547 | /// instructions. For access to extra instruction properties, use the mutators |
| 2548 | /// (e.g. setVolatile) on the instructions after they have been |
| 2549 | /// created. Convenience state exists to specify fast-math flags and fp-math |
| 2550 | /// tags. |
| 2551 | /// |
| 2552 | /// The first template argument specifies a class to use for creating constants. |
| 2553 | /// This defaults to creating minimally folded constants. The second template |
| 2554 | /// argument allows clients to specify custom insertion hooks that are called on |
| 2555 | /// every newly created insertion. |
| 2556 | template <typename FolderTy = ConstantFolder, |
| 2557 | typename InserterTy = IRBuilderDefaultInserter> |
| 2558 | class IRBuilder : public IRBuilderBase { |
| 2559 | private: |
| 2560 | FolderTy Folder; |
| 2561 | InserterTy Inserter; |
| 2562 | |
| 2563 | public: |
| 2564 | IRBuilder(LLVMContext &C, FolderTy Folder, InserterTy Inserter = InserterTy(), |
| 2565 | MDNode *FPMathTag = nullptr, |
| 2566 | ArrayRef<OperandBundleDef> OpBundles = std::nullopt) |
| 2567 | : IRBuilderBase(C, this->Folder, this->Inserter, FPMathTag, OpBundles), |
| 2568 | Folder(Folder), Inserter(Inserter) {} |
| 2569 | |
| 2570 | explicit IRBuilder(LLVMContext &C, MDNode *FPMathTag = nullptr, |
| 2571 | ArrayRef<OperandBundleDef> OpBundles = std::nullopt) |
| 2572 | : IRBuilderBase(C, this->Folder, this->Inserter, FPMathTag, OpBundles) {} |
| 2573 | |
| 2574 | explicit IRBuilder(BasicBlock *TheBB, FolderTy Folder, |
| 2575 | MDNode *FPMathTag = nullptr, |
| 2576 | ArrayRef<OperandBundleDef> OpBundles = std::nullopt) |
| 2577 | : IRBuilderBase(TheBB->getContext(), this->Folder, this->Inserter, |
| 2578 | FPMathTag, OpBundles), |
| 2579 | Folder(Folder) { |
| 2580 | SetInsertPoint(TheBB); |
| 2581 | } |
| 2582 | |
| 2583 | explicit IRBuilder(BasicBlock *TheBB, MDNode *FPMathTag = nullptr, |
| 2584 | ArrayRef<OperandBundleDef> OpBundles = std::nullopt) |
| 2585 | : IRBuilderBase(TheBB->getContext(), this->Folder, this->Inserter, |
| 2586 | FPMathTag, OpBundles) { |
| 2587 | SetInsertPoint(TheBB); |
| 2588 | } |
| 2589 | |
| 2590 | explicit IRBuilder(Instruction *IP, MDNode *FPMathTag = nullptr, |
| 2591 | ArrayRef<OperandBundleDef> OpBundles = std::nullopt) |
| 2592 | : IRBuilderBase(IP->getContext(), this->Folder, this->Inserter, FPMathTag, |
| 2593 | OpBundles) { |
| 2594 | SetInsertPoint(IP); |
| 2595 | } |
| 2596 | |
| 2597 | IRBuilder(BasicBlock *TheBB, BasicBlock::iterator IP, FolderTy Folder, |
| 2598 | MDNode *FPMathTag = nullptr, |
| 2599 | ArrayRef<OperandBundleDef> OpBundles = std::nullopt) |
| 2600 | : IRBuilderBase(TheBB->getContext(), this->Folder, this->Inserter, |
| 2601 | FPMathTag, OpBundles), |
| 2602 | Folder(Folder) { |
| 2603 | SetInsertPoint(TheBB, IP); |
| 2604 | } |
| 2605 | |
| 2606 | IRBuilder(BasicBlock *TheBB, BasicBlock::iterator IP, |
| 2607 | MDNode *FPMathTag = nullptr, |
| 2608 | ArrayRef<OperandBundleDef> OpBundles = std::nullopt) |
| 2609 | : IRBuilderBase(TheBB->getContext(), this->Folder, this->Inserter, |
| 2610 | FPMathTag, OpBundles) { |
| 2611 | SetInsertPoint(TheBB, IP); |
| 2612 | } |
| 2613 | |
| 2614 | /// Avoid copying the full IRBuilder. Prefer using InsertPointGuard |
| 2615 | /// or FastMathFlagGuard instead. |
| 2616 | IRBuilder(const IRBuilder &) = delete; |
| 2617 | |
| 2618 | InserterTy &getInserter() { return Inserter; } |
| 2619 | }; |
| 2620 | |
| 2621 | template <typename FolderTy, typename InserterTy> |
| 2622 | IRBuilder(LLVMContext &, FolderTy, InserterTy, MDNode *, |
| 2623 | ArrayRef<OperandBundleDef>) -> IRBuilder<FolderTy, InserterTy>; |
| 2624 | IRBuilder(LLVMContext &, MDNode *, ArrayRef<OperandBundleDef>) -> IRBuilder<>; |
| 2625 | template <typename FolderTy> |
| 2626 | IRBuilder(BasicBlock *, FolderTy, MDNode *, ArrayRef<OperandBundleDef>) |
| 2627 | -> IRBuilder<FolderTy>; |
| 2628 | IRBuilder(BasicBlock *, MDNode *, ArrayRef<OperandBundleDef>) -> IRBuilder<>; |
| 2629 | IRBuilder(Instruction *, MDNode *, ArrayRef<OperandBundleDef>) -> IRBuilder<>; |
| 2630 | template <typename FolderTy> |
| 2631 | IRBuilder(BasicBlock *, BasicBlock::iterator, FolderTy, MDNode *, |
| 2632 | ArrayRef<OperandBundleDef>) -> IRBuilder<FolderTy>; |
| 2633 | IRBuilder(BasicBlock *, BasicBlock::iterator, MDNode *, |
| 2634 | ArrayRef<OperandBundleDef>) -> IRBuilder<>; |
| 2635 | |
| 2636 | |
| 2637 | // Create wrappers for C Binding types (see CBindingWrapping.h). |
| 2638 | DEFINE_SIMPLE_CONVERSION_FUNCTIONS(IRBuilder<>, LLVMBuilderRef)inline IRBuilder<> *unwrap(LLVMBuilderRef P) { return reinterpret_cast <IRBuilder<>*>(P); } inline LLVMBuilderRef wrap(const IRBuilder<> *P) { return reinterpret_cast<LLVMBuilderRef >(const_cast<IRBuilder<>*>(P)); } |
| 2639 | |
| 2640 | } // end namespace llvm |
| 2641 | |
| 2642 | #endif // LLVM_IR_IRBUILDER_H |