| 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/APInt.h" |
| 16 | #include "llvm/ADT/ArrayRef.h" |
| 17 | #include "llvm/ADT/FloatingPointMode.h" |
| 18 | #include "llvm/ADT/STLExtras.h" |
| 19 | #include "llvm/ADT/ScopeExit.h" |
| 20 | #include "llvm/ADT/SmallPtrSet.h" |
| 21 | #include "llvm/ADT/SmallVector.h" |
| 22 | #include "llvm/ADT/StringRef.h" |
| 23 | #include "llvm/ADT/iterator_range.h" |
| 24 | #include "llvm/Analysis/AliasAnalysis.h" |
| 25 | #include "llvm/Analysis/AssumeBundleQueries.h" |
| 26 | #include "llvm/Analysis/AssumptionCache.h" |
| 27 | #include "llvm/Analysis/ConstantFolding.h" |
| 28 | #include "llvm/Analysis/DomConditionCache.h" |
| 29 | #include "llvm/Analysis/FloatingPointPredicateUtils.h" |
| 30 | #include "llvm/Analysis/GuardUtils.h" |
| 31 | #include "llvm/Analysis/InstructionSimplify.h" |
| 32 | #include "llvm/Analysis/Loads.h" |
| 33 | #include "llvm/Analysis/LoopInfo.h" |
| 34 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 35 | #include "llvm/Analysis/VectorUtils.h" |
| 36 | #include "llvm/Analysis/WithCache.h" |
| 37 | #include "llvm/IR/Argument.h" |
| 38 | #include "llvm/IR/Attributes.h" |
| 39 | #include "llvm/IR/BasicBlock.h" |
| 40 | #include "llvm/IR/BundleAttributes.h" |
| 41 | #include "llvm/IR/Constant.h" |
| 42 | #include "llvm/IR/ConstantFPRange.h" |
| 43 | #include "llvm/IR/ConstantRange.h" |
| 44 | #include "llvm/IR/Constants.h" |
| 45 | #include "llvm/IR/DerivedTypes.h" |
| 46 | #include "llvm/IR/DiagnosticInfo.h" |
| 47 | #include "llvm/IR/Dominators.h" |
| 48 | #include "llvm/IR/EHPersonalities.h" |
| 49 | #include "llvm/IR/Function.h" |
| 50 | #include "llvm/IR/GetElementPtrTypeIterator.h" |
| 51 | #include "llvm/IR/GlobalAlias.h" |
| 52 | #include "llvm/IR/GlobalValue.h" |
| 53 | #include "llvm/IR/GlobalVariable.h" |
| 54 | #include "llvm/IR/InstrTypes.h" |
| 55 | #include "llvm/IR/Instruction.h" |
| 56 | #include "llvm/IR/Instructions.h" |
| 57 | #include "llvm/IR/IntrinsicInst.h" |
| 58 | #include "llvm/IR/Intrinsics.h" |
| 59 | #include "llvm/IR/IntrinsicsAArch64.h" |
| 60 | #include "llvm/IR/IntrinsicsAMDGPU.h" |
| 61 | #include "llvm/IR/IntrinsicsRISCV.h" |
| 62 | #include "llvm/IR/IntrinsicsX86.h" |
| 63 | #include "llvm/IR/LLVMContext.h" |
| 64 | #include "llvm/IR/Metadata.h" |
| 65 | #include "llvm/IR/Module.h" |
| 66 | #include "llvm/IR/Operator.h" |
| 67 | #include "llvm/IR/PatternMatch.h" |
| 68 | #include "llvm/IR/Type.h" |
| 69 | #include "llvm/IR/User.h" |
| 70 | #include "llvm/IR/Value.h" |
| 71 | #include "llvm/Support/Casting.h" |
| 72 | #include "llvm/Support/CommandLine.h" |
| 73 | #include "llvm/Support/Compiler.h" |
| 74 | #include "llvm/Support/ErrorHandling.h" |
| 75 | #include "llvm/Support/KnownBits.h" |
| 76 | #include "llvm/Support/KnownFPClass.h" |
| 77 | #include "llvm/Support/MathExtras.h" |
| 78 | #include "llvm/Support/UndefPoison.h" |
| 79 | #include "llvm/TargetParser/RISCVTargetParser.h" |
| 80 | #include <algorithm> |
| 81 | #include <cassert> |
| 82 | #include <cstdint> |
| 83 | #include <optional> |
| 84 | #include <utility> |
| 85 | |
| 86 | using namespace llvm; |
| 87 | using namespace llvm::PatternMatch; |
| 88 | |
| 89 | // Controls the number of uses of the value searched for possible |
| 90 | // dominating comparisons. |
| 91 | static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses" , |
| 92 | cl::Hidden, cl::init(Val: 20)); |
| 93 | |
| 94 | /// Maximum number of instructions to check between assume and context |
| 95 | /// instruction. |
| 96 | static constexpr unsigned MaxInstrsToCheckForFree = 32; |
| 97 | |
| 98 | template <typename InstTy> |
| 99 | static bool matchTwoInputRecurrence(const PHINode *PN, InstTy *&Inst, |
| 100 | Value *&Init, Value *&OtherOp); |
| 101 | |
| 102 | /// Returns the bitwidth of the given scalar or pointer type. For vector types, |
| 103 | /// returns the element type's bitwidth. |
| 104 | static unsigned getBitWidth(Type *Ty, const DataLayout &DL) { |
| 105 | if (unsigned BitWidth = Ty->getScalarSizeInBits()) |
| 106 | return BitWidth; |
| 107 | |
| 108 | return DL.getPointerTypeSizeInBits(Ty); |
| 109 | } |
| 110 | |
| 111 | // Given the provided Value and, potentially, a context instruction, return |
| 112 | // the preferred context instruction (if any). |
| 113 | static const Instruction *safeCtxI(const Value *V, const Instruction *CtxI) { |
| 114 | // If we've been provided with a context instruction, then use that (provided |
| 115 | // it has been inserted). |
| 116 | if (CtxI && CtxI->getParent()) |
| 117 | return CtxI; |
| 118 | |
| 119 | // If the value is really an already-inserted instruction, then use that. |
| 120 | CtxI = dyn_cast<Instruction>(Val: V); |
| 121 | if (CtxI && CtxI->getParent()) |
| 122 | return CtxI; |
| 123 | |
| 124 | return nullptr; |
| 125 | } |
| 126 | |
| 127 | static bool getShuffleDemandedElts(const ShuffleVectorInst *Shuf, |
| 128 | const APInt &DemandedElts, |
| 129 | APInt &DemandedLHS, APInt &DemandedRHS) { |
| 130 | if (isa<ScalableVectorType>(Val: Shuf->getType())) { |
| 131 | assert(DemandedElts == APInt(1,1)); |
| 132 | DemandedLHS = DemandedRHS = DemandedElts; |
| 133 | return true; |
| 134 | } |
| 135 | |
| 136 | int NumElts = |
| 137 | cast<FixedVectorType>(Val: Shuf->getOperand(i_nocapture: 0)->getType())->getNumElements(); |
| 138 | return llvm::getShuffleDemandedElts(SrcWidth: NumElts, Mask: Shuf->getShuffleMask(), |
| 139 | DemandedElts, DemandedLHS, DemandedRHS); |
| 140 | } |
| 141 | |
| 142 | static void computeKnownBits(const Value *V, const APInt &DemandedElts, |
| 143 | KnownBits &Known, const SimplifyQuery &Q, |
| 144 | unsigned Depth); |
| 145 | |
| 146 | void llvm::computeKnownBits(const Value *V, KnownBits &Known, |
| 147 | const SimplifyQuery &Q, unsigned Depth) { |
| 148 | // Since the number of lanes in a scalable vector is unknown at compile time, |
| 149 | // we track one bit which is implicitly broadcast to all lanes. This means |
| 150 | // that all lanes in a scalable vector are considered demanded. |
| 151 | auto *FVTy = dyn_cast<FixedVectorType>(Val: V->getType()); |
| 152 | APInt DemandedElts = |
| 153 | FVTy ? APInt::getAllOnes(numBits: FVTy->getNumElements()) : APInt(1, 1); |
| 154 | ::computeKnownBits(V, DemandedElts, Known, Q, Depth); |
| 155 | } |
| 156 | |
| 157 | void llvm::computeKnownBits(const Value *V, KnownBits &Known, |
| 158 | const DataLayout &DL, AssumptionCache *AC, |
| 159 | const Instruction *CtxI, const DominatorTree *DT, |
| 160 | bool UseInstrInfo, unsigned Depth) { |
| 161 | computeKnownBits(V, Known, |
| 162 | Q: SimplifyQuery(DL, DT, AC, safeCtxI(V, CtxI), UseInstrInfo), |
| 163 | Depth); |
| 164 | } |
| 165 | |
| 166 | KnownBits llvm::computeKnownBits(const Value *V, const DataLayout &DL, |
| 167 | AssumptionCache *AC, const Instruction *CtxI, |
| 168 | const DominatorTree *DT, bool UseInstrInfo, |
| 169 | unsigned Depth) { |
| 170 | return computeKnownBits( |
| 171 | V, Q: SimplifyQuery(DL, DT, AC, safeCtxI(V, CtxI), UseInstrInfo), Depth); |
| 172 | } |
| 173 | |
| 174 | static NoCommonBitsSetResult |
| 175 | haveNoCommonBitsSetSpecialCases(const Value *LHS, const Value *RHS, |
| 176 | const SimplifyQuery &SQ) { |
| 177 | // Look for an inverted mask: (X & ~M) op (Y & M). |
| 178 | { |
| 179 | Value *M; |
| 180 | if (match(V: LHS, P: m_c_And(L: m_Not(V: m_Value(V&: M)), R: m_Value())) && |
| 181 | match(V: RHS, P: m_c_And(L: m_Specific(V: M), R: m_Value()))) |
| 182 | return isGuaranteedNotToBeUndef(V: M, AC: SQ.AC, CtxI: SQ.CtxI, DT: SQ.DT) |
| 183 | ? NoCommonBitsSetResult::Known |
| 184 | : NoCommonBitsSetResult::OnlyIfUndefIgnored; |
| 185 | } |
| 186 | |
| 187 | // X op (Y & ~X) |
| 188 | if (match(V: RHS, P: m_c_And(L: m_Not(V: m_Specific(V: LHS)), R: m_Value()))) |
| 189 | return isGuaranteedNotToBeUndef(V: LHS, AC: SQ.AC, CtxI: SQ.CtxI, DT: SQ.DT) |
| 190 | ? NoCommonBitsSetResult::Known |
| 191 | : NoCommonBitsSetResult::OnlyIfUndefIgnored; |
| 192 | |
| 193 | // X op ((X & Y) ^ Y) -- this is the canonical form of the previous pattern |
| 194 | // for constant Y. |
| 195 | Value *Y; |
| 196 | if (match(V: RHS, |
| 197 | P: m_c_Xor(L: m_c_And(L: m_Specific(V: LHS), R: m_Value(V&: Y)), R: m_Deferred(V: Y)))) { |
| 198 | bool IsNoUndef = isGuaranteedNotToBeUndef(V: LHS, AC: SQ.AC, CtxI: SQ.CtxI, DT: SQ.DT) && |
| 199 | isGuaranteedNotToBeUndef(V: Y, AC: SQ.AC, CtxI: SQ.CtxI, DT: SQ.DT); |
| 200 | return IsNoUndef ? NoCommonBitsSetResult::Known |
| 201 | : NoCommonBitsSetResult::OnlyIfUndefIgnored; |
| 202 | } |
| 203 | |
| 204 | // Peek through extends to find a 'not' of the other side: |
| 205 | // (ext Y) op ext(~Y) |
| 206 | if (match(V: LHS, P: m_ZExtOrSExt(Op: m_Value(V&: Y))) && |
| 207 | match(V: RHS, P: m_ZExtOrSExt(Op: m_Not(V: m_Specific(V: Y))))) |
| 208 | return isGuaranteedNotToBeUndef(V: Y, AC: SQ.AC, CtxI: SQ.CtxI, DT: SQ.DT) |
| 209 | ? NoCommonBitsSetResult::Known |
| 210 | : NoCommonBitsSetResult::OnlyIfUndefIgnored; |
| 211 | |
| 212 | // Look for: (A & B) op ~(A | B) |
| 213 | { |
| 214 | Value *A, *B; |
| 215 | if (match(V: LHS, P: m_And(L: m_Value(V&: A), R: m_Value(V&: B))) && |
| 216 | match(V: RHS, P: m_Not(V: m_c_Or(L: m_Specific(V: A), R: m_Specific(V: B))))) { |
| 217 | bool IsNoUndef = isGuaranteedNotToBeUndef(V: A, AC: SQ.AC, CtxI: SQ.CtxI, DT: SQ.DT) && |
| 218 | isGuaranteedNotToBeUndef(V: B, AC: SQ.AC, CtxI: SQ.CtxI, DT: SQ.DT); |
| 219 | return IsNoUndef ? NoCommonBitsSetResult::Known |
| 220 | : NoCommonBitsSetResult::OnlyIfUndefIgnored; |
| 221 | } |
| 222 | } |
| 223 | |
| 224 | // Look for: (X << V) op (Y >> (BitWidth - V)) |
| 225 | // or (X >> V) op (Y << (BitWidth - V)) |
| 226 | { |
| 227 | const Value *V; |
| 228 | const APInt *R; |
| 229 | if (((match(V: RHS, P: m_Shl(L: m_Value(), R: m_Sub(L: m_APInt(Res&: R), R: m_Value(V)))) && |
| 230 | match(V: LHS, P: m_LShr(L: m_Value(), R: m_Specific(V)))) || |
| 231 | (match(V: RHS, P: m_LShr(L: m_Value(), R: m_Sub(L: m_APInt(Res&: R), R: m_Value(V)))) && |
| 232 | match(V: LHS, P: m_Shl(L: m_Value(), R: m_Specific(V))))) && |
| 233 | R->uge(RHS: LHS->getType()->getScalarSizeInBits())) |
| 234 | return NoCommonBitsSetResult::Known; |
| 235 | } |
| 236 | |
| 237 | return NoCommonBitsSetResult::Unknown; |
| 238 | } |
| 239 | |
| 240 | NoCommonBitsSetResult |
| 241 | llvm::getNoCommonBitsSetResult(const WithCache<const Value *> &LHSCache, |
| 242 | const WithCache<const Value *> &RHSCache, |
| 243 | const SimplifyQuery &SQ) { |
| 244 | const Value *LHS = LHSCache.getValue(); |
| 245 | const Value *RHS = RHSCache.getValue(); |
| 246 | |
| 247 | assert(LHS->getType() == RHS->getType() && |
| 248 | "LHS and RHS should have the same type" ); |
| 249 | assert(LHS->getType()->isIntOrIntVectorTy() && |
| 250 | "LHS and RHS should be integers" ); |
| 251 | |
| 252 | NoCommonBitsSetResult Result = haveNoCommonBitsSetSpecialCases(LHS, RHS, SQ); |
| 253 | if (Result == NoCommonBitsSetResult::Known) |
| 254 | return NoCommonBitsSetResult::Known; |
| 255 | |
| 256 | NoCommonBitsSetResult CommuteResult = |
| 257 | haveNoCommonBitsSetSpecialCases(LHS: RHS, RHS: LHS, SQ); |
| 258 | if (CommuteResult == NoCommonBitsSetResult::Known) |
| 259 | return NoCommonBitsSetResult::Known; |
| 260 | |
| 261 | if (KnownBits::haveNoCommonBitsSet(LHS: LHSCache.getKnownBits(Q: SQ), |
| 262 | RHS: RHSCache.getKnownBits(Q: SQ))) |
| 263 | return NoCommonBitsSetResult::Known; |
| 264 | |
| 265 | if (Result == NoCommonBitsSetResult::OnlyIfUndefIgnored || |
| 266 | CommuteResult == NoCommonBitsSetResult::OnlyIfUndefIgnored) |
| 267 | return NoCommonBitsSetResult::OnlyIfUndefIgnored; |
| 268 | |
| 269 | return NoCommonBitsSetResult::Unknown; |
| 270 | } |
| 271 | |
| 272 | bool llvm::haveNoCommonBitsSet(const WithCache<const Value *> &LHSCache, |
| 273 | const WithCache<const Value *> &RHSCache, |
| 274 | const SimplifyQuery &SQ) { |
| 275 | NoCommonBitsSetResult Result = |
| 276 | getNoCommonBitsSetResult(LHSCache, RHSCache, SQ); |
| 277 | return Result == NoCommonBitsSetResult::Known; |
| 278 | } |
| 279 | |
| 280 | bool llvm::isOnlyUsedInZeroComparison(const Instruction *I) { |
| 281 | return !I->user_empty() && |
| 282 | all_of(Range: I->users(), P: match_fn(P: m_ICmp(L: m_Value(), R: m_Zero()))); |
| 283 | } |
| 284 | |
| 285 | bool llvm::isOnlyUsedInZeroEqualityComparison(const Instruction *I) { |
| 286 | return !I->user_empty() && all_of(Range: I->users(), P: [](const User *U) { |
| 287 | CmpPredicate P; |
| 288 | return match(V: U, P: m_ICmp(Pred&: P, L: m_Value(), R: m_Zero())) && ICmpInst::isEquality(P); |
| 289 | }); |
| 290 | } |
| 291 | |
| 292 | bool llvm::isKnownToBeAPowerOfTwo(const Value *V, const DataLayout &DL, |
| 293 | bool OrZero, AssumptionCache *AC, |
| 294 | const Instruction *CtxI, |
| 295 | const DominatorTree *DT, bool UseInstrInfo, |
| 296 | unsigned Depth) { |
| 297 | return ::isKnownToBeAPowerOfTwo( |
| 298 | V, OrZero, Q: SimplifyQuery(DL, DT, AC, safeCtxI(V, CtxI), UseInstrInfo), |
| 299 | Depth); |
| 300 | } |
| 301 | |
| 302 | static bool isKnownNonZero(const Value *V, const APInt &DemandedElts, |
| 303 | const SimplifyQuery &Q, unsigned Depth); |
| 304 | |
| 305 | bool llvm::isKnownNonNegative(const Value *V, const SimplifyQuery &SQ, |
| 306 | unsigned Depth) { |
| 307 | return computeKnownBits(V, Q: SQ, Depth).isNonNegative(); |
| 308 | } |
| 309 | |
| 310 | bool llvm::isKnownPositive(const WithCache<const Value *> &V, |
| 311 | const SimplifyQuery &SQ, unsigned Depth) { |
| 312 | if (auto *CI = dyn_cast<ConstantInt>(Val: V.getValue())) |
| 313 | return CI->getValue().isStrictlyPositive(); |
| 314 | |
| 315 | // If `isKnownNonNegative` ever becomes more sophisticated, make sure to keep |
| 316 | // this updated. |
| 317 | KnownBits Known = Depth == 0 ? V.getKnownBits(Q: SQ) |
| 318 | : computeKnownBits(V: V.getValue(), Q: SQ, Depth); |
| 319 | return Known.isNonNegative() && |
| 320 | (Known.isNonZero() || isKnownNonZero(V: V.getValue(), Q: SQ, Depth)); |
| 321 | } |
| 322 | |
| 323 | bool llvm::isKnownNegative(const WithCache<const Value *> &V, |
| 324 | const SimplifyQuery &SQ, unsigned Depth) { |
| 325 | if (Depth == 0) |
| 326 | return V.getKnownBits(Q: SQ).isNegative(); |
| 327 | return computeKnownBits(V: V.getValue(), Q: SQ, Depth).isNegative(); |
| 328 | } |
| 329 | |
| 330 | static bool isKnownNonEqual(const Value *V1, const Value *V2, |
| 331 | const APInt &DemandedElts, const SimplifyQuery &Q, |
| 332 | unsigned Depth); |
| 333 | |
| 334 | static bool isTruePredicate(CmpInst::Predicate Pred, const Value *LHS, |
| 335 | const Value *RHS); |
| 336 | |
| 337 | bool llvm::isKnownNonEqual(const Value *V1, const Value *V2, |
| 338 | const SimplifyQuery &Q, unsigned Depth) { |
| 339 | // We don't support looking through casts. |
| 340 | if (V1 == V2 || V1->getType() != V2->getType()) |
| 341 | return false; |
| 342 | auto *FVTy = dyn_cast<FixedVectorType>(Val: V1->getType()); |
| 343 | APInt DemandedElts = |
| 344 | FVTy ? APInt::getAllOnes(numBits: FVTy->getNumElements()) : APInt(1, 1); |
| 345 | return ::isKnownNonEqual(V1, V2, DemandedElts, Q, Depth); |
| 346 | } |
| 347 | |
| 348 | bool llvm::MaskedValueIsZero(const Value *V, const APInt &Mask, |
| 349 | const SimplifyQuery &SQ, unsigned Depth) { |
| 350 | KnownBits Known(Mask.getBitWidth()); |
| 351 | computeKnownBits(V, Known, Q: SQ, Depth); |
| 352 | return Mask.isSubsetOf(RHS: Known.Zero); |
| 353 | } |
| 354 | |
| 355 | static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts, |
| 356 | const SimplifyQuery &Q, unsigned Depth); |
| 357 | |
| 358 | static unsigned ComputeNumSignBits(const Value *V, const SimplifyQuery &Q, |
| 359 | unsigned Depth = 0) { |
| 360 | auto *FVTy = dyn_cast<FixedVectorType>(Val: V->getType()); |
| 361 | APInt DemandedElts = |
| 362 | FVTy ? APInt::getAllOnes(numBits: FVTy->getNumElements()) : APInt(1, 1); |
| 363 | return ComputeNumSignBits(V, DemandedElts, Q, Depth); |
| 364 | } |
| 365 | |
| 366 | unsigned llvm::ComputeNumSignBits(const Value *V, const DataLayout &DL, |
| 367 | AssumptionCache *AC, const Instruction *CtxI, |
| 368 | const DominatorTree *DT, bool UseInstrInfo, |
| 369 | unsigned Depth) { |
| 370 | return ::ComputeNumSignBits( |
| 371 | V, Q: SimplifyQuery(DL, DT, AC, safeCtxI(V, CtxI), UseInstrInfo), Depth); |
| 372 | } |
| 373 | |
| 374 | unsigned llvm::ComputeMaxSignificantBits(const Value *V, const DataLayout &DL, |
| 375 | AssumptionCache *AC, |
| 376 | const Instruction *CtxI, |
| 377 | const DominatorTree *DT, |
| 378 | unsigned Depth) { |
| 379 | unsigned SignBits = ComputeNumSignBits(V, DL, AC, CtxI, DT, UseInstrInfo: Depth); |
| 380 | return V->getType()->getScalarSizeInBits() - SignBits + 1; |
| 381 | } |
| 382 | |
| 383 | /// Try to detect the lerp pattern: a * (b - c) + c * d |
| 384 | /// where a >= 0, b >= 0, c >= 0, d >= 0, and b >= c. |
| 385 | /// |
| 386 | /// In that particular case, we can use the following chain of reasoning: |
| 387 | /// |
| 388 | /// a * (b - c) + c * d <= a' * (b - c) + a' * c = a' * b where a' = max(a, d) |
| 389 | /// |
| 390 | /// Since that is true for arbitrary a, b, c and d within our constraints, we |
| 391 | /// can conclude that: |
| 392 | /// |
| 393 | /// max(a * (b - c) + c * d) <= max(max(a), max(d)) * max(b) = U |
| 394 | /// |
| 395 | /// Considering that any result of the lerp would be less or equal to U, it |
| 396 | /// would have at least the number of leading 0s as in U. |
| 397 | /// |
| 398 | /// While being quite a specific situation, it is fairly common in computer |
| 399 | /// graphics in the shape of alpha blending. |
| 400 | /// |
| 401 | /// Modifies given KnownOut in-place with the inferred information. |
| 402 | static void computeKnownBitsFromLerpPattern(const Value *Op0, const Value *Op1, |
| 403 | const APInt &DemandedElts, |
| 404 | KnownBits &KnownOut, |
| 405 | const SimplifyQuery &Q, |
| 406 | unsigned Depth) { |
| 407 | |
| 408 | Type *Ty = Op0->getType(); |
| 409 | const unsigned BitWidth = Ty->getScalarSizeInBits(); |
| 410 | |
| 411 | // Only handle scalar types for now |
| 412 | if (Ty->isVectorTy()) |
| 413 | return; |
| 414 | |
| 415 | // Try to match: a * (b - c) + c * d. |
| 416 | // When a == 1 => A == nullptr, the same applies to d/D as well. |
| 417 | const Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr; |
| 418 | const Instruction *SubBC = nullptr; |
| 419 | |
| 420 | const auto MatchSubBC = [&]() { |
| 421 | // (b - c) can have two forms that interest us: |
| 422 | // |
| 423 | // 1. sub nuw %b, %c |
| 424 | // 2. xor %c, %b |
| 425 | // |
| 426 | // For the first case, nuw flag guarantees our requirement b >= c. |
| 427 | // |
| 428 | // The second case might happen when the analysis can infer that b is a mask |
| 429 | // for c and we can transform sub operation into xor (that is usually true |
| 430 | // for constant b's). Even though xor is symmetrical, canonicalization |
| 431 | // ensures that the constant will be the RHS. We have additional checks |
| 432 | // later on to ensure that this xor operation is equivalent to subtraction. |
| 433 | return m_Instruction(I&: SubBC, P: m_CombineOr(Ps: m_NUWSub(L: m_Value(V&: B), R: m_Value(V&: C)), |
| 434 | Ps: m_Xor(L: m_Value(V&: C), R: m_Value(V&: B)))); |
| 435 | }; |
| 436 | |
| 437 | const auto MatchASubBC = [&]() { |
| 438 | // Cases: |
| 439 | // - a * (b - c) |
| 440 | // - (b - c) * a |
| 441 | // - (b - c) <- a implicitly equals 1 |
| 442 | return m_CombineOr(Ps: m_c_Mul(L: m_Value(V&: A), R: MatchSubBC()), Ps: MatchSubBC()); |
| 443 | }; |
| 444 | |
| 445 | const auto MatchCD = [&]() { |
| 446 | // Cases: |
| 447 | // - d * c |
| 448 | // - c * d |
| 449 | // - c <- d implicitly equals 1 |
| 450 | return m_CombineOr(Ps: m_c_Mul(L: m_Value(V&: D), R: m_Specific(V: C)), Ps: m_Specific(V: C)); |
| 451 | }; |
| 452 | |
| 453 | const auto Match = [&](const Value *LHS, const Value *RHS) { |
| 454 | // We do use m_Specific(C) in MatchCD, so we have to make sure that |
| 455 | // it's bound to anything and match(LHS, MatchASubBC()) absolutely |
| 456 | // has to evaluate first and return true. |
| 457 | // |
| 458 | // If Match returns true, it is guaranteed that B != nullptr, C != nullptr. |
| 459 | return match(V: LHS, P: MatchASubBC()) && match(V: RHS, P: MatchCD()); |
| 460 | }; |
| 461 | |
| 462 | if (!Match(Op0, Op1) && !Match(Op1, Op0)) |
| 463 | return; |
| 464 | |
| 465 | const auto ComputeKnownBitsOrOne = [&](const Value *V) { |
| 466 | // For some of the values we use the convention of leaving |
| 467 | // it nullptr to signify an implicit constant 1. |
| 468 | return V ? computeKnownBits(V, DemandedElts, Q, Depth: Depth + 1) |
| 469 | : KnownBits::makeConstant(C: APInt(BitWidth, 1)); |
| 470 | }; |
| 471 | |
| 472 | // Check that all operands are non-negative |
| 473 | const KnownBits KnownA = ComputeKnownBitsOrOne(A); |
| 474 | if (!KnownA.isNonNegative()) |
| 475 | return; |
| 476 | |
| 477 | const KnownBits KnownD = ComputeKnownBitsOrOne(D); |
| 478 | if (!KnownD.isNonNegative()) |
| 479 | return; |
| 480 | |
| 481 | const KnownBits KnownB = computeKnownBits(V: B, DemandedElts, Q, Depth: Depth + 1); |
| 482 | if (!KnownB.isNonNegative()) |
| 483 | return; |
| 484 | |
| 485 | const KnownBits KnownC = computeKnownBits(V: C, DemandedElts, Q, Depth: Depth + 1); |
| 486 | if (!KnownC.isNonNegative()) |
| 487 | return; |
| 488 | |
| 489 | // If we matched subtraction as xor, we need to actually check that xor |
| 490 | // is semantically equivalent to subtraction. |
| 491 | // |
| 492 | // For that to be true, b has to be a mask for c or that b's known |
| 493 | // ones cover all known and possible ones of c. |
| 494 | if (SubBC->getOpcode() == Instruction::Xor && |
| 495 | !KnownC.getMaxValue().isSubsetOf(RHS: KnownB.getMinValue())) |
| 496 | return; |
| 497 | |
| 498 | const APInt MaxA = KnownA.getMaxValue(); |
| 499 | const APInt MaxD = KnownD.getMaxValue(); |
| 500 | const APInt MaxAD = APIntOps::umax(A: MaxA, B: MaxD); |
| 501 | const APInt MaxB = KnownB.getMaxValue(); |
| 502 | |
| 503 | // We can't infer leading zeros info if the upper-bound estimate wraps. |
| 504 | bool Overflow; |
| 505 | const APInt UpperBound = MaxAD.umul_ov(RHS: MaxB, Overflow); |
| 506 | |
| 507 | if (Overflow) |
| 508 | return; |
| 509 | |
| 510 | // If we know that x <= y and both are positive than x has at least the same |
| 511 | // number of leading zeros as y. |
| 512 | const unsigned MinimumNumberOfLeadingZeros = UpperBound.countl_zero(); |
| 513 | KnownOut.Zero.setHighBits(MinimumNumberOfLeadingZeros); |
| 514 | } |
| 515 | |
| 516 | /// Infer non-negativity for R + (R < 0 ? A : 0), where R = srem X, Y. |
| 517 | /// A must be a non-negative Y or abs(Y). |
| 518 | static void computeKnownBitsFromEuclideanRem(const Value *Op0, const Value *Op1, |
| 519 | KnownBits &KnownOut, |
| 520 | const SimplifyQuery &Q, |
| 521 | unsigned Depth) { |
| 522 | Type *Ty = Op0->getType(); |
| 523 | |
| 524 | // Only handle scalars for now. |
| 525 | if (Ty->isVectorTy()) |
| 526 | return; |
| 527 | |
| 528 | const Value *Divisor = nullptr, *Cond = nullptr, *Remainder = nullptr, |
| 529 | *AdjustmentValue = nullptr; |
| 530 | CmpPredicate Pred; |
| 531 | const APInt *Threshold = nullptr; |
| 532 | bool TrueIfNegative = false; |
| 533 | |
| 534 | const auto Match = [&](const Value *Rem, const Value *Adjustment) { |
| 535 | Remainder = Rem; |
| 536 | |
| 537 | // Match an adjustment added only when this remainder is negative. |
| 538 | return match(V: Rem, P: m_SRem(L: m_Value(), R: m_Value(V&: Divisor))) && |
| 539 | match(V: Adjustment, |
| 540 | P: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: AdjustmentValue), R: m_Zero())) && |
| 541 | match(V: Cond, P: m_ICmp(Pred, L: m_Specific(V: Rem), R: m_APInt(Res&: Threshold))) && |
| 542 | isSignBitCheck(Pred, RHS: *Threshold, TrueIfSigned&: TrueIfNegative) && TrueIfNegative; |
| 543 | }; |
| 544 | |
| 545 | if (!Match(Op0, Op1) && !Match(Op1, Op0)) |
| 546 | return; |
| 547 | |
| 548 | // Accept Y itself if non-negative, or abs(Y). |
| 549 | bool IsPositiveDivisorCase = |
| 550 | Divisor == AdjustmentValue && isKnownNonNegative(V: Divisor, SQ: Q, Depth: Depth + 1); |
| 551 | bool IsAbsoluteValueCase = |
| 552 | match(V: AdjustmentValue, |
| 553 | P: m_Intrinsic<Intrinsic::abs>(Ops: m_Specific(V: Divisor), Ops: m_Value())); |
| 554 | |
| 555 | if (!IsPositiveDivisorCase && !IsAbsoluteValueCase) |
| 556 | return; |
| 557 | |
| 558 | // Repeated uses must agree. An undef value could break this relationship. |
| 559 | if (!isGuaranteedNotToBeUndef(V: Remainder, AC: Q.AC, CtxI: Q.CtxI, DT: Q.DT, Depth: Depth + 1) || |
| 560 | !isGuaranteedNotToBeUndef(V: Divisor, AC: Q.AC, CtxI: Q.CtxI, DT: Q.DT, Depth: Depth + 1)) |
| 561 | return; |
| 562 | |
| 563 | // For Y other than signed minimum, adding abs(Y) to a negative remainder |
| 564 | // gives a non-negative result without overflow. For signed minimum, the |
| 565 | // wrapping addition clears the remainder's sign bit. Poison-producing cases |
| 566 | // do not invalidate this deduction. |
| 567 | KnownOut.makeNonNegative(); |
| 568 | } |
| 569 | |
| 570 | static void computeKnownBitsAddSub(bool Add, const Value *Op0, const Value *Op1, |
| 571 | bool NSW, bool NUW, |
| 572 | const APInt &DemandedElts, |
| 573 | KnownBits &KnownOut, KnownBits &Known2, |
| 574 | const SimplifyQuery &Q, unsigned Depth) { |
| 575 | computeKnownBits(V: Op1, DemandedElts, Known&: KnownOut, Q, Depth: Depth + 1); |
| 576 | |
| 577 | if (KnownOut.isUnknown() && !NSW && !NUW) { |
| 578 | if (Add) { |
| 579 | // Independent operand bits give no result, but the Euclidean remainder |
| 580 | // pattern can establish non-negativity through the operands' |
| 581 | // relationship. |
| 582 | computeKnownBitsFromEuclideanRem(Op0, Op1, KnownOut, Q, Depth); |
| 583 | } |
| 584 | return; |
| 585 | } |
| 586 | |
| 587 | computeKnownBits(V: Op0, DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 588 | KnownOut = KnownBits::computeForAddSub(Add, NSW, NUW, LHS: Known2, RHS: KnownOut); |
| 589 | |
| 590 | if (!Add && NSW && !KnownOut.isNonNegative() && |
| 591 | (isImpliedByDomCondition(Pred: ICmpInst::ICMP_SLE, LHS: Op1, RHS: Op0, ContextI: Q.CtxI, DL: Q.DL) |
| 592 | .value_or(u: false) || |
| 593 | match(V: Op1, P: m_c_SMin(L: m_Specific(V: Op0), R: m_Value())))) |
| 594 | KnownOut.makeNonNegative(); |
| 595 | |
| 596 | if (Add) { |
| 597 | // Try to match lerp pattern and combine results |
| 598 | computeKnownBitsFromLerpPattern(Op0, Op1, DemandedElts, KnownOut, Q, Depth); |
| 599 | |
| 600 | // Use the relationship between the remainder and its adjustment. |
| 601 | computeKnownBitsFromEuclideanRem(Op0, Op1, KnownOut, Q, Depth); |
| 602 | } |
| 603 | } |
| 604 | |
| 605 | static void computeKnownBitsMul(const Value *Op0, const Value *Op1, bool NSW, |
| 606 | bool NUW, const APInt &DemandedElts, |
| 607 | KnownBits &Known, KnownBits &Known2, |
| 608 | const SimplifyQuery &Q, unsigned Depth) { |
| 609 | computeKnownBits(V: Op1, DemandedElts, Known, Q, Depth: Depth + 1); |
| 610 | computeKnownBits(V: Op0, DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 611 | |
| 612 | bool isKnownNegative = false; |
| 613 | bool isKnownNonNegative = false; |
| 614 | // If the multiplication is known not to overflow, compute the sign bit. |
| 615 | if (NSW) { |
| 616 | if (Op0 == Op1) { |
| 617 | // The product of a number with itself is non-negative. |
| 618 | isKnownNonNegative = true; |
| 619 | } else { |
| 620 | bool isKnownNonNegativeOp1 = Known.isNonNegative(); |
| 621 | bool isKnownNonNegativeOp0 = Known2.isNonNegative(); |
| 622 | bool isKnownNegativeOp1 = Known.isNegative(); |
| 623 | bool isKnownNegativeOp0 = Known2.isNegative(); |
| 624 | // The product of two numbers with the same sign is non-negative. |
| 625 | isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) || |
| 626 | (isKnownNonNegativeOp1 && isKnownNonNegativeOp0); |
| 627 | if (!isKnownNonNegative && NUW) { |
| 628 | // mul nuw nsw with a factor > 1 is non-negative. |
| 629 | KnownBits One = KnownBits::makeConstant(C: APInt(Known.getBitWidth(), 1)); |
| 630 | isKnownNonNegative = KnownBits::sgt(LHS: Known, RHS: One).value_or(u: false) || |
| 631 | KnownBits::sgt(LHS: Known2, RHS: One).value_or(u: false); |
| 632 | } |
| 633 | |
| 634 | // The product of a negative number and a non-negative number is either |
| 635 | // negative or zero. |
| 636 | if (!isKnownNonNegative) |
| 637 | isKnownNegative = |
| 638 | (isKnownNegativeOp1 && isKnownNonNegativeOp0 && |
| 639 | Known2.isNonZero()) || |
| 640 | (isKnownNegativeOp0 && isKnownNonNegativeOp1 && Known.isNonZero()); |
| 641 | } |
| 642 | } |
| 643 | |
| 644 | bool SelfMultiply = Op0 == Op1; |
| 645 | if (SelfMultiply) |
| 646 | SelfMultiply &= |
| 647 | isGuaranteedNotToBeUndef(V: Op0, AC: Q.AC, CtxI: Q.CtxI, DT: Q.DT, Depth: Depth + 1); |
| 648 | Known = KnownBits::mul(LHS: Known, RHS: Known2, NoUndefSelfMultiply: SelfMultiply); |
| 649 | |
| 650 | if (SelfMultiply) { |
| 651 | unsigned SignBits = ComputeNumSignBits(V: Op0, DemandedElts, Q, Depth: Depth + 1); |
| 652 | unsigned TyBits = Op0->getType()->getScalarSizeInBits(); |
| 653 | unsigned OutValidBits = 2 * (TyBits - SignBits + 1); |
| 654 | |
| 655 | if (OutValidBits < TyBits) { |
| 656 | APInt KnownZeroMask = |
| 657 | APInt::getHighBitsSet(numBits: TyBits, hiBitsSet: TyBits - OutValidBits + 1); |
| 658 | Known.Zero |= KnownZeroMask; |
| 659 | } |
| 660 | } |
| 661 | |
| 662 | // Only make use of no-wrap flags if we failed to compute the sign bit |
| 663 | // directly. This matters if the multiplication always overflows, in |
| 664 | // which case we prefer to follow the result of the direct computation, |
| 665 | // though as the program is invoking undefined behaviour we can choose |
| 666 | // whatever we like here. |
| 667 | if (isKnownNonNegative && !Known.isNegative()) |
| 668 | Known.makeNonNegative(); |
| 669 | else if (isKnownNegative && !Known.isNonNegative()) |
| 670 | Known.makeNegative(); |
| 671 | } |
| 672 | |
| 673 | void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges, |
| 674 | KnownBits &Known) { |
| 675 | unsigned BitWidth = Known.getBitWidth(); |
| 676 | unsigned NumRanges = Ranges.getNumOperands() / 2; |
| 677 | assert(NumRanges >= 1); |
| 678 | |
| 679 | Known.setAllConflict(); |
| 680 | |
| 681 | for (unsigned i = 0; i < NumRanges; ++i) { |
| 682 | ConstantInt *Lower = |
| 683 | mdconst::extract<ConstantInt>(MD: Ranges.getOperand(I: 2 * i + 0)); |
| 684 | ConstantInt *Upper = |
| 685 | mdconst::extract<ConstantInt>(MD: Ranges.getOperand(I: 2 * i + 1)); |
| 686 | ConstantRange Range(Lower->getValue(), Upper->getValue()); |
| 687 | // BitWidth must equal the Ranges BitWidth for the correct number of high |
| 688 | // bits to be set. |
| 689 | assert(BitWidth == Range.getBitWidth() && |
| 690 | "Known bit width must match range bit width!" ); |
| 691 | |
| 692 | // The first CommonPrefixBits of all values in Range are equal. |
| 693 | unsigned CommonPrefixBits = |
| 694 | (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countl_zero(); |
| 695 | APInt Mask = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: CommonPrefixBits); |
| 696 | APInt UnsignedMax = Range.getUnsignedMax().zextOrTrunc(width: BitWidth); |
| 697 | Known.One &= UnsignedMax & Mask; |
| 698 | Known.Zero &= ~UnsignedMax & Mask; |
| 699 | } |
| 700 | } |
| 701 | |
| 702 | static bool isEphemeralValueOf(const Instruction *I, const Value *E) { |
| 703 | // The instruction defining an assumption's condition itself is always |
| 704 | // considered ephemeral to that assumption (even if it has other |
| 705 | // non-ephemeral users). See r246696's test case for an example. |
| 706 | if (is_contained(Range: I->operands(), Element: E)) |
| 707 | return true; |
| 708 | |
| 709 | const auto *EI = dyn_cast<Instruction>(Val: E); |
| 710 | if (!EI) |
| 711 | return false; |
| 712 | |
| 713 | if (EI == I) |
| 714 | return true; |
| 715 | |
| 716 | SmallPtrSet<const Instruction *, 16> Visited; |
| 717 | SmallVector<const Instruction *, 16> WorkList; |
| 718 | Visited.insert(Ptr: EI); |
| 719 | WorkList.push_back(Elt: EI); |
| 720 | bool ReachesI = false; |
| 721 | while (!WorkList.empty()) { |
| 722 | const Instruction *V = WorkList.pop_back_val(); |
| 723 | for (const User *U : V->users()) { |
| 724 | const auto *UI = cast<Instruction>(Val: U); |
| 725 | if (UI == I) { |
| 726 | ReachesI = true; |
| 727 | continue; |
| 728 | } |
| 729 | if (UI->mayHaveSideEffects() || UI->isTerminator()) |
| 730 | return false; |
| 731 | if (Visited.insert(Ptr: UI).second) |
| 732 | WorkList.push_back(Elt: UI); |
| 733 | } |
| 734 | } |
| 735 | return ReachesI; |
| 736 | } |
| 737 | |
| 738 | // Is this an intrinsic that cannot be speculated but also cannot trap? |
| 739 | bool llvm::isAssumeLikeIntrinsic(const Instruction *I) { |
| 740 | if (const IntrinsicInst *CI = dyn_cast<IntrinsicInst>(Val: I)) |
| 741 | return CI->isAssumeLikeIntrinsic(); |
| 742 | |
| 743 | return false; |
| 744 | } |
| 745 | |
| 746 | bool llvm::isValidAssumeForContext(const Instruction *Inv, |
| 747 | const Instruction *CtxI, |
| 748 | const DominatorTree *DT, |
| 749 | bool AllowEphemerals) { |
| 750 | // There are two restrictions on the use of an assume: |
| 751 | // 1. The assume must dominate the context (or the control flow must |
| 752 | // reach the assume whenever it reaches the context). |
| 753 | // 2. The context must not be in the assume's set of ephemeral values |
| 754 | // (otherwise we will use the assume to prove that the condition |
| 755 | // feeding the assume is trivially true, thus causing the removal of |
| 756 | // the assume). |
| 757 | |
| 758 | // If the checked instruction and the assume are the same, it's valid exactly |
| 759 | // if ephemeral values are allowed. |
| 760 | if (Inv == CtxI) |
| 761 | return AllowEphemerals; |
| 762 | |
| 763 | if (Inv->getParent() == CtxI->getParent()) { |
| 764 | // If Inv and CtxI are in the same block, check if the assume (Inv) is first |
| 765 | // in the BB. |
| 766 | if (Inv->comesBefore(Other: CtxI)) |
| 767 | return true; |
| 768 | |
| 769 | // The context comes first, but they're both in the same block. |
| 770 | // Make sure there is nothing in between that might interrupt |
| 771 | // the control flow, not even CtxI itself. |
| 772 | // We limit the scan distance between the assume and its context instruction |
| 773 | // to avoid a compile-time explosion. This limit is chosen arbitrarily, so |
| 774 | // it can be adjusted if needed (could be turned into a cl::opt). |
| 775 | auto Range = make_range(x: CtxI->getIterator(), y: Inv->getIterator()); |
| 776 | if (!isGuaranteedToTransferExecutionToSuccessor(Range, ScanLimit: 15)) |
| 777 | return false; |
| 778 | |
| 779 | return AllowEphemerals || !isEphemeralValueOf(I: Inv, E: CtxI); |
| 780 | } |
| 781 | |
| 782 | // Inv and CtxI are in different blocks. |
| 783 | if (DT) { |
| 784 | if (DT->dominates(Def: Inv, User: CtxI)) |
| 785 | return true; |
| 786 | } else if (Inv->getParent() == CtxI->getParent()->getSinglePredecessor() || |
| 787 | Inv->getParent()->isEntryBlock()) { |
| 788 | // We don't have a DT, but this trivially dominates. |
| 789 | return true; |
| 790 | } |
| 791 | |
| 792 | return false; |
| 793 | } |
| 794 | |
| 795 | static bool hasNoFreeInRange(BasicBlock::const_iterator Begin, |
| 796 | BasicBlock::const_iterator End, |
| 797 | unsigned &NumChecked) { |
| 798 | for (const Instruction &I : make_range(x: Begin, y: End)) { |
| 799 | if (NumChecked++ > MaxInstrsToCheckForFree) |
| 800 | return false; |
| 801 | if (auto *CB = dyn_cast<CallBase>(Val: &I)) { |
| 802 | if (!CB->hasFnAttr(Kind: Attribute::NoFree)) |
| 803 | return false; |
| 804 | } else if (I.maySynchronize()) { |
| 805 | return false; |
| 806 | } |
| 807 | } |
| 808 | return true; |
| 809 | } |
| 810 | |
| 811 | bool llvm::willNotFreeBetween(const Instruction *Assume, |
| 812 | const Instruction *CtxI, |
| 813 | const DominatorTree *DT) { |
| 814 | const BasicBlock *CtxBB = CtxI->getParent(); |
| 815 | const BasicBlock *AssumeBB = Assume->getParent(); |
| 816 | unsigned NumChecked = 0; |
| 817 | BasicBlock::const_iterator CtxIter = CtxI->getIterator(); |
| 818 | if (CtxBB == AssumeBB) { |
| 819 | if (Assume != CtxI && !Assume->comesBefore(Other: CtxI)) |
| 820 | return false; |
| 821 | return hasNoFreeInRange(Begin: Assume->getIterator(), End: CtxIter, NumChecked); |
| 822 | } |
| 823 | if (DT && !DT->dominates(Def: Assume, User: CtxI)) |
| 824 | return false; |
| 825 | if (!hasNoFreeInRange(Begin: CtxBB->begin(), End: CtxIter, NumChecked)) |
| 826 | return false; |
| 827 | if (pred_empty(BB: CtxBB)) |
| 828 | return false; |
| 829 | |
| 830 | // Note: CtxBB is NOT pre-inserted into Visited to ensure that loop |
| 831 | // backedges returning to CtxBB are enqueued and checked correctly. |
| 832 | SmallVector<const BasicBlock *, 16> Worklist(predecessors(BB: CtxBB)); |
| 833 | SmallPtrSet<const BasicBlock *, 16> Visited; |
| 834 | while (!Worklist.empty()) { |
| 835 | const BasicBlock *CurBB = Worklist.pop_back_val(); |
| 836 | if (!Visited.insert(Ptr: CurBB).second) |
| 837 | continue; |
| 838 | |
| 839 | if (CurBB == AssumeBB) { |
| 840 | if (!hasNoFreeInRange(Begin: Assume->getIterator(), End: AssumeBB->end(), NumChecked)) |
| 841 | return false; |
| 842 | continue; |
| 843 | } |
| 844 | assert((!DT || DT->dominates(AssumeBB, CurBB)) && |
| 845 | "Blocks between Assume and CtxI must be dominated by AssumeBB" ); |
| 846 | |
| 847 | if (pred_empty(BB: CurBB)) { |
| 848 | // Reaching the entry block means there is a path to CtxI that does not |
| 849 | // pass through Assume. |
| 850 | if (CurBB->isEntryBlock()) |
| 851 | return false; |
| 852 | |
| 853 | // CurBB is unreachable (for example, emptied by SimplifyCFG but not yet |
| 854 | // removed), so nothing in it can execute a free. |
| 855 | continue; |
| 856 | } |
| 857 | |
| 858 | // If CurBB == CtxBB (due to a loop backedge targeting CtxBB), check |
| 859 | // instructions from CtxIter to the end of CtxBB (instructions before |
| 860 | // CtxIter were checked above). Otherwise, check the entire block. |
| 861 | auto StartIt = (CurBB == CtxBB) ? CtxIter : CurBB->begin(); |
| 862 | if (!hasNoFreeInRange(Begin: StartIt, End: CurBB->end(), NumChecked)) |
| 863 | return false; |
| 864 | append_range(C&: Worklist, R: predecessors(BB: CurBB)); |
| 865 | } |
| 866 | return true; |
| 867 | } |
| 868 | |
| 869 | // TODO: cmpExcludesZero misses many cases where `RHS` is non-constant but |
| 870 | // we still have enough information about `RHS` to conclude non-zero. For |
| 871 | // example Pred=EQ, RHS=isKnownNonZero. cmpExcludesZero is called in loops |
| 872 | // so the extra compile time may not be worth it, but possibly a second API |
| 873 | // should be created for use outside of loops. |
| 874 | static bool cmpExcludesZero(CmpInst::Predicate Pred, const Value *RHS) { |
| 875 | // v u> y implies v != 0. |
| 876 | if (Pred == ICmpInst::ICMP_UGT) |
| 877 | return true; |
| 878 | |
| 879 | // Special-case v != 0 to also handle v != null. |
| 880 | if (Pred == ICmpInst::ICMP_NE) |
| 881 | return match(V: RHS, P: m_Zero()); |
| 882 | |
| 883 | // All other predicates - rely on generic ConstantRange handling. |
| 884 | const APInt *C; |
| 885 | auto Zero = APInt::getZero(numBits: RHS->getType()->getScalarSizeInBits()); |
| 886 | if (match(V: RHS, P: m_APInt(Res&: C))) { |
| 887 | ConstantRange TrueValues = ConstantRange::makeExactICmpRegion(Pred, Other: *C); |
| 888 | return !TrueValues.contains(Val: Zero); |
| 889 | } |
| 890 | |
| 891 | auto *VC = dyn_cast<ConstantDataVector>(Val: RHS); |
| 892 | if (VC == nullptr) |
| 893 | return false; |
| 894 | |
| 895 | for (unsigned ElemIdx = 0, NElem = VC->getNumElements(); ElemIdx < NElem; |
| 896 | ++ElemIdx) { |
| 897 | ConstantRange TrueValues = ConstantRange::makeExactICmpRegion( |
| 898 | Pred, Other: VC->getElementAsAPInt(i: ElemIdx)); |
| 899 | if (TrueValues.contains(Val: Zero)) |
| 900 | return false; |
| 901 | } |
| 902 | return true; |
| 903 | } |
| 904 | |
| 905 | static void breakSelfRecursivePHI(const Use *U, const PHINode *PHI, |
| 906 | Value *&ValOut, Instruction *&CtxIOut, |
| 907 | const PHINode **PhiOut = nullptr) { |
| 908 | ValOut = U->get(); |
| 909 | if (ValOut == PHI) |
| 910 | return; |
| 911 | CtxIOut = PHI->getIncomingBlock(U: *U)->getTerminator(); |
| 912 | if (PhiOut) |
| 913 | *PhiOut = PHI; |
| 914 | Value *V; |
| 915 | // If the Use is a select of this phi, compute analysis on other arm to break |
| 916 | // recursion. |
| 917 | // TODO: Min/Max |
| 918 | if (match(V: ValOut, P: m_Select(C: m_Value(), L: m_Specific(V: PHI), R: m_Value(V))) || |
| 919 | match(V: ValOut, P: m_Select(C: m_Value(), L: m_Value(V), R: m_Specific(V: PHI)))) |
| 920 | ValOut = V; |
| 921 | |
| 922 | // Same for select, if this phi is 2-operand phi, compute analysis on other |
| 923 | // incoming value to break recursion. |
| 924 | // TODO: We could handle any number of incoming edges as long as we only have |
| 925 | // two unique values. |
| 926 | if (auto *IncPhi = dyn_cast<PHINode>(Val: ValOut); |
| 927 | IncPhi && IncPhi->getNumIncomingValues() == 2) { |
| 928 | for (int Idx = 0; Idx < 2; ++Idx) { |
| 929 | if (IncPhi->getIncomingValue(i: Idx) == PHI) { |
| 930 | ValOut = IncPhi->getIncomingValue(i: 1 - Idx); |
| 931 | if (PhiOut) |
| 932 | *PhiOut = IncPhi; |
| 933 | CtxIOut = IncPhi->getIncomingBlock(i: 1 - Idx)->getTerminator(); |
| 934 | break; |
| 935 | } |
| 936 | } |
| 937 | } |
| 938 | } |
| 939 | |
| 940 | static bool isKnownNonZeroFromAssume(const Value *V, const SimplifyQuery &Q) { |
| 941 | // Use of assumptions is context-sensitive. If we don't have a context, we |
| 942 | // cannot use them! |
| 943 | if (!Q.AC || !Q.CtxI) |
| 944 | return false; |
| 945 | |
| 946 | for (AssumptionCache::ResultElem &Elem : Q.AC->assumptionsFor(V)) { |
| 947 | if (!Elem.Assume) |
| 948 | continue; |
| 949 | |
| 950 | AssumeInst *I = cast<AssumeInst>(Val&: Elem.Assume); |
| 951 | assert(I->getFunction() == Q.CtxI->getFunction() && |
| 952 | "Got assumption for the wrong function!" ); |
| 953 | |
| 954 | if (Elem.Index != AssumptionCache::ExprResultIdx) { |
| 955 | if (assumeBundleImpliesNonNull(Val: V, Context: Q.CtxI->getFunction(), |
| 956 | OBU: I->getOperandBundleAt(Index: Elem.Index)) && |
| 957 | isValidAssumeForContext(I, Q)) |
| 958 | return true; |
| 959 | continue; |
| 960 | } |
| 961 | |
| 962 | // Warning: This loop can end up being somewhat performance sensitive. |
| 963 | // We're running this loop for once for each value queried resulting in a |
| 964 | // runtime of ~O(#assumes * #values). |
| 965 | |
| 966 | if (match(V: I->getArgOperand(i: 0), P: m_Trunc(Op: m_Specific(V))) && |
| 967 | isValidAssumeForContext(I, Q)) |
| 968 | return true; |
| 969 | |
| 970 | Value *RHS; |
| 971 | CmpPredicate Pred; |
| 972 | auto m_V = m_CombineOr(Ps: m_Specific(V), Ps: m_PtrToInt(Op: m_Specific(V))); |
| 973 | if (!match(V: I->getArgOperand(i: 0), P: m_c_ICmp(Pred, L: m_V, R: m_Value(V&: RHS)))) |
| 974 | continue; |
| 975 | |
| 976 | if (cmpExcludesZero(Pred, RHS) && isValidAssumeForContext(I, Q)) |
| 977 | return true; |
| 978 | } |
| 979 | |
| 980 | return false; |
| 981 | } |
| 982 | |
| 983 | static void computeKnownBitsFromCmp(const Value *V, CmpInst::Predicate Pred, |
| 984 | Value *LHS, Value *RHS, KnownBits &Known, |
| 985 | const SimplifyQuery &Q) { |
| 986 | if (RHS->getType()->isPointerTy()) { |
| 987 | // Handle comparison of pointer to null explicitly, as it will not be |
| 988 | // covered by the m_APInt() logic below. |
| 989 | if (LHS == V && match(V: RHS, P: m_Zero())) { |
| 990 | switch (Pred) { |
| 991 | case ICmpInst::ICMP_EQ: |
| 992 | Known.setAllZero(); |
| 993 | break; |
| 994 | case ICmpInst::ICMP_SGE: |
| 995 | case ICmpInst::ICMP_SGT: |
| 996 | Known.makeNonNegative(); |
| 997 | break; |
| 998 | case ICmpInst::ICMP_SLT: |
| 999 | Known.makeNegative(); |
| 1000 | break; |
| 1001 | default: |
| 1002 | break; |
| 1003 | } |
| 1004 | } |
| 1005 | return; |
| 1006 | } |
| 1007 | |
| 1008 | unsigned BitWidth = Known.getBitWidth(); |
| 1009 | auto m_V = |
| 1010 | m_CombineOr(Ps: m_Specific(V), Ps: m_PtrToIntSameSize(DL: Q.DL, Op: m_Specific(V))); |
| 1011 | |
| 1012 | Value *Y; |
| 1013 | const APInt *Mask, *C; |
| 1014 | if (!match(V: RHS, P: m_APInt(Res&: C))) |
| 1015 | return; |
| 1016 | |
| 1017 | uint64_t ShAmt; |
| 1018 | switch (Pred) { |
| 1019 | case ICmpInst::ICMP_EQ: |
| 1020 | // assume(V = C) |
| 1021 | if (match(V: LHS, P: m_V)) { |
| 1022 | Known = Known.unionWith(RHS: KnownBits::makeConstant(C: *C)); |
| 1023 | // assume(V & Mask = C) |
| 1024 | } else if (match(V: LHS, P: m_c_And(L: m_V, R: m_Value(V&: Y)))) { |
| 1025 | // For one bits in Mask, we can propagate bits from C to V. |
| 1026 | Known.One |= *C; |
| 1027 | if (match(V: Y, P: m_APInt(Res&: Mask))) |
| 1028 | Known.Zero |= ~*C & *Mask; |
| 1029 | // assume(V | Mask = C) |
| 1030 | } else if (match(V: LHS, P: m_c_Or(L: m_V, R: m_Value(V&: Y)))) { |
| 1031 | // For zero bits in Mask, we can propagate bits from C to V. |
| 1032 | Known.Zero |= ~*C; |
| 1033 | if (match(V: Y, P: m_APInt(Res&: Mask))) |
| 1034 | Known.One |= *C & ~*Mask; |
| 1035 | // assume(V << ShAmt = C) |
| 1036 | } else if (match(V: LHS, P: m_Shl(L: m_V, R: m_ConstantInt(V&: ShAmt))) && |
| 1037 | ShAmt < BitWidth) { |
| 1038 | // For those bits in C that are known, we can propagate them to known |
| 1039 | // bits in V shifted to the right by ShAmt. |
| 1040 | KnownBits RHSKnown = KnownBits::makeConstant(C: *C); |
| 1041 | RHSKnown >>= ShAmt; |
| 1042 | Known = Known.unionWith(RHS: RHSKnown); |
| 1043 | // assume(V >> ShAmt = C) |
| 1044 | } else if (match(V: LHS, P: m_Shr(L: m_V, R: m_ConstantInt(V&: ShAmt))) && |
| 1045 | ShAmt < BitWidth) { |
| 1046 | // For those bits in RHS that are known, we can propagate them to known |
| 1047 | // bits in V shifted to the right by C. |
| 1048 | KnownBits RHSKnown = KnownBits::makeConstant(C: *C); |
| 1049 | RHSKnown <<= ShAmt; |
| 1050 | Known = Known.unionWith(RHS: RHSKnown); |
| 1051 | } |
| 1052 | break; |
| 1053 | case ICmpInst::ICMP_NE: { |
| 1054 | // assume (V & B != 0) where B is a power of 2 |
| 1055 | const APInt *BPow2; |
| 1056 | if (C->isZero() && match(V: LHS, P: m_And(L: m_V, R: m_Power2(V&: BPow2)))) |
| 1057 | Known.One |= *BPow2; |
| 1058 | break; |
| 1059 | } |
| 1060 | default: { |
| 1061 | const APInt *Offset = nullptr; |
| 1062 | if (match(V: LHS, P: m_CombineOr(Ps: m_V, Ps: m_AddLike(L: m_V, R: m_APInt(Res&: Offset))))) { |
| 1063 | ConstantRange LHSRange = ConstantRange::makeAllowedICmpRegion(Pred, Other: *C); |
| 1064 | if (Offset) |
| 1065 | LHSRange = LHSRange.sub(Other: *Offset); |
| 1066 | Known = Known.unionWith(RHS: LHSRange.toKnownBits()); |
| 1067 | } |
| 1068 | if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) { |
| 1069 | // X & Y u> C -> X u> C && Y u> C |
| 1070 | // X nuw- Y u> C -> X u> C |
| 1071 | if (match(V: LHS, P: m_c_And(L: m_V, R: m_Value())) || |
| 1072 | match(V: LHS, P: m_NUWSub(L: m_V, R: m_Value()))) |
| 1073 | Known.One.setHighBits( |
| 1074 | (*C + (Pred == ICmpInst::ICMP_UGT)).countLeadingOnes()); |
| 1075 | } |
| 1076 | if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) { |
| 1077 | // X | Y u< C -> X u< C && Y u< C |
| 1078 | // X nuw+ Y u< C -> X u< C && Y u< C |
| 1079 | if (match(V: LHS, P: m_c_Or(L: m_V, R: m_Value())) || |
| 1080 | match(V: LHS, P: m_c_NUWAdd(L: m_V, R: m_Value()))) { |
| 1081 | Known.Zero.setHighBits( |
| 1082 | (*C - (Pred == ICmpInst::ICMP_ULT)).countLeadingZeros()); |
| 1083 | } |
| 1084 | } |
| 1085 | } break; |
| 1086 | } |
| 1087 | } |
| 1088 | |
| 1089 | static void computeKnownBitsFromICmpCond(const Value *V, ICmpInst *Cmp, |
| 1090 | KnownBits &Known, |
| 1091 | const SimplifyQuery &SQ, bool Invert) { |
| 1092 | ICmpInst::Predicate Pred = |
| 1093 | Invert ? Cmp->getInversePredicate() : Cmp->getPredicate(); |
| 1094 | Value *LHS = Cmp->getOperand(i_nocapture: 0); |
| 1095 | Value *RHS = Cmp->getOperand(i_nocapture: 1); |
| 1096 | |
| 1097 | // Handle icmp pred (trunc V), C |
| 1098 | if (match(V: LHS, P: m_Trunc(Op: m_Specific(V)))) { |
| 1099 | KnownBits DstKnown(LHS->getType()->getScalarSizeInBits()); |
| 1100 | computeKnownBitsFromCmp(V: LHS, Pred, LHS, RHS, Known&: DstKnown, Q: SQ); |
| 1101 | if (cast<TruncInst>(Val: LHS)->hasNoUnsignedWrap()) |
| 1102 | Known = Known.unionWith(RHS: DstKnown.zext(BitWidth: Known.getBitWidth())); |
| 1103 | else |
| 1104 | Known = Known.unionWith(RHS: DstKnown.anyext(BitWidth: Known.getBitWidth())); |
| 1105 | return; |
| 1106 | } |
| 1107 | |
| 1108 | computeKnownBitsFromCmp(V, Pred, LHS, RHS, Known, Q: SQ); |
| 1109 | } |
| 1110 | |
| 1111 | static void computeKnownBitsFromCond(const Value *V, Value *Cond, |
| 1112 | KnownBits &Known, const SimplifyQuery &SQ, |
| 1113 | bool Invert, unsigned Depth) { |
| 1114 | Value *A, *B; |
| 1115 | if (Depth < MaxAnalysisRecursionDepth && |
| 1116 | match(V: Cond, P: m_LogicalOp(L: m_Value(V&: A), R: m_Value(V&: B)))) { |
| 1117 | KnownBits Known2(Known.getBitWidth()); |
| 1118 | KnownBits Known3(Known.getBitWidth()); |
| 1119 | computeKnownBitsFromCond(V, Cond: A, Known&: Known2, SQ, Invert, Depth: Depth + 1); |
| 1120 | computeKnownBitsFromCond(V, Cond: B, Known&: Known3, SQ, Invert, Depth: Depth + 1); |
| 1121 | if (Invert ? match(V: Cond, P: m_LogicalOr(L: m_Value(), R: m_Value())) |
| 1122 | : match(V: Cond, P: m_LogicalAnd(L: m_Value(), R: m_Value()))) |
| 1123 | Known2 = Known2.unionWith(RHS: Known3); |
| 1124 | else |
| 1125 | Known2 = Known2.intersectWith(RHS: Known3); |
| 1126 | Known = Known.unionWith(RHS: Known2); |
| 1127 | return; |
| 1128 | } |
| 1129 | |
| 1130 | if (auto *Cmp = dyn_cast<ICmpInst>(Val: Cond)) { |
| 1131 | computeKnownBitsFromICmpCond(V, Cmp, Known, SQ, Invert); |
| 1132 | return; |
| 1133 | } |
| 1134 | |
| 1135 | if (match(V: Cond, P: m_Trunc(Op: m_Specific(V)))) { |
| 1136 | KnownBits DstKnown(1); |
| 1137 | if (Invert) { |
| 1138 | DstKnown.setAllZero(); |
| 1139 | } else { |
| 1140 | DstKnown.setAllOnes(); |
| 1141 | } |
| 1142 | if (cast<TruncInst>(Val: Cond)->hasNoUnsignedWrap()) { |
| 1143 | Known = Known.unionWith(RHS: DstKnown.zext(BitWidth: Known.getBitWidth())); |
| 1144 | return; |
| 1145 | } |
| 1146 | Known = Known.unionWith(RHS: DstKnown.anyext(BitWidth: Known.getBitWidth())); |
| 1147 | return; |
| 1148 | } |
| 1149 | |
| 1150 | if (Depth < MaxAnalysisRecursionDepth && match(V: Cond, P: m_Not(V: m_Value(V&: A)))) |
| 1151 | computeKnownBitsFromCond(V, Cond: A, Known, SQ, Invert: !Invert, Depth: Depth + 1); |
| 1152 | } |
| 1153 | |
| 1154 | void llvm::computeKnownBitsFromContext(const Value *V, KnownBits &Known, |
| 1155 | const SimplifyQuery &Q, unsigned Depth) { |
| 1156 | // Handle injected condition. |
| 1157 | if (Q.CC && Q.CC->AffectedValues.contains(Ptr: V)) |
| 1158 | computeKnownBitsFromCond(V, Cond: Q.CC->Cond, Known, SQ: Q, Invert: Q.CC->Invert, Depth); |
| 1159 | |
| 1160 | if (!Q.CtxI) |
| 1161 | return; |
| 1162 | |
| 1163 | if (Q.DC && Q.DT) { |
| 1164 | // Handle dominating conditions. |
| 1165 | for (CondBrInst *BI : Q.DC->conditionsFor(V)) { |
| 1166 | BasicBlockEdge Edge0(BI->getParent(), BI->getSuccessor(i: 0)); |
| 1167 | if (Q.DT->dominates(BBE: Edge0, BB: Q.CtxI->getParent())) |
| 1168 | computeKnownBitsFromCond(V, Cond: BI->getCondition(), Known, SQ: Q, |
| 1169 | /*Invert*/ false, Depth); |
| 1170 | |
| 1171 | BasicBlockEdge Edge1(BI->getParent(), BI->getSuccessor(i: 1)); |
| 1172 | if (Q.DT->dominates(BBE: Edge1, BB: Q.CtxI->getParent())) |
| 1173 | computeKnownBitsFromCond(V, Cond: BI->getCondition(), Known, SQ: Q, |
| 1174 | /*Invert*/ true, Depth); |
| 1175 | } |
| 1176 | |
| 1177 | if (Known.hasConflict()) |
| 1178 | Known.resetAll(); |
| 1179 | } |
| 1180 | |
| 1181 | if (!Q.AC) |
| 1182 | return; |
| 1183 | |
| 1184 | unsigned BitWidth = Known.getBitWidth(); |
| 1185 | |
| 1186 | // Note that the patterns below need to be kept in sync with the code |
| 1187 | // in AssumptionCache::updateAffectedValues. |
| 1188 | |
| 1189 | for (AssumptionCache::ResultElem &Elem : Q.AC->assumptionsFor(V)) { |
| 1190 | if (!Elem.Assume) |
| 1191 | continue; |
| 1192 | |
| 1193 | AssumeInst *I = cast<AssumeInst>(Val&: Elem.Assume); |
| 1194 | assert(I->getParent()->getParent() == Q.CtxI->getParent()->getParent() && |
| 1195 | "Got assumption for the wrong function!" ); |
| 1196 | |
| 1197 | if (Elem.Index != AssumptionCache::ExprResultIdx) { |
| 1198 | auto OBU = I->getOperandBundleAt(Index: Elem.Index); |
| 1199 | switch (getBundleAttrFromOBU(OBU)) { |
| 1200 | case BundleAttr::Align: { |
| 1201 | auto [Ptr, _, _2, Alignment, Offset] = getAssumeAlignInfo(OBU); |
| 1202 | if (Ptr == V && Alignment && Offset && isPowerOf2_64(Value: *Alignment) && |
| 1203 | isValidAssumeForContext(I, Q)) { |
| 1204 | Known.Zero |= (*Alignment - 1) & ~*Offset; |
| 1205 | Known.One |= (*Alignment - 1) & *Offset; |
| 1206 | } |
| 1207 | } break; |
| 1208 | |
| 1209 | case BundleAttr::Range: { |
| 1210 | auto [Val, Range] = getAssumeRangeInfo(OBU); |
| 1211 | if (Val == V && isValidAssumeForContext(I, Q)) |
| 1212 | Known = Known.unionWith(RHS: Range.toKnownBits()); |
| 1213 | } break; |
| 1214 | |
| 1215 | default: |
| 1216 | break; |
| 1217 | } |
| 1218 | continue; |
| 1219 | } |
| 1220 | |
| 1221 | // Warning: This loop can end up being somewhat performance sensitive. |
| 1222 | // We're running this loop for once for each value queried resulting in a |
| 1223 | // runtime of ~O(#assumes * #values). |
| 1224 | |
| 1225 | Value *Arg = I->getArgOperand(i: 0); |
| 1226 | |
| 1227 | if (Arg == V && isValidAssumeForContext(I, Q)) { |
| 1228 | assert(BitWidth == 1 && "assume operand is not i1?" ); |
| 1229 | (void)BitWidth; |
| 1230 | Known.setAllOnes(); |
| 1231 | return; |
| 1232 | } |
| 1233 | if (match(V: Arg, P: m_Not(V: m_Specific(V))) && |
| 1234 | isValidAssumeForContext(I, Q)) { |
| 1235 | assert(BitWidth == 1 && "assume operand is not i1?" ); |
| 1236 | (void)BitWidth; |
| 1237 | Known.setAllZero(); |
| 1238 | return; |
| 1239 | } |
| 1240 | auto *Trunc = dyn_cast<TruncInst>(Val: Arg); |
| 1241 | if (Trunc && Trunc->getOperand(i_nocapture: 0) == V && |
| 1242 | isValidAssumeForContext(I, Q)) { |
| 1243 | if (Trunc->hasNoUnsignedWrap()) { |
| 1244 | Known = KnownBits::makeConstant(C: APInt(BitWidth, 1)); |
| 1245 | return; |
| 1246 | } |
| 1247 | Known.One.setBit(0); |
| 1248 | return; |
| 1249 | } |
| 1250 | |
| 1251 | // The remaining tests are all recursive, so bail out if we hit the limit. |
| 1252 | if (Depth == MaxAnalysisRecursionDepth) |
| 1253 | continue; |
| 1254 | |
| 1255 | ICmpInst *Cmp = dyn_cast<ICmpInst>(Val: Arg); |
| 1256 | if (!Cmp) |
| 1257 | continue; |
| 1258 | |
| 1259 | if (!isValidAssumeForContext(I, Q)) |
| 1260 | continue; |
| 1261 | |
| 1262 | computeKnownBitsFromICmpCond(V, Cmp, Known, SQ: Q, /*Invert=*/false); |
| 1263 | } |
| 1264 | |
| 1265 | // Conflicting assumption: Undefined behavior will occur on this execution |
| 1266 | // path. |
| 1267 | if (Known.hasConflict()) |
| 1268 | Known.resetAll(); |
| 1269 | } |
| 1270 | |
| 1271 | /// Compute known bits from a shift operator, including those with a |
| 1272 | /// non-constant shift amount. Known is the output of this function. Known2 is a |
| 1273 | /// pre-allocated temporary with the same bit width as Known and on return |
| 1274 | /// contains the known bit of the shift value source. KF is an |
| 1275 | /// operator-specific function that, given the known-bits and a shift amount, |
| 1276 | /// compute the implied known-bits of the shift operator's result respectively |
| 1277 | /// for that shift amount. The results from calling KF are conservatively |
| 1278 | /// combined for all permitted shift amounts. |
| 1279 | static void computeKnownBitsFromShiftOperator( |
| 1280 | const Operator *I, const APInt &DemandedElts, KnownBits &Known, |
| 1281 | KnownBits &Known2, const SimplifyQuery &Q, unsigned Depth, |
| 1282 | function_ref<KnownBits(const KnownBits &, const KnownBits &, bool)> KF) { |
| 1283 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 1284 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known, Q, Depth: Depth + 1); |
| 1285 | // To limit compile-time impact, only query isKnownNonZero() if we know at |
| 1286 | // least something about the shift amount. |
| 1287 | bool ShAmtNonZero = |
| 1288 | Known.isNonZero() || |
| 1289 | (Known.getMaxValue().ult(RHS: Known.getBitWidth()) && |
| 1290 | isKnownNonZero(V: I->getOperand(i: 1), DemandedElts, Q, Depth: Depth + 1)); |
| 1291 | Known = KF(Known2, Known, ShAmtNonZero); |
| 1292 | } |
| 1293 | |
| 1294 | static KnownBits |
| 1295 | getKnownBitsFromAndXorOr(const Operator *I, const APInt &DemandedElts, |
| 1296 | const KnownBits &KnownLHS, const KnownBits &KnownRHS, |
| 1297 | const SimplifyQuery &Q, unsigned Depth) { |
| 1298 | unsigned BitWidth = KnownLHS.getBitWidth(); |
| 1299 | KnownBits KnownOut(BitWidth); |
| 1300 | bool IsAnd = false; |
| 1301 | bool HasKnownOne = !KnownLHS.One.isZero() || !KnownRHS.One.isZero(); |
| 1302 | Value *X = nullptr, *Y = nullptr; |
| 1303 | |
| 1304 | switch (I->getOpcode()) { |
| 1305 | case Instruction::And: |
| 1306 | KnownOut = KnownLHS & KnownRHS; |
| 1307 | IsAnd = true; |
| 1308 | // and(x, -x) is common idioms that will clear all but lowest set |
| 1309 | // bit. If we have a single known bit in x, we can clear all bits |
| 1310 | // above it. |
| 1311 | // TODO: instcombine often reassociates independent `and` which can hide |
| 1312 | // this pattern. Try to match and(x, and(-x, y)) / and(and(x, y), -x). |
| 1313 | if (HasKnownOne && match(V: I, P: m_c_And(L: m_Value(V&: X), R: m_Neg(V: m_Deferred(V: X))))) { |
| 1314 | // -(-x) == x so using whichever (LHS/RHS) gets us a better result. |
| 1315 | if (KnownLHS.countMaxTrailingZeros() <= KnownRHS.countMaxTrailingZeros()) |
| 1316 | KnownOut = KnownLHS.blsi(); |
| 1317 | else |
| 1318 | KnownOut = KnownRHS.blsi(); |
| 1319 | } |
| 1320 | break; |
| 1321 | case Instruction::Or: |
| 1322 | KnownOut = KnownLHS | KnownRHS; |
| 1323 | break; |
| 1324 | case Instruction::Xor: |
| 1325 | KnownOut = KnownLHS ^ KnownRHS; |
| 1326 | // xor(x, x-1) is common idioms that will clear all but lowest set |
| 1327 | // bit. If we have a single known bit in x, we can clear all bits |
| 1328 | // above it. |
| 1329 | // TODO: xor(x, x-1) is often rewritting as xor(x, x-C) where C != |
| 1330 | // -1 but for the purpose of demanded bits (xor(x, x-C) & |
| 1331 | // Demanded) == (xor(x, x-1) & Demanded). Extend the xor pattern |
| 1332 | // to use arbitrary C if xor(x, x-C) as the same as xor(x, x-1). |
| 1333 | if (HasKnownOne && |
| 1334 | match(V: I, P: m_c_Xor(L: m_Value(V&: X), R: m_Add(L: m_Deferred(V: X), R: m_AllOnes())))) { |
| 1335 | const KnownBits &XBits = I->getOperand(i: 0) == X ? KnownLHS : KnownRHS; |
| 1336 | KnownOut = XBits.blsmsk(); |
| 1337 | } |
| 1338 | break; |
| 1339 | default: |
| 1340 | llvm_unreachable("Invalid Op used in 'analyzeKnownBitsFromAndXorOr'" ); |
| 1341 | } |
| 1342 | |
| 1343 | // and(x, add (x, -1)) is a common idiom that always clears the low bit; |
| 1344 | // xor/or(x, add (x, -1)) is an idiom that will always set the low bit. |
| 1345 | // here we handle the more general case of adding any odd number by |
| 1346 | // matching the form and/xor/or(x, add(x, y)) where y is odd. |
| 1347 | // TODO: This could be generalized to clearing any bit set in y where the |
| 1348 | // following bit is known to be unset in y. |
| 1349 | if (!KnownOut.Zero[0] && !KnownOut.One[0] && |
| 1350 | (match(V: I, P: m_c_BinOp(L: m_Value(V&: X), R: m_c_Add(L: m_Deferred(V: X), R: m_Value(V&: Y)))) || |
| 1351 | match(V: I, P: m_c_BinOp(L: m_Value(V&: X), R: m_Sub(L: m_Deferred(V: X), R: m_Value(V&: Y)))) || |
| 1352 | match(V: I, P: m_c_BinOp(L: m_Value(V&: X), R: m_Sub(L: m_Value(V&: Y), R: m_Deferred(V: X)))))) { |
| 1353 | KnownBits KnownY(BitWidth); |
| 1354 | computeKnownBits(V: Y, DemandedElts, Known&: KnownY, Q, Depth: Depth + 1); |
| 1355 | if (KnownY.countMinTrailingOnes() > 0) { |
| 1356 | if (IsAnd) |
| 1357 | KnownOut.Zero.setBit(0); |
| 1358 | else |
| 1359 | KnownOut.One.setBit(0); |
| 1360 | } |
| 1361 | } |
| 1362 | return KnownOut; |
| 1363 | } |
| 1364 | |
| 1365 | static KnownBits computeKnownBitsForHorizontalOperation( |
| 1366 | const Operator *I, const APInt &DemandedElts, const SimplifyQuery &Q, |
| 1367 | unsigned Depth, |
| 1368 | const function_ref<KnownBits(const KnownBits &, const KnownBits &)> |
| 1369 | KnownBitsFunc) { |
| 1370 | APInt DemandedEltsLHS, DemandedEltsRHS; |
| 1371 | getHorizDemandedEltsForFirstOperand(VectorBitWidth: Q.DL.getTypeSizeInBits(Ty: I->getType()), |
| 1372 | DemandedElts, DemandedLHS&: DemandedEltsLHS, |
| 1373 | DemandedRHS&: DemandedEltsRHS); |
| 1374 | |
| 1375 | const auto ComputeForSingleOpFunc = |
| 1376 | [Depth, &Q, KnownBitsFunc](const Value *Op, APInt &DemandedEltsOp) { |
| 1377 | return KnownBitsFunc( |
| 1378 | computeKnownBits(V: Op, DemandedElts: DemandedEltsOp, Q, Depth: Depth + 1), |
| 1379 | computeKnownBits(V: Op, DemandedElts: DemandedEltsOp << 1, Q, Depth: Depth + 1)); |
| 1380 | }; |
| 1381 | |
| 1382 | if (DemandedEltsRHS.isZero()) |
| 1383 | return ComputeForSingleOpFunc(I->getOperand(i: 0), DemandedEltsLHS); |
| 1384 | if (DemandedEltsLHS.isZero()) |
| 1385 | return ComputeForSingleOpFunc(I->getOperand(i: 1), DemandedEltsRHS); |
| 1386 | |
| 1387 | return ComputeForSingleOpFunc(I->getOperand(i: 0), DemandedEltsLHS) |
| 1388 | .intersectWith(RHS: ComputeForSingleOpFunc(I->getOperand(i: 1), DemandedEltsRHS)); |
| 1389 | } |
| 1390 | |
| 1391 | // Public so this can be used in `SimplifyDemandedUseBits`. |
| 1392 | KnownBits llvm::analyzeKnownBitsFromAndXorOr(const Operator *I, |
| 1393 | const KnownBits &KnownLHS, |
| 1394 | const KnownBits &KnownRHS, |
| 1395 | const SimplifyQuery &SQ, |
| 1396 | unsigned Depth) { |
| 1397 | auto *FVTy = dyn_cast<FixedVectorType>(Val: I->getType()); |
| 1398 | APInt DemandedElts = |
| 1399 | FVTy ? APInt::getAllOnes(numBits: FVTy->getNumElements()) : APInt(1, 1); |
| 1400 | |
| 1401 | return getKnownBitsFromAndXorOr(I, DemandedElts, KnownLHS, KnownRHS, Q: SQ, |
| 1402 | Depth); |
| 1403 | } |
| 1404 | |
| 1405 | ConstantRange llvm::getVScaleRange(const Function *F, unsigned BitWidth) { |
| 1406 | Attribute Attr = F->getFnAttribute(Kind: Attribute::VScaleRange); |
| 1407 | // Without vscale_range, we only know that vscale is non-zero. |
| 1408 | if (!Attr.isValid()) |
| 1409 | return ConstantRange(APInt(BitWidth, 1), APInt::getZero(numBits: BitWidth)); |
| 1410 | |
| 1411 | unsigned AttrMin = Attr.getVScaleRangeMin(); |
| 1412 | // Minimum is larger than vscale width, result is always poison. |
| 1413 | if ((unsigned)llvm::bit_width(Value: AttrMin) > BitWidth) |
| 1414 | return ConstantRange::getEmpty(BitWidth); |
| 1415 | |
| 1416 | APInt Min(BitWidth, AttrMin); |
| 1417 | std::optional<unsigned> AttrMax = Attr.getVScaleRangeMax(); |
| 1418 | if (!AttrMax || (unsigned)llvm::bit_width(Value: *AttrMax) > BitWidth) |
| 1419 | return ConstantRange(Min, APInt::getZero(numBits: BitWidth)); |
| 1420 | |
| 1421 | return ConstantRange(Min, APInt(BitWidth, *AttrMax) + 1); |
| 1422 | } |
| 1423 | |
| 1424 | /// Return true if \p II reads a register named "vlenb". On RISC-V this is the |
| 1425 | /// VLENB CSR, which holds VLEN/8: a non-zero power of two bounded by the |
| 1426 | /// target's VLEN range. Callers must ensure the target is RISC-V. |
| 1427 | static bool isReadVLENB(const IntrinsicInst &II) { |
| 1428 | auto *MAV = dyn_cast<MetadataAsValue>(Val: II.getArgOperand(i: 0)); |
| 1429 | if (!MAV) |
| 1430 | return false; |
| 1431 | auto *MD = dyn_cast<MDNode>(Val: MAV->getMetadata()); |
| 1432 | if (!MD || MD->getNumOperands() != 1) |
| 1433 | return false; |
| 1434 | auto *RegName = dyn_cast<MDString>(Val: MD->getOperand(I: 0)); |
| 1435 | return RegName && RegName->getString() == "vlenb" ; |
| 1436 | } |
| 1437 | |
| 1438 | /// Return the value range of a RISC-V vlenb CSR read. RVV requires VLEN to be a |
| 1439 | /// power of two in [32, 65536] (Zvl32b is the smallest vector extension), so |
| 1440 | /// VLENB = VLEN/8 is in [4, 8192]. This architectural bound is independent of |
| 1441 | /// any function attribute and stays sound for Zvl32b, whose VLEN (32) is not |
| 1442 | /// representable as an integer vscale (VLEN / RVVBitsPerBlock). A vscale_range |
| 1443 | /// attribute, when present, pins the subtarget's VLEN in units of |
| 1444 | /// RVVBitsPerBlock (64 bits) and so gives a tighter VLENB = vscale * |
| 1445 | /// RVVBytesPerBlock. |
| 1446 | static ConstantRange getRISCVVLENBRange(const IntrinsicInst &II, |
| 1447 | unsigned Width) { |
| 1448 | // Architectural bounds: VLEN in [32, 65536] => VLENB in [4, 8192]. |
| 1449 | ConstantRange Range(APInt(Width, 32 / 8), APInt(Width, 65536 / 8) + 1); |
| 1450 | |
| 1451 | const Function *F = II.getFunction(); |
| 1452 | if (F->getFnAttribute(Kind: Attribute::VScaleRange).isValid()) { |
| 1453 | ConstantRange VScale = getVScaleRange(F, BitWidth: Width); |
| 1454 | Range = Range.intersectWith( |
| 1455 | CR: VScale.multiply(Other: ConstantRange(APInt(Width, RISCV::RVVBytesPerBlock)))); |
| 1456 | } |
| 1457 | return Range; |
| 1458 | } |
| 1459 | |
| 1460 | void llvm::adjustKnownBitsForSelectArm(KnownBits &Known, Value *Cond, |
| 1461 | Value *Arm, bool Invert, |
| 1462 | const SimplifyQuery &Q, unsigned Depth) { |
| 1463 | // If we have a constant arm, we are done. |
| 1464 | if (Known.isConstant()) |
| 1465 | return; |
| 1466 | |
| 1467 | // See what condition implies about the bits of the select arm. |
| 1468 | KnownBits CondRes(Known.getBitWidth()); |
| 1469 | computeKnownBitsFromCond(V: Arm, Cond, Known&: CondRes, SQ: Q, Invert, Depth: Depth + 1); |
| 1470 | // If we don't get any information from the condition, no reason to |
| 1471 | // proceed. |
| 1472 | if (CondRes.isUnknown()) |
| 1473 | return; |
| 1474 | |
| 1475 | // We can have conflict if the condition is dead. I.e if we have |
| 1476 | // (x | 64) < 32 ? (x | 64) : y |
| 1477 | // we will have conflict at bit 6 from the condition/the `or`. |
| 1478 | // In that case just return. Its not particularly important |
| 1479 | // what we do, as this select is going to be simplified soon. |
| 1480 | CondRes = CondRes.unionWith(RHS: Known); |
| 1481 | if (CondRes.hasConflict()) |
| 1482 | return; |
| 1483 | |
| 1484 | // Finally make sure the information we found is valid. This is relatively |
| 1485 | // expensive so it's left for the very end. |
| 1486 | if (!isGuaranteedNotToBeUndef(V: Arm, AC: Q.AC, CtxI: Q.CtxI, DT: Q.DT, Depth: Depth + 1)) |
| 1487 | return; |
| 1488 | |
| 1489 | // Finally, we know we get information from the condition and its valid, |
| 1490 | // so return it. |
| 1491 | Known = std::move(CondRes); |
| 1492 | } |
| 1493 | |
| 1494 | // Match a signed min+max clamp pattern like smax(smin(In, CHigh), CLow). |
| 1495 | // Returns the input and lower/upper bounds. |
| 1496 | static bool isSignedMinMaxClamp(const Value *Select, const Value *&In, |
| 1497 | const APInt *&CLow, const APInt *&CHigh) { |
| 1498 | assert(isa<Operator>(Select) && |
| 1499 | cast<Operator>(Select)->getOpcode() == Instruction::Select && |
| 1500 | "Input should be a Select!" ); |
| 1501 | |
| 1502 | const Value *LHS = nullptr, *RHS = nullptr; |
| 1503 | SelectPatternFlavor SPF = matchSelectPattern(V: Select, LHS, RHS).Flavor; |
| 1504 | if (SPF != SPF_SMAX && SPF != SPF_SMIN) |
| 1505 | return false; |
| 1506 | |
| 1507 | if (!match(V: RHS, P: m_APInt(Res&: CLow))) |
| 1508 | return false; |
| 1509 | |
| 1510 | const Value *LHS2 = nullptr, *RHS2 = nullptr; |
| 1511 | SelectPatternFlavor SPF2 = matchSelectPattern(V: LHS, LHS&: LHS2, RHS&: RHS2).Flavor; |
| 1512 | if (getInverseMinMaxFlavor(SPF) != SPF2) |
| 1513 | return false; |
| 1514 | |
| 1515 | if (!match(V: RHS2, P: m_APInt(Res&: CHigh))) |
| 1516 | return false; |
| 1517 | |
| 1518 | if (SPF == SPF_SMIN) |
| 1519 | std::swap(a&: CLow, b&: CHigh); |
| 1520 | |
| 1521 | In = LHS2; |
| 1522 | return CLow->sle(RHS: *CHigh); |
| 1523 | } |
| 1524 | |
| 1525 | static bool isSignedMinMaxIntrinsicClamp(const IntrinsicInst *II, |
| 1526 | const APInt *&CLow, |
| 1527 | const APInt *&CHigh) { |
| 1528 | assert((II->getIntrinsicID() == Intrinsic::smin || |
| 1529 | II->getIntrinsicID() == Intrinsic::smax) && |
| 1530 | "Must be smin/smax" ); |
| 1531 | |
| 1532 | Intrinsic::ID InverseID = getInverseMinMaxIntrinsic(MinMaxID: II->getIntrinsicID()); |
| 1533 | auto *InnerII = dyn_cast<IntrinsicInst>(Val: II->getArgOperand(i: 0)); |
| 1534 | if (!InnerII || InnerII->getIntrinsicID() != InverseID || |
| 1535 | !match(V: II->getArgOperand(i: 1), P: m_APInt(Res&: CLow)) || |
| 1536 | !match(V: InnerII->getArgOperand(i: 1), P: m_APInt(Res&: CHigh))) |
| 1537 | return false; |
| 1538 | |
| 1539 | if (II->getIntrinsicID() == Intrinsic::smin) |
| 1540 | std::swap(a&: CLow, b&: CHigh); |
| 1541 | return CLow->sle(RHS: *CHigh); |
| 1542 | } |
| 1543 | |
| 1544 | static void unionWithMinMaxIntrinsicClamp(const IntrinsicInst *II, |
| 1545 | KnownBits &Known) { |
| 1546 | const APInt *CLow, *CHigh; |
| 1547 | if (isSignedMinMaxIntrinsicClamp(II, CLow, CHigh)) |
| 1548 | Known = Known.unionWith( |
| 1549 | RHS: ConstantRange::getNonEmpty(Lower: *CLow, Upper: *CHigh + 1).toKnownBits()); |
| 1550 | } |
| 1551 | |
| 1552 | static void computeKnownBitsForRecurrenceOperands( |
| 1553 | const PHINode *P, Value *Start, Value *Step, const APInt &DemandedElts, |
| 1554 | KnownBits &KnownStart, KnownBits &KnownStep, const SimplifyQuery &Q, |
| 1555 | unsigned Depth) { |
| 1556 | // Change the context instruction to the "edge" that flows into the phi. This |
| 1557 | // is important because that is where the value is actually "evaluated" even |
| 1558 | // though it is used later somewhere else. (see also D69571). |
| 1559 | SimplifyQuery RecQ = Q.getWithoutCondContext(); |
| 1560 | unsigned OpNum = P->getOperand(i_nocapture: 0) == Start ? 0 : 1; |
| 1561 | |
| 1562 | RecQ.CtxI = P->getIncomingBlock(i: OpNum)->getTerminator(); |
| 1563 | computeKnownBits(V: Start, DemandedElts, Known&: KnownStart, Q: RecQ, Depth: Depth + 1); |
| 1564 | |
| 1565 | RecQ.CtxI = P->getIncomingBlock(i: 1 - OpNum)->getTerminator(); |
| 1566 | computeKnownBits(V: Step, DemandedElts, Known&: KnownStep, Q: RecQ, Depth: Depth + 1); |
| 1567 | } |
| 1568 | |
| 1569 | static void computeKnownBitsFromOperator(const Operator *I, |
| 1570 | const APInt &DemandedElts, |
| 1571 | KnownBits &Known, |
| 1572 | const SimplifyQuery &Q, |
| 1573 | unsigned Depth) { |
| 1574 | unsigned BitWidth = Known.getBitWidth(); |
| 1575 | |
| 1576 | KnownBits Known2(BitWidth); |
| 1577 | switch (I->getOpcode()) { |
| 1578 | default: break; |
| 1579 | case Instruction::Load: |
| 1580 | if (MDNode *MD = |
| 1581 | Q.IIQ.getMetadata(I: cast<LoadInst>(Val: I), KindID: LLVMContext::MD_range)) |
| 1582 | computeKnownBitsFromRangeMetadata(Ranges: *MD, Known); |
| 1583 | break; |
| 1584 | case Instruction::And: |
| 1585 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known, Q, Depth: Depth + 1); |
| 1586 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 1587 | |
| 1588 | Known = getKnownBitsFromAndXorOr(I, DemandedElts, KnownLHS: Known2, KnownRHS: Known, Q, Depth); |
| 1589 | break; |
| 1590 | case Instruction::Or: |
| 1591 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known, Q, Depth: Depth + 1); |
| 1592 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 1593 | |
| 1594 | Known = getKnownBitsFromAndXorOr(I, DemandedElts, KnownLHS: Known2, KnownRHS: Known, Q, Depth); |
| 1595 | break; |
| 1596 | case Instruction::Xor: |
| 1597 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known, Q, Depth: Depth + 1); |
| 1598 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 1599 | |
| 1600 | Known = getKnownBitsFromAndXorOr(I, DemandedElts, KnownLHS: Known2, KnownRHS: Known, Q, Depth); |
| 1601 | break; |
| 1602 | case Instruction::Mul: { |
| 1603 | bool NSW = Q.IIQ.hasNoSignedWrap(Op: cast<OverflowingBinaryOperator>(Val: I)); |
| 1604 | bool NUW = Q.IIQ.hasNoUnsignedWrap(Op: cast<OverflowingBinaryOperator>(Val: I)); |
| 1605 | computeKnownBitsMul(Op0: I->getOperand(i: 0), Op1: I->getOperand(i: 1), NSW, NUW, |
| 1606 | DemandedElts, Known, Known2, Q, Depth); |
| 1607 | break; |
| 1608 | } |
| 1609 | case Instruction::UDiv: { |
| 1610 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 1611 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 1612 | Known = |
| 1613 | KnownBits::udiv(LHS: Known, RHS: Known2, Exact: Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I))); |
| 1614 | break; |
| 1615 | } |
| 1616 | case Instruction::SDiv: { |
| 1617 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 1618 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 1619 | Known = |
| 1620 | KnownBits::sdiv(LHS: Known, RHS: Known2, Exact: Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I))); |
| 1621 | break; |
| 1622 | } |
| 1623 | case Instruction::Select: { |
| 1624 | auto ComputeForArm = [&](Value *Arm, bool Invert) { |
| 1625 | KnownBits Res(Known.getBitWidth()); |
| 1626 | computeKnownBits(V: Arm, DemandedElts, Known&: Res, Q, Depth: Depth + 1); |
| 1627 | adjustKnownBitsForSelectArm(Known&: Res, Cond: I->getOperand(i: 0), Arm, Invert, Q, Depth); |
| 1628 | return Res; |
| 1629 | }; |
| 1630 | // Only known if known in both the LHS and RHS. |
| 1631 | Known = |
| 1632 | ComputeForArm(I->getOperand(i: 1), /*Invert=*/false) |
| 1633 | .intersectWith(RHS: ComputeForArm(I->getOperand(i: 2), /*Invert=*/true)); |
| 1634 | break; |
| 1635 | } |
| 1636 | case Instruction::FPToSI: { |
| 1637 | // fptosi is poison if the rounded value doesn't fit in the result type, |
| 1638 | // so we can assume the conversion is well-defined and rounds towards |
| 1639 | // zero. +-Inf can never fit in an integer type, so it is always poison, |
| 1640 | // like NaN. Negative subnormals and negative zero round to 0. That |
| 1641 | // leaves negative normals as the only class that can produce a defined |
| 1642 | // negative result. |
| 1643 | KnownFPClass SrcFPClass = computeKnownFPClass( |
| 1644 | V: I->getOperand(i: 0), DemandedElts, InterestedClasses: fcNegNormal, SQ: Q, Depth: Depth + 1); |
| 1645 | if (SrcFPClass.isKnownNever(Mask: fcNegNormal)) |
| 1646 | Known.makeNonNegative(); |
| 1647 | break; |
| 1648 | } |
| 1649 | case Instruction::FPTrunc: |
| 1650 | case Instruction::FPExt: |
| 1651 | case Instruction::FPToUI: |
| 1652 | case Instruction::SIToFP: |
| 1653 | case Instruction::UIToFP: |
| 1654 | break; // Can't work with floating point. |
| 1655 | case Instruction::PtrToInt: |
| 1656 | case Instruction::PtrToAddr: |
| 1657 | case Instruction::IntToPtr: |
| 1658 | // Fall through and handle them the same as zext/trunc. |
| 1659 | [[fallthrough]]; |
| 1660 | case Instruction::ZExt: |
| 1661 | case Instruction::Trunc: { |
| 1662 | Type *SrcTy = I->getOperand(i: 0)->getType(); |
| 1663 | |
| 1664 | unsigned SrcBitWidth; |
| 1665 | // Note that we handle pointer operands here because of inttoptr/ptrtoint |
| 1666 | // which fall through here. |
| 1667 | Type *ScalarTy = SrcTy->getScalarType(); |
| 1668 | SrcBitWidth = ScalarTy->isPointerTy() ? |
| 1669 | Q.DL.getPointerTypeSizeInBits(ScalarTy) : |
| 1670 | Q.DL.getTypeSizeInBits(Ty: ScalarTy); |
| 1671 | |
| 1672 | assert(SrcBitWidth && "SrcBitWidth can't be zero" ); |
| 1673 | Known = Known.anyextOrTrunc(BitWidth: SrcBitWidth); |
| 1674 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 1675 | if (auto *Inst = dyn_cast<PossiblyNonNegInst>(Val: I); |
| 1676 | Inst && Inst->hasNonNeg() && !Known.isNegative()) |
| 1677 | Known.makeNonNegative(); |
| 1678 | Known = Known.zextOrTrunc(BitWidth); |
| 1679 | break; |
| 1680 | } |
| 1681 | case Instruction::BitCast: { |
| 1682 | Type *SrcTy = I->getOperand(i: 0)->getType(); |
| 1683 | if (SrcTy->isIntOrPtrTy() && |
| 1684 | // TODO: For now, not handling conversions like: |
| 1685 | // (bitcast i64 %x to <2 x i32>) |
| 1686 | !I->getType()->isVectorTy()) { |
| 1687 | computeKnownBits(V: I->getOperand(i: 0), Known, Q, Depth: Depth + 1); |
| 1688 | break; |
| 1689 | } |
| 1690 | |
| 1691 | const Value *V; |
| 1692 | // Handle bitcast from floating point to integer. |
| 1693 | if (match(V: I, P: m_ElementWiseBitCast(Op: m_Value(V))) && |
| 1694 | V->getType()->isFPOrFPVectorTy()) { |
| 1695 | Type *FPType = V->getType()->getScalarType(); |
| 1696 | KnownFPClass Result = |
| 1697 | computeKnownFPClass(V, DemandedElts, InterestedClasses: fcAllFlags, SQ: Q, Depth: Depth + 1); |
| 1698 | |
| 1699 | Known = Result.toKnownBits(FltSemantics: FPType->getFltSemantics()); |
| 1700 | |
| 1701 | break; |
| 1702 | } |
| 1703 | |
| 1704 | // Handle cast from vector integer type to scalar or vector integer. |
| 1705 | auto *SrcVecTy = dyn_cast<FixedVectorType>(Val: SrcTy); |
| 1706 | if (!SrcVecTy || !SrcVecTy->getElementType()->isIntegerTy() || |
| 1707 | !I->getType()->isIntOrIntVectorTy() || |
| 1708 | isa<ScalableVectorType>(Val: I->getType())) |
| 1709 | break; |
| 1710 | |
| 1711 | unsigned NumElts = DemandedElts.getBitWidth(); |
| 1712 | bool IsLE = Q.DL.isLittleEndian(); |
| 1713 | // Look through a cast from narrow vector elements to wider type. |
| 1714 | // Examples: v4i32 -> v2i64, v3i8 -> v24 |
| 1715 | unsigned SubBitWidth = SrcVecTy->getScalarSizeInBits(); |
| 1716 | if (BitWidth % SubBitWidth == 0) { |
| 1717 | // Known bits are automatically intersected across demanded elements of a |
| 1718 | // vector. So for example, if a bit is computed as known zero, it must be |
| 1719 | // zero across all demanded elements of the vector. |
| 1720 | // |
| 1721 | // For this bitcast, each demanded element of the output is sub-divided |
| 1722 | // across a set of smaller vector elements in the source vector. To get |
| 1723 | // the known bits for an entire element of the output, compute the known |
| 1724 | // bits for each sub-element sequentially. This is done by shifting the |
| 1725 | // one-set-bit demanded elements parameter across the sub-elements for |
| 1726 | // consecutive calls to computeKnownBits. We are using the demanded |
| 1727 | // elements parameter as a mask operator. |
| 1728 | // |
| 1729 | // The known bits of each sub-element are then inserted into place |
| 1730 | // (dependent on endian) to form the full result of known bits. |
| 1731 | unsigned SubScale = BitWidth / SubBitWidth; |
| 1732 | APInt SubDemandedElts = APInt::getZero(numBits: NumElts * SubScale); |
| 1733 | for (unsigned i = 0; i != NumElts; ++i) { |
| 1734 | if (DemandedElts[i]) |
| 1735 | SubDemandedElts.setBit(i * SubScale); |
| 1736 | } |
| 1737 | |
| 1738 | KnownBits KnownSrc(SubBitWidth); |
| 1739 | for (unsigned i = 0; i != SubScale; ++i) { |
| 1740 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts: SubDemandedElts.shl(shiftAmt: i), Known&: KnownSrc, Q, |
| 1741 | Depth: Depth + 1); |
| 1742 | unsigned ShiftElt = IsLE ? i : SubScale - 1 - i; |
| 1743 | Known.insertBits(SubBits: KnownSrc, BitPosition: ShiftElt * SubBitWidth); |
| 1744 | } |
| 1745 | } |
| 1746 | // Look through a cast from wider vector elements to narrow type. |
| 1747 | // Examples: v2i64 -> v4i32 |
| 1748 | if (SubBitWidth % BitWidth == 0) { |
| 1749 | unsigned SubScale = SubBitWidth / BitWidth; |
| 1750 | KnownBits KnownSrc(SubBitWidth); |
| 1751 | APInt SubDemandedElts = |
| 1752 | APIntOps::ScaleBitMask(A: DemandedElts, NewBitWidth: NumElts / SubScale); |
| 1753 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts: SubDemandedElts, Known&: KnownSrc, Q, |
| 1754 | Depth: Depth + 1); |
| 1755 | |
| 1756 | Known.setAllConflict(); |
| 1757 | for (unsigned i = 0; i != NumElts; ++i) { |
| 1758 | if (DemandedElts[i]) { |
| 1759 | unsigned Shifts = IsLE ? i : NumElts - 1 - i; |
| 1760 | unsigned Offset = (Shifts % SubScale) * BitWidth; |
| 1761 | Known = Known.intersectWith(RHS: KnownSrc.extractBits(NumBits: BitWidth, BitPosition: Offset)); |
| 1762 | if (Known.isUnknown()) |
| 1763 | break; |
| 1764 | } |
| 1765 | } |
| 1766 | } |
| 1767 | break; |
| 1768 | } |
| 1769 | case Instruction::SExt: { |
| 1770 | // Compute the bits in the result that are not present in the input. |
| 1771 | unsigned SrcBitWidth = I->getOperand(i: 0)->getType()->getScalarSizeInBits(); |
| 1772 | |
| 1773 | Known = Known.trunc(BitWidth: SrcBitWidth); |
| 1774 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 1775 | // If the sign bit of the input is known set or clear, then we know the |
| 1776 | // top bits of the result. |
| 1777 | Known = Known.sext(BitWidth); |
| 1778 | break; |
| 1779 | } |
| 1780 | case Instruction::Shl: { |
| 1781 | bool NUW = Q.IIQ.hasNoUnsignedWrap(Op: cast<OverflowingBinaryOperator>(Val: I)); |
| 1782 | bool NSW = Q.IIQ.hasNoSignedWrap(Op: cast<OverflowingBinaryOperator>(Val: I)); |
| 1783 | auto KF = [NUW, NSW](const KnownBits &KnownVal, const KnownBits &KnownAmt, |
| 1784 | bool ShAmtNonZero) { |
| 1785 | return KnownBits::shl(LHS: KnownVal, RHS: KnownAmt, NUW, NSW, ShAmtNonZero); |
| 1786 | }; |
| 1787 | computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Q, Depth, |
| 1788 | KF); |
| 1789 | // X << X is even: X is either zero or shifts by at least one bit. Both |
| 1790 | // uses must observe the same value, so X must not be undef. |
| 1791 | if (I->getOperand(i: 0) == I->getOperand(i: 1) && |
| 1792 | isGuaranteedNotToBeUndef(V: I->getOperand(i: 0), AC: Q.AC, CtxI: Q.CtxI, DT: Q.DT, |
| 1793 | Depth: Depth + 1)) |
| 1794 | Known.Zero.setBit(0); |
| 1795 | |
| 1796 | // Trailing zeros of a right-shifted constant never decrease. |
| 1797 | const APInt *C; |
| 1798 | if (match(V: I->getOperand(i: 0), P: m_APInt(Res&: C))) |
| 1799 | Known.Zero.setLowBits(C->countr_zero()); |
| 1800 | |
| 1801 | // shl X, sub(Y, xor(ctlz(X, true), BitWidth-1)) shifts X so that its MSB |
| 1802 | // lands at bit Y, when BitWidth is a power of 2. |
| 1803 | const APInt *YC; |
| 1804 | Value *X = I->getOperand(i: 0); |
| 1805 | if (isPowerOf2_32(Value: BitWidth) && |
| 1806 | match(V: I->getOperand(i: 1), |
| 1807 | P: m_Sub(L: m_APInt(Res&: YC), R: m_Xor(L: m_Ctlz(Op0: m_Specific(V: X), Op1: m_One()), |
| 1808 | R: m_SpecificInt(V: BitWidth - 1)))) && |
| 1809 | YC->ult(RHS: BitWidth - 1)) { |
| 1810 | unsigned Y = YC->getZExtValue(); |
| 1811 | Known.One.setBit(Y); |
| 1812 | Known.Zero.setBitsFrom(Y + 1); |
| 1813 | } |
| 1814 | break; |
| 1815 | } |
| 1816 | case Instruction::LShr: { |
| 1817 | bool Exact = Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I)); |
| 1818 | auto KF = [Exact](const KnownBits &KnownVal, const KnownBits &KnownAmt, |
| 1819 | bool ShAmtNonZero) { |
| 1820 | return KnownBits::lshr(LHS: KnownVal, RHS: KnownAmt, ShAmtNonZero, Exact); |
| 1821 | }; |
| 1822 | computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Q, Depth, |
| 1823 | KF); |
| 1824 | // Leading zeros of a left-shifted constant never decrease. |
| 1825 | const APInt *C; |
| 1826 | if (match(V: I->getOperand(i: 0), P: m_APInt(Res&: C))) |
| 1827 | Known.Zero.setHighBits(C->countl_zero()); |
| 1828 | break; |
| 1829 | } |
| 1830 | case Instruction::AShr: { |
| 1831 | bool Exact = Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I)); |
| 1832 | auto KF = [Exact](const KnownBits &KnownVal, const KnownBits &KnownAmt, |
| 1833 | bool ShAmtNonZero) { |
| 1834 | return KnownBits::ashr(LHS: KnownVal, RHS: KnownAmt, ShAmtNonZero, Exact); |
| 1835 | }; |
| 1836 | computeKnownBitsFromShiftOperator(I, DemandedElts, Known, Known2, Q, Depth, |
| 1837 | KF); |
| 1838 | break; |
| 1839 | } |
| 1840 | case Instruction::Sub: { |
| 1841 | bool NSW = Q.IIQ.hasNoSignedWrap(Op: cast<OverflowingBinaryOperator>(Val: I)); |
| 1842 | bool NUW = Q.IIQ.hasNoUnsignedWrap(Op: cast<OverflowingBinaryOperator>(Val: I)); |
| 1843 | computeKnownBitsAddSub(Add: false, Op0: I->getOperand(i: 0), Op1: I->getOperand(i: 1), NSW, NUW, |
| 1844 | DemandedElts, KnownOut&: Known, Known2, Q, Depth); |
| 1845 | break; |
| 1846 | } |
| 1847 | case Instruction::Add: { |
| 1848 | bool NSW = Q.IIQ.hasNoSignedWrap(Op: cast<OverflowingBinaryOperator>(Val: I)); |
| 1849 | bool NUW = Q.IIQ.hasNoUnsignedWrap(Op: cast<OverflowingBinaryOperator>(Val: I)); |
| 1850 | computeKnownBitsAddSub(Add: true, Op0: I->getOperand(i: 0), Op1: I->getOperand(i: 1), NSW, NUW, |
| 1851 | DemandedElts, KnownOut&: Known, Known2, Q, Depth); |
| 1852 | break; |
| 1853 | } |
| 1854 | case Instruction::SRem: |
| 1855 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 1856 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 1857 | Known = KnownBits::srem(LHS: Known, RHS: Known2); |
| 1858 | break; |
| 1859 | |
| 1860 | case Instruction::URem: |
| 1861 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 1862 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 1863 | Known = KnownBits::urem(LHS: Known, RHS: Known2); |
| 1864 | break; |
| 1865 | case Instruction::Alloca: |
| 1866 | Known.Zero.setLowBits(Log2(A: cast<AllocaInst>(Val: I)->getAlign())); |
| 1867 | break; |
| 1868 | case Instruction::GetElementPtr: { |
| 1869 | // Analyze all of the subscripts of this getelementptr instruction |
| 1870 | // to determine if we can prove known low zero bits. |
| 1871 | computeKnownBits(V: I->getOperand(i: 0), Known, Q, Depth: Depth + 1); |
| 1872 | // Accumulate the constant indices in a separate variable |
| 1873 | // to minimize the number of calls to computeForAddSub. |
| 1874 | unsigned IndexWidth = Q.DL.getIndexTypeSizeInBits(Ty: I->getType()); |
| 1875 | APInt AccConstIndices(IndexWidth, 0); |
| 1876 | |
| 1877 | auto AddIndexToKnown = [&](KnownBits IndexBits) { |
| 1878 | if (IndexWidth == BitWidth) { |
| 1879 | // Note that inbounds does *not* guarantee nsw for the addition, as only |
| 1880 | // the offset is signed, while the base address is unsigned. |
| 1881 | Known = KnownBits::add(LHS: Known, RHS: IndexBits); |
| 1882 | } else { |
| 1883 | // If the index width is smaller than the pointer width, only add the |
| 1884 | // value to the low bits. |
| 1885 | assert(IndexWidth < BitWidth && |
| 1886 | "Index width can't be larger than pointer width" ); |
| 1887 | Known.insertBits(SubBits: KnownBits::add(LHS: Known.trunc(BitWidth: IndexWidth), RHS: IndexBits), BitPosition: 0); |
| 1888 | } |
| 1889 | }; |
| 1890 | |
| 1891 | gep_type_iterator GTI = gep_type_begin(GEP: I); |
| 1892 | for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) { |
| 1893 | // TrailZ can only become smaller, short-circuit if we hit zero. |
| 1894 | if (Known.isUnknown()) |
| 1895 | break; |
| 1896 | |
| 1897 | Value *Index = I->getOperand(i); |
| 1898 | |
| 1899 | // Handle case when index is zero. |
| 1900 | Constant *CIndex = dyn_cast<Constant>(Val: Index); |
| 1901 | if (CIndex && CIndex->isNullValue()) |
| 1902 | continue; |
| 1903 | |
| 1904 | if (StructType *STy = GTI.getStructTypeOrNull()) { |
| 1905 | // Handle struct member offset arithmetic. |
| 1906 | |
| 1907 | assert(CIndex && |
| 1908 | "Access to structure field must be known at compile time" ); |
| 1909 | |
| 1910 | if (CIndex->getType()->isVectorTy()) |
| 1911 | Index = CIndex->getSplatValue(); |
| 1912 | |
| 1913 | unsigned Idx = cast<ConstantInt>(Val: Index)->getZExtValue(); |
| 1914 | const StructLayout *SL = Q.DL.getStructLayout(Ty: STy); |
| 1915 | uint64_t Offset = SL->getElementOffset(Idx); |
| 1916 | AccConstIndices += Offset; |
| 1917 | continue; |
| 1918 | } |
| 1919 | |
| 1920 | // Handle array index arithmetic. |
| 1921 | Type *IndexedTy = GTI.getIndexedType(); |
| 1922 | if (!IndexedTy->isSized()) { |
| 1923 | Known.resetAll(); |
| 1924 | break; |
| 1925 | } |
| 1926 | |
| 1927 | TypeSize Stride = GTI.getSequentialElementStride(DL: Q.DL); |
| 1928 | uint64_t StrideInBytes = Stride.getKnownMinValue(); |
| 1929 | if (!Stride.isScalable()) { |
| 1930 | // Fast path for constant offset. |
| 1931 | if (auto *CI = dyn_cast<ConstantInt>(Val: Index)) { |
| 1932 | AccConstIndices += |
| 1933 | CI->getValue().sextOrTrunc(width: IndexWidth) * StrideInBytes; |
| 1934 | continue; |
| 1935 | } |
| 1936 | } |
| 1937 | |
| 1938 | KnownBits IndexBits = |
| 1939 | computeKnownBits(V: Index, Q, Depth: Depth + 1).sextOrTrunc(BitWidth: IndexWidth); |
| 1940 | KnownBits ScalingFactor(IndexWidth); |
| 1941 | // Multiply by current sizeof type. |
| 1942 | // &A[i] == A + i * sizeof(*A[i]). |
| 1943 | if (Stride.isScalable()) { |
| 1944 | // For scalable types the only thing we know about sizeof is |
| 1945 | // that this is a multiple of the minimum size. |
| 1946 | ScalingFactor.Zero.setLowBits(llvm::countr_zero(Val: StrideInBytes)); |
| 1947 | } else { |
| 1948 | ScalingFactor = |
| 1949 | KnownBits::makeConstant(C: APInt(IndexWidth, StrideInBytes)); |
| 1950 | } |
| 1951 | AddIndexToKnown(KnownBits::mul(LHS: IndexBits, RHS: ScalingFactor)); |
| 1952 | } |
| 1953 | if (!Known.isUnknown() && !AccConstIndices.isZero()) |
| 1954 | AddIndexToKnown(KnownBits::makeConstant(C: AccConstIndices)); |
| 1955 | break; |
| 1956 | } |
| 1957 | case Instruction::PHI: { |
| 1958 | const PHINode *P = cast<PHINode>(Val: I); |
| 1959 | BinaryOperator *BO = nullptr; |
| 1960 | Value *Start = nullptr, *Step = nullptr; |
| 1961 | KnownBits &KnownStart = Known2; |
| 1962 | if (matchSimpleRecurrence(P, BO, Start, Step)) { |
| 1963 | // Handle the case of a simple two-predecessor recurrence PHI. |
| 1964 | // There's a lot more that could theoretically be done here, but |
| 1965 | // this is sufficient to catch some interesting cases. |
| 1966 | unsigned Opcode = BO->getOpcode(); |
| 1967 | |
| 1968 | switch (Opcode) { |
| 1969 | // If this is a shift recurrence, we know the bits being shifted in. We |
| 1970 | // can combine that with information about the start value of the |
| 1971 | // recurrence to conclude facts about the result. If this is a udiv |
| 1972 | // recurrence, we know that the result can never exceed either the |
| 1973 | // numerator or the start value, whichever is greater. |
| 1974 | case Instruction::LShr: |
| 1975 | case Instruction::AShr: |
| 1976 | case Instruction::Shl: |
| 1977 | case Instruction::UDiv: |
| 1978 | if (BO->getOperand(i_nocapture: 0) != I) |
| 1979 | break; |
| 1980 | [[fallthrough]]; |
| 1981 | |
| 1982 | // For a urem recurrence, the result can never exceed the start value. The |
| 1983 | // phi could either be the numerator or the denominator. |
| 1984 | case Instruction::URem: { |
| 1985 | // We have matched a recurrence of the form: |
| 1986 | // %iv = [R, %entry], [%iv.next, %backedge] |
| 1987 | // %iv.next = shift_op %iv, L |
| 1988 | |
| 1989 | // Recurse with the phi context to avoid concern about whether facts |
| 1990 | // inferred hold at original context instruction. TODO: It may be |
| 1991 | // correct to use the original context. IF warranted, explore and |
| 1992 | // add sufficient tests to cover. |
| 1993 | SimplifyQuery RecQ = Q.getWithoutCondContext(); |
| 1994 | RecQ.CtxI = P; |
| 1995 | computeKnownBits(V: Start, DemandedElts, Known&: KnownStart, Q: RecQ, Depth: Depth + 1); |
| 1996 | switch (Opcode) { |
| 1997 | case Instruction::Shl: |
| 1998 | // A shl recurrence will only increase the tailing zeros |
| 1999 | Known.Zero.setLowBits(KnownStart.countMinTrailingZeros()); |
| 2000 | break; |
| 2001 | case Instruction::LShr: |
| 2002 | case Instruction::UDiv: |
| 2003 | case Instruction::URem: |
| 2004 | // lshr, udiv, and urem recurrences will preserve the leading zeros of |
| 2005 | // the start value. |
| 2006 | Known.Zero.setHighBits(KnownStart.countMinLeadingZeros()); |
| 2007 | break; |
| 2008 | case Instruction::AShr: |
| 2009 | // An ashr recurrence will extend the initial sign bit |
| 2010 | Known.Zero.setHighBits(KnownStart.countMinLeadingZeros()); |
| 2011 | Known.One.setHighBits(KnownStart.countMinLeadingOnes()); |
| 2012 | break; |
| 2013 | } |
| 2014 | break; |
| 2015 | } |
| 2016 | |
| 2017 | case Instruction::And: { |
| 2018 | // Bits that are zero in the start value stay zero, and bits that are |
| 2019 | // one in both the start value and the step stay one. |
| 2020 | KnownBits KnownStep(BitWidth); |
| 2021 | computeKnownBitsForRecurrenceOperands(P, Start, Step, DemandedElts, |
| 2022 | KnownStart, KnownStep, Q, Depth); |
| 2023 | Known.Zero |= KnownStart.Zero; |
| 2024 | Known.One |= KnownStart.One & KnownStep.One; |
| 2025 | break; |
| 2026 | } |
| 2027 | |
| 2028 | case Instruction::Or: { |
| 2029 | // Bits that are zero in both the start value and the step stay zero, |
| 2030 | // and bits that are one in the start value stay one. |
| 2031 | KnownBits KnownStep(BitWidth); |
| 2032 | computeKnownBitsForRecurrenceOperands(P, Start, Step, DemandedElts, |
| 2033 | KnownStart, KnownStep, Q, Depth); |
| 2034 | Known.Zero |= KnownStart.Zero & KnownStep.Zero; |
| 2035 | Known.One |= KnownStart.One; |
| 2036 | break; |
| 2037 | } |
| 2038 | |
| 2039 | // Check for operations that have the property that if |
| 2040 | // both their operands have low zero bits, the result |
| 2041 | // will have low zero bits. |
| 2042 | case Instruction::Add: |
| 2043 | case Instruction::Sub: |
| 2044 | case Instruction::Mul: { |
| 2045 | // Ok, we have a recurrence of the form {Start,op,Step}. Check for low |
| 2046 | // zero bits. |
| 2047 | KnownBits KnownStep(BitWidth); |
| 2048 | computeKnownBitsForRecurrenceOperands(P, Start, Step, DemandedElts, |
| 2049 | KnownStart, KnownStep, Q, Depth); |
| 2050 | |
| 2051 | Known.Zero.setLowBits(std::min(a: KnownStart.countMinTrailingZeros(), |
| 2052 | b: KnownStep.countMinTrailingZeros())); |
| 2053 | |
| 2054 | auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(Val: BO); |
| 2055 | if (!OverflowOp || !Q.IIQ.hasNoSignedWrap(Op: OverflowOp)) |
| 2056 | break; |
| 2057 | |
| 2058 | switch (Opcode) { |
| 2059 | // If initial value of recurrence is nonnegative, and we are adding |
| 2060 | // a nonnegative number with nsw, the result can only be nonnegative |
| 2061 | // or poison value regardless of the number of times we execute the |
| 2062 | // add in phi recurrence. If initial value is negative and we are |
| 2063 | // adding a negative number with nsw, the result can only be |
| 2064 | // negative or poison value. Similar arguments apply to sub and mul. |
| 2065 | // |
| 2066 | // (add non-negative, non-negative) --> non-negative |
| 2067 | // (add negative, negative) --> negative |
| 2068 | case Instruction::Add: { |
| 2069 | if (KnownStart.isNonNegative() && KnownStep.isNonNegative()) |
| 2070 | Known.makeNonNegative(); |
| 2071 | else if (KnownStart.isNegative() && KnownStep.isNegative()) |
| 2072 | Known.makeNegative(); |
| 2073 | break; |
| 2074 | } |
| 2075 | |
| 2076 | // (sub nsw non-negative, negative) --> non-negative |
| 2077 | // (sub nsw negative, non-negative) --> negative |
| 2078 | case Instruction::Sub: { |
| 2079 | if (BO->getOperand(i_nocapture: 0) != I) |
| 2080 | break; |
| 2081 | if (KnownStart.isNonNegative() && KnownStep.isNegative()) |
| 2082 | Known.makeNonNegative(); |
| 2083 | else if (KnownStart.isNegative() && KnownStep.isNonNegative()) |
| 2084 | Known.makeNegative(); |
| 2085 | break; |
| 2086 | } |
| 2087 | |
| 2088 | // (mul nsw non-negative, non-negative) --> non-negative |
| 2089 | case Instruction::Mul: |
| 2090 | if (KnownStart.isNonNegative() && KnownStep.isNonNegative()) |
| 2091 | Known.makeNonNegative(); |
| 2092 | break; |
| 2093 | |
| 2094 | default: |
| 2095 | break; |
| 2096 | } |
| 2097 | break; |
| 2098 | } |
| 2099 | |
| 2100 | default: |
| 2101 | break; |
| 2102 | } |
| 2103 | } else { |
| 2104 | IntrinsicInst *II = nullptr; |
| 2105 | GEPOperator *GEP = nullptr; |
| 2106 | if (matchTwoInputRecurrence<IntrinsicInst>(PN: P, Inst&: II, Init&: Start, OtherOp&: Step)) { |
| 2107 | // %iv = [<Start>, %entry], [%iv.next, %backedge] |
| 2108 | // |
| 2109 | // %iv.next = <II>(%iv, <Step>) |
| 2110 | // or |
| 2111 | // %iv.next = <II>(<Step>, %iv) |
| 2112 | Intrinsic::ID IntrinsicID = II->getIntrinsicID(); |
| 2113 | if (IntrinsicID == Intrinsic::umin || IntrinsicID == Intrinsic::umax) { |
| 2114 | KnownBits KnownStep(BitWidth); |
| 2115 | computeKnownBitsForRecurrenceOperands( |
| 2116 | P, Start, Step, DemandedElts, KnownStart, KnownStep, Q, Depth); |
| 2117 | |
| 2118 | if (IntrinsicID == Intrinsic::umin) { |
| 2119 | Known.Zero.setHighBits(KnownStart.countMinLeadingZeros()); |
| 2120 | Known.One.setHighBits(std::min(a: KnownStart.countMinLeadingOnes(), |
| 2121 | b: KnownStep.countMinLeadingOnes())); |
| 2122 | } else { |
| 2123 | // umax |
| 2124 | Known.Zero.setHighBits(std::min(a: KnownStart.countMinLeadingZeros(), |
| 2125 | b: KnownStep.countMinLeadingZeros())); |
| 2126 | Known.One.setHighBits(KnownStart.countMinLeadingOnes()); |
| 2127 | } |
| 2128 | } |
| 2129 | } else if (Known.isUnknown() && I->getType()->isPointerTy() && |
| 2130 | matchTwoInputRecurrence<GEPOperator>(PN: P, Inst&: GEP, Init&: Start, OtherOp&: Step) && |
| 2131 | GEP->getPointerOperand() == P) { |
| 2132 | // Handle a pointer induction variable, whose step is a getelementptr: |
| 2133 | // %p = phi ptr [ %start, %preheader ], [ %next, %latch ] |
| 2134 | // %next = getelementptr <ty>, ptr %p, <indices> |
| 2135 | // Every value %p takes is %start plus a whole number of steps, and |
| 2136 | // tz(a + b) >= min(tz(a), tz(b)), so the trailing zeros of %p - and |
| 2137 | // hence its alignment - are at least the smaller of those of %start and |
| 2138 | // of the step. |
| 2139 | |
| 2140 | // Evaluate the start value on its own incoming edge. |
| 2141 | unsigned StartIdx = P->getIncomingValue(i: 0) == Start ? 0 : 1; |
| 2142 | SimplifyQuery RecQ = Q.getWithoutCondContext(); |
| 2143 | RecQ.CtxI = P->getIncomingBlock(i: StartIdx)->getTerminator(); |
| 2144 | computeKnownBits(V: Start, DemandedElts, Known&: KnownStart, Q: RecQ, Depth: Depth + 1); |
| 2145 | Known.Zero.setLowBits( |
| 2146 | std::min(a: KnownStart.countMinTrailingZeros(), |
| 2147 | b: Log2(A: GEP->getMaxPreservedAlignment(DL: Q.DL)))); |
| 2148 | } |
| 2149 | } |
| 2150 | |
| 2151 | // Unreachable blocks may have zero-operand PHI nodes. |
| 2152 | if (P->getNumIncomingValues() == 0) |
| 2153 | break; |
| 2154 | |
| 2155 | // Otherwise take the unions of the known bit sets of the operands, |
| 2156 | // taking conservative care to avoid excessive recursion. |
| 2157 | if (Depth < MaxAnalysisRecursionDepth - 1 && Known.isUnknown()) { |
| 2158 | // Skip if every incoming value references to ourself. |
| 2159 | if (isa_and_nonnull<UndefValue>(Val: P->hasConstantValue())) |
| 2160 | break; |
| 2161 | |
| 2162 | Known.setAllConflict(); |
| 2163 | for (const Use &U : P->operands()) { |
| 2164 | Value *IncValue; |
| 2165 | const PHINode *CtxPhi; |
| 2166 | Instruction *CtxI; |
| 2167 | breakSelfRecursivePHI(U: &U, PHI: P, ValOut&: IncValue, CtxIOut&: CtxI, PhiOut: &CtxPhi); |
| 2168 | // Skip direct self references. |
| 2169 | if (IncValue == P) |
| 2170 | continue; |
| 2171 | |
| 2172 | // Change the context instruction to the "edge" that flows into the |
| 2173 | // phi. This is important because that is where the value is actually |
| 2174 | // "evaluated" even though it is used later somewhere else. (see also |
| 2175 | // D69571). |
| 2176 | SimplifyQuery RecQ = Q.getWithoutCondContext().getWithInstruction(I: CtxI); |
| 2177 | |
| 2178 | Known2 = KnownBits(BitWidth); |
| 2179 | |
| 2180 | // Recurse, but cap the recursion to one level, because we don't |
| 2181 | // want to waste time spinning around in loops. |
| 2182 | // TODO: See if we can base recursion limiter on number of incoming phi |
| 2183 | // edges so we don't overly clamp analysis. |
| 2184 | computeKnownBits(V: IncValue, DemandedElts, Known&: Known2, Q: RecQ, |
| 2185 | Depth: MaxAnalysisRecursionDepth - 1); |
| 2186 | |
| 2187 | // See if we can further use a conditional branch into the phi |
| 2188 | // to help us determine the range of the value. |
| 2189 | if (!Known2.isConstant()) { |
| 2190 | CmpPredicate Pred; |
| 2191 | Value *RHSValue = nullptr; |
| 2192 | BasicBlock *TrueSucc, *FalseSucc; |
| 2193 | if (match( |
| 2194 | V: RecQ.CtxI, |
| 2195 | P: m_Br(C: m_c_ICmp(Pred, L: m_Specific(V: IncValue), R: m_Value(V&: RHSValue)), |
| 2196 | T: m_BasicBlock(V&: TrueSucc), F: m_BasicBlock(V&: FalseSucc)))) { |
| 2197 | // Check for cases of duplicate successors. |
| 2198 | if ((TrueSucc == CtxPhi->getParent()) != |
| 2199 | (FalseSucc == CtxPhi->getParent())) { |
| 2200 | // If we're using the false successor, invert the predicate. |
| 2201 | if (FalseSucc == CtxPhi->getParent()) |
| 2202 | Pred = CmpInst::getInversePredicate(pred: Pred); |
| 2203 | // Get the knownbits implied by the incoming phi condition. |
| 2204 | KnownBits RHSKnown = computeKnownBits( |
| 2205 | V: RHSValue, DemandedElts, Q: RecQ, Depth: MaxAnalysisRecursionDepth - 1); |
| 2206 | if (!RHSKnown.isUnknown()) { |
| 2207 | ConstantRange RHSRange = ConstantRange::fromKnownBits( |
| 2208 | Known: RHSKnown, /*IsSigned=*/ICmpInst::isSigned(Pred)); |
| 2209 | auto CR = ConstantRange::makeAllowedICmpRegion(Pred, Other: RHSRange); |
| 2210 | KnownBits KnownUnion = Known2.unionWith(RHS: CR.toKnownBits()); |
| 2211 | // We can have conflicts here if we are analyzing deadcode (its |
| 2212 | // impossible for us reach this BB based the icmp). |
| 2213 | if (KnownUnion.hasConflict()) { |
| 2214 | // No reason to continue analyzing in a known dead region, so |
| 2215 | // just resetAll and break. This will cause us to also exit |
| 2216 | // the outer loop. |
| 2217 | Known.resetAll(); |
| 2218 | break; |
| 2219 | } |
| 2220 | Known2 = KnownUnion; |
| 2221 | } |
| 2222 | } |
| 2223 | } |
| 2224 | } |
| 2225 | |
| 2226 | Known = Known.intersectWith(RHS: Known2); |
| 2227 | // If all bits have been ruled out, there's no need to check |
| 2228 | // more operands. |
| 2229 | if (Known.isUnknown()) |
| 2230 | break; |
| 2231 | } |
| 2232 | } |
| 2233 | break; |
| 2234 | } |
| 2235 | case Instruction::Call: |
| 2236 | case Instruction::Invoke: { |
| 2237 | // If range metadata is attached to this call, set known bits from that, |
| 2238 | // and then intersect with known bits based on other properties of the |
| 2239 | // function. |
| 2240 | if (MDNode *MD = |
| 2241 | Q.IIQ.getMetadata(I: cast<Instruction>(Val: I), KindID: LLVMContext::MD_range)) |
| 2242 | computeKnownBitsFromRangeMetadata(Ranges: *MD, Known); |
| 2243 | |
| 2244 | const auto *CB = cast<CallBase>(Val: I); |
| 2245 | |
| 2246 | if (std::optional<ConstantRange> Range = CB->getRange()) |
| 2247 | Known = Known.unionWith(RHS: Range->toKnownBits()); |
| 2248 | |
| 2249 | if (const Value *RV = CB->getReturnedArgOperand()) { |
| 2250 | if (RV->getType() == I->getType()) { |
| 2251 | computeKnownBits(V: RV, Known&: Known2, Q, Depth: Depth + 1); |
| 2252 | Known = Known.unionWith(RHS: Known2); |
| 2253 | // If the function doesn't return properly for all input values |
| 2254 | // (e.g. unreachable exits) then there might be conflicts between the |
| 2255 | // argument value and the range metadata. Simply discard the known bits |
| 2256 | // in case of conflicts. |
| 2257 | if (Known.hasConflict()) |
| 2258 | Known.resetAll(); |
| 2259 | } |
| 2260 | } |
| 2261 | if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: I)) { |
| 2262 | switch (II->getIntrinsicID()) { |
| 2263 | default: |
| 2264 | break; |
| 2265 | case Intrinsic::abs: { |
| 2266 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2267 | bool IntMinIsPoison = match(V: II->getArgOperand(i: 1), P: m_One()); |
| 2268 | Known = Known.unionWith(RHS: Known2.abs(IntMinIsPoison)); |
| 2269 | break; |
| 2270 | } |
| 2271 | case Intrinsic::bitreverse: |
| 2272 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2273 | Known = Known.unionWith(RHS: Known2.reverseBits()); |
| 2274 | break; |
| 2275 | case Intrinsic::bswap: |
| 2276 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2277 | Known = Known.unionWith(RHS: Known2.byteSwap()); |
| 2278 | break; |
| 2279 | case Intrinsic::ctlz: { |
| 2280 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2281 | // If we have a known 1, its position is our upper bound. |
| 2282 | unsigned PossibleLZ = Known2.countMaxLeadingZeros(); |
| 2283 | // If this call is poison for 0 input, the result will be less than 2^n. |
| 2284 | if (II->getArgOperand(i: 1) == ConstantInt::getTrue(Context&: II->getContext())) |
| 2285 | PossibleLZ = std::min(a: PossibleLZ, b: BitWidth - 1); |
| 2286 | unsigned LowBits = llvm::bit_width(Value: PossibleLZ); |
| 2287 | Known.Zero.setBitsFrom(LowBits); |
| 2288 | break; |
| 2289 | } |
| 2290 | case Intrinsic::cttz: { |
| 2291 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2292 | // If we have a known 1, its position is our upper bound. |
| 2293 | unsigned PossibleTZ = Known2.countMaxTrailingZeros(); |
| 2294 | // If this call is poison for 0 input, the result will be less than 2^n. |
| 2295 | if (II->getArgOperand(i: 1) == ConstantInt::getTrue(Context&: II->getContext())) |
| 2296 | PossibleTZ = std::min(a: PossibleTZ, b: BitWidth - 1); |
| 2297 | unsigned LowBits = llvm::bit_width(Value: PossibleTZ); |
| 2298 | Known.Zero.setBitsFrom(LowBits); |
| 2299 | break; |
| 2300 | } |
| 2301 | case Intrinsic::ctpop: { |
| 2302 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2303 | // We can bound the space the count needs. Also, bits known to be zero |
| 2304 | // can't contribute to the population. |
| 2305 | unsigned BitsPossiblySet = Known2.countMaxPopulation(); |
| 2306 | unsigned LowBits = llvm::bit_width(Value: BitsPossiblySet); |
| 2307 | Known.Zero.setBitsFrom(LowBits); |
| 2308 | // TODO: we could bound KnownOne using the lower bound on the number |
| 2309 | // of bits which might be set provided by popcnt KnownOne2. |
| 2310 | break; |
| 2311 | } |
| 2312 | case Intrinsic::fshr: |
| 2313 | case Intrinsic::fshl: { |
| 2314 | const APInt *SA; |
| 2315 | if (!match(V: I->getOperand(i: 2), P: m_APInt(Res&: SA))) |
| 2316 | break; |
| 2317 | |
| 2318 | KnownBits Known3(BitWidth); |
| 2319 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2320 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known3, Q, Depth: Depth + 1); |
| 2321 | Known = II->getIntrinsicID() == Intrinsic::fshl |
| 2322 | ? KnownBits::fshl(LHS: Known2, RHS: Known3, Amt: *SA) |
| 2323 | : KnownBits::fshr(LHS: Known2, RHS: Known3, Amt: *SA); |
| 2324 | break; |
| 2325 | } |
| 2326 | case Intrinsic::clmul: |
| 2327 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 2328 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2329 | Known = KnownBits::clmul(LHS: Known, RHS: Known2); |
| 2330 | break; |
| 2331 | case Intrinsic::pext: |
| 2332 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 2333 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2334 | Known = KnownBits::pext(Val: Known, Mask: Known2); |
| 2335 | break; |
| 2336 | case Intrinsic::pdep: |
| 2337 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 2338 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2339 | Known = KnownBits::pdep(Val: Known, Mask: Known2); |
| 2340 | break; |
| 2341 | case Intrinsic::smulh: |
| 2342 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 2343 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2344 | Known = KnownBits::mulhs(LHS: Known, RHS: Known2); |
| 2345 | break; |
| 2346 | case Intrinsic::umulh: |
| 2347 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 2348 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2349 | Known = KnownBits::mulhu(LHS: Known, RHS: Known2); |
| 2350 | break; |
| 2351 | case Intrinsic::uadd_sat: |
| 2352 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 2353 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2354 | Known = KnownBits::uadd_sat(LHS: Known, RHS: Known2); |
| 2355 | break; |
| 2356 | case Intrinsic::usub_sat: |
| 2357 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 2358 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2359 | Known = KnownBits::usub_sat(LHS: Known, RHS: Known2); |
| 2360 | break; |
| 2361 | case Intrinsic::sadd_sat: |
| 2362 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 2363 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2364 | Known = KnownBits::sadd_sat(LHS: Known, RHS: Known2); |
| 2365 | break; |
| 2366 | case Intrinsic::ssub_sat: |
| 2367 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 2368 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2369 | Known = KnownBits::ssub_sat(LHS: Known, RHS: Known2); |
| 2370 | break; |
| 2371 | // Vec reverse preserves bits from input vec. |
| 2372 | case Intrinsic::vector_reverse: |
| 2373 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts: DemandedElts.reverseBits(), Known, Q, |
| 2374 | Depth: Depth + 1); |
| 2375 | break; |
| 2376 | // for min/max/and/or reduce, any bit common to each element in the |
| 2377 | // input vec is set in the output. |
| 2378 | case Intrinsic::vector_reduce_and: |
| 2379 | case Intrinsic::vector_reduce_or: |
| 2380 | case Intrinsic::vector_reduce_umax: |
| 2381 | case Intrinsic::vector_reduce_umin: |
| 2382 | case Intrinsic::vector_reduce_smax: |
| 2383 | case Intrinsic::vector_reduce_smin: |
| 2384 | computeKnownBits(V: I->getOperand(i: 0), Known, Q, Depth: Depth + 1); |
| 2385 | break; |
| 2386 | case Intrinsic::vector_reduce_xor: { |
| 2387 | computeKnownBits(V: I->getOperand(i: 0), Known, Q, Depth: Depth + 1); |
| 2388 | // The zeros common to all vecs are zero in the output. |
| 2389 | // If the number of elements is odd, then the common ones remain. If the |
| 2390 | // number of elements is even, then the common ones becomes zeros. |
| 2391 | auto *VecTy = cast<VectorType>(Val: I->getOperand(i: 0)->getType()); |
| 2392 | // Even, so the ones become zeros. |
| 2393 | bool EvenCnt = VecTy->getElementCount().isKnownEven(); |
| 2394 | if (EvenCnt) |
| 2395 | Known.Zero |= Known.One; |
| 2396 | // Maybe even element count so need to clear ones. |
| 2397 | if (VecTy->isScalableTy() || EvenCnt) |
| 2398 | Known.One.clearAllBits(); |
| 2399 | break; |
| 2400 | } |
| 2401 | case Intrinsic::vector_reduce_add: { |
| 2402 | auto *VecTy = dyn_cast<FixedVectorType>(Val: I->getOperand(i: 0)->getType()); |
| 2403 | if (!VecTy) |
| 2404 | break; |
| 2405 | computeKnownBits(V: I->getOperand(i: 0), Known, Q, Depth: Depth + 1); |
| 2406 | Known = Known.reduceAdd(NumElts: VecTy->getNumElements()); |
| 2407 | break; |
| 2408 | } |
| 2409 | case Intrinsic::umin: |
| 2410 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 2411 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2412 | Known = KnownBits::umin(LHS: Known, RHS: Known2); |
| 2413 | break; |
| 2414 | case Intrinsic::umax: |
| 2415 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 2416 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2417 | Known = KnownBits::umax(LHS: Known, RHS: Known2); |
| 2418 | break; |
| 2419 | case Intrinsic::smin: |
| 2420 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 2421 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2422 | Known = KnownBits::smin(LHS: Known, RHS: Known2); |
| 2423 | unionWithMinMaxIntrinsicClamp(II, Known); |
| 2424 | break; |
| 2425 | case Intrinsic::smax: |
| 2426 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 2427 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2428 | Known = KnownBits::smax(LHS: Known, RHS: Known2); |
| 2429 | unionWithMinMaxIntrinsicClamp(II, Known); |
| 2430 | break; |
| 2431 | case Intrinsic::ptrmask: { |
| 2432 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 2433 | |
| 2434 | const Value *Mask = I->getOperand(i: 1); |
| 2435 | Known2 = KnownBits(Mask->getType()->getScalarSizeInBits()); |
| 2436 | computeKnownBits(V: Mask, DemandedElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2437 | // TODO: 1-extend would be more precise. |
| 2438 | Known &= Known2.anyextOrTrunc(BitWidth); |
| 2439 | break; |
| 2440 | } |
| 2441 | case Intrinsic::x86_sse42_crc32_64_64: |
| 2442 | Known.Zero.setBitsFrom(32); |
| 2443 | break; |
| 2444 | case Intrinsic::x86_ssse3_phadd_d_128: |
| 2445 | case Intrinsic::x86_ssse3_phadd_w_128: |
| 2446 | case Intrinsic::x86_avx2_phadd_d: |
| 2447 | case Intrinsic::x86_avx2_phadd_w: { |
| 2448 | Known = computeKnownBitsForHorizontalOperation( |
| 2449 | I, DemandedElts, Q, Depth, |
| 2450 | KnownBitsFunc: [](const KnownBits &KnownLHS, const KnownBits &KnownRHS) { |
| 2451 | return KnownBits::add(LHS: KnownLHS, RHS: KnownRHS); |
| 2452 | }); |
| 2453 | break; |
| 2454 | } |
| 2455 | case Intrinsic::x86_ssse3_phadd_sw_128: |
| 2456 | case Intrinsic::x86_avx2_phadd_sw: { |
| 2457 | Known = computeKnownBitsForHorizontalOperation( |
| 2458 | I, DemandedElts, Q, Depth, KnownBitsFunc: KnownBits::sadd_sat); |
| 2459 | break; |
| 2460 | } |
| 2461 | case Intrinsic::x86_ssse3_phsub_d_128: |
| 2462 | case Intrinsic::x86_ssse3_phsub_w_128: |
| 2463 | case Intrinsic::x86_avx2_phsub_d: |
| 2464 | case Intrinsic::x86_avx2_phsub_w: { |
| 2465 | Known = computeKnownBitsForHorizontalOperation( |
| 2466 | I, DemandedElts, Q, Depth, |
| 2467 | KnownBitsFunc: [](const KnownBits &KnownLHS, const KnownBits &KnownRHS) { |
| 2468 | return KnownBits::sub(LHS: KnownLHS, RHS: KnownRHS); |
| 2469 | }); |
| 2470 | break; |
| 2471 | } |
| 2472 | case Intrinsic::x86_ssse3_phsub_sw_128: |
| 2473 | case Intrinsic::x86_avx2_phsub_sw: { |
| 2474 | Known = computeKnownBitsForHorizontalOperation( |
| 2475 | I, DemandedElts, Q, Depth, KnownBitsFunc: KnownBits::ssub_sat); |
| 2476 | break; |
| 2477 | } |
| 2478 | case Intrinsic::riscv_vsetvli: |
| 2479 | case Intrinsic::riscv_vsetvlimax: { |
| 2480 | bool HasAVL = II->getIntrinsicID() == Intrinsic::riscv_vsetvli; |
| 2481 | const ConstantRange Range = getVScaleRange(F: II->getFunction(), BitWidth); |
| 2482 | uint64_t SEW = RISCVVType::decodeVSEW( |
| 2483 | VSEW: cast<ConstantInt>(Val: II->getArgOperand(i: HasAVL))->getZExtValue()); |
| 2484 | RISCVVType::VLMUL VLMUL = static_cast<RISCVVType::VLMUL>( |
| 2485 | cast<ConstantInt>(Val: II->getArgOperand(i: 1 + HasAVL))->getZExtValue()); |
| 2486 | uint64_t MaxVLEN = |
| 2487 | Range.getUnsignedMax().getZExtValue() * RISCV::RVVBitsPerBlock; |
| 2488 | uint64_t MaxVL = MaxVLEN / RISCVVType::getSEWLMULRatio(SEW, VLMul: VLMUL); |
| 2489 | |
| 2490 | // Result of vsetvli must be not larger than AVL. |
| 2491 | if (HasAVL) |
| 2492 | if (auto *CI = dyn_cast<ConstantInt>(Val: II->getArgOperand(i: 0))) |
| 2493 | MaxVL = std::min(a: MaxVL, b: CI->getZExtValue()); |
| 2494 | |
| 2495 | unsigned KnownZeroFirstBit = Log2_32(Value: MaxVL) + 1; |
| 2496 | if (BitWidth > KnownZeroFirstBit) |
| 2497 | Known.Zero.setBitsFrom(KnownZeroFirstBit); |
| 2498 | break; |
| 2499 | } |
| 2500 | case Intrinsic::amdgcn_mbcnt_hi: |
| 2501 | case Intrinsic::amdgcn_mbcnt_lo: { |
| 2502 | // Wave64 mbcnt_lo returns at most 32 + src1. Otherwise these return at |
| 2503 | // most 31 + src1. |
| 2504 | KnownBits MbcntKnown(BitWidth); |
| 2505 | MbcntKnown.Zero.setBitsFrom( |
| 2506 | II->getIntrinsicID() == Intrinsic::amdgcn_mbcnt_lo ? 6 : 5); |
| 2507 | computeKnownBits(V: I->getOperand(i: 1), Known&: Known2, Q, Depth: Depth + 1); |
| 2508 | Known = Known.unionWith(RHS: KnownBits::add(LHS: MbcntKnown, RHS: Known2)); |
| 2509 | break; |
| 2510 | } |
| 2511 | case Intrinsic::vscale: { |
| 2512 | if (!II->getParent() || !II->getFunction()) |
| 2513 | break; |
| 2514 | |
| 2515 | Known = getVScaleRange(F: II->getFunction(), BitWidth).toKnownBits(); |
| 2516 | break; |
| 2517 | } |
| 2518 | case Intrinsic::stepvector: { |
| 2519 | auto *VecTy = cast<VectorType>(Val: II->getType()); |
| 2520 | unsigned MinNumElts = VecTy->getElementCount().getKnownMinValue(); |
| 2521 | if (!isUIntN(N: BitWidth, x: MinNumElts)) |
| 2522 | break; |
| 2523 | |
| 2524 | bool Overflow = false; |
| 2525 | APInt MaxNumElts(BitWidth, MinNumElts); |
| 2526 | if (VecTy->isScalableTy()) { |
| 2527 | if (!II->getParent() || !II->getFunction()) |
| 2528 | break; |
| 2529 | MaxNumElts = getVScaleRange(F: II->getFunction(), BitWidth) |
| 2530 | .getUnsignedMax() |
| 2531 | .umul_ov(RHS: MaxNumElts, Overflow); |
| 2532 | } |
| 2533 | |
| 2534 | // Give up if the lane count could wrap. Stepvector truncates lane |
| 2535 | // indices that do not fit in the element type. |
| 2536 | if (Overflow) |
| 2537 | break; |
| 2538 | |
| 2539 | Known.Zero.setHighBits((MaxNumElts - 1).countl_zero()); |
| 2540 | break; |
| 2541 | } |
| 2542 | } |
| 2543 | } |
| 2544 | break; |
| 2545 | } |
| 2546 | case Instruction::ShuffleVector: { |
| 2547 | if (auto *Splat = getSplatValue(V: I)) { |
| 2548 | computeKnownBits(V: Splat, Known, Q, Depth: Depth + 1); |
| 2549 | break; |
| 2550 | } |
| 2551 | |
| 2552 | auto *Shuf = dyn_cast<ShuffleVectorInst>(Val: I); |
| 2553 | // FIXME: Do we need to handle ConstantExpr involving shufflevectors? |
| 2554 | if (!Shuf) { |
| 2555 | Known.resetAll(); |
| 2556 | return; |
| 2557 | } |
| 2558 | // For undef elements, we don't know anything about the common state of |
| 2559 | // the shuffle result. |
| 2560 | APInt DemandedLHS, DemandedRHS; |
| 2561 | if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) { |
| 2562 | Known.resetAll(); |
| 2563 | return; |
| 2564 | } |
| 2565 | Known.setAllConflict(); |
| 2566 | if (!!DemandedLHS) { |
| 2567 | const Value *LHS = Shuf->getOperand(i_nocapture: 0); |
| 2568 | computeKnownBits(V: LHS, DemandedElts: DemandedLHS, Known, Q, Depth: Depth + 1); |
| 2569 | // If we don't know any bits, early out. |
| 2570 | if (Known.isUnknown()) |
| 2571 | break; |
| 2572 | } |
| 2573 | if (!!DemandedRHS) { |
| 2574 | const Value *RHS = Shuf->getOperand(i_nocapture: 1); |
| 2575 | computeKnownBits(V: RHS, DemandedElts: DemandedRHS, Known&: Known2, Q, Depth: Depth + 1); |
| 2576 | Known = Known.intersectWith(RHS: Known2); |
| 2577 | } |
| 2578 | break; |
| 2579 | } |
| 2580 | case Instruction::InsertElement: { |
| 2581 | if (isa<ScalableVectorType>(Val: I->getType())) { |
| 2582 | Known.resetAll(); |
| 2583 | return; |
| 2584 | } |
| 2585 | const Value *Vec = I->getOperand(i: 0); |
| 2586 | const Value *Elt = I->getOperand(i: 1); |
| 2587 | auto *CIdx = dyn_cast<ConstantInt>(Val: I->getOperand(i: 2)); |
| 2588 | unsigned NumElts = DemandedElts.getBitWidth(); |
| 2589 | APInt DemandedVecElts = DemandedElts; |
| 2590 | bool NeedsElt = true; |
| 2591 | // If we know the index we are inserting too, clear it from Vec check. |
| 2592 | if (CIdx && CIdx->getValue().ult(RHS: NumElts)) { |
| 2593 | DemandedVecElts.clearBit(BitPosition: CIdx->getZExtValue()); |
| 2594 | NeedsElt = DemandedElts[CIdx->getZExtValue()]; |
| 2595 | } |
| 2596 | |
| 2597 | Known.setAllConflict(); |
| 2598 | if (NeedsElt) { |
| 2599 | computeKnownBits(V: Elt, Known, Q, Depth: Depth + 1); |
| 2600 | // If we don't know any bits, early out. |
| 2601 | if (Known.isUnknown()) |
| 2602 | break; |
| 2603 | } |
| 2604 | |
| 2605 | if (!DemandedVecElts.isZero()) { |
| 2606 | computeKnownBits(V: Vec, DemandedElts: DemandedVecElts, Known&: Known2, Q, Depth: Depth + 1); |
| 2607 | Known = Known.intersectWith(RHS: Known2); |
| 2608 | } |
| 2609 | break; |
| 2610 | } |
| 2611 | case Instruction::ExtractElement: { |
| 2612 | // Look through extract element. If the index is non-constant or |
| 2613 | // out-of-range demand all elements, otherwise just the extracted element. |
| 2614 | const Value *Vec = I->getOperand(i: 0); |
| 2615 | const Value *Idx = I->getOperand(i: 1); |
| 2616 | auto *CIdx = dyn_cast<ConstantInt>(Val: Idx); |
| 2617 | if (isa<ScalableVectorType>(Val: Vec->getType())) { |
| 2618 | // FIXME: there's probably *something* we can do with scalable vectors |
| 2619 | Known.resetAll(); |
| 2620 | break; |
| 2621 | } |
| 2622 | unsigned NumElts = cast<FixedVectorType>(Val: Vec->getType())->getNumElements(); |
| 2623 | APInt DemandedVecElts = APInt::getAllOnes(numBits: NumElts); |
| 2624 | if (CIdx && CIdx->getValue().ult(RHS: NumElts)) |
| 2625 | DemandedVecElts = APInt::getOneBitSet(numBits: NumElts, BitNo: CIdx->getZExtValue()); |
| 2626 | computeKnownBits(V: Vec, DemandedElts: DemandedVecElts, Known, Q, Depth: Depth + 1); |
| 2627 | break; |
| 2628 | } |
| 2629 | case Instruction::ExtractValue: |
| 2630 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: I->getOperand(i: 0))) { |
| 2631 | const ExtractValueInst *EVI = cast<ExtractValueInst>(Val: I); |
| 2632 | if (EVI->getNumIndices() != 1) break; |
| 2633 | if (EVI->getIndices()[0] == 0) { |
| 2634 | switch (II->getIntrinsicID()) { |
| 2635 | default: break; |
| 2636 | case Intrinsic::uadd_with_overflow: |
| 2637 | case Intrinsic::sadd_with_overflow: |
| 2638 | computeKnownBitsAddSub( |
| 2639 | Add: true, Op0: II->getArgOperand(i: 0), Op1: II->getArgOperand(i: 1), /*NSW=*/false, |
| 2640 | /* NUW=*/false, DemandedElts, KnownOut&: Known, Known2, Q, Depth); |
| 2641 | break; |
| 2642 | case Intrinsic::usub_with_overflow: |
| 2643 | case Intrinsic::ssub_with_overflow: |
| 2644 | computeKnownBitsAddSub( |
| 2645 | Add: false, Op0: II->getArgOperand(i: 0), Op1: II->getArgOperand(i: 1), /*NSW=*/false, |
| 2646 | /* NUW=*/false, DemandedElts, KnownOut&: Known, Known2, Q, Depth); |
| 2647 | break; |
| 2648 | case Intrinsic::umul_with_overflow: |
| 2649 | case Intrinsic::smul_with_overflow: |
| 2650 | computeKnownBitsMul(Op0: II->getArgOperand(i: 0), Op1: II->getArgOperand(i: 1), NSW: false, |
| 2651 | NUW: false, DemandedElts, Known, Known2, Q, Depth); |
| 2652 | break; |
| 2653 | } |
| 2654 | } |
| 2655 | } |
| 2656 | break; |
| 2657 | case Instruction::Freeze: |
| 2658 | if (isGuaranteedNotToBePoison(V: I->getOperand(i: 0), AC: Q.AC, CtxI: Q.CtxI, DT: Q.DT, |
| 2659 | Depth: Depth + 1)) |
| 2660 | computeKnownBits(V: I->getOperand(i: 0), Known, Q, Depth: Depth + 1); |
| 2661 | break; |
| 2662 | } |
| 2663 | } |
| 2664 | |
| 2665 | /// Determine which bits of V are known to be either zero or one and return |
| 2666 | /// them. |
| 2667 | KnownBits llvm::computeKnownBits(const Value *V, const APInt &DemandedElts, |
| 2668 | const SimplifyQuery &Q, unsigned Depth) { |
| 2669 | KnownBits Known(getBitWidth(Ty: V->getType(), DL: Q.DL)); |
| 2670 | ::computeKnownBits(V, DemandedElts, Known, Q, Depth); |
| 2671 | return Known; |
| 2672 | } |
| 2673 | |
| 2674 | /// Determine which bits of V are known to be either zero or one and return |
| 2675 | /// them. |
| 2676 | KnownBits llvm::computeKnownBits(const Value *V, const SimplifyQuery &Q, |
| 2677 | unsigned Depth) { |
| 2678 | KnownBits Known(getBitWidth(Ty: V->getType(), DL: Q.DL)); |
| 2679 | computeKnownBits(V, Known, Q, Depth); |
| 2680 | return Known; |
| 2681 | } |
| 2682 | |
| 2683 | /// Determine which bits of V are known to be either zero or one and return |
| 2684 | /// them in the Known bit set. |
| 2685 | /// |
| 2686 | /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that |
| 2687 | /// we cannot optimize based on the assumption that it is zero without changing |
| 2688 | /// it to be an explicit zero. If we don't change it to zero, other code could |
| 2689 | /// optimized based on the contradictory assumption that it is non-zero. |
| 2690 | /// Because instcombine aggressively folds operations with undef args anyway, |
| 2691 | /// this won't lose us code quality. |
| 2692 | /// |
| 2693 | /// This function is defined on values with integer type, values with pointer |
| 2694 | /// type, and vectors of integers. In the case |
| 2695 | /// where V is a vector, known zero, and known one values are the |
| 2696 | /// same width as the vector element, and the bit is set only if it is true |
| 2697 | /// for all of the demanded elements in the vector specified by DemandedElts. |
| 2698 | void computeKnownBits(const Value *V, const APInt &DemandedElts, |
| 2699 | KnownBits &Known, const SimplifyQuery &Q, |
| 2700 | unsigned Depth) { |
| 2701 | if (!DemandedElts) { |
| 2702 | // No demanded elts, better to assume we don't know anything. |
| 2703 | Known.resetAll(); |
| 2704 | return; |
| 2705 | } |
| 2706 | |
| 2707 | assert(V && "No Value?" ); |
| 2708 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth" ); |
| 2709 | |
| 2710 | #ifndef NDEBUG |
| 2711 | Type *Ty = V->getType(); |
| 2712 | unsigned BitWidth = Known.getBitWidth(); |
| 2713 | |
| 2714 | assert((Ty->isIntOrIntVectorTy(BitWidth) || Ty->isPtrOrPtrVectorTy()) && |
| 2715 | "Not integer or pointer type!" ); |
| 2716 | |
| 2717 | if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) { |
| 2718 | assert( |
| 2719 | FVTy->getNumElements() == DemandedElts.getBitWidth() && |
| 2720 | "DemandedElt width should equal the fixed vector number of elements" ); |
| 2721 | } else { |
| 2722 | assert(DemandedElts == APInt(1, 1) && |
| 2723 | "DemandedElt width should be 1 for scalars or scalable vectors" ); |
| 2724 | } |
| 2725 | |
| 2726 | Type *ScalarTy = Ty->getScalarType(); |
| 2727 | if (ScalarTy->isPointerTy()) { |
| 2728 | assert(BitWidth == Q.DL.getPointerTypeSizeInBits(ScalarTy) && |
| 2729 | "V and Known should have same BitWidth" ); |
| 2730 | } else { |
| 2731 | assert(BitWidth == Q.DL.getTypeSizeInBits(ScalarTy) && |
| 2732 | "V and Known should have same BitWidth" ); |
| 2733 | } |
| 2734 | #endif |
| 2735 | |
| 2736 | const APInt *C; |
| 2737 | if (match(V, P: m_APInt(Res&: C))) { |
| 2738 | // We know all of the bits for a scalar constant or a splat vector constant! |
| 2739 | Known = KnownBits::makeConstant(C: *C); |
| 2740 | return; |
| 2741 | } |
| 2742 | // Null and aggregate-zero are all-zeros. |
| 2743 | if (isa<ConstantPointerNull>(Val: V) || isa<ConstantAggregateZero>(Val: V)) { |
| 2744 | Known.setAllZero(); |
| 2745 | return; |
| 2746 | } |
| 2747 | // Handle a constant vector by taking the intersection of the known bits of |
| 2748 | // each element. |
| 2749 | if (const ConstantDataVector *CDV = dyn_cast<ConstantDataVector>(Val: V)) { |
| 2750 | assert(!isa<ScalableVectorType>(V->getType())); |
| 2751 | // We know that CDV must be a vector of integers. Take the intersection of |
| 2752 | // each element. |
| 2753 | Known.setAllConflict(); |
| 2754 | for (unsigned i = 0, e = CDV->getNumElements(); i != e; ++i) { |
| 2755 | if (!DemandedElts[i]) |
| 2756 | continue; |
| 2757 | APInt Elt = CDV->getElementAsAPInt(i); |
| 2758 | Known.Zero &= ~Elt; |
| 2759 | Known.One &= Elt; |
| 2760 | } |
| 2761 | if (Known.hasConflict()) |
| 2762 | Known.resetAll(); |
| 2763 | return; |
| 2764 | } |
| 2765 | |
| 2766 | if (const auto *CV = dyn_cast<ConstantVector>(Val: V)) { |
| 2767 | assert(!isa<ScalableVectorType>(V->getType())); |
| 2768 | // We know that CV must be a vector of integers. Take the intersection of |
| 2769 | // each element. |
| 2770 | Known.setAllConflict(); |
| 2771 | for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) { |
| 2772 | if (!DemandedElts[i]) |
| 2773 | continue; |
| 2774 | Constant *Element = CV->getAggregateElement(Elt: i); |
| 2775 | if (isa<PoisonValue>(Val: Element)) |
| 2776 | continue; |
| 2777 | auto *ElementCI = dyn_cast_or_null<ConstantInt>(Val: Element); |
| 2778 | if (!ElementCI) { |
| 2779 | Known.resetAll(); |
| 2780 | return; |
| 2781 | } |
| 2782 | const APInt &Elt = ElementCI->getValue(); |
| 2783 | Known.Zero &= ~Elt; |
| 2784 | Known.One &= Elt; |
| 2785 | } |
| 2786 | if (Known.hasConflict()) |
| 2787 | Known.resetAll(); |
| 2788 | return; |
| 2789 | } |
| 2790 | |
| 2791 | // Start out not knowing anything. |
| 2792 | Known.resetAll(); |
| 2793 | |
| 2794 | // We can't imply anything about undefs. |
| 2795 | if (isa<UndefValue>(Val: V)) |
| 2796 | return; |
| 2797 | |
| 2798 | // There's no point in looking through other users of ConstantData for |
| 2799 | // assumptions. Confirm that we've handled them all. |
| 2800 | assert(!isa<ConstantData>(V) && "Unhandled constant data!" ); |
| 2801 | |
| 2802 | if (const auto *A = dyn_cast<Argument>(Val: V)) |
| 2803 | if (std::optional<ConstantRange> Range = A->getRange()) |
| 2804 | Known = Range->toKnownBits(); |
| 2805 | |
| 2806 | // All recursive calls that increase depth must come after this. |
| 2807 | if (Depth == MaxAnalysisRecursionDepth) |
| 2808 | return; |
| 2809 | |
| 2810 | // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has |
| 2811 | // the bits of its aliasee. |
| 2812 | if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(Val: V)) { |
| 2813 | if (!GA->isInterposable()) |
| 2814 | computeKnownBits(V: GA->getAliasee(), Known, Q, Depth: Depth + 1); |
| 2815 | return; |
| 2816 | } |
| 2817 | |
| 2818 | if (const Operator *I = dyn_cast<Operator>(Val: V)) |
| 2819 | computeKnownBitsFromOperator(I, DemandedElts, Known, Q, Depth); |
| 2820 | else if (const GlobalValue *GV = dyn_cast<GlobalValue>(Val: V)) { |
| 2821 | if (std::optional<ConstantRange> CR = GV->getAbsoluteSymbolRange()) |
| 2822 | Known = CR->toKnownBits(); |
| 2823 | } |
| 2824 | |
| 2825 | // Aligned pointers have trailing zeros - refine Known.Zero set |
| 2826 | if (isa<PointerType>(Val: V->getType())) { |
| 2827 | Align Alignment = V->getPointerAlignment(DL: Q.DL); |
| 2828 | Known.Zero.setLowBits(Log2(A: Alignment)); |
| 2829 | } |
| 2830 | |
| 2831 | // computeKnownBitsFromContext strictly refines Known. |
| 2832 | // Therefore, we run them after computeKnownBitsFromOperator. |
| 2833 | |
| 2834 | // Check whether we can determine known bits from context such as assumes. |
| 2835 | computeKnownBitsFromContext(V, Known, Q, Depth); |
| 2836 | } |
| 2837 | |
| 2838 | /// Try to detect a recurrence that the value of the induction variable is |
| 2839 | /// always a power of two (or zero). |
| 2840 | static bool isPowerOfTwoRecurrence(const PHINode *PN, bool OrZero, |
| 2841 | SimplifyQuery &Q, unsigned Depth) { |
| 2842 | BinaryOperator *BO = nullptr; |
| 2843 | Value *Start = nullptr, *Step = nullptr; |
| 2844 | if (!matchSimpleRecurrence(P: PN, BO, Start, Step)) |
| 2845 | return false; |
| 2846 | |
| 2847 | // Initial value must be a power of two. |
| 2848 | for (const Use &U : PN->operands()) { |
| 2849 | if (U.get() == Start) { |
| 2850 | // Initial value comes from a different BB, need to adjust context |
| 2851 | // instruction for analysis. |
| 2852 | Q.CtxI = PN->getIncomingBlock(U)->getTerminator(); |
| 2853 | if (!isKnownToBeAPowerOfTwo(V: Start, OrZero, Q, Depth)) |
| 2854 | return false; |
| 2855 | } |
| 2856 | } |
| 2857 | |
| 2858 | // Except for Mul, the induction variable must be on the left side of the |
| 2859 | // increment expression, otherwise its value can be arbitrary. |
| 2860 | if (BO->getOpcode() != Instruction::Mul && BO->getOperand(i_nocapture: 1) != Step) |
| 2861 | return false; |
| 2862 | |
| 2863 | Q.CtxI = BO->getParent()->getTerminator(); |
| 2864 | switch (BO->getOpcode()) { |
| 2865 | case Instruction::Mul: |
| 2866 | // Power of two is closed under multiplication. |
| 2867 | return (OrZero || Q.IIQ.hasNoUnsignedWrap(Op: BO) || |
| 2868 | Q.IIQ.hasNoSignedWrap(Op: BO)) && |
| 2869 | isKnownToBeAPowerOfTwo(V: Step, OrZero, Q, Depth); |
| 2870 | case Instruction::SDiv: |
| 2871 | // Start value must not be signmask for signed division, so simply being a |
| 2872 | // power of two is not sufficient, and it has to be a constant. |
| 2873 | if (!match(V: Start, P: m_Power2()) || match(V: Start, P: m_SignMask())) |
| 2874 | return false; |
| 2875 | [[fallthrough]]; |
| 2876 | case Instruction::UDiv: |
| 2877 | // Divisor must be a power of two. |
| 2878 | // If OrZero is false, cannot guarantee induction variable is non-zero after |
| 2879 | // division, same for Shr, unless it is exact division. |
| 2880 | return (OrZero || Q.IIQ.isExact(Op: BO)) && |
| 2881 | isKnownToBeAPowerOfTwo(V: Step, OrZero: false, Q, Depth); |
| 2882 | case Instruction::Shl: |
| 2883 | return OrZero || Q.IIQ.hasNoUnsignedWrap(Op: BO) || Q.IIQ.hasNoSignedWrap(Op: BO); |
| 2884 | case Instruction::AShr: |
| 2885 | if (!match(V: Start, P: m_Power2()) || match(V: Start, P: m_SignMask())) |
| 2886 | return false; |
| 2887 | [[fallthrough]]; |
| 2888 | case Instruction::LShr: |
| 2889 | return OrZero || Q.IIQ.isExact(Op: BO); |
| 2890 | default: |
| 2891 | return false; |
| 2892 | } |
| 2893 | } |
| 2894 | |
| 2895 | /// Return true if we can infer that \p V is known to be a power of 2 from |
| 2896 | /// dominating condition \p Cond (e.g., ctpop(V) == 1). |
| 2897 | static bool isImpliedToBeAPowerOfTwoFromCond(const Value *V, bool OrZero, |
| 2898 | const Value *Cond, |
| 2899 | bool CondIsTrue) { |
| 2900 | CmpPredicate Pred; |
| 2901 | const APInt *RHSC; |
| 2902 | if (!match(V: Cond, P: m_ICmp(Pred, L: m_Ctpop(Op0: m_Specific(V)), R: m_APInt(Res&: RHSC)))) |
| 2903 | return false; |
| 2904 | if (!CondIsTrue) |
| 2905 | Pred = ICmpInst::getInversePredicate(pred: Pred); |
| 2906 | // ctpop(V) u< 2 |
| 2907 | if (OrZero && Pred == ICmpInst::ICMP_ULT && *RHSC == 2) |
| 2908 | return true; |
| 2909 | // ctpop(V) == 1 |
| 2910 | return Pred == ICmpInst::ICMP_EQ && *RHSC == 1; |
| 2911 | } |
| 2912 | |
| 2913 | /// Return true if the given value is known to have exactly one |
| 2914 | /// bit set when defined. For vectors return true if every element is known to |
| 2915 | /// be a power of two when defined. Supports values with integer or pointer |
| 2916 | /// types and vectors of integers. |
| 2917 | bool llvm::isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, |
| 2918 | const SimplifyQuery &Q, unsigned Depth) { |
| 2919 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth" ); |
| 2920 | |
| 2921 | if (isa<Constant>(Val: V)) |
| 2922 | return OrZero ? match(V, P: m_Power2OrZero()) : match(V, P: m_Power2()); |
| 2923 | |
| 2924 | // i1 is by definition a power of 2 or zero. |
| 2925 | if (OrZero && V->getType()->getScalarSizeInBits() == 1) |
| 2926 | return true; |
| 2927 | |
| 2928 | // Try to infer from assumptions. |
| 2929 | if (Q.AC && Q.CtxI) { |
| 2930 | for (auto &AssumeVH : Q.AC->assumptionsFor(V)) { |
| 2931 | if (!AssumeVH) |
| 2932 | continue; |
| 2933 | CallInst *I = cast<CallInst>(Val&: AssumeVH); |
| 2934 | if (isImpliedToBeAPowerOfTwoFromCond(V, OrZero, Cond: I->getArgOperand(i: 0), |
| 2935 | /*CondIsTrue=*/true) && |
| 2936 | isValidAssumeForContext(I, Q)) |
| 2937 | return true; |
| 2938 | } |
| 2939 | } |
| 2940 | |
| 2941 | // Handle dominating conditions. |
| 2942 | if (Q.DC && Q.CtxI && Q.DT) { |
| 2943 | for (CondBrInst *BI : Q.DC->conditionsFor(V)) { |
| 2944 | Value *Cond = BI->getCondition(); |
| 2945 | |
| 2946 | BasicBlockEdge Edge0(BI->getParent(), BI->getSuccessor(i: 0)); |
| 2947 | if (isImpliedToBeAPowerOfTwoFromCond(V, OrZero, Cond, |
| 2948 | /*CondIsTrue=*/true) && |
| 2949 | Q.DT->dominates(BBE: Edge0, BB: Q.CtxI->getParent())) |
| 2950 | return true; |
| 2951 | |
| 2952 | BasicBlockEdge Edge1(BI->getParent(), BI->getSuccessor(i: 1)); |
| 2953 | if (isImpliedToBeAPowerOfTwoFromCond(V, OrZero, Cond, |
| 2954 | /*CondIsTrue=*/false) && |
| 2955 | Q.DT->dominates(BBE: Edge1, BB: Q.CtxI->getParent())) |
| 2956 | return true; |
| 2957 | } |
| 2958 | } |
| 2959 | |
| 2960 | auto *I = dyn_cast<Instruction>(Val: V); |
| 2961 | if (!I) |
| 2962 | return false; |
| 2963 | |
| 2964 | if (Q.CtxI && match(V, P: m_VScale())) { |
| 2965 | const Function *F = Q.CtxI->getFunction(); |
| 2966 | // The vscale_range indicates vscale is a power-of-two. |
| 2967 | return F->hasFnAttribute(Kind: Attribute::VScaleRange); |
| 2968 | } |
| 2969 | |
| 2970 | // 1 << X is clearly a power of two if the one is not shifted off the end. If |
| 2971 | // it is shifted off the end then the result is undefined. |
| 2972 | if (match(V: I, P: m_Shl(L: m_One(), R: m_Value()))) |
| 2973 | return true; |
| 2974 | |
| 2975 | // (signmask) >>l X is clearly a power of two if the one is not shifted off |
| 2976 | // the bottom. If it is shifted off the bottom then the result is undefined. |
| 2977 | if (match(V: I, P: m_LShr(L: m_SignMask(), R: m_Value()))) |
| 2978 | return true; |
| 2979 | |
| 2980 | // The remaining tests are all recursive, so bail out if we hit the limit. |
| 2981 | if (Depth++ == MaxAnalysisRecursionDepth) |
| 2982 | return false; |
| 2983 | |
| 2984 | switch (I->getOpcode()) { |
| 2985 | case Instruction::ZExt: |
| 2986 | return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), OrZero, Q, Depth); |
| 2987 | case Instruction::Trunc: |
| 2988 | return OrZero && isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), OrZero, Q, Depth); |
| 2989 | case Instruction::Shl: |
| 2990 | if (OrZero || Q.IIQ.hasNoUnsignedWrap(Op: I) || Q.IIQ.hasNoSignedWrap(Op: I)) |
| 2991 | return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), OrZero, Q, Depth); |
| 2992 | return false; |
| 2993 | case Instruction::LShr: |
| 2994 | if (OrZero || Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I))) |
| 2995 | return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), OrZero, Q, Depth); |
| 2996 | return false; |
| 2997 | case Instruction::UDiv: |
| 2998 | if (Q.IIQ.isExact(Op: cast<BinaryOperator>(Val: I))) |
| 2999 | return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), OrZero, Q, Depth); |
| 3000 | return false; |
| 3001 | case Instruction::Mul: |
| 3002 | return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 1), OrZero, Q, Depth) && |
| 3003 | isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), OrZero, Q, Depth) && |
| 3004 | (OrZero || isKnownNonZero(V: I, Q, Depth)); |
| 3005 | case Instruction::And: |
| 3006 | // A power of two and'd with anything is a power of two or zero. |
| 3007 | if (OrZero && |
| 3008 | (isKnownToBeAPowerOfTwo(V: I->getOperand(i: 1), /*OrZero*/ true, Q, Depth) || |
| 3009 | isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), /*OrZero*/ true, Q, Depth))) |
| 3010 | return true; |
| 3011 | // X & (-X) is always a power of two or zero. |
| 3012 | if (match(V: I->getOperand(i: 0), P: m_Neg(V: m_Specific(V: I->getOperand(i: 1)))) || |
| 3013 | match(V: I->getOperand(i: 1), P: m_Neg(V: m_Specific(V: I->getOperand(i: 0))))) |
| 3014 | return OrZero || isKnownNonZero(V: I->getOperand(i: 0), Q, Depth); |
| 3015 | return false; |
| 3016 | case Instruction::Add: { |
| 3017 | // Adding a power-of-two or zero to the same power-of-two or zero yields |
| 3018 | // either the original power-of-two, a larger power-of-two or zero. |
| 3019 | const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(Val: V); |
| 3020 | if (OrZero || Q.IIQ.hasNoUnsignedWrap(Op: VOBO) || |
| 3021 | Q.IIQ.hasNoSignedWrap(Op: VOBO)) { |
| 3022 | if (match(V: I->getOperand(i: 0), |
| 3023 | P: m_c_And(L: m_Specific(V: I->getOperand(i: 1)), R: m_Value())) && |
| 3024 | isKnownToBeAPowerOfTwo(V: I->getOperand(i: 1), OrZero, Q, Depth)) |
| 3025 | return true; |
| 3026 | if (match(V: I->getOperand(i: 1), |
| 3027 | P: m_c_And(L: m_Specific(V: I->getOperand(i: 0)), R: m_Value())) && |
| 3028 | isKnownToBeAPowerOfTwo(V: I->getOperand(i: 0), OrZero, Q, Depth)) |
| 3029 | return true; |
| 3030 | |
| 3031 | unsigned BitWidth = V->getType()->getScalarSizeInBits(); |
| 3032 | KnownBits LHSBits(BitWidth); |
| 3033 | computeKnownBits(V: I->getOperand(i: 0), Known&: LHSBits, Q, Depth); |
| 3034 | |
| 3035 | KnownBits RHSBits(BitWidth); |
| 3036 | computeKnownBits(V: I->getOperand(i: 1), Known&: RHSBits, Q, Depth); |
| 3037 | // If i8 V is a power of two or zero: |
| 3038 | // ZeroBits: 1 1 1 0 1 1 1 1 |
| 3039 | // ~ZeroBits: 0 0 0 1 0 0 0 0 |
| 3040 | if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2()) |
| 3041 | // If OrZero isn't set, we cannot give back a zero result. |
| 3042 | // Make sure either the LHS or RHS has a bit set. |
| 3043 | if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue()) |
| 3044 | return true; |
| 3045 | } |
| 3046 | |
| 3047 | // LShr(UINT_MAX, Y) + 1 is a power of two (if add is nuw) or zero. |
| 3048 | if (OrZero || Q.IIQ.hasNoUnsignedWrap(Op: VOBO)) |
| 3049 | if (match(V: I, P: m_Add(L: m_LShr(L: m_AllOnes(), R: m_Value()), R: m_One()))) |
| 3050 | return true; |
| 3051 | return false; |
| 3052 | } |
| 3053 | case Instruction::Select: |
| 3054 | return isKnownToBeAPowerOfTwo(V: I->getOperand(i: 1), OrZero, Q, Depth) && |
| 3055 | isKnownToBeAPowerOfTwo(V: I->getOperand(i: 2), OrZero, Q, Depth); |
| 3056 | case Instruction::PHI: { |
| 3057 | // A PHI node is power of two if all incoming values are power of two, or if |
| 3058 | // it is an induction variable where in each step its value is a power of |
| 3059 | // two. |
| 3060 | auto *PN = cast<PHINode>(Val: I); |
| 3061 | SimplifyQuery RecQ = Q.getWithoutCondContext(); |
| 3062 | |
| 3063 | // Check if it is an induction variable and always power of two. |
| 3064 | if (isPowerOfTwoRecurrence(PN, OrZero, Q&: RecQ, Depth)) |
| 3065 | return true; |
| 3066 | |
| 3067 | // Recursively check all incoming values. Limit recursion to 2 levels, so |
| 3068 | // that search complexity is limited to number of operands^2. |
| 3069 | unsigned NewDepth = std::max(a: Depth, b: MaxAnalysisRecursionDepth - 1); |
| 3070 | return llvm::all_of(Range: PN->operands(), P: [&](const Use &U) { |
| 3071 | // Value is power of 2 if it is coming from PHI node itself by induction. |
| 3072 | if (U.get() == PN) |
| 3073 | return true; |
| 3074 | |
| 3075 | // Change the context instruction to the incoming block where it is |
| 3076 | // evaluated. |
| 3077 | RecQ.CtxI = PN->getIncomingBlock(U)->getTerminator(); |
| 3078 | return isKnownToBeAPowerOfTwo(V: U.get(), OrZero, Q: RecQ, Depth: NewDepth); |
| 3079 | }); |
| 3080 | } |
| 3081 | case Instruction::Invoke: |
| 3082 | case Instruction::Call: { |
| 3083 | if (auto *II = dyn_cast<IntrinsicInst>(Val: I)) { |
| 3084 | switch (II->getIntrinsicID()) { |
| 3085 | case Intrinsic::umax: |
| 3086 | case Intrinsic::smax: |
| 3087 | case Intrinsic::umin: |
| 3088 | case Intrinsic::smin: |
| 3089 | return isKnownToBeAPowerOfTwo(V: II->getArgOperand(i: 1), OrZero, Q, Depth) && |
| 3090 | isKnownToBeAPowerOfTwo(V: II->getArgOperand(i: 0), OrZero, Q, Depth); |
| 3091 | // bswap/bitreverse just move around bits, but don't change any 1s/0s |
| 3092 | // thus dont change pow2/non-pow2 status. |
| 3093 | case Intrinsic::bitreverse: |
| 3094 | case Intrinsic::bswap: |
| 3095 | return isKnownToBeAPowerOfTwo(V: II->getArgOperand(i: 0), OrZero, Q, Depth); |
| 3096 | case Intrinsic::fshr: |
| 3097 | case Intrinsic::fshl: |
| 3098 | // If Op0 == Op1, this is a rotate. is_pow2(rotate(x, y)) == is_pow2(x) |
| 3099 | if (II->getArgOperand(i: 0) == II->getArgOperand(i: 1)) |
| 3100 | return isKnownToBeAPowerOfTwo(V: II->getArgOperand(i: 0), OrZero, Q, Depth); |
| 3101 | break; |
| 3102 | case Intrinsic::riscv_vsetvlimax: |
| 3103 | // VLMAX is VLEN * LMUL / SEW, which is always a non-zero power of two |
| 3104 | // for any valid vtype, so it is a power of two regardless of OrZero. |
| 3105 | return true; |
| 3106 | case Intrinsic::read_register: |
| 3107 | case Intrinsic::read_volatile_register: { |
| 3108 | // The RISC-V vlenb CSR holds VLEN/8, which is always a non-zero power |
| 3109 | // of two, so it is a power of two regardless of OrZero. |
| 3110 | const Module *M = II->getModule(); |
| 3111 | if (!M || !M->getTargetTriple().isRISCV()) |
| 3112 | break; |
| 3113 | return isReadVLENB(II: *II); |
| 3114 | } |
| 3115 | default: |
| 3116 | break; |
| 3117 | } |
| 3118 | } |
| 3119 | return false; |
| 3120 | } |
| 3121 | default: |
| 3122 | return false; |
| 3123 | } |
| 3124 | } |
| 3125 | |
| 3126 | /// Test whether a GEP's result is known to be non-null. |
| 3127 | /// |
| 3128 | /// Uses properties inherent in a GEP to try to determine whether it is known |
| 3129 | /// to be non-null. |
| 3130 | /// |
| 3131 | /// Currently this routine does not support vector GEPs. |
| 3132 | static bool isGEPKnownNonNull(const GEPOperator *GEP, const SimplifyQuery &Q, |
| 3133 | unsigned Depth) { |
| 3134 | const Function *F = nullptr; |
| 3135 | if (const Instruction *I = dyn_cast<Instruction>(Val: GEP)) |
| 3136 | F = I->getFunction(); |
| 3137 | |
| 3138 | // If the gep is nuw or inbounds with invalid null pointer, then the GEP |
| 3139 | // may be null iff the base pointer is null and the offset is zero. |
| 3140 | if (!GEP->hasNoUnsignedWrap() && |
| 3141 | !(GEP->isInBounds() && |
| 3142 | !NullPointerIsDefined(F, AS: GEP->getPointerAddressSpace()))) |
| 3143 | return false; |
| 3144 | |
| 3145 | // FIXME: Support vector-GEPs. |
| 3146 | assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP" ); |
| 3147 | |
| 3148 | // If the base pointer is non-null, we cannot walk to a null address with an |
| 3149 | // inbounds GEP in address space zero. |
| 3150 | if (isKnownNonZero(V: GEP->getPointerOperand(), Q, Depth)) |
| 3151 | return true; |
| 3152 | |
| 3153 | // Walk the GEP operands and see if any operand introduces a non-zero offset. |
| 3154 | // If so, then the GEP cannot produce a null pointer, as doing so would |
| 3155 | // inherently violate the inbounds contract within address space zero. |
| 3156 | for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP); |
| 3157 | GTI != GTE; ++GTI) { |
| 3158 | // Struct types are easy -- they must always be indexed by a constant. |
| 3159 | if (StructType *STy = GTI.getStructTypeOrNull()) { |
| 3160 | ConstantInt *OpC = cast<ConstantInt>(Val: GTI.getOperand()); |
| 3161 | unsigned ElementIdx = OpC->getZExtValue(); |
| 3162 | const StructLayout *SL = Q.DL.getStructLayout(Ty: STy); |
| 3163 | uint64_t ElementOffset = SL->getElementOffset(Idx: ElementIdx); |
| 3164 | if (ElementOffset > 0) |
| 3165 | return true; |
| 3166 | continue; |
| 3167 | } |
| 3168 | |
| 3169 | // If we have a zero-sized type, the index doesn't matter. Keep looping. |
| 3170 | if (GTI.getSequentialElementStride(DL: Q.DL).isZero()) |
| 3171 | continue; |
| 3172 | |
| 3173 | // Fast path the constant operand case both for efficiency and so we don't |
| 3174 | // increment Depth when just zipping down an all-constant GEP. |
| 3175 | if (ConstantInt *OpC = dyn_cast<ConstantInt>(Val: GTI.getOperand())) { |
| 3176 | if (!OpC->isZero()) |
| 3177 | return true; |
| 3178 | continue; |
| 3179 | } |
| 3180 | |
| 3181 | // We post-increment Depth here because while isKnownNonZero increments it |
| 3182 | // as well, when we pop back up that increment won't persist. We don't want |
| 3183 | // to recurse 10k times just because we have 10k GEP operands. We don't |
| 3184 | // bail completely out because we want to handle constant GEPs regardless |
| 3185 | // of depth. |
| 3186 | if (Depth++ >= MaxAnalysisRecursionDepth) |
| 3187 | continue; |
| 3188 | |
| 3189 | if (isKnownNonZero(V: GTI.getOperand(), Q, Depth)) |
| 3190 | return true; |
| 3191 | } |
| 3192 | |
| 3193 | return false; |
| 3194 | } |
| 3195 | |
| 3196 | static bool isKnownNonNullFromDominatingCondition(const Value *V, |
| 3197 | const Instruction *CtxI, |
| 3198 | const DominatorTree *DT) { |
| 3199 | assert(!isa<Constant>(V) && "Called for constant?" ); |
| 3200 | |
| 3201 | if (!CtxI || !DT) |
| 3202 | return false; |
| 3203 | |
| 3204 | unsigned NumUsesExplored = 0; |
| 3205 | for (auto &U : V->uses()) { |
| 3206 | // Avoid massive lists |
| 3207 | if (NumUsesExplored >= DomConditionsMaxUses) |
| 3208 | break; |
| 3209 | NumUsesExplored++; |
| 3210 | |
| 3211 | const Instruction *UI = cast<Instruction>(Val: U.getUser()); |
| 3212 | // If the value is used as an argument to a call or invoke, then argument |
| 3213 | // attributes may provide an answer about null-ness. |
| 3214 | if (V->getType()->isPointerTy()) { |
| 3215 | if (const auto *CB = dyn_cast<CallBase>(Val: UI)) { |
| 3216 | if (CB->isArgOperand(U: &U) && |
| 3217 | CB->paramHasNonNullAttr(ArgNo: CB->getArgOperandNo(U: &U), |
| 3218 | /*AllowUndefOrPoison=*/false) && |
| 3219 | DT->dominates(Def: CB, User: CtxI)) |
| 3220 | return true; |
| 3221 | } |
| 3222 | } |
| 3223 | |
| 3224 | // If the value is used as a load/store, then the pointer must be non null. |
| 3225 | if (V == getLoadStorePointerOperand(V: UI)) { |
| 3226 | if (!NullPointerIsDefined(F: UI->getFunction(), |
| 3227 | AS: V->getType()->getPointerAddressSpace()) && |
| 3228 | DT->dominates(Def: UI, User: CtxI)) |
| 3229 | return true; |
| 3230 | } |
| 3231 | |
| 3232 | if ((match(V: UI, P: m_IDiv(L: m_Value(), R: m_Specific(V))) || |
| 3233 | match(V: UI, P: m_IRem(L: m_Value(), R: m_Specific(V)))) && |
| 3234 | isValidAssumeForContext(Inv: UI, CtxI, DT)) |
| 3235 | return true; |
| 3236 | |
| 3237 | // Consider only compare instructions uniquely controlling a branch |
| 3238 | Value *RHS; |
| 3239 | CmpPredicate Pred; |
| 3240 | bool NonNullIfTrue; |
| 3241 | if (match(V: UI, P: m_c_ICmp(Pred, L: m_Specific(V), R: m_Value(V&: RHS)))) { |
| 3242 | if (cmpExcludesZero(Pred, RHS)) |
| 3243 | NonNullIfTrue = true; |
| 3244 | else if (cmpExcludesZero(Pred: CmpInst::getInversePredicate(pred: Pred), RHS)) |
| 3245 | NonNullIfTrue = false; |
| 3246 | else |
| 3247 | continue; |
| 3248 | } else if (UI->getType()->isIntegerTy(BitWidth: 1) && |
| 3249 | match(V: UI, P: m_Trunc(Op: m_Specific(V)))) { |
| 3250 | NonNullIfTrue = true; |
| 3251 | } else |
| 3252 | continue; |
| 3253 | |
| 3254 | SmallVector<const User *, 4> WorkList; |
| 3255 | SmallPtrSet<const User *, 4> Visited; |
| 3256 | for (const auto *CmpU : UI->users()) { |
| 3257 | assert(WorkList.empty() && "Should be!" ); |
| 3258 | if (Visited.insert(Ptr: CmpU).second) |
| 3259 | WorkList.push_back(Elt: CmpU); |
| 3260 | |
| 3261 | while (!WorkList.empty()) { |
| 3262 | auto *Curr = WorkList.pop_back_val(); |
| 3263 | |
| 3264 | // If a user is an AND, add all its users to the work list. We only |
| 3265 | // propagate "pred != null" condition through AND because it is only |
| 3266 | // correct to assume that all conditions of AND are met in true branch. |
| 3267 | // TODO: Support similar logic of OR and EQ predicate? |
| 3268 | if (NonNullIfTrue) |
| 3269 | if (match(V: Curr, P: m_LogicalAnd(L: m_Value(), R: m_Value()))) { |
| 3270 | for (const auto *CurrU : Curr->users()) |
| 3271 | if (Visited.insert(Ptr: CurrU).second) |
| 3272 | WorkList.push_back(Elt: CurrU); |
| 3273 | continue; |
| 3274 | } |
| 3275 | |
| 3276 | if (const CondBrInst *BI = dyn_cast<CondBrInst>(Val: Curr)) { |
| 3277 | BasicBlock *NonNullSuccessor = |
| 3278 | BI->getSuccessor(i: NonNullIfTrue ? 0 : 1); |
| 3279 | BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor); |
| 3280 | if (DT->dominates(BBE: Edge, BB: CtxI->getParent())) |
| 3281 | return true; |
| 3282 | } else if (NonNullIfTrue && isGuard(U: Curr) && |
| 3283 | DT->dominates(Def: cast<Instruction>(Val: Curr), User: CtxI)) { |
| 3284 | return true; |
| 3285 | } |
| 3286 | } |
| 3287 | } |
| 3288 | } |
| 3289 | |
| 3290 | return false; |
| 3291 | } |
| 3292 | |
| 3293 | /// Does the 'Range' metadata (which must be a valid MD_range operand list) |
| 3294 | /// ensure that the value it's attached to is never Value? 'RangeType' is |
| 3295 | /// is the type of the value described by the range. |
| 3296 | static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) { |
| 3297 | const unsigned NumRanges = Ranges->getNumOperands() / 2; |
| 3298 | assert(NumRanges >= 1); |
| 3299 | for (unsigned i = 0; i < NumRanges; ++i) { |
| 3300 | ConstantInt *Lower = |
| 3301 | mdconst::extract<ConstantInt>(MD: Ranges->getOperand(I: 2 * i + 0)); |
| 3302 | ConstantInt *Upper = |
| 3303 | mdconst::extract<ConstantInt>(MD: Ranges->getOperand(I: 2 * i + 1)); |
| 3304 | ConstantRange Range(Lower->getValue(), Upper->getValue()); |
| 3305 | if (Range.contains(Val: Value)) |
| 3306 | return false; |
| 3307 | } |
| 3308 | return true; |
| 3309 | } |
| 3310 | |
| 3311 | /// Try to detect a recurrence that monotonically increases/decreases from a |
| 3312 | /// non-zero starting value. These are common as induction variables. |
| 3313 | static bool isNonZeroRecurrence(const PHINode *PN) { |
| 3314 | BinaryOperator *BO = nullptr; |
| 3315 | Value *Start = nullptr, *Step = nullptr; |
| 3316 | const APInt *StartC, *StepC; |
| 3317 | if (!matchSimpleRecurrence(P: PN, BO, Start, Step) || |
| 3318 | !match(V: Start, P: m_APInt(Res&: StartC)) || StartC->isZero()) |
| 3319 | return false; |
| 3320 | |
| 3321 | switch (BO->getOpcode()) { |
| 3322 | case Instruction::Add: |
| 3323 | // Starting from non-zero and stepping away from zero can never wrap back |
| 3324 | // to zero. |
| 3325 | return BO->hasNoUnsignedWrap() || |
| 3326 | (BO->hasNoSignedWrap() && match(V: Step, P: m_APInt(Res&: StepC)) && |
| 3327 | StartC->isNegative() == StepC->isNegative()); |
| 3328 | case Instruction::Mul: |
| 3329 | return (BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap()) && |
| 3330 | match(V: Step, P: m_APInt(Res&: StepC)) && !StepC->isZero(); |
| 3331 | case Instruction::Shl: |
| 3332 | return BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap(); |
| 3333 | case Instruction::AShr: |
| 3334 | case Instruction::LShr: |
| 3335 | return BO->isExact(); |
| 3336 | case Instruction::Or: |
| 3337 | return true; |
| 3338 | default: |
| 3339 | return false; |
| 3340 | } |
| 3341 | } |
| 3342 | |
| 3343 | static bool matchOpWithOpEqZero(Value *Op0, Value *Op1) { |
| 3344 | return match(V: Op0, P: m_ZExtOrSExt(Op: m_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, |
| 3345 | L: m_Specific(V: Op1), R: m_Zero()))) || |
| 3346 | match(V: Op1, P: m_ZExtOrSExt(Op: m_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, |
| 3347 | L: m_Specific(V: Op0), R: m_Zero()))); |
| 3348 | } |
| 3349 | |
| 3350 | /// Return true if "addr(P) + Off" cannot be zero. The object P points to |
| 3351 | /// cannot cross the unsigned address space boundary, including the address one |
| 3352 | /// past its end, so an offset that stays within the object cannot wrap around |
| 3353 | /// to zero. This holds even if the object starts at the zero address, which is |
| 3354 | /// all that null_pointer_is_valid permits, so a non-zero offset is enough. |
| 3355 | static bool isNonZeroAddrPlusOffset(const SimplifyQuery &Q, Value *X, |
| 3356 | Value *Y) { |
| 3357 | const APInt *Off; |
| 3358 | if (!match(V: Y, P: m_APInt(Res&: Off)) || Off->isZero()) |
| 3359 | return false; |
| 3360 | |
| 3361 | // ptrtoaddr always produces the address. ptrtoint only does so when its |
| 3362 | // result is at least as wide as the address: a narrower result is truncated |
| 3363 | // and may wrap to zero, while bits above the address can only make the |
| 3364 | // result more non-zero. |
| 3365 | Value *P; |
| 3366 | if (!match(V: X, P: m_PtrToAddr(Op: m_Value(V&: P))) && |
| 3367 | !(match(V: X, P: m_PtrToInt(Op: m_Value(V&: P))) && |
| 3368 | X->getType()->getScalarSizeInBits() >= |
| 3369 | Q.DL.getAddressSizeInBits(Ty: P->getType()))) |
| 3370 | return false; |
| 3371 | |
| 3372 | // A vector of pointers has no single object whose size can be taken, and |
| 3373 | // getPointerDereferenceableBytes() requires a plain pointer. |
| 3374 | if (!P->getType()->isPointerTy()) |
| 3375 | return false; |
| 3376 | |
| 3377 | bool CanBeNull; |
| 3378 | uint64_t DerefBytes = P->getPointerDereferenceableBytes( |
| 3379 | DL: Q.DL, CanBeNull, /*CanBeFreed=*/nullptr); |
| 3380 | if (CanBeNull || DerefBytes == 0 || !Off->ule(RHS: DerefBytes)) |
| 3381 | return false; |
| 3382 | |
| 3383 | return true; |
| 3384 | } |
| 3385 | |
| 3386 | static bool isNonZeroAdd(const APInt &DemandedElts, const SimplifyQuery &Q, |
| 3387 | unsigned BitWidth, Value *X, Value *Y, bool NSW, |
| 3388 | bool NUW, unsigned Depth) { |
| 3389 | if (isNonZeroAddrPlusOffset(Q, X, Y) || isNonZeroAddrPlusOffset(Q, X: Y, Y: X)) |
| 3390 | return true; |
| 3391 | |
| 3392 | // (X + (X != 0)) is non zero |
| 3393 | if (matchOpWithOpEqZero(Op0: X, Op1: Y)) |
| 3394 | return true; |
| 3395 | |
| 3396 | if (NUW) |
| 3397 | return isKnownNonZero(V: Y, DemandedElts, Q, Depth) || |
| 3398 | isKnownNonZero(V: X, DemandedElts, Q, Depth); |
| 3399 | |
| 3400 | KnownBits XKnown = computeKnownBits(V: X, DemandedElts, Q, Depth); |
| 3401 | KnownBits YKnown = computeKnownBits(V: Y, DemandedElts, Q, Depth); |
| 3402 | |
| 3403 | // If X and Y are both non-negative (as signed values) then their sum is not |
| 3404 | // zero unless both X and Y are zero. |
| 3405 | if (XKnown.isNonNegative() && YKnown.isNonNegative()) |
| 3406 | if (isKnownNonZero(V: Y, DemandedElts, Q, Depth) || |
| 3407 | isKnownNonZero(V: X, DemandedElts, Q, Depth)) |
| 3408 | return true; |
| 3409 | |
| 3410 | // If X and Y are both negative (as signed values) then their sum is not |
| 3411 | // zero unless both X and Y equal INT_MIN. |
| 3412 | if (XKnown.isNegative() && YKnown.isNegative()) { |
| 3413 | APInt Mask = APInt::getSignedMaxValue(numBits: BitWidth); |
| 3414 | // The sign bit of X is set. If some other bit is set then X is not equal |
| 3415 | // to INT_MIN. |
| 3416 | if (XKnown.One.intersects(RHS: Mask)) |
| 3417 | return true; |
| 3418 | // The sign bit of Y is set. If some other bit is set then Y is not equal |
| 3419 | // to INT_MIN. |
| 3420 | if (YKnown.One.intersects(RHS: Mask)) |
| 3421 | return true; |
| 3422 | } |
| 3423 | |
| 3424 | // The sum of a non-negative number and a power of two is not zero. |
| 3425 | if (XKnown.isNonNegative() && |
| 3426 | isKnownToBeAPowerOfTwo(V: Y, /*OrZero*/ false, Q, Depth)) |
| 3427 | return true; |
| 3428 | if (YKnown.isNonNegative() && |
| 3429 | isKnownToBeAPowerOfTwo(V: X, /*OrZero*/ false, Q, Depth)) |
| 3430 | return true; |
| 3431 | |
| 3432 | return KnownBits::add(LHS: XKnown, RHS: YKnown, NSW, NUW).isNonZero(); |
| 3433 | } |
| 3434 | |
| 3435 | static bool isNonZeroSub(const APInt &DemandedElts, const SimplifyQuery &Q, |
| 3436 | unsigned BitWidth, Value *X, Value *Y, |
| 3437 | unsigned Depth) { |
| 3438 | // (X - (X != 0)) is non zero |
| 3439 | // ((X != 0) - X) is non zero |
| 3440 | if (matchOpWithOpEqZero(Op0: X, Op1: Y)) |
| 3441 | return true; |
| 3442 | |
| 3443 | // TODO: Move this case into isKnownNonEqual(). |
| 3444 | if (auto *C = dyn_cast<Constant>(Val: X)) |
| 3445 | if (C->isNullValue() && isKnownNonZero(V: Y, DemandedElts, Q, Depth)) |
| 3446 | return true; |
| 3447 | |
| 3448 | return ::isKnownNonEqual(V1: X, V2: Y, DemandedElts, Q, Depth); |
| 3449 | } |
| 3450 | |
| 3451 | static bool isNonZeroMul(const APInt &DemandedElts, const SimplifyQuery &Q, |
| 3452 | unsigned BitWidth, Value *X, Value *Y, bool NSW, |
| 3453 | bool NUW, unsigned Depth) { |
| 3454 | // If X and Y are non-zero then so is X * Y as long as the multiplication |
| 3455 | // does not overflow. |
| 3456 | if (NSW || NUW) |
| 3457 | return isKnownNonZero(V: X, DemandedElts, Q, Depth) && |
| 3458 | isKnownNonZero(V: Y, DemandedElts, Q, Depth); |
| 3459 | |
| 3460 | // If either X or Y is odd, then if the other is non-zero the result can't |
| 3461 | // be zero. |
| 3462 | KnownBits XKnown = computeKnownBits(V: X, DemandedElts, Q, Depth); |
| 3463 | if (XKnown.One[0]) |
| 3464 | return isKnownNonZero(V: Y, DemandedElts, Q, Depth); |
| 3465 | |
| 3466 | KnownBits YKnown = computeKnownBits(V: Y, DemandedElts, Q, Depth); |
| 3467 | if (YKnown.One[0]) |
| 3468 | return XKnown.isNonZero() || isKnownNonZero(V: X, DemandedElts, Q, Depth); |
| 3469 | |
| 3470 | // If there exists any subset of X (sX) and subset of Y (sY) s.t sX * sY is |
| 3471 | // non-zero, then X * Y is non-zero. We can find sX and sY by just taking |
| 3472 | // the lowest known One of X and Y. If they are non-zero, the result |
| 3473 | // must be non-zero. We can check if LSB(X) * LSB(Y) != 0 by doing |
| 3474 | // X.CountLeadingZeros + Y.CountLeadingZeros < BitWidth. |
| 3475 | return (XKnown.countMaxTrailingZeros() + YKnown.countMaxTrailingZeros()) < |
| 3476 | BitWidth; |
| 3477 | } |
| 3478 | |
| 3479 | static bool isNonZeroShift(const Operator *I, const APInt &DemandedElts, |
| 3480 | const SimplifyQuery &Q, const KnownBits &KnownVal, |
| 3481 | unsigned Depth) { |
| 3482 | auto ShiftOp = [&](const APInt &Lhs, const APInt &Rhs) { |
| 3483 | switch (I->getOpcode()) { |
| 3484 | case Instruction::Shl: |
| 3485 | return Lhs.shl(ShiftAmt: Rhs); |
| 3486 | case Instruction::LShr: |
| 3487 | return Lhs.lshr(ShiftAmt: Rhs); |
| 3488 | case Instruction::AShr: |
| 3489 | return Lhs.ashr(ShiftAmt: Rhs); |
| 3490 | default: |
| 3491 | llvm_unreachable("Unknown Shift Opcode" ); |
| 3492 | } |
| 3493 | }; |
| 3494 | |
| 3495 | auto InvShiftOp = [&](const APInt &Lhs, const APInt &Rhs) { |
| 3496 | switch (I->getOpcode()) { |
| 3497 | case Instruction::Shl: |
| 3498 | return Lhs.lshr(ShiftAmt: Rhs); |
| 3499 | case Instruction::LShr: |
| 3500 | case Instruction::AShr: |
| 3501 | return Lhs.shl(ShiftAmt: Rhs); |
| 3502 | default: |
| 3503 | llvm_unreachable("Unknown Shift Opcode" ); |
| 3504 | } |
| 3505 | }; |
| 3506 | |
| 3507 | if (KnownVal.isUnknown()) |
| 3508 | return false; |
| 3509 | |
| 3510 | KnownBits KnownCnt = |
| 3511 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Q, Depth); |
| 3512 | APInt MaxShift = KnownCnt.getMaxValue(); |
| 3513 | unsigned NumBits = KnownVal.getBitWidth(); |
| 3514 | if (MaxShift.uge(RHS: NumBits)) |
| 3515 | return false; |
| 3516 | |
| 3517 | if (!ShiftOp(KnownVal.One, MaxShift).isZero()) |
| 3518 | return true; |
| 3519 | |
| 3520 | // If all of the bits shifted out are known to be zero, and Val is known |
| 3521 | // non-zero then at least one non-zero bit must remain. |
| 3522 | if (InvShiftOp(KnownVal.Zero, NumBits - MaxShift) |
| 3523 | .eq(RHS: InvShiftOp(APInt::getAllOnes(numBits: NumBits), NumBits - MaxShift)) && |
| 3524 | isKnownNonZero(V: I->getOperand(i: 0), DemandedElts, Q, Depth)) |
| 3525 | return true; |
| 3526 | |
| 3527 | return false; |
| 3528 | } |
| 3529 | |
| 3530 | static bool isKnownNonZeroFromOperator(const Operator *I, |
| 3531 | const APInt &DemandedElts, |
| 3532 | const SimplifyQuery &Q, unsigned Depth) { |
| 3533 | unsigned BitWidth = getBitWidth(Ty: I->getType()->getScalarType(), DL: Q.DL); |
| 3534 | switch (I->getOpcode()) { |
| 3535 | case Instruction::Alloca: |
| 3536 | // Alloca never returns null, malloc might. |
| 3537 | return I->getType()->getPointerAddressSpace() == 0; |
| 3538 | case Instruction::GetElementPtr: |
| 3539 | if (I->getType()->isPointerTy()) |
| 3540 | return isGEPKnownNonNull(GEP: cast<GEPOperator>(Val: I), Q, Depth); |
| 3541 | break; |
| 3542 | case Instruction::BitCast: { |
| 3543 | // We need to be a bit careful here. We can only peek through the bitcast |
| 3544 | // if the scalar size of elements in the operand are smaller than and a |
| 3545 | // multiple of the size they are casting too. Take three cases: |
| 3546 | // |
| 3547 | // 1) Unsafe: |
| 3548 | // bitcast <2 x i16> %NonZero to <4 x i8> |
| 3549 | // |
| 3550 | // %NonZero can have 2 non-zero i16 elements, but isKnownNonZero on a |
| 3551 | // <4 x i8> requires that all 4 i8 elements be non-zero which isn't |
| 3552 | // guranteed (imagine just sign bit set in the 2 i16 elements). |
| 3553 | // |
| 3554 | // 2) Unsafe: |
| 3555 | // bitcast <4 x i3> %NonZero to <3 x i4> |
| 3556 | // |
| 3557 | // Even though the scalar size of the src (`i3`) is smaller than the |
| 3558 | // scalar size of the dst `i4`, because `i3` is not a multiple of `i4` |
| 3559 | // its possible for the `3 x i4` elements to be zero because there are |
| 3560 | // some elements in the destination that don't contain any full src |
| 3561 | // element. |
| 3562 | // |
| 3563 | // 3) Safe: |
| 3564 | // bitcast <4 x i8> %NonZero to <2 x i16> |
| 3565 | // |
| 3566 | // This is always safe as non-zero in the 4 i8 elements implies |
| 3567 | // non-zero in the combination of any two adjacent ones. Since i8 is a |
| 3568 | // multiple of i16, each i16 is guranteed to have 2 full i8 elements. |
| 3569 | // This all implies the 2 i16 elements are non-zero. |
| 3570 | Type *FromTy = I->getOperand(i: 0)->getType(); |
| 3571 | if ((FromTy->isIntOrIntVectorTy() || FromTy->isPtrOrPtrVectorTy()) && |
| 3572 | (BitWidth % getBitWidth(Ty: FromTy->getScalarType(), DL: Q.DL)) == 0) |
| 3573 | return isKnownNonZero(V: I->getOperand(i: 0), Q, Depth); |
| 3574 | } break; |
| 3575 | case Instruction::IntToPtr: |
| 3576 | // Note that we have to take special care to avoid looking through |
| 3577 | // truncating casts, e.g., int2ptr/ptr2int with appropriate sizes, as well |
| 3578 | // as casts that can alter the value, e.g., AddrSpaceCasts. |
| 3579 | if (!isa<ScalableVectorType>(Val: I->getType()) && |
| 3580 | Q.DL.getTypeSizeInBits(Ty: I->getOperand(i: 0)->getType()).getFixedValue() <= |
| 3581 | Q.DL.getTypeSizeInBits(Ty: I->getType()).getFixedValue()) |
| 3582 | return isKnownNonZero(V: I->getOperand(i: 0), DemandedElts, Q, Depth); |
| 3583 | break; |
| 3584 | case Instruction::PtrToAddr: |
| 3585 | // isKnownNonZero() for pointers refers to the address bits being non-zero, |
| 3586 | // so we can directly forward. |
| 3587 | return isKnownNonZero(V: I->getOperand(i: 0), DemandedElts, Q, Depth); |
| 3588 | case Instruction::PtrToInt: |
| 3589 | // For inttoptr, make sure the result size is >= the address size. If the |
| 3590 | // address is non-zero, any larger value is also non-zero. |
| 3591 | if (Q.DL.getAddressSizeInBits(Ty: I->getOperand(i: 0)->getType()) <= |
| 3592 | I->getType()->getScalarSizeInBits()) |
| 3593 | return isKnownNonZero(V: I->getOperand(i: 0), DemandedElts, Q, Depth); |
| 3594 | break; |
| 3595 | case Instruction::Trunc: |
| 3596 | // nuw/nsw trunc preserves zero/non-zero status of input. |
| 3597 | if (auto *TI = dyn_cast<TruncInst>(Val: I)) |
| 3598 | if (TI->hasNoSignedWrap() || TI->hasNoUnsignedWrap()) |
| 3599 | return isKnownNonZero(V: TI->getOperand(i_nocapture: 0), DemandedElts, Q, Depth); |
| 3600 | break; |
| 3601 | |
| 3602 | // Iff x - y != 0, then x ^ y != 0 |
| 3603 | // Therefore we can do the same exact checks |
| 3604 | case Instruction::Xor: |
| 3605 | case Instruction::Sub: |
| 3606 | return isNonZeroSub(DemandedElts, Q, BitWidth, X: I->getOperand(i: 0), |
| 3607 | Y: I->getOperand(i: 1), Depth); |
| 3608 | case Instruction::Or: |
| 3609 | // (X | (X != 0)) is non zero |
| 3610 | if (matchOpWithOpEqZero(Op0: I->getOperand(i: 0), Op1: I->getOperand(i: 1))) |
| 3611 | return true; |
| 3612 | // X | Y != 0 if X != Y. |
| 3613 | if (isKnownNonEqual(V1: I->getOperand(i: 0), V2: I->getOperand(i: 1), DemandedElts, Q, |
| 3614 | Depth)) |
| 3615 | return true; |
| 3616 | // X | Y != 0 if X != 0 or Y != 0. |
| 3617 | return isKnownNonZero(V: I->getOperand(i: 1), DemandedElts, Q, Depth) || |
| 3618 | isKnownNonZero(V: I->getOperand(i: 0), DemandedElts, Q, Depth); |
| 3619 | case Instruction::SExt: |
| 3620 | case Instruction::ZExt: |
| 3621 | // ext X != 0 if X != 0. |
| 3622 | return isKnownNonZero(V: I->getOperand(i: 0), DemandedElts, Q, Depth); |
| 3623 | |
| 3624 | case Instruction::Shl: { |
| 3625 | // shl nsw/nuw can't remove any non-zero bits. |
| 3626 | const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(Val: I); |
| 3627 | if (Q.IIQ.hasNoUnsignedWrap(Op: BO) || Q.IIQ.hasNoSignedWrap(Op: BO)) |
| 3628 | return isKnownNonZero(V: I->getOperand(i: 0), DemandedElts, Q, Depth); |
| 3629 | |
| 3630 | // shl X, Y != 0 if X is odd. Note that the value of the shift is undefined |
| 3631 | // if the lowest bit is shifted off the end. |
| 3632 | KnownBits Known(BitWidth); |
| 3633 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Known, Q, Depth); |
| 3634 | if (Known.One[0]) |
| 3635 | return true; |
| 3636 | |
| 3637 | return isNonZeroShift(I, DemandedElts, Q, KnownVal: Known, Depth); |
| 3638 | } |
| 3639 | case Instruction::LShr: |
| 3640 | case Instruction::AShr: { |
| 3641 | // shr exact can only shift out zero bits. |
| 3642 | const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(Val: I); |
| 3643 | if (BO->isExact()) |
| 3644 | return isKnownNonZero(V: I->getOperand(i: 0), DemandedElts, Q, Depth); |
| 3645 | |
| 3646 | // shr X, Y != 0 if X is negative. Note that the value of the shift is not |
| 3647 | // defined if the sign bit is shifted off the end. |
| 3648 | KnownBits Known = |
| 3649 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Q, Depth); |
| 3650 | if (Known.isNegative()) |
| 3651 | return true; |
| 3652 | |
| 3653 | // shr (add nuw A, B), C is non-zero if A or B has a known-one bit at |
| 3654 | // position >= C, because the sum >= max(A, B). |
| 3655 | Value *A, *B; |
| 3656 | const APInt *C; |
| 3657 | if (Depth + 1 < MaxAnalysisRecursionDepth && |
| 3658 | match(V: I->getOperand(i: 0), P: m_NUWAdd(L: m_Value(V&: A), R: m_Value(V&: B))) && |
| 3659 | match(V: I->getOperand(i: 1), P: m_APInt(Res&: C)) && C->ult(RHS: BitWidth)) { |
| 3660 | KnownBits KnownA = computeKnownBits(V: A, DemandedElts, Q, Depth: Depth + 1); |
| 3661 | if (!KnownA.One.lshr(ShiftAmt: *C).isZero()) |
| 3662 | return true; |
| 3663 | KnownBits KnownB = computeKnownBits(V: B, DemandedElts, Q, Depth: Depth + 1); |
| 3664 | if (!KnownB.One.lshr(ShiftAmt: *C).isZero()) |
| 3665 | return true; |
| 3666 | } |
| 3667 | |
| 3668 | return isNonZeroShift(I, DemandedElts, Q, KnownVal: Known, Depth); |
| 3669 | } |
| 3670 | case Instruction::UDiv: |
| 3671 | case Instruction::SDiv: { |
| 3672 | // X / Y |
| 3673 | // div exact can only produce a zero if the dividend is zero. |
| 3674 | if (cast<PossiblyExactOperator>(Val: I)->isExact()) |
| 3675 | return isKnownNonZero(V: I->getOperand(i: 0), DemandedElts, Q, Depth); |
| 3676 | |
| 3677 | KnownBits XKnown = |
| 3678 | computeKnownBits(V: I->getOperand(i: 0), DemandedElts, Q, Depth); |
| 3679 | // If X is fully unknown we won't be able to figure anything out so don't |
| 3680 | // both computing knownbits for Y. |
| 3681 | if (XKnown.isUnknown()) |
| 3682 | return false; |
| 3683 | |
| 3684 | KnownBits YKnown = |
| 3685 | computeKnownBits(V: I->getOperand(i: 1), DemandedElts, Q, Depth); |
| 3686 | if (I->getOpcode() == Instruction::SDiv) { |
| 3687 | // For signed division need to compare abs value of the operands. |
| 3688 | XKnown = XKnown.abs(/*IntMinIsPoison*/ false); |
| 3689 | YKnown = YKnown.abs(/*IntMinIsPoison*/ false); |
| 3690 | } |
| 3691 | // If X u>= Y then div is non zero (0/0 is UB). |
| 3692 | std::optional<bool> XUgeY = KnownBits::uge(LHS: XKnown, RHS: YKnown); |
| 3693 | // If X is total unknown or X u< Y we won't be able to prove non-zero |
| 3694 | // with compute known bits so just return early. |
| 3695 | return XUgeY && *XUgeY; |
| 3696 | } |
| 3697 | case Instruction::Add: { |
| 3698 | // X + Y. |
| 3699 | |
| 3700 | // If Add has nuw wrap flag, then if either X or Y is non-zero the result is |
| 3701 | // non-zero. |
| 3702 | auto *BO = cast<OverflowingBinaryOperator>(Val: I); |
| 3703 | return isNonZeroAdd(DemandedElts, Q, BitWidth, X: I->getOperand(i: 0), |
| 3704 | Y: I->getOperand(i: 1), NSW: Q.IIQ.hasNoSignedWrap(Op: BO), |
| 3705 | NUW: Q.IIQ.hasNoUnsignedWrap(Op: BO), Depth); |
| 3706 | } |
| 3707 | case Instruction::Mul: { |
| 3708 | const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(Val: I); |
| 3709 | return isNonZeroMul(DemandedElts, Q, BitWidth, X: I->getOperand(i: 0), |
| 3710 | Y: I->getOperand(i: 1), NSW: Q.IIQ.hasNoSignedWrap(Op: BO), |
| 3711 | NUW: Q.IIQ.hasNoUnsignedWrap(Op: BO), Depth); |
| 3712 | } |
| 3713 | case Instruction::Select: { |
| 3714 | // (C ? X : Y) != 0 if X != 0 and Y != 0. |
| 3715 | |
| 3716 | // First check if the arm is non-zero using `isKnownNonZero`. If that fails, |
| 3717 | // then see if the select condition implies the arm is non-zero. For example |
| 3718 | // (X != 0 ? X : Y), we know the true arm is non-zero as the `X` "return" is |
| 3719 | // dominated by `X != 0`. |
| 3720 | auto SelectArmIsNonZero = [&](bool IsTrueArm) { |
| 3721 | Value *Op; |
| 3722 | Op = IsTrueArm ? I->getOperand(i: 1) : I->getOperand(i: 2); |
| 3723 | // Op is trivially non-zero. |
| 3724 | if (isKnownNonZero(V: Op, DemandedElts, Q, Depth)) |
| 3725 | return true; |
| 3726 | |
| 3727 | // The condition of the select dominates the true/false arm. Check if the |
| 3728 | // condition implies that a given arm is non-zero. |
| 3729 | Value *X; |
| 3730 | CmpPredicate Pred; |
| 3731 | if (!match(V: I->getOperand(i: 0), P: m_c_ICmp(Pred, L: m_Specific(V: Op), R: m_Value(V&: X)))) |
| 3732 | return false; |
| 3733 | |
| 3734 | if (!IsTrueArm) |
| 3735 | Pred = ICmpInst::getInversePredicate(pred: Pred); |
| 3736 | |
| 3737 | return cmpExcludesZero(Pred, RHS: X); |
| 3738 | }; |
| 3739 | |
| 3740 | if (SelectArmIsNonZero(/* IsTrueArm */ true) && |
| 3741 | SelectArmIsNonZero(/* IsTrueArm */ false)) |
| 3742 | return true; |
| 3743 | break; |
| 3744 | } |
| 3745 | case Instruction::PHI: { |
| 3746 | auto *PN = cast<PHINode>(Val: I); |
| 3747 | if (Q.IIQ.UseInstrInfo && isNonZeroRecurrence(PN)) |
| 3748 | return true; |
| 3749 | |
| 3750 | // Check if all incoming values are non-zero using recursion. |
| 3751 | SimplifyQuery RecQ = Q.getWithoutCondContext(); |
| 3752 | unsigned NewDepth = std::max(a: Depth, b: MaxAnalysisRecursionDepth - 1); |
| 3753 | return llvm::all_of(Range: PN->operands(), P: [&](const Use &U) { |
| 3754 | if (U.get() == PN) |
| 3755 | return true; |
| 3756 | RecQ.CtxI = PN->getIncomingBlock(U)->getTerminator(); |
| 3757 | // Check if the branch on the phi excludes zero. |
| 3758 | CmpPredicate Pred; |
| 3759 | Value *X; |
| 3760 | BasicBlock *TrueSucc, *FalseSucc; |
| 3761 | if (match(V: RecQ.CtxI, |
| 3762 | P: m_Br(C: m_c_ICmp(Pred, L: m_Specific(V: U.get()), R: m_Value(V&: X)), |
| 3763 | T: m_BasicBlock(V&: TrueSucc), F: m_BasicBlock(V&: FalseSucc)))) { |
| 3764 | // Check for cases of duplicate successors. |
| 3765 | if ((TrueSucc == PN->getParent()) != (FalseSucc == PN->getParent())) { |
| 3766 | // If we're using the false successor, invert the predicate. |
| 3767 | if (FalseSucc == PN->getParent()) |
| 3768 | Pred = CmpInst::getInversePredicate(pred: Pred); |
| 3769 | if (cmpExcludesZero(Pred, RHS: X)) |
| 3770 | return true; |
| 3771 | } |
| 3772 | } |
| 3773 | // Finally recurse on the edge and check it directly. |
| 3774 | return isKnownNonZero(V: U.get(), DemandedElts, Q: RecQ, Depth: NewDepth); |
| 3775 | }); |
| 3776 | } |
| 3777 | case Instruction::InsertElement: { |
| 3778 | if (isa<ScalableVectorType>(Val: I->getType())) |
| 3779 | break; |
| 3780 | |
| 3781 | const Value *Vec = I->getOperand(i: 0); |
| 3782 | const Value *Elt = I->getOperand(i: 1); |
| 3783 | auto *CIdx = dyn_cast<ConstantInt>(Val: I->getOperand(i: 2)); |
| 3784 | |
| 3785 | unsigned NumElts = DemandedElts.getBitWidth(); |
| 3786 | APInt DemandedVecElts = DemandedElts; |
| 3787 | bool SkipElt = false; |
| 3788 | // If we know the index we are inserting too, clear it from Vec check. |
| 3789 | if (CIdx && CIdx->getValue().ult(RHS: NumElts)) { |
| 3790 | DemandedVecElts.clearBit(BitPosition: CIdx->getZExtValue()); |
| 3791 | SkipElt = !DemandedElts[CIdx->getZExtValue()]; |
| 3792 | } |
| 3793 | |
| 3794 | // Result is zero if Elt is non-zero and rest of the demanded elts in Vec |
| 3795 | // are non-zero. |
| 3796 | return (SkipElt || isKnownNonZero(V: Elt, Q, Depth)) && |
| 3797 | (DemandedVecElts.isZero() || |
| 3798 | isKnownNonZero(V: Vec, DemandedElts: DemandedVecElts, Q, Depth)); |
| 3799 | } |
| 3800 | case Instruction::ExtractElement: |
| 3801 | if (const auto *EEI = dyn_cast<ExtractElementInst>(Val: I)) { |
| 3802 | const Value *Vec = EEI->getVectorOperand(); |
| 3803 | const Value *Idx = EEI->getIndexOperand(); |
| 3804 | auto *CIdx = dyn_cast<ConstantInt>(Val: Idx); |
| 3805 | if (auto *VecTy = dyn_cast<FixedVectorType>(Val: Vec->getType())) { |
| 3806 | unsigned NumElts = VecTy->getNumElements(); |
| 3807 | APInt DemandedVecElts = APInt::getAllOnes(numBits: NumElts); |
| 3808 | if (CIdx && CIdx->getValue().ult(RHS: NumElts)) |
| 3809 | DemandedVecElts = APInt::getOneBitSet(numBits: NumElts, BitNo: CIdx->getZExtValue()); |
| 3810 | return isKnownNonZero(V: Vec, DemandedElts: DemandedVecElts, Q, Depth); |
| 3811 | } |
| 3812 | } |
| 3813 | break; |
| 3814 | case Instruction::ShuffleVector: { |
| 3815 | auto *Shuf = dyn_cast<ShuffleVectorInst>(Val: I); |
| 3816 | if (!Shuf) |
| 3817 | break; |
| 3818 | APInt DemandedLHS, DemandedRHS; |
| 3819 | // For undef elements, we don't know anything about the common state of |
| 3820 | // the shuffle result. |
| 3821 | if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) |
| 3822 | break; |
| 3823 | // If demanded elements for both vecs are non-zero, the shuffle is non-zero. |
| 3824 | return (DemandedRHS.isZero() || |
| 3825 | isKnownNonZero(V: Shuf->getOperand(i_nocapture: 1), DemandedElts: DemandedRHS, Q, Depth)) && |
| 3826 | (DemandedLHS.isZero() || |
| 3827 | isKnownNonZero(V: Shuf->getOperand(i_nocapture: 0), DemandedElts: DemandedLHS, Q, Depth)); |
| 3828 | } |
| 3829 | case Instruction::Freeze: |
| 3830 | return isKnownNonZero(V: I->getOperand(i: 0), Q, Depth) && |
| 3831 | isGuaranteedNotToBePoison(V: I->getOperand(i: 0), AC: Q.AC, CtxI: Q.CtxI, DT: Q.DT, |
| 3832 | Depth); |
| 3833 | case Instruction::Load: { |
| 3834 | auto *LI = cast<LoadInst>(Val: I); |
| 3835 | // A Load tagged with nonnull or dereferenceable with null pointer undefined |
| 3836 | // is never null. |
| 3837 | if (auto *PtrT = dyn_cast<PointerType>(Val: I->getType())) { |
| 3838 | if (Q.IIQ.getMetadata(I: LI, KindID: LLVMContext::MD_nonnull) || |
| 3839 | (Q.IIQ.getMetadata(I: LI, KindID: LLVMContext::MD_dereferenceable) && |
| 3840 | !NullPointerIsDefined(F: LI->getFunction(), AS: PtrT->getAddressSpace()))) |
| 3841 | return true; |
| 3842 | } else if (MDNode *Ranges = Q.IIQ.getMetadata(I: LI, KindID: LLVMContext::MD_range)) { |
| 3843 | return rangeMetadataExcludesValue(Ranges, Value: APInt::getZero(numBits: BitWidth)); |
| 3844 | } |
| 3845 | |
| 3846 | // No need to fall through to computeKnownBits as range metadata is already |
| 3847 | // handled in isKnownNonZero. |
| 3848 | return false; |
| 3849 | } |
| 3850 | case Instruction::ExtractValue: { |
| 3851 | const WithOverflowInst *WO; |
| 3852 | if (match(V: I, P: m_ExtractValue<0>(V: m_WithOverflowInst(I&: WO)))) { |
| 3853 | switch (WO->getBinaryOp()) { |
| 3854 | default: |
| 3855 | break; |
| 3856 | case Instruction::Add: |
| 3857 | return isNonZeroAdd(DemandedElts, Q, BitWidth, X: WO->getArgOperand(i: 0), |
| 3858 | Y: WO->getArgOperand(i: 1), |
| 3859 | /*NSW=*/false, |
| 3860 | /*NUW=*/false, Depth); |
| 3861 | case Instruction::Sub: |
| 3862 | return isNonZeroSub(DemandedElts, Q, BitWidth, X: WO->getArgOperand(i: 0), |
| 3863 | Y: WO->getArgOperand(i: 1), Depth); |
| 3864 | case Instruction::Mul: |
| 3865 | return isNonZeroMul(DemandedElts, Q, BitWidth, X: WO->getArgOperand(i: 0), |
| 3866 | Y: WO->getArgOperand(i: 1), |
| 3867 | /*NSW=*/false, /*NUW=*/false, Depth); |
| 3868 | break; |
| 3869 | } |
| 3870 | } |
| 3871 | break; |
| 3872 | } |
| 3873 | case Instruction::Call: |
| 3874 | case Instruction::Invoke: { |
| 3875 | const auto *Call = cast<CallBase>(Val: I); |
| 3876 | if (I->getType()->isPointerTy()) { |
| 3877 | if (Call->isReturnNonNull()) |
| 3878 | return true; |
| 3879 | if (const auto *RP = getArgumentAliasingToReturnedPointer( |
| 3880 | Call, /*MustPreserveOffset=*/true)) |
| 3881 | return isKnownNonZero(V: RP, Q, Depth); |
| 3882 | } else { |
| 3883 | if (MDNode *Ranges = Q.IIQ.getMetadata(I: Call, KindID: LLVMContext::MD_range)) |
| 3884 | return rangeMetadataExcludesValue(Ranges, Value: APInt::getZero(numBits: BitWidth)); |
| 3885 | if (std::optional<ConstantRange> Range = Call->getRange()) { |
| 3886 | const APInt ZeroValue(Range->getBitWidth(), 0); |
| 3887 | if (!Range->contains(Val: ZeroValue)) |
| 3888 | return true; |
| 3889 | } |
| 3890 | if (const Value *RV = Call->getReturnedArgOperand()) |
| 3891 | if (RV->getType() == I->getType() && isKnownNonZero(V: RV, Q, Depth)) |
| 3892 | return true; |
| 3893 | } |
| 3894 | |
| 3895 | if (auto *II = dyn_cast<IntrinsicInst>(Val: I)) { |
| 3896 | switch (II->getIntrinsicID()) { |
| 3897 | case Intrinsic::sshl_sat: |
| 3898 | case Intrinsic::ushl_sat: |
| 3899 | case Intrinsic::abs: |
| 3900 | case Intrinsic::bitreverse: |
| 3901 | case Intrinsic::bswap: |
| 3902 | case Intrinsic::ctpop: |
| 3903 | return isKnownNonZero(V: II->getArgOperand(i: 0), DemandedElts, Q, Depth); |
| 3904 | // NB: We don't do usub_sat here as in any case we can prove its |
| 3905 | // non-zero, we will fold it to `sub nuw` in InstCombine. |
| 3906 | case Intrinsic::ssub_sat: |
| 3907 | // For most types, if x != y then ssub.sat x, y != 0. But |
| 3908 | // ssub.sat.i1 0, -1 = 0, because 1 saturates to 0. This means |
| 3909 | // isNonZeroSub will do the wrong thing for ssub.sat.i1. |
| 3910 | if (BitWidth == 1) |
| 3911 | return false; |
| 3912 | return isNonZeroSub(DemandedElts, Q, BitWidth, X: II->getArgOperand(i: 0), |
| 3913 | Y: II->getArgOperand(i: 1), Depth); |
| 3914 | case Intrinsic::sadd_sat: |
| 3915 | return isNonZeroAdd(DemandedElts, Q, BitWidth, X: II->getArgOperand(i: 0), |
| 3916 | Y: II->getArgOperand(i: 1), |
| 3917 | /*NSW=*/true, /* NUW=*/false, Depth); |
| 3918 | // Vec reverse preserves zero/non-zero status from input vec. |
| 3919 | case Intrinsic::vector_reverse: |
| 3920 | return isKnownNonZero(V: II->getArgOperand(i: 0), DemandedElts: DemandedElts.reverseBits(), |
| 3921 | Q, Depth); |
| 3922 | // umin/smin/smax/smin/or of all non-zero elements is always non-zero. |
| 3923 | case Intrinsic::vector_reduce_or: |
| 3924 | case Intrinsic::vector_reduce_umax: |
| 3925 | case Intrinsic::vector_reduce_umin: |
| 3926 | case Intrinsic::vector_reduce_smax: |
| 3927 | case Intrinsic::vector_reduce_smin: |
| 3928 | return isKnownNonZero(V: II->getArgOperand(i: 0), Q, Depth); |
| 3929 | case Intrinsic::umax: |
| 3930 | case Intrinsic::uadd_sat: |
| 3931 | // umax(X, (X != 0)) is non zero |
| 3932 | // X +usat (X != 0) is non zero |
| 3933 | if (matchOpWithOpEqZero(Op0: II->getArgOperand(i: 0), Op1: II->getArgOperand(i: 1))) |
| 3934 | return true; |
| 3935 | |
| 3936 | return isKnownNonZero(V: II->getArgOperand(i: 1), DemandedElts, Q, Depth) || |
| 3937 | isKnownNonZero(V: II->getArgOperand(i: 0), DemandedElts, Q, Depth); |
| 3938 | case Intrinsic::smax: { |
| 3939 | // If either arg is strictly positive the result is non-zero. Otherwise |
| 3940 | // the result is non-zero if both ops are non-zero. |
| 3941 | auto IsNonZero = [&](Value *Op, std::optional<bool> &OpNonZero, |
| 3942 | const KnownBits &OpKnown) { |
| 3943 | if (!OpNonZero.has_value()) |
| 3944 | OpNonZero = OpKnown.isNonZero() || |
| 3945 | isKnownNonZero(V: Op, DemandedElts, Q, Depth); |
| 3946 | return *OpNonZero; |
| 3947 | }; |
| 3948 | // Avoid re-computing isKnownNonZero. |
| 3949 | std::optional<bool> Op0NonZero, Op1NonZero; |
| 3950 | KnownBits Op1Known = |
| 3951 | computeKnownBits(V: II->getArgOperand(i: 1), DemandedElts, Q, Depth); |
| 3952 | if (Op1Known.isNonNegative() && |
| 3953 | IsNonZero(II->getArgOperand(i: 1), Op1NonZero, Op1Known)) |
| 3954 | return true; |
| 3955 | KnownBits Op0Known = |
| 3956 | computeKnownBits(V: II->getArgOperand(i: 0), DemandedElts, Q, Depth); |
| 3957 | if (Op0Known.isNonNegative() && |
| 3958 | IsNonZero(II->getArgOperand(i: 0), Op0NonZero, Op0Known)) |
| 3959 | return true; |
| 3960 | return IsNonZero(II->getArgOperand(i: 1), Op1NonZero, Op1Known) && |
| 3961 | IsNonZero(II->getArgOperand(i: 0), Op0NonZero, Op0Known); |
| 3962 | } |
| 3963 | case Intrinsic::smin: { |
| 3964 | // If either arg is negative the result is non-zero. Otherwise |
| 3965 | // the result is non-zero if both ops are non-zero. |
| 3966 | KnownBits Op1Known = |
| 3967 | computeKnownBits(V: II->getArgOperand(i: 1), DemandedElts, Q, Depth); |
| 3968 | if (Op1Known.isNegative()) |
| 3969 | return true; |
| 3970 | KnownBits Op0Known = |
| 3971 | computeKnownBits(V: II->getArgOperand(i: 0), DemandedElts, Q, Depth); |
| 3972 | if (Op0Known.isNegative()) |
| 3973 | return true; |
| 3974 | |
| 3975 | if (Op1Known.isNonZero() && Op0Known.isNonZero()) |
| 3976 | return true; |
| 3977 | } |
| 3978 | [[fallthrough]]; |
| 3979 | case Intrinsic::umin: |
| 3980 | return isKnownNonZero(V: II->getArgOperand(i: 0), DemandedElts, Q, Depth) && |
| 3981 | isKnownNonZero(V: II->getArgOperand(i: 1), DemandedElts, Q, Depth); |
| 3982 | case Intrinsic::cttz: |
| 3983 | return computeKnownBits(V: II->getArgOperand(i: 0), DemandedElts, Q, Depth) |
| 3984 | .Zero[0]; |
| 3985 | case Intrinsic::ctlz: |
| 3986 | return computeKnownBits(V: II->getArgOperand(i: 0), DemandedElts, Q, Depth) |
| 3987 | .isNonNegative(); |
| 3988 | case Intrinsic::fshr: |
| 3989 | case Intrinsic::fshl: |
| 3990 | // If Op0 == Op1, this is a rotate. rotate(x, y) != 0 iff x != 0. |
| 3991 | if (II->getArgOperand(i: 0) == II->getArgOperand(i: 1)) |
| 3992 | return isKnownNonZero(V: II->getArgOperand(i: 0), DemandedElts, Q, Depth); |
| 3993 | break; |
| 3994 | case Intrinsic::vscale: |
| 3995 | return true; |
| 3996 | case Intrinsic::experimental_get_vector_length: |
| 3997 | return isKnownNonZero(V: I->getOperand(i: 0), Q, Depth); |
| 3998 | default: |
| 3999 | break; |
| 4000 | } |
| 4001 | break; |
| 4002 | } |
| 4003 | |
| 4004 | return false; |
| 4005 | } |
| 4006 | } |
| 4007 | |
| 4008 | KnownBits Known(BitWidth); |
| 4009 | computeKnownBits(V: I, DemandedElts, Known, Q, Depth); |
| 4010 | return Known.One != 0; |
| 4011 | } |
| 4012 | |
| 4013 | /// Return true if the given value is known to be non-zero when defined. For |
| 4014 | /// vectors, return true if every demanded element is known to be non-zero when |
| 4015 | /// defined. For pointers, if the context instruction and dominator tree are |
| 4016 | /// specified, perform context-sensitive analysis and return true if the |
| 4017 | /// pointer couldn't possibly be null at the specified instruction. |
| 4018 | /// Supports values with integer or pointer type and vectors of integers. |
| 4019 | bool isKnownNonZero(const Value *V, const APInt &DemandedElts, |
| 4020 | const SimplifyQuery &Q, unsigned Depth) { |
| 4021 | Type *Ty = V->getType(); |
| 4022 | |
| 4023 | #ifndef NDEBUG |
| 4024 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth" ); |
| 4025 | |
| 4026 | if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) { |
| 4027 | assert( |
| 4028 | FVTy->getNumElements() == DemandedElts.getBitWidth() && |
| 4029 | "DemandedElt width should equal the fixed vector number of elements" ); |
| 4030 | } else { |
| 4031 | assert(DemandedElts == APInt(1, 1) && |
| 4032 | "DemandedElt width should be 1 for scalars" ); |
| 4033 | } |
| 4034 | #endif |
| 4035 | |
| 4036 | if (auto *C = dyn_cast<Constant>(Val: V)) { |
| 4037 | if (C->isNullValue()) |
| 4038 | return false; |
| 4039 | if (isa<ConstantInt>(Val: C)) |
| 4040 | // Must be non-zero due to null test above. |
| 4041 | return true; |
| 4042 | |
| 4043 | // For constant vectors, check that all elements are poison or known |
| 4044 | // non-zero to determine that the whole vector is known non-zero. |
| 4045 | if (auto *VecTy = dyn_cast<FixedVectorType>(Val: Ty)) { |
| 4046 | for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) { |
| 4047 | if (!DemandedElts[i]) |
| 4048 | continue; |
| 4049 | Constant *Elt = C->getAggregateElement(Elt: i); |
| 4050 | if (!Elt || Elt->isNullValue()) |
| 4051 | return false; |
| 4052 | if (!isa<PoisonValue>(Val: Elt) && !isa<ConstantInt>(Val: Elt)) |
| 4053 | return false; |
| 4054 | } |
| 4055 | return true; |
| 4056 | } |
| 4057 | |
| 4058 | // Constant ptrauth can be null, iff the base pointer can be. |
| 4059 | if (auto *CPA = dyn_cast<ConstantPtrAuth>(Val: V)) |
| 4060 | return isKnownNonZero(V: CPA->getPointer(), DemandedElts, Q, Depth); |
| 4061 | |
| 4062 | // A global variable is non null unless extern weak or an absolute symbol |
| 4063 | // reference. An alias is also excluded, as it may resolve to a weak |
| 4064 | // undefined symbol and thus to the null address. Where null is a valid |
| 4065 | // address an allocation may start there, so nothing can be assumed; that |
| 4066 | // covers both null_pointer_is_valid and address spaces other than zero. |
| 4067 | if (const GlobalValue *GV = dyn_cast<GlobalValue>(Val: V)) { |
| 4068 | if (!NullPointerIsDefined(F: Q.getFunction(), |
| 4069 | AS: GV->getType()->getAddressSpace()) && |
| 4070 | !GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() && |
| 4071 | !isa<GlobalAlias>(Val: GV)) |
| 4072 | return true; |
| 4073 | } |
| 4074 | |
| 4075 | // For constant expressions, fall through to the Operator code below. |
| 4076 | if (!isa<ConstantExpr>(Val: V)) |
| 4077 | return false; |
| 4078 | } |
| 4079 | |
| 4080 | if (const auto *A = dyn_cast<Argument>(Val: V)) |
| 4081 | if (std::optional<ConstantRange> Range = A->getRange()) { |
| 4082 | const APInt ZeroValue(Range->getBitWidth(), 0); |
| 4083 | if (!Range->contains(Val: ZeroValue)) |
| 4084 | return true; |
| 4085 | } |
| 4086 | |
| 4087 | if (!isa<Constant>(Val: V) && isKnownNonZeroFromAssume(V, Q)) |
| 4088 | return true; |
| 4089 | |
| 4090 | // Some of the tests below are recursive, so bail out if we hit the limit. |
| 4091 | if (Depth++ >= MaxAnalysisRecursionDepth) |
| 4092 | return false; |
| 4093 | |
| 4094 | // Check for pointer simplifications. |
| 4095 | |
| 4096 | if (PointerType *PtrTy = dyn_cast<PointerType>(Val: Ty)) { |
| 4097 | // A byval, inalloca may not be null in a non-default addres space. A |
| 4098 | // nonnull argument is assumed never 0. |
| 4099 | if (const Argument *A = dyn_cast<Argument>(Val: V)) { |
| 4100 | if (((A->hasPassPointeeByValueCopyAttr() && |
| 4101 | !NullPointerIsDefined(F: A->getParent(), AS: PtrTy->getAddressSpace())) || |
| 4102 | A->hasNonNullAttr())) |
| 4103 | return true; |
| 4104 | } |
| 4105 | } |
| 4106 | |
| 4107 | if (const auto *I = dyn_cast<Operator>(Val: V)) |
| 4108 | if (isKnownNonZeroFromOperator(I, DemandedElts, Q, Depth)) |
| 4109 | return true; |
| 4110 | |
| 4111 | if (!isa<Constant>(Val: V) && |
| 4112 | isKnownNonNullFromDominatingCondition(V, CtxI: Q.CtxI, DT: Q.DT)) |
| 4113 | return true; |
| 4114 | |
| 4115 | if (const Value *Stripped = stripNullTest(V)) |
| 4116 | return isKnownNonZero(V: Stripped, DemandedElts, Q, Depth); |
| 4117 | |
| 4118 | return false; |
| 4119 | } |
| 4120 | |
| 4121 | bool llvm::isKnownNonZero(const Value *V, const SimplifyQuery &Q, |
| 4122 | unsigned Depth) { |
| 4123 | auto *FVTy = dyn_cast<FixedVectorType>(Val: V->getType()); |
| 4124 | APInt DemandedElts = |
| 4125 | FVTy ? APInt::getAllOnes(numBits: FVTy->getNumElements()) : APInt(1, 1); |
| 4126 | return ::isKnownNonZero(V, DemandedElts, Q, Depth); |
| 4127 | } |
| 4128 | |
| 4129 | /// If the pair of operators are the same invertible function, return the |
| 4130 | /// the operands of the function corresponding to each input. Otherwise, |
| 4131 | /// return std::nullopt. An invertible function is one that is 1-to-1 and maps |
| 4132 | /// every input value to exactly one output value. This is equivalent to |
| 4133 | /// saying that Op1 and Op2 are equal exactly when the specified pair of |
| 4134 | /// operands are equal, (except that Op1 and Op2 may be poison more often.) |
| 4135 | static std::optional<std::pair<Value*, Value*>> |
| 4136 | getInvertibleOperands(const Operator *Op1, |
| 4137 | const Operator *Op2) { |
| 4138 | if (Op1->getOpcode() != Op2->getOpcode()) |
| 4139 | return std::nullopt; |
| 4140 | |
| 4141 | auto getOperands = [&](unsigned OpNum) -> auto { |
| 4142 | return std::make_pair(x: Op1->getOperand(i: OpNum), y: Op2->getOperand(i: OpNum)); |
| 4143 | }; |
| 4144 | |
| 4145 | switch (Op1->getOpcode()) { |
| 4146 | default: |
| 4147 | break; |
| 4148 | case Instruction::Or: |
| 4149 | if (!cast<PossiblyDisjointInst>(Val: Op1)->isDisjoint() || |
| 4150 | !cast<PossiblyDisjointInst>(Val: Op2)->isDisjoint()) |
| 4151 | break; |
| 4152 | [[fallthrough]]; |
| 4153 | case Instruction::Xor: |
| 4154 | case Instruction::Add: { |
| 4155 | Value *Other; |
| 4156 | if (match(V: Op2, P: m_c_BinOp(L: m_Specific(V: Op1->getOperand(i: 0)), R: m_Value(V&: Other)))) |
| 4157 | return std::make_pair(x: Op1->getOperand(i: 1), y&: Other); |
| 4158 | if (match(V: Op2, P: m_c_BinOp(L: m_Specific(V: Op1->getOperand(i: 1)), R: m_Value(V&: Other)))) |
| 4159 | return std::make_pair(x: Op1->getOperand(i: 0), y&: Other); |
| 4160 | break; |
| 4161 | } |
| 4162 | case Instruction::Sub: |
| 4163 | if (Op1->getOperand(i: 0) == Op2->getOperand(i: 0)) |
| 4164 | return getOperands(1); |
| 4165 | if (Op1->getOperand(i: 1) == Op2->getOperand(i: 1)) |
| 4166 | return getOperands(0); |
| 4167 | break; |
| 4168 | case Instruction::Mul: { |
| 4169 | // invertible if A * B == (A * B) mod 2^N where A, and B are integers |
| 4170 | // and N is the bitwdith. The nsw case is non-obvious, but proven by |
| 4171 | // alive2: https://alive2.llvm.org/ce/z/Z6D5qK |
| 4172 | auto *OBO1 = cast<OverflowingBinaryOperator>(Val: Op1); |
| 4173 | auto *OBO2 = cast<OverflowingBinaryOperator>(Val: Op2); |
| 4174 | if ((!OBO1->hasNoUnsignedWrap() || !OBO2->hasNoUnsignedWrap()) && |
| 4175 | (!OBO1->hasNoSignedWrap() || !OBO2->hasNoSignedWrap())) |
| 4176 | break; |
| 4177 | |
| 4178 | // Assume operand order has been canonicalized |
| 4179 | if (Op1->getOperand(i: 1) == Op2->getOperand(i: 1) && |
| 4180 | isa<ConstantInt>(Val: Op1->getOperand(i: 1)) && |
| 4181 | !cast<ConstantInt>(Val: Op1->getOperand(i: 1))->isZero()) |
| 4182 | return getOperands(0); |
| 4183 | break; |
| 4184 | } |
| 4185 | case Instruction::Shl: { |
| 4186 | // Same as multiplies, with the difference that we don't need to check |
| 4187 | // for a non-zero multiply. Shifts always multiply by non-zero. |
| 4188 | auto *OBO1 = cast<OverflowingBinaryOperator>(Val: Op1); |
| 4189 | auto *OBO2 = cast<OverflowingBinaryOperator>(Val: Op2); |
| 4190 | if ((!OBO1->hasNoUnsignedWrap() || !OBO2->hasNoUnsignedWrap()) && |
| 4191 | (!OBO1->hasNoSignedWrap() || !OBO2->hasNoSignedWrap())) |
| 4192 | break; |
| 4193 | |
| 4194 | if (Op1->getOperand(i: 1) == Op2->getOperand(i: 1)) |
| 4195 | return getOperands(0); |
| 4196 | break; |
| 4197 | } |
| 4198 | case Instruction::AShr: |
| 4199 | case Instruction::LShr: { |
| 4200 | auto *PEO1 = cast<PossiblyExactOperator>(Val: Op1); |
| 4201 | auto *PEO2 = cast<PossiblyExactOperator>(Val: Op2); |
| 4202 | if (!PEO1->isExact() || !PEO2->isExact()) |
| 4203 | break; |
| 4204 | |
| 4205 | if (Op1->getOperand(i: 1) == Op2->getOperand(i: 1)) |
| 4206 | return getOperands(0); |
| 4207 | break; |
| 4208 | } |
| 4209 | case Instruction::SExt: |
| 4210 | case Instruction::ZExt: |
| 4211 | if (Op1->getOperand(i: 0)->getType() == Op2->getOperand(i: 0)->getType()) |
| 4212 | return getOperands(0); |
| 4213 | break; |
| 4214 | case Instruction::PHI: { |
| 4215 | const PHINode *PN1 = cast<PHINode>(Val: Op1); |
| 4216 | const PHINode *PN2 = cast<PHINode>(Val: Op2); |
| 4217 | |
| 4218 | // If PN1 and PN2 are both recurrences, can we prove the entire recurrences |
| 4219 | // are a single invertible function of the start values? Note that repeated |
| 4220 | // application of an invertible function is also invertible |
| 4221 | BinaryOperator *BO1 = nullptr; |
| 4222 | Value *Start1 = nullptr, *Step1 = nullptr; |
| 4223 | BinaryOperator *BO2 = nullptr; |
| 4224 | Value *Start2 = nullptr, *Step2 = nullptr; |
| 4225 | if (PN1->getParent() != PN2->getParent() || |
| 4226 | !matchSimpleRecurrence(P: PN1, BO&: BO1, Start&: Start1, Step&: Step1) || |
| 4227 | !matchSimpleRecurrence(P: PN2, BO&: BO2, Start&: Start2, Step&: Step2)) |
| 4228 | break; |
| 4229 | |
| 4230 | auto Values = getInvertibleOperands(Op1: cast<Operator>(Val: BO1), |
| 4231 | Op2: cast<Operator>(Val: BO2)); |
| 4232 | if (!Values) |
| 4233 | break; |
| 4234 | |
| 4235 | // We have to be careful of mutually defined recurrences here. Ex: |
| 4236 | // * X_i = X_(i-1) OP Y_(i-1), and Y_i = X_(i-1) OP V |
| 4237 | // * X_i = Y_i = X_(i-1) OP Y_(i-1) |
| 4238 | // The invertibility of these is complicated, and not worth reasoning |
| 4239 | // about (yet?). |
| 4240 | if (Values->first != PN1 || Values->second != PN2) |
| 4241 | break; |
| 4242 | |
| 4243 | return std::make_pair(x&: Start1, y&: Start2); |
| 4244 | } |
| 4245 | } |
| 4246 | return std::nullopt; |
| 4247 | } |
| 4248 | |
| 4249 | /// Return true if V1 == (binop V2, X), where X is known non-zero. |
| 4250 | /// Only handle a small subset of binops where (binop V2, X) with non-zero X |
| 4251 | /// implies V2 != V1. |
| 4252 | static bool isModifyingBinopOfNonZero(const Value *V1, const Value *V2, |
| 4253 | const APInt &DemandedElts, |
| 4254 | const SimplifyQuery &Q, unsigned Depth) { |
| 4255 | const BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: V1); |
| 4256 | if (!BO) |
| 4257 | return false; |
| 4258 | switch (BO->getOpcode()) { |
| 4259 | default: |
| 4260 | break; |
| 4261 | case Instruction::Or: |
| 4262 | if (!cast<PossiblyDisjointInst>(Val: V1)->isDisjoint()) |
| 4263 | break; |
| 4264 | [[fallthrough]]; |
| 4265 | case Instruction::Xor: |
| 4266 | case Instruction::Add: |
| 4267 | Value *Op = nullptr; |
| 4268 | if (V2 == BO->getOperand(i_nocapture: 0)) |
| 4269 | Op = BO->getOperand(i_nocapture: 1); |
| 4270 | else if (V2 == BO->getOperand(i_nocapture: 1)) |
| 4271 | Op = BO->getOperand(i_nocapture: 0); |
| 4272 | else |
| 4273 | return false; |
| 4274 | return isKnownNonZero(V: Op, DemandedElts, Q, Depth: Depth + 1); |
| 4275 | } |
| 4276 | return false; |
| 4277 | } |
| 4278 | |
| 4279 | /// Return true if V2 == V1 * C, where V1 is known non-zero, C is not 0/1 and |
| 4280 | /// the multiplication is nuw or nsw. |
| 4281 | static bool isNonEqualMul(const Value *V1, const Value *V2, |
| 4282 | const APInt &DemandedElts, const SimplifyQuery &Q, |
| 4283 | unsigned Depth) { |
| 4284 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(Val: V2)) { |
| 4285 | const APInt *C; |
| 4286 | return match(V: OBO, P: m_Mul(L: m_Specific(V: V1), R: m_APInt(Res&: C))) && |
| 4287 | (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) && |
| 4288 | !C->isZero() && !C->isOne() && |
| 4289 | isKnownNonZero(V: V1, DemandedElts, Q, Depth: Depth + 1); |
| 4290 | } |
| 4291 | return false; |
| 4292 | } |
| 4293 | |
| 4294 | /// Return true if V2 == V1 << C, where V1 is known non-zero, C is not 0 and |
| 4295 | /// the shift is nuw or nsw. |
| 4296 | static bool isNonEqualShl(const Value *V1, const Value *V2, |
| 4297 | const APInt &DemandedElts, const SimplifyQuery &Q, |
| 4298 | unsigned Depth) { |
| 4299 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(Val: V2)) { |
| 4300 | const APInt *C; |
| 4301 | return match(V: OBO, P: m_Shl(L: m_Specific(V: V1), R: m_APInt(Res&: C))) && |
| 4302 | (OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap()) && |
| 4303 | !C->isZero() && isKnownNonZero(V: V1, DemandedElts, Q, Depth: Depth + 1); |
| 4304 | } |
| 4305 | return false; |
| 4306 | } |
| 4307 | |
| 4308 | static bool isNonEqualPHIs(const PHINode *PN1, const PHINode *PN2, |
| 4309 | const APInt &DemandedElts, const SimplifyQuery &Q, |
| 4310 | unsigned Depth) { |
| 4311 | // Check two PHIs are in same block. |
| 4312 | if (PN1->getParent() != PN2->getParent()) |
| 4313 | return false; |
| 4314 | |
| 4315 | SmallPtrSet<const BasicBlock *, 8> VisitedBBs; |
| 4316 | bool UsedFullRecursion = false; |
| 4317 | for (const BasicBlock *IncomBB : PN1->blocks()) { |
| 4318 | if (!VisitedBBs.insert(Ptr: IncomBB).second) |
| 4319 | continue; // Don't reprocess blocks that we have dealt with already. |
| 4320 | const Value *IV1 = PN1->getIncomingValueForBlock(BB: IncomBB); |
| 4321 | const Value *IV2 = PN2->getIncomingValueForBlock(BB: IncomBB); |
| 4322 | const APInt *C1, *C2; |
| 4323 | if (match(V: IV1, P: m_APInt(Res&: C1)) && match(V: IV2, P: m_APInt(Res&: C2)) && *C1 != *C2) |
| 4324 | continue; |
| 4325 | |
| 4326 | // Only one pair of phi operands is allowed for full recursion. |
| 4327 | if (UsedFullRecursion) |
| 4328 | return false; |
| 4329 | |
| 4330 | SimplifyQuery RecQ = Q.getWithoutCondContext(); |
| 4331 | RecQ.CtxI = IncomBB->getTerminator(); |
| 4332 | if (!isKnownNonEqual(V1: IV1, V2: IV2, DemandedElts, Q: RecQ, Depth: Depth + 1)) |
| 4333 | return false; |
| 4334 | UsedFullRecursion = true; |
| 4335 | } |
| 4336 | return true; |
| 4337 | } |
| 4338 | |
| 4339 | static bool isNonEqualSelect(const Value *V1, const Value *V2, |
| 4340 | const APInt &DemandedElts, const SimplifyQuery &Q, |
| 4341 | unsigned Depth) { |
| 4342 | const SelectInst *SI1 = dyn_cast<SelectInst>(Val: V1); |
| 4343 | if (!SI1) |
| 4344 | return false; |
| 4345 | |
| 4346 | if (const SelectInst *SI2 = dyn_cast<SelectInst>(Val: V2)) { |
| 4347 | const Value *Cond1 = SI1->getCondition(); |
| 4348 | const Value *Cond2 = SI2->getCondition(); |
| 4349 | if (Cond1 == Cond2) |
| 4350 | return isKnownNonEqual(V1: SI1->getTrueValue(), V2: SI2->getTrueValue(), |
| 4351 | DemandedElts, Q, Depth: Depth + 1) && |
| 4352 | isKnownNonEqual(V1: SI1->getFalseValue(), V2: SI2->getFalseValue(), |
| 4353 | DemandedElts, Q, Depth: Depth + 1); |
| 4354 | } |
| 4355 | return isKnownNonEqual(V1: SI1->getTrueValue(), V2, DemandedElts, Q, Depth: Depth + 1) && |
| 4356 | isKnownNonEqual(V1: SI1->getFalseValue(), V2, DemandedElts, Q, Depth: Depth + 1); |
| 4357 | } |
| 4358 | |
| 4359 | // Check to see if A is both a GEP and is the incoming value for a PHI in the |
| 4360 | // loop, and B is either a ptr or another GEP. If the PHI has 2 incoming values, |
| 4361 | // one of them being the recursive GEP A and the other a ptr at same base and at |
| 4362 | // the same/higher offset than B we are only incrementing the pointer further in |
| 4363 | // loop if offset of recursive GEP is greater than 0. |
| 4364 | static bool isNonEqualPointersWithRecursiveGEP(const Value *A, const Value *B, |
| 4365 | const SimplifyQuery &Q) { |
| 4366 | if (!A->getType()->isPointerTy() || !B->getType()->isPointerTy()) |
| 4367 | return false; |
| 4368 | |
| 4369 | auto *GEPA = dyn_cast<GEPOperator>(Val: A); |
| 4370 | if (!GEPA || GEPA->getNumIndices() != 1 || !isa<Constant>(Val: GEPA->idx_begin())) |
| 4371 | return false; |
| 4372 | |
| 4373 | // Handle 2 incoming PHI values with one being a recursive GEP. |
| 4374 | auto *PN = dyn_cast<PHINode>(Val: GEPA->getPointerOperand()); |
| 4375 | if (!PN || PN->getNumIncomingValues() != 2) |
| 4376 | return false; |
| 4377 | |
| 4378 | // Search for the recursive GEP as an incoming operand, and record that as |
| 4379 | // Step. |
| 4380 | Value *Start = nullptr; |
| 4381 | Value *Step = const_cast<Value *>(A); |
| 4382 | if (PN->getIncomingValue(i: 0) == Step) |
| 4383 | Start = PN->getIncomingValue(i: 1); |
| 4384 | else if (PN->getIncomingValue(i: 1) == Step) |
| 4385 | Start = PN->getIncomingValue(i: 0); |
| 4386 | else |
| 4387 | return false; |
| 4388 | |
| 4389 | // Other incoming node base should match the B base. |
| 4390 | // StartOffset >= OffsetB && StepOffset > 0? |
| 4391 | // StartOffset <= OffsetB && StepOffset < 0? |
| 4392 | // Is non-equal if above are true. |
| 4393 | // We use stripAndAccumulateInBoundsConstantOffsets to restrict the |
| 4394 | // optimisation to inbounds GEPs only. |
| 4395 | unsigned IndexWidth = Q.DL.getIndexTypeSizeInBits(Ty: Start->getType()); |
| 4396 | APInt StartOffset(IndexWidth, 0); |
| 4397 | Start = Start->stripAndAccumulateInBoundsConstantOffsets(DL: Q.DL, Offset&: StartOffset); |
| 4398 | APInt StepOffset(IndexWidth, 0); |
| 4399 | Step = Step->stripAndAccumulateInBoundsConstantOffsets(DL: Q.DL, Offset&: StepOffset); |
| 4400 | |
| 4401 | // Check if Base Pointer of Step matches the PHI. |
| 4402 | if (Step != PN) |
| 4403 | return false; |
| 4404 | APInt OffsetB(IndexWidth, 0); |
| 4405 | B = B->stripAndAccumulateInBoundsConstantOffsets(DL: Q.DL, Offset&: OffsetB); |
| 4406 | return Start == B && |
| 4407 | ((StartOffset.sge(RHS: OffsetB) && StepOffset.isStrictlyPositive()) || |
| 4408 | (StartOffset.sle(RHS: OffsetB) && StepOffset.isNegative())); |
| 4409 | } |
| 4410 | |
| 4411 | static bool isKnownNonEqualFromContext(const Value *V1, const Value *V2, |
| 4412 | const SimplifyQuery &Q, unsigned Depth) { |
| 4413 | if (!Q.CtxI) |
| 4414 | return false; |
| 4415 | |
| 4416 | // Try to infer NonEqual based on information from dominating conditions. |
| 4417 | if (Q.DC && Q.DT) { |
| 4418 | auto IsKnownNonEqualFromDominatingCondition = [&](const Value *V) { |
| 4419 | for (CondBrInst *BI : Q.DC->conditionsFor(V)) { |
| 4420 | Value *Cond = BI->getCondition(); |
| 4421 | BasicBlockEdge Edge0(BI->getParent(), BI->getSuccessor(i: 0)); |
| 4422 | if (Q.DT->dominates(BBE: Edge0, BB: Q.CtxI->getParent()) && |
| 4423 | isImpliedCondition(LHS: Cond, RHSPred: ICmpInst::ICMP_NE, RHSOp0: V1, RHSOp1: V2, DL: Q.DL, |
| 4424 | /*LHSIsTrue=*/true, Depth) |
| 4425 | .value_or(u: false)) |
| 4426 | return true; |
| 4427 | |
| 4428 | BasicBlockEdge Edge1(BI->getParent(), BI->getSuccessor(i: 1)); |
| 4429 | if (Q.DT->dominates(BBE: Edge1, BB: Q.CtxI->getParent()) && |
| 4430 | isImpliedCondition(LHS: Cond, RHSPred: ICmpInst::ICMP_NE, RHSOp0: V1, RHSOp1: V2, DL: Q.DL, |
| 4431 | /*LHSIsTrue=*/false, Depth) |
| 4432 | .value_or(u: false)) |
| 4433 | return true; |
| 4434 | } |
| 4435 | |
| 4436 | return false; |
| 4437 | }; |
| 4438 | |
| 4439 | if (IsKnownNonEqualFromDominatingCondition(V1) || |
| 4440 | IsKnownNonEqualFromDominatingCondition(V2)) |
| 4441 | return true; |
| 4442 | } |
| 4443 | |
| 4444 | if (!Q.AC) |
| 4445 | return false; |
| 4446 | |
| 4447 | // Try to infer NonEqual based on information from assumptions. |
| 4448 | for (auto &AssumeVH : Q.AC->assumptionsFor(V: V1)) { |
| 4449 | if (!AssumeVH) |
| 4450 | continue; |
| 4451 | CallInst *I = cast<CallInst>(Val&: AssumeVH); |
| 4452 | |
| 4453 | assert(I->getFunction() == Q.CtxI->getFunction() && |
| 4454 | "Got assumption for the wrong function!" ); |
| 4455 | assert(I->getIntrinsicID() == Intrinsic::assume && |
| 4456 | "must be an assume intrinsic" ); |
| 4457 | |
| 4458 | if (isImpliedCondition(LHS: I->getArgOperand(i: 0), RHSPred: ICmpInst::ICMP_NE, RHSOp0: V1, RHSOp1: V2, DL: Q.DL, |
| 4459 | /*LHSIsTrue=*/true, Depth) |
| 4460 | .value_or(u: false) && |
| 4461 | isValidAssumeForContext(I, Q)) |
| 4462 | return true; |
| 4463 | } |
| 4464 | |
| 4465 | return false; |
| 4466 | } |
| 4467 | |
| 4468 | static bool isNonEqualURem(const Value *X, const Value *Rem, |
| 4469 | const SimplifyQuery &Q) { |
| 4470 | const Value *Y; |
| 4471 | if (!match(V: Rem, P: m_URem(L: m_Specific(V: X), R: m_Value(V&: Y)))) |
| 4472 | return false; |
| 4473 | |
| 4474 | // For a defined urem, X != X urem Y exactly when X u>= Y. |
| 4475 | // isTruePredicate does not handle UGE, so use the equivalent Y u<= X. |
| 4476 | if (isTruePredicate(Pred: ICmpInst::ICMP_ULE, LHS: Y, RHS: X)) |
| 4477 | return true; |
| 4478 | |
| 4479 | std::optional<bool> Implied = |
| 4480 | isImpliedByDomCondition(Pred: ICmpInst::ICMP_UGE, LHS: X, RHS: Y, ContextI: Q.CtxI, DL: Q.DL); |
| 4481 | return Implied && *Implied; |
| 4482 | } |
| 4483 | |
| 4484 | /// Return true if it is known that V1 != V2. |
| 4485 | static bool isKnownNonEqual(const Value *V1, const Value *V2, |
| 4486 | const APInt &DemandedElts, const SimplifyQuery &Q, |
| 4487 | unsigned Depth) { |
| 4488 | if (V1 == V2) |
| 4489 | return false; |
| 4490 | if (V1->getType() != V2->getType()) |
| 4491 | // We can't look through casts yet. |
| 4492 | return false; |
| 4493 | |
| 4494 | if (Depth >= MaxAnalysisRecursionDepth) |
| 4495 | return false; |
| 4496 | |
| 4497 | // See if we can recurse through (exactly one of) our operands. This |
| 4498 | // requires our operation be 1-to-1 and map every input value to exactly |
| 4499 | // one output value. Such an operation is invertible. |
| 4500 | auto *O1 = dyn_cast<Operator>(Val: V1); |
| 4501 | auto *O2 = dyn_cast<Operator>(Val: V2); |
| 4502 | if (O1 && O2 && O1->getOpcode() == O2->getOpcode()) { |
| 4503 | if (auto Values = getInvertibleOperands(Op1: O1, Op2: O2)) |
| 4504 | return isKnownNonEqual(V1: Values->first, V2: Values->second, DemandedElts, Q, |
| 4505 | Depth: Depth + 1); |
| 4506 | |
| 4507 | if (const PHINode *PN1 = dyn_cast<PHINode>(Val: V1)) { |
| 4508 | const PHINode *PN2 = cast<PHINode>(Val: V2); |
| 4509 | // FIXME: This is missing a generalization to handle the case where one is |
| 4510 | // a PHI and another one isn't. |
| 4511 | if (isNonEqualPHIs(PN1, PN2, DemandedElts, Q, Depth)) |
| 4512 | return true; |
| 4513 | }; |
| 4514 | } |
| 4515 | |
| 4516 | if (isModifyingBinopOfNonZero(V1, V2, DemandedElts, Q, Depth) || |
| 4517 | isModifyingBinopOfNonZero(V1: V2, V2: V1, DemandedElts, Q, Depth)) |
| 4518 | return true; |
| 4519 | |
| 4520 | if (isNonEqualMul(V1, V2, DemandedElts, Q, Depth) || |
| 4521 | isNonEqualMul(V1: V2, V2: V1, DemandedElts, Q, Depth)) |
| 4522 | return true; |
| 4523 | |
| 4524 | if (isNonEqualShl(V1, V2, DemandedElts, Q, Depth) || |
| 4525 | isNonEqualShl(V1: V2, V2: V1, DemandedElts, Q, Depth)) |
| 4526 | return true; |
| 4527 | |
| 4528 | if (V1->getType()->isIntOrIntVectorTy()) { |
| 4529 | // Are any known bits in V1 contradictory to known bits in V2? If V1 |
| 4530 | // has a known zero where V2 has a known one, they must not be equal. |
| 4531 | KnownBits Known1 = computeKnownBits(V: V1, DemandedElts, Q, Depth); |
| 4532 | if (!Known1.isUnknown()) { |
| 4533 | KnownBits Known2 = computeKnownBits(V: V2, DemandedElts, Q, Depth); |
| 4534 | if (Known1.Zero.intersects(RHS: Known2.One) || |
| 4535 | Known2.Zero.intersects(RHS: Known1.One)) |
| 4536 | return true; |
| 4537 | } |
| 4538 | } |
| 4539 | |
| 4540 | if (isNonEqualSelect(V1, V2, DemandedElts, Q, Depth) || |
| 4541 | isNonEqualSelect(V1: V2, V2: V1, DemandedElts, Q, Depth)) |
| 4542 | return true; |
| 4543 | |
| 4544 | if (isNonEqualPointersWithRecursiveGEP(A: V1, B: V2, Q) || |
| 4545 | isNonEqualPointersWithRecursiveGEP(A: V2, B: V1, Q)) |
| 4546 | return true; |
| 4547 | |
| 4548 | Value *A, *B; |
| 4549 | // PtrToInts are NonEqual if their Ptrs are NonEqual. |
| 4550 | // Check PtrToInt type matches the pointer size. |
| 4551 | if (match(V: V1, P: m_PtrToIntSameSize(DL: Q.DL, Op: m_Value(V&: A))) && |
| 4552 | match(V: V2, P: m_PtrToIntSameSize(DL: Q.DL, Op: m_Value(V&: B)))) |
| 4553 | return isKnownNonEqual(V1: A, V2: B, DemandedElts, Q, Depth: Depth + 1); |
| 4554 | |
| 4555 | if (isNonEqualURem(X: V1, Rem: V2, Q) || isNonEqualURem(X: V2, Rem: V1, Q)) |
| 4556 | return true; |
| 4557 | |
| 4558 | if (isKnownNonEqualFromContext(V1, V2, Q, Depth)) |
| 4559 | return true; |
| 4560 | |
| 4561 | return false; |
| 4562 | } |
| 4563 | |
| 4564 | /// For vector constants, loop over the elements and find the constant with the |
| 4565 | /// minimum number of sign bits. Return 0 if the value is not a vector constant |
| 4566 | /// or if any element was not analyzed; otherwise, return the count for the |
| 4567 | /// element with the minimum number of sign bits. |
| 4568 | static unsigned computeNumSignBitsVectorConstant(const Value *V, |
| 4569 | const APInt &DemandedElts, |
| 4570 | unsigned TyBits) { |
| 4571 | const auto *CV = dyn_cast<Constant>(Val: V); |
| 4572 | if (!CV || !isa<FixedVectorType>(Val: CV->getType())) |
| 4573 | return 0; |
| 4574 | |
| 4575 | unsigned MinSignBits = TyBits; |
| 4576 | unsigned NumElts = cast<FixedVectorType>(Val: CV->getType())->getNumElements(); |
| 4577 | for (unsigned i = 0; i != NumElts; ++i) { |
| 4578 | if (!DemandedElts[i]) |
| 4579 | continue; |
| 4580 | // If we find a non-ConstantInt, bail out. |
| 4581 | auto *Elt = dyn_cast_or_null<ConstantInt>(Val: CV->getAggregateElement(Elt: i)); |
| 4582 | if (!Elt) |
| 4583 | return 0; |
| 4584 | |
| 4585 | MinSignBits = std::min(a: MinSignBits, b: Elt->getValue().getNumSignBits()); |
| 4586 | } |
| 4587 | |
| 4588 | return MinSignBits; |
| 4589 | } |
| 4590 | |
| 4591 | static unsigned ComputeNumSignBitsImpl(const Value *V, |
| 4592 | const APInt &DemandedElts, |
| 4593 | const SimplifyQuery &Q, unsigned Depth); |
| 4594 | |
| 4595 | static unsigned ComputeNumSignBits(const Value *V, const APInt &DemandedElts, |
| 4596 | const SimplifyQuery &Q, unsigned Depth) { |
| 4597 | unsigned Result = ComputeNumSignBitsImpl(V, DemandedElts, Q, Depth); |
| 4598 | assert(Result > 0 && "At least one sign bit needs to be present!" ); |
| 4599 | return Result; |
| 4600 | } |
| 4601 | |
| 4602 | /// Return the number of times the sign bit of the register is replicated into |
| 4603 | /// the other bits. We know that at least 1 bit is always equal to the sign bit |
| 4604 | /// (itself), but other cases can give us information. For example, immediately |
| 4605 | /// after an "ashr X, 2", we know that the top 3 bits are all equal to each |
| 4606 | /// other, so we return 3. For vectors, return the number of sign bits for the |
| 4607 | /// vector element with the minimum number of known sign bits of the demanded |
| 4608 | /// elements in the vector specified by DemandedElts. |
| 4609 | static unsigned ComputeNumSignBitsImpl(const Value *V, |
| 4610 | const APInt &DemandedElts, |
| 4611 | const SimplifyQuery &Q, unsigned Depth) { |
| 4612 | Type *Ty = V->getType(); |
| 4613 | #ifndef NDEBUG |
| 4614 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth" ); |
| 4615 | |
| 4616 | if (auto *FVTy = dyn_cast<FixedVectorType>(Ty)) { |
| 4617 | assert( |
| 4618 | FVTy->getNumElements() == DemandedElts.getBitWidth() && |
| 4619 | "DemandedElt width should equal the fixed vector number of elements" ); |
| 4620 | } else { |
| 4621 | assert(DemandedElts == APInt(1, 1) && |
| 4622 | "DemandedElt width should be 1 for scalars" ); |
| 4623 | } |
| 4624 | #endif |
| 4625 | |
| 4626 | // We return the minimum number of sign bits that are guaranteed to be present |
| 4627 | // in V, so for undef we have to conservatively return 1. We don't have the |
| 4628 | // same behavior for poison though -- that's a FIXME today. |
| 4629 | |
| 4630 | Type *ScalarTy = Ty->getScalarType(); |
| 4631 | unsigned TyBits = ScalarTy->isPointerTy() ? |
| 4632 | Q.DL.getPointerTypeSizeInBits(ScalarTy) : |
| 4633 | Q.DL.getTypeSizeInBits(Ty: ScalarTy); |
| 4634 | |
| 4635 | unsigned Tmp, Tmp2; |
| 4636 | unsigned FirstAnswer = 1; |
| 4637 | |
| 4638 | // Note that ConstantInt is handled by the general computeKnownBits case |
| 4639 | // below. |
| 4640 | |
| 4641 | if (Depth == MaxAnalysisRecursionDepth) |
| 4642 | return 1; |
| 4643 | |
| 4644 | if (auto *U = dyn_cast<Operator>(Val: V)) { |
| 4645 | switch (Operator::getOpcode(V)) { |
| 4646 | default: break; |
| 4647 | case Instruction::BitCast: { |
| 4648 | Value *Src = U->getOperand(i: 0); |
| 4649 | Type *SrcTy = Src->getType(); |
| 4650 | |
| 4651 | // Skip if the source type is not an integer or integer vector type |
| 4652 | // This ensures we only process integer-like types |
| 4653 | if (!SrcTy->isIntOrIntVectorTy()) |
| 4654 | break; |
| 4655 | |
| 4656 | unsigned SrcBits = SrcTy->getScalarSizeInBits(); |
| 4657 | |
| 4658 | // Bitcast 'large element' scalar/vector to 'small element' vector. |
| 4659 | if ((SrcBits % TyBits) != 0) |
| 4660 | break; |
| 4661 | |
| 4662 | // Only proceed if the destination type is a fixed-size vector |
| 4663 | if (isa<FixedVectorType>(Val: Ty)) { |
| 4664 | // Fast case - sign splat can be simply split across the small elements. |
| 4665 | // This works for both vector and scalar sources |
| 4666 | Tmp = ComputeNumSignBits(V: Src, Q, Depth: Depth + 1); |
| 4667 | if (Tmp == SrcBits) |
| 4668 | return TyBits; |
| 4669 | } |
| 4670 | break; |
| 4671 | } |
| 4672 | case Instruction::SExt: |
| 4673 | Tmp = TyBits - U->getOperand(i: 0)->getType()->getScalarSizeInBits(); |
| 4674 | return ComputeNumSignBits(V: U->getOperand(i: 0), DemandedElts, Q, Depth: Depth + 1) + |
| 4675 | Tmp; |
| 4676 | |
| 4677 | case Instruction::SDiv: { |
| 4678 | const APInt *Denominator; |
| 4679 | // sdiv X, C -> adds log(C) sign bits. |
| 4680 | if (match(V: U->getOperand(i: 1), P: m_APInt(Res&: Denominator))) { |
| 4681 | |
| 4682 | // Ignore non-positive denominator. |
| 4683 | if (!Denominator->isStrictlyPositive()) |
| 4684 | break; |
| 4685 | |
| 4686 | // Calculate the incoming numerator bits. |
| 4687 | unsigned NumBits = |
| 4688 | ComputeNumSignBits(V: U->getOperand(i: 0), DemandedElts, Q, Depth: Depth + 1); |
| 4689 | |
| 4690 | // Add floor(log(C)) bits to the numerator bits. |
| 4691 | return std::min(a: TyBits, b: NumBits + Denominator->logBase2()); |
| 4692 | } |
| 4693 | break; |
| 4694 | } |
| 4695 | |
| 4696 | case Instruction::SRem: { |
| 4697 | Tmp = ComputeNumSignBits(V: U->getOperand(i: 0), DemandedElts, Q, Depth: Depth + 1); |
| 4698 | |
| 4699 | const APInt *Denominator; |
| 4700 | // srem X, C -> we know that the result is within [-C+1,C) when C is a |
| 4701 | // positive constant. This let us put a lower bound on the number of sign |
| 4702 | // bits. |
| 4703 | if (match(V: U->getOperand(i: 1), P: m_APInt(Res&: Denominator))) { |
| 4704 | |
| 4705 | // Ignore non-positive denominator. |
| 4706 | if (Denominator->isStrictlyPositive()) { |
| 4707 | // Calculate the leading sign bit constraints by examining the |
| 4708 | // denominator. Given that the denominator is positive, there are two |
| 4709 | // cases: |
| 4710 | // |
| 4711 | // 1. The numerator is positive. The result range is [0,C) and |
| 4712 | // [0,C) u< (1 << ceilLogBase2(C)). |
| 4713 | // |
| 4714 | // 2. The numerator is negative. Then the result range is (-C,0] and |
| 4715 | // integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)). |
| 4716 | // |
| 4717 | // Thus a lower bound on the number of sign bits is `TyBits - |
| 4718 | // ceilLogBase2(C)`. |
| 4719 | |
| 4720 | unsigned ResBits = TyBits - Denominator->ceilLogBase2(); |
| 4721 | Tmp = std::max(a: Tmp, b: ResBits); |
| 4722 | } |
| 4723 | } |
| 4724 | return Tmp; |
| 4725 | } |
| 4726 | |
| 4727 | case Instruction::AShr: { |
| 4728 | Tmp = ComputeNumSignBits(V: U->getOperand(i: 0), DemandedElts, Q, Depth: Depth + 1); |
| 4729 | // ashr X, C -> adds C sign bits. Vectors too. |
| 4730 | const APInt *ShAmt; |
| 4731 | if (match(V: U->getOperand(i: 1), P: m_APInt(Res&: ShAmt))) { |
| 4732 | if (ShAmt->uge(RHS: TyBits)) |
| 4733 | break; // Bad shift. |
| 4734 | unsigned ShAmtLimited = ShAmt->getZExtValue(); |
| 4735 | Tmp += ShAmtLimited; |
| 4736 | if (Tmp > TyBits) Tmp = TyBits; |
| 4737 | } |
| 4738 | return Tmp; |
| 4739 | } |
| 4740 | case Instruction::Shl: { |
| 4741 | const APInt *ShAmt; |
| 4742 | Value *X = nullptr; |
| 4743 | if (match(V: U->getOperand(i: 1), P: m_APInt(Res&: ShAmt))) { |
| 4744 | // shl destroys sign bits. |
| 4745 | if (ShAmt->uge(RHS: TyBits)) |
| 4746 | break; // Bad shift. |
| 4747 | // We can look through a zext (more or less treating it as a sext) if |
| 4748 | // all extended bits are shifted out. |
| 4749 | if (match(V: U->getOperand(i: 0), P: m_ZExt(Op: m_Value(V&: X))) && |
| 4750 | ShAmt->uge(RHS: TyBits - X->getType()->getScalarSizeInBits())) { |
| 4751 | Tmp = ComputeNumSignBits(V: X, DemandedElts, Q, Depth: Depth + 1); |
| 4752 | Tmp += TyBits - X->getType()->getScalarSizeInBits(); |
| 4753 | } else |
| 4754 | Tmp = |
| 4755 | ComputeNumSignBits(V: U->getOperand(i: 0), DemandedElts, Q, Depth: Depth + 1); |
| 4756 | if (ShAmt->uge(RHS: Tmp)) |
| 4757 | break; // Shifted all sign bits out. |
| 4758 | Tmp2 = ShAmt->getZExtValue(); |
| 4759 | return Tmp - Tmp2; |
| 4760 | } |
| 4761 | break; |
| 4762 | } |
| 4763 | case Instruction::And: |
| 4764 | case Instruction::Or: |
| 4765 | case Instruction::Xor: // NOT is handled here. |
| 4766 | // Logical binary ops preserve the number of sign bits at the worst. |
| 4767 | Tmp = ComputeNumSignBits(V: U->getOperand(i: 0), DemandedElts, Q, Depth: Depth + 1); |
| 4768 | if (Tmp != 1) { |
| 4769 | Tmp2 = ComputeNumSignBits(V: U->getOperand(i: 1), DemandedElts, Q, Depth: Depth + 1); |
| 4770 | FirstAnswer = std::min(a: Tmp, b: Tmp2); |
| 4771 | // We computed what we know about the sign bits as our first |
| 4772 | // answer. Now proceed to the generic code that uses |
| 4773 | // computeKnownBits, and pick whichever answer is better. |
| 4774 | } |
| 4775 | break; |
| 4776 | |
| 4777 | case Instruction::Select: { |
| 4778 | // If we have a clamp pattern, we know that the number of sign bits will |
| 4779 | // be the minimum of the clamp min/max range. |
| 4780 | const Value *X; |
| 4781 | const APInt *CLow, *CHigh; |
| 4782 | if (isSignedMinMaxClamp(Select: U, In&: X, CLow, CHigh)) |
| 4783 | return std::min(a: CLow->getNumSignBits(), b: CHigh->getNumSignBits()); |
| 4784 | |
| 4785 | Tmp = ComputeNumSignBits(V: U->getOperand(i: 1), DemandedElts, Q, Depth: Depth + 1); |
| 4786 | if (Tmp == 1) |
| 4787 | break; |
| 4788 | Tmp2 = ComputeNumSignBits(V: U->getOperand(i: 2), DemandedElts, Q, Depth: Depth + 1); |
| 4789 | return std::min(a: Tmp, b: Tmp2); |
| 4790 | } |
| 4791 | |
| 4792 | case Instruction::Add: |
| 4793 | // Add can have at most one carry bit. Thus we know that the output |
| 4794 | // is, at worst, one more bit than the inputs. |
| 4795 | Tmp = ComputeNumSignBits(V: U->getOperand(i: 0), Q, Depth: Depth + 1); |
| 4796 | if (Tmp == 1) break; |
| 4797 | |
| 4798 | // Special case decrementing a value (ADD X, -1): |
| 4799 | if (const auto *CRHS = dyn_cast<Constant>(Val: U->getOperand(i: 1))) |
| 4800 | if (CRHS->isAllOnesValue()) { |
| 4801 | KnownBits Known(TyBits); |
| 4802 | computeKnownBits(V: U->getOperand(i: 0), DemandedElts, Known, Q, Depth: Depth + 1); |
| 4803 | |
| 4804 | // If the input is known to be 0 or 1, the output is 0/-1, which is |
| 4805 | // all sign bits set. |
| 4806 | if ((Known.Zero | 1).isAllOnes()) |
| 4807 | return TyBits; |
| 4808 | |
| 4809 | // If we are subtracting one from a positive number, there is no carry |
| 4810 | // out of the result. |
| 4811 | if (Known.isNonNegative()) |
| 4812 | return Tmp; |
| 4813 | } |
| 4814 | |
| 4815 | Tmp2 = ComputeNumSignBits(V: U->getOperand(i: 1), DemandedElts, Q, Depth: Depth + 1); |
| 4816 | if (Tmp2 == 1) |
| 4817 | break; |
| 4818 | return std::min(a: Tmp, b: Tmp2) - 1; |
| 4819 | |
| 4820 | case Instruction::Sub: |
| 4821 | Tmp2 = ComputeNumSignBits(V: U->getOperand(i: 1), DemandedElts, Q, Depth: Depth + 1); |
| 4822 | if (Tmp2 == 1) |
| 4823 | break; |
| 4824 | |
| 4825 | // Handle NEG. |
| 4826 | if (const auto *CLHS = dyn_cast<Constant>(Val: U->getOperand(i: 0))) |
| 4827 | if (CLHS->isNullValue()) { |
| 4828 | KnownBits Known(TyBits); |
| 4829 | computeKnownBits(V: U->getOperand(i: 1), DemandedElts, Known, Q, Depth: Depth + 1); |
| 4830 | // If the input is known to be 0 or 1, the output is 0/-1, which is |
| 4831 | // all sign bits set. |
| 4832 | if ((Known.Zero | 1).isAllOnes()) |
| 4833 | return TyBits; |
| 4834 | |
| 4835 | // If the input is known to be positive (the sign bit is known clear), |
| 4836 | // the output of the NEG has the same number of sign bits as the |
| 4837 | // input. |
| 4838 | if (Known.isNonNegative()) |
| 4839 | return Tmp2; |
| 4840 | |
| 4841 | // Otherwise, we treat this like a SUB. |
| 4842 | } |
| 4843 | |
| 4844 | // Sub can have at most one carry bit. Thus we know that the output |
| 4845 | // is, at worst, one more bit than the inputs. |
| 4846 | Tmp = ComputeNumSignBits(V: U->getOperand(i: 0), DemandedElts, Q, Depth: Depth + 1); |
| 4847 | if (Tmp == 1) |
| 4848 | break; |
| 4849 | return std::min(a: Tmp, b: Tmp2) - 1; |
| 4850 | |
| 4851 | case Instruction::Mul: { |
| 4852 | // The output of the Mul can be at most twice the valid bits in the |
| 4853 | // inputs. |
| 4854 | unsigned SignBitsOp0 = |
| 4855 | ComputeNumSignBits(V: U->getOperand(i: 0), DemandedElts, Q, Depth: Depth + 1); |
| 4856 | if (SignBitsOp0 == 1) |
| 4857 | break; |
| 4858 | unsigned SignBitsOp1 = |
| 4859 | ComputeNumSignBits(V: U->getOperand(i: 1), DemandedElts, Q, Depth: Depth + 1); |
| 4860 | if (SignBitsOp1 == 1) |
| 4861 | break; |
| 4862 | unsigned OutValidBits = |
| 4863 | (TyBits - SignBitsOp0 + 1) + (TyBits - SignBitsOp1 + 1); |
| 4864 | return OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1; |
| 4865 | } |
| 4866 | |
| 4867 | case Instruction::PHI: { |
| 4868 | const PHINode *PN = cast<PHINode>(Val: U); |
| 4869 | unsigned NumIncomingValues = PN->getNumIncomingValues(); |
| 4870 | // Don't analyze large in-degree PHIs. |
| 4871 | if (NumIncomingValues > 4) break; |
| 4872 | // Unreachable blocks may have zero-operand PHI nodes. |
| 4873 | if (NumIncomingValues == 0) break; |
| 4874 | |
| 4875 | // Take the minimum of all incoming values. This can't infinitely loop |
| 4876 | // because of our depth threshold. |
| 4877 | SimplifyQuery RecQ = Q.getWithoutCondContext(); |
| 4878 | Tmp = TyBits; |
| 4879 | for (unsigned i = 0, e = NumIncomingValues; i != e; ++i) { |
| 4880 | if (Tmp == 1) return Tmp; |
| 4881 | RecQ.CtxI = PN->getIncomingBlock(i)->getTerminator(); |
| 4882 | Tmp = std::min(a: Tmp, b: ComputeNumSignBits(V: PN->getIncomingValue(i), |
| 4883 | DemandedElts, Q: RecQ, Depth: Depth + 1)); |
| 4884 | } |
| 4885 | return Tmp; |
| 4886 | } |
| 4887 | |
| 4888 | case Instruction::Trunc: { |
| 4889 | // If the input contained enough sign bits that some remain after the |
| 4890 | // truncation, then we can make use of that. Otherwise we don't know |
| 4891 | // anything. |
| 4892 | Tmp = ComputeNumSignBits(V: U->getOperand(i: 0), Q, Depth: Depth + 1); |
| 4893 | unsigned OperandTyBits = U->getOperand(i: 0)->getType()->getScalarSizeInBits(); |
| 4894 | if (Tmp > (OperandTyBits - TyBits)) |
| 4895 | return Tmp - (OperandTyBits - TyBits); |
| 4896 | |
| 4897 | return 1; |
| 4898 | } |
| 4899 | |
| 4900 | case Instruction::ExtractElement: |
| 4901 | // Look through extract element. At the moment we keep this simple and |
| 4902 | // skip tracking the specific element. But at least we might find |
| 4903 | // information valid for all elements of the vector (for example if vector |
| 4904 | // is sign extended, shifted, etc). |
| 4905 | return ComputeNumSignBits(V: U->getOperand(i: 0), Q, Depth: Depth + 1); |
| 4906 | |
| 4907 | case Instruction::ShuffleVector: { |
| 4908 | // Collect the minimum number of sign bits that are shared by every vector |
| 4909 | // element referenced by the shuffle. |
| 4910 | auto *Shuf = dyn_cast<ShuffleVectorInst>(Val: U); |
| 4911 | if (!Shuf) { |
| 4912 | // FIXME: Add support for shufflevector constant expressions. |
| 4913 | return 1; |
| 4914 | } |
| 4915 | APInt DemandedLHS, DemandedRHS; |
| 4916 | // For undef elements, we don't know anything about the common state of |
| 4917 | // the shuffle result. |
| 4918 | if (!getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) |
| 4919 | return 1; |
| 4920 | Tmp = std::numeric_limits<unsigned>::max(); |
| 4921 | if (!!DemandedLHS) { |
| 4922 | const Value *LHS = Shuf->getOperand(i_nocapture: 0); |
| 4923 | Tmp = ComputeNumSignBits(V: LHS, DemandedElts: DemandedLHS, Q, Depth: Depth + 1); |
| 4924 | } |
| 4925 | // If we don't know anything, early out and try computeKnownBits |
| 4926 | // fall-back. |
| 4927 | if (Tmp == 1) |
| 4928 | break; |
| 4929 | if (!!DemandedRHS) { |
| 4930 | const Value *RHS = Shuf->getOperand(i_nocapture: 1); |
| 4931 | Tmp2 = ComputeNumSignBits(V: RHS, DemandedElts: DemandedRHS, Q, Depth: Depth + 1); |
| 4932 | Tmp = std::min(a: Tmp, b: Tmp2); |
| 4933 | } |
| 4934 | // If we don't know anything, early out and try computeKnownBits |
| 4935 | // fall-back. |
| 4936 | if (Tmp == 1) |
| 4937 | break; |
| 4938 | assert(Tmp <= TyBits && "Failed to determine minimum sign bits" ); |
| 4939 | return Tmp; |
| 4940 | } |
| 4941 | case Instruction::Call: { |
| 4942 | if (const auto *II = dyn_cast<IntrinsicInst>(Val: U)) { |
| 4943 | switch (II->getIntrinsicID()) { |
| 4944 | default: |
| 4945 | break; |
| 4946 | case Intrinsic::abs: |
| 4947 | Tmp = |
| 4948 | ComputeNumSignBits(V: U->getOperand(i: 0), DemandedElts, Q, Depth: Depth + 1); |
| 4949 | if (Tmp == 1) |
| 4950 | break; |
| 4951 | |
| 4952 | // Absolute value reduces number of sign bits by at most 1. |
| 4953 | return Tmp - 1; |
| 4954 | case Intrinsic::smin: |
| 4955 | case Intrinsic::smax: { |
| 4956 | const APInt *CLow, *CHigh; |
| 4957 | if (isSignedMinMaxIntrinsicClamp(II, CLow, CHigh)) |
| 4958 | return std::min(a: CLow->getNumSignBits(), b: CHigh->getNumSignBits()); |
| 4959 | } |
| 4960 | } |
| 4961 | } |
| 4962 | } |
| 4963 | } |
| 4964 | } |
| 4965 | |
| 4966 | // Finally, if we can prove that the top bits of the result are 0's or 1's, |
| 4967 | // use this information. |
| 4968 | |
| 4969 | // If we can examine all elements of a vector constant successfully, we're |
| 4970 | // done (we can't do any better than that). If not, keep trying. |
| 4971 | if (unsigned VecSignBits = |
| 4972 | computeNumSignBitsVectorConstant(V, DemandedElts, TyBits)) |
| 4973 | return VecSignBits; |
| 4974 | |
| 4975 | KnownBits Known(TyBits); |
| 4976 | computeKnownBits(V, DemandedElts, Known, Q, Depth); |
| 4977 | |
| 4978 | // If we know that the sign bit is either zero or one, determine the number of |
| 4979 | // identical bits in the top of the input value. |
| 4980 | return std::max(a: FirstAnswer, b: Known.countMinSignBits()); |
| 4981 | } |
| 4982 | |
| 4983 | Intrinsic::ID llvm::getIntrinsicForCallSite(const CallBase &CB, |
| 4984 | const TargetLibraryInfo *TLI) { |
| 4985 | const Function *F = CB.getCalledFunction(); |
| 4986 | if (!F) |
| 4987 | return Intrinsic::not_intrinsic; |
| 4988 | |
| 4989 | if (F->isIntrinsic()) |
| 4990 | return F->getIntrinsicID(); |
| 4991 | |
| 4992 | // We are going to infer semantics of a library function based on mapping it |
| 4993 | // to an LLVM intrinsic. Check that the library function is available from |
| 4994 | // this callbase and in this environment. |
| 4995 | if (F->hasLocalLinkage() || !TLI || !CB.onlyReadsMemory()) |
| 4996 | return Intrinsic::not_intrinsic; |
| 4997 | |
| 4998 | LibFunc Func = TLI->getLibFunc(CB); |
| 4999 | if (Func == NotLibFunc) |
| 5000 | return Intrinsic::not_intrinsic; |
| 5001 | |
| 5002 | switch (Func) { |
| 5003 | default: |
| 5004 | break; |
| 5005 | case LibFunc_sin: |
| 5006 | case LibFunc_sinf: |
| 5007 | case LibFunc_sinl: |
| 5008 | return Intrinsic::sin; |
| 5009 | case LibFunc_cos: |
| 5010 | case LibFunc_cosf: |
| 5011 | case LibFunc_cosl: |
| 5012 | return Intrinsic::cos; |
| 5013 | case LibFunc_tan: |
| 5014 | case LibFunc_tanf: |
| 5015 | case LibFunc_tanl: |
| 5016 | return Intrinsic::tan; |
| 5017 | case LibFunc_asin: |
| 5018 | case LibFunc_asinf: |
| 5019 | case LibFunc_asinl: |
| 5020 | return Intrinsic::asin; |
| 5021 | case LibFunc_acos: |
| 5022 | case LibFunc_acosf: |
| 5023 | case LibFunc_acosl: |
| 5024 | return Intrinsic::acos; |
| 5025 | case LibFunc_atan: |
| 5026 | case LibFunc_atanf: |
| 5027 | case LibFunc_atanl: |
| 5028 | return Intrinsic::atan; |
| 5029 | case LibFunc_atan2: |
| 5030 | case LibFunc_atan2f: |
| 5031 | case LibFunc_atan2l: |
| 5032 | return Intrinsic::atan2; |
| 5033 | case LibFunc_sinh: |
| 5034 | case LibFunc_sinhf: |
| 5035 | case LibFunc_sinhl: |
| 5036 | return Intrinsic::sinh; |
| 5037 | case LibFunc_cosh: |
| 5038 | case LibFunc_coshf: |
| 5039 | case LibFunc_coshl: |
| 5040 | return Intrinsic::cosh; |
| 5041 | case LibFunc_tanh: |
| 5042 | case LibFunc_tanhf: |
| 5043 | case LibFunc_tanhl: |
| 5044 | return Intrinsic::tanh; |
| 5045 | case LibFunc_exp: |
| 5046 | case LibFunc_expf: |
| 5047 | case LibFunc_expl: |
| 5048 | return Intrinsic::exp; |
| 5049 | case LibFunc_exp2: |
| 5050 | case LibFunc_exp2f: |
| 5051 | case LibFunc_exp2l: |
| 5052 | return Intrinsic::exp2; |
| 5053 | case LibFunc_exp10: |
| 5054 | case LibFunc_exp10f: |
| 5055 | case LibFunc_exp10l: |
| 5056 | return Intrinsic::exp10; |
| 5057 | case LibFunc_log: |
| 5058 | case LibFunc_logf: |
| 5059 | case LibFunc_logl: |
| 5060 | return Intrinsic::log; |
| 5061 | case LibFunc_log10: |
| 5062 | case LibFunc_log10f: |
| 5063 | case LibFunc_log10l: |
| 5064 | return Intrinsic::log10; |
| 5065 | case LibFunc_log2: |
| 5066 | case LibFunc_log2f: |
| 5067 | case LibFunc_log2l: |
| 5068 | return Intrinsic::log2; |
| 5069 | case LibFunc_fabs: |
| 5070 | case LibFunc_fabsf: |
| 5071 | case LibFunc_fabsl: |
| 5072 | return Intrinsic::fabs; |
| 5073 | case LibFunc_fmin: |
| 5074 | case LibFunc_fminf: |
| 5075 | case LibFunc_fminl: |
| 5076 | return Intrinsic::minnum; |
| 5077 | case LibFunc_fmax: |
| 5078 | case LibFunc_fmaxf: |
| 5079 | case LibFunc_fmaxl: |
| 5080 | return Intrinsic::maxnum; |
| 5081 | case LibFunc_copysign: |
| 5082 | case LibFunc_copysignf: |
| 5083 | case LibFunc_copysignl: |
| 5084 | return Intrinsic::copysign; |
| 5085 | case LibFunc_floor: |
| 5086 | case LibFunc_floorf: |
| 5087 | case LibFunc_floorl: |
| 5088 | return Intrinsic::floor; |
| 5089 | case LibFunc_ceil: |
| 5090 | case LibFunc_ceilf: |
| 5091 | case LibFunc_ceill: |
| 5092 | return Intrinsic::ceil; |
| 5093 | case LibFunc_trunc: |
| 5094 | case LibFunc_truncf: |
| 5095 | case LibFunc_truncl: |
| 5096 | return Intrinsic::trunc; |
| 5097 | case LibFunc_rint: |
| 5098 | case LibFunc_rintf: |
| 5099 | case LibFunc_rintl: |
| 5100 | return Intrinsic::rint; |
| 5101 | case LibFunc_nearbyint: |
| 5102 | case LibFunc_nearbyintf: |
| 5103 | case LibFunc_nearbyintl: |
| 5104 | return Intrinsic::nearbyint; |
| 5105 | case LibFunc_round: |
| 5106 | case LibFunc_roundf: |
| 5107 | case LibFunc_roundl: |
| 5108 | return Intrinsic::round; |
| 5109 | case LibFunc_roundeven: |
| 5110 | case LibFunc_roundevenf: |
| 5111 | case LibFunc_roundevenl: |
| 5112 | return Intrinsic::roundeven; |
| 5113 | case LibFunc_pow: |
| 5114 | case LibFunc_powf: |
| 5115 | case LibFunc_powl: |
| 5116 | return Intrinsic::pow; |
| 5117 | case LibFunc_sqrt: |
| 5118 | case LibFunc_sqrtf: |
| 5119 | case LibFunc_sqrtl: |
| 5120 | return Intrinsic::sqrt; |
| 5121 | } |
| 5122 | |
| 5123 | return Intrinsic::not_intrinsic; |
| 5124 | } |
| 5125 | |
| 5126 | /// Given an exploded icmp instruction, return true if the comparison only |
| 5127 | /// checks the sign bit. If it only checks the sign bit, set TrueIfSigned if |
| 5128 | /// the result of the comparison is true when the input value is signed. |
| 5129 | bool llvm::isSignBitCheck(ICmpInst::Predicate Pred, const APInt &RHS, |
| 5130 | bool &TrueIfSigned) { |
| 5131 | switch (Pred) { |
| 5132 | case ICmpInst::ICMP_SLT: // True if LHS s< 0 |
| 5133 | TrueIfSigned = true; |
| 5134 | return RHS.isZero(); |
| 5135 | case ICmpInst::ICMP_SLE: // True if LHS s<= -1 |
| 5136 | TrueIfSigned = true; |
| 5137 | return RHS.isAllOnes(); |
| 5138 | case ICmpInst::ICMP_SGT: // True if LHS s> -1 |
| 5139 | TrueIfSigned = false; |
| 5140 | return RHS.isAllOnes(); |
| 5141 | case ICmpInst::ICMP_SGE: // True if LHS s>= 0 |
| 5142 | TrueIfSigned = false; |
| 5143 | return RHS.isZero(); |
| 5144 | case ICmpInst::ICMP_UGT: |
| 5145 | // True if LHS u> RHS and RHS == sign-bit-mask - 1 |
| 5146 | TrueIfSigned = true; |
| 5147 | return RHS.isMaxSignedValue(); |
| 5148 | case ICmpInst::ICMP_UGE: |
| 5149 | // True if LHS u>= RHS and RHS == sign-bit-mask (2^7, 2^15, 2^31, etc) |
| 5150 | TrueIfSigned = true; |
| 5151 | return RHS.isMinSignedValue(); |
| 5152 | case ICmpInst::ICMP_ULT: |
| 5153 | // True if LHS u< RHS and RHS == sign-bit-mask (2^7, 2^15, 2^31, etc) |
| 5154 | TrueIfSigned = false; |
| 5155 | return RHS.isMinSignedValue(); |
| 5156 | case ICmpInst::ICMP_ULE: |
| 5157 | // True if LHS u<= RHS and RHS == sign-bit-mask - 1 |
| 5158 | TrueIfSigned = false; |
| 5159 | return RHS.isMaxSignedValue(); |
| 5160 | default: |
| 5161 | return false; |
| 5162 | } |
| 5163 | } |
| 5164 | |
| 5165 | static void computeKnownFPClassFromCond(const Value *V, Value *Cond, |
| 5166 | bool CondIsTrue, |
| 5167 | const Instruction *CtxI, |
| 5168 | KnownFPClass &KnownFromContext, |
| 5169 | unsigned Depth = 0) { |
| 5170 | Value *A, *B; |
| 5171 | if (Depth < MaxAnalysisRecursionDepth && |
| 5172 | (CondIsTrue ? match(V: Cond, P: m_LogicalAnd(L: m_Value(V&: A), R: m_Value(V&: B))) |
| 5173 | : match(V: Cond, P: m_LogicalOr(L: m_Value(V&: A), R: m_Value(V&: B))))) { |
| 5174 | computeKnownFPClassFromCond(V, Cond: A, CondIsTrue, CtxI, KnownFromContext, |
| 5175 | Depth: Depth + 1); |
| 5176 | computeKnownFPClassFromCond(V, Cond: B, CondIsTrue, CtxI, KnownFromContext, |
| 5177 | Depth: Depth + 1); |
| 5178 | return; |
| 5179 | } |
| 5180 | if (Depth < MaxAnalysisRecursionDepth && match(V: Cond, P: m_Not(V: m_Value(V&: A)))) { |
| 5181 | computeKnownFPClassFromCond(V, Cond: A, CondIsTrue: !CondIsTrue, CtxI, KnownFromContext, |
| 5182 | Depth: Depth + 1); |
| 5183 | return; |
| 5184 | } |
| 5185 | CmpPredicate Pred; |
| 5186 | Value *LHS; |
| 5187 | uint64_t ClassVal = 0; |
| 5188 | const APFloat *CRHS; |
| 5189 | const APInt *RHS; |
| 5190 | if (match(V: Cond, P: m_FCmp(Pred, L: m_Value(V&: LHS), R: m_APFloat(Res&: CRHS)))) { |
| 5191 | auto [CmpVal, MaskIfTrue, MaskIfFalse] = fcmpImpliesClass( |
| 5192 | Pred, F: *cast<Instruction>(Val: Cond)->getParent()->getParent(), LHS, ConstRHS: *CRHS, |
| 5193 | LookThroughSrc: LHS != V); |
| 5194 | if (CmpVal == V) |
| 5195 | KnownFromContext.knownNot(RuleOut: ~(CondIsTrue ? MaskIfTrue : MaskIfFalse)); |
| 5196 | } else if (match(V: Cond, P: m_Intrinsic<Intrinsic::is_fpclass>( |
| 5197 | Ops: m_Specific(V), Ops: m_ConstantInt(V&: ClassVal)))) { |
| 5198 | FPClassTest Mask = static_cast<FPClassTest>(ClassVal); |
| 5199 | KnownFromContext.knownNot(RuleOut: CondIsTrue ? ~Mask : Mask); |
| 5200 | } else if (match(V: Cond, P: m_ICmp(Pred, L: m_ElementWiseBitCast(Op: m_Specific(V)), |
| 5201 | R: m_APInt(Res&: RHS)))) { |
| 5202 | bool TrueIfSigned; |
| 5203 | if (!isSignBitCheck(Pred, RHS: *RHS, TrueIfSigned)) |
| 5204 | return; |
| 5205 | if (TrueIfSigned == CondIsTrue) |
| 5206 | KnownFromContext.signBitMustBeOne(); |
| 5207 | else |
| 5208 | KnownFromContext.signBitMustBeZero(); |
| 5209 | } |
| 5210 | } |
| 5211 | |
| 5212 | /// Compute the minimum and maximum values (inclusive) for the exponent of \p V, |
| 5213 | /// assuming it is not nan. Returns {min, max, max-assuming-nonzero}. A value |
| 5214 | /// frexp(0) = 0, so the tighter max-assuming-nonzero bound is only usable when |
| 5215 | /// \p V is known not to be a logical zero (e.g., for fabs(x) < 0.25, the non-0 |
| 5216 | /// exponent range is [-149, -2], but the 0 edge case is above this range). |
| 5217 | static std::tuple<int, int, int> |
| 5218 | computeKnownExponentRangeFromContext(const Value *V, const SimplifyQuery &Q) { |
| 5219 | if (!Q.CtxI || !Q.DC || !Q.DT) |
| 5220 | return {APFloat::IEK_NaN, APFloat::IEK_Inf, APFloat::IEK_Inf}; |
| 5221 | |
| 5222 | // Intersect the bounds implied by every dominating condition, keeping the |
| 5223 | // tightest maximum. A value may participate in multiple compares |
| 5224 | // (e.g. fabs(x) < 2.0 and fabs(x) < 1.0), and the tighter one wins. |
| 5225 | int MaxExp = APFloat::IEK_Inf; |
| 5226 | int MaxExpNonZero = APFloat::IEK_Inf; |
| 5227 | |
| 5228 | for (CondBrInst *BI : Q.DC->conditionsFor(V)) { |
| 5229 | CmpPredicate Pred; |
| 5230 | const APFloat *LimitC; |
| 5231 | if (!match(V: BI->getCondition(), |
| 5232 | P: m_FCmp(Pred, L: m_FAbs(Op0: m_Specific(V)), R: m_Finite(V&: LimitC)))) |
| 5233 | continue; |
| 5234 | |
| 5235 | if (Pred == FCmpInst::FCMP_ORD || Pred == FCmpInst::FCMP_UNO || |
| 5236 | Pred == FCmpInst::FCMP_TRUE || Pred == FCmpInst::FCMP_FALSE) |
| 5237 | continue; |
| 5238 | |
| 5239 | // If fabs(x) <= K, implies the exponent min exp range. |
| 5240 | // if fabs(x) >= K, swap the successor |
| 5241 | bool IsLessEqual = |
| 5242 | Pred == FCmpInst::FCMP_OLT || Pred == FCmpInst::FCMP_OLE || |
| 5243 | Pred == FCmpInst::FCMP_ULT || Pred == FCmpInst::FCMP_ULE || |
| 5244 | Pred == FCmpInst::FCMP_OEQ || Pred == FCmpInst::FCMP_UEQ; |
| 5245 | |
| 5246 | bool KnownStrictlyLess = |
| 5247 | Pred == FCmpInst::FCMP_OLT || Pred == FCmpInst::FCMP_ULT || |
| 5248 | Pred == FCmpInst::FCMP_OGE || Pred == FCmpInst::FCMP_UGE; |
| 5249 | |
| 5250 | BasicBlockEdge Edge1(BI->getParent(), |
| 5251 | BI->getSuccessor(i: IsLessEqual ? 0 : 1)); |
| 5252 | if (Q.DT->dominates(BBE: Edge1, BB: Q.CtxI->getParent())) { |
| 5253 | // frexp returns an exponent one greater than ilogb. |
| 5254 | int Exp = ilogb(Arg: *LimitC) + 1; |
| 5255 | |
| 5256 | // A strict bound fabs(V) < 2^n forces ilogb(V) <= n - 1, so the max frexp |
| 5257 | // exponent drops by one when K is exact power of two. |
| 5258 | if (KnownStrictlyLess && LimitC->getExactLog2Abs() != INT_MIN) |
| 5259 | --Exp; |
| 5260 | |
| 5261 | // frexp(0) = 0, which the bound above (assuming a normal nonzero value) |
| 5262 | // may exclude. |
| 5263 | |
| 5264 | // TODO: Figure out lower bound to detect no-underflow. |
| 5265 | MaxExpNonZero = std::min(a: MaxExpNonZero, b: Exp); |
| 5266 | MaxExp = std::min(a: MaxExp, b: std::max(a: Exp, b: 0)); |
| 5267 | } |
| 5268 | } |
| 5269 | |
| 5270 | return {APFloat::IEK_NaN, MaxExp, MaxExpNonZero}; |
| 5271 | } |
| 5272 | |
| 5273 | static KnownFPClass computeKnownFPClassFromContext(const Value *V, |
| 5274 | const SimplifyQuery &Q) { |
| 5275 | KnownFPClass KnownFromContext; |
| 5276 | |
| 5277 | if (Q.CC && Q.CC->AffectedValues.contains(Ptr: V)) |
| 5278 | computeKnownFPClassFromCond(V, Cond: Q.CC->Cond, CondIsTrue: !Q.CC->Invert, CtxI: Q.CtxI, |
| 5279 | KnownFromContext); |
| 5280 | |
| 5281 | if (!Q.CtxI) |
| 5282 | return KnownFromContext; |
| 5283 | |
| 5284 | if (Q.DC && Q.DT) { |
| 5285 | // Handle dominating conditions. |
| 5286 | for (CondBrInst *BI : Q.DC->conditionsFor(V)) { |
| 5287 | Value *Cond = BI->getCondition(); |
| 5288 | |
| 5289 | BasicBlockEdge Edge0(BI->getParent(), BI->getSuccessor(i: 0)); |
| 5290 | if (Q.DT->dominates(BBE: Edge0, BB: Q.CtxI->getParent())) |
| 5291 | computeKnownFPClassFromCond(V, Cond, /*CondIsTrue=*/true, CtxI: Q.CtxI, |
| 5292 | KnownFromContext); |
| 5293 | |
| 5294 | BasicBlockEdge Edge1(BI->getParent(), BI->getSuccessor(i: 1)); |
| 5295 | if (Q.DT->dominates(BBE: Edge1, BB: Q.CtxI->getParent())) |
| 5296 | computeKnownFPClassFromCond(V, Cond, /*CondIsTrue=*/false, CtxI: Q.CtxI, |
| 5297 | KnownFromContext); |
| 5298 | } |
| 5299 | } |
| 5300 | |
| 5301 | if (!Q.AC) |
| 5302 | return KnownFromContext; |
| 5303 | |
| 5304 | // Try to restrict the floating-point classes based on information from |
| 5305 | // assumptions. |
| 5306 | for (auto &AssumeVH : Q.AC->assumptionsFor(V)) { |
| 5307 | if (!AssumeVH) |
| 5308 | continue; |
| 5309 | CallInst *I = cast<CallInst>(Val&: AssumeVH); |
| 5310 | |
| 5311 | assert(I->getFunction() == Q.CtxI->getParent()->getParent() && |
| 5312 | "Got assumption for the wrong function!" ); |
| 5313 | assert(I->getIntrinsicID() == Intrinsic::assume && |
| 5314 | "must be an assume intrinsic" ); |
| 5315 | |
| 5316 | if (!isValidAssumeForContext(I, Q)) |
| 5317 | continue; |
| 5318 | |
| 5319 | computeKnownFPClassFromCond(V, Cond: I->getArgOperand(i: 0), |
| 5320 | /*CondIsTrue=*/true, CtxI: Q.CtxI, KnownFromContext); |
| 5321 | } |
| 5322 | |
| 5323 | return KnownFromContext; |
| 5324 | } |
| 5325 | |
| 5326 | void llvm::adjustKnownFPClassForSelectArm(KnownFPClass &Known, Value *Cond, |
| 5327 | Value *Arm, bool Invert, |
| 5328 | const SimplifyQuery &SQ, |
| 5329 | unsigned Depth) { |
| 5330 | |
| 5331 | KnownFPClass KnownSrc; |
| 5332 | computeKnownFPClassFromCond(V: Arm, Cond, |
| 5333 | /*CondIsTrue=*/!Invert, CtxI: SQ.CtxI, KnownFromContext&: KnownSrc, |
| 5334 | Depth: Depth + 1); |
| 5335 | KnownSrc = KnownSrc.unionWith(RHS: Known); |
| 5336 | if (KnownSrc.isUnknown()) |
| 5337 | return; |
| 5338 | |
| 5339 | if (isGuaranteedNotToBeUndef(V: Arm, AC: SQ.AC, CtxI: SQ.CtxI, DT: SQ.DT, Depth: Depth + 1)) |
| 5340 | Known = KnownSrc; |
| 5341 | } |
| 5342 | |
| 5343 | void computeKnownFPClass(const Value *V, const APInt &DemandedElts, |
| 5344 | FPClassTest InterestedClasses, KnownFPClass &Known, |
| 5345 | const SimplifyQuery &Q, unsigned Depth); |
| 5346 | |
| 5347 | static void computeKnownFPClass(const Value *V, KnownFPClass &Known, |
| 5348 | FPClassTest InterestedClasses, |
| 5349 | const SimplifyQuery &Q, unsigned Depth) { |
| 5350 | auto *FVTy = dyn_cast<FixedVectorType>(Val: V->getType()); |
| 5351 | APInt DemandedElts = |
| 5352 | FVTy ? APInt::getAllOnes(numBits: FVTy->getNumElements()) : APInt(1, 1); |
| 5353 | computeKnownFPClass(V, DemandedElts, InterestedClasses, Known, Q, Depth); |
| 5354 | } |
| 5355 | |
| 5356 | static void computeKnownFPClassForFPTrunc(const Operator *Op, |
| 5357 | const APInt &DemandedElts, |
| 5358 | FPClassTest InterestedClasses, |
| 5359 | KnownFPClass &Known, |
| 5360 | const SimplifyQuery &Q, |
| 5361 | unsigned Depth) { |
| 5362 | if ((InterestedClasses & |
| 5363 | (KnownFPClass::OrderedLessThanZeroMask | fcNan)) == fcNone) |
| 5364 | return; |
| 5365 | |
| 5366 | KnownFPClass KnownSrc; |
| 5367 | computeKnownFPClass(V: Op->getOperand(i: 0), DemandedElts, InterestedClasses, |
| 5368 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 5369 | |
| 5370 | const fltSemantics &DstTy = Op->getType()->getScalarType()->getFltSemantics(); |
| 5371 | const fltSemantics &SrcTy = |
| 5372 | Op->getOperand(i: 0)->getType()->getScalarType()->getFltSemantics(); |
| 5373 | |
| 5374 | const auto *I = dyn_cast<Instruction>(Val: Op); |
| 5375 | const Function *F = I ? I->getFunction() : nullptr; |
| 5376 | DenormalMode DstMode = |
| 5377 | F ? F->getDenormalMode(FPType: DstTy) : DenormalMode::getDynamic(); |
| 5378 | DenormalMode SrcMode = |
| 5379 | F ? F->getDenormalMode(FPType: SrcTy) : DenormalMode::getDynamic(); |
| 5380 | |
| 5381 | Known = KnownFPClass::fptrunc(KnownSrc, DstTy, SrcTy, DstMode, SrcMode); |
| 5382 | } |
| 5383 | |
| 5384 | static void |
| 5385 | computeKnownFPClassForFAddSub(const Operator *Op, const APInt &DemandedElts, |
| 5386 | FPClassTest InterestedClasses, |
| 5387 | KnownFPClass &Known, const SimplifyQuery &Q, |
| 5388 | unsigned Depth, bool IsAdd, bool IsStrict) { |
| 5389 | const Value *LHS = Op->getOperand(i: 0); |
| 5390 | const Value *RHS = Op->getOperand(i: 1); |
| 5391 | |
| 5392 | const RoundingMode RM = |
| 5393 | IsStrict ? cast<ConstrainedFPIntrinsic>(Val: Op)->getRoundingMode().value_or( |
| 5394 | u: RoundingMode::Dynamic) |
| 5395 | : RoundingMode::NearestTiesToEven; |
| 5396 | |
| 5397 | KnownFPClass KnownLHS, KnownRHS; |
| 5398 | bool WantNegative = |
| 5399 | IsAdd && |
| 5400 | (InterestedClasses & KnownFPClass::OrderedLessThanZeroMask) != fcNone; |
| 5401 | bool WantNaN = (InterestedClasses & fcNan) != fcNone; |
| 5402 | bool WantNegZero = (InterestedClasses & fcNegZero) != fcNone; |
| 5403 | bool WantSubnormal = (InterestedClasses & fcSubnormal) != fcNone; |
| 5404 | |
| 5405 | if (!WantNaN && !WantNegative && !WantNegZero && !WantSubnormal) |
| 5406 | return; |
| 5407 | |
| 5408 | FPClassTest InterestedSrcs = InterestedClasses; |
| 5409 | if (WantNegative) |
| 5410 | InterestedSrcs |= KnownFPClass::OrderedLessThanZeroMask; |
| 5411 | if (InterestedClasses & fcNan) |
| 5412 | InterestedSrcs |= fcInf; |
| 5413 | computeKnownFPClass(V: RHS, DemandedElts, InterestedClasses: InterestedSrcs, Known&: KnownRHS, Q, |
| 5414 | Depth: Depth + 1); |
| 5415 | |
| 5416 | // Special case fadd x, x, which is the canonical form of fmul x, 2. |
| 5417 | bool Self = LHS == RHS && |
| 5418 | isGuaranteedNotToBeUndef(V: LHS, AC: Q.AC, CtxI: Q.CtxI, DT: Q.DT, Depth: Depth + 1); |
| 5419 | if (Self) |
| 5420 | KnownLHS = KnownRHS; |
| 5421 | |
| 5422 | if ((WantNaN && KnownRHS.isKnownNeverNaN()) || |
| 5423 | (WantNegative && KnownRHS.cannotBeOrderedLessThanZero()) || WantNegZero || |
| 5424 | WantSubnormal || !IsAdd) { |
| 5425 | |
| 5426 | // FIXME: Context function should always be passed in separately |
| 5427 | const Function *F = cast<Instruction>(Val: Op)->getFunction(); |
| 5428 | const fltSemantics &FltSem = |
| 5429 | Op->getType()->getScalarType()->getFltSemantics(); |
| 5430 | DenormalMode Mode = |
| 5431 | F ? F->getDenormalMode(FPType: FltSem) : DenormalMode::getDynamic(); |
| 5432 | |
| 5433 | if (Self && IsAdd) { |
| 5434 | Known = KnownFPClass::fadd_self(Src: KnownLHS, RM, Mode); |
| 5435 | } else { |
| 5436 | // RHS is canonically cheaper to compute. Skip inspecting the LHS if |
| 5437 | // there's no point. |
| 5438 | |
| 5439 | if (!Self) { |
| 5440 | computeKnownFPClass(V: LHS, DemandedElts, InterestedClasses: InterestedSrcs, Known&: KnownLHS, Q, |
| 5441 | Depth: Depth + 1); |
| 5442 | } |
| 5443 | |
| 5444 | if (const APFloat *CRHS; IsAdd && match(V: RHS, P: m_APFloat(Res&: CRHS))) { |
| 5445 | Known = KnownFPClass::fadd(LHS: KnownLHS, RHS: *CRHS, RM, Mode); |
| 5446 | return; |
| 5447 | } |
| 5448 | |
| 5449 | Known = IsAdd ? KnownFPClass::fadd(LHS: KnownLHS, RHS: KnownRHS, RM, Mode) |
| 5450 | : KnownFPClass::fsub(LHS: KnownLHS, RHS: KnownRHS, RM, Mode); |
| 5451 | } |
| 5452 | } |
| 5453 | } |
| 5454 | |
| 5455 | static void computeKnownFPClassForFRem(const Operator *Op, |
| 5456 | const APInt &DemandedElts, |
| 5457 | FPClassTest InterestedClasses, |
| 5458 | KnownFPClass &Known, |
| 5459 | const SimplifyQuery &Q, unsigned Depth) { |
| 5460 | const Value *LHS = Op->getOperand(i: 0); |
| 5461 | const Value *RHS = Op->getOperand(i: 1); |
| 5462 | |
| 5463 | FPClassTest InterestedLHS = fcNone; |
| 5464 | FPClassTest InterestedRHS = fcNone; |
| 5465 | |
| 5466 | // NaN is also generated for frem(Inf, x) and frem(x, 0.0). |
| 5467 | if (InterestedClasses & fcNan) { |
| 5468 | InterestedLHS |= fcNan | fcInf; |
| 5469 | InterestedRHS |= fcNan | fcZero | fcSubnormal; |
| 5470 | } |
| 5471 | |
| 5472 | // The sign for frem is the same as the first operand. |
| 5473 | if (InterestedClasses & (fcPosNormal | fcPosSubnormal)) |
| 5474 | InterestedLHS |= fcPosNormal | fcPosSubnormal; |
| 5475 | if (InterestedClasses & (fcNegNormal | fcNegSubnormal)) |
| 5476 | InterestedLHS |= fcNegNormal | fcNegSubnormal; |
| 5477 | |
| 5478 | // A negative zero result requires a negative finite first operand. |
| 5479 | if (InterestedClasses & fcNegZero) |
| 5480 | InterestedLHS |= fcNegFinite; |
| 5481 | |
| 5482 | // A positive zero result can additionally come from a negative finite |
| 5483 | // result being flushed to positive zero. |
| 5484 | if (InterestedClasses & fcPosZero) |
| 5485 | InterestedLHS |= fcPosFinite | fcNegNormal | fcNegSubnormal; |
| 5486 | |
| 5487 | const Function *F = cast<Instruction>(Val: Op)->getFunction(); |
| 5488 | DenormalMode Mode = |
| 5489 | F ? F->getDenormalMode(FPType: Op->getType()->getScalarType()->getFltSemantics()) |
| 5490 | : DenormalMode::getDynamic(); |
| 5491 | |
| 5492 | if (LHS == RHS && isGuaranteedNotToBeUndef(V: LHS, AC: Q.AC, CtxI: Q.CtxI, DT: Q.DT)) { |
| 5493 | // X % X is always exactly [+-]0.0 or a NaN. |
| 5494 | FPClassTest InterestedSrcs = InterestedLHS | InterestedRHS; |
| 5495 | KnownFPClass KnownSrc; |
| 5496 | if (InterestedSrcs != fcNone) |
| 5497 | computeKnownFPClass(V: LHS, DemandedElts, InterestedClasses: InterestedSrcs, Known&: KnownSrc, Q, |
| 5498 | Depth: Depth + 1); |
| 5499 | Known = KnownFPClass::frem_self(Src: KnownSrc, Mode); |
| 5500 | return; |
| 5501 | } |
| 5502 | |
| 5503 | KnownFPClass KnownLHS; |
| 5504 | if (InterestedLHS != fcNone) |
| 5505 | computeKnownFPClass(V: LHS, DemandedElts, InterestedClasses: InterestedLHS, Known&: KnownLHS, Q, |
| 5506 | Depth: Depth + 1); |
| 5507 | |
| 5508 | KnownFPClass KnownRHS; |
| 5509 | // RHS is only useful for refining NaN classes. |
| 5510 | if (InterestedRHS != fcNone && KnownLHS.isKnownNever(Mask: fcSNan)) |
| 5511 | computeKnownFPClass(V: RHS, DemandedElts, InterestedClasses: InterestedRHS, Known&: KnownRHS, Q, |
| 5512 | Depth: Depth + 1); |
| 5513 | |
| 5514 | Known = KnownFPClass::frem(LHS: KnownLHS, RHS: KnownRHS, Mode); |
| 5515 | } |
| 5516 | |
| 5517 | static constexpr KnownFPClass::MinMaxKind getMinMaxKind(Intrinsic::ID IID) { |
| 5518 | switch (IID) { |
| 5519 | case Intrinsic::minimum: |
| 5520 | return KnownFPClass::MinMaxKind::minimum; |
| 5521 | case Intrinsic::maximum: |
| 5522 | return KnownFPClass::MinMaxKind::maximum; |
| 5523 | case Intrinsic::minimumnum: |
| 5524 | return KnownFPClass::MinMaxKind::minimumnum; |
| 5525 | case Intrinsic::maximumnum: |
| 5526 | return KnownFPClass::MinMaxKind::maximumnum; |
| 5527 | case Intrinsic::minnum: |
| 5528 | return KnownFPClass::MinMaxKind::minnum; |
| 5529 | case Intrinsic::maxnum: |
| 5530 | return KnownFPClass::MinMaxKind::maxnum; |
| 5531 | default: |
| 5532 | llvm_unreachable("not a floating-point min-max intrinsic" ); |
| 5533 | } |
| 5534 | } |
| 5535 | |
| 5536 | /// \return true if this is a floating point value that is known to have a |
| 5537 | /// magnitude smaller than 1. i.e., fabs(X) <= 1.0 or is nan. |
| 5538 | static bool isAbsoluteValueULEOne(const Value *V) { |
| 5539 | // TODO: Handle frexp |
| 5540 | // TODO: Other rounding intrinsics? |
| 5541 | // TODO: Try computeKnownExponentRangeFromContext |
| 5542 | |
| 5543 | // fabs(x - floor(x)) <= 1 |
| 5544 | const Value *SubFloorX; |
| 5545 | if (match(V, P: m_FSub(L: m_Value(V&: SubFloorX), |
| 5546 | R: m_Intrinsic<Intrinsic::floor>(Ops: m_Deferred(V: SubFloorX))))) |
| 5547 | return true; |
| 5548 | |
| 5549 | return match(V, P: m_Intrinsic<Intrinsic::amdgcn_trig_preop>(Ops: m_Value())) || |
| 5550 | match(V, P: m_Intrinsic<Intrinsic::amdgcn_fract>(Ops: m_Value())); |
| 5551 | } |
| 5552 | |
| 5553 | void computeKnownFPClass(const Value *V, const APInt &DemandedElts, |
| 5554 | FPClassTest InterestedClasses, KnownFPClass &Known, |
| 5555 | const SimplifyQuery &Q, unsigned Depth) { |
| 5556 | assert(Known.isUnknown() && "should not be called with known information" ); |
| 5557 | |
| 5558 | if (!DemandedElts) { |
| 5559 | // No demanded elts, better to assume we don't know anything. |
| 5560 | Known.resetAll(); |
| 5561 | return; |
| 5562 | } |
| 5563 | |
| 5564 | assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth" ); |
| 5565 | |
| 5566 | if (auto *CFP = dyn_cast<ConstantFP>(Val: V)) { |
| 5567 | Known = KnownFPClass(CFP->getValueAPF()); |
| 5568 | return; |
| 5569 | } |
| 5570 | |
| 5571 | if (isa<ConstantAggregateZero>(Val: V)) { |
| 5572 | Known.setKnownFPClasses(fcPosZero); |
| 5573 | return; |
| 5574 | } |
| 5575 | |
| 5576 | if (isa<PoisonValue>(Val: V)) { |
| 5577 | Known = KnownFPClass::getEmpty(); |
| 5578 | return; |
| 5579 | } |
| 5580 | |
| 5581 | // Try to handle fixed width vector constants |
| 5582 | auto *VFVTy = dyn_cast<FixedVectorType>(Val: V->getType()); |
| 5583 | const Constant *CV = dyn_cast<Constant>(Val: V); |
| 5584 | if (VFVTy && CV) { |
| 5585 | Known = KnownFPClass::getEmpty(); |
| 5586 | |
| 5587 | // For vectors, verify that each element is not NaN. |
| 5588 | unsigned NumElts = VFVTy->getNumElements(); |
| 5589 | for (unsigned i = 0; i != NumElts; ++i) { |
| 5590 | if (!DemandedElts[i]) |
| 5591 | continue; |
| 5592 | |
| 5593 | Constant *Elt = CV->getAggregateElement(Elt: i); |
| 5594 | if (!Elt) { |
| 5595 | Known = KnownFPClass(); |
| 5596 | return; |
| 5597 | } |
| 5598 | if (isa<PoisonValue>(Val: Elt)) |
| 5599 | continue; |
| 5600 | auto *CElt = dyn_cast<ConstantFP>(Val: Elt); |
| 5601 | if (!CElt) { |
| 5602 | Known = KnownFPClass(); |
| 5603 | return; |
| 5604 | } |
| 5605 | |
| 5606 | Known |= KnownFPClass(CElt->getValueAPF()); |
| 5607 | } |
| 5608 | return; |
| 5609 | } |
| 5610 | |
| 5611 | if (const auto *CDS = dyn_cast<ConstantDataSequential>(Val: V)) { |
| 5612 | Known = KnownFPClass::getEmpty(); |
| 5613 | for (size_t I = 0, E = CDS->getNumElements(); I != E; ++I) |
| 5614 | Known |= KnownFPClass(CDS->getElementAsAPFloat(i: I)); |
| 5615 | return; |
| 5616 | } |
| 5617 | |
| 5618 | if (const auto *CA = dyn_cast<ConstantAggregate>(Val: V)) { |
| 5619 | // TODO: Handle complex aggregates |
| 5620 | Known = KnownFPClass::getEmpty(); |
| 5621 | for (const Use &Op : CA->operands()) { |
| 5622 | auto *CFP = dyn_cast<ConstantFP>(Val: Op.get()); |
| 5623 | if (!CFP) { |
| 5624 | Known = KnownFPClass(); |
| 5625 | return; |
| 5626 | } |
| 5627 | |
| 5628 | Known |= KnownFPClass(CFP->getValueAPF()); |
| 5629 | } |
| 5630 | |
| 5631 | return; |
| 5632 | } |
| 5633 | |
| 5634 | FPClassTest KnownNotFromFlags = fcNone; |
| 5635 | if (const auto *CB = dyn_cast<CallBase>(Val: V)) |
| 5636 | KnownNotFromFlags |= CB->getRetNoFPClass(); |
| 5637 | else if (const auto *Arg = dyn_cast<Argument>(Val: V)) |
| 5638 | KnownNotFromFlags |= Arg->getNoFPClass(); |
| 5639 | |
| 5640 | const Operator *Op = dyn_cast<Operator>(Val: V); |
| 5641 | if (const FPMathOperator *FPOp = dyn_cast_or_null<FPMathOperator>(Val: Op)) { |
| 5642 | if (FPOp->hasNoNaNs()) |
| 5643 | KnownNotFromFlags |= fcNan; |
| 5644 | if (FPOp->hasNoInfs()) |
| 5645 | KnownNotFromFlags |= fcInf; |
| 5646 | } |
| 5647 | |
| 5648 | KnownFPClass AssumedClasses = computeKnownFPClassFromContext(V, Q); |
| 5649 | KnownNotFromFlags |= ~AssumedClasses.getKnownFPClasses(); |
| 5650 | |
| 5651 | // We no longer need to find out about these bits from inputs if we can |
| 5652 | // assume this from flags/attributes. |
| 5653 | InterestedClasses &= ~KnownNotFromFlags; |
| 5654 | |
| 5655 | llvm::scope_exit ClearClassesFromFlags([=, &Known] { |
| 5656 | Known.knownNot(RuleOut: KnownNotFromFlags); |
| 5657 | if (!Known.getSignBit() && AssumedClasses.getSignBit()) { |
| 5658 | if (*AssumedClasses.getSignBit()) |
| 5659 | Known.signBitMustBeOne(); |
| 5660 | else |
| 5661 | Known.signBitMustBeZero(); |
| 5662 | } |
| 5663 | }); |
| 5664 | |
| 5665 | if (!Op) |
| 5666 | return; |
| 5667 | |
| 5668 | // All recursive calls that increase depth must come after this. |
| 5669 | if (Depth == MaxAnalysisRecursionDepth) |
| 5670 | return; |
| 5671 | |
| 5672 | const unsigned Opc = Op->getOpcode(); |
| 5673 | switch (Opc) { |
| 5674 | case Instruction::FNeg: { |
| 5675 | computeKnownFPClass(V: Op->getOperand(i: 0), DemandedElts, InterestedClasses, |
| 5676 | Known, Q, Depth: Depth + 1); |
| 5677 | Known.fneg(); |
| 5678 | break; |
| 5679 | } |
| 5680 | case Instruction::Select: { |
| 5681 | auto ComputeForArm = [&](Value *Arm, bool Invert) { |
| 5682 | KnownFPClass Res; |
| 5683 | computeKnownFPClass(V: Arm, DemandedElts, InterestedClasses, Known&: Res, Q, |
| 5684 | Depth: Depth + 1); |
| 5685 | adjustKnownFPClassForSelectArm(Known&: Res, Cond: Op->getOperand(i: 0), Arm, Invert, SQ: Q, |
| 5686 | Depth); |
| 5687 | return Res; |
| 5688 | }; |
| 5689 | // Only known if known in both the LHS and RHS. |
| 5690 | Known = |
| 5691 | ComputeForArm(Op->getOperand(i: 1), /*Invert=*/false) |
| 5692 | .intersectWith(RHS: ComputeForArm(Op->getOperand(i: 2), /*Invert=*/true)); |
| 5693 | break; |
| 5694 | } |
| 5695 | case Instruction::Load: { |
| 5696 | const MDNode *NoFPClass = |
| 5697 | cast<LoadInst>(Val: Op)->getMetadata(KindID: LLVMContext::MD_nofpclass); |
| 5698 | if (!NoFPClass) |
| 5699 | break; |
| 5700 | |
| 5701 | ConstantInt *MaskVal = |
| 5702 | mdconst::extract<ConstantInt>(MD: NoFPClass->getOperand(I: 0)); |
| 5703 | Known.knownNot(RuleOut: static_cast<FPClassTest>(MaskVal->getZExtValue())); |
| 5704 | break; |
| 5705 | } |
| 5706 | case Instruction::Call: { |
| 5707 | const CallInst *II = cast<CallInst>(Val: Op); |
| 5708 | const Intrinsic::ID IID = II->getIntrinsicID(); |
| 5709 | switch (IID) { |
| 5710 | case Intrinsic::fabs: { |
| 5711 | if ((InterestedClasses & (fcNan | fcPositive)) != fcNone) { |
| 5712 | // If we only care about the sign bit we don't need to inspect the |
| 5713 | // operand. |
| 5714 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, |
| 5715 | InterestedClasses, Known, Q, Depth: Depth + 1); |
| 5716 | } |
| 5717 | |
| 5718 | Known.fabs(); |
| 5719 | break; |
| 5720 | } |
| 5721 | case Intrinsic::copysign: { |
| 5722 | KnownFPClass KnownSign; |
| 5723 | |
| 5724 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
| 5725 | Known, Q, Depth: Depth + 1); |
| 5726 | computeKnownFPClass(V: II->getArgOperand(i: 1), DemandedElts, InterestedClasses, |
| 5727 | Known&: KnownSign, Q, Depth: Depth + 1); |
| 5728 | Known.copysign(Sign: KnownSign); |
| 5729 | break; |
| 5730 | } |
| 5731 | case Intrinsic::fma: |
| 5732 | case Intrinsic::fmuladd: { |
| 5733 | if ((InterestedClasses & (fcNegative | fcNan)) == fcNone) |
| 5734 | break; |
| 5735 | |
| 5736 | const bool IsFMA = IID == Intrinsic::fma; |
| 5737 | |
| 5738 | // FIXME: This should check isGuaranteedNotToBeUndef |
| 5739 | if (II->getArgOperand(i: 0) == II->getArgOperand(i: 1)) { |
| 5740 | FPClassTest InterestedSquared = InterestedClasses; |
| 5741 | FPClassTest InterestedAddend = InterestedClasses; |
| 5742 | // Only fma(+inf, +inf, -inf) and fma(-inf, -inf, -inf) produce NaN when |
| 5743 | // squaring. |
| 5744 | if (InterestedClasses & fcQNan) { |
| 5745 | InterestedSquared |= fcSNan | fcInf; |
| 5746 | InterestedAddend |= fcSNan | fcNegInf; |
| 5747 | } |
| 5748 | |
| 5749 | KnownFPClass KnownSrc, KnownAddend; |
| 5750 | computeKnownFPClass(V: II->getArgOperand(i: 2), DemandedElts, |
| 5751 | InterestedClasses: InterestedAddend, Known&: KnownAddend, Q, Depth: Depth + 1); |
| 5752 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, |
| 5753 | InterestedClasses: InterestedSquared, Known&: KnownSrc, Q, Depth: Depth + 1); |
| 5754 | |
| 5755 | const Function *F = II->getFunction(); |
| 5756 | const fltSemantics &FltSem = |
| 5757 | II->getType()->getScalarType()->getFltSemantics(); |
| 5758 | DenormalMode Mode = |
| 5759 | F ? F->getDenormalMode(FPType: FltSem) : DenormalMode::getDynamic(); |
| 5760 | |
| 5761 | if ((KnownNotFromFlags & fcNan) == fcNan) { |
| 5762 | KnownSrc.knownNot(RuleOut: fcNan); |
| 5763 | KnownAddend.knownNot(RuleOut: fcNan); |
| 5764 | } |
| 5765 | |
| 5766 | if ((KnownNotFromFlags & fcInf) == fcInf) { |
| 5767 | KnownSrc.knownNot(RuleOut: fcInf); |
| 5768 | KnownAddend.knownNot(RuleOut: fcInf); |
| 5769 | } |
| 5770 | |
| 5771 | Known = IsFMA |
| 5772 | ? KnownFPClass::fma_square(Squared: KnownSrc, Addend: KnownAddend, Mode) |
| 5773 | : KnownFPClass::fmuladd_square(Squared: KnownSrc, Addend: KnownAddend, Mode); |
| 5774 | break; |
| 5775 | } |
| 5776 | |
| 5777 | KnownFPClass KnownSrc[3]; |
| 5778 | for (int I = 0; I != 3; ++I) { |
| 5779 | computeKnownFPClass(V: II->getArgOperand(i: I), DemandedElts, |
| 5780 | InterestedClasses, Known&: KnownSrc[I], Q, Depth: Depth + 1); |
| 5781 | if (KnownSrc[I].isUnknown()) |
| 5782 | return; |
| 5783 | |
| 5784 | if ((KnownNotFromFlags & fcNan) == fcNan) |
| 5785 | KnownSrc[I].knownNot(RuleOut: fcNan); |
| 5786 | if ((KnownNotFromFlags & fcInf) == fcInf) |
| 5787 | KnownSrc[I].knownNot(RuleOut: fcInf); |
| 5788 | } |
| 5789 | |
| 5790 | const Function *F = II->getFunction(); |
| 5791 | const fltSemantics &FltSem = |
| 5792 | II->getType()->getScalarType()->getFltSemantics(); |
| 5793 | DenormalMode Mode = |
| 5794 | F ? F->getDenormalMode(FPType: FltSem) : DenormalMode::getDynamic(); |
| 5795 | Known = |
| 5796 | IsFMA ? KnownFPClass::fma(LHS: KnownSrc[0], RHS: KnownSrc[1], Addend: KnownSrc[2], Mode) |
| 5797 | : KnownFPClass::fmuladd(LHS: KnownSrc[0], RHS: KnownSrc[1], Addend: KnownSrc[2], |
| 5798 | Mode); |
| 5799 | break; |
| 5800 | } |
| 5801 | case Intrinsic::experimental_constrained_fadd: |
| 5802 | case Intrinsic::experimental_constrained_fsub: { |
| 5803 | computeKnownFPClassForFAddSub( |
| 5804 | Op, DemandedElts, InterestedClasses, Known, Q, Depth, |
| 5805 | /*IsAdd=*/IID == Intrinsic::experimental_constrained_fadd, |
| 5806 | /*IsStrict=*/true); |
| 5807 | break; |
| 5808 | } |
| 5809 | case Intrinsic::experimental_constrained_frem: { |
| 5810 | computeKnownFPClassForFRem(Op, DemandedElts, InterestedClasses, Known, Q, |
| 5811 | Depth); |
| 5812 | break; |
| 5813 | } |
| 5814 | case Intrinsic::sqrt: |
| 5815 | case Intrinsic::experimental_constrained_sqrt: { |
| 5816 | KnownFPClass KnownSrc; |
| 5817 | FPClassTest InterestedSrcs = InterestedClasses; |
| 5818 | if (InterestedClasses & fcQNan) |
| 5819 | InterestedSrcs |= KnownFPClass::OrderedLessThanZeroMask | fcSNan; |
| 5820 | // Subnormals may flush to zero. |
| 5821 | if (InterestedClasses & fcZero) |
| 5822 | InterestedSrcs |= fcSubnormal; |
| 5823 | |
| 5824 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses: InterestedSrcs, |
| 5825 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 5826 | |
| 5827 | DenormalMode Mode = DenormalMode::getDynamic(); |
| 5828 | |
| 5829 | bool HasNSZ = Q.IIQ.hasNoSignedZeros(Op: II); |
| 5830 | if (!HasNSZ) { |
| 5831 | const Function *F = II->getFunction(); |
| 5832 | const fltSemantics &FltSem = |
| 5833 | II->getType()->getScalarType()->getFltSemantics(); |
| 5834 | Mode = F ? F->getDenormalMode(FPType: FltSem) : DenormalMode::getDynamic(); |
| 5835 | } |
| 5836 | |
| 5837 | Known = KnownFPClass::sqrt(Src: KnownSrc, Mode); |
| 5838 | if (HasNSZ) |
| 5839 | Known.knownNot(RuleOut: fcNegZero); |
| 5840 | |
| 5841 | break; |
| 5842 | } |
| 5843 | case Intrinsic::sin: { |
| 5844 | KnownFPClass KnownSrc; |
| 5845 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
| 5846 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 5847 | Known = KnownFPClass::sin(Src: KnownSrc); |
| 5848 | break; |
| 5849 | } |
| 5850 | case Intrinsic::cos: { |
| 5851 | KnownFPClass KnownSrc; |
| 5852 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
| 5853 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 5854 | Known = KnownFPClass::cos(Src: KnownSrc); |
| 5855 | break; |
| 5856 | } |
| 5857 | case Intrinsic::tan: { |
| 5858 | KnownFPClass KnownSrc; |
| 5859 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
| 5860 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 5861 | Known = KnownFPClass::tan(Src: KnownSrc); |
| 5862 | break; |
| 5863 | } |
| 5864 | case Intrinsic::sinh: { |
| 5865 | KnownFPClass KnownSrc; |
| 5866 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
| 5867 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 5868 | Known = KnownFPClass::sinh(Src: KnownSrc); |
| 5869 | break; |
| 5870 | } |
| 5871 | case Intrinsic::cosh: { |
| 5872 | KnownFPClass KnownSrc; |
| 5873 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
| 5874 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 5875 | Known = KnownFPClass::cosh(Src: KnownSrc); |
| 5876 | break; |
| 5877 | } |
| 5878 | case Intrinsic::tanh: { |
| 5879 | KnownFPClass KnownSrc; |
| 5880 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
| 5881 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 5882 | Known = KnownFPClass::tanh(Src: KnownSrc); |
| 5883 | break; |
| 5884 | } |
| 5885 | case Intrinsic::asin: { |
| 5886 | KnownFPClass KnownSrc; |
| 5887 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
| 5888 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 5889 | Known = KnownFPClass::asin(Src: KnownSrc); |
| 5890 | break; |
| 5891 | } |
| 5892 | case Intrinsic::acos: { |
| 5893 | KnownFPClass KnownSrc; |
| 5894 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
| 5895 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 5896 | Known = KnownFPClass::acos(Src: KnownSrc); |
| 5897 | break; |
| 5898 | } |
| 5899 | case Intrinsic::atan: { |
| 5900 | KnownFPClass KnownSrc; |
| 5901 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
| 5902 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 5903 | Known = KnownFPClass::atan(Src: KnownSrc); |
| 5904 | break; |
| 5905 | } |
| 5906 | case Intrinsic::atan2: { |
| 5907 | FPClassTest InterestedY = InterestedClasses; |
| 5908 | FPClassTest InterestedX = InterestedClasses; |
| 5909 | |
| 5910 | // We can rule out negative values if y cannot have a negative value. |
| 5911 | if ((InterestedClasses & fcNegFinite) != fcNone) |
| 5912 | InterestedY |= fcNegative; |
| 5913 | |
| 5914 | // We can rule out positive values if y cannot have a positive value. |
| 5915 | if ((InterestedClasses & fcPosFinite) != fcNone) |
| 5916 | InterestedY |= fcPositive | fcNegSubnormal; |
| 5917 | |
| 5918 | // We can rule out zero and subnormal if x cannot have a positive value. |
| 5919 | if ((InterestedClasses & (fcZero | fcSubnormal)) != fcNone) |
| 5920 | InterestedX |= fcPositive | fcNegSubnormal; |
| 5921 | |
| 5922 | KnownFPClass KnownY, KnownX; |
| 5923 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses: InterestedY, |
| 5924 | Known&: KnownY, Q, Depth: Depth + 1); |
| 5925 | computeKnownFPClass(V: II->getArgOperand(i: 1), DemandedElts, InterestedClasses: InterestedX, |
| 5926 | Known&: KnownX, Q, Depth: Depth + 1); |
| 5927 | |
| 5928 | const Function *F = II->getFunction(); |
| 5929 | DenormalMode Mode = |
| 5930 | F ? F->getDenormalMode( |
| 5931 | FPType: II->getType()->getScalarType()->getFltSemantics()) |
| 5932 | : DenormalMode::getDynamic(); |
| 5933 | Known = KnownFPClass::atan2(LHS: KnownY, RHS: KnownX, Mode); |
| 5934 | break; |
| 5935 | } |
| 5936 | case Intrinsic::maxnum: |
| 5937 | case Intrinsic::minnum: |
| 5938 | case Intrinsic::minimum: |
| 5939 | case Intrinsic::maximum: |
| 5940 | case Intrinsic::minimumnum: |
| 5941 | case Intrinsic::maximumnum: { |
| 5942 | KnownFPClass KnownLHS, KnownRHS; |
| 5943 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
| 5944 | Known&: KnownLHS, Q, Depth: Depth + 1); |
| 5945 | computeKnownFPClass(V: II->getArgOperand(i: 1), DemandedElts, InterestedClasses, |
| 5946 | Known&: KnownRHS, Q, Depth: Depth + 1); |
| 5947 | |
| 5948 | const Function *F = II->getFunction(); |
| 5949 | |
| 5950 | DenormalMode Mode = |
| 5951 | F ? F->getDenormalMode( |
| 5952 | FPType: II->getType()->getScalarType()->getFltSemantics()) |
| 5953 | : DenormalMode::getDynamic(); |
| 5954 | |
| 5955 | Known = KnownFPClass::minMaxLike(LHS: KnownLHS, RHS: KnownRHS, Kind: getMinMaxKind(IID), |
| 5956 | Mode); |
| 5957 | break; |
| 5958 | } |
| 5959 | case Intrinsic::canonicalize: { |
| 5960 | KnownFPClass KnownSrc; |
| 5961 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
| 5962 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 5963 | |
| 5964 | const Function *F = II->getFunction(); |
| 5965 | const fltSemantics &Flt = |
| 5966 | II->getType()->getScalarType()->getFltSemantics(); |
| 5967 | DenormalMode Mode = |
| 5968 | F ? F->getDenormalMode(FPType: Flt) : DenormalMode::getDynamic(); |
| 5969 | Known = KnownFPClass::canonicalize(Src: KnownSrc, Flt, Mode); |
| 5970 | break; |
| 5971 | } |
| 5972 | case Intrinsic::vector_reduce_fmax: |
| 5973 | case Intrinsic::vector_reduce_fmin: |
| 5974 | case Intrinsic::vector_reduce_fmaximum: |
| 5975 | case Intrinsic::vector_reduce_fminimum: |
| 5976 | case Intrinsic::vector_reduce_fmaximumnum: |
| 5977 | case Intrinsic::vector_reduce_fminimumnum: { |
| 5978 | // reduce min/max will choose an element from one of the vector elements, |
| 5979 | // so we can infer and class information that is common to all elements. |
| 5980 | Known = computeKnownFPClass(V: II->getArgOperand(i: 0), FMF: II->getFastMathFlags(), |
| 5981 | InterestedClasses, SQ: Q, Depth: Depth + 1); |
| 5982 | // Can only propagate sign if output is never NaN. |
| 5983 | if (!Known.isKnownNeverNaN()) |
| 5984 | Known.setSignBit(std::nullopt); |
| 5985 | break; |
| 5986 | } |
| 5987 | // reverse preserves all characteristics of the input vec's element. |
| 5988 | case Intrinsic::vector_reverse: |
| 5989 | Known = computeKnownFPClass( |
| 5990 | V: II->getArgOperand(i: 0), DemandedElts: DemandedElts.reverseBits(), |
| 5991 | FMF: II->getFastMathFlags(), InterestedClasses, SQ: Q, Depth: Depth + 1); |
| 5992 | break; |
| 5993 | case Intrinsic::trunc: |
| 5994 | case Intrinsic::floor: |
| 5995 | case Intrinsic::ceil: |
| 5996 | case Intrinsic::rint: |
| 5997 | case Intrinsic::nearbyint: |
| 5998 | case Intrinsic::round: |
| 5999 | case Intrinsic::roundeven: { |
| 6000 | KnownFPClass KnownSrc; |
| 6001 | FPClassTest InterestedSrcs = InterestedClasses; |
| 6002 | |
| 6003 | // Negative round ups towards zero produce negative zero. |
| 6004 | if (InterestedSrcs & fcNegFinite) |
| 6005 | InterestedSrcs |= fcNegFinite; |
| 6006 | |
| 6007 | // Negative subnormals may flush to positive zero. |
| 6008 | if (InterestedSrcs & fcPosFinite) |
| 6009 | InterestedSrcs |= fcPosFinite | fcNegSubnormal; |
| 6010 | |
| 6011 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses: InterestedSrcs, |
| 6012 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 6013 | |
| 6014 | const Function *F = II->getFunction(); |
| 6015 | DenormalMode Mode = |
| 6016 | F ? F->getDenormalMode( |
| 6017 | FPType: II->getType()->getScalarType()->getFltSemantics()) |
| 6018 | : DenormalMode::getDynamic(); |
| 6019 | const bool IsMultiUnitFPType = |
| 6020 | V->getType()->getScalarType()->isMultiUnitFPType(); |
| 6021 | |
| 6022 | switch (IID) { |
| 6023 | case Intrinsic::trunc: |
| 6024 | Known = KnownFPClass::trunc(Src: KnownSrc, Mode); |
| 6025 | break; |
| 6026 | case Intrinsic::floor: |
| 6027 | Known = KnownFPClass::floor(Src: KnownSrc, IsMultiUnitFPType, Mode); |
| 6028 | break; |
| 6029 | case Intrinsic::ceil: |
| 6030 | Known = KnownFPClass::ceil(Src: KnownSrc, IsMultiUnitFPType, Mode); |
| 6031 | break; |
| 6032 | default: |
| 6033 | Known = |
| 6034 | KnownFPClass::roundToIntegral(Src: KnownSrc, IsMultiUnitFPType, Mode); |
| 6035 | break; |
| 6036 | } |
| 6037 | break; |
| 6038 | } |
| 6039 | case Intrinsic::exp: |
| 6040 | case Intrinsic::exp2: |
| 6041 | case Intrinsic::exp10: |
| 6042 | case Intrinsic::amdgcn_exp2: { |
| 6043 | KnownFPClass KnownSrc; |
| 6044 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
| 6045 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 6046 | |
| 6047 | Known = KnownFPClass::exp(Src: KnownSrc); |
| 6048 | |
| 6049 | Type *EltTy = II->getType()->getScalarType(); |
| 6050 | if (IID == Intrinsic::amdgcn_exp2 && EltTy->isFloatTy()) |
| 6051 | Known.knownNot(RuleOut: fcSubnormal); |
| 6052 | |
| 6053 | break; |
| 6054 | } |
| 6055 | case Intrinsic::fptrunc_round: { |
| 6056 | computeKnownFPClassForFPTrunc(Op, DemandedElts, InterestedClasses, Known, |
| 6057 | Q, Depth); |
| 6058 | break; |
| 6059 | } |
| 6060 | case Intrinsic::log: |
| 6061 | case Intrinsic::log10: |
| 6062 | case Intrinsic::log2: |
| 6063 | case Intrinsic::experimental_constrained_log: |
| 6064 | case Intrinsic::experimental_constrained_log10: |
| 6065 | case Intrinsic::experimental_constrained_log2: |
| 6066 | case Intrinsic::amdgcn_log: { |
| 6067 | FPClassTest InterestedSrcs = fcNone; |
| 6068 | |
| 6069 | // log(negative) produces NaN. |
| 6070 | if ((InterestedClasses & fcNan) != fcNone) |
| 6071 | InterestedSrcs |= fcNan | fcNegative; |
| 6072 | |
| 6073 | // log(logical-zero) produces negative infinity. |
| 6074 | if ((InterestedClasses & fcNegInf) != fcNone) |
| 6075 | InterestedSrcs |= fcZero | fcSubnormal; |
| 6076 | |
| 6077 | // log(x) < -0.0 if x < +1.0 |
| 6078 | if ((InterestedClasses & fcNegNormal) != fcNone) |
| 6079 | InterestedSrcs |= fcPosSubnormal | fcPosNormal; |
| 6080 | |
| 6081 | // log(x) >= +0.0 if x >= +1.0 |
| 6082 | if ((InterestedClasses & (fcPosZero | fcPosNormal)) != fcNone) |
| 6083 | InterestedSrcs |= fcPosNormal; |
| 6084 | |
| 6085 | // log(x) is positive infinity iff x is positive infinity. |
| 6086 | if ((InterestedClasses & fcPosInf) != fcNone) |
| 6087 | InterestedSrcs |= fcPosInf; |
| 6088 | |
| 6089 | KnownFPClass KnownSrc; |
| 6090 | if (InterestedSrcs != fcNone) |
| 6091 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses: InterestedSrcs, |
| 6092 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 6093 | const Function *F = II->getFunction(); |
| 6094 | DenormalMode Mode = |
| 6095 | F ? F->getDenormalMode( |
| 6096 | FPType: II->getType()->getScalarType()->getFltSemantics()) |
| 6097 | : DenormalMode::getDynamic(); |
| 6098 | const bool IsMultiUnitFPType = |
| 6099 | V->getType()->getScalarType()->isMultiUnitFPType(); |
| 6100 | Known = KnownFPClass::log(Src: KnownSrc, Mode, IsMultiUnitFPType); |
| 6101 | break; |
| 6102 | } |
| 6103 | case Intrinsic::pow: { |
| 6104 | const bool WantNaN = (InterestedClasses & fcNan) != fcNone; |
| 6105 | const bool WantNegative = (InterestedClasses & fcNegative) != fcNone; |
| 6106 | if (!WantNaN && !WantNegative) |
| 6107 | break; |
| 6108 | |
| 6109 | FPClassTest InterestedLHS = fcNone; |
| 6110 | FPClassTest InterestedRHS = fcNone; |
| 6111 | if (WantNaN) { |
| 6112 | // pow may return NaN if one of the arguments is NaN. NaN may also be |
| 6113 | // produced from a negative, non-zero finite base and a non-integer |
| 6114 | // exponent. |
| 6115 | InterestedLHS |= fcNan | fcNegNormal | fcNegSubnormal; |
| 6116 | InterestedRHS |= fcNan; |
| 6117 | } |
| 6118 | if (WantNegative) { |
| 6119 | // A negative value is returned when a negative base is raised to an odd |
| 6120 | // integer power. Only normal values can be odd integers. |
| 6121 | InterestedLHS |= fcNegative; |
| 6122 | InterestedRHS |= fcNormal; |
| 6123 | } |
| 6124 | |
| 6125 | KnownFPClass KnownLHS; |
| 6126 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses: InterestedLHS, |
| 6127 | Known&: KnownLHS, Q, Depth: Depth + 1); |
| 6128 | |
| 6129 | // If the LHS is unknown, then querying the RHS is only useful for rare |
| 6130 | // edge cases. |
| 6131 | if (KnownLHS.isUnknown()) |
| 6132 | break; |
| 6133 | |
| 6134 | KnownFPClass KnownRHS; |
| 6135 | computeKnownFPClass(V: II->getArgOperand(i: 1), DemandedElts, InterestedClasses: InterestedRHS, |
| 6136 | Known&: KnownRHS, Q, Depth: Depth + 1); |
| 6137 | Known = KnownFPClass::pow(LHS: KnownLHS, RHS: KnownRHS); |
| 6138 | break; |
| 6139 | } |
| 6140 | case Intrinsic::powi: { |
| 6141 | if ((InterestedClasses & (fcNan | fcInf | fcNegative)) == fcNone) |
| 6142 | break; |
| 6143 | |
| 6144 | // The exponent is always a scalar, even when raising a vector to a power. |
| 6145 | const Value *Exp = II->getArgOperand(i: 1); |
| 6146 | unsigned BitWidth = Exp->getType()->getIntegerBitWidth(); |
| 6147 | KnownBits ExponentKnownBits(BitWidth); |
| 6148 | computeKnownBits(V: Exp, DemandedElts: APInt(1, 1), Known&: ExponentKnownBits, Q, Depth: Depth + 1); |
| 6149 | |
| 6150 | FPClassTest InterestedSrcs = fcNone; |
| 6151 | if (InterestedClasses & fcNan) |
| 6152 | InterestedSrcs |= fcNan; |
| 6153 | if (!ExponentKnownBits.isZero()) { |
| 6154 | if (InterestedClasses & fcInf) |
| 6155 | InterestedSrcs |= fcFinite | fcInf; |
| 6156 | if ((InterestedClasses & fcNegative) && !ExponentKnownBits.isEven()) |
| 6157 | InterestedSrcs |= fcNegative; |
| 6158 | } |
| 6159 | |
| 6160 | KnownFPClass KnownSrc; |
| 6161 | if (InterestedSrcs != fcNone) |
| 6162 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses: InterestedSrcs, |
| 6163 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 6164 | |
| 6165 | Known = KnownFPClass::powi(Src: KnownSrc, N: ExponentKnownBits); |
| 6166 | break; |
| 6167 | } |
| 6168 | case Intrinsic::ldexp: { |
| 6169 | KnownFPClass KnownSrc; |
| 6170 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
| 6171 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 6172 | // Can refine inf/zero handling based on the exponent operand. |
| 6173 | const FPClassTest ExpInfoMask = fcZero | fcSubnormal | fcInf; |
| 6174 | |
| 6175 | const Value *ExpArg = II->getArgOperand(i: 1); |
| 6176 | ConstantRange ExpKnownRange = |
| 6177 | ((KnownSrc.getKnownFPClasses() & ExpInfoMask) != fcNone) |
| 6178 | ? computeConstantRange(V: ExpArg, /*ForSigned=*/true, SQ: Q, Depth: Depth + 1) |
| 6179 | : ConstantRange::getFull( |
| 6180 | BitWidth: ExpArg->getType()->getScalarSizeInBits()); |
| 6181 | |
| 6182 | const fltSemantics &Flt = |
| 6183 | II->getType()->getScalarType()->getFltSemantics(); |
| 6184 | |
| 6185 | const Function *F = II->getFunction(); |
| 6186 | DenormalMode Mode = |
| 6187 | F ? F->getDenormalMode(FPType: Flt) : DenormalMode::getDynamic(); |
| 6188 | |
| 6189 | Known = KnownFPClass::ldexp(Src: KnownSrc, ConstantRangeMin: ExpKnownRange.getSignedMin(), |
| 6190 | ConstantRangeMax: ExpKnownRange.getSignedMax(), Flt, Mode); |
| 6191 | break; |
| 6192 | } |
| 6193 | case Intrinsic::arithmetic_fence: { |
| 6194 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
| 6195 | Known, Q, Depth: Depth + 1); |
| 6196 | break; |
| 6197 | } |
| 6198 | case Intrinsic::experimental_constrained_sitofp: |
| 6199 | case Intrinsic::experimental_constrained_uitofp: |
| 6200 | // Cannot produce nan |
| 6201 | Known.knownNot(RuleOut: fcNan); |
| 6202 | |
| 6203 | // sitofp and uitofp turn into +0.0 for zero. |
| 6204 | Known.knownNot(RuleOut: fcNegZero); |
| 6205 | |
| 6206 | // Integers cannot be subnormal |
| 6207 | Known.knownNot(RuleOut: fcSubnormal); |
| 6208 | |
| 6209 | if (IID == Intrinsic::experimental_constrained_uitofp) |
| 6210 | Known.signBitMustBeZero(); |
| 6211 | |
| 6212 | // TODO: Copy inf handling from instructions |
| 6213 | break; |
| 6214 | |
| 6215 | case Intrinsic::amdgcn_fract: { |
| 6216 | Known.knownNot(RuleOut: fcInf); |
| 6217 | |
| 6218 | if (InterestedClasses & fcNan) { |
| 6219 | KnownFPClass KnownSrc; |
| 6220 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, |
| 6221 | InterestedClasses, Known&: KnownSrc, Q, Depth: Depth + 1); |
| 6222 | |
| 6223 | if (KnownSrc.isKnownNeverInfOrNaN()) |
| 6224 | Known.knownNot(RuleOut: fcNan); |
| 6225 | else if (KnownSrc.isKnownNever(Mask: fcSNan)) |
| 6226 | Known.knownNot(RuleOut: fcSNan); |
| 6227 | } |
| 6228 | |
| 6229 | break; |
| 6230 | } |
| 6231 | case Intrinsic::amdgcn_rcp: { |
| 6232 | KnownFPClass KnownSrc; |
| 6233 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
| 6234 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 6235 | |
| 6236 | Known.propagateNonNaN(Src: KnownSrc); |
| 6237 | |
| 6238 | Type *EltTy = II->getType()->getScalarType(); |
| 6239 | |
| 6240 | // f32 denormal always flushed. |
| 6241 | if (EltTy->isFloatTy()) { |
| 6242 | Known.knownNot(RuleOut: fcSubnormal); |
| 6243 | KnownSrc.knownNot(RuleOut: fcSubnormal); |
| 6244 | } |
| 6245 | |
| 6246 | if (KnownSrc.isKnownNever(Mask: fcNegative)) |
| 6247 | Known.knownNot(RuleOut: fcNegative); |
| 6248 | if (KnownSrc.isKnownNever(Mask: fcPositive)) |
| 6249 | Known.knownNot(RuleOut: fcPositive); |
| 6250 | |
| 6251 | if (const Function *F = II->getFunction()) { |
| 6252 | DenormalMode Mode = F->getDenormalMode(FPType: EltTy->getFltSemantics()); |
| 6253 | if (KnownSrc.isKnownNeverLogicalPosZero(Mode)) |
| 6254 | Known.knownNot(RuleOut: fcPosInf); |
| 6255 | if (KnownSrc.isKnownNeverLogicalNegZero(Mode)) |
| 6256 | Known.knownNot(RuleOut: fcNegInf); |
| 6257 | } |
| 6258 | |
| 6259 | break; |
| 6260 | } |
| 6261 | case Intrinsic::amdgcn_rsq: { |
| 6262 | KnownFPClass KnownSrc; |
| 6263 | // The only negative value that can be returned is -inf for -0 inputs. |
| 6264 | Known.knownNot(RuleOut: fcNegZero | fcNegSubnormal | fcNegNormal); |
| 6265 | |
| 6266 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, InterestedClasses, |
| 6267 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 6268 | |
| 6269 | // Negative -> nan |
| 6270 | if (KnownSrc.isKnownNeverNaN() && KnownSrc.cannotBeOrderedLessThanZero()) |
| 6271 | Known.knownNot(RuleOut: fcNan); |
| 6272 | else if (KnownSrc.isKnownNever(Mask: fcSNan)) |
| 6273 | Known.knownNot(RuleOut: fcSNan); |
| 6274 | |
| 6275 | // +inf -> +0 |
| 6276 | if (KnownSrc.isKnownNeverPosInfinity()) |
| 6277 | Known.knownNot(RuleOut: fcPosZero); |
| 6278 | |
| 6279 | Type *EltTy = II->getType()->getScalarType(); |
| 6280 | |
| 6281 | // f32 denormal always flushed. |
| 6282 | if (EltTy->isFloatTy()) |
| 6283 | Known.knownNot(RuleOut: fcPosSubnormal); |
| 6284 | |
| 6285 | if (const Function *F = II->getFunction()) { |
| 6286 | DenormalMode Mode = F->getDenormalMode(FPType: EltTy->getFltSemantics()); |
| 6287 | |
| 6288 | // -0 -> -inf |
| 6289 | if (KnownSrc.isKnownNeverLogicalNegZero(Mode)) |
| 6290 | Known.knownNot(RuleOut: fcNegInf); |
| 6291 | |
| 6292 | // +0 -> +inf |
| 6293 | if (KnownSrc.isKnownNeverLogicalPosZero(Mode)) |
| 6294 | Known.knownNot(RuleOut: fcPosInf); |
| 6295 | } |
| 6296 | |
| 6297 | break; |
| 6298 | } |
| 6299 | case Intrinsic::amdgcn_trig_preop: { |
| 6300 | // Always returns a value [0, 1) |
| 6301 | Known.knownNot(RuleOut: fcNan | fcInf | fcNegative); |
| 6302 | break; |
| 6303 | } |
| 6304 | case Intrinsic::convert_from_arbitrary_fp: { |
| 6305 | auto *MD = cast<MetadataAsValue>(Val: II->getArgOperand(i: 1))->getMetadata(); |
| 6306 | StringRef FormatStr = cast<MDString>(Val: MD)->getString(); |
| 6307 | |
| 6308 | const fltSemantics *SrcSemantics = |
| 6309 | APFloat::getArbitraryFPSemantics(Format: FormatStr); |
| 6310 | if (!SrcSemantics) |
| 6311 | break; |
| 6312 | |
| 6313 | const fltSemantics DstSemantics = |
| 6314 | II->getType()->getScalarType()->getFltSemantics(); |
| 6315 | |
| 6316 | if (!APFloat::semanticsHasNaN(*SrcSemantics)) |
| 6317 | Known.knownNot(RuleOut: fcNan); |
| 6318 | |
| 6319 | // fcInf can only be cleared if the source format has no Inf encoding |
| 6320 | // and the dst max exp can accommodate src max exp. |
| 6321 | if (!APFloat::semanticsHasInf(*SrcSemantics) && |
| 6322 | APFloat::semanticsMaxExponent(*SrcSemantics) <= |
| 6323 | APFloat::semanticsMaxExponent(DstSemantics)) |
| 6324 | Known.knownNot(RuleOut: fcInf); |
| 6325 | |
| 6326 | // Check and clear all neg flags for formats that do not have signed |
| 6327 | // representation. |
| 6328 | if (!APFloat::semanticsHasSignedRepr(*SrcSemantics)) |
| 6329 | Known.knownNot(RuleOut: fcNegative); |
| 6330 | |
| 6331 | // Check if format has no zero at all (Float8E8M0FNU), or no negative |
| 6332 | // zero. |
| 6333 | if (!APFloat::semanticsHasZero(*SrcSemantics)) |
| 6334 | Known.knownNot(RuleOut: fcZero); |
| 6335 | else if (SrcSemantics->nanEncoding == fltNanEncoding::NegativeZero) |
| 6336 | Known.knownNot(RuleOut: fcNegZero); |
| 6337 | |
| 6338 | // If src lands normally in dest, the result can never be subnormal. |
| 6339 | if (APFloat::isRepresentableAsNormalIn(Src: *SrcSemantics, Dst: DstSemantics)) |
| 6340 | Known.knownNot(RuleOut: fcSubnormal); |
| 6341 | break; |
| 6342 | } |
| 6343 | default: |
| 6344 | break; |
| 6345 | } |
| 6346 | |
| 6347 | break; |
| 6348 | } |
| 6349 | case Instruction::FAdd: |
| 6350 | case Instruction::FSub: { |
| 6351 | computeKnownFPClassForFAddSub(Op, DemandedElts, InterestedClasses, Known, Q, |
| 6352 | Depth, /*IsAdd=*/Opc == Instruction::FAdd, |
| 6353 | /*IsStrict=*/false); |
| 6354 | break; |
| 6355 | } |
| 6356 | case Instruction::FMul: { |
| 6357 | const Function *F = cast<Instruction>(Val: Op)->getFunction(); |
| 6358 | DenormalMode Mode = |
| 6359 | F ? F->getDenormalMode( |
| 6360 | FPType: Op->getType()->getScalarType()->getFltSemantics()) |
| 6361 | : DenormalMode::getDynamic(); |
| 6362 | |
| 6363 | Value *LHS = Op->getOperand(i: 0); |
| 6364 | Value *RHS = Op->getOperand(i: 1); |
| 6365 | // X * X is always non-negative or a NaN. |
| 6366 | // FIXME: Should check isGuaranteedNotToBeUndef |
| 6367 | if (LHS == RHS) { |
| 6368 | KnownFPClass KnownSrc; |
| 6369 | computeKnownFPClass(V: LHS, DemandedElts, InterestedClasses: fcAllFlags, Known&: KnownSrc, Q, |
| 6370 | Depth: Depth + 1); |
| 6371 | Known = KnownFPClass::square(Src: KnownSrc, Mode); |
| 6372 | break; |
| 6373 | } |
| 6374 | |
| 6375 | KnownFPClass KnownLHS, KnownRHS; |
| 6376 | |
| 6377 | const APFloat *CRHS; |
| 6378 | if (match(V: RHS, P: m_APFloat(Res&: CRHS))) { |
| 6379 | computeKnownFPClass(V: LHS, DemandedElts, InterestedClasses: fcAllFlags, Known&: KnownLHS, Q, |
| 6380 | Depth: Depth + 1); |
| 6381 | Known = KnownFPClass::fmul(LHS: KnownLHS, RHS: *CRHS, Mode); |
| 6382 | } else { |
| 6383 | computeKnownFPClass(V: RHS, DemandedElts, InterestedClasses: fcAllFlags, Known&: KnownRHS, Q, |
| 6384 | Depth: Depth + 1); |
| 6385 | // TODO: Improve accuracy in unfused FMA pattern. We can prove an |
| 6386 | // additional not-nan if the addend is known-not negative infinity if the |
| 6387 | // multiply is known-not infinity. |
| 6388 | |
| 6389 | computeKnownFPClass(V: LHS, DemandedElts, InterestedClasses: fcAllFlags, Known&: KnownLHS, Q, |
| 6390 | Depth: Depth + 1); |
| 6391 | Known = KnownFPClass::fmul(LHS: KnownLHS, RHS: KnownRHS, Mode); |
| 6392 | } |
| 6393 | |
| 6394 | /// Propgate no-infs if the other source is known smaller than one, such |
| 6395 | /// that this cannot introduce overflow. |
| 6396 | if (KnownLHS.isKnownNever(Mask: fcInf) && isAbsoluteValueULEOne(V: RHS)) |
| 6397 | Known.knownNot(RuleOut: fcInf); |
| 6398 | else if (KnownRHS.isKnownNever(Mask: fcInf) && isAbsoluteValueULEOne(V: LHS)) |
| 6399 | Known.knownNot(RuleOut: fcInf); |
| 6400 | |
| 6401 | break; |
| 6402 | } |
| 6403 | case Instruction::FDiv: { |
| 6404 | const bool WantNan = (InterestedClasses & fcNan) != fcNone; |
| 6405 | |
| 6406 | const Function *F = cast<Instruction>(Val: Op)->getFunction(); |
| 6407 | const fltSemantics &FltSem = |
| 6408 | Op->getType()->getScalarType()->getFltSemantics(); |
| 6409 | DenormalMode Mode = |
| 6410 | F ? F->getDenormalMode(FPType: FltSem) : DenormalMode::getDynamic(); |
| 6411 | |
| 6412 | if (Op->getOperand(i: 0) == Op->getOperand(i: 1) && |
| 6413 | isGuaranteedNotToBeUndef(V: Op->getOperand(i: 0), AC: Q.AC, CtxI: Q.CtxI, DT: Q.DT)) { |
| 6414 | // X / X is always exactly 1.0 or a NaN. |
| 6415 | Known.setKnownFPClasses(fcNan | fcPosNormal); |
| 6416 | |
| 6417 | if (!WantNan) |
| 6418 | break; |
| 6419 | |
| 6420 | KnownFPClass KnownSrc; |
| 6421 | computeKnownFPClass(V: Op->getOperand(i: 0), DemandedElts, |
| 6422 | InterestedClasses: fcNan | fcInf | fcZero | fcSubnormal, Known&: KnownSrc, Q, |
| 6423 | Depth: Depth + 1); |
| 6424 | |
| 6425 | Known = KnownFPClass::fdiv_self(Src: KnownSrc, Mode); |
| 6426 | break; |
| 6427 | } |
| 6428 | |
| 6429 | const bool WantNegative = (InterestedClasses & fcNegative) != fcNone; |
| 6430 | const bool WantPositive = (InterestedClasses & fcPositive) != fcNone; |
| 6431 | if (!WantNan && !WantNegative && !WantPositive) |
| 6432 | break; |
| 6433 | |
| 6434 | KnownFPClass KnownLHS, KnownRHS; |
| 6435 | computeKnownFPClass(V: Op->getOperand(i: 1), DemandedElts, InterestedClasses: fcAllFlags, Known&: KnownRHS, |
| 6436 | Q, Depth: Depth + 1); |
| 6437 | |
| 6438 | bool KnowSomethingUseful = |
| 6439 | KnownRHS.isKnownNever(Mask: fcSNan) || |
| 6440 | KnownRHS.isKnownNever(Mask: fcNegNormal | fcNegSubnormal) || |
| 6441 | KnownRHS.isKnownNever(Mask: fcPosNormal | fcPosSubnormal); |
| 6442 | |
| 6443 | if (KnowSomethingUseful) |
| 6444 | computeKnownFPClass(V: Op->getOperand(i: 0), DemandedElts, InterestedClasses: fcAllFlags, Known&: KnownLHS, |
| 6445 | Q, Depth: Depth + 1); |
| 6446 | |
| 6447 | Known = KnownFPClass::fdiv(LHS: KnownLHS, RHS: KnownRHS, Mode); |
| 6448 | break; |
| 6449 | } |
| 6450 | case Instruction::FRem: { |
| 6451 | computeKnownFPClassForFRem(Op, DemandedElts, InterestedClasses, Known, Q, |
| 6452 | Depth); |
| 6453 | break; |
| 6454 | } |
| 6455 | case Instruction::FPExt: { |
| 6456 | KnownFPClass KnownSrc; |
| 6457 | computeKnownFPClass(V: Op->getOperand(i: 0), DemandedElts, InterestedClasses, |
| 6458 | Known&: KnownSrc, Q, Depth: Depth + 1); |
| 6459 | |
| 6460 | const fltSemantics &DstTy = |
| 6461 | Op->getType()->getScalarType()->getFltSemantics(); |
| 6462 | const fltSemantics &SrcTy = |
| 6463 | Op->getOperand(i: 0)->getType()->getScalarType()->getFltSemantics(); |
| 6464 | |
| 6465 | const auto *I = dyn_cast<Instruction>(Val: Op); |
| 6466 | const Function *F = I ? I->getFunction() : nullptr; |
| 6467 | DenormalMode DstMode = |
| 6468 | F ? F->getDenormalMode(FPType: DstTy) : DenormalMode::getDynamic(); |
| 6469 | DenormalMode SrcMode = |
| 6470 | F ? F->getDenormalMode(FPType: SrcTy) : DenormalMode::getDynamic(); |
| 6471 | |
| 6472 | Known = KnownFPClass::fpext(KnownSrc, DstTy, SrcTy, DstMode, SrcMode); |
| 6473 | break; |
| 6474 | } |
| 6475 | case Instruction::FPTrunc: { |
| 6476 | computeKnownFPClassForFPTrunc(Op, DemandedElts, InterestedClasses, Known, Q, |
| 6477 | Depth); |
| 6478 | break; |
| 6479 | } |
| 6480 | case Instruction::SIToFP: |
| 6481 | case Instruction::UIToFP: { |
| 6482 | // Cannot produce nan |
| 6483 | Known.knownNot(RuleOut: fcNan); |
| 6484 | |
| 6485 | // Integers cannot be subnormal |
| 6486 | Known.knownNot(RuleOut: fcSubnormal); |
| 6487 | |
| 6488 | // sitofp and uitofp turn into +0.0 for zero. |
| 6489 | Known.knownNot(RuleOut: fcNegZero); |
| 6490 | |
| 6491 | // UIToFP is always non-negative regardless of known bits. |
| 6492 | if (Op->getOpcode() == Instruction::UIToFP) |
| 6493 | Known.signBitMustBeZero(); |
| 6494 | |
| 6495 | // Only compute known bits if we can learn something useful from them. |
| 6496 | if (!(InterestedClasses & (fcPosZero | fcNormal | fcInf))) |
| 6497 | break; |
| 6498 | |
| 6499 | KnownBits IntKnown = |
| 6500 | computeKnownBits(V: Op->getOperand(i: 0), DemandedElts, Q, Depth: Depth + 1); |
| 6501 | |
| 6502 | // If the integer is non-zero, the result cannot be +0.0 |
| 6503 | if (IntKnown.isNonZero()) |
| 6504 | Known.knownNot(RuleOut: fcPosZero); |
| 6505 | |
| 6506 | if (Op->getOpcode() == Instruction::SIToFP) { |
| 6507 | // If the signed integer is known non-negative, the result is |
| 6508 | // non-negative. If the signed integer is known negative, the result is |
| 6509 | // negative. |
| 6510 | if (IntKnown.isNonNegative()) { |
| 6511 | Known.signBitMustBeZero(); |
| 6512 | } else if (IntKnown.isNegative()) { |
| 6513 | Known.signBitMustBeOne(); |
| 6514 | } |
| 6515 | } |
| 6516 | |
| 6517 | // Guard kept for ilogb() |
| 6518 | if (InterestedClasses & fcInf) { |
| 6519 | // Get width of largest magnitude integer known. |
| 6520 | // This still works for a signed minimum value because the largest FP |
| 6521 | // value is scaled by some fraction close to 2.0 (1.0 + 0.xxxx). |
| 6522 | int IntSize = IntKnown.getBitWidth(); |
| 6523 | if (Op->getOpcode() == Instruction::UIToFP) |
| 6524 | IntSize -= IntKnown.countMinLeadingZeros(); |
| 6525 | else if (Op->getOpcode() == Instruction::SIToFP) |
| 6526 | IntSize -= IntKnown.countMinSignBits(); |
| 6527 | |
| 6528 | // If the exponent of the largest finite FP value can hold the largest |
| 6529 | // integer, the result of the cast must be finite. |
| 6530 | Type *FPTy = Op->getType()->getScalarType(); |
| 6531 | if (ilogb(Arg: APFloat::getLargest(Sem: FPTy->getFltSemantics())) >= IntSize) |
| 6532 | Known.knownNot(RuleOut: fcInf); |
| 6533 | } |
| 6534 | |
| 6535 | break; |
| 6536 | } |
| 6537 | case Instruction::ExtractElement: { |
| 6538 | // Look through extract element. If the index is non-constant or |
| 6539 | // out-of-range demand all elements, otherwise just the extracted element. |
| 6540 | const Value *Vec = Op->getOperand(i: 0); |
| 6541 | |
| 6542 | APInt DemandedVecElts; |
| 6543 | if (auto *VecTy = dyn_cast<FixedVectorType>(Val: Vec->getType())) { |
| 6544 | unsigned NumElts = VecTy->getNumElements(); |
| 6545 | DemandedVecElts = APInt::getAllOnes(numBits: NumElts); |
| 6546 | auto *CIdx = dyn_cast<ConstantInt>(Val: Op->getOperand(i: 1)); |
| 6547 | if (CIdx && CIdx->getValue().ult(RHS: NumElts)) |
| 6548 | DemandedVecElts = APInt::getOneBitSet(numBits: NumElts, BitNo: CIdx->getZExtValue()); |
| 6549 | } else { |
| 6550 | DemandedVecElts = APInt(1, 1); |
| 6551 | } |
| 6552 | |
| 6553 | return computeKnownFPClass(V: Vec, DemandedElts: DemandedVecElts, InterestedClasses, Known, |
| 6554 | Q, Depth: Depth + 1); |
| 6555 | } |
| 6556 | case Instruction::InsertElement: { |
| 6557 | if (isa<ScalableVectorType>(Val: Op->getType())) |
| 6558 | return; |
| 6559 | |
| 6560 | const Value *Vec = Op->getOperand(i: 0); |
| 6561 | const Value *Elt = Op->getOperand(i: 1); |
| 6562 | auto *CIdx = dyn_cast<ConstantInt>(Val: Op->getOperand(i: 2)); |
| 6563 | unsigned NumElts = DemandedElts.getBitWidth(); |
| 6564 | APInt DemandedVecElts = DemandedElts; |
| 6565 | bool NeedsElt = true; |
| 6566 | // If we know the index we are inserting to, clear it from Vec check. |
| 6567 | if (CIdx && CIdx->getValue().ult(RHS: NumElts)) { |
| 6568 | DemandedVecElts.clearBit(BitPosition: CIdx->getZExtValue()); |
| 6569 | NeedsElt = DemandedElts[CIdx->getZExtValue()]; |
| 6570 | } |
| 6571 | |
| 6572 | // Do we demand the inserted element? |
| 6573 | if (NeedsElt) { |
| 6574 | computeKnownFPClass(V: Elt, Known, InterestedClasses, Q, Depth: Depth + 1); |
| 6575 | // If we don't know any bits, early out. |
| 6576 | if (Known.isUnknown()) |
| 6577 | break; |
| 6578 | } else { |
| 6579 | Known = KnownFPClass::getEmpty(); |
| 6580 | } |
| 6581 | |
| 6582 | // Do we need anymore elements from Vec? |
| 6583 | if (!DemandedVecElts.isZero()) { |
| 6584 | KnownFPClass Known2; |
| 6585 | computeKnownFPClass(V: Vec, DemandedElts: DemandedVecElts, InterestedClasses, Known&: Known2, Q, |
| 6586 | Depth: Depth + 1); |
| 6587 | Known |= Known2; |
| 6588 | } |
| 6589 | |
| 6590 | break; |
| 6591 | } |
| 6592 | case Instruction::ShuffleVector: { |
| 6593 | // Handle vector splat idiom |
| 6594 | if (Value *Splat = getSplatValue(V)) { |
| 6595 | computeKnownFPClass(V: Splat, Known, InterestedClasses, Q, Depth: Depth + 1); |
| 6596 | break; |
| 6597 | } |
| 6598 | |
| 6599 | // For undef elements, we don't know anything about the common state of |
| 6600 | // the shuffle result. |
| 6601 | APInt DemandedLHS, DemandedRHS; |
| 6602 | auto *Shuf = dyn_cast<ShuffleVectorInst>(Val: Op); |
| 6603 | if (!Shuf || !getShuffleDemandedElts(Shuf, DemandedElts, DemandedLHS, DemandedRHS)) |
| 6604 | return; |
| 6605 | |
| 6606 | if (!!DemandedLHS) { |
| 6607 | const Value *LHS = Shuf->getOperand(i_nocapture: 0); |
| 6608 | computeKnownFPClass(V: LHS, DemandedElts: DemandedLHS, InterestedClasses, Known, Q, |
| 6609 | Depth: Depth + 1); |
| 6610 | |
| 6611 | // If we don't know any bits, early out. |
| 6612 | if (Known.isUnknown()) |
| 6613 | break; |
| 6614 | } else { |
| 6615 | Known = KnownFPClass::getEmpty(); |
| 6616 | } |
| 6617 | |
| 6618 | if (!!DemandedRHS) { |
| 6619 | KnownFPClass Known2; |
| 6620 | const Value *RHS = Shuf->getOperand(i_nocapture: 1); |
| 6621 | computeKnownFPClass(V: RHS, DemandedElts: DemandedRHS, InterestedClasses, Known&: Known2, Q, |
| 6622 | Depth: Depth + 1); |
| 6623 | Known |= Known2; |
| 6624 | } |
| 6625 | |
| 6626 | break; |
| 6627 | } |
| 6628 | case Instruction::ExtractValue: { |
| 6629 | const ExtractValueInst * = cast<ExtractValueInst>(Val: Op); |
| 6630 | ArrayRef<unsigned> Indices = Extract->getIndices(); |
| 6631 | const Value *Src = Extract->getAggregateOperand(); |
| 6632 | if (isa<StructType>(Val: Src->getType()) && Indices.size() == 1 && |
| 6633 | Indices[0] == 0) { |
| 6634 | if (const auto *II = dyn_cast<IntrinsicInst>(Val: Src)) { |
| 6635 | switch (II->getIntrinsicID()) { |
| 6636 | case Intrinsic::frexp: { |
| 6637 | FPClassTest InterestedSrcs = InterestedClasses; |
| 6638 | |
| 6639 | // Positive subnormals and negative subnormals could become positive |
| 6640 | // zero. |
| 6641 | if (InterestedClasses & fcPosZero) |
| 6642 | InterestedSrcs |= fcSubnormal; |
| 6643 | |
| 6644 | // Negative subnormals could become negative zero. |
| 6645 | if (InterestedClasses & fcNegZero) |
| 6646 | InterestedSrcs |= fcNegSubnormal; |
| 6647 | |
| 6648 | if (InterestedClasses & fcPosNormal) |
| 6649 | InterestedSrcs |= fcPosSubnormal; |
| 6650 | |
| 6651 | if (InterestedClasses & fcNegNormal) |
| 6652 | InterestedSrcs |= fcNegSubnormal; |
| 6653 | |
| 6654 | KnownFPClass KnownSrc; |
| 6655 | computeKnownFPClass(V: II->getArgOperand(i: 0), DemandedElts, |
| 6656 | InterestedClasses: InterestedSrcs, Known&: KnownSrc, Q, Depth: Depth + 1); |
| 6657 | |
| 6658 | const Function *F = cast<Instruction>(Val: Op)->getFunction(); |
| 6659 | const fltSemantics &FltSem = |
| 6660 | Op->getType()->getScalarType()->getFltSemantics(); |
| 6661 | |
| 6662 | DenormalMode Mode = |
| 6663 | F ? F->getDenormalMode(FPType: FltSem) : DenormalMode::getDynamic(); |
| 6664 | Known = KnownFPClass::frexp_mant(Src: KnownSrc, Mode); |
| 6665 | return; |
| 6666 | } |
| 6667 | default: |
| 6668 | break; |
| 6669 | } |
| 6670 | } |
| 6671 | } |
| 6672 | |
| 6673 | computeKnownFPClass(V: Src, DemandedElts, InterestedClasses, Known, Q, |
| 6674 | Depth: Depth + 1); |
| 6675 | break; |
| 6676 | } |
| 6677 | case Instruction::PHI: { |
| 6678 | const PHINode *P = cast<PHINode>(Val: Op); |
| 6679 | // Unreachable blocks may have zero-operand PHI nodes. |
| 6680 | if (P->getNumIncomingValues() == 0) |
| 6681 | break; |
| 6682 | |
| 6683 | // Otherwise take the unions of the known bit sets of the operands, |
| 6684 | // taking conservative care to avoid excessive recursion. |
| 6685 | const unsigned PhiRecursionLimit = MaxAnalysisRecursionDepth - 2; |
| 6686 | |
| 6687 | if (Depth < PhiRecursionLimit) { |
| 6688 | // Skip if every incoming value references to ourself. |
| 6689 | if (isa_and_nonnull<UndefValue>(Val: P->hasConstantValue())) |
| 6690 | break; |
| 6691 | |
| 6692 | bool First = true; |
| 6693 | |
| 6694 | for (const Use &U : P->operands()) { |
| 6695 | Value *IncValue; |
| 6696 | Instruction *CtxI; |
| 6697 | breakSelfRecursivePHI(U: &U, PHI: P, ValOut&: IncValue, CtxIOut&: CtxI); |
| 6698 | // Skip direct self references. |
| 6699 | if (IncValue == P) |
| 6700 | continue; |
| 6701 | |
| 6702 | KnownFPClass KnownSrc; |
| 6703 | // Recurse, but cap the recursion to two levels, because we don't want |
| 6704 | // to waste time spinning around in loops. We need at least depth 2 to |
| 6705 | // detect known sign bits. |
| 6706 | computeKnownFPClass(V: IncValue, DemandedElts, InterestedClasses, Known&: KnownSrc, |
| 6707 | Q: Q.getWithoutCondContext().getWithInstruction(I: CtxI), |
| 6708 | Depth: PhiRecursionLimit); |
| 6709 | |
| 6710 | if (First) { |
| 6711 | Known = KnownSrc; |
| 6712 | First = false; |
| 6713 | } else { |
| 6714 | Known |= KnownSrc; |
| 6715 | } |
| 6716 | |
| 6717 | if (Known.getKnownFPClasses() == fcAllFlags) |
| 6718 | break; |
| 6719 | } |
| 6720 | } |
| 6721 | |
| 6722 | // Look for the case of a for loop which has a positive |
| 6723 | // initial value and is incremented by a squared value. |
| 6724 | // This will propagate sign information out of such loops. |
| 6725 | if (P->getNumIncomingValues() != 2 || Known.cannotBeOrderedLessThanZero()) |
| 6726 | break; |
| 6727 | for (unsigned I = 0; I < 2; I++) { |
| 6728 | Value *RecurValue = P->getIncomingValue(i: 1 - I); |
| 6729 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: RecurValue); |
| 6730 | if (!II) |
| 6731 | continue; |
| 6732 | Value *R, *L, *Init; |
| 6733 | PHINode *PN; |
| 6734 | if (matchSimpleTernaryIntrinsicRecurrence(I: II, P&: PN, Init, OtherOp0&: L, OtherOp1&: R) && |
| 6735 | PN == P) { |
| 6736 | switch (II->getIntrinsicID()) { |
| 6737 | case Intrinsic::fma: |
| 6738 | case Intrinsic::fmuladd: { |
| 6739 | KnownFPClass KnownStart; |
| 6740 | computeKnownFPClass(V: Init, DemandedElts, InterestedClasses, Known&: KnownStart, |
| 6741 | Q, Depth: Depth + 1); |
| 6742 | if (KnownStart.cannotBeOrderedLessThanZero() && L == R && |
| 6743 | isGuaranteedNotToBeUndef(V: L, AC: Q.AC, CtxI: Q.CtxI, DT: Q.DT, Depth: Depth + 1)) |
| 6744 | Known.knownNot(RuleOut: KnownFPClass::OrderedLessThanZeroMask); |
| 6745 | break; |
| 6746 | } |
| 6747 | } |
| 6748 | } |
| 6749 | } |
| 6750 | break; |
| 6751 | } |
| 6752 | case Instruction::BitCast: { |
| 6753 | const Value *Src; |
| 6754 | if (!match(V: Op, P: m_ElementWiseBitCast(Op: m_Value(V&: Src))) || |
| 6755 | !Src->getType()->isIntOrIntVectorTy()) |
| 6756 | break; |
| 6757 | |
| 6758 | const Type *Ty = Op->getType(); |
| 6759 | |
| 6760 | Value *CastLHS, *CastRHS; |
| 6761 | |
| 6762 | // Match bitcast(umax(bitcast(a), bitcast(b))) |
| 6763 | if (match(V: Src, P: m_c_MaxOrMin(L: m_BitCast(Op: m_Value(V&: CastLHS)), |
| 6764 | R: m_BitCast(Op: m_Value(V&: CastRHS)))) && |
| 6765 | CastLHS->getType() == Ty && CastRHS->getType() == Ty) { |
| 6766 | KnownFPClass KnownLHS, KnownRHS; |
| 6767 | computeKnownFPClass(V: CastRHS, DemandedElts, InterestedClasses, Known&: KnownRHS, Q, |
| 6768 | Depth: Depth + 1); |
| 6769 | if (!KnownRHS.isUnknown()) { |
| 6770 | computeKnownFPClass(V: CastLHS, DemandedElts, InterestedClasses, Known&: KnownLHS, |
| 6771 | Q, Depth: Depth + 1); |
| 6772 | Known = KnownLHS | KnownRHS; |
| 6773 | } |
| 6774 | |
| 6775 | return; |
| 6776 | } |
| 6777 | |
| 6778 | const Type *EltTy = Ty->getScalarType(); |
| 6779 | KnownBits Bits(EltTy->getPrimitiveSizeInBits()); |
| 6780 | computeKnownBits(V: Src, DemandedElts, Known&: Bits, Q, Depth: Depth + 1); |
| 6781 | |
| 6782 | Known = KnownFPClass::bitcast(FltSemantics: EltTy->getFltSemantics(), Bits); |
| 6783 | break; |
| 6784 | } |
| 6785 | default: |
| 6786 | break; |
| 6787 | } |
| 6788 | } |
| 6789 | |
| 6790 | KnownFPClass llvm::computeKnownFPClass(const Value *V, |
| 6791 | const APInt &DemandedElts, |
| 6792 | FPClassTest InterestedClasses, |
| 6793 | const SimplifyQuery &SQ, |
| 6794 | unsigned Depth) { |
| 6795 | KnownFPClass KnownClasses; |
| 6796 | ::computeKnownFPClass(V, DemandedElts, InterestedClasses, Known&: KnownClasses, Q: SQ, |
| 6797 | Depth); |
| 6798 | return KnownClasses; |
| 6799 | } |
| 6800 | |
| 6801 | KnownFPClass llvm::computeKnownFPClass(const Value *V, |
| 6802 | FPClassTest InterestedClasses, |
| 6803 | const SimplifyQuery &SQ, |
| 6804 | unsigned Depth) { |
| 6805 | KnownFPClass Known; |
| 6806 | ::computeKnownFPClass(V, Known, InterestedClasses, Q: SQ, Depth); |
| 6807 | return Known; |
| 6808 | } |
| 6809 | |
| 6810 | KnownFPClass llvm::computeKnownFPClass( |
| 6811 | const Value *V, const DataLayout &DL, FPClassTest InterestedClasses, |
| 6812 | const TargetLibraryInfo *TLI, AssumptionCache *AC, const Instruction *CtxI, |
| 6813 | const DominatorTree *DT, bool UseInstrInfo, unsigned Depth) { |
| 6814 | return computeKnownFPClass(V, InterestedClasses, |
| 6815 | SQ: SimplifyQuery(DL, TLI, DT, AC, CtxI, UseInstrInfo), |
| 6816 | Depth); |
| 6817 | } |
| 6818 | |
| 6819 | KnownFPClass |
| 6820 | llvm::computeKnownFPClass(const Value *V, const APInt &DemandedElts, |
| 6821 | FastMathFlags FMF, FPClassTest InterestedClasses, |
| 6822 | const SimplifyQuery &SQ, unsigned Depth) { |
| 6823 | if (FMF.noNaNs()) |
| 6824 | InterestedClasses &= ~fcNan; |
| 6825 | if (FMF.noInfs()) |
| 6826 | InterestedClasses &= ~fcInf; |
| 6827 | |
| 6828 | KnownFPClass Result = |
| 6829 | computeKnownFPClass(V, DemandedElts, InterestedClasses, SQ, Depth); |
| 6830 | |
| 6831 | if (FMF.noNaNs()) |
| 6832 | Result.setKnownFPClasses(Result.getKnownFPClasses() & ~fcNan); |
| 6833 | if (FMF.noInfs()) |
| 6834 | Result.setKnownFPClasses(Result.getKnownFPClasses() & ~fcInf); |
| 6835 | return Result; |
| 6836 | } |
| 6837 | |
| 6838 | KnownFPClass llvm::computeKnownFPClass(const Value *V, FastMathFlags FMF, |
| 6839 | FPClassTest InterestedClasses, |
| 6840 | const SimplifyQuery &SQ, |
| 6841 | unsigned Depth) { |
| 6842 | auto *FVTy = dyn_cast<FixedVectorType>(Val: V->getType()); |
| 6843 | APInt DemandedElts = |
| 6844 | FVTy ? APInt::getAllOnes(numBits: FVTy->getNumElements()) : APInt(1, 1); |
| 6845 | return computeKnownFPClass(V, DemandedElts, FMF, InterestedClasses, SQ, |
| 6846 | Depth); |
| 6847 | } |
| 6848 | |
| 6849 | bool llvm::cannotBeNegativeZero(const Value *V, const SimplifyQuery &SQ, |
| 6850 | unsigned Depth) { |
| 6851 | KnownFPClass Known = computeKnownFPClass(V, InterestedClasses: fcNegZero, SQ, Depth); |
| 6852 | return Known.isKnownNever(Mask: fcNegZero); |
| 6853 | } |
| 6854 | |
| 6855 | bool llvm::cannotBeOrderedLessThanZero(const Value *V, const SimplifyQuery &SQ, |
| 6856 | unsigned Depth) { |
| 6857 | KnownFPClass Known = |
| 6858 | computeKnownFPClass(V, InterestedClasses: KnownFPClass::OrderedLessThanZeroMask, SQ, Depth); |
| 6859 | return Known.cannotBeOrderedLessThanZero(); |
| 6860 | } |
| 6861 | |
| 6862 | bool llvm::isKnownNeverInfinity(const Value *V, const SimplifyQuery &SQ, |
| 6863 | unsigned Depth) { |
| 6864 | KnownFPClass Known = computeKnownFPClass(V, InterestedClasses: fcInf, SQ, Depth); |
| 6865 | return Known.isKnownNeverInfinity(); |
| 6866 | } |
| 6867 | |
| 6868 | /// Return true if the floating-point value can never contain a NaN or infinity. |
| 6869 | bool llvm::isKnownNeverInfOrNaN(const Value *V, const SimplifyQuery &SQ, |
| 6870 | unsigned Depth) { |
| 6871 | KnownFPClass Known = computeKnownFPClass(V, InterestedClasses: fcInf | fcNan, SQ, Depth); |
| 6872 | return Known.isKnownNeverNaN() && Known.isKnownNeverInfinity(); |
| 6873 | } |
| 6874 | |
| 6875 | /// Return true if the floating-point scalar value is not a NaN or if the |
| 6876 | /// floating-point vector value has no NaN elements. Return false if a value |
| 6877 | /// could ever be NaN. |
| 6878 | bool llvm::isKnownNeverNaN(const Value *V, const SimplifyQuery &SQ, |
| 6879 | unsigned Depth) { |
| 6880 | KnownFPClass Known = computeKnownFPClass(V, InterestedClasses: fcNan, SQ, Depth); |
| 6881 | return Known.isKnownNeverNaN(); |
| 6882 | } |
| 6883 | |
| 6884 | /// Return false if we can prove that the specified FP value's sign bit is 0. |
| 6885 | /// Return true if we can prove that the specified FP value's sign bit is 1. |
| 6886 | /// Otherwise return std::nullopt. |
| 6887 | std::optional<bool> llvm::computeKnownFPSignBit(const Value *V, |
| 6888 | const SimplifyQuery &SQ, |
| 6889 | unsigned Depth) { |
| 6890 | KnownFPClass Known = computeKnownFPClass(V, InterestedClasses: fcAllFlags, SQ, Depth); |
| 6891 | return Known.getSignBit(); |
| 6892 | } |
| 6893 | |
| 6894 | bool llvm::canIgnoreSignBitOfZero(const Use &U) { |
| 6895 | auto *User = cast<Instruction>(Val: U.getUser()); |
| 6896 | if (auto *FPOp = dyn_cast<FPMathOperator>(Val: User)) { |
| 6897 | if (FPOp->hasNoSignedZeros()) |
| 6898 | return true; |
| 6899 | } |
| 6900 | |
| 6901 | switch (User->getOpcode()) { |
| 6902 | case Instruction::FPToSI: |
| 6903 | case Instruction::FPToUI: |
| 6904 | return true; |
| 6905 | case Instruction::FRem: |
| 6906 | // frem(x, +-0.0) == NaN, and more generally frem(x, y) == frem(x, -y). |
| 6907 | return U.getOperandNo() == 1; |
| 6908 | case Instruction::FCmp: |
| 6909 | // fcmp treats both positive and negative zero as equal. |
| 6910 | return true; |
| 6911 | case Instruction::Call: |
| 6912 | if (auto *II = dyn_cast<IntrinsicInst>(Val: User)) { |
| 6913 | switch (II->getIntrinsicID()) { |
| 6914 | case Intrinsic::fabs: |
| 6915 | case Intrinsic::exp: |
| 6916 | case Intrinsic::exp2: |
| 6917 | case Intrinsic::exp10: |
| 6918 | case Intrinsic::log: |
| 6919 | case Intrinsic::log2: |
| 6920 | case Intrinsic::log10: |
| 6921 | case Intrinsic::cos: |
| 6922 | case Intrinsic::acos: |
| 6923 | case Intrinsic::cosh: |
| 6924 | case Intrinsic::fptosi_sat: |
| 6925 | case Intrinsic::fptoui_sat: |
| 6926 | case Intrinsic::lround: |
| 6927 | case Intrinsic::llround: |
| 6928 | case Intrinsic::lrint: |
| 6929 | case Intrinsic::llrint: |
| 6930 | return true; |
| 6931 | case Intrinsic::copysign: |
| 6932 | return U.getOperandNo() == 0; |
| 6933 | case Intrinsic::pow: |
| 6934 | // pow(x, +-0.0) == +1.0 or NaN |
| 6935 | return U.getOperandNo() == 1; |
| 6936 | case Intrinsic::is_fpclass: { |
| 6937 | auto Test = |
| 6938 | static_cast<FPClassTest>( |
| 6939 | cast<ConstantInt>(Val: II->getArgOperand(i: 1))->getZExtValue()) & |
| 6940 | FPClassTest::fcZero; |
| 6941 | return Test == FPClassTest::fcZero || Test == FPClassTest::fcNone; |
| 6942 | } |
| 6943 | default: |
| 6944 | return false; |
| 6945 | } |
| 6946 | } |
| 6947 | return false; |
| 6948 | default: |
| 6949 | return false; |
| 6950 | } |
| 6951 | } |
| 6952 | |
| 6953 | bool llvm::canIgnoreSignBitOfNaN(const Use &U) { |
| 6954 | auto *User = cast<Instruction>(Val: U.getUser()); |
| 6955 | if (auto *FPOp = dyn_cast<FPMathOperator>(Val: User)) { |
| 6956 | if (FPOp->hasNoNaNs()) |
| 6957 | return true; |
| 6958 | } |
| 6959 | |
| 6960 | switch (User->getOpcode()) { |
| 6961 | case Instruction::FPToSI: |
| 6962 | case Instruction::FPToUI: |
| 6963 | return true; |
| 6964 | // Proper FP math operations ignore the sign bit of NaN. |
| 6965 | case Instruction::FAdd: |
| 6966 | case Instruction::FSub: |
| 6967 | case Instruction::FMul: |
| 6968 | case Instruction::FDiv: |
| 6969 | case Instruction::FRem: |
| 6970 | case Instruction::FPTrunc: |
| 6971 | case Instruction::FPExt: |
| 6972 | case Instruction::FCmp: |
| 6973 | return true; |
| 6974 | // Bitwise FP operations should preserve the sign bit of NaN. |
| 6975 | case Instruction::FNeg: |
| 6976 | case Instruction::Select: |
| 6977 | case Instruction::PHI: |
| 6978 | return false; |
| 6979 | case Instruction::Ret: |
| 6980 | return User->getFunction()->getAttributes().getRetNoFPClass() & |
| 6981 | FPClassTest::fcNan; |
| 6982 | case Instruction::Call: |
| 6983 | case Instruction::Invoke: { |
| 6984 | if (auto *II = dyn_cast<IntrinsicInst>(Val: User)) { |
| 6985 | switch (II->getIntrinsicID()) { |
| 6986 | case Intrinsic::fabs: |
| 6987 | return true; |
| 6988 | case Intrinsic::copysign: |
| 6989 | return U.getOperandNo() == 0; |
| 6990 | // Other proper FP math intrinsics ignore the sign bit of NaN. |
| 6991 | case Intrinsic::maxnum: |
| 6992 | case Intrinsic::minnum: |
| 6993 | case Intrinsic::maximum: |
| 6994 | case Intrinsic::minimum: |
| 6995 | case Intrinsic::maximumnum: |
| 6996 | case Intrinsic::minimumnum: |
| 6997 | case Intrinsic::canonicalize: |
| 6998 | case Intrinsic::fma: |
| 6999 | case Intrinsic::fmuladd: |
| 7000 | case Intrinsic::sqrt: |
| 7001 | case Intrinsic::pow: |
| 7002 | case Intrinsic::powi: |
| 7003 | case Intrinsic::fptoui_sat: |
| 7004 | case Intrinsic::fptosi_sat: |
| 7005 | case Intrinsic::is_fpclass: |
| 7006 | return true; |
| 7007 | default: |
| 7008 | return false; |
| 7009 | } |
| 7010 | } |
| 7011 | |
| 7012 | FPClassTest NoFPClass = |
| 7013 | cast<CallBase>(Val: User)->getParamNoFPClass(i: U.getOperandNo()); |
| 7014 | return NoFPClass & FPClassTest::fcNan; |
| 7015 | } |
| 7016 | default: |
| 7017 | return false; |
| 7018 | } |
| 7019 | } |
| 7020 | |
| 7021 | bool llvm::isKnownIntegral(const Value *V, const SimplifyQuery &SQ, |
| 7022 | FastMathFlags FMF) { |
| 7023 | if (isa<PoisonValue>(Val: V)) |
| 7024 | return true; |
| 7025 | if (isa<UndefValue>(Val: V)) |
| 7026 | return false; |
| 7027 | |
| 7028 | if (match(V, P: m_CheckedFp(CheckFn: [](const APFloat &Val) { return Val.isInteger(); }))) |
| 7029 | return true; |
| 7030 | |
| 7031 | const Instruction *I = dyn_cast<Instruction>(Val: V); |
| 7032 | if (!I) |
| 7033 | return false; |
| 7034 | |
| 7035 | switch (I->getOpcode()) { |
| 7036 | case Instruction::SIToFP: |
| 7037 | case Instruction::UIToFP: |
| 7038 | // TODO: Could check nofpclass(inf) on incoming argument |
| 7039 | if (FMF.noInfs()) |
| 7040 | return true; |
| 7041 | |
| 7042 | // Need to check int size cannot produce infinity, which computeKnownFPClass |
| 7043 | // knows how to do already. |
| 7044 | return isKnownNeverInfinity(V: I, SQ); |
| 7045 | case Instruction::Call: { |
| 7046 | const CallInst *CI = cast<CallInst>(Val: I); |
| 7047 | switch (CI->getIntrinsicID()) { |
| 7048 | case Intrinsic::trunc: |
| 7049 | case Intrinsic::floor: |
| 7050 | case Intrinsic::ceil: |
| 7051 | case Intrinsic::rint: |
| 7052 | case Intrinsic::nearbyint: |
| 7053 | case Intrinsic::round: |
| 7054 | case Intrinsic::roundeven: |
| 7055 | return (FMF.noInfs() && FMF.noNaNs()) || isKnownNeverInfOrNaN(V: I, SQ); |
| 7056 | default: |
| 7057 | break; |
| 7058 | } |
| 7059 | |
| 7060 | break; |
| 7061 | } |
| 7062 | default: |
| 7063 | break; |
| 7064 | } |
| 7065 | |
| 7066 | return false; |
| 7067 | } |
| 7068 | |
| 7069 | Value *llvm::isBytewiseValue(Value *V, const DataLayout &DL) { |
| 7070 | |
| 7071 | // All byte-wide stores are splatable, even of arbitrary variables. |
| 7072 | if (V->getType()->isIntegerTy(BitWidth: 8)) |
| 7073 | return V; |
| 7074 | |
| 7075 | LLVMContext &Ctx = V->getContext(); |
| 7076 | |
| 7077 | // Undef don't care. |
| 7078 | auto *UndefInt8 = UndefValue::get(T: Type::getInt8Ty(C&: Ctx)); |
| 7079 | if (isa<UndefValue>(Val: V)) |
| 7080 | return UndefInt8; |
| 7081 | |
| 7082 | // Return poison for zero-sized type. |
| 7083 | if (DL.getTypeStoreSize(Ty: V->getType()).isZero()) |
| 7084 | return PoisonValue::get(T: Type::getInt8Ty(C&: Ctx)); |
| 7085 | |
| 7086 | Constant *C = dyn_cast<Constant>(Val: V); |
| 7087 | if (!C) { |
| 7088 | // Conceptually, we could handle things like: |
| 7089 | // %a = zext i8 %X to i16 |
| 7090 | // %b = shl i16 %a, 8 |
| 7091 | // %c = or i16 %a, %b |
| 7092 | // but until there is an example that actually needs this, it doesn't seem |
| 7093 | // worth worrying about. |
| 7094 | return nullptr; |
| 7095 | } |
| 7096 | |
| 7097 | // Handle 'null' ConstantArrayZero etc. |
| 7098 | if (C->isNullValue()) |
| 7099 | return Constant::getNullValue(Ty: Type::getInt8Ty(C&: Ctx)); |
| 7100 | |
| 7101 | // Constant floating-point values can be handled as integer values if the |
| 7102 | // corresponding integer value is "byteable". An important case is 0.0. |
| 7103 | if (ConstantFP *CFP = dyn_cast<ConstantFP>(Val: C)) { |
| 7104 | Type *ScalarTy = CFP->getType()->getScalarType(); |
| 7105 | if (ScalarTy->isHalfTy() || ScalarTy->isFloatTy() || ScalarTy->isDoubleTy()) |
| 7106 | return isBytewiseValue( |
| 7107 | V: ConstantInt::get(Context&: Ctx, V: CFP->getValue().bitcastToAPInt()), DL); |
| 7108 | |
| 7109 | // Don't handle long double formats, which have strange constraints. |
| 7110 | return nullptr; |
| 7111 | } |
| 7112 | |
| 7113 | // We can handle constant integers that are multiple of 8 bits. |
| 7114 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: C)) { |
| 7115 | if (CI->getBitWidth() % 8 == 0) { |
| 7116 | if (!CI->getValue().isSplat(SplatSizeInBits: 8)) |
| 7117 | return nullptr; |
| 7118 | return ConstantInt::get(Context&: Ctx, V: CI->getValue().trunc(width: 8)); |
| 7119 | } |
| 7120 | } |
| 7121 | |
| 7122 | if (auto *CE = dyn_cast<ConstantExpr>(Val: C)) { |
| 7123 | if (CE->getOpcode() == Instruction::IntToPtr) { |
| 7124 | if (auto *PtrTy = dyn_cast<PointerType>(Val: CE->getType())) { |
| 7125 | unsigned BitWidth = DL.getPointerSizeInBits(AS: PtrTy->getAddressSpace()); |
| 7126 | if (Constant *Op = ConstantFoldIntegerCast( |
| 7127 | C: CE->getOperand(i_nocapture: 0), DestTy: Type::getIntNTy(C&: Ctx, N: BitWidth), IsSigned: false, DL)) |
| 7128 | return isBytewiseValue(V: Op, DL); |
| 7129 | } |
| 7130 | } |
| 7131 | } |
| 7132 | |
| 7133 | auto Merge = [&](Value *LHS, Value *RHS) -> Value * { |
| 7134 | if (LHS == RHS) |
| 7135 | return LHS; |
| 7136 | if (!LHS || !RHS) |
| 7137 | return nullptr; |
| 7138 | if (LHS == UndefInt8) |
| 7139 | return RHS; |
| 7140 | if (RHS == UndefInt8) |
| 7141 | return LHS; |
| 7142 | return nullptr; |
| 7143 | }; |
| 7144 | |
| 7145 | if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(Val: C)) { |
| 7146 | Value *Val = UndefInt8; |
| 7147 | for (uint64_t I = 0, E = CA->getNumElements(); I != E; ++I) |
| 7148 | if (!(Val = Merge(Val, isBytewiseValue(V: CA->getElementAsConstant(i: I), DL)))) |
| 7149 | return nullptr; |
| 7150 | return Val; |
| 7151 | } |
| 7152 | |
| 7153 | if (isa<ConstantAggregate>(Val: C)) { |
| 7154 | Value *Val = UndefInt8; |
| 7155 | for (Value *Op : C->operands()) |
| 7156 | if (!(Val = Merge(Val, isBytewiseValue(V: Op, DL)))) |
| 7157 | return nullptr; |
| 7158 | return Val; |
| 7159 | } |
| 7160 | |
| 7161 | // Don't try to handle the handful of other constants. |
| 7162 | return nullptr; |
| 7163 | } |
| 7164 | |
| 7165 | // This is the recursive version of BuildSubAggregate. It takes a few different |
| 7166 | // arguments. Idxs is the index within the nested struct From that we are |
| 7167 | // looking at now (which is of type IndexedType). IdxSkip is the number of |
| 7168 | // indices from Idxs that should be left out when inserting into the resulting |
| 7169 | // struct. To is the result struct built so far, new insertvalue instructions |
| 7170 | // build on that. |
| 7171 | static Value *BuildSubAggregate(Value *From, Value *To, Type *IndexedType, |
| 7172 | SmallVectorImpl<unsigned> &Idxs, |
| 7173 | unsigned IdxSkip, |
| 7174 | BasicBlock::iterator InsertBefore) { |
| 7175 | StructType *STy = dyn_cast<StructType>(Val: IndexedType); |
| 7176 | if (STy) { |
| 7177 | // Save the original To argument so we can modify it |
| 7178 | Value *OrigTo = To; |
| 7179 | // General case, the type indexed by Idxs is a struct |
| 7180 | for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { |
| 7181 | // Process each struct element recursively |
| 7182 | Idxs.push_back(Elt: i); |
| 7183 | Value *PrevTo = To; |
| 7184 | To = BuildSubAggregate(From, To, IndexedType: STy->getElementType(N: i), Idxs, IdxSkip, |
| 7185 | InsertBefore); |
| 7186 | Idxs.pop_back(); |
| 7187 | if (!To) { |
| 7188 | // Couldn't find any inserted value for this index? Cleanup |
| 7189 | while (PrevTo != OrigTo) { |
| 7190 | InsertValueInst* Del = cast<InsertValueInst>(Val: PrevTo); |
| 7191 | PrevTo = Del->getAggregateOperand(); |
| 7192 | Del->eraseFromParent(); |
| 7193 | } |
| 7194 | // Stop processing elements |
| 7195 | break; |
| 7196 | } |
| 7197 | } |
| 7198 | // If we successfully found a value for each of our subaggregates |
| 7199 | if (To) |
| 7200 | return To; |
| 7201 | } |
| 7202 | // Base case, the type indexed by SourceIdxs is not a struct, or not all of |
| 7203 | // the struct's elements had a value that was inserted directly. In the latter |
| 7204 | // case, perhaps we can't determine each of the subelements individually, but |
| 7205 | // we might be able to find the complete struct somewhere. |
| 7206 | |
| 7207 | // Find the value that is at that particular spot |
| 7208 | Value *V = FindInsertedValue(V: From, idx_range: Idxs); |
| 7209 | |
| 7210 | if (!V) |
| 7211 | return nullptr; |
| 7212 | |
| 7213 | // Insert the value in the new (sub) aggregate |
| 7214 | return InsertValueInst::Create(Agg: To, Val: V, Idxs: ArrayRef(Idxs).slice(N: IdxSkip), NameStr: "tmp" , |
| 7215 | InsertBefore); |
| 7216 | } |
| 7217 | |
| 7218 | // This helper takes a nested struct and extracts a part of it (which is again a |
| 7219 | // struct) into a new value. For example, given the struct: |
| 7220 | // { a, { b, { c, d }, e } } |
| 7221 | // and the indices "1, 1" this returns |
| 7222 | // { c, d }. |
| 7223 | // |
| 7224 | // It does this by inserting an insertvalue for each element in the resulting |
| 7225 | // struct, as opposed to just inserting a single struct. This will only work if |
| 7226 | // each of the elements of the substruct are known (ie, inserted into From by an |
| 7227 | // insertvalue instruction somewhere). |
| 7228 | // |
| 7229 | // All inserted insertvalue instructions are inserted before InsertBefore |
| 7230 | static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range, |
| 7231 | BasicBlock::iterator InsertBefore) { |
| 7232 | Type *IndexedType = ExtractValueInst::getIndexedType(Agg: From->getType(), |
| 7233 | Idxs: idx_range); |
| 7234 | Value *To = PoisonValue::get(T: IndexedType); |
| 7235 | SmallVector<unsigned, 10> Idxs(idx_range); |
| 7236 | unsigned IdxSkip = Idxs.size(); |
| 7237 | |
| 7238 | return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore); |
| 7239 | } |
| 7240 | |
| 7241 | /// Given an aggregate and a sequence of indices, see if the scalar value |
| 7242 | /// indexed is already around as a register, for example if it was inserted |
| 7243 | /// directly into the aggregate. |
| 7244 | /// |
| 7245 | /// If InsertBefore is not null, this function will duplicate (modified) |
| 7246 | /// insertvalues when a part of a nested struct is extracted. |
| 7247 | Value * |
| 7248 | llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range, |
| 7249 | std::optional<BasicBlock::iterator> InsertBefore) { |
| 7250 | // Nothing to index? Just return V then (this is useful at the end of our |
| 7251 | // recursion). |
| 7252 | if (idx_range.empty()) |
| 7253 | return V; |
| 7254 | // We have indices, so V should have an indexable type. |
| 7255 | assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) && |
| 7256 | "Not looking at a struct or array?" ); |
| 7257 | assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) && |
| 7258 | "Invalid indices for type?" ); |
| 7259 | |
| 7260 | if (Constant *C = dyn_cast<Constant>(Val: V)) { |
| 7261 | C = C->getAggregateElement(Elt: idx_range[0]); |
| 7262 | if (!C) return nullptr; |
| 7263 | return FindInsertedValue(V: C, idx_range: idx_range.slice(N: 1), InsertBefore); |
| 7264 | } |
| 7265 | |
| 7266 | if (InsertValueInst *I = dyn_cast<InsertValueInst>(Val: V)) { |
| 7267 | // Loop the indices for the insertvalue instruction in parallel with the |
| 7268 | // requested indices |
| 7269 | const unsigned *req_idx = idx_range.begin(); |
| 7270 | for (const unsigned *i = I->idx_begin(), *e = I->idx_end(); |
| 7271 | i != e; ++i, ++req_idx) { |
| 7272 | if (req_idx == idx_range.end()) { |
| 7273 | // We can't handle this without inserting insertvalues |
| 7274 | if (!InsertBefore) |
| 7275 | return nullptr; |
| 7276 | |
| 7277 | // The requested index identifies a part of a nested aggregate. Handle |
| 7278 | // this specially. For example, |
| 7279 | // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0 |
| 7280 | // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1 |
| 7281 | // %C = extractvalue {i32, { i32, i32 } } %B, 1 |
| 7282 | // This can be changed into |
| 7283 | // %A = insertvalue {i32, i32 } undef, i32 10, 0 |
| 7284 | // %C = insertvalue {i32, i32 } %A, i32 11, 1 |
| 7285 | // which allows the unused 0,0 element from the nested struct to be |
| 7286 | // removed. |
| 7287 | return BuildSubAggregate(From: V, idx_range: ArrayRef(idx_range.begin(), req_idx), |
| 7288 | InsertBefore: *InsertBefore); |
| 7289 | } |
| 7290 | |
| 7291 | // This insert value inserts something else than what we are looking for. |
| 7292 | // See if the (aggregate) value inserted into has the value we are |
| 7293 | // looking for, then. |
| 7294 | if (*req_idx != *i) |
| 7295 | return FindInsertedValue(V: I->getAggregateOperand(), idx_range, |
| 7296 | InsertBefore); |
| 7297 | } |
| 7298 | // If we end up here, the indices of the insertvalue match with those |
| 7299 | // requested (though possibly only partially). Now we recursively look at |
| 7300 | // the inserted value, passing any remaining indices. |
| 7301 | return FindInsertedValue(V: I->getInsertedValueOperand(), |
| 7302 | idx_range: ArrayRef(req_idx, idx_range.end()), InsertBefore); |
| 7303 | } |
| 7304 | |
| 7305 | if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(Val: V)) { |
| 7306 | // If we're extracting a value from an aggregate that was extracted from |
| 7307 | // something else, we can extract from that something else directly instead. |
| 7308 | // However, we will need to chain I's indices with the requested indices. |
| 7309 | |
| 7310 | // Calculate the number of indices required |
| 7311 | unsigned size = I->getNumIndices() + idx_range.size(); |
| 7312 | // Allocate some space to put the new indices in |
| 7313 | SmallVector<unsigned, 5> Idxs; |
| 7314 | Idxs.reserve(N: size); |
| 7315 | // Add indices from the extract value instruction |
| 7316 | Idxs.append(in_start: I->idx_begin(), in_end: I->idx_end()); |
| 7317 | |
| 7318 | // Add requested indices |
| 7319 | Idxs.append(in_start: idx_range.begin(), in_end: idx_range.end()); |
| 7320 | |
| 7321 | assert(Idxs.size() == size |
| 7322 | && "Number of indices added not correct?" ); |
| 7323 | |
| 7324 | return FindInsertedValue(V: I->getAggregateOperand(), idx_range: Idxs, InsertBefore); |
| 7325 | } |
| 7326 | // Otherwise, we don't know (such as, extracting from a function return value |
| 7327 | // or load instruction) |
| 7328 | return nullptr; |
| 7329 | } |
| 7330 | |
| 7331 | // If V refers to an initialized global constant, set Slice either to |
| 7332 | // its initializer if the size of its elements equals ElementSize, or, |
| 7333 | // for ElementSize == 8, to its representation as an array of unsiged |
| 7334 | // char. Return true on success. |
| 7335 | // Offset is in the unit "nr of ElementSize sized elements". |
| 7336 | bool llvm::getConstantDataArrayInfo(const Value *V, |
| 7337 | ConstantDataArraySlice &Slice, |
| 7338 | unsigned ElementSize, uint64_t Offset) { |
| 7339 | assert(V && "V should not be null." ); |
| 7340 | assert((ElementSize % 8) == 0 && |
| 7341 | "ElementSize expected to be a multiple of the size of a byte." ); |
| 7342 | unsigned ElementSizeInBytes = ElementSize / 8; |
| 7343 | |
| 7344 | // Drill down into the pointer expression V, ignoring any intervening |
| 7345 | // casts, and determine the identity of the object it references along |
| 7346 | // with the cumulative byte offset into it. |
| 7347 | const GlobalVariable *GV = |
| 7348 | dyn_cast<GlobalVariable>(Val: getUnderlyingObject(V)); |
| 7349 | if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer()) |
| 7350 | // Fail if V is not based on constant global object. |
| 7351 | return false; |
| 7352 | |
| 7353 | const DataLayout &DL = GV->getDataLayout(); |
| 7354 | APInt Off(DL.getIndexTypeSizeInBits(Ty: V->getType()), 0); |
| 7355 | |
| 7356 | if (GV != V->stripAndAccumulateConstantOffsets(DL, Offset&: Off, |
| 7357 | /*AllowNonInbounds*/ true)) |
| 7358 | // Fail if a constant offset could not be determined. |
| 7359 | return false; |
| 7360 | |
| 7361 | uint64_t StartIdx = Off.getLimitedValue(); |
| 7362 | if (StartIdx == UINT64_MAX) |
| 7363 | // Fail if the constant offset is excessive. |
| 7364 | return false; |
| 7365 | |
| 7366 | // Off/StartIdx is in the unit of bytes. So we need to convert to number of |
| 7367 | // elements. Simply bail out if that isn't possible. |
| 7368 | if ((StartIdx % ElementSizeInBytes) != 0) |
| 7369 | return false; |
| 7370 | |
| 7371 | Offset += StartIdx / ElementSizeInBytes; |
| 7372 | ConstantDataArray *Array = nullptr; |
| 7373 | ArrayType *ArrayTy = nullptr; |
| 7374 | |
| 7375 | if (GV->getInitializer()->isNullValue()) { |
| 7376 | Type *GVTy = GV->getValueType(); |
| 7377 | uint64_t SizeInBytes = DL.getTypeStoreSize(Ty: GVTy).getFixedValue(); |
| 7378 | uint64_t Length = SizeInBytes / ElementSizeInBytes; |
| 7379 | |
| 7380 | Slice.Array = nullptr; |
| 7381 | Slice.Offset = 0; |
| 7382 | // Return an empty Slice for undersized constants to let callers |
| 7383 | // transform even undefined library calls into simpler, well-defined |
| 7384 | // expressions. This is preferable to making the calls although it |
| 7385 | // prevents sanitizers from detecting such calls. |
| 7386 | Slice.Length = Length < Offset ? 0 : Length - Offset; |
| 7387 | return true; |
| 7388 | } |
| 7389 | |
| 7390 | auto *Init = const_cast<Constant *>(GV->getInitializer()); |
| 7391 | if (auto *ArrayInit = dyn_cast<ConstantDataArray>(Val: Init)) { |
| 7392 | Type *InitElTy = ArrayInit->getElementType(); |
| 7393 | if (InitElTy->isIntegerTy(BitWidth: ElementSize)) { |
| 7394 | // If Init is an initializer for an array of the expected type |
| 7395 | // and size, use it as is. |
| 7396 | Array = ArrayInit; |
| 7397 | ArrayTy = ArrayInit->getType(); |
| 7398 | } |
| 7399 | } |
| 7400 | |
| 7401 | if (!Array) { |
| 7402 | if (ElementSize != 8) |
| 7403 | // TODO: Handle conversions to larger integral types. |
| 7404 | return false; |
| 7405 | |
| 7406 | // Otherwise extract the portion of the initializer starting |
| 7407 | // at Offset as an array of bytes, and reset Offset. |
| 7408 | Init = ReadByteArrayFromGlobal(GV, Offset); |
| 7409 | if (!Init) |
| 7410 | return false; |
| 7411 | |
| 7412 | Offset = 0; |
| 7413 | Array = dyn_cast<ConstantDataArray>(Val: Init); |
| 7414 | ArrayTy = dyn_cast<ArrayType>(Val: Init->getType()); |
| 7415 | } |
| 7416 | |
| 7417 | uint64_t NumElts = ArrayTy->getArrayNumElements(); |
| 7418 | if (Offset > NumElts) |
| 7419 | return false; |
| 7420 | |
| 7421 | Slice.Array = Array; |
| 7422 | Slice.Offset = Offset; |
| 7423 | Slice.Length = NumElts - Offset; |
| 7424 | return true; |
| 7425 | } |
| 7426 | |
| 7427 | /// Extract bytes from the initializer of the constant array V, which need |
| 7428 | /// not be a nul-terminated string. On success, store the bytes in Str and |
| 7429 | /// return true. When TrimAtNul is set, Str will contain only the bytes up |
| 7430 | /// to but not including the first nul. Return false on failure. |
| 7431 | bool llvm::getConstantStringInfo(const Value *V, StringRef &Str, |
| 7432 | bool TrimAtNul) { |
| 7433 | ConstantDataArraySlice Slice; |
| 7434 | if (!getConstantDataArrayInfo(V, Slice, ElementSize: 8)) |
| 7435 | return false; |
| 7436 | |
| 7437 | if (Slice.Array == nullptr) { |
| 7438 | if (TrimAtNul) { |
| 7439 | // Return a nul-terminated string even for an empty Slice. This is |
| 7440 | // safe because all existing SimplifyLibcalls callers require string |
| 7441 | // arguments and the behavior of the functions they fold is undefined |
| 7442 | // otherwise. Folding the calls this way is preferable to making |
| 7443 | // the undefined library calls, even though it prevents sanitizers |
| 7444 | // from reporting such calls. |
| 7445 | Str = StringRef(); |
| 7446 | return true; |
| 7447 | } |
| 7448 | if (Slice.Length == 1) { |
| 7449 | Str = StringRef("" , 1); |
| 7450 | return true; |
| 7451 | } |
| 7452 | // We cannot instantiate a StringRef as we do not have an appropriate string |
| 7453 | // of 0s at hand. |
| 7454 | return false; |
| 7455 | } |
| 7456 | |
| 7457 | // Start out with the entire array in the StringRef. |
| 7458 | Str = Slice.Array->getAsString(); |
| 7459 | // Skip over 'offset' bytes. |
| 7460 | Str = Str.substr(Start: Slice.Offset); |
| 7461 | |
| 7462 | if (TrimAtNul) { |
| 7463 | // Trim off the \0 and anything after it. If the array is not nul |
| 7464 | // terminated, we just return the whole end of string. The client may know |
| 7465 | // some other way that the string is length-bound. |
| 7466 | Str = Str.substr(Start: 0, N: Str.find(C: '\0')); |
| 7467 | } |
| 7468 | return true; |
| 7469 | } |
| 7470 | |
| 7471 | // These next two are very similar to the above, but also look through PHI |
| 7472 | // nodes. |
| 7473 | // TODO: See if we can integrate these two together. |
| 7474 | |
| 7475 | /// If we can compute the length of the string pointed to by |
| 7476 | /// the specified pointer, return 'len+1'. If we can't, return 0. |
| 7477 | static uint64_t GetStringLengthH(const Value *V, |
| 7478 | SmallPtrSetImpl<const PHINode*> &PHIs, |
| 7479 | unsigned CharSize) { |
| 7480 | // Look through noop bitcast instructions. |
| 7481 | V = V->stripPointerCasts(); |
| 7482 | |
| 7483 | // If this is a PHI node, there are two cases: either we have already seen it |
| 7484 | // or we haven't. |
| 7485 | if (const PHINode *PN = dyn_cast<PHINode>(Val: V)) { |
| 7486 | if (!PHIs.insert(Ptr: PN).second) |
| 7487 | return ~0ULL; // already in the set. |
| 7488 | |
| 7489 | // If it was new, see if all the input strings are the same length. |
| 7490 | uint64_t LenSoFar = ~0ULL; |
| 7491 | for (Value *IncValue : PN->incoming_values()) { |
| 7492 | uint64_t Len = GetStringLengthH(V: IncValue, PHIs, CharSize); |
| 7493 | if (Len == 0) return 0; // Unknown length -> unknown. |
| 7494 | |
| 7495 | if (Len == ~0ULL) continue; |
| 7496 | |
| 7497 | if (Len != LenSoFar && LenSoFar != ~0ULL) |
| 7498 | return 0; // Disagree -> unknown. |
| 7499 | LenSoFar = Len; |
| 7500 | } |
| 7501 | |
| 7502 | // Success, all agree. |
| 7503 | return LenSoFar; |
| 7504 | } |
| 7505 | |
| 7506 | // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y) |
| 7507 | if (const SelectInst *SI = dyn_cast<SelectInst>(Val: V)) { |
| 7508 | uint64_t Len1 = GetStringLengthH(V: SI->getTrueValue(), PHIs, CharSize); |
| 7509 | if (Len1 == 0) return 0; |
| 7510 | uint64_t Len2 = GetStringLengthH(V: SI->getFalseValue(), PHIs, CharSize); |
| 7511 | if (Len2 == 0) return 0; |
| 7512 | if (Len1 == ~0ULL) return Len2; |
| 7513 | if (Len2 == ~0ULL) return Len1; |
| 7514 | if (Len1 != Len2) return 0; |
| 7515 | return Len1; |
| 7516 | } |
| 7517 | |
| 7518 | // Otherwise, see if we can read the string. |
| 7519 | ConstantDataArraySlice Slice; |
| 7520 | if (!getConstantDataArrayInfo(V, Slice, ElementSize: CharSize)) |
| 7521 | return 0; |
| 7522 | |
| 7523 | if (Slice.Array == nullptr) |
| 7524 | // Zeroinitializer (including an empty one). |
| 7525 | return 1; |
| 7526 | |
| 7527 | // Search for the first nul character. Return a conservative result even |
| 7528 | // when there is no nul. This is safe since otherwise the string function |
| 7529 | // being folded such as strlen is undefined, and can be preferable to |
| 7530 | // making the undefined library call. |
| 7531 | unsigned NullIndex = 0; |
| 7532 | for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) { |
| 7533 | if (Slice.Array->getElementAsInteger(i: Slice.Offset + NullIndex) == 0) |
| 7534 | break; |
| 7535 | } |
| 7536 | |
| 7537 | return NullIndex + 1; |
| 7538 | } |
| 7539 | |
| 7540 | /// If we can compute the length of the string pointed to by |
| 7541 | /// the specified pointer, return 'len+1'. If we can't, return 0. |
| 7542 | uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) { |
| 7543 | if (!V->getType()->isPointerTy()) |
| 7544 | return 0; |
| 7545 | |
| 7546 | SmallPtrSet<const PHINode*, 32> PHIs; |
| 7547 | uint64_t Len = GetStringLengthH(V, PHIs, CharSize); |
| 7548 | // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return |
| 7549 | // an empty string as a length. |
| 7550 | return Len == ~0ULL ? 1 : Len; |
| 7551 | } |
| 7552 | |
| 7553 | const Value * |
| 7554 | llvm::getArgumentAliasingToReturnedPointer(const CallBase *Call, |
| 7555 | bool MustPreserveOffset, |
| 7556 | bool MustPreserveProvenance) { |
| 7557 | assert(Call && |
| 7558 | "getArgumentAliasingToReturnedPointer only works on nonnull calls" ); |
| 7559 | if (const Value *RV = Call->getReturnedArgOperand()) |
| 7560 | return RV; |
| 7561 | // This can be used only as a aliasing property. |
| 7562 | if (isIntrinsicReturningPointerAliasingArgumentWithoutCapturing( |
| 7563 | Call, MustPreserveOffset, MustPreserveProvenance)) |
| 7564 | return Call->getArgOperand(i: 0); |
| 7565 | return nullptr; |
| 7566 | } |
| 7567 | |
| 7568 | bool llvm::isIntrinsicReturningPointerAliasingArgumentWithoutCapturing( |
| 7569 | const CallBase *Call, bool MustPreserveOffset, |
| 7570 | bool MustPreserveProvenance) { |
| 7571 | switch (Call->getIntrinsicID()) { |
| 7572 | case Intrinsic::launder_invariant_group: |
| 7573 | case Intrinsic::aarch64_irg: |
| 7574 | case Intrinsic::aarch64_tagp: |
| 7575 | // The amdgcn_make_buffer_rsrc function does not alter the address of the |
| 7576 | // input pointer (and thus preserves the byte offset, which is the property |
| 7577 | // the MustPreserveOffset flag selects). However, it will not necessarily |
| 7578 | // map ptr addrspace(N) null to ptr addrspace(8) null, aka the "null |
| 7579 | // descriptor", which has "all loads return 0, all stores are dropped" |
| 7580 | // semantics. Given the context of this intrinsic list, no one should be |
| 7581 | // relying on such a strict bit-exact null mapping (and, at time of |
| 7582 | // writing, they are not), but we document this fact out of an abundance |
| 7583 | // of caution. |
| 7584 | case Intrinsic::amdgcn_make_buffer_rsrc: |
| 7585 | return !MustPreserveProvenance; |
| 7586 | case Intrinsic::ptrmask: |
| 7587 | return !MustPreserveOffset; |
| 7588 | case Intrinsic::threadlocal_address: |
| 7589 | // The underlying variable changes with thread ID. The Thread ID may change |
| 7590 | // at coroutine suspend points. |
| 7591 | return !Call->getParent()->getParent()->isPresplitCoroutine(); |
| 7592 | default: |
| 7593 | return false; |
| 7594 | } |
| 7595 | } |
| 7596 | |
| 7597 | /// \p PN defines a loop-variant pointer to an object. Check if the |
| 7598 | /// previous iteration of the loop was referring to the same object as \p PN. |
| 7599 | static bool isSameUnderlyingObjectInLoop(const PHINode *PN, |
| 7600 | const LoopInfo *LI) { |
| 7601 | // Find the loop-defined value. |
| 7602 | Loop *L = LI->getLoopFor(BB: PN->getParent()); |
| 7603 | if (PN->getNumIncomingValues() != 2) |
| 7604 | return true; |
| 7605 | |
| 7606 | // Find the value from previous iteration. |
| 7607 | auto *PrevValue = dyn_cast<Instruction>(Val: PN->getIncomingValue(i: 0)); |
| 7608 | if (!PrevValue || LI->getLoopFor(BB: PrevValue->getParent()) != L) |
| 7609 | PrevValue = dyn_cast<Instruction>(Val: PN->getIncomingValue(i: 1)); |
| 7610 | if (!PrevValue || LI->getLoopFor(BB: PrevValue->getParent()) != L) |
| 7611 | return true; |
| 7612 | |
| 7613 | // If a new pointer is loaded in the loop, the pointer references a different |
| 7614 | // object in every iteration. E.g.: |
| 7615 | // for (i) |
| 7616 | // int *p = a[i]; |
| 7617 | // ... |
| 7618 | if (auto *Load = dyn_cast<LoadInst>(Val: PrevValue)) |
| 7619 | if (!L->isLoopInvariant(V: Load->getPointerOperand())) |
| 7620 | return false; |
| 7621 | return true; |
| 7622 | } |
| 7623 | |
| 7624 | const Value *llvm::getUnderlyingObject(const Value *V, unsigned MaxLookup, |
| 7625 | bool MustPreserveProvenance) { |
| 7626 | for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) { |
| 7627 | if (auto *GEP = dyn_cast<GEPOperator>(Val: V)) { |
| 7628 | const Value *PtrOp = GEP->getPointerOperand(); |
| 7629 | if (!PtrOp->getType()->isPointerTy()) // Only handle scalar pointer base. |
| 7630 | return V; |
| 7631 | V = PtrOp; |
| 7632 | } else if (Operator::getOpcode(V) == Instruction::BitCast || |
| 7633 | Operator::getOpcode(V) == Instruction::AddrSpaceCast) { |
| 7634 | Value *NewV = cast<Operator>(Val: V)->getOperand(i: 0); |
| 7635 | if (!NewV->getType()->isPointerTy()) |
| 7636 | return V; |
| 7637 | V = NewV; |
| 7638 | } else if (auto *GA = dyn_cast<GlobalAlias>(Val: V)) { |
| 7639 | if (GA->isInterposable()) |
| 7640 | return V; |
| 7641 | V = GA->getAliasee(); |
| 7642 | } else { |
| 7643 | if (auto *PHI = dyn_cast<PHINode>(Val: V)) { |
| 7644 | // Look through single-arg phi nodes created by LCSSA. |
| 7645 | if (PHI->getNumIncomingValues() == 1) { |
| 7646 | V = PHI->getIncomingValue(i: 0); |
| 7647 | continue; |
| 7648 | } |
| 7649 | } else if (auto *Call = dyn_cast<CallBase>(Val: V)) { |
| 7650 | // CaptureTracking can know about special capturing properties of some |
| 7651 | // intrinsics like launder.invariant.group, that can't be expressed with |
| 7652 | // the attributes, but have properties like returning aliasing pointer. |
| 7653 | // Because some analysis may assume that nocaptured pointer is not |
| 7654 | // returned from some special intrinsic (because function would have to |
| 7655 | // be marked with returns attribute), it is crucial to use this function |
| 7656 | // because it should be in sync with CaptureTracking. Not using it may |
| 7657 | // cause weird miscompilations where 2 aliasing pointers are assumed to |
| 7658 | // noalias. |
| 7659 | if (auto *RP = getArgumentAliasingToReturnedPointer( |
| 7660 | Call, /*MustPreserveOffset=*/false, MustPreserveProvenance)) { |
| 7661 | V = RP; |
| 7662 | continue; |
| 7663 | } |
| 7664 | } |
| 7665 | |
| 7666 | return V; |
| 7667 | } |
| 7668 | assert(V->getType()->isPointerTy() && "Unexpected operand type!" ); |
| 7669 | } |
| 7670 | return V; |
| 7671 | } |
| 7672 | |
| 7673 | void llvm::getUnderlyingObjects(const Value *V, |
| 7674 | SmallVectorImpl<const Value *> &Objects, |
| 7675 | const LoopInfo *LI, unsigned MaxLookup) { |
| 7676 | SmallPtrSet<const Value *, 4> Visited; |
| 7677 | SmallVector<const Value *, 4> Worklist; |
| 7678 | Worklist.push_back(Elt: V); |
| 7679 | do { |
| 7680 | const Value *P = Worklist.pop_back_val(); |
| 7681 | P = getUnderlyingObject(V: P, MaxLookup); |
| 7682 | |
| 7683 | if (!Visited.insert(Ptr: P).second) |
| 7684 | continue; |
| 7685 | |
| 7686 | if (auto *SI = dyn_cast<SelectInst>(Val: P)) { |
| 7687 | Worklist.push_back(Elt: SI->getTrueValue()); |
| 7688 | Worklist.push_back(Elt: SI->getFalseValue()); |
| 7689 | continue; |
| 7690 | } |
| 7691 | |
| 7692 | if (auto *PN = dyn_cast<PHINode>(Val: P)) { |
| 7693 | // If this PHI changes the underlying object in every iteration of the |
| 7694 | // loop, don't look through it. Consider: |
| 7695 | // int **A; |
| 7696 | // for (i) { |
| 7697 | // Prev = Curr; // Prev = PHI (Prev_0, Curr) |
| 7698 | // Curr = A[i]; |
| 7699 | // *Prev, *Curr; |
| 7700 | // |
| 7701 | // Prev is tracking Curr one iteration behind so they refer to different |
| 7702 | // underlying objects. |
| 7703 | if (!LI || !LI->isLoopHeader(BB: PN->getParent()) || |
| 7704 | isSameUnderlyingObjectInLoop(PN, LI)) |
| 7705 | append_range(C&: Worklist, R: PN->incoming_values()); |
| 7706 | else |
| 7707 | Objects.push_back(Elt: P); |
| 7708 | continue; |
| 7709 | } |
| 7710 | |
| 7711 | Objects.push_back(Elt: P); |
| 7712 | } while (!Worklist.empty()); |
| 7713 | } |
| 7714 | |
| 7715 | const Value *llvm::getUnderlyingObjectAggressive(const Value *V, |
| 7716 | bool MustPreserveProvenance) { |
| 7717 | const unsigned MaxVisited = 8; |
| 7718 | |
| 7719 | SmallPtrSet<const Value *, 8> Visited; |
| 7720 | SmallVector<const Value *, 8> Worklist; |
| 7721 | Worklist.push_back(Elt: V); |
| 7722 | const Value *Object = nullptr; |
| 7723 | // Used as fallback if we can't find a common underlying object through |
| 7724 | // recursion. |
| 7725 | bool First = true; |
| 7726 | const Value *FirstObject = |
| 7727 | getUnderlyingObject(V, MaxLookup: MaxLookupSearchDepth, MustPreserveProvenance); |
| 7728 | do { |
| 7729 | const Value *P = Worklist.pop_back_val(); |
| 7730 | P = First ? FirstObject |
| 7731 | : getUnderlyingObject(V: P, MaxLookup: MaxLookupSearchDepth, |
| 7732 | MustPreserveProvenance); |
| 7733 | First = false; |
| 7734 | |
| 7735 | if (!Visited.insert(Ptr: P).second) |
| 7736 | continue; |
| 7737 | |
| 7738 | if (Visited.size() == MaxVisited) |
| 7739 | return FirstObject; |
| 7740 | |
| 7741 | if (auto *SI = dyn_cast<SelectInst>(Val: P)) { |
| 7742 | Worklist.push_back(Elt: SI->getTrueValue()); |
| 7743 | Worklist.push_back(Elt: SI->getFalseValue()); |
| 7744 | continue; |
| 7745 | } |
| 7746 | |
| 7747 | if (auto *PN = dyn_cast<PHINode>(Val: P)) { |
| 7748 | append_range(C&: Worklist, R: PN->incoming_values()); |
| 7749 | continue; |
| 7750 | } |
| 7751 | |
| 7752 | if (!Object) |
| 7753 | Object = P; |
| 7754 | else if (Object != P) |
| 7755 | return FirstObject; |
| 7756 | } while (!Worklist.empty()); |
| 7757 | |
| 7758 | return Object ? Object : FirstObject; |
| 7759 | } |
| 7760 | |
| 7761 | /// This is the function that does the work of looking through basic |
| 7762 | /// ptrtoint+arithmetic+inttoptr sequences. |
| 7763 | static const Value *getUnderlyingObjectFromInt(const Value *V) { |
| 7764 | do { |
| 7765 | if (const Operator *U = dyn_cast<Operator>(Val: V)) { |
| 7766 | // If we find a ptrtoint, we can transfer control back to the |
| 7767 | // regular getUnderlyingObjectFromInt. |
| 7768 | if (U->getOpcode() == Instruction::PtrToInt) |
| 7769 | return U->getOperand(i: 0); |
| 7770 | // If we find an add of a constant, a multiplied value, or a phi, it's |
| 7771 | // likely that the other operand will lead us to the base |
| 7772 | // object. We don't have to worry about the case where the |
| 7773 | // object address is somehow being computed by the multiply, |
| 7774 | // because our callers only care when the result is an |
| 7775 | // identifiable object. |
| 7776 | if (U->getOpcode() != Instruction::Add || |
| 7777 | (!isa<ConstantInt>(Val: U->getOperand(i: 1)) && |
| 7778 | Operator::getOpcode(V: U->getOperand(i: 1)) != Instruction::Mul && |
| 7779 | !isa<PHINode>(Val: U->getOperand(i: 1)))) |
| 7780 | return V; |
| 7781 | V = U->getOperand(i: 0); |
| 7782 | } else { |
| 7783 | return V; |
| 7784 | } |
| 7785 | assert(V->getType()->isIntegerTy() && "Unexpected operand type!" ); |
| 7786 | } while (true); |
| 7787 | } |
| 7788 | |
| 7789 | /// This is a wrapper around getUnderlyingObjects and adds support for basic |
| 7790 | /// ptrtoint+arithmetic+inttoptr sequences. |
| 7791 | /// It returns false if unidentified object is found in getUnderlyingObjects. |
| 7792 | bool llvm::getUnderlyingObjectsForCodeGen(const Value *V, |
| 7793 | SmallVectorImpl<Value *> &Objects) { |
| 7794 | SmallPtrSet<const Value *, 16> Visited; |
| 7795 | SmallVector<const Value *, 4> Working(1, V); |
| 7796 | bool AllObjectsIdentified = true; |
| 7797 | do { |
| 7798 | V = Working.pop_back_val(); |
| 7799 | |
| 7800 | SmallVector<const Value *, 4> Objs; |
| 7801 | getUnderlyingObjects(V, Objects&: Objs); |
| 7802 | |
| 7803 | for (const Value *V : Objs) { |
| 7804 | if (!Visited.insert(Ptr: V).second) |
| 7805 | continue; |
| 7806 | if (Operator::getOpcode(V) == Instruction::IntToPtr) { |
| 7807 | const Value *O = |
| 7808 | getUnderlyingObjectFromInt(V: cast<User>(Val: V)->getOperand(i: 0)); |
| 7809 | if (O->getType()->isPointerTy()) { |
| 7810 | Working.push_back(Elt: O); |
| 7811 | continue; |
| 7812 | } |
| 7813 | } |
| 7814 | AllObjectsIdentified &= isIdentifiedObject(V); |
| 7815 | Objects.push_back(Elt: const_cast<Value *>(V)); |
| 7816 | } |
| 7817 | } while (!Working.empty()); |
| 7818 | return AllObjectsIdentified; |
| 7819 | } |
| 7820 | |
| 7821 | AllocaInst *llvm::findAllocaForValue(Value *V, bool OffsetZero) { |
| 7822 | AllocaInst *Result = nullptr; |
| 7823 | SmallPtrSet<Value *, 4> Visited; |
| 7824 | SmallVector<Value *, 4> Worklist; |
| 7825 | |
| 7826 | auto AddWork = [&](Value *V) { |
| 7827 | if (Visited.insert(Ptr: V).second) |
| 7828 | Worklist.push_back(Elt: V); |
| 7829 | }; |
| 7830 | |
| 7831 | AddWork(V); |
| 7832 | do { |
| 7833 | V = Worklist.pop_back_val(); |
| 7834 | assert(Visited.count(V)); |
| 7835 | |
| 7836 | if (AllocaInst *AI = dyn_cast<AllocaInst>(Val: V)) { |
| 7837 | if (Result && Result != AI) |
| 7838 | return nullptr; |
| 7839 | Result = AI; |
| 7840 | } else if (CastInst *CI = dyn_cast<CastInst>(Val: V)) { |
| 7841 | AddWork(CI->getOperand(i_nocapture: 0)); |
| 7842 | } else if (PHINode *PN = dyn_cast<PHINode>(Val: V)) { |
| 7843 | for (Value *IncValue : PN->incoming_values()) |
| 7844 | AddWork(IncValue); |
| 7845 | } else if (auto *SI = dyn_cast<SelectInst>(Val: V)) { |
| 7846 | AddWork(SI->getTrueValue()); |
| 7847 | AddWork(SI->getFalseValue()); |
| 7848 | } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Val: V)) { |
| 7849 | if (OffsetZero && !GEP->hasAllZeroIndices()) |
| 7850 | return nullptr; |
| 7851 | AddWork(GEP->getPointerOperand()); |
| 7852 | } else if (CallBase *CB = dyn_cast<CallBase>(Val: V)) { |
| 7853 | Value *Returned = CB->getReturnedArgOperand(); |
| 7854 | if (Returned) |
| 7855 | AddWork(Returned); |
| 7856 | else |
| 7857 | return nullptr; |
| 7858 | } else { |
| 7859 | return nullptr; |
| 7860 | } |
| 7861 | } while (!Worklist.empty()); |
| 7862 | |
| 7863 | return Result; |
| 7864 | } |
| 7865 | |
| 7866 | bool llvm::isNotCrossLaneOperation(const Instruction *I) { |
| 7867 | if (auto *II = dyn_cast<IntrinsicInst>(Val: I)) |
| 7868 | return isTriviallyVectorizable(ID: II->getIntrinsicID()); |
| 7869 | auto *Shuffle = dyn_cast<ShuffleVectorInst>(Val: I); |
| 7870 | return (!Shuffle || Shuffle->isSelect()) && |
| 7871 | !isa<CallBase, BitCastInst, ExtractElementInst>(Val: I); |
| 7872 | } |
| 7873 | |
| 7874 | bool llvm::isSafeToSpeculativelyExecute( |
| 7875 | const Instruction *Inst, const Instruction *CtxI, AssumptionCache *AC, |
| 7876 | const DominatorTree *DT, const TargetLibraryInfo *TLI, bool UseVariableInfo, |
| 7877 | bool IgnoreUBImplyingAttrs) { |
| 7878 | return isSafeToSpeculativelyExecuteWithOpcode(Opcode: Inst->getOpcode(), Inst, CtxI, |
| 7879 | AC, DT, TLI, UseVariableInfo, |
| 7880 | IgnoreUBImplyingAttrs); |
| 7881 | } |
| 7882 | |
| 7883 | bool llvm::isSafeToSpeculativelyExecuteWithOpcode( |
| 7884 | unsigned Opcode, const Instruction *Inst, const Instruction *CtxI, |
| 7885 | AssumptionCache *AC, const DominatorTree *DT, const TargetLibraryInfo *TLI, |
| 7886 | bool UseVariableInfo, bool IgnoreUBImplyingAttrs) { |
| 7887 | #ifndef NDEBUG |
| 7888 | if (Inst->getOpcode() != Opcode) { |
| 7889 | // Check that the operands are actually compatible with the Opcode override. |
| 7890 | auto hasEqualReturnAndLeadingOperandTypes = |
| 7891 | [](const Instruction *Inst, unsigned NumLeadingOperands) { |
| 7892 | if (Inst->getNumOperands() < NumLeadingOperands) |
| 7893 | return false; |
| 7894 | const Type *ExpectedType = Inst->getType(); |
| 7895 | for (unsigned ItOp = 0; ItOp < NumLeadingOperands; ++ItOp) |
| 7896 | if (Inst->getOperand(ItOp)->getType() != ExpectedType) |
| 7897 | return false; |
| 7898 | return true; |
| 7899 | }; |
| 7900 | assert(!Instruction::isBinaryOp(Opcode) || |
| 7901 | hasEqualReturnAndLeadingOperandTypes(Inst, 2)); |
| 7902 | assert(!Instruction::isUnaryOp(Opcode) || |
| 7903 | hasEqualReturnAndLeadingOperandTypes(Inst, 1)); |
| 7904 | } |
| 7905 | #endif |
| 7906 | |
| 7907 | switch (Opcode) { |
| 7908 | default: |
| 7909 | return true; |
| 7910 | case Instruction::UDiv: |
| 7911 | case Instruction::URem: { |
| 7912 | // x / y is undefined if y == 0. |
| 7913 | const APInt *V; |
| 7914 | if (match(V: Inst->getOperand(i: 1), P: m_APInt(Res&: V))) |
| 7915 | return *V != 0; |
| 7916 | return false; |
| 7917 | } |
| 7918 | case Instruction::SDiv: |
| 7919 | case Instruction::SRem: { |
| 7920 | // x / y is undefined if y == 0 or x == INT_MIN and y == -1 |
| 7921 | const APInt *Numerator, *Denominator; |
| 7922 | if (!match(V: Inst->getOperand(i: 1), P: m_APInt(Res&: Denominator))) |
| 7923 | return false; |
| 7924 | // We cannot hoist this division if the denominator is 0. |
| 7925 | if (*Denominator == 0) |
| 7926 | return false; |
| 7927 | // It's safe to hoist if the denominator is not 0 or -1. |
| 7928 | if (!Denominator->isAllOnes()) |
| 7929 | return true; |
| 7930 | // At this point we know that the denominator is -1. It is safe to hoist as |
| 7931 | // long we know that the numerator is not INT_MIN. |
| 7932 | if (match(V: Inst->getOperand(i: 0), P: m_APInt(Res&: Numerator))) |
| 7933 | return !Numerator->isMinSignedValue(); |
| 7934 | // The numerator *might* be MinSignedValue. |
| 7935 | return false; |
| 7936 | } |
| 7937 | case Instruction::Load: { |
| 7938 | if (!UseVariableInfo) |
| 7939 | return false; |
| 7940 | |
| 7941 | const LoadInst *LI = dyn_cast<LoadInst>(Val: Inst); |
| 7942 | if (!LI) |
| 7943 | return false; |
| 7944 | if (mustSuppressSpeculation(LI: *LI)) |
| 7945 | return false; |
| 7946 | const DataLayout &DL = LI->getDataLayout(); |
| 7947 | return isDereferenceableAndAlignedPointer( |
| 7948 | V: LI->getPointerOperand(), Ty: LI->getType(), Alignment: LI->getAlign(), |
| 7949 | Q: SimplifyQuery(DL, TLI, DT, AC, CtxI)); |
| 7950 | } |
| 7951 | case Instruction::Call: { |
| 7952 | auto *CI = dyn_cast<const CallInst>(Val: Inst); |
| 7953 | if (!CI) |
| 7954 | return false; |
| 7955 | const Function *Callee = CI->getCalledFunction(); |
| 7956 | |
| 7957 | // The called function could have undefined behavior or side-effects, even |
| 7958 | // if marked readnone nounwind. |
| 7959 | if (!Callee || !Callee->isSpeculatable()) |
| 7960 | return false; |
| 7961 | // Since the operands may be changed after hoisting, undefined behavior may |
| 7962 | // be triggered by some UB-implying attributes. |
| 7963 | return IgnoreUBImplyingAttrs || !CI->hasUBImplyingAttrs(); |
| 7964 | } |
| 7965 | case Instruction::VAArg: |
| 7966 | case Instruction::Alloca: |
| 7967 | case Instruction::Invoke: |
| 7968 | case Instruction::CallBr: |
| 7969 | case Instruction::PHI: |
| 7970 | case Instruction::Store: |
| 7971 | case Instruction::Ret: |
| 7972 | case Instruction::UncondBr: |
| 7973 | case Instruction::CondBr: |
| 7974 | case Instruction::IndirectBr: |
| 7975 | case Instruction::Switch: |
| 7976 | case Instruction::Unreachable: |
| 7977 | case Instruction::Fence: |
| 7978 | case Instruction::AtomicRMW: |
| 7979 | case Instruction::AtomicCmpXchg: |
| 7980 | case Instruction::LandingPad: |
| 7981 | case Instruction::Resume: |
| 7982 | case Instruction::CatchSwitch: |
| 7983 | case Instruction::CatchPad: |
| 7984 | case Instruction::CatchRet: |
| 7985 | case Instruction::CleanupPad: |
| 7986 | case Instruction::CleanupRet: |
| 7987 | return false; // Misc instructions which have effects |
| 7988 | } |
| 7989 | } |
| 7990 | |
| 7991 | bool llvm::mayHaveNonDefUseDependency(const Instruction &I) { |
| 7992 | if (I.mayReadOrWriteMemory()) |
| 7993 | // Memory dependency possible |
| 7994 | return true; |
| 7995 | if (!isSafeToSpeculativelyExecute(Inst: &I)) |
| 7996 | // Can't move above a maythrow call or infinite loop. Or if an |
| 7997 | // inalloca alloca, above a stacksave call. |
| 7998 | return true; |
| 7999 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &I)) |
| 8000 | // 1) Can't reorder two inf-loop calls, even if readonly |
| 8001 | // 2) Also can't reorder an inf-loop call below a instruction which isn't |
| 8002 | // safe to speculative execute. (Inverse of above) |
| 8003 | return true; |
| 8004 | return false; |
| 8005 | } |
| 8006 | |
| 8007 | /// Convert ConstantRange OverflowResult into ValueTracking OverflowResult. |
| 8008 | static OverflowResult mapOverflowResult(ConstantRange::OverflowResult OR) { |
| 8009 | switch (OR) { |
| 8010 | case ConstantRange::OverflowResult::MayOverflow: |
| 8011 | return OverflowResult::MayOverflow; |
| 8012 | case ConstantRange::OverflowResult::AlwaysOverflowsLow: |
| 8013 | return OverflowResult::AlwaysOverflowsLow; |
| 8014 | case ConstantRange::OverflowResult::AlwaysOverflowsHigh: |
| 8015 | return OverflowResult::AlwaysOverflowsHigh; |
| 8016 | case ConstantRange::OverflowResult::NeverOverflows: |
| 8017 | return OverflowResult::NeverOverflows; |
| 8018 | } |
| 8019 | llvm_unreachable("Unknown OverflowResult" ); |
| 8020 | } |
| 8021 | |
| 8022 | /// Combine constant ranges from computeConstantRange() and computeKnownBits(). |
| 8023 | ConstantRange |
| 8024 | llvm::computeConstantRangeIncludingKnownBits(const WithCache<const Value *> &V, |
| 8025 | bool ForSigned, |
| 8026 | const SimplifyQuery &SQ) { |
| 8027 | ConstantRange CR1 = |
| 8028 | ConstantRange::fromKnownBits(Known: V.getKnownBits(Q: SQ), IsSigned: ForSigned); |
| 8029 | ConstantRange CR2 = computeConstantRange(V, ForSigned, SQ); |
| 8030 | ConstantRange::PreferredRangeType RangeType = |
| 8031 | ForSigned ? ConstantRange::Signed : ConstantRange::Unsigned; |
| 8032 | return CR1.intersectWith(CR: CR2, Type: RangeType); |
| 8033 | } |
| 8034 | |
| 8035 | OverflowResult llvm::computeOverflowForUnsignedMul(const Value *LHS, |
| 8036 | const Value *RHS, |
| 8037 | const SimplifyQuery &SQ, |
| 8038 | bool IsNSW) { |
| 8039 | ConstantRange LHSRange = |
| 8040 | computeConstantRangeIncludingKnownBits(V: LHS, /*ForSigned=*/false, SQ); |
| 8041 | ConstantRange RHSRange = |
| 8042 | computeConstantRangeIncludingKnownBits(V: RHS, /*ForSigned=*/false, SQ); |
| 8043 | |
| 8044 | // mul nsw of two non-negative numbers is also nuw. |
| 8045 | if (IsNSW && LHSRange.isAllNonNegative() && RHSRange.isAllNonNegative()) |
| 8046 | return OverflowResult::NeverOverflows; |
| 8047 | |
| 8048 | return mapOverflowResult(OR: LHSRange.unsignedMulMayOverflow(Other: RHSRange)); |
| 8049 | } |
| 8050 | |
| 8051 | OverflowResult llvm::computeOverflowForSignedMul(const Value *LHS, |
| 8052 | const Value *RHS, |
| 8053 | const SimplifyQuery &SQ) { |
| 8054 | // Multiplying n * m significant bits yields a result of n + m significant |
| 8055 | // bits. If the total number of significant bits does not exceed the |
| 8056 | // result bit width (minus 1), there is no overflow. |
| 8057 | // This means if we have enough leading sign bits in the operands |
| 8058 | // we can guarantee that the result does not overflow. |
| 8059 | // Ref: "Hacker's Delight" by Henry Warren |
| 8060 | unsigned BitWidth = LHS->getType()->getScalarSizeInBits(); |
| 8061 | |
| 8062 | // Note that underestimating the number of sign bits gives a more |
| 8063 | // conservative answer. |
| 8064 | unsigned SignBits = |
| 8065 | ::ComputeNumSignBits(V: LHS, Q: SQ) + ::ComputeNumSignBits(V: RHS, Q: SQ); |
| 8066 | |
| 8067 | // First handle the easy case: if we have enough sign bits there's |
| 8068 | // definitely no overflow. |
| 8069 | if (SignBits > BitWidth + 1) |
| 8070 | return OverflowResult::NeverOverflows; |
| 8071 | |
| 8072 | // There are two ambiguous cases where there can be no overflow: |
| 8073 | // SignBits == BitWidth + 1 and |
| 8074 | // SignBits == BitWidth |
| 8075 | // The second case is difficult to check, therefore we only handle the |
| 8076 | // first case. |
| 8077 | if (SignBits == BitWidth + 1) { |
| 8078 | // It overflows only when both arguments are negative and the true |
| 8079 | // product is exactly the minimum negative number. |
| 8080 | // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000 |
| 8081 | // For simplicity we just check if at least one side is not negative. |
| 8082 | KnownBits LHSKnown = computeKnownBits(V: LHS, Q: SQ); |
| 8083 | KnownBits RHSKnown = computeKnownBits(V: RHS, Q: SQ); |
| 8084 | if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative()) |
| 8085 | return OverflowResult::NeverOverflows; |
| 8086 | } |
| 8087 | return OverflowResult::MayOverflow; |
| 8088 | } |
| 8089 | |
| 8090 | OverflowResult |
| 8091 | llvm::computeOverflowForUnsignedAdd(const WithCache<const Value *> &LHS, |
| 8092 | const WithCache<const Value *> &RHS, |
| 8093 | const SimplifyQuery &SQ) { |
| 8094 | ConstantRange LHSRange = |
| 8095 | computeConstantRangeIncludingKnownBits(V: LHS, /*ForSigned=*/false, SQ); |
| 8096 | ConstantRange RHSRange = |
| 8097 | computeConstantRangeIncludingKnownBits(V: RHS, /*ForSigned=*/false, SQ); |
| 8098 | return mapOverflowResult(OR: LHSRange.unsignedAddMayOverflow(Other: RHSRange)); |
| 8099 | } |
| 8100 | |
| 8101 | static OverflowResult |
| 8102 | computeOverflowForSignedAdd(const WithCache<const Value *> &LHS, |
| 8103 | const WithCache<const Value *> &RHS, |
| 8104 | const AddOperator *Add, const SimplifyQuery &SQ) { |
| 8105 | if (Add && Add->hasNoSignedWrap()) { |
| 8106 | return OverflowResult::NeverOverflows; |
| 8107 | } |
| 8108 | |
| 8109 | // If LHS and RHS each have at least two sign bits, the addition will look |
| 8110 | // like |
| 8111 | // |
| 8112 | // XX..... + |
| 8113 | // YY..... |
| 8114 | // |
| 8115 | // If the carry into the most significant position is 0, X and Y can't both |
| 8116 | // be 1 and therefore the carry out of the addition is also 0. |
| 8117 | // |
| 8118 | // If the carry into the most significant position is 1, X and Y can't both |
| 8119 | // be 0 and therefore the carry out of the addition is also 1. |
| 8120 | // |
| 8121 | // Since the carry into the most significant position is always equal to |
| 8122 | // the carry out of the addition, there is no signed overflow. |
| 8123 | if (::ComputeNumSignBits(V: LHS, Q: SQ) > 1 && ::ComputeNumSignBits(V: RHS, Q: SQ) > 1) |
| 8124 | return OverflowResult::NeverOverflows; |
| 8125 | |
| 8126 | ConstantRange LHSRange = |
| 8127 | computeConstantRangeIncludingKnownBits(V: LHS, /*ForSigned=*/true, SQ); |
| 8128 | ConstantRange RHSRange = |
| 8129 | computeConstantRangeIncludingKnownBits(V: RHS, /*ForSigned=*/true, SQ); |
| 8130 | OverflowResult OR = |
| 8131 | mapOverflowResult(OR: LHSRange.signedAddMayOverflow(Other: RHSRange)); |
| 8132 | if (OR != OverflowResult::MayOverflow) |
| 8133 | return OR; |
| 8134 | |
| 8135 | // The remaining code needs Add to be available. Early returns if not so. |
| 8136 | if (!Add) |
| 8137 | return OverflowResult::MayOverflow; |
| 8138 | |
| 8139 | // If the sign of Add is the same as at least one of the operands, this add |
| 8140 | // CANNOT overflow. If this can be determined from the known bits of the |
| 8141 | // operands the above signedAddMayOverflow() check will have already done so. |
| 8142 | // The only other way to improve on the known bits is from an assumption, so |
| 8143 | // call computeKnownBitsFromContext() directly. |
| 8144 | bool LHSOrRHSKnownNonNegative = |
| 8145 | (LHSRange.isAllNonNegative() || RHSRange.isAllNonNegative()); |
| 8146 | bool LHSOrRHSKnownNegative = |
| 8147 | (LHSRange.isAllNegative() || RHSRange.isAllNegative()); |
| 8148 | if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) { |
| 8149 | KnownBits AddKnown(LHSRange.getBitWidth()); |
| 8150 | computeKnownBitsFromContext(V: Add, Known&: AddKnown, Q: SQ); |
| 8151 | if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) || |
| 8152 | (AddKnown.isNegative() && LHSOrRHSKnownNegative)) |
| 8153 | return OverflowResult::NeverOverflows; |
| 8154 | } |
| 8155 | |
| 8156 | return OverflowResult::MayOverflow; |
| 8157 | } |
| 8158 | |
| 8159 | OverflowResult llvm::computeOverflowForUnsignedSub(const Value *LHS, |
| 8160 | const Value *RHS, |
| 8161 | const SimplifyQuery &SQ) { |
| 8162 | // X - (X % ?) |
| 8163 | // The remainder of a value can't have greater magnitude than itself, |
| 8164 | // so the subtraction can't overflow. |
| 8165 | |
| 8166 | // X - (X -nuw ?) |
| 8167 | // In the minimal case, this would simplify to "?", so there's no subtract |
| 8168 | // at all. But if this analysis is used to peek through casts, for example, |
| 8169 | // then determining no-overflow may allow other transforms. |
| 8170 | |
| 8171 | // TODO: There are other patterns like this. |
| 8172 | // See simplifyICmpWithBinOpOnLHS() for candidates. |
| 8173 | if (match(V: RHS, P: m_URem(L: m_Specific(V: LHS), R: m_Value())) || |
| 8174 | match(V: RHS, P: m_NUWSub(L: m_Specific(V: LHS), R: m_Value()))) |
| 8175 | if (isGuaranteedNotToBeUndef(V: LHS, AC: SQ.AC, CtxI: SQ.CtxI, DT: SQ.DT)) |
| 8176 | return OverflowResult::NeverOverflows; |
| 8177 | |
| 8178 | if (auto C = isImpliedByDomCondition(Pred: CmpInst::ICMP_UGE, LHS, RHS, ContextI: SQ.CtxI, |
| 8179 | DL: SQ.DL)) { |
| 8180 | if (*C) |
| 8181 | return OverflowResult::NeverOverflows; |
| 8182 | return OverflowResult::AlwaysOverflowsLow; |
| 8183 | } |
| 8184 | |
| 8185 | ConstantRange LHSRange = |
| 8186 | computeConstantRangeIncludingKnownBits(V: LHS, /*ForSigned=*/false, SQ); |
| 8187 | ConstantRange RHSRange = |
| 8188 | computeConstantRangeIncludingKnownBits(V: RHS, /*ForSigned=*/false, SQ); |
| 8189 | return mapOverflowResult(OR: LHSRange.unsignedSubMayOverflow(Other: RHSRange)); |
| 8190 | } |
| 8191 | |
| 8192 | OverflowResult llvm::computeOverflowForSignedSub(const Value *LHS, |
| 8193 | const Value *RHS, |
| 8194 | const SimplifyQuery &SQ) { |
| 8195 | // X - (X % ?) |
| 8196 | // The remainder of a value can't have greater magnitude than itself, |
| 8197 | // so the subtraction can't overflow. |
| 8198 | |
| 8199 | // X - (X -nsw ?) |
| 8200 | // In the minimal case, this would simplify to "?", so there's no subtract |
| 8201 | // at all. But if this analysis is used to peek through casts, for example, |
| 8202 | // then determining no-overflow may allow other transforms. |
| 8203 | if (match(V: RHS, P: m_SRem(L: m_Specific(V: LHS), R: m_Value())) || |
| 8204 | match(V: RHS, P: m_NSWSub(L: m_Specific(V: LHS), R: m_Value()))) |
| 8205 | if (isGuaranteedNotToBeUndef(V: LHS, AC: SQ.AC, CtxI: SQ.CtxI, DT: SQ.DT)) |
| 8206 | return OverflowResult::NeverOverflows; |
| 8207 | |
| 8208 | // If LHS and RHS each have at least two sign bits, the subtraction |
| 8209 | // cannot overflow. |
| 8210 | if (::ComputeNumSignBits(V: LHS, Q: SQ) > 1 && ::ComputeNumSignBits(V: RHS, Q: SQ) > 1) |
| 8211 | return OverflowResult::NeverOverflows; |
| 8212 | |
| 8213 | ConstantRange LHSRange = |
| 8214 | computeConstantRangeIncludingKnownBits(V: LHS, /*ForSigned=*/true, SQ); |
| 8215 | ConstantRange RHSRange = |
| 8216 | computeConstantRangeIncludingKnownBits(V: RHS, /*ForSigned=*/true, SQ); |
| 8217 | return mapOverflowResult(OR: LHSRange.signedSubMayOverflow(Other: RHSRange)); |
| 8218 | } |
| 8219 | |
| 8220 | bool llvm::isOverflowIntrinsicNoWrap(const WithOverflowInst *WO, |
| 8221 | const DominatorTree &DT) { |
| 8222 | SmallVector<const CondBrInst *, 2> GuardingBranches; |
| 8223 | SmallVector<const ExtractValueInst *, 2> Results; |
| 8224 | |
| 8225 | for (const User *U : WO->users()) { |
| 8226 | if (const auto *EVI = dyn_cast<ExtractValueInst>(Val: U)) { |
| 8227 | assert(EVI->getNumIndices() == 1 && "Obvious from CI's type" ); |
| 8228 | |
| 8229 | if (EVI->getIndices()[0] == 0) |
| 8230 | Results.push_back(Elt: EVI); |
| 8231 | else { |
| 8232 | assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type" ); |
| 8233 | |
| 8234 | for (const auto *U : EVI->users()) |
| 8235 | if (const auto *B = dyn_cast<CondBrInst>(Val: U)) |
| 8236 | GuardingBranches.push_back(Elt: B); |
| 8237 | } |
| 8238 | } else { |
| 8239 | // We are using the aggregate directly in a way we don't want to analyze |
| 8240 | // here (storing it to a global, say). |
| 8241 | return false; |
| 8242 | } |
| 8243 | } |
| 8244 | |
| 8245 | auto AllUsesGuardedByBranch = [&](const CondBrInst *BI) { |
| 8246 | BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(i: 1)); |
| 8247 | |
| 8248 | // Check if all users of the add are provably no-wrap. |
| 8249 | for (const auto *Result : Results) { |
| 8250 | // If the extractvalue itself is not executed on overflow, the we don't |
| 8251 | // need to check each use separately, since domination is transitive. |
| 8252 | if (DT.dominates(BBE: NoWrapEdge, BB: Result->getParent())) |
| 8253 | continue; |
| 8254 | |
| 8255 | for (const auto &RU : Result->uses()) |
| 8256 | if (!DT.dominates(BBE: NoWrapEdge, U: RU)) |
| 8257 | return false; |
| 8258 | } |
| 8259 | |
| 8260 | return true; |
| 8261 | }; |
| 8262 | |
| 8263 | return llvm::any_of(Range&: GuardingBranches, P: AllUsesGuardedByBranch); |
| 8264 | } |
| 8265 | |
| 8266 | /// Shifts return poison if shiftwidth is larger than the bitwidth. |
| 8267 | static bool shiftAmountKnownInRange(const Value *ShiftAmount) { |
| 8268 | auto *C = dyn_cast<Constant>(Val: ShiftAmount); |
| 8269 | if (!C) |
| 8270 | return false; |
| 8271 | |
| 8272 | // Shifts return poison if shiftwidth is larger than the bitwidth. |
| 8273 | SmallVector<const Constant *, 4> ShiftAmounts; |
| 8274 | if (auto *FVTy = dyn_cast<FixedVectorType>(Val: C->getType())) { |
| 8275 | unsigned NumElts = FVTy->getNumElements(); |
| 8276 | for (unsigned i = 0; i < NumElts; ++i) |
| 8277 | ShiftAmounts.push_back(Elt: C->getAggregateElement(Elt: i)); |
| 8278 | } else if (isa<ScalableVectorType>(Val: C->getType())) |
| 8279 | return false; // Can't tell, just return false to be safe |
| 8280 | else |
| 8281 | ShiftAmounts.push_back(Elt: C); |
| 8282 | |
| 8283 | bool Safe = llvm::all_of(Range&: ShiftAmounts, P: [](const Constant *C) { |
| 8284 | auto *CI = dyn_cast_or_null<ConstantInt>(Val: C); |
| 8285 | return CI && CI->getValue().ult(RHS: C->getType()->getIntegerBitWidth()); |
| 8286 | }); |
| 8287 | |
| 8288 | return Safe; |
| 8289 | } |
| 8290 | |
| 8291 | static bool canCreateUndefOrPoison(const Operator *Op, UndefPoisonKind Kind, |
| 8292 | bool ConsiderFlagsAndMetadata) { |
| 8293 | const auto *CB = dyn_cast<CallBase>(Val: Op); |
| 8294 | bool CBHasNoUndef = CB && CB->hasRetAttr(Kind: Attribute::NoUndef); |
| 8295 | if (ConsiderFlagsAndMetadata && includesPoison(Kind) && |
| 8296 | Op->hasPoisonGeneratingAnnotations() && !CBHasNoUndef) |
| 8297 | return true; |
| 8298 | |
| 8299 | unsigned Opcode = Op->getOpcode(); |
| 8300 | |
| 8301 | // Check whether opcode is a poison/undef-generating operation |
| 8302 | switch (Opcode) { |
| 8303 | case Instruction::Shl: |
| 8304 | case Instruction::AShr: |
| 8305 | case Instruction::LShr: |
| 8306 | return includesPoison(Kind) && !shiftAmountKnownInRange(ShiftAmount: Op->getOperand(i: 1)); |
| 8307 | case Instruction::FPToSI: |
| 8308 | case Instruction::FPToUI: |
| 8309 | // fptosi/ui yields poison if the resulting value does not fit in the |
| 8310 | // destination type. |
| 8311 | return true; |
| 8312 | case Instruction::Call: |
| 8313 | if (auto *II = dyn_cast<IntrinsicInst>(Val: Op)) { |
| 8314 | switch (II->getIntrinsicID()) { |
| 8315 | // NOTE: Use IntrNoCreateUndefOrPoison when possible. |
| 8316 | case Intrinsic::ctlz: |
| 8317 | case Intrinsic::cttz: |
| 8318 | case Intrinsic::abs: |
| 8319 | // We're not considering flags so it is safe to just return false. |
| 8320 | return false; |
| 8321 | case Intrinsic::sshl_sat: |
| 8322 | case Intrinsic::ushl_sat: |
| 8323 | if (!includesPoison(Kind) || |
| 8324 | shiftAmountKnownInRange(ShiftAmount: II->getArgOperand(i: 1))) |
| 8325 | return false; |
| 8326 | break; |
| 8327 | } |
| 8328 | } |
| 8329 | [[fallthrough]]; |
| 8330 | case Instruction::CallBr: |
| 8331 | case Instruction::Invoke: { |
| 8332 | return !CBHasNoUndef && !CB->hasFnAttr(Kind: Attribute::NoCreateUndefOrPoison); |
| 8333 | } |
| 8334 | case Instruction::InsertElement: |
| 8335 | case Instruction::ExtractElement: { |
| 8336 | // If index exceeds the length of the vector, it returns poison |
| 8337 | auto *VTy = cast<VectorType>(Val: Op->getOperand(i: 0)->getType()); |
| 8338 | unsigned IdxOp = Op->getOpcode() == Instruction::InsertElement ? 2 : 1; |
| 8339 | auto *Idx = dyn_cast<ConstantInt>(Val: Op->getOperand(i: IdxOp)); |
| 8340 | if (includesPoison(Kind)) |
| 8341 | return !Idx || |
| 8342 | Idx->getValue().uge(RHS: VTy->getElementCount().getKnownMinValue()); |
| 8343 | return false; |
| 8344 | } |
| 8345 | case Instruction::ShuffleVector: { |
| 8346 | ArrayRef<int> Mask = isa<ConstantExpr>(Val: Op) |
| 8347 | ? cast<ConstantExpr>(Val: Op)->getShuffleMask() |
| 8348 | : cast<ShuffleVectorInst>(Val: Op)->getShuffleMask(); |
| 8349 | return includesPoison(Kind) && is_contained(Range&: Mask, Element: PoisonMaskElem); |
| 8350 | } |
| 8351 | case Instruction::FNeg: |
| 8352 | case Instruction::PHI: |
| 8353 | case Instruction::Select: |
| 8354 | case Instruction::ExtractValue: |
| 8355 | case Instruction::InsertValue: |
| 8356 | case Instruction::Freeze: |
| 8357 | case Instruction::ICmp: |
| 8358 | case Instruction::FCmp: |
| 8359 | case Instruction::GetElementPtr: |
| 8360 | return false; |
| 8361 | case Instruction::AddrSpaceCast: |
| 8362 | return true; |
| 8363 | default: { |
| 8364 | const auto *CE = dyn_cast<ConstantExpr>(Val: Op); |
| 8365 | if (isa<CastInst>(Val: Op) || (CE && CE->isCast())) |
| 8366 | return false; |
| 8367 | else if (Instruction::isBinaryOp(Opcode)) |
| 8368 | return false; |
| 8369 | // Be conservative and return true. |
| 8370 | return true; |
| 8371 | } |
| 8372 | } |
| 8373 | } |
| 8374 | |
| 8375 | bool llvm::canCreateUndefOrPoison(const Operator *Op, |
| 8376 | bool ConsiderFlagsAndMetadata) { |
| 8377 | return ::canCreateUndefOrPoison(Op, Kind: UndefPoisonKind::UndefOrPoison, |
| 8378 | ConsiderFlagsAndMetadata); |
| 8379 | } |
| 8380 | |
| 8381 | bool llvm::canCreatePoison(const Operator *Op, bool ConsiderFlagsAndMetadata) { |
| 8382 | return ::canCreateUndefOrPoison(Op, Kind: UndefPoisonKind::PoisonOnly, |
| 8383 | ConsiderFlagsAndMetadata); |
| 8384 | } |
| 8385 | |
| 8386 | static bool directlyImpliesPoison(const Value *ValAssumedPoison, const Value *V, |
| 8387 | unsigned Depth) { |
| 8388 | if (ValAssumedPoison == V) |
| 8389 | return true; |
| 8390 | |
| 8391 | const unsigned MaxDepth = 2; |
| 8392 | if (Depth >= MaxDepth) |
| 8393 | return false; |
| 8394 | |
| 8395 | if (const auto *I = dyn_cast<Instruction>(Val: V)) { |
| 8396 | if (any_of(Range: I->operands(), P: [=](const Use &Op) { |
| 8397 | return propagatesPoison(PoisonOp: Op) && |
| 8398 | directlyImpliesPoison(ValAssumedPoison, V: Op, Depth: Depth + 1); |
| 8399 | })) |
| 8400 | return true; |
| 8401 | |
| 8402 | // V = extractvalue V0, idx |
| 8403 | // V2 = extractvalue V0, idx2 |
| 8404 | // V0's elements are all poison or not. (e.g., add_with_overflow) |
| 8405 | const WithOverflowInst *II; |
| 8406 | if (match(V: I, P: m_ExtractValue(V: m_WithOverflowInst(I&: II))) && |
| 8407 | (match(V: ValAssumedPoison, P: m_ExtractValue(V: m_Specific(V: II))) || |
| 8408 | llvm::is_contained(Range: II->args(), Element: ValAssumedPoison))) |
| 8409 | return true; |
| 8410 | } |
| 8411 | return false; |
| 8412 | } |
| 8413 | |
| 8414 | static bool impliesPoison(const Value *ValAssumedPoison, const Value *V, |
| 8415 | unsigned Depth) { |
| 8416 | if (isGuaranteedNotToBePoison(V: ValAssumedPoison)) |
| 8417 | return true; |
| 8418 | |
| 8419 | if (directlyImpliesPoison(ValAssumedPoison, V, /* Depth */ 0)) |
| 8420 | return true; |
| 8421 | |
| 8422 | const unsigned MaxDepth = 2; |
| 8423 | if (Depth >= MaxDepth) |
| 8424 | return false; |
| 8425 | |
| 8426 | const auto *I = dyn_cast<Instruction>(Val: ValAssumedPoison); |
| 8427 | if (I && !canCreatePoison(Op: cast<Operator>(Val: I))) { |
| 8428 | return all_of(Range: I->operands(), P: [=](const Value *Op) { |
| 8429 | return impliesPoison(ValAssumedPoison: Op, V, Depth: Depth + 1); |
| 8430 | }); |
| 8431 | } |
| 8432 | return false; |
| 8433 | } |
| 8434 | |
| 8435 | bool llvm::impliesPoison(const Value *ValAssumedPoison, const Value *V) { |
| 8436 | return ::impliesPoison(ValAssumedPoison, V, /* Depth */ 0); |
| 8437 | } |
| 8438 | |
| 8439 | static bool programUndefinedIfUndefOrPoison(const Value *V, bool PoisonOnly); |
| 8440 | |
| 8441 | static bool isGuaranteedNotToBeUndefOrPoison( |
| 8442 | const Value *V, AssumptionCache *AC, const Instruction *CtxI, |
| 8443 | const DominatorTree *DT, unsigned Depth, UndefPoisonKind Kind) { |
| 8444 | if (Depth >= MaxAnalysisRecursionDepth) |
| 8445 | return false; |
| 8446 | |
| 8447 | if (isa<MetadataAsValue>(Val: V)) |
| 8448 | return false; |
| 8449 | |
| 8450 | if (const auto *A = dyn_cast<Argument>(Val: V)) { |
| 8451 | if (A->hasAttribute(Kind: Attribute::NoUndef) || |
| 8452 | A->hasAttribute(Kind: Attribute::Dereferenceable) || |
| 8453 | A->hasAttribute(Kind: Attribute::DereferenceableOrNull)) |
| 8454 | return true; |
| 8455 | } |
| 8456 | |
| 8457 | if (auto *C = dyn_cast<Constant>(Val: V)) { |
| 8458 | if (isa<PoisonValue>(Val: C)) |
| 8459 | return !includesPoison(Kind); |
| 8460 | |
| 8461 | if (isa<UndefValue>(Val: C)) |
| 8462 | return !includesUndef(Kind); |
| 8463 | |
| 8464 | if (isa<ConstantInt>(Val: C) || isa<GlobalVariable>(Val: C) || isa<ConstantFP>(Val: C) || |
| 8465 | isa<ConstantPointerNull>(Val: C) || isa<Function>(Val: C) || isa<ConstantData>(Val: C)) |
| 8466 | return true; |
| 8467 | |
| 8468 | if (C->getType()->isVectorTy() || C->getType()->isAggregateType()) { |
| 8469 | if (isa<ConstantExpr>(Val: C)) { |
| 8470 | // Scalable vectors can use a ConstantExpr to build a splat. |
| 8471 | if (Constant *SplatC = C->getSplatValue()) |
| 8472 | if (isa<ConstantInt>(Val: SplatC) || isa<ConstantFP>(Val: SplatC)) |
| 8473 | return true; |
| 8474 | } else if (auto *CA = dyn_cast<ConstantAggregate>(Val: C)) { |
| 8475 | return all_of(Range: CA->operands(), P: [&](const Use &Op) { |
| 8476 | return isGuaranteedNotToBeUndefOrPoison(V: Op.get(), AC, CtxI, DT, |
| 8477 | Depth: Depth + 1, Kind); |
| 8478 | }); |
| 8479 | } |
| 8480 | } |
| 8481 | } |
| 8482 | |
| 8483 | // Strip cast operations from a pointer value. |
| 8484 | // Note that stripPointerCastsSameRepresentation can strip off getelementptr |
| 8485 | // inbounds with zero offset. To guarantee that the result isn't poison, the |
| 8486 | // stripped pointer is checked as it has to be pointing into an allocated |
| 8487 | // object or be null `null` to ensure `inbounds` getelement pointers with a |
| 8488 | // zero offset could not produce poison. |
| 8489 | // It can strip off addrspacecast that do not change bit representation as |
| 8490 | // well. We believe that such addrspacecast is equivalent to no-op. |
| 8491 | auto *StrippedV = V->stripPointerCastsSameRepresentation(); |
| 8492 | if (isa<AllocaInst>(Val: StrippedV) || isa<GlobalVariable>(Val: StrippedV) || |
| 8493 | isa<Function>(Val: StrippedV) || isa<ConstantPointerNull>(Val: StrippedV)) |
| 8494 | return true; |
| 8495 | |
| 8496 | auto OpCheck = [&](const Value *V) { |
| 8497 | return isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth: Depth + 1, Kind); |
| 8498 | }; |
| 8499 | |
| 8500 | if (auto *Opr = dyn_cast<Operator>(Val: V)) { |
| 8501 | // If the value is a freeze instruction, then it can never |
| 8502 | // be undef or poison. |
| 8503 | if (isa<FreezeInst>(Val: V)) |
| 8504 | return true; |
| 8505 | |
| 8506 | if (const auto *CB = dyn_cast<CallBase>(Val: V)) { |
| 8507 | if (CB->hasRetAttr(Kind: Attribute::NoUndef) || |
| 8508 | CB->hasRetAttr(Kind: Attribute::Dereferenceable) || |
| 8509 | CB->hasRetAttr(Kind: Attribute::DereferenceableOrNull)) |
| 8510 | return true; |
| 8511 | } |
| 8512 | |
| 8513 | if (!::canCreateUndefOrPoison(Op: Opr, Kind, |
| 8514 | /*ConsiderFlagsAndMetadata=*/true)) { |
| 8515 | if (const auto *PN = dyn_cast<PHINode>(Val: V)) { |
| 8516 | unsigned Num = PN->getNumIncomingValues(); |
| 8517 | bool IsWellDefined = true; |
| 8518 | for (unsigned i = 0; i < Num; ++i) { |
| 8519 | if (PN == PN->getIncomingValue(i)) |
| 8520 | continue; |
| 8521 | auto *TI = PN->getIncomingBlock(i)->getTerminator(); |
| 8522 | if (!isGuaranteedNotToBeUndefOrPoison(V: PN->getIncomingValue(i), AC, CtxI: TI, |
| 8523 | DT, Depth: Depth + 1, Kind)) { |
| 8524 | IsWellDefined = false; |
| 8525 | break; |
| 8526 | } |
| 8527 | } |
| 8528 | if (IsWellDefined) |
| 8529 | return true; |
| 8530 | } else if (auto *Splat = isa<ShuffleVectorInst>(Val: Opr) ? getSplatValue(V: Opr) |
| 8531 | : nullptr) { |
| 8532 | // For splats we only need to check the value being splatted. |
| 8533 | if (OpCheck(Splat)) |
| 8534 | return true; |
| 8535 | } else if (all_of(Range: Opr->operands(), P: OpCheck)) |
| 8536 | return true; |
| 8537 | } |
| 8538 | } |
| 8539 | |
| 8540 | if (auto *I = dyn_cast<LoadInst>(Val: V)) |
| 8541 | if (I->hasMetadata(KindID: LLVMContext::MD_noundef) || |
| 8542 | I->hasMetadata(KindID: LLVMContext::MD_dereferenceable) || |
| 8543 | I->hasMetadata(KindID: LLVMContext::MD_dereferenceable_or_null)) |
| 8544 | return true; |
| 8545 | |
| 8546 | if (programUndefinedIfUndefOrPoison(V, PoisonOnly: !includesUndef(Kind))) |
| 8547 | return true; |
| 8548 | |
| 8549 | // CtxI may be null or a cloned instruction. |
| 8550 | if (!CtxI || !CtxI->getParent() || !DT) |
| 8551 | return false; |
| 8552 | |
| 8553 | auto *DNode = DT->getNode(BB: CtxI->getParent()); |
| 8554 | if (!DNode) |
| 8555 | // Unreachable block |
| 8556 | return false; |
| 8557 | |
| 8558 | // If V is used as a branch condition before reaching CtxI, V cannot be |
| 8559 | // undef or poison. |
| 8560 | // br V, BB1, BB2 |
| 8561 | // BB1: |
| 8562 | // CtxI ; V cannot be undef or poison here |
| 8563 | auto *Dominator = DNode->getIDom(); |
| 8564 | // This check is purely for compile time reasons: we can skip the IDom walk |
| 8565 | // if what we are checking for includes undef and the value is not an integer. |
| 8566 | if (!includesUndef(Kind) || V->getType()->isIntegerTy()) |
| 8567 | while (Dominator) { |
| 8568 | auto *TI = Dominator->getBlock()->getTerminatorOrNull(); |
| 8569 | |
| 8570 | Value *Cond = nullptr; |
| 8571 | if (auto BI = dyn_cast_or_null<CondBrInst>(Val: TI)) { |
| 8572 | Cond = BI->getCondition(); |
| 8573 | } else if (auto SI = dyn_cast_or_null<SwitchInst>(Val: TI)) { |
| 8574 | Cond = SI->getCondition(); |
| 8575 | } |
| 8576 | |
| 8577 | if (Cond) { |
| 8578 | if (Cond == V) |
| 8579 | return true; |
| 8580 | else if (!includesUndef(Kind) && isa<Operator>(Val: Cond)) { |
| 8581 | // For poison, we can analyze further |
| 8582 | auto *Opr = cast<Operator>(Val: Cond); |
| 8583 | if (any_of(Range: Opr->operands(), P: [V](const Use &U) { |
| 8584 | return V == U && propagatesPoison(PoisonOp: U); |
| 8585 | })) |
| 8586 | return true; |
| 8587 | } |
| 8588 | } |
| 8589 | |
| 8590 | Dominator = Dominator->getIDom(); |
| 8591 | } |
| 8592 | |
| 8593 | if (AC && getKnowledgeValidInContext(V, AttrKinds: {Attribute::NoUndef}, AC&: *AC, CtxI, DT)) |
| 8594 | return true; |
| 8595 | |
| 8596 | return false; |
| 8597 | } |
| 8598 | |
| 8599 | bool llvm::isGuaranteedNotToBeUndefOrPoison(const Value *V, AssumptionCache *AC, |
| 8600 | const Instruction *CtxI, |
| 8601 | const DominatorTree *DT, |
| 8602 | unsigned Depth) { |
| 8603 | return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, |
| 8604 | Kind: UndefPoisonKind::UndefOrPoison); |
| 8605 | } |
| 8606 | |
| 8607 | bool llvm::isGuaranteedNotToBePoison(const Value *V, AssumptionCache *AC, |
| 8608 | const Instruction *CtxI, |
| 8609 | const DominatorTree *DT, unsigned Depth) { |
| 8610 | return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, |
| 8611 | Kind: UndefPoisonKind::PoisonOnly); |
| 8612 | } |
| 8613 | |
| 8614 | bool llvm::isGuaranteedNotToBeUndef(const Value *V, AssumptionCache *AC, |
| 8615 | const Instruction *CtxI, |
| 8616 | const DominatorTree *DT, unsigned Depth) { |
| 8617 | return ::isGuaranteedNotToBeUndefOrPoison(V, AC, CtxI, DT, Depth, |
| 8618 | Kind: UndefPoisonKind::UndefOnly); |
| 8619 | } |
| 8620 | |
| 8621 | /// Return true if undefined behavior would provably be executed on the path to |
| 8622 | /// OnPathTo if Root produced a posion result. Note that this doesn't say |
| 8623 | /// anything about whether OnPathTo is actually executed or whether Root is |
| 8624 | /// actually poison. This can be used to assess whether a new use of Root can |
| 8625 | /// be added at a location which is control equivalent with OnPathTo (such as |
| 8626 | /// immediately before it) without introducing UB which didn't previously |
| 8627 | /// exist. Note that a false result conveys no information. |
| 8628 | bool llvm::mustExecuteUBIfPoisonOnPathTo(Instruction *Root, |
| 8629 | Instruction *OnPathTo, |
| 8630 | DominatorTree *DT) { |
| 8631 | // Basic approach is to assume Root is poison, propagate poison forward |
| 8632 | // through all users we can easily track, and then check whether any of those |
| 8633 | // users are provable UB and must execute before out exiting block might |
| 8634 | // exit. |
| 8635 | |
| 8636 | // The set of all recursive users we've visited (which are assumed to all be |
| 8637 | // poison because of said visit) |
| 8638 | SmallPtrSet<const Value *, 16> KnownPoison; |
| 8639 | SmallVector<const Instruction*, 16> Worklist; |
| 8640 | Worklist.push_back(Elt: Root); |
| 8641 | while (!Worklist.empty()) { |
| 8642 | const Instruction *I = Worklist.pop_back_val(); |
| 8643 | |
| 8644 | // If we know this must trigger UB on a path leading our target. |
| 8645 | if (mustTriggerUB(I, KnownPoison) && DT->dominates(Def: I, User: OnPathTo)) |
| 8646 | return true; |
| 8647 | |
| 8648 | // If we can't analyze propagation through this instruction, just skip it |
| 8649 | // and transitive users. Safe as false is a conservative result. |
| 8650 | if (I != Root && !any_of(Range: I->operands(), P: [&KnownPoison](const Use &U) { |
| 8651 | return KnownPoison.contains(Ptr: U) && propagatesPoison(PoisonOp: U); |
| 8652 | })) |
| 8653 | continue; |
| 8654 | |
| 8655 | if (KnownPoison.insert(Ptr: I).second) |
| 8656 | for (const User *User : I->users()) |
| 8657 | Worklist.push_back(Elt: cast<Instruction>(Val: User)); |
| 8658 | } |
| 8659 | |
| 8660 | // Might be non-UB, or might have a path we couldn't prove must execute on |
| 8661 | // way to exiting bb. |
| 8662 | return false; |
| 8663 | } |
| 8664 | |
| 8665 | OverflowResult llvm::computeOverflowForSignedAdd(const AddOperator *Add, |
| 8666 | const SimplifyQuery &SQ) { |
| 8667 | return ::computeOverflowForSignedAdd(LHS: Add->getOperand(i_nocapture: 0), RHS: Add->getOperand(i_nocapture: 1), |
| 8668 | Add, SQ); |
| 8669 | } |
| 8670 | |
| 8671 | OverflowResult |
| 8672 | llvm::computeOverflowForSignedAdd(const WithCache<const Value *> &LHS, |
| 8673 | const WithCache<const Value *> &RHS, |
| 8674 | const SimplifyQuery &SQ) { |
| 8675 | return ::computeOverflowForSignedAdd(LHS, RHS, Add: nullptr, SQ); |
| 8676 | } |
| 8677 | |
| 8678 | bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) { |
| 8679 | // Note: An atomic operation isn't guaranteed to return in a reasonable amount |
| 8680 | // of time because it's possible for another thread to interfere with it for an |
| 8681 | // arbitrary length of time, but programs aren't allowed to rely on that. |
| 8682 | |
| 8683 | // If there is no successor, then execution can't transfer to it. |
| 8684 | if (isa<ReturnInst>(Val: I)) |
| 8685 | return false; |
| 8686 | if (isa<UnreachableInst>(Val: I)) |
| 8687 | return false; |
| 8688 | |
| 8689 | // Note: Do not add new checks here; instead, change Instruction::mayThrow or |
| 8690 | // Instruction::willReturn. |
| 8691 | // |
| 8692 | // FIXME: Move this check into Instruction::willReturn. |
| 8693 | if (isa<CatchPadInst>(Val: I)) { |
| 8694 | switch (classifyEHPersonality(Pers: I->getFunction()->getPersonalityFn())) { |
| 8695 | default: |
| 8696 | // A catchpad may invoke exception object constructors and such, which |
| 8697 | // in some languages can be arbitrary code, so be conservative by default. |
| 8698 | return false; |
| 8699 | case EHPersonality::CoreCLR: |
| 8700 | // For CoreCLR, it just involves a type test. |
| 8701 | return true; |
| 8702 | } |
| 8703 | } |
| 8704 | |
| 8705 | // An instruction that returns without throwing must transfer control flow |
| 8706 | // to a successor. |
| 8707 | return !I->mayThrow() && I->willReturn(); |
| 8708 | } |
| 8709 | |
| 8710 | bool llvm::isGuaranteedToTransferExecutionToSuccessor(const BasicBlock *BB) { |
| 8711 | // TODO: This is slightly conservative for invoke instruction since exiting |
| 8712 | // via an exception *is* normal control for them. |
| 8713 | for (const Instruction &I : *BB) |
| 8714 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &I)) |
| 8715 | return false; |
| 8716 | return true; |
| 8717 | } |
| 8718 | |
| 8719 | bool llvm::isGuaranteedToTransferExecutionToSuccessor( |
| 8720 | BasicBlock::const_iterator Begin, BasicBlock::const_iterator End, |
| 8721 | unsigned ScanLimit) { |
| 8722 | return isGuaranteedToTransferExecutionToSuccessor(Range: make_range(x: Begin, y: End), |
| 8723 | ScanLimit); |
| 8724 | } |
| 8725 | |
| 8726 | bool llvm::isGuaranteedToTransferExecutionToSuccessor( |
| 8727 | iterator_range<BasicBlock::const_iterator> Range, unsigned ScanLimit) { |
| 8728 | assert(ScanLimit && "scan limit must be non-zero" ); |
| 8729 | for (const Instruction &I : Range) { |
| 8730 | if (--ScanLimit == 0) |
| 8731 | return false; |
| 8732 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &I)) |
| 8733 | return false; |
| 8734 | } |
| 8735 | return true; |
| 8736 | } |
| 8737 | |
| 8738 | bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I, |
| 8739 | const Loop *L) { |
| 8740 | // The loop header is guaranteed to be executed for every iteration. |
| 8741 | // |
| 8742 | // FIXME: Relax this constraint to cover all basic blocks that are |
| 8743 | // guaranteed to be executed at every iteration. |
| 8744 | if (I->getParent() != L->getHeader()) return false; |
| 8745 | |
| 8746 | for (const Instruction &LI : *L->getHeader()) { |
| 8747 | if (&LI == I) return true; |
| 8748 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &LI)) return false; |
| 8749 | } |
| 8750 | llvm_unreachable("Instruction not contained in its own parent basic block." ); |
| 8751 | } |
| 8752 | |
| 8753 | bool llvm::intrinsicPropagatesPoison(Intrinsic::ID IID) { |
| 8754 | switch (IID) { |
| 8755 | // TODO: Add more intrinsics. |
| 8756 | case Intrinsic::sadd_with_overflow: |
| 8757 | case Intrinsic::ssub_with_overflow: |
| 8758 | case Intrinsic::smul_with_overflow: |
| 8759 | case Intrinsic::uadd_with_overflow: |
| 8760 | case Intrinsic::usub_with_overflow: |
| 8761 | case Intrinsic::umul_with_overflow: |
| 8762 | // If an input is a vector containing a poison element, the |
| 8763 | // two output vectors (calculated results, overflow bits)' |
| 8764 | // corresponding lanes are poison. |
| 8765 | return true; |
| 8766 | case Intrinsic::ctpop: |
| 8767 | case Intrinsic::ctlz: |
| 8768 | case Intrinsic::cttz: |
| 8769 | case Intrinsic::abs: |
| 8770 | case Intrinsic::smax: |
| 8771 | case Intrinsic::smin: |
| 8772 | case Intrinsic::umax: |
| 8773 | case Intrinsic::umin: |
| 8774 | case Intrinsic::scmp: |
| 8775 | case Intrinsic::smulh: |
| 8776 | case Intrinsic::umulh: |
| 8777 | case Intrinsic::is_fpclass: |
| 8778 | case Intrinsic::ptrmask: |
| 8779 | case Intrinsic::ucmp: |
| 8780 | case Intrinsic::bitreverse: |
| 8781 | case Intrinsic::bswap: |
| 8782 | case Intrinsic::sadd_sat: |
| 8783 | case Intrinsic::ssub_sat: |
| 8784 | case Intrinsic::sshl_sat: |
| 8785 | case Intrinsic::uadd_sat: |
| 8786 | case Intrinsic::usub_sat: |
| 8787 | case Intrinsic::ushl_sat: |
| 8788 | case Intrinsic::smul_fix: |
| 8789 | case Intrinsic::smul_fix_sat: |
| 8790 | case Intrinsic::umul_fix: |
| 8791 | case Intrinsic::umul_fix_sat: |
| 8792 | case Intrinsic::pow: |
| 8793 | case Intrinsic::experimental_constrained_pow: |
| 8794 | case Intrinsic::powi: |
| 8795 | case Intrinsic::experimental_constrained_powi: |
| 8796 | case Intrinsic::sin: |
| 8797 | case Intrinsic::experimental_constrained_sin: |
| 8798 | case Intrinsic::sinh: |
| 8799 | case Intrinsic::experimental_constrained_sinh: |
| 8800 | case Intrinsic::cos: |
| 8801 | case Intrinsic::experimental_constrained_cos: |
| 8802 | case Intrinsic::cosh: |
| 8803 | case Intrinsic::experimental_constrained_cosh: |
| 8804 | case Intrinsic::sincos: |
| 8805 | case Intrinsic::sincospi: |
| 8806 | case Intrinsic::tan: |
| 8807 | case Intrinsic::experimental_constrained_tan: |
| 8808 | case Intrinsic::tanh: |
| 8809 | case Intrinsic::experimental_constrained_tanh: |
| 8810 | case Intrinsic::asin: |
| 8811 | case Intrinsic::experimental_constrained_asin: |
| 8812 | case Intrinsic::acos: |
| 8813 | case Intrinsic::experimental_constrained_acos: |
| 8814 | case Intrinsic::atan: |
| 8815 | case Intrinsic::experimental_constrained_atan: |
| 8816 | case Intrinsic::atan2: |
| 8817 | case Intrinsic::experimental_constrained_atan2: |
| 8818 | case Intrinsic::canonicalize: |
| 8819 | case Intrinsic::sqrt: |
| 8820 | case Intrinsic::experimental_constrained_sqrt: |
| 8821 | case Intrinsic::fma: |
| 8822 | case Intrinsic::experimental_constrained_fma: |
| 8823 | case Intrinsic::fmuladd: |
| 8824 | case Intrinsic::experimental_constrained_fmuladd: |
| 8825 | case Intrinsic::exp: |
| 8826 | case Intrinsic::experimental_constrained_exp: |
| 8827 | case Intrinsic::exp2: |
| 8828 | case Intrinsic::experimental_constrained_exp2: |
| 8829 | case Intrinsic::exp10: |
| 8830 | case Intrinsic::log: |
| 8831 | case Intrinsic::experimental_constrained_log: |
| 8832 | case Intrinsic::log2: |
| 8833 | case Intrinsic::experimental_constrained_log2: |
| 8834 | case Intrinsic::log10: |
| 8835 | case Intrinsic::experimental_constrained_log10: |
| 8836 | case Intrinsic::modf: |
| 8837 | case Intrinsic::floor: |
| 8838 | case Intrinsic::experimental_constrained_floor: |
| 8839 | case Intrinsic::ceil: |
| 8840 | case Intrinsic::experimental_constrained_ceil: |
| 8841 | case Intrinsic::trunc: |
| 8842 | case Intrinsic::experimental_constrained_trunc: |
| 8843 | case Intrinsic::rint: |
| 8844 | case Intrinsic::experimental_constrained_rint: |
| 8845 | case Intrinsic::nearbyint: |
| 8846 | case Intrinsic::experimental_constrained_nearbyint: |
| 8847 | case Intrinsic::round: |
| 8848 | case Intrinsic::experimental_constrained_round: |
| 8849 | case Intrinsic::roundeven: |
| 8850 | case Intrinsic::experimental_constrained_roundeven: |
| 8851 | case Intrinsic::lrint: |
| 8852 | case Intrinsic::experimental_constrained_lrint: |
| 8853 | case Intrinsic::llrint: |
| 8854 | case Intrinsic::experimental_constrained_llrint: |
| 8855 | case Intrinsic::lround: |
| 8856 | case Intrinsic::experimental_constrained_lround: |
| 8857 | case Intrinsic::llround: |
| 8858 | case Intrinsic::experimental_constrained_llround: |
| 8859 | case Intrinsic::fshl: |
| 8860 | case Intrinsic::fshr: |
| 8861 | case Intrinsic::frexp: |
| 8862 | case Intrinsic::ldexp: |
| 8863 | case Intrinsic::experimental_constrained_ldexp: |
| 8864 | case Intrinsic::get_active_lane_mask: |
| 8865 | case Intrinsic::mask_beforefirst: |
| 8866 | return true; |
| 8867 | default: |
| 8868 | return false; |
| 8869 | } |
| 8870 | } |
| 8871 | |
| 8872 | bool llvm::propagatesPoison(const Use &PoisonOp) { |
| 8873 | const Operator *I = cast<Operator>(Val: PoisonOp.getUser()); |
| 8874 | switch (I->getOpcode()) { |
| 8875 | case Instruction::Freeze: |
| 8876 | case Instruction::PHI: |
| 8877 | case Instruction::Invoke: |
| 8878 | return false; |
| 8879 | case Instruction::Select: |
| 8880 | return PoisonOp.getOperandNo() == 0; |
| 8881 | case Instruction::Call: |
| 8882 | if (auto *II = dyn_cast<IntrinsicInst>(Val: I)) |
| 8883 | return intrinsicPropagatesPoison(IID: II->getIntrinsicID()); |
| 8884 | return false; |
| 8885 | case Instruction::ICmp: |
| 8886 | case Instruction::FCmp: |
| 8887 | case Instruction::GetElementPtr: |
| 8888 | return true; |
| 8889 | default: |
| 8890 | if (isa<BinaryOperator>(Val: I) || isa<UnaryOperator>(Val: I) || isa<CastInst>(Val: I)) |
| 8891 | return true; |
| 8892 | |
| 8893 | // Be conservative and return false. |
| 8894 | return false; |
| 8895 | } |
| 8896 | } |
| 8897 | |
| 8898 | /// Enumerates all operands of \p I that are guaranteed to not be undef or |
| 8899 | /// poison. If the callback \p Handle returns true, stop processing and return |
| 8900 | /// true. Otherwise, return false. |
| 8901 | template <typename CallableT> |
| 8902 | static bool handleGuaranteedWellDefinedOps(const Instruction *I, |
| 8903 | const CallableT &Handle) { |
| 8904 | switch (I->getOpcode()) { |
| 8905 | case Instruction::Store: |
| 8906 | if (Handle(cast<StoreInst>(Val: I)->getPointerOperand())) |
| 8907 | return true; |
| 8908 | break; |
| 8909 | |
| 8910 | case Instruction::Load: |
| 8911 | if (Handle(cast<LoadInst>(Val: I)->getPointerOperand())) |
| 8912 | return true; |
| 8913 | break; |
| 8914 | |
| 8915 | // Since dereferenceable attribute imply noundef, atomic operations |
| 8916 | // also implicitly have noundef pointers too |
| 8917 | case Instruction::AtomicCmpXchg: |
| 8918 | if (Handle(cast<AtomicCmpXchgInst>(Val: I)->getPointerOperand())) |
| 8919 | return true; |
| 8920 | break; |
| 8921 | |
| 8922 | case Instruction::AtomicRMW: |
| 8923 | if (Handle(cast<AtomicRMWInst>(Val: I)->getPointerOperand())) |
| 8924 | return true; |
| 8925 | break; |
| 8926 | |
| 8927 | case Instruction::Call: |
| 8928 | case Instruction::Invoke: { |
| 8929 | const CallBase *CB = cast<CallBase>(Val: I); |
| 8930 | if (CB->isIndirectCall() && Handle(CB->getCalledOperand())) |
| 8931 | return true; |
| 8932 | for (unsigned i = 0; i < CB->arg_size(); ++i) |
| 8933 | if ((CB->paramHasAttr(ArgNo: i, Kind: Attribute::NoUndef) || |
| 8934 | CB->paramHasAttr(ArgNo: i, Kind: Attribute::Dereferenceable) || |
| 8935 | CB->paramHasAttr(ArgNo: i, Kind: Attribute::DereferenceableOrNull)) && |
| 8936 | Handle(CB->getArgOperand(i))) |
| 8937 | return true; |
| 8938 | break; |
| 8939 | } |
| 8940 | case Instruction::Ret: |
| 8941 | if (I->getFunction()->hasRetAttribute(Kind: Attribute::NoUndef) && |
| 8942 | Handle(I->getOperand(i: 0))) |
| 8943 | return true; |
| 8944 | break; |
| 8945 | case Instruction::Switch: |
| 8946 | if (Handle(cast<SwitchInst>(Val: I)->getCondition())) |
| 8947 | return true; |
| 8948 | break; |
| 8949 | case Instruction::CondBr: |
| 8950 | if (Handle(cast<CondBrInst>(Val: I)->getCondition())) |
| 8951 | return true; |
| 8952 | break; |
| 8953 | default: |
| 8954 | break; |
| 8955 | } |
| 8956 | |
| 8957 | return false; |
| 8958 | } |
| 8959 | |
| 8960 | /// Enumerates all operands of \p I that are guaranteed to not be poison. |
| 8961 | template <typename CallableT> |
| 8962 | static bool handleGuaranteedNonPoisonOps(const Instruction *I, |
| 8963 | const CallableT &Handle) { |
| 8964 | if (handleGuaranteedWellDefinedOps(I, Handle)) |
| 8965 | return true; |
| 8966 | switch (I->getOpcode()) { |
| 8967 | // Divisors of these operations are allowed to be partially undef. |
| 8968 | case Instruction::UDiv: |
| 8969 | case Instruction::SDiv: |
| 8970 | case Instruction::URem: |
| 8971 | case Instruction::SRem: |
| 8972 | return Handle(I->getOperand(i: 1)); |
| 8973 | default: |
| 8974 | return false; |
| 8975 | } |
| 8976 | } |
| 8977 | |
| 8978 | bool llvm::mustTriggerUB(const Instruction *I, |
| 8979 | const SmallPtrSetImpl<const Value *> &KnownPoison) { |
| 8980 | return handleGuaranteedNonPoisonOps( |
| 8981 | I, Handle: [&](const Value *V) { return KnownPoison.count(Ptr: V); }); |
| 8982 | } |
| 8983 | |
| 8984 | static bool programUndefinedIfUndefOrPoison(const Value *V, |
| 8985 | bool PoisonOnly) { |
| 8986 | // We currently only look for uses of values within the same basic |
| 8987 | // block, as that makes it easier to guarantee that the uses will be |
| 8988 | // executed given that Inst is executed. |
| 8989 | // |
| 8990 | // FIXME: Expand this to consider uses beyond the same basic block. To do |
| 8991 | // this, look out for the distinction between post-dominance and strong |
| 8992 | // post-dominance. |
| 8993 | const BasicBlock *BB = nullptr; |
| 8994 | BasicBlock::const_iterator Begin; |
| 8995 | if (const auto *Inst = dyn_cast<Instruction>(Val: V)) { |
| 8996 | BB = Inst->getParent(); |
| 8997 | Begin = Inst->getIterator(); |
| 8998 | Begin++; |
| 8999 | } else if (const auto *Arg = dyn_cast<Argument>(Val: V)) { |
| 9000 | if (Arg->getParent()->isDeclaration()) |
| 9001 | return false; |
| 9002 | BB = &Arg->getParent()->getEntryBlock(); |
| 9003 | Begin = BB->begin(); |
| 9004 | } else { |
| 9005 | return false; |
| 9006 | } |
| 9007 | |
| 9008 | // Limit number of instructions we look at, to avoid scanning through large |
| 9009 | // blocks. The current limit is chosen arbitrarily. |
| 9010 | unsigned ScanLimit = 32; |
| 9011 | BasicBlock::const_iterator End = BB->end(); |
| 9012 | |
| 9013 | if (!PoisonOnly) { |
| 9014 | // Since undef does not propagate eagerly, be conservative & just check |
| 9015 | // whether a value is directly passed to an instruction that must take |
| 9016 | // well-defined operands. |
| 9017 | |
| 9018 | for (const auto &I : make_range(x: Begin, y: End)) { |
| 9019 | if (--ScanLimit == 0) |
| 9020 | break; |
| 9021 | |
| 9022 | if (handleGuaranteedWellDefinedOps(I: &I, Handle: [V](const Value *WellDefinedOp) { |
| 9023 | return WellDefinedOp == V; |
| 9024 | })) |
| 9025 | return true; |
| 9026 | |
| 9027 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &I)) |
| 9028 | break; |
| 9029 | } |
| 9030 | return false; |
| 9031 | } |
| 9032 | |
| 9033 | // Set of instructions that we have proved will yield poison if Inst |
| 9034 | // does. |
| 9035 | SmallPtrSet<const Value *, 16> YieldsPoison; |
| 9036 | SmallPtrSet<const BasicBlock *, 4> Visited; |
| 9037 | |
| 9038 | YieldsPoison.insert(Ptr: V); |
| 9039 | Visited.insert(Ptr: BB); |
| 9040 | |
| 9041 | while (true) { |
| 9042 | for (const auto &I : make_range(x: Begin, y: End)) { |
| 9043 | if (--ScanLimit == 0) |
| 9044 | return false; |
| 9045 | if (mustTriggerUB(I: &I, KnownPoison: YieldsPoison)) |
| 9046 | return true; |
| 9047 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &I)) |
| 9048 | return false; |
| 9049 | |
| 9050 | // If an operand is poison and propagates it, mark I as yielding poison. |
| 9051 | for (const Use &Op : I.operands()) { |
| 9052 | if (YieldsPoison.count(Ptr: Op) && propagatesPoison(PoisonOp: Op)) { |
| 9053 | YieldsPoison.insert(Ptr: &I); |
| 9054 | break; |
| 9055 | } |
| 9056 | } |
| 9057 | |
| 9058 | // Special handling for select, which returns poison if its operand 0 is |
| 9059 | // poison (handled in the loop above) *or* if both its true/false operands |
| 9060 | // are poison (handled here). |
| 9061 | if (I.getOpcode() == Instruction::Select && |
| 9062 | YieldsPoison.count(Ptr: I.getOperand(i: 1)) && |
| 9063 | YieldsPoison.count(Ptr: I.getOperand(i: 2))) { |
| 9064 | YieldsPoison.insert(Ptr: &I); |
| 9065 | } |
| 9066 | } |
| 9067 | |
| 9068 | BB = BB->getSingleSuccessor(); |
| 9069 | if (!BB || !Visited.insert(Ptr: BB).second) |
| 9070 | break; |
| 9071 | |
| 9072 | Begin = BB->getFirstNonPHIIt(); |
| 9073 | End = BB->end(); |
| 9074 | } |
| 9075 | return false; |
| 9076 | } |
| 9077 | |
| 9078 | bool llvm::programUndefinedIfUndefOrPoison(const Instruction *Inst) { |
| 9079 | return ::programUndefinedIfUndefOrPoison(V: Inst, PoisonOnly: false); |
| 9080 | } |
| 9081 | |
| 9082 | bool llvm::programUndefinedIfPoison(const Instruction *Inst) { |
| 9083 | return ::programUndefinedIfUndefOrPoison(V: Inst, PoisonOnly: true); |
| 9084 | } |
| 9085 | |
| 9086 | static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) { |
| 9087 | if (FMF.noNaNs()) |
| 9088 | return true; |
| 9089 | |
| 9090 | if (auto *C = dyn_cast<ConstantFP>(Val: V)) |
| 9091 | return !C->isNaN(); |
| 9092 | |
| 9093 | if (auto *C = dyn_cast<ConstantDataVector>(Val: V)) { |
| 9094 | if (!C->getElementType()->isFloatingPointTy()) |
| 9095 | return false; |
| 9096 | for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) { |
| 9097 | if (C->getElementAsAPFloat(i: I).isNaN()) |
| 9098 | return false; |
| 9099 | } |
| 9100 | return true; |
| 9101 | } |
| 9102 | |
| 9103 | if (isa<ConstantAggregateZero>(Val: V)) |
| 9104 | return true; |
| 9105 | |
| 9106 | return false; |
| 9107 | } |
| 9108 | |
| 9109 | static bool isKnownNonZero(const Value *V) { |
| 9110 | if (auto *C = dyn_cast<ConstantFP>(Val: V)) |
| 9111 | return !C->isZero(); |
| 9112 | |
| 9113 | if (auto *C = dyn_cast<ConstantDataVector>(Val: V)) { |
| 9114 | if (!C->getElementType()->isFloatingPointTy()) |
| 9115 | return false; |
| 9116 | for (unsigned I = 0, E = C->getNumElements(); I < E; ++I) { |
| 9117 | if (C->getElementAsAPFloat(i: I).isZero()) |
| 9118 | return false; |
| 9119 | } |
| 9120 | return true; |
| 9121 | } |
| 9122 | |
| 9123 | return false; |
| 9124 | } |
| 9125 | |
| 9126 | /// Match clamp pattern for float types without care about NaNs or signed zeros. |
| 9127 | /// Given non-min/max outer cmp/select from the clamp pattern this |
| 9128 | /// function recognizes if it can be substitued by a "canonical" min/max |
| 9129 | /// pattern. |
| 9130 | static SelectPatternResult matchFastFloatClamp(CmpInst::Predicate Pred, |
| 9131 | Value *CmpLHS, Value *CmpRHS, |
| 9132 | Value *TrueVal, Value *FalseVal, |
| 9133 | Value *&LHS, Value *&RHS) { |
| 9134 | // Try to match |
| 9135 | // X < C1 ? C1 : Min(X, C2) --> Max(C1, Min(X, C2)) |
| 9136 | // X > C1 ? C1 : Max(X, C2) --> Min(C1, Max(X, C2)) |
| 9137 | // and return description of the outer Max/Min. |
| 9138 | |
| 9139 | // First, check if select has inverse order: |
| 9140 | if (CmpRHS == FalseVal) { |
| 9141 | std::swap(a&: TrueVal, b&: FalseVal); |
| 9142 | Pred = CmpInst::getInversePredicate(pred: Pred); |
| 9143 | } |
| 9144 | |
| 9145 | // Assume success now. If there's no match, callers should not use these anyway. |
| 9146 | LHS = TrueVal; |
| 9147 | RHS = FalseVal; |
| 9148 | |
| 9149 | const APFloat *FC1; |
| 9150 | if (CmpRHS != TrueVal || !match(V: CmpRHS, P: m_APFloat(Res&: FC1)) || !FC1->isFinite()) |
| 9151 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9152 | |
| 9153 | const APFloat *FC2; |
| 9154 | switch (Pred) { |
| 9155 | case CmpInst::FCMP_OLT: |
| 9156 | case CmpInst::FCMP_OLE: |
| 9157 | case CmpInst::FCMP_ULT: |
| 9158 | case CmpInst::FCMP_ULE: |
| 9159 | if (match(V: FalseVal, P: m_OrdOrUnordFMin(L: m_Specific(V: CmpLHS), R: m_APFloat(Res&: FC2))) && |
| 9160 | *FC1 < *FC2) |
| 9161 | return {.Flavor: SPF_FMAXNUM, .NaNBehavior: SPNB_RETURNS_ANY, .Ordered: false}; |
| 9162 | if (match(V: FalseVal, P: m_FMinNum(Op0: m_Specific(V: CmpLHS), Op1: m_APFloat(Res&: FC2))) && |
| 9163 | *FC1 < *FC2) |
| 9164 | return {.Flavor: SPF_FMAXNUM, .NaNBehavior: SPNB_RETURNS_ANY, .Ordered: false}; |
| 9165 | break; |
| 9166 | case CmpInst::FCMP_OGT: |
| 9167 | case CmpInst::FCMP_OGE: |
| 9168 | case CmpInst::FCMP_UGT: |
| 9169 | case CmpInst::FCMP_UGE: |
| 9170 | if (match(V: FalseVal, P: m_OrdOrUnordFMax(L: m_Specific(V: CmpLHS), R: m_APFloat(Res&: FC2))) && |
| 9171 | *FC1 > *FC2) |
| 9172 | return {.Flavor: SPF_FMINNUM, .NaNBehavior: SPNB_RETURNS_ANY, .Ordered: false}; |
| 9173 | if (match(V: FalseVal, P: m_FMaxNum(Op0: m_Specific(V: CmpLHS), Op1: m_APFloat(Res&: FC2))) && |
| 9174 | *FC1 > *FC2) |
| 9175 | return {.Flavor: SPF_FMINNUM, .NaNBehavior: SPNB_RETURNS_ANY, .Ordered: false}; |
| 9176 | break; |
| 9177 | default: |
| 9178 | break; |
| 9179 | } |
| 9180 | |
| 9181 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9182 | } |
| 9183 | |
| 9184 | /// Recognize variations of: |
| 9185 | /// CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v))) |
| 9186 | static SelectPatternResult matchClamp(CmpInst::Predicate Pred, |
| 9187 | Value *CmpLHS, Value *CmpRHS, |
| 9188 | Value *TrueVal, Value *FalseVal) { |
| 9189 | // Swap the select operands and predicate to match the patterns below. |
| 9190 | if (CmpRHS != TrueVal) { |
| 9191 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
| 9192 | std::swap(a&: TrueVal, b&: FalseVal); |
| 9193 | } |
| 9194 | const APInt *C1; |
| 9195 | if (CmpRHS == TrueVal && match(V: CmpRHS, P: m_APInt(Res&: C1))) { |
| 9196 | const APInt *C2; |
| 9197 | // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1) |
| 9198 | if (match(V: FalseVal, P: m_SMin(Op0: m_Specific(V: CmpLHS), Op1: m_APInt(Res&: C2))) && |
| 9199 | C1->slt(RHS: *C2) && Pred == CmpInst::ICMP_SLT) |
| 9200 | return {.Flavor: SPF_SMAX, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9201 | |
| 9202 | // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1) |
| 9203 | if (match(V: FalseVal, P: m_SMax(Op0: m_Specific(V: CmpLHS), Op1: m_APInt(Res&: C2))) && |
| 9204 | C1->sgt(RHS: *C2) && Pred == CmpInst::ICMP_SGT) |
| 9205 | return {.Flavor: SPF_SMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9206 | |
| 9207 | // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1) |
| 9208 | if (match(V: FalseVal, P: m_UMin(Op0: m_Specific(V: CmpLHS), Op1: m_APInt(Res&: C2))) && |
| 9209 | C1->ult(RHS: *C2) && Pred == CmpInst::ICMP_ULT) |
| 9210 | return {.Flavor: SPF_UMAX, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9211 | |
| 9212 | // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1) |
| 9213 | if (match(V: FalseVal, P: m_UMax(Op0: m_Specific(V: CmpLHS), Op1: m_APInt(Res&: C2))) && |
| 9214 | C1->ugt(RHS: *C2) && Pred == CmpInst::ICMP_UGT) |
| 9215 | return {.Flavor: SPF_UMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9216 | } |
| 9217 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9218 | } |
| 9219 | |
| 9220 | /// If the input value is the result of a 'not' op, constant integer, or vector |
| 9221 | /// splat of a constant integer, return the bitwise-not source value. |
| 9222 | /// TODO: This could be extended to handle non-splat vector integer constants. |
| 9223 | static Value *getNotValue(Value *V) { |
| 9224 | Value *NotV; |
| 9225 | if (match(V, P: m_Not(V: m_Value(V&: NotV)))) |
| 9226 | return NotV; |
| 9227 | |
| 9228 | const APInt *C; |
| 9229 | if (match(V, P: m_APInt(Res&: C))) |
| 9230 | return ConstantInt::get(Ty: V->getType(), V: ~(*C)); |
| 9231 | |
| 9232 | return nullptr; |
| 9233 | } |
| 9234 | |
| 9235 | /// Match non-obvious integer minimum and maximum sequences. |
| 9236 | static SelectPatternResult matchMinMax(CmpInst::Predicate Pred, Value *CmpLHS, |
| 9237 | Value *CmpRHS, Value *TrueVal, |
| 9238 | Value *FalseVal, Value *&LHS, |
| 9239 | Value *&RHS) { |
| 9240 | // Assume success. If there's no match, callers should not use these anyway. |
| 9241 | LHS = TrueVal; |
| 9242 | RHS = FalseVal; |
| 9243 | |
| 9244 | SelectPatternResult SPR = matchClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal); |
| 9245 | if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN) |
| 9246 | return SPR; |
| 9247 | |
| 9248 | // Look through 'not' ops to find disguised min/max. |
| 9249 | // (X > Y) ? ~X : ~Y ==> (~X < ~Y) ? ~X : ~Y ==> MIN(~X, ~Y) |
| 9250 | // (X < Y) ? ~X : ~Y ==> (~X > ~Y) ? ~X : ~Y ==> MAX(~X, ~Y) |
| 9251 | if (CmpLHS == getNotValue(V: TrueVal) && CmpRHS == getNotValue(V: FalseVal)) { |
| 9252 | switch (Pred) { |
| 9253 | case CmpInst::ICMP_SGT: return {.Flavor: SPF_SMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9254 | case CmpInst::ICMP_SLT: return {.Flavor: SPF_SMAX, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9255 | case CmpInst::ICMP_UGT: return {.Flavor: SPF_UMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9256 | case CmpInst::ICMP_ULT: return {.Flavor: SPF_UMAX, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9257 | default: break; |
| 9258 | } |
| 9259 | } |
| 9260 | |
| 9261 | // (X > Y) ? ~Y : ~X ==> (~X < ~Y) ? ~Y : ~X ==> MAX(~Y, ~X) |
| 9262 | // (X < Y) ? ~Y : ~X ==> (~X > ~Y) ? ~Y : ~X ==> MIN(~Y, ~X) |
| 9263 | if (CmpLHS == getNotValue(V: FalseVal) && CmpRHS == getNotValue(V: TrueVal)) { |
| 9264 | switch (Pred) { |
| 9265 | case CmpInst::ICMP_SGT: return {.Flavor: SPF_SMAX, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9266 | case CmpInst::ICMP_SLT: return {.Flavor: SPF_SMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9267 | case CmpInst::ICMP_UGT: return {.Flavor: SPF_UMAX, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9268 | case CmpInst::ICMP_ULT: return {.Flavor: SPF_UMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9269 | default: break; |
| 9270 | } |
| 9271 | } |
| 9272 | |
| 9273 | if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT) |
| 9274 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9275 | |
| 9276 | const APInt *C1; |
| 9277 | if (!match(V: CmpRHS, P: m_APInt(Res&: C1))) |
| 9278 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9279 | |
| 9280 | // An unsigned min/max can be written with a signed compare. |
| 9281 | const APInt *C2; |
| 9282 | if ((CmpLHS == TrueVal && match(V: FalseVal, P: m_APInt(Res&: C2))) || |
| 9283 | (CmpLHS == FalseVal && match(V: TrueVal, P: m_APInt(Res&: C2)))) { |
| 9284 | // Is the sign bit set? |
| 9285 | // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX |
| 9286 | // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN |
| 9287 | if (Pred == CmpInst::ICMP_SLT && C1->isZero() && C2->isMaxSignedValue()) |
| 9288 | return {.Flavor: CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9289 | |
| 9290 | // Is the sign bit clear? |
| 9291 | // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX |
| 9292 | // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN |
| 9293 | if (Pred == CmpInst::ICMP_SGT && C1->isAllOnes() && C2->isMinSignedValue()) |
| 9294 | return {.Flavor: CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9295 | } |
| 9296 | |
| 9297 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9298 | } |
| 9299 | |
| 9300 | bool llvm::isKnownNegation(const Value *X, const Value *Y, bool NeedNSW, |
| 9301 | bool AllowPoison) { |
| 9302 | assert(X && Y && "Invalid operand" ); |
| 9303 | |
| 9304 | auto IsNegationOf = [&](const Value *X, const Value *Y) { |
| 9305 | if (!match(V: X, P: m_Neg(V: m_Specific(V: Y)))) |
| 9306 | return false; |
| 9307 | |
| 9308 | auto *BO = cast<BinaryOperator>(Val: X); |
| 9309 | if (NeedNSW && !BO->hasNoSignedWrap()) |
| 9310 | return false; |
| 9311 | |
| 9312 | auto *Zero = cast<Constant>(Val: BO->getOperand(i_nocapture: 0)); |
| 9313 | if (!AllowPoison && !Zero->isNullValue()) |
| 9314 | return false; |
| 9315 | |
| 9316 | return true; |
| 9317 | }; |
| 9318 | |
| 9319 | // X = -Y or Y = -X |
| 9320 | if (IsNegationOf(X, Y) || IsNegationOf(Y, X)) |
| 9321 | return true; |
| 9322 | |
| 9323 | // X = sub (A, B), Y = sub (B, A) || X = sub nsw (A, B), Y = sub nsw (B, A) |
| 9324 | Value *A, *B; |
| 9325 | return (!NeedNSW && (match(V: X, P: m_Sub(L: m_Value(V&: A), R: m_Value(V&: B))) && |
| 9326 | match(V: Y, P: m_Sub(L: m_Specific(V: B), R: m_Specific(V: A))))) || |
| 9327 | (NeedNSW && (match(V: X, P: m_NSWSub(L: m_Value(V&: A), R: m_Value(V&: B))) && |
| 9328 | match(V: Y, P: m_NSWSub(L: m_Specific(V: B), R: m_Specific(V: A))))); |
| 9329 | } |
| 9330 | |
| 9331 | bool llvm::isKnownInversion(const Value *X, const Value *Y) { |
| 9332 | // Handle X = icmp pred A, B, Y = icmp pred A, C. |
| 9333 | Value *A, *B, *C; |
| 9334 | CmpPredicate Pred1, Pred2; |
| 9335 | if (!match(V: X, P: m_ICmp(Pred&: Pred1, L: m_Value(V&: A), R: m_Value(V&: B))) || |
| 9336 | !match(V: Y, P: m_c_ICmp(Pred&: Pred2, L: m_Specific(V: A), R: m_Value(V&: C)))) |
| 9337 | return false; |
| 9338 | |
| 9339 | // They must both have samesign flag or not. |
| 9340 | if (Pred1.hasSameSign() != Pred2.hasSameSign()) |
| 9341 | return false; |
| 9342 | |
| 9343 | if (B == C) |
| 9344 | return Pred1 == ICmpInst::getInversePredicate(pred: Pred2); |
| 9345 | |
| 9346 | // Try to infer the relationship from constant ranges. |
| 9347 | const APInt *RHSC1, *RHSC2; |
| 9348 | if (!match(V: B, P: m_APInt(Res&: RHSC1)) || !match(V: C, P: m_APInt(Res&: RHSC2))) |
| 9349 | return false; |
| 9350 | |
| 9351 | // Sign bits of two RHSCs should match. |
| 9352 | if (Pred1.hasSameSign() && RHSC1->isNonNegative() != RHSC2->isNonNegative()) |
| 9353 | return false; |
| 9354 | |
| 9355 | const auto CR1 = ConstantRange::makeExactICmpRegion(Pred: Pred1, Other: *RHSC1); |
| 9356 | const auto CR2 = ConstantRange::makeExactICmpRegion(Pred: Pred2, Other: *RHSC2); |
| 9357 | |
| 9358 | return CR1.inverse() == CR2; |
| 9359 | } |
| 9360 | |
| 9361 | SelectPatternResult llvm::getSelectPattern(CmpInst::Predicate Pred, |
| 9362 | SelectPatternNaNBehavior NaNBehavior, |
| 9363 | bool Ordered) { |
| 9364 | switch (Pred) { |
| 9365 | default: |
| 9366 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; // Equality. |
| 9367 | case ICmpInst::ICMP_UGT: |
| 9368 | case ICmpInst::ICMP_UGE: |
| 9369 | return {.Flavor: SPF_UMAX, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9370 | case ICmpInst::ICMP_SGT: |
| 9371 | case ICmpInst::ICMP_SGE: |
| 9372 | return {.Flavor: SPF_SMAX, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9373 | case ICmpInst::ICMP_ULT: |
| 9374 | case ICmpInst::ICMP_ULE: |
| 9375 | return {.Flavor: SPF_UMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9376 | case ICmpInst::ICMP_SLT: |
| 9377 | case ICmpInst::ICMP_SLE: |
| 9378 | return {.Flavor: SPF_SMIN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9379 | case FCmpInst::FCMP_UGT: |
| 9380 | case FCmpInst::FCMP_UGE: |
| 9381 | case FCmpInst::FCMP_OGT: |
| 9382 | case FCmpInst::FCMP_OGE: |
| 9383 | return {.Flavor: SPF_FMAXNUM, .NaNBehavior: NaNBehavior, .Ordered: Ordered}; |
| 9384 | case FCmpInst::FCMP_ULT: |
| 9385 | case FCmpInst::FCMP_ULE: |
| 9386 | case FCmpInst::FCMP_OLT: |
| 9387 | case FCmpInst::FCMP_OLE: |
| 9388 | return {.Flavor: SPF_FMINNUM, .NaNBehavior: NaNBehavior, .Ordered: Ordered}; |
| 9389 | } |
| 9390 | } |
| 9391 | |
| 9392 | std::optional<std::pair<CmpPredicate, Constant *>> |
| 9393 | llvm::getFlippedStrictnessPredicateAndConstant(CmpPredicate Pred, Constant *C) { |
| 9394 | assert(ICmpInst::isRelational(Pred) && ICmpInst::isIntPredicate(Pred) && |
| 9395 | "Only for relational integer predicates." ); |
| 9396 | if (isa<UndefValue>(Val: C)) |
| 9397 | return std::nullopt; |
| 9398 | |
| 9399 | Type *Type = C->getType(); |
| 9400 | bool IsSigned = ICmpInst::isSigned(Pred); |
| 9401 | |
| 9402 | CmpInst::Predicate UnsignedPred = ICmpInst::getUnsignedPredicate(Pred); |
| 9403 | bool WillIncrement = |
| 9404 | UnsignedPred == ICmpInst::ICMP_ULE || UnsignedPred == ICmpInst::ICMP_UGT; |
| 9405 | |
| 9406 | // Check if the constant operand can be safely incremented/decremented |
| 9407 | // without overflowing/underflowing. |
| 9408 | auto ConstantIsOk = [Pred, WillIncrement, IsSigned](ConstantInt *C) { |
| 9409 | if (WillIncrement ? C->isMaxValue(IsSigned) : C->isMinValue(IsSigned)) |
| 9410 | return false; |
| 9411 | |
| 9412 | if (!Pred.hasSameSign()) |
| 9413 | return true; |
| 9414 | |
| 9415 | // Crossing the corresponding boundary in the other ordering changes the |
| 9416 | // sign bit, and therefore changes the poison domain. |
| 9417 | return WillIncrement ? !C->isMaxValue(IsSigned: !IsSigned) |
| 9418 | : !C->isMinValue(IsSigned: !IsSigned); |
| 9419 | }; |
| 9420 | |
| 9421 | Constant *SafeReplacementConstant = nullptr; |
| 9422 | if (auto *CI = dyn_cast<ConstantInt>(Val: C)) { |
| 9423 | // Bail out if the constant can't be safely incremented/decremented. |
| 9424 | if (!ConstantIsOk(CI)) |
| 9425 | return std::nullopt; |
| 9426 | } else if (auto *FVTy = dyn_cast<FixedVectorType>(Val: Type)) { |
| 9427 | unsigned NumElts = FVTy->getNumElements(); |
| 9428 | for (unsigned i = 0; i != NumElts; ++i) { |
| 9429 | Constant *Elt = C->getAggregateElement(Elt: i); |
| 9430 | if (!Elt) |
| 9431 | return std::nullopt; |
| 9432 | |
| 9433 | if (isa<UndefValue>(Val: Elt)) |
| 9434 | continue; |
| 9435 | |
| 9436 | // Bail out if we can't determine if this constant is min/max or if we |
| 9437 | // know that this constant is min/max. |
| 9438 | auto *CI = dyn_cast<ConstantInt>(Val: Elt); |
| 9439 | if (!CI || !ConstantIsOk(CI)) |
| 9440 | return std::nullopt; |
| 9441 | |
| 9442 | if (!SafeReplacementConstant) |
| 9443 | SafeReplacementConstant = CI; |
| 9444 | } |
| 9445 | } else if (isa<VectorType>(Val: C->getType())) { |
| 9446 | // Handle scalable splat |
| 9447 | Value *SplatC = C->getSplatValue(); |
| 9448 | auto *CI = dyn_cast_or_null<ConstantInt>(Val: SplatC); |
| 9449 | // Bail out if the constant can't be safely incremented/decremented. |
| 9450 | if (!CI || !ConstantIsOk(CI)) |
| 9451 | return std::nullopt; |
| 9452 | } else { |
| 9453 | // ConstantExpr? |
| 9454 | return std::nullopt; |
| 9455 | } |
| 9456 | |
| 9457 | // It may not be safe to change a compare predicate in the presence of |
| 9458 | // undefined elements, so replace those elements with the first safe constant |
| 9459 | // that we found. |
| 9460 | // TODO: in case of poison, it is safe; let's replace undefs only. |
| 9461 | if (C->containsUndefOrPoisonElement()) { |
| 9462 | assert(SafeReplacementConstant && "Replacement constant not set" ); |
| 9463 | C = Constant::replaceUndefsWith(C, Replacement: SafeReplacementConstant); |
| 9464 | } |
| 9465 | |
| 9466 | CmpPredicate NewPred(CmpInst::getFlippedStrictnessPredicate(pred: Pred), |
| 9467 | Pred.hasSameSign()); |
| 9468 | |
| 9469 | // Increment or decrement the constant. |
| 9470 | Constant *OneOrNegOne = ConstantInt::get(Ty: Type, V: WillIncrement ? 1 : -1, IsSigned: true); |
| 9471 | Constant *NewC = ConstantExpr::getAdd(C1: C, C2: OneOrNegOne); |
| 9472 | |
| 9473 | return std::make_pair(x&: NewPred, y&: NewC); |
| 9474 | } |
| 9475 | |
| 9476 | static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred, |
| 9477 | FastMathFlags FMF, Value *CmpLHS, |
| 9478 | Value *CmpRHS, Value *TrueVal, |
| 9479 | Value *FalseVal, Value *&LHS, |
| 9480 | Value *&RHS) { |
| 9481 | if (CmpInst::isFPPredicate(P: Pred)) { |
| 9482 | // IEEE-754 ignores the sign of 0.0 in comparisons. So if the select has one |
| 9483 | // 0.0 operand, set the compare's 0.0 operands to that same value for the |
| 9484 | // purpose of identifying min/max. Disregard vector constants with undefined |
| 9485 | // elements because those can not be back-propagated for analysis. |
| 9486 | Value *OutputZeroVal = nullptr; |
| 9487 | if (match(V: TrueVal, P: m_AnyZeroFP()) && !match(V: FalseVal, P: m_AnyZeroFP()) && |
| 9488 | !cast<Constant>(Val: TrueVal)->containsUndefOrPoisonElement()) |
| 9489 | OutputZeroVal = TrueVal; |
| 9490 | else if (match(V: FalseVal, P: m_AnyZeroFP()) && !match(V: TrueVal, P: m_AnyZeroFP()) && |
| 9491 | !cast<Constant>(Val: FalseVal)->containsUndefOrPoisonElement()) |
| 9492 | OutputZeroVal = FalseVal; |
| 9493 | |
| 9494 | if (OutputZeroVal) { |
| 9495 | if (match(V: CmpLHS, P: m_AnyZeroFP()) && CmpLHS != OutputZeroVal) |
| 9496 | CmpLHS = OutputZeroVal; |
| 9497 | if (match(V: CmpRHS, P: m_AnyZeroFP()) && CmpRHS != OutputZeroVal) |
| 9498 | CmpRHS = OutputZeroVal; |
| 9499 | } |
| 9500 | } |
| 9501 | |
| 9502 | LHS = CmpLHS; |
| 9503 | RHS = CmpRHS; |
| 9504 | |
| 9505 | // Signed zero may return inconsistent results between implementations. |
| 9506 | // (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0 |
| 9507 | // minNum(0.0, -0.0) // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1) |
| 9508 | // Therefore, we behave conservatively and only proceed if at least one of the |
| 9509 | // operands is known to not be zero or if we don't care about signed zero. |
| 9510 | if (CmpInst::isFPPredicate(P: Pred)) { |
| 9511 | if (!FMF.noSignedZeros() && !isKnownNonZero(V: CmpLHS) && |
| 9512 | !isKnownNonZero(V: CmpRHS)) |
| 9513 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9514 | } |
| 9515 | |
| 9516 | SelectPatternNaNBehavior NaNBehavior = SPNB_NA; |
| 9517 | bool Ordered = false; |
| 9518 | |
| 9519 | // When given one NaN and one non-NaN input: |
| 9520 | // - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input. |
| 9521 | // - A simple C99 (a < b ? a : b) construction will return 'b' (as the |
| 9522 | // ordered comparison fails), which could be NaN or non-NaN. |
| 9523 | // so here we discover exactly what NaN behavior is required/accepted. |
| 9524 | if (CmpInst::isFPPredicate(P: Pred)) { |
| 9525 | bool LHSSafe = isKnownNonNaN(V: CmpLHS, FMF); |
| 9526 | bool RHSSafe = isKnownNonNaN(V: CmpRHS, FMF); |
| 9527 | |
| 9528 | if (LHSSafe && RHSSafe) { |
| 9529 | // Both operands are known non-NaN. |
| 9530 | NaNBehavior = SPNB_RETURNS_ANY; |
| 9531 | Ordered = CmpInst::isOrdered(predicate: Pred); |
| 9532 | } else if (CmpInst::isOrdered(predicate: Pred)) { |
| 9533 | // An ordered comparison will return false when given a NaN, so it |
| 9534 | // returns the RHS. |
| 9535 | Ordered = true; |
| 9536 | if (LHSSafe) |
| 9537 | // LHS is non-NaN, so if RHS is NaN then NaN will be returned. |
| 9538 | NaNBehavior = SPNB_RETURNS_NAN; |
| 9539 | else if (RHSSafe) |
| 9540 | NaNBehavior = SPNB_RETURNS_OTHER; |
| 9541 | else |
| 9542 | // Completely unsafe. |
| 9543 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9544 | } else { |
| 9545 | Ordered = false; |
| 9546 | // An unordered comparison will return true when given a NaN, so it |
| 9547 | // returns the LHS. |
| 9548 | if (LHSSafe) |
| 9549 | // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned. |
| 9550 | NaNBehavior = SPNB_RETURNS_OTHER; |
| 9551 | else if (RHSSafe) |
| 9552 | NaNBehavior = SPNB_RETURNS_NAN; |
| 9553 | else |
| 9554 | // Completely unsafe. |
| 9555 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9556 | } |
| 9557 | } |
| 9558 | |
| 9559 | if (TrueVal == CmpRHS && FalseVal == CmpLHS) { |
| 9560 | std::swap(a&: CmpLHS, b&: CmpRHS); |
| 9561 | Pred = CmpInst::getSwappedPredicate(pred: Pred); |
| 9562 | if (NaNBehavior == SPNB_RETURNS_NAN) |
| 9563 | NaNBehavior = SPNB_RETURNS_OTHER; |
| 9564 | else if (NaNBehavior == SPNB_RETURNS_OTHER) |
| 9565 | NaNBehavior = SPNB_RETURNS_NAN; |
| 9566 | Ordered = !Ordered; |
| 9567 | } |
| 9568 | |
| 9569 | // ([if]cmp X, Y) ? X : Y |
| 9570 | if (TrueVal == CmpLHS && FalseVal == CmpRHS) |
| 9571 | return getSelectPattern(Pred, NaNBehavior, Ordered); |
| 9572 | |
| 9573 | if (isKnownNegation(X: TrueVal, Y: FalseVal)) { |
| 9574 | // Sign-extending LHS does not change its sign, so TrueVal/FalseVal can |
| 9575 | // match against either LHS or sign-preserving operations on LHS, like |
| 9576 | // sext(LHS), or binary ops that do not wrap in signed sense. |
| 9577 | auto CmpLHSOrSExt = |
| 9578 | m_CombineOr(Ps: m_Specific(V: CmpLHS), Ps: m_SExt(Op: m_Specific(V: CmpLHS))); |
| 9579 | auto MaybeSExtOrMulCmpLHS = |
| 9580 | m_CombineOr(Ps: CmpLHSOrSExt, Ps: m_NSWMul(L: CmpLHSOrSExt, R: m_StrictlyPositive()), |
| 9581 | Ps: m_NSWShl(L: CmpLHSOrSExt, R: m_Value())); |
| 9582 | auto ZeroOrAllOnes = m_CombineOr(Ps: m_ZeroInt(), Ps: m_AllOnes()); |
| 9583 | auto ZeroOrOne = m_CombineOr(Ps: m_ZeroInt(), Ps: m_One()); |
| 9584 | if (match(V: TrueVal, P: MaybeSExtOrMulCmpLHS)) { |
| 9585 | // Set the return values. If the compare uses the negated value (-X >s 0), |
| 9586 | // swap the return values because the negated value is always 'RHS'. |
| 9587 | LHS = TrueVal; |
| 9588 | RHS = FalseVal; |
| 9589 | if (match(V: CmpLHS, P: m_Neg(V: m_Specific(V: FalseVal)))) |
| 9590 | std::swap(a&: LHS, b&: RHS); |
| 9591 | |
| 9592 | // (X >s 0) ? X : -X or (X >s -1) ? X : -X --> ABS(X) |
| 9593 | // (-X >s 0) ? -X : X or (-X >s -1) ? -X : X --> ABS(X) |
| 9594 | if (Pred == ICmpInst::ICMP_SGT && match(V: CmpRHS, P: ZeroOrAllOnes)) |
| 9595 | return {.Flavor: SPF_ABS, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9596 | |
| 9597 | // (X >=s 0) ? X : -X or (X >=s 1) ? X : -X --> ABS(X) |
| 9598 | if (Pred == ICmpInst::ICMP_SGE && match(V: CmpRHS, P: ZeroOrOne)) |
| 9599 | return {.Flavor: SPF_ABS, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9600 | |
| 9601 | // (X <s 0) ? X : -X or (X <s 1) ? X : -X --> NABS(X) |
| 9602 | // (-X <s 0) ? -X : X or (-X <s 1) ? -X : X --> NABS(X) |
| 9603 | if (Pred == ICmpInst::ICMP_SLT && match(V: CmpRHS, P: ZeroOrOne)) |
| 9604 | return {.Flavor: SPF_NABS, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9605 | } else if (match(V: FalseVal, P: MaybeSExtOrMulCmpLHS)) { |
| 9606 | // Set the return values. If the compare uses the negated value (-X >s 0), |
| 9607 | // swap the return values because the negated value is always 'RHS'. |
| 9608 | LHS = FalseVal; |
| 9609 | RHS = TrueVal; |
| 9610 | if (match(V: CmpLHS, P: m_Neg(V: m_Specific(V: TrueVal)))) |
| 9611 | std::swap(a&: LHS, b&: RHS); |
| 9612 | |
| 9613 | // (X >s 0) ? -X : X or (X >s -1) ? -X : X --> NABS(X) |
| 9614 | // (-X >s 0) ? X : -X or (-X >s -1) ? X : -X --> NABS(X) |
| 9615 | if (Pred == ICmpInst::ICMP_SGT && match(V: CmpRHS, P: ZeroOrAllOnes)) |
| 9616 | return {.Flavor: SPF_NABS, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9617 | |
| 9618 | // (X <s 0) ? -X : X or (X <s 1) ? -X : X --> ABS(X) |
| 9619 | // (-X <s 0) ? X : -X or (-X <s 1) ? X : -X --> ABS(X) |
| 9620 | if (Pred == ICmpInst::ICMP_SLT && match(V: CmpRHS, P: ZeroOrOne)) |
| 9621 | return {.Flavor: SPF_ABS, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9622 | } |
| 9623 | } |
| 9624 | |
| 9625 | if (CmpInst::isIntPredicate(P: Pred)) |
| 9626 | return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS); |
| 9627 | |
| 9628 | // According to (IEEE 754-2008 5.3.1), minNum(0.0, -0.0) and similar |
| 9629 | // may return either -0.0 or 0.0, so fcmp/select pair has stricter |
| 9630 | // semantics than minNum. Be conservative in such case. |
| 9631 | if (NaNBehavior != SPNB_RETURNS_ANY || |
| 9632 | (!FMF.noSignedZeros() && !isKnownNonZero(V: CmpLHS) && |
| 9633 | !isKnownNonZero(V: CmpRHS))) |
| 9634 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9635 | |
| 9636 | return matchFastFloatClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS); |
| 9637 | } |
| 9638 | |
| 9639 | static Value *lookThroughCastConst(CmpInst *CmpI, Type *SrcTy, Constant *C, |
| 9640 | Instruction::CastOps *CastOp) { |
| 9641 | const DataLayout &DL = CmpI->getDataLayout(); |
| 9642 | |
| 9643 | Constant *CastedTo = nullptr; |
| 9644 | switch (*CastOp) { |
| 9645 | case Instruction::ZExt: |
| 9646 | if (CmpI->isUnsigned()) |
| 9647 | CastedTo = ConstantExpr::getTrunc(C, Ty: SrcTy); |
| 9648 | break; |
| 9649 | case Instruction::SExt: |
| 9650 | if (CmpI->isSigned()) |
| 9651 | CastedTo = ConstantExpr::getTrunc(C, Ty: SrcTy, OnlyIfReduced: true); |
| 9652 | break; |
| 9653 | case Instruction::Trunc: |
| 9654 | Constant *CmpConst; |
| 9655 | if (match(V: CmpI->getOperand(i_nocapture: 1), P: m_Constant(C&: CmpConst)) && |
| 9656 | CmpConst->getType() == SrcTy) { |
| 9657 | // Here we have the following case: |
| 9658 | // |
| 9659 | // %cond = cmp iN %x, CmpConst |
| 9660 | // %tr = trunc iN %x to iK |
| 9661 | // %narrowsel = select i1 %cond, iK %t, iK C |
| 9662 | // |
| 9663 | // We can always move trunc after select operation: |
| 9664 | // |
| 9665 | // %cond = cmp iN %x, CmpConst |
| 9666 | // %widesel = select i1 %cond, iN %x, iN CmpConst |
| 9667 | // %tr = trunc iN %widesel to iK |
| 9668 | // |
| 9669 | // Note that C could be extended in any way because we don't care about |
| 9670 | // upper bits after truncation. It can't be abs pattern, because it would |
| 9671 | // look like: |
| 9672 | // |
| 9673 | // select i1 %cond, x, -x. |
| 9674 | // |
| 9675 | // So only min/max pattern could be matched. Such match requires widened C |
| 9676 | // == CmpConst. That is why set widened C = CmpConst, condition trunc |
| 9677 | // CmpConst == C is checked below. |
| 9678 | CastedTo = CmpConst; |
| 9679 | } else { |
| 9680 | unsigned ExtOp = CmpI->isSigned() ? Instruction::SExt : Instruction::ZExt; |
| 9681 | CastedTo = ConstantFoldCastOperand(Opcode: ExtOp, C, DestTy: SrcTy, DL); |
| 9682 | } |
| 9683 | break; |
| 9684 | case Instruction::FPTrunc: |
| 9685 | CastedTo = ConstantFoldCastOperand(Opcode: Instruction::FPExt, C, DestTy: SrcTy, DL); |
| 9686 | break; |
| 9687 | case Instruction::FPExt: |
| 9688 | CastedTo = ConstantFoldCastOperand(Opcode: Instruction::FPTrunc, C, DestTy: SrcTy, DL); |
| 9689 | break; |
| 9690 | case Instruction::FPToUI: |
| 9691 | CastedTo = ConstantFoldCastOperand(Opcode: Instruction::UIToFP, C, DestTy: SrcTy, DL); |
| 9692 | break; |
| 9693 | case Instruction::FPToSI: |
| 9694 | CastedTo = ConstantFoldCastOperand(Opcode: Instruction::SIToFP, C, DestTy: SrcTy, DL); |
| 9695 | break; |
| 9696 | case Instruction::UIToFP: |
| 9697 | CastedTo = ConstantFoldCastOperand(Opcode: Instruction::FPToUI, C, DestTy: SrcTy, DL); |
| 9698 | break; |
| 9699 | case Instruction::SIToFP: |
| 9700 | CastedTo = ConstantFoldCastOperand(Opcode: Instruction::FPToSI, C, DestTy: SrcTy, DL); |
| 9701 | break; |
| 9702 | default: |
| 9703 | break; |
| 9704 | } |
| 9705 | |
| 9706 | if (!CastedTo) |
| 9707 | return nullptr; |
| 9708 | |
| 9709 | // Make sure the cast doesn't lose any information. |
| 9710 | Constant *CastedBack = |
| 9711 | ConstantFoldCastOperand(Opcode: *CastOp, C: CastedTo, DestTy: C->getType(), DL); |
| 9712 | if (CastedBack && CastedBack != C) |
| 9713 | return nullptr; |
| 9714 | |
| 9715 | return CastedTo; |
| 9716 | } |
| 9717 | |
| 9718 | /// Helps to match a select pattern in case of a type mismatch. |
| 9719 | /// |
| 9720 | /// The function processes the case when type of true and false values of a |
| 9721 | /// select instruction differs from type of the cmp instruction operands because |
| 9722 | /// of a cast instruction. The function checks if it is legal to move the cast |
| 9723 | /// operation after "select". If yes, it returns the new second value of |
| 9724 | /// "select" (with the assumption that cast is moved): |
| 9725 | /// 1. As operand of cast instruction when both values of "select" are same cast |
| 9726 | /// instructions. |
| 9727 | /// 2. As restored constant (by applying reverse cast operation) when the first |
| 9728 | /// value of the "select" is a cast operation and the second value is a |
| 9729 | /// constant. It is implemented in lookThroughCastConst(). |
| 9730 | /// 3. As one operand is cast instruction and the other is not. The operands in |
| 9731 | /// sel(cmp) are in different type integer. |
| 9732 | /// NOTE: We return only the new second value because the first value could be |
| 9733 | /// accessed as operand of cast instruction. |
| 9734 | static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2, |
| 9735 | Instruction::CastOps *CastOp) { |
| 9736 | auto *Cast1 = dyn_cast<CastInst>(Val: V1); |
| 9737 | if (!Cast1) |
| 9738 | return nullptr; |
| 9739 | |
| 9740 | *CastOp = Cast1->getOpcode(); |
| 9741 | Type *SrcTy = Cast1->getSrcTy(); |
| 9742 | if (auto *Cast2 = dyn_cast<CastInst>(Val: V2)) { |
| 9743 | // If V1 and V2 are both the same cast from the same type, look through V1. |
| 9744 | if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy()) |
| 9745 | return Cast2->getOperand(i_nocapture: 0); |
| 9746 | return nullptr; |
| 9747 | } |
| 9748 | |
| 9749 | auto *C = dyn_cast<Constant>(Val: V2); |
| 9750 | if (C) |
| 9751 | return lookThroughCastConst(CmpI, SrcTy, C, CastOp); |
| 9752 | |
| 9753 | Value *CastedTo = nullptr; |
| 9754 | if (*CastOp == Instruction::Trunc) { |
| 9755 | if (match(V: CmpI->getOperand(i_nocapture: 1), P: m_ZExtOrSExt(Op: m_Specific(V: V2)))) { |
| 9756 | // Here we have the following case: |
| 9757 | // %y_ext = sext iK %y to iN |
| 9758 | // %cond = cmp iN %x, %y_ext |
| 9759 | // %tr = trunc iN %x to iK |
| 9760 | // %narrowsel = select i1 %cond, iK %tr, iK %y |
| 9761 | // |
| 9762 | // We can always move trunc after select operation: |
| 9763 | // %y_ext = sext iK %y to iN |
| 9764 | // %cond = cmp iN %x, %y_ext |
| 9765 | // %widesel = select i1 %cond, iN %x, iN %y_ext |
| 9766 | // %tr = trunc iN %widesel to iK |
| 9767 | assert(V2->getType() == Cast1->getType() && |
| 9768 | "V2 and Cast1 should be the same type." ); |
| 9769 | CastedTo = CmpI->getOperand(i_nocapture: 1); |
| 9770 | } |
| 9771 | } |
| 9772 | |
| 9773 | return CastedTo; |
| 9774 | } |
| 9775 | SelectPatternResult llvm::matchSelectPattern(Value *V, Value *&LHS, Value *&RHS, |
| 9776 | Instruction::CastOps *CastOp) { |
| 9777 | SelectInst *SI = dyn_cast<SelectInst>(Val: V); |
| 9778 | if (!SI) return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9779 | |
| 9780 | CmpInst *CmpI = dyn_cast<CmpInst>(Val: SI->getCondition()); |
| 9781 | if (!CmpI) return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9782 | |
| 9783 | Value *TrueVal = SI->getTrueValue(); |
| 9784 | Value *FalseVal = SI->getFalseValue(); |
| 9785 | |
| 9786 | return llvm::matchDecomposedSelectPattern( |
| 9787 | CmpI, TrueVal, FalseVal, LHS, RHS, FMF: SI->getFastMathFlagsOrNone(), CastOp); |
| 9788 | } |
| 9789 | |
| 9790 | SelectPatternResult llvm::matchDecomposedSelectPattern( |
| 9791 | CmpInst *CmpI, Value *TrueVal, Value *FalseVal, Value *&LHS, Value *&RHS, |
| 9792 | FastMathFlags FMF, Instruction::CastOps *CastOp) { |
| 9793 | CmpInst::Predicate Pred = CmpI->getPredicate(); |
| 9794 | Value *CmpLHS = CmpI->getOperand(i_nocapture: 0); |
| 9795 | Value *CmpRHS = CmpI->getOperand(i_nocapture: 1); |
| 9796 | if (isa<FPMathOperator>(Val: CmpI) && CmpI->hasNoNaNs()) |
| 9797 | FMF.setNoNaNs(); |
| 9798 | |
| 9799 | // Bail out early. |
| 9800 | if (CmpI->isEquality()) |
| 9801 | return {.Flavor: SPF_UNKNOWN, .NaNBehavior: SPNB_NA, .Ordered: false}; |
| 9802 | |
| 9803 | // Deal with type mismatches. |
| 9804 | if (CastOp && CmpLHS->getType() != TrueVal->getType()) { |
| 9805 | if (Value *C = lookThroughCast(CmpI, V1: TrueVal, V2: FalseVal, CastOp)) { |
| 9806 | // If this is a potential fmin/fmax with a cast to integer, then ignore |
| 9807 | // -0.0 because there is no corresponding integer value. |
| 9808 | if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI) |
| 9809 | FMF.setNoSignedZeros(); |
| 9810 | return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, |
| 9811 | TrueVal: cast<CastInst>(Val: TrueVal)->getOperand(i_nocapture: 0), FalseVal: C, |
| 9812 | LHS, RHS); |
| 9813 | } |
| 9814 | if (Value *C = lookThroughCast(CmpI, V1: FalseVal, V2: TrueVal, CastOp)) { |
| 9815 | // If this is a potential fmin/fmax with a cast to integer, then ignore |
| 9816 | // -0.0 because there is no corresponding integer value. |
| 9817 | if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI) |
| 9818 | FMF.setNoSignedZeros(); |
| 9819 | return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal: C, |
| 9820 | FalseVal: cast<CastInst>(Val: FalseVal)->getOperand(i_nocapture: 0), LHS, |
| 9821 | RHS); |
| 9822 | } |
| 9823 | } |
| 9824 | return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, |
| 9825 | RHS); |
| 9826 | } |
| 9827 | |
| 9828 | CmpInst::Predicate llvm::getMinMaxPred(SelectPatternFlavor SPF, bool Ordered) { |
| 9829 | if (SPF == SPF_SMIN) return ICmpInst::ICMP_SLT; |
| 9830 | if (SPF == SPF_UMIN) return ICmpInst::ICMP_ULT; |
| 9831 | if (SPF == SPF_SMAX) return ICmpInst::ICMP_SGT; |
| 9832 | if (SPF == SPF_UMAX) return ICmpInst::ICMP_UGT; |
| 9833 | if (SPF == SPF_FMINNUM) |
| 9834 | return Ordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; |
| 9835 | if (SPF == SPF_FMAXNUM) |
| 9836 | return Ordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; |
| 9837 | llvm_unreachable("unhandled!" ); |
| 9838 | } |
| 9839 | |
| 9840 | Intrinsic::ID llvm::getMinMaxIntrinsic(SelectPatternFlavor SPF) { |
| 9841 | switch (SPF) { |
| 9842 | case SelectPatternFlavor::SPF_UMIN: |
| 9843 | return Intrinsic::umin; |
| 9844 | case SelectPatternFlavor::SPF_UMAX: |
| 9845 | return Intrinsic::umax; |
| 9846 | case SelectPatternFlavor::SPF_SMIN: |
| 9847 | return Intrinsic::smin; |
| 9848 | case SelectPatternFlavor::SPF_SMAX: |
| 9849 | return Intrinsic::smax; |
| 9850 | default: |
| 9851 | llvm_unreachable("Unexpected SPF" ); |
| 9852 | } |
| 9853 | } |
| 9854 | |
| 9855 | SelectPatternFlavor llvm::getInverseMinMaxFlavor(SelectPatternFlavor SPF) { |
| 9856 | if (SPF == SPF_SMIN) return SPF_SMAX; |
| 9857 | if (SPF == SPF_UMIN) return SPF_UMAX; |
| 9858 | if (SPF == SPF_SMAX) return SPF_SMIN; |
| 9859 | if (SPF == SPF_UMAX) return SPF_UMIN; |
| 9860 | llvm_unreachable("unhandled!" ); |
| 9861 | } |
| 9862 | |
| 9863 | Intrinsic::ID llvm::getInverseMinMaxIntrinsic(Intrinsic::ID MinMaxID) { |
| 9864 | switch (MinMaxID) { |
| 9865 | case Intrinsic::smax: return Intrinsic::smin; |
| 9866 | case Intrinsic::smin: return Intrinsic::smax; |
| 9867 | case Intrinsic::umax: return Intrinsic::umin; |
| 9868 | case Intrinsic::umin: return Intrinsic::umax; |
| 9869 | // Please note that next four intrinsics may produce the same result for |
| 9870 | // original and inverted case even if X != Y due to NaN is handled specially. |
| 9871 | case Intrinsic::maximum: return Intrinsic::minimum; |
| 9872 | case Intrinsic::minimum: return Intrinsic::maximum; |
| 9873 | case Intrinsic::maxnum: return Intrinsic::minnum; |
| 9874 | case Intrinsic::minnum: return Intrinsic::maxnum; |
| 9875 | case Intrinsic::maximumnum: |
| 9876 | return Intrinsic::minimumnum; |
| 9877 | case Intrinsic::minimumnum: |
| 9878 | return Intrinsic::maximumnum; |
| 9879 | default: llvm_unreachable("Unexpected intrinsic" ); |
| 9880 | } |
| 9881 | } |
| 9882 | |
| 9883 | APInt llvm::getMinMaxLimit(SelectPatternFlavor SPF, unsigned BitWidth) { |
| 9884 | switch (SPF) { |
| 9885 | case SPF_SMAX: return APInt::getSignedMaxValue(numBits: BitWidth); |
| 9886 | case SPF_SMIN: return APInt::getSignedMinValue(numBits: BitWidth); |
| 9887 | case SPF_UMAX: return APInt::getMaxValue(numBits: BitWidth); |
| 9888 | case SPF_UMIN: return APInt::getMinValue(numBits: BitWidth); |
| 9889 | default: llvm_unreachable("Unexpected flavor" ); |
| 9890 | } |
| 9891 | } |
| 9892 | |
| 9893 | std::pair<Intrinsic::ID, bool> |
| 9894 | llvm::canConvertToMinOrMaxIntrinsic(ArrayRef<Value *> VL) { |
| 9895 | // Check if VL contains select instructions that can be folded into a min/max |
| 9896 | // vector intrinsic and return the intrinsic if it is possible. |
| 9897 | // TODO: Support floating point min/max. |
| 9898 | bool AllCmpSingleUse = true; |
| 9899 | SelectPatternResult SelectPattern; |
| 9900 | SelectPattern.Flavor = SPF_UNKNOWN; |
| 9901 | if (all_of(Range&: VL, P: [&SelectPattern, &AllCmpSingleUse](Value *I) { |
| 9902 | Value *LHS, *RHS; |
| 9903 | auto CurrentPattern = matchSelectPattern(V: I, LHS, RHS); |
| 9904 | if (!SelectPatternResult::isMinOrMax(SPF: CurrentPattern.Flavor)) |
| 9905 | return false; |
| 9906 | if (SelectPattern.Flavor != SPF_UNKNOWN && |
| 9907 | SelectPattern.Flavor != CurrentPattern.Flavor) |
| 9908 | return false; |
| 9909 | SelectPattern = CurrentPattern; |
| 9910 | AllCmpSingleUse &= |
| 9911 | match(V: I, P: m_Select(C: m_OneUse(SubPattern: m_Value()), L: m_Value(), R: m_Value())); |
| 9912 | return true; |
| 9913 | })) { |
| 9914 | switch (SelectPattern.Flavor) { |
| 9915 | case SPF_SMIN: |
| 9916 | return {Intrinsic::smin, AllCmpSingleUse}; |
| 9917 | case SPF_UMIN: |
| 9918 | return {Intrinsic::umin, AllCmpSingleUse}; |
| 9919 | case SPF_SMAX: |
| 9920 | return {Intrinsic::smax, AllCmpSingleUse}; |
| 9921 | case SPF_UMAX: |
| 9922 | return {Intrinsic::umax, AllCmpSingleUse}; |
| 9923 | case SPF_FMAXNUM: |
| 9924 | return {Intrinsic::maxnum, AllCmpSingleUse}; |
| 9925 | case SPF_FMINNUM: |
| 9926 | return {Intrinsic::minnum, AllCmpSingleUse}; |
| 9927 | default: |
| 9928 | llvm_unreachable("unexpected select pattern flavor" ); |
| 9929 | } |
| 9930 | } |
| 9931 | return {Intrinsic::not_intrinsic, false}; |
| 9932 | } |
| 9933 | |
| 9934 | template <typename InstTy> |
| 9935 | static bool matchTwoInputRecurrence(const PHINode *PN, InstTy *&Inst, |
| 9936 | Value *&Init, Value *&OtherOp) { |
| 9937 | // Handle the case of a simple two-predecessor recurrence PHI. |
| 9938 | // There's a lot more that could theoretically be done here, but |
| 9939 | // this is sufficient to catch some interesting cases. |
| 9940 | // TODO: Expand list -- gep, uadd.sat etc. |
| 9941 | if (PN->getNumIncomingValues() != 2) |
| 9942 | return false; |
| 9943 | |
| 9944 | for (unsigned I = 0; I != 2; ++I) { |
| 9945 | if (auto *Operation = dyn_cast<InstTy>(PN->getIncomingValue(i: I)); |
| 9946 | Operation && Operation->getNumOperands() >= 2) { |
| 9947 | Value *LHS = Operation->getOperand(0); |
| 9948 | Value *RHS = Operation->getOperand(1); |
| 9949 | if (LHS != PN && RHS != PN) |
| 9950 | continue; |
| 9951 | |
| 9952 | Inst = Operation; |
| 9953 | Init = PN->getIncomingValue(i: !I); |
| 9954 | OtherOp = (LHS == PN) ? RHS : LHS; |
| 9955 | return true; |
| 9956 | } |
| 9957 | } |
| 9958 | return false; |
| 9959 | } |
| 9960 | |
| 9961 | template <typename InstTy> |
| 9962 | static bool matchThreeInputRecurrence(const PHINode *PN, InstTy *&Inst, |
| 9963 | Value *&Init, Value *&OtherOp0, |
| 9964 | Value *&OtherOp1) { |
| 9965 | if (PN->getNumIncomingValues() != 2) |
| 9966 | return false; |
| 9967 | |
| 9968 | for (unsigned I = 0; I != 2; ++I) { |
| 9969 | if (auto *Operation = dyn_cast<InstTy>(PN->getIncomingValue(i: I)); |
| 9970 | Operation && Operation->getNumOperands() >= 3) { |
| 9971 | Value *Op0 = Operation->getOperand(0); |
| 9972 | Value *Op1 = Operation->getOperand(1); |
| 9973 | Value *Op2 = Operation->getOperand(2); |
| 9974 | |
| 9975 | if (Op0 != PN && Op1 != PN && Op2 != PN) |
| 9976 | continue; |
| 9977 | |
| 9978 | Inst = Operation; |
| 9979 | Init = PN->getIncomingValue(i: !I); |
| 9980 | if (Op0 == PN) { |
| 9981 | OtherOp0 = Op1; |
| 9982 | OtherOp1 = Op2; |
| 9983 | } else if (Op1 == PN) { |
| 9984 | OtherOp0 = Op0; |
| 9985 | OtherOp1 = Op2; |
| 9986 | } else { |
| 9987 | OtherOp0 = Op0; |
| 9988 | OtherOp1 = Op1; |
| 9989 | } |
| 9990 | return true; |
| 9991 | } |
| 9992 | } |
| 9993 | return false; |
| 9994 | } |
| 9995 | bool llvm::matchSimpleRecurrence(const PHINode *P, BinaryOperator *&BO, |
| 9996 | Value *&Start, Value *&Step) { |
| 9997 | // We try to match a recurrence of the form: |
| 9998 | // %iv = [Start, %entry], [%iv.next, %backedge] |
| 9999 | // %iv.next = binop %iv, Step |
| 10000 | // Or: |
| 10001 | // %iv = [Start, %entry], [%iv.next, %backedge] |
| 10002 | // %iv.next = binop Step, %iv |
| 10003 | return matchTwoInputRecurrence(PN: P, Inst&: BO, Init&: Start, OtherOp&: Step); |
| 10004 | } |
| 10005 | |
| 10006 | bool llvm::matchSimpleRecurrence(const BinaryOperator *I, PHINode *&P, |
| 10007 | Value *&Start, Value *&Step) { |
| 10008 | BinaryOperator *BO = nullptr; |
| 10009 | return match(V: I, P: m_c_BinOp(L: m_Phi(PN&: P), R: m_Value())) && |
| 10010 | matchSimpleRecurrence(P, BO, Start, Step) && BO == I; |
| 10011 | } |
| 10012 | |
| 10013 | bool llvm::matchSimpleBinaryIntrinsicRecurrence(const IntrinsicInst *I, |
| 10014 | PHINode *&P, Value *&Init, |
| 10015 | Value *&OtherOp) { |
| 10016 | // Binary intrinsics only supported for now. |
| 10017 | if (I->arg_size() != 2 || I->getType() != I->getArgOperand(i: 0)->getType() || |
| 10018 | I->getType() != I->getArgOperand(i: 1)->getType()) |
| 10019 | return false; |
| 10020 | |
| 10021 | IntrinsicInst *II = nullptr; |
| 10022 | P = dyn_cast<PHINode>(Val: I->getArgOperand(i: 0)); |
| 10023 | if (!P) |
| 10024 | P = dyn_cast<PHINode>(Val: I->getArgOperand(i: 1)); |
| 10025 | |
| 10026 | return P && matchTwoInputRecurrence(PN: P, Inst&: II, Init, OtherOp) && II == I; |
| 10027 | } |
| 10028 | |
| 10029 | bool llvm::matchSimpleTernaryIntrinsicRecurrence(const IntrinsicInst *I, |
| 10030 | PHINode *&P, Value *&Init, |
| 10031 | Value *&OtherOp0, |
| 10032 | Value *&OtherOp1) { |
| 10033 | if (I->arg_size() != 3 || I->getType() != I->getArgOperand(i: 0)->getType() || |
| 10034 | I->getType() != I->getArgOperand(i: 1)->getType() || |
| 10035 | I->getType() != I->getArgOperand(i: 2)->getType()) |
| 10036 | return false; |
| 10037 | IntrinsicInst *II = nullptr; |
| 10038 | P = dyn_cast<PHINode>(Val: I->getArgOperand(i: 0)); |
| 10039 | if (!P) { |
| 10040 | P = dyn_cast<PHINode>(Val: I->getArgOperand(i: 1)); |
| 10041 | if (!P) |
| 10042 | P = dyn_cast<PHINode>(Val: I->getArgOperand(i: 2)); |
| 10043 | } |
| 10044 | return P && matchThreeInputRecurrence(PN: P, Inst&: II, Init, OtherOp0, OtherOp1) && |
| 10045 | II == I; |
| 10046 | } |
| 10047 | |
| 10048 | /// Return true if "icmp Pred LHS RHS" is always true. |
| 10049 | static bool isTruePredicate(CmpInst::Predicate Pred, const Value *LHS, |
| 10050 | const Value *RHS) { |
| 10051 | if (ICmpInst::isTrueWhenEqual(predicate: Pred) && LHS == RHS) |
| 10052 | return true; |
| 10053 | |
| 10054 | switch (Pred) { |
| 10055 | default: |
| 10056 | return false; |
| 10057 | |
| 10058 | case CmpInst::ICMP_SLE: { |
| 10059 | const APInt *C; |
| 10060 | |
| 10061 | // LHS s<= LHS +_{nsw} C if C >= 0 |
| 10062 | // LHS s<= LHS | C if C >= 0 |
| 10063 | if (match(V: RHS, P: m_NSWAdd(L: m_Specific(V: LHS), R: m_APInt(Res&: C))) || |
| 10064 | match(V: RHS, P: m_Or(L: m_Specific(V: LHS), R: m_APInt(Res&: C)))) |
| 10065 | return !C->isNegative(); |
| 10066 | |
| 10067 | // LHS s<= smax(LHS, V) for any V |
| 10068 | if (match(V: RHS, P: m_c_SMax(L: m_Specific(V: LHS), R: m_Value()))) |
| 10069 | return true; |
| 10070 | |
| 10071 | // smin(RHS, V) s<= RHS for any V |
| 10072 | if (match(V: LHS, P: m_c_SMin(L: m_Specific(V: RHS), R: m_Value()))) |
| 10073 | return true; |
| 10074 | |
| 10075 | // Match A to (X +_{nsw} CA) and B to (X +_{nsw} CB) |
| 10076 | const Value *X; |
| 10077 | const APInt *CLHS, *CRHS; |
| 10078 | if (match(V: LHS, P: m_NSWAddLike(L: m_Value(V&: X), R: m_APInt(Res&: CLHS))) && |
| 10079 | match(V: RHS, P: m_NSWAddLike(L: m_Specific(V: X), R: m_APInt(Res&: CRHS)))) |
| 10080 | return CLHS->sle(RHS: *CRHS); |
| 10081 | |
| 10082 | return false; |
| 10083 | } |
| 10084 | |
| 10085 | case CmpInst::ICMP_ULE: { |
| 10086 | // LHS u<= LHS +_{nuw} V for any V |
| 10087 | if (match(V: RHS, P: m_c_Add(L: m_Specific(V: LHS), R: m_Value())) && |
| 10088 | cast<OverflowingBinaryOperator>(Val: RHS)->hasNoUnsignedWrap()) |
| 10089 | return true; |
| 10090 | |
| 10091 | // LHS u<= LHS | V for any V |
| 10092 | if (match(V: RHS, P: m_c_Or(L: m_Specific(V: LHS), R: m_Value()))) |
| 10093 | return true; |
| 10094 | |
| 10095 | // LHS u<= umax(LHS, V) for any V |
| 10096 | if (match(V: RHS, P: m_c_UMax(L: m_Specific(V: LHS), R: m_Value()))) |
| 10097 | return true; |
| 10098 | |
| 10099 | // RHS >> V u<= RHS for any V |
| 10100 | if (match(V: LHS, P: m_LShr(L: m_Specific(V: RHS), R: m_Value()))) |
| 10101 | return true; |
| 10102 | |
| 10103 | // RHS u/ C_ugt_1 u<= RHS |
| 10104 | const APInt *C; |
| 10105 | if (match(V: LHS, P: m_UDiv(L: m_Specific(V: RHS), R: m_APInt(Res&: C))) && C->ugt(RHS: 1)) |
| 10106 | return true; |
| 10107 | |
| 10108 | // RHS & V u<= RHS for any V |
| 10109 | if (match(V: LHS, P: m_c_And(L: m_Specific(V: RHS), R: m_Value()))) |
| 10110 | return true; |
| 10111 | |
| 10112 | // umin(RHS, V) u<= RHS for any V |
| 10113 | if (match(V: LHS, P: m_c_UMin(L: m_Specific(V: RHS), R: m_Value()))) |
| 10114 | return true; |
| 10115 | |
| 10116 | // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB) |
| 10117 | const Value *X; |
| 10118 | const APInt *CLHS, *CRHS; |
| 10119 | if (match(V: LHS, P: m_NUWAddLike(L: m_Value(V&: X), R: m_APInt(Res&: CLHS))) && |
| 10120 | match(V: RHS, P: m_NUWAddLike(L: m_Specific(V: X), R: m_APInt(Res&: CRHS)))) |
| 10121 | return CLHS->ule(RHS: *CRHS); |
| 10122 | |
| 10123 | return false; |
| 10124 | } |
| 10125 | } |
| 10126 | } |
| 10127 | |
| 10128 | /// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred |
| 10129 | /// ALHS ARHS" is true. Otherwise, return std::nullopt. |
| 10130 | static std::optional<bool> |
| 10131 | isImpliedCondOperands(CmpInst::Predicate Pred, const Value *ALHS, |
| 10132 | const Value *ARHS, const Value *BLHS, const Value *BRHS) { |
| 10133 | switch (Pred) { |
| 10134 | default: |
| 10135 | return std::nullopt; |
| 10136 | |
| 10137 | case CmpInst::ICMP_SLT: |
| 10138 | case CmpInst::ICMP_SLE: |
| 10139 | if (isTruePredicate(Pred: CmpInst::ICMP_SLE, LHS: BLHS, RHS: ALHS) && |
| 10140 | isTruePredicate(Pred: CmpInst::ICMP_SLE, LHS: ARHS, RHS: BRHS)) |
| 10141 | return true; |
| 10142 | return std::nullopt; |
| 10143 | |
| 10144 | case CmpInst::ICMP_SGT: |
| 10145 | case CmpInst::ICMP_SGE: |
| 10146 | if (isTruePredicate(Pred: CmpInst::ICMP_SLE, LHS: ALHS, RHS: BLHS) && |
| 10147 | isTruePredicate(Pred: CmpInst::ICMP_SLE, LHS: BRHS, RHS: ARHS)) |
| 10148 | return true; |
| 10149 | return std::nullopt; |
| 10150 | |
| 10151 | case CmpInst::ICMP_ULT: |
| 10152 | case CmpInst::ICMP_ULE: |
| 10153 | if (isTruePredicate(Pred: CmpInst::ICMP_ULE, LHS: BLHS, RHS: ALHS) && |
| 10154 | isTruePredicate(Pred: CmpInst::ICMP_ULE, LHS: ARHS, RHS: BRHS)) |
| 10155 | return true; |
| 10156 | return std::nullopt; |
| 10157 | |
| 10158 | case CmpInst::ICMP_UGT: |
| 10159 | case CmpInst::ICMP_UGE: |
| 10160 | if (isTruePredicate(Pred: CmpInst::ICMP_ULE, LHS: ALHS, RHS: BLHS) && |
| 10161 | isTruePredicate(Pred: CmpInst::ICMP_ULE, LHS: BRHS, RHS: ARHS)) |
| 10162 | return true; |
| 10163 | return std::nullopt; |
| 10164 | } |
| 10165 | } |
| 10166 | |
| 10167 | /// Return true if "icmp LPred X, LCR" implies "icmp RPred X, RCR" is true. |
| 10168 | /// Return false if "icmp LPred X, LCR" implies "icmp RPred X, RCR" is false. |
| 10169 | /// Otherwise, return std::nullopt if we can't infer anything. |
| 10170 | static std::optional<bool> |
| 10171 | isImpliedCondCommonOperandWithCR(CmpPredicate LPred, const ConstantRange &LCR, |
| 10172 | CmpPredicate RPred, const ConstantRange &RCR) { |
| 10173 | auto CRImpliesPred = [&](ConstantRange CR, |
| 10174 | CmpInst::Predicate Pred) -> std::optional<bool> { |
| 10175 | // If all true values for lhs and true for rhs, lhs implies rhs |
| 10176 | if (CR.icmp(Pred, Other: RCR)) |
| 10177 | return true; |
| 10178 | |
| 10179 | // If there is no overlap, lhs implies not rhs |
| 10180 | if (CR.icmp(Pred: CmpInst::getInversePredicate(pred: Pred), Other: RCR)) |
| 10181 | return false; |
| 10182 | |
| 10183 | return std::nullopt; |
| 10184 | }; |
| 10185 | if (auto Res = CRImpliesPred(ConstantRange::makeAllowedICmpRegion(Pred: LPred, Other: LCR), |
| 10186 | RPred)) |
| 10187 | return Res; |
| 10188 | if (LPred.hasSameSign() ^ RPred.hasSameSign()) { |
| 10189 | LPred = LPred.hasSameSign() ? ICmpInst::getFlippedSignednessPredicate(Pred: LPred) |
| 10190 | : LPred.dropSameSign(); |
| 10191 | RPred = RPred.hasSameSign() ? ICmpInst::getFlippedSignednessPredicate(Pred: RPred) |
| 10192 | : RPred.dropSameSign(); |
| 10193 | return CRImpliesPred(ConstantRange::makeAllowedICmpRegion(Pred: LPred, Other: LCR), |
| 10194 | RPred); |
| 10195 | } |
| 10196 | return std::nullopt; |
| 10197 | } |
| 10198 | |
| 10199 | /// Return true if LHS implies RHS (expanded to its components as "R0 RPred R1") |
| 10200 | /// is true. Return false if LHS implies RHS is false. Otherwise, return |
| 10201 | /// std::nullopt if we can't infer anything. |
| 10202 | static std::optional<bool> |
| 10203 | isImpliedCondICmps(CmpPredicate LPred, const Value *L0, const Value *L1, |
| 10204 | CmpPredicate RPred, const Value *R0, const Value *R1, |
| 10205 | const DataLayout &DL, bool LHSIsTrue) { |
| 10206 | // The rest of the logic assumes the LHS condition is true. If that's not the |
| 10207 | // case, invert the predicate to make it so. |
| 10208 | if (!LHSIsTrue) |
| 10209 | LPred = ICmpInst::getInverseCmpPredicate(Pred: LPred); |
| 10210 | |
| 10211 | // We can have non-canonical operands, so try to normalize any common operand |
| 10212 | // to L0/R0. |
| 10213 | if (L0 == R1) { |
| 10214 | std::swap(a&: R0, b&: R1); |
| 10215 | RPred = ICmpInst::getSwappedCmpPredicate(Pred: RPred); |
| 10216 | } |
| 10217 | if (R0 == L1) { |
| 10218 | std::swap(a&: L0, b&: L1); |
| 10219 | LPred = ICmpInst::getSwappedCmpPredicate(Pred: LPred); |
| 10220 | } |
| 10221 | if (L1 == R1) { |
| 10222 | // If we have L0 == R0 and L1 == R1, then make L1/R1 the constants. |
| 10223 | if (L0 != R0 || match(V: L0, P: m_ImmConstant())) { |
| 10224 | std::swap(a&: L0, b&: L1); |
| 10225 | LPred = ICmpInst::getSwappedCmpPredicate(Pred: LPred); |
| 10226 | std::swap(a&: R0, b&: R1); |
| 10227 | RPred = ICmpInst::getSwappedCmpPredicate(Pred: RPred); |
| 10228 | } |
| 10229 | } |
| 10230 | |
| 10231 | // See if we can infer anything if operand-0 matches and we have at least one |
| 10232 | // constant. |
| 10233 | const APInt *Unused; |
| 10234 | if (L0 == R0 && (match(V: L1, P: m_APInt(Res&: Unused)) || match(V: R1, P: m_APInt(Res&: Unused)))) { |
| 10235 | // Potential TODO: We could also further use the constant range of L0/R0 to |
| 10236 | // further constraint the constant ranges. At the moment this leads to |
| 10237 | // several regressions related to not transforming `multi_use(A + C0) eq/ne |
| 10238 | // C1` (see discussion: D58633). |
| 10239 | SimplifyQuery SQ(DL); |
| 10240 | ConstantRange LCR = computeConstantRange(V: L1, ForSigned: ICmpInst::isSigned(Pred: LPred), SQ, |
| 10241 | Depth: MaxAnalysisRecursionDepth - 1); |
| 10242 | ConstantRange RCR = computeConstantRange(V: R1, ForSigned: ICmpInst::isSigned(Pred: RPred), SQ, |
| 10243 | Depth: MaxAnalysisRecursionDepth - 1); |
| 10244 | |
| 10245 | // Even if L1/R1 are not both constant, we can still sometimes deduce |
| 10246 | // relationship from a single constant. For example X u> Y implies X != 0. |
| 10247 | if (auto R = isImpliedCondCommonOperandWithCR(LPred, LCR, RPred, RCR)) |
| 10248 | return R; |
| 10249 | // If both L1/R1 were exact constant ranges and we didn't get anything |
| 10250 | // here, we won't be able to deduce this. |
| 10251 | if (match(V: L1, P: m_APInt(Res&: Unused)) && match(V: R1, P: m_APInt(Res&: Unused))) |
| 10252 | return std::nullopt; |
| 10253 | } |
| 10254 | |
| 10255 | // Can we infer anything when the two compares have matching operands? |
| 10256 | if (L0 == R0 && L1 == R1) |
| 10257 | return ICmpInst::isImpliedByMatchingCmp(Pred1: LPred, Pred2: RPred); |
| 10258 | |
| 10259 | // It only really makes sense in the context of signed comparison for "X - Y |
| 10260 | // must be positive if X >= Y and no overflow". |
| 10261 | // Take SGT as an example: L0:x > L1:y and C >= 0 |
| 10262 | // ==> R0:(x -nsw y) < R1:(-C) is false |
| 10263 | CmpInst::Predicate SignedLPred = LPred.getPreferredSignedPredicate(); |
| 10264 | if ((SignedLPred == ICmpInst::ICMP_SGT || |
| 10265 | SignedLPred == ICmpInst::ICMP_SGE) && |
| 10266 | match(V: R0, P: m_NSWSub(L: m_Specific(V: L0), R: m_Specific(V: L1)))) { |
| 10267 | if (match(V: R1, P: m_NonPositive()) && |
| 10268 | ICmpInst::isImpliedByMatchingCmp(Pred1: SignedLPred, Pred2: RPred) == false) |
| 10269 | return false; |
| 10270 | } |
| 10271 | |
| 10272 | // Take SLT as an example: L0:x < L1:y and C <= 0 |
| 10273 | // ==> R0:(x -nsw y) < R1:(-C) is true |
| 10274 | if ((SignedLPred == ICmpInst::ICMP_SLT || |
| 10275 | SignedLPred == ICmpInst::ICMP_SLE) && |
| 10276 | match(V: R0, P: m_NSWSub(L: m_Specific(V: L0), R: m_Specific(V: L1)))) { |
| 10277 | if (match(V: R1, P: m_NonNegative()) && |
| 10278 | ICmpInst::isImpliedByMatchingCmp(Pred1: SignedLPred, Pred2: RPred) == true) |
| 10279 | return true; |
| 10280 | } |
| 10281 | |
| 10282 | // (a - b) pred C -> a != b, if 0 pred C is false |
| 10283 | // (ptrtoint(a) - ptrtoint(b)) pred C -> a != b, if 0 pred C is false |
| 10284 | Value *A, *B; |
| 10285 | if (ICmpInst::isEquality(P: RPred) && cmpExcludesZero(Pred: LPred, RHS: L1) && |
| 10286 | match(V: L0, P: m_Sub(L: m_Value(V&: A), R: m_Value(V&: B))) && |
| 10287 | ((A == R0 && B == R1) || (A == R1 && B == R0) || |
| 10288 | (match(V: A, P: m_PtrToIntOrAddr(Op: m_Specific(V: R0))) && |
| 10289 | match(V: B, P: m_PtrToIntOrAddr(Op: m_Specific(V: R1)))) || |
| 10290 | (match(V: A, P: m_PtrToIntOrAddr(Op: m_Specific(V: R1))) && |
| 10291 | match(V: B, P: m_PtrToIntOrAddr(Op: m_Specific(V: R0)))))) { |
| 10292 | return RPred.dropSameSign() == ICmpInst::ICMP_NE; |
| 10293 | } |
| 10294 | |
| 10295 | // L0 = R0 = L1 + R1, L0 >=u L1 implies R0 >=u R1, L0 <u L1 implies R0 <u R1 |
| 10296 | if (L0 == R0 && |
| 10297 | (LPred == ICmpInst::ICMP_ULT || LPred == ICmpInst::ICMP_UGE) && |
| 10298 | (RPred == ICmpInst::ICMP_ULT || RPred == ICmpInst::ICMP_UGE) && |
| 10299 | match(V: L0, P: m_c_Add(L: m_Specific(V: L1), R: m_Specific(V: R1)))) |
| 10300 | return CmpPredicate::getMatching(A: LPred, B: RPred).has_value(); |
| 10301 | |
| 10302 | if (auto P = CmpPredicate::getMatching(A: LPred, B: RPred)) |
| 10303 | return isImpliedCondOperands(Pred: *P, ALHS: L0, ARHS: L1, BLHS: R0, BRHS: R1); |
| 10304 | |
| 10305 | // L0 u< C sets limits to L0's bits which may imply (L0 & Mask) pred RC |
| 10306 | // Example: L0 u< 13 => (L0 & 16) == 0 |
| 10307 | const APInt *LC, *RC, *MaskC; |
| 10308 | if (match(V: L1, P: m_APInt(Res&: LC)) && match(V: R1, P: m_APInt(Res&: RC)) && |
| 10309 | match(V: R0, P: m_And(L: m_Specific(V: L0), R: m_APInt(Res&: MaskC)))) { |
| 10310 | ConstantRange LCRange = ConstantRange::makeExactICmpRegion(Pred: LPred, Other: *LC); |
| 10311 | ConstantRange MaskedCRange = LCRange.binaryAnd(Other: *MaskC); |
| 10312 | if (MaskedCRange.icmp(Pred: RPred, Other: ConstantRange(*RC))) |
| 10313 | return true; |
| 10314 | if (MaskedCRange.icmp(Pred: ICmpInst::getInversePredicate(pred: RPred), |
| 10315 | Other: ConstantRange(*RC))) |
| 10316 | return false; |
| 10317 | } |
| 10318 | |
| 10319 | return std::nullopt; |
| 10320 | } |
| 10321 | |
| 10322 | /// Return true if LHS implies RHS (expanded to its components as "R0 RPred R1") |
| 10323 | /// is true. Return false if LHS implies RHS is false. Otherwise, return |
| 10324 | /// std::nullopt if we can't infer anything. |
| 10325 | static std::optional<bool> |
| 10326 | isImpliedCondFCmps(FCmpInst::Predicate LPred, const Value *L0, const Value *L1, |
| 10327 | FCmpInst::Predicate RPred, const Value *R0, const Value *R1, |
| 10328 | const DataLayout &DL, bool LHSIsTrue) { |
| 10329 | // The rest of the logic assumes the LHS condition is true. If that's not the |
| 10330 | // case, invert the predicate to make it so. |
| 10331 | if (!LHSIsTrue) |
| 10332 | LPred = FCmpInst::getInversePredicate(pred: LPred); |
| 10333 | |
| 10334 | // We can have non-canonical operands, so try to normalize any common operand |
| 10335 | // to L0/R0. |
| 10336 | if (L0 == R1) { |
| 10337 | std::swap(a&: R0, b&: R1); |
| 10338 | RPred = FCmpInst::getSwappedPredicate(pred: RPred); |
| 10339 | } |
| 10340 | if (R0 == L1) { |
| 10341 | std::swap(a&: L0, b&: L1); |
| 10342 | LPred = FCmpInst::getSwappedPredicate(pred: LPred); |
| 10343 | } |
| 10344 | if (L1 == R1) { |
| 10345 | // If we have L0 == R0 and L1 == R1, then make L1/R1 the constants. |
| 10346 | if (L0 != R0 || match(V: L0, P: m_ImmConstant())) { |
| 10347 | std::swap(a&: L0, b&: L1); |
| 10348 | LPred = ICmpInst::getSwappedCmpPredicate(Pred: LPred); |
| 10349 | std::swap(a&: R0, b&: R1); |
| 10350 | RPred = ICmpInst::getSwappedCmpPredicate(Pred: RPred); |
| 10351 | } |
| 10352 | } |
| 10353 | |
| 10354 | // Can we infer anything when the two compares have matching operands? |
| 10355 | if (L0 == R0 && L1 == R1) { |
| 10356 | if ((LPred & RPred) == LPred) |
| 10357 | return true; |
| 10358 | if ((LPred & ~RPred) == LPred) |
| 10359 | return false; |
| 10360 | } |
| 10361 | |
| 10362 | // See if we can infer anything if operand-0 matches and we have at least one |
| 10363 | // constant. |
| 10364 | const APFloat *L1C, *R1C; |
| 10365 | if (L0 == R0 && match(V: L1, P: m_APFloat(Res&: L1C)) && match(V: R1, P: m_APFloat(Res&: R1C))) { |
| 10366 | if (std::optional<ConstantFPRange> DomCR = |
| 10367 | ConstantFPRange::makeExactFCmpRegion(Pred: LPred, Other: *L1C)) { |
| 10368 | if (std::optional<ConstantFPRange> ImpliedCR = |
| 10369 | ConstantFPRange::makeExactFCmpRegion(Pred: RPred, Other: *R1C)) { |
| 10370 | if (ImpliedCR->contains(CR: *DomCR)) |
| 10371 | return true; |
| 10372 | } |
| 10373 | if (std::optional<ConstantFPRange> ImpliedCR = |
| 10374 | ConstantFPRange::makeExactFCmpRegion( |
| 10375 | Pred: FCmpInst::getInversePredicate(pred: RPred), Other: *R1C)) { |
| 10376 | if (ImpliedCR->contains(CR: *DomCR)) |
| 10377 | return false; |
| 10378 | } |
| 10379 | } |
| 10380 | } |
| 10381 | |
| 10382 | return std::nullopt; |
| 10383 | } |
| 10384 | |
| 10385 | /// Return true if LHS implies RHS is true. Return false if LHS implies RHS is |
| 10386 | /// false. Otherwise, return std::nullopt if we can't infer anything. We |
| 10387 | /// expect the RHS to be an icmp and the LHS to be an 'and', 'or', or a 'select' |
| 10388 | /// instruction. |
| 10389 | static std::optional<bool> |
| 10390 | isImpliedCondAndOr(const Instruction *LHS, CmpPredicate RHSPred, |
| 10391 | const Value *RHSOp0, const Value *RHSOp1, |
| 10392 | const DataLayout &DL, bool LHSIsTrue, unsigned Depth) { |
| 10393 | // The LHS must be an 'or', 'and', or a 'select' instruction. |
| 10394 | assert((LHS->getOpcode() == Instruction::And || |
| 10395 | LHS->getOpcode() == Instruction::Or || |
| 10396 | LHS->getOpcode() == Instruction::Select) && |
| 10397 | "Expected LHS to be 'and', 'or', or 'select'." ); |
| 10398 | |
| 10399 | assert(Depth <= MaxAnalysisRecursionDepth && "Hit recursion limit" ); |
| 10400 | |
| 10401 | // If the result of an 'or' is false, then we know both legs of the 'or' are |
| 10402 | // false. Similarly, if the result of an 'and' is true, then we know both |
| 10403 | // legs of the 'and' are true. |
| 10404 | const Value *ALHS, *ARHS; |
| 10405 | if ((!LHSIsTrue && match(V: LHS, P: m_LogicalOr(L: m_Value(V&: ALHS), R: m_Value(V&: ARHS)))) || |
| 10406 | (LHSIsTrue && match(V: LHS, P: m_LogicalAnd(L: m_Value(V&: ALHS), R: m_Value(V&: ARHS))))) { |
| 10407 | // FIXME: Make this non-recursion. |
| 10408 | if (std::optional<bool> Implication = isImpliedCondition( |
| 10409 | LHS: ALHS, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, Depth: Depth + 1)) |
| 10410 | return Implication; |
| 10411 | if (std::optional<bool> Implication = isImpliedCondition( |
| 10412 | LHS: ARHS, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, Depth: Depth + 1)) |
| 10413 | return Implication; |
| 10414 | return std::nullopt; |
| 10415 | } |
| 10416 | return std::nullopt; |
| 10417 | } |
| 10418 | |
| 10419 | std::optional<bool> |
| 10420 | llvm::isImpliedCondition(const Value *LHS, CmpPredicate RHSPred, |
| 10421 | const Value *RHSOp0, const Value *RHSOp1, |
| 10422 | const DataLayout &DL, bool LHSIsTrue, unsigned Depth) { |
| 10423 | // Bail out when we hit the limit. |
| 10424 | if (Depth == MaxAnalysisRecursionDepth) |
| 10425 | return std::nullopt; |
| 10426 | |
| 10427 | // A mismatch occurs when we compare a scalar cmp to a vector cmp, for |
| 10428 | // example. |
| 10429 | if (RHSOp0->getType()->isVectorTy() != LHS->getType()->isVectorTy()) |
| 10430 | return std::nullopt; |
| 10431 | |
| 10432 | assert(LHS->getType()->isIntOrIntVectorTy(1) && |
| 10433 | "Expected integer type only!" ); |
| 10434 | |
| 10435 | // Match not |
| 10436 | if (match(V: LHS, P: m_Not(V: m_Value(V&: LHS)))) |
| 10437 | LHSIsTrue = !LHSIsTrue; |
| 10438 | |
| 10439 | // Both LHS and RHS are icmps. |
| 10440 | if (RHSOp0->getType()->getScalarType()->isIntOrPtrTy()) { |
| 10441 | CmpPredicate LHSPred; |
| 10442 | Value *LHSOp0, *LHSOp1; |
| 10443 | if (match(V: LHS, P: m_ICmpLike(Pred&: LHSPred, L: m_Value(V&: LHSOp0), R: m_Value(V&: LHSOp1)))) |
| 10444 | return isImpliedCondICmps(LPred: LHSPred, L0: LHSOp0, L1: LHSOp1, RPred: RHSPred, R0: RHSOp0, |
| 10445 | R1: RHSOp1, DL, LHSIsTrue); |
| 10446 | } else { |
| 10447 | assert(RHSOp0->getType()->isFPOrFPVectorTy() && |
| 10448 | "Expected floating point type only!" ); |
| 10449 | if (const auto *LHSCmp = dyn_cast<FCmpInst>(Val: LHS)) |
| 10450 | return isImpliedCondFCmps(LPred: LHSCmp->getPredicate(), L0: LHSCmp->getOperand(i_nocapture: 0), |
| 10451 | L1: LHSCmp->getOperand(i_nocapture: 1), RPred: RHSPred, R0: RHSOp0, R1: RHSOp1, |
| 10452 | DL, LHSIsTrue); |
| 10453 | } |
| 10454 | |
| 10455 | /// The LHS should be an 'or', 'and', or a 'select' instruction. We expect |
| 10456 | /// the RHS to be an icmp. |
| 10457 | /// FIXME: Add support for and/or/select on the RHS. |
| 10458 | if (const Instruction *LHSI = dyn_cast<Instruction>(Val: LHS)) { |
| 10459 | if ((LHSI->getOpcode() == Instruction::And || |
| 10460 | LHSI->getOpcode() == Instruction::Or || |
| 10461 | LHSI->getOpcode() == Instruction::Select)) |
| 10462 | return isImpliedCondAndOr(LHS: LHSI, RHSPred, RHSOp0, RHSOp1, DL, LHSIsTrue, |
| 10463 | Depth); |
| 10464 | } |
| 10465 | return std::nullopt; |
| 10466 | } |
| 10467 | |
| 10468 | std::optional<bool> llvm::isImpliedCondition(const Value *LHS, const Value *RHS, |
| 10469 | const DataLayout &DL, |
| 10470 | bool LHSIsTrue, unsigned Depth) { |
| 10471 | // LHS ==> RHS by definition |
| 10472 | if (LHS == RHS) |
| 10473 | return LHSIsTrue; |
| 10474 | |
| 10475 | // Match not |
| 10476 | bool InvertRHS = false; |
| 10477 | if (match(V: RHS, P: m_Not(V: m_Value(V&: RHS)))) { |
| 10478 | if (LHS == RHS) |
| 10479 | return !LHSIsTrue; |
| 10480 | InvertRHS = true; |
| 10481 | } |
| 10482 | |
| 10483 | CmpPredicate RHSPred; |
| 10484 | Value *RHSOp0, *RHSOp1; |
| 10485 | if (match(V: RHS, P: m_ICmpLike(Pred&: RHSPred, L: m_Value(V&: RHSOp0), R: m_Value(V&: RHSOp1)))) { |
| 10486 | if (auto Implied = isImpliedCondition(LHS, RHSPred, RHSOp0, RHSOp1, DL, |
| 10487 | LHSIsTrue, Depth)) |
| 10488 | return InvertRHS ? !*Implied : *Implied; |
| 10489 | return std::nullopt; |
| 10490 | } |
| 10491 | if (const FCmpInst *RHSCmp = dyn_cast<FCmpInst>(Val: RHS)) { |
| 10492 | if (auto Implied = isImpliedCondition( |
| 10493 | LHS, RHSPred: RHSCmp->getPredicate(), RHSOp0: RHSCmp->getOperand(i_nocapture: 0), |
| 10494 | RHSOp1: RHSCmp->getOperand(i_nocapture: 1), DL, LHSIsTrue, Depth)) |
| 10495 | return InvertRHS ? !*Implied : *Implied; |
| 10496 | return std::nullopt; |
| 10497 | } |
| 10498 | |
| 10499 | if (Depth == MaxAnalysisRecursionDepth) |
| 10500 | return std::nullopt; |
| 10501 | |
| 10502 | // LHS ==> (RHS1 || RHS2) if LHS ==> RHS1 or LHS ==> RHS2 |
| 10503 | // LHS ==> !(RHS1 && RHS2) if LHS ==> !RHS1 or LHS ==> !RHS2 |
| 10504 | const Value *RHS1, *RHS2; |
| 10505 | if (match(V: RHS, P: m_LogicalOr(L: m_Value(V&: RHS1), R: m_Value(V&: RHS2)))) { |
| 10506 | if (std::optional<bool> Imp = |
| 10507 | isImpliedCondition(LHS, RHS: RHS1, DL, LHSIsTrue, Depth: Depth + 1)) |
| 10508 | if (*Imp == true) |
| 10509 | return !InvertRHS; |
| 10510 | if (std::optional<bool> Imp = |
| 10511 | isImpliedCondition(LHS, RHS: RHS2, DL, LHSIsTrue, Depth: Depth + 1)) |
| 10512 | if (*Imp == true) |
| 10513 | return !InvertRHS; |
| 10514 | } |
| 10515 | if (match(V: RHS, P: m_LogicalAnd(L: m_Value(V&: RHS1), R: m_Value(V&: RHS2)))) { |
| 10516 | if (std::optional<bool> Imp = |
| 10517 | isImpliedCondition(LHS, RHS: RHS1, DL, LHSIsTrue, Depth: Depth + 1)) |
| 10518 | if (*Imp == false) |
| 10519 | return InvertRHS; |
| 10520 | if (std::optional<bool> Imp = |
| 10521 | isImpliedCondition(LHS, RHS: RHS2, DL, LHSIsTrue, Depth: Depth + 1)) |
| 10522 | if (*Imp == false) |
| 10523 | return InvertRHS; |
| 10524 | } |
| 10525 | |
| 10526 | return std::nullopt; |
| 10527 | } |
| 10528 | |
| 10529 | // Returns a pair (Condition, ConditionIsTrue), where Condition is a branch |
| 10530 | // condition dominating ContextI or nullptr, if no condition is found. |
| 10531 | static std::pair<Value *, bool> |
| 10532 | getDomPredecessorCondition(const Instruction *ContextI) { |
| 10533 | if (!ContextI || !ContextI->getParent()) |
| 10534 | return {nullptr, false}; |
| 10535 | |
| 10536 | // TODO: This is a poor/cheap way to determine dominance. Should we use a |
| 10537 | // dominator tree (eg, from a SimplifyQuery) instead? |
| 10538 | const BasicBlock *ContextBB = ContextI->getParent(); |
| 10539 | const BasicBlock *PredBB = ContextBB->getSinglePredecessor(); |
| 10540 | if (!PredBB) |
| 10541 | return {nullptr, false}; |
| 10542 | |
| 10543 | // We need a conditional branch in the predecessor. |
| 10544 | Value *PredCond; |
| 10545 | BasicBlock *TrueBB, *FalseBB; |
| 10546 | if (!match(V: PredBB->getTerminator(), P: m_Br(C: m_Value(V&: PredCond), T&: TrueBB, F&: FalseBB))) |
| 10547 | return {nullptr, false}; |
| 10548 | |
| 10549 | // The branch should get simplified. Don't bother simplifying this condition. |
| 10550 | if (TrueBB == FalseBB) |
| 10551 | return {nullptr, false}; |
| 10552 | |
| 10553 | assert((TrueBB == ContextBB || FalseBB == ContextBB) && |
| 10554 | "Predecessor block does not point to successor?" ); |
| 10555 | |
| 10556 | // Is this condition implied by the predecessor condition? |
| 10557 | return {PredCond, TrueBB == ContextBB}; |
| 10558 | } |
| 10559 | |
| 10560 | std::optional<bool> llvm::isImpliedByDomCondition(const Value *Cond, |
| 10561 | const Instruction *ContextI, |
| 10562 | const DataLayout &DL) { |
| 10563 | assert(Cond->getType()->isIntOrIntVectorTy(1) && "Condition must be bool" ); |
| 10564 | auto PredCond = getDomPredecessorCondition(ContextI); |
| 10565 | if (PredCond.first) |
| 10566 | return isImpliedCondition(LHS: PredCond.first, RHS: Cond, DL, LHSIsTrue: PredCond.second); |
| 10567 | return std::nullopt; |
| 10568 | } |
| 10569 | |
| 10570 | std::optional<bool> llvm::isImpliedByDomCondition(CmpPredicate Pred, |
| 10571 | const Value *LHS, |
| 10572 | const Value *RHS, |
| 10573 | const Instruction *ContextI, |
| 10574 | const DataLayout &DL) { |
| 10575 | auto PredCond = getDomPredecessorCondition(ContextI); |
| 10576 | if (PredCond.first) |
| 10577 | return isImpliedCondition(LHS: PredCond.first, RHSPred: Pred, RHSOp0: LHS, RHSOp1: RHS, DL, |
| 10578 | LHSIsTrue: PredCond.second); |
| 10579 | return std::nullopt; |
| 10580 | } |
| 10581 | |
| 10582 | static void setLimitsForBinOp(const BinaryOperator &BO, APInt &Lower, |
| 10583 | APInt &Upper, const InstrInfoQuery &IIQ, |
| 10584 | bool PreferSignedRange) { |
| 10585 | unsigned Width = Lower.getBitWidth(); |
| 10586 | const APInt *C; |
| 10587 | switch (BO.getOpcode()) { |
| 10588 | case Instruction::Sub: |
| 10589 | if (match(V: BO.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) { |
| 10590 | bool HasNSW = IIQ.hasNoSignedWrap(Op: &BO); |
| 10591 | bool HasNUW = IIQ.hasNoUnsignedWrap(Op: &BO); |
| 10592 | |
| 10593 | // If the caller expects a signed compare, then try to use a signed range. |
| 10594 | // Otherwise if both no-wraps are set, use the unsigned range because it |
| 10595 | // is never larger than the signed range. Example: |
| 10596 | // "sub nuw nsw i8 -2, x" is unsigned [0, 254] vs. signed [-128, 126]. |
| 10597 | // "sub nuw nsw i8 2, x" is unsigned [0, 2] vs. signed [-125, 127]. |
| 10598 | if (PreferSignedRange && HasNSW && HasNUW) |
| 10599 | HasNUW = false; |
| 10600 | |
| 10601 | if (HasNUW) { |
| 10602 | // 'sub nuw c, x' produces [0, C]. |
| 10603 | Upper = *C + 1; |
| 10604 | } else if (HasNSW) { |
| 10605 | if (C->isNegative()) { |
| 10606 | // 'sub nsw -C, x' produces [SINT_MIN, -C - SINT_MIN]. |
| 10607 | Lower = APInt::getSignedMinValue(numBits: Width); |
| 10608 | Upper = *C - APInt::getSignedMaxValue(numBits: Width); |
| 10609 | } else { |
| 10610 | // Note that sub 0, INT_MIN is not NSW. It techically is a signed wrap |
| 10611 | // 'sub nsw C, x' produces [C - SINT_MAX, SINT_MAX]. |
| 10612 | Lower = *C - APInt::getSignedMaxValue(numBits: Width); |
| 10613 | Upper = APInt::getSignedMinValue(numBits: Width); |
| 10614 | } |
| 10615 | } |
| 10616 | } |
| 10617 | break; |
| 10618 | case Instruction::Add: |
| 10619 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) && !C->isZero()) { |
| 10620 | bool HasNSW = IIQ.hasNoSignedWrap(Op: &BO); |
| 10621 | bool HasNUW = IIQ.hasNoUnsignedWrap(Op: &BO); |
| 10622 | |
| 10623 | // If the caller expects a signed compare, then try to use a signed |
| 10624 | // range. Otherwise if both no-wraps are set, use the unsigned range |
| 10625 | // because it is never larger than the signed range. Example: "add nuw |
| 10626 | // nsw i8 X, -2" is unsigned [254,255] vs. signed [-128, 125]. |
| 10627 | if (PreferSignedRange && HasNSW && HasNUW) |
| 10628 | HasNUW = false; |
| 10629 | |
| 10630 | if (HasNUW) { |
| 10631 | // 'add nuw x, C' produces [C, UINT_MAX]. |
| 10632 | Lower = *C; |
| 10633 | } else if (HasNSW) { |
| 10634 | if (C->isNegative()) { |
| 10635 | // 'add nsw x, -C' produces [SINT_MIN, SINT_MAX - C]. |
| 10636 | Lower = APInt::getSignedMinValue(numBits: Width); |
| 10637 | Upper = APInt::getSignedMaxValue(numBits: Width) + *C + 1; |
| 10638 | } else { |
| 10639 | // 'add nsw x, +C' produces [SINT_MIN + C, SINT_MAX]. |
| 10640 | Lower = APInt::getSignedMinValue(numBits: Width) + *C; |
| 10641 | Upper = APInt::getSignedMaxValue(numBits: Width) + 1; |
| 10642 | } |
| 10643 | } |
| 10644 | } |
| 10645 | break; |
| 10646 | |
| 10647 | case Instruction::And: |
| 10648 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) |
| 10649 | // 'and x, C' produces [0, C]. |
| 10650 | Upper = *C + 1; |
| 10651 | // X & -X is a power of two or zero. So we can cap the value at max power of |
| 10652 | // two. |
| 10653 | if (match(V: BO.getOperand(i_nocapture: 0), P: m_Neg(V: m_Specific(V: BO.getOperand(i_nocapture: 1)))) || |
| 10654 | match(V: BO.getOperand(i_nocapture: 1), P: m_Neg(V: m_Specific(V: BO.getOperand(i_nocapture: 0))))) |
| 10655 | Upper = APInt::getSignedMinValue(numBits: Width) + 1; |
| 10656 | break; |
| 10657 | |
| 10658 | case Instruction::Or: |
| 10659 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) |
| 10660 | // 'or x, C' produces [C, UINT_MAX]. |
| 10661 | Lower = *C; |
| 10662 | break; |
| 10663 | |
| 10664 | case Instruction::AShr: |
| 10665 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) && C->ult(RHS: Width)) { |
| 10666 | // 'ashr x, C' produces [INT_MIN >> C, INT_MAX >> C]. |
| 10667 | Lower = APInt::getSignedMinValue(numBits: Width).ashr(ShiftAmt: *C); |
| 10668 | Upper = APInt::getSignedMaxValue(numBits: Width).ashr(ShiftAmt: *C) + 1; |
| 10669 | } else if (match(V: BO.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) { |
| 10670 | unsigned ShiftAmount = Width - 1; |
| 10671 | if (!C->isZero() && IIQ.isExact(Op: &BO)) |
| 10672 | ShiftAmount = C->countr_zero(); |
| 10673 | if (C->isNegative()) { |
| 10674 | // 'ashr C, x' produces [C, C >> (Width-1)] |
| 10675 | Lower = *C; |
| 10676 | Upper = C->ashr(ShiftAmt: ShiftAmount) + 1; |
| 10677 | } else { |
| 10678 | // 'ashr C, x' produces [C >> (Width-1), C] |
| 10679 | Lower = C->ashr(ShiftAmt: ShiftAmount); |
| 10680 | Upper = *C + 1; |
| 10681 | } |
| 10682 | } |
| 10683 | break; |
| 10684 | |
| 10685 | case Instruction::LShr: |
| 10686 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) && C->ult(RHS: Width)) { |
| 10687 | // 'lshr x, C' produces [0, UINT_MAX >> C]. |
| 10688 | Upper = APInt::getAllOnes(numBits: Width).lshr(ShiftAmt: *C) + 1; |
| 10689 | } else if (match(V: BO.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) { |
| 10690 | // 'lshr C, x' produces [C >> (Width-1), C]. |
| 10691 | unsigned ShiftAmount = Width - 1; |
| 10692 | if (!C->isZero() && IIQ.isExact(Op: &BO)) |
| 10693 | ShiftAmount = C->countr_zero(); |
| 10694 | Lower = C->lshr(shiftAmt: ShiftAmount); |
| 10695 | Upper = *C + 1; |
| 10696 | } |
| 10697 | break; |
| 10698 | |
| 10699 | case Instruction::Shl: |
| 10700 | if (match(V: BO.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) { |
| 10701 | if (IIQ.hasNoUnsignedWrap(Op: &BO)) { |
| 10702 | // 'shl nuw C, x' produces [C, C << CLZ(C)] |
| 10703 | Lower = *C; |
| 10704 | Upper = Lower.shl(shiftAmt: Lower.countl_zero()) + 1; |
| 10705 | } else if (BO.hasNoSignedWrap()) { // TODO: What if both nuw+nsw? |
| 10706 | if (C->isNegative()) { |
| 10707 | // 'shl nsw C, x' produces [C << CLO(C)-1, C] |
| 10708 | unsigned ShiftAmount = C->countl_one() - 1; |
| 10709 | Lower = C->shl(shiftAmt: ShiftAmount); |
| 10710 | Upper = *C + 1; |
| 10711 | } else { |
| 10712 | // 'shl nsw C, x' produces [C, C << CLZ(C)-1] |
| 10713 | unsigned ShiftAmount = C->countl_zero() - 1; |
| 10714 | Lower = *C; |
| 10715 | Upper = C->shl(shiftAmt: ShiftAmount) + 1; |
| 10716 | } |
| 10717 | } else { |
| 10718 | // If lowbit is set, value can never be zero. |
| 10719 | if ((*C)[0]) |
| 10720 | Lower = APInt::getOneBitSet(numBits: Width, BitNo: 0); |
| 10721 | // If we are shifting a constant the largest it can be is if the longest |
| 10722 | // sequence of consecutive ones is shifted to the highbits (breaking |
| 10723 | // ties for which sequence is higher). At the moment we take a liberal |
| 10724 | // upper bound on this by just popcounting the constant. |
| 10725 | // TODO: There may be a bitwise trick for it longest/highest |
| 10726 | // consecutative sequence of ones (naive method is O(Width) loop). |
| 10727 | Upper = APInt::getHighBitsSet(numBits: Width, hiBitsSet: C->popcount()) + 1; |
| 10728 | } |
| 10729 | } else if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) && C->ult(RHS: Width)) { |
| 10730 | Upper = APInt::getBitsSetFrom(numBits: Width, loBit: C->getZExtValue()) + 1; |
| 10731 | } |
| 10732 | break; |
| 10733 | |
| 10734 | case Instruction::SDiv: |
| 10735 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) { |
| 10736 | APInt IntMin = APInt::getSignedMinValue(numBits: Width); |
| 10737 | APInt IntMax = APInt::getSignedMaxValue(numBits: Width); |
| 10738 | if (C->isAllOnes()) { |
| 10739 | // 'sdiv x, -1' produces [INT_MIN + 1, INT_MAX] |
| 10740 | // where C != -1 and C != 0 and C != 1 |
| 10741 | Lower = IntMin + 1; |
| 10742 | Upper = IntMax + 1; |
| 10743 | } else if (C->countl_zero() < Width - 1) { |
| 10744 | // 'sdiv x, C' produces [INT_MIN / C, INT_MAX / C] |
| 10745 | // where C != -1 and C != 0 and C != 1 |
| 10746 | Lower = IntMin.sdiv(RHS: *C); |
| 10747 | Upper = IntMax.sdiv(RHS: *C); |
| 10748 | if (Lower.sgt(RHS: Upper)) |
| 10749 | std::swap(a&: Lower, b&: Upper); |
| 10750 | Upper = Upper + 1; |
| 10751 | assert(Upper != Lower && "Upper part of range has wrapped!" ); |
| 10752 | } |
| 10753 | } else if (match(V: BO.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) { |
| 10754 | if (C->isMinSignedValue()) { |
| 10755 | // 'sdiv INT_MIN, x' produces [INT_MIN, INT_MIN / -2]. |
| 10756 | Lower = *C; |
| 10757 | Upper = Lower.lshr(shiftAmt: 1) + 1; |
| 10758 | } else { |
| 10759 | // 'sdiv C, x' produces [-|C|, |C|]. |
| 10760 | Upper = C->abs() + 1; |
| 10761 | Lower = (-Upper) + 1; |
| 10762 | } |
| 10763 | } |
| 10764 | break; |
| 10765 | |
| 10766 | case Instruction::UDiv: |
| 10767 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) && !C->isZero()) { |
| 10768 | // 'udiv x, C' produces [0, UINT_MAX / C]. |
| 10769 | Upper = APInt::getMaxValue(numBits: Width).udiv(RHS: *C) + 1; |
| 10770 | } else if (match(V: BO.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) { |
| 10771 | // 'udiv C, x' produces [0, C]. |
| 10772 | Upper = *C + 1; |
| 10773 | } |
| 10774 | break; |
| 10775 | |
| 10776 | case Instruction::SRem: |
| 10777 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) { |
| 10778 | // 'srem x, C' produces (-|C|, |C|). |
| 10779 | Upper = C->abs(); |
| 10780 | Lower = (-Upper) + 1; |
| 10781 | } else if (match(V: BO.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) { |
| 10782 | if (C->isNegative()) { |
| 10783 | // 'srem -|C|, x' produces [-|C|, 0]. |
| 10784 | Upper = 1; |
| 10785 | Lower = *C; |
| 10786 | } else { |
| 10787 | // 'srem |C|, x' produces [0, |C|]. |
| 10788 | Upper = *C + 1; |
| 10789 | } |
| 10790 | } |
| 10791 | break; |
| 10792 | |
| 10793 | case Instruction::URem: |
| 10794 | if (match(V: BO.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) |
| 10795 | // 'urem x, C' produces [0, C). |
| 10796 | Upper = *C; |
| 10797 | else if (match(V: BO.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) |
| 10798 | // 'urem C, x' produces [0, C]. |
| 10799 | Upper = *C + 1; |
| 10800 | break; |
| 10801 | |
| 10802 | default: |
| 10803 | break; |
| 10804 | } |
| 10805 | } |
| 10806 | |
| 10807 | static ConstantRange getRangeForIntrinsic(const IntrinsicInst &II, |
| 10808 | bool UseInstrInfo) { |
| 10809 | unsigned Width = II.getType()->getScalarSizeInBits(); |
| 10810 | const APInt *C; |
| 10811 | switch (II.getIntrinsicID()) { |
| 10812 | case Intrinsic::ctlz: |
| 10813 | case Intrinsic::cttz: { |
| 10814 | APInt Upper(Width, Width); |
| 10815 | if (!UseInstrInfo || !match(V: II.getArgOperand(i: 1), P: m_One())) |
| 10816 | Upper += 1; |
| 10817 | // Maximum of set/clear bits is the bit width. |
| 10818 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: Width), Upper); |
| 10819 | } |
| 10820 | case Intrinsic::ctpop: |
| 10821 | // Maximum of set/clear bits is the bit width. |
| 10822 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: Width), |
| 10823 | Upper: APInt(Width, Width) + 1); |
| 10824 | case Intrinsic::uadd_sat: |
| 10825 | // uadd.sat(x, C) produces [C, UINT_MAX]. |
| 10826 | if (match(V: II.getOperand(i_nocapture: 0), P: m_APInt(Res&: C)) || |
| 10827 | match(V: II.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) |
| 10828 | return ConstantRange::getNonEmpty(Lower: *C, Upper: APInt::getZero(numBits: Width)); |
| 10829 | break; |
| 10830 | case Intrinsic::sadd_sat: |
| 10831 | if (match(V: II.getOperand(i_nocapture: 0), P: m_APInt(Res&: C)) || |
| 10832 | match(V: II.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) { |
| 10833 | if (C->isNegative()) |
| 10834 | // sadd.sat(x, -C) produces [SINT_MIN, SINT_MAX + (-C)]. |
| 10835 | return ConstantRange::getNonEmpty(Lower: APInt::getSignedMinValue(numBits: Width), |
| 10836 | Upper: APInt::getSignedMaxValue(numBits: Width) + *C + |
| 10837 | 1); |
| 10838 | |
| 10839 | // sadd.sat(x, +C) produces [SINT_MIN + C, SINT_MAX]. |
| 10840 | return ConstantRange::getNonEmpty(Lower: APInt::getSignedMinValue(numBits: Width) + *C, |
| 10841 | Upper: APInt::getSignedMaxValue(numBits: Width) + 1); |
| 10842 | } |
| 10843 | break; |
| 10844 | case Intrinsic::usub_sat: |
| 10845 | // usub.sat(C, x) produces [0, C]. |
| 10846 | if (match(V: II.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) |
| 10847 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: Width), Upper: *C + 1); |
| 10848 | |
| 10849 | // usub.sat(x, C) produces [0, UINT_MAX - C]. |
| 10850 | if (match(V: II.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) |
| 10851 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: Width), |
| 10852 | Upper: APInt::getMaxValue(numBits: Width) - *C + 1); |
| 10853 | break; |
| 10854 | case Intrinsic::ssub_sat: |
| 10855 | if (match(V: II.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) { |
| 10856 | if (C->isNegative()) |
| 10857 | // ssub.sat(-C, x) produces [SINT_MIN, -SINT_MIN + (-C)]. |
| 10858 | return ConstantRange::getNonEmpty(Lower: APInt::getSignedMinValue(numBits: Width), |
| 10859 | Upper: *C - APInt::getSignedMinValue(numBits: Width) + |
| 10860 | 1); |
| 10861 | |
| 10862 | // ssub.sat(+C, x) produces [-SINT_MAX + C, SINT_MAX]. |
| 10863 | return ConstantRange::getNonEmpty(Lower: *C - APInt::getSignedMaxValue(numBits: Width), |
| 10864 | Upper: APInt::getSignedMaxValue(numBits: Width) + 1); |
| 10865 | } else if (match(V: II.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) { |
| 10866 | if (C->isNegative()) |
| 10867 | // ssub.sat(x, -C) produces [SINT_MIN - (-C), SINT_MAX]: |
| 10868 | return ConstantRange::getNonEmpty(Lower: APInt::getSignedMinValue(numBits: Width) - *C, |
| 10869 | Upper: APInt::getSignedMaxValue(numBits: Width) + 1); |
| 10870 | |
| 10871 | // ssub.sat(x, +C) produces [SINT_MIN, SINT_MAX - C]. |
| 10872 | return ConstantRange::getNonEmpty(Lower: APInt::getSignedMinValue(numBits: Width), |
| 10873 | Upper: APInt::getSignedMaxValue(numBits: Width) - *C + |
| 10874 | 1); |
| 10875 | } |
| 10876 | break; |
| 10877 | case Intrinsic::ushl_sat: |
| 10878 | if (match(V: II.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) |
| 10879 | return ConstantRange::getNonEmpty(Lower: *C, Upper: APInt::getZero(numBits: Width)); |
| 10880 | break; |
| 10881 | case Intrinsic::sshl_sat: |
| 10882 | if (match(V: II.getOperand(i_nocapture: 0), P: m_APInt(Res&: C))) { |
| 10883 | if (C->isNegative()) |
| 10884 | return ConstantRange::getNonEmpty(Lower: APInt::getSignedMinValue(numBits: Width), |
| 10885 | Upper: *C + 1); |
| 10886 | return ConstantRange::getNonEmpty(Lower: *C, |
| 10887 | Upper: APInt::getSignedMaxValue(numBits: Width) + 1); |
| 10888 | } |
| 10889 | break; |
| 10890 | case Intrinsic::umin: |
| 10891 | case Intrinsic::umax: |
| 10892 | case Intrinsic::smin: |
| 10893 | case Intrinsic::smax: |
| 10894 | if (!match(V: II.getOperand(i_nocapture: 0), P: m_APInt(Res&: C)) && |
| 10895 | !match(V: II.getOperand(i_nocapture: 1), P: m_APInt(Res&: C))) |
| 10896 | break; |
| 10897 | |
| 10898 | switch (II.getIntrinsicID()) { |
| 10899 | case Intrinsic::umin: |
| 10900 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: Width), Upper: *C + 1); |
| 10901 | case Intrinsic::umax: |
| 10902 | return ConstantRange::getNonEmpty(Lower: *C, Upper: APInt::getZero(numBits: Width)); |
| 10903 | case Intrinsic::smin: |
| 10904 | return ConstantRange::getNonEmpty(Lower: APInt::getSignedMinValue(numBits: Width), |
| 10905 | Upper: *C + 1); |
| 10906 | case Intrinsic::smax: |
| 10907 | return ConstantRange::getNonEmpty(Lower: *C, |
| 10908 | Upper: APInt::getSignedMaxValue(numBits: Width) + 1); |
| 10909 | default: |
| 10910 | llvm_unreachable("Must be min/max intrinsic" ); |
| 10911 | } |
| 10912 | break; |
| 10913 | case Intrinsic::abs: |
| 10914 | // If abs of SIGNED_MIN is poison, then the result is [0..SIGNED_MAX], |
| 10915 | // otherwise it is [0..SIGNED_MIN], as -SIGNED_MIN == SIGNED_MIN. |
| 10916 | if (match(V: II.getOperand(i_nocapture: 1), P: m_One())) |
| 10917 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: Width), |
| 10918 | Upper: APInt::getSignedMaxValue(numBits: Width) + 1); |
| 10919 | |
| 10920 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: Width), |
| 10921 | Upper: APInt::getSignedMinValue(numBits: Width) + 1); |
| 10922 | case Intrinsic::vscale: |
| 10923 | if (!II.getParent() || !II.getFunction()) |
| 10924 | break; |
| 10925 | return getVScaleRange(F: II.getFunction(), BitWidth: Width); |
| 10926 | case Intrinsic::read_register: |
| 10927 | case Intrinsic::read_volatile_register: { |
| 10928 | const Module *M = II.getModule(); |
| 10929 | if (!M || !M->getTargetTriple().isRISCV()) |
| 10930 | break; |
| 10931 | if (II.getFunction() && isReadVLENB(II)) |
| 10932 | return getRISCVVLENBRange(II, Width); |
| 10933 | break; |
| 10934 | } |
| 10935 | default: |
| 10936 | break; |
| 10937 | } |
| 10938 | |
| 10939 | return ConstantRange::getFull(BitWidth: Width); |
| 10940 | } |
| 10941 | |
| 10942 | static ConstantRange getRangeForSelectPattern(const SelectInst &SI, |
| 10943 | const InstrInfoQuery &IIQ) { |
| 10944 | unsigned BitWidth = SI.getType()->getScalarSizeInBits(); |
| 10945 | const Value *LHS = nullptr, *RHS = nullptr; |
| 10946 | SelectPatternResult R = matchSelectPattern(V: &SI, LHS, RHS); |
| 10947 | if (R.Flavor == SPF_UNKNOWN) |
| 10948 | return ConstantRange::getFull(BitWidth); |
| 10949 | |
| 10950 | if (R.Flavor == SelectPatternFlavor::SPF_ABS) { |
| 10951 | // If the negation part of the abs (in RHS) has the NSW flag, |
| 10952 | // then the result of abs(X) is [0..SIGNED_MAX], |
| 10953 | // otherwise it is [0..SIGNED_MIN], as -SIGNED_MIN == SIGNED_MIN. |
| 10954 | if (match(V: RHS, P: m_Neg(V: m_Specific(V: LHS))) && |
| 10955 | IIQ.hasNoSignedWrap(Op: cast<Instruction>(Val: RHS))) |
| 10956 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: BitWidth), |
| 10957 | Upper: APInt::getSignedMaxValue(numBits: BitWidth) + 1); |
| 10958 | |
| 10959 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: BitWidth), |
| 10960 | Upper: APInt::getSignedMinValue(numBits: BitWidth) + 1); |
| 10961 | } |
| 10962 | |
| 10963 | if (R.Flavor == SelectPatternFlavor::SPF_NABS) { |
| 10964 | // The result of -abs(X) is <= 0. |
| 10965 | return ConstantRange::getNonEmpty(Lower: APInt::getSignedMinValue(numBits: BitWidth), |
| 10966 | Upper: APInt(BitWidth, 1)); |
| 10967 | } |
| 10968 | |
| 10969 | const APInt *C; |
| 10970 | if (!match(V: LHS, P: m_APInt(Res&: C)) && !match(V: RHS, P: m_APInt(Res&: C))) |
| 10971 | return ConstantRange::getFull(BitWidth); |
| 10972 | |
| 10973 | switch (R.Flavor) { |
| 10974 | case SPF_UMIN: |
| 10975 | return ConstantRange::getNonEmpty(Lower: APInt::getZero(numBits: BitWidth), Upper: *C + 1); |
| 10976 | case SPF_UMAX: |
| 10977 | return ConstantRange::getNonEmpty(Lower: *C, Upper: APInt::getZero(numBits: BitWidth)); |
| 10978 | case SPF_SMIN: |
| 10979 | return ConstantRange::getNonEmpty(Lower: APInt::getSignedMinValue(numBits: BitWidth), |
| 10980 | Upper: *C + 1); |
| 10981 | case SPF_SMAX: |
| 10982 | return ConstantRange::getNonEmpty(Lower: *C, |
| 10983 | Upper: APInt::getSignedMaxValue(numBits: BitWidth) + 1); |
| 10984 | default: |
| 10985 | return ConstantRange::getFull(BitWidth); |
| 10986 | } |
| 10987 | } |
| 10988 | |
| 10989 | static void setLimitForFPToI(const Instruction *I, APInt &Lower, APInt &Upper) { |
| 10990 | // The maximum representable value of a half is 65504. For floats the maximum |
| 10991 | // value is 3.4e38 which requires roughly 129 bits. |
| 10992 | unsigned BitWidth = I->getType()->getScalarSizeInBits(); |
| 10993 | if (!I->getOperand(i: 0)->getType()->getScalarType()->isHalfTy()) |
| 10994 | return; |
| 10995 | if (isa<FPToSIInst>(Val: I) && BitWidth >= 17) { |
| 10996 | Lower = APInt(BitWidth, -65504, true); |
| 10997 | Upper = APInt(BitWidth, 65505); |
| 10998 | } |
| 10999 | |
| 11000 | if (isa<FPToUIInst>(Val: I) && BitWidth >= 16) { |
| 11001 | // For a fptoui the lower limit is left as 0. |
| 11002 | Upper = APInt(BitWidth, 65505); |
| 11003 | } |
| 11004 | } |
| 11005 | |
| 11006 | ConstantRange llvm::computeConstantRange(const Value *V, bool ForSigned, |
| 11007 | const SimplifyQuery &SQ, |
| 11008 | unsigned Depth) { |
| 11009 | assert(V->getType()->isIntOrIntVectorTy() && "Expected integer instruction" ); |
| 11010 | |
| 11011 | if (Depth == MaxAnalysisRecursionDepth) |
| 11012 | return ConstantRange::getFull(BitWidth: V->getType()->getScalarSizeInBits()); |
| 11013 | |
| 11014 | if (auto *C = dyn_cast<Constant>(Val: V)) |
| 11015 | return C->toConstantRange(); |
| 11016 | |
| 11017 | unsigned BitWidth = V->getType()->getScalarSizeInBits(); |
| 11018 | ConstantRange CR = ConstantRange::getFull(BitWidth); |
| 11019 | if (auto *BO = dyn_cast<BinaryOperator>(Val: V)) { |
| 11020 | APInt Lower = APInt(BitWidth, 0); |
| 11021 | APInt Upper = APInt(BitWidth, 0); |
| 11022 | // TODO: Return ConstantRange. |
| 11023 | setLimitsForBinOp(BO: *BO, Lower, Upper, IIQ: SQ.IIQ, PreferSignedRange: ForSigned); |
| 11024 | CR = ConstantRange::getNonEmpty(Lower, Upper); |
| 11025 | } else if (auto *II = dyn_cast<IntrinsicInst>(Val: V)) |
| 11026 | CR = getRangeForIntrinsic(II: *II, UseInstrInfo: SQ.IIQ.UseInstrInfo); |
| 11027 | else if (auto *SI = dyn_cast<SelectInst>(Val: V)) { |
| 11028 | ConstantRange CRTrue = |
| 11029 | computeConstantRange(V: SI->getTrueValue(), ForSigned, SQ, Depth: Depth + 1); |
| 11030 | ConstantRange CRFalse = |
| 11031 | computeConstantRange(V: SI->getFalseValue(), ForSigned, SQ, Depth: Depth + 1); |
| 11032 | CR = CRTrue.unionWith(CR: CRFalse, Type: ForSigned ? ConstantRange::Signed |
| 11033 | : ConstantRange::Unsigned); |
| 11034 | CR = CR.intersectWith(CR: getRangeForSelectPattern(SI: *SI, IIQ: SQ.IIQ)); |
| 11035 | } else if (auto *TI = dyn_cast<TruncInst>(Val: V)) { |
| 11036 | ConstantRange SrcCR = |
| 11037 | computeConstantRange(V: TI->getOperand(i_nocapture: 0), ForSigned, SQ, Depth: Depth + 1); |
| 11038 | CR = SrcCR.truncate(BitWidth); |
| 11039 | } else if (auto *ZExt = dyn_cast<ZExtInst>(Val: V)) { |
| 11040 | ConstantRange SrcCR = |
| 11041 | computeConstantRange(V: ZExt->getOperand(i_nocapture: 0), ForSigned, SQ, Depth: Depth + 1); |
| 11042 | CR = SrcCR.zeroExtend(BitWidth); |
| 11043 | } else if (auto *SExt = dyn_cast<SExtInst>(Val: V)) { |
| 11044 | ConstantRange SrcCR = |
| 11045 | computeConstantRange(V: SExt->getOperand(i_nocapture: 0), ForSigned, SQ, Depth: Depth + 1); |
| 11046 | CR = SrcCR.signExtend(BitWidth); |
| 11047 | } else if (isa<FPToUIInst>(Val: V) || isa<FPToSIInst>(Val: V)) { |
| 11048 | APInt Lower = APInt(BitWidth, 0); |
| 11049 | APInt Upper = APInt(BitWidth, 0); |
| 11050 | // TODO: Return ConstantRange. |
| 11051 | setLimitForFPToI(I: cast<Instruction>(Val: V), Lower, Upper); |
| 11052 | CR = ConstantRange::getNonEmpty(Lower, Upper); |
| 11053 | } else if (const auto *A = dyn_cast<Argument>(Val: V)) |
| 11054 | if (std::optional<ConstantRange> Range = A->getRange()) |
| 11055 | CR = *Range; |
| 11056 | |
| 11057 | if (auto *I = dyn_cast<Instruction>(Val: V)) { |
| 11058 | if (auto *Range = SQ.IIQ.getMetadata(I, KindID: LLVMContext::MD_range)) |
| 11059 | CR = CR.intersectWith(CR: getConstantRangeFromMetadata(RangeMD: *Range)); |
| 11060 | |
| 11061 | Value *FrexpSrc; |
| 11062 | if (const auto *CB = dyn_cast<CallBase>(Val: V)) { |
| 11063 | if (std::optional<ConstantRange> Range = CB->getRange()) |
| 11064 | CR = CR.intersectWith(CR: *Range); |
| 11065 | } else if (match(V: I, P: m_ExtractValue<1>(V: m_Intrinsic<Intrinsic::frexp>( |
| 11066 | Ops: m_Value(V&: FrexpSrc))))) { |
| 11067 | const fltSemantics &FltSem = |
| 11068 | FrexpSrc->getType()->getScalarType()->getFltSemantics(); |
| 11069 | // It should be possible to implement this for any type, but this logic |
| 11070 | // only computes the range assuming standard subnormal handling. |
| 11071 | if (APFloat::isIEEELikeFP(FltSem)) { |
| 11072 | KnownFPClass KnownSrc = computeKnownFPClass( |
| 11073 | V: FrexpSrc, InterestedClasses: fcSubnormal | fcZero | fcNan | fcInf, SQ, Depth: Depth + 1); |
| 11074 | |
| 11075 | // The exponent of frexp(NaN) and frexp(Inf) is unspecified. Only |
| 11076 | // constrain its range when the source can be neither. |
| 11077 | if (KnownSrc.isKnownNeverInfOrNaN()) { |
| 11078 | int MinExp = APFloat::semanticsMinExponent(FltSem) + 1; |
| 11079 | |
| 11080 | // Offset to find the true minimum exponent value for a denormal. |
| 11081 | if (!KnownSrc.isKnownNeverSubnormal()) |
| 11082 | MinExp -= (APFloat::semanticsPrecision(FltSem) - 1); |
| 11083 | |
| 11084 | int MaxExp = APFloat::semanticsMaxExponent(FltSem) + 1; |
| 11085 | |
| 11086 | auto [AdjustedMin, AdjustedMax, AdjustedMaxNonZero] = |
| 11087 | computeKnownExponentRangeFromContext(V: FrexpSrc, Q: SQ); |
| 11088 | |
| 11089 | DenormalMode Mode = I->getFunction()->getDenormalMode(FPType: FltSem); |
| 11090 | bool NeverLogicalZero = KnownSrc.isKnownNeverLogicalZero(Mode); |
| 11091 | |
| 11092 | MinExp = std::max(a: AdjustedMin, b: MinExp); |
| 11093 | MaxExp = std::min(a: NeverLogicalZero ? AdjustedMaxNonZero : AdjustedMax, |
| 11094 | b: MaxExp); |
| 11095 | |
| 11096 | CR = ConstantRange::getNonEmpty( |
| 11097 | Lower: APInt(BitWidth, static_cast<int64_t>(MinExp), /*isSigned=*/true), |
| 11098 | Upper: APInt(BitWidth, static_cast<int64_t>(MaxExp) + 1, |
| 11099 | /*isSigned=*/true)); |
| 11100 | } |
| 11101 | } |
| 11102 | } |
| 11103 | } |
| 11104 | |
| 11105 | if (SQ.CtxI && SQ.AC) { |
| 11106 | // Try to restrict the range based on information from assumptions. |
| 11107 | for (auto &AssumeVH : SQ.AC->assumptionsFor(V)) { |
| 11108 | if (!AssumeVH) |
| 11109 | continue; |
| 11110 | CallInst *I = cast<CallInst>(Val&: AssumeVH); |
| 11111 | assert(I->getParent()->getParent() == SQ.CtxI->getParent()->getParent() && |
| 11112 | "Got assumption for the wrong function!" ); |
| 11113 | assert(I->getIntrinsicID() == Intrinsic::assume && |
| 11114 | "must be an assume intrinsic" ); |
| 11115 | |
| 11116 | if (AssumeVH.Index != AssumptionCache::ExprResultIdx) { |
| 11117 | auto OBU = I->getOperandBundleAt(Index: AssumeVH.Index); |
| 11118 | if (getBundleAttrFromOBU(OBU) == BundleAttr::Range) { |
| 11119 | auto [Val, Range] = getAssumeRangeInfo(OBU); |
| 11120 | if (Val == V && isValidAssumeForContext(I, Q: SQ)) |
| 11121 | CR = CR.intersectWith(CR: Range); |
| 11122 | } |
| 11123 | continue; |
| 11124 | } |
| 11125 | |
| 11126 | if (!isValidAssumeForContext(I, Q: SQ)) |
| 11127 | continue; |
| 11128 | Value *Arg = I->getArgOperand(i: 0); |
| 11129 | ICmpInst *Cmp = dyn_cast<ICmpInst>(Val: Arg); |
| 11130 | // Currently we just use information from comparisons. |
| 11131 | if (!Cmp || Cmp->getOperand(i_nocapture: 0) != V) |
| 11132 | continue; |
| 11133 | // TODO: Set "ForSigned" parameter via Cmp->isSigned()? |
| 11134 | ConstantRange RHS = |
| 11135 | computeConstantRange(V: Cmp->getOperand(i_nocapture: 1), /*ForSigned=*/false, |
| 11136 | SQ: SQ.getWithInstruction(I), Depth: Depth + 1); |
| 11137 | CR = CR.intersectWith( |
| 11138 | CR: ConstantRange::makeAllowedICmpRegion(Pred: Cmp->getCmpPredicate(), Other: RHS)); |
| 11139 | } |
| 11140 | } |
| 11141 | |
| 11142 | return CR; |
| 11143 | } |
| 11144 | |
| 11145 | static void |
| 11146 | addValueAffectedByCondition(Value *V, |
| 11147 | function_ref<void(Value *)> InsertAffected) { |
| 11148 | assert(V != nullptr); |
| 11149 | if (isa<Argument>(Val: V) || isa<GlobalValue>(Val: V)) { |
| 11150 | InsertAffected(V); |
| 11151 | } else if (auto *I = dyn_cast<Instruction>(Val: V)) { |
| 11152 | InsertAffected(V); |
| 11153 | |
| 11154 | // Peek through unary operators to find the source of the condition. |
| 11155 | Value *Op; |
| 11156 | if (match(V: I, P: m_CombineOr(Ps: m_PtrToIntOrAddr(Op: m_Value(V&: Op)), |
| 11157 | Ps: m_Trunc(Op: m_Value(V&: Op))))) { |
| 11158 | if (isa<Instruction>(Val: Op) || isa<Argument>(Val: Op)) |
| 11159 | InsertAffected(Op); |
| 11160 | } |
| 11161 | } |
| 11162 | } |
| 11163 | |
| 11164 | void llvm::findValuesAffectedByCondition( |
| 11165 | Value *Cond, bool IsAssume, function_ref<void(Value *)> InsertAffected) { |
| 11166 | auto AddAffected = [&InsertAffected](Value *V) { |
| 11167 | addValueAffectedByCondition(V, InsertAffected); |
| 11168 | }; |
| 11169 | |
| 11170 | auto AddCmpOperands = [&AddAffected, IsAssume](Value *LHS, Value *RHS) { |
| 11171 | if (IsAssume) { |
| 11172 | AddAffected(LHS); |
| 11173 | AddAffected(RHS); |
| 11174 | } else if (match(V: RHS, P: m_Constant())) |
| 11175 | AddAffected(LHS); |
| 11176 | }; |
| 11177 | |
| 11178 | SmallVector<Value *, 8> Worklist; |
| 11179 | SmallPtrSet<Value *, 8> Visited; |
| 11180 | Worklist.push_back(Elt: Cond); |
| 11181 | while (!Worklist.empty()) { |
| 11182 | Value *V = Worklist.pop_back_val(); |
| 11183 | if (!Visited.insert(Ptr: V).second) |
| 11184 | continue; |
| 11185 | |
| 11186 | CmpPredicate Pred; |
| 11187 | Value *A, *B, *X; |
| 11188 | |
| 11189 | if (IsAssume) { |
| 11190 | AddAffected(V); |
| 11191 | if (match(V, P: m_Not(V: m_Value(V&: X)))) |
| 11192 | AddAffected(X); |
| 11193 | } |
| 11194 | |
| 11195 | if (match(V, P: m_LogicalOp(L: m_Value(V&: A), R: m_Value(V&: B)))) { |
| 11196 | // assume(A && B) is split to -> assume(A); assume(B); |
| 11197 | // assume(!(A || B)) is split to -> assume(!A); assume(!B); |
| 11198 | // Finally, assume(A || B) / assume(!(A && B)) generally don't provide |
| 11199 | // enough information to be worth handling (intersection of information as |
| 11200 | // opposed to union). |
| 11201 | if (!IsAssume) { |
| 11202 | Worklist.push_back(Elt: A); |
| 11203 | Worklist.push_back(Elt: B); |
| 11204 | } |
| 11205 | } else if (match(V, P: m_ICmp(Pred, L: m_Value(V&: A), R: m_Value(V&: B)))) { |
| 11206 | bool HasRHSC = match(V: B, P: m_ConstantInt()); |
| 11207 | if (ICmpInst::isEquality(P: Pred)) { |
| 11208 | AddAffected(A); |
| 11209 | if (IsAssume) |
| 11210 | AddAffected(B); |
| 11211 | if (HasRHSC) { |
| 11212 | Value *Y; |
| 11213 | // (X << C) or (X >>_s C) or (X >>_u C). |
| 11214 | if (match(V: A, P: m_Shift(L: m_Value(V&: X), R: m_ConstantInt()))) |
| 11215 | AddAffected(X); |
| 11216 | // (X & C) or (X | C). |
| 11217 | else if (match(V: A, P: m_And(L: m_Value(V&: X), R: m_Value(V&: Y))) || |
| 11218 | match(V: A, P: m_Or(L: m_Value(V&: X), R: m_Value(V&: Y)))) { |
| 11219 | AddAffected(X); |
| 11220 | AddAffected(Y); |
| 11221 | } |
| 11222 | // X - Y |
| 11223 | else if (match(V: A, P: m_Sub(L: m_Value(V&: X), R: m_Value(V&: Y)))) { |
| 11224 | AddAffected(X); |
| 11225 | AddAffected(Y); |
| 11226 | } |
| 11227 | } |
| 11228 | } else { |
| 11229 | AddCmpOperands(A, B); |
| 11230 | if (HasRHSC) { |
| 11231 | Value *Y; |
| 11232 | // Handle (A + C1) u< C2, which is the canonical form of |
| 11233 | // A > C3 && A < C4. |
| 11234 | if (match(V: A, P: m_AddLike(L: m_Value(V&: X), R: m_ConstantInt()))) |
| 11235 | AddAffected(X); |
| 11236 | // X - Y pred C, which implies X != Y if 0 pred C is false. |
| 11237 | else if (match(V: A, P: m_Sub(L: m_Value(V&: X), R: m_Value(V&: Y)))) { |
| 11238 | AddAffected(X); |
| 11239 | AddAffected(Y); |
| 11240 | } |
| 11241 | |
| 11242 | if (ICmpInst::isUnsigned(Pred)) { |
| 11243 | // X & Y u> C -> X >u C && Y >u C |
| 11244 | // X | Y u< C -> X u< C && Y u< C |
| 11245 | // X nuw+ Y u< C -> X u< C && Y u< C |
| 11246 | if (match(V: A, P: m_And(L: m_Value(V&: X), R: m_Value(V&: Y))) || |
| 11247 | match(V: A, P: m_Or(L: m_Value(V&: X), R: m_Value(V&: Y))) || |
| 11248 | match(V: A, P: m_NUWAdd(L: m_Value(V&: X), R: m_Value(V&: Y)))) { |
| 11249 | AddAffected(X); |
| 11250 | AddAffected(Y); |
| 11251 | } |
| 11252 | } |
| 11253 | } |
| 11254 | |
| 11255 | // Handle icmp slt/sgt (bitcast X to int), 0/-1, which is supported |
| 11256 | // by computeKnownFPClass(). |
| 11257 | if (match(V: A, P: m_ElementWiseBitCast(Op: m_Value(V&: X)))) { |
| 11258 | if (Pred == ICmpInst::ICMP_SLT && match(V: B, P: m_Zero())) |
| 11259 | InsertAffected(X); |
| 11260 | else if (Pred == ICmpInst::ICMP_SGT && match(V: B, P: m_AllOnes())) |
| 11261 | InsertAffected(X); |
| 11262 | } |
| 11263 | } |
| 11264 | |
| 11265 | auto AddNuwSquareOperand = [&AddAffected](Value *Op) { |
| 11266 | Value *SquareOp = nullptr; |
| 11267 | if (match(V: Op, P: m_NUWMul(L: m_Value(V&: SquareOp), R: m_Deferred(V: SquareOp)))) |
| 11268 | AddAffected(SquareOp); |
| 11269 | }; |
| 11270 | AddNuwSquareOperand(A); |
| 11271 | AddNuwSquareOperand(B); |
| 11272 | |
| 11273 | if (HasRHSC && match(V: A, P: m_Ctpop(Op0: m_Value(V&: X)))) |
| 11274 | AddAffected(X); |
| 11275 | } else if (match(V, P: m_FCmp(Pred, L: m_Value(V&: A), R: m_Value(V&: B)))) { |
| 11276 | AddCmpOperands(A, B); |
| 11277 | |
| 11278 | // fcmp fneg(x), y |
| 11279 | // fcmp fabs(x), y |
| 11280 | // fcmp fneg(fabs(x)), y |
| 11281 | // TODO: Confirm whether m_FNegOrFSubNegZero is correct. |
| 11282 | if (match(V: A, P: m_FNegOrFSubNegZero(X: m_Value(V&: A)))) |
| 11283 | AddAffected(A); |
| 11284 | if (match(V: A, P: m_FAbs(Op0: m_Value(V&: A)))) |
| 11285 | AddAffected(A); |
| 11286 | |
| 11287 | } else if (match(V, P: m_Intrinsic<Intrinsic::is_fpclass>(Ops: m_Value(V&: A), |
| 11288 | Ops: m_Value()))) { |
| 11289 | // Handle patterns that computeKnownFPClass() support. |
| 11290 | AddAffected(A); |
| 11291 | } else if (!IsAssume && match(V, P: m_Trunc(Op: m_Value(V&: X)))) { |
| 11292 | // Assume is checked here as X is already added above for assumes in |
| 11293 | // addValueAffectedByCondition |
| 11294 | AddAffected(X); |
| 11295 | } else if (!IsAssume && match(V, P: m_Not(V: m_Value(V&: X)))) { |
| 11296 | // Assume is checked here to avoid issues with ephemeral values |
| 11297 | Worklist.push_back(Elt: X); |
| 11298 | } |
| 11299 | } |
| 11300 | } |
| 11301 | |
| 11302 | const Value *llvm::stripNullTest(const Value *V) { |
| 11303 | // (X >> C) or/add (X & mask(C) != 0) |
| 11304 | if (const auto *BO = dyn_cast<BinaryOperator>(Val: V)) { |
| 11305 | if (BO->getOpcode() == Instruction::Add || |
| 11306 | BO->getOpcode() == Instruction::Or) { |
| 11307 | const Value *X; |
| 11308 | const APInt *C1, *C2; |
| 11309 | if (match(V: BO, P: m_c_BinOp(L: m_LShr(L: m_Value(V&: X), R: m_APInt(Res&: C1)), |
| 11310 | R: m_ZExt(Op: m_SpecificICmp( |
| 11311 | MatchPred: ICmpInst::ICMP_NE, |
| 11312 | L: m_And(L: m_Deferred(V: X), R: m_LowBitMask(V&: C2)), |
| 11313 | R: m_Zero())))) && |
| 11314 | C2->popcount() == C1->getZExtValue()) |
| 11315 | return X; |
| 11316 | } |
| 11317 | } |
| 11318 | |
| 11319 | // select(cond, X, -X) |
| 11320 | { |
| 11321 | const Value *X; |
| 11322 | if (match(V, P: m_c_Select(L: m_Value(V&: X), R: m_Neg(V: m_Deferred(V: X))))) |
| 11323 | return X; |
| 11324 | } |
| 11325 | |
| 11326 | return nullptr; |
| 11327 | } |
| 11328 | |
| 11329 | Value *llvm::stripNullTest(Value *V) { |
| 11330 | return const_cast<Value *>(stripNullTest(V: const_cast<const Value *>(V))); |
| 11331 | } |
| 11332 | |
| 11333 | bool llvm::collectPossibleValues(const Value *V, |
| 11334 | SmallPtrSetImpl<const Constant *> &Constants, |
| 11335 | unsigned MaxCount, bool AllowUndefOrPoison) { |
| 11336 | SmallPtrSet<const Instruction *, 8> Visited; |
| 11337 | SmallVector<const Instruction *, 8> Worklist; |
| 11338 | auto Push = [&](const Value *V) -> bool { |
| 11339 | Constant *C; |
| 11340 | if (match(V: const_cast<Value *>(V), P: m_ImmConstant(C))) { |
| 11341 | if (!AllowUndefOrPoison && !isGuaranteedNotToBeUndefOrPoison(V: C)) |
| 11342 | return false; |
| 11343 | // Check existence first to avoid unnecessary allocations. |
| 11344 | if (Constants.contains(Ptr: C)) |
| 11345 | return true; |
| 11346 | if (Constants.size() == MaxCount) |
| 11347 | return false; |
| 11348 | Constants.insert(Ptr: C); |
| 11349 | return true; |
| 11350 | } |
| 11351 | |
| 11352 | if (auto *Inst = dyn_cast<Instruction>(Val: V)) { |
| 11353 | if (Visited.insert(Ptr: Inst).second) |
| 11354 | Worklist.push_back(Elt: Inst); |
| 11355 | return true; |
| 11356 | } |
| 11357 | return false; |
| 11358 | }; |
| 11359 | if (!Push(V)) |
| 11360 | return false; |
| 11361 | while (!Worklist.empty()) { |
| 11362 | const Instruction *CurInst = Worklist.pop_back_val(); |
| 11363 | switch (CurInst->getOpcode()) { |
| 11364 | case Instruction::Select: |
| 11365 | if (!Push(CurInst->getOperand(i: 1))) |
| 11366 | return false; |
| 11367 | if (!Push(CurInst->getOperand(i: 2))) |
| 11368 | return false; |
| 11369 | break; |
| 11370 | case Instruction::PHI: |
| 11371 | for (Value *IncomingValue : cast<PHINode>(Val: CurInst)->incoming_values()) { |
| 11372 | // Fast path for recurrence PHI. |
| 11373 | if (IncomingValue == CurInst) |
| 11374 | continue; |
| 11375 | if (!Push(IncomingValue)) |
| 11376 | return false; |
| 11377 | } |
| 11378 | break; |
| 11379 | default: |
| 11380 | return false; |
| 11381 | } |
| 11382 | } |
| 11383 | return true; |
| 11384 | } |
| 11385 | |