| 1 | //===-- Operator.cpp - Implement the LLVM operators -----------------------===// |
| 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 implements the non-inline methods for the LLVM Operator classes. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "llvm/IR/Operator.h" |
| 14 | #include "llvm/IR/DataLayout.h" |
| 15 | #include "llvm/IR/GetElementPtrTypeIterator.h" |
| 16 | #include "llvm/IR/Instructions.h" |
| 17 | #include "llvm/IR/IntrinsicInst.h" |
| 18 | |
| 19 | #include "ConstantsContext.h" |
| 20 | |
| 21 | using namespace llvm; |
| 22 | |
| 23 | bool Operator::hasPoisonGeneratingFlags() const { |
| 24 | switch (getOpcode()) { |
| 25 | case Instruction::Add: |
| 26 | case Instruction::Sub: |
| 27 | case Instruction::Mul: |
| 28 | case Instruction::Shl: { |
| 29 | auto *OBO = cast<OverflowingBinaryOperator>(Val: this); |
| 30 | return OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap(); |
| 31 | } |
| 32 | case Instruction::Trunc: { |
| 33 | if (auto *TI = dyn_cast<TruncInst>(Val: this)) |
| 34 | return TI->hasNoUnsignedWrap() || TI->hasNoSignedWrap(); |
| 35 | return false; |
| 36 | } |
| 37 | case Instruction::UDiv: |
| 38 | case Instruction::SDiv: |
| 39 | case Instruction::AShr: |
| 40 | case Instruction::LShr: |
| 41 | return cast<PossiblyExactOperator>(Val: this)->isExact(); |
| 42 | case Instruction::Or: |
| 43 | return cast<PossiblyDisjointInst>(Val: this)->isDisjoint(); |
| 44 | case Instruction::GetElementPtr: { |
| 45 | auto *GEP = cast<GEPOperator>(Val: this); |
| 46 | // Note: inrange exists on constexpr only |
| 47 | return GEP->getNoWrapFlags() != GEPNoWrapFlags::none() || |
| 48 | GEP->getInRange() != std::nullopt; |
| 49 | } |
| 50 | case Instruction::UIToFP: |
| 51 | case Instruction::ZExt: |
| 52 | if (auto *NNI = dyn_cast<PossiblyNonNegInst>(Val: this)) |
| 53 | return NNI->hasNonNeg(); |
| 54 | return false; |
| 55 | case Instruction::ICmp: |
| 56 | return cast<ICmpInst>(Val: this)->hasSameSign(); |
| 57 | case Instruction::Call: |
| 58 | if (auto *II = dyn_cast<IntrinsicInst>(Val: this)) { |
| 59 | switch (II->getIntrinsicID()) { |
| 60 | case Intrinsic::ctlz: |
| 61 | case Intrinsic::cttz: |
| 62 | case Intrinsic::abs: |
| 63 | return cast<ConstantInt>(Val: II->getArgOperand(i: 1))->isOneValue(); |
| 64 | } |
| 65 | } |
| 66 | [[fallthrough]]; |
| 67 | default: |
| 68 | if (const auto *FP = dyn_cast<FPMathOperator>(Val: this)) |
| 69 | return FP->hasNoNaNs() || FP->hasNoInfs(); |
| 70 | return false; |
| 71 | } |
| 72 | } |
| 73 | |
| 74 | bool Operator::hasPoisonGeneratingAnnotations() const { |
| 75 | if (hasPoisonGeneratingFlags()) |
| 76 | return true; |
| 77 | auto *I = dyn_cast<Instruction>(Val: this); |
| 78 | return I && (I->hasPoisonGeneratingAttributes() || |
| 79 | I->hasPoisonGeneratingMetadata()); |
| 80 | } |
| 81 | |
| 82 | Type *GEPOperator::getSourceElementType() const { |
| 83 | if (auto *I = dyn_cast<GetElementPtrInst>(Val: this)) |
| 84 | return I->getSourceElementType(); |
| 85 | return cast<GetElementPtrConstantExpr>(Val: this)->getSourceElementType(); |
| 86 | } |
| 87 | |
| 88 | Type *GEPOperator::getResultElementType() const { |
| 89 | if (auto *I = dyn_cast<GetElementPtrInst>(Val: this)) |
| 90 | return I->getResultElementType(); |
| 91 | return cast<GetElementPtrConstantExpr>(Val: this)->getResultElementType(); |
| 92 | } |
| 93 | |
| 94 | std::optional<ConstantRange> GEPOperator::getInRange() const { |
| 95 | if (auto *CE = dyn_cast<GetElementPtrConstantExpr>(Val: this)) |
| 96 | return CE->getInRange(); |
| 97 | return std::nullopt; |
| 98 | } |
| 99 | |
| 100 | Align GEPOperator::getMaxPreservedAlignment(const DataLayout &DL) const { |
| 101 | /// compute the worse possible offset for every level of the GEP et accumulate |
| 102 | /// the minimum alignment into Result. |
| 103 | |
| 104 | Align Result = Align(llvm::Value::MaximumAlignment); |
| 105 | for (gep_type_iterator GTI = gep_type_begin(GEP: this), GTE = gep_type_end(GEP: this); |
| 106 | GTI != GTE; ++GTI) { |
| 107 | uint64_t Offset; |
| 108 | ConstantInt *OpC = dyn_cast<ConstantInt>(Val: GTI.getOperand()); |
| 109 | |
| 110 | if (StructType *STy = GTI.getStructTypeOrNull()) { |
| 111 | const StructLayout *SL = DL.getStructLayout(Ty: STy); |
| 112 | Offset = |
| 113 | SL->getElementOffset(Idx: OpC->getValue().getLoBits(numBits: 32).getZExtValue()); |
| 114 | } else { |
| 115 | assert(GTI.isSequential() && "should be sequencial" ); |
| 116 | /// If the index isn't known, we take 1 because it is the index that will |
| 117 | /// give the worse alignment of the offset. |
| 118 | const uint64_t ElemCount = OpC ? OpC->getLimitedValue() : 1; |
| 119 | Offset = GTI.getSequentialElementStride(DL) * ElemCount; |
| 120 | } |
| 121 | Result = Align(MinAlign(A: Offset, B: Result.value())); |
| 122 | } |
| 123 | return Result; |
| 124 | } |
| 125 | |
| 126 | bool GEPOperator::accumulateConstantOffset( |
| 127 | const DataLayout &DL, APInt &Offset, |
| 128 | function_ref<bool(Value &, APInt &)> ExternalAnalysis) const { |
| 129 | assert(Offset.getBitWidth() == |
| 130 | DL.getIndexSizeInBits(getPointerAddressSpace()) && |
| 131 | "The offset bit width does not match DL specification." ); |
| 132 | SmallVector<const Value *> Index(llvm::drop_begin(RangeOrContainer: operand_values())); |
| 133 | return GEPOperator::accumulateConstantOffset(SourceType: getSourceElementType(), Index, |
| 134 | DL, Offset, ExternalAnalysis); |
| 135 | } |
| 136 | |
| 137 | bool GEPOperator::accumulateConstantOffset( |
| 138 | Type *SourceType, ArrayRef<const Value *> Index, const DataLayout &DL, |
| 139 | APInt &Offset, function_ref<bool(Value &, APInt &)> ExternalAnalysis) { |
| 140 | // Fast path for canonical getelementptr i8 form. |
| 141 | if (SourceType->isIntegerTy(BitWidth: 8) && !Index.empty() && !ExternalAnalysis) { |
| 142 | auto *CI = dyn_cast<ConstantInt>(Val: Index.front()); |
| 143 | if (CI && CI->getType()->isIntegerTy()) { |
| 144 | Offset += CI->getValue().sextOrTrunc(width: Offset.getBitWidth()); |
| 145 | return true; |
| 146 | } |
| 147 | return false; |
| 148 | } |
| 149 | |
| 150 | bool UsedExternalAnalysis = false; |
| 151 | auto AccumulateOffset = [&](APInt Index, uint64_t Size) -> bool { |
| 152 | Index = Index.sextOrTrunc(width: Offset.getBitWidth()); |
| 153 | // Truncate if type size exceeds index space. |
| 154 | APInt IndexedSize(Offset.getBitWidth(), Size, /*isSigned=*/false, |
| 155 | /*implcitTrunc=*/true); |
| 156 | // For array or vector indices, scale the index by the size of the type. |
| 157 | if (!UsedExternalAnalysis) { |
| 158 | Offset += Index * IndexedSize; |
| 159 | } else { |
| 160 | // External Analysis can return a result higher/lower than the value |
| 161 | // represents. We need to detect overflow/underflow. |
| 162 | bool Overflow = false; |
| 163 | APInt OffsetPlus = Index.smul_ov(RHS: IndexedSize, Overflow); |
| 164 | if (Overflow) |
| 165 | return false; |
| 166 | Offset = Offset.sadd_ov(RHS: OffsetPlus, Overflow); |
| 167 | if (Overflow) |
| 168 | return false; |
| 169 | } |
| 170 | return true; |
| 171 | }; |
| 172 | auto begin = generic_gep_type_iterator<decltype(Index.begin())>::begin( |
| 173 | Ty: SourceType, It: Index.begin()); |
| 174 | auto end = generic_gep_type_iterator<decltype(Index.end())>::end(It: Index.end()); |
| 175 | for (auto GTI = begin, GTE = end; GTI != GTE; ++GTI) { |
| 176 | // Scalable vectors are multiplied by a runtime constant. |
| 177 | bool ScalableType = GTI.getIndexedType()->isScalableTy(); |
| 178 | |
| 179 | Value *V = GTI.getOperand(); |
| 180 | StructType *STy = GTI.getStructTypeOrNull(); |
| 181 | // Handle ConstantInt if possible. |
| 182 | auto *ConstOffset = dyn_cast<ConstantInt>(Val: V); |
| 183 | if (ConstOffset && ConstOffset->getType()->isIntegerTy()) { |
| 184 | if (ConstOffset->isZero()) |
| 185 | continue; |
| 186 | // if the type is scalable and the constant is not zero (vscale * n * 0 = |
| 187 | // 0) bailout. |
| 188 | if (ScalableType) |
| 189 | return false; |
| 190 | // Handle a struct index, which adds its field offset to the pointer. |
| 191 | if (STy) { |
| 192 | unsigned ElementIdx = ConstOffset->getZExtValue(); |
| 193 | const StructLayout *SL = DL.getStructLayout(Ty: STy); |
| 194 | // Element offset is in bytes. |
| 195 | if (!AccumulateOffset( |
| 196 | APInt(Offset.getBitWidth(), SL->getElementOffset(Idx: ElementIdx)), |
| 197 | 1)) |
| 198 | return false; |
| 199 | continue; |
| 200 | } |
| 201 | if (!AccumulateOffset(ConstOffset->getValue(), |
| 202 | GTI.getSequentialElementStride(DL))) |
| 203 | return false; |
| 204 | continue; |
| 205 | } |
| 206 | |
| 207 | // The operand is not constant, check if an external analysis was provided. |
| 208 | // External analsis is not applicable to a struct type. |
| 209 | if (!ExternalAnalysis || STy || ScalableType) |
| 210 | return false; |
| 211 | APInt AnalysisIndex; |
| 212 | if (!ExternalAnalysis(*V, AnalysisIndex)) |
| 213 | return false; |
| 214 | UsedExternalAnalysis = true; |
| 215 | if (!AccumulateOffset(AnalysisIndex, GTI.getSequentialElementStride(DL))) |
| 216 | return false; |
| 217 | } |
| 218 | return true; |
| 219 | } |
| 220 | |
| 221 | bool GEPOperator::collectOffset( |
| 222 | const DataLayout &DL, unsigned BitWidth, |
| 223 | SmallMapVector<Value *, APInt, 4> &VariableOffsets, |
| 224 | APInt &ConstantOffset) const { |
| 225 | assert(BitWidth == DL.getIndexSizeInBits(getPointerAddressSpace()) && |
| 226 | "The offset bit width does not match DL specification." ); |
| 227 | |
| 228 | auto CollectConstantOffset = [&](APInt Index, uint64_t Size) { |
| 229 | Index = Index.sextOrTrunc(width: BitWidth); |
| 230 | // Truncate if type size exceeds index space. |
| 231 | APInt IndexedSize(BitWidth, Size, /*isSigned=*/false, |
| 232 | /*implcitTrunc=*/true); |
| 233 | ConstantOffset += Index * IndexedSize; |
| 234 | }; |
| 235 | |
| 236 | for (gep_type_iterator GTI = gep_type_begin(GEP: this), GTE = gep_type_end(GEP: this); |
| 237 | GTI != GTE; ++GTI) { |
| 238 | // Scalable vectors are multiplied by a runtime constant. |
| 239 | bool ScalableType = GTI.getIndexedType()->isScalableTy(); |
| 240 | |
| 241 | Value *V = GTI.getOperand(); |
| 242 | StructType *STy = GTI.getStructTypeOrNull(); |
| 243 | // Handle ConstantInt if possible. |
| 244 | auto *ConstOffset = dyn_cast<ConstantInt>(Val: V); |
| 245 | if (ConstOffset && ConstOffset->getType()->isIntegerTy()) { |
| 246 | if (ConstOffset->isZero()) |
| 247 | continue; |
| 248 | // If the type is scalable and the constant is not zero (vscale * n * 0 = |
| 249 | // 0) bailout. |
| 250 | // TODO: If the runtime value is accessible at any point before DWARF |
| 251 | // emission, then we could potentially keep a forward reference to it |
| 252 | // in the debug value to be filled in later. |
| 253 | if (ScalableType) |
| 254 | return false; |
| 255 | // Handle a struct index, which adds its field offset to the pointer. |
| 256 | if (STy) { |
| 257 | unsigned ElementIdx = ConstOffset->getZExtValue(); |
| 258 | const StructLayout *SL = DL.getStructLayout(Ty: STy); |
| 259 | // Element offset is in bytes. |
| 260 | CollectConstantOffset(APInt(BitWidth, SL->getElementOffset(Idx: ElementIdx)), |
| 261 | 1); |
| 262 | continue; |
| 263 | } |
| 264 | CollectConstantOffset(ConstOffset->getValue(), |
| 265 | GTI.getSequentialElementStride(DL)); |
| 266 | continue; |
| 267 | } |
| 268 | |
| 269 | if (STy || ScalableType) |
| 270 | return false; |
| 271 | // Truncate if type size exceeds index space. |
| 272 | APInt IndexedSize(BitWidth, GTI.getSequentialElementStride(DL), |
| 273 | /*isSigned=*/false, /*implicitTrunc=*/true); |
| 274 | // Insert an initial offset of 0 for V iff none exists already, then |
| 275 | // increment the offset by IndexedSize. |
| 276 | if (!IndexedSize.isZero()) { |
| 277 | auto *It = VariableOffsets.insert(KV: {V, APInt(BitWidth, 0)}).first; |
| 278 | It->second += IndexedSize; |
| 279 | } |
| 280 | } |
| 281 | return true; |
| 282 | } |
| 283 | |
| 284 | void FastMathFlags::print(raw_ostream &O) const { |
| 285 | if (all()) |
| 286 | O << " fast" ; |
| 287 | else { |
| 288 | if (allowReassoc()) |
| 289 | O << " reassoc" ; |
| 290 | if (noNaNs()) |
| 291 | O << " nnan" ; |
| 292 | if (noInfs()) |
| 293 | O << " ninf" ; |
| 294 | if (noSignedZeros()) |
| 295 | O << " nsz" ; |
| 296 | if (allowReciprocal()) |
| 297 | O << " arcp" ; |
| 298 | if (allowContract()) |
| 299 | O << " contract" ; |
| 300 | if (approxFunc()) |
| 301 | O << " afn" ; |
| 302 | } |
| 303 | } |
| 304 | |
| 305 | FastMathFlags &FPMathOperator::getFastMathFlagsImpl() { |
| 306 | auto *I = cast<Instruction>(Val: this); |
| 307 | |
| 308 | if (FastMathFlagsStorage *Op = dyn_cast<FPUnaryOperator>(Val: I)) |
| 309 | return Op->FMF; |
| 310 | if (FastMathFlagsStorage *Op = dyn_cast<FPBinaryOperator>(Val: I)) |
| 311 | return Op->FMF; |
| 312 | if (FastMathFlagsStorage *Op = dyn_cast<FPTruncInst>(Val: I)) |
| 313 | return Op->FMF; |
| 314 | if (FastMathFlagsStorage *Op = dyn_cast<FPExtInst>(Val: I)) |
| 315 | return Op->FMF; |
| 316 | if (FastMathFlagsStorage *Op = dyn_cast<FCmpInst>(Val: I)) |
| 317 | return Op->FMF; |
| 318 | if (FastMathFlagsStorage *Op = dyn_cast<PHINode>(Val: I)) |
| 319 | return Op->FMF; |
| 320 | if (FastMathFlagsStorage *Op = dyn_cast<SelectInst>(Val: I)) |
| 321 | return Op->FMF; |
| 322 | if (FastMathFlagsStorage *Op = dyn_cast<CallInst>(Val: I)) |
| 323 | return Op->FMF; |
| 324 | if (FastMathFlagsStorage *Op = dyn_cast<UIToFPInst>(Val: I)) |
| 325 | return Op->FMF; |
| 326 | if (FastMathFlagsStorage *Op = dyn_cast<SIToFPInst>(Val: I)) |
| 327 | return Op->FMF; |
| 328 | |
| 329 | llvm_unreachable("Unknown FPMathOperator!" ); |
| 330 | } |
| 331 | |