| File: | build/source/llvm/lib/Analysis/ValueTracking.cpp |
| Warning: | line 5641, column 22 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===- ValueTracking.cpp - Walk computations to compute properties --------===// | |||
| 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 contains routines that help analyze properties that chains of | |||
| 10 | // computations have. | |||
| 11 | // | |||
| 12 | //===----------------------------------------------------------------------===// | |||
| 13 | ||||
| 14 | #include "llvm/Analysis/ValueTracking.h" | |||
| 15 | #include "llvm/ADT/APFloat.h" | |||
| 16 | #include "llvm/ADT/APInt.h" | |||
| 17 | #include "llvm/ADT/ArrayRef.h" | |||
| 18 | #include "llvm/ADT/STLExtras.h" | |||
| 19 | #include "llvm/ADT/ScopeExit.h" | |||
| 20 | #include "llvm/ADT/SmallPtrSet.h" | |||
| 21 | #include "llvm/ADT/SmallSet.h" | |||
| 22 | #include "llvm/ADT/SmallVector.h" | |||
| 23 | #include "llvm/ADT/StringRef.h" | |||
| 24 | #include "llvm/ADT/iterator_range.h" | |||
| 25 | #include "llvm/Analysis/AliasAnalysis.h" | |||
| 26 | #include "llvm/Analysis/AssumeBundleQueries.h" | |||
| 27 | #include "llvm/Analysis/AssumptionCache.h" | |||
| 28 | #include "llvm/Analysis/ConstantFolding.h" | |||
| 29 | #include "llvm/Analysis/GuardUtils.h" | |||
| 30 | #include "llvm/Analysis/InstructionSimplify.h" | |||
| 31 | #include "llvm/Analysis/Loads.h" | |||
| 32 | #include "llvm/Analysis/LoopInfo.h" | |||
| 33 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | |||
| 34 | #include "llvm/Analysis/TargetLibraryInfo.h" | |||
| 35 | #include "llvm/Analysis/VectorUtils.h" | |||
| 36 | #include "llvm/IR/Argument.h" | |||
| 37 | #include "llvm/IR/Attributes.h" | |||
| 38 | #include "llvm/IR/BasicBlock.h" | |||
| 39 | #include "llvm/IR/Constant.h" | |||
| 40 | #include "llvm/IR/ConstantRange.h" | |||
| 41 | #include "llvm/IR/Constants.h" | |||
| 42 | #include "llvm/IR/DerivedTypes.h" | |||
| 43 | #include "llvm/IR/DiagnosticInfo.h" | |||
| 44 | #include "llvm/IR/Dominators.h" | |||
| 45 | #include "llvm/IR/EHPersonalities.h" | |||
| 46 | #include "llvm/IR/Function.h" | |||
| 47 | #include "llvm/IR/GetElementPtrTypeIterator.h" | |||
| 48 | #include "llvm/IR/GlobalAlias.h" | |||
| 49 | #include "llvm/IR/GlobalValue.h" | |||
| 50 | #include "llvm/IR/GlobalVariable.h" | |||
| 51 | #include "llvm/IR/InstrTypes.h" | |||
| 52 | #include "llvm/IR/Instruction.h" | |||
| 53 | #include "llvm/IR/Instructions.h" | |||
| 54 | #include "llvm/IR/IntrinsicInst.h" | |||
| 55 | #include "llvm/IR/Intrinsics.h" | |||
| 56 | #include "llvm/IR/IntrinsicsAArch64.h" | |||
| 57 | #include "llvm/IR/IntrinsicsRISCV.h" | |||
| 58 | #include "llvm/IR/IntrinsicsX86.h" | |||
| 59 | #include "llvm/IR/LLVMContext.h" | |||
| 60 | #include "llvm/IR/Metadata.h" | |||
| 61 | #include "llvm/IR/Module.h" | |||
| 62 | #include "llvm/IR/Operator.h" | |||
| 63 | #include "llvm/IR/PatternMatch.h" | |||
| 64 | #include "llvm/IR/Type.h" | |||
| 65 | #include "llvm/IR/User.h" | |||
| 66 | #include "llvm/IR/Value.h" | |||
| 67 | #include "llvm/Support/Casting.h" | |||
| 68 | #include "llvm/Support/CommandLine.h" | |||
| 69 | #include "llvm/Support/Compiler.h" | |||
| 70 | #include "llvm/Support/ErrorHandling.h" | |||
| 71 | #include "llvm/Support/KnownBits.h" | |||
| 72 | #include "llvm/Support/MathExtras.h" | |||
| 73 | #include <algorithm> | |||
| 74 | #include <cassert> | |||
| 75 | #include <cstdint> | |||
| 76 | #include <optional> | |||
| 77 | #include <utility> | |||
| 78 | ||||
| 79 | using namespace llvm; | |||
| 80 | using namespace llvm::PatternMatch; | |||
| 81 | ||||
| 82 | // Controls the number of uses of the value searched for possible | |||
| 83 | // dominating comparisons. | |||
| 84 | static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses", | |||
| 85 | cl::Hidden, cl::init(20)); | |||
| 86 | ||||
| 87 | ||||
| 88 | /// Returns the bitwidth of the given scalar or pointer type. For vector types, | |||
| 89 | /// returns the element type's bitwidth. | |||
| 90 | static unsigned getBitWidth(Type *Ty, const DataLayout &DL) { | |||
| 91 | if (unsigned BitWidth = Ty->getScalarSizeInBits()) | |||
| 92 | return BitWidth; | |||
| 93 | ||||
| 94 | return DL.getPointerTypeSizeInBits(Ty); | |||
| 95 | } | |||
| 96 | ||||
| 97 | namespace { | |||
| 98 | ||||
| 99 | // Simplifying using an assume can only be done in a particular control-flow | |||
| 100 | // context (the context instruction provides that context). If an assume and | |||
| 101 | // the context instruction are not in the same block then the DT helps in | |||
| 102 | // figuring out if we can use it. | |||
| 103 | struct Query { | |||
| 104 | const DataLayout &DL; | |||
| 105 | AssumptionCache *AC; | |||
| 106 | const Instruction *CxtI; | |||
| 107 | const DominatorTree *DT; | |||
| 108 | ||||
| 109 | // Unlike the other analyses, this may be a nullptr because not all clients | |||
| 110 | // provide it currently. | |||
| 111 | OptimizationRemarkEmitter *ORE; | |||
| 112 | ||||
| 113 | /// If true, it is safe to use metadata during simplification. | |||
| 114 | InstrInfoQuery IIQ; | |||
| 115 | ||||
| 116 | Query(const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI, | |||
| 117 | const DominatorTree *DT, bool UseInstrInfo, | |||
| 118 | OptimizationRemarkEmitter *ORE = nullptr) | |||
| 119 | : DL(DL), AC(AC), CxtI(CxtI), DT(DT), ORE(ORE), IIQ(UseInstrInfo) {} | |||
| 120 | }; | |||
| 121 | ||||
| 122 | } // end anonymous namespace | |||
| 123 | ||||
| 124 | // Given the provided Value and, potentially, a context instruction, return | |||
| 125 | // the preferred context instruction (if any). | |||
| 126 | static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) { | |||
| 127 | // If we've been provided with a context instruction, then use that (provided | |||
| 128 | // it has been inserted). | |||
| 129 | if (CxtI && CxtI->getParent()) | |||
| 130 | return CxtI; | |||
| 131 | ||||
| 132 | // If the value is really an already-inserted instruction, then use that. | |||
| 133 | CxtI = dyn_cast<Instruction>(V); | |||
| 134 | if (CxtI && CxtI->getParent()) | |||
| 135 | return CxtI; | |||
| 136 | ||||
| 137 | return nullptr; | |||
| 138 | } | |||
| 139 | ||||
| 140 | static const Instruction *safeCxtI(const Value *V1, const Value *V2, const Instruction *CxtI) { | |||
| 141 | // If we've been provided with a context instruction, then use that (provided | |||
| 142 | // it has been inserted). | |||
| 143 | if (CxtI && CxtI->getParent()) | |||
| 144 | return CxtI; | |||
| 145 | ||||
| 146 | // If the value is really an already-inserted instruction, then use that. | |||
| 147 | CxtI = dyn_cast<Instruction>(V1); | |||
| 148 | if (CxtI && CxtI->getParent()) | |||
| 149 | return CxtI; | |||
| 150 | ||||
| 151 | CxtI = dyn_cast<Instruction>(V2); | |||
| 152 | if (CxtI && CxtI->getParent()) | |||
| 153 | return CxtI; | |||
| 154 | ||||
| 155 | return nullptr; | |||
| 156 | } | |||
| 157 | ||||
| 158 | static bool getShuffleDemandedElts(const ShuffleVectorInst *Shuf, | |||
| 159 | const APInt &DemandedElts, | |||
| 160 | APInt &DemandedLHS, APInt &DemandedRHS) { | |||
| 161 | if (isa<ScalableVectorType>(Shuf->getType())) { | |||
| 162 | assert(DemandedElts == APInt(1,1))(static_cast <bool> (DemandedElts == APInt(1,1)) ? void (0) : __assert_fail ("DemandedElts == APInt(1,1)", "llvm/lib/Analysis/ValueTracking.cpp" , 162, __extension__ __PRETTY_FUNCTION__)); | |||
| 163 | DemandedLHS = DemandedRHS = DemandedElts; | |||
| 164 | return true; | |||
| 165 | } | |||
| 166 | ||||
| 167 | int NumElts = | |||
| 168 | cast<FixedVectorType>(Shuf->getOperand(0)->getType())->getNumElements(); | |||
| 169 | return llvm::getShuffleDemandedElts(NumElts, Shuf->getShuffleMask(), | |||
| 170 | DemandedElts, DemandedLHS, DemandedRHS); | |||
| 171 | } | |||
| 172 | ||||
| 173 | static void computeKnownBits(const Value *V, const APInt &DemandedElts, | |||
| 174 | KnownBits &Known, unsigned Depth, const Query &Q); | |||
| 175 | ||||
| 176 | static void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth, | |||
| 177 | const Query &Q) { | |||
| 178 | // Since the number of lanes in a scalable vector is unknown at compile time, | |||
| 179 | // we track one bit which is implicitly broadcast to all lanes. This means | |||
| 180 | // that all lanes in a scalable vector are considered demanded. | |||
| 181 | auto *FVTy = dyn_cast<FixedVectorType>(V->getType()); | |||
| 182 | APInt DemandedElts = | |||
| 183 | FVTy ? APInt::getAllOnes(FVTy->getNumElements()) : APInt(1, 1); | |||
| 184 | computeKnownBits(V, DemandedElts, Known, Depth, Q); | |||
| 185 | } | |||
| 186 | ||||
| 187 | void llvm::computeKnownBits(const Value *V, KnownBits &Known, | |||
| 188 | const DataLayout &DL, unsigned Depth, | |||
| 189 | AssumptionCache *AC, const Instruction *CxtI, | |||
| 190 | const DominatorTree *DT, | |||
| 191 | OptimizationRemarkEmitter *ORE, bool UseInstrInfo) { | |||
| 192 | ::computeKnownBits(V, Known, Depth, | |||
| 193 | Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); | |||
| 194 | } | |||
| 195 | ||||
| 196 | void llvm::computeKnownBits(const Value *V, const APInt &DemandedElts, | |||
| 197 | KnownBits &Known, const DataLayout &DL, | |||
| 198 | unsigned Depth, AssumptionCache *AC, | |||
| 199 | const Instruction *CxtI, const DominatorTree *DT, | |||
| 200 | OptimizationRemarkEmitter *ORE, bool UseInstrInfo) { | |||
| 201 | ::computeKnownBits(V, DemandedElts, Known, Depth, | |||
| 202 | Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); | |||
| 203 | } | |||
| 204 | ||||
| 205 | static KnownBits computeKnownBits(const Value *V, const APInt &DemandedElts, | |||
| 206 | unsigned Depth, const Query &Q); | |||
| 207 | ||||
| 208 | static KnownBits computeKnownBits(const Value *V, unsigned Depth, | |||
| 209 | const Query &Q); | |||
| 210 | ||||
| 211 | KnownBits llvm::computeKnownBits(const Value *V, const DataLayout &DL, | |||
| 212 | unsigned Depth, AssumptionCache *AC, | |||
| 213 | const Instruction *CxtI, | |||
| 214 | const DominatorTree *DT, | |||
| 215 | OptimizationRemarkEmitter *ORE, | |||
| 216 | bool UseInstrInfo) { | |||
| 217 | return ::computeKnownBits( | |||
| 218 | V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); | |||
| 219 | } | |||
| 220 | ||||
| 221 | KnownBits llvm::computeKnownBits(const Value *V, const APInt &DemandedElts, | |||
| 222 | const DataLayout &DL, unsigned Depth, | |||
| 223 | AssumptionCache *AC, const Instruction *CxtI, | |||
| 224 | const DominatorTree *DT, | |||
| 225 | OptimizationRemarkEmitter *ORE, | |||
| 226 | bool UseInstrInfo) { | |||
| 227 | return ::computeKnownBits( | |||
| 228 | V, DemandedElts, Depth, | |||
| 229 | Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE)); | |||
| 230 | } | |||
| 231 | ||||
| 232 | bool llvm::haveNoCommonBitsSet(const Value *LHS, const Value *RHS, | |||
| 233 | const DataLayout &DL, AssumptionCache *AC, | |||
| 234 | const Instruction *CxtI, const DominatorTree *DT, | |||
| 235 | bool UseInstrInfo) { | |||
| 236 | assert(LHS->getType() == RHS->getType() &&(static_cast <bool> (LHS->getType() == RHS->getType () && "LHS and RHS should have the same type") ? void (0) : __assert_fail ("LHS->getType() == RHS->getType() && \"LHS and RHS should have the same type\"" , "llvm/lib/Analysis/ValueTracking.cpp", 237, __extension__ __PRETTY_FUNCTION__ )) | |||
| 237 | "LHS and RHS should have the same type")(static_cast <bool> (LHS->getType() == RHS->getType () && "LHS and RHS should have the same type") ? void (0) : __assert_fail ("LHS->getType() == RHS->getType() && \"LHS and RHS should have the same type\"" , "llvm/lib/Analysis/ValueTracking.cpp", 237, __extension__ __PRETTY_FUNCTION__ )); | |||
| 238 | assert(LHS->getType()->isIntOrIntVectorTy() &&(static_cast <bool> (LHS->getType()->isIntOrIntVectorTy () && "LHS and RHS should be integers") ? void (0) : __assert_fail ("LHS->getType()->isIntOrIntVectorTy() && \"LHS and RHS should be integers\"" , "llvm/lib/Analysis/ValueTracking.cpp", 239, __extension__ __PRETTY_FUNCTION__ )) | |||
| 239 | "LHS and RHS should be integers")(static_cast <bool> (LHS->getType()->isIntOrIntVectorTy () && "LHS and RHS should be integers") ? void (0) : __assert_fail ("LHS->getType()->isIntOrIntVectorTy() && \"LHS and RHS should be integers\"" , "llvm/lib/Analysis/ValueTracking.cpp", 239, __extension__ __PRETTY_FUNCTION__ )); | |||
| 240 | // Look for an inverted mask: (X & ~M) op (Y & M). | |||
| 241 | { | |||
| 242 | Value *M; | |||
| 243 | if (match(LHS, m_c_And(m_Not(m_Value(M)), m_Value())) && | |||
| 244 | match(RHS, m_c_And(m_Specific(M), m_Value()))) | |||
| 245 | return true; | |||
| 246 | if (match(RHS, m_c_And(m_Not(m_Value(M)), m_Value())) && | |||
| 247 | match(LHS, m_c_And(m_Specific(M), m_Value()))) | |||
| 248 | return true; | |||
| 249 | } | |||
| 250 | ||||
| 251 | // X op (Y & ~X) | |||
| 252 | if (match(RHS, m_c_And(m_Not(m_Specific(LHS)), m_Value())) || | |||
| 253 | match(LHS, m_c_And(m_Not(m_Specific(RHS)), m_Value()))) | |||
| 254 | return true; | |||
| 255 | ||||
| 256 | // X op ((X & Y) ^ Y) -- this is the canonical form of the previous pattern | |||
| 257 | // for constant Y. | |||
| 258 | Value *Y; | |||
| 259 | if (match(RHS, | |||
| 260 | m_c_Xor(m_c_And(m_Specific(LHS), m_Value(Y)), m_Deferred(Y))) || | |||
| 261 | match(LHS, m_c_Xor(m_c_And(m_Specific(RHS), m_Value(Y)), m_Deferred(Y)))) | |||
| 262 | return true; | |||
| 263 | ||||
| 264 | // Peek through extends to find a 'not' of the other side: | |||
| 265 | // (ext Y) op ext(~Y) | |||
| 266 | // (ext ~Y) op ext(Y) | |||
| 267 | if ((match(LHS, m_ZExtOrSExt(m_Value(Y))) && | |||
| 268 | match(RHS, m_ZExtOrSExt(m_Not(m_Specific(Y))))) || | |||
| 269 | (match(RHS, m_ZExtOrSExt(m_Value(Y))) && | |||
| 270 | match(LHS, m_ZExtOrSExt(m_Not(m_Specific(Y)))))) | |||
| 271 | return true; | |||
| 272 | ||||
| 273 | // Look for: (A & B) op ~(A | B) | |||
| 274 | { | |||
| 275 | Value *A, *B; | |||
| 276 | if (match(LHS, m_And(m_Value(A), m_Value(B))) && | |||
| 277 | match(RHS, m_Not(m_c_Or(m_Specific(A), m_Specific(B))))) | |||
| 278 | return true; | |||
| 279 | if (match(RHS, m_And(m_Value(A), m_Value(B))) && | |||
| 280 | match(LHS, m_Not(m_c_Or(m_Specific(A), m_Specific(B))))) | |||
| 281 | return true; | |||
| 282 | } | |||
| 283 | IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType()); | |||
| 284 | KnownBits LHSKnown(IT->getBitWidth()); | |||
| 285 | KnownBits RHSKnown(IT->getBitWidth()); | |||
| 286 | computeKnownBits(LHS, LHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo); | |||
| 287 | computeKnownBits(RHS, RHSKnown, DL, 0, AC, CxtI, DT, nullptr, UseInstrInfo); | |||
| 288 | return KnownBits::haveNoCommonBitsSet(LHSKnown, RHSKnown); | |||
| 289 | } | |||
| 290 | ||||
| 291 | bool llvm::isOnlyUsedInZeroEqualityComparison(const Instruction *I) { | |||
| 292 | return !I->user_empty() && all_of(I->users(), [](const User *U) { | |||
| 293 | ICmpInst::Predicate P; | |||
| 294 | return match(U, m_ICmp(P, m_Value(), m_Zero())) && ICmpInst::isEquality(P); | |||
| 295 | }); | |||
| 296 | } | |||
| 297 | ||||
| 298 | static bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth, | |||
| 299 | const Query &Q); | |||
| 300 | ||||
| 301 | bool llvm::isKnownToBeAPowerOfTwo(const Value *V, const DataLayout &DL, | |||
| 302 | bool OrZero, unsigned Depth, | |||
| 303 | AssumptionCache *AC, const Instruction *CxtI, | |||
| 304 | const DominatorTree *DT, bool UseInstrInfo) { | |||
| 305 | return ::isKnownToBeAPowerOfTwo( | |||
| 306 | V, OrZero, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); | |||
| 307 | } | |||
| 308 | ||||
| 309 | static bool isKnownNonZero(const Value *V, const APInt &DemandedElts, | |||
| 310 | unsigned Depth, const Query &Q); | |||
| 311 | ||||
| 312 | static bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q); | |||
| 313 | ||||
| 314 | bool llvm::isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth, | |||
| 315 | AssumptionCache *AC, const Instruction *CxtI, | |||
| 316 | const DominatorTree *DT, bool UseInstrInfo) { | |||
| 317 | return ::isKnownNonZero(V, Depth, | |||
| 318 | Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); | |||
| 319 | } | |||
| 320 | ||||
| 321 | bool llvm::isKnownNonNegative(const Value *V, const DataLayout &DL, | |||
| 322 | unsigned Depth, AssumptionCache *AC, | |||
| 323 | const Instruction *CxtI, const DominatorTree *DT, | |||
| 324 | bool UseInstrInfo) { | |||
| 325 | KnownBits Known = | |||
| 326 | computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo); | |||
| 327 | return Known.isNonNegative(); | |||
| 328 | } | |||
| 329 | ||||
| 330 | bool llvm::isKnownPositive(const Value *V, const DataLayout &DL, unsigned Depth, | |||
| 331 | AssumptionCache *AC, const Instruction *CxtI, | |||
| 332 | const DominatorTree *DT, bool UseInstrInfo) { | |||
| 333 | if (auto *CI = dyn_cast<ConstantInt>(V)) | |||
| 334 | return CI->getValue().isStrictlyPositive(); | |||
| 335 | ||||
| 336 | // TODO: We'd doing two recursive queries here. We should factor this such | |||
| 337 | // that only a single query is needed. | |||
| 338 | return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT, UseInstrInfo) && | |||
| 339 | isKnownNonZero(V, DL, Depth, AC, CxtI, DT, UseInstrInfo); | |||
| 340 | } | |||
| 341 | ||||
| 342 | bool llvm::isKnownNegative(const Value *V, const DataLayout &DL, unsigned Depth, | |||
| 343 | AssumptionCache *AC, const Instruction *CxtI, | |||
| 344 | const DominatorTree *DT, bool UseInstrInfo) { | |||
| 345 | KnownBits Known = | |||
| 346 | computeKnownBits(V, DL, Depth, AC, CxtI, DT, nullptr, UseInstrInfo); | |||
| 347 | return Known.isNegative(); | |||
| 348 | } | |||
| 349 | ||||
| 350 | static bool isKnownNonEqual(const Value *V1, const Value *V2, unsigned Depth, | |||
| 351 | const Query &Q); | |||
| 352 | ||||
| 353 | bool llvm::isKnownNonEqual(const Value *V1, const Value *V2, | |||
| 354 | const DataLayout &DL, AssumptionCache *AC, | |||
| 355 | const Instruction *CxtI, const DominatorTree *DT, | |||
| 356 | bool UseInstrInfo) { | |||
| 357 | return ::isKnownNonEqual(V1, V2, 0, | |||
| 358 | Query(DL, AC, safeCxtI(V2, V1, CxtI), DT, | |||
| 359 | UseInstrInfo, /*ORE=*/nullptr)); | |||
| 360 | } | |||
| 361 | ||||
| 362 | static bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth, | |||
| 363 | const Query &Q); | |||
| 364 | ||||
| 365 | bool llvm::MaskedValueIsZero(const Value *V, const APInt &Mask, | |||
| 366 | const DataLayout &DL, unsigned Depth, | |||
| 367 | AssumptionCache *AC, const Instruction *CxtI, | |||
| 368 | const DominatorTree *DT, bool UseInstrInfo) { | |||
| 369 | return ::MaskedValueIsZero( | |||
| 370 | V, Mask, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); | |||
| 371 | } | |||
| 372 | ||||
| 373 | static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts, | |||
| 374 | unsigned Depth, const Query &Q); | |||
| 375 | ||||
| 376 | static unsigned ComputeNumSignBits(const Value *V, unsigned Depth, | |||
| 377 | const Query &Q) { | |||
| 378 | auto *FVTy = dyn_cast<FixedVectorType>(V->getType()); | |||
| 379 | APInt DemandedElts = | |||
| 380 | FVTy ? APInt::getAllOnes(FVTy->getNumElements()) : APInt(1, 1); | |||
| 381 | return ComputeNumSignBits(V, DemandedElts, Depth, Q); | |||
| 382 | } | |||
| 383 | ||||
| 384 | unsigned llvm::ComputeNumSignBits(const Value *V, const DataLayout &DL, | |||
| 385 | unsigned Depth, AssumptionCache *AC, | |||
| 386 | const Instruction *CxtI, | |||
| 387 | const DominatorTree *DT, bool UseInstrInfo) { | |||
| 388 | return ::ComputeNumSignBits( | |||
| 389 | V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo)); | |||
| 390 | } | |||
| 391 | ||||
| 392 | unsigned llvm::ComputeMaxSignificantBits(const Value *V, const DataLayout &DL, | |||
| 393 | unsigned Depth, AssumptionCache *AC, | |||
| 394 | const Instruction *CxtI, | |||
| 395 | const DominatorTree *DT) { | |||
| 396 | unsigned SignBits = ComputeNumSignBits(V, DL, Depth, AC, CxtI, DT); | |||
| 397 | return V->getType()->getScalarSizeInBits() - SignBits + 1; | |||
| 398 | } | |||
| 399 | ||||
| 400 | static void computeKnownBitsAddSub(bool Add, const Value *Op0, const Value *Op1, | |||
| 401 | bool NSW, const APInt &DemandedElts, | |||
| 402 | KnownBits &KnownOut, KnownBits &Known2, | |||
| 403 | unsigned Depth, const Query &Q) { | |||
| 404 | computeKnownBits(Op1, DemandedElts, KnownOut, Depth + 1, Q); | |||
| 405 | ||||
| 406 | // If one operand is unknown and we have no nowrap information, | |||
| 407 | // the result will be unknown independently of the second operand. | |||
| 408 | if (KnownOut.isUnknown() && !NSW) | |||
| 409 | return; | |||
| 410 | ||||
| 411 | computeKnownBits(Op0, DemandedElts, Known2, Depth + 1, Q); | |||
| 412 | KnownOut = KnownBits::computeForAddSub(Add, NSW, Known2, KnownOut); | |||
| 413 | } | |||
| 414 | ||||
| 415 | static void computeKnownBitsMul(const Value *Op0, const Value *Op1, bool NSW, | |||
| 416 | const APInt &DemandedElts, KnownBits &Known, | |||
| 417 | KnownBits &Known2, unsigned Depth, | |||
| 418 | const Query &Q) { | |||
| 419 | computeKnownBits(Op1, DemandedElts, Known, Depth + 1, Q); | |||
| 420 | computeKnownBits(Op0, DemandedElts, Known2, Depth + 1, Q); | |||
| 421 | ||||
| 422 | bool isKnownNegative = false; | |||
| 423 | bool isKnownNonNegative = false; | |||
| 424 | // If the multiplication is known not to overflow, compute the sign bit. | |||
| 425 | if (NSW) { | |||
| 426 | if (Op0 == Op1) { | |||
| 427 | // The product of a number with itself is non-negative. | |||
| 428 | isKnownNonNegative = true; | |||
| 429 | } else { | |||
| 430 | bool isKnownNonNegativeOp1 = Known.isNonNegative(); | |||
| 431 | bool isKnownNonNegativeOp0 = Known2.isNonNegative(); | |||
| 432 | bool isKnownNegativeOp1 = Known.isNegative(); | |||
| 433 | bool isKnownNegativeOp0 = Known2.isNegative(); | |||
| 434 | // The product of two numbers with the same sign is non-negative. | |||
| 435 | isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) || | |||
| 436 | (isKnownNonNegativeOp1 && isKnownNonNegativeOp0); | |||
| 437 | // The product of a negative number and a non-negative number is either | |||
| 438 | // negative or zero. | |||
| 439 | if (!isKnownNonNegative) | |||
| 440 | isKnownNegative = | |||
| 441 | (isKnownNegativeOp1 && isKnownNonNegativeOp0 && | |||
| 442 | Known2.isNonZero()) || | |||
| 443 | (isKnownNegativeOp0 && isKnownNonNegativeOp1 && Known.isNonZero()); | |||
| 444 | } | |||
| 445 | } | |||
| 446 | ||||
| 447 | bool SelfMultiply = Op0 == Op1; | |||
| 448 | // TODO: SelfMultiply can be poison, but not undef. | |||
| 449 | if (SelfMultiply) | |||
| 450 | SelfMultiply &= | |||
| 451 | isGuaranteedNotToBeUndefOrPoison(Op0, Q.AC, Q.CxtI, Q.DT, Depth + 1); | |||
| 452 | Known = KnownBits::mul(Known, Known2, SelfMultiply); | |||
| 453 | ||||
| 454 | // Only make use of no-wrap flags if we failed to compute the sign bit | |||
| 455 | // directly. This matters if the multiplication always overflows, in | |||
| 456 | // which case we prefer to follow the result of the direct computation, | |||
| 457 | // though as the program is invoking undefined behaviour we can choose | |||
| 458 | // whatever we like here. | |||
| 459 | if (isKnownNonNegative && !Known.isNegative()) | |||
| 460 | Known.makeNonNegative(); | |||
| 461 | else if (isKnownNegative && !Known.isNonNegative()) | |||
| 462 | Known.makeNegative(); | |||
| 463 | } | |||
| 464 | ||||
| 465 | void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges, | |||
| 466 | KnownBits &Known) { | |||
| 467 | unsigned BitWidth = Known.getBitWidth(); | |||
| 468 | unsigned NumRanges = Ranges.getNumOperands() / 2; | |||
| 469 | assert(NumRanges >= 1)(static_cast <bool> (NumRanges >= 1) ? void (0) : __assert_fail ("NumRanges >= 1", "llvm/lib/Analysis/ValueTracking.cpp", 469, __extension__ __PRETTY_FUNCTION__)); | |||
| 470 | ||||
| 471 | Known.Zero.setAllBits(); | |||
| 472 | Known.One.setAllBits(); | |||
| 473 | ||||
| 474 | for (unsigned i = 0; i < NumRanges; ++i) { | |||
| 475 | ConstantInt *Lower = | |||
| 476 | mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0)); | |||
| 477 | ConstantInt *Upper = | |||
| 478 | mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1)); | |||
| 479 | ConstantRange Range(Lower->getValue(), Upper->getValue()); | |||
| 480 | ||||
| 481 | // The first CommonPrefixBits of all values in Range are equal. | |||
| 482 | unsigned CommonPrefixBits = | |||
| 483 | (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countl_zero(); | |||
| 484 | APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits); | |||
| 485 | APInt UnsignedMax = Range.getUnsignedMax().zextOrTrunc(BitWidth); | |||
| 486 | Known.One &= UnsignedMax & Mask; | |||
| 487 | Known.Zero &= ~UnsignedMax & Mask; | |||
| 488 | } | |||
| 489 | } | |||
| 490 | ||||
| 491 | static bool isEphemeralValueOf(const Instruction *I, const Value *E) { | |||
| 492 | SmallVector<const Value *, 16> WorkSet(1, I); | |||
| 493 | SmallPtrSet<const Value *, 32> Visited; | |||
| 494 | SmallPtrSet<const Value *, 16> EphValues; | |||
| 495 | ||||
| 496 | // The instruction defining an assumption's condition itself is always | |||
| 497 | // considered ephemeral to that assumption (even if it has other | |||
| 498 | // non-ephemeral users). See r246696's test case for an example. | |||
| 499 | if (is_contained(I->operands(), E)) | |||
| 500 | return true; | |||
| 501 | ||||
| 502 | while (!WorkSet.empty()) { | |||
| 503 | const Value *V = WorkSet.pop_back_val(); | |||
| 504 | if (!Visited.insert(V).second) | |||
| 505 | continue; | |||
| 506 | ||||
| 507 | // If all uses of this value are ephemeral, then so is this value. | |||
| 508 | if (llvm::all_of(V->users(), [&](const User *U) { | |||
| 509 | return EphValues.count(U); | |||
| 510 | })) { | |||
| 511 | if (V == E) | |||
| 512 | return true; | |||
| 513 | ||||
| 514 | if (V == I || (isa<Instruction>(V) && | |||
| 515 | !cast<Instruction>(V)->mayHaveSideEffects() && | |||
| 516 | !cast<Instruction>(V)->isTerminator())) { | |||
| 517 | EphValues.insert(V); | |||
| 518 | if (const User *U = dyn_cast<User>(V)) | |||
| 519 | append_range(WorkSet, U->operands()); | |||
| 520 | } | |||
| 521 | } | |||
| 522 | } | |||
| 523 | ||||
| 524 | return false; | |||
| 525 | } | |||
| 526 | ||||
| 527 | // Is this an intrinsic that cannot be speculated but also cannot trap? | |||
| 528 | bool llvm::isAssumeLikeIntrinsic(const Instruction *I) { | |||
| 529 | if (const IntrinsicInst *CI = dyn_cast<IntrinsicInst>(I)) | |||
| 530 | return CI->isAssumeLikeIntrinsic(); | |||
| 531 | ||||
| 532 | return false; | |||
| 533 | } | |||
| 534 | ||||
| 535 | bool llvm::isValidAssumeForContext(const Instruction *Inv, | |||
| 536 | const Instruction *CxtI, | |||
| 537 | const DominatorTree *DT) { | |||
| 538 | // There are two restrictions on the use of an assume: | |||
| 539 | // 1. The assume must dominate the context (or the control flow must | |||
| 540 | // reach the assume whenever it reaches the context). | |||
| 541 | // 2. The context must not be in the assume's set of ephemeral values | |||
| 542 | // (otherwise we will use the assume to prove that the condition | |||
| 543 | // feeding the assume is trivially true, thus causing the removal of | |||
| 544 | // the assume). | |||
| 545 | ||||
| 546 | if (Inv->getParent() == CxtI->getParent()) { | |||
| 547 | // If Inv and CtxI are in the same block, check if the assume (Inv) is first | |||
| 548 | // in the BB. | |||
| 549 | if (Inv->comesBefore(CxtI)) | |||
| 550 | return true; | |||
| 551 | ||||
| 552 | // Don't let an assume affect itself - this would cause the problems | |||
| 553 | // `isEphemeralValueOf` is trying to prevent, and it would also make | |||
| 554 | // the loop below go out of bounds. | |||
| 555 | if (Inv == CxtI) | |||
| 556 | return false; | |||
| 557 | ||||
| 558 | // The context comes first, but they're both in the same block. | |||
| 559 | // Make sure there is nothing in between that might interrupt | |||
| 560 | // the control flow, not even CxtI itself. | |||
| 561 | // We limit the scan distance between the assume and its context instruction | |||
| 562 | // to avoid a compile-time explosion. This limit is chosen arbitrarily, so | |||
| 563 | // it can be adjusted if needed (could be turned into a cl::opt). | |||
| 564 | auto Range = make_range(CxtI->getIterator(), Inv->getIterator()); | |||
| 565 | if (!isGuaranteedToTransferExecutionToSuccessor(Range, 15)) | |||
| 566 | return false; | |||
| 567 | ||||
| 568 | return !isEphemeralValueOf(Inv, CxtI); | |||
| 569 | } | |||
| 570 | ||||
| 571 | // Inv and CxtI are in different blocks. | |||
| 572 | if (DT) { | |||
| 573 | if (DT->dominates(Inv, CxtI)) | |||
| 574 | return true; | |||
| 575 | } else if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) { | |||
| 576 | // We don't have a DT, but this trivially dominates. | |||
| 577 | return true; | |||
| 578 | } | |||
| 579 | ||||
| 580 | return false; | |||
| 581 | } | |||
| 582 | ||||
| 583 | static bool cmpExcludesZero(CmpInst::Predicate Pred, const Value *RHS) { | |||
| 584 | // v u> y implies v != 0. | |||
| 585 | if (Pred == ICmpInst::ICMP_UGT) | |||
| 586 | return true; | |||
| 587 | ||||
| 588 | // Special-case v != 0 to also handle v != null. | |||
| 589 | if (Pred == ICmpInst::ICMP_NE) | |||
| 590 | return match(RHS, m_Zero()); | |||
| 591 | ||||
| 592 | // All other predicates - rely on generic ConstantRange handling. | |||
| 593 | const APInt *C; | |||
| 594 | if (!match(RHS, m_APInt(C))) | |||
| 595 | return false; | |||
| 596 | ||||
| 597 | ConstantRange TrueValues = ConstantRange::makeExactICmpRegion(Pred, *C); | |||
| 598 | return !TrueValues.contains(APInt::getZero(C->getBitWidth())); | |||
| 599 | } | |||
| 600 | ||||
| 601 | static bool isKnownNonZeroFromAssume(const Value *V, const Query &Q) { | |||
| 602 | // Use of assumptions is context-sensitive. If we don't have a context, we | |||
| 603 | // cannot use them! | |||
| 604 | if (!Q.AC || !Q.CxtI) | |||
| 605 | return false; | |||
| 606 | ||||
| 607 | if (Q.CxtI && V->getType()->isPointerTy()) { | |||
| 608 | SmallVector<Attribute::AttrKind, 2> AttrKinds{Attribute::NonNull}; | |||
| 609 | if (!NullPointerIsDefined(Q.CxtI->getFunction(), | |||
| 610 | V->getType()->getPointerAddressSpace())) | |||
| 611 | AttrKinds.push_back(Attribute::Dereferenceable); | |||
| 612 | ||||
| 613 | if (getKnowledgeValidInContext(V, AttrKinds, Q.CxtI, Q.DT, Q.AC)) | |||
| 614 | return true; | |||
| 615 | } | |||
| 616 | ||||
| 617 | for (auto &AssumeVH : Q.AC->assumptionsFor(V)) { | |||
| 618 | if (!AssumeVH) | |||
| 619 | continue; | |||
| 620 | CallInst *I = cast<CallInst>(AssumeVH); | |||
| 621 | assert(I->getFunction() == Q.CxtI->getFunction() &&(static_cast <bool> (I->getFunction() == Q.CxtI-> getFunction() && "Got assumption for the wrong function!" ) ? void (0) : __assert_fail ("I->getFunction() == Q.CxtI->getFunction() && \"Got assumption for the wrong function!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 622, __extension__ __PRETTY_FUNCTION__ )) | |||
| 622 | "Got assumption for the wrong function!")(static_cast <bool> (I->getFunction() == Q.CxtI-> getFunction() && "Got assumption for the wrong function!" ) ? void (0) : __assert_fail ("I->getFunction() == Q.CxtI->getFunction() && \"Got assumption for the wrong function!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 622, __extension__ __PRETTY_FUNCTION__ )); | |||
| 623 | ||||
| 624 | // Warning: This loop can end up being somewhat performance sensitive. | |||
| 625 | // We're running this loop for once for each value queried resulting in a | |||
| 626 | // runtime of ~O(#assumes * #values). | |||
| 627 | ||||
| 628 | assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&(static_cast <bool> (I->getCalledFunction()->getIntrinsicID () == Intrinsic::assume && "must be an assume intrinsic" ) ? void (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\"" , "llvm/lib/Analysis/ValueTracking.cpp", 629, __extension__ __PRETTY_FUNCTION__ )) | |||
| 629 | "must be an assume intrinsic")(static_cast <bool> (I->getCalledFunction()->getIntrinsicID () == Intrinsic::assume && "must be an assume intrinsic" ) ? void (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\"" , "llvm/lib/Analysis/ValueTracking.cpp", 629, __extension__ __PRETTY_FUNCTION__ )); | |||
| 630 | ||||
| 631 | Value *RHS; | |||
| 632 | CmpInst::Predicate Pred; | |||
| 633 | auto m_V = m_CombineOr(m_Specific(V), m_PtrToInt(m_Specific(V))); | |||
| 634 | if (!match(I->getArgOperand(0), m_c_ICmp(Pred, m_V, m_Value(RHS)))) | |||
| 635 | return false; | |||
| 636 | ||||
| 637 | if (cmpExcludesZero(Pred, RHS) && isValidAssumeForContext(I, Q.CxtI, Q.DT)) | |||
| 638 | return true; | |||
| 639 | } | |||
| 640 | ||||
| 641 | return false; | |||
| 642 | } | |||
| 643 | ||||
| 644 | static void computeKnownBitsFromCmp(const Value *V, const ICmpInst *Cmp, | |||
| 645 | KnownBits &Known, unsigned Depth, | |||
| 646 | const Query &Q) { | |||
| 647 | unsigned BitWidth = Known.getBitWidth(); | |||
| 648 | // We are attempting to compute known bits for the operands of an assume. | |||
| 649 | // Do not try to use other assumptions for those recursive calls because | |||
| 650 | // that can lead to mutual recursion and a compile-time explosion. | |||
| 651 | // An example of the mutual recursion: computeKnownBits can call | |||
| 652 | // isKnownNonZero which calls computeKnownBitsFromAssume (this function) | |||
| 653 | // and so on. | |||
| 654 | Query QueryNoAC = Q; | |||
| 655 | QueryNoAC.AC = nullptr; | |||
| 656 | ||||
| 657 | // Note that ptrtoint may change the bitwidth. | |||
| 658 | Value *A, *B; | |||
| 659 | auto m_V = m_CombineOr(m_Specific(V), m_PtrToInt(m_Specific(V))); | |||
| 660 | ||||
| 661 | CmpInst::Predicate Pred; | |||
| 662 | uint64_t C; | |||
| 663 | switch (Cmp->getPredicate()) { | |||
| 664 | default: | |||
| 665 | break; | |||
| 666 | case ICmpInst::ICMP_EQ: | |||
| 667 | // assume(v = a) | |||
| 668 | if (match(Cmp, m_c_ICmp(Pred, m_V, m_Value(A)))) { | |||
| 669 | KnownBits RHSKnown = | |||
| 670 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 671 | Known.Zero |= RHSKnown.Zero; | |||
| 672 | Known.One |= RHSKnown.One; | |||
| 673 | // assume(v & b = a) | |||
| 674 | } else if (match(Cmp, | |||
| 675 | m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A)))) { | |||
| 676 | KnownBits RHSKnown = | |||
| 677 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 678 | KnownBits MaskKnown = | |||
| 679 | computeKnownBits(B, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 680 | ||||
| 681 | // For those bits in the mask that are known to be one, we can propagate | |||
| 682 | // known bits from the RHS to V. | |||
| 683 | Known.Zero |= RHSKnown.Zero & MaskKnown.One; | |||
| 684 | Known.One |= RHSKnown.One & MaskKnown.One; | |||
| 685 | // assume(~(v & b) = a) | |||
| 686 | } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))), | |||
| 687 | m_Value(A)))) { | |||
| 688 | KnownBits RHSKnown = | |||
| 689 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 690 | KnownBits MaskKnown = | |||
| 691 | computeKnownBits(B, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 692 | ||||
| 693 | // For those bits in the mask that are known to be one, we can propagate | |||
| 694 | // inverted known bits from the RHS to V. | |||
| 695 | Known.Zero |= RHSKnown.One & MaskKnown.One; | |||
| 696 | Known.One |= RHSKnown.Zero & MaskKnown.One; | |||
| 697 | // assume(v | b = a) | |||
| 698 | } else if (match(Cmp, | |||
| 699 | m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A)))) { | |||
| 700 | KnownBits RHSKnown = | |||
| 701 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 702 | KnownBits BKnown = | |||
| 703 | computeKnownBits(B, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 704 | ||||
| 705 | // For those bits in B that are known to be zero, we can propagate known | |||
| 706 | // bits from the RHS to V. | |||
| 707 | Known.Zero |= RHSKnown.Zero & BKnown.Zero; | |||
| 708 | Known.One |= RHSKnown.One & BKnown.Zero; | |||
| 709 | // assume(~(v | b) = a) | |||
| 710 | } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))), | |||
| 711 | m_Value(A)))) { | |||
| 712 | KnownBits RHSKnown = | |||
| 713 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 714 | KnownBits BKnown = | |||
| 715 | computeKnownBits(B, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 716 | ||||
| 717 | // For those bits in B that are known to be zero, we can propagate | |||
| 718 | // inverted known bits from the RHS to V. | |||
| 719 | Known.Zero |= RHSKnown.One & BKnown.Zero; | |||
| 720 | Known.One |= RHSKnown.Zero & BKnown.Zero; | |||
| 721 | // assume(v ^ b = a) | |||
| 722 | } else if (match(Cmp, | |||
| 723 | m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A)))) { | |||
| 724 | KnownBits RHSKnown = | |||
| 725 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 726 | KnownBits BKnown = | |||
| 727 | computeKnownBits(B, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 728 | ||||
| 729 | // For those bits in B that are known to be zero, we can propagate known | |||
| 730 | // bits from the RHS to V. For those bits in B that are known to be one, | |||
| 731 | // we can propagate inverted known bits from the RHS to V. | |||
| 732 | Known.Zero |= RHSKnown.Zero & BKnown.Zero; | |||
| 733 | Known.One |= RHSKnown.One & BKnown.Zero; | |||
| 734 | Known.Zero |= RHSKnown.One & BKnown.One; | |||
| 735 | Known.One |= RHSKnown.Zero & BKnown.One; | |||
| 736 | // assume(~(v ^ b) = a) | |||
| 737 | } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))), | |||
| 738 | m_Value(A)))) { | |||
| 739 | KnownBits RHSKnown = | |||
| 740 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 741 | KnownBits BKnown = | |||
| 742 | computeKnownBits(B, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 743 | ||||
| 744 | // For those bits in B that are known to be zero, we can propagate | |||
| 745 | // inverted known bits from the RHS to V. For those bits in B that are | |||
| 746 | // known to be one, we can propagate known bits from the RHS to V. | |||
| 747 | Known.Zero |= RHSKnown.One & BKnown.Zero; | |||
| 748 | Known.One |= RHSKnown.Zero & BKnown.Zero; | |||
| 749 | Known.Zero |= RHSKnown.Zero & BKnown.One; | |||
| 750 | Known.One |= RHSKnown.One & BKnown.One; | |||
| 751 | // assume(v << c = a) | |||
| 752 | } else if (match(Cmp, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)), | |||
| 753 | m_Value(A))) && | |||
| 754 | C < BitWidth) { | |||
| 755 | KnownBits RHSKnown = | |||
| 756 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 757 | ||||
| 758 | // For those bits in RHS that are known, we can propagate them to known | |||
| 759 | // bits in V shifted to the right by C. | |||
| 760 | RHSKnown.Zero.lshrInPlace(C); | |||
| 761 | Known.Zero |= RHSKnown.Zero; | |||
| 762 | RHSKnown.One.lshrInPlace(C); | |||
| 763 | Known.One |= RHSKnown.One; | |||
| 764 | // assume(~(v << c) = a) | |||
| 765 | } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))), | |||
| 766 | m_Value(A))) && | |||
| 767 | C < BitWidth) { | |||
| 768 | KnownBits RHSKnown = | |||
| 769 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 770 | // For those bits in RHS that are known, we can propagate them inverted | |||
| 771 | // to known bits in V shifted to the right by C. | |||
| 772 | RHSKnown.One.lshrInPlace(C); | |||
| 773 | Known.Zero |= RHSKnown.One; | |||
| 774 | RHSKnown.Zero.lshrInPlace(C); | |||
| 775 | Known.One |= RHSKnown.Zero; | |||
| 776 | // assume(v >> c = a) | |||
| 777 | } else if (match(Cmp, m_c_ICmp(Pred, m_Shr(m_V, m_ConstantInt(C)), | |||
| 778 | m_Value(A))) && | |||
| 779 | C < BitWidth) { | |||
| 780 | KnownBits RHSKnown = | |||
| 781 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 782 | // For those bits in RHS that are known, we can propagate them to known | |||
| 783 | // bits in V shifted to the right by C. | |||
| 784 | Known.Zero |= RHSKnown.Zero << C; | |||
| 785 | Known.One |= RHSKnown.One << C; | |||
| 786 | // assume(~(v >> c) = a) | |||
| 787 | } else if (match(Cmp, m_c_ICmp(Pred, m_Not(m_Shr(m_V, m_ConstantInt(C))), | |||
| 788 | m_Value(A))) && | |||
| 789 | C < BitWidth) { | |||
| 790 | KnownBits RHSKnown = | |||
| 791 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 792 | // For those bits in RHS that are known, we can propagate them inverted | |||
| 793 | // to known bits in V shifted to the right by C. | |||
| 794 | Known.Zero |= RHSKnown.One << C; | |||
| 795 | Known.One |= RHSKnown.Zero << C; | |||
| 796 | } | |||
| 797 | break; | |||
| 798 | case ICmpInst::ICMP_SGE: | |||
| 799 | // assume(v >=_s c) where c is non-negative | |||
| 800 | if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A)))) { | |||
| 801 | KnownBits RHSKnown = | |||
| 802 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 803 | ||||
| 804 | if (RHSKnown.isNonNegative()) { | |||
| 805 | // We know that the sign bit is zero. | |||
| 806 | Known.makeNonNegative(); | |||
| 807 | } | |||
| 808 | } | |||
| 809 | break; | |||
| 810 | case ICmpInst::ICMP_SGT: | |||
| 811 | // assume(v >_s c) where c is at least -1. | |||
| 812 | if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A)))) { | |||
| 813 | KnownBits RHSKnown = | |||
| 814 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 815 | ||||
| 816 | if (RHSKnown.isAllOnes() || RHSKnown.isNonNegative()) { | |||
| 817 | // We know that the sign bit is zero. | |||
| 818 | Known.makeNonNegative(); | |||
| 819 | } | |||
| 820 | } | |||
| 821 | break; | |||
| 822 | case ICmpInst::ICMP_SLE: | |||
| 823 | // assume(v <=_s c) where c is negative | |||
| 824 | if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A)))) { | |||
| 825 | KnownBits RHSKnown = | |||
| 826 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 827 | ||||
| 828 | if (RHSKnown.isNegative()) { | |||
| 829 | // We know that the sign bit is one. | |||
| 830 | Known.makeNegative(); | |||
| 831 | } | |||
| 832 | } | |||
| 833 | break; | |||
| 834 | case ICmpInst::ICMP_SLT: | |||
| 835 | // assume(v <_s c) where c is non-positive | |||
| 836 | if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A)))) { | |||
| 837 | KnownBits RHSKnown = | |||
| 838 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 839 | ||||
| 840 | if (RHSKnown.isZero() || RHSKnown.isNegative()) { | |||
| 841 | // We know that the sign bit is one. | |||
| 842 | Known.makeNegative(); | |||
| 843 | } | |||
| 844 | } | |||
| 845 | break; | |||
| 846 | case ICmpInst::ICMP_ULE: | |||
| 847 | // assume(v <=_u c) | |||
| 848 | if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A)))) { | |||
| 849 | KnownBits RHSKnown = | |||
| 850 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 851 | ||||
| 852 | // Whatever high bits in c are zero are known to be zero. | |||
| 853 | Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros()); | |||
| 854 | } | |||
| 855 | break; | |||
| 856 | case ICmpInst::ICMP_ULT: | |||
| 857 | // assume(v <_u c) | |||
| 858 | if (match(Cmp, m_ICmp(Pred, m_V, m_Value(A)))) { | |||
| 859 | KnownBits RHSKnown = | |||
| 860 | computeKnownBits(A, Depth + 1, QueryNoAC).anyextOrTrunc(BitWidth); | |||
| 861 | ||||
| 862 | // If the RHS is known zero, then this assumption must be wrong (nothing | |||
| 863 | // is unsigned less than zero). Signal a conflict and get out of here. | |||
| 864 | if (RHSKnown.isZero()) { | |||
| 865 | Known.Zero.setAllBits(); | |||
| 866 | Known.One.setAllBits(); | |||
| 867 | break; | |||
| 868 | } | |||
| 869 | ||||
| 870 | // Whatever high bits in c are zero are known to be zero (if c is a power | |||
| 871 | // of 2, then one more). | |||
| 872 | if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, QueryNoAC)) | |||
| 873 | Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros() + 1); | |||
| 874 | else | |||
| 875 | Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros()); | |||
| 876 | } | |||
| 877 | break; | |||
| 878 | case ICmpInst::ICMP_NE: { | |||
| 879 | // assume (v & b != 0) where b is a power of 2 | |||
| 880 | const APInt *BPow2; | |||
| 881 | if (match(Cmp, m_ICmp(Pred, m_c_And(m_V, m_Power2(BPow2)), m_Zero()))) { | |||
| 882 | Known.One |= BPow2->zextOrTrunc(BitWidth); | |||
| 883 | } | |||
| 884 | } break; | |||
| 885 | } | |||
| 886 | } | |||
| 887 | ||||
| 888 | static void computeKnownBitsFromAssume(const Value *V, KnownBits &Known, | |||
| 889 | unsigned Depth, const Query &Q) { | |||
| 890 | // Use of assumptions is context-sensitive. If we don't have a context, we | |||
| 891 | // cannot use them! | |||
| 892 | if (!Q.AC || !Q.CxtI) | |||
| 893 | return; | |||
| 894 | ||||
| 895 | unsigned BitWidth = Known.getBitWidth(); | |||
| 896 | ||||
| 897 | // Refine Known set if the pointer alignment is set by assume bundles. | |||
| 898 | if (V->getType()->isPointerTy()) { | |||
| 899 | if (RetainedKnowledge RK = getKnowledgeValidInContext( | |||
| 900 | V, { Attribute::Alignment }, Q.CxtI, Q.DT, Q.AC)) { | |||
| 901 | if (isPowerOf2_64(RK.ArgValue)) | |||
| 902 | Known.Zero.setLowBits(Log2_64(RK.ArgValue)); | |||
| 903 | } | |||
| 904 | } | |||
| 905 | ||||
| 906 | // Note that the patterns below need to be kept in sync with the code | |||
| 907 | // in AssumptionCache::updateAffectedValues. | |||
| 908 | ||||
| 909 | for (auto &AssumeVH : Q.AC->assumptionsFor(V)) { | |||
| 910 | if (!AssumeVH) | |||
| 911 | continue; | |||
| 912 | CallInst *I = cast<CallInst>(AssumeVH); | |||
| 913 | assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&(static_cast <bool> (I->getParent()->getParent() == Q.CxtI->getParent()->getParent() && "Got assumption for the wrong function!" ) ? void (0) : __assert_fail ("I->getParent()->getParent() == Q.CxtI->getParent()->getParent() && \"Got assumption for the wrong function!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 914, __extension__ __PRETTY_FUNCTION__ )) | |||
| 914 | "Got assumption for the wrong function!")(static_cast <bool> (I->getParent()->getParent() == Q.CxtI->getParent()->getParent() && "Got assumption for the wrong function!" ) ? void (0) : __assert_fail ("I->getParent()->getParent() == Q.CxtI->getParent()->getParent() && \"Got assumption for the wrong function!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 914, __extension__ __PRETTY_FUNCTION__ )); | |||
| 915 | ||||
| 916 | // Warning: This loop can end up being somewhat performance sensitive. | |||
| 917 | // We're running this loop for once for each value queried resulting in a | |||
| 918 | // runtime of ~O(#assumes * #values). | |||
| 919 | ||||
| 920 | assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&(static_cast <bool> (I->getCalledFunction()->getIntrinsicID () == Intrinsic::assume && "must be an assume intrinsic" ) ? void (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\"" , "llvm/lib/Analysis/ValueTracking.cpp", 921, __extension__ __PRETTY_FUNCTION__ )) | |||
| 921 | "must be an assume intrinsic")(static_cast <bool> (I->getCalledFunction()->getIntrinsicID () == Intrinsic::assume && "must be an assume intrinsic" ) ? void (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\"" , "llvm/lib/Analysis/ValueTracking.cpp", 921, __extension__ __PRETTY_FUNCTION__ )); | |||
| 922 | ||||
| 923 | Value *Arg = I->getArgOperand(0); | |||
| 924 | ||||
| 925 | if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | |||
| 926 | assert(BitWidth == 1 && "assume operand is not i1?")(static_cast <bool> (BitWidth == 1 && "assume operand is not i1?" ) ? void (0) : __assert_fail ("BitWidth == 1 && \"assume operand is not i1?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 926, __extension__ __PRETTY_FUNCTION__ )); | |||
| 927 | (void)BitWidth; | |||
| 928 | Known.setAllOnes(); | |||
| 929 | return; | |||
| 930 | } | |||
| 931 | if (match(Arg, m_Not(m_Specific(V))) && | |||
| 932 | isValidAssumeForContext(I, Q.CxtI, Q.DT)) { | |||
| 933 | assert(BitWidth == 1 && "assume operand is not i1?")(static_cast <bool> (BitWidth == 1 && "assume operand is not i1?" ) ? void (0) : __assert_fail ("BitWidth == 1 && \"assume operand is not i1?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 933, __extension__ __PRETTY_FUNCTION__ )); | |||
| 934 | (void)BitWidth; | |||
| 935 | Known.setAllZero(); | |||
| 936 | return; | |||
| 937 | } | |||
| 938 | ||||
| 939 | // The remaining tests are all recursive, so bail out if we hit the limit. | |||
| 940 | if (Depth == MaxAnalysisRecursionDepth) | |||
| 941 | continue; | |||
| 942 | ||||
| 943 | ICmpInst *Cmp = dyn_cast<ICmpInst>(Arg); | |||
| 944 | if (!Cmp) | |||
| 945 | continue; | |||
| 946 | ||||
| 947 | if (!isValidAssumeForContext(I, Q.CxtI, Q.DT)) | |||
| 948 | continue; | |||
| 949 | ||||
| 950 | computeKnownBitsFromCmp(V, Cmp, Known, Depth, Q); | |||
| 951 | } | |||
| 952 | ||||
| 953 | // If assumptions conflict with each other or previous known bits, then we | |||
| 954 | // have a logical fallacy. It's possible that the assumption is not reachable, | |||
| 955 | // so this isn't a real bug. On the other hand, the program may have undefined | |||
| 956 | // behavior, or we might have a bug in the compiler. We can't assert/crash, so | |||
| 957 | // clear out the known bits, try to warn the user, and hope for the best. | |||
| 958 | if (Known.Zero.intersects(Known.One)) { | |||
| 959 | Known.resetAll(); | |||
| 960 | ||||
| 961 | if (Q.ORE) | |||
| 962 | Q.ORE->emit([&]() { | |||
| 963 | auto *CxtI = const_cast<Instruction *>(Q.CxtI); | |||
| 964 | return OptimizationRemarkAnalysis("value-tracking", "BadAssumption", | |||
| 965 | CxtI) | |||
| 966 | << "Detected conflicting code assumptions. Program may " | |||
| 967 | "have undefined behavior, or compiler may have " | |||
| 968 | "internal error."; | |||
| 969 | }); | |||
| 970 | } | |||
| 971 | } | |||
| 972 | ||||
| 973 | /// Compute known bits from a shift operator, including those with a | |||
| 974 | /// non-constant shift amount. Known is the output of this function. Known2 is a | |||
| 975 | /// pre-allocated temporary with the same bit width as Known and on return | |||
| 976 | /// contains the known bit of the shift value source. KF is an | |||
| 977 | /// operator-specific function that, given the known-bits and a shift amount, | |||
| 978 | /// compute the implied known-bits of the shift operator's result respectively | |||
| 979 | /// for that shift amount. The results from calling KF are conservatively | |||
| 980 | /// combined for all permitted shift amounts. | |||
| 981 | static void computeKnownBitsFromShiftOperator( | |||
| 982 | const Operator *I, const APInt &DemandedElts, KnownBits &Known, | |||
| 983 | KnownBits &Known2, unsigned Depth, const Query &Q, | |||
| 984 | function_ref<KnownBits(const KnownBits &, const KnownBits &)> KF) { | |||
| 985 | unsigned BitWidth = Known.getBitWidth(); | |||
| 986 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | |||
| 987 | computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); | |||
| 988 | ||||
| 989 | // Note: We cannot use Known.Zero.getLimitedValue() here, because if | |||
| 990 | // BitWidth > 64 and any upper bits are known, we'll end up returning the | |||
| 991 | // limit value (which implies all bits are known). | |||
| 992 | uint64_t ShiftAmtKZ = Known.Zero.zextOrTrunc(64).getZExtValue(); | |||
| 993 | uint64_t ShiftAmtKO = Known.One.zextOrTrunc(64).getZExtValue(); | |||
| 994 | bool ShiftAmtIsConstant = Known.isConstant(); | |||
| 995 | bool MaxShiftAmtIsOutOfRange = Known.getMaxValue().uge(BitWidth); | |||
| 996 | ||||
| 997 | if (ShiftAmtIsConstant) { | |||
| 998 | Known = KF(Known2, Known); | |||
| 999 | ||||
| 1000 | // If the known bits conflict, this must be an overflowing left shift, so | |||
| 1001 | // the shift result is poison. We can return anything we want. Choose 0 for | |||
| 1002 | // the best folding opportunity. | |||
| 1003 | if (Known.hasConflict()) | |||
| 1004 | Known.setAllZero(); | |||
| 1005 | ||||
| 1006 | return; | |||
| 1007 | } | |||
| 1008 | ||||
| 1009 | // If the shift amount could be greater than or equal to the bit-width of the | |||
| 1010 | // LHS, the value could be poison, but bail out because the check below is | |||
| 1011 | // expensive. | |||
| 1012 | // TODO: Should we just carry on? | |||
| 1013 | if (MaxShiftAmtIsOutOfRange) { | |||
| 1014 | Known.resetAll(); | |||
| 1015 | return; | |||
| 1016 | } | |||
| 1017 | ||||
| 1018 | // It would be more-clearly correct to use the two temporaries for this | |||
| 1019 | // calculation. Reusing the APInts here to prevent unnecessary allocations. | |||
| 1020 | Known.resetAll(); | |||
| 1021 | ||||
| 1022 | // If we know the shifter operand is nonzero, we can sometimes infer more | |||
| 1023 | // known bits. However this is expensive to compute, so be lazy about it and | |||
| 1024 | // only compute it when absolutely necessary. | |||
| 1025 | std::optional<bool> ShifterOperandIsNonZero; | |||
| 1026 | ||||
| 1027 | // Early exit if we can't constrain any well-defined shift amount. | |||
| 1028 | if (!(ShiftAmtKZ & (PowerOf2Ceil(BitWidth) - 1)) && | |||
| 1029 | !(ShiftAmtKO & (PowerOf2Ceil(BitWidth) - 1))) { | |||
| 1030 | ShifterOperandIsNonZero = | |||
| 1031 | isKnownNonZero(I->getOperand(1), DemandedElts, Depth + 1, Q); | |||
| 1032 | if (!*ShifterOperandIsNonZero) | |||
| 1033 | return; | |||
| 1034 | } | |||
| 1035 | ||||
| 1036 | Known.Zero.setAllBits(); | |||
| 1037 | Known.One.setAllBits(); | |||
| 1038 | for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) { | |||
| 1039 | // Combine the shifted known input bits only for those shift amounts | |||
| 1040 | // compatible with its known constraints. | |||
| 1041 | if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt) | |||
| 1042 | continue; | |||
| 1043 | if ((ShiftAmt | ShiftAmtKO) != ShiftAmt) | |||
| 1044 | continue; | |||
| 1045 | // If we know the shifter is nonzero, we may be able to infer more known | |||
| 1046 | // bits. This check is sunk down as far as possible to avoid the expensive | |||
| 1047 | // call to isKnownNonZero if the cheaper checks above fail. | |||
| 1048 | if (ShiftAmt == 0) { | |||
| 1049 | if (!ShifterOperandIsNonZero) | |||
| 1050 | ShifterOperandIsNonZero = | |||
| 1051 | isKnownNonZero(I->getOperand(1), DemandedElts, Depth + 1, Q); | |||
| 1052 | if (*ShifterOperandIsNonZero) | |||
| 1053 | continue; | |||
| 1054 | } | |||
| 1055 | ||||
| 1056 | Known = KnownBits::commonBits( | |||
| 1057 | Known, KF(Known2, KnownBits::makeConstant(APInt(32, ShiftAmt)))); | |||
| 1058 | } | |||
| 1059 | ||||
| 1060 | // If the known bits conflict, the result is poison. Return a 0 and hope the | |||
| 1061 | // caller can further optimize that. | |||
| 1062 | if (Known.hasConflict()) | |||
| 1063 | Known.setAllZero(); | |||
| 1064 | } | |||
| 1065 | ||||
| 1066 | static KnownBits getKnownBitsFromAndXorOr(const Operator *I, | |||
| 1067 | const APInt &DemandedElts, | |||
| 1068 | const KnownBits &KnownLHS, | |||
| 1069 | const KnownBits &KnownRHS, | |||
| 1070 | unsigned Depth, const Query &Q) { | |||
| 1071 | unsigned BitWidth = KnownLHS.getBitWidth(); | |||
| 1072 | KnownBits KnownOut(BitWidth); | |||
| 1073 | bool IsAnd = false; | |||
| 1074 | bool HasKnownOne = !KnownLHS.One.isZero() || !KnownRHS.One.isZero(); | |||
| 1075 | Value *X = nullptr, *Y = nullptr; | |||
| 1076 | ||||
| 1077 | switch (I->getOpcode()) { | |||
| 1078 | case Instruction::And: | |||
| 1079 | KnownOut = KnownLHS & KnownRHS; | |||
| 1080 | IsAnd = true; | |||
| 1081 | // and(x, -x) is common idioms that will clear all but lowest set | |||
| 1082 | // bit. If we have a single known bit in x, we can clear all bits | |||
| 1083 | // above it. | |||
| 1084 | // TODO: instcombine often reassociates independent `and` which can hide | |||
| 1085 | // this pattern. Try to match and(x, and(-x, y)) / and(and(x, y), -x). | |||
| 1086 | if (HasKnownOne && match(I, m_c_And(m_Value(X), m_Neg(m_Deferred(X))))) { | |||
| 1087 | // -(-x) == x so using whichever (LHS/RHS) gets us a better result. | |||
| 1088 | if (KnownLHS.countMaxTrailingZeros() <= KnownRHS.countMaxTrailingZeros()) | |||
| 1089 | KnownOut = KnownLHS.blsi(); | |||
| 1090 | else | |||
| 1091 | KnownOut = KnownRHS.blsi(); | |||
| 1092 | } | |||
| 1093 | break; | |||
| 1094 | case Instruction::Or: | |||
| 1095 | KnownOut = KnownLHS | KnownRHS; | |||
| 1096 | break; | |||
| 1097 | case Instruction::Xor: | |||
| 1098 | KnownOut = KnownLHS ^ KnownRHS; | |||
| 1099 | // xor(x, x-1) is common idioms that will clear all but lowest set | |||
| 1100 | // bit. If we have a single known bit in x, we can clear all bits | |||
| 1101 | // above it. | |||
| 1102 | // TODO: xor(x, x-1) is often rewritting as xor(x, x-C) where C != | |||
| 1103 | // -1 but for the purpose of demanded bits (xor(x, x-C) & | |||
| 1104 | // Demanded) == (xor(x, x-1) & Demanded). Extend the xor pattern | |||
| 1105 | // to use arbitrary C if xor(x, x-C) as the same as xor(x, x-1). | |||
| 1106 | if (HasKnownOne && | |||
| 1107 | match(I, m_c_Xor(m_Value(X), m_c_Add(m_Deferred(X), m_AllOnes())))) { | |||
| 1108 | const KnownBits &XBits = I->getOperand(0) == X ? KnownLHS : KnownRHS; | |||
| 1109 | KnownOut = XBits.blsmsk(); | |||
| 1110 | } | |||
| 1111 | break; | |||
| 1112 | default: | |||
| 1113 | llvm_unreachable("Invalid Op used in 'analyzeKnownBitsFromAndXorOr'")::llvm::llvm_unreachable_internal("Invalid Op used in 'analyzeKnownBitsFromAndXorOr'" , "llvm/lib/Analysis/ValueTracking.cpp", 1113); | |||
| 1114 | } | |||
| 1115 | ||||
| 1116 | // and(x, add (x, -1)) is a common idiom that always clears the low bit; | |||
| 1117 | // xor/or(x, add (x, -1)) is an idiom that will always set the low bit. | |||
| 1118 | // here we handle the more general case of adding any odd number by | |||
| 1119 | // matching the form and/xor/or(x, add(x, y)) where y is odd. | |||
| 1120 | // TODO: This could be generalized to clearing any bit set in y where the | |||
| 1121 | // following bit is known to be unset in y. | |||
| 1122 | if (!KnownOut.Zero[0] && !KnownOut.One[0] && | |||
| 1123 | (match(I, m_c_BinOp(m_Value(X), m_c_Add(m_Deferred(X), m_Value(Y)))) || | |||
| 1124 | match(I, m_c_BinOp(m_Value(X), m_Sub(m_Deferred(X), m_Value(Y)))) || | |||
| 1125 | match(I, m_c_BinOp(m_Value(X), m_Sub(m_Value(Y), m_Deferred(X)))))) { | |||
| 1126 | KnownBits KnownY(BitWidth); | |||
| 1127 | computeKnownBits(Y, DemandedElts, KnownY, Depth + 1, Q); | |||
| 1128 | if (KnownY.countMinTrailingOnes() > 0) { | |||
| 1129 | if (IsAnd) | |||
| 1130 | KnownOut.Zero.setBit(0); | |||
| 1131 | else | |||
| 1132 | KnownOut.One.setBit(0); | |||
| 1133 | } | |||
| 1134 | } | |||
| 1135 | return KnownOut; | |||
| 1136 | } | |||
| 1137 | ||||
| 1138 | // Public so this can be used in `SimplifyDemandedUseBits`. | |||
| 1139 | KnownBits llvm::analyzeKnownBitsFromAndXorOr( | |||
| 1140 | const Operator *I, const KnownBits &KnownLHS, const KnownBits &KnownRHS, | |||
| 1141 | unsigned Depth, const DataLayout &DL, AssumptionCache *AC, | |||
| 1142 | const Instruction *CxtI, const DominatorTree *DT, | |||
| 1143 | OptimizationRemarkEmitter *ORE, bool UseInstrInfo) { | |||
| 1144 | auto *FVTy = dyn_cast<FixedVectorType>(I->getType()); | |||
| 1145 | APInt DemandedElts = | |||
| 1146 | FVTy ? APInt::getAllOnes(FVTy->getNumElements()) : APInt(1, 1); | |||
| 1147 | ||||
| 1148 | return getKnownBitsFromAndXorOr( | |||
| 1149 | I, DemandedElts, KnownLHS, KnownRHS, Depth, | |||
| 1150 | Query(DL, AC, safeCxtI(I, CxtI), DT, UseInstrInfo, ORE)); | |||
| 1151 | } | |||
| 1152 | ||||
| 1153 | ConstantRange llvm::getVScaleRange(const Function *F, unsigned BitWidth) { | |||
| 1154 | Attribute Attr = F->getFnAttribute(Attribute::VScaleRange); | |||
| 1155 | // Without vscale_range, we only know that vscale is non-zero. | |||
| 1156 | if (!Attr.isValid()) | |||
| 1157 | return ConstantRange(APInt(BitWidth, 1), APInt::getZero(BitWidth)); | |||
| 1158 | ||||
| 1159 | unsigned AttrMin = Attr.getVScaleRangeMin(); | |||
| 1160 | // Minimum is larger than vscale width, result is always poison. | |||
| 1161 | if ((unsigned)llvm::bit_width(AttrMin) > BitWidth) | |||
| 1162 | return ConstantRange::getEmpty(BitWidth); | |||
| 1163 | ||||
| 1164 | APInt Min(BitWidth, AttrMin); | |||
| 1165 | std::optional<unsigned> AttrMax = Attr.getVScaleRangeMax(); | |||
| 1166 | if (!AttrMax || (unsigned)llvm::bit_width(*AttrMax) > BitWidth) | |||
| 1167 | return ConstantRange(Min, APInt::getZero(BitWidth)); | |||
| 1168 | ||||
| 1169 | return ConstantRange(Min, APInt(BitWidth, *AttrMax) + 1); | |||
| 1170 | } | |||
| 1171 | ||||
| 1172 | static void computeKnownBitsFromOperator(const Operator *I, | |||
| 1173 | const APInt &DemandedElts, | |||
| 1174 | KnownBits &Known, unsigned Depth, | |||
| 1175 | const Query &Q) { | |||
| 1176 | unsigned BitWidth = Known.getBitWidth(); | |||
| 1177 | ||||
| 1178 | KnownBits Known2(BitWidth); | |||
| 1179 | switch (I->getOpcode()) { | |||
| 1180 | default: break; | |||
| 1181 | case Instruction::Load: | |||
| 1182 | if (MDNode *MD = | |||
| 1183 | Q.IIQ.getMetadata(cast<LoadInst>(I), LLVMContext::MD_range)) | |||
| 1184 | computeKnownBitsFromRangeMetadata(*MD, Known); | |||
| 1185 | break; | |||
| 1186 | case Instruction::And: | |||
| 1187 | computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); | |||
| 1188 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | |||
| 1189 | ||||
| 1190 | Known = getKnownBitsFromAndXorOr(I, DemandedElts, Known2, Known, Depth, Q); | |||
| 1191 | break; | |||
| 1192 | case Instruction::Or: | |||
| 1193 | computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); | |||
| 1194 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | |||
| 1195 | ||||
| 1196 | Known = getKnownBitsFromAndXorOr(I, DemandedElts, Known2, Known, Depth, Q); | |||
| 1197 | break; | |||
| 1198 | case Instruction::Xor: | |||
| 1199 | computeKnownBits(I->getOperand(1), DemandedElts, Known, Depth + 1, Q); | |||
| 1200 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | |||
| 1201 | ||||
| 1202 | Known = getKnownBitsFromAndXorOr(I, DemandedElts, Known2, Known, Depth, Q); | |||
| 1203 | break; | |||
| 1204 | case Instruction::Mul: { | |||
| 1205 | bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); | |||
| 1206 | computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, DemandedElts, | |||
| 1207 | Known, Known2, Depth, Q); | |||
| 1208 | break; | |||
| 1209 | } | |||
| 1210 | case Instruction::UDiv: { | |||
| 1211 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
| 1212 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
| 1213 | Known = KnownBits::udiv(Known, Known2); | |||
| 1214 | break; | |||
| 1215 | } | |||
| 1216 | case Instruction::Select: { | |||
| 1217 | const Value *LHS = nullptr, *RHS = nullptr; | |||
| 1218 | SelectPatternFlavor SPF = matchSelectPattern(I, LHS, RHS).Flavor; | |||
| 1219 | if (SelectPatternResult::isMinOrMax(SPF)) { | |||
| 1220 | computeKnownBits(RHS, Known, Depth + 1, Q); | |||
| 1221 | computeKnownBits(LHS, Known2, Depth + 1, Q); | |||
| 1222 | switch (SPF) { | |||
| 1223 | default: | |||
| 1224 | llvm_unreachable("Unhandled select pattern flavor!")::llvm::llvm_unreachable_internal("Unhandled select pattern flavor!" , "llvm/lib/Analysis/ValueTracking.cpp", 1224); | |||
| 1225 | case SPF_SMAX: | |||
| 1226 | Known = KnownBits::smax(Known, Known2); | |||
| 1227 | break; | |||
| 1228 | case SPF_SMIN: | |||
| 1229 | Known = KnownBits::smin(Known, Known2); | |||
| 1230 | break; | |||
| 1231 | case SPF_UMAX: | |||
| 1232 | Known = KnownBits::umax(Known, Known2); | |||
| 1233 | break; | |||
| 1234 | case SPF_UMIN: | |||
| 1235 | Known = KnownBits::umin(Known, Known2); | |||
| 1236 | break; | |||
| 1237 | } | |||
| 1238 | break; | |||
| 1239 | } | |||
| 1240 | ||||
| 1241 | computeKnownBits(I->getOperand(2), Known, Depth + 1, Q); | |||
| 1242 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
| 1243 | ||||
| 1244 | // Only known if known in both the LHS and RHS. | |||
| 1245 | Known = KnownBits::commonBits(Known, Known2); | |||
| 1246 | ||||
| 1247 | if (SPF == SPF_ABS) { | |||
| 1248 | // RHS from matchSelectPattern returns the negation part of abs pattern. | |||
| 1249 | // If the negate has an NSW flag we can assume the sign bit of the result | |||
| 1250 | // will be 0 because that makes abs(INT_MIN) undefined. | |||
| 1251 | if (match(RHS, m_Neg(m_Specific(LHS))) && | |||
| 1252 | Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(RHS))) | |||
| 1253 | Known.Zero.setSignBit(); | |||
| 1254 | } | |||
| 1255 | ||||
| 1256 | break; | |||
| 1257 | } | |||
| 1258 | case Instruction::FPTrunc: | |||
| 1259 | case Instruction::FPExt: | |||
| 1260 | case Instruction::FPToUI: | |||
| 1261 | case Instruction::FPToSI: | |||
| 1262 | case Instruction::SIToFP: | |||
| 1263 | case Instruction::UIToFP: | |||
| 1264 | break; // Can't work with floating point. | |||
| 1265 | case Instruction::PtrToInt: | |||
| 1266 | case Instruction::IntToPtr: | |||
| 1267 | // Fall through and handle them the same as zext/trunc. | |||
| 1268 | [[fallthrough]]; | |||
| 1269 | case Instruction::ZExt: | |||
| 1270 | case Instruction::Trunc: { | |||
| 1271 | Type *SrcTy = I->getOperand(0)->getType(); | |||
| 1272 | ||||
| 1273 | unsigned SrcBitWidth; | |||
| 1274 | // Note that we handle pointer operands here because of inttoptr/ptrtoint | |||
| 1275 | // which fall through here. | |||
| 1276 | Type *ScalarTy = SrcTy->getScalarType(); | |||
| 1277 | SrcBitWidth = ScalarTy->isPointerTy() ? | |||
| 1278 | Q.DL.getPointerTypeSizeInBits(ScalarTy) : | |||
| 1279 | Q.DL.getTypeSizeInBits(ScalarTy); | |||
| 1280 | ||||
| 1281 | assert(SrcBitWidth && "SrcBitWidth can't be zero")(static_cast <bool> (SrcBitWidth && "SrcBitWidth can't be zero" ) ? void (0) : __assert_fail ("SrcBitWidth && \"SrcBitWidth can't be zero\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1281, __extension__ __PRETTY_FUNCTION__ )); | |||
| 1282 | Known = Known.anyextOrTrunc(SrcBitWidth); | |||
| 1283 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
| 1284 | Known = Known.zextOrTrunc(BitWidth); | |||
| 1285 | break; | |||
| 1286 | } | |||
| 1287 | case Instruction::BitCast: { | |||
| 1288 | Type *SrcTy = I->getOperand(0)->getType(); | |||
| 1289 | if (SrcTy->isIntOrPtrTy() && | |||
| 1290 | // TODO: For now, not handling conversions like: | |||
| 1291 | // (bitcast i64 %x to <2 x i32>) | |||
| 1292 | !I->getType()->isVectorTy()) { | |||
| 1293 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
| 1294 | break; | |||
| 1295 | } | |||
| 1296 | ||||
| 1297 | // Handle cast from vector integer type to scalar or vector integer. | |||
| 1298 | auto *SrcVecTy = dyn_cast<FixedVectorType>(SrcTy); | |||
| 1299 | if (!SrcVecTy || !SrcVecTy->getElementType()->isIntegerTy() || | |||
| 1300 | !I->getType()->isIntOrIntVectorTy() || | |||
| 1301 | isa<ScalableVectorType>(I->getType())) | |||
| 1302 | break; | |||
| 1303 | ||||
| 1304 | // Look through a cast from narrow vector elements to wider type. | |||
| 1305 | // Examples: v4i32 -> v2i64, v3i8 -> v24 | |||
| 1306 | unsigned SubBitWidth = SrcVecTy->getScalarSizeInBits(); | |||
| 1307 | if (BitWidth % SubBitWidth == 0) { | |||
| 1308 | // Known bits are automatically intersected across demanded elements of a | |||
| 1309 | // vector. So for example, if a bit is computed as known zero, it must be | |||
| 1310 | // zero across all demanded elements of the vector. | |||
| 1311 | // | |||
| 1312 | // For this bitcast, each demanded element of the output is sub-divided | |||
| 1313 | // across a set of smaller vector elements in the source vector. To get | |||
| 1314 | // the known bits for an entire element of the output, compute the known | |||
| 1315 | // bits for each sub-element sequentially. This is done by shifting the | |||
| 1316 | // one-set-bit demanded elements parameter across the sub-elements for | |||
| 1317 | // consecutive calls to computeKnownBits. We are using the demanded | |||
| 1318 | // elements parameter as a mask operator. | |||
| 1319 | // | |||
| 1320 | // The known bits of each sub-element are then inserted into place | |||
| 1321 | // (dependent on endian) to form the full result of known bits. | |||
| 1322 | unsigned NumElts = DemandedElts.getBitWidth(); | |||
| 1323 | unsigned SubScale = BitWidth / SubBitWidth; | |||
| 1324 | APInt SubDemandedElts = APInt::getZero(NumElts * SubScale); | |||
| 1325 | for (unsigned i = 0; i != NumElts; ++i) { | |||
| 1326 | if (DemandedElts[i]) | |||
| 1327 | SubDemandedElts.setBit(i * SubScale); | |||
| 1328 | } | |||
| 1329 | ||||
| 1330 | KnownBits KnownSrc(SubBitWidth); | |||
| 1331 | for (unsigned i = 0; i != SubScale; ++i) { | |||
| 1332 | computeKnownBits(I->getOperand(0), SubDemandedElts.shl(i), KnownSrc, | |||
| 1333 | Depth + 1, Q); | |||
| 1334 | unsigned ShiftElt = Q.DL.isLittleEndian() ? i : SubScale - 1 - i; | |||
| 1335 | Known.insertBits(KnownSrc, ShiftElt * SubBitWidth); | |||
| 1336 | } | |||
| 1337 | } | |||
| 1338 | break; | |||
| 1339 | } | |||
| 1340 | case Instruction::SExt: { | |||
| 1341 | // Compute the bits in the result that are not present in the input. | |||
| 1342 | unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits(); | |||
| 1343 | ||||
| 1344 | Known = Known.trunc(SrcBitWidth); | |||
| 1345 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
| 1346 | // If the sign bit of the input is known set or clear, then we know the | |||
| 1347 | // top bits of the result. | |||
| 1348 | Known = Known.sext(BitWidth); | |||
| 1349 | break; | |||
| 1350 | } | |||
| 1351 | case Instruction::Shl: { | |||
| 1352 | bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); | |||
| 1353 | auto KF = [NSW](const KnownBits &KnownVal, const KnownBits &KnownAmt) { | |||
| 1354 | KnownBits Result = KnownBits::shl(KnownVal, KnownAmt); | |||
| 1355 | // If this shift has "nsw" keyword, then the result is either a poison | |||
| 1356 | // value or has the same sign bit as the first operand. | |||
| 1357 | if (NSW) { | |||
| 1358 | if (KnownVal.Zero.isSignBitSet()) | |||
| 1359 | Result.Zero.setSignBit(); | |||
| 1360 | if (KnownVal.One.isSignBitSet()) | |||
| 1361 | Result.One.setSignBit(); | |||
| 1362 | } | |||
| 1363 | return Result; | |||
| 1364 | }; | |||
| 1365 | computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, | |||
| 1366 | KF); | |||
| 1367 | // Trailing zeros of a right-shifted constant never decrease. | |||
| 1368 | const APInt *C; | |||
| 1369 | if (match(I->getOperand(0), m_APInt(C))) | |||
| 1370 | Known.Zero.setLowBits(C->countr_zero()); | |||
| 1371 | break; | |||
| 1372 | } | |||
| 1373 | case Instruction::LShr: { | |||
| 1374 | auto KF = [](const KnownBits &KnownVal, const KnownBits &KnownAmt) { | |||
| 1375 | return KnownBits::lshr(KnownVal, KnownAmt); | |||
| 1376 | }; | |||
| 1377 | computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, | |||
| 1378 | KF); | |||
| 1379 | // Leading zeros of a left-shifted constant never decrease. | |||
| 1380 | const APInt *C; | |||
| 1381 | if (match(I->getOperand(0), m_APInt(C))) | |||
| 1382 | Known.Zero.setHighBits(C->countl_zero()); | |||
| 1383 | break; | |||
| 1384 | } | |||
| 1385 | case Instruction::AShr: { | |||
| 1386 | auto KF = [](const KnownBits &KnownVal, const KnownBits &KnownAmt) { | |||
| 1387 | return KnownBits::ashr(KnownVal, KnownAmt); | |||
| 1388 | }; | |||
| 1389 | computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Depth, Q, | |||
| 1390 | KF); | |||
| 1391 | break; | |||
| 1392 | } | |||
| 1393 | case Instruction::Sub: { | |||
| 1394 | bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); | |||
| 1395 | computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW, | |||
| 1396 | DemandedElts, Known, Known2, Depth, Q); | |||
| 1397 | break; | |||
| 1398 | } | |||
| 1399 | case Instruction::Add: { | |||
| 1400 | bool NSW = Q.IIQ.hasNoSignedWrap(cast<OverflowingBinaryOperator>(I)); | |||
| 1401 | computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW, | |||
| 1402 | DemandedElts, Known, Known2, Depth, Q); | |||
| 1403 | break; | |||
| 1404 | } | |||
| 1405 | case Instruction::SRem: | |||
| 1406 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
| 1407 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
| 1408 | Known = KnownBits::srem(Known, Known2); | |||
| 1409 | break; | |||
| 1410 | ||||
| 1411 | case Instruction::URem: | |||
| 1412 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
| 1413 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
| 1414 | Known = KnownBits::urem(Known, Known2); | |||
| 1415 | break; | |||
| 1416 | case Instruction::Alloca: | |||
| 1417 | Known.Zero.setLowBits(Log2(cast<AllocaInst>(I)->getAlign())); | |||
| 1418 | break; | |||
| 1419 | case Instruction::GetElementPtr: { | |||
| 1420 | // Analyze all of the subscripts of this getelementptr instruction | |||
| 1421 | // to determine if we can prove known low zero bits. | |||
| 1422 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
| 1423 | // Accumulate the constant indices in a separate variable | |||
| 1424 | // to minimize the number of calls to computeForAddSub. | |||
| 1425 | APInt AccConstIndices(BitWidth, 0, /*IsSigned*/ true); | |||
| 1426 | ||||
| 1427 | gep_type_iterator GTI = gep_type_begin(I); | |||
| 1428 | for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) { | |||
| 1429 | // TrailZ can only become smaller, short-circuit if we hit zero. | |||
| 1430 | if (Known.isUnknown()) | |||
| 1431 | break; | |||
| 1432 | ||||
| 1433 | Value *Index = I->getOperand(i); | |||
| 1434 | ||||
| 1435 | // Handle case when index is zero. | |||
| 1436 | Constant *CIndex = dyn_cast<Constant>(Index); | |||
| 1437 | if (CIndex && CIndex->isZeroValue()) | |||
| 1438 | continue; | |||
| 1439 | ||||
| 1440 | if (StructType *STy = GTI.getStructTypeOrNull()) { | |||
| 1441 | // Handle struct member offset arithmetic. | |||
| 1442 | ||||
| 1443 | assert(CIndex &&(static_cast <bool> (CIndex && "Access to structure field must be known at compile time" ) ? void (0) : __assert_fail ("CIndex && \"Access to structure field must be known at compile time\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1444, __extension__ __PRETTY_FUNCTION__ )) | |||
| 1444 | "Access to structure field must be known at compile time")(static_cast <bool> (CIndex && "Access to structure field must be known at compile time" ) ? void (0) : __assert_fail ("CIndex && \"Access to structure field must be known at compile time\"" , "llvm/lib/Analysis/ValueTracking.cpp", 1444, __extension__ __PRETTY_FUNCTION__ )); | |||
| 1445 | ||||
| 1446 | if (CIndex->getType()->isVectorTy()) | |||
| 1447 | Index = CIndex->getSplatValue(); | |||
| 1448 | ||||
| 1449 | unsigned Idx = cast<ConstantInt>(Index)->getZExtValue(); | |||
| 1450 | const StructLayout *SL = Q.DL.getStructLayout(STy); | |||
| 1451 | uint64_t Offset = SL->getElementOffset(Idx); | |||
| 1452 | AccConstIndices += Offset; | |||
| 1453 | continue; | |||
| 1454 | } | |||
| 1455 | ||||
| 1456 | // Handle array index arithmetic. | |||
| 1457 | Type *IndexedTy = GTI.getIndexedType(); | |||
| 1458 | if (!IndexedTy->isSized()) { | |||
| 1459 | Known.resetAll(); | |||
| 1460 | break; | |||
| 1461 | } | |||
| 1462 | ||||
| 1463 | unsigned IndexBitWidth = Index->getType()->getScalarSizeInBits(); | |||
| 1464 | KnownBits IndexBits(IndexBitWidth); | |||
| 1465 | computeKnownBits(Index, IndexBits, Depth + 1, Q); | |||
| 1466 | TypeSize IndexTypeSize = Q.DL.getTypeAllocSize(IndexedTy); | |||
| 1467 | uint64_t TypeSizeInBytes = IndexTypeSize.getKnownMinValue(); | |||
| 1468 | KnownBits ScalingFactor(IndexBitWidth); | |||
| 1469 | // Multiply by current sizeof type. | |||
| 1470 | // &A[i] == A + i * sizeof(*A[i]). | |||
| 1471 | if (IndexTypeSize.isScalable()) { | |||
| 1472 | // For scalable types the only thing we know about sizeof is | |||
| 1473 | // that this is a multiple of the minimum size. | |||
| 1474 | ScalingFactor.Zero.setLowBits(llvm::countr_zero(TypeSizeInBytes)); | |||
| 1475 | } else if (IndexBits.isConstant()) { | |||
| 1476 | APInt IndexConst = IndexBits.getConstant(); | |||
| 1477 | APInt ScalingFactor(IndexBitWidth, TypeSizeInBytes); | |||
| 1478 | IndexConst *= ScalingFactor; | |||
| 1479 | AccConstIndices += IndexConst.sextOrTrunc(BitWidth); | |||
| 1480 | continue; | |||
| 1481 | } else { | |||
| 1482 | ScalingFactor = | |||
| 1483 | KnownBits::makeConstant(APInt(IndexBitWidth, TypeSizeInBytes)); | |||
| 1484 | } | |||
| 1485 | IndexBits = KnownBits::mul(IndexBits, ScalingFactor); | |||
| 1486 | ||||
| 1487 | // If the offsets have a different width from the pointer, according | |||
| 1488 | // to the language reference we need to sign-extend or truncate them | |||
| 1489 | // to the width of the pointer. | |||
| 1490 | IndexBits = IndexBits.sextOrTrunc(BitWidth); | |||
| 1491 | ||||
| 1492 | // Note that inbounds does *not* guarantee nsw for the addition, as only | |||
| 1493 | // the offset is signed, while the base address is unsigned. | |||
| 1494 | Known = KnownBits::computeForAddSub( | |||
| 1495 | /*Add=*/true, /*NSW=*/false, Known, IndexBits); | |||
| 1496 | } | |||
| 1497 | if (!Known.isUnknown() && !AccConstIndices.isZero()) { | |||
| 1498 | KnownBits Index = KnownBits::makeConstant(AccConstIndices); | |||
| 1499 | Known = KnownBits::computeForAddSub( | |||
| 1500 | /*Add=*/true, /*NSW=*/false, Known, Index); | |||
| 1501 | } | |||
| 1502 | break; | |||
| 1503 | } | |||
| 1504 | case Instruction::PHI: { | |||
| 1505 | const PHINode *P = cast<PHINode>(I); | |||
| 1506 | BinaryOperator *BO = nullptr; | |||
| 1507 | Value *R = nullptr, *L = nullptr; | |||
| 1508 | if (matchSimpleRecurrence(P, BO, R, L)) { | |||
| 1509 | // Handle the case of a simple two-predecessor recurrence PHI. | |||
| 1510 | // There's a lot more that could theoretically be done here, but | |||
| 1511 | // this is sufficient to catch some interesting cases. | |||
| 1512 | unsigned Opcode = BO->getOpcode(); | |||
| 1513 | ||||
| 1514 | // If this is a shift recurrence, we know the bits being shifted in. | |||
| 1515 | // We can combine that with information about the start value of the | |||
| 1516 | // recurrence to conclude facts about the result. | |||
| 1517 | if ((Opcode == Instruction::LShr || Opcode == Instruction::AShr || | |||
| 1518 | Opcode == Instruction::Shl) && | |||
| 1519 | BO->getOperand(0) == I) { | |||
| 1520 | ||||
| 1521 | // We have matched a recurrence of the form: | |||
| 1522 | // %iv = [R, %entry], [%iv.next, %backedge] | |||
| 1523 | // %iv.next = shift_op %iv, L | |||
| 1524 | ||||
| 1525 | // Recurse with the phi context to avoid concern about whether facts | |||
| 1526 | // inferred hold at original context instruction. TODO: It may be | |||
| 1527 | // correct to use the original context. IF warranted, explore and | |||
| 1528 | // add sufficient tests to cover. | |||
| 1529 | Query RecQ = Q; | |||
| 1530 | RecQ.CxtI = P; | |||
| 1531 | computeKnownBits(R, DemandedElts, Known2, Depth + 1, RecQ); | |||
| 1532 | switch (Opcode) { | |||
| 1533 | case Instruction::Shl: | |||
| 1534 | // A shl recurrence will only increase the tailing zeros | |||
| 1535 | Known.Zero.setLowBits(Known2.countMinTrailingZeros()); | |||
| 1536 | break; | |||
| 1537 | case Instruction::LShr: | |||
| 1538 | // A lshr recurrence will preserve the leading zeros of the | |||
| 1539 | // start value | |||
| 1540 | Known.Zero.setHighBits(Known2.countMinLeadingZeros()); | |||
| 1541 | break; | |||
| 1542 | case Instruction::AShr: | |||
| 1543 | // An ashr recurrence will extend the initial sign bit | |||
| 1544 | Known.Zero.setHighBits(Known2.countMinLeadingZeros()); | |||
| 1545 | Known.One.setHighBits(Known2.countMinLeadingOnes()); | |||
| 1546 | break; | |||
| 1547 | }; | |||
| 1548 | } | |||
| 1549 | ||||
| 1550 | // Check for operations that have the property that if | |||
| 1551 | // both their operands have low zero bits, the result | |||
| 1552 | // will have low zero bits. | |||
| 1553 | if (Opcode == Instruction::Add || | |||
| 1554 | Opcode == Instruction::Sub || | |||
| 1555 | Opcode == Instruction::And || | |||
| 1556 | Opcode == Instruction::Or || | |||
| 1557 | Opcode == Instruction::Mul) { | |||
| 1558 | // Change the context instruction to the "edge" that flows into the | |||
| 1559 | // phi. This is important because that is where the value is actually | |||
| 1560 | // "evaluated" even though it is used later somewhere else. (see also | |||
| 1561 | // D69571). | |||
| 1562 | Query RecQ = Q; | |||
| 1563 | ||||
| 1564 | unsigned OpNum = P->getOperand(0) == R ? 0 : 1; | |||
| 1565 | Instruction *RInst = P->getIncomingBlock(OpNum)->getTerminator(); | |||
| 1566 | Instruction *LInst = P->getIncomingBlock(1-OpNum)->getTerminator(); | |||
| 1567 | ||||
| 1568 | // Ok, we have a PHI of the form L op= R. Check for low | |||
| 1569 | // zero bits. | |||
| 1570 | RecQ.CxtI = RInst; | |||
| 1571 | computeKnownBits(R, Known2, Depth + 1, RecQ); | |||
| 1572 | ||||
| 1573 | // We need to take the minimum number of known bits | |||
| 1574 | KnownBits Known3(BitWidth); | |||
| 1575 | RecQ.CxtI = LInst; | |||
| 1576 | computeKnownBits(L, Known3, Depth + 1, RecQ); | |||
| 1577 | ||||
| 1578 | Known.Zero.setLowBits(std::min(Known2.countMinTrailingZeros(), | |||
| 1579 | Known3.countMinTrailingZeros())); | |||
| 1580 | ||||
| 1581 | auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(BO); | |||
| 1582 | if (OverflowOp && Q.IIQ.hasNoSignedWrap(OverflowOp)) { | |||
| 1583 | // If initial value of recurrence is nonnegative, and we are adding | |||
| 1584 | // a nonnegative number with nsw, the result can only be nonnegative | |||
| 1585 | // or poison value regardless of the number of times we execute the | |||
| 1586 | // add in phi recurrence. If initial value is negative and we are | |||
| 1587 | // adding a negative number with nsw, the result can only be | |||
| 1588 | // negative or poison value. Similar arguments apply to sub and mul. | |||
| 1589 | // | |||
| 1590 | // (add non-negative, non-negative) --> non-negative | |||
| 1591 | // (add negative, negative) --> negative | |||
| 1592 | if (Opcode == Instruction::Add) { | |||
| 1593 | if (Known2.isNonNegative() && Known3.isNonNegative()) | |||
| 1594 | Known.makeNonNegative(); | |||
| 1595 | else if (Known2.isNegative() && Known3.isNegative()) | |||
| 1596 | Known.makeNegative(); | |||
| 1597 | } | |||
| 1598 | ||||
| 1599 | // (sub nsw non-negative, negative) --> non-negative | |||
| 1600 | // (sub nsw negative, non-negative) --> negative | |||
| 1601 | else if (Opcode == Instruction::Sub && BO->getOperand(0) == I) { | |||
| 1602 | if (Known2.isNonNegative() && Known3.isNegative()) | |||
| 1603 | Known.makeNonNegative(); | |||
| 1604 | else if (Known2.isNegative() && Known3.isNonNegative()) | |||
| 1605 | Known.makeNegative(); | |||
| 1606 | } | |||
| 1607 | ||||
| 1608 | // (mul nsw non-negative, non-negative) --> non-negative | |||
| 1609 | else if (Opcode == Instruction::Mul && Known2.isNonNegative() && | |||
| 1610 | Known3.isNonNegative()) | |||
| 1611 | Known.makeNonNegative(); | |||
| 1612 | } | |||
| 1613 | ||||
| 1614 | break; | |||
| 1615 | } | |||
| 1616 | } | |||
| 1617 | ||||
| 1618 | // Unreachable blocks may have zero-operand PHI nodes. | |||
| 1619 | if (P->getNumIncomingValues() == 0) | |||
| 1620 | break; | |||
| 1621 | ||||
| 1622 | // Otherwise take the unions of the known bit sets of the operands, | |||
| 1623 | // taking conservative care to avoid excessive recursion. | |||
| 1624 | if (Depth < MaxAnalysisRecursionDepth - 1 && !Known.Zero && !Known.One) { | |||
| 1625 | // Skip if every incoming value references to ourself. | |||
| 1626 | if (isa_and_nonnull<UndefValue>(P->hasConstantValue())) | |||
| 1627 | break; | |||
| 1628 | ||||
| 1629 | Known.Zero.setAllBits(); | |||
| 1630 | Known.One.setAllBits(); | |||
| 1631 | for (unsigned u = 0, e = P->getNumIncomingValues(); u < e; ++u) { | |||
| 1632 | Value *IncValue = P->getIncomingValue(u); | |||
| 1633 | // Skip direct self references. | |||
| 1634 | if (IncValue == P) continue; | |||
| 1635 | ||||
| 1636 | // Change the context instruction to the "edge" that flows into the | |||
| 1637 | // phi. This is important because that is where the value is actually | |||
| 1638 | // "evaluated" even though it is used later somewhere else. (see also | |||
| 1639 | // D69571). | |||
| 1640 | Query RecQ = Q; | |||
| 1641 | RecQ.CxtI = P->getIncomingBlock(u)->getTerminator(); | |||
| 1642 | ||||
| 1643 | Known2 = KnownBits(BitWidth); | |||
| 1644 | ||||
| 1645 | // Recurse, but cap the recursion to one level, because we don't | |||
| 1646 | // want to waste time spinning around in loops. | |||
| 1647 | computeKnownBits(IncValue, Known2, MaxAnalysisRecursionDepth - 1, RecQ); | |||
| 1648 | ||||
| 1649 | // If this failed, see if we can use a conditional branch into the phi | |||
| 1650 | // to help us determine the range of the value. | |||
| 1651 | if (Known2.isUnknown()) { | |||
| 1652 | ICmpInst::Predicate Pred; | |||
| 1653 | const APInt *RHSC; | |||
| 1654 | BasicBlock *TrueSucc, *FalseSucc; | |||
| 1655 | // TODO: Use RHS Value and compute range from its known bits. | |||
| 1656 | if (match(RecQ.CxtI, | |||
| 1657 | m_Br(m_c_ICmp(Pred, m_Specific(IncValue), m_APInt(RHSC)), | |||
| 1658 | m_BasicBlock(TrueSucc), m_BasicBlock(FalseSucc)))) { | |||
| 1659 | // Check for cases of duplicate successors. | |||
| 1660 | if ((TrueSucc == P->getParent()) != (FalseSucc == P->getParent())) { | |||
| 1661 | // If we're using the false successor, invert the predicate. | |||
| 1662 | if (FalseSucc == P->getParent()) | |||
| 1663 | Pred = CmpInst::getInversePredicate(Pred); | |||
| 1664 | ||||
| 1665 | switch (Pred) { | |||
| 1666 | case CmpInst::Predicate::ICMP_EQ: | |||
| 1667 | Known2 = KnownBits::makeConstant(*RHSC); | |||
| 1668 | break; | |||
| 1669 | case CmpInst::Predicate::ICMP_ULE: | |||
| 1670 | Known2.Zero.setHighBits(RHSC->countl_zero()); | |||
| 1671 | break; | |||
| 1672 | case CmpInst::Predicate::ICMP_ULT: | |||
| 1673 | Known2.Zero.setHighBits((*RHSC - 1).countl_zero()); | |||
| 1674 | break; | |||
| 1675 | default: | |||
| 1676 | // TODO - add additional integer predicate handling. | |||
| 1677 | break; | |||
| 1678 | } | |||
| 1679 | } | |||
| 1680 | } | |||
| 1681 | } | |||
| 1682 | ||||
| 1683 | Known = KnownBits::commonBits(Known, Known2); | |||
| 1684 | // If all bits have been ruled out, there's no need to check | |||
| 1685 | // more operands. | |||
| 1686 | if (Known.isUnknown()) | |||
| 1687 | break; | |||
| 1688 | } | |||
| 1689 | } | |||
| 1690 | break; | |||
| 1691 | } | |||
| 1692 | case Instruction::Call: | |||
| 1693 | case Instruction::Invoke: | |||
| 1694 | // If range metadata is attached to this call, set known bits from that, | |||
| 1695 | // and then intersect with known bits based on other properties of the | |||
| 1696 | // function. | |||
| 1697 | if (MDNode *MD = | |||
| 1698 | Q.IIQ.getMetadata(cast<Instruction>(I), LLVMContext::MD_range)) | |||
| 1699 | computeKnownBitsFromRangeMetadata(*MD, Known); | |||
| 1700 | if (const Value *RV = cast<CallBase>(I)->getReturnedArgOperand()) { | |||
| 1701 | computeKnownBits(RV, Known2, Depth + 1, Q); | |||
| 1702 | Known.Zero |= Known2.Zero; | |||
| 1703 | Known.One |= Known2.One; | |||
| 1704 | } | |||
| 1705 | if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { | |||
| 1706 | switch (II->getIntrinsicID()) { | |||
| 1707 | default: break; | |||
| 1708 | case Intrinsic::abs: { | |||
| 1709 | computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); | |||
| 1710 | bool IntMinIsPoison = match(II->getArgOperand(1), m_One()); | |||
| 1711 | Known = Known2.abs(IntMinIsPoison); | |||
| 1712 | break; | |||
| 1713 | } | |||
| 1714 | case Intrinsic::bitreverse: | |||
| 1715 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | |||
| 1716 | Known.Zero |= Known2.Zero.reverseBits(); | |||
| 1717 | Known.One |= Known2.One.reverseBits(); | |||
| 1718 | break; | |||
| 1719 | case Intrinsic::bswap: | |||
| 1720 | computeKnownBits(I->getOperand(0), DemandedElts, Known2, Depth + 1, Q); | |||
| 1721 | Known.Zero |= Known2.Zero.byteSwap(); | |||
| 1722 | Known.One |= Known2.One.byteSwap(); | |||
| 1723 | break; | |||
| 1724 | case Intrinsic::ctlz: { | |||
| 1725 | computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); | |||
| 1726 | // If we have a known 1, its position is our upper bound. | |||
| 1727 | unsigned PossibleLZ = Known2.countMaxLeadingZeros(); | |||
| 1728 | // If this call is poison for 0 input, the result will be less than 2^n. | |||
| 1729 | if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext())) | |||
| 1730 | PossibleLZ = std::min(PossibleLZ, BitWidth - 1); | |||
| 1731 | unsigned LowBits = llvm::bit_width(PossibleLZ); | |||
| 1732 | Known.Zero.setBitsFrom(LowBits); | |||
| 1733 | break; | |||
| 1734 | } | |||
| 1735 | case Intrinsic::cttz: { | |||
| 1736 | computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); | |||
| 1737 | // If we have a known 1, its position is our upper bound. | |||
| 1738 | unsigned PossibleTZ = Known2.countMaxTrailingZeros(); | |||
| 1739 | // If this call is poison for 0 input, the result will be less than 2^n. | |||
| 1740 | if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext())) | |||
| 1741 | PossibleTZ = std::min(PossibleTZ, BitWidth - 1); | |||
| 1742 | unsigned LowBits = llvm::bit_width(PossibleTZ); | |||
| 1743 | Known.Zero.setBitsFrom(LowBits); | |||
| 1744 | break; | |||
| 1745 | } | |||
| 1746 | case Intrinsic::ctpop: { | |||
| 1747 | computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); | |||
| 1748 | // We can bound the space the count needs. Also, bits known to be zero | |||
| 1749 | // can't contribute to the population. | |||
| 1750 | unsigned BitsPossiblySet = Known2.countMaxPopulation(); | |||
| 1751 | unsigned LowBits = llvm::bit_width(BitsPossiblySet); | |||
| 1752 | Known.Zero.setBitsFrom(LowBits); | |||
| 1753 | // TODO: we could bound KnownOne using the lower bound on the number | |||
| 1754 | // of bits which might be set provided by popcnt KnownOne2. | |||
| 1755 | break; | |||
| 1756 | } | |||
| 1757 | case Intrinsic::fshr: | |||
| 1758 | case Intrinsic::fshl: { | |||
| 1759 | const APInt *SA; | |||
| 1760 | if (!match(I->getOperand(2), m_APInt(SA))) | |||
| 1761 | break; | |||
| 1762 | ||||
| 1763 | // Normalize to funnel shift left. | |||
| 1764 | uint64_t ShiftAmt = SA->urem(BitWidth); | |||
| 1765 | if (II->getIntrinsicID() == Intrinsic::fshr) | |||
| 1766 | ShiftAmt = BitWidth - ShiftAmt; | |||
| 1767 | ||||
| 1768 | KnownBits Known3(BitWidth); | |||
| 1769 | computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q); | |||
| 1770 | computeKnownBits(I->getOperand(1), Known3, Depth + 1, Q); | |||
| 1771 | ||||
| 1772 | Known.Zero = | |||
| 1773 | Known2.Zero.shl(ShiftAmt) | Known3.Zero.lshr(BitWidth - ShiftAmt); | |||
| 1774 | Known.One = | |||
| 1775 | Known2.One.shl(ShiftAmt) | Known3.One.lshr(BitWidth - ShiftAmt); | |||
| 1776 | break; | |||
| 1777 | } | |||
| 1778 | case Intrinsic::uadd_sat: | |||
| 1779 | case Intrinsic::usub_sat: { | |||
| 1780 | bool IsAdd = II->getIntrinsicID() == Intrinsic::uadd_sat; | |||
| 1781 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
| 1782 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
| 1783 | ||||
| 1784 | // Add: Leading ones of either operand are preserved. | |||
| 1785 | // Sub: Leading zeros of LHS and leading ones of RHS are preserved | |||
| 1786 | // as leading zeros in the result. | |||
| 1787 | unsigned LeadingKnown; | |||
| 1788 | if (IsAdd) | |||
| 1789 | LeadingKnown = std::max(Known.countMinLeadingOnes(), | |||
| 1790 | Known2.countMinLeadingOnes()); | |||
| 1791 | else | |||
| 1792 | LeadingKnown = std::max(Known.countMinLeadingZeros(), | |||
| 1793 | Known2.countMinLeadingOnes()); | |||
| 1794 | ||||
| 1795 | Known = KnownBits::computeForAddSub( | |||
| 1796 | IsAdd, /* NSW */ false, Known, Known2); | |||
| 1797 | ||||
| 1798 | // We select between the operation result and all-ones/zero | |||
| 1799 | // respectively, so we can preserve known ones/zeros. | |||
| 1800 | if (IsAdd) { | |||
| 1801 | Known.One.setHighBits(LeadingKnown); | |||
| 1802 | Known.Zero.clearAllBits(); | |||
| 1803 | } else { | |||
| 1804 | Known.Zero.setHighBits(LeadingKnown); | |||
| 1805 | Known.One.clearAllBits(); | |||
| 1806 | } | |||
| 1807 | break; | |||
| 1808 | } | |||
| 1809 | case Intrinsic::umin: | |||
| 1810 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
| 1811 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
| 1812 | Known = KnownBits::umin(Known, Known2); | |||
| 1813 | break; | |||
| 1814 | case Intrinsic::umax: | |||
| 1815 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
| 1816 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
| 1817 | Known = KnownBits::umax(Known, Known2); | |||
| 1818 | break; | |||
| 1819 | case Intrinsic::smin: | |||
| 1820 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
| 1821 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
| 1822 | Known = KnownBits::smin(Known, Known2); | |||
| 1823 | break; | |||
| 1824 | case Intrinsic::smax: | |||
| 1825 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
| 1826 | computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q); | |||
| 1827 | Known = KnownBits::smax(Known, Known2); | |||
| 1828 | break; | |||
| 1829 | case Intrinsic::x86_sse42_crc32_64_64: | |||
| 1830 | Known.Zero.setBitsFrom(32); | |||
| 1831 | break; | |||
| 1832 | case Intrinsic::riscv_vsetvli: | |||
| 1833 | case Intrinsic::riscv_vsetvlimax: | |||
| 1834 | // Assume that VL output is >= 65536. | |||
| 1835 | // TODO: Take SEW and LMUL into account. | |||
| 1836 | if (BitWidth > 17) | |||
| 1837 | Known.Zero.setBitsFrom(17); | |||
| 1838 | break; | |||
| 1839 | case Intrinsic::vscale: { | |||
| 1840 | if (!II->getParent() || !II->getFunction()) | |||
| 1841 | break; | |||
| 1842 | ||||
| 1843 | Known = getVScaleRange(II->getFunction(), BitWidth).toKnownBits(); | |||
| 1844 | break; | |||
| 1845 | } | |||
| 1846 | } | |||
| 1847 | } | |||
| 1848 | break; | |||
| 1849 | case Instruction::ShuffleVector: { | |||
| 1850 | auto *Shuf = dyn_cast<ShuffleVectorInst>(I); | |||
| 1851 | // FIXME: Do we need to handle ConstantExpr involving shufflevectors? | |||
| 1852 | if (!Shuf) { | |||
| 1853 | Known.resetAll(); | |||
| 1854 | return; | |||
| 1855 | } | |||
| 1856 | // For undef elements, we don't know anything about the common state of | |||
| 1857 | // the shuffle result. | |||
| 1858 | APInt DemandedLHS, DemandedRHS; | |||
| 1859 | if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) { | |||
| 1860 | Known.resetAll(); | |||
| 1861 | return; | |||
| 1862 | } | |||
| 1863 | Known.One.setAllBits(); | |||
| 1864 | Known.Zero.setAllBits(); | |||
| 1865 | if (!!DemandedLHS) { | |||
| 1866 | const Value *LHS = Shuf->getOperand(0); | |||
| 1867 | computeKnownBits(LHS, DemandedLHS, Known, Depth + 1, Q); | |||
| 1868 | // If we don't know any bits, early out. | |||
| 1869 | if (Known.isUnknown()) | |||
| 1870 | break; | |||
| 1871 | } | |||
| 1872 | if (!!DemandedRHS) { | |||
| 1873 | const Value *RHS = Shuf->getOperand(1); | |||
| 1874 | computeKnownBits(RHS, DemandedRHS, Known2, Depth + 1, Q); | |||
| 1875 | Known = KnownBits::commonBits(Known, Known2); | |||
| 1876 | } | |||
| 1877 | break; | |||
| 1878 | } | |||
| 1879 | case Instruction::InsertElement: { | |||
| 1880 | if (isa<ScalableVectorType>(I->getType())) { | |||
| 1881 | Known.resetAll(); | |||
| 1882 | return; | |||
| 1883 | } | |||
| 1884 | const Value *Vec = I->getOperand(0); | |||
| 1885 | const Value *Elt = I->getOperand(1); | |||
| 1886 | auto *CIdx = dyn_cast<ConstantInt>(I->getOperand(2)); | |||
| 1887 | // Early out if the index is non-constant or out-of-range. | |||
| 1888 | unsigned NumElts = DemandedElts.getBitWidth(); | |||
| 1889 | if (!CIdx || CIdx->getValue().uge(NumElts)) { | |||
| 1890 | Known.resetAll(); | |||
| 1891 | return; | |||
| 1892 | } | |||
| 1893 | Known.One.setAllBits(); | |||
| 1894 | Known.Zero.setAllBits(); | |||
| 1895 | unsigned EltIdx = CIdx->getZExtValue(); | |||
| 1896 | // Do we demand the inserted element? | |||
| 1897 | if (DemandedElts[EltIdx]) { | |||
| 1898 | computeKnownBits(Elt, Known, Depth + 1, Q); | |||
| 1899 | // If we don't know any bits, early out. | |||
| 1900 | if (Known.isUnknown()) | |||
| 1901 | break; | |||
| 1902 | } | |||
| 1903 | // We don't need the base vector element that has been inserted. | |||
| 1904 | APInt DemandedVecElts = DemandedElts; | |||
| 1905 | DemandedVecElts.clearBit(EltIdx); | |||
| 1906 | if (!!DemandedVecElts) { | |||
| 1907 | computeKnownBits(Vec, DemandedVecElts, Known2, Depth + 1, Q); | |||
| 1908 | Known = KnownBits::commonBits(Known, Known2); | |||
| 1909 | } | |||
| 1910 | break; | |||
| 1911 | } | |||
| 1912 | case Instruction::ExtractElement: { | |||
| 1913 | // Look through extract element. If the index is non-constant or | |||
| 1914 | // out-of-range demand all elements, otherwise just the extracted element. | |||
| 1915 | const Value *Vec = I->getOperand(0); | |||
| 1916 | const Value *Idx = I->getOperand(1); | |||
| 1917 | auto *CIdx = dyn_cast<ConstantInt>(Idx); | |||
| 1918 | if (isa<ScalableVectorType>(Vec->getType())) { | |||
| 1919 | // FIXME: there's probably *something* we can do with scalable vectors | |||
| 1920 | Known.resetAll(); | |||
| 1921 | break; | |||
| 1922 | } | |||
| 1923 | unsigned NumElts = cast<FixedVectorType>(Vec->getType())->getNumElements(); | |||
| 1924 | APInt DemandedVecElts = APInt::getAllOnes(NumElts); | |||
| 1925 | if (CIdx && CIdx->getValue().ult(NumElts)) | |||
| 1926 | DemandedVecElts = APInt::getOneBitSet(NumElts, CIdx->getZExtValue()); | |||
| 1927 | computeKnownBits(Vec, DemandedVecElts, Known, Depth + 1, Q); | |||
| 1928 | break; | |||
| 1929 | } | |||
| 1930 | case Instruction::ExtractValue: | |||
| 1931 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) { | |||
| 1932 | const ExtractValueInst *EVI = cast<ExtractValueInst>(I); | |||
| 1933 | if (EVI->getNumIndices() != 1) break; | |||
| 1934 | if (EVI->getIndices()[0] == 0) { | |||
| 1935 | switch (II->getIntrinsicID()) { | |||
| 1936 | default: break; | |||
| 1937 | case Intrinsic::uadd_with_overflow: | |||
| 1938 | case Intrinsic::sadd_with_overflow: | |||
| 1939 | computeKnownBitsAddSub(true, II->getArgOperand(0), | |||
| 1940 | II->getArgOperand(1), false, DemandedElts, | |||
| 1941 | Known, Known2, Depth, Q); | |||
| 1942 | break; | |||
| 1943 | case Intrinsic::usub_with_overflow: | |||
| 1944 | case Intrinsic::ssub_with_overflow: | |||
| 1945 | computeKnownBitsAddSub(false, II->getArgOperand(0), | |||
| 1946 | II->getArgOperand(1), false, DemandedElts, | |||
| 1947 | Known, Known2, Depth, Q); | |||
| 1948 | break; | |||
| 1949 | case Intrinsic::umul_with_overflow: | |||
| 1950 | case Intrinsic::smul_with_overflow: | |||
| 1951 | computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false, | |||
| 1952 | DemandedElts, Known, Known2, Depth, Q); | |||
| 1953 | break; | |||
| 1954 | } | |||
| 1955 | } | |||
| 1956 | } | |||
| 1957 | break; | |||
| 1958 | case Instruction::Freeze: | |||
| 1959 | if (isGuaranteedNotToBePoison(I->getOperand(0), Q.AC, Q.CxtI, Q.DT, | |||
| 1960 | Depth + 1)) | |||
| 1961 | computeKnownBits(I->getOperand(0), Known, Depth + 1, Q); | |||
| 1962 | break; | |||
| 1963 | } | |||
| 1964 | } | |||
| 1965 | ||||
| 1966 | /// Determine which bits of V are known to be either zero or one and return | |||
| 1967 | /// them. | |||
| 1968 | KnownBits computeKnownBits(const Value *V, const APInt &DemandedElts, | |||
| 1969 | unsigned Depth, const Query &Q) { | |||
| 1970 | KnownBits Known(getBitWidth(V->getType(), Q.DL)); | |||
| 1971 | computeKnownBits(V, DemandedElts, Known, Depth, Q); | |||
| 1972 | return Known; | |||
| 1973 | } | |||
| 1974 | ||||
| 1975 | /// Determine which bits of V are known to be either zero or one and return | |||
| 1976 | /// them. | |||
| 1977 | KnownBits computeKnownBits(const Value *V, unsigned Depth, const Query &Q) { | |||
| 1978 | KnownBits Known(getBitWidth(V->getType(), Q.DL)); | |||
| 1979 | computeKnownBits(V, Known, Depth, Q); | |||
| 1980 | return Known; | |||
| 1981 | } | |||
| 1982 | ||||
| 1983 | /// Determine which bits of V are known to be either zero or one and return | |||
| 1984 | /// them in the Known bit set. | |||
| 1985 | /// | |||
| 1986 | /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that | |||
| 1987 | /// we cannot optimize based on the assumption that it is zero without changing | |||
| 1988 | /// it to be an explicit zero. If we don't change it to zero, other code could | |||
| 1989 | /// optimized based on the contradictory assumption that it is non-zero. | |||
| 1990 | /// Because instcombine aggressively folds operations with undef args anyway, | |||
| 1991 | /// this won't lose us code quality. | |||
| 1992 | /// | |||
| 1993 | /// This function is defined on values with integer type, values with pointer | |||
| 1994 | /// type, and vectors of integers. In the case | |||
| 1995 | /// where V is a vector, known zero, and known one values are the | |||
| 1996 | /// same width as the vector element, and the bit is set only if it is true | |||
| 1997 | /// for all of the demanded elements in the vector specified by DemandedElts. | |||
| 1998 | void computeKnownBits(const Value *V, const APInt &DemandedElts, | |||
| 1999 | KnownBits &Known, unsigned Depth, const Query &Q) { | |||
| 2000 | if (!DemandedElts) { | |||
| 2001 | // No demanded elts, better to assume we don't know anything. | |||
| 2002 | Known.resetAll(); | |||
| 2003 | return; | |||
| 2004 | } | |||
| 2005 | ||||
| 2006 | assert(V && "No Value?")(static_cast <bool> (V && "No Value?") ? void ( 0) : __assert_fail ("V && \"No Value?\"", "llvm/lib/Analysis/ValueTracking.cpp" , 2006, __extension__ __PRETTY_FUNCTION__)); | |||
| 2007 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth")(static_cast <bool> (Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth") ? void (0) : __assert_fail ( "Depth <= MaxAnalysisRecursionDepth && \"Limit Search Depth\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2007, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2008 | ||||
| 2009 | #ifndef NDEBUG | |||
| 2010 | Type *Ty = V->getType(); | |||
| 2011 | unsigned BitWidth = Known.getBitWidth(); | |||
| 2012 | ||||
| 2013 | assert((Ty->isIntOrIntVectorTy(BitWidth) || Ty->isPtrOrPtrVectorTy()) &&(static_cast <bool> ((Ty->isIntOrIntVectorTy(BitWidth ) || Ty->isPtrOrPtrVectorTy()) && "Not integer or pointer type!" ) ? void (0) : __assert_fail ("(Ty->isIntOrIntVectorTy(BitWidth) || Ty->isPtrOrPtrVectorTy()) && \"Not integer or pointer type!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2014, __extension__ __PRETTY_FUNCTION__ )) | |||
| 2014 | "Not integer or pointer type!")(static_cast <bool> ((Ty->isIntOrIntVectorTy(BitWidth ) || Ty->isPtrOrPtrVectorTy()) && "Not integer or pointer type!" ) ? void (0) : __assert_fail ("(Ty->isIntOrIntVectorTy(BitWidth) || Ty->isPtrOrPtrVectorTy()) && \"Not integer or pointer type!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2014, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2015 | ||||
| 2016 | if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) { | |||
| 2017 | assert((static_cast <bool> (FVTy->getNumElements() == DemandedElts .getBitWidth() && "DemandedElt width should equal the fixed vector number of elements" ) ? void (0) : __assert_fail ("FVTy->getNumElements() == DemandedElts.getBitWidth() && \"DemandedElt width should equal the fixed vector number of elements\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2019, __extension__ __PRETTY_FUNCTION__ )) | |||
| 2018 | FVTy->getNumElements() == DemandedElts.getBitWidth() &&(static_cast <bool> (FVTy->getNumElements() == DemandedElts .getBitWidth() && "DemandedElt width should equal the fixed vector number of elements" ) ? void (0) : __assert_fail ("FVTy->getNumElements() == DemandedElts.getBitWidth() && \"DemandedElt width should equal the fixed vector number of elements\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2019, __extension__ __PRETTY_FUNCTION__ )) | |||
| 2019 | "DemandedElt width should equal the fixed vector number of elements")(static_cast <bool> (FVTy->getNumElements() == DemandedElts .getBitWidth() && "DemandedElt width should equal the fixed vector number of elements" ) ? void (0) : __assert_fail ("FVTy->getNumElements() == DemandedElts.getBitWidth() && \"DemandedElt width should equal the fixed vector number of elements\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2019, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2020 | } else { | |||
| 2021 | assert(DemandedElts == APInt(1, 1) &&(static_cast <bool> (DemandedElts == APInt(1, 1) && "DemandedElt width should be 1 for scalars or scalable vectors" ) ? void (0) : __assert_fail ("DemandedElts == APInt(1, 1) && \"DemandedElt width should be 1 for scalars or scalable vectors\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2022, __extension__ __PRETTY_FUNCTION__ )) | |||
| 2022 | "DemandedElt width should be 1 for scalars or scalable vectors")(static_cast <bool> (DemandedElts == APInt(1, 1) && "DemandedElt width should be 1 for scalars or scalable vectors" ) ? void (0) : __assert_fail ("DemandedElts == APInt(1, 1) && \"DemandedElt width should be 1 for scalars or scalable vectors\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2022, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2023 | } | |||
| 2024 | ||||
| 2025 | Type *ScalarTy = Ty->getScalarType(); | |||
| 2026 | if (ScalarTy->isPointerTy()) { | |||
| 2027 | assert(BitWidth == Q.DL.getPointerTypeSizeInBits(ScalarTy) &&(static_cast <bool> (BitWidth == Q.DL.getPointerTypeSizeInBits (ScalarTy) && "V and Known should have same BitWidth" ) ? void (0) : __assert_fail ("BitWidth == Q.DL.getPointerTypeSizeInBits(ScalarTy) && \"V and Known should have same BitWidth\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2028, __extension__ __PRETTY_FUNCTION__ )) | |||
| 2028 | "V and Known should have same BitWidth")(static_cast <bool> (BitWidth == Q.DL.getPointerTypeSizeInBits (ScalarTy) && "V and Known should have same BitWidth" ) ? void (0) : __assert_fail ("BitWidth == Q.DL.getPointerTypeSizeInBits(ScalarTy) && \"V and Known should have same BitWidth\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2028, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2029 | } else { | |||
| 2030 | assert(BitWidth == Q.DL.getTypeSizeInBits(ScalarTy) &&(static_cast <bool> (BitWidth == Q.DL.getTypeSizeInBits (ScalarTy) && "V and Known should have same BitWidth" ) ? void (0) : __assert_fail ("BitWidth == Q.DL.getTypeSizeInBits(ScalarTy) && \"V and Known should have same BitWidth\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2031, __extension__ __PRETTY_FUNCTION__ )) | |||
| 2031 | "V and Known should have same BitWidth")(static_cast <bool> (BitWidth == Q.DL.getTypeSizeInBits (ScalarTy) && "V and Known should have same BitWidth" ) ? void (0) : __assert_fail ("BitWidth == Q.DL.getTypeSizeInBits(ScalarTy) && \"V and Known should have same BitWidth\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2031, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2032 | } | |||
| 2033 | #endif | |||
| 2034 | ||||
| 2035 | const APInt *C; | |||
| 2036 | if (match(V, m_APInt(C))) { | |||
| 2037 | // We know all of the bits for a scalar constant or a splat vector constant! | |||
| 2038 | Known = KnownBits::makeConstant(*C); | |||
| 2039 | return; | |||
| 2040 | } | |||
| 2041 | // Null and aggregate-zero are all-zeros. | |||
| 2042 | if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) { | |||
| 2043 | Known.setAllZero(); | |||
| 2044 | return; | |||
| 2045 | } | |||
| 2046 | // Handle a constant vector by taking the intersection of the known bits of | |||
| 2047 | // each element. | |||
| 2048 | if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(V)) { | |||
| 2049 | assert(!isa<ScalableVectorType>(V->getType()))(static_cast <bool> (!isa<ScalableVectorType>(V-> getType())) ? void (0) : __assert_fail ("!isa<ScalableVectorType>(V->getType())" , "llvm/lib/Analysis/ValueTracking.cpp", 2049, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2050 | // We know that CDV must be a vector of integers. Take the intersection of | |||
| 2051 | // each element. | |||
| 2052 | Known.Zero.setAllBits(); Known.One.setAllBits(); | |||
| 2053 | for (unsigned i = 0, e = CDV->getNumElements(); i != e; ++i) { | |||
| 2054 | if (!DemandedElts[i]) | |||
| 2055 | continue; | |||
| 2056 | APInt Elt = CDV->getElementAsAPInt(i); | |||
| 2057 | Known.Zero &= ~Elt; | |||
| 2058 | Known.One &= Elt; | |||
| 2059 | } | |||
| 2060 | return; | |||
| 2061 | } | |||
| 2062 | ||||
| 2063 | if (const auto *CV = dyn_cast<ConstantVector>(V)) { | |||
| 2064 | assert(!isa<ScalableVectorType>(V->getType()))(static_cast <bool> (!isa<ScalableVectorType>(V-> getType())) ? void (0) : __assert_fail ("!isa<ScalableVectorType>(V->getType())" , "llvm/lib/Analysis/ValueTracking.cpp", 2064, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2065 | // We know that CV must be a vector of integers. Take the intersection of | |||
| 2066 | // each element. | |||
| 2067 | Known.Zero.setAllBits(); Known.One.setAllBits(); | |||
| 2068 | for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) { | |||
| 2069 | if (!DemandedElts[i]) | |||
| 2070 | continue; | |||
| 2071 | Constant *Element = CV->getAggregateElement(i); | |||
| 2072 | auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element); | |||
| 2073 | if (!ElementCI) { | |||
| 2074 | Known.resetAll(); | |||
| 2075 | return; | |||
| 2076 | } | |||
| 2077 | const APInt &Elt = ElementCI->getValue(); | |||
| 2078 | Known.Zero &= ~Elt; | |||
| 2079 | Known.One &= Elt; | |||
| 2080 | } | |||
| 2081 | return; | |||
| 2082 | } | |||
| 2083 | ||||
| 2084 | // Start out not knowing anything. | |||
| 2085 | Known.resetAll(); | |||
| 2086 | ||||
| 2087 | // We can't imply anything about undefs. | |||
| 2088 | if (isa<UndefValue>(V)) | |||
| 2089 | return; | |||
| 2090 | ||||
| 2091 | // There's no point in looking through other users of ConstantData for | |||
| 2092 | // assumptions. Confirm that we've handled them all. | |||
| 2093 | assert(!isa<ConstantData>(V) && "Unhandled constant data!")(static_cast <bool> (!isa<ConstantData>(V) && "Unhandled constant data!") ? void (0) : __assert_fail ("!isa<ConstantData>(V) && \"Unhandled constant data!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2093, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2094 | ||||
| 2095 | // All recursive calls that increase depth must come after this. | |||
| 2096 | if (Depth == MaxAnalysisRecursionDepth) | |||
| 2097 | return; | |||
| 2098 | ||||
| 2099 | // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has | |||
| 2100 | // the bits of its aliasee. | |||
| 2101 | if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { | |||
| 2102 | if (!GA->isInterposable()) | |||
| 2103 | computeKnownBits(GA->getAliasee(), Known, Depth + 1, Q); | |||
| 2104 | return; | |||
| 2105 | } | |||
| 2106 | ||||
| 2107 | if (const Operator *I = dyn_cast<Operator>(V)) | |||
| 2108 | computeKnownBitsFromOperator(I, DemandedElts, Known, Depth, Q); | |||
| 2109 | ||||
| 2110 | // Aligned pointers have trailing zeros - refine Known.Zero set | |||
| 2111 | if (isa<PointerType>(V->getType())) { | |||
| 2112 | Align Alignment = V->getPointerAlignment(Q.DL); | |||
| 2113 | Known.Zero.setLowBits(Log2(Alignment)); | |||
| 2114 | } | |||
| 2115 | ||||
| 2116 | // computeKnownBitsFromAssume strictly refines Known. | |||
| 2117 | // Therefore, we run them after computeKnownBitsFromOperator. | |||
| 2118 | ||||
| 2119 | // Check whether a nearby assume intrinsic can determine some known bits. | |||
| 2120 | computeKnownBitsFromAssume(V, Known, Depth, Q); | |||
| 2121 | ||||
| 2122 | assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?")(static_cast <bool> ((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?") ? void (0) : __assert_fail ("(Known.Zero & Known.One) == 0 && \"Bits known to be one AND zero?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2122, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2123 | } | |||
| 2124 | ||||
| 2125 | /// Try to detect a recurrence that the value of the induction variable is | |||
| 2126 | /// always a power of two (or zero). | |||
| 2127 | static bool isPowerOfTwoRecurrence(const PHINode *PN, bool OrZero, | |||
| 2128 | unsigned Depth, Query &Q) { | |||
| 2129 | BinaryOperator *BO = nullptr; | |||
| 2130 | Value *Start = nullptr, *Step = nullptr; | |||
| 2131 | if (!matchSimpleRecurrence(PN, BO, Start, Step)) | |||
| 2132 | return false; | |||
| 2133 | ||||
| 2134 | // Initial value must be a power of two. | |||
| 2135 | for (const Use &U : PN->operands()) { | |||
| 2136 | if (U.get() == Start) { | |||
| 2137 | // Initial value comes from a different BB, need to adjust context | |||
| 2138 | // instruction for analysis. | |||
| 2139 | Q.CxtI = PN->getIncomingBlock(U)->getTerminator(); | |||
| 2140 | if (!isKnownToBeAPowerOfTwo(Start, OrZero, Depth, Q)) | |||
| 2141 | return false; | |||
| 2142 | } | |||
| 2143 | } | |||
| 2144 | ||||
| 2145 | // Except for Mul, the induction variable must be on the left side of the | |||
| 2146 | // increment expression, otherwise its value can be arbitrary. | |||
| 2147 | if (BO->getOpcode() != Instruction::Mul && BO->getOperand(1) != Step) | |||
| 2148 | return false; | |||
| 2149 | ||||
| 2150 | Q.CxtI = BO->getParent()->getTerminator(); | |||
| 2151 | switch (BO->getOpcode()) { | |||
| 2152 | case Instruction::Mul: | |||
| 2153 | // Power of two is closed under multiplication. | |||
| 2154 | return (OrZero || Q.IIQ.hasNoUnsignedWrap(BO) || | |||
| 2155 | Q.IIQ.hasNoSignedWrap(BO)) && | |||
| 2156 | isKnownToBeAPowerOfTwo(Step, OrZero, Depth, Q); | |||
| 2157 | case Instruction::SDiv: | |||
| 2158 | // Start value must not be signmask for signed division, so simply being a | |||
| 2159 | // power of two is not sufficient, and it has to be a constant. | |||
| 2160 | if (!match(Start, m_Power2()) || match(Start, m_SignMask())) | |||
| 2161 | return false; | |||
| 2162 | [[fallthrough]]; | |||
| 2163 | case Instruction::UDiv: | |||
| 2164 | // Divisor must be a power of two. | |||
| 2165 | // If OrZero is false, cannot guarantee induction variable is non-zero after | |||
| 2166 | // division, same for Shr, unless it is exact division. | |||
| 2167 | return (OrZero || Q.IIQ.isExact(BO)) && | |||
| 2168 | isKnownToBeAPowerOfTwo(Step, false, Depth, Q); | |||
| 2169 | case Instruction::Shl: | |||
| 2170 | return OrZero || Q.IIQ.hasNoUnsignedWrap(BO) || Q.IIQ.hasNoSignedWrap(BO); | |||
| 2171 | case Instruction::AShr: | |||
| 2172 | if (!match(Start, m_Power2()) || match(Start, m_SignMask())) | |||
| 2173 | return false; | |||
| 2174 | [[fallthrough]]; | |||
| 2175 | case Instruction::LShr: | |||
| 2176 | return OrZero || Q.IIQ.isExact(BO); | |||
| 2177 | default: | |||
| 2178 | return false; | |||
| 2179 | } | |||
| 2180 | } | |||
| 2181 | ||||
| 2182 | /// Return true if the given value is known to have exactly one | |||
| 2183 | /// bit set when defined. For vectors return true if every element is known to | |||
| 2184 | /// be a power of two when defined. Supports values with integer or pointer | |||
| 2185 | /// types and vectors of integers. | |||
| 2186 | bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth, | |||
| 2187 | const Query &Q) { | |||
| 2188 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth")(static_cast <bool> (Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth") ? void (0) : __assert_fail ( "Depth <= MaxAnalysisRecursionDepth && \"Limit Search Depth\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2188, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2189 | ||||
| 2190 | // Attempt to match against constants. | |||
| 2191 | if (OrZero && match(V, m_Power2OrZero())) | |||
| 2192 | return true; | |||
| 2193 | if (match(V, m_Power2())) | |||
| 2194 | return true; | |||
| 2195 | ||||
| 2196 | // 1 << X is clearly a power of two if the one is not shifted off the end. If | |||
| 2197 | // it is shifted off the end then the result is undefined. | |||
| 2198 | if (match(V, m_Shl(m_One(), m_Value()))) | |||
| 2199 | return true; | |||
| 2200 | ||||
| 2201 | // (signmask) >>l X is clearly a power of two if the one is not shifted off | |||
| 2202 | // the bottom. If it is shifted off the bottom then the result is undefined. | |||
| 2203 | if (match(V, m_LShr(m_SignMask(), m_Value()))) | |||
| 2204 | return true; | |||
| 2205 | ||||
| 2206 | // The remaining tests are all recursive, so bail out if we hit the limit. | |||
| 2207 | if (Depth++ == MaxAnalysisRecursionDepth) | |||
| 2208 | return false; | |||
| 2209 | ||||
| 2210 | Value *X = nullptr, *Y = nullptr; | |||
| 2211 | // A shift left or a logical shift right of a power of two is a power of two | |||
| 2212 | // or zero. | |||
| 2213 | if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) || | |||
| 2214 | match(V, m_LShr(m_Value(X), m_Value())))) | |||
| 2215 | return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q); | |||
| 2216 | ||||
| 2217 | if (const ZExtInst *ZI = dyn_cast<ZExtInst>(V)) | |||
| 2218 | return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q); | |||
| 2219 | ||||
| 2220 | if (const SelectInst *SI = dyn_cast<SelectInst>(V)) | |||
| 2221 | return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) && | |||
| 2222 | isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q); | |||
| 2223 | ||||
| 2224 | // Peek through min/max. | |||
| 2225 | if (match(V, m_MaxOrMin(m_Value(X), m_Value(Y)))) { | |||
| 2226 | return isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q) && | |||
| 2227 | isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q); | |||
| 2228 | } | |||
| 2229 | ||||
| 2230 | if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) { | |||
| 2231 | // A power of two and'd with anything is a power of two or zero. | |||
| 2232 | if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) || | |||
| 2233 | isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q)) | |||
| 2234 | return true; | |||
| 2235 | // X & (-X) is always a power of two or zero. | |||
| 2236 | if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X)))) | |||
| 2237 | return true; | |||
| 2238 | return false; | |||
| 2239 | } | |||
| 2240 | ||||
| 2241 | // Adding a power-of-two or zero to the same power-of-two or zero yields | |||
| 2242 | // either the original power-of-two, a larger power-of-two or zero. | |||
| 2243 | if (match(V, m_Add(m_Value(X), m_Value(Y)))) { | |||
| 2244 | const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V); | |||
| 2245 | if (OrZero || Q.IIQ.hasNoUnsignedWrap(VOBO) || | |||
| 2246 | Q.IIQ.hasNoSignedWrap(VOBO)) { | |||
| 2247 | if (match(X, m_And(m_Specific(Y), m_Value())) || | |||
| 2248 | match(X, m_And(m_Value(), m_Specific(Y)))) | |||
| 2249 | if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q)) | |||
| 2250 | return true; | |||
| 2251 | if (match(Y, m_And(m_Specific(X), m_Value())) || | |||
| 2252 | match(Y, m_And(m_Value(), m_Specific(X)))) | |||
| 2253 | if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q)) | |||
| 2254 | return true; | |||
| 2255 | ||||
| 2256 | unsigned BitWidth = V->getType()->getScalarSizeInBits(); | |||
| 2257 | KnownBits LHSBits(BitWidth); | |||
| 2258 | computeKnownBits(X, LHSBits, Depth, Q); | |||
| 2259 | ||||
| 2260 | KnownBits RHSBits(BitWidth); | |||
| 2261 | computeKnownBits(Y, RHSBits, Depth, Q); | |||
| 2262 | // If i8 V is a power of two or zero: | |||
| 2263 | // ZeroBits: 1 1 1 0 1 1 1 1 | |||
| 2264 | // ~ZeroBits: 0 0 0 1 0 0 0 0 | |||
| 2265 | if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2()) | |||
| 2266 | // If OrZero isn't set, we cannot give back a zero result. | |||
| 2267 | // Make sure either the LHS or RHS has a bit set. | |||
| 2268 | if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue()) | |||
| 2269 | return true; | |||
| 2270 | } | |||
| 2271 | } | |||
| 2272 | ||||
| 2273 | // A PHI node is power of two if all incoming values are power of two, or if | |||
| 2274 | // it is an induction variable where in each step its value is a power of two. | |||
| 2275 | if (const PHINode *PN = dyn_cast<PHINode>(V)) { | |||
| 2276 | Query RecQ = Q; | |||
| 2277 | ||||
| 2278 | // Check if it is an induction variable and always power of two. | |||
| 2279 | if (isPowerOfTwoRecurrence(PN, OrZero, Depth, RecQ)) | |||
| 2280 | return true; | |||
| 2281 | ||||
| 2282 | // Recursively check all incoming values. Limit recursion to 2 levels, so | |||
| 2283 | // that search complexity is limited to number of operands^2. | |||
| 2284 | unsigned NewDepth = std::max(Depth, MaxAnalysisRecursionDepth - 1); | |||
| 2285 | return llvm::all_of(PN->operands(), [&](const Use &U) { | |||
| 2286 | // Value is power of 2 if it is coming from PHI node itself by induction. | |||
| 2287 | if (U.get() == PN) | |||
| 2288 | return true; | |||
| 2289 | ||||
| 2290 | // Change the context instruction to the incoming block where it is | |||
| 2291 | // evaluated. | |||
| 2292 | RecQ.CxtI = PN->getIncomingBlock(U)->getTerminator(); | |||
| 2293 | return isKnownToBeAPowerOfTwo(U.get(), OrZero, NewDepth, RecQ); | |||
| 2294 | }); | |||
| 2295 | } | |||
| 2296 | ||||
| 2297 | // An exact divide or right shift can only shift off zero bits, so the result | |||
| 2298 | // is a power of two only if the first operand is a power of two and not | |||
| 2299 | // copying a sign bit (sdiv int_min, 2). | |||
| 2300 | if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) || | |||
| 2301 | match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) { | |||
| 2302 | return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero, | |||
| 2303 | Depth, Q); | |||
| 2304 | } | |||
| 2305 | ||||
| 2306 | return false; | |||
| 2307 | } | |||
| 2308 | ||||
| 2309 | /// Test whether a GEP's result is known to be non-null. | |||
| 2310 | /// | |||
| 2311 | /// Uses properties inherent in a GEP to try to determine whether it is known | |||
| 2312 | /// to be non-null. | |||
| 2313 | /// | |||
| 2314 | /// Currently this routine does not support vector GEPs. | |||
| 2315 | static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth, | |||
| 2316 | const Query &Q) { | |||
| 2317 | const Function *F = nullptr; | |||
| 2318 | if (const Instruction *I = dyn_cast<Instruction>(GEP)) | |||
| 2319 | F = I->getFunction(); | |||
| 2320 | ||||
| 2321 | if (!GEP->isInBounds() || | |||
| 2322 | NullPointerIsDefined(F, GEP->getPointerAddressSpace())) | |||
| 2323 | return false; | |||
| 2324 | ||||
| 2325 | // FIXME: Support vector-GEPs. | |||
| 2326 | assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP")(static_cast <bool> (GEP->getType()->isPointerTy( ) && "We only support plain pointer GEP") ? void (0) : __assert_fail ("GEP->getType()->isPointerTy() && \"We only support plain pointer GEP\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2326, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2327 | ||||
| 2328 | // If the base pointer is non-null, we cannot walk to a null address with an | |||
| 2329 | // inbounds GEP in address space zero. | |||
| 2330 | if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q)) | |||
| 2331 | return true; | |||
| 2332 | ||||
| 2333 | // Walk the GEP operands and see if any operand introduces a non-zero offset. | |||
| 2334 | // If so, then the GEP cannot produce a null pointer, as doing so would | |||
| 2335 | // inherently violate the inbounds contract within address space zero. | |||
| 2336 | for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP); | |||
| 2337 | GTI != GTE; ++GTI) { | |||
| 2338 | // Struct types are easy -- they must always be indexed by a constant. | |||
| 2339 | if (StructType *STy = GTI.getStructTypeOrNull()) { | |||
| 2340 | ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand()); | |||
| 2341 | unsigned ElementIdx = OpC->getZExtValue(); | |||
| 2342 | const StructLayout *SL = Q.DL.getStructLayout(STy); | |||
| 2343 | uint64_t ElementOffset = SL->getElementOffset(ElementIdx); | |||
| 2344 | if (ElementOffset > 0) | |||
| 2345 | return true; | |||
| 2346 | continue; | |||
| 2347 | } | |||
| 2348 | ||||
| 2349 | // If we have a zero-sized type, the index doesn't matter. Keep looping. | |||
| 2350 | if (Q.DL.getTypeAllocSize(GTI.getIndexedType()).isZero()) | |||
| 2351 | continue; | |||
| 2352 | ||||
| 2353 | // Fast path the constant operand case both for efficiency and so we don't | |||
| 2354 | // increment Depth when just zipping down an all-constant GEP. | |||
| 2355 | if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) { | |||
| 2356 | if (!OpC->isZero()) | |||
| 2357 | return true; | |||
| 2358 | continue; | |||
| 2359 | } | |||
| 2360 | ||||
| 2361 | // We post-increment Depth here because while isKnownNonZero increments it | |||
| 2362 | // as well, when we pop back up that increment won't persist. We don't want | |||
| 2363 | // to recurse 10k times just because we have 10k GEP operands. We don't | |||
| 2364 | // bail completely out because we want to handle constant GEPs regardless | |||
| 2365 | // of depth. | |||
| 2366 | if (Depth++ >= MaxAnalysisRecursionDepth) | |||
| 2367 | continue; | |||
| 2368 | ||||
| 2369 | if (isKnownNonZero(GTI.getOperand(), Depth, Q)) | |||
| 2370 | return true; | |||
| 2371 | } | |||
| 2372 | ||||
| 2373 | return false; | |||
| 2374 | } | |||
| 2375 | ||||
| 2376 | static bool isKnownNonNullFromDominatingCondition(const Value *V, | |||
| 2377 | const Instruction *CtxI, | |||
| 2378 | const DominatorTree *DT) { | |||
| 2379 | assert(!isa<Constant>(V) && "Called for constant?")(static_cast <bool> (!isa<Constant>(V) && "Called for constant?") ? void (0) : __assert_fail ("!isa<Constant>(V) && \"Called for constant?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2379, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2380 | ||||
| 2381 | if (!CtxI || !DT) | |||
| 2382 | return false; | |||
| 2383 | ||||
| 2384 | unsigned NumUsesExplored = 0; | |||
| 2385 | for (const auto *U : V->users()) { | |||
| 2386 | // Avoid massive lists | |||
| 2387 | if (NumUsesExplored >= DomConditionsMaxUses) | |||
| 2388 | break; | |||
| 2389 | NumUsesExplored++; | |||
| 2390 | ||||
| 2391 | // If the value is used as an argument to a call or invoke, then argument | |||
| 2392 | // attributes may provide an answer about null-ness. | |||
| 2393 | if (const auto *CB = dyn_cast<CallBase>(U)) | |||
| 2394 | if (auto *CalledFunc = CB->getCalledFunction()) | |||
| 2395 | for (const Argument &Arg : CalledFunc->args()) | |||
| 2396 | if (CB->getArgOperand(Arg.getArgNo()) == V && | |||
| 2397 | Arg.hasNonNullAttr(/* AllowUndefOrPoison */ false) && | |||
| 2398 | DT->dominates(CB, CtxI)) | |||
| 2399 | return true; | |||
| 2400 | ||||
| 2401 | // If the value is used as a load/store, then the pointer must be non null. | |||
| 2402 | if (V == getLoadStorePointerOperand(U)) { | |||
| 2403 | const Instruction *I = cast<Instruction>(U); | |||
| 2404 | if (!NullPointerIsDefined(I->getFunction(), | |||
| 2405 | V->getType()->getPointerAddressSpace()) && | |||
| 2406 | DT->dominates(I, CtxI)) | |||
| 2407 | return true; | |||
| 2408 | } | |||
| 2409 | ||||
| 2410 | // Consider only compare instructions uniquely controlling a branch | |||
| 2411 | Value *RHS; | |||
| 2412 | CmpInst::Predicate Pred; | |||
| 2413 | if (!match(U, m_c_ICmp(Pred, m_Specific(V), m_Value(RHS)))) | |||
| 2414 | continue; | |||
| 2415 | ||||
| 2416 | bool NonNullIfTrue; | |||
| 2417 | if (cmpExcludesZero(Pred, RHS)) | |||
| 2418 | NonNullIfTrue = true; | |||
| 2419 | else if (cmpExcludesZero(CmpInst::getInversePredicate(Pred), RHS)) | |||
| 2420 | NonNullIfTrue = false; | |||
| 2421 | else | |||
| 2422 | continue; | |||
| 2423 | ||||
| 2424 | SmallVector<const User *, 4> WorkList; | |||
| 2425 | SmallPtrSet<const User *, 4> Visited; | |||
| 2426 | for (const auto *CmpU : U->users()) { | |||
| 2427 | assert(WorkList.empty() && "Should be!")(static_cast <bool> (WorkList.empty() && "Should be!" ) ? void (0) : __assert_fail ("WorkList.empty() && \"Should be!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2427, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2428 | if (Visited.insert(CmpU).second) | |||
| 2429 | WorkList.push_back(CmpU); | |||
| 2430 | ||||
| 2431 | while (!WorkList.empty()) { | |||
| 2432 | auto *Curr = WorkList.pop_back_val(); | |||
| 2433 | ||||
| 2434 | // If a user is an AND, add all its users to the work list. We only | |||
| 2435 | // propagate "pred != null" condition through AND because it is only | |||
| 2436 | // correct to assume that all conditions of AND are met in true branch. | |||
| 2437 | // TODO: Support similar logic of OR and EQ predicate? | |||
| 2438 | if (NonNullIfTrue) | |||
| 2439 | if (match(Curr, m_LogicalAnd(m_Value(), m_Value()))) { | |||
| 2440 | for (const auto *CurrU : Curr->users()) | |||
| 2441 | if (Visited.insert(CurrU).second) | |||
| 2442 | WorkList.push_back(CurrU); | |||
| 2443 | continue; | |||
| 2444 | } | |||
| 2445 | ||||
| 2446 | if (const BranchInst *BI = dyn_cast<BranchInst>(Curr)) { | |||
| 2447 | assert(BI->isConditional() && "uses a comparison!")(static_cast <bool> (BI->isConditional() && "uses a comparison!" ) ? void (0) : __assert_fail ("BI->isConditional() && \"uses a comparison!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2447, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2448 | ||||
| 2449 | BasicBlock *NonNullSuccessor = | |||
| 2450 | BI->getSuccessor(NonNullIfTrue ? 0 : 1); | |||
| 2451 | BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor); | |||
| 2452 | if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent())) | |||
| 2453 | return true; | |||
| 2454 | } else if (NonNullIfTrue && isGuard(Curr) && | |||
| 2455 | DT->dominates(cast<Instruction>(Curr), CtxI)) { | |||
| 2456 | return true; | |||
| 2457 | } | |||
| 2458 | } | |||
| 2459 | } | |||
| 2460 | } | |||
| 2461 | ||||
| 2462 | return false; | |||
| 2463 | } | |||
| 2464 | ||||
| 2465 | /// Does the 'Range' metadata (which must be a valid MD_range operand list) | |||
| 2466 | /// ensure that the value it's attached to is never Value? 'RangeType' is | |||
| 2467 | /// is the type of the value described by the range. | |||
| 2468 | static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) { | |||
| 2469 | const unsigned NumRanges = Ranges->getNumOperands() / 2; | |||
| 2470 | assert(NumRanges >= 1)(static_cast <bool> (NumRanges >= 1) ? void (0) : __assert_fail ("NumRanges >= 1", "llvm/lib/Analysis/ValueTracking.cpp", 2470, __extension__ __PRETTY_FUNCTION__)); | |||
| 2471 | for (unsigned i = 0; i < NumRanges; ++i) { | |||
| 2472 | ConstantInt *Lower = | |||
| 2473 | mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0)); | |||
| 2474 | ConstantInt *Upper = | |||
| 2475 | mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1)); | |||
| 2476 | ConstantRange Range(Lower->getValue(), Upper->getValue()); | |||
| 2477 | if (Range.contains(Value)) | |||
| 2478 | return false; | |||
| 2479 | } | |||
| 2480 | return true; | |||
| 2481 | } | |||
| 2482 | ||||
| 2483 | /// Try to detect a recurrence that monotonically increases/decreases from a | |||
| 2484 | /// non-zero starting value. These are common as induction variables. | |||
| 2485 | static bool isNonZeroRecurrence(const PHINode *PN) { | |||
| 2486 | BinaryOperator *BO = nullptr; | |||
| 2487 | Value *Start = nullptr, *Step = nullptr; | |||
| 2488 | const APInt *StartC, *StepC; | |||
| 2489 | if (!matchSimpleRecurrence(PN, BO, Start, Step) || | |||
| 2490 | !match(Start, m_APInt(StartC)) || StartC->isZero()) | |||
| 2491 | return false; | |||
| 2492 | ||||
| 2493 | switch (BO->getOpcode()) { | |||
| 2494 | case Instruction::Add: | |||
| 2495 | // Starting from non-zero and stepping away from zero can never wrap back | |||
| 2496 | // to zero. | |||
| 2497 | return BO->hasNoUnsignedWrap() || | |||
| 2498 | (BO->hasNoSignedWrap() && match(Step, m_APInt(StepC)) && | |||
| 2499 | StartC->isNegative() == StepC->isNegative()); | |||
| 2500 | case Instruction::Mul: | |||
| 2501 | return (BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap()) && | |||
| 2502 | match(Step, m_APInt(StepC)) && !StepC->isZero(); | |||
| 2503 | case Instruction::Shl: | |||
| 2504 | return BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap(); | |||
| 2505 | case Instruction::AShr: | |||
| 2506 | case Instruction::LShr: | |||
| 2507 | return BO->isExact(); | |||
| 2508 | default: | |||
| 2509 | return false; | |||
| 2510 | } | |||
| 2511 | } | |||
| 2512 | ||||
| 2513 | static bool isNonZeroAdd(const APInt &DemandedElts, unsigned Depth, | |||
| 2514 | const Query &Q, unsigned BitWidth, Value *X, Value *Y, | |||
| 2515 | bool NSW) { | |||
| 2516 | KnownBits XKnown = computeKnownBits(X, DemandedElts, Depth, Q); | |||
| 2517 | KnownBits YKnown = computeKnownBits(Y, DemandedElts, Depth, Q); | |||
| 2518 | ||||
| 2519 | // If X and Y are both non-negative (as signed values) then their sum is not | |||
| 2520 | // zero unless both X and Y are zero. | |||
| 2521 | if (XKnown.isNonNegative() && YKnown.isNonNegative()) | |||
| 2522 | if (isKnownNonZero(Y, DemandedElts, Depth, Q) || | |||
| 2523 | isKnownNonZero(X, DemandedElts, Depth, Q)) | |||
| 2524 | return true; | |||
| 2525 | ||||
| 2526 | // If X and Y are both negative (as signed values) then their sum is not | |||
| 2527 | // zero unless both X and Y equal INT_MIN. | |||
| 2528 | if (XKnown.isNegative() && YKnown.isNegative()) { | |||
| 2529 | APInt Mask = APInt::getSignedMaxValue(BitWidth); | |||
| 2530 | // The sign bit of X is set. If some other bit is set then X is not equal | |||
| 2531 | // to INT_MIN. | |||
| 2532 | if (XKnown.One.intersects(Mask)) | |||
| 2533 | return true; | |||
| 2534 | // The sign bit of Y is set. If some other bit is set then Y is not equal | |||
| 2535 | // to INT_MIN. | |||
| 2536 | if (YKnown.One.intersects(Mask)) | |||
| 2537 | return true; | |||
| 2538 | } | |||
| 2539 | ||||
| 2540 | // The sum of a non-negative number and a power of two is not zero. | |||
| 2541 | if (XKnown.isNonNegative() && | |||
| 2542 | isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q)) | |||
| 2543 | return true; | |||
| 2544 | if (YKnown.isNonNegative() && | |||
| 2545 | isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q)) | |||
| 2546 | return true; | |||
| 2547 | ||||
| 2548 | return KnownBits::computeForAddSub(/*Add*/ true, NSW, XKnown, YKnown) | |||
| 2549 | .isNonZero(); | |||
| 2550 | } | |||
| 2551 | ||||
| 2552 | static bool isNonZeroSub(const APInt &DemandedElts, unsigned Depth, | |||
| 2553 | const Query &Q, unsigned BitWidth, Value *X, | |||
| 2554 | Value *Y) { | |||
| 2555 | if (auto *C = dyn_cast<Constant>(X)) | |||
| 2556 | if (C->isNullValue() && isKnownNonZero(Y, DemandedElts, Depth, Q)) | |||
| 2557 | return true; | |||
| 2558 | ||||
| 2559 | KnownBits XKnown = computeKnownBits(X, DemandedElts, Depth, Q); | |||
| 2560 | if (XKnown.isUnknown()) | |||
| 2561 | return false; | |||
| 2562 | KnownBits YKnown = computeKnownBits(Y, DemandedElts, Depth, Q); | |||
| 2563 | // If X != Y then X - Y is non zero. | |||
| 2564 | std::optional<bool> ne = KnownBits::ne(XKnown, YKnown); | |||
| 2565 | // If we are unable to compute if X != Y, we won't be able to do anything | |||
| 2566 | // computing the knownbits of the sub expression so just return here. | |||
| 2567 | return ne && *ne; | |||
| 2568 | } | |||
| 2569 | ||||
| 2570 | static bool isNonZeroShift(const Operator *I, const APInt &DemandedElts, | |||
| 2571 | unsigned Depth, const Query &Q, | |||
| 2572 | const KnownBits &KnownVal) { | |||
| 2573 | auto ShiftOp = [&](const APInt &Lhs, const APInt &Rhs) { | |||
| 2574 | switch (I->getOpcode()) { | |||
| 2575 | case Instruction::Shl: | |||
| 2576 | return Lhs.shl(Rhs); | |||
| 2577 | case Instruction::LShr: | |||
| 2578 | return Lhs.lshr(Rhs); | |||
| 2579 | case Instruction::AShr: | |||
| 2580 | return Lhs.ashr(Rhs); | |||
| 2581 | default: | |||
| 2582 | llvm_unreachable("Unknown Shift Opcode")::llvm::llvm_unreachable_internal("Unknown Shift Opcode", "llvm/lib/Analysis/ValueTracking.cpp" , 2582); | |||
| 2583 | } | |||
| 2584 | }; | |||
| 2585 | ||||
| 2586 | auto InvShiftOp = [&](const APInt &Lhs, const APInt &Rhs) { | |||
| 2587 | switch (I->getOpcode()) { | |||
| 2588 | case Instruction::Shl: | |||
| 2589 | return Lhs.lshr(Rhs); | |||
| 2590 | case Instruction::LShr: | |||
| 2591 | case Instruction::AShr: | |||
| 2592 | return Lhs.shl(Rhs); | |||
| 2593 | default: | |||
| 2594 | llvm_unreachable("Unknown Shift Opcode")::llvm::llvm_unreachable_internal("Unknown Shift Opcode", "llvm/lib/Analysis/ValueTracking.cpp" , 2594); | |||
| 2595 | } | |||
| 2596 | }; | |||
| 2597 | ||||
| 2598 | if (KnownVal.isUnknown()) | |||
| 2599 | return false; | |||
| 2600 | ||||
| 2601 | KnownBits KnownCnt = | |||
| 2602 | computeKnownBits(I->getOperand(1), DemandedElts, Depth, Q); | |||
| 2603 | APInt MaxShift = KnownCnt.getMaxValue(); | |||
| 2604 | unsigned NumBits = KnownVal.getBitWidth(); | |||
| 2605 | if (MaxShift.uge(NumBits)) | |||
| 2606 | return false; | |||
| 2607 | ||||
| 2608 | if (!ShiftOp(KnownVal.One, MaxShift).isZero()) | |||
| 2609 | return true; | |||
| 2610 | ||||
| 2611 | // If all of the bits shifted out are known to be zero, and Val is known | |||
| 2612 | // non-zero then at least one non-zero bit must remain. | |||
| 2613 | if (InvShiftOp(KnownVal.Zero, NumBits - MaxShift) | |||
| 2614 | .eq(InvShiftOp(APInt::getAllOnes(NumBits), NumBits - MaxShift)) && | |||
| 2615 | isKnownNonZero(I->getOperand(0), DemandedElts, Depth, Q)) | |||
| 2616 | return true; | |||
| 2617 | ||||
| 2618 | return false; | |||
| 2619 | } | |||
| 2620 | ||||
| 2621 | /// Return true if the given value is known to be non-zero when defined. For | |||
| 2622 | /// vectors, return true if every demanded element is known to be non-zero when | |||
| 2623 | /// defined. For pointers, if the context instruction and dominator tree are | |||
| 2624 | /// specified, perform context-sensitive analysis and return true if the | |||
| 2625 | /// pointer couldn't possibly be null at the specified instruction. | |||
| 2626 | /// Supports values with integer or pointer type and vectors of integers. | |||
| 2627 | bool isKnownNonZero(const Value *V, const APInt &DemandedElts, unsigned Depth, | |||
| 2628 | const Query &Q) { | |||
| 2629 | ||||
| 2630 | #ifndef NDEBUG | |||
| 2631 | Type *Ty = V->getType(); | |||
| 2632 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth")(static_cast <bool> (Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth") ? void (0) : __assert_fail ( "Depth <= MaxAnalysisRecursionDepth && \"Limit Search Depth\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2632, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2633 | ||||
| 2634 | if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) { | |||
| 2635 | assert((static_cast <bool> (FVTy->getNumElements() == DemandedElts .getBitWidth() && "DemandedElt width should equal the fixed vector number of elements" ) ? void (0) : __assert_fail ("FVTy->getNumElements() == DemandedElts.getBitWidth() && \"DemandedElt width should equal the fixed vector number of elements\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2637, __extension__ __PRETTY_FUNCTION__ )) | |||
| 2636 | FVTy->getNumElements() == DemandedElts.getBitWidth() &&(static_cast <bool> (FVTy->getNumElements() == DemandedElts .getBitWidth() && "DemandedElt width should equal the fixed vector number of elements" ) ? void (0) : __assert_fail ("FVTy->getNumElements() == DemandedElts.getBitWidth() && \"DemandedElt width should equal the fixed vector number of elements\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2637, __extension__ __PRETTY_FUNCTION__ )) | |||
| 2637 | "DemandedElt width should equal the fixed vector number of elements")(static_cast <bool> (FVTy->getNumElements() == DemandedElts .getBitWidth() && "DemandedElt width should equal the fixed vector number of elements" ) ? void (0) : __assert_fail ("FVTy->getNumElements() == DemandedElts.getBitWidth() && \"DemandedElt width should equal the fixed vector number of elements\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2637, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2638 | } else { | |||
| 2639 | assert(DemandedElts == APInt(1, 1) &&(static_cast <bool> (DemandedElts == APInt(1, 1) && "DemandedElt width should be 1 for scalars") ? void (0) : __assert_fail ("DemandedElts == APInt(1, 1) && \"DemandedElt width should be 1 for scalars\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2640, __extension__ __PRETTY_FUNCTION__ )) | |||
| 2640 | "DemandedElt width should be 1 for scalars")(static_cast <bool> (DemandedElts == APInt(1, 1) && "DemandedElt width should be 1 for scalars") ? void (0) : __assert_fail ("DemandedElts == APInt(1, 1) && \"DemandedElt width should be 1 for scalars\"" , "llvm/lib/Analysis/ValueTracking.cpp", 2640, __extension__ __PRETTY_FUNCTION__ )); | |||
| 2641 | } | |||
| 2642 | #endif | |||
| 2643 | ||||
| 2644 | if (auto *C = dyn_cast<Constant>(V)) { | |||
| 2645 | if (C->isNullValue()) | |||
| 2646 | return false; | |||
| 2647 | if (isa<ConstantInt>(C)) | |||
| 2648 | // Must be non-zero due to null test above. | |||
| 2649 | return true; | |||
| 2650 | ||||
| 2651 | // For constant vectors, check that all elements are undefined or known | |||
| 2652 | // non-zero to determine that the whole vector is known non-zero. | |||
| 2653 | if (auto *VecTy = dyn_cast<FixedVectorType>(C->getType())) { | |||
| 2654 | for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) { | |||
| 2655 | if (!DemandedElts[i]) | |||
| 2656 | continue; | |||
| 2657 | Constant *Elt = C->getAggregateElement(i); | |||
| 2658 | if (!Elt || Elt->isNullValue()) | |||
| 2659 | return false; | |||
| 2660 | if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt)) | |||
| 2661 | return false; | |||
| 2662 | } | |||
| 2663 | return true; | |||
| 2664 | } | |||
| 2665 | ||||
| 2666 | // A global variable in address space 0 is non null unless extern weak | |||
| 2667 | // or an absolute symbol reference. Other address spaces may have null as a | |||
| 2668 | // valid address for a global, so we can't assume anything. | |||
| 2669 | if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) { | |||
| 2670 | if (!GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() && | |||
| 2671 | GV->getType()->getAddressSpace() == 0) | |||
| 2672 | return true; | |||
| 2673 | } | |||
| 2674 | ||||
| 2675 | // For constant expressions, fall through to the Operator code below. | |||
| 2676 | if (!isa<ConstantExpr>(V)) | |||
| 2677 | return false; | |||
| 2678 | } | |||
| 2679 | ||||
| 2680 | if (auto *I = dyn_cast<Instruction>(V)) { | |||
| 2681 | if (MDNode *Ranges = Q.IIQ.getMetadata(I, LLVMContext::MD_range)) { | |||
| 2682 | // If the possible ranges don't contain zero, then the value is | |||
| 2683 | // definitely non-zero. | |||
| 2684 | if (auto *Ty = dyn_cast<IntegerType>(V->getType())) { | |||
| 2685 | const APInt ZeroValue(Ty->getBitWidth(), 0); | |||
| 2686 | if (rangeMetadataExcludesValue(Ranges, ZeroValue)) | |||
| 2687 | return true; | |||
| 2688 | } | |||
| 2689 | } | |||
| 2690 | } | |||
| 2691 | ||||
| 2692 | if (!isa<Constant>(V) && isKnownNonZeroFromAssume(V, Q)) | |||
| 2693 | return true; | |||
| 2694 | ||||
| 2695 | // Some of the tests below are recursive, so bail out if we hit the limit. | |||
| 2696 | if (Depth++ >= MaxAnalysisRecursionDepth) | |||
| 2697 | return false; | |||
| 2698 | ||||
| 2699 | // Check for pointer simplifications. | |||
| 2700 | ||||
| 2701 | if (PointerType *PtrTy = dyn_cast<PointerType>(V->getType())) { | |||
| 2702 | // Alloca never returns null, malloc might. | |||
| 2703 | if (isa<AllocaInst>(V) && Q.DL.getAllocaAddrSpace() == 0) | |||
| 2704 | return true; | |||
| 2705 | ||||
| 2706 | // A byval, inalloca may not be null in a non-default addres space. A | |||
| 2707 | // nonnull argument is assumed never 0. | |||
| 2708 | if (const Argument *A = dyn_cast<Argument>(V)) { | |||
| 2709 | if (((A->hasPassPointeeByValueCopyAttr() && | |||
| 2710 | !NullPointerIsDefined(A->getParent(), PtrTy->getAddressSpace())) || | |||
| 2711 | A->hasNonNullAttr())) | |||
| 2712 | return true; | |||
| 2713 | } | |||
| 2714 | ||||
| 2715 | // A Load tagged with nonnull metadata is never null. | |||
| 2716 | if (const LoadInst *LI = dyn_cast<LoadInst>(V)) | |||
| 2717 | if (Q.IIQ.getMetadata(LI, LLVMContext::MD_nonnull)) | |||
| 2718 | return true; | |||
| 2719 | ||||
| 2720 | if (const auto *Call = dyn_cast<CallBase>(V)) { | |||
| 2721 | if (Call->isReturnNonNull()) | |||
| 2722 | return true; | |||
| 2723 | if (const auto *RP = getArgumentAliasingToReturnedPointer(Call, true)) | |||
| 2724 | return isKnownNonZero(RP, Depth, Q); | |||
| 2725 | } | |||
| 2726 | } | |||
| 2727 | ||||
| 2728 | if (!isa<Constant>(V) && | |||
| 2729 | isKnownNonNullFromDominatingCondition(V, Q.CxtI, Q.DT)) | |||
| 2730 | return true; | |||
| 2731 | ||||
| 2732 | const Operator *I = dyn_cast<Operator>(V); | |||
| 2733 | if (!I) | |||
| 2734 | return false; | |||
| 2735 | ||||
| 2736 | unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL); | |||
| 2737 | switch (I->getOpcode()) { | |||
| 2738 | case Instruction::GetElementPtr: | |||
| 2739 | if (I->getType()->isPointerTy()) | |||
| 2740 | return isGEPKnownNonNull(cast<GEPOperator>(I), Depth, Q); | |||
| 2741 | break; | |||
| 2742 | case Instruction::BitCast: { | |||
| 2743 | // We need to be a bit careful here. We can only peek through the bitcast | |||
| 2744 | // if the scalar size of elements in the operand are smaller than and a | |||
| 2745 | // multiple of the size they are casting too. Take three cases: | |||
| 2746 | // | |||
| 2747 | // 1) Unsafe: | |||
| 2748 | // bitcast <2 x i16> %NonZero to <4 x i8> | |||
| 2749 | // | |||
| 2750 | // %NonZero can have 2 non-zero i16 elements, but isKnownNonZero on a | |||
| 2751 | // <4 x i8> requires that all 4 i8 elements be non-zero which isn't | |||
| 2752 | // guranteed (imagine just sign bit set in the 2 i16 elements). | |||
| 2753 | // | |||
| 2754 | // 2) Unsafe: | |||
| 2755 | // bitcast <4 x i3> %NonZero to <3 x i4> | |||
| 2756 | // | |||
| 2757 | // Even though the scalar size of the src (`i3`) is smaller than the | |||
| 2758 | // scalar size of the dst `i4`, because `i3` is not a multiple of `i4` | |||
| 2759 | // its possible for the `3 x i4` elements to be zero because there are | |||
| 2760 | // some elements in the destination that don't contain any full src | |||
| 2761 | // element. | |||
| 2762 | // | |||
| 2763 | // 3) Safe: | |||
| 2764 | // bitcast <4 x i8> %NonZero to <2 x i16> | |||
| 2765 | // | |||
| 2766 | // This is always safe as non-zero in the 4 i8 elements implies | |||
| 2767 | // non-zero in the combination of any two adjacent ones. Since i8 is a | |||
| 2768 | // multiple of i16, each i16 is guranteed to have 2 full i8 elements. | |||
| 2769 | // This all implies the 2 i16 elements are non-zero. | |||
| 2770 | Type *FromTy = I->getOperand(0)->getType(); | |||
| 2771 | if ((FromTy->isIntOrIntVectorTy() || FromTy->isPtrOrPtrVectorTy()) && | |||
| 2772 | (BitWidth % getBitWidth(FromTy->getScalarType(), Q.DL)) == 0) | |||
| 2773 | return isKnownNonZero(I->getOperand(0), Depth, Q); | |||
| 2774 | } break; | |||
| 2775 | case Instruction::IntToPtr: | |||
| 2776 | // Note that we have to take special care to avoid looking through | |||
| 2777 | // truncating casts, e.g., int2ptr/ptr2int with appropriate sizes, as well | |||
| 2778 | // as casts that can alter the value, e.g., AddrSpaceCasts. | |||
| 2779 | if (!isa<ScalableVectorType>(I->getType()) && | |||
| 2780 | Q.DL.getTypeSizeInBits(I->getOperand(0)->getType()).getFixedValue() <= | |||
| 2781 | Q.DL.getTypeSizeInBits(I->getType()).getFixedValue()) | |||
| 2782 | return isKnownNonZero(I->getOperand(0), Depth, Q); | |||
| 2783 | break; | |||
| 2784 | case Instruction::PtrToInt: | |||
| 2785 | // Similar to int2ptr above, we can look through ptr2int here if the cast | |||
| 2786 | // is a no-op or an extend and not a truncate. | |||
| 2787 | if (!isa<ScalableVectorType>(I->getType()) && | |||
| 2788 | Q.DL.getTypeSizeInBits(I->getOperand(0)->getType()).getFixedValue() <= | |||
| 2789 | Q.DL.getTypeSizeInBits(I->getType()).getFixedValue()) | |||
| 2790 | return isKnownNonZero(I->getOperand(0), Depth, Q); | |||
| 2791 | break; | |||
| 2792 | case Instruction::Sub: | |||
| 2793 | return isNonZeroSub(DemandedElts, Depth, Q, BitWidth, I->getOperand(0), | |||
| 2794 | I->getOperand(1)); | |||
| 2795 | case Instruction::Or: | |||
| 2796 | // X | Y != 0 if X != 0 or Y != 0. | |||
| 2797 | return isKnownNonZero(I->getOperand(1), DemandedElts, Depth, Q) || | |||
| 2798 | isKnownNonZero(I->getOperand(0), DemandedElts, Depth, Q); | |||
| 2799 | case Instruction::SExt: | |||
| 2800 | case Instruction::ZExt: | |||
| 2801 | // ext X != 0 if X != 0. | |||
| 2802 | return isKnownNonZero(I->getOperand(0), Depth, Q); | |||
| 2803 | ||||
| 2804 | case Instruction::Shl: { | |||
| 2805 | // shl nsw/nuw can't remove any non-zero bits. | |||
| 2806 | const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V); | |||
| 2807 | if (Q.IIQ.hasNoUnsignedWrap(BO) || Q.IIQ.hasNoSignedWrap(BO)) | |||
| 2808 | return isKnownNonZero(I->getOperand(0), Depth, Q); | |||
| 2809 | ||||
| 2810 | // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined | |||
| 2811 | // if the lowest bit is shifted off the end. | |||
| 2812 | KnownBits Known(BitWidth); | |||
| 2813 | computeKnownBits(I->getOperand(0), DemandedElts, Known, Depth, Q); | |||
| 2814 | if (Known.One[0]) | |||
| 2815 | return true; | |||
| 2816 | ||||
| 2817 | return isNonZeroShift(I, DemandedElts, Depth, Q, Known); | |||
| 2818 | } | |||
| 2819 | case Instruction::LShr: | |||
| 2820 | case Instruction::AShr: { | |||
| 2821 | // shr exact can only shift out zero bits. | |||
| 2822 | const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V); | |||
| 2823 | if (BO->isExact()) | |||
| 2824 | return isKnownNonZero(I->getOperand(0), Depth, Q); | |||
| 2825 | ||||
| 2826 | // shr X, Y != 0 if X is negative. Note that the value of the shift is not | |||
| 2827 | // defined if the sign bit is shifted off the end. | |||
| 2828 | KnownBits Known = | |||
| 2829 | computeKnownBits(I->getOperand(0), DemandedElts, Depth, Q); | |||
| 2830 | if (Known.isNegative()) | |||
| 2831 | return true; | |||
| 2832 | ||||
| 2833 | return isNonZeroShift(I, DemandedElts, Depth, Q, Known); | |||
| 2834 | } | |||
| 2835 | case Instruction::UDiv: | |||
| 2836 | case Instruction::SDiv: | |||
| 2837 | // X / Y | |||
| 2838 | // div exact can only produce a zero if the dividend is zero. | |||
| 2839 | if (cast<PossiblyExactOperator>(I)->isExact()) | |||
| 2840 | return isKnownNonZero(I->getOperand(0), DemandedElts, Depth, Q); | |||
| 2841 | if (I->getOpcode() == Instruction::UDiv) { | |||
| 2842 | std::optional<bool> XUgeY; | |||
| 2843 | KnownBits XKnown = | |||
| 2844 | computeKnownBits(I->getOperand(0), DemandedElts, Depth, Q); | |||
| 2845 | if (!XKnown.isUnknown()) { | |||
| 2846 | KnownBits YKnown = | |||
| 2847 | computeKnownBits(I->getOperand(1), DemandedElts, Depth, Q); | |||
| 2848 | // If X u>= Y then div is non zero (0/0 is UB). | |||
| 2849 | XUgeY = KnownBits::uge(XKnown, YKnown); | |||
| 2850 | } | |||
| 2851 | // If X is total unknown or X u< Y we won't be able to prove non-zero | |||
| 2852 | // with compute known bits so just return early. | |||
| 2853 | return XUgeY && *XUgeY; | |||
| 2854 | } | |||
| 2855 | break; | |||
| 2856 | case Instruction::Add: { | |||
| 2857 | // X + Y. | |||
| 2858 | ||||
| 2859 | // If Add has nuw wrap flag, then if either X or Y is non-zero the result is | |||
| 2860 | // non-zero. | |||
| 2861 | auto *BO = cast<OverflowingBinaryOperator>(V); | |||
| 2862 | if (Q.IIQ.hasNoUnsignedWrap(BO)) | |||
| 2863 | return isKnownNonZero(I->getOperand(1), DemandedElts, Depth, Q) || | |||
| 2864 | isKnownNonZero(I->getOperand(0), DemandedElts, Depth, Q); | |||
| 2865 | ||||
| 2866 | return isNonZeroAdd(DemandedElts, Depth, Q, BitWidth, I->getOperand(0), | |||
| 2867 | I->getOperand(1), Q.IIQ.hasNoSignedWrap(BO)); | |||
| 2868 | } | |||
| 2869 | case Instruction::Mul: { | |||
| 2870 | // If X and Y are non-zero then so is X * Y as long as the multiplication | |||
| 2871 | // does not overflow. | |||
| 2872 | const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V); | |||
| 2873 | if (Q.IIQ.hasNoSignedWrap(BO) || Q.IIQ.hasNoUnsignedWrap(BO)) | |||
| 2874 | return isKnownNonZero(I->getOperand(0), DemandedElts, Depth, Q) && | |||
| 2875 | isKnownNonZero(I->getOperand(1), DemandedElts, Depth, Q); | |||
| 2876 | ||||
| 2877 | // If either X or Y is odd, then if the other is non-zero the result can't | |||
| 2878 | // be zero. | |||
| 2879 | KnownBits XKnown = | |||
| 2880 | computeKnownBits(I->getOperand(0), DemandedElts, Depth, Q); | |||
| 2881 | if (XKnown.One[0]) | |||
| 2882 | return isKnownNonZero(I->getOperand(1), DemandedElts, Depth, Q); | |||
| 2883 | ||||
| 2884 | KnownBits YKnown = | |||
| 2885 | computeKnownBits(I->getOperand(1), DemandedElts, Depth, Q); | |||
| 2886 | if (YKnown.One[0]) | |||
| 2887 | return XKnown.isNonZero() || | |||
| 2888 | isKnownNonZero(I->getOperand(0), DemandedElts, Depth, Q); | |||
| 2889 | ||||
| 2890 | return KnownBits::mul(XKnown, YKnown).isNonZero(); | |||
| 2891 | } | |||
| 2892 | case Instruction::Select: | |||
| 2893 | // (C ? X : Y) != 0 if X != 0 and Y != 0. | |||
| 2894 | if (isKnownNonZero(I->getOperand(1), DemandedElts, Depth, Q) && | |||
| 2895 | isKnownNonZero(I->getOperand(2), DemandedElts, Depth, Q)) | |||
| 2896 | return true; | |||
| 2897 | break; | |||
| 2898 | case Instruction::PHI: { | |||
| 2899 | auto *PN = cast<PHINode>(I); | |||
| 2900 | if (Q.IIQ.UseInstrInfo && isNonZeroRecurrence(PN)) | |||
| 2901 | return true; | |||
| 2902 | ||||
| 2903 | // Check if all incoming values are non-zero using recursion. | |||
| 2904 | Query RecQ = Q; | |||
| 2905 | unsigned NewDepth = std::max(Depth, MaxAnalysisRecursionDepth - 1); | |||
| 2906 | return llvm::all_of(PN->operands(), [&](const Use &U) { | |||
| 2907 | if (U.get() == PN) | |||
| 2908 | return true; | |||
| 2909 | RecQ.CxtI = PN->getIncomingBlock(U)->getTerminator(); | |||
| 2910 | return isKnownNonZero(U.get(), DemandedElts, NewDepth, RecQ); | |||
| 2911 | }); | |||
| 2912 | } | |||
| 2913 | case Instruction::ExtractElement: | |||
| 2914 | if (const auto *EEI = dyn_cast<ExtractElementInst>(V)) { | |||
| 2915 | const Value *Vec = EEI->getVectorOperand(); | |||
| 2916 | const Value *Idx = EEI->getIndexOperand(); | |||
| 2917 | auto *CIdx = dyn_cast<ConstantInt>(Idx); | |||
| 2918 | if (auto *VecTy = dyn_cast<FixedVectorType>(Vec->getType())) { | |||
| 2919 | unsigned NumElts = VecTy->getNumElements(); | |||
| 2920 | APInt DemandedVecElts = APInt::getAllOnes(NumElts); | |||
| 2921 | if (CIdx && CIdx->getValue().ult(NumElts)) | |||
| 2922 | DemandedVecElts = APInt::getOneBitSet(NumElts, CIdx->getZExtValue()); | |||
| 2923 | return isKnownNonZero(Vec, DemandedVecElts, Depth, Q); | |||
| 2924 | } | |||
| 2925 | } | |||
| 2926 | break; | |||
| 2927 | case Instruction::Freeze: | |||
| 2928 | return isKnownNonZero(I->getOperand(0), Depth, Q) && | |||
| 2929 | isGuaranteedNotToBePoison(I->getOperand(0), Q.AC, Q.CxtI, Q.DT, | |||
| 2930 | Depth); | |||
| 2931 | case Instruction::Call: | |||
| 2932 | if (auto *II = dyn_cast<IntrinsicInst>(I)) { | |||
| 2933 | switch (II->getIntrinsicID()) { | |||
| 2934 | case Intrinsic::sshl_sat: | |||
| 2935 | case Intrinsic::ushl_sat: | |||
| 2936 | case Intrinsic::abs: | |||
| 2937 | case Intrinsic::bitreverse: | |||
| 2938 | case Intrinsic::bswap: | |||
| 2939 | case Intrinsic::ctpop: | |||
| 2940 | return isKnownNonZero(II->getArgOperand(0), DemandedElts, Depth, Q); | |||
| 2941 | case Intrinsic::ssub_sat: | |||
| 2942 | return isNonZeroSub(DemandedElts, Depth, Q, BitWidth, | |||
| 2943 | II->getArgOperand(0), II->getArgOperand(1)); | |||
| 2944 | case Intrinsic::sadd_sat: | |||
| 2945 | return isNonZeroAdd(DemandedElts, Depth, Q, BitWidth, | |||
| 2946 | II->getArgOperand(0), II->getArgOperand(1), | |||
| 2947 | /*NSW*/ true); | |||
| 2948 | case Intrinsic::umax: | |||
| 2949 | case Intrinsic::uadd_sat: | |||
| 2950 | return isKnownNonZero(II->getArgOperand(1), DemandedElts, Depth, Q) || | |||
| 2951 | isKnownNonZero(II->getArgOperand(0), DemandedElts, Depth, Q); | |||
| 2952 | case Intrinsic::smin: | |||
| 2953 | case Intrinsic::smax: { | |||
| 2954 | auto KnownOpImpliesNonZero = [&](const KnownBits &K) { | |||
| 2955 | return II->getIntrinsicID() == Intrinsic::smin | |||
| 2956 | ? K.isNegative() | |||
| 2957 | : K.isStrictlyPositive(); | |||
| 2958 | }; | |||
| 2959 | KnownBits XKnown = | |||
| 2960 | computeKnownBits(II->getArgOperand(0), DemandedElts, Depth, Q); | |||
| 2961 | if (KnownOpImpliesNonZero(XKnown)) | |||
| 2962 | return true; | |||
| 2963 | KnownBits YKnown = | |||
| 2964 | computeKnownBits(II->getArgOperand(1), DemandedElts, Depth, Q); | |||
| 2965 | if (KnownOpImpliesNonZero(YKnown)) | |||
| 2966 | return true; | |||
| 2967 | ||||
| 2968 | if (XKnown.isNonZero() && YKnown.isNonZero()) | |||
| 2969 | return true; | |||
| 2970 | } | |||
| 2971 | [[fallthrough]]; | |||
| 2972 | case Intrinsic::umin: | |||
| 2973 | return isKnownNonZero(II->getArgOperand(0), DemandedElts, Depth, Q) && | |||
| 2974 | isKnownNonZero(II->getArgOperand(1), DemandedElts, Depth, Q); | |||
| 2975 | case Intrinsic::cttz: | |||
| 2976 | return computeKnownBits(II->getArgOperand(0), DemandedElts, Depth, Q) | |||
| 2977 | .Zero[0]; | |||
| 2978 | case Intrinsic::ctlz: | |||
| 2979 | return computeKnownBits(II->getArgOperand(0), DemandedElts, Depth, Q) | |||
| 2980 | .isNonNegative(); | |||
| 2981 | case Intrinsic::fshr: | |||
| 2982 | case Intrinsic::fshl: | |||
| 2983 | // If Op0 == Op1, this is a rotate. rotate(x, y) != 0 iff x != 0. | |||
| 2984 | if (II->getArgOperand(0) == II->getArgOperand(1)) | |||
| 2985 | return isKnownNonZero(II->getArgOperand(0), DemandedElts, Depth, Q); | |||
| 2986 | break; | |||
| 2987 | case Intrinsic::vscale: | |||
| 2988 | return true; | |||
| 2989 | default: | |||
| 2990 | break; | |||
| 2991 | } | |||
| 2992 | } | |||
| 2993 | break; | |||
| 2994 | } | |||
| 2995 | ||||
| 2996 | KnownBits Known(BitWidth); | |||
| 2997 | computeKnownBits(V, DemandedElts, Known, Depth, Q); | |||
| 2998 | return Known.One != 0; | |||
| 2999 | } | |||
| 3000 | ||||
| 3001 | bool isKnownNonZero(const Value* V, unsigned Depth, const Query& Q) { | |||
| 3002 | auto *FVTy = dyn_cast<FixedVectorType>(V->getType()); | |||
| 3003 | APInt DemandedElts = | |||
| 3004 | FVTy ? APInt::getAllOnes(FVTy->getNumElements()) : APInt(1, 1); | |||
| 3005 | return isKnownNonZero(V, DemandedElts, Depth, Q); | |||
| 3006 | } | |||
| 3007 | ||||
| 3008 | /// If the pair of operators are the same invertible function, return the | |||
| 3009 | /// the operands of the function corresponding to each input. Otherwise, | |||
| 3010 | /// return std::nullopt. An invertible function is one that is 1-to-1 and maps | |||
| 3011 | /// every input value to exactly one output value. This is equivalent to | |||
| 3012 | /// saying that Op1 and Op2 are equal exactly when the specified pair of | |||
| 3013 | /// operands are equal, (except that Op1 and Op2 may be poison more often.) | |||
| 3014 | static std::optional<std::pair<Value*, Value*>> | |||
| 3015 | getInvertibleOperands(const Operator *Op1, | |||
| 3016 | const Operator *Op2) { | |||
| 3017 | if (Op1->getOpcode() != Op2->getOpcode()) | |||
| 3018 | return std::nullopt; | |||
| 3019 | ||||
| 3020 | auto getOperands = [&](unsigned OpNum) -> auto { | |||
| 3021 | return std::make_pair(Op1->getOperand(OpNum), Op2->getOperand(OpNum)); | |||
| 3022 | }; | |||
| 3023 | ||||
| 3024 | switch (Op1->getOpcode()) { | |||
| 3025 | default: | |||
| 3026 | break; | |||
| 3027 | case Instruction::Add: | |||
| 3028 | case Instruction::Sub: | |||
| 3029 | if (Op1->getOperand(0) == Op2->getOperand(0)) | |||
| 3030 | return getOperands(1); | |||
| 3031 | if (Op1->getOperand(1) == Op2->getOperand(1)) | |||
| 3032 | return getOperands(0); | |||
| 3033 | break; | |||
| 3034 | case Instruction::Mul: { | |||
| 3035 | // invertible if A * B == (A * B) mod 2^N where A, and B are integers | |||
| 3036 | // and N is the bitwdith. The nsw case is non-obvious, but proven by | |||
| 3037 | // alive2: https://alive2.llvm.org/ce/z/Z6D5qK | |||
| 3038 | auto *OBO1 = cast<OverflowingBinaryOperator>(Op1); | |||
| 3039 | auto *OBO2 = cast<OverflowingBinaryOperator>(Op2); | |||
| 3040 | if ((!OBO1->hasNoUnsignedWrap() || !OBO2->hasNoUnsignedWrap()) && | |||
| 3041 | (!OBO1->hasNoSignedWrap() || !OBO2->hasNoSignedWrap())) | |||
| 3042 | break; | |||
| 3043 | ||||
| 3044 | // Assume operand order has been canonicalized | |||
| 3045 | if (Op1->getOperand(1) == Op2->getOperand(1) && | |||
| 3046 | isa<ConstantInt>(Op1->getOperand(1)) && | |||
| 3047 | !cast<ConstantInt>(Op1->getOperand(1))->isZero()) | |||
| 3048 | return getOperands(0); | |||
| 3049 | break; | |||
| 3050 | } | |||
| 3051 | case Instruction::Shl: { | |||
| 3052 | // Same as multiplies, with the difference that we don't need to check | |||
| 3053 | // for a non-zero multiply. Shifts always multiply by non-zero. | |||
| 3054 | auto *OBO1 = cast<OverflowingBinaryOperator>(Op1); | |||
| 3055 | auto *OBO2 = cast<OverflowingBinaryOperator>(Op2); | |||
| 3056 | if ((!OBO1->hasNoUnsignedWrap() || !OBO2->hasNoUnsignedWrap()) && | |||
| 3057 | (!OBO1->hasNoSignedWrap() || !OBO2->hasNoSignedWrap())) | |||
| 3058 | break; | |||
| 3059 | ||||
| 3060 | if (Op1->getOperand(1) == Op2->getOperand(1)) | |||
| 3061 | return getOperands(0); | |||
| 3062 | break; | |||
| 3063 | } | |||
| 3064 | case Instruction::AShr: | |||
| 3065 | case Instruction::LShr: { | |||
| 3066 | auto *PEO1 = cast<PossiblyExactOperator>(Op1); | |||
| 3067 | auto *PEO2 = cast<PossiblyExactOperator>(Op2); | |||
| 3068 | if (!PEO1->isExact() || !PEO2->isExact()) | |||
| 3069 | break; | |||
| 3070 | ||||
| 3071 | if (Op1->getOperand(1) == Op2->getOperand(1)) | |||
| 3072 | return getOperands(0); | |||
| 3073 | break; | |||
| 3074 | } | |||
| 3075 | case Instruction::SExt: | |||
| 3076 | case Instruction::ZExt: | |||
| 3077 | if (Op1->getOperand(0)->getType() == Op2->getOperand(0)->getType()) | |||
| 3078 | return getOperands(0); | |||
| 3079 | break; | |||
| 3080 | case Instruction::PHI: { | |||
| 3081 | const PHINode *PN1 = cast<PHINode>(Op1); | |||
| 3082 | const PHINode *PN2 = cast<PHINode>(Op2); | |||
| 3083 | ||||
| 3084 | // If PN1 and PN2 are both recurrences, can we prove the entire recurrences | |||
| 3085 | // are a single invertible function of the start values? Note that repeated | |||
| 3086 | // application of an invertible function is also invertible | |||
| 3087 | BinaryOperator *BO1 = nullptr; | |||
| 3088 | Value *Start1 = nullptr, *Step1 = nullptr; | |||
| 3089 | BinaryOperator *BO2 = nullptr; | |||
| 3090 | Value *Start2 = nullptr, *Step2 = nullptr; | |||
| 3091 | if (PN1->getParent() != PN2->getParent() || | |||
| 3092 | !matchSimpleRecurrence(PN1, BO1, Start1, Step1) || | |||
| 3093 | !matchSimpleRecurrence(PN2, BO2, Start2, Step2)) | |||
| 3094 | break; | |||
| 3095 | ||||
| 3096 | auto Values = getInvertibleOperands(cast<Operator>(BO1), | |||
| 3097 | cast<Operator>(BO2)); | |||
| 3098 | if (!Values) | |||
| 3099 | break; | |||
| 3100 | ||||
| 3101 | // We have to be careful of mutually defined recurrences here. Ex: | |||
| 3102 | // * X_i = X_(i-1) OP Y_(i-1), and Y_i = X_(i-1) OP V | |||
| 3103 | // * X_i = Y_i = X_(i-1) OP Y_(i-1) | |||
| 3104 | // The invertibility of these is complicated, and not worth reasoning | |||
| 3105 | // about (yet?). | |||
| 3106 | if (Values->first != PN1 || Values->second != PN2) | |||
| 3107 | break; | |||
| 3108 | ||||
| 3109 | return std::make_pair(Start1, Start2); | |||
| 3110 | } | |||
| 3111 | } | |||
| 3112 | return std::nullopt; | |||
| 3113 | } | |||
| 3114 | ||||
| 3115 | /// Return true if V2 == V1 + X, where X is known non-zero. | |||
| 3116 | static bool isAddOfNonZero(const Value *V1, const Value *V2, unsigned Depth, | |||
| 3117 | const Query &Q) { | |||
| 3118 | const BinaryOperator *BO = dyn_cast<BinaryOperator>(V1); | |||
| 3119 | if (!BO || BO->getOpcode() != Instruction::Add) | |||
| 3120 | return false; | |||
| 3121 | Value *Op = nullptr; | |||
| 3122 | if (V2 == BO->getOperand(0)) | |||
| 3123 | Op = BO->getOperand(1); | |||
| 3124 | else if (V2 == BO->getOperand(1)) | |||
| 3125 | Op = BO->getOperand(0); | |||
| 3126 | else | |||
| 3127 | return false; | |||
| 3128 | return isKnownNonZero(Op, Depth + 1, Q); | |||
| 3129 | } | |||
| 3130 | ||||
| 3131 | /// Return true if V2 == V1 * C, where V1 is known non-zero, C is not 0/1 and | |||
| 3132 | /// the multiplication is nuw or nsw. | |||
| 3133 | static bool isNonEqualMul(const Value *V1, const Value *V2, unsigned Depth, | |||
| 3134 | const Query &Q) { | |||
| 3135 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(V2)) { | |||
| 3136 | const APInt *C; | |||
| 3137 | return match(OBO, m_Mul(m_Specific(V1), m_APInt(C))) && | |||
| 3138 | (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) && | |||
| 3139 | !C->isZero() && !C->isOne() && isKnownNonZero(V1, Depth + 1, Q); | |||
| 3140 | } | |||
| 3141 | return false; | |||
| 3142 | } | |||
| 3143 | ||||
| 3144 | /// Return true if V2 == V1 << C, where V1 is known non-zero, C is not 0 and | |||
| 3145 | /// the shift is nuw or nsw. | |||
| 3146 | static bool isNonEqualShl(const Value *V1, const Value *V2, unsigned Depth, | |||
| 3147 | const Query &Q) { | |||
| 3148 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(V2)) { | |||
| 3149 | const APInt *C; | |||
| 3150 | return match(OBO, m_Shl(m_Specific(V1), m_APInt(C))) && | |||
| 3151 | (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) && | |||
| 3152 | !C->isZero() && isKnownNonZero(V1, Depth + 1, Q); | |||
| 3153 | } | |||
| 3154 | return false; | |||
| 3155 | } | |||
| 3156 | ||||
| 3157 | static bool isNonEqualPHIs(const PHINode *PN1, const PHINode *PN2, | |||
| 3158 | unsigned Depth, const Query &Q) { | |||
| 3159 | // Check two PHIs are in same block. | |||
| 3160 | if (PN1->getParent() != PN2->getParent()) | |||
| 3161 | return false; | |||
| 3162 | ||||
| 3163 | SmallPtrSet<const BasicBlock *, 8> VisitedBBs; | |||
| 3164 | bool UsedFullRecursion = false; | |||
| 3165 | for (const BasicBlock *IncomBB : PN1->blocks()) { | |||
| 3166 | if (!VisitedBBs.insert(IncomBB).second) | |||
| 3167 | continue; // Don't reprocess blocks that we have dealt with already. | |||
| 3168 | const Value *IV1 = PN1->getIncomingValueForBlock(IncomBB); | |||
| 3169 | const Value *IV2 = PN2->getIncomingValueForBlock(IncomBB); | |||
| 3170 | const APInt *C1, *C2; | |||
| 3171 | if (match(IV1, m_APInt(C1)) && match(IV2, m_APInt(C2)) && *C1 != *C2) | |||
| 3172 | continue; | |||
| 3173 | ||||
| 3174 | // Only one pair of phi operands is allowed for full recursion. | |||
| 3175 | if (UsedFullRecursion) | |||
| 3176 | return false; | |||
| 3177 | ||||
| 3178 | Query RecQ = Q; | |||
| 3179 | RecQ.CxtI = IncomBB->getTerminator(); | |||
| 3180 | if (!isKnownNonEqual(IV1, IV2, Depth + 1, RecQ)) | |||
| 3181 | return false; | |||
| 3182 | UsedFullRecursion = true; | |||
| 3183 | } | |||
| 3184 | return true; | |||
| 3185 | } | |||
| 3186 | ||||
| 3187 | /// Return true if it is known that V1 != V2. | |||
| 3188 | static bool isKnownNonEqual(const Value *V1, const Value *V2, unsigned Depth, | |||
| 3189 | const Query &Q) { | |||
| 3190 | if (V1 == V2) | |||
| 3191 | return false; | |||
| 3192 | if (V1->getType() != V2->getType()) | |||
| 3193 | // We can't look through casts yet. | |||
| 3194 | return false; | |||
| 3195 | ||||
| 3196 | if (Depth >= MaxAnalysisRecursionDepth) | |||
| 3197 | return false; | |||
| 3198 | ||||
| 3199 | // See if we can recurse through (exactly one of) our operands. This | |||
| 3200 | // requires our operation be 1-to-1 and map every input value to exactly | |||
| 3201 | // one output value. Such an operation is invertible. | |||
| 3202 | auto *O1 = dyn_cast<Operator>(V1); | |||
| 3203 | auto *O2 = dyn_cast<Operator>(V2); | |||
| 3204 | if (O1 && O2 && O1->getOpcode() == O2->getOpcode()) { | |||
| 3205 | if (auto Values = getInvertibleOperands(O1, O2)) | |||
| 3206 | return isKnownNonEqual(Values->first, Values->second, Depth + 1, Q); | |||
| 3207 | ||||
| 3208 | if (const PHINode *PN1 = dyn_cast<PHINode>(V1)) { | |||
| 3209 | const PHINode *PN2 = cast<PHINode>(V2); | |||
| 3210 | // FIXME: This is missing a generalization to handle the case where one is | |||
| 3211 | // a PHI and another one isn't. | |||
| 3212 | if (isNonEqualPHIs(PN1, PN2, Depth, Q)) | |||
| 3213 | return true; | |||
| 3214 | }; | |||
| 3215 | } | |||
| 3216 | ||||
| 3217 | if (isAddOfNonZero(V1, V2, Depth, Q) || isAddOfNonZero(V2, V1, Depth, Q)) | |||
| 3218 | return true; | |||
| 3219 | ||||
| 3220 | if (isNonEqualMul(V1, V2, Depth, Q) || isNonEqualMul(V2, V1, Depth, Q)) | |||
| 3221 | return true; | |||
| 3222 | ||||
| 3223 | if (isNonEqualShl(V1, V2, Depth, Q) || isNonEqualShl(V2, V1, Depth, Q)) | |||
| 3224 | return true; | |||
| 3225 | ||||
| 3226 | if (V1->getType()->isIntOrIntVectorTy()) { | |||
| 3227 | // Are any known bits in V1 contradictory to known bits in V2? If V1 | |||
| 3228 | // has a known zero where V2 has a known one, they must not be equal. | |||
| 3229 | KnownBits Known1 = computeKnownBits(V1, Depth, Q); | |||
| 3230 | KnownBits Known2 = computeKnownBits(V2, Depth, Q); | |||
| 3231 | ||||
| 3232 | if (Known1.Zero.intersects(Known2.One) || | |||
| 3233 | Known2.Zero.intersects(Known1.One)) | |||
| 3234 | return true; | |||
| 3235 | } | |||
| 3236 | return false; | |||
| 3237 | } | |||
| 3238 | ||||
| 3239 | /// Return true if 'V & Mask' is known to be zero. We use this predicate to | |||
| 3240 | /// simplify operations downstream. Mask is known to be zero for bits that V | |||
| 3241 | /// cannot have. | |||
| 3242 | /// | |||
| 3243 | /// This function is defined on values with integer type, values with pointer | |||
| 3244 | /// type, and vectors of integers. In the case | |||
| 3245 | /// where V is a vector, the mask, known zero, and known one values are the | |||
| 3246 | /// same width as the vector element, and the bit is set only if it is true | |||
| 3247 | /// for all of the elements in the vector. | |||
| 3248 | bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth, | |||
| 3249 | const Query &Q) { | |||
| 3250 | KnownBits Known(Mask.getBitWidth()); | |||
| 3251 | computeKnownBits(V, Known, Depth, Q); | |||
| 3252 | return Mask.isSubsetOf(Known.Zero); | |||
| 3253 | } | |||
| 3254 | ||||
| 3255 | // Match a signed min+max clamp pattern like smax(smin(In, CHigh), CLow). | |||
| 3256 | // Returns the input and lower/upper bounds. | |||
| 3257 | static bool isSignedMinMaxClamp(const Value *Select, const Value *&In, | |||
| 3258 | const APInt *&CLow, const APInt *&CHigh) { | |||
| 3259 | assert(isa<Operator>(Select) &&(static_cast <bool> (isa<Operator>(Select) && cast<Operator>(Select)->getOpcode() == Instruction:: Select && "Input should be a Select!") ? void (0) : __assert_fail ("isa<Operator>(Select) && cast<Operator>(Select)->getOpcode() == Instruction::Select && \"Input should be a Select!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3261, __extension__ __PRETTY_FUNCTION__ )) | |||
| 3260 | cast<Operator>(Select)->getOpcode() == Instruction::Select &&(static_cast <bool> (isa<Operator>(Select) && cast<Operator>(Select)->getOpcode() == Instruction:: Select && "Input should be a Select!") ? void (0) : __assert_fail ("isa<Operator>(Select) && cast<Operator>(Select)->getOpcode() == Instruction::Select && \"Input should be a Select!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3261, __extension__ __PRETTY_FUNCTION__ )) | |||
| 3261 | "Input should be a Select!")(static_cast <bool> (isa<Operator>(Select) && cast<Operator>(Select)->getOpcode() == Instruction:: Select && "Input should be a Select!") ? void (0) : __assert_fail ("isa<Operator>(Select) && cast<Operator>(Select)->getOpcode() == Instruction::Select && \"Input should be a Select!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3261, __extension__ __PRETTY_FUNCTION__ )); | |||
| 3262 | ||||
| 3263 | const Value *LHS = nullptr, *RHS = nullptr; | |||
| 3264 | SelectPatternFlavor SPF = matchSelectPattern(Select, LHS, RHS).Flavor; | |||
| 3265 | if (SPF != SPF_SMAX && SPF != SPF_SMIN) | |||
| 3266 | return false; | |||
| 3267 | ||||
| 3268 | if (!match(RHS, m_APInt(CLow))) | |||
| 3269 | return false; | |||
| 3270 | ||||
| 3271 | const Value *LHS2 = nullptr, *RHS2 = nullptr; | |||
| 3272 | SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor; | |||
| 3273 | if (getInverseMinMaxFlavor(SPF) != SPF2) | |||
| 3274 | return false; | |||
| 3275 | ||||
| 3276 | if (!match(RHS2, m_APInt(CHigh))) | |||
| 3277 | return false; | |||
| 3278 | ||||
| 3279 | if (SPF == SPF_SMIN) | |||
| 3280 | std::swap(CLow, CHigh); | |||
| 3281 | ||||
| 3282 | In = LHS2; | |||
| 3283 | return CLow->sle(*CHigh); | |||
| 3284 | } | |||
| 3285 | ||||
| 3286 | static bool isSignedMinMaxIntrinsicClamp(const IntrinsicInst *II, | |||
| 3287 | const APInt *&CLow, | |||
| 3288 | const APInt *&CHigh) { | |||
| 3289 | assert((II->getIntrinsicID() == Intrinsic::smin ||(static_cast <bool> ((II->getIntrinsicID() == Intrinsic ::smin || II->getIntrinsicID() == Intrinsic::smax) && "Must be smin/smax") ? void (0) : __assert_fail ("(II->getIntrinsicID() == Intrinsic::smin || II->getIntrinsicID() == Intrinsic::smax) && \"Must be smin/smax\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3290, __extension__ __PRETTY_FUNCTION__ )) | |||
| 3290 | II->getIntrinsicID() == Intrinsic::smax) && "Must be smin/smax")(static_cast <bool> ((II->getIntrinsicID() == Intrinsic ::smin || II->getIntrinsicID() == Intrinsic::smax) && "Must be smin/smax") ? void (0) : __assert_fail ("(II->getIntrinsicID() == Intrinsic::smin || II->getIntrinsicID() == Intrinsic::smax) && \"Must be smin/smax\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3290, __extension__ __PRETTY_FUNCTION__ )); | |||
| 3291 | ||||
| 3292 | Intrinsic::ID InverseID = getInverseMinMaxIntrinsic(II->getIntrinsicID()); | |||
| 3293 | auto *InnerII = dyn_cast<IntrinsicInst>(II->getArgOperand(0)); | |||
| 3294 | if (!InnerII || InnerII->getIntrinsicID() != InverseID || | |||
| 3295 | !match(II->getArgOperand(1), m_APInt(CLow)) || | |||
| 3296 | !match(InnerII->getArgOperand(1), m_APInt(CHigh))) | |||
| 3297 | return false; | |||
| 3298 | ||||
| 3299 | if (II->getIntrinsicID() == Intrinsic::smin) | |||
| 3300 | std::swap(CLow, CHigh); | |||
| 3301 | return CLow->sle(*CHigh); | |||
| 3302 | } | |||
| 3303 | ||||
| 3304 | /// For vector constants, loop over the elements and find the constant with the | |||
| 3305 | /// minimum number of sign bits. Return 0 if the value is not a vector constant | |||
| 3306 | /// or if any element was not analyzed; otherwise, return the count for the | |||
| 3307 | /// element with the minimum number of sign bits. | |||
| 3308 | static unsigned computeNumSignBitsVectorConstant(const Value *V, | |||
| 3309 | const APInt &DemandedElts, | |||
| 3310 | unsigned TyBits) { | |||
| 3311 | const auto *CV = dyn_cast<Constant>(V); | |||
| 3312 | if (!CV || !isa<FixedVectorType>(CV->getType())) | |||
| 3313 | return 0; | |||
| 3314 | ||||
| 3315 | unsigned MinSignBits = TyBits; | |||
| 3316 | unsigned NumElts = cast<FixedVectorType>(CV->getType())->getNumElements(); | |||
| 3317 | for (unsigned i = 0; i != NumElts; ++i) { | |||
| 3318 | if (!DemandedElts[i]) | |||
| 3319 | continue; | |||
| 3320 | // If we find a non-ConstantInt, bail out. | |||
| 3321 | auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i)); | |||
| 3322 | if (!Elt) | |||
| 3323 | return 0; | |||
| 3324 | ||||
| 3325 | MinSignBits = std::min(MinSignBits, Elt->getValue().getNumSignBits()); | |||
| 3326 | } | |||
| 3327 | ||||
| 3328 | return MinSignBits; | |||
| 3329 | } | |||
| 3330 | ||||
| 3331 | static unsigned ComputeNumSignBitsImpl(const Value *V, | |||
| 3332 | const APInt &DemandedElts, | |||
| 3333 | unsigned Depth, const Query &Q); | |||
| 3334 | ||||
| 3335 | static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts, | |||
| 3336 | unsigned Depth, const Query &Q) { | |||
| 3337 | unsigned Result = ComputeNumSignBitsImpl(V, DemandedElts, Depth, Q); | |||
| 3338 | assert(Result > 0 && "At least one sign bit needs to be present!")(static_cast <bool> (Result > 0 && "At least one sign bit needs to be present!" ) ? void (0) : __assert_fail ("Result > 0 && \"At least one sign bit needs to be present!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3338, __extension__ __PRETTY_FUNCTION__ )); | |||
| 3339 | return Result; | |||
| 3340 | } | |||
| 3341 | ||||
| 3342 | /// Return the number of times the sign bit of the register is replicated into | |||
| 3343 | /// the other bits. We know that at least 1 bit is always equal to the sign bit | |||
| 3344 | /// (itself), but other cases can give us information. For example, immediately | |||
| 3345 | /// after an "ashr X, 2", we know that the top 3 bits are all equal to each | |||
| 3346 | /// other, so we return 3. For vectors, return the number of sign bits for the | |||
| 3347 | /// vector element with the minimum number of known sign bits of the demanded | |||
| 3348 | /// elements in the vector specified by DemandedElts. | |||
| 3349 | static unsigned ComputeNumSignBitsImpl(const Value *V, | |||
| 3350 | const APInt &DemandedElts, | |||
| 3351 | unsigned Depth, const Query &Q) { | |||
| 3352 | Type *Ty = V->getType(); | |||
| 3353 | #ifndef NDEBUG | |||
| 3354 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth")(static_cast <bool> (Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth") ? void (0) : __assert_fail ( "Depth <= MaxAnalysisRecursionDepth && \"Limit Search Depth\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3354, __extension__ __PRETTY_FUNCTION__ )); | |||
| 3355 | ||||
| 3356 | if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) { | |||
| 3357 | assert((static_cast <bool> (FVTy->getNumElements() == DemandedElts .getBitWidth() && "DemandedElt width should equal the fixed vector number of elements" ) ? void (0) : __assert_fail ("FVTy->getNumElements() == DemandedElts.getBitWidth() && \"DemandedElt width should equal the fixed vector number of elements\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3359, __extension__ __PRETTY_FUNCTION__ )) | |||
| 3358 | FVTy->getNumElements() == DemandedElts.getBitWidth() &&(static_cast <bool> (FVTy->getNumElements() == DemandedElts .getBitWidth() && "DemandedElt width should equal the fixed vector number of elements" ) ? void (0) : __assert_fail ("FVTy->getNumElements() == DemandedElts.getBitWidth() && \"DemandedElt width should equal the fixed vector number of elements\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3359, __extension__ __PRETTY_FUNCTION__ )) | |||
| 3359 | "DemandedElt width should equal the fixed vector number of elements")(static_cast <bool> (FVTy->getNumElements() == DemandedElts .getBitWidth() && "DemandedElt width should equal the fixed vector number of elements" ) ? void (0) : __assert_fail ("FVTy->getNumElements() == DemandedElts.getBitWidth() && \"DemandedElt width should equal the fixed vector number of elements\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3359, __extension__ __PRETTY_FUNCTION__ )); | |||
| 3360 | } else { | |||
| 3361 | assert(DemandedElts == APInt(1, 1) &&(static_cast <bool> (DemandedElts == APInt(1, 1) && "DemandedElt width should be 1 for scalars") ? void (0) : __assert_fail ("DemandedElts == APInt(1, 1) && \"DemandedElt width should be 1 for scalars\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3362, __extension__ __PRETTY_FUNCTION__ )) | |||
| 3362 | "DemandedElt width should be 1 for scalars")(static_cast <bool> (DemandedElts == APInt(1, 1) && "DemandedElt width should be 1 for scalars") ? void (0) : __assert_fail ("DemandedElts == APInt(1, 1) && \"DemandedElt width should be 1 for scalars\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3362, __extension__ __PRETTY_FUNCTION__ )); | |||
| 3363 | } | |||
| 3364 | #endif | |||
| 3365 | ||||
| 3366 | // We return the minimum number of sign bits that are guaranteed to be present | |||
| 3367 | // in V, so for undef we have to conservatively return 1. We don't have the | |||
| 3368 | // same behavior for poison though -- that's a FIXME today. | |||
| 3369 | ||||
| 3370 | Type *ScalarTy = Ty->getScalarType(); | |||
| 3371 | unsigned TyBits = ScalarTy->isPointerTy() ? | |||
| 3372 | Q.DL.getPointerTypeSizeInBits(ScalarTy) : | |||
| 3373 | Q.DL.getTypeSizeInBits(ScalarTy); | |||
| 3374 | ||||
| 3375 | unsigned Tmp, Tmp2; | |||
| 3376 | unsigned FirstAnswer = 1; | |||
| 3377 | ||||
| 3378 | // Note that ConstantInt is handled by the general computeKnownBits case | |||
| 3379 | // below. | |||
| 3380 | ||||
| 3381 | if (Depth == MaxAnalysisRecursionDepth) | |||
| 3382 | return 1; | |||
| 3383 | ||||
| 3384 | if (auto *U = dyn_cast<Operator>(V)) { | |||
| 3385 | switch (Operator::getOpcode(V)) { | |||
| 3386 | default: break; | |||
| 3387 | case Instruction::SExt: | |||
| 3388 | Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits(); | |||
| 3389 | return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp; | |||
| 3390 | ||||
| 3391 | case Instruction::SDiv: { | |||
| 3392 | const APInt *Denominator; | |||
| 3393 | // sdiv X, C -> adds log(C) sign bits. | |||
| 3394 | if (match(U->getOperand(1), m_APInt(Denominator))) { | |||
| 3395 | ||||
| 3396 | // Ignore non-positive denominator. | |||
| 3397 | if (!Denominator->isStrictlyPositive()) | |||
| 3398 | break; | |||
| 3399 | ||||
| 3400 | // Calculate the incoming numerator bits. | |||
| 3401 | unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
| 3402 | ||||
| 3403 | // Add floor(log(C)) bits to the numerator bits. | |||
| 3404 | return std::min(TyBits, NumBits + Denominator->logBase2()); | |||
| 3405 | } | |||
| 3406 | break; | |||
| 3407 | } | |||
| 3408 | ||||
| 3409 | case Instruction::SRem: { | |||
| 3410 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
| 3411 | ||||
| 3412 | const APInt *Denominator; | |||
| 3413 | // srem X, C -> we know that the result is within [-C+1,C) when C is a | |||
| 3414 | // positive constant. This let us put a lower bound on the number of sign | |||
| 3415 | // bits. | |||
| 3416 | if (match(U->getOperand(1), m_APInt(Denominator))) { | |||
| 3417 | ||||
| 3418 | // Ignore non-positive denominator. | |||
| 3419 | if (Denominator->isStrictlyPositive()) { | |||
| 3420 | // Calculate the leading sign bit constraints by examining the | |||
| 3421 | // denominator. Given that the denominator is positive, there are two | |||
| 3422 | // cases: | |||
| 3423 | // | |||
| 3424 | // 1. The numerator is positive. The result range is [0,C) and | |||
| 3425 | // [0,C) u< (1 << ceilLogBase2(C)). | |||
| 3426 | // | |||
| 3427 | // 2. The numerator is negative. Then the result range is (-C,0] and | |||
| 3428 | // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)). | |||
| 3429 | // | |||
| 3430 | // Thus a lower bound on the number of sign bits is `TyBits - | |||
| 3431 | // ceilLogBase2(C)`. | |||
| 3432 | ||||
| 3433 | unsigned ResBits = TyBits - Denominator->ceilLogBase2(); | |||
| 3434 | Tmp = std::max(Tmp, ResBits); | |||
| 3435 | } | |||
| 3436 | } | |||
| 3437 | return Tmp; | |||
| 3438 | } | |||
| 3439 | ||||
| 3440 | case Instruction::AShr: { | |||
| 3441 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
| 3442 | // ashr X, C -> adds C sign bits. Vectors too. | |||
| 3443 | const APInt *ShAmt; | |||
| 3444 | if (match(U->getOperand(1), m_APInt(ShAmt))) { | |||
| 3445 | if (ShAmt->uge(TyBits)) | |||
| 3446 | break; // Bad shift. | |||
| 3447 | unsigned ShAmtLimited = ShAmt->getZExtValue(); | |||
| 3448 | Tmp += ShAmtLimited; | |||
| 3449 | if (Tmp > TyBits) Tmp = TyBits; | |||
| 3450 | } | |||
| 3451 | return Tmp; | |||
| 3452 | } | |||
| 3453 | case Instruction::Shl: { | |||
| 3454 | const APInt *ShAmt; | |||
| 3455 | if (match(U->getOperand(1), m_APInt(ShAmt))) { | |||
| 3456 | // shl destroys sign bits. | |||
| 3457 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
| 3458 | if (ShAmt->uge(TyBits) || // Bad shift. | |||
| 3459 | ShAmt->uge(Tmp)) break; // Shifted all sign bits out. | |||
| 3460 | Tmp2 = ShAmt->getZExtValue(); | |||
| 3461 | return Tmp - Tmp2; | |||
| 3462 | } | |||
| 3463 | break; | |||
| 3464 | } | |||
| 3465 | case Instruction::And: | |||
| 3466 | case Instruction::Or: | |||
| 3467 | case Instruction::Xor: // NOT is handled here. | |||
| 3468 | // Logical binary ops preserve the number of sign bits at the worst. | |||
| 3469 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
| 3470 | if (Tmp != 1) { | |||
| 3471 | Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); | |||
| 3472 | FirstAnswer = std::min(Tmp, Tmp2); | |||
| 3473 | // We computed what we know about the sign bits as our first | |||
| 3474 | // answer. Now proceed to the generic code that uses | |||
| 3475 | // computeKnownBits, and pick whichever answer is better. | |||
| 3476 | } | |||
| 3477 | break; | |||
| 3478 | ||||
| 3479 | case Instruction::Select: { | |||
| 3480 | // If we have a clamp pattern, we know that the number of sign bits will | |||
| 3481 | // be the minimum of the clamp min/max range. | |||
| 3482 | const Value *X; | |||
| 3483 | const APInt *CLow, *CHigh; | |||
| 3484 | if (isSignedMinMaxClamp(U, X, CLow, CHigh)) | |||
| 3485 | return std::min(CLow->getNumSignBits(), CHigh->getNumSignBits()); | |||
| 3486 | ||||
| 3487 | Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); | |||
| 3488 | if (Tmp == 1) break; | |||
| 3489 | Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q); | |||
| 3490 | return std::min(Tmp, Tmp2); | |||
| 3491 | } | |||
| 3492 | ||||
| 3493 | case Instruction::Add: | |||
| 3494 | // Add can have at most one carry bit. Thus we know that the output | |||
| 3495 | // is, at worst, one more bit than the inputs. | |||
| 3496 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
| 3497 | if (Tmp == 1) break; | |||
| 3498 | ||||
| 3499 | // Special case decrementing a value (ADD X, -1): | |||
| 3500 | if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1))) | |||
| 3501 | if (CRHS->isAllOnesValue()) { | |||
| 3502 | KnownBits Known(TyBits); | |||
| 3503 | computeKnownBits(U->getOperand(0), Known, Depth + 1, Q); | |||
| 3504 | ||||
| 3505 | // If the input is known to be 0 or 1, the output is 0/-1, which is | |||
| 3506 | // all sign bits set. | |||
| 3507 | if ((Known.Zero | 1).isAllOnes()) | |||
| 3508 | return TyBits; | |||
| 3509 | ||||
| 3510 | // If we are subtracting one from a positive number, there is no carry | |||
| 3511 | // out of the result. | |||
| 3512 | if (Known.isNonNegative()) | |||
| 3513 | return Tmp; | |||
| 3514 | } | |||
| 3515 | ||||
| 3516 | Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); | |||
| 3517 | if (Tmp2 == 1) break; | |||
| 3518 | return std::min(Tmp, Tmp2) - 1; | |||
| 3519 | ||||
| 3520 | case Instruction::Sub: | |||
| 3521 | Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); | |||
| 3522 | if (Tmp2 == 1) break; | |||
| 3523 | ||||
| 3524 | // Handle NEG. | |||
| 3525 | if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0))) | |||
| 3526 | if (CLHS->isNullValue()) { | |||
| 3527 | KnownBits Known(TyBits); | |||
| 3528 | computeKnownBits(U->getOperand(1), Known, Depth + 1, Q); | |||
| 3529 | // If the input is known to be 0 or 1, the output is 0/-1, which is | |||
| 3530 | // all sign bits set. | |||
| 3531 | if ((Known.Zero | 1).isAllOnes()) | |||
| 3532 | return TyBits; | |||
| 3533 | ||||
| 3534 | // If the input is known to be positive (the sign bit is known clear), | |||
| 3535 | // the output of the NEG has the same number of sign bits as the | |||
| 3536 | // input. | |||
| 3537 | if (Known.isNonNegative()) | |||
| 3538 | return Tmp2; | |||
| 3539 | ||||
| 3540 | // Otherwise, we treat this like a SUB. | |||
| 3541 | } | |||
| 3542 | ||||
| 3543 | // Sub can have at most one carry bit. Thus we know that the output | |||
| 3544 | // is, at worst, one more bit than the inputs. | |||
| 3545 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
| 3546 | if (Tmp == 1) break; | |||
| 3547 | return std::min(Tmp, Tmp2) - 1; | |||
| 3548 | ||||
| 3549 | case Instruction::Mul: { | |||
| 3550 | // The output of the Mul can be at most twice the valid bits in the | |||
| 3551 | // inputs. | |||
| 3552 | unsigned SignBitsOp0 = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
| 3553 | if (SignBitsOp0 == 1) break; | |||
| 3554 | unsigned SignBitsOp1 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q); | |||
| 3555 | if (SignBitsOp1 == 1) break; | |||
| 3556 | unsigned OutValidBits = | |||
| 3557 | (TyBits - SignBitsOp0 + 1) + (TyBits - SignBitsOp1 + 1); | |||
| 3558 | return OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1; | |||
| 3559 | } | |||
| 3560 | ||||
| 3561 | case Instruction::PHI: { | |||
| 3562 | const PHINode *PN = cast<PHINode>(U); | |||
| 3563 | unsigned NumIncomingValues = PN->getNumIncomingValues(); | |||
| 3564 | // Don't analyze large in-degree PHIs. | |||
| 3565 | if (NumIncomingValues > 4) break; | |||
| 3566 | // Unreachable blocks may have zero-operand PHI nodes. | |||
| 3567 | if (NumIncomingValues == 0) break; | |||
| 3568 | ||||
| 3569 | // Take the minimum of all incoming values. This can't infinitely loop | |||
| 3570 | // because of our depth threshold. | |||
| 3571 | Query RecQ = Q; | |||
| 3572 | Tmp = TyBits; | |||
| 3573 | for (unsigned i = 0, e = NumIncomingValues; i != e; ++i) { | |||
| 3574 | if (Tmp == 1) return Tmp; | |||
| 3575 | RecQ.CxtI = PN->getIncomingBlock(i)->getTerminator(); | |||
| 3576 | Tmp = std::min( | |||
| 3577 | Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, RecQ)); | |||
| 3578 | } | |||
| 3579 | return Tmp; | |||
| 3580 | } | |||
| 3581 | ||||
| 3582 | case Instruction::Trunc: { | |||
| 3583 | // If the input contained enough sign bits that some remain after the | |||
| 3584 | // truncation, then we can make use of that. Otherwise we don't know | |||
| 3585 | // anything. | |||
| 3586 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
| 3587 | unsigned OperandTyBits = U->getOperand(0)->getType()->getScalarSizeInBits(); | |||
| 3588 | if (Tmp > (OperandTyBits - TyBits)) | |||
| 3589 | return Tmp - (OperandTyBits - TyBits); | |||
| 3590 | ||||
| 3591 | return 1; | |||
| 3592 | } | |||
| 3593 | ||||
| 3594 | case Instruction::ExtractElement: | |||
| 3595 | // Look through extract element. At the moment we keep this simple and | |||
| 3596 | // skip tracking the specific element. But at least we might find | |||
| 3597 | // information valid for all elements of the vector (for example if vector | |||
| 3598 | // is sign extended, shifted, etc). | |||
| 3599 | return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
| 3600 | ||||
| 3601 | case Instruction::ShuffleVector: { | |||
| 3602 | // Collect the minimum number of sign bits that are shared by every vector | |||
| 3603 | // element referenced by the shuffle. | |||
| 3604 | auto *Shuf = dyn_cast<ShuffleVectorInst>(U); | |||
| 3605 | if (!Shuf) { | |||
| 3606 | // FIXME: Add support for shufflevector constant expressions. | |||
| 3607 | return 1; | |||
| 3608 | } | |||
| 3609 | APInt DemandedLHS, DemandedRHS; | |||
| 3610 | // For undef elements, we don't know anything about the common state of | |||
| 3611 | // the shuffle result. | |||
| 3612 | if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) | |||
| 3613 | return 1; | |||
| 3614 | Tmp = std::numeric_limits<unsigned>::max(); | |||
| 3615 | if (!!DemandedLHS) { | |||
| 3616 | const Value *LHS = Shuf->getOperand(0); | |||
| 3617 | Tmp = ComputeNumSignBits(LHS, DemandedLHS, Depth + 1, Q); | |||
| 3618 | } | |||
| 3619 | // If we don't know anything, early out and try computeKnownBits | |||
| 3620 | // fall-back. | |||
| 3621 | if (Tmp == 1) | |||
| 3622 | break; | |||
| 3623 | if (!!DemandedRHS) { | |||
| 3624 | const Value *RHS = Shuf->getOperand(1); | |||
| 3625 | Tmp2 = ComputeNumSignBits(RHS, DemandedRHS, Depth + 1, Q); | |||
| 3626 | Tmp = std::min(Tmp, Tmp2); | |||
| 3627 | } | |||
| 3628 | // If we don't know anything, early out and try computeKnownBits | |||
| 3629 | // fall-back. | |||
| 3630 | if (Tmp == 1) | |||
| 3631 | break; | |||
| 3632 | assert(Tmp <= TyBits && "Failed to determine minimum sign bits")(static_cast <bool> (Tmp <= TyBits && "Failed to determine minimum sign bits" ) ? void (0) : __assert_fail ("Tmp <= TyBits && \"Failed to determine minimum sign bits\"" , "llvm/lib/Analysis/ValueTracking.cpp", 3632, __extension__ __PRETTY_FUNCTION__ )); | |||
| 3633 | return Tmp; | |||
| 3634 | } | |||
| 3635 | case Instruction::Call: { | |||
| 3636 | if (const auto *II = dyn_cast<IntrinsicInst>(U)) { | |||
| 3637 | switch (II->getIntrinsicID()) { | |||
| 3638 | default: break; | |||
| 3639 | case Intrinsic::abs: | |||
| 3640 | Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q); | |||
| 3641 | if (Tmp == 1) break; | |||
| 3642 | ||||
| 3643 | // Absolute value reduces number of sign bits by at most 1. | |||
| 3644 | return Tmp - 1; | |||
| 3645 | case Intrinsic::smin: | |||
| 3646 | case Intrinsic::smax: { | |||
| 3647 | const APInt *CLow, *CHigh; | |||
| 3648 | if (isSignedMinMaxIntrinsicClamp(II, CLow, CHigh)) | |||
| 3649 | return std::min(CLow->getNumSignBits(), CHigh->getNumSignBits()); | |||
| 3650 | } | |||
| 3651 | } | |||
| 3652 | } | |||
| 3653 | } | |||
| 3654 | } | |||
| 3655 | } | |||
| 3656 | ||||
| 3657 | // Finally, if we can prove that the top bits of the result are 0's or 1's, | |||
| 3658 | // use this information. | |||
| 3659 | ||||
| 3660 | // If we can examine all elements of a vector constant successfully, we're | |||
| 3661 | // done (we can't do any better than that). If not, keep trying. | |||
| 3662 | if (unsigned VecSignBits = | |||
| 3663 | computeNumSignBitsVectorConstant(V, DemandedElts, TyBits)) | |||
| 3664 | return VecSignBits; | |||
| 3665 | ||||
| 3666 | KnownBits Known(TyBits); | |||
| 3667 | computeKnownBits(V, DemandedElts, Known, Depth, Q); | |||
| 3668 | ||||
| 3669 | // If we know that the sign bit is either zero or one, determine the number of | |||
| 3670 | // identical bits in the top of the input value. | |||
| 3671 | return std::max(FirstAnswer, Known.countMinSignBits()); | |||
| 3672 | } | |||
| 3673 | ||||
| 3674 | Intrinsic::ID llvm::getIntrinsicForCallSite(const CallBase &CB, | |||
| 3675 | const TargetLibraryInfo *TLI) { | |||
| 3676 | const Function *F = CB.getCalledFunction(); | |||
| 3677 | if (!F) | |||
| 3678 | return Intrinsic::not_intrinsic; | |||
| 3679 | ||||
| 3680 | if (F->isIntrinsic()) | |||
| 3681 | return F->getIntrinsicID(); | |||
| 3682 | ||||
| 3683 | // We are going to infer semantics of a library function based on mapping it | |||
| 3684 | // to an LLVM intrinsic. Check that the library function is available from | |||
| 3685 | // this callbase and in this environment. | |||
| 3686 | LibFunc Func; | |||
| 3687 | if (F->hasLocalLinkage() || !TLI || !TLI->getLibFunc(CB, Func) || | |||
| 3688 | !CB.onlyReadsMemory()) | |||
| 3689 | return Intrinsic::not_intrinsic; | |||
| 3690 | ||||
| 3691 | switch (Func) { | |||
| 3692 | default: | |||
| 3693 | break; | |||
| 3694 | case LibFunc_sin: | |||
| 3695 | case LibFunc_sinf: | |||
| 3696 | case LibFunc_sinl: | |||
| 3697 | return Intrinsic::sin; | |||
| 3698 | case LibFunc_cos: | |||
| 3699 | case LibFunc_cosf: | |||
| 3700 | case LibFunc_cosl: | |||
| 3701 | return Intrinsic::cos; | |||
| 3702 | case LibFunc_exp: | |||
| 3703 | case LibFunc_expf: | |||
| 3704 | case LibFunc_expl: | |||
| 3705 | return Intrinsic::exp; | |||
| 3706 | case LibFunc_exp2: | |||
| 3707 | case LibFunc_exp2f: | |||
| 3708 | case LibFunc_exp2l: | |||
| 3709 | return Intrinsic::exp2; | |||
| 3710 | case LibFunc_log: | |||
| 3711 | case LibFunc_logf: | |||
| 3712 | case LibFunc_logl: | |||
| 3713 | return Intrinsic::log; | |||
| 3714 | case LibFunc_log10: | |||
| 3715 | case LibFunc_log10f: | |||
| 3716 | case LibFunc_log10l: | |||
| 3717 | return Intrinsic::log10; | |||
| 3718 | case LibFunc_log2: | |||
| 3719 | case LibFunc_log2f: | |||
| 3720 | case LibFunc_log2l: | |||
| 3721 | return Intrinsic::log2; | |||
| 3722 | case LibFunc_fabs: | |||
| 3723 | case LibFunc_fabsf: | |||
| 3724 | case LibFunc_fabsl: | |||
| 3725 | return Intrinsic::fabs; | |||
| 3726 | case LibFunc_fmin: | |||
| 3727 | case LibFunc_fminf: | |||
| 3728 | case LibFunc_fminl: | |||
| 3729 | return Intrinsic::minnum; | |||
| 3730 | case LibFunc_fmax: | |||
| 3731 | case LibFunc_fmaxf: | |||
| 3732 | case LibFunc_fmaxl: | |||
| 3733 | return Intrinsic::maxnum; | |||
| 3734 | case LibFunc_copysign: | |||
| 3735 | case LibFunc_copysignf: | |||
| 3736 | case LibFunc_copysignl: | |||
| 3737 | return Intrinsic::copysign; | |||
| 3738 | case LibFunc_floor: | |||
| 3739 | case LibFunc_floorf: | |||
| 3740 | case LibFunc_floorl: | |||
| 3741 | return Intrinsic::floor; | |||
| 3742 | case LibFunc_ceil: | |||
| 3743 | case LibFunc_ceilf: | |||
| 3744 | case LibFunc_ceill: | |||
| 3745 | return Intrinsic::ceil; | |||
| 3746 | case LibFunc_trunc: | |||
| 3747 | case LibFunc_truncf: | |||
| 3748 | case LibFunc_truncl: | |||
| 3749 | return Intrinsic::trunc; | |||
| 3750 | case LibFunc_rint: | |||
| 3751 | case LibFunc_rintf: | |||
| 3752 | case LibFunc_rintl: | |||
| 3753 | return Intrinsic::rint; | |||
| 3754 | case LibFunc_nearbyint: | |||
| 3755 | case LibFunc_nearbyintf: | |||
| 3756 | case LibFunc_nearbyintl: | |||
| 3757 | return Intrinsic::nearbyint; | |||
| 3758 | case LibFunc_round: | |||
| 3759 | case LibFunc_roundf: | |||
| 3760 | case LibFunc_roundl: | |||
| 3761 | return Intrinsic::round; | |||
| 3762 | case LibFunc_roundeven: | |||
| 3763 | case LibFunc_roundevenf: | |||
| 3764 | case LibFunc_roundevenl: | |||
| 3765 | return Intrinsic::roundeven; | |||
| 3766 | case LibFunc_pow: | |||
| 3767 | case LibFunc_powf: | |||
| 3768 | case LibFunc_powl: | |||
| 3769 | return Intrinsic::pow; | |||
| 3770 | case LibFunc_sqrt: | |||
| 3771 | case LibFunc_sqrtf: | |||
| 3772 | case LibFunc_sqrtl: | |||
| 3773 | return Intrinsic::sqrt; | |||
| 3774 | } | |||
| 3775 | ||||
| 3776 | return Intrinsic::not_intrinsic; | |||
| 3777 | } | |||
| 3778 | ||||
| 3779 | /// Return true if we can prove that the specified FP value is never equal to | |||
| 3780 | /// -0.0. | |||
| 3781 | /// NOTE: Do not check 'nsz' here because that fast-math-flag does not guarantee | |||
| 3782 | /// that a value is not -0.0. It only guarantees that -0.0 may be treated | |||
| 3783 | /// the same as +0.0 in floating-point ops. | |||
| 3784 | bool llvm::CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI, | |||
| 3785 | unsigned Depth) { | |||
| 3786 | if (auto *CFP = dyn_cast<ConstantFP>(V)) | |||
| 3787 | return !CFP->getValueAPF().isNegZero(); | |||
| 3788 | ||||
| 3789 | if (Depth == MaxAnalysisRecursionDepth) | |||
| 3790 | return false; | |||
| 3791 | ||||
| 3792 | auto *Op = dyn_cast<Operator>(V); | |||
| 3793 | if (!Op) | |||
| 3794 | return false; | |||
| 3795 | ||||
| 3796 | // (fadd x, 0.0) is guaranteed to return +0.0, not -0.0. | |||
| 3797 | if (match(Op, m_FAdd(m_Value(), m_PosZeroFP()))) | |||
| 3798 | return true; | |||
| 3799 | ||||
| 3800 | // sitofp and uitofp turn into +0.0 for zero. | |||
| 3801 | if (isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op)) | |||
| 3802 | return true; | |||
| 3803 | ||||
| 3804 | if (auto *Call = dyn_cast<CallInst>(Op)) { | |||
| 3805 | Intrinsic::ID IID = getIntrinsicForCallSite(*Call, TLI); | |||
| 3806 | switch (IID) { | |||
| 3807 | default: | |||
| 3808 | break; | |||
| 3809 | // sqrt(-0.0) = -0.0, no other negative results are possible. | |||
| 3810 | case Intrinsic::sqrt: | |||
| 3811 | case Intrinsic::canonicalize: | |||
| 3812 | return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1); | |||
| 3813 | case Intrinsic::experimental_constrained_sqrt: { | |||
| 3814 | // NOTE: This rounding mode restriction may be too strict. | |||
| 3815 | const auto *CI = cast<ConstrainedFPIntrinsic>(Call); | |||
| 3816 | if (CI->getRoundingMode() == RoundingMode::NearestTiesToEven) | |||
| 3817 | return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1); | |||
| 3818 | else | |||
| 3819 | return false; | |||
| 3820 | } | |||
| 3821 | // fabs(x) != -0.0 | |||
| 3822 | case Intrinsic::fabs: | |||
| 3823 | return true; | |||
| 3824 | // sitofp and uitofp turn into +0.0 for zero. | |||
| 3825 | case Intrinsic::experimental_constrained_sitofp: | |||
| 3826 | case Intrinsic::experimental_constrained_uitofp: | |||
| 3827 | return true; | |||
| 3828 | } | |||
| 3829 | } | |||
| 3830 | ||||
| 3831 | return false; | |||
| 3832 | } | |||
| 3833 | ||||
| 3834 | /// If \p SignBitOnly is true, test for a known 0 sign bit rather than a | |||
| 3835 | /// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign | |||
| 3836 | /// bit despite comparing equal. | |||
| 3837 | static bool cannotBeOrderedLessThanZeroImpl(const Value *V, | |||
| 3838 | const TargetLibraryInfo *TLI, | |||
| 3839 | bool SignBitOnly, | |||
| 3840 | unsigned Depth) { | |||
| 3841 | // TODO: This function does not do the right thing when SignBitOnly is true | |||
| 3842 | // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform | |||
| 3843 | // which flips the sign bits of NaNs. See | |||
| 3844 | // https://llvm.org/bugs/show_bug.cgi?id=31702. | |||
| 3845 | ||||
| 3846 | if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { | |||
| 3847 | return !CFP->getValueAPF().isNegative() || | |||
| 3848 | (!SignBitOnly && CFP->getValueAPF().isZero()); | |||
| 3849 | } | |||
| 3850 | ||||
| 3851 | // Handle vector of constants. | |||
| 3852 | if (auto *CV = dyn_cast<Constant>(V)) { | |||
| 3853 | if (auto *CVFVTy = dyn_cast<FixedVectorType>(CV->getType())) { | |||
| 3854 | unsigned NumElts = CVFVTy->getNumElements(); | |||
| 3855 | for (unsigned i = 0; i != NumElts; ++i) { | |||
| 3856 | auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i)); | |||
| 3857 | if (!CFP) | |||
| 3858 | return false; | |||
| 3859 | if (CFP->getValueAPF().isNegative() && | |||
| 3860 | (SignBitOnly || !CFP->getValueAPF().isZero())) | |||
| 3861 | return false; | |||
| 3862 | } | |||
| 3863 | ||||
| 3864 | // All non-negative ConstantFPs. | |||
| 3865 | return true; | |||
| 3866 | } | |||
| 3867 | } | |||
| 3868 | ||||
| 3869 | if (Depth == MaxAnalysisRecursionDepth) | |||
| 3870 | return false; | |||
| 3871 | ||||
| 3872 | const Operator *I = dyn_cast<Operator>(V); | |||
| 3873 | if (!I) | |||
| 3874 | return false; | |||
| 3875 | ||||
| 3876 | switch (I->getOpcode()) { | |||
| 3877 | default: | |||
| 3878 | break; | |||
| 3879 | // Unsigned integers are always nonnegative. | |||
| 3880 | case Instruction::UIToFP: | |||
| 3881 | return true; | |||
| 3882 | case Instruction::FDiv: | |||
| 3883 | // X / X is always exactly 1.0 or a NaN. | |||
| 3884 | if (I->getOperand(0) == I->getOperand(1) && | |||
| 3885 | (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs())) | |||
| 3886 | return true; | |||
| 3887 | ||||
| 3888 | // Set SignBitOnly for RHS, because X / -0.0 is -Inf (or NaN). | |||
| 3889 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | |||
| 3890 | Depth + 1) && | |||
| 3891 | cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, | |||
| 3892 | /*SignBitOnly*/ true, Depth + 1); | |||
| 3893 | case Instruction::FMul: | |||
| 3894 | // X * X is always non-negative or a NaN. | |||
| 3895 | if (I->getOperand(0) == I->getOperand(1) && | |||
| 3896 | (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs())) | |||
| 3897 | return true; | |||
| 3898 | ||||
| 3899 | [[fallthrough]]; | |||
| 3900 | case Instruction::FAdd: | |||
| 3901 | case Instruction::FRem: | |||
| 3902 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | |||
| 3903 | Depth + 1) && | |||
| 3904 | cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, | |||
| 3905 | Depth + 1); | |||
| 3906 | case Instruction::Select: | |||
| 3907 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, | |||
| 3908 | Depth + 1) && | |||
| 3909 | cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly, | |||
| 3910 | Depth + 1); | |||
| 3911 | case Instruction::FPExt: | |||
| 3912 | case Instruction::FPTrunc: | |||
| 3913 | // Widening/narrowing never change sign. | |||
| 3914 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | |||
| 3915 | Depth + 1); | |||
| 3916 | case Instruction::ExtractElement: | |||
| 3917 | // Look through extract element. At the moment we keep this simple and skip | |||
| 3918 | // tracking the specific element. But at least we might find information | |||
| 3919 | // valid for all elements of the vector. | |||
| 3920 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | |||
| 3921 | Depth + 1); | |||
| 3922 | case Instruction::Call: | |||
| 3923 | const auto *CI = cast<CallInst>(I); | |||
| 3924 | Intrinsic::ID IID = getIntrinsicForCallSite(*CI, TLI); | |||
| 3925 | switch (IID) { | |||
| 3926 | default: | |||
| 3927 | break; | |||
| 3928 | case Intrinsic::canonicalize: | |||
| 3929 | case Intrinsic::arithmetic_fence: | |||
| 3930 | case Intrinsic::floor: | |||
| 3931 | case Intrinsic::ceil: | |||
| 3932 | case Intrinsic::trunc: | |||
| 3933 | case Intrinsic::rint: | |||
| 3934 | case Intrinsic::nearbyint: | |||
| 3935 | case Intrinsic::round: | |||
| 3936 | case Intrinsic::roundeven: | |||
| 3937 | case Intrinsic::fptrunc_round: | |||
| 3938 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, Depth + 1); | |||
| 3939 | case Intrinsic::maxnum: { | |||
| 3940 | Value *V0 = I->getOperand(0), *V1 = I->getOperand(1); | |||
| 3941 | auto isPositiveNum = [&](Value *V) { | |||
| 3942 | if (SignBitOnly) { | |||
| 3943 | // With SignBitOnly, this is tricky because the result of | |||
| 3944 | // maxnum(+0.0, -0.0) is unspecified. Just check if the operand is | |||
| 3945 | // a constant strictly greater than 0.0. | |||
| 3946 | const APFloat *C; | |||
| 3947 | return match(V, m_APFloat(C)) && | |||
| 3948 | *C > APFloat::getZero(C->getSemantics()); | |||
| 3949 | } | |||
| 3950 | ||||
| 3951 | // -0.0 compares equal to 0.0, so if this operand is at least -0.0, | |||
| 3952 | // maxnum can't be ordered-less-than-zero. | |||
| 3953 | return isKnownNeverNaN(V, TLI) && | |||
| 3954 | cannotBeOrderedLessThanZeroImpl(V, TLI, false, Depth + 1); | |||
| 3955 | }; | |||
| 3956 | ||||
| 3957 | // TODO: This could be improved. We could also check that neither operand | |||
| 3958 | // has its sign bit set (and at least 1 is not-NAN?). | |||
| 3959 | return isPositiveNum(V0) || isPositiveNum(V1); | |||
| 3960 | } | |||
| 3961 | ||||
| 3962 | case Intrinsic::maximum: | |||
| 3963 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | |||
| 3964 | Depth + 1) || | |||
| 3965 | cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, | |||
| 3966 | Depth + 1); | |||
| 3967 | case Intrinsic::minnum: | |||
| 3968 | case Intrinsic::minimum: | |||
| 3969 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | |||
| 3970 | Depth + 1) && | |||
| 3971 | cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly, | |||
| 3972 | Depth + 1); | |||
| 3973 | case Intrinsic::exp: | |||
| 3974 | case Intrinsic::exp2: | |||
| 3975 | case Intrinsic::fabs: | |||
| 3976 | return true; | |||
| 3977 | case Intrinsic::copysign: | |||
| 3978 | // Only the sign operand matters. | |||
| 3979 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, true, | |||
| 3980 | Depth + 1); | |||
| 3981 | case Intrinsic::sqrt: | |||
| 3982 | // sqrt(x) is always >= -0 or NaN. Moreover, sqrt(x) == -0 iff x == -0. | |||
| 3983 | if (!SignBitOnly) | |||
| 3984 | return true; | |||
| 3985 | return CI->hasNoNaNs() && (CI->hasNoSignedZeros() || | |||
| 3986 | CannotBeNegativeZero(CI->getOperand(0), TLI)); | |||
| 3987 | ||||
| 3988 | case Intrinsic::powi: | |||
| 3989 | if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) { | |||
| 3990 | // powi(x,n) is non-negative if n is even. | |||
| 3991 | if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0) | |||
| 3992 | return true; | |||
| 3993 | } | |||
| 3994 | // TODO: This is not correct. Given that exp is an integer, here are the | |||
| 3995 | // ways that pow can return a negative value: | |||
| 3996 | // | |||
| 3997 | // pow(x, exp) --> negative if exp is odd and x is negative. | |||
| 3998 | // pow(-0, exp) --> -inf if exp is negative odd. | |||
| 3999 | // pow(-0, exp) --> -0 if exp is positive odd. | |||
| 4000 | // pow(-inf, exp) --> -0 if exp is negative odd. | |||
| 4001 | // pow(-inf, exp) --> -inf if exp is positive odd. | |||
| 4002 | // | |||
| 4003 | // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN, | |||
| 4004 | // but we must return false if x == -0. Unfortunately we do not currently | |||
| 4005 | // have a way of expressing this constraint. See details in | |||
| 4006 | // https://llvm.org/bugs/show_bug.cgi?id=31702. | |||
| 4007 | return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly, | |||
| 4008 | Depth + 1); | |||
| 4009 | ||||
| 4010 | case Intrinsic::fma: | |||
| 4011 | case Intrinsic::fmuladd: | |||
| 4012 | // x*x+y is non-negative if y is non-negative. | |||
| 4013 | return I->getOperand(0) == I->getOperand(1) && | |||
| 4014 | (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) && | |||
| 4015 | cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly, | |||
| 4016 | Depth + 1); | |||
| 4017 | } | |||
| 4018 | break; | |||
| 4019 | } | |||
| 4020 | return false; | |||
| 4021 | } | |||
| 4022 | ||||
| 4023 | bool llvm::CannotBeOrderedLessThanZero(const Value *V, | |||
| 4024 | const TargetLibraryInfo *TLI) { | |||
| 4025 | return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0); | |||
| 4026 | } | |||
| 4027 | ||||
| 4028 | bool llvm::SignBitMustBeZero(const Value *V, const TargetLibraryInfo *TLI) { | |||
| 4029 | return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0); | |||
| 4030 | } | |||
| 4031 | ||||
| 4032 | bool llvm::isKnownNeverInfinity(const Value *V, const TargetLibraryInfo *TLI, | |||
| 4033 | unsigned Depth) { | |||
| 4034 | assert(V->getType()->isFPOrFPVectorTy() && "Querying for Inf on non-FP type")(static_cast <bool> (V->getType()->isFPOrFPVectorTy () && "Querying for Inf on non-FP type") ? void (0) : __assert_fail ("V->getType()->isFPOrFPVectorTy() && \"Querying for Inf on non-FP type\"" , "llvm/lib/Analysis/ValueTracking.cpp", 4034, __extension__ __PRETTY_FUNCTION__ )); | |||
| 4035 | ||||
| 4036 | // If we're told that infinities won't happen, assume they won't. | |||
| 4037 | if (auto *FPMathOp = dyn_cast<FPMathOperator>(V)) | |||
| 4038 | if (FPMathOp->hasNoInfs()) | |||
| 4039 | return true; | |||
| 4040 | ||||
| 4041 | if (const auto *Arg = dyn_cast<Argument>(V)) { | |||
| 4042 | if ((Arg->getNoFPClass() & fcInf) == fcInf) | |||
| 4043 | return true; | |||
| 4044 | } | |||
| 4045 | ||||
| 4046 | // TODO: Use fpclass like API for isKnown queries and distinguish +inf from | |||
| 4047 | // -inf. | |||
| 4048 | if (const auto *CB = dyn_cast<CallBase>(V)) { | |||
| 4049 | if ((CB->getRetNoFPClass() & fcInf) == fcInf) | |||
| 4050 | return true; | |||
| 4051 | } | |||
| 4052 | ||||
| 4053 | // Handle scalar constants. | |||
| 4054 | if (auto *CFP = dyn_cast<ConstantFP>(V)) | |||
| 4055 | return !CFP->isInfinity(); | |||
| 4056 | ||||
| 4057 | if (Depth == MaxAnalysisRecursionDepth) | |||
| 4058 | return false; | |||
| 4059 | ||||
| 4060 | if (auto *Inst = dyn_cast<Instruction>(V)) { | |||
| 4061 | switch (Inst->getOpcode()) { | |||
| 4062 | case Instruction::Select: { | |||
| 4063 | return isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1) && | |||
| 4064 | isKnownNeverInfinity(Inst->getOperand(2), TLI, Depth + 1); | |||
| 4065 | } | |||
| 4066 | case Instruction::SIToFP: | |||
| 4067 | case Instruction::UIToFP: { | |||
| 4068 | // Get width of largest magnitude integer (remove a bit if signed). | |||
| 4069 | // This still works for a signed minimum value because the largest FP | |||
| 4070 | // value is scaled by some fraction close to 2.0 (1.0 + 0.xxxx). | |||
| 4071 | int IntSize = Inst->getOperand(0)->getType()->getScalarSizeInBits(); | |||
| 4072 | if (Inst->getOpcode() == Instruction::SIToFP) | |||
| 4073 | --IntSize; | |||
| 4074 | ||||
| 4075 | // If the exponent of the largest finite FP value can hold the largest | |||
| 4076 | // integer, the result of the cast must be finite. | |||
| 4077 | Type *FPTy = Inst->getType()->getScalarType(); | |||
| 4078 | return ilogb(APFloat::getLargest(FPTy->getFltSemantics())) >= IntSize; | |||
| 4079 | } | |||
| 4080 | case Instruction::FNeg: | |||
| 4081 | case Instruction::FPExt: { | |||
| 4082 | // Peek through to source op. If it is not infinity, this is not infinity. | |||
| 4083 | return isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1); | |||
| 4084 | } | |||
| 4085 | case Instruction::FPTrunc: { | |||
| 4086 | // Need a range check. | |||
| 4087 | return false; | |||
| 4088 | } | |||
| 4089 | default: | |||
| 4090 | break; | |||
| 4091 | } | |||
| 4092 | ||||
| 4093 | if (const auto *II = dyn_cast<IntrinsicInst>(V)) { | |||
| 4094 | switch (II->getIntrinsicID()) { | |||
| 4095 | case Intrinsic::sin: | |||
| 4096 | case Intrinsic::cos: | |||
| 4097 | // Return NaN on infinite inputs. | |||
| 4098 | return true; | |||
| 4099 | case Intrinsic::fabs: | |||
| 4100 | case Intrinsic::sqrt: | |||
| 4101 | case Intrinsic::canonicalize: | |||
| 4102 | case Intrinsic::copysign: | |||
| 4103 | case Intrinsic::arithmetic_fence: | |||
| 4104 | case Intrinsic::trunc: | |||
| 4105 | return isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1); | |||
| 4106 | case Intrinsic::floor: | |||
| 4107 | case Intrinsic::ceil: | |||
| 4108 | case Intrinsic::rint: | |||
| 4109 | case Intrinsic::nearbyint: | |||
| 4110 | case Intrinsic::round: | |||
| 4111 | case Intrinsic::roundeven: | |||
| 4112 | // PPC_FP128 is a special case. | |||
| 4113 | if (V->getType()->isMultiUnitFPType()) | |||
| 4114 | return false; | |||
| 4115 | return isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1); | |||
| 4116 | case Intrinsic::fptrunc_round: | |||
| 4117 | // Requires knowing the value range. | |||
| 4118 | return false; | |||
| 4119 | case Intrinsic::minnum: | |||
| 4120 | case Intrinsic::maxnum: | |||
| 4121 | case Intrinsic::minimum: | |||
| 4122 | case Intrinsic::maximum: | |||
| 4123 | return isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) && | |||
| 4124 | isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1); | |||
| 4125 | case Intrinsic::log: | |||
| 4126 | case Intrinsic::log10: | |||
| 4127 | case Intrinsic::log2: | |||
| 4128 | // log(+inf) -> +inf | |||
| 4129 | // log([+-]0.0) -> -inf | |||
| 4130 | // log(-inf) -> nan | |||
| 4131 | // log(-x) -> nan | |||
| 4132 | // TODO: We lack API to check the == 0 case. | |||
| 4133 | return false; | |||
| 4134 | case Intrinsic::exp: | |||
| 4135 | case Intrinsic::exp2: | |||
| 4136 | case Intrinsic::pow: | |||
| 4137 | case Intrinsic::powi: | |||
| 4138 | case Intrinsic::fma: | |||
| 4139 | case Intrinsic::fmuladd: | |||
| 4140 | // These can return infinities on overflow cases, so it's hard to prove | |||
| 4141 | // anything about it. | |||
| 4142 | return false; | |||
| 4143 | default: | |||
| 4144 | break; | |||
| 4145 | } | |||
| 4146 | } | |||
| 4147 | } | |||
| 4148 | ||||
| 4149 | // try to handle fixed width vector constants | |||
| 4150 | auto *VFVTy = dyn_cast<FixedVectorType>(V->getType()); | |||
| 4151 | if (VFVTy && isa<Constant>(V)) { | |||
| 4152 | // For vectors, verify that each element is not infinity. | |||
| 4153 | unsigned NumElts = VFVTy->getNumElements(); | |||
| 4154 | for (unsigned i = 0; i != NumElts; ++i) { | |||
| 4155 | Constant *Elt = cast<Constant>(V)->getAggregateElement(i); | |||
| 4156 | if (!Elt) | |||
| 4157 | return false; | |||
| 4158 | if (isa<UndefValue>(Elt)) | |||
| 4159 | continue; | |||
| 4160 | auto *CElt = dyn_cast<ConstantFP>(Elt); | |||
| 4161 | if (!CElt || CElt->isInfinity()) | |||
| 4162 | return false; | |||
| 4163 | } | |||
| 4164 | // All elements were confirmed non-infinity or undefined. | |||
| 4165 | return true; | |||
| 4166 | } | |||
| 4167 | ||||
| 4168 | // was not able to prove that V never contains infinity | |||
| 4169 | return false; | |||
| 4170 | } | |||
| 4171 | ||||
| 4172 | bool llvm::isKnownNeverNaN(const Value *V, const TargetLibraryInfo *TLI, | |||
| 4173 | unsigned Depth) { | |||
| 4174 | assert(V->getType()->isFPOrFPVectorTy() && "Querying for NaN on non-FP type")(static_cast <bool> (V->getType()->isFPOrFPVectorTy () && "Querying for NaN on non-FP type") ? void (0) : __assert_fail ("V->getType()->isFPOrFPVectorTy() && \"Querying for NaN on non-FP type\"" , "llvm/lib/Analysis/ValueTracking.cpp", 4174, __extension__ __PRETTY_FUNCTION__ )); | |||
| 4175 | ||||
| 4176 | // If we're told that NaNs won't happen, assume they won't. | |||
| 4177 | if (auto *FPMathOp = dyn_cast<FPMathOperator>(V)) | |||
| 4178 | if (FPMathOp->hasNoNaNs()) | |||
| 4179 | return true; | |||
| 4180 | ||||
| 4181 | if (const auto *Arg = dyn_cast<Argument>(V)) { | |||
| 4182 | if ((Arg->getNoFPClass() & fcNan) == fcNan) | |||
| 4183 | return true; | |||
| 4184 | } | |||
| 4185 | ||||
| 4186 | // TODO: Use fpclass like API for isKnown queries and distinguish snan from | |||
| 4187 | // qnan. | |||
| 4188 | if (const auto *CB = dyn_cast<CallBase>(V)) { | |||
| 4189 | FPClassTest Mask = CB->getRetNoFPClass(); | |||
| 4190 | if ((Mask & fcNan) == fcNan) | |||
| 4191 | return true; | |||
| 4192 | } | |||
| 4193 | ||||
| 4194 | // Handle scalar constants. | |||
| 4195 | if (auto *CFP = dyn_cast<ConstantFP>(V)) | |||
| 4196 | return !CFP->isNaN(); | |||
| 4197 | ||||
| 4198 | if (Depth == MaxAnalysisRecursionDepth) | |||
| 4199 | return false; | |||
| 4200 | ||||
| 4201 | if (auto *Inst = dyn_cast<Instruction>(V)) { | |||
| 4202 | switch (Inst->getOpcode()) { | |||
| 4203 | case Instruction::FAdd: | |||
| 4204 | case Instruction::FSub: | |||
| 4205 | // Adding positive and negative infinity produces NaN. | |||
| 4206 | return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) && | |||
| 4207 | isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) && | |||
| 4208 | (isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) || | |||
| 4209 | isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1)); | |||
| 4210 | ||||
| 4211 | case Instruction::FMul: | |||
| 4212 | // Zero multiplied with infinity produces NaN. | |||
| 4213 | // FIXME: If neither side can be zero fmul never produces NaN. | |||
| 4214 | return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1) && | |||
| 4215 | isKnownNeverInfinity(Inst->getOperand(0), TLI, Depth + 1) && | |||
| 4216 | isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) && | |||
| 4217 | isKnownNeverInfinity(Inst->getOperand(1), TLI, Depth + 1); | |||
| 4218 | ||||
| 4219 | case Instruction::FDiv: | |||
| 4220 | case Instruction::FRem: | |||
| 4221 | // FIXME: Only 0/0, Inf/Inf, Inf REM x and x REM 0 produce NaN. | |||
| 4222 | return false; | |||
| 4223 | ||||
| 4224 | case Instruction::Select: { | |||
| 4225 | return isKnownNeverNaN(Inst->getOperand(1), TLI, Depth + 1) && | |||
| 4226 | isKnownNeverNaN(Inst->getOperand(2), TLI, Depth + 1); | |||
| 4227 | } | |||
| 4228 | case Instruction::SIToFP: | |||
| 4229 | case Instruction::UIToFP: | |||
| 4230 | return true; | |||
| 4231 | case Instruction::FPTrunc: | |||
| 4232 | case Instruction::FPExt: | |||
| 4233 | case Instruction::FNeg: | |||
| 4234 | return isKnownNeverNaN(Inst->getOperand(0), TLI, Depth + 1); | |||
| 4235 | default: | |||
| 4236 | break; | |||
| 4237 | } | |||
| 4238 | } | |||
| 4239 | ||||
| 4240 | if (const auto *II = dyn_cast<IntrinsicInst>(V)) { | |||
| 4241 | switch (II->getIntrinsicID()) { | |||
| 4242 | case Intrinsic::canonicalize: | |||
| 4243 | case Intrinsic::fabs: | |||
| 4244 | case Intrinsic::copysign: | |||
| 4245 | case Intrinsic::exp: | |||
| 4246 | case Intrinsic::exp2: | |||
| 4247 | case Intrinsic::floor: | |||
| 4248 | case Intrinsic::ceil: | |||
| 4249 | case Intrinsic::trunc: | |||
| 4250 | case Intrinsic::rint: | |||
| 4251 | case Intrinsic::nearbyint: | |||
| 4252 | case Intrinsic::round: | |||
| 4253 | case Intrinsic::roundeven: | |||
| 4254 | case Intrinsic::arithmetic_fence: | |||
| 4255 | return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1); | |||
| 4256 | case Intrinsic::sqrt: | |||
| 4257 | return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) && | |||
| 4258 | CannotBeOrderedLessThanZero(II->getArgOperand(0), TLI); | |||
| 4259 | case Intrinsic::minnum: | |||
| 4260 | case Intrinsic::maxnum: | |||
| 4261 | // If either operand is not NaN, the result is not NaN. | |||
| 4262 | return isKnownNeverNaN(II->getArgOperand(0), TLI, Depth + 1) || | |||
| 4263 | isKnownNeverNaN(II->getArgOperand(1), TLI, Depth + 1); | |||
| 4264 | default: | |||
| 4265 | return false; | |||
| 4266 | } | |||
| 4267 | } | |||
| 4268 | ||||
| 4269 | // Try to handle fixed width vector constants | |||
| 4270 | auto *VFVTy = dyn_cast<FixedVectorType>(V->getType()); | |||
| 4271 | if (VFVTy && isa<Constant>(V)) { | |||
| 4272 | // For vectors, verify that each element is not NaN. | |||
| 4273 | unsigned NumElts = VFVTy->getNumElements(); | |||
| 4274 | for (unsigned i = 0; i != NumElts; ++i) { | |||
| 4275 | Constant *Elt = cast<Constant>(V)->getAggregateElement(i); | |||
| 4276 | if (!Elt) | |||
| 4277 | return false; | |||
| 4278 | if (isa<UndefValue>(Elt)) | |||
| 4279 | continue; | |||
| 4280 | auto *CElt = dyn_cast<ConstantFP>(Elt); | |||
| 4281 | if (!CElt || CElt->isNaN()) | |||
| 4282 | return false; | |||
| 4283 | } | |||
| 4284 | // All elements were confirmed not-NaN or undefined. | |||
| 4285 | return true; | |||
| 4286 | } | |||
| 4287 | ||||
| 4288 | // Was not able to prove that V never contains NaN | |||
| 4289 | return false; | |||
| 4290 | } | |||
| 4291 | ||||
| 4292 | /// Return true if it's possible to assume IEEE treatment of input denormals in | |||
| 4293 | /// \p F for \p Val. | |||
| 4294 | static bool inputDenormalIsIEEE(const Function &F, const Type *Ty) { | |||
| 4295 | Ty = Ty->getScalarType(); | |||
| 4296 | return F.getDenormalMode(Ty->getFltSemantics()).Input == DenormalMode::IEEE; | |||
| 4297 | } | |||
| 4298 | ||||
| 4299 | bool KnownFPClass::isKnownNeverLogicalZero(const Function &F, Type *Ty) const { | |||
| 4300 | return isKnownNeverZero() && | |||
| 4301 | (isKnownNeverSubnormal() || inputDenormalIsIEEE(F, Ty)); | |||
| 4302 | } | |||
| 4303 | ||||
| 4304 | /// Returns a pair of values, which if passed to llvm.is.fpclass, returns the | |||
| 4305 | /// same result as an fcmp with the given operands. | |||
| 4306 | std::pair<Value *, FPClassTest> llvm::fcmpToClassTest(FCmpInst::Predicate Pred, | |||
| 4307 | const Function &F, | |||
| 4308 | Value *LHS, Value *RHS, | |||
| 4309 | bool LookThroughSrc) { | |||
| 4310 | const APFloat *ConstRHS; | |||
| 4311 | if (!match(RHS, m_APFloat(ConstRHS))) | |||
| 4312 | return {nullptr, fcNone}; | |||
| 4313 | ||||
| 4314 | if (ConstRHS->isZero()) { | |||
| 4315 | // Compares with fcNone are only exactly equal to fcZero if input denormals | |||
| 4316 | // are not flushed. | |||
| 4317 | // TODO: Handle DAZ by expanding masks to cover subnormal cases. | |||
| 4318 | if (Pred != FCmpInst::FCMP_ORD && Pred != FCmpInst::FCMP_UNO && | |||
| 4319 | !inputDenormalIsIEEE(F, LHS->getType())) | |||
| 4320 | return {nullptr, fcNone}; | |||
| 4321 | ||||
| 4322 | switch (Pred) { | |||
| 4323 | case FCmpInst::FCMP_OEQ: // Match x == 0.0 | |||
| 4324 | return {LHS, fcZero}; | |||
| 4325 | case FCmpInst::FCMP_UEQ: // Match isnan(x) || (x == 0.0) | |||
| 4326 | return {LHS, fcZero | fcNan}; | |||
| 4327 | case FCmpInst::FCMP_UNE: // Match (x != 0.0) | |||
| 4328 | return {LHS, ~fcZero}; | |||
| 4329 | case FCmpInst::FCMP_ONE: // Match !isnan(x) && x != 0.0 | |||
| 4330 | return {LHS, ~fcNan & ~fcZero}; | |||
| 4331 | case FCmpInst::FCMP_ORD: | |||
| 4332 | // Canonical form of ord/uno is with a zero. We could also handle | |||
| 4333 | // non-canonical other non-NaN constants or LHS == RHS. | |||
| 4334 | return {LHS, ~fcNan}; | |||
| 4335 | case FCmpInst::FCMP_UNO: | |||
| 4336 | return {LHS, fcNan}; | |||
| 4337 | case FCmpInst::FCMP_OGT: // x > 0 | |||
| 4338 | return {LHS, fcPosSubnormal | fcPosNormal | fcPosInf}; | |||
| 4339 | case FCmpInst::FCMP_UGT: // isnan(x) || x > 0 | |||
| 4340 | return {LHS, fcPosSubnormal | fcPosNormal | fcPosInf | fcNan}; | |||
| 4341 | case FCmpInst::FCMP_OGE: // x >= 0 | |||
| 4342 | return {LHS, fcPositive | fcNegZero}; | |||
| 4343 | case FCmpInst::FCMP_UGE: // isnan(x) || x >= 0 | |||
| 4344 | return {LHS, fcPositive | fcNegZero | fcNan}; | |||
| 4345 | case FCmpInst::FCMP_OLT: // x < 0 | |||
| 4346 | return {LHS, fcNegSubnormal | fcNegNormal | fcNegInf}; | |||
| 4347 | case FCmpInst::FCMP_ULT: // isnan(x) || x < 0 | |||
| 4348 | return {LHS, fcNegSubnormal | fcNegNormal | fcNegInf | fcNan}; | |||
| 4349 | case FCmpInst::FCMP_OLE: // x <= 0 | |||
| 4350 | return {LHS, fcNegative | fcPosZero}; | |||
| 4351 | case FCmpInst::FCMP_ULE: // isnan(x) || x <= 0 | |||
| 4352 | return {LHS, fcNegative | fcPosZero | fcNan}; | |||
| 4353 | default: | |||
| 4354 | break; | |||
| 4355 | } | |||
| 4356 | ||||
| 4357 | return {nullptr, fcNone}; | |||
| 4358 | } | |||
| 4359 | ||||
| 4360 | Value *Src = LHS; | |||
| 4361 | const bool IsFabs = LookThroughSrc && match(LHS, m_FAbs(m_Value(Src))); | |||
| 4362 | ||||
| 4363 | // Compute the test mask that would return true for the ordered comparisons. | |||
| 4364 | FPClassTest Mask; | |||
| 4365 | ||||
| 4366 | if (ConstRHS->isInfinity()) { | |||
| 4367 | switch (Pred) { | |||
| 4368 | case FCmpInst::FCMP_OEQ: | |||
| 4369 | case FCmpInst::FCMP_UNE: { | |||
| 4370 | // Match __builtin_isinf patterns | |||
| 4371 | // | |||
| 4372 | // fcmp oeq x, +inf -> is_fpclass x, fcPosInf | |||
| 4373 | // fcmp oeq fabs(x), +inf -> is_fpclass x, fcInf | |||
| 4374 | // fcmp oeq x, -inf -> is_fpclass x, fcNegInf | |||
| 4375 | // fcmp oeq fabs(x), -inf -> is_fpclass x, 0 -> false | |||
| 4376 | // | |||
| 4377 | // fcmp une x, +inf -> is_fpclass x, ~fcPosInf | |||
| 4378 | // fcmp une fabs(x), +inf -> is_fpclass x, ~fcInf | |||
| 4379 | // fcmp une x, -inf -> is_fpclass x, ~fcNegInf | |||
| 4380 | // fcmp une fabs(x), -inf -> is_fpclass x, fcAllFlags -> true | |||
| 4381 | ||||
| 4382 | if (ConstRHS->isNegative()) { | |||
| 4383 | Mask = fcNegInf; | |||
| 4384 | if (IsFabs) | |||
| 4385 | Mask = fcNone; | |||
| 4386 | } else { | |||
| 4387 | Mask = fcPosInf; | |||
| 4388 | if (IsFabs) | |||
| 4389 | Mask |= fcNegInf; | |||
| 4390 | } | |||
| 4391 | ||||
| 4392 | break; | |||
| 4393 | } | |||
| 4394 | case FCmpInst::FCMP_ONE: | |||
| 4395 | case FCmpInst::FCMP_UEQ: { | |||
| 4396 | // Match __builtin_isinf patterns | |||
| 4397 | // fcmp one x, -inf -> is_fpclass x, fcNegInf | |||
| 4398 | // fcmp one fabs(x), -inf -> is_fpclass x, ~fcNegInf & ~fcNan | |||
| 4399 | // fcmp one x, +inf -> is_fpclass x, ~fcNegInf & ~fcNan | |||
| 4400 | // fcmp one fabs(x), +inf -> is_fpclass x, ~fcInf & fcNan | |||
| 4401 | // | |||
| 4402 | // fcmp ueq x, +inf -> is_fpclass x, fcPosInf|fcNan | |||
| 4403 | // fcmp ueq (fabs x), +inf -> is_fpclass x, fcInf|fcNan | |||
| 4404 | // fcmp ueq x, -inf -> is_fpclass x, fcNegInf|fcNan | |||
| 4405 | // fcmp ueq fabs(x), -inf -> is_fpclass x, fcNan | |||
| 4406 | if (ConstRHS->isNegative()) { | |||
| 4407 | Mask = ~fcNegInf & ~fcNan; | |||
| 4408 | if (IsFabs) | |||
| 4409 | Mask = ~fcNan; | |||
| 4410 | } else { | |||
| 4411 | Mask = ~fcPosInf & ~fcNan; | |||
| 4412 | if (IsFabs) | |||
| 4413 | Mask &= ~fcNegInf; | |||
| 4414 | } | |||
| 4415 | ||||
| 4416 | break; | |||
| 4417 | } | |||
| 4418 | case FCmpInst::FCMP_OLT: | |||
| 4419 | case FCmpInst::FCMP_UGE: { | |||
| 4420 | if (ConstRHS->isNegative()) // TODO | |||
| 4421 | return {nullptr, fcNone}; | |||
| 4422 | ||||
| 4423 | // fcmp olt fabs(x), +inf -> fcFinite | |||
| 4424 | // fcmp uge fabs(x), +inf -> ~fcFinite | |||
| 4425 | // fcmp olt x, +inf -> fcFinite|fcNegInf | |||
| 4426 | // fcmp uge x, +inf -> ~(fcFinite|fcNegInf) | |||
| 4427 | Mask = fcFinite; | |||
| 4428 | if (!IsFabs) | |||
| 4429 | Mask |= fcNegInf; | |||
| 4430 | break; | |||
| 4431 | } | |||
| 4432 | case FCmpInst::FCMP_OGE: | |||
| 4433 | case FCmpInst::FCMP_ULT: { | |||
| 4434 | if (ConstRHS->isNegative()) // TODO | |||
| 4435 | return {nullptr, fcNone}; | |||
| 4436 | ||||
| 4437 | // fcmp oge fabs(x), +inf -> fcInf | |||
| 4438 | // fcmp oge x, +inf -> fcPosInf | |||
| 4439 | // fcmp ult fabs(x), +inf -> ~fcInf | |||
| 4440 | // fcmp ult x, +inf -> ~fcPosInf | |||
| 4441 | Mask = fcPosInf; | |||
| 4442 | if (IsFabs) | |||
| 4443 | Mask |= fcNegInf; | |||
| 4444 | break; | |||
| 4445 | } | |||
| 4446 | default: | |||
| 4447 | return {nullptr, fcNone}; | |||
| 4448 | } | |||
| 4449 | } else if (ConstRHS->isSmallestNormalized() && !ConstRHS->isNegative()) { | |||
| 4450 | // Match pattern that's used in __builtin_isnormal. | |||
| 4451 | switch (Pred) { | |||
| 4452 | case FCmpInst::FCMP_OLT: | |||
| 4453 | case FCmpInst::FCMP_UGE: { | |||
| 4454 | // fcmp olt x, smallest_normal -> fcNegInf|fcNegNormal|fcSubnormal|fcZero | |||
| 4455 | // fcmp olt fabs(x), smallest_normal -> fcSubnormal|fcZero | |||
| 4456 | // fcmp uge x, smallest_normal -> fcNan|fcPosNormal|fcPosInf | |||
| 4457 | // fcmp uge fabs(x), smallest_normal -> ~(fcSubnormal|fcZero) | |||
| 4458 | Mask = fcZero | fcSubnormal; | |||
| 4459 | if (!IsFabs) | |||
| 4460 | Mask |= fcNegNormal | fcNegInf; | |||
| 4461 | ||||
| 4462 | break; | |||
| 4463 | } | |||
| 4464 | case FCmpInst::FCMP_OGE: | |||
| 4465 | case FCmpInst::FCMP_ULT: { | |||
| 4466 | // fcmp oge x, smallest_normal -> fcPosNormal | fcPosInf | |||
| 4467 | // fcmp oge fabs(x), smallest_normal -> fcInf | fcNormal | |||
| 4468 | // fcmp ult x, smallest_normal -> ~(fcPosNormal | fcPosInf) | |||
| 4469 | // fcmp ult fabs(x), smallest_normal -> ~(fcInf | fcNormal) | |||
| 4470 | Mask = fcPosInf | fcPosNormal; | |||
| 4471 | if (IsFabs) | |||
| 4472 | Mask |= fcNegInf | fcNegNormal; | |||
| 4473 | break; | |||
| 4474 | } | |||
| 4475 | default: | |||
| 4476 | return {nullptr, fcNone}; | |||
| 4477 | } | |||
| 4478 | } else | |||
| 4479 | return {nullptr, fcNone}; | |||
| 4480 | ||||
| 4481 | // Invert the comparison for the unordered cases. | |||
| 4482 | if (FCmpInst::isUnordered(Pred)) | |||
| 4483 | Mask = ~Mask; | |||
| 4484 | ||||
| 4485 | return {Src, Mask}; | |||
| 4486 | } | |||
| 4487 | ||||
| 4488 | static FPClassTest computeKnownFPClassFromAssumes(const Value *V, | |||
| 4489 | const Query &Q) { | |||
| 4490 | FPClassTest KnownFromAssume = fcAllFlags; | |||
| 4491 | ||||
| 4492 | // Try to restrict the floating-point classes based on information from | |||
| 4493 | // assumptions. | |||
| 4494 | for (auto &AssumeVH : Q.AC->assumptionsFor(V)) { | |||
| 4495 | if (!AssumeVH) | |||
| 4496 | continue; | |||
| 4497 | CallInst *I = cast<CallInst>(AssumeVH); | |||
| 4498 | const Function *F = I->getFunction(); | |||
| 4499 | ||||
| 4500 | assert(F == Q.CxtI->getParent()->getParent() &&(static_cast <bool> (F == Q.CxtI->getParent()->getParent () && "Got assumption for the wrong function!") ? void (0) : __assert_fail ("F == Q.CxtI->getParent()->getParent() && \"Got assumption for the wrong function!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 4501, __extension__ __PRETTY_FUNCTION__ )) | |||
| 4501 | "Got assumption for the wrong function!")(static_cast <bool> (F == Q.CxtI->getParent()->getParent () && "Got assumption for the wrong function!") ? void (0) : __assert_fail ("F == Q.CxtI->getParent()->getParent() && \"Got assumption for the wrong function!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 4501, __extension__ __PRETTY_FUNCTION__ )); | |||
| 4502 | assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&(static_cast <bool> (I->getCalledFunction()->getIntrinsicID () == Intrinsic::assume && "must be an assume intrinsic" ) ? void (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\"" , "llvm/lib/Analysis/ValueTracking.cpp", 4503, __extension__ __PRETTY_FUNCTION__ )) | |||
| 4503 | "must be an assume intrinsic")(static_cast <bool> (I->getCalledFunction()->getIntrinsicID () == Intrinsic::assume && "must be an assume intrinsic" ) ? void (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\"" , "llvm/lib/Analysis/ValueTracking.cpp", 4503, __extension__ __PRETTY_FUNCTION__ )); | |||
| 4504 | ||||
| 4505 | if (!isValidAssumeForContext(I, Q.CxtI, Q.DT)) | |||
| 4506 | continue; | |||
| 4507 | ||||
| 4508 | CmpInst::Predicate Pred; | |||
| 4509 | Value *LHS, *RHS; | |||
| 4510 | uint64_t ClassVal = 0; | |||
| 4511 | if (match(I->getArgOperand(0), m_FCmp(Pred, m_Value(LHS), m_Value(RHS)))) { | |||
| 4512 | auto [TestedValue, TestedMask] = | |||
| 4513 | fcmpToClassTest(Pred, *F, LHS, RHS, true); | |||
| 4514 | // First see if we can fold in fabs/fneg into the test. | |||
| 4515 | if (TestedValue == V) | |||
| 4516 | KnownFromAssume &= TestedMask; | |||
| 4517 | else { | |||
| 4518 | // Try again without the lookthrough if we found a different source | |||
| 4519 | // value. | |||
| 4520 | auto [TestedValue, TestedMask] = | |||
| 4521 | fcmpToClassTest(Pred, *F, LHS, RHS, false); | |||
| 4522 | if (TestedValue == V) | |||
| 4523 | KnownFromAssume &= TestedMask; | |||
| 4524 | } | |||
| 4525 | } else if (match(I->getArgOperand(0), | |||
| 4526 | m_Intrinsic<Intrinsic::is_fpclass>( | |||
| 4527 | m_Value(LHS), m_ConstantInt(ClassVal)))) { | |||
| 4528 | KnownFromAssume &= static_cast<FPClassTest>(ClassVal); | |||
| 4529 | } | |||
| 4530 | } | |||
| 4531 | ||||
| 4532 | return KnownFromAssume; | |||
| 4533 | } | |||
| 4534 | ||||
| 4535 | void computeKnownFPClass(const Value *V, const APInt &DemandedElts, | |||
| 4536 | FPClassTest InterestedClasses, KnownFPClass &Known, | |||
| 4537 | unsigned Depth, const Query &Q, | |||
| 4538 | const TargetLibraryInfo *TLI); | |||
| 4539 | ||||
| 4540 | static void computeKnownFPClass(const Value *V, KnownFPClass &Known, | |||
| 4541 | FPClassTest InterestedClasses, unsigned Depth, | |||
| 4542 | const Query &Q, const TargetLibraryInfo *TLI) { | |||
| 4543 | auto *FVTy = dyn_cast<FixedVectorType>(V->getType()); | |||
| 4544 | APInt DemandedElts = | |||
| 4545 | FVTy ? APInt::getAllOnes(FVTy->getNumElements()) : APInt(1, 1); | |||
| 4546 | computeKnownFPClass(V, DemandedElts, InterestedClasses, Known, Depth, Q, TLI); | |||
| 4547 | } | |||
| 4548 | ||||
| 4549 | static void computeKnownFPClassForFPTrunc(const Operator *Op, | |||
| 4550 | const APInt &DemandedElts, | |||
| 4551 | FPClassTest InterestedClasses, | |||
| 4552 | KnownFPClass &Known, unsigned Depth, | |||
| 4553 | const Query &Q, | |||
| 4554 | const TargetLibraryInfo *TLI) { | |||
| 4555 | if ((InterestedClasses & fcNan) == fcNone) | |||
| 4556 | return; | |||
| 4557 | ||||
| 4558 | KnownFPClass KnownSrc; | |||
| 4559 | computeKnownFPClass(Op->getOperand(0), DemandedElts, InterestedClasses, | |||
| 4560 | KnownSrc, Depth + 1, Q, TLI); | |||
| 4561 | if (KnownSrc.isKnownNeverNaN()) | |||
| 4562 | Known.knownNot(fcNan); | |||
| 4563 | ||||
| 4564 | // Infinity needs a range check. | |||
| 4565 | // TODO: Sign bit should be preserved | |||
| 4566 | } | |||
| 4567 | ||||
| 4568 | // TODO: Merge implementations of isKnownNeverNaN, isKnownNeverInfinity, | |||
| 4569 | // CannotBeNegativeZero, cannotBeOrderedLessThanZero into here. | |||
| 4570 | ||||
| 4571 | void computeKnownFPClass(const Value *V, const APInt &DemandedElts, | |||
| 4572 | FPClassTest InterestedClasses, KnownFPClass &Known, | |||
| 4573 | unsigned Depth, const Query &Q, | |||
| 4574 | const TargetLibraryInfo *TLI) { | |||
| 4575 | assert(Known.isUnknown() && "should not be called with known information")(static_cast <bool> (Known.isUnknown() && "should not be called with known information" ) ? void (0) : __assert_fail ("Known.isUnknown() && \"should not be called with known information\"" , "llvm/lib/Analysis/ValueTracking.cpp", 4575, __extension__ __PRETTY_FUNCTION__ )); | |||
| 4576 | ||||
| 4577 | if (!DemandedElts) { | |||
| 4578 | // No demanded elts, better to assume we don't know anything. | |||
| 4579 | Known.resetAll(); | |||
| 4580 | return; | |||
| 4581 | } | |||
| 4582 | ||||
| 4583 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth")(static_cast <bool> (Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth") ? void (0) : __assert_fail ( "Depth <= MaxAnalysisRecursionDepth && \"Limit Search Depth\"" , "llvm/lib/Analysis/ValueTracking.cpp", 4583, __extension__ __PRETTY_FUNCTION__ )); | |||
| 4584 | ||||
| 4585 | if (auto *CFP = dyn_cast_or_null<ConstantFP>(V)) { | |||
| 4586 | Known.KnownFPClasses = CFP->getValueAPF().classify(); | |||
| 4587 | Known.SignBit = CFP->isNegative(); | |||
| 4588 | return; | |||
| 4589 | } | |||
| 4590 | ||||
| 4591 | // Try to handle fixed width vector constants | |||
| 4592 | auto *VFVTy = dyn_cast<FixedVectorType>(V->getType()); | |||
| 4593 | const Constant *CV = dyn_cast<Constant>(V); | |||
| 4594 | if (VFVTy && CV) { | |||
| 4595 | Known.KnownFPClasses = fcNone; | |||
| 4596 | ||||
| 4597 | // For vectors, verify that each element is not NaN. | |||
| 4598 | unsigned NumElts = VFVTy->getNumElements(); | |||
| 4599 | for (unsigned i = 0; i != NumElts; ++i) { | |||
| 4600 | Constant *Elt = CV->getAggregateElement(i); | |||
| 4601 | if (!Elt) { | |||
| 4602 | Known = KnownFPClass(); | |||
| 4603 | return; | |||
| 4604 | } | |||
| 4605 | if (isa<UndefValue>(Elt)) | |||
| 4606 | continue; | |||
| 4607 | auto *CElt = dyn_cast<ConstantFP>(Elt); | |||
| 4608 | if (!CElt) { | |||
| 4609 | Known = KnownFPClass(); | |||
| 4610 | return; | |||
| 4611 | } | |||
| 4612 | ||||
| 4613 | KnownFPClass KnownElt{CElt->getValueAPF().classify(), CElt->isNegative()}; | |||
| 4614 | Known |= KnownElt; | |||
| 4615 | } | |||
| 4616 | ||||
| 4617 | return; | |||
| 4618 | } | |||
| 4619 | ||||
| 4620 | FPClassTest KnownNotFromFlags = fcNone; | |||
| 4621 | if (const auto *CB = dyn_cast<CallBase>(V)) | |||
| 4622 | KnownNotFromFlags |= CB->getRetNoFPClass(); | |||
| 4623 | else if (const auto *Arg = dyn_cast<Argument>(V)) | |||
| 4624 | KnownNotFromFlags |= Arg->getNoFPClass(); | |||
| 4625 | ||||
| 4626 | const Operator *Op = dyn_cast<Operator>(V); | |||
| 4627 | if (const FPMathOperator *FPOp = dyn_cast_or_null<FPMathOperator>(Op)) { | |||
| 4628 | if (FPOp->hasNoNaNs()) | |||
| 4629 | KnownNotFromFlags |= fcNan; | |||
| 4630 | if (FPOp->hasNoInfs()) | |||
| 4631 | KnownNotFromFlags |= fcInf; | |||
| 4632 | } | |||
| 4633 | ||||
| 4634 | if (Q.AC) { | |||
| 4635 | FPClassTest AssumedClasses = computeKnownFPClassFromAssumes(V, Q); | |||
| 4636 | KnownNotFromFlags |= ~AssumedClasses; | |||
| 4637 | } | |||
| 4638 | ||||
| 4639 | // We no longer need to find out about these bits from inputs if we can | |||
| 4640 | // assume this from flags/attributes. | |||
| 4641 | InterestedClasses &= ~KnownNotFromFlags; | |||
| 4642 | ||||
| 4643 | auto ClearClassesFromFlags = make_scope_exit([=, &Known] { | |||
| 4644 | Known.knownNot(KnownNotFromFlags); | |||
| 4645 | }); | |||
| 4646 | ||||
| 4647 | if (!Op) | |||
| 4648 | return; | |||
| 4649 | ||||
| 4650 | // All recursive calls that increase depth must come after this. | |||
| 4651 | if (Depth == MaxAnalysisRecursionDepth) | |||
| 4652 | return; | |||
| 4653 | ||||
| 4654 | const unsigned Opc = Op->getOpcode(); | |||
| 4655 | switch (Opc) { | |||
| 4656 | case Instruction::FNeg: { | |||
| 4657 | computeKnownFPClass(Op->getOperand(0), DemandedElts, InterestedClasses, | |||
| 4658 | Known, Depth + 1, Q, TLI); | |||
| 4659 | Known.fneg(); | |||
| 4660 | break; | |||
| 4661 | } | |||
| 4662 | case Instruction::Select: { | |||
| 4663 | KnownFPClass Known2; | |||
| 4664 | computeKnownFPClass(Op->getOperand(1), DemandedElts, InterestedClasses, | |||
| 4665 | Known, Depth + 1, Q, TLI); | |||
| 4666 | computeKnownFPClass(Op->getOperand(2), DemandedElts, InterestedClasses, | |||
| 4667 | Known2, Depth + 1, Q, TLI); | |||
| 4668 | Known |= Known2; | |||
| 4669 | break; | |||
| 4670 | } | |||
| 4671 | case Instruction::Call: { | |||
| 4672 | if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op)) { | |||
| 4673 | const Intrinsic::ID IID = II->getIntrinsicID(); | |||
| 4674 | switch (IID) { | |||
| 4675 | case Intrinsic::fabs: { | |||
| 4676 | if ((InterestedClasses & (fcNan | fcPositive)) != fcNone) { | |||
| 4677 | // If we only care about the sign bit we don't need to inspect the | |||
| 4678 | // operand. | |||
| 4679 | computeKnownFPClass(II->getArgOperand(0), DemandedElts, | |||
| 4680 | InterestedClasses, Known, Depth + 1, Q, TLI); | |||
| 4681 | } | |||
| 4682 | ||||
| 4683 | Known.fabs(); | |||
| 4684 | break; | |||
| 4685 | } | |||
| 4686 | case Intrinsic::copysign: { | |||
| 4687 | KnownFPClass KnownSign; | |||
| 4688 | ||||
| 4689 | computeKnownFPClass(II->getArgOperand(0), DemandedElts, | |||
| 4690 | InterestedClasses, Known, Depth + 1, Q, TLI); | |||
| 4691 | computeKnownFPClass(II->getArgOperand(1), DemandedElts, | |||
| 4692 | InterestedClasses, KnownSign, Depth + 1, Q, TLI); | |||
| 4693 | Known.copysign(KnownSign); | |||
| 4694 | break; | |||
| 4695 | } | |||
| 4696 | case Intrinsic::fma: | |||
| 4697 | case Intrinsic::fmuladd: { | |||
| 4698 | if ((InterestedClasses & fcNegative) == fcNone) | |||
| 4699 | break; | |||
| 4700 | ||||
| 4701 | if (II->getArgOperand(0) != II->getArgOperand(1)) | |||
| 4702 | break; | |||
| 4703 | ||||
| 4704 | // The multiply cannot be -0 and therefore the add can't be -0 | |||
| 4705 | Known.knownNot(fcNegZero); | |||
| 4706 | ||||
| 4707 | // x * x + y is non-negative if y is non-negative. | |||
| 4708 | KnownFPClass KnownAddend; | |||
| 4709 | computeKnownFPClass(II->getArgOperand(2), DemandedElts, | |||
| 4710 | InterestedClasses, KnownAddend, Depth + 1, Q, TLI); | |||
| 4711 | ||||
| 4712 | // TODO: Known sign bit with no nans | |||
| 4713 | if (KnownAddend.cannotBeOrderedLessThanZero()) | |||
| 4714 | Known.knownNot(fcNegative); | |||
| 4715 | break; | |||
| 4716 | } | |||
| 4717 | case Intrinsic::sin: | |||
| 4718 | case Intrinsic::cos: { | |||
| 4719 | // Return NaN on infinite inputs. | |||
| 4720 | KnownFPClass KnownSrc; | |||
| 4721 | computeKnownFPClass(II->getArgOperand(0), DemandedElts, | |||
| 4722 | InterestedClasses, KnownSrc, Depth + 1, Q, TLI); | |||
| 4723 | Known.knownNot(fcInf); | |||
| 4724 | if (KnownSrc.isKnownNeverNaN() && KnownSrc.isKnownNeverInfinity()) | |||
| 4725 | Known.knownNot(fcNan); | |||
| 4726 | break; | |||
| 4727 | } | |||
| 4728 | ||||
| 4729 | case Intrinsic::maxnum: | |||
| 4730 | case Intrinsic::minnum: | |||
| 4731 | case Intrinsic::minimum: | |||
| 4732 | case Intrinsic::maximum: { | |||
| 4733 | KnownFPClass KnownLHS, KnownRHS; | |||
| 4734 | computeKnownFPClass(II->getArgOperand(0), DemandedElts, | |||
| 4735 | InterestedClasses, KnownLHS, Depth + 1, Q, TLI); | |||
| 4736 | computeKnownFPClass(II->getArgOperand(1), DemandedElts, | |||
| 4737 | InterestedClasses, KnownRHS, Depth + 1, Q, TLI); | |||
| 4738 | ||||
| 4739 | bool NeverNaN = | |||
| 4740 | KnownLHS.isKnownNeverNaN() || KnownRHS.isKnownNeverNaN(); | |||
| 4741 | Known = KnownLHS | KnownRHS; | |||
| 4742 | ||||
| 4743 | // If either operand is not NaN, the result is not NaN. | |||
| 4744 | if (NeverNaN && (IID == Intrinsic::minnum || IID == Intrinsic::maxnum)) | |||
| 4745 | Known.knownNot(fcNan); | |||
| 4746 | ||||
| 4747 | if (IID == Intrinsic::maxnum) { | |||
| 4748 | // If at least one operand is known to be positive, the result must be | |||
| 4749 | // positive. | |||
| 4750 | if ((KnownLHS.cannotBeOrderedLessThanZero() && | |||
| 4751 | KnownLHS.isKnownNeverNaN()) || | |||
| 4752 | (KnownRHS.cannotBeOrderedLessThanZero() && | |||
| 4753 | KnownRHS.isKnownNeverNaN())) | |||
| 4754 | Known.knownNot(KnownFPClass::OrderedLessThanZeroMask); | |||
| 4755 | } else if (IID == Intrinsic::maximum) { | |||
| 4756 | // If at least one operand is known to be positive, the result must be | |||
| 4757 | // positive. | |||
| 4758 | if (KnownLHS.cannotBeOrderedLessThanZero() || | |||
| 4759 | KnownRHS.cannotBeOrderedLessThanZero()) | |||
| 4760 | Known.knownNot(KnownFPClass::OrderedLessThanZeroMask); | |||
| 4761 | } else if (IID == Intrinsic::minnum) { | |||
| 4762 | // If at least one operand is known to be negative, the result must be | |||
| 4763 | // negative. | |||
| 4764 | if ((KnownLHS.cannotBeOrderedGreaterThanZero() && | |||
| 4765 | KnownLHS.isKnownNeverNaN()) || | |||
| 4766 | (KnownRHS.cannotBeOrderedGreaterThanZero() && | |||
| 4767 | KnownRHS.isKnownNeverNaN())) | |||
| 4768 | Known.knownNot(KnownFPClass::OrderedGreaterThanZeroMask); | |||
| 4769 | } else { | |||
| 4770 | // If at least one operand is known to be negative, the result must be | |||
| 4771 | // negative. | |||
| 4772 | if (KnownLHS.cannotBeOrderedGreaterThanZero() || | |||
| 4773 | KnownRHS.cannotBeOrderedGreaterThanZero()) | |||
| 4774 | Known.knownNot(KnownFPClass::OrderedGreaterThanZeroMask); | |||
| 4775 | } | |||
| 4776 | ||||
| 4777 | // Fixup zero handling if denormals could be returned as a zero. | |||
| 4778 | // | |||
| 4779 | // As there's no spec for denormal flushing, be conservative with the | |||
| 4780 | // treatment of denormals that could be flushed to zero. For older | |||
| 4781 | // subtargets on AMDGPU the min/max instructions would not flush the | |||
| 4782 | // output and return the original value. | |||
| 4783 | // | |||
| 4784 | // TODO: This could be refined based on the sign | |||
| 4785 | if ((Known.KnownFPClasses & fcZero) != fcNone && | |||
| 4786 | !Known.isKnownNeverSubnormal()) { | |||
| 4787 | const Function *Parent = II->getFunction(); | |||
| 4788 | DenormalMode Mode = Parent->getDenormalMode( | |||
| 4789 | II->getType()->getScalarType()->getFltSemantics()); | |||
| 4790 | if (Mode != DenormalMode::getIEEE()) | |||
| 4791 | Known.KnownFPClasses |= fcZero; | |||
| 4792 | } | |||
| 4793 | ||||
| 4794 | break; | |||
| 4795 | } | |||
| 4796 | case Intrinsic::canonicalize: { | |||
| 4797 | computeKnownFPClass(II->getArgOperand(0), DemandedElts, | |||
| 4798 | InterestedClasses, Known, Depth + 1, Q, TLI); | |||
| 4799 | // Canonicalize is guaranteed to quiet signaling nans. | |||
| 4800 | Known.knownNot(fcSNan); | |||
| 4801 | ||||
| 4802 | // If the parent function flushes denormals, the canonical output cannot | |||
| 4803 | // be a denormal. | |||
| 4804 | const fltSemantics &FPType = II->getType()->getFltSemantics(); | |||
| 4805 | DenormalMode DenormMode = II->getFunction()->getDenormalMode(FPType); | |||
| 4806 | if (DenormMode.inputsAreZero() || DenormMode.outputsAreZero()) | |||
| 4807 | Known.knownNot(fcSubnormal); | |||
| 4808 | ||||
| 4809 | if (DenormMode.Input == DenormalMode::PositiveZero || | |||
| 4810 | (DenormMode.Output == DenormalMode::PositiveZero && | |||
| 4811 | DenormMode.Input == DenormalMode::IEEE)) | |||
| 4812 | Known.knownNot(fcNegZero); | |||
| 4813 | ||||
| 4814 | break; | |||
| 4815 | } | |||
| 4816 | case Intrinsic::trunc: { | |||
| 4817 | KnownFPClass KnownSrc; | |||
| 4818 | ||||
| 4819 | FPClassTest InterestedSrcs = InterestedClasses; | |||
| 4820 | if (InterestedClasses & fcZero) | |||
| 4821 | InterestedClasses |= fcNormal | fcSubnormal; | |||
| 4822 | ||||
| 4823 | computeKnownFPClass(II->getArgOperand(0), DemandedElts, InterestedSrcs, | |||
| 4824 | KnownSrc, Depth + 1, Q, TLI); | |||
| 4825 | ||||
| 4826 | // Integer results cannot be subnormal. | |||
| 4827 | Known.knownNot(fcSubnormal); | |||
| 4828 | ||||
| 4829 | // trunc passes through infinities. | |||
| 4830 | if (KnownSrc.isKnownNeverPosInfinity()) | |||
| 4831 | Known.knownNot(fcPosInf); | |||
| 4832 | if (KnownSrc.isKnownNeverNegInfinity()) | |||
| 4833 | Known.knownNot(fcNegInf); | |||
| 4834 | ||||
| 4835 | // Non-constrained intrinsics do not guarantee signaling nan quieting. | |||
| 4836 | if (KnownSrc.isKnownNeverNaN()) | |||
| 4837 | Known.knownNot(fcNan); | |||
| 4838 | ||||
| 4839 | if (KnownSrc.isKnownNever(fcPosNormal)) | |||
| 4840 | Known.knownNot(fcPosNormal); | |||
| 4841 | ||||
| 4842 | if (KnownSrc.isKnownNever(fcNegNormal)) | |||
| 4843 | Known.knownNot(fcNegNormal); | |||
| 4844 | ||||
| 4845 | if (KnownSrc.isKnownNever(fcPosZero | fcPosSubnormal | fcPosNormal)) | |||
| 4846 | Known.knownNot(fcPosZero); | |||
| 4847 | ||||
| 4848 | if (KnownSrc.isKnownNever(fcNegZero | fcNegSubnormal | fcNegNormal)) | |||
| 4849 | Known.knownNot(fcNegZero); | |||
| 4850 | ||||
| 4851 | // Sign should be preserved | |||
| 4852 | Known.SignBit = KnownSrc.SignBit; | |||
| 4853 | break; | |||
| 4854 | } | |||
| 4855 | case Intrinsic::exp: | |||
| 4856 | case Intrinsic::exp2: { | |||
| 4857 | Known.knownNot(fcNegative); | |||
| 4858 | if ((InterestedClasses & fcNan) == fcNone) | |||
| 4859 | break; | |||
| 4860 | ||||
| 4861 | KnownFPClass KnownSrc; | |||
| 4862 | computeKnownFPClass(II->getArgOperand(0), DemandedElts, | |||
| 4863 | InterestedClasses, KnownSrc, Depth + 1, Q, TLI); | |||
| 4864 | if (KnownSrc.isKnownNeverNaN()) { | |||
| 4865 | Known.knownNot(fcNan); | |||
| 4866 | Known.SignBit = false; | |||
| 4867 | } | |||
| 4868 | ||||
| 4869 | break; | |||
| 4870 | } | |||
| 4871 | case Intrinsic::fptrunc_round: { | |||
| 4872 | computeKnownFPClassForFPTrunc(Op, DemandedElts, InterestedClasses, | |||
| 4873 | Known, Depth, Q, TLI); | |||
| 4874 | break; | |||
| 4875 | } | |||
| 4876 | case Intrinsic::log: | |||
| 4877 | case Intrinsic::log10: | |||
| 4878 | case Intrinsic::log2: | |||
| 4879 | case Intrinsic::experimental_constrained_log: | |||
| 4880 | case Intrinsic::experimental_constrained_log10: | |||
| 4881 | case Intrinsic::experimental_constrained_log2: { | |||
| 4882 | // log(+inf) -> +inf | |||
| 4883 | // log([+-]0.0) -> -inf | |||
| 4884 | // log(-inf) -> nan | |||
| 4885 | // log(-x) -> nan | |||
| 4886 | if ((InterestedClasses & (fcNan | fcInf)) == fcNone) | |||
| 4887 | break; | |||
| 4888 | ||||
| 4889 | FPClassTest InterestedSrcs = InterestedClasses; | |||
| 4890 | if ((InterestedClasses & fcNegInf) != fcNone) | |||
| 4891 | InterestedSrcs |= fcZero | fcSubnormal; | |||
| 4892 | if ((InterestedClasses & fcNan) != fcNone) | |||
| 4893 | InterestedSrcs |= fcNan | (fcNegative & ~fcNan); | |||
| 4894 | ||||
| 4895 | KnownFPClass KnownSrc; | |||
| 4896 | computeKnownFPClass(II->getArgOperand(0), DemandedElts, InterestedSrcs, | |||
| 4897 | KnownSrc, Depth + 1, Q, TLI); | |||
| 4898 | ||||
| 4899 | if (KnownSrc.isKnownNeverPosInfinity()) | |||
| 4900 | Known.knownNot(fcPosInf); | |||
| 4901 | ||||
| 4902 | if (KnownSrc.isKnownNeverNaN() && | |||
| 4903 | KnownSrc.cannotBeOrderedLessThanZero()) | |||
| 4904 | Known.knownNot(fcNan); | |||
| 4905 | ||||
| 4906 | if (KnownSrc.isKnownNeverLogicalZero(*II->getFunction(), II->getType())) | |||
| 4907 | Known.knownNot(fcNegInf); | |||
| 4908 | ||||
| 4909 | break; | |||
| 4910 | } | |||
| 4911 | case Intrinsic::powi: { | |||
| 4912 | if ((InterestedClasses & fcNegative) == fcNone) | |||
| 4913 | break; | |||
| 4914 | ||||
| 4915 | const Value *Exp = II->getArgOperand(1); | |||
| 4916 | unsigned BitWidth = | |||
| 4917 | Exp->getType()->getScalarType()->getIntegerBitWidth(); | |||
| 4918 | KnownBits ExponentKnownBits(BitWidth); | |||
| 4919 | computeKnownBits(Exp, DemandedElts, ExponentKnownBits, Depth + 1, Q); | |||
| 4920 | ||||
| 4921 | if (ExponentKnownBits.Zero[0]) { // Is even | |||
| 4922 | Known.knownNot(fcNegative); | |||
| 4923 | break; | |||
| 4924 | } | |||
| 4925 | ||||
| 4926 | // Given that exp is an integer, here are the | |||
| 4927 | // ways that pow can return a negative value: | |||
| 4928 | // | |||
| 4929 | // pow(-x, exp) --> negative if exp is odd and x is negative. | |||
| 4930 | // pow(-0, exp) --> -inf if exp is negative odd. | |||
| 4931 | // pow(-0, exp) --> -0 if exp is positive odd. | |||
| 4932 | // pow(-inf, exp) --> -0 if exp is negative odd. | |||
| 4933 | // pow(-inf, exp) --> -inf if exp is positive odd. | |||
| 4934 | KnownFPClass KnownSrc; | |||
| 4935 | computeKnownFPClass(II->getArgOperand(0), DemandedElts, fcNegative, | |||
| 4936 | KnownSrc, Depth + 1, Q, TLI); | |||
| 4937 | if (KnownSrc.isKnownNever(fcNegative)) | |||
| 4938 | Known.knownNot(fcNegative); | |||
| 4939 | break; | |||
| 4940 | } | |||
| 4941 | case Intrinsic::arithmetic_fence: { | |||
| 4942 | computeKnownFPClass(II->getArgOperand(0), DemandedElts, | |||
| 4943 | InterestedClasses, Known, Depth + 1, Q, TLI); | |||
| 4944 | break; | |||
| 4945 | } | |||
| 4946 | case Intrinsic::experimental_constrained_sitofp: | |||
| 4947 | case Intrinsic::experimental_constrained_uitofp: | |||
| 4948 | // Cannot produce nan | |||
| 4949 | Known.knownNot(fcNan); | |||
| 4950 | ||||
| 4951 | // sitofp and uitofp turn into +0.0 for zero. | |||
| 4952 | Known.knownNot(fcNegZero); | |||
| 4953 | ||||
| 4954 | // Integers cannot be subnormal | |||
| 4955 | Known.knownNot(fcSubnormal); | |||
| 4956 | ||||
| 4957 | if (IID == Intrinsic::experimental_constrained_uitofp) | |||
| 4958 | Known.signBitMustBeZero(); | |||
| 4959 | ||||
| 4960 | // TODO: Copy inf handling from instructions | |||
| 4961 | break; | |||
| 4962 | default: | |||
| 4963 | break; | |||
| 4964 | } | |||
| 4965 | } | |||
| 4966 | ||||
| 4967 | break; | |||
| 4968 | } | |||
| 4969 | case Instruction::FAdd: | |||
| 4970 | case Instruction::FSub: { | |||
| 4971 | KnownFPClass KnownLHS, KnownRHS; | |||
| 4972 | computeKnownFPClass(Op->getOperand(1), DemandedElts, fcNan | fcInf, | |||
| 4973 | KnownRHS, Depth + 1, Q, TLI); | |||
| 4974 | if (KnownRHS.isKnownNeverNaN()) { | |||
| 4975 | // RHS is canonically cheaper to compute. Skip inspecting the LHS if | |||
| 4976 | // there's no point. | |||
| 4977 | computeKnownFPClass(Op->getOperand(0), DemandedElts, fcNan | fcInf, | |||
| 4978 | KnownLHS, Depth + 1, Q, TLI); | |||
| 4979 | // Adding positive and negative infinity produces NaN. | |||
| 4980 | // TODO: Check sign of infinities. | |||
| 4981 | if (KnownLHS.isKnownNeverNaN() && | |||
| 4982 | (KnownLHS.isKnownNeverInfinity() || KnownRHS.isKnownNeverInfinity())) | |||
| 4983 | Known.knownNot(fcNan); | |||
| 4984 | } | |||
| 4985 | ||||
| 4986 | break; | |||
| 4987 | } | |||
| 4988 | case Instruction::FMul: { | |||
| 4989 | // X * X is always non-negative or a NaN. | |||
| 4990 | if (Op->getOperand(0) == Op->getOperand(1)) | |||
| 4991 | Known.knownNot(fcNegative); | |||
| 4992 | ||||
| 4993 | if ((InterestedClasses & fcNan) != fcNan) | |||
| 4994 | break; | |||
| 4995 | ||||
| 4996 | KnownFPClass KnownLHS, KnownRHS; | |||
| 4997 | computeKnownFPClass(Op->getOperand(1), DemandedElts, | |||
| 4998 | fcNan | fcInf | fcZero | fcSubnormal, KnownRHS, | |||
| 4999 | Depth + 1, Q, TLI); | |||
| 5000 | if (KnownRHS.isKnownNeverNaN() && | |||
| 5001 | (KnownRHS.isKnownNeverInfinity() || KnownRHS.isKnownNeverZero())) { | |||
| 5002 | computeKnownFPClass(Op->getOperand(0), DemandedElts, | |||
| 5003 | fcNan | fcInf | fcZero, KnownLHS, Depth + 1, Q, TLI); | |||
| 5004 | if (!KnownLHS.isKnownNeverNaN()) | |||
| 5005 | break; | |||
| 5006 | ||||
| 5007 | const Function *F = cast<Instruction>(Op)->getFunction(); | |||
| 5008 | ||||
| 5009 | // If neither side can be zero (or nan) fmul never produces NaN. | |||
| 5010 | // TODO: Check operand combinations. | |||
| 5011 | // e.g. fmul nofpclass(inf nan zero), nofpclass(nan) -> nofpclass(nan) | |||
| 5012 | if ((KnownLHS.isKnownNeverInfinity() || | |||
| 5013 | KnownLHS.isKnownNeverLogicalZero(*F, Op->getType())) && | |||
| 5014 | (KnownRHS.isKnownNeverInfinity() || | |||
| 5015 | KnownRHS.isKnownNeverLogicalZero(*F, Op->getType()))) | |||
| 5016 | Known.knownNot(fcNan); | |||
| 5017 | } | |||
| 5018 | ||||
| 5019 | break; | |||
| 5020 | } | |||
| 5021 | case Instruction::FDiv: | |||
| 5022 | case Instruction::FRem: { | |||
| 5023 | if (Op->getOperand(0) == Op->getOperand(1)) { | |||
| 5024 | // TODO: Could filter out snan if we inspect the operand | |||
| 5025 | if (Op->getOpcode() == Instruction::FDiv) { | |||
| 5026 | // X / X is always exactly 1.0 or a NaN. | |||
| 5027 | Known.KnownFPClasses = fcNan | fcPosNormal; | |||
| 5028 | } else { | |||
| 5029 | // X % X is always exactly [+-]0.0 or a NaN. | |||
| 5030 | Known.KnownFPClasses = fcNan | fcZero; | |||
| 5031 | } | |||
| 5032 | ||||
| 5033 | break; | |||
| 5034 | } | |||
| 5035 | ||||
| 5036 | const bool WantNan = (InterestedClasses & fcNan) != fcNone; | |||
| 5037 | const bool WantNegative = (InterestedClasses & fcNegative) != fcNone; | |||
| 5038 | const bool WantPositive = | |||
| 5039 | Opc == Instruction::FRem && (InterestedClasses & fcPositive) != fcNone; | |||
| 5040 | if (!WantNan && !WantNegative && !WantPositive) | |||
| 5041 | break; | |||
| 5042 | ||||
| 5043 | KnownFPClass KnownLHS, KnownRHS; | |||
| 5044 | ||||
| 5045 | computeKnownFPClass(Op->getOperand(1), DemandedElts, | |||
| 5046 | fcNan | fcInf | fcZero | fcNegative, KnownRHS, | |||
| 5047 | Depth + 1, Q, TLI); | |||
| 5048 | ||||
| 5049 | bool KnowSomethingUseful = | |||
| 5050 | KnownRHS.isKnownNeverNaN() || KnownRHS.isKnownNever(fcNegative); | |||
| 5051 | ||||
| 5052 | if (KnowSomethingUseful || WantPositive) { | |||
| 5053 | const FPClassTest InterestedLHS = | |||
| 5054 | WantPositive ? fcAllFlags | |||
| 5055 | : fcNan | fcInf | fcZero | fcSubnormal | fcNegative; | |||
| 5056 | ||||
| 5057 | computeKnownFPClass(Op->getOperand(0), DemandedElts, | |||
| 5058 | InterestedClasses & InterestedLHS, KnownLHS, | |||
| 5059 | Depth + 1, Q, TLI); | |||
| 5060 | } | |||
| 5061 | ||||
| 5062 | const Function *F = cast<Instruction>(Op)->getFunction(); | |||
| 5063 | ||||
| 5064 | if (Op->getOpcode() == Instruction::FDiv) { | |||
| 5065 | // Only 0/0, Inf/Inf produce NaN. | |||
| 5066 | if (KnownLHS.isKnownNeverNaN() && KnownRHS.isKnownNeverNaN() && | |||
| 5067 | (KnownLHS.isKnownNeverInfinity() || | |||
| 5068 | KnownRHS.isKnownNeverInfinity()) && | |||
| 5069 | (KnownLHS.isKnownNeverLogicalZero(*F, Op->getType()) || | |||
| 5070 | KnownRHS.isKnownNeverLogicalZero(*F, Op->getType()))) { | |||
| 5071 | Known.knownNot(fcNan); | |||
| 5072 | } | |||
| 5073 | ||||
| 5074 | // X / -0.0 is -Inf (or NaN). | |||
| 5075 | // +X / +X is +X | |||
| 5076 | if (KnownLHS.isKnownNever(fcNegative) && KnownRHS.isKnownNever(fcNegative)) | |||
| 5077 | Known.knownNot(fcNegative); | |||
| 5078 | } else { | |||
| 5079 | // Inf REM x and x REM 0 produce NaN. | |||
| 5080 | if (KnownLHS.isKnownNeverNaN() && KnownRHS.isKnownNeverNaN() && | |||
| 5081 | KnownLHS.isKnownNeverInfinity() && | |||
| 5082 | KnownRHS.isKnownNeverLogicalZero(*F, Op->getType())) { | |||
| 5083 | Known.knownNot(fcNan); | |||
| 5084 | } | |||
| 5085 | ||||
| 5086 | // The sign for frem is the same as the first operand. | |||
| 5087 | if (KnownLHS.isKnownNever(fcNegative)) | |||
| 5088 | Known.knownNot(fcNegative); | |||
| 5089 | if (KnownLHS.isKnownNever(fcPositive)) | |||
| 5090 | Known.knownNot(fcPositive); | |||
| 5091 | } | |||
| 5092 | ||||
| 5093 | break; | |||
| 5094 | } | |||
| 5095 | case Instruction::FPExt: { | |||
| 5096 | // Infinity, nan and zero propagate from source. | |||
| 5097 | computeKnownFPClass(Op->getOperand(0), DemandedElts, InterestedClasses, | |||
| 5098 | Known, Depth + 1, Q, TLI); | |||
| 5099 | ||||
| 5100 | const fltSemantics &DstTy = | |||
| 5101 | Op->getType()->getScalarType()->getFltSemantics(); | |||
| 5102 | const fltSemantics &SrcTy = | |||
| 5103 | Op->getOperand(0)->getType()->getScalarType()->getFltSemantics(); | |||
| 5104 | ||||
| 5105 | // All subnormal inputs should be in the normal range in the result type. | |||
| 5106 | if (APFloat::isRepresentableAsNormalIn(SrcTy, DstTy)) | |||
| 5107 | Known.knownNot(fcSubnormal); | |||
| 5108 | ||||
| 5109 | // Sign bit of a nan isn't guaranteed. | |||
| 5110 | if (!Known.isKnownNeverNaN()) | |||
| 5111 | Known.SignBit = std::nullopt; | |||
| 5112 | break; | |||
| 5113 | } | |||
| 5114 | case Instruction::FPTrunc: { | |||
| 5115 | computeKnownFPClassForFPTrunc(Op, DemandedElts, InterestedClasses, Known, | |||
| 5116 | Depth, Q, TLI); | |||
| 5117 | break; | |||
| 5118 | } | |||
| 5119 | case Instruction::SIToFP: | |||
| 5120 | case Instruction::UIToFP: { | |||
| 5121 | // Cannot produce nan | |||
| 5122 | Known.knownNot(fcNan); | |||
| 5123 | ||||
| 5124 | // Integers cannot be subnormal | |||
| 5125 | Known.knownNot(fcSubnormal); | |||
| 5126 | ||||
| 5127 | // sitofp and uitofp turn into +0.0 for zero. | |||
| 5128 | Known.knownNot(fcNegZero); | |||
| 5129 | if (Op->getOpcode() == Instruction::UIToFP) | |||
| 5130 | Known.signBitMustBeZero(); | |||
| 5131 | ||||
| 5132 | if (InterestedClasses & fcInf) { | |||
| 5133 | // Get width of largest magnitude integer (remove a bit if signed). | |||
| 5134 | // This still works for a signed minimum value because the largest FP | |||
| 5135 | // value is scaled by some fraction close to 2.0 (1.0 + 0.xxxx). | |||
| 5136 | int IntSize = Op->getOperand(0)->getType()->getScalarSizeInBits(); | |||
| 5137 | if (Op->getOpcode() == Instruction::SIToFP) | |||
| 5138 | --IntSize; | |||
| 5139 | ||||
| 5140 | // If the exponent of the largest finite FP value can hold the largest | |||
| 5141 | // integer, the result of the cast must be finite. | |||
| 5142 | Type *FPTy = Op->getType()->getScalarType(); | |||
| 5143 | if (ilogb(APFloat::getLargest(FPTy->getFltSemantics())) >= IntSize) | |||
| 5144 | Known.knownNot(fcInf); | |||
| 5145 | } | |||
| 5146 | ||||
| 5147 | break; | |||
| 5148 | } | |||
| 5149 | case Instruction::ExtractElement: { | |||
| 5150 | // Look through extract element. If the index is non-constant or | |||
| 5151 | // out-of-range demand all elements, otherwise just the extracted element. | |||
| 5152 | const Value *Vec = Op->getOperand(0); | |||
| 5153 | const Value *Idx = Op->getOperand(1); | |||
| 5154 | auto *CIdx = dyn_cast<ConstantInt>(Idx); | |||
| 5155 | ||||
| 5156 | if (auto *VecTy = dyn_cast<FixedVectorType>(Vec->getType())) { | |||
| 5157 | unsigned NumElts = VecTy->getNumElements(); | |||
| 5158 | APInt DemandedVecElts = APInt::getAllOnes(NumElts); | |||
| 5159 | if (CIdx && CIdx->getValue().ult(NumElts)) | |||
| 5160 | DemandedVecElts = APInt::getOneBitSet(NumElts, CIdx->getZExtValue()); | |||
| 5161 | return computeKnownFPClass(Vec, DemandedVecElts, InterestedClasses, Known, | |||
| 5162 | Depth + 1, Q, TLI); | |||
| 5163 | } | |||
| 5164 | ||||
| 5165 | break; | |||
| 5166 | } | |||
| 5167 | case Instruction::InsertElement: { | |||
| 5168 | if (isa<ScalableVectorType>(Op->getType())) | |||
| 5169 | return; | |||
| 5170 | ||||
| 5171 | const Value *Vec = Op->getOperand(0); | |||
| 5172 | const Value *Elt = Op->getOperand(1); | |||
| 5173 | auto *CIdx = dyn_cast<ConstantInt>(Op->getOperand(2)); | |||
| 5174 | // Early out if the index is non-constant or out-of-range. | |||
| 5175 | unsigned NumElts = DemandedElts.getBitWidth(); | |||
| 5176 | if (!CIdx || CIdx->getValue().uge(NumElts)) | |||
| 5177 | return; | |||
| 5178 | ||||
| 5179 | unsigned EltIdx = CIdx->getZExtValue(); | |||
| 5180 | // Do we demand the inserted element? | |||
| 5181 | if (DemandedElts[EltIdx]) { | |||
| 5182 | computeKnownFPClass(Elt, Known, InterestedClasses, Depth + 1, Q, TLI); | |||
| 5183 | // If we don't know any bits, early out. | |||
| 5184 | if (Known.isUnknown()) | |||
| 5185 | break; | |||
| 5186 | } else { | |||
| 5187 | Known.KnownFPClasses = fcNone; | |||
| 5188 | } | |||
| 5189 | ||||
| 5190 | // We don't need the base vector element that has been inserted. | |||
| 5191 | APInt DemandedVecElts = DemandedElts; | |||
| 5192 | DemandedVecElts.clearBit(EltIdx); | |||
| 5193 | if (!!DemandedVecElts) { | |||
| 5194 | KnownFPClass Known2; | |||
| 5195 | computeKnownFPClass(Vec, DemandedVecElts, InterestedClasses, Known2, | |||
| 5196 | Depth + 1, Q, TLI); | |||
| 5197 | Known |= Known2; | |||
| 5198 | } | |||
| 5199 | ||||
| 5200 | break; | |||
| 5201 | } | |||
| 5202 | case Instruction::ShuffleVector: { | |||
| 5203 | // For undef elements, we don't know anything about the common state of | |||
| 5204 | // the shuffle result. | |||
| 5205 | APInt DemandedLHS, DemandedRHS; | |||
| 5206 | auto *Shuf = dyn_cast<ShuffleVectorInst>(Op); | |||
| 5207 | if (!Shuf || !getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) | |||
| 5208 | return; | |||
| 5209 | ||||
| 5210 | if (!!DemandedLHS) { | |||
| 5211 | const Value *LHS = Shuf->getOperand(0); | |||
| 5212 | computeKnownFPClass(LHS, DemandedLHS, InterestedClasses, Known, | |||
| 5213 | Depth + 1, Q, TLI); | |||
| 5214 | ||||
| 5215 | // If we don't know any bits, early out. | |||
| 5216 | if (Known.isUnknown()) | |||
| 5217 | break; | |||
| 5218 | } else { | |||
| 5219 | Known.KnownFPClasses = fcNone; | |||
| 5220 | } | |||
| 5221 | ||||
| 5222 | if (!!DemandedRHS) { | |||
| 5223 | KnownFPClass Known2; | |||
| 5224 | const Value *RHS = Shuf->getOperand(1); | |||
| 5225 | computeKnownFPClass(RHS, DemandedRHS, InterestedClasses, Known2, | |||
| 5226 | Depth + 1, Q, TLI); | |||
| 5227 | Known |= Known2; | |||
| 5228 | } | |||
| 5229 | ||||
| 5230 | break; | |||
| 5231 | } | |||
| 5232 | case Instruction::ExtractValue: { | |||
| 5233 | computeKnownFPClass(Op->getOperand(0), DemandedElts, InterestedClasses, | |||
| 5234 | Known, Depth + 1, Q, TLI); | |||
| 5235 | break; | |||
| 5236 | } | |||
| 5237 | default: | |||
| 5238 | break; | |||
| 5239 | } | |||
| 5240 | } | |||
| 5241 | ||||
| 5242 | KnownFPClass llvm::computeKnownFPClass( | |||
| 5243 | const Value *V, const APInt &DemandedElts, const DataLayout &DL, | |||
| 5244 | FPClassTest InterestedClasses, unsigned Depth, const TargetLibraryInfo *TLI, | |||
| 5245 | AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT, | |||
| 5246 | OptimizationRemarkEmitter *ORE, bool UseInstrInfo) { | |||
| 5247 | KnownFPClass KnownClasses; | |||
| 5248 | ::computeKnownFPClass(V, DemandedElts, InterestedClasses, KnownClasses, Depth, | |||
| 5249 | Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE), | |||
| 5250 | TLI); | |||
| 5251 | return KnownClasses; | |||
| 5252 | } | |||
| 5253 | ||||
| 5254 | KnownFPClass | |||
| 5255 | llvm::computeKnownFPClass(const Value *V, const DataLayout &DL, | |||
| 5256 | FPClassTest InterestedClasses, unsigned Depth, | |||
| 5257 | const TargetLibraryInfo *TLI, AssumptionCache *AC, | |||
| 5258 | const Instruction *CxtI, const DominatorTree *DT, | |||
| 5259 | OptimizationRemarkEmitter *ORE, bool UseInstrInfo) { | |||
| 5260 | KnownFPClass Known; | |||
| 5261 | ::computeKnownFPClass(V, Known, InterestedClasses, Depth, | |||
| 5262 | Query(DL, AC, safeCxtI(V, CxtI), DT, UseInstrInfo, ORE), | |||
| 5263 | TLI); | |||
| 5264 | return Known; | |||
| 5265 | } | |||
| 5266 | ||||
| 5267 | Value *llvm::isBytewiseValue(Value *V, const DataLayout &DL) { | |||
| 5268 | ||||
| 5269 | // All byte-wide stores are splatable, even of arbitrary variables. | |||
| 5270 | if (V->getType()->isIntegerTy(8)) | |||
| 5271 | return V; | |||
| 5272 | ||||
| 5273 | LLVMContext &Ctx = V->getContext(); | |||
| 5274 | ||||
| 5275 | // Undef don't care. | |||
| 5276 | auto *UndefInt8 = UndefValue::get(Type::getInt8Ty(Ctx)); | |||
| 5277 | if (isa<UndefValue>(V)) | |||
| 5278 | return UndefInt8; | |||
| 5279 | ||||
| 5280 | // Return Undef for zero-sized type. | |||
| 5281 | if (!DL.getTypeStoreSize(V->getType()).isNonZero()) | |||
| 5282 | return UndefInt8; | |||
| 5283 | ||||
| 5284 | Constant *C = dyn_cast<Constant>(V); | |||
| 5285 | if (!C) { | |||
| 5286 | // Conceptually, we could handle things like: | |||
| 5287 | // %a = zext i8 %X to i16 | |||
| 5288 | // %b = shl i16 %a, 8 | |||
| 5289 | // %c = or i16 %a, %b | |||
| 5290 | // but until there is an example that actually needs this, it doesn't seem | |||
| 5291 | // worth worrying about. | |||
| 5292 | return nullptr; | |||
| 5293 | } | |||
| 5294 | ||||
| 5295 | // Handle 'null' ConstantArrayZero etc. | |||
| 5296 | if (C->isNullValue()) | |||
| 5297 | return Constant::getNullValue(Type::getInt8Ty(Ctx)); | |||
| 5298 | ||||
| 5299 | // Constant floating-point values can be handled as integer values if the | |||
| 5300 | // corresponding integer value is "byteable". An important case is 0.0. | |||
| 5301 | if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { | |||
| 5302 | Type *Ty = nullptr; | |||
| 5303 | if (CFP->getType()->isHalfTy()) | |||
| 5304 | Ty = Type::getInt16Ty(Ctx); | |||
| 5305 | else if (CFP->getType()->isFloatTy()) | |||
| 5306 | Ty = Type::getInt32Ty(Ctx); | |||
| 5307 | else if (CFP->getType()->isDoubleTy()) | |||
| 5308 | Ty = Type::getInt64Ty(Ctx); | |||
| 5309 | // Don't handle long double formats, which have strange constraints. | |||
| 5310 | return Ty ? isBytewiseValue(ConstantExpr::getBitCast(CFP, Ty), DL) | |||
| 5311 | : nullptr; | |||
| 5312 | } | |||
| 5313 | ||||
| 5314 | // We can handle constant integers that are multiple of 8 bits. | |||
| 5315 | if (ConstantInt *CI = dyn_cast<ConstantInt>(C)) { | |||
| 5316 | if (CI->getBitWidth() % 8 == 0) { | |||
| 5317 | assert(CI->getBitWidth() > 8 && "8 bits should be handled above!")(static_cast <bool> (CI->getBitWidth() > 8 && "8 bits should be handled above!") ? void (0) : __assert_fail ("CI->getBitWidth() > 8 && \"8 bits should be handled above!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5317, __extension__ __PRETTY_FUNCTION__ )); | |||
| 5318 | if (!CI->getValue().isSplat(8)) | |||
| 5319 | return nullptr; | |||
| 5320 | return ConstantInt::get(Ctx, CI->getValue().trunc(8)); | |||
| 5321 | } | |||
| 5322 | } | |||
| 5323 | ||||
| 5324 | if (auto *CE = dyn_cast<ConstantExpr>(C)) { | |||
| 5325 | if (CE->getOpcode() == Instruction::IntToPtr) { | |||
| 5326 | if (auto *PtrTy = dyn_cast<PointerType>(CE->getType())) { | |||
| 5327 | unsigned BitWidth = DL.getPointerSizeInBits(PtrTy->getAddressSpace()); | |||
| 5328 | return isBytewiseValue( | |||
| 5329 | ConstantExpr::getIntegerCast(CE->getOperand(0), | |||
| 5330 | Type::getIntNTy(Ctx, BitWidth), false), | |||
| 5331 | DL); | |||
| 5332 | } | |||
| 5333 | } | |||
| 5334 | } | |||
| 5335 | ||||
| 5336 | auto Merge = [&](Value *LHS, Value *RHS) -> Value * { | |||
| 5337 | if (LHS == RHS) | |||
| 5338 | return LHS; | |||
| 5339 | if (!LHS || !RHS) | |||
| 5340 | return nullptr; | |||
| 5341 | if (LHS == UndefInt8) | |||
| 5342 | return RHS; | |||
| 5343 | if (RHS == UndefInt8) | |||
| 5344 | return LHS; | |||
| 5345 | return nullptr; | |||
| 5346 | }; | |||
| 5347 | ||||
| 5348 | if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(C)) { | |||
| 5349 | Value *Val = UndefInt8; | |||
| 5350 | for (unsigned I = 0, E = CA->getNumElements(); I != E; ++I) | |||
| 5351 | if (!(Val = Merge(Val, isBytewiseValue(CA->getElementAsConstant(I), DL)))) | |||
| 5352 | return nullptr; | |||
| 5353 | return Val; | |||
| 5354 | } | |||
| 5355 | ||||
| 5356 | if (isa<ConstantAggregate>(C)) { | |||
| 5357 | Value *Val = UndefInt8; | |||
| 5358 | for (unsigned I = 0, E = C->getNumOperands(); I != E; ++I) | |||
| 5359 | if (!(Val = Merge(Val, isBytewiseValue(C->getOperand(I), DL)))) | |||
| 5360 | return nullptr; | |||
| 5361 | return Val; | |||
| 5362 | } | |||
| 5363 | ||||
| 5364 | // Don't try to handle the handful of other constants. | |||
| 5365 | return nullptr; | |||
| 5366 | } | |||
| 5367 | ||||
| 5368 | // This is the recursive version of BuildSubAggregate. It takes a few different | |||
| 5369 | // arguments. Idxs is the index within the nested struct From that we are | |||
| 5370 | // looking at now (which is of type IndexedType). IdxSkip is the number of | |||
| 5371 | // indices from Idxs that should be left out when inserting into the resulting | |||
| 5372 | // struct. To is the result struct built so far, new insertvalue instructions | |||
| 5373 | // build on that. | |||
| 5374 | static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType, | |||
| 5375 | SmallVectorImpl<unsigned> &Idxs, | |||
| 5376 | unsigned IdxSkip, | |||
| 5377 | Instruction *InsertBefore) { | |||
| 5378 | StructType *STy = dyn_cast<StructType>(IndexedType); | |||
| 5379 | if (STy) { | |||
| 5380 | // Save the original To argument so we can modify it | |||
| 5381 | Value *OrigTo = To; | |||
| 5382 | // General case, the type indexed by Idxs is a struct | |||
| 5383 | for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { | |||
| 5384 | // Process each struct element recursively | |||
| 5385 | Idxs.push_back(i); | |||
| 5386 | Value *PrevTo = To; | |||
| 5387 | To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip, | |||
| 5388 | InsertBefore); | |||
| 5389 | Idxs.pop_back(); | |||
| 5390 | if (!To) { | |||
| 5391 | // Couldn't find any inserted value for this index? Cleanup | |||
| 5392 | while (PrevTo != OrigTo) { | |||
| 5393 | InsertValueInst* Del = cast<InsertValueInst>(PrevTo); | |||
| 5394 | PrevTo = Del->getAggregateOperand(); | |||
| 5395 | Del->eraseFromParent(); | |||
| 5396 | } | |||
| 5397 | // Stop processing elements | |||
| 5398 | break; | |||
| 5399 | } | |||
| 5400 | } | |||
| 5401 | // If we successfully found a value for each of our subaggregates | |||
| 5402 | if (To) | |||
| 5403 | return To; | |||
| 5404 | } | |||
| 5405 | // Base case, the type indexed by SourceIdxs is not a struct, or not all of | |||
| 5406 | // the struct's elements had a value that was inserted directly. In the latter | |||
| 5407 | // case, perhaps we can't determine each of the subelements individually, but | |||
| 5408 | // we might be able to find the complete struct somewhere. | |||
| 5409 | ||||
| 5410 | // Find the value that is at that particular spot | |||
| 5411 | Value *V = FindInsertedValue(From, Idxs); | |||
| 5412 | ||||
| 5413 | if (!V) | |||
| 5414 | return nullptr; | |||
| 5415 | ||||
| 5416 | // Insert the value in the new (sub) aggregate | |||
| 5417 | return InsertValueInst::Create(To, V, ArrayRef(Idxs).slice(IdxSkip), "tmp", | |||
| 5418 | InsertBefore); | |||
| 5419 | } | |||
| 5420 | ||||
| 5421 | // This helper takes a nested struct and extracts a part of it (which is again a | |||
| 5422 | // struct) into a new value. For example, given the struct: | |||
| 5423 | // { a, { b, { c, d }, e } } | |||
| 5424 | // and the indices "1, 1" this returns | |||
| 5425 | // { c, d }. | |||
| 5426 | // | |||
| 5427 | // It does this by inserting an insertvalue for each element in the resulting | |||
| 5428 | // struct, as opposed to just inserting a single struct. This will only work if | |||
| 5429 | // each of the elements of the substruct are known (ie, inserted into From by an | |||
| 5430 | // insertvalue instruction somewhere). | |||
| 5431 | // | |||
| 5432 | // All inserted insertvalue instructions are inserted before InsertBefore | |||
| 5433 | static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range, | |||
| 5434 | Instruction *InsertBefore) { | |||
| 5435 | assert(InsertBefore && "Must have someplace to insert!")(static_cast <bool> (InsertBefore && "Must have someplace to insert!" ) ? void (0) : __assert_fail ("InsertBefore && \"Must have someplace to insert!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5435, __extension__ __PRETTY_FUNCTION__ )); | |||
| 5436 | Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(), | |||
| 5437 | idx_range); | |||
| 5438 | Value *To = PoisonValue::get(IndexedType); | |||
| 5439 | SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end()); | |||
| 5440 | unsigned IdxSkip = Idxs.size(); | |||
| 5441 | ||||
| 5442 | return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore); | |||
| 5443 | } | |||
| 5444 | ||||
| 5445 | /// Given an aggregate and a sequence of indices, see if the scalar value | |||
| 5446 | /// indexed is already around as a register, for example if it was inserted | |||
| 5447 | /// directly into the aggregate. | |||
| 5448 | /// | |||
| 5449 | /// If InsertBefore is not null, this function will duplicate (modified) | |||
| 5450 | /// insertvalues when a part of a nested struct is extracted. | |||
| 5451 | Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range, | |||
| 5452 | Instruction *InsertBefore) { | |||
| 5453 | // Nothing to index? Just return V then (this is useful at the end of our | |||
| 5454 | // recursion). | |||
| 5455 | if (idx_range.empty()) | |||
| 5456 | return V; | |||
| 5457 | // We have indices, so V should have an indexable type. | |||
| 5458 | assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&(static_cast <bool> ((V->getType()->isStructTy() || V->getType()->isArrayTy()) && "Not looking at a struct or array?" ) ? void (0) : __assert_fail ("(V->getType()->isStructTy() || V->getType()->isArrayTy()) && \"Not looking at a struct or array?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5459, __extension__ __PRETTY_FUNCTION__ )) | |||
| 5459 | "Not looking at a struct or array?")(static_cast <bool> ((V->getType()->isStructTy() || V->getType()->isArrayTy()) && "Not looking at a struct or array?" ) ? void (0) : __assert_fail ("(V->getType()->isStructTy() || V->getType()->isArrayTy()) && \"Not looking at a struct or array?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5459, __extension__ __PRETTY_FUNCTION__ )); | |||
| 5460 | assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&(static_cast <bool> (ExtractValueInst::getIndexedType(V ->getType(), idx_range) && "Invalid indices for type?" ) ? void (0) : __assert_fail ("ExtractValueInst::getIndexedType(V->getType(), idx_range) && \"Invalid indices for type?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5461, __extension__ __PRETTY_FUNCTION__ )) | |||
| 5461 | "Invalid indices for type?")(static_cast <bool> (ExtractValueInst::getIndexedType(V ->getType(), idx_range) && "Invalid indices for type?" ) ? void (0) : __assert_fail ("ExtractValueInst::getIndexedType(V->getType(), idx_range) && \"Invalid indices for type?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5461, __extension__ __PRETTY_FUNCTION__ )); | |||
| 5462 | ||||
| 5463 | if (Constant *C = dyn_cast<Constant>(V)) { | |||
| 5464 | C = C->getAggregateElement(idx_range[0]); | |||
| 5465 | if (!C) return nullptr; | |||
| 5466 | return FindInsertedValue(C, idx_range.slice(1), InsertBefore); | |||
| 5467 | } | |||
| 5468 | ||||
| 5469 | if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) { | |||
| 5470 | // Loop the indices for the insertvalue instruction in parallel with the | |||
| 5471 | // requested indices | |||
| 5472 | const unsigned *req_idx = idx_range.begin(); | |||
| 5473 | for (const unsigned *i = I->idx_begin(), *e = I->idx_end(); | |||
| 5474 | i != e; ++i, ++req_idx) { | |||
| 5475 | if (req_idx == idx_range.end()) { | |||
| 5476 | // We can't handle this without inserting insertvalues | |||
| 5477 | if (!InsertBefore) | |||
| 5478 | return nullptr; | |||
| 5479 | ||||
| 5480 | // The requested index identifies a part of a nested aggregate. Handle | |||
| 5481 | // this specially. For example, | |||
| 5482 | // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0 | |||
| 5483 | // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1 | |||
| 5484 | // %C = extractvalue {i32, { i32, i32 } } %B, 1 | |||
| 5485 | // This can be changed into | |||
| 5486 | // %A = insertvalue {i32, i32 } undef, i32 10, 0 | |||
| 5487 | // %C = insertvalue {i32, i32 } %A, i32 11, 1 | |||
| 5488 | // which allows the unused 0,0 element from the nested struct to be | |||
| 5489 | // removed. | |||
| 5490 | return BuildSubAggregate(V, ArrayRef(idx_range.begin(), req_idx), | |||
| 5491 | InsertBefore); | |||
| 5492 | } | |||
| 5493 | ||||
| 5494 | // This insert value inserts something else than what we are looking for. | |||
| 5495 | // See if the (aggregate) value inserted into has the value we are | |||
| 5496 | // looking for, then. | |||
| 5497 | if (*req_idx != *i) | |||
| 5498 | return FindInsertedValue(I->getAggregateOperand(), idx_range, | |||
| 5499 | InsertBefore); | |||
| 5500 | } | |||
| 5501 | // If we end up here, the indices of the insertvalue match with those | |||
| 5502 | // requested (though possibly only partially). Now we recursively look at | |||
| 5503 | // the inserted value, passing any remaining indices. | |||
| 5504 | return FindInsertedValue(I->getInsertedValueOperand(), | |||
| 5505 | ArrayRef(req_idx, idx_range.end()), InsertBefore); | |||
| 5506 | } | |||
| 5507 | ||||
| 5508 | if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) { | |||
| 5509 | // If we're extracting a value from an aggregate that was extracted from | |||
| 5510 | // something else, we can extract from that something else directly instead. | |||
| 5511 | // However, we will need to chain I's indices with the requested indices. | |||
| 5512 | ||||
| 5513 | // Calculate the number of indices required | |||
| 5514 | unsigned size = I->getNumIndices() + idx_range.size(); | |||
| 5515 | // Allocate some space to put the new indices in | |||
| 5516 | SmallVector<unsigned, 5> Idxs; | |||
| 5517 | Idxs.reserve(size); | |||
| 5518 | // Add indices from the extract value instruction | |||
| 5519 | Idxs.append(I->idx_begin(), I->idx_end()); | |||
| 5520 | ||||
| 5521 | // Add requested indices | |||
| 5522 | Idxs.append(idx_range.begin(), idx_range.end()); | |||
| 5523 | ||||
| 5524 | assert(Idxs.size() == size(static_cast <bool> (Idxs.size() == size && "Number of indices added not correct?" ) ? void (0) : __assert_fail ("Idxs.size() == size && \"Number of indices added not correct?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5525, __extension__ __PRETTY_FUNCTION__ )) | |||
| 5525 | && "Number of indices added not correct?")(static_cast <bool> (Idxs.size() == size && "Number of indices added not correct?" ) ? void (0) : __assert_fail ("Idxs.size() == size && \"Number of indices added not correct?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5525, __extension__ __PRETTY_FUNCTION__ )); | |||
| 5526 | ||||
| 5527 | return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore); | |||
| 5528 | } | |||
| 5529 | // Otherwise, we don't know (such as, extracting from a function return value | |||
| 5530 | // or load instruction) | |||
| 5531 | return nullptr; | |||
| 5532 | } | |||
| 5533 | ||||
| 5534 | bool llvm::isGEPBasedOnPointerToString(const GEPOperator *GEP, | |||
| 5535 | unsigned CharSize) { | |||
| 5536 | // Make sure the GEP has exactly three arguments. | |||
| 5537 | if (GEP->getNumOperands() != 3) | |||
| 5538 | return false; | |||
| 5539 | ||||
| 5540 | // Make sure the index-ee is a pointer to array of \p CharSize integers. | |||
| 5541 | // CharSize. | |||
| 5542 | ArrayType *AT = dyn_cast<ArrayType>(GEP->getSourceElementType()); | |||
| 5543 | if (!AT || !AT->getElementType()->isIntegerTy(CharSize)) | |||
| 5544 | return false; | |||
| 5545 | ||||
| 5546 | // Check to make sure that the first operand of the GEP is an integer and | |||
| 5547 | // has value 0 so that we are sure we're indexing into the initializer. | |||
| 5548 | const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1)); | |||
| 5549 | if (!FirstIdx || !FirstIdx->isZero()) | |||
| 5550 | return false; | |||
| 5551 | ||||
| 5552 | return true; | |||
| 5553 | } | |||
| 5554 | ||||
| 5555 | // If V refers to an initialized global constant, set Slice either to | |||
| 5556 | // its initializer if the size of its elements equals ElementSize, or, | |||
| 5557 | // for ElementSize == 8, to its representation as an array of unsiged | |||
| 5558 | // char. Return true on success. | |||
| 5559 | // Offset is in the unit "nr of ElementSize sized elements". | |||
| 5560 | bool llvm::getConstantDataArrayInfo(const Value *V, | |||
| 5561 | ConstantDataArraySlice &Slice, | |||
| 5562 | unsigned ElementSize, uint64_t Offset) { | |||
| 5563 | assert(V && "V should not be null.")(static_cast <bool> (V && "V should not be null." ) ? void (0) : __assert_fail ("V && \"V should not be null.\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5563, __extension__ __PRETTY_FUNCTION__ )); | |||
| 5564 | assert((ElementSize % 8) == 0 &&(static_cast <bool> ((ElementSize % 8) == 0 && "ElementSize expected to be a multiple of the size of a byte." ) ? void (0) : __assert_fail ("(ElementSize % 8) == 0 && \"ElementSize expected to be a multiple of the size of a byte.\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5565, __extension__ __PRETTY_FUNCTION__ )) | |||
| 5565 | "ElementSize expected to be a multiple of the size of a byte.")(static_cast <bool> ((ElementSize % 8) == 0 && "ElementSize expected to be a multiple of the size of a byte." ) ? void (0) : __assert_fail ("(ElementSize % 8) == 0 && \"ElementSize expected to be a multiple of the size of a byte.\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5565, __extension__ __PRETTY_FUNCTION__ )); | |||
| 5566 | unsigned ElementSizeInBytes = ElementSize / 8; | |||
| 5567 | ||||
| 5568 | // Drill down into the pointer expression V, ignoring any intervening | |||
| 5569 | // casts, and determine the identity of the object it references along | |||
| 5570 | // with the cumulative byte offset into it. | |||
| 5571 | const GlobalVariable *GV = | |||
| 5572 | dyn_cast<GlobalVariable>(getUnderlyingObject(V)); | |||
| 5573 | if (!GV
| |||
| 5574 | // Fail if V is not based on constant global object. | |||
| 5575 | return false; | |||
| 5576 | ||||
| 5577 | const DataLayout &DL = GV->getParent()->getDataLayout(); | |||
| 5578 | APInt Off(DL.getIndexTypeSizeInBits(V->getType()), 0); | |||
| 5579 | ||||
| 5580 | if (GV != V->stripAndAccumulateConstantOffsets(DL, Off, | |||
| 5581 | /*AllowNonInbounds*/ true)) | |||
| 5582 | // Fail if a constant offset could not be determined. | |||
| 5583 | return false; | |||
| 5584 | ||||
| 5585 | uint64_t StartIdx = Off.getLimitedValue(); | |||
| 5586 | if (StartIdx == UINT64_MAX(18446744073709551615UL)) | |||
| 5587 | // Fail if the constant offset is excessive. | |||
| 5588 | return false; | |||
| 5589 | ||||
| 5590 | // Off/StartIdx is in the unit of bytes. So we need to convert to number of | |||
| 5591 | // elements. Simply bail out if that isn't possible. | |||
| 5592 | if ((StartIdx % ElementSizeInBytes) != 0) | |||
| 5593 | return false; | |||
| 5594 | ||||
| 5595 | Offset += StartIdx / ElementSizeInBytes; | |||
| 5596 | ConstantDataArray *Array = nullptr; | |||
| 5597 | ArrayType *ArrayTy = nullptr; | |||
| 5598 | ||||
| 5599 | if (GV->getInitializer()->isNullValue()) { | |||
| 5600 | Type *GVTy = GV->getValueType(); | |||
| 5601 | uint64_t SizeInBytes = DL.getTypeStoreSize(GVTy).getFixedValue(); | |||
| 5602 | uint64_t Length = SizeInBytes / ElementSizeInBytes; | |||
| 5603 | ||||
| 5604 | Slice.Array = nullptr; | |||
| 5605 | Slice.Offset = 0; | |||
| 5606 | // Return an empty Slice for undersized constants to let callers | |||
| 5607 | // transform even undefined library calls into simpler, well-defined | |||
| 5608 | // expressions. This is preferable to making the calls although it | |||
| 5609 | // prevents sanitizers from detecting such calls. | |||
| 5610 | Slice.Length = Length < Offset ? 0 : Length - Offset; | |||
| 5611 | return true; | |||
| 5612 | } | |||
| 5613 | ||||
| 5614 | auto *Init = const_cast<Constant *>(GV->getInitializer()); | |||
| 5615 | if (auto *ArrayInit
| |||
| 5616 | Type *InitElTy = ArrayInit->getElementType(); | |||
| 5617 | if (InitElTy->isIntegerTy(ElementSize)) { | |||
| 5618 | // If Init is an initializer for an array of the expected type | |||
| 5619 | // and size, use it as is. | |||
| 5620 | Array = ArrayInit; | |||
| 5621 | ArrayTy = ArrayInit->getType(); | |||
| 5622 | } | |||
| 5623 | } | |||
| 5624 | ||||
| 5625 | if (!Array
| |||
| 5626 | if (ElementSize
| |||
| 5627 | // TODO: Handle conversions to larger integral types. | |||
| 5628 | return false; | |||
| 5629 | ||||
| 5630 | // Otherwise extract the portion of the initializer starting | |||
| 5631 | // at Offset as an array of bytes, and reset Offset. | |||
| 5632 | Init = ReadByteArrayFromGlobal(GV, Offset); | |||
| 5633 | if (!Init) | |||
| 5634 | return false; | |||
| 5635 | ||||
| 5636 | Offset = 0; | |||
| 5637 | Array = dyn_cast<ConstantDataArray>(Init); | |||
| 5638 | ArrayTy = dyn_cast<ArrayType>(Init->getType()); | |||
| 5639 | } | |||
| 5640 | ||||
| 5641 | uint64_t NumElts = ArrayTy->getArrayNumElements(); | |||
| ||||
| 5642 | if (Offset > NumElts) | |||
| 5643 | return false; | |||
| 5644 | ||||
| 5645 | Slice.Array = Array; | |||
| 5646 | Slice.Offset = Offset; | |||
| 5647 | Slice.Length = NumElts - Offset; | |||
| 5648 | return true; | |||
| 5649 | } | |||
| 5650 | ||||
| 5651 | /// Extract bytes from the initializer of the constant array V, which need | |||
| 5652 | /// not be a nul-terminated string. On success, store the bytes in Str and | |||
| 5653 | /// return true. When TrimAtNul is set, Str will contain only the bytes up | |||
| 5654 | /// to but not including the first nul. Return false on failure. | |||
| 5655 | bool llvm::getConstantStringInfo(const Value *V, StringRef &Str, | |||
| 5656 | bool TrimAtNul) { | |||
| 5657 | ConstantDataArraySlice Slice; | |||
| 5658 | if (!getConstantDataArrayInfo(V, Slice, 8)) | |||
| ||||
| 5659 | return false; | |||
| 5660 | ||||
| 5661 | if (Slice.Array == nullptr) { | |||
| 5662 | if (TrimAtNul) { | |||
| 5663 | // Return a nul-terminated string even for an empty Slice. This is | |||
| 5664 | // safe because all existing SimplifyLibcalls callers require string | |||
| 5665 | // arguments and the behavior of the functions they fold is undefined | |||
| 5666 | // otherwise. Folding the calls this way is preferable to making | |||
| 5667 | // the undefined library calls, even though it prevents sanitizers | |||
| 5668 | // from reporting such calls. | |||
| 5669 | Str = StringRef(); | |||
| 5670 | return true; | |||
| 5671 | } | |||
| 5672 | if (Slice.Length == 1) { | |||
| 5673 | Str = StringRef("", 1); | |||
| 5674 | return true; | |||
| 5675 | } | |||
| 5676 | // We cannot instantiate a StringRef as we do not have an appropriate string | |||
| 5677 | // of 0s at hand. | |||
| 5678 | return false; | |||
| 5679 | } | |||
| 5680 | ||||
| 5681 | // Start out with the entire array in the StringRef. | |||
| 5682 | Str = Slice.Array->getAsString(); | |||
| 5683 | // Skip over 'offset' bytes. | |||
| 5684 | Str = Str.substr(Slice.Offset); | |||
| 5685 | ||||
| 5686 | if (TrimAtNul) { | |||
| 5687 | // Trim off the \0 and anything after it. If the array is not nul | |||
| 5688 | // terminated, we just return the whole end of string. The client may know | |||
| 5689 | // some other way that the string is length-bound. | |||
| 5690 | Str = Str.substr(0, Str.find('\0')); | |||
| 5691 | } | |||
| 5692 | return true; | |||
| 5693 | } | |||
| 5694 | ||||
| 5695 | // These next two are very similar to the above, but also look through PHI | |||
| 5696 | // nodes. | |||
| 5697 | // TODO: See if we can integrate these two together. | |||
| 5698 | ||||
| 5699 | /// If we can compute the length of the string pointed to by | |||
| 5700 | /// the specified pointer, return 'len+1'. If we can't, return 0. | |||
| 5701 | static uint64_t GetStringLengthH(const Value *V, | |||
| 5702 | SmallPtrSetImpl<const PHINode*> &PHIs, | |||
| 5703 | unsigned CharSize) { | |||
| 5704 | // Look through noop bitcast instructions. | |||
| 5705 | V = V->stripPointerCasts(); | |||
| 5706 | ||||
| 5707 | // If this is a PHI node, there are two cases: either we have already seen it | |||
| 5708 | // or we haven't. | |||
| 5709 | if (const PHINode *PN = dyn_cast<PHINode>(V)) { | |||
| 5710 | if (!PHIs.insert(PN).second) | |||
| 5711 | return ~0ULL; // already in the set. | |||
| 5712 | ||||
| 5713 | // If it was new, see if all the input strings are the same length. | |||
| 5714 | uint64_t LenSoFar = ~0ULL; | |||
| 5715 | for (Value *IncValue : PN->incoming_values()) { | |||
| 5716 | uint64_t Len = GetStringLengthH(IncValue, PHIs, CharSize); | |||
| 5717 | if (Len == 0) return 0; // Unknown length -> unknown. | |||
| 5718 | ||||
| 5719 | if (Len == ~0ULL) continue; | |||
| 5720 | ||||
| 5721 | if (Len != LenSoFar && LenSoFar != ~0ULL) | |||
| 5722 | return 0; // Disagree -> unknown. | |||
| 5723 | LenSoFar = Len; | |||
| 5724 | } | |||
| 5725 | ||||
| 5726 | // Success, all agree. | |||
| 5727 | return LenSoFar; | |||
| 5728 | } | |||
| 5729 | ||||
| 5730 | // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y) | |||
| 5731 | if (const SelectInst *SI = dyn_cast<SelectInst>(V)) { | |||
| 5732 | uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs, CharSize); | |||
| 5733 | if (Len1 == 0) return 0; | |||
| 5734 | uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs, CharSize); | |||
| 5735 | if (Len2 == 0) return 0; | |||
| 5736 | if (Len1 == ~0ULL) return Len2; | |||
| 5737 | if (Len2 == ~0ULL) return Len1; | |||
| 5738 | if (Len1 != Len2) return 0; | |||
| 5739 | return Len1; | |||
| 5740 | } | |||
| 5741 | ||||
| 5742 | // Otherwise, see if we can read the string. | |||
| 5743 | ConstantDataArraySlice Slice; | |||
| 5744 | if (!getConstantDataArrayInfo(V, Slice, CharSize)) | |||
| 5745 | return 0; | |||
| 5746 | ||||
| 5747 | if (Slice.Array == nullptr) | |||
| 5748 | // Zeroinitializer (including an empty one). | |||
| 5749 | return 1; | |||
| 5750 | ||||
| 5751 | // Search for the first nul character. Return a conservative result even | |||
| 5752 | // when there is no nul. This is safe since otherwise the string function | |||
| 5753 | // being folded such as strlen is undefined, and can be preferable to | |||
| 5754 | // making the undefined library call. | |||
| 5755 | unsigned NullIndex = 0; | |||
| 5756 | for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) { | |||
| 5757 | if (Slice.Array->getElementAsInteger(Slice.Offset + NullIndex) == 0) | |||
| 5758 | break; | |||
| 5759 | } | |||
| 5760 | ||||
| 5761 | return NullIndex + 1; | |||
| 5762 | } | |||
| 5763 | ||||
| 5764 | /// If we can compute the length of the string pointed to by | |||
| 5765 | /// the specified pointer, return 'len+1'. If we can't, return 0. | |||
| 5766 | uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) { | |||
| 5767 | if (!V->getType()->isPointerTy()) | |||
| 5768 | return 0; | |||
| 5769 | ||||
| 5770 | SmallPtrSet<const PHINode*, 32> PHIs; | |||
| 5771 | uint64_t Len = GetStringLengthH(V, PHIs, CharSize); | |||
| 5772 | // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return | |||
| 5773 | // an empty string as a length. | |||
| 5774 | return Len == ~0ULL ? 1 : Len; | |||
| 5775 | } | |||
| 5776 | ||||
| 5777 | const Value * | |||
| 5778 | llvm::getArgumentAliasingToReturnedPointer(const CallBase *Call, | |||
| 5779 | bool MustPreserveNullness) { | |||
| 5780 | assert(Call &&(static_cast <bool> (Call && "getArgumentAliasingToReturnedPointer only works on nonnull calls" ) ? void (0) : __assert_fail ("Call && \"getArgumentAliasingToReturnedPointer only works on nonnull calls\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5781, __extension__ __PRETTY_FUNCTION__ )) | |||
| 5781 | "getArgumentAliasingToReturnedPointer only works on nonnull calls")(static_cast <bool> (Call && "getArgumentAliasingToReturnedPointer only works on nonnull calls" ) ? void (0) : __assert_fail ("Call && \"getArgumentAliasingToReturnedPointer only works on nonnull calls\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5781, __extension__ __PRETTY_FUNCTION__ )); | |||
| 5782 | if (const Value *RV = Call->getReturnedArgOperand()) | |||
| 5783 | return RV; | |||
| 5784 | // This can be used only as a aliasing property. | |||
| 5785 | if (isIntrinsicReturningPointerAliasingArgumentWithoutCapturing( | |||
| 5786 | Call, MustPreserveNullness)) | |||
| 5787 | return Call->getArgOperand(0); | |||
| 5788 | return nullptr; | |||
| 5789 | } | |||
| 5790 | ||||
| 5791 | bool llvm::isIntrinsicReturningPointerAliasingArgumentWithoutCapturing( | |||
| 5792 | const CallBase *Call, bool MustPreserveNullness) { | |||
| 5793 | switch (Call->getIntrinsicID()) { | |||
| 5794 | case Intrinsic::launder_invariant_group: | |||
| 5795 | case Intrinsic::strip_invariant_group: | |||
| 5796 | case Intrinsic::aarch64_irg: | |||
| 5797 | case Intrinsic::aarch64_tagp: | |||
| 5798 | return true; | |||
| 5799 | case Intrinsic::ptrmask: | |||
| 5800 | return !MustPreserveNullness; | |||
| 5801 | default: | |||
| 5802 | return false; | |||
| 5803 | } | |||
| 5804 | } | |||
| 5805 | ||||
| 5806 | /// \p PN defines a loop-variant pointer to an object. Check if the | |||
| 5807 | /// previous iteration of the loop was referring to the same object as \p PN. | |||
| 5808 | static bool isSameUnderlyingObjectInLoop(const PHINode *PN, | |||
| 5809 | const LoopInfo *LI) { | |||
| 5810 | // Find the loop-defined value. | |||
| 5811 | Loop *L = LI->getLoopFor(PN->getParent()); | |||
| 5812 | if (PN->getNumIncomingValues() != 2) | |||
| 5813 | return true; | |||
| 5814 | ||||
| 5815 | // Find the value from previous iteration. | |||
| 5816 | auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0)); | |||
| 5817 | if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L) | |||
| 5818 | PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1)); | |||
| 5819 | if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L) | |||
| 5820 | return true; | |||
| 5821 | ||||
| 5822 | // If a new pointer is loaded in the loop, the pointer references a different | |||
| 5823 | // object in every iteration. E.g.: | |||
| 5824 | // for (i) | |||
| 5825 | // int *p = a[i]; | |||
| 5826 | // ... | |||
| 5827 | if (auto *Load = dyn_cast<LoadInst>(PrevValue)) | |||
| 5828 | if (!L->isLoopInvariant(Load->getPointerOperand())) | |||
| 5829 | return false; | |||
| 5830 | return true; | |||
| 5831 | } | |||
| 5832 | ||||
| 5833 | const Value *llvm::getUnderlyingObject(const Value *V, unsigned MaxLookup) { | |||
| 5834 | if (!V->getType()->isPointerTy()) | |||
| 5835 | return V; | |||
| 5836 | for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) { | |||
| 5837 | if (auto *GEP = dyn_cast<GEPOperator>(V)) { | |||
| 5838 | V = GEP->getPointerOperand(); | |||
| 5839 | } else if (Operator::getOpcode(V) == Instruction::BitCast || | |||
| 5840 | Operator::getOpcode(V) == Instruction::AddrSpaceCast) { | |||
| 5841 | V = cast<Operator>(V)->getOperand(0); | |||
| 5842 | if (!V->getType()->isPointerTy()) | |||
| 5843 | return V; | |||
| 5844 | } else if (auto *GA = dyn_cast<GlobalAlias>(V)) { | |||
| 5845 | if (GA->isInterposable()) | |||
| 5846 | return V; | |||
| 5847 | V = GA->getAliasee(); | |||
| 5848 | } else { | |||
| 5849 | if (auto *PHI = dyn_cast<PHINode>(V)) { | |||
| 5850 | // Look through single-arg phi nodes created by LCSSA. | |||
| 5851 | if (PHI->getNumIncomingValues() == 1) { | |||
| 5852 | V = PHI->getIncomingValue(0); | |||
| 5853 | continue; | |||
| 5854 | } | |||
| 5855 | } else if (auto *Call = dyn_cast<CallBase>(V)) { | |||
| 5856 | // CaptureTracking can know about special capturing properties of some | |||
| 5857 | // intrinsics like launder.invariant.group, that can't be expressed with | |||
| 5858 | // the attributes, but have properties like returning aliasing pointer. | |||
| 5859 | // Because some analysis may assume that nocaptured pointer is not | |||
| 5860 | // returned from some special intrinsic (because function would have to | |||
| 5861 | // be marked with returns attribute), it is crucial to use this function | |||
| 5862 | // because it should be in sync with CaptureTracking. Not using it may | |||
| 5863 | // cause weird miscompilations where 2 aliasing pointers are assumed to | |||
| 5864 | // noalias. | |||
| 5865 | if (auto *RP = getArgumentAliasingToReturnedPointer(Call, false)) { | |||
| 5866 | V = RP; | |||
| 5867 | continue; | |||
| 5868 | } | |||
| 5869 | } | |||
| 5870 | ||||
| 5871 | return V; | |||
| 5872 | } | |||
| 5873 | assert(V->getType()->isPointerTy() && "Unexpected operand type!")(static_cast <bool> (V->getType()->isPointerTy() && "Unexpected operand type!") ? void (0) : __assert_fail ("V->getType()->isPointerTy() && \"Unexpected operand type!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5873, __extension__ __PRETTY_FUNCTION__ )); | |||
| 5874 | } | |||
| 5875 | return V; | |||
| 5876 | } | |||
| 5877 | ||||
| 5878 | void llvm::getUnderlyingObjects(const Value *V, | |||
| 5879 | SmallVectorImpl<const Value *> &Objects, | |||
| 5880 | LoopInfo *LI, unsigned MaxLookup) { | |||
| 5881 | SmallPtrSet<const Value *, 4> Visited; | |||
| 5882 | SmallVector<const Value *, 4> Worklist; | |||
| 5883 | Worklist.push_back(V); | |||
| 5884 | do { | |||
| 5885 | const Value *P = Worklist.pop_back_val(); | |||
| 5886 | P = getUnderlyingObject(P, MaxLookup); | |||
| 5887 | ||||
| 5888 | if (!Visited.insert(P).second) | |||
| 5889 | continue; | |||
| 5890 | ||||
| 5891 | if (auto *SI = dyn_cast<SelectInst>(P)) { | |||
| 5892 | Worklist.push_back(SI->getTrueValue()); | |||
| 5893 | Worklist.push_back(SI->getFalseValue()); | |||
| 5894 | continue; | |||
| 5895 | } | |||
| 5896 | ||||
| 5897 | if (auto *PN = dyn_cast<PHINode>(P)) { | |||
| 5898 | // If this PHI changes the underlying object in every iteration of the | |||
| 5899 | // loop, don't look through it. Consider: | |||
| 5900 | // int **A; | |||
| 5901 | // for (i) { | |||
| 5902 | // Prev = Curr; // Prev = PHI (Prev_0, Curr) | |||
| 5903 | // Curr = A[i]; | |||
| 5904 | // *Prev, *Curr; | |||
| 5905 | // | |||
| 5906 | // Prev is tracking Curr one iteration behind so they refer to different | |||
| 5907 | // underlying objects. | |||
| 5908 | if (!LI || !LI->isLoopHeader(PN->getParent()) || | |||
| 5909 | isSameUnderlyingObjectInLoop(PN, LI)) | |||
| 5910 | append_range(Worklist, PN->incoming_values()); | |||
| 5911 | continue; | |||
| 5912 | } | |||
| 5913 | ||||
| 5914 | Objects.push_back(P); | |||
| 5915 | } while (!Worklist.empty()); | |||
| 5916 | } | |||
| 5917 | ||||
| 5918 | /// This is the function that does the work of looking through basic | |||
| 5919 | /// ptrtoint+arithmetic+inttoptr sequences. | |||
| 5920 | static const Value *getUnderlyingObjectFromInt(const Value *V) { | |||
| 5921 | do { | |||
| 5922 | if (const Operator *U = dyn_cast<Operator>(V)) { | |||
| 5923 | // If we find a ptrtoint, we can transfer control back to the | |||
| 5924 | // regular getUnderlyingObjectFromInt. | |||
| 5925 | if (U->getOpcode() == Instruction::PtrToInt) | |||
| 5926 | return U->getOperand(0); | |||
| 5927 | // If we find an add of a constant, a multiplied value, or a phi, it's | |||
| 5928 | // likely that the other operand will lead us to the base | |||
| 5929 | // object. We don't have to worry about the case where the | |||
| 5930 | // object address is somehow being computed by the multiply, | |||
| 5931 | // because our callers only care when the result is an | |||
| 5932 | // identifiable object. | |||
| 5933 | if (U->getOpcode() != Instruction::Add || | |||
| 5934 | (!isa<ConstantInt>(U->getOperand(1)) && | |||
| 5935 | Operator::getOpcode(U->getOperand(1)) != Instruction::Mul && | |||
| 5936 | !isa<PHINode>(U->getOperand(1)))) | |||
| 5937 | return V; | |||
| 5938 | V = U->getOperand(0); | |||
| 5939 | } else { | |||
| 5940 | return V; | |||
| 5941 | } | |||
| 5942 | assert(V->getType()->isIntegerTy() && "Unexpected operand type!")(static_cast <bool> (V->getType()->isIntegerTy() && "Unexpected operand type!") ? void (0) : __assert_fail ("V->getType()->isIntegerTy() && \"Unexpected operand type!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 5942, __extension__ __PRETTY_FUNCTION__ )); | |||
| 5943 | } while (true); | |||
| 5944 | } | |||
| 5945 | ||||
| 5946 | /// This is a wrapper around getUnderlyingObjects and adds support for basic | |||
| 5947 | /// ptrtoint+arithmetic+inttoptr sequences. | |||
| 5948 | /// It returns false if unidentified object is found in getUnderlyingObjects. | |||
| 5949 | bool llvm::getUnderlyingObjectsForCodeGen(const Value *V, | |||
| 5950 | SmallVectorImpl<Value *> &Objects) { | |||
| 5951 | SmallPtrSet<const Value *, 16> Visited; | |||
| 5952 | SmallVector<const Value *, 4> Working(1, V); | |||
| 5953 | do { | |||
| 5954 | V = Working.pop_back_val(); | |||
| 5955 | ||||
| 5956 | SmallVector<const Value *, 4> Objs; | |||
| 5957 | getUnderlyingObjects(V, Objs); | |||
| 5958 | ||||
| 5959 | for (const Value *V : Objs) { | |||
| 5960 | if (!Visited.insert(V).second) | |||
| 5961 | continue; | |||
| 5962 | if (Operator::getOpcode(V) == Instruction::IntToPtr) { | |||
| 5963 | const Value *O = | |||
| 5964 | getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0)); | |||
| 5965 | if (O->getType()->isPointerTy()) { | |||
| 5966 | Working.push_back(O); | |||
| 5967 | continue; | |||
| 5968 | } | |||
| 5969 | } | |||
| 5970 | // If getUnderlyingObjects fails to find an identifiable object, | |||
| 5971 | // getUnderlyingObjectsForCodeGen also fails for safety. | |||
| 5972 | if (!isIdentifiedObject(V)) { | |||
| 5973 | Objects.clear(); | |||
| 5974 | return false; | |||
| 5975 | } | |||
| 5976 | Objects.push_back(const_cast<Value *>(V)); | |||
| 5977 | } | |||
| 5978 | } while (!Working.empty()); | |||
| 5979 | return true; | |||
| 5980 | } | |||
| 5981 | ||||
| 5982 | AllocaInst *llvm::findAllocaForValue(Value *V, bool OffsetZero) { | |||
| 5983 | AllocaInst *Result = nullptr; | |||
| 5984 | SmallPtrSet<Value *, 4> Visited; | |||
| 5985 | SmallVector<Value *, 4> Worklist; | |||
| 5986 | ||||
| 5987 | auto AddWork = [&](Value *V) { | |||
| 5988 | if (Visited.insert(V).second) | |||
| 5989 | Worklist.push_back(V); | |||
| 5990 | }; | |||
| 5991 | ||||
| 5992 | AddWork(V); | |||
| 5993 | do { | |||
| 5994 | V = Worklist.pop_back_val(); | |||
| 5995 | assert(Visited.count(V))(static_cast <bool> (Visited.count(V)) ? void (0) : __assert_fail ("Visited.count(V)", "llvm/lib/Analysis/ValueTracking.cpp", 5995 , __extension__ __PRETTY_FUNCTION__)); | |||
| 5996 | ||||
| 5997 | if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) { | |||
| 5998 | if (Result && Result != AI) | |||
| 5999 | return nullptr; | |||
| 6000 | Result = AI; | |||
| 6001 | } else if (CastInst *CI = dyn_cast<CastInst>(V)) { | |||
| 6002 | AddWork(CI->getOperand(0)); | |||
| 6003 | } else if (PHINode *PN = dyn_cast<PHINode>(V)) { | |||
| 6004 | for (Value *IncValue : PN->incoming_values()) | |||
| 6005 | AddWork(IncValue); | |||
| 6006 | } else if (auto *SI = dyn_cast<SelectInst>(V)) { | |||
| 6007 | AddWork(SI->getTrueValue()); | |||
| 6008 | AddWork(SI->getFalseValue()); | |||
| 6009 | } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) { | |||
| 6010 | if (OffsetZero && !GEP->hasAllZeroIndices()) | |||
| 6011 | return nullptr; | |||
| 6012 | AddWork(GEP->getPointerOperand()); | |||
| 6013 | } else if (CallBase *CB = dyn_cast<CallBase>(V)) { | |||
| 6014 | Value *Returned = CB->getReturnedArgOperand(); | |||
| 6015 | if (Returned) | |||
| 6016 | AddWork(Returned); | |||
| 6017 | else | |||
| 6018 | return nullptr; | |||
| 6019 | } else { | |||
| 6020 | return nullptr; | |||
| 6021 | } | |||
| 6022 | } while (!Worklist.empty()); | |||
| 6023 | ||||
| 6024 | return Result; | |||
| 6025 | } | |||
| 6026 | ||||
| 6027 | static bool onlyUsedByLifetimeMarkersOrDroppableInstsHelper( | |||
| 6028 | const Value *V, bool AllowLifetime, bool AllowDroppable) { | |||
| 6029 | for (const User *U : V->users()) { | |||
| 6030 | const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U); | |||
| 6031 | if (!II) | |||
| 6032 | return false; | |||
| 6033 | ||||
| 6034 | if (AllowLifetime && II->isLifetimeStartOrEnd()) | |||
| 6035 | continue; | |||
| 6036 | ||||
| 6037 | if (AllowDroppable && II->isDroppable()) | |||
| 6038 | continue; | |||
| 6039 | ||||
| 6040 | return false; | |||
| 6041 | } | |||
| 6042 | return true; | |||
| 6043 | } | |||
| 6044 | ||||
| 6045 | bool llvm::onlyUsedByLifetimeMarkers(const Value *V) { | |||
| 6046 | return onlyUsedByLifetimeMarkersOrDroppableInstsHelper( | |||
| 6047 | V, /* AllowLifetime */ true, /* AllowDroppable */ false); | |||
| 6048 | } | |||
| 6049 | bool llvm::onlyUsedByLifetimeMarkersOrDroppableInsts(const Value *V) { | |||
| 6050 | return onlyUsedByLifetimeMarkersOrDroppableInstsHelper( | |||
| 6051 | V, /* AllowLifetime */ true, /* AllowDroppable */ true); | |||
| 6052 | } | |||
| 6053 | ||||
| 6054 | bool llvm::mustSuppressSpeculation(const LoadInst &LI) { | |||
| 6055 | if (!LI.isUnordered()) | |||
| 6056 | return true; | |||
| 6057 | const Function &F = *LI.getFunction(); | |||
| 6058 | // Speculative load may create a race that did not exist in the source. | |||
| 6059 | return F.hasFnAttribute(Attribute::SanitizeThread) || | |||
| 6060 | // Speculative load may load data from dirty regions. | |||
| 6061 | F.hasFnAttribute(Attribute::SanitizeAddress) || | |||
| 6062 | F.hasFnAttribute(Attribute::SanitizeHWAddress); | |||
| 6063 | } | |||
| 6064 | ||||
| 6065 | bool llvm::isSafeToSpeculativelyExecute(const Instruction *Inst, | |||
| 6066 | const Instruction *CtxI, | |||
| 6067 | AssumptionCache *AC, | |||
| 6068 | const DominatorTree *DT, | |||
| 6069 | const TargetLibraryInfo *TLI) { | |||
| 6070 | return isSafeToSpeculativelyExecuteWithOpcode(Inst->getOpcode(), Inst, CtxI, | |||
| 6071 | AC, DT, TLI); | |||
| 6072 | } | |||
| 6073 | ||||
| 6074 | bool llvm::isSafeToSpeculativelyExecuteWithOpcode( | |||
| 6075 | unsigned Opcode, const Instruction *Inst, const Instruction *CtxI, | |||
| 6076 | AssumptionCache *AC, const DominatorTree *DT, | |||
| 6077 | const TargetLibraryInfo *TLI) { | |||
| 6078 | #ifndef NDEBUG | |||
| 6079 | if (Inst->getOpcode() != Opcode) { | |||
| 6080 | // Check that the operands are actually compatible with the Opcode override. | |||
| 6081 | auto hasEqualReturnAndLeadingOperandTypes = | |||
| 6082 | [](const Instruction *Inst, unsigned NumLeadingOperands) { | |||
| 6083 | if (Inst->getNumOperands() < NumLeadingOperands) | |||
| 6084 | return false; | |||
| 6085 | const Type *ExpectedType = Inst->getType(); | |||
| 6086 | for (unsigned ItOp = 0; ItOp < NumLeadingOperands; ++ItOp) | |||
| 6087 | if (Inst->getOperand(ItOp)->getType() != ExpectedType) | |||
| 6088 | return false; | |||
| 6089 | return true; | |||
| 6090 | }; | |||
| 6091 | assert(!Instruction::isBinaryOp(Opcode) ||(static_cast <bool> (!Instruction::isBinaryOp(Opcode) || hasEqualReturnAndLeadingOperandTypes(Inst, 2)) ? void (0) : __assert_fail ("!Instruction::isBinaryOp(Opcode) || hasEqualReturnAndLeadingOperandTypes(Inst, 2)" , "llvm/lib/Analysis/ValueTracking.cpp", 6092, __extension__ __PRETTY_FUNCTION__ )) | |||
| 6092 | hasEqualReturnAndLeadingOperandTypes(Inst, 2))(static_cast <bool> (!Instruction::isBinaryOp(Opcode) || hasEqualReturnAndLeadingOperandTypes(Inst, 2)) ? void (0) : __assert_fail ("!Instruction::isBinaryOp(Opcode) || hasEqualReturnAndLeadingOperandTypes(Inst, 2)" , "llvm/lib/Analysis/ValueTracking.cpp", 6092, __extension__ __PRETTY_FUNCTION__ )); | |||
| 6093 | assert(!Instruction::isUnaryOp(Opcode) ||(static_cast <bool> (!Instruction::isUnaryOp(Opcode) || hasEqualReturnAndLeadingOperandTypes(Inst, 1)) ? void (0) : __assert_fail ("!Instruction::isUnaryOp(Opcode) || hasEqualReturnAndLeadingOperandTypes(Inst, 1)" , "llvm/lib/Analysis/ValueTracking.cpp", 6094, __extension__ __PRETTY_FUNCTION__ )) | |||
| 6094 | hasEqualReturnAndLeadingOperandTypes(Inst, 1))(static_cast <bool> (!Instruction::isUnaryOp(Opcode) || hasEqualReturnAndLeadingOperandTypes(Inst, 1)) ? void (0) : __assert_fail ("!Instruction::isUnaryOp(Opcode) || hasEqualReturnAndLeadingOperandTypes(Inst, 1)" , "llvm/lib/Analysis/ValueTracking.cpp", 6094, __extension__ __PRETTY_FUNCTION__ )); | |||
| 6095 | } | |||
| 6096 | #endif | |||
| 6097 | ||||
| 6098 | switch (Opcode) { | |||
| 6099 | default: | |||
| 6100 | return true; | |||
| 6101 | case Instruction::UDiv: | |||
| 6102 | case Instruction::URem: { | |||
| 6103 | // x / y is undefined if y == 0. | |||
| 6104 | const APInt *V; | |||
| 6105 | if (match(Inst->getOperand(1), m_APInt(V))) | |||
| 6106 | return *V != 0; | |||
| 6107 | return false; | |||
| 6108 | } | |||
| 6109 | case Instruction::SDiv: | |||
| 6110 | case Instruction::SRem: { | |||
| 6111 | // x / y is undefined if y == 0 or x == INT_MIN and y == -1 | |||
| 6112 | const APInt *Numerator, *Denominator; | |||
| 6113 | if (!match(Inst->getOperand(1), m_APInt(Denominator))) | |||
| 6114 | return false; | |||
| 6115 | // We cannot hoist this division if the denominator is 0. | |||
| 6116 | if (*Denominator == 0) | |||
| 6117 | return false; | |||
| 6118 | // It's safe to hoist if the denominator is not 0 or -1. | |||
| 6119 | if (!Denominator->isAllOnes()) | |||
| 6120 | return true; | |||
| 6121 | // At this point we know that the denominator is -1. It is safe to hoist as | |||
| 6122 | // long we know that the numerator is not INT_MIN. | |||
| 6123 | if (match(Inst->getOperand(0), m_APInt(Numerator))) | |||
| 6124 | return !Numerator->isMinSignedValue(); | |||
| 6125 | // The numerator *might* be MinSignedValue. | |||
| 6126 | return false; | |||
| 6127 | } | |||
| 6128 | case Instruction::Load: { | |||
| 6129 | const LoadInst *LI = dyn_cast<LoadInst>(Inst); | |||
| 6130 | if (!LI) | |||
| 6131 | return false; | |||
| 6132 | if (mustSuppressSpeculation(*LI)) | |||
| 6133 | return false; | |||
| 6134 | const DataLayout &DL = LI->getModule()->getDataLayout(); | |||
| 6135 | return isDereferenceableAndAlignedPointer(LI->getPointerOperand(), | |||
| 6136 | LI->getType(), LI->getAlign(), DL, | |||
| 6137 | CtxI, AC, DT, TLI); | |||
| 6138 | } | |||
| 6139 | case Instruction::Call: { | |||
| 6140 | auto *CI = dyn_cast<const CallInst>(Inst); | |||
| 6141 | if (!CI) | |||
| 6142 | return false; | |||
| 6143 | const Function *Callee = CI->getCalledFunction(); | |||
| 6144 | ||||
| 6145 | // The called function could have undefined behavior or side-effects, even | |||
| 6146 | // if marked readnone nounwind. | |||
| 6147 | return Callee && Callee->isSpeculatable(); | |||
| 6148 | } | |||
| 6149 | case Instruction::VAArg: | |||
| 6150 | case Instruction::Alloca: | |||
| 6151 | case Instruction::Invoke: | |||
| 6152 | case Instruction::CallBr: | |||
| 6153 | case Instruction::PHI: | |||
| 6154 | case Instruction::Store: | |||
| 6155 | case Instruction::Ret: | |||
| 6156 | case Instruction::Br: | |||
| 6157 | case Instruction::IndirectBr: | |||
| 6158 | case Instruction::Switch: | |||
| 6159 | case Instruction::Unreachable: | |||
| 6160 | case Instruction::Fence: | |||
| 6161 | case Instruction::AtomicRMW: | |||
| 6162 | case Instruction::AtomicCmpXchg: | |||
| 6163 | case Instruction::LandingPad: | |||
| 6164 | case Instruction::Resume: | |||
| 6165 | case Instruction::CatchSwitch: | |||
| 6166 | case Instruction::CatchPad: | |||
| 6167 | case Instruction::CatchRet: | |||
| 6168 | case Instruction::CleanupPad: | |||
| 6169 | case Instruction::CleanupRet: | |||
| 6170 | return false; // Misc instructions which have effects | |||
| 6171 | } | |||
| 6172 | } | |||
| 6173 | ||||
| 6174 | bool llvm::mayHaveNonDefUseDependency(const Instruction &I) { | |||
| 6175 | if (I.mayReadOrWriteMemory()) | |||
| 6176 | // Memory dependency possible | |||
| 6177 | return true; | |||
| 6178 | if (!isSafeToSpeculativelyExecute(&I)) | |||
| 6179 | // Can't move above a maythrow call or infinite loop. Or if an | |||
| 6180 | // inalloca alloca, above a stacksave call. | |||
| 6181 | return true; | |||
| 6182 | if (!isGuaranteedToTransferExecutionToSuccessor(&I)) | |||
| 6183 | // 1) Can't reorder two inf-loop calls, even if readonly | |||
| 6184 | // 2) Also can't reorder an inf-loop call below a instruction which isn't | |||
| 6185 | // safe to speculative execute. (Inverse of above) | |||
| 6186 | return true; | |||
| 6187 | return false; | |||
| 6188 | } | |||
| 6189 | ||||
| 6190 | /// Convert ConstantRange OverflowResult into ValueTracking OverflowResult. | |||
| 6191 | static OverflowResult mapOverflowResult(ConstantRange::OverflowResult OR) { | |||
| 6192 | switch (OR) { | |||
| 6193 | case ConstantRange::OverflowResult::MayOverflow: | |||
| 6194 | return OverflowResult::MayOverflow; | |||
| 6195 | case ConstantRange::OverflowResult::AlwaysOverflowsLow: | |||
| 6196 | return OverflowResult::AlwaysOverflowsLow; | |||
| 6197 | case ConstantRange::OverflowResult::AlwaysOverflowsHigh: | |||
| 6198 | return OverflowResult::AlwaysOverflowsHigh; | |||
| 6199 | case ConstantRange::OverflowResult::NeverOverflows: | |||
| 6200 | return OverflowResult::NeverOverflows; | |||
| 6201 | } | |||
| 6202 | llvm_unreachable("Unknown OverflowResult")::llvm::llvm_unreachable_internal("Unknown OverflowResult", "llvm/lib/Analysis/ValueTracking.cpp" , 6202); | |||
| 6203 | } | |||
| 6204 | ||||
| 6205 | /// Combine constant ranges from computeConstantRange() and computeKnownBits(). | |||
| 6206 | static ConstantRange computeConstantRangeIncludingKnownBits( | |||
| 6207 | const Value *V, bool ForSigned, const DataLayout &DL, unsigned Depth, | |||
| 6208 | AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT, | |||
| 6209 | OptimizationRemarkEmitter *ORE = nullptr, bool UseInstrInfo = true) { | |||
| 6210 | KnownBits Known = computeKnownBits( | |||
| 6211 | V, DL, Depth, AC, CxtI, DT, ORE, UseInstrInfo); | |||
| 6212 | ConstantRange CR1 = ConstantRange::fromKnownBits(Known, ForSigned); | |||
| 6213 | ConstantRange CR2 = computeConstantRange(V, ForSigned, UseInstrInfo); | |||
| 6214 | ConstantRange::PreferredRangeType RangeType = | |||
| 6215 | ForSigned ? ConstantRange::Signed : ConstantRange::Unsigned; | |||
| 6216 | return CR1.intersectWith(CR2, RangeType); | |||
| 6217 | } | |||
| 6218 | ||||
| 6219 | OverflowResult llvm::computeOverflowForUnsignedMul( | |||
| 6220 | const Value *LHS, const Value *RHS, const DataLayout &DL, | |||
| 6221 | AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT, | |||
| 6222 | bool UseInstrInfo) { | |||
| 6223 | KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT, | |||
| 6224 | nullptr, UseInstrInfo); | |||
| 6225 | KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT, | |||
| 6226 | nullptr, UseInstrInfo); | |||
| 6227 | ConstantRange LHSRange = ConstantRange::fromKnownBits(LHSKnown, false); | |||
| 6228 | ConstantRange RHSRange = ConstantRange::fromKnownBits(RHSKnown, false); | |||
| 6229 | return mapOverflowResult(LHSRange.unsignedMulMayOverflow(RHSRange)); | |||
| 6230 | } | |||
| 6231 | ||||
| 6232 | OverflowResult | |||
| 6233 | llvm::computeOverflowForSignedMul(const Value *LHS, const Value *RHS, | |||
| 6234 | const DataLayout &DL, AssumptionCache *AC, | |||
| 6235 | const Instruction *CxtI, | |||
| 6236 | const DominatorTree *DT, bool UseInstrInfo) { | |||
| 6237 | // Multiplying n * m significant bits yields a result of n + m significant | |||
| 6238 | // bits. If the total number of significant bits does not exceed the | |||
| 6239 | // result bit width (minus 1), there is no overflow. | |||
| 6240 | // This means if we have enough leading sign bits in the operands | |||
| 6241 | // we can guarantee that the result does not overflow. | |||
| 6242 | // Ref: "Hacker's Delight" by Henry Warren | |||
| 6243 | unsigned BitWidth = LHS->getType()->getScalarSizeInBits(); | |||
| 6244 | ||||
| 6245 | // Note that underestimating the number of sign bits gives a more | |||
| 6246 | // conservative answer. | |||
| 6247 | unsigned SignBits = ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) + | |||
| 6248 | ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT); | |||
| 6249 | ||||
| 6250 | // First handle the easy case: if we have enough sign bits there's | |||
| 6251 | // definitely no overflow. | |||
| 6252 | if (SignBits > BitWidth + 1) | |||
| 6253 | return OverflowResult::NeverOverflows; | |||
| 6254 | ||||
| 6255 | // There are two ambiguous cases where there can be no overflow: | |||
| 6256 | // SignBits == BitWidth + 1 and | |||
| 6257 | // SignBits == BitWidth | |||
| 6258 | // The second case is difficult to check, therefore we only handle the | |||
| 6259 | // first case. | |||
| 6260 | if (SignBits == BitWidth + 1) { | |||
| 6261 | // It overflows only when both arguments are negative and the true | |||
| 6262 | // product is exactly the minimum negative number. | |||
| 6263 | // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000 | |||
| 6264 | // For simplicity we just check if at least one side is not negative. | |||
| 6265 | KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT, | |||
| 6266 | nullptr, UseInstrInfo); | |||
| 6267 | KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT, | |||
| 6268 | nullptr, UseInstrInfo); | |||
| 6269 | if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative()) | |||
| 6270 | return OverflowResult::NeverOverflows; | |||
| 6271 | } | |||
| 6272 | return OverflowResult::MayOverflow; | |||
| 6273 | } | |||
| 6274 | ||||
| 6275 | OverflowResult llvm::computeOverflowForUnsignedAdd( | |||
| 6276 | const Value *LHS, const Value *RHS, const DataLayout &DL, | |||
| 6277 | AssumptionCache *AC, const Instruction *CxtI, const DominatorTree *DT, | |||
| 6278 | bool UseInstrInfo) { | |||
| 6279 | ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( | |||
| 6280 | LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT, | |||
| 6281 | nullptr, UseInstrInfo); | |||
| 6282 | ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( | |||
| 6283 | RHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT, | |||
| 6284 | nullptr, UseInstrInfo); | |||
| 6285 | return mapOverflowResult(LHSRange.unsignedAddMayOverflow(RHSRange)); | |||
| 6286 | } | |||
| 6287 | ||||
| 6288 | static OverflowResult computeOverflowForSignedAdd(const Value *LHS, | |||
| 6289 | const Value *RHS, | |||
| 6290 | const AddOperator *Add, | |||
| 6291 | const DataLayout &DL, | |||
| 6292 | AssumptionCache *AC, | |||
| 6293 | const Instruction *CxtI, | |||
| 6294 | const DominatorTree *DT) { | |||
| 6295 | if (Add && Add->hasNoSignedWrap()) { | |||
| 6296 | return OverflowResult::NeverOverflows; | |||
| 6297 | } | |||
| 6298 | ||||
| 6299 | // If LHS and RHS each have at least two sign bits, the addition will look | |||
| 6300 | // like | |||
| 6301 | // | |||
| 6302 | // XX..... + | |||
| 6303 | // YY..... | |||
| 6304 | // | |||
| 6305 | // If the carry into the most significant position is 0, X and Y can't both | |||
| 6306 | // be 1 and therefore the carry out of the addition is also 0. | |||
| 6307 | // | |||
| 6308 | // If the carry into the most significant position is 1, X and Y can't both | |||
| 6309 | // be 0 and therefore the carry out of the addition is also 1. | |||
| 6310 | // | |||
| 6311 | // Since the carry into the most significant position is always equal to | |||
| 6312 | // the carry out of the addition, there is no signed overflow. | |||
| 6313 | if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 && | |||
| 6314 | ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1) | |||
| 6315 | return OverflowResult::NeverOverflows; | |||
| 6316 | ||||
| 6317 | ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( | |||
| 6318 | LHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); | |||
| 6319 | ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( | |||
| 6320 | RHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); | |||
| 6321 | OverflowResult OR = | |||
| 6322 | mapOverflowResult(LHSRange.signedAddMayOverflow(RHSRange)); | |||
| 6323 | if (OR != OverflowResult::MayOverflow) | |||
| 6324 | return OR; | |||
| 6325 | ||||
| 6326 | // The remaining code needs Add to be available. Early returns if not so. | |||
| 6327 | if (!Add) | |||
| 6328 | return OverflowResult::MayOverflow; | |||
| 6329 | ||||
| 6330 | // If the sign of Add is the same as at least one of the operands, this add | |||
| 6331 | // CANNOT overflow. If this can be determined from the known bits of the | |||
| 6332 | // operands the above signedAddMayOverflow() check will have already done so. | |||
| 6333 | // The only other way to improve on the known bits is from an assumption, so | |||
| 6334 | // call computeKnownBitsFromAssume() directly. | |||
| 6335 | bool LHSOrRHSKnownNonNegative = | |||
| 6336 | (LHSRange.isAllNonNegative() || RHSRange.isAllNonNegative()); | |||
| 6337 | bool LHSOrRHSKnownNegative = | |||
| 6338 | (LHSRange.isAllNegative() || RHSRange.isAllNegative()); | |||
| 6339 | if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) { | |||
| 6340 | KnownBits AddKnown(LHSRange.getBitWidth()); | |||
| 6341 | computeKnownBitsFromAssume( | |||
| 6342 | Add, AddKnown, /*Depth=*/0, Query(DL, AC, CxtI, DT, true)); | |||
| 6343 | if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) || | |||
| 6344 | (AddKnown.isNegative() && LHSOrRHSKnownNegative)) | |||
| 6345 | return OverflowResult::NeverOverflows; | |||
| 6346 | } | |||
| 6347 | ||||
| 6348 | return OverflowResult::MayOverflow; | |||
| 6349 | } | |||
| 6350 | ||||
| 6351 | OverflowResult llvm::computeOverflowForUnsignedSub(const Value *LHS, | |||
| 6352 | const Value *RHS, | |||
| 6353 | const DataLayout &DL, | |||
| 6354 | AssumptionCache *AC, | |||
| 6355 | const Instruction *CxtI, | |||
| 6356 | const DominatorTree *DT) { | |||
| 6357 | // X - (X % ?) | |||
| 6358 | // The remainder of a value can't have greater magnitude than itself, | |||
| 6359 | // so the subtraction can't overflow. | |||
| 6360 | ||||
| 6361 | // X - (X -nuw ?) | |||
| 6362 | // In the minimal case, this would simplify to "?", so there's no subtract | |||
| 6363 | // at all. But if this analysis is used to peek through casts, for example, | |||
| 6364 | // then determining no-overflow may allow other transforms. | |||
| 6365 | ||||
| 6366 | // TODO: There are other patterns like this. | |||
| 6367 | // See simplifyICmpWithBinOpOnLHS() for candidates. | |||
| 6368 | if (match(RHS, m_URem(m_Specific(LHS), m_Value())) || | |||
| 6369 | match(RHS, m_NUWSub(m_Specific(LHS), m_Value()))) | |||
| 6370 | if (isGuaranteedNotToBeUndefOrPoison(LHS, AC, CxtI, DT)) | |||
| 6371 | return OverflowResult::NeverOverflows; | |||
| 6372 | ||||
| 6373 | // Checking for conditions implied by dominating conditions may be expensive. | |||
| 6374 | // Limit it to usub_with_overflow calls for now. | |||
| 6375 | if (match(CxtI, | |||
| 6376 | m_Intrinsic<Intrinsic::usub_with_overflow>(m_Value(), m_Value()))) | |||
| 6377 | if (auto C = | |||
| 6378 | isImpliedByDomCondition(CmpInst::ICMP_UGE, LHS, RHS, CxtI, DL)) { | |||
| 6379 | if (*C) | |||
| 6380 | return OverflowResult::NeverOverflows; | |||
| 6381 | return OverflowResult::AlwaysOverflowsLow; | |||
| 6382 | } | |||
| 6383 | ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( | |||
| 6384 | LHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT); | |||
| 6385 | ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( | |||
| 6386 | RHS, /*ForSigned=*/false, DL, /*Depth=*/0, AC, CxtI, DT); | |||
| 6387 | return mapOverflowResult(LHSRange.unsignedSubMayOverflow(RHSRange)); | |||
| 6388 | } | |||
| 6389 | ||||
| 6390 | OverflowResult llvm::computeOverflowForSignedSub(const Value *LHS, | |||
| 6391 | const Value *RHS, | |||
| 6392 | const DataLayout &DL, | |||
| 6393 | AssumptionCache *AC, | |||
| 6394 | const Instruction *CxtI, | |||
| 6395 | const DominatorTree *DT) { | |||
| 6396 | // X - (X % ?) | |||
| 6397 | // The remainder of a value can't have greater magnitude than itself, | |||
| 6398 | // so the subtraction can't overflow. | |||
| 6399 | ||||
| 6400 | // X - (X -nsw ?) | |||
| 6401 | // In the minimal case, this would simplify to "?", so there's no subtract | |||
| 6402 | // at all. But if this analysis is used to peek through casts, for example, | |||
| 6403 | // then determining no-overflow may allow other transforms. | |||
| 6404 | if (match(RHS, m_SRem(m_Specific(LHS), m_Value())) || | |||
| 6405 | match(RHS, m_NSWSub(m_Specific(LHS), m_Value()))) | |||
| 6406 | if (isGuaranteedNotToBeUndefOrPoison(LHS, AC, CxtI, DT)) | |||
| 6407 | return OverflowResult::NeverOverflows; | |||
| 6408 | ||||
| 6409 | // If LHS and RHS each have at least two sign bits, the subtraction | |||
| 6410 | // cannot overflow. | |||
| 6411 | if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 && | |||
| 6412 | ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1) | |||
| 6413 | return OverflowResult::NeverOverflows; | |||
| 6414 | ||||
| 6415 | ConstantRange LHSRange = computeConstantRangeIncludingKnownBits( | |||
| 6416 | LHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); | |||
| 6417 | ConstantRange RHSRange = computeConstantRangeIncludingKnownBits( | |||
| 6418 | RHS, /*ForSigned=*/true, DL, /*Depth=*/0, AC, CxtI, DT); | |||
| 6419 | return mapOverflowResult(LHSRange.signedSubMayOverflow(RHSRange)); | |||
| 6420 | } | |||
| 6421 | ||||
| 6422 | bool llvm::isOverflowIntrinsicNoWrap(const WithOverflowInst *WO, | |||
| 6423 | const DominatorTree &DT) { | |||
| 6424 | SmallVector<const BranchInst *, 2> GuardingBranches; | |||
| 6425 | SmallVector<const ExtractValueInst *, 2> Results; | |||
| 6426 | ||||
| 6427 | for (const User *U : WO->users()) { | |||
| 6428 | if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) { | |||
| 6429 | assert(EVI->getNumIndices() == 1 && "Obvious from CI's type")(static_cast <bool> (EVI->getNumIndices() == 1 && "Obvious from CI's type") ? void (0) : __assert_fail ("EVI->getNumIndices() == 1 && \"Obvious from CI's type\"" , "llvm/lib/Analysis/ValueTracking.cpp", 6429, __extension__ __PRETTY_FUNCTION__ )); | |||
| 6430 | ||||
| 6431 | if (EVI->getIndices()[0] == 0) | |||
| 6432 | Results.push_back(EVI); | |||
| 6433 | else { | |||
| 6434 | assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type")(static_cast <bool> (EVI->getIndices()[0] == 1 && "Obvious from CI's type") ? void (0) : __assert_fail ("EVI->getIndices()[0] == 1 && \"Obvious from CI's type\"" , "llvm/lib/Analysis/ValueTracking.cpp", 6434, __extension__ __PRETTY_FUNCTION__ )); | |||
| 6435 | ||||
| 6436 | for (const auto *U : EVI->users()) | |||
| 6437 | if (const auto *B = dyn_cast<BranchInst>(U)) { | |||
| 6438 | assert(B->isConditional() && "How else is it using an i1?")(static_cast <bool> (B->isConditional() && "How else is it using an i1?" ) ? void (0) : __assert_fail ("B->isConditional() && \"How else is it using an i1?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 6438, __extension__ __PRETTY_FUNCTION__ )); | |||
| 6439 | GuardingBranches.push_back(B); | |||
| 6440 | } | |||
| 6441 | } | |||
| 6442 | } else { | |||
| 6443 | // We are using the aggregate directly in a way we don't want to analyze | |||
| 6444 | // here (storing it to a global, say). | |||
| 6445 | return false; | |||
| 6446 | } | |||
| 6447 | } | |||
| 6448 | ||||
| 6449 | auto AllUsesGuardedByBranch = [&](const BranchInst *BI) { | |||
| 6450 | BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1)); | |||
| 6451 | if (!NoWrapEdge.isSingleEdge()) | |||
| 6452 | return false; | |||
| 6453 | ||||
| 6454 | // Check if all users of the add are provably no-wrap. | |||
| 6455 | for (const auto *Result : Results) { | |||
| 6456 | // If the extractvalue itself is not executed on overflow, the we don't | |||
| 6457 | // need to check each use separately, since domination is transitive. | |||
| 6458 | if (DT.dominates(NoWrapEdge, Result->getParent())) | |||
| 6459 | continue; | |||
| 6460 | ||||
| 6461 | for (const auto &RU : Result->uses()) | |||
| 6462 | if (!DT.dominates(NoWrapEdge, RU)) | |||
| 6463 | return false; | |||
| 6464 | } | |||
| 6465 | ||||
| 6466 | return true; | |||
| 6467 | }; | |||
| 6468 | ||||
| 6469 | return llvm::any_of(GuardingBranches, AllUsesGuardedByBranch); | |||
| 6470 | } | |||
| 6471 | ||||
| 6472 | /// Shifts return poison if shiftwidth is larger than the bitwidth. | |||
| 6473 | static bool shiftAmountKnownInRange(const Value *ShiftAmount) { | |||
| 6474 | auto *C = dyn_cast<Constant>(ShiftAmount); | |||
| 6475 | if (!C) | |||
| 6476 | return false; | |||
| 6477 | ||||
| 6478 | // Shifts return poison if shiftwidth is larger than the bitwidth. | |||
| 6479 | SmallVector<const Constant *, 4> ShiftAmounts; | |||
| 6480 | if (auto *FVTy = dyn_cast<FixedVectorType>(C->getType())) { | |||
| 6481 | unsigned NumElts = FVTy->getNumElements(); | |||
| 6482 | for (unsigned i = 0; i < NumElts; ++i) | |||
| 6483 | ShiftAmounts.push_back(C->getAggregateElement(i)); | |||
| 6484 | } else if (isa<ScalableVectorType>(C->getType())) | |||
| 6485 | return false; // Can't tell, just return false to be safe | |||
| 6486 | else | |||
| 6487 | ShiftAmounts.push_back(C); | |||
| 6488 | ||||
| 6489 | bool Safe = llvm::all_of(ShiftAmounts, [](const Constant *C) { | |||
| 6490 | auto *CI = dyn_cast_or_null<ConstantInt>(C); | |||
| 6491 | return CI && CI->getValue().ult(C->getType()->getIntegerBitWidth()); | |||
| 6492 | }); | |||
| 6493 | ||||
| 6494 | return Safe; | |||
| 6495 | } | |||
| 6496 | ||||
| 6497 | static bool canCreateUndefOrPoison(const Operator *Op, bool PoisonOnly, | |||
| 6498 | bool ConsiderFlagsAndMetadata) { | |||
| 6499 | ||||
| 6500 | if (ConsiderFlagsAndMetadata && Op->hasPoisonGeneratingFlagsOrMetadata()) | |||
| 6501 | return true; | |||
| 6502 | ||||
| 6503 | unsigned Opcode = Op->getOpcode(); | |||
| 6504 | ||||
| 6505 | // Check whether opcode is a poison/undef-generating operation | |||
| 6506 | switch (Opcode) { | |||
| 6507 | case Instruction::Shl: | |||
| 6508 | case Instruction::AShr: | |||
| 6509 | case Instruction::LShr: | |||
| 6510 | return !shiftAmountKnownInRange(Op->getOperand(1)); | |||
| 6511 | case Instruction::FPToSI: | |||
| 6512 | case Instruction::FPToUI: | |||
| 6513 | // fptosi/ui yields poison if the resulting value does not fit in the | |||
| 6514 | // destination type. | |||
| 6515 | return true; | |||
| 6516 | case Instruction::Call: | |||
| 6517 | if (auto *II = dyn_cast<IntrinsicInst>(Op)) { | |||
| 6518 | switch (II->getIntrinsicID()) { | |||
| 6519 | // TODO: Add more intrinsics. | |||
| 6520 | case Intrinsic::ctlz: | |||
| 6521 | case Intrinsic::cttz: | |||
| 6522 | case Intrinsic::abs: | |||
| 6523 | if (cast<ConstantInt>(II->getArgOperand(1))->isNullValue()) | |||
| 6524 | return false; | |||
| 6525 | break; | |||
| 6526 | case Intrinsic::ctpop: | |||
| 6527 | case Intrinsic::bswap: | |||
| 6528 | case Intrinsic::bitreverse: | |||
| 6529 | case Intrinsic::fshl: | |||
| 6530 | case Intrinsic::fshr: | |||
| 6531 | case Intrinsic::smax: | |||
| 6532 | case Intrinsic::smin: | |||
| 6533 | case Intrinsic::umax: | |||
| 6534 | case Intrinsic::umin: | |||
| 6535 | case Intrinsic::ptrmask: | |||
| 6536 | case Intrinsic::fptoui_sat: | |||
| 6537 | case Intrinsic::fptosi_sat: | |||
| 6538 | case Intrinsic::sadd_with_overflow: | |||
| 6539 | case Intrinsic::ssub_with_overflow: | |||
| 6540 | case Intrinsic::smul_with_overflow: | |||
| 6541 | case Intrinsic::uadd_with_overflow: | |||
| 6542 | case Intrinsic::usub_with_overflow: | |||
| 6543 | case Intrinsic::umul_with_overflow: | |||
| 6544 | case Intrinsic::sadd_sat: | |||
| 6545 | case Intrinsic::uadd_sat: | |||
| 6546 | case Intrinsic::ssub_sat: | |||
| 6547 | case Intrinsic::usub_sat: | |||
| 6548 | return false; | |||
| 6549 | case Intrinsic::sshl_sat: | |||
| 6550 | case Intrinsic::ushl_sat: | |||
| 6551 | return !shiftAmountKnownInRange(II->getArgOperand(1)); | |||
| 6552 | case Intrinsic::fma: | |||
| 6553 | case Intrinsic::fmuladd: | |||
| 6554 | case Intrinsic::sqrt: | |||
| 6555 | case Intrinsic::powi: | |||
| 6556 | case Intrinsic::sin: | |||
| 6557 | case Intrinsic::cos: | |||
| 6558 | case Intrinsic::pow: | |||
| 6559 | case Intrinsic::log: | |||
| 6560 | case Intrinsic::log10: | |||
| 6561 | case Intrinsic::log2: | |||
| 6562 | case Intrinsic::exp: | |||
| 6563 | case Intrinsic::exp2: | |||
| 6564 | case Intrinsic::fabs: | |||
| 6565 | case Intrinsic::copysign: | |||
| 6566 | case Intrinsic::floor: | |||
| 6567 | case Intrinsic::ceil: | |||
| 6568 | case Intrinsic::trunc: | |||
| 6569 | case Intrinsic::rint: | |||
| 6570 | case Intrinsic::nearbyint: | |||
| 6571 | case Intrinsic::round: | |||
| 6572 | case Intrinsic::roundeven: | |||
| 6573 | case Intrinsic::fptrunc_round: | |||
| 6574 | case Intrinsic::canonicalize: | |||
| 6575 | case Intrinsic::arithmetic_fence: | |||
| 6576 | case Intrinsic::minnum: | |||
| 6577 | case Intrinsic::maxnum: | |||
| 6578 | case Intrinsic::minimum: | |||
| 6579 | case Intrinsic::maximum: | |||
| 6580 | case Intrinsic::is_fpclass: | |||
| 6581 | return false; | |||
| 6582 | case Intrinsic::lround: | |||
| 6583 | case Intrinsic::llround: | |||
| 6584 | case Intrinsic::lrint: | |||
| 6585 | case Intrinsic::llrint: | |||
| 6586 | // If the value doesn't fit an unspecified value is returned (but this | |||
| 6587 | // is not poison). | |||
| 6588 | return false; | |||
| 6589 | } | |||
| 6590 | } | |||
| 6591 | [[fallthrough]]; | |||
| 6592 | case Instruction::CallBr: | |||
| 6593 | case Instruction::Invoke: { | |||
| 6594 | const auto *CB = cast<CallBase>(Op); | |||
| 6595 | return !CB->hasRetAttr(Attribute::NoUndef); | |||
| 6596 | } | |||
| 6597 | case Instruction::InsertElement: | |||
| 6598 | case Instruction::ExtractElement: { | |||
| 6599 | // If index exceeds the length of the vector, it returns poison | |||
| 6600 | auto *VTy = cast<VectorType>(Op->getOperand(0)->getType()); | |||
| 6601 | unsigned IdxOp = Op->getOpcode() == Instruction::InsertElement ? 2 : 1; | |||
| 6602 | auto *Idx = dyn_cast<ConstantInt>(Op->getOperand(IdxOp)); | |||
| 6603 | if (!Idx || Idx->getValue().uge(VTy->getElementCount().getKnownMinValue())) | |||
| 6604 | return true; | |||
| 6605 | return false; | |||
| 6606 | } | |||
| 6607 | case Instruction::ShuffleVector: { | |||
| 6608 | // shufflevector may return undef. | |||
| 6609 | if (PoisonOnly) | |||
| 6610 | return false; | |||
| 6611 | ArrayRef<int> Mask = isa<ConstantExpr>(Op) | |||
| 6612 | ? cast<ConstantExpr>(Op)->getShuffleMask() | |||
| 6613 | : cast<ShuffleVectorInst>(Op)->getShuffleMask(); | |||
| 6614 | return is_contained(Mask, PoisonMaskElem); | |||
| 6615 | } | |||
| 6616 | case Instruction::FNeg: | |||
| 6617 | case Instruction::PHI: | |||
| 6618 | case Instruction::Select: | |||
| 6619 | case Instruction::URem: | |||
| 6620 | case Instruction::SRem: | |||
| 6621 | case Instruction::ExtractValue: | |||
| 6622 | case Instruction::InsertValue: | |||
| 6623 | case Instruction::Freeze: | |||
| 6624 | case Instruction::ICmp: | |||
| 6625 | case Instruction::FCmp: | |||
| 6626 | case Instruction::FAdd: | |||
| 6627 | case Instruction::FSub: | |||
| 6628 | case Instruction::FMul: | |||
| 6629 | case Instruction::FDiv: | |||
| 6630 | case Instruction::FRem: | |||
| 6631 | return false; | |||
| 6632 | case Instruction::GetElementPtr: | |||
| 6633 | // inbounds is handled above | |||
| 6634 | // TODO: what about inrange on constexpr? | |||
| 6635 | return false; | |||
| 6636 | default: { | |||
| 6637 | const auto *CE = dyn_cast<ConstantExpr>(Op); | |||
| 6638 | if (isa<CastInst>(Op) || (CE && CE->isCast())) | |||
| 6639 | return false; | |||
| 6640 | else if (Instruction::isBinaryOp(Opcode)) | |||
| 6641 | return false; | |||
| 6642 | // Be conservative and return true. | |||
| 6643 | return true; | |||
| 6644 | } | |||
| 6645 | } | |||
| 6646 | } | |||
| 6647 | ||||
| 6648 | bool llvm::canCreateUndefOrPoison(const Operator *Op, | |||
| 6649 | bool ConsiderFlagsAndMetadata) { | |||
| 6650 | return ::canCreateUndefOrPoison(Op, /*PoisonOnly=*/false, | |||
| 6651 | ConsiderFlagsAndMetadata); | |||
| 6652 | } | |||
| 6653 | ||||
| 6654 | bool llvm::canCreatePoison(const Operator *Op, bool ConsiderFlagsAndMetadata) { | |||
| 6655 | return ::canCreateUndefOrPoison(Op, /*PoisonOnly=*/true, | |||
| 6656 | ConsiderFlagsAndMetadata); | |||
| 6657 | } | |||
| 6658 | ||||
| 6659 | static bool directlyImpliesPoison(const Value *ValAssumedPoison, | |||
| 6660 | const Value *V, unsigned Depth) { | |||
| 6661 | if (ValAssumedPoison == V) | |||
| 6662 | return true; | |||
| 6663 | ||||
| 6664 | const unsigned MaxDepth = 2; | |||
| 6665 | if (Depth >= MaxDepth) | |||
| 6666 | return false; | |||
| 6667 | ||||
| 6668 | if (const auto *I = dyn_cast<Instruction>(V)) { | |||
| 6669 | if (any_of(I->operands(), [=](const Use &Op) { | |||
| 6670 | return propagatesPoison(Op) && | |||
| 6671 | directlyImpliesPoison(ValAssumedPoison, Op, Depth + 1); | |||
| 6672 | })) | |||
| 6673 | return true; | |||
| 6674 | ||||
| 6675 | // V = extractvalue V0, idx | |||
| 6676 | // V2 = extractvalue V0, idx2 | |||
| 6677 | // V0's elements are all poison or not. (e.g., add_with_overflow) | |||
| 6678 | const WithOverflowInst *II; | |||
| 6679 | if (match(I, m_ExtractValue(m_WithOverflowInst(II))) && | |||
| 6680 | (match(ValAssumedPoison, m_ExtractValue(m_Specific(II))) || | |||
| 6681 | llvm::is_contained(II->args(), ValAssumedPoison))) | |||
| 6682 | return true; | |||
| 6683 | } | |||
| 6684 | return false; | |||
| 6685 | } | |||
| 6686 | ||||
| 6687 | static bool impliesPoison(const Value *ValAssumedPoison, const Value *V, | |||
| 6688 | unsigned Depth) { | |||
| 6689 | if (isGuaranteedNotToBePoison(ValAssumedPoison)) | |||
| 6690 | return true; | |||
| 6691 | ||||
| 6692 | if (directlyImpliesPoison(ValAssumedPoison, V, /* Depth */ 0)) | |||
| 6693 | return true; | |||
| 6694 | ||||
| 6695 | const unsigned MaxDepth = 2; | |||
| 6696 | if (Depth >= MaxDepth) | |||
| 6697 | return false; | |||
| 6698 | ||||
| 6699 | const auto *I = dyn_cast<Instruction>(ValAssumedPoison); | |||
| 6700 | if (I && !canCreatePoison(cast<Operator>(I))) { | |||
| 6701 | return all_of(I->operands(), [=](const Value *Op) { | |||
| 6702 | return impliesPoison(Op, V, Depth + 1); | |||
| 6703 | }); | |||
| 6704 | } | |||
| 6705 | return false; | |||
| 6706 | } | |||
| 6707 | ||||
| 6708 | bool llvm::impliesPoison(const Value *ValAssumedPoison, const Value *V) { | |||
| 6709 | return ::impliesPoison(ValAssumedPoison, V, /* Depth */ 0); | |||
| 6710 | } | |||
| 6711 | ||||
| 6712 | static bool programUndefinedIfUndefOrPoison(const Value *V, | |||
| 6713 | bool PoisonOnly); | |||
| 6714 | ||||
| 6715 | static bool isGuaranteedNotToBeUndefOrPoison(const Value *V, | |||
| 6716 | AssumptionCache *AC, | |||
| 6717 | const Instruction *CtxI, | |||
| 6718 | const DominatorTree *DT, | |||
| 6719 | unsigned Depth, bool PoisonOnly) { | |||
| 6720 | if (Depth >= MaxAnalysisRecursionDepth) | |||
| 6721 | return false; | |||
| 6722 | ||||
| 6723 | if (isa<MetadataAsValue>(V)) | |||
| 6724 | return false; | |||
| 6725 | ||||
| 6726 | if (const auto *A = dyn_cast<Argument>(V)) { | |||
| 6727 | if (A->hasAttribute(Attribute::NoUndef) || | |||
| 6728 | A->hasAttribute(Attribute::Dereferenceable) || | |||
| 6729 | A->hasAttribute(Attribute::DereferenceableOrNull)) | |||
| 6730 | return true; | |||
| 6731 | } | |||
| 6732 | ||||
| 6733 | if (auto *C = dyn_cast<Constant>(V)) { | |||
| 6734 | if (isa<UndefValue>(C)) | |||
| 6735 | return PoisonOnly && !isa<PoisonValue>(C); | |||
| 6736 | ||||
| 6737 | if (isa<ConstantInt>(C) || isa<GlobalVariable>(C) || isa<ConstantFP>(V) || | |||
| 6738 | isa<ConstantPointerNull>(C) || isa<Function>(C)) | |||
| 6739 | return true; | |||
| 6740 | ||||
| 6741 | if (C->getType()->isVectorTy() && !isa<ConstantExpr>(C)) | |||
| 6742 | return (PoisonOnly ? !C->containsPoisonElement() | |||
| 6743 | : !C->containsUndefOrPoisonElement()) && | |||
| 6744 | !C->containsConstantExpression(); | |||
| 6745 | } | |||
| 6746 | ||||
| 6747 | // Strip cast operations from a pointer value. | |||
| 6748 | // Note that stripPointerCastsSameRepresentation can strip off getelementptr | |||
| 6749 | // inbounds with zero offset. To guarantee that the result isn't poison, the | |||
| 6750 | // stripped pointer is checked as it has to be pointing into an allocated | |||
| 6751 | // object or be null `null` to ensure `inbounds` getelement pointers with a | |||
| 6752 | // zero offset could not produce poison. | |||
| 6753 | // It can strip off addrspacecast that do not change bit representation as | |||
| 6754 | // well. We believe that such addrspacecast is equivalent to no-op. | |||
| 6755 | auto *StrippedV = V->stripPointerCastsSameRepresentation(); | |||
| 6756 | if (isa<AllocaInst>(StrippedV) || isa<GlobalVariable>(StrippedV) || | |||
| 6757 | isa<Function>(StrippedV) || isa<ConstantPointerNull>(StrippedV)) | |||
| 6758 | return true; | |||
| 6759 | ||||
| 6760 | auto OpCheck = [&](const Value *V) { | |||
| 6761 | return isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth + 1, | |||
| 6762 | PoisonOnly); | |||
| 6763 | }; | |||
| 6764 | ||||
| 6765 | if (auto *Opr = dyn_cast<Operator>(V)) { | |||
| 6766 | // If the value is a freeze instruction, then it can never | |||
| 6767 | // be undef or poison. | |||
| 6768 | if (isa<FreezeInst>(V)) | |||
| 6769 | return true; | |||
| 6770 | ||||
| 6771 | if (const auto *CB = dyn_cast<CallBase>(V)) { | |||
| 6772 | if (CB->hasRetAttr(Attribute::NoUndef)) | |||
| 6773 | return true; | |||
| 6774 | } | |||
| 6775 | ||||
| 6776 | if (const auto *PN = dyn_cast<PHINode>(V)) { | |||
| 6777 | unsigned Num = PN->getNumIncomingValues(); | |||
| 6778 | bool IsWellDefined = true; | |||
| 6779 | for (unsigned i = 0; i < Num; ++i) { | |||
| 6780 | auto *TI = PN->getIncomingBlock(i)->getTerminator(); | |||
| 6781 | if (!isGuaranteedNotToBeUndefOrPoison(PN->getIncomingValue(i), AC, TI, | |||
| 6782 | DT, Depth + 1, PoisonOnly)) { | |||
| 6783 | IsWellDefined = false; | |||
| 6784 | break; | |||
| 6785 | } | |||
| 6786 | } | |||
| 6787 | if (IsWellDefined) | |||
| 6788 | return true; | |||
| 6789 | } else if (!canCreateUndefOrPoison(Opr) && all_of(Opr->operands(), OpCheck)) | |||
| 6790 | return true; | |||
| 6791 | } | |||
| 6792 | ||||
| 6793 | if (auto *I = dyn_cast<LoadInst>(V)) | |||
| 6794 | if (I->hasMetadata(LLVMContext::MD_noundef) || | |||
| 6795 | I->hasMetadata(LLVMContext::MD_dereferenceable) || | |||
| 6796 | I->hasMetadata(LLVMContext::MD_dereferenceable_or_null)) | |||
| 6797 | return true; | |||
| 6798 | ||||
| 6799 | if (programUndefinedIfUndefOrPoison(V, PoisonOnly)) | |||
| 6800 | return true; | |||
| 6801 | ||||
| 6802 | // CxtI may be null or a cloned instruction. | |||
| 6803 | if (!CtxI || !CtxI->getParent() || !DT) | |||
| 6804 | return false; | |||
| 6805 | ||||
| 6806 | auto *DNode = DT->getNode(CtxI->getParent()); | |||
| 6807 | if (!DNode) | |||
| 6808 | // Unreachable block | |||
| 6809 | return false; | |||
| 6810 | ||||
| 6811 | // If V is used as a branch condition before reaching CtxI, V cannot be | |||
| 6812 | // undef or poison. | |||
| 6813 | // br V, BB1, BB2 | |||
| 6814 | // BB1: | |||
| 6815 | // CtxI ; V cannot be undef or poison here | |||
| 6816 | auto *Dominator = DNode->getIDom(); | |||
| 6817 | while (Dominator) { | |||
| 6818 | auto *TI = Dominator->getBlock()->getTerminator(); | |||
| 6819 | ||||
| 6820 | Value *Cond = nullptr; | |||
| 6821 | if (auto BI = dyn_cast_or_null<BranchInst>(TI)) { | |||
| 6822 | if (BI->isConditional()) | |||
| 6823 | Cond = BI->getCondition(); | |||
| 6824 | } else if (auto SI = dyn_cast_or_null<SwitchInst>(TI)) { | |||
| 6825 | Cond = SI->getCondition(); | |||
| 6826 | } | |||
| 6827 | ||||
| 6828 | if (Cond) { | |||
| 6829 | if (Cond == V) | |||
| 6830 | return true; | |||
| 6831 | else if (PoisonOnly && isa<Operator>(Cond)) { | |||
| 6832 | // For poison, we can analyze further | |||
| 6833 | auto *Opr = cast<Operator>(Cond); | |||
| 6834 | if (any_of(Opr->operands(), | |||
| 6835 | [V](const Use &U) { return V == U && propagatesPoison(U); })) | |||
| 6836 | return true; | |||
| 6837 | } | |||
| 6838 | } | |||
| 6839 | ||||
| 6840 | Dominator = Dominator->getIDom(); | |||
| 6841 | } | |||
| 6842 | ||||
| 6843 | if (getKnowledgeValidInContext(V, {Attribute::NoUndef}, CtxI, DT, AC)) | |||
| 6844 | return true; | |||
| 6845 | ||||
| 6846 | return false; | |||
| 6847 | } | |||
| 6848 | ||||
| 6849 | bool llvm::isGuaranteedNotToBeUndefOrPoison(const Value *V, AssumptionCache *AC, | |||
| 6850 | const Instruction *CtxI, | |||
| 6851 | const DominatorTree *DT, | |||
| 6852 | unsigned Depth) { | |||
| 6853 | return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, false); | |||
| 6854 | } | |||
| 6855 | ||||
| 6856 | bool llvm::isGuaranteedNotToBePoison(const Value *V, AssumptionCache *AC, | |||
| 6857 | const Instruction *CtxI, | |||
| 6858 | const DominatorTree *DT, unsigned Depth) { | |||
| 6859 | return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, true); | |||
| 6860 | } | |||
| 6861 | ||||
| 6862 | /// Return true if undefined behavior would provably be executed on the path to | |||
| 6863 | /// OnPathTo if Root produced a posion result. Note that this doesn't say | |||
| 6864 | /// anything about whether OnPathTo is actually executed or whether Root is | |||
| 6865 | /// actually poison. This can be used to assess whether a new use of Root can | |||
| 6866 | /// be added at a location which is control equivalent with OnPathTo (such as | |||
| 6867 | /// immediately before it) without introducing UB which didn't previously | |||
| 6868 | /// exist. Note that a false result conveys no information. | |||
| 6869 | bool llvm::mustExecuteUBIfPoisonOnPathTo(Instruction *Root, | |||
| 6870 | Instruction *OnPathTo, | |||
| 6871 | DominatorTree *DT) { | |||
| 6872 | // Basic approach is to assume Root is poison, propagate poison forward | |||
| 6873 | // through all users we can easily track, and then check whether any of those | |||
| 6874 | // users are provable UB and must execute before out exiting block might | |||
| 6875 | // exit. | |||
| 6876 | ||||
| 6877 | // The set of all recursive users we've visited (which are assumed to all be | |||
| 6878 | // poison because of said visit) | |||
| 6879 | SmallSet<const Value *, 16> KnownPoison; | |||
| 6880 | SmallVector<const Instruction*, 16> Worklist; | |||
| 6881 | Worklist.push_back(Root); | |||
| 6882 | while (!Worklist.empty()) { | |||
| 6883 | const Instruction *I = Worklist.pop_back_val(); | |||
| 6884 | ||||
| 6885 | // If we know this must trigger UB on a path leading our target. | |||
| 6886 | if (mustTriggerUB(I, KnownPoison) && DT->dominates(I, OnPathTo)) | |||
| 6887 | return true; | |||
| 6888 | ||||
| 6889 | // If we can't analyze propagation through this instruction, just skip it | |||
| 6890 | // and transitive users. Safe as false is a conservative result. | |||
| 6891 | if (I != Root && !any_of(I->operands(), [&KnownPoison](const Use &U) { | |||
| 6892 | return KnownPoison.contains(U) && propagatesPoison(U); | |||
| 6893 | })) | |||
| 6894 | continue; | |||
| 6895 | ||||
| 6896 | if (KnownPoison.insert(I).second) | |||
| 6897 | for (const User *User : I->users()) | |||
| 6898 | Worklist.push_back(cast<Instruction>(User)); | |||
| 6899 | } | |||
| 6900 | ||||
| 6901 | // Might be non-UB, or might have a path we couldn't prove must execute on | |||
| 6902 | // way to exiting bb. | |||
| 6903 | return false; | |||
| 6904 | } | |||
| 6905 | ||||
| 6906 | OverflowResult llvm::computeOverflowForSignedAdd(const AddOperator *Add, | |||
| 6907 | const DataLayout &DL, | |||
| 6908 | AssumptionCache *AC, | |||
| 6909 | const Instruction *CxtI, | |||
| 6910 | const DominatorTree *DT) { | |||
| 6911 | return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1), | |||
| 6912 | Add, DL, AC, CxtI, DT); | |||
| 6913 | } | |||
| 6914 | ||||
| 6915 | OverflowResult llvm::computeOverflowForSignedAdd(const Value *LHS, | |||
| 6916 | const Value *RHS, | |||
| 6917 | const DataLayout &DL, | |||
| 6918 | AssumptionCache *AC, | |||
| 6919 | const Instruction *CxtI, | |||
| 6920 | const DominatorTree *DT) { | |||
| 6921 | return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT); | |||
| 6922 | } | |||
| 6923 | ||||
| 6924 | bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) { | |||
| 6925 | // Note: An atomic operation isn't guaranteed to return in a reasonable amount | |||
| 6926 | // of time because it's possible for another thread to interfere with it for an | |||
| 6927 | // arbitrary length of time, but programs aren't allowed to rely on that. | |||
| 6928 | ||||
| 6929 | // If there is no successor, then execution can't transfer to it. | |||
| 6930 | if (isa<ReturnInst>(I)) | |||
| 6931 | return false; | |||
| 6932 | if (isa<UnreachableInst>(I)) | |||
| 6933 | return false; | |||
| 6934 | ||||
| 6935 | // Note: Do not add new checks here; instead, change Instruction::mayThrow or | |||
| 6936 | // Instruction::willReturn. | |||
| 6937 | // | |||
| 6938 | // FIXME: Move this check into Instruction::willReturn. | |||
| 6939 | if (isa<CatchPadInst>(I)) { | |||
| 6940 | switch (classifyEHPersonality(I->getFunction()->getPersonalityFn())) { | |||
| 6941 | default: | |||
| 6942 | // A catchpad may invoke exception object constructors and such, which | |||
| 6943 | // in some languages can be arbitrary code, so be conservative by default. | |||
| 6944 | return false; | |||
| 6945 | case EHPersonality::CoreCLR: | |||
| 6946 | // For CoreCLR, it just involves a type test. | |||
| 6947 | return true; | |||
| 6948 | } | |||
| 6949 | } | |||
| 6950 | ||||
| 6951 | // An instruction that returns without throwing must transfer control flow | |||
| 6952 | // to a successor. | |||
| 6953 | return !I->mayThrow() && I->willReturn(); | |||
| 6954 | } | |||
| 6955 | ||||
| 6956 | bool llvm::isGuaranteedToTransferExecutionToSuccessor(const BasicBlock *BB) { | |||
| 6957 | // TODO: This is slightly conservative for invoke instruction since exiting | |||
| 6958 | // via an exception *is* normal control for them. | |||
| 6959 | for (const Instruction &I : *BB) | |||
| 6960 | if (!isGuaranteedToTransferExecutionToSuccessor(&I)) | |||
| 6961 | return false; | |||
| 6962 | return true; | |||
| 6963 | } | |||
| 6964 | ||||
| 6965 | bool llvm::isGuaranteedToTransferExecutionToSuccessor( | |||
| 6966 | BasicBlock::const_iterator Begin, BasicBlock::const_iterator End, | |||
| 6967 | unsigned ScanLimit) { | |||
| 6968 | return isGuaranteedToTransferExecutionToSuccessor(make_range(Begin, End), | |||
| 6969 | ScanLimit); | |||
| 6970 | } | |||
| 6971 | ||||
| 6972 | bool llvm::isGuaranteedToTransferExecutionToSuccessor( | |||
| 6973 | iterator_range<BasicBlock::const_iterator> Range, unsigned ScanLimit) { | |||
| 6974 | assert(ScanLimit && "scan limit must be non-zero")(static_cast <bool> (ScanLimit && "scan limit must be non-zero" ) ? void (0) : __assert_fail ("ScanLimit && \"scan limit must be non-zero\"" , "llvm/lib/Analysis/ValueTracking.cpp", 6974, __extension__ __PRETTY_FUNCTION__ )); | |||
| 6975 | for (const Instruction &I : Range) { | |||
| 6976 | if (isa<DbgInfoIntrinsic>(I)) | |||
| 6977 | continue; | |||
| 6978 | if (--ScanLimit == 0) | |||
| 6979 | return false; | |||
| 6980 | if (!isGuaranteedToTransferExecutionToSuccessor(&I)) | |||
| 6981 | return false; | |||
| 6982 | } | |||
| 6983 | return true; | |||
| 6984 | } | |||
| 6985 | ||||
| 6986 | bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I, | |||
| 6987 | const Loop *L) { | |||
| 6988 | // The loop header is guaranteed to be executed for every iteration. | |||
| 6989 | // | |||
| 6990 | // FIXME: Relax this constraint to cover all basic blocks that are | |||
| 6991 | // guaranteed to be executed at every iteration. | |||
| 6992 | if (I->getParent() != L->getHeader()) return false; | |||
| 6993 | ||||
| 6994 | for (const Instruction &LI : *L->getHeader()) { | |||
| 6995 | if (&LI == I) return true; | |||
| 6996 | if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false; | |||
| 6997 | } | |||
| 6998 | llvm_unreachable("Instruction not contained in its own parent basic block.")::llvm::llvm_unreachable_internal("Instruction not contained in its own parent basic block." , "llvm/lib/Analysis/ValueTracking.cpp", 6998); | |||
| 6999 | } | |||
| 7000 | ||||
| 7001 | bool llvm::propagatesPoison(const Use &PoisonOp) { | |||
| 7002 | const Operator *I = cast<Operator>(PoisonOp.getUser()); | |||
| 7003 | switch (I->getOpcode()) { | |||
| 7004 | case Instruction::Freeze: | |||
| 7005 | case Instruction::PHI: | |||
| 7006 | case Instruction::Invoke: | |||
| 7007 | return false; | |||
| 7008 | case Instruction::Select: | |||
| 7009 | return PoisonOp.getOperandNo() == 0; | |||
| 7010 | case Instruction::Call: | |||
| 7011 | if (auto *II = dyn_cast<IntrinsicInst>(I)) { | |||
| 7012 | switch (II->getIntrinsicID()) { | |||
| 7013 | // TODO: Add more intrinsics. | |||
| 7014 | case Intrinsic::sadd_with_overflow: | |||
| 7015 | case Intrinsic::ssub_with_overflow: | |||
| 7016 | case Intrinsic::smul_with_overflow: | |||
| 7017 | case Intrinsic::uadd_with_overflow: | |||
| 7018 | case Intrinsic::usub_with_overflow: | |||
| 7019 | case Intrinsic::umul_with_overflow: | |||
| 7020 | // If an input is a vector containing a poison element, the | |||
| 7021 | // two output vectors (calculated results, overflow bits)' | |||
| 7022 | // corresponding lanes are poison. | |||
| 7023 | return true; | |||
| 7024 | case Intrinsic::ctpop: | |||
| 7025 | return true; | |||
| 7026 | } | |||
| 7027 | } | |||
| 7028 | return false; | |||
| 7029 | case Instruction::ICmp: | |||
| 7030 | case Instruction::FCmp: | |||
| 7031 | case Instruction::GetElementPtr: | |||
| 7032 | return true; | |||
| 7033 | default: | |||
| 7034 | if (isa<BinaryOperator>(I) || isa<UnaryOperator>(I) || isa<CastInst>(I)) | |||
| 7035 | return true; | |||
| 7036 | ||||
| 7037 | // Be conservative and return false. | |||
| 7038 | return false; | |||
| 7039 | } | |||
| 7040 | } | |||
| 7041 | ||||
| 7042 | void llvm::getGuaranteedWellDefinedOps( | |||
| 7043 | const Instruction *I, SmallVectorImpl<const Value *> &Operands) { | |||
| 7044 | switch (I->getOpcode()) { | |||
| 7045 | case Instruction::Store: | |||
| 7046 | Operands.push_back(cast<StoreInst>(I)->getPointerOperand()); | |||
| 7047 | break; | |||
| 7048 | ||||
| 7049 | case Instruction::Load: | |||
| 7050 | Operands.push_back(cast<LoadInst>(I)->getPointerOperand()); | |||
| 7051 | break; | |||
| 7052 | ||||
| 7053 | // Since dereferenceable attribute imply noundef, atomic operations | |||
| 7054 | // also implicitly have noundef pointers too | |||
| 7055 | case Instruction::AtomicCmpXchg: | |||
| 7056 | Operands.push_back(cast<AtomicCmpXchgInst>(I)->getPointerOperand()); | |||
| 7057 | break; | |||
| 7058 | ||||
| 7059 | case Instruction::AtomicRMW: | |||
| 7060 | Operands.push_back(cast<AtomicRMWInst>(I)->getPointerOperand()); | |||
| 7061 | break; | |||
| 7062 | ||||
| 7063 | case Instruction::Call: | |||
| 7064 | case Instruction::Invoke: { | |||
| 7065 | const CallBase *CB = cast<CallBase>(I); | |||
| 7066 | if (CB->isIndirectCall()) | |||
| 7067 | Operands.push_back(CB->getCalledOperand()); | |||
| 7068 | for (unsigned i = 0; i < CB->arg_size(); ++i) { | |||
| 7069 | if (CB->paramHasAttr(i, Attribute::NoUndef) || | |||
| 7070 | CB->paramHasAttr(i, Attribute::Dereferenceable)) | |||
| 7071 | Operands.push_back(CB->getArgOperand(i)); | |||
| 7072 | } | |||
| 7073 | break; | |||
| 7074 | } | |||
| 7075 | case Instruction::Ret: | |||
| 7076 | if (I->getFunction()->hasRetAttribute(Attribute::NoUndef)) | |||
| 7077 | Operands.push_back(I->getOperand(0)); | |||
| 7078 | break; | |||
| 7079 | case Instruction::Switch: | |||
| 7080 | Operands.push_back(cast<SwitchInst>(I)->getCondition()); | |||
| 7081 | break; | |||
| 7082 | case Instruction::Br: { | |||
| 7083 | auto *BR = cast<BranchInst>(I); | |||
| 7084 | if (BR->isConditional()) | |||
| 7085 | Operands.push_back(BR->getCondition()); | |||
| 7086 | break; | |||
| 7087 | } | |||
| 7088 | default: | |||
| 7089 | break; | |||
| 7090 | } | |||
| 7091 | } | |||
| 7092 | ||||
| 7093 | void llvm::getGuaranteedNonPoisonOps(const Instruction *I, | |||
| 7094 | SmallVectorImpl<const Value *> &Operands) { | |||
| 7095 | getGuaranteedWellDefinedOps(I, Operands); | |||
| 7096 | switch (I->getOpcode()) { | |||
| 7097 | // Divisors of these operations are allowed to be partially undef. | |||
| 7098 | case Instruction::UDiv: | |||
| 7099 | case Instruction::SDiv: | |||
| 7100 | case Instruction::URem: | |||
| 7101 | case Instruction::SRem: | |||
| 7102 | Operands.push_back(I->getOperand(1)); | |||
| 7103 | break; | |||
| 7104 | default: | |||
| 7105 | break; | |||
| 7106 | } | |||
| 7107 | } | |||
| 7108 | ||||
| 7109 | bool llvm::mustTriggerUB(const Instruction *I, | |||
| 7110 | const SmallPtrSetImpl<const Value *> &KnownPoison) { | |||
| 7111 | SmallVector<const Value *, 4> NonPoisonOps; | |||
| 7112 | getGuaranteedNonPoisonOps(I, NonPoisonOps); | |||
| 7113 | ||||
| 7114 | for (const auto *V : NonPoisonOps) | |||
| 7115 | if (KnownPoison.count(V)) | |||
| 7116 | return true; | |||
| 7117 | ||||
| 7118 | return false; | |||
| 7119 | } | |||
| 7120 | ||||
| 7121 | static bool programUndefinedIfUndefOrPoison(const Value *V, | |||
| 7122 | bool PoisonOnly) { | |||
| 7123 | // We currently only look for uses of values within the same basic | |||
| 7124 | // block, as that makes it easier to guarantee that the uses will be | |||
| 7125 | // executed given that Inst is executed. | |||
| 7126 | // | |||
| 7127 | // FIXME: Expand this to consider uses beyond the same basic block. To do | |||
| 7128 | // this, look out for the distinction between post-dominance and strong | |||
| 7129 | // post-dominance. | |||
| 7130 | const BasicBlock *BB = nullptr; | |||
| 7131 | BasicBlock::const_iterator Begin; | |||
| 7132 | if (const auto *Inst = dyn_cast<Instruction>(V)) { | |||
| 7133 | BB = Inst->getParent(); | |||
| 7134 | Begin = Inst->getIterator(); | |||
| 7135 | Begin++; | |||
| 7136 | } else if (const auto *Arg = dyn_cast<Argument>(V)) { | |||
| 7137 | BB = &Arg->getParent()->getEntryBlock(); | |||
| 7138 | Begin = BB->begin(); | |||
| 7139 | } else { | |||
| 7140 | return false; | |||
| 7141 | } | |||
| 7142 | ||||
| 7143 | // Limit number of instructions we look at, to avoid scanning through large | |||
| 7144 | // blocks. The current limit is chosen arbitrarily. | |||
| 7145 | unsigned ScanLimit = 32; | |||
| 7146 | BasicBlock::const_iterator End = BB->end(); | |||
| 7147 | ||||
| 7148 | if (!PoisonOnly) { | |||
| 7149 | // Since undef does not propagate eagerly, be conservative & just check | |||
| 7150 | // whether a value is directly passed to an instruction that must take | |||
| 7151 | // well-defined operands. | |||
| 7152 | ||||
| 7153 | for (const auto &I : make_range(Begin, End)) { | |||
| 7154 | if (isa<DbgInfoIntrinsic>(I)) | |||
| 7155 | continue; | |||
| 7156 | if (--ScanLimit == 0) | |||
| 7157 | break; | |||
| 7158 | ||||
| 7159 | SmallVector<const Value *, 4> WellDefinedOps; | |||
| 7160 | getGuaranteedWellDefinedOps(&I, WellDefinedOps); | |||
| 7161 | if (is_contained(WellDefinedOps, V)) | |||
| 7162 | return true; | |||
| 7163 | ||||
| 7164 | if (!isGuaranteedToTransferExecutionToSuccessor(&I)) | |||
| 7165 | break; | |||
| 7166 | } | |||
| 7167 | return false; | |||
| 7168 | } | |||
| 7169 | ||||
| 7170 | // Set of instructions that we have proved will yield poison if Inst | |||
| 7171 | // does. | |||
| 7172 | SmallSet<const Value *, 16> YieldsPoison; | |||
| 7173 | SmallSet<const BasicBlock *, 4> Visited; | |||
| 7174 | ||||
| 7175 | YieldsPoison.insert(V); | |||
| 7176 | Visited.insert(BB); | |||
| 7177 | ||||
| 7178 | while (true) { | |||
| 7179 | for (const auto &I : make_range(Begin, End)) { | |||
| 7180 | if (isa<DbgInfoIntrinsic>(I)) | |||
| 7181 | continue; | |||
| 7182 | if (--ScanLimit == 0) | |||
| 7183 | return false; | |||
| 7184 | if (mustTriggerUB(&I, YieldsPoison)) | |||
| 7185 | return true; | |||
| 7186 | if (!isGuaranteedToTransferExecutionToSuccessor(&I)) | |||
| 7187 | return false; | |||
| 7188 | ||||
| 7189 | // If an operand is poison and propagates it, mark I as yielding poison. | |||
| 7190 | for (const Use &Op : I.operands()) { | |||
| 7191 | if (YieldsPoison.count(Op) && propagatesPoison(Op)) { | |||
| 7192 | YieldsPoison.insert(&I); | |||
| 7193 | break; | |||
| 7194 | } | |||
| 7195 | } | |||
| 7196 | ||||
| 7197 | // Special handling for select, which returns poison if its operand 0 is | |||
| 7198 | // poison (handled in the loop above) *or* if both its true/false operands | |||
| 7199 | // are poison (handled here). | |||
| 7200 | if (I.getOpcode() == Instruction::Select && | |||
| 7201 | YieldsPoison.count(I.getOperand(1)) && | |||
| 7202 | YieldsPoison.count(I.getOperand(2))) { | |||
| 7203 | YieldsPoison.insert(&I); | |||
| 7204 | } | |||
| 7205 | } | |||
| 7206 | ||||
| 7207 | BB = BB->getSingleSuccessor(); | |||
| 7208 | if (!BB || !Visited.insert(BB).second) | |||
| 7209 | break; | |||
| 7210 | ||||
| 7211 | Begin = BB->getFirstNonPHI()->getIterator(); | |||
| 7212 | End = BB->end(); | |||
| 7213 | } | |||
| 7214 | return false; | |||
| 7215 | } | |||
| 7216 | ||||
| 7217 | bool llvm::programUndefinedIfUndefOrPoison(const Instruction *Inst) { | |||
| 7218 | return ::programUndefinedIfUndefOrPoison(Inst, false); | |||
| 7219 | } | |||
| 7220 | ||||
| 7221 | bool llvm::programUndefinedIfPoison(const Instruction *Inst) { | |||
| 7222 | return ::programUndefinedIfUndefOrPoison(Inst, true); | |||
| 7223 | } | |||
| 7224 | ||||
| 7225 | static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) { | |||
| 7226 | if (FMF.noNaNs()) | |||
| 7227 | return true; | |||
| 7228 | ||||
| 7229 | if (auto *C = dyn_cast<ConstantFP>(V)) | |||
| 7230 | return !C->isNaN(); | |||
| 7231 | ||||
| 7232 | if (auto *C = dyn_cast<ConstantDataVector>(V)) { | |||
| 7233 | if (!C->getElementType()->isFloatingPointTy()) | |||
| 7234 | return false; | |||
| 7235 | for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) { | |||
| 7236 | if (C->getElementAsAPFloat(I).isNaN()) | |||
| 7237 | return false; | |||
| 7238 | } | |||
| 7239 | return true; | |||
| 7240 | } | |||
| 7241 | ||||
| 7242 | if (isa<ConstantAggregateZero>(V)) | |||
| 7243 | return true; | |||
| 7244 | ||||
| 7245 | return false; | |||
| 7246 | } | |||
| 7247 | ||||
| 7248 | static bool isKnownNonZero(const Value *V) { | |||
| 7249 | if (auto *C = dyn_cast<ConstantFP>(V)) | |||
| 7250 | return !C->isZero(); | |||
| 7251 | ||||
| 7252 | if (auto *C = dyn_cast<ConstantDataVector>(V)) { | |||
| 7253 | if (!C->getElementType()->isFloatingPointTy()) | |||
| 7254 | return false; | |||
| 7255 | for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) { | |||
| 7256 | if (C->getElementAsAPFloat(I).isZero()) | |||
| 7257 | return false; | |||
| 7258 | } | |||
| 7259 | return true; | |||
| 7260 | } | |||
| 7261 | ||||
| 7262 | return false; | |||
| 7263 | } | |||
| 7264 | ||||
| 7265 | /// Match clamp pattern for float types without care about NaNs or signed zeros. | |||
| 7266 | /// Given non-min/max outer cmp/select from the clamp pattern this | |||
| 7267 | /// function recognizes if it can be substitued by a "canonical" min/max | |||
| 7268 | /// pattern. | |||
| 7269 | static SelectPatternResult matchFastFloatClamp(CmpInst::Predicate Pred, | |||
| 7270 | Value *CmpLHS, Value *CmpRHS, | |||
| 7271 | Value *TrueVal, Value *FalseVal, | |||
| 7272 | Value *&LHS, Value *&RHS) { | |||
| 7273 | // Try to match | |||
| 7274 | // X < C1 ? C1 : Min(X, C2) --> Max(C1, Min(X, C2)) | |||
| 7275 | // X > C1 ? C1 : Max(X, C2) --> Min(C1, Max(X, C2)) | |||
| 7276 | // and return description of the outer Max/Min. | |||
| 7277 | ||||
| 7278 | // First, check if select has inverse order: | |||
| 7279 | if (CmpRHS == FalseVal) { | |||
| 7280 | std::swap(TrueVal, FalseVal); | |||
| 7281 | Pred = CmpInst::getInversePredicate(Pred); | |||
| 7282 | } | |||
| 7283 | ||||
| 7284 | // Assume success now. If there's no match, callers should not use these anyway. | |||
| 7285 | LHS = TrueVal; | |||
| 7286 | RHS = FalseVal; | |||
| 7287 | ||||
| 7288 | const APFloat *FC1; | |||
| 7289 | if (CmpRHS != TrueVal || !match(CmpRHS, m_APFloat(FC1)) || !FC1->isFinite()) | |||
| 7290 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7291 | ||||
| 7292 | const APFloat *FC2; | |||
| 7293 | switch (Pred) { | |||
| 7294 | case CmpInst::FCMP_OLT: | |||
| 7295 | case CmpInst::FCMP_OLE: | |||
| 7296 | case CmpInst::FCMP_ULT: | |||
| 7297 | case CmpInst::FCMP_ULE: | |||
| 7298 | if (match(FalseVal, | |||
| 7299 | m_CombineOr(m_OrdFMin(m_Specific(CmpLHS), m_APFloat(FC2)), | |||
| 7300 | m_UnordFMin(m_Specific(CmpLHS), m_APFloat(FC2)))) && | |||
| 7301 | *FC1 < *FC2) | |||
| 7302 | return {SPF_FMAXNUM, SPNB_RETURNS_ANY, false}; | |||
| 7303 | break; | |||
| 7304 | case CmpInst::FCMP_OGT: | |||
| 7305 | case CmpInst::FCMP_OGE: | |||
| 7306 | case CmpInst::FCMP_UGT: | |||
| 7307 | case CmpInst::FCMP_UGE: | |||
| 7308 | if (match(FalseVal, | |||
| 7309 | m_CombineOr(m_OrdFMax(m_Specific(CmpLHS), m_APFloat(FC2)), | |||
| 7310 | m_UnordFMax(m_Specific(CmpLHS), m_APFloat(FC2)))) && | |||
| 7311 | *FC1 > *FC2) | |||
| 7312 | return {SPF_FMINNUM, SPNB_RETURNS_ANY, false}; | |||
| 7313 | break; | |||
| 7314 | default: | |||
| 7315 | break; | |||
| 7316 | } | |||
| 7317 | ||||
| 7318 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7319 | } | |||
| 7320 | ||||
| 7321 | /// Recognize variations of: | |||
| 7322 | /// CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v))) | |||
| 7323 | static SelectPatternResult matchClamp(CmpInst::Predicate Pred, | |||
| 7324 | Value *CmpLHS, Value *CmpRHS, | |||
| 7325 | Value *TrueVal, Value *FalseVal) { | |||
| 7326 | // Swap the select operands and predicate to match the patterns below. | |||
| 7327 | if (CmpRHS != TrueVal) { | |||
| 7328 | Pred = ICmpInst::getSwappedPredicate(Pred); | |||
| 7329 | std::swap(TrueVal, FalseVal); | |||
| 7330 | } | |||
| 7331 | const APInt *C1; | |||
| 7332 | if (CmpRHS == TrueVal && match(CmpRHS, m_APInt(C1))) { | |||
| 7333 | const APInt *C2; | |||
| 7334 | // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1) | |||
| 7335 | if (match(FalseVal, m_SMin(m_Specific(CmpLHS), m_APInt(C2))) && | |||
| 7336 | C1->slt(*C2) && Pred == CmpInst::ICMP_SLT) | |||
| 7337 | return {SPF_SMAX, SPNB_NA, false}; | |||
| 7338 | ||||
| 7339 | // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1) | |||
| 7340 | if (match(FalseVal, m_SMax(m_Specific(CmpLHS), m_APInt(C2))) && | |||
| 7341 | C1->sgt(*C2) && Pred == CmpInst::ICMP_SGT) | |||
| 7342 | return {SPF_SMIN, SPNB_NA, false}; | |||
| 7343 | ||||
| 7344 | // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1) | |||
| 7345 | if (match(FalseVal, m_UMin(m_Specific(CmpLHS), m_APInt(C2))) && | |||
| 7346 | C1->ult(*C2) && Pred == CmpInst::ICMP_ULT) | |||
| 7347 | return {SPF_UMAX, SPNB_NA, false}; | |||
| 7348 | ||||
| 7349 | // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1) | |||
| 7350 | if (match(FalseVal, m_UMax(m_Specific(CmpLHS), m_APInt(C2))) && | |||
| 7351 | C1->ugt(*C2) && Pred == CmpInst::ICMP_UGT) | |||
| 7352 | return {SPF_UMIN, SPNB_NA, false}; | |||
| 7353 | } | |||
| 7354 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7355 | } | |||
| 7356 | ||||
| 7357 | /// Recognize variations of: | |||
| 7358 | /// a < c ? min(a,b) : min(b,c) ==> min(min(a,b),min(b,c)) | |||
| 7359 | static SelectPatternResult matchMinMaxOfMinMax(CmpInst::Predicate Pred, | |||
| 7360 | Value *CmpLHS, Value *CmpRHS, | |||
| 7361 | Value *TVal, Value *FVal, | |||
| 7362 | unsigned Depth) { | |||
| 7363 | // TODO: Allow FP min/max with nnan/nsz. | |||
| 7364 | assert(CmpInst::isIntPredicate(Pred) && "Expected integer comparison")(static_cast <bool> (CmpInst::isIntPredicate(Pred) && "Expected integer comparison") ? void (0) : __assert_fail ("CmpInst::isIntPredicate(Pred) && \"Expected integer comparison\"" , "llvm/lib/Analysis/ValueTracking.cpp", 7364, __extension__ __PRETTY_FUNCTION__ )); | |||
| 7365 | ||||
| 7366 | Value *A = nullptr, *B = nullptr; | |||
| 7367 | SelectPatternResult L = matchSelectPattern(TVal, A, B, nullptr, Depth + 1); | |||
| 7368 | if (!SelectPatternResult::isMinOrMax(L.Flavor)) | |||
| 7369 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7370 | ||||
| 7371 | Value *C = nullptr, *D = nullptr; | |||
| 7372 | SelectPatternResult R = matchSelectPattern(FVal, C, D, nullptr, Depth + 1); | |||
| 7373 | if (L.Flavor != R.Flavor) | |||
| 7374 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7375 | ||||
| 7376 | // We have something like: x Pred y ? min(a, b) : min(c, d). | |||
| 7377 | // Try to match the compare to the min/max operations of the select operands. | |||
| 7378 | // First, make sure we have the right compare predicate. | |||
| 7379 | switch (L.Flavor) { | |||
| 7380 | case SPF_SMIN: | |||
| 7381 | if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) { | |||
| 7382 | Pred = ICmpInst::getSwappedPredicate(Pred); | |||
| 7383 | std::swap(CmpLHS, CmpRHS); | |||
| 7384 | } | |||
| 7385 | if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) | |||
| 7386 | break; | |||
| 7387 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7388 | case SPF_SMAX: | |||
| 7389 | if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) { | |||
| 7390 | Pred = ICmpInst::getSwappedPredicate(Pred); | |||
| 7391 | std::swap(CmpLHS, CmpRHS); | |||
| 7392 | } | |||
| 7393 | if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) | |||
| 7394 | break; | |||
| 7395 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7396 | case SPF_UMIN: | |||
| 7397 | if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) { | |||
| 7398 | Pred = ICmpInst::getSwappedPredicate(Pred); | |||
| 7399 | std::swap(CmpLHS, CmpRHS); | |||
| 7400 | } | |||
| 7401 | if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) | |||
| 7402 | break; | |||
| 7403 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7404 | case SPF_UMAX: | |||
| 7405 | if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) { | |||
| 7406 | Pred = ICmpInst::getSwappedPredicate(Pred); | |||
| 7407 | std::swap(CmpLHS, CmpRHS); | |||
| 7408 | } | |||
| 7409 | if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) | |||
| 7410 | break; | |||
| 7411 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7412 | default: | |||
| 7413 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7414 | } | |||
| 7415 | ||||
| 7416 | // If there is a common operand in the already matched min/max and the other | |||
| 7417 | // min/max operands match the compare operands (either directly or inverted), | |||
| 7418 | // then this is min/max of the same flavor. | |||
| 7419 | ||||
| 7420 | // a pred c ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b)) | |||
| 7421 | // ~c pred ~a ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b)) | |||
| 7422 | if (D == B) { | |||
| 7423 | if ((CmpLHS == A && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) && | |||
| 7424 | match(A, m_Not(m_Specific(CmpRHS))))) | |||
| 7425 | return {L.Flavor, SPNB_NA, false}; | |||
| 7426 | } | |||
| 7427 | // a pred d ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d)) | |||
| 7428 | // ~d pred ~a ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d)) | |||
| 7429 | if (C == B) { | |||
| 7430 | if ((CmpLHS == A && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) && | |||
| 7431 | match(A, m_Not(m_Specific(CmpRHS))))) | |||
| 7432 | return {L.Flavor, SPNB_NA, false}; | |||
| 7433 | } | |||
| 7434 | // b pred c ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a)) | |||
| 7435 | // ~c pred ~b ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a)) | |||
| 7436 | if (D == A) { | |||
| 7437 | if ((CmpLHS == B && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) && | |||
| 7438 | match(B, m_Not(m_Specific(CmpRHS))))) | |||
| 7439 | return {L.Flavor, SPNB_NA, false}; | |||
| 7440 | } | |||
| 7441 | // b pred d ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d)) | |||
| 7442 | // ~d pred ~b ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d)) | |||
| 7443 | if (C == A) { | |||
| 7444 | if ((CmpLHS == B && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) && | |||
| 7445 | match(B, m_Not(m_Specific(CmpRHS))))) | |||
| 7446 | return {L.Flavor, SPNB_NA, false}; | |||
| 7447 | } | |||
| 7448 | ||||
| 7449 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7450 | } | |||
| 7451 | ||||
| 7452 | /// If the input value is the result of a 'not' op, constant integer, or vector | |||
| 7453 | /// splat of a constant integer, return the bitwise-not source value. | |||
| 7454 | /// TODO: This could be extended to handle non-splat vector integer constants. | |||
| 7455 | static Value *getNotValue(Value *V) { | |||
| 7456 | Value *NotV; | |||
| 7457 | if (match(V, m_Not(m_Value(NotV)))) | |||
| 7458 | return NotV; | |||
| 7459 | ||||
| 7460 | const APInt *C; | |||
| 7461 | if (match(V, m_APInt(C))) | |||
| 7462 | return ConstantInt::get(V->getType(), ~(*C)); | |||
| 7463 | ||||
| 7464 | return nullptr; | |||
| 7465 | } | |||
| 7466 | ||||
| 7467 | /// Match non-obvious integer minimum and maximum sequences. | |||
| 7468 | static SelectPatternResult matchMinMax(CmpInst::Predicate Pred, | |||
| 7469 | Value *CmpLHS, Value *CmpRHS, | |||
| 7470 | Value *TrueVal, Value *FalseVal, | |||
| 7471 | Value *&LHS, Value *&RHS, | |||
| 7472 | unsigned Depth) { | |||
| 7473 | // Assume success. If there's no match, callers should not use these anyway. | |||
| 7474 | LHS = TrueVal; | |||
| 7475 | RHS = FalseVal; | |||
| 7476 | ||||
| 7477 | SelectPatternResult SPR = matchClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal); | |||
| 7478 | if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN) | |||
| 7479 | return SPR; | |||
| 7480 | ||||
| 7481 | SPR = matchMinMaxOfMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, Depth); | |||
| 7482 | if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN) | |||
| 7483 | return SPR; | |||
| 7484 | ||||
| 7485 | // Look through 'not' ops to find disguised min/max. | |||
| 7486 | // (X > Y) ? ~X : ~Y ==> (~X < ~Y) ? ~X : ~Y ==> MIN(~X, ~Y) | |||
| 7487 | // (X < Y) ? ~X : ~Y ==> (~X > ~Y) ? ~X : ~Y ==> MAX(~X, ~Y) | |||
| 7488 | if (CmpLHS == getNotValue(TrueVal) && CmpRHS == getNotValue(FalseVal)) { | |||
| 7489 | switch (Pred) { | |||
| 7490 | case CmpInst::ICMP_SGT: return {SPF_SMIN, SPNB_NA, false}; | |||
| 7491 | case CmpInst::ICMP_SLT: return {SPF_SMAX, SPNB_NA, false}; | |||
| 7492 | case CmpInst::ICMP_UGT: return {SPF_UMIN, SPNB_NA, false}; | |||
| 7493 | case CmpInst::ICMP_ULT: return {SPF_UMAX, SPNB_NA, false}; | |||
| 7494 | default: break; | |||
| 7495 | } | |||
| 7496 | } | |||
| 7497 | ||||
| 7498 | // (X > Y) ? ~Y : ~X ==> (~X < ~Y) ? ~Y : ~X ==> MAX(~Y, ~X) | |||
| 7499 | // (X < Y) ? ~Y : ~X ==> (~X > ~Y) ? ~Y : ~X ==> MIN(~Y, ~X) | |||
| 7500 | if (CmpLHS == getNotValue(FalseVal) && CmpRHS == getNotValue(TrueVal)) { | |||
| 7501 | switch (Pred) { | |||
| 7502 | case CmpInst::ICMP_SGT: return {SPF_SMAX, SPNB_NA, false}; | |||
| 7503 | case CmpInst::ICMP_SLT: return {SPF_SMIN, SPNB_NA, false}; | |||
| 7504 | case CmpInst::ICMP_UGT: return {SPF_UMAX, SPNB_NA, false}; | |||
| 7505 | case CmpInst::ICMP_ULT: return {SPF_UMIN, SPNB_NA, false}; | |||
| 7506 | default: break; | |||
| 7507 | } | |||
| 7508 | } | |||
| 7509 | ||||
| 7510 | if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT) | |||
| 7511 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7512 | ||||
| 7513 | const APInt *C1; | |||
| 7514 | if (!match(CmpRHS, m_APInt(C1))) | |||
| 7515 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7516 | ||||
| 7517 | // An unsigned min/max can be written with a signed compare. | |||
| 7518 | const APInt *C2; | |||
| 7519 | if ((CmpLHS == TrueVal && match(FalseVal, m_APInt(C2))) || | |||
| 7520 | (CmpLHS == FalseVal && match(TrueVal, m_APInt(C2)))) { | |||
| 7521 | // Is the sign bit set? | |||
| 7522 | // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX | |||
| 7523 | // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN | |||
| 7524 | if (Pred == CmpInst::ICMP_SLT && C1->isZero() && C2->isMaxSignedValue()) | |||
| 7525 | return {CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false}; | |||
| 7526 | ||||
| 7527 | // Is the sign bit clear? | |||
| 7528 | // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX | |||
| 7529 | // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN | |||
| 7530 | if (Pred == CmpInst::ICMP_SGT && C1->isAllOnes() && C2->isMinSignedValue()) | |||
| 7531 | return {CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false}; | |||
| 7532 | } | |||
| 7533 | ||||
| 7534 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7535 | } | |||
| 7536 | ||||
| 7537 | bool llvm::isKnownNegation(const Value *X, const Value *Y, bool NeedNSW) { | |||
| 7538 | assert(X && Y && "Invalid operand")(static_cast <bool> (X && Y && "Invalid operand" ) ? void (0) : __assert_fail ("X && Y && \"Invalid operand\"" , "llvm/lib/Analysis/ValueTracking.cpp", 7538, __extension__ __PRETTY_FUNCTION__ )); | |||
| 7539 | ||||
| 7540 | // X = sub (0, Y) || X = sub nsw (0, Y) | |||
| 7541 | if ((!NeedNSW && match(X, m_Sub(m_ZeroInt(), m_Specific(Y)))) || | |||
| 7542 | (NeedNSW && match(X, m_NSWSub(m_ZeroInt(), m_Specific(Y))))) | |||
| 7543 | return true; | |||
| 7544 | ||||
| 7545 | // Y = sub (0, X) || Y = sub nsw (0, X) | |||
| 7546 | if ((!NeedNSW && match(Y, m_Sub(m_ZeroInt(), m_Specific(X)))) || | |||
| 7547 | (NeedNSW && match(Y, m_NSWSub(m_ZeroInt(), m_Specific(X))))) | |||
| 7548 | return true; | |||
| 7549 | ||||
| 7550 | // X = sub (A, B), Y = sub (B, A) || X = sub nsw (A, B), Y = sub nsw (B, A) | |||
| 7551 | Value *A, *B; | |||
| 7552 | return (!NeedNSW && (match(X, m_Sub(m_Value(A), m_Value(B))) && | |||
| 7553 | match(Y, m_Sub(m_Specific(B), m_Specific(A))))) || | |||
| 7554 | (NeedNSW && (match(X, m_NSWSub(m_Value(A), m_Value(B))) && | |||
| 7555 | match(Y, m_NSWSub(m_Specific(B), m_Specific(A))))); | |||
| 7556 | } | |||
| 7557 | ||||
| 7558 | static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred, | |||
| 7559 | FastMathFlags FMF, | |||
| 7560 | Value *CmpLHS, Value *CmpRHS, | |||
| 7561 | Value *TrueVal, Value *FalseVal, | |||
| 7562 | Value *&LHS, Value *&RHS, | |||
| 7563 | unsigned Depth) { | |||
| 7564 | bool HasMismatchedZeros = false; | |||
| 7565 | if (CmpInst::isFPPredicate(Pred)) { | |||
| 7566 | // IEEE-754 ignores the sign of 0.0 in comparisons. So if the select has one | |||
| 7567 | // 0.0 operand, set the compare's 0.0 operands to that same value for the | |||
| 7568 | // purpose of identifying min/max. Disregard vector constants with undefined | |||
| 7569 | // elements because those can not be back-propagated for analysis. | |||
| 7570 | Value *OutputZeroVal = nullptr; | |||
| 7571 | if (match(TrueVal, m_AnyZeroFP()) && !match(FalseVal, m_AnyZeroFP()) && | |||
| 7572 | !cast<Constant>(TrueVal)->containsUndefOrPoisonElement()) | |||
| 7573 | OutputZeroVal = TrueVal; | |||
| 7574 | else if (match(FalseVal, m_AnyZeroFP()) && !match(TrueVal, m_AnyZeroFP()) && | |||
| 7575 | !cast<Constant>(FalseVal)->containsUndefOrPoisonElement()) | |||
| 7576 | OutputZeroVal = FalseVal; | |||
| 7577 | ||||
| 7578 | if (OutputZeroVal) { | |||
| 7579 | if (match(CmpLHS, m_AnyZeroFP()) && CmpLHS != OutputZeroVal) { | |||
| 7580 | HasMismatchedZeros = true; | |||
| 7581 | CmpLHS = OutputZeroVal; | |||
| 7582 | } | |||
| 7583 | if (match(CmpRHS, m_AnyZeroFP()) && CmpRHS != OutputZeroVal) { | |||
| 7584 | HasMismatchedZeros = true; | |||
| 7585 | CmpRHS = OutputZeroVal; | |||
| 7586 | } | |||
| 7587 | } | |||
| 7588 | } | |||
| 7589 | ||||
| 7590 | LHS = CmpLHS; | |||
| 7591 | RHS = CmpRHS; | |||
| 7592 | ||||
| 7593 | // Signed zero may return inconsistent results between implementations. | |||
| 7594 | // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0 | |||
| 7595 | // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1) | |||
| 7596 | // Therefore, we behave conservatively and only proceed if at least one of the | |||
| 7597 | // operands is known to not be zero or if we don't care about signed zero. | |||
| 7598 | switch (Pred) { | |||
| 7599 | default: break; | |||
| 7600 | case CmpInst::FCMP_OGT: case CmpInst::FCMP_OLT: | |||
| 7601 | case CmpInst::FCMP_UGT: case CmpInst::FCMP_ULT: | |||
| 7602 | if (!HasMismatchedZeros) | |||
| 7603 | break; | |||
| 7604 | [[fallthrough]]; | |||
| 7605 | case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE: | |||
| 7606 | case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE: | |||
| 7607 | if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) && | |||
| 7608 | !isKnownNonZero(CmpRHS)) | |||
| 7609 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7610 | } | |||
| 7611 | ||||
| 7612 | SelectPatternNaNBehavior NaNBehavior = SPNB_NA; | |||
| 7613 | bool Ordered = false; | |||
| 7614 | ||||
| 7615 | // When given one NaN and one non-NaN input: | |||
| 7616 | // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input. | |||
| 7617 | // - A simple C99 (a < b ? a : b) construction will return 'b' (as the | |||
| 7618 | // ordered comparison fails), which could be NaN or non-NaN. | |||
| 7619 | // so here we discover exactly what NaN behavior is required/accepted. | |||
| 7620 | if (CmpInst::isFPPredicate(Pred)) { | |||
| 7621 | bool LHSSafe = isKnownNonNaN(CmpLHS, FMF); | |||
| 7622 | bool RHSSafe = isKnownNonNaN(CmpRHS, FMF); | |||
| 7623 | ||||
| 7624 | if (LHSSafe && RHSSafe) { | |||
| 7625 | // Both operands are known non-NaN. | |||
| 7626 | NaNBehavior = SPNB_RETURNS_ANY; | |||
| 7627 | } else if (CmpInst::isOrdered(Pred)) { | |||
| 7628 | // An ordered comparison will return false when given a NaN, so it | |||
| 7629 | // returns the RHS. | |||
| 7630 | Ordered = true; | |||
| 7631 | if (LHSSafe) | |||
| 7632 | // LHS is non-NaN, so if RHS is NaN then NaN will be returned. | |||
| 7633 | NaNBehavior = SPNB_RETURNS_NAN; | |||
| 7634 | else if (RHSSafe) | |||
| 7635 | NaNBehavior = SPNB_RETURNS_OTHER; | |||
| 7636 | else | |||
| 7637 | // Completely unsafe. | |||
| 7638 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7639 | } else { | |||
| 7640 | Ordered = false; | |||
| 7641 | // An unordered comparison will return true when given a NaN, so it | |||
| 7642 | // returns the LHS. | |||
| 7643 | if (LHSSafe) | |||
| 7644 | // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned. | |||
| 7645 | NaNBehavior = SPNB_RETURNS_OTHER; | |||
| 7646 | else if (RHSSafe) | |||
| 7647 | NaNBehavior = SPNB_RETURNS_NAN; | |||
| 7648 | else | |||
| 7649 | // Completely unsafe. | |||
| 7650 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7651 | } | |||
| 7652 | } | |||
| 7653 | ||||
| 7654 | if (TrueVal == CmpRHS && FalseVal == CmpLHS) { | |||
| 7655 | std::swap(CmpLHS, CmpRHS); | |||
| 7656 | Pred = CmpInst::getSwappedPredicate(Pred); | |||
| 7657 | if (NaNBehavior == SPNB_RETURNS_NAN) | |||
| 7658 | NaNBehavior = SPNB_RETURNS_OTHER; | |||
| 7659 | else if (NaNBehavior == SPNB_RETURNS_OTHER) | |||
| 7660 | NaNBehavior = SPNB_RETURNS_NAN; | |||
| 7661 | Ordered = !Ordered; | |||
| 7662 | } | |||
| 7663 | ||||
| 7664 | // ([if]cmp X, Y) ? X : Y | |||
| 7665 | if (TrueVal == CmpLHS && FalseVal == CmpRHS) { | |||
| 7666 | switch (Pred) { | |||
| 7667 | default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality. | |||
| 7668 | case ICmpInst::ICMP_UGT: | |||
| 7669 | case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false}; | |||
| 7670 | case ICmpInst::ICMP_SGT: | |||
| 7671 | case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false}; | |||
| 7672 | case ICmpInst::ICMP_ULT: | |||
| 7673 | case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false}; | |||
| 7674 | case ICmpInst::ICMP_SLT: | |||
| 7675 | case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false}; | |||
| 7676 | case FCmpInst::FCMP_UGT: | |||
| 7677 | case FCmpInst::FCMP_UGE: | |||
| 7678 | case FCmpInst::FCMP_OGT: | |||
| 7679 | case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered}; | |||
| 7680 | case FCmpInst::FCMP_ULT: | |||
| 7681 | case FCmpInst::FCMP_ULE: | |||
| 7682 | case FCmpInst::FCMP_OLT: | |||
| 7683 | case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered}; | |||
| 7684 | } | |||
| 7685 | } | |||
| 7686 | ||||
| 7687 | if (isKnownNegation(TrueVal, FalseVal)) { | |||
| 7688 | // Sign-extending LHS does not change its sign, so TrueVal/FalseVal can | |||
| 7689 | // match against either LHS or sext(LHS). | |||
| 7690 | auto MaybeSExtCmpLHS = | |||
| 7691 | m_CombineOr(m_Specific(CmpLHS), m_SExt(m_Specific(CmpLHS))); | |||
| 7692 | auto ZeroOrAllOnes = m_CombineOr(m_ZeroInt(), m_AllOnes()); | |||
| 7693 | auto ZeroOrOne = m_CombineOr(m_ZeroInt(), m_One()); | |||
| 7694 | if (match(TrueVal, MaybeSExtCmpLHS)) { | |||
| 7695 | // Set the return values. If the compare uses the negated value (-X >s 0), | |||
| 7696 | // swap the return values because the negated value is always 'RHS'. | |||
| 7697 | LHS = TrueVal; | |||
| 7698 | RHS = FalseVal; | |||
| 7699 | if (match(CmpLHS, m_Neg(m_Specific(FalseVal)))) | |||
| 7700 | std::swap(LHS, RHS); | |||
| 7701 | ||||
| 7702 | // (X >s 0) ? X : -X or (X >s -1) ? X : -X --> ABS(X) | |||
| 7703 | // (-X >s 0) ? -X : X or (-X >s -1) ? -X : X --> ABS(X) | |||
| 7704 | if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes)) | |||
| 7705 | return {SPF_ABS, SPNB_NA, false}; | |||
| 7706 | ||||
| 7707 | // (X >=s 0) ? X : -X or (X >=s 1) ? X : -X --> ABS(X) | |||
| 7708 | if (Pred == ICmpInst::ICMP_SGE && match(CmpRHS, ZeroOrOne)) | |||
| 7709 | return {SPF_ABS, SPNB_NA, false}; | |||
| 7710 | ||||
| 7711 | // (X <s 0) ? X : -X or (X <s 1) ? X : -X --> NABS(X) | |||
| 7712 | // (-X <s 0) ? -X : X or (-X <s 1) ? -X : X --> NABS(X) | |||
| 7713 | if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne)) | |||
| 7714 | return {SPF_NABS, SPNB_NA, false}; | |||
| 7715 | } | |||
| 7716 | else if (match(FalseVal, MaybeSExtCmpLHS)) { | |||
| 7717 | // Set the return values. If the compare uses the negated value (-X >s 0), | |||
| 7718 | // swap the return values because the negated value is always 'RHS'. | |||
| 7719 | LHS = FalseVal; | |||
| 7720 | RHS = TrueVal; | |||
| 7721 | if (match(CmpLHS, m_Neg(m_Specific(TrueVal)))) | |||
| 7722 | std::swap(LHS, RHS); | |||
| 7723 | ||||
| 7724 | // (X >s 0) ? -X : X or (X >s -1) ? -X : X --> NABS(X) | |||
| 7725 | // (-X >s 0) ? X : -X or (-X >s -1) ? X : -X --> NABS(X) | |||
| 7726 | if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes)) | |||
| 7727 | return {SPF_NABS, SPNB_NA, false}; | |||
| 7728 | ||||
| 7729 | // (X <s 0) ? -X : X or (X <s 1) ? -X : X --> ABS(X) | |||
| 7730 | // (-X <s 0) ? X : -X or (-X <s 1) ? X : -X --> ABS(X) | |||
| 7731 | if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne)) | |||
| 7732 | return {SPF_ABS, SPNB_NA, false}; | |||
| 7733 | } | |||
| 7734 | } | |||
| 7735 | ||||
| 7736 | if (CmpInst::isIntPredicate(Pred)) | |||
| 7737 | return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS, Depth); | |||
| 7738 | ||||
| 7739 | // According to (IEEE 754-2008 5.3.1), minNum(0.0, -0.0) and similar | |||
| 7740 | // may return either -0.0 or 0.0, so fcmp/select pair has stricter | |||
| 7741 | // semantics than minNum. Be conservative in such case. | |||
| 7742 | if (NaNBehavior != SPNB_RETURNS_ANY || | |||
| 7743 | (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) && | |||
| 7744 | !isKnownNonZero(CmpRHS))) | |||
| 7745 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7746 | ||||
| 7747 | return matchFastFloatClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS); | |||
| 7748 | } | |||
| 7749 | ||||
| 7750 | /// Helps to match a select pattern in case of a type mismatch. | |||
| 7751 | /// | |||
| 7752 | /// The function processes the case when type of true and false values of a | |||
| 7753 | /// select instruction differs from type of the cmp instruction operands because | |||
| 7754 | /// of a cast instruction. The function checks if it is legal to move the cast | |||
| 7755 | /// operation after "select". If yes, it returns the new second value of | |||
| 7756 | /// "select" (with the assumption that cast is moved): | |||
| 7757 | /// 1. As operand of cast instruction when both values of "select" are same cast | |||
| 7758 | /// instructions. | |||
| 7759 | /// 2. As restored constant (by applying reverse cast operation) when the first | |||
| 7760 | /// value of the "select" is a cast operation and the second value is a | |||
| 7761 | /// constant. | |||
| 7762 | /// NOTE: We return only the new second value because the first value could be | |||
| 7763 | /// accessed as operand of cast instruction. | |||
| 7764 | static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2, | |||
| 7765 | Instruction::CastOps *CastOp) { | |||
| 7766 | auto *Cast1 = dyn_cast<CastInst>(V1); | |||
| 7767 | if (!Cast1) | |||
| 7768 | return nullptr; | |||
| 7769 | ||||
| 7770 | *CastOp = Cast1->getOpcode(); | |||
| 7771 | Type *SrcTy = Cast1->getSrcTy(); | |||
| 7772 | if (auto *Cast2 = dyn_cast<CastInst>(V2)) { | |||
| 7773 | // If V1 and V2 are both the same cast from the same type, look through V1. | |||
| 7774 | if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy()) | |||
| 7775 | return Cast2->getOperand(0); | |||
| 7776 | return nullptr; | |||
| 7777 | } | |||
| 7778 | ||||
| 7779 | auto *C = dyn_cast<Constant>(V2); | |||
| 7780 | if (!C) | |||
| 7781 | return nullptr; | |||
| 7782 | ||||
| 7783 | Constant *CastedTo = nullptr; | |||
| 7784 | switch (*CastOp) { | |||
| 7785 | case Instruction::ZExt: | |||
| 7786 | if (CmpI->isUnsigned()) | |||
| 7787 | CastedTo = ConstantExpr::getTrunc(C, SrcTy); | |||
| 7788 | break; | |||
| 7789 | case Instruction::SExt: | |||
| 7790 | if (CmpI->isSigned()) | |||
| 7791 | CastedTo = ConstantExpr::getTrunc(C, SrcTy, true); | |||
| 7792 | break; | |||
| 7793 | case Instruction::Trunc: | |||
| 7794 | Constant *CmpConst; | |||
| 7795 | if (match(CmpI->getOperand(1), m_Constant(CmpConst)) && | |||
| 7796 | CmpConst->getType() == SrcTy) { | |||
| 7797 | // Here we have the following case: | |||
| 7798 | // | |||
| 7799 | // %cond = cmp iN %x, CmpConst | |||
| 7800 | // %tr = trunc iN %x to iK | |||
| 7801 | // %narrowsel = select i1 %cond, iK %t, iK C | |||
| 7802 | // | |||
| 7803 | // We can always move trunc after select operation: | |||
| 7804 | // | |||
| 7805 | // %cond = cmp iN %x, CmpConst | |||
| 7806 | // %widesel = select i1 %cond, iN %x, iN CmpConst | |||
| 7807 | // %tr = trunc iN %widesel to iK | |||
| 7808 | // | |||
| 7809 | // Note that C could be extended in any way because we don't care about | |||
| 7810 | // upper bits after truncation. It can't be abs pattern, because it would | |||
| 7811 | // look like: | |||
| 7812 | // | |||
| 7813 | // select i1 %cond, x, -x. | |||
| 7814 | // | |||
| 7815 | // So only min/max pattern could be matched. Such match requires widened C | |||
| 7816 | // == CmpConst. That is why set widened C = CmpConst, condition trunc | |||
| 7817 | // CmpConst == C is checked below. | |||
| 7818 | CastedTo = CmpConst; | |||
| 7819 | } else { | |||
| 7820 | CastedTo = ConstantExpr::getIntegerCast(C, SrcTy, CmpI->isSigned()); | |||
| 7821 | } | |||
| 7822 | break; | |||
| 7823 | case Instruction::FPTrunc: | |||
| 7824 | CastedTo = ConstantExpr::getFPExtend(C, SrcTy, true); | |||
| 7825 | break; | |||
| 7826 | case Instruction::FPExt: | |||
| 7827 | CastedTo = ConstantExpr::getFPTrunc(C, SrcTy, true); | |||
| 7828 | break; | |||
| 7829 | case Instruction::FPToUI: | |||
| 7830 | CastedTo = ConstantExpr::getUIToFP(C, SrcTy, true); | |||
| 7831 | break; | |||
| 7832 | case Instruction::FPToSI: | |||
| 7833 | CastedTo = ConstantExpr::getSIToFP(C, SrcTy, true); | |||
| 7834 | break; | |||
| 7835 | case Instruction::UIToFP: | |||
| 7836 | CastedTo = ConstantExpr::getFPToUI(C, SrcTy, true); | |||
| 7837 | break; | |||
| 7838 | case Instruction::SIToFP: | |||
| 7839 | CastedTo = ConstantExpr::getFPToSI(C, SrcTy, true); | |||
| 7840 | break; | |||
| 7841 | default: | |||
| 7842 | break; | |||
| 7843 | } | |||
| 7844 | ||||
| 7845 | if (!CastedTo) | |||
| 7846 | return nullptr; | |||
| 7847 | ||||
| 7848 | // Make sure the cast doesn't lose any information. | |||
| 7849 | Constant *CastedBack = | |||
| 7850 | ConstantExpr::getCast(*CastOp, CastedTo, C->getType(), true); | |||
| 7851 | if (CastedBack != C) | |||
| 7852 | return nullptr; | |||
| 7853 | ||||
| 7854 | return CastedTo; | |||
| 7855 | } | |||
| 7856 | ||||
| 7857 | SelectPatternResult llvm::matchSelectPattern(Value *V, Value *&LHS, Value *&RHS, | |||
| 7858 | Instruction::CastOps *CastOp, | |||
| 7859 | unsigned Depth) { | |||
| 7860 | if (Depth >= MaxAnalysisRecursionDepth) | |||
| 7861 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7862 | ||||
| 7863 | SelectInst *SI = dyn_cast<SelectInst>(V); | |||
| 7864 | if (!SI) return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7865 | ||||
| 7866 | CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition()); | |||
| 7867 | if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7868 | ||||
| 7869 | Value *TrueVal = SI->getTrueValue(); | |||
| 7870 | Value *FalseVal = SI->getFalseValue(); | |||
| 7871 | ||||
| 7872 | return llvm::matchDecomposedSelectPattern(CmpI, TrueVal, FalseVal, LHS, RHS, | |||
| 7873 | CastOp, Depth); | |||
| 7874 | } | |||
| 7875 | ||||
| 7876 | SelectPatternResult llvm::matchDecomposedSelectPattern( | |||
| 7877 | CmpInst *CmpI, Value *TrueVal, Value *FalseVal, Value *&LHS, Value *&RHS, | |||
| 7878 | Instruction::CastOps *CastOp, unsigned Depth) { | |||
| 7879 | CmpInst::Predicate Pred = CmpI->getPredicate(); | |||
| 7880 | Value *CmpLHS = CmpI->getOperand(0); | |||
| 7881 | Value *CmpRHS = CmpI->getOperand(1); | |||
| 7882 | FastMathFlags FMF; | |||
| 7883 | if (isa<FPMathOperator>(CmpI)) | |||
| 7884 | FMF = CmpI->getFastMathFlags(); | |||
| 7885 | ||||
| 7886 | // Bail out early. | |||
| 7887 | if (CmpI->isEquality()) | |||
| 7888 | return {SPF_UNKNOWN, SPNB_NA, false}; | |||
| 7889 | ||||
| 7890 | // Deal with type mismatches. | |||
| 7891 | if (CastOp && CmpLHS->getType() != TrueVal->getType()) { | |||
| 7892 | if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp)) { | |||
| 7893 | // If this is a potential fmin/fmax with a cast to integer, then ignore | |||
| 7894 | // -0.0 because there is no corresponding integer value. | |||
| 7895 | if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI) | |||
| 7896 | FMF.setNoSignedZeros(); | |||
| 7897 | return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, | |||
| 7898 | cast<CastInst>(TrueVal)->getOperand(0), C, | |||
| 7899 | LHS, RHS, Depth); | |||
| 7900 | } | |||
| 7901 | if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp)) { | |||
| 7902 | // If this is a potential fmin/fmax with a cast to integer, then ignore | |||
| 7903 | // -0.0 because there is no corresponding integer value. | |||
| 7904 | if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI) | |||
| 7905 | FMF.setNoSignedZeros(); | |||
| 7906 | return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, | |||
| 7907 | C, cast<CastInst>(FalseVal)->getOperand(0), | |||
| 7908 | LHS, RHS, Depth); | |||
| 7909 | } | |||
| 7910 | } | |||
| 7911 | return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal, | |||
| 7912 | LHS, RHS, Depth); | |||
| 7913 | } | |||
| 7914 | ||||
| 7915 | CmpInst::Predicate llvm::getMinMaxPred(SelectPatternFlavor SPF, bool Ordered) { | |||
| 7916 | if (SPF == SPF_SMIN) return ICmpInst::ICMP_SLT; | |||
| 7917 | if (SPF == SPF_UMIN) return ICmpInst::ICMP_ULT; | |||
| 7918 | if (SPF == SPF_SMAX) return ICmpInst::ICMP_SGT; | |||
| 7919 | if (SPF == SPF_UMAX) return ICmpInst::ICMP_UGT; | |||
| 7920 | if (SPF == SPF_FMINNUM) | |||
| 7921 | return Ordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; | |||
| 7922 | if (SPF == SPF_FMAXNUM) | |||
| 7923 | return Ordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; | |||
| 7924 | llvm_unreachable("unhandled!")::llvm::llvm_unreachable_internal("unhandled!", "llvm/lib/Analysis/ValueTracking.cpp" , 7924); | |||
| 7925 | } | |||
| 7926 | ||||
| 7927 | SelectPatternFlavor llvm::getInverseMinMaxFlavor(SelectPatternFlavor SPF) { | |||
| 7928 | if (SPF == SPF_SMIN) return SPF_SMAX; | |||
| 7929 | if (SPF == SPF_UMIN) return SPF_UMAX; | |||
| 7930 | if (SPF == SPF_SMAX) return SPF_SMIN; | |||
| 7931 | if (SPF == SPF_UMAX) return SPF_UMIN; | |||
| 7932 | llvm_unreachable("unhandled!")::llvm::llvm_unreachable_internal("unhandled!", "llvm/lib/Analysis/ValueTracking.cpp" , 7932); | |||
| 7933 | } | |||
| 7934 | ||||
| 7935 | Intrinsic::ID llvm::getInverseMinMaxIntrinsic(Intrinsic::ID MinMaxID) { | |||
| 7936 | switch (MinMaxID) { | |||
| 7937 | case Intrinsic::smax: return Intrinsic::smin; | |||
| 7938 | case Intrinsic::smin: return Intrinsic::smax; | |||
| 7939 | case Intrinsic::umax: return Intrinsic::umin; | |||
| 7940 | case Intrinsic::umin: return Intrinsic::umax; | |||
| 7941 | // Please note that next four intrinsics may produce the same result for | |||
| 7942 | // original and inverted case even if X != Y due to NaN is handled specially. | |||
| 7943 | case Intrinsic::maximum: return Intrinsic::minimum; | |||
| 7944 | case Intrinsic::minimum: return Intrinsic::maximum; | |||
| 7945 | case Intrinsic::maxnum: return Intrinsic::minnum; | |||
| 7946 | case Intrinsic::minnum: return Intrinsic::maxnum; | |||
| 7947 | default: llvm_unreachable("Unexpected intrinsic")::llvm::llvm_unreachable_internal("Unexpected intrinsic", "llvm/lib/Analysis/ValueTracking.cpp" , 7947); | |||
| 7948 | } | |||
| 7949 | } | |||
| 7950 | ||||
| 7951 | APInt llvm::getMinMaxLimit(SelectPatternFlavor SPF, unsigned BitWidth) { | |||
| 7952 | switch (SPF) { | |||
| 7953 | case SPF_SMAX: return APInt::getSignedMaxValue(BitWidth); | |||
| 7954 | case SPF_SMIN: return APInt::getSignedMinValue(BitWidth); | |||
| 7955 | case SPF_UMAX: return APInt::getMaxValue(BitWidth); | |||
| 7956 | case SPF_UMIN: return APInt::getMinValue(BitWidth); | |||
| 7957 | default: llvm_unreachable("Unexpected flavor")::llvm::llvm_unreachable_internal("Unexpected flavor", "llvm/lib/Analysis/ValueTracking.cpp" , 7957); | |||
| 7958 | } | |||
| 7959 | } | |||
| 7960 | ||||
| 7961 | std::pair<Intrinsic::ID, bool> | |||
| 7962 | llvm::canConvertToMinOrMaxIntrinsic(ArrayRef<Value *> VL) { | |||
| 7963 | // Check if VL contains select instructions that can be folded into a min/max | |||
| 7964 | // vector intrinsic and return the intrinsic if it is possible. | |||
| 7965 | // TODO: Support floating point min/max. | |||
| 7966 | bool AllCmpSingleUse = true; | |||
| 7967 | SelectPatternResult SelectPattern; | |||
| 7968 | SelectPattern.Flavor = SPF_UNKNOWN; | |||
| 7969 | if (all_of(VL, [&SelectPattern, &AllCmpSingleUse](Value *I) { | |||
| 7970 | Value *LHS, *RHS; | |||
| 7971 | auto CurrentPattern = matchSelectPattern(I, LHS, RHS); | |||
| 7972 | if (!SelectPatternResult::isMinOrMax(CurrentPattern.Flavor) || | |||
| 7973 | CurrentPattern.Flavor == SPF_FMINNUM || | |||
| 7974 | CurrentPattern.Flavor == SPF_FMAXNUM || | |||
| 7975 | !I->getType()->isIntOrIntVectorTy()) | |||
| 7976 | return false; | |||
| 7977 | if (SelectPattern.Flavor != SPF_UNKNOWN && | |||
| 7978 | SelectPattern.Flavor != CurrentPattern.Flavor) | |||
| 7979 | return false; | |||
| 7980 | SelectPattern = CurrentPattern; | |||
| 7981 | AllCmpSingleUse &= | |||
| 7982 | match(I, m_Select(m_OneUse(m_Value()), m_Value(), m_Value())); | |||
| 7983 | return true; | |||
| 7984 | })) { | |||
| 7985 | switch (SelectPattern.Flavor) { | |||
| 7986 | case SPF_SMIN: | |||
| 7987 | return {Intrinsic::smin, AllCmpSingleUse}; | |||
| 7988 | case SPF_UMIN: | |||
| 7989 | return {Intrinsic::umin, AllCmpSingleUse}; | |||
| 7990 | case SPF_SMAX: | |||
| 7991 | return {Intrinsic::smax, AllCmpSingleUse}; | |||
| 7992 | case SPF_UMAX: | |||
| 7993 | return {Intrinsic::umax, AllCmpSingleUse}; | |||
| 7994 | default: | |||
| 7995 | llvm_unreachable("unexpected select pattern flavor")::llvm::llvm_unreachable_internal("unexpected select pattern flavor" , "llvm/lib/Analysis/ValueTracking.cpp", 7995); | |||
| 7996 | } | |||
| 7997 | } | |||
| 7998 | return {Intrinsic::not_intrinsic, false}; | |||
| 7999 | } | |||
| 8000 | ||||
| 8001 | bool llvm::matchSimpleRecurrence(const PHINode *P, BinaryOperator *&BO, | |||
| 8002 | Value *&Start, Value *&Step) { | |||
| 8003 | // Handle the case of a simple two-predecessor recurrence PHI. | |||
| 8004 | // There's a lot more that could theoretically be done here, but | |||
| 8005 | // this is sufficient to catch some interesting cases. | |||
| 8006 | if (P->getNumIncomingValues() != 2) | |||
| 8007 | return false; | |||
| 8008 | ||||
| 8009 | for (unsigned i = 0; i != 2; ++i) { | |||
| 8010 | Value *L = P->getIncomingValue(i); | |||
| 8011 | Value *R = P->getIncomingValue(!i); | |||
| 8012 | Operator *LU = dyn_cast<Operator>(L); | |||
| 8013 | if (!LU) | |||
| 8014 | continue; | |||
| 8015 | unsigned Opcode = LU->getOpcode(); | |||
| 8016 | ||||
| 8017 | switch (Opcode) { | |||
| 8018 | default: | |||
| 8019 | continue; | |||
| 8020 | // TODO: Expand list -- xor, div, gep, uaddo, etc.. | |||
| 8021 | case Instruction::LShr: | |||
| 8022 | case Instruction::AShr: | |||
| 8023 | case Instruction::Shl: | |||
| 8024 | case Instruction::Add: | |||
| 8025 | case Instruction::Sub: | |||
| 8026 | case Instruction::And: | |||
| 8027 | case Instruction::Or: | |||
| 8028 | case Instruction::Mul: | |||
| 8029 | case Instruction::FMul: { | |||
| 8030 | Value *LL = LU->getOperand(0); | |||
| 8031 | Value *LR = LU->getOperand(1); | |||
| 8032 | // Find a recurrence. | |||
| 8033 | if (LL == P) | |||
| 8034 | L = LR; | |||
| 8035 | else if (LR == P) | |||
| 8036 | L = LL; | |||
| 8037 | else | |||
| 8038 | continue; // Check for recurrence with L and R flipped. | |||
| 8039 | ||||
| 8040 | break; // Match! | |||
| 8041 | } | |||
| 8042 | }; | |||
| 8043 | ||||
| 8044 | // We have matched a recurrence of the form: | |||
| 8045 | // %iv = [R, %entry], [%iv.next, %backedge] | |||
| 8046 | // %iv.next = binop %iv, L | |||
| 8047 | // OR | |||
| 8048 | // %iv = [R, %entry], [%iv.next, %backedge] | |||
| 8049 | // %iv.next = binop L, %iv | |||
| 8050 | BO = cast<BinaryOperator>(LU); | |||
| 8051 | Start = R; | |||
| 8052 | Step = L; | |||
| 8053 | return true; | |||
| 8054 | } | |||
| 8055 | return false; | |||
| 8056 | } | |||
| 8057 | ||||
| 8058 | bool llvm::matchSimpleRecurrence(const BinaryOperator *I, PHINode *&P, | |||
| 8059 | Value *&Start, Value *&Step) { | |||
| 8060 | BinaryOperator *BO = nullptr; | |||
| 8061 | P = dyn_cast<PHINode>(I->getOperand(0)); | |||
| 8062 | if (!P) | |||
| 8063 | P = dyn_cast<PHINode>(I->getOperand(1)); | |||
| 8064 | return P && matchSimpleRecurrence(P, BO, Start, Step) && BO == I; | |||
| 8065 | } | |||
| 8066 | ||||
| 8067 | /// Return true if "icmp Pred LHS RHS" is always true. | |||
| 8068 | static bool isTruePredicate(CmpInst::Predicate Pred, const Value *LHS, | |||
| 8069 | const Value *RHS, const DataLayout &DL, | |||
| 8070 | unsigned Depth) { | |||
| 8071 | if (ICmpInst::isTrueWhenEqual(Pred) && LHS == RHS) | |||
| 8072 | return true; | |||
| 8073 | ||||
| 8074 | switch (Pred) { | |||
| 8075 | default: | |||
| 8076 | return false; | |||
| 8077 | ||||
| 8078 | case CmpInst::ICMP_SLE: { | |||
| 8079 | const APInt *C; | |||
| 8080 | ||||
| 8081 | // LHS s<= LHS +_{nsw} C if C >= 0 | |||
| 8082 | if (match(RHS, m_NSWAdd(m_Specific(LHS), m_APInt(C)))) | |||
| 8083 | return !C->isNegative(); | |||
| 8084 | return false; | |||
| 8085 | } | |||
| 8086 | ||||
| 8087 | case CmpInst::ICMP_ULE: { | |||
| 8088 | const APInt *C; | |||
| 8089 | ||||
| 8090 | // LHS u<= LHS +_{nuw} C for any C | |||
| 8091 | if (match(RHS, m_NUWAdd(m_Specific(LHS), m_APInt(C)))) | |||
| 8092 | return true; | |||
| 8093 | ||||
| 8094 | // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB) | |||
| 8095 | auto MatchNUWAddsToSameValue = [&](const Value *A, const Value *B, | |||
| 8096 | const Value *&X, | |||
| 8097 | const APInt *&CA, const APInt *&CB) { | |||
| 8098 | if (match(A, m_NUWAdd(m_Value(X), m_APInt(CA))) && | |||
| 8099 | match(B, m_NUWAdd(m_Specific(X), m_APInt(CB)))) | |||
| 8100 | return true; | |||
| 8101 | ||||
| 8102 | // If X & C == 0 then (X | C) == X +_{nuw} C | |||
| 8103 | if (match(A, m_Or(m_Value(X), m_APInt(CA))) && | |||
| 8104 | match(B, m_Or(m_Specific(X), m_APInt(CB)))) { | |||
| 8105 | KnownBits Known(CA->getBitWidth()); | |||
| 8106 | computeKnownBits(X, Known, DL, Depth + 1, /*AC*/ nullptr, | |||
| 8107 | /*CxtI*/ nullptr, /*DT*/ nullptr); | |||
| 8108 | if (CA->isSubsetOf(Known.Zero) && CB->isSubsetOf(Known.Zero)) | |||
| 8109 | return true; | |||
| 8110 | } | |||
| 8111 | ||||
| 8112 | return false; | |||
| 8113 | }; | |||
| 8114 | ||||
| 8115 | const Value *X; | |||
| 8116 | const APInt *CLHS, *CRHS; | |||
| 8117 | if (MatchNUWAddsToSameValue(LHS, RHS, X, CLHS, CRHS)) | |||
| 8118 | return CLHS->ule(*CRHS); | |||
| 8119 | ||||
| 8120 | return false; | |||
| 8121 | } | |||
| 8122 | } | |||
| 8123 | } | |||
| 8124 | ||||
| 8125 | /// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred | |||
| 8126 | /// ALHS ARHS" is true. Otherwise, return std::nullopt. | |||
| 8127 | static std::optional<bool> | |||
| 8128 | isImpliedCondOperands(CmpInst::Predicate Pred, const Value *ALHS, | |||
| 8129 | const Value *ARHS, const Value *BLHS, const Value *BRHS, | |||
| 8130 | const DataLayout &DL, unsigned Depth) { | |||
| 8131 | switch (Pred) { | |||
| 8132 | default: | |||
| 8133 | return std::nullopt; | |||
| 8134 | ||||
| 8135 | case CmpInst::ICMP_SLT: | |||
| 8136 | case CmpInst::ICMP_SLE: | |||
| 8137 | if (isTruePredicate(CmpInst::ICMP_SLE, BLHS, ALHS, DL, Depth) && | |||
| 8138 | isTruePredicate(CmpInst::ICMP_SLE, ARHS, BRHS, DL, Depth)) | |||
| 8139 | return true; | |||
| 8140 | return std::nullopt; | |||
| 8141 | ||||
| 8142 | case CmpInst::ICMP_SGT: | |||
| 8143 | case CmpInst::ICMP_SGE: | |||
| 8144 | if (isTruePredicate(CmpInst::ICMP_SLE, ALHS, BLHS, DL, Depth) && | |||
| 8145 | isTruePredicate(CmpInst::ICMP_SLE, BRHS, ARHS, DL, Depth)) | |||
| 8146 | return true; | |||
| 8147 | return std::nullopt; | |||
| 8148 | ||||
| 8149 | case CmpInst::ICMP_ULT: | |||
| 8150 | case CmpInst::ICMP_ULE: | |||
| 8151 | if (isTruePredicate(CmpInst::ICMP_ULE, BLHS, ALHS, DL, Depth) && | |||
| 8152 | isTruePredicate(CmpInst::ICMP_ULE, ARHS, BRHS, DL, Depth)) | |||
| 8153 | return true; | |||
| 8154 | return std::nullopt; | |||
| 8155 | ||||
| 8156 | case CmpInst::ICMP_UGT: | |||
| 8157 | case CmpInst::ICMP_UGE: | |||
| 8158 | if (isTruePredicate(CmpInst::ICMP_ULE, ALHS, BLHS, DL, Depth) && | |||
| 8159 | isTruePredicate(CmpInst::ICMP_ULE, BRHS, ARHS, DL, Depth)) | |||
| 8160 | return true; | |||
| 8161 | return std::nullopt; | |||
| 8162 | } | |||
| 8163 | } | |||
| 8164 | ||||
| 8165 | /// Return true if the operands of two compares (expanded as "L0 pred L1" and | |||
| 8166 | /// "R0 pred R1") match. IsSwappedOps is true when the operands match, but are | |||
| 8167 | /// swapped. | |||
| 8168 | static bool areMatchingOperands(const Value *L0, const Value *L1, const Value *R0, | |||
| 8169 | const Value *R1, bool &AreSwappedOps) { | |||
| 8170 | bool AreMatchingOps = (L0 == R0 && L1 == R1); | |||
| 8171 | AreSwappedOps = (L0 == R1 && L1 == R0); | |||
| 8172 | return AreMatchingOps || AreSwappedOps; | |||
| 8173 | } | |||
| 8174 | ||||
| 8175 | /// Return true if "icmp1 LPred X, Y" implies "icmp2 RPred X, Y" is true. | |||
| 8176 | /// Return false if "icmp1 LPred X, Y" implies "icmp2 RPred X, Y" is false. | |||
| 8177 | /// Otherwise, return std::nullopt if we can't infer anything. | |||
| 8178 | static std::optional<bool> | |||
| 8179 | isImpliedCondMatchingOperands(CmpInst::Predicate LPred, | |||
| 8180 | CmpInst::Predicate RPred, bool AreSwappedOps) { | |||
| 8181 | // Canonicalize the predicate as if the operands were not commuted. | |||
| 8182 | if (AreSwappedOps) | |||
| 8183 | RPred = ICmpInst::getSwappedPredicate(RPred); | |||
| 8184 | ||||
| 8185 | if (CmpInst::isImpliedTrueByMatchingCmp(LPred, RPred)) | |||
| 8186 | return true; | |||
| 8187 | if (CmpInst::isImpliedFalseByMatchingCmp(LPred, RPred)) | |||
| 8188 | return false; | |||
| 8189 | ||||
| 8190 | return std::nullopt; | |||
| 8191 | } | |||
| 8192 | ||||
| 8193 | /// Return true if "icmp LPred X, LC" implies "icmp RPred X, RC" is true. | |||
| 8194 | /// Return false if "icmp LPred X, LC" implies "icmp RPred X, RC" is false. | |||
| 8195 | /// Otherwise, return std::nullopt if we can't infer anything. | |||
| 8196 | static std::optional<bool> isImpliedCondCommonOperandWithConstants( | |||
| 8197 | CmpInst::Predicate LPred, const APInt &LC, CmpInst::Predicate RPred, | |||
| 8198 | const APInt &RC) { | |||
| 8199 | ConstantRange DomCR = ConstantRange::makeExactICmpRegion(LPred, LC); | |||
| 8200 | ConstantRange CR = ConstantRange::makeExactICmpRegion(RPred, RC); | |||
| 8201 | ConstantRange Intersection = DomCR.intersectWith(CR); | |||
| 8202 | ConstantRange Difference = DomCR.difference(CR); | |||
| 8203 | if (Intersection.isEmptySet()) | |||
| 8204 | return false; | |||
| 8205 | if (Difference.isEmptySet()) | |||
| 8206 | return true; | |||
| 8207 | return std::nullopt; | |||
| 8208 | } | |||
| 8209 | ||||
| 8210 | /// Return true if LHS implies RHS (expanded to its components as "R0 RPred R1") | |||
| 8211 | /// is true. Return false if LHS implies RHS is false. Otherwise, return | |||
| 8212 | /// std::nullopt if we can't infer anything. | |||
| 8213 | static std::optional<bool> isImpliedCondICmps(const ICmpInst *LHS, | |||
| 8214 | CmpInst::Predicate RPred, | |||
| 8215 | const Value *R0, const Value *R1, | |||
| 8216 | const DataLayout &DL, | |||
| 8217 | bool LHSIsTrue, unsigned Depth) { | |||
| 8218 | Value *L0 = LHS->getOperand(0); | |||
| 8219 | Value *L1 = LHS->getOperand(1); | |||
| 8220 | ||||
| 8221 | // The rest of the logic assumes the LHS condition is true. If that's not the | |||
| 8222 | // case, invert the predicate to make it so. | |||
| 8223 | CmpInst::Predicate LPred = | |||
| 8224 | LHSIsTrue ? LHS->getPredicate() : LHS->getInversePredicate(); | |||
| 8225 | ||||
| 8226 | // Can we infer anything when the two compares have matching operands? | |||
| 8227 | bool AreSwappedOps; | |||
| 8228 | if (areMatchingOperands(L0, L1, R0, R1, AreSwappedOps)) | |||
| 8229 | return isImpliedCondMatchingOperands(LPred, RPred, AreSwappedOps); | |||
| 8230 | ||||
| 8231 | // Can we infer anything when the 0-operands match and the 1-operands are | |||
| 8232 | // constants (not necessarily matching)? | |||
| 8233 | const APInt *LC, *RC; | |||
| 8234 | if (L0 == R0 && match(L1, m_APInt(LC)) && match(R1, m_APInt(RC))) | |||
| 8235 | return isImpliedCondCommonOperandWithConstants(LPred, *LC, RPred, *RC); | |||
| 8236 | ||||
| 8237 | if (LPred == RPred) | |||
| 8238 | return isImpliedCondOperands(LPred, L0, L1, R0, R1, DL, Depth); | |||
| 8239 | ||||
| 8240 | return std::nullopt; | |||
| 8241 | } | |||
| 8242 | ||||
| 8243 | /// Return true if LHS implies RHS is true. Return false if LHS implies RHS is | |||
| 8244 | /// false. Otherwise, return std::nullopt if we can't infer anything. We | |||
| 8245 | /// expect the RHS to be an icmp and the LHS to be an 'and', 'or', or a 'select' | |||
| 8246 | /// instruction. | |||
| 8247 | static std::optional<bool> | |||
| 8248 | isImpliedCondAndOr(const Instruction *LHS, CmpInst::Predicate RHSPred, | |||
| 8249 | const Value *RHSOp0, const Value *RHSOp1, | |||
| 8250 | const DataLayout &DL, bool LHSIsTrue, unsigned Depth) { | |||
| 8251 | // The LHS must be an 'or', 'and', or a 'select' instruction. | |||
| 8252 | assert((LHS->getOpcode() == Instruction::And ||(static_cast <bool> ((LHS->getOpcode() == Instruction ::And || LHS->getOpcode() == Instruction::Or || LHS->getOpcode () == Instruction::Select) && "Expected LHS to be 'and', 'or', or 'select'." ) ? void (0) : __assert_fail ("(LHS->getOpcode() == Instruction::And || LHS->getOpcode() == Instruction::Or || LHS->getOpcode() == Instruction::Select) && \"Expected LHS to be 'and', 'or', or 'select'.\"" , "llvm/lib/Analysis/ValueTracking.cpp", 8255, __extension__ __PRETTY_FUNCTION__ )) | |||
| 8253 | LHS->getOpcode() == Instruction::Or ||(static_cast <bool> ((LHS->getOpcode() == Instruction ::And || LHS->getOpcode() == Instruction::Or || LHS->getOpcode () == Instruction::Select) && "Expected LHS to be 'and', 'or', or 'select'." ) ? void (0) : __assert_fail ("(LHS->getOpcode() == Instruction::And || LHS->getOpcode() == Instruction::Or || LHS->getOpcode() == Instruction::Select) && \"Expected LHS to be 'and', 'or', or 'select'.\"" , "llvm/lib/Analysis/ValueTracking.cpp", 8255, __extension__ __PRETTY_FUNCTION__ )) | |||
| 8254 | LHS->getOpcode() == Instruction::Select) &&(static_cast <bool> ((LHS->getOpcode() == Instruction ::And || LHS->getOpcode() == Instruction::Or || LHS->getOpcode () == Instruction::Select) && "Expected LHS to be 'and', 'or', or 'select'." ) ? void (0) : __assert_fail ("(LHS->getOpcode() == Instruction::And || LHS->getOpcode() == Instruction::Or || LHS->getOpcode() == Instruction::Select) && \"Expected LHS to be 'and', 'or', or 'select'.\"" , "llvm/lib/Analysis/ValueTracking.cpp", 8255, __extension__ __PRETTY_FUNCTION__ )) | |||
| 8255 | "Expected LHS to be 'and', 'or', or 'select'.")(static_cast <bool> ((LHS->getOpcode() == Instruction ::And || LHS->getOpcode() == Instruction::Or || LHS->getOpcode () == Instruction::Select) && "Expected LHS to be 'and', 'or', or 'select'." ) ? void (0) : __assert_fail ("(LHS->getOpcode() == Instruction::And || LHS->getOpcode() == Instruction::Or || LHS->getOpcode() == Instruction::Select) && \"Expected LHS to be 'and', 'or', or 'select'.\"" , "llvm/lib/Analysis/ValueTracking.cpp", 8255, __extension__ __PRETTY_FUNCTION__ )); | |||
| 8256 | ||||
| 8257 | assert(Depth <= MaxAnalysisRecursionDepth && "Hit recursion limit")(static_cast <bool> (Depth <= MaxAnalysisRecursionDepth && "Hit recursion limit") ? void (0) : __assert_fail ("Depth <= MaxAnalysisRecursionDepth && \"Hit recursion limit\"" , "llvm/lib/Analysis/ValueTracking.cpp", 8257, __extension__ __PRETTY_FUNCTION__ )); | |||
| 8258 | ||||
| 8259 | // If the result of an 'or' is false, then we know both legs of the 'or' are | |||
| 8260 | // false. Similarly, if the result of an 'and' is true, then we know both | |||
| 8261 | // legs of the 'and' are true. | |||
| 8262 | const Value *ALHS, *ARHS; | |||
| 8263 | if ((!LHSIsTrue && match(LHS, m_LogicalOr(m_Value(ALHS), m_Value(ARHS)))) || | |||
| 8264 | (LHSIsTrue && match(LHS, m_LogicalAnd(m_Value(ALHS), m_Value(ARHS))))) { | |||
| 8265 | // FIXME: Make this non-recursion. | |||
| 8266 | if (std::optional<bool> Implication = isImpliedCondition( | |||
| 8267 | ALHS, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, Depth + 1)) | |||
| 8268 | return Implication; | |||
| 8269 | if (std::optional<bool> Implication = isImpliedCondition( | |||
| 8270 | ARHS, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, Depth + 1)) | |||
| 8271 | return Implication; | |||
| 8272 | return std::nullopt; | |||
| 8273 | } | |||
| 8274 | return std::nullopt; | |||
| 8275 | } | |||
| 8276 | ||||
| 8277 | std::optional<bool> | |||
| 8278 | llvm::isImpliedCondition(const Value *LHS, CmpInst::Predicate RHSPred, | |||
| 8279 | const Value *RHSOp0, const Value *RHSOp1, | |||
| 8280 | const DataLayout &DL, bool LHSIsTrue, unsigned Depth) { | |||
| 8281 | // Bail out when we hit the limit. | |||
| 8282 | if (Depth == MaxAnalysisRecursionDepth) | |||
| 8283 | return std::nullopt; | |||
| 8284 | ||||
| 8285 | // A mismatch occurs when we compare a scalar cmp to a vector cmp, for | |||
| 8286 | // example. | |||
| 8287 | if (RHSOp0->getType()->isVectorTy() != LHS->getType()->isVectorTy()) | |||
| 8288 | return std::nullopt; | |||
| 8289 | ||||
| 8290 | assert(LHS->getType()->isIntOrIntVectorTy(1) &&(static_cast <bool> (LHS->getType()->isIntOrIntVectorTy (1) && "Expected integer type only!") ? void (0) : __assert_fail ("LHS->getType()->isIntOrIntVectorTy(1) && \"Expected integer type only!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 8291, __extension__ __PRETTY_FUNCTION__ )) | |||
| 8291 | "Expected integer type only!")(static_cast <bool> (LHS->getType()->isIntOrIntVectorTy (1) && "Expected integer type only!") ? void (0) : __assert_fail ("LHS->getType()->isIntOrIntVectorTy(1) && \"Expected integer type only!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 8291, __extension__ __PRETTY_FUNCTION__ )); | |||
| 8292 | ||||
| 8293 | // Both LHS and RHS are icmps. | |||
| 8294 | const ICmpInst *LHSCmp = dyn_cast<ICmpInst>(LHS); | |||
| 8295 | if (LHSCmp) | |||
| 8296 | return isImpliedCondICmps(LHSCmp, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, | |||
| 8297 | Depth); | |||
| 8298 | ||||
| 8299 | /// The LHS should be an 'or', 'and', or a 'select' instruction. We expect | |||
| 8300 | /// the RHS to be an icmp. | |||
| 8301 | /// FIXME: Add support for and/or/select on the RHS. | |||
| 8302 | if (const Instruction *LHSI = dyn_cast<Instruction>(LHS)) { | |||
| 8303 | if ((LHSI->getOpcode() == Instruction::And || | |||
| 8304 | LHSI->getOpcode() == Instruction::Or || | |||
| 8305 | LHSI->getOpcode() == Instruction::Select)) | |||
| 8306 | return isImpliedCondAndOr(LHSI, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, | |||
| 8307 | Depth); | |||
| 8308 | } | |||
| 8309 | return std::nullopt; | |||
| 8310 | } | |||
| 8311 | ||||
| 8312 | std::optional<bool> llvm::isImpliedCondition(const Value *LHS, const Value *RHS, | |||
| 8313 | const DataLayout &DL, | |||
| 8314 | bool LHSIsTrue, unsigned Depth) { | |||
| 8315 | // LHS ==> RHS by definition | |||
| 8316 | if (LHS == RHS) | |||
| 8317 | return LHSIsTrue; | |||
| 8318 | ||||
| 8319 | if (const ICmpInst *RHSCmp = dyn_cast<ICmpInst>(RHS)) | |||
| 8320 | return isImpliedCondition(LHS, RHSCmp->getPredicate(), | |||
| 8321 | RHSCmp->getOperand(0), RHSCmp->getOperand(1), DL, | |||
| 8322 | LHSIsTrue, Depth); | |||
| 8323 | ||||
| 8324 | if (Depth == MaxAnalysisRecursionDepth) | |||
| 8325 | return std::nullopt; | |||
| 8326 | ||||
| 8327 | // LHS ==> (RHS1 || RHS2) if LHS ==> RHS1 or LHS ==> RHS2 | |||
| 8328 | // LHS ==> !(RHS1 && RHS2) if LHS ==> !RHS1 or LHS ==> !RHS2 | |||
| 8329 | const Value *RHS1, *RHS2; | |||
| 8330 | if (match(RHS, m_LogicalOr(m_Value(RHS1), m_Value(RHS2)))) { | |||
| 8331 | if (std::optional<bool> Imp = | |||
| 8332 | isImpliedCondition(LHS, RHS1, DL, LHSIsTrue, Depth + 1)) | |||
| 8333 | if (*Imp == true) | |||
| 8334 | return true; | |||
| 8335 | if (std::optional<bool> Imp = | |||
| 8336 | isImpliedCondition(LHS, RHS2, DL, LHSIsTrue, Depth + 1)) | |||
| 8337 | if (*Imp == true) | |||
| 8338 | return true; | |||
| 8339 | } | |||
| 8340 | if (match(RHS, m_LogicalAnd(m_Value(RHS1), m_Value(RHS2)))) { | |||
| 8341 | if (std::optional<bool> Imp = | |||
| 8342 | isImpliedCondition(LHS, RHS1, DL, LHSIsTrue, Depth + 1)) | |||
| 8343 | if (*Imp == false) | |||
| 8344 | return false; | |||
| 8345 | if (std::optional<bool> Imp = | |||
| 8346 | isImpliedCondition(LHS, RHS2, DL, LHSIsTrue, Depth + 1)) | |||
| 8347 | if (*Imp == false) | |||
| 8348 | return false; | |||
| 8349 | } | |||
| 8350 | ||||
| 8351 | return std::nullopt; | |||
| 8352 | } | |||
| 8353 | ||||
| 8354 | // Returns a pair (Condition, ConditionIsTrue), where Condition is a branch | |||
| 8355 | // condition dominating ContextI or nullptr, if no condition is found. | |||
| 8356 | static std::pair<Value *, bool> | |||
| 8357 | getDomPredecessorCondition(const Instruction *ContextI) { | |||
| 8358 | if (!ContextI || !ContextI->getParent()) | |||
| 8359 | return {nullptr, false}; | |||
| 8360 | ||||
| 8361 | // TODO: This is a poor/cheap way to determine dominance. Should we use a | |||
| 8362 | // dominator tree (eg, from a SimplifyQuery) instead? | |||
| 8363 | const BasicBlock *ContextBB = ContextI->getParent(); | |||
| 8364 | const BasicBlock *PredBB = ContextBB->getSinglePredecessor(); | |||
| 8365 | if (!PredBB) | |||
| 8366 | return {nullptr, false}; | |||
| 8367 | ||||
| 8368 | // We need a conditional branch in the predecessor. | |||
| 8369 | Value *PredCond; | |||
| 8370 | BasicBlock *TrueBB, *FalseBB; | |||
| 8371 | if (!match(PredBB->getTerminator(), m_Br(m_Value(PredCond), TrueBB, FalseBB))) | |||
| 8372 | return {nullptr, false}; | |||
| 8373 | ||||
| 8374 | // The branch should get simplified. Don't bother simplifying this condition. | |||
| 8375 | if (TrueBB == FalseBB) | |||
| 8376 | return {nullptr, false}; | |||
| 8377 | ||||
| 8378 | assert((TrueBB == ContextBB || FalseBB == ContextBB) &&(static_cast <bool> ((TrueBB == ContextBB || FalseBB == ContextBB) && "Predecessor block does not point to successor?" ) ? void (0) : __assert_fail ("(TrueBB == ContextBB || FalseBB == ContextBB) && \"Predecessor block does not point to successor?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 8379, __extension__ __PRETTY_FUNCTION__ )) | |||
| 8379 | "Predecessor block does not point to successor?")(static_cast <bool> ((TrueBB == ContextBB || FalseBB == ContextBB) && "Predecessor block does not point to successor?" ) ? void (0) : __assert_fail ("(TrueBB == ContextBB || FalseBB == ContextBB) && \"Predecessor block does not point to successor?\"" , "llvm/lib/Analysis/ValueTracking.cpp", 8379, __extension__ __PRETTY_FUNCTION__ )); | |||
| 8380 | ||||
| 8381 | // Is this condition implied by the predecessor condition? | |||
| 8382 | return {PredCond, TrueBB == ContextBB}; | |||
| 8383 | } | |||
| 8384 | ||||
| 8385 | std::optional<bool> llvm::isImpliedByDomCondition(const Value *Cond, | |||
| 8386 | const Instruction *ContextI, | |||
| 8387 | const DataLayout &DL) { | |||
| 8388 | assert(Cond->getType()->isIntOrIntVectorTy(1) && "Condition must be bool")(static_cast <bool> (Cond->getType()->isIntOrIntVectorTy (1) && "Condition must be bool") ? void (0) : __assert_fail ("Cond->getType()->isIntOrIntVectorTy(1) && \"Condition must be bool\"" , "llvm/lib/Analysis/ValueTracking.cpp", 8388, __extension__ __PRETTY_FUNCTION__ )); | |||
| 8389 | auto PredCond = getDomPredecessorCondition(ContextI); | |||
| 8390 | if (PredCond.first) | |||
| 8391 | return isImpliedCondition(PredCond.first, Cond, DL, PredCond.second); | |||
| 8392 | return std::nullopt; | |||
| 8393 | } | |||
| 8394 | ||||
| 8395 | std::optional<bool> llvm::isImpliedByDomCondition(CmpInst::Predicate Pred, | |||
| 8396 | const Value *LHS, | |||
| 8397 | const Value *RHS, | |||
| 8398 | const Instruction *ContextI, | |||
| 8399 | const DataLayout &DL) { | |||
| 8400 | auto PredCond = getDomPredecessorCondition(ContextI); | |||
| 8401 | if (PredCond.first) | |||
| 8402 | return isImpliedCondition(PredCond.first, Pred, LHS, RHS, DL, | |||
| 8403 | PredCond.second); | |||
| 8404 | return std::nullopt; | |||
| 8405 | } | |||
| 8406 | ||||
| 8407 | static void setLimitsForBinOp(const BinaryOperator &BO, APInt &Lower, | |||
| 8408 | APInt &Upper, const InstrInfoQuery &IIQ, | |||
| 8409 | bool PreferSignedRange) { | |||
| 8410 | unsigned Width = Lower.getBitWidth(); | |||
| 8411 | const APInt *C; | |||
| 8412 | switch (BO.getOpcode()) { | |||
| 8413 | case Instruction::Add: | |||
| 8414 | if (match(BO.getOperand(1), m_APInt(C)) && !C->isZero()) { | |||
| 8415 | bool HasNSW = IIQ.hasNoSignedWrap(&BO); | |||
| 8416 | bool HasNUW = IIQ.hasNoUnsignedWrap(&BO); | |||
| 8417 | ||||
| 8418 | // If the caller expects a signed compare, then try to use a signed range. | |||
| 8419 | // Otherwise if both no-wraps are set, use the unsigned range because it | |||
| 8420 | // is never larger than the signed range. Example: | |||
| 8421 | // "add nuw nsw i8 X, -2" is unsigned [254,255] vs. signed [-128, 125]. | |||
| 8422 | if (PreferSignedRange && HasNSW && HasNUW) | |||
| 8423 | HasNUW = false; | |||
| 8424 | ||||
| 8425 | if (HasNUW) { | |||
| 8426 | // 'add nuw x, C' produces [C, UINT_MAX]. | |||
| 8427 | Lower = *C; | |||
| 8428 | } else if (HasNSW) { | |||
| 8429 | if (C->isNegative()) { | |||
| 8430 | // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C]. | |||
| 8431 | Lower = APInt::getSignedMinValue(Width); | |||
| 8432 | Upper = APInt::getSignedMaxValue(Width) + *C + 1; | |||
| 8433 | } else { | |||
| 8434 | // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX]. | |||
| 8435 | Lower = APInt::getSignedMinValue(Width) + *C; | |||
| 8436 | Upper = APInt::getSignedMaxValue(Width) + 1; | |||
| 8437 | } | |||
| 8438 | } | |||
| 8439 | } | |||
| 8440 | break; | |||
| 8441 | ||||
| 8442 | case Instruction::And: | |||
| 8443 | if (match(BO.getOperand(1), m_APInt(C))) | |||
| 8444 | // 'and x, C' produces [0, C]. | |||
| 8445 | Upper = *C + 1; | |||
| 8446 | break; | |||
| 8447 | ||||
| 8448 | case Instruction::Or: | |||
| 8449 | if (match(BO.getOperand(1), m_APInt(C))) | |||
| 8450 | // 'or x, C' produces [C, UINT_MAX]. | |||
| 8451 | Lower = *C; | |||
| 8452 | break; | |||
| 8453 | ||||
| 8454 | case Instruction::AShr: | |||
| 8455 | if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) { | |||
| 8456 | // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C]. | |||
| 8457 | Lower = APInt::getSignedMinValue(Width).ashr(*C); | |||
| 8458 | Upper = APInt::getSignedMaxValue(Width).ashr(*C) + 1; | |||
| 8459 | } else if (match(BO.getOperand(0), m_APInt(C))) { | |||
| 8460 | unsigned ShiftAmount = Width - 1; | |||
| 8461 | if (!C->isZero() && IIQ.isExact(&BO)) | |||
| 8462 | ShiftAmount = C->countr_zero(); | |||
| 8463 | if (C->isNegative()) { | |||
| 8464 | // 'ashr C, x' produces [C, C >> (Width-1)] | |||
| 8465 | Lower = *C; | |||
| 8466 | Upper = C->ashr(ShiftAmount) + 1; | |||
| 8467 | } else { | |||
| 8468 | // 'ashr C, x' produces [C >> (Width-1), C] | |||
| 8469 | Lower = C->ashr(ShiftAmount); | |||
| 8470 | Upper = *C + 1; | |||
| 8471 | } | |||
| 8472 | } | |||
| 8473 | break; | |||
| 8474 | ||||
| 8475 | case Instruction::LShr: | |||
| 8476 | if (match(BO.getOperand(1), m_APInt(C)) && C->ult(Width)) { | |||
| 8477 | // 'lshr x, C' produces [0, UINT_MAX >> C]. | |||
| 8478 | Upper = APInt::getAllOnes(Width).lshr(*C) + 1; | |||
| 8479 | } else if (match(BO.getOperand(0), m_APInt(C))) { | |||
| 8480 | // 'lshr C, x' produces [C >> (Width-1), C]. | |||
| 8481 | unsigned ShiftAmount = Width - 1; | |||
| 8482 | if (!C->isZero() && IIQ.isExact(&BO)) | |||
| 8483 | ShiftAmount = C->countr_zero(); | |||
| 8484 | Lower = C->lshr(ShiftAmount); | |||
| 8485 | Upper = *C + 1; | |||
| 8486 | } | |||
| 8487 | break; | |||
| 8488 | ||||
| 8489 | case Instruction::Shl: | |||
| 8490 | if (match(BO.getOperand(0), m_APInt(C))) { | |||
| 8491 | if (IIQ.hasNoUnsignedWrap(&BO)) { | |||
| 8492 | // 'shl nuw C, x' produces [C, C << CLZ(C)] | |||
| 8493 | Lower = *C; | |||
| 8494 | Upper = Lower.shl(Lower.countl_zero()) + 1; | |||
| 8495 | } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw? | |||
| 8496 | if (C->isNegative()) { | |||
| 8497 | // 'shl nsw C, x' produces [C << CLO(C)-1, C] | |||
| 8498 | unsigned ShiftAmount = C->countl_one() - 1; | |||
| 8499 | Lower = C->shl(ShiftAmount); | |||
| 8500 | Upper = *C + 1; | |||
| 8501 | } else { | |||
| 8502 | // 'shl nsw C, x' produces [C, C << CLZ(C)-1] | |||
| 8503 | unsigned ShiftAmount = C->countl_zero() - 1; | |||
| 8504 | Lower = *C; | |||
| 8505 | Upper = C->shl(ShiftAmount) + 1; | |||
| 8506 | } | |||
| 8507 | } | |||
| 8508 | } | |||
| 8509 | break; | |||
| 8510 | ||||
| 8511 | case Instruction::SDiv: | |||
| 8512 | if (match(BO.getOperand(1), m_APInt(C))) { | |||
| 8513 | APInt IntMin = APInt::getSignedMinValue(Width); | |||
| 8514 | APInt IntMax = APInt::getSignedMaxValue(Width); | |||
| 8515 | if (C->isAllOnes()) { | |||
| 8516 | // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX] | |||
| 8517 | // where C != -1 and C != 0 and C != 1 | |||
| 8518 | Lower = IntMin + 1; | |||
| 8519 | Upper = IntMax + 1; | |||
| 8520 | } else if (C->countl_zero() < Width - 1) { | |||
| 8521 | // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C] | |||
| 8522 | // where C != -1 and C != 0 and C != 1 | |||
| 8523 | Lower = IntMin.sdiv(*C); | |||
| 8524 | Upper = IntMax.sdiv(*C); | |||
| 8525 | if (Lower.sgt(Upper)) | |||
| 8526 | std::swap(Lower, Upper); | |||
| 8527 | Upper = Upper + 1; | |||
| 8528 | assert(Upper != Lower && "Upper part of range has wrapped!")(static_cast <bool> (Upper != Lower && "Upper part of range has wrapped!" ) ? void (0) : __assert_fail ("Upper != Lower && \"Upper part of range has wrapped!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 8528, __extension__ __PRETTY_FUNCTION__ )); | |||
| 8529 | } | |||
| 8530 | } else if (match(BO.getOperand(0), m_APInt(C))) { | |||
| 8531 | if (C->isMinSignedValue()) { | |||
| 8532 | // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2]. | |||
| 8533 | Lower = *C; | |||
| 8534 | Upper = Lower.lshr(1) + 1; | |||
| 8535 | } else { | |||
| 8536 | // 'sdiv C, x' produces [-|C|, |C|]. | |||
| 8537 | Upper = C->abs() + 1; | |||
| 8538 | Lower = (-Upper) + 1; | |||
| 8539 | } | |||
| 8540 | } | |||
| 8541 | break; | |||
| 8542 | ||||
| 8543 | case Instruction::UDiv: | |||
| 8544 | if (match(BO.getOperand(1), m_APInt(C)) && !C->isZero()) { | |||
| 8545 | // 'udiv x, C' produces [0, UINT_MAX / C]. | |||
| 8546 | Upper = APInt::getMaxValue(Width).udiv(*C) + 1; | |||
| 8547 | } else if (match(BO.getOperand(0), m_APInt(C))) { | |||
| 8548 | // 'udiv C, x' produces [0, C]. | |||
| 8549 | Upper = *C + 1; | |||
| 8550 | } | |||
| 8551 | break; | |||
| 8552 | ||||
| 8553 | case Instruction::SRem: | |||
| 8554 | if (match(BO.getOperand(1), m_APInt(C))) { | |||
| 8555 | // 'srem x, C' produces (-|C|, |C|). | |||
| 8556 | Upper = C->abs(); | |||
| 8557 | Lower = (-Upper) + 1; | |||
| 8558 | } | |||
| 8559 | break; | |||
| 8560 | ||||
| 8561 | case Instruction::URem: | |||
| 8562 | if (match(BO.getOperand(1), m_APInt(C))) | |||
| 8563 | // 'urem x, C' produces [0, C). | |||
| 8564 | Upper = *C; | |||
| 8565 | break; | |||
| 8566 | ||||
| 8567 | default: | |||
| 8568 | break; | |||
| 8569 | } | |||
| 8570 | } | |||
| 8571 | ||||
| 8572 | static ConstantRange getRangeForIntrinsic(const IntrinsicInst &II) { | |||
| 8573 | unsigned Width = II.getType()->getScalarSizeInBits(); | |||
| 8574 | const APInt *C; | |||
| 8575 | switch (II.getIntrinsicID()) { | |||
| 8576 | case Intrinsic::ctpop: | |||
| 8577 | case Intrinsic::ctlz: | |||
| 8578 | case Intrinsic::cttz: | |||
| 8579 | // Maximum of set/clear bits is the bit width. | |||
| 8580 | return ConstantRange(APInt::getZero(Width), APInt(Width, Width + 1)); | |||
| 8581 | case Intrinsic::uadd_sat: | |||
| 8582 | // uadd.sat(x, C) produces [C, UINT_MAX]. | |||
| 8583 | if (match(II.getOperand(0), m_APInt(C)) || | |||
| 8584 | match(II.getOperand(1), m_APInt(C))) | |||
| 8585 | return ConstantRange::getNonEmpty(*C, APInt::getZero(Width)); | |||
| 8586 | break; | |||
| 8587 | case Intrinsic::sadd_sat: | |||
| 8588 | if (match(II.getOperand(0), m_APInt(C)) || | |||
| 8589 | match(II.getOperand(1), m_APInt(C))) { | |||
| 8590 | if (C->isNegative()) | |||
| 8591 | // sadd.sat(x, -C) produces [SINT_MIN, SINT_MAX + (-C)]. | |||
| 8592 | return ConstantRange::getNonEmpty(APInt::getSignedMinValue(Width), | |||
| 8593 | APInt::getSignedMaxValue(Width) + *C + | |||
| 8594 | 1); | |||
| 8595 | ||||
| 8596 | // sadd.sat(x, +C) produces [SINT_MIN + C, SINT_MAX]. | |||
| 8597 | return ConstantRange::getNonEmpty(APInt::getSignedMinValue(Width) + *C, | |||
| 8598 | APInt::getSignedMaxValue(Width) + 1); | |||
| 8599 | } | |||
| 8600 | break; | |||
| 8601 | case Intrinsic::usub_sat: | |||
| 8602 | // usub.sat(C, x) produces [0, C]. | |||
| 8603 | if (match(II.getOperand(0), m_APInt(C))) | |||
| 8604 | return ConstantRange::getNonEmpty(APInt::getZero(Width), *C + 1); | |||
| 8605 | ||||
| 8606 | // usub.sat(x, C) produces [0, UINT_MAX - C]. | |||
| 8607 | if (match(II.getOperand(1), m_APInt(C))) | |||
| 8608 | return ConstantRange::getNonEmpty(APInt::getZero(Width), | |||
| 8609 | APInt::getMaxValue(Width) - *C + 1); | |||
| 8610 | break; | |||
| 8611 | case Intrinsic::ssub_sat: | |||
| 8612 | if (match(II.getOperand(0), m_APInt(C))) { | |||
| 8613 | if (C->isNegative()) | |||
| 8614 | // ssub.sat(-C, x) produces [SINT_MIN, -SINT_MIN + (-C)]. | |||
| 8615 | return ConstantRange::getNonEmpty(APInt::getSignedMinValue(Width), | |||
| 8616 | *C - APInt::getSignedMinValue(Width) + | |||
| 8617 | 1); | |||
| 8618 | ||||
| 8619 | // ssub.sat(+C, x) produces [-SINT_MAX + C, SINT_MAX]. | |||
| 8620 | return ConstantRange::getNonEmpty(*C - APInt::getSignedMaxValue(Width), | |||
| 8621 | APInt::getSignedMaxValue(Width) + 1); | |||
| 8622 | } else if (match(II.getOperand(1), m_APInt(C))) { | |||
| 8623 | if (C->isNegative()) | |||
| 8624 | // ssub.sat(x, -C) produces [SINT_MIN - (-C), SINT_MAX]: | |||
| 8625 | return ConstantRange::getNonEmpty(APInt::getSignedMinValue(Width) - *C, | |||
| 8626 | APInt::getSignedMaxValue(Width) + 1); | |||
| 8627 | ||||
| 8628 | // ssub.sat(x, +C) produces [SINT_MIN, SINT_MAX - C]. | |||
| 8629 | return ConstantRange::getNonEmpty(APInt::getSignedMinValue(Width), | |||
| 8630 | APInt::getSignedMaxValue(Width) - *C + | |||
| 8631 | 1); | |||
| 8632 | } | |||
| 8633 | break; | |||
| 8634 | case Intrinsic::umin: | |||
| 8635 | case Intrinsic::umax: | |||
| 8636 | case Intrinsic::smin: | |||
| 8637 | case Intrinsic::smax: | |||
| 8638 | if (!match(II.getOperand(0), m_APInt(C)) && | |||
| 8639 | !match(II.getOperand(1), m_APInt(C))) | |||
| 8640 | break; | |||
| 8641 | ||||
| 8642 | switch (II.getIntrinsicID()) { | |||
| 8643 | case Intrinsic::umin: | |||
| 8644 | return ConstantRange::getNonEmpty(APInt::getZero(Width), *C + 1); | |||
| 8645 | case Intrinsic::umax: | |||
| 8646 | return ConstantRange::getNonEmpty(*C, APInt::getZero(Width)); | |||
| 8647 | case Intrinsic::smin: | |||
| 8648 | return ConstantRange::getNonEmpty(APInt::getSignedMinValue(Width), | |||
| 8649 | *C + 1); | |||
| 8650 | case Intrinsic::smax: | |||
| 8651 | return ConstantRange::getNonEmpty(*C, | |||
| 8652 | APInt::getSignedMaxValue(Width) + 1); | |||
| 8653 | default: | |||
| 8654 | llvm_unreachable("Must be min/max intrinsic")::llvm::llvm_unreachable_internal("Must be min/max intrinsic" , "llvm/lib/Analysis/ValueTracking.cpp", 8654); | |||
| 8655 | } | |||
| 8656 | break; | |||
| 8657 | case Intrinsic::abs: | |||
| 8658 | // If abs of SIGNED_MIN is poison, then the result is [0..SIGNED_MAX], | |||
| 8659 | // otherwise it is [0..SIGNED_MIN], as -SIGNED_MIN == SIGNED_MIN. | |||
| 8660 | if (match(II.getOperand(1), m_One())) | |||
| 8661 | return ConstantRange(APInt::getZero(Width), | |||
| 8662 | APInt::getSignedMaxValue(Width) + 1); | |||
| 8663 | ||||
| 8664 | return ConstantRange(APInt::getZero(Width), | |||
| 8665 | APInt::getSignedMinValue(Width) + 1); | |||
| 8666 | case Intrinsic::vscale: | |||
| 8667 | if (!II.getParent() || !II.getFunction()) | |||
| 8668 | break; | |||
| 8669 | return getVScaleRange(II.getFunction(), Width); | |||
| 8670 | default: | |||
| 8671 | break; | |||
| 8672 | } | |||
| 8673 | ||||
| 8674 | return ConstantRange::getFull(Width); | |||
| 8675 | } | |||
| 8676 | ||||
| 8677 | static void setLimitsForSelectPattern(const SelectInst &SI, APInt &Lower, | |||
| 8678 | APInt &Upper, const InstrInfoQuery &IIQ) { | |||
| 8679 | const Value *LHS = nullptr, *RHS = nullptr; | |||
| 8680 | SelectPatternResult R = matchSelectPattern(&SI, LHS, RHS); | |||
| 8681 | if (R.Flavor == SPF_UNKNOWN) | |||
| 8682 | return; | |||
| 8683 | ||||
| 8684 | unsigned BitWidth = SI.getType()->getScalarSizeInBits(); | |||
| 8685 | ||||
| 8686 | if (R.Flavor == SelectPatternFlavor::SPF_ABS) { | |||
| 8687 | // If the negation part of the abs (in RHS) has the NSW flag, | |||
| 8688 | // then the result of abs(X) is [0..SIGNED_MAX], | |||
| 8689 | // otherwise it is [0..SIGNED_MIN], as -SIGNED_MIN == SIGNED_MIN. | |||
| 8690 | Lower = APInt::getZero(BitWidth); | |||
| 8691 | if (match(RHS, m_Neg(m_Specific(LHS))) && | |||
| 8692 | IIQ.hasNoSignedWrap(cast<Instruction>(RHS))) | |||
| 8693 | Upper = APInt::getSignedMaxValue(BitWidth) + 1; | |||
| 8694 | else | |||
| 8695 | Upper = APInt::getSignedMinValue(BitWidth) + 1; | |||
| 8696 | return; | |||
| 8697 | } | |||
| 8698 | ||||
| 8699 | if (R.Flavor == SelectPatternFlavor::SPF_NABS) { | |||
| 8700 | // The result of -abs(X) is <= 0. | |||
| 8701 | Lower = APInt::getSignedMinValue(BitWidth); | |||
| 8702 | Upper = APInt(BitWidth, 1); | |||
| 8703 | return; | |||
| 8704 | } | |||
| 8705 | ||||
| 8706 | const APInt *C; | |||
| 8707 | if (!match(LHS, m_APInt(C)) && !match(RHS, m_APInt(C))) | |||
| 8708 | return; | |||
| 8709 | ||||
| 8710 | switch (R.Flavor) { | |||
| 8711 | case SPF_UMIN: | |||
| 8712 | Upper = *C + 1; | |||
| 8713 | break; | |||
| 8714 | case SPF_UMAX: | |||
| 8715 | Lower = *C; | |||
| 8716 | break; | |||
| 8717 | case SPF_SMIN: | |||
| 8718 | Lower = APInt::getSignedMinValue(BitWidth); | |||
| 8719 | Upper = *C + 1; | |||
| 8720 | break; | |||
| 8721 | case SPF_SMAX: | |||
| 8722 | Lower = *C; | |||
| 8723 | Upper = APInt::getSignedMaxValue(BitWidth) + 1; | |||
| 8724 | break; | |||
| 8725 | default: | |||
| 8726 | break; | |||
| 8727 | } | |||
| 8728 | } | |||
| 8729 | ||||
| 8730 | static void setLimitForFPToI(const Instruction *I, APInt &Lower, APInt &Upper) { | |||
| 8731 | // The maximum representable value of a half is 65504. For floats the maximum | |||
| 8732 | // value is 3.4e38 which requires roughly 129 bits. | |||
| 8733 | unsigned BitWidth = I->getType()->getScalarSizeInBits(); | |||
| 8734 | if (!I->getOperand(0)->getType()->getScalarType()->isHalfTy()) | |||
| 8735 | return; | |||
| 8736 | if (isa<FPToSIInst>(I) && BitWidth >= 17) { | |||
| 8737 | Lower = APInt(BitWidth, -65504); | |||
| 8738 | Upper = APInt(BitWidth, 65505); | |||
| 8739 | } | |||
| 8740 | ||||
| 8741 | if (isa<FPToUIInst>(I) && BitWidth >= 16) { | |||
| 8742 | // For a fptoui the lower limit is left as 0. | |||
| 8743 | Upper = APInt(BitWidth, 65505); | |||
| 8744 | } | |||
| 8745 | } | |||
| 8746 | ||||
| 8747 | ConstantRange llvm::computeConstantRange(const Value *V, bool ForSigned, | |||
| 8748 | bool UseInstrInfo, AssumptionCache *AC, | |||
| 8749 | const Instruction *CtxI, | |||
| 8750 | const DominatorTree *DT, | |||
| 8751 | unsigned Depth) { | |||
| 8752 | assert(V->getType()->isIntOrIntVectorTy() && "Expected integer instruction")(static_cast <bool> (V->getType()->isIntOrIntVectorTy () && "Expected integer instruction") ? void (0) : __assert_fail ("V->getType()->isIntOrIntVectorTy() && \"Expected integer instruction\"" , "llvm/lib/Analysis/ValueTracking.cpp", 8752, __extension__ __PRETTY_FUNCTION__ )); | |||
| 8753 | ||||
| 8754 | if (Depth == MaxAnalysisRecursionDepth) | |||
| 8755 | return ConstantRange::getFull(V->getType()->getScalarSizeInBits()); | |||
| 8756 | ||||
| 8757 | const APInt *C; | |||
| 8758 | if (match(V, m_APInt(C))) | |||
| 8759 | return ConstantRange(*C); | |||
| 8760 | ||||
| 8761 | InstrInfoQuery IIQ(UseInstrInfo); | |||
| 8762 | unsigned BitWidth = V->getType()->getScalarSizeInBits(); | |||
| 8763 | ConstantRange CR = ConstantRange::getFull(BitWidth); | |||
| 8764 | if (auto *BO = dyn_cast<BinaryOperator>(V)) { | |||
| 8765 | APInt Lower = APInt(BitWidth, 0); | |||
| 8766 | APInt Upper = APInt(BitWidth, 0); | |||
| 8767 | // TODO: Return ConstantRange. | |||
| 8768 | setLimitsForBinOp(*BO, Lower, Upper, IIQ, ForSigned); | |||
| 8769 | CR = ConstantRange::getNonEmpty(Lower, Upper); | |||
| 8770 | } else if (auto *II = dyn_cast<IntrinsicInst>(V)) | |||
| 8771 | CR = getRangeForIntrinsic(*II); | |||
| 8772 | else if (auto *SI = dyn_cast<SelectInst>(V)) { | |||
| 8773 | APInt Lower = APInt(BitWidth, 0); | |||
| 8774 | APInt Upper = APInt(BitWidth, 0); | |||
| 8775 | // TODO: Return ConstantRange. | |||
| 8776 | setLimitsForSelectPattern(*SI, Lower, Upper, IIQ); | |||
| 8777 | CR = ConstantRange::getNonEmpty(Lower, Upper); | |||
| 8778 | } else if (isa<FPToUIInst>(V) || isa<FPToSIInst>(V)) { | |||
| 8779 | APInt Lower = APInt(BitWidth, 0); | |||
| 8780 | APInt Upper = APInt(BitWidth, 0); | |||
| 8781 | // TODO: Return ConstantRange. | |||
| 8782 | setLimitForFPToI(cast<Instruction>(V), Lower, Upper); | |||
| 8783 | CR = ConstantRange::getNonEmpty(Lower, Upper); | |||
| 8784 | } | |||
| 8785 | ||||
| 8786 | if (auto *I = dyn_cast<Instruction>(V)) | |||
| 8787 | if (auto *Range = IIQ.getMetadata(I, LLVMContext::MD_range)) | |||
| 8788 | CR = CR.intersectWith(getConstantRangeFromMetadata(*Range)); | |||
| 8789 | ||||
| 8790 | if (CtxI && AC) { | |||
| 8791 | // Try to restrict the range based on information from assumptions. | |||
| 8792 | for (auto &AssumeVH : AC->assumptionsFor(V)) { | |||
| 8793 | if (!AssumeVH) | |||
| 8794 | continue; | |||
| 8795 | CallInst *I = cast<CallInst>(AssumeVH); | |||
| 8796 | assert(I->getParent()->getParent() == CtxI->getParent()->getParent() &&(static_cast <bool> (I->getParent()->getParent() == CtxI->getParent()->getParent() && "Got assumption for the wrong function!" ) ? void (0) : __assert_fail ("I->getParent()->getParent() == CtxI->getParent()->getParent() && \"Got assumption for the wrong function!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 8797, __extension__ __PRETTY_FUNCTION__ )) | |||
| 8797 | "Got assumption for the wrong function!")(static_cast <bool> (I->getParent()->getParent() == CtxI->getParent()->getParent() && "Got assumption for the wrong function!" ) ? void (0) : __assert_fail ("I->getParent()->getParent() == CtxI->getParent()->getParent() && \"Got assumption for the wrong function!\"" , "llvm/lib/Analysis/ValueTracking.cpp", 8797, __extension__ __PRETTY_FUNCTION__ )); | |||
| 8798 | assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&(static_cast <bool> (I->getCalledFunction()->getIntrinsicID () == Intrinsic::assume && "must be an assume intrinsic" ) ? void (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\"" , "llvm/lib/Analysis/ValueTracking.cpp", 8799, __extension__ __PRETTY_FUNCTION__ )) | |||
| 8799 | "must be an assume intrinsic")(static_cast <bool> (I->getCalledFunction()->getIntrinsicID () == Intrinsic::assume && "must be an assume intrinsic" ) ? void (0) : __assert_fail ("I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume && \"must be an assume intrinsic\"" , "llvm/lib/Analysis/ValueTracking.cpp", 8799, __extension__ __PRETTY_FUNCTION__ )); | |||
| 8800 | ||||
| 8801 | if (!isValidAssumeForContext(I, CtxI, DT)) | |||
| 8802 | continue; | |||
| 8803 | Value *Arg = I->getArgOperand(0); | |||
| 8804 | ICmpInst *Cmp = dyn_cast<ICmpInst>(Arg); | |||
| 8805 | // Currently we just use information from comparisons. | |||
| 8806 | if (!Cmp || Cmp->getOperand(0) != V) | |||
| 8807 | continue; | |||
| 8808 | // TODO: Set "ForSigned" parameter via Cmp->isSigned()? | |||
| 8809 | ConstantRange RHS = | |||
| 8810 | computeConstantRange(Cmp->getOperand(1), /* ForSigned */ false, | |||
| 8811 | UseInstrInfo, AC, I, DT, Depth + 1); | |||
| 8812 | CR = CR.intersectWith( | |||
| 8813 | ConstantRange::makeAllowedICmpRegion(Cmp->getPredicate(), RHS)); | |||
| 8814 | } | |||
| 8815 | } | |||
| 8816 | ||||
| 8817 | return CR; | |||
| 8818 | } |