1 //===-- SimplifyIndVar.cpp - Induction variable simplification ------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements induction variable simplification. It does 10 // not define any actual pass or policy, but provides a single function to 11 // simplify a loop's induction variables based on ScalarEvolution. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "llvm/Transforms/Utils/SimplifyIndVar.h" 16 #include "llvm/ADT/SmallVector.h" 17 #include "llvm/ADT/Statistic.h" 18 #include "llvm/Analysis/LoopInfo.h" 19 #include "llvm/Analysis/ValueTracking.h" 20 #include "llvm/IR/Dominators.h" 21 #include "llvm/IR/IRBuilder.h" 22 #include "llvm/IR/Instructions.h" 23 #include "llvm/IR/IntrinsicInst.h" 24 #include "llvm/IR/PatternMatch.h" 25 #include "llvm/Support/Debug.h" 26 #include "llvm/Support/raw_ostream.h" 27 #include "llvm/Transforms/Utils/Local.h" 28 #include "llvm/Transforms/Utils/LoopUtils.h" 29 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" 30 31 using namespace llvm; 32 using namespace llvm::PatternMatch; 33 34 #define DEBUG_TYPE "indvars" 35 36 STATISTIC(NumElimIdentity, "Number of IV identities eliminated"); 37 STATISTIC(NumElimOperand, "Number of IV operands folded into a use"); 38 STATISTIC(NumFoldedUser, "Number of IV users folded into a constant"); 39 STATISTIC(NumElimRem , "Number of IV remainder operations eliminated"); 40 STATISTIC( 41 NumSimplifiedSDiv, 42 "Number of IV signed division operations converted to unsigned division"); 43 STATISTIC( 44 NumSimplifiedSRem, 45 "Number of IV signed remainder operations converted to unsigned remainder"); 46 STATISTIC(NumElimCmp , "Number of IV comparisons eliminated"); 47 48 namespace { 49 /// This is a utility for simplifying induction variables 50 /// based on ScalarEvolution. It is the primary instrument of the 51 /// IndvarSimplify pass, but it may also be directly invoked to cleanup after 52 /// other loop passes that preserve SCEV. 53 class SimplifyIndvar { 54 Loop *L; 55 LoopInfo *LI; 56 ScalarEvolution *SE; 57 DominatorTree *DT; 58 const TargetTransformInfo *TTI; 59 SCEVExpander &Rewriter; 60 SmallVectorImpl<WeakTrackingVH> &DeadInsts; 61 62 bool Changed = false; 63 bool RunUnswitching = false; 64 65 public: 66 SimplifyIndvar(Loop *Loop, ScalarEvolution *SE, DominatorTree *DT, 67 LoopInfo *LI, const TargetTransformInfo *TTI, 68 SCEVExpander &Rewriter, 69 SmallVectorImpl<WeakTrackingVH> &Dead) 70 : L(Loop), LI(LI), SE(SE), DT(DT), TTI(TTI), Rewriter(Rewriter), 71 DeadInsts(Dead) { 72 assert(LI && "IV simplification requires LoopInfo"); 73 } 74 75 bool hasChanged() const { return Changed; } 76 bool runUnswitching() const { return RunUnswitching; } 77 78 /// Iteratively perform simplification on a worklist of users of the 79 /// specified induction variable. This is the top-level driver that applies 80 /// all simplifications to users of an IV. 81 void simplifyUsers(PHINode *CurrIV, IVVisitor *V = nullptr); 82 83 void pushIVUsers(Instruction *Def, 84 SmallPtrSet<Instruction *, 16> &Simplified, 85 SmallVectorImpl<std::pair<Instruction *, Instruction *>> 86 &SimpleIVUsers); 87 88 Value *foldIVUser(Instruction *UseInst, Instruction *IVOperand); 89 90 bool eliminateIdentitySCEV(Instruction *UseInst, Instruction *IVOperand); 91 bool replaceIVUserWithLoopInvariant(Instruction *UseInst); 92 bool replaceFloatIVWithIntegerIV(Instruction *UseInst); 93 94 bool eliminateOverflowIntrinsic(WithOverflowInst *WO); 95 bool eliminateSaturatingIntrinsic(SaturatingInst *SI); 96 bool eliminateTrunc(TruncInst *TI); 97 bool eliminateIVUser(Instruction *UseInst, Instruction *IVOperand); 98 bool makeIVComparisonInvariant(ICmpInst *ICmp, Instruction *IVOperand); 99 void eliminateIVComparison(ICmpInst *ICmp, Instruction *IVOperand); 100 void simplifyIVRemainder(BinaryOperator *Rem, Instruction *IVOperand, 101 bool IsSigned); 102 void replaceRemWithNumerator(BinaryOperator *Rem); 103 void replaceRemWithNumeratorOrZero(BinaryOperator *Rem); 104 void replaceSRemWithURem(BinaryOperator *Rem); 105 bool eliminateSDiv(BinaryOperator *SDiv); 106 bool strengthenBinaryOp(BinaryOperator *BO, Instruction *IVOperand); 107 bool strengthenOverflowingOperation(BinaryOperator *OBO, 108 Instruction *IVOperand); 109 bool strengthenRightShift(BinaryOperator *BO, Instruction *IVOperand); 110 }; 111 } 112 113 /// Find a point in code which dominates all given instructions. We can safely 114 /// assume that, whatever fact we can prove at the found point, this fact is 115 /// also true for each of the given instructions. 116 static Instruction *findCommonDominator(ArrayRef<Instruction *> Instructions, 117 DominatorTree &DT) { 118 Instruction *CommonDom = nullptr; 119 for (auto *Insn : Instructions) 120 CommonDom = 121 CommonDom ? DT.findNearestCommonDominator(CommonDom, Insn) : Insn; 122 assert(CommonDom && "Common dominator not found?"); 123 return CommonDom; 124 } 125 126 /// Fold an IV operand into its use. This removes increments of an 127 /// aligned IV when used by a instruction that ignores the low bits. 128 /// 129 /// IVOperand is guaranteed SCEVable, but UseInst may not be. 130 /// 131 /// Return the operand of IVOperand for this induction variable if IVOperand can 132 /// be folded (in case more folding opportunities have been exposed). 133 /// Otherwise return null. 134 Value *SimplifyIndvar::foldIVUser(Instruction *UseInst, Instruction *IVOperand) { 135 Value *IVSrc = nullptr; 136 const unsigned OperIdx = 0; 137 const SCEV *FoldedExpr = nullptr; 138 bool MustDropExactFlag = false; 139 switch (UseInst->getOpcode()) { 140 default: 141 return nullptr; 142 case Instruction::UDiv: 143 case Instruction::LShr: 144 // We're only interested in the case where we know something about 145 // the numerator and have a constant denominator. 146 if (IVOperand != UseInst->getOperand(OperIdx) || 147 !isa<ConstantInt>(UseInst->getOperand(1))) 148 return nullptr; 149 150 // Attempt to fold a binary operator with constant operand. 151 // e.g. ((I + 1) >> 2) => I >> 2 152 if (!isa<BinaryOperator>(IVOperand) 153 || !isa<ConstantInt>(IVOperand->getOperand(1))) 154 return nullptr; 155 156 IVSrc = IVOperand->getOperand(0); 157 // IVSrc must be the (SCEVable) IV, since the other operand is const. 158 assert(SE->isSCEVable(IVSrc->getType()) && "Expect SCEVable IV operand"); 159 160 ConstantInt *D = cast<ConstantInt>(UseInst->getOperand(1)); 161 if (UseInst->getOpcode() == Instruction::LShr) { 162 // Get a constant for the divisor. See createSCEV. 163 uint32_t BitWidth = cast<IntegerType>(UseInst->getType())->getBitWidth(); 164 if (D->getValue().uge(BitWidth)) 165 return nullptr; 166 167 D = ConstantInt::get(UseInst->getContext(), 168 APInt::getOneBitSet(BitWidth, D->getZExtValue())); 169 } 170 const auto *LHS = SE->getSCEV(IVSrc); 171 const auto *RHS = SE->getSCEV(D); 172 FoldedExpr = SE->getUDivExpr(LHS, RHS); 173 // We might have 'exact' flag set at this point which will no longer be 174 // correct after we make the replacement. 175 if (UseInst->isExact() && LHS != SE->getMulExpr(FoldedExpr, RHS)) 176 MustDropExactFlag = true; 177 } 178 // We have something that might fold it's operand. Compare SCEVs. 179 if (!SE->isSCEVable(UseInst->getType())) 180 return nullptr; 181 182 // Bypass the operand if SCEV can prove it has no effect. 183 if (SE->getSCEV(UseInst) != FoldedExpr) 184 return nullptr; 185 186 LLVM_DEBUG(dbgs() << "INDVARS: Eliminated IV operand: " << *IVOperand 187 << " -> " << *UseInst << '\n'); 188 189 UseInst->setOperand(OperIdx, IVSrc); 190 assert(SE->getSCEV(UseInst) == FoldedExpr && "bad SCEV with folded oper"); 191 192 if (MustDropExactFlag) 193 UseInst->dropPoisonGeneratingFlags(); 194 195 ++NumElimOperand; 196 Changed = true; 197 if (IVOperand->use_empty()) 198 DeadInsts.emplace_back(IVOperand); 199 return IVSrc; 200 } 201 202 bool SimplifyIndvar::makeIVComparisonInvariant(ICmpInst *ICmp, 203 Instruction *IVOperand) { 204 auto *Preheader = L->getLoopPreheader(); 205 if (!Preheader) 206 return false; 207 unsigned IVOperIdx = 0; 208 ICmpInst::Predicate Pred = ICmp->getPredicate(); 209 if (IVOperand != ICmp->getOperand(0)) { 210 // Swapped 211 assert(IVOperand == ICmp->getOperand(1) && "Can't find IVOperand"); 212 IVOperIdx = 1; 213 Pred = ICmpInst::getSwappedPredicate(Pred); 214 } 215 216 // Get the SCEVs for the ICmp operands (in the specific context of the 217 // current loop) 218 const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent()); 219 const SCEV *S = SE->getSCEVAtScope(ICmp->getOperand(IVOperIdx), ICmpLoop); 220 const SCEV *X = SE->getSCEVAtScope(ICmp->getOperand(1 - IVOperIdx), ICmpLoop); 221 auto LIP = SE->getLoopInvariantPredicate(Pred, S, X, L, ICmp); 222 if (!LIP) 223 return false; 224 ICmpInst::Predicate InvariantPredicate = LIP->Pred; 225 const SCEV *InvariantLHS = LIP->LHS; 226 const SCEV *InvariantRHS = LIP->RHS; 227 228 // Do not generate something ridiculous. 229 auto *PHTerm = Preheader->getTerminator(); 230 if (Rewriter.isHighCostExpansion({InvariantLHS, InvariantRHS}, L, 231 2 * SCEVCheapExpansionBudget, TTI, PHTerm) || 232 !Rewriter.isSafeToExpandAt(InvariantLHS, PHTerm) || 233 !Rewriter.isSafeToExpandAt(InvariantRHS, PHTerm)) 234 return false; 235 auto *NewLHS = 236 Rewriter.expandCodeFor(InvariantLHS, IVOperand->getType(), PHTerm); 237 auto *NewRHS = 238 Rewriter.expandCodeFor(InvariantRHS, IVOperand->getType(), PHTerm); 239 LLVM_DEBUG(dbgs() << "INDVARS: Simplified comparison: " << *ICmp << '\n'); 240 ICmp->setPredicate(InvariantPredicate); 241 ICmp->setOperand(0, NewLHS); 242 ICmp->setOperand(1, NewRHS); 243 RunUnswitching = true; 244 return true; 245 } 246 247 /// SimplifyIVUsers helper for eliminating useless 248 /// comparisons against an induction variable. 249 void SimplifyIndvar::eliminateIVComparison(ICmpInst *ICmp, 250 Instruction *IVOperand) { 251 unsigned IVOperIdx = 0; 252 ICmpInst::Predicate Pred = ICmp->getPredicate(); 253 ICmpInst::Predicate OriginalPred = Pred; 254 if (IVOperand != ICmp->getOperand(0)) { 255 // Swapped 256 assert(IVOperand == ICmp->getOperand(1) && "Can't find IVOperand"); 257 IVOperIdx = 1; 258 Pred = ICmpInst::getSwappedPredicate(Pred); 259 } 260 261 // Get the SCEVs for the ICmp operands (in the specific context of the 262 // current loop) 263 const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent()); 264 const SCEV *S = SE->getSCEVAtScope(ICmp->getOperand(IVOperIdx), ICmpLoop); 265 const SCEV *X = SE->getSCEVAtScope(ICmp->getOperand(1 - IVOperIdx), ICmpLoop); 266 267 // If the condition is always true or always false in the given context, 268 // replace it with a constant value. 269 SmallVector<Instruction *, 4> Users; 270 for (auto *U : ICmp->users()) 271 Users.push_back(cast<Instruction>(U)); 272 const Instruction *CtxI = findCommonDominator(Users, *DT); 273 if (auto Ev = SE->evaluatePredicateAt(Pred, S, X, CtxI)) { 274 SE->forgetValue(ICmp); 275 ICmp->replaceAllUsesWith(ConstantInt::getBool(ICmp->getContext(), *Ev)); 276 DeadInsts.emplace_back(ICmp); 277 LLVM_DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n'); 278 } else if (makeIVComparisonInvariant(ICmp, IVOperand)) { 279 // fallthrough to end of function 280 } else if (ICmpInst::isSigned(OriginalPred) && 281 SE->isKnownNonNegative(S) && SE->isKnownNonNegative(X)) { 282 // If we were unable to make anything above, all we can is to canonicalize 283 // the comparison hoping that it will open the doors for other 284 // optimizations. If we find out that we compare two non-negative values, 285 // we turn the instruction's predicate to its unsigned version. Note that 286 // we cannot rely on Pred here unless we check if we have swapped it. 287 assert(ICmp->getPredicate() == OriginalPred && "Predicate changed?"); 288 LLVM_DEBUG(dbgs() << "INDVARS: Turn to unsigned comparison: " << *ICmp 289 << '\n'); 290 ICmp->setPredicate(ICmpInst::getUnsignedPredicate(OriginalPred)); 291 } else 292 return; 293 294 ++NumElimCmp; 295 Changed = true; 296 } 297 298 bool SimplifyIndvar::eliminateSDiv(BinaryOperator *SDiv) { 299 // Get the SCEVs for the ICmp operands. 300 auto *N = SE->getSCEV(SDiv->getOperand(0)); 301 auto *D = SE->getSCEV(SDiv->getOperand(1)); 302 303 // Simplify unnecessary loops away. 304 const Loop *L = LI->getLoopFor(SDiv->getParent()); 305 N = SE->getSCEVAtScope(N, L); 306 D = SE->getSCEVAtScope(D, L); 307 308 // Replace sdiv by udiv if both of the operands are non-negative 309 if (SE->isKnownNonNegative(N) && SE->isKnownNonNegative(D)) { 310 auto *UDiv = BinaryOperator::Create( 311 BinaryOperator::UDiv, SDiv->getOperand(0), SDiv->getOperand(1), 312 SDiv->getName() + ".udiv", SDiv->getIterator()); 313 UDiv->setIsExact(SDiv->isExact()); 314 SDiv->replaceAllUsesWith(UDiv); 315 UDiv->setDebugLoc(SDiv->getDebugLoc()); 316 LLVM_DEBUG(dbgs() << "INDVARS: Simplified sdiv: " << *SDiv << '\n'); 317 ++NumSimplifiedSDiv; 318 Changed = true; 319 DeadInsts.push_back(SDiv); 320 return true; 321 } 322 323 return false; 324 } 325 326 // i %s n -> i %u n if i >= 0 and n >= 0 327 void SimplifyIndvar::replaceSRemWithURem(BinaryOperator *Rem) { 328 auto *N = Rem->getOperand(0), *D = Rem->getOperand(1); 329 auto *URem = BinaryOperator::Create(BinaryOperator::URem, N, D, 330 Rem->getName() + ".urem", Rem->getIterator()); 331 Rem->replaceAllUsesWith(URem); 332 URem->setDebugLoc(Rem->getDebugLoc()); 333 LLVM_DEBUG(dbgs() << "INDVARS: Simplified srem: " << *Rem << '\n'); 334 ++NumSimplifiedSRem; 335 Changed = true; 336 DeadInsts.emplace_back(Rem); 337 } 338 339 // i % n --> i if i is in [0,n). 340 void SimplifyIndvar::replaceRemWithNumerator(BinaryOperator *Rem) { 341 Rem->replaceAllUsesWith(Rem->getOperand(0)); 342 LLVM_DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n'); 343 ++NumElimRem; 344 Changed = true; 345 DeadInsts.emplace_back(Rem); 346 } 347 348 // (i+1) % n --> (i+1)==n?0:(i+1) if i is in [0,n). 349 void SimplifyIndvar::replaceRemWithNumeratorOrZero(BinaryOperator *Rem) { 350 auto *T = Rem->getType(); 351 auto *N = Rem->getOperand(0), *D = Rem->getOperand(1); 352 ICmpInst *ICmp = new ICmpInst(Rem->getIterator(), ICmpInst::ICMP_EQ, N, D); 353 SelectInst *Sel = 354 SelectInst::Create(ICmp, ConstantInt::get(T, 0), N, "iv.rem", Rem->getIterator()); 355 Rem->replaceAllUsesWith(Sel); 356 Sel->setDebugLoc(Rem->getDebugLoc()); 357 LLVM_DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n'); 358 ++NumElimRem; 359 Changed = true; 360 DeadInsts.emplace_back(Rem); 361 } 362 363 /// SimplifyIVUsers helper for eliminating useless remainder operations 364 /// operating on an induction variable or replacing srem by urem. 365 void SimplifyIndvar::simplifyIVRemainder(BinaryOperator *Rem, 366 Instruction *IVOperand, 367 bool IsSigned) { 368 auto *NValue = Rem->getOperand(0); 369 auto *DValue = Rem->getOperand(1); 370 // We're only interested in the case where we know something about 371 // the numerator, unless it is a srem, because we want to replace srem by urem 372 // in general. 373 bool UsedAsNumerator = IVOperand == NValue; 374 if (!UsedAsNumerator && !IsSigned) 375 return; 376 377 const SCEV *N = SE->getSCEV(NValue); 378 379 // Simplify unnecessary loops away. 380 const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent()); 381 N = SE->getSCEVAtScope(N, ICmpLoop); 382 383 bool IsNumeratorNonNegative = !IsSigned || SE->isKnownNonNegative(N); 384 385 // Do not proceed if the Numerator may be negative 386 if (!IsNumeratorNonNegative) 387 return; 388 389 const SCEV *D = SE->getSCEV(DValue); 390 D = SE->getSCEVAtScope(D, ICmpLoop); 391 392 if (UsedAsNumerator) { 393 auto LT = IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 394 if (SE->isKnownPredicate(LT, N, D)) { 395 replaceRemWithNumerator(Rem); 396 return; 397 } 398 399 auto *T = Rem->getType(); 400 const auto *NLessOne = SE->getMinusSCEV(N, SE->getOne(T)); 401 if (SE->isKnownPredicate(LT, NLessOne, D)) { 402 replaceRemWithNumeratorOrZero(Rem); 403 return; 404 } 405 } 406 407 // Try to replace SRem with URem, if both N and D are known non-negative. 408 // Since we had already check N, we only need to check D now 409 if (!IsSigned || !SE->isKnownNonNegative(D)) 410 return; 411 412 replaceSRemWithURem(Rem); 413 } 414 415 bool SimplifyIndvar::eliminateOverflowIntrinsic(WithOverflowInst *WO) { 416 const SCEV *LHS = SE->getSCEV(WO->getLHS()); 417 const SCEV *RHS = SE->getSCEV(WO->getRHS()); 418 if (!SE->willNotOverflow(WO->getBinaryOp(), WO->isSigned(), LHS, RHS)) 419 return false; 420 421 // Proved no overflow, nuke the overflow check and, if possible, the overflow 422 // intrinsic as well. 423 424 BinaryOperator *NewResult = BinaryOperator::Create( 425 WO->getBinaryOp(), WO->getLHS(), WO->getRHS(), "", WO->getIterator()); 426 427 if (WO->isSigned()) 428 NewResult->setHasNoSignedWrap(true); 429 else 430 NewResult->setHasNoUnsignedWrap(true); 431 432 SmallVector<ExtractValueInst *, 4> ToDelete; 433 434 for (auto *U : WO->users()) { 435 if (auto *EVI = dyn_cast<ExtractValueInst>(U)) { 436 if (EVI->getIndices()[0] == 1) 437 EVI->replaceAllUsesWith(ConstantInt::getFalse(WO->getContext())); 438 else { 439 assert(EVI->getIndices()[0] == 0 && "Only two possibilities!"); 440 EVI->replaceAllUsesWith(NewResult); 441 NewResult->setDebugLoc(EVI->getDebugLoc()); 442 } 443 ToDelete.push_back(EVI); 444 } 445 } 446 447 for (auto *EVI : ToDelete) 448 EVI->eraseFromParent(); 449 450 if (WO->use_empty()) 451 WO->eraseFromParent(); 452 453 Changed = true; 454 return true; 455 } 456 457 bool SimplifyIndvar::eliminateSaturatingIntrinsic(SaturatingInst *SI) { 458 const SCEV *LHS = SE->getSCEV(SI->getLHS()); 459 const SCEV *RHS = SE->getSCEV(SI->getRHS()); 460 if (!SE->willNotOverflow(SI->getBinaryOp(), SI->isSigned(), LHS, RHS)) 461 return false; 462 463 BinaryOperator *BO = BinaryOperator::Create( 464 SI->getBinaryOp(), SI->getLHS(), SI->getRHS(), SI->getName(), SI->getIterator()); 465 if (SI->isSigned()) 466 BO->setHasNoSignedWrap(); 467 else 468 BO->setHasNoUnsignedWrap(); 469 470 SI->replaceAllUsesWith(BO); 471 BO->setDebugLoc(SI->getDebugLoc()); 472 DeadInsts.emplace_back(SI); 473 Changed = true; 474 return true; 475 } 476 477 bool SimplifyIndvar::eliminateTrunc(TruncInst *TI) { 478 // It is always legal to replace 479 // icmp <pred> i32 trunc(iv), n 480 // with 481 // icmp <pred> i64 sext(trunc(iv)), sext(n), if pred is signed predicate. 482 // Or with 483 // icmp <pred> i64 zext(trunc(iv)), zext(n), if pred is unsigned predicate. 484 // Or with either of these if pred is an equality predicate. 485 // 486 // If we can prove that iv == sext(trunc(iv)) or iv == zext(trunc(iv)) for 487 // every comparison which uses trunc, it means that we can replace each of 488 // them with comparison of iv against sext/zext(n). We no longer need trunc 489 // after that. 490 // 491 // TODO: Should we do this if we can widen *some* comparisons, but not all 492 // of them? Sometimes it is enough to enable other optimizations, but the 493 // trunc instruction will stay in the loop. 494 Value *IV = TI->getOperand(0); 495 Type *IVTy = IV->getType(); 496 const SCEV *IVSCEV = SE->getSCEV(IV); 497 const SCEV *TISCEV = SE->getSCEV(TI); 498 499 // Check if iv == zext(trunc(iv)) and if iv == sext(trunc(iv)). If so, we can 500 // get rid of trunc 501 bool DoesSExtCollapse = false; 502 bool DoesZExtCollapse = false; 503 if (IVSCEV == SE->getSignExtendExpr(TISCEV, IVTy)) 504 DoesSExtCollapse = true; 505 if (IVSCEV == SE->getZeroExtendExpr(TISCEV, IVTy)) 506 DoesZExtCollapse = true; 507 508 // If neither sext nor zext does collapse, it is not profitable to do any 509 // transform. Bail. 510 if (!DoesSExtCollapse && !DoesZExtCollapse) 511 return false; 512 513 // Collect users of the trunc that look like comparisons against invariants. 514 // Bail if we find something different. 515 SmallVector<ICmpInst *, 4> ICmpUsers; 516 for (auto *U : TI->users()) { 517 // We don't care about users in unreachable blocks. 518 if (isa<Instruction>(U) && 519 !DT->isReachableFromEntry(cast<Instruction>(U)->getParent())) 520 continue; 521 ICmpInst *ICI = dyn_cast<ICmpInst>(U); 522 if (!ICI) return false; 523 assert(L->contains(ICI->getParent()) && "LCSSA form broken?"); 524 if (!(ICI->getOperand(0) == TI && L->isLoopInvariant(ICI->getOperand(1))) && 525 !(ICI->getOperand(1) == TI && L->isLoopInvariant(ICI->getOperand(0)))) 526 return false; 527 // If we cannot get rid of trunc, bail. 528 if (ICI->isSigned() && !DoesSExtCollapse) 529 return false; 530 if (ICI->isUnsigned() && !DoesZExtCollapse) 531 return false; 532 // For equality, either signed or unsigned works. 533 ICmpUsers.push_back(ICI); 534 } 535 536 auto CanUseZExt = [&](ICmpInst *ICI) { 537 // Unsigned comparison can be widened as unsigned. 538 if (ICI->isUnsigned()) 539 return true; 540 // Is it profitable to do zext? 541 if (!DoesZExtCollapse) 542 return false; 543 // For equality, we can safely zext both parts. 544 if (ICI->isEquality()) 545 return true; 546 // Otherwise we can only use zext when comparing two non-negative or two 547 // negative values. But in practice, we will never pass DoesZExtCollapse 548 // check for a negative value, because zext(trunc(x)) is non-negative. So 549 // it only make sense to check for non-negativity here. 550 const SCEV *SCEVOP1 = SE->getSCEV(ICI->getOperand(0)); 551 const SCEV *SCEVOP2 = SE->getSCEV(ICI->getOperand(1)); 552 return SE->isKnownNonNegative(SCEVOP1) && SE->isKnownNonNegative(SCEVOP2); 553 }; 554 // Replace all comparisons against trunc with comparisons against IV. 555 for (auto *ICI : ICmpUsers) { 556 bool IsSwapped = L->isLoopInvariant(ICI->getOperand(0)); 557 auto *Op1 = IsSwapped ? ICI->getOperand(0) : ICI->getOperand(1); 558 IRBuilder<> Builder(ICI); 559 Value *Ext = nullptr; 560 // For signed/unsigned predicate, replace the old comparison with comparison 561 // of immediate IV against sext/zext of the invariant argument. If we can 562 // use either sext or zext (i.e. we are dealing with equality predicate), 563 // then prefer zext as a more canonical form. 564 // TODO: If we see a signed comparison which can be turned into unsigned, 565 // we can do it here for canonicalization purposes. 566 ICmpInst::Predicate Pred = ICI->getPredicate(); 567 if (IsSwapped) Pred = ICmpInst::getSwappedPredicate(Pred); 568 if (CanUseZExt(ICI)) { 569 assert(DoesZExtCollapse && "Unprofitable zext?"); 570 Ext = Builder.CreateZExt(Op1, IVTy, "zext"); 571 Pred = ICmpInst::getUnsignedPredicate(Pred); 572 } else { 573 assert(DoesSExtCollapse && "Unprofitable sext?"); 574 Ext = Builder.CreateSExt(Op1, IVTy, "sext"); 575 assert(Pred == ICmpInst::getSignedPredicate(Pred) && "Must be signed!"); 576 } 577 bool Changed; 578 L->makeLoopInvariant(Ext, Changed); 579 (void)Changed; 580 auto *NewCmp = Builder.CreateICmp(Pred, IV, Ext); 581 ICI->replaceAllUsesWith(NewCmp); 582 DeadInsts.emplace_back(ICI); 583 } 584 585 // Trunc no longer needed. 586 TI->replaceAllUsesWith(PoisonValue::get(TI->getType())); 587 DeadInsts.emplace_back(TI); 588 return true; 589 } 590 591 /// Eliminate an operation that consumes a simple IV and has no observable 592 /// side-effect given the range of IV values. IVOperand is guaranteed SCEVable, 593 /// but UseInst may not be. 594 bool SimplifyIndvar::eliminateIVUser(Instruction *UseInst, 595 Instruction *IVOperand) { 596 if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) { 597 eliminateIVComparison(ICmp, IVOperand); 598 return true; 599 } 600 if (BinaryOperator *Bin = dyn_cast<BinaryOperator>(UseInst)) { 601 bool IsSRem = Bin->getOpcode() == Instruction::SRem; 602 if (IsSRem || Bin->getOpcode() == Instruction::URem) { 603 simplifyIVRemainder(Bin, IVOperand, IsSRem); 604 return true; 605 } 606 607 if (Bin->getOpcode() == Instruction::SDiv) 608 return eliminateSDiv(Bin); 609 } 610 611 if (auto *WO = dyn_cast<WithOverflowInst>(UseInst)) 612 if (eliminateOverflowIntrinsic(WO)) 613 return true; 614 615 if (auto *SI = dyn_cast<SaturatingInst>(UseInst)) 616 if (eliminateSaturatingIntrinsic(SI)) 617 return true; 618 619 if (auto *TI = dyn_cast<TruncInst>(UseInst)) 620 if (eliminateTrunc(TI)) 621 return true; 622 623 if (eliminateIdentitySCEV(UseInst, IVOperand)) 624 return true; 625 626 return false; 627 } 628 629 static Instruction *GetLoopInvariantInsertPosition(Loop *L, Instruction *Hint) { 630 if (auto *BB = L->getLoopPreheader()) 631 return BB->getTerminator(); 632 633 return Hint; 634 } 635 636 /// Replace the UseInst with a loop invariant expression if it is safe. 637 bool SimplifyIndvar::replaceIVUserWithLoopInvariant(Instruction *I) { 638 if (!SE->isSCEVable(I->getType())) 639 return false; 640 641 // Get the symbolic expression for this instruction. 642 const SCEV *S = SE->getSCEV(I); 643 644 if (!SE->isLoopInvariant(S, L)) 645 return false; 646 647 // Do not generate something ridiculous even if S is loop invariant. 648 if (Rewriter.isHighCostExpansion(S, L, SCEVCheapExpansionBudget, TTI, I)) 649 return false; 650 651 auto *IP = GetLoopInvariantInsertPosition(L, I); 652 653 if (!Rewriter.isSafeToExpandAt(S, IP)) { 654 LLVM_DEBUG(dbgs() << "INDVARS: Can not replace IV user: " << *I 655 << " with non-speculable loop invariant: " << *S << '\n'); 656 return false; 657 } 658 659 auto *Invariant = Rewriter.expandCodeFor(S, I->getType(), IP); 660 bool NeedToEmitLCSSAPhis = false; 661 if (!LI->replacementPreservesLCSSAForm(I, Invariant)) 662 NeedToEmitLCSSAPhis = true; 663 664 I->replaceAllUsesWith(Invariant); 665 LLVM_DEBUG(dbgs() << "INDVARS: Replace IV user: " << *I 666 << " with loop invariant: " << *S << '\n'); 667 668 if (NeedToEmitLCSSAPhis) { 669 SmallVector<Instruction *, 1> NeedsLCSSAPhis; 670 NeedsLCSSAPhis.push_back(cast<Instruction>(Invariant)); 671 formLCSSAForInstructions(NeedsLCSSAPhis, *DT, *LI, SE); 672 LLVM_DEBUG(dbgs() << " INDVARS: Replacement breaks LCSSA form" 673 << " inserting LCSSA Phis" << '\n'); 674 } 675 ++NumFoldedUser; 676 Changed = true; 677 DeadInsts.emplace_back(I); 678 return true; 679 } 680 681 /// Eliminate redundant type cast between integer and float. 682 bool SimplifyIndvar::replaceFloatIVWithIntegerIV(Instruction *UseInst) { 683 if (UseInst->getOpcode() != CastInst::SIToFP && 684 UseInst->getOpcode() != CastInst::UIToFP) 685 return false; 686 687 Instruction *IVOperand = cast<Instruction>(UseInst->getOperand(0)); 688 // Get the symbolic expression for this instruction. 689 const SCEV *IV = SE->getSCEV(IVOperand); 690 int MaskBits; 691 if (UseInst->getOpcode() == CastInst::SIToFP) 692 MaskBits = (int)SE->getSignedRange(IV).getMinSignedBits(); 693 else 694 MaskBits = (int)SE->getUnsignedRange(IV).getActiveBits(); 695 int DestNumSigBits = UseInst->getType()->getFPMantissaWidth(); 696 if (MaskBits <= DestNumSigBits) { 697 for (User *U : UseInst->users()) { 698 // Match for fptosi/fptoui of sitofp and with same type. 699 auto *CI = dyn_cast<CastInst>(U); 700 if (!CI) 701 continue; 702 703 CastInst::CastOps Opcode = CI->getOpcode(); 704 if (Opcode != CastInst::FPToSI && Opcode != CastInst::FPToUI) 705 continue; 706 707 Value *Conv = nullptr; 708 if (IVOperand->getType() != CI->getType()) { 709 IRBuilder<> Builder(CI); 710 StringRef Name = IVOperand->getName(); 711 // To match InstCombine logic, we only need sext if both fptosi and 712 // sitofp are used. If one of them is unsigned, then we can use zext. 713 if (SE->getTypeSizeInBits(IVOperand->getType()) > 714 SE->getTypeSizeInBits(CI->getType())) { 715 Conv = Builder.CreateTrunc(IVOperand, CI->getType(), Name + ".trunc"); 716 } else if (Opcode == CastInst::FPToUI || 717 UseInst->getOpcode() == CastInst::UIToFP) { 718 Conv = Builder.CreateZExt(IVOperand, CI->getType(), Name + ".zext"); 719 } else { 720 Conv = Builder.CreateSExt(IVOperand, CI->getType(), Name + ".sext"); 721 } 722 } else 723 Conv = IVOperand; 724 725 CI->replaceAllUsesWith(Conv); 726 DeadInsts.push_back(CI); 727 LLVM_DEBUG(dbgs() << "INDVARS: Replace IV user: " << *CI 728 << " with: " << *Conv << '\n'); 729 730 ++NumFoldedUser; 731 Changed = true; 732 } 733 } 734 735 return Changed; 736 } 737 738 /// Eliminate any operation that SCEV can prove is an identity function. 739 bool SimplifyIndvar::eliminateIdentitySCEV(Instruction *UseInst, 740 Instruction *IVOperand) { 741 if (!SE->isSCEVable(UseInst->getType()) || 742 UseInst->getType() != IVOperand->getType()) 743 return false; 744 745 const SCEV *UseSCEV = SE->getSCEV(UseInst); 746 if (UseSCEV != SE->getSCEV(IVOperand)) 747 return false; 748 749 // getSCEV(X) == getSCEV(Y) does not guarantee that X and Y are related in the 750 // dominator tree, even if X is an operand to Y. For instance, in 751 // 752 // %iv = phi i32 {0,+,1} 753 // br %cond, label %left, label %merge 754 // 755 // left: 756 // %X = add i32 %iv, 0 757 // br label %merge 758 // 759 // merge: 760 // %M = phi (%X, %iv) 761 // 762 // getSCEV(%M) == getSCEV(%X) == {0,+,1}, but %X does not dominate %M, and 763 // %M.replaceAllUsesWith(%X) would be incorrect. 764 765 if (isa<PHINode>(UseInst)) 766 // If UseInst is not a PHI node then we know that IVOperand dominates 767 // UseInst directly from the legality of SSA. 768 if (!DT || !DT->dominates(IVOperand, UseInst)) 769 return false; 770 771 if (!LI->replacementPreservesLCSSAForm(UseInst, IVOperand)) 772 return false; 773 774 // Make sure the operand is not more poisonous than the instruction. 775 if (!impliesPoison(IVOperand, UseInst)) { 776 SmallVector<Instruction *> DropPoisonGeneratingInsts; 777 if (!SE->canReuseInstruction(UseSCEV, IVOperand, DropPoisonGeneratingInsts)) 778 return false; 779 780 for (Instruction *I : DropPoisonGeneratingInsts) 781 I->dropPoisonGeneratingAnnotations(); 782 } 783 784 LLVM_DEBUG(dbgs() << "INDVARS: Eliminated identity: " << *UseInst << '\n'); 785 786 SE->forgetValue(UseInst); 787 UseInst->replaceAllUsesWith(IVOperand); 788 ++NumElimIdentity; 789 Changed = true; 790 DeadInsts.emplace_back(UseInst); 791 return true; 792 } 793 794 bool SimplifyIndvar::strengthenBinaryOp(BinaryOperator *BO, 795 Instruction *IVOperand) { 796 return (isa<OverflowingBinaryOperator>(BO) && 797 strengthenOverflowingOperation(BO, IVOperand)) || 798 (isa<ShlOperator>(BO) && strengthenRightShift(BO, IVOperand)); 799 } 800 801 /// Annotate BO with nsw / nuw if it provably does not signed-overflow / 802 /// unsigned-overflow. Returns true if anything changed, false otherwise. 803 bool SimplifyIndvar::strengthenOverflowingOperation(BinaryOperator *BO, 804 Instruction *IVOperand) { 805 auto Flags = SE->getStrengthenedNoWrapFlagsFromBinOp( 806 cast<OverflowingBinaryOperator>(BO)); 807 808 if (!Flags) 809 return false; 810 811 BO->setHasNoUnsignedWrap(ScalarEvolution::maskFlags(*Flags, SCEV::FlagNUW) == 812 SCEV::FlagNUW); 813 BO->setHasNoSignedWrap(ScalarEvolution::maskFlags(*Flags, SCEV::FlagNSW) == 814 SCEV::FlagNSW); 815 816 // The getStrengthenedNoWrapFlagsFromBinOp() check inferred additional nowrap 817 // flags on addrecs while performing zero/sign extensions. We could call 818 // forgetValue() here to make sure those flags also propagate to any other 819 // SCEV expressions based on the addrec. However, this can have pathological 820 // compile-time impact, see https://bugs.llvm.org/show_bug.cgi?id=50384. 821 return true; 822 } 823 824 /// Annotate the Shr in (X << IVOperand) >> C as exact using the 825 /// information from the IV's range. Returns true if anything changed, false 826 /// otherwise. 827 bool SimplifyIndvar::strengthenRightShift(BinaryOperator *BO, 828 Instruction *IVOperand) { 829 if (BO->getOpcode() == Instruction::Shl) { 830 bool Changed = false; 831 ConstantRange IVRange = SE->getUnsignedRange(SE->getSCEV(IVOperand)); 832 for (auto *U : BO->users()) { 833 const APInt *C; 834 if (match(U, 835 m_AShr(m_Shl(m_Value(), m_Specific(IVOperand)), m_APInt(C))) || 836 match(U, 837 m_LShr(m_Shl(m_Value(), m_Specific(IVOperand)), m_APInt(C)))) { 838 BinaryOperator *Shr = cast<BinaryOperator>(U); 839 if (!Shr->isExact() && IVRange.getUnsignedMin().uge(*C)) { 840 Shr->setIsExact(true); 841 Changed = true; 842 } 843 } 844 } 845 return Changed; 846 } 847 848 return false; 849 } 850 851 /// Add all uses of Def to the current IV's worklist. 852 void SimplifyIndvar::pushIVUsers( 853 Instruction *Def, SmallPtrSet<Instruction *, 16> &Simplified, 854 SmallVectorImpl<std::pair<Instruction *, Instruction *>> &SimpleIVUsers) { 855 for (User *U : Def->users()) { 856 Instruction *UI = cast<Instruction>(U); 857 858 // Avoid infinite or exponential worklist processing. 859 // Also ensure unique worklist users. 860 // If Def is a LoopPhi, it may not be in the Simplified set, so check for 861 // self edges first. 862 if (UI == Def) 863 continue; 864 865 // Only change the current Loop, do not change the other parts (e.g. other 866 // Loops). 867 if (!L->contains(UI)) 868 continue; 869 870 // Do not push the same instruction more than once. 871 if (!Simplified.insert(UI).second) 872 continue; 873 874 SimpleIVUsers.push_back(std::make_pair(UI, Def)); 875 } 876 } 877 878 /// Return true if this instruction generates a simple SCEV 879 /// expression in terms of that IV. 880 /// 881 /// This is similar to IVUsers' isInteresting() but processes each instruction 882 /// non-recursively when the operand is already known to be a simpleIVUser. 883 /// 884 static bool isSimpleIVUser(Instruction *I, const Loop *L, ScalarEvolution *SE) { 885 if (!SE->isSCEVable(I->getType())) 886 return false; 887 888 // Get the symbolic expression for this instruction. 889 const SCEV *S = SE->getSCEV(I); 890 891 // Only consider affine recurrences. 892 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S); 893 if (AR && AR->getLoop() == L) 894 return true; 895 896 return false; 897 } 898 899 /// Iteratively perform simplification on a worklist of users 900 /// of the specified induction variable. Each successive simplification may push 901 /// more users which may themselves be candidates for simplification. 902 /// 903 /// This algorithm does not require IVUsers analysis. Instead, it simplifies 904 /// instructions in-place during analysis. Rather than rewriting induction 905 /// variables bottom-up from their users, it transforms a chain of IVUsers 906 /// top-down, updating the IR only when it encounters a clear optimization 907 /// opportunity. 908 /// 909 /// Once DisableIVRewrite is default, LSR will be the only client of IVUsers. 910 /// 911 void SimplifyIndvar::simplifyUsers(PHINode *CurrIV, IVVisitor *V) { 912 if (!SE->isSCEVable(CurrIV->getType())) 913 return; 914 915 // Instructions processed by SimplifyIndvar for CurrIV. 916 SmallPtrSet<Instruction*,16> Simplified; 917 918 // Use-def pairs if IV users waiting to be processed for CurrIV. 919 SmallVector<std::pair<Instruction*, Instruction*>, 8> SimpleIVUsers; 920 921 // Push users of the current LoopPhi. In rare cases, pushIVUsers may be 922 // called multiple times for the same LoopPhi. This is the proper thing to 923 // do for loop header phis that use each other. 924 pushIVUsers(CurrIV, Simplified, SimpleIVUsers); 925 926 while (!SimpleIVUsers.empty()) { 927 std::pair<Instruction*, Instruction*> UseOper = 928 SimpleIVUsers.pop_back_val(); 929 Instruction *UseInst = UseOper.first; 930 931 // If a user of the IndVar is trivially dead, we prefer just to mark it dead 932 // rather than try to do some complex analysis or transformation (such as 933 // widening) basing on it. 934 // TODO: Propagate TLI and pass it here to handle more cases. 935 if (isInstructionTriviallyDead(UseInst, /* TLI */ nullptr)) { 936 DeadInsts.emplace_back(UseInst); 937 continue; 938 } 939 940 // Bypass back edges to avoid extra work. 941 if (UseInst == CurrIV) continue; 942 943 // Try to replace UseInst with a loop invariant before any other 944 // simplifications. 945 if (replaceIVUserWithLoopInvariant(UseInst)) 946 continue; 947 948 // Go further for the bitcast 'prtoint ptr to i64' or if the cast is done 949 // by truncation 950 if ((isa<PtrToIntInst>(UseInst)) || (isa<TruncInst>(UseInst))) 951 for (Use &U : UseInst->uses()) { 952 Instruction *User = cast<Instruction>(U.getUser()); 953 if (replaceIVUserWithLoopInvariant(User)) 954 break; // done replacing 955 } 956 957 Instruction *IVOperand = UseOper.second; 958 for (unsigned N = 0; IVOperand; ++N) { 959 assert(N <= Simplified.size() && "runaway iteration"); 960 (void) N; 961 962 Value *NewOper = foldIVUser(UseInst, IVOperand); 963 if (!NewOper) 964 break; // done folding 965 IVOperand = dyn_cast<Instruction>(NewOper); 966 } 967 if (!IVOperand) 968 continue; 969 970 if (eliminateIVUser(UseInst, IVOperand)) { 971 pushIVUsers(IVOperand, Simplified, SimpleIVUsers); 972 continue; 973 } 974 975 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(UseInst)) { 976 if (strengthenBinaryOp(BO, IVOperand)) { 977 // re-queue uses of the now modified binary operator and fall 978 // through to the checks that remain. 979 pushIVUsers(IVOperand, Simplified, SimpleIVUsers); 980 } 981 } 982 983 // Try to use integer induction for FPToSI of float induction directly. 984 if (replaceFloatIVWithIntegerIV(UseInst)) { 985 // Re-queue the potentially new direct uses of IVOperand. 986 pushIVUsers(IVOperand, Simplified, SimpleIVUsers); 987 continue; 988 } 989 990 CastInst *Cast = dyn_cast<CastInst>(UseInst); 991 if (V && Cast) { 992 V->visitCast(Cast); 993 continue; 994 } 995 if (isSimpleIVUser(UseInst, L, SE)) { 996 pushIVUsers(UseInst, Simplified, SimpleIVUsers); 997 } 998 } 999 } 1000 1001 namespace llvm { 1002 1003 void IVVisitor::anchor() { } 1004 1005 /// Simplify instructions that use this induction variable 1006 /// by using ScalarEvolution to analyze the IV's recurrence. 1007 /// Returns a pair where the first entry indicates that the function makes 1008 /// changes and the second entry indicates that it introduced new opportunities 1009 /// for loop unswitching. 1010 std::pair<bool, bool> simplifyUsersOfIV(PHINode *CurrIV, ScalarEvolution *SE, 1011 DominatorTree *DT, LoopInfo *LI, 1012 const TargetTransformInfo *TTI, 1013 SmallVectorImpl<WeakTrackingVH> &Dead, 1014 SCEVExpander &Rewriter, IVVisitor *V) { 1015 SimplifyIndvar SIV(LI->getLoopFor(CurrIV->getParent()), SE, DT, LI, TTI, 1016 Rewriter, Dead); 1017 SIV.simplifyUsers(CurrIV, V); 1018 return {SIV.hasChanged(), SIV.runUnswitching()}; 1019 } 1020 1021 /// Simplify users of induction variables within this 1022 /// loop. This does not actually change or add IVs. 1023 bool simplifyLoopIVs(Loop *L, ScalarEvolution *SE, DominatorTree *DT, 1024 LoopInfo *LI, const TargetTransformInfo *TTI, 1025 SmallVectorImpl<WeakTrackingVH> &Dead) { 1026 SCEVExpander Rewriter(*SE, SE->getDataLayout(), "indvars"); 1027 #ifndef NDEBUG 1028 Rewriter.setDebugType(DEBUG_TYPE); 1029 #endif 1030 bool Changed = false; 1031 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { 1032 const auto &[C, _] = 1033 simplifyUsersOfIV(cast<PHINode>(I), SE, DT, LI, TTI, Dead, Rewriter); 1034 Changed |= C; 1035 } 1036 return Changed; 1037 } 1038 1039 } // namespace llvm 1040 1041 namespace { 1042 //===----------------------------------------------------------------------===// 1043 // Widen Induction Variables - Extend the width of an IV to cover its 1044 // widest uses. 1045 //===----------------------------------------------------------------------===// 1046 1047 class WidenIV { 1048 // Parameters 1049 PHINode *OrigPhi; 1050 Type *WideType; 1051 1052 // Context 1053 LoopInfo *LI; 1054 Loop *L; 1055 ScalarEvolution *SE; 1056 DominatorTree *DT; 1057 1058 // Does the module have any calls to the llvm.experimental.guard intrinsic 1059 // at all? If not we can avoid scanning instructions looking for guards. 1060 bool HasGuards; 1061 1062 bool UsePostIncrementRanges; 1063 1064 // Statistics 1065 unsigned NumElimExt = 0; 1066 unsigned NumWidened = 0; 1067 1068 // Result 1069 PHINode *WidePhi = nullptr; 1070 Instruction *WideInc = nullptr; 1071 const SCEV *WideIncExpr = nullptr; 1072 SmallVectorImpl<WeakTrackingVH> &DeadInsts; 1073 1074 SmallPtrSet<Instruction *,16> Widened; 1075 1076 enum class ExtendKind { Zero, Sign, Unknown }; 1077 1078 // A map tracking the kind of extension used to widen each narrow IV 1079 // and narrow IV user. 1080 // Key: pointer to a narrow IV or IV user. 1081 // Value: the kind of extension used to widen this Instruction. 1082 DenseMap<AssertingVH<Instruction>, ExtendKind> ExtendKindMap; 1083 1084 using DefUserPair = std::pair<AssertingVH<Value>, AssertingVH<Instruction>>; 1085 1086 // A map with control-dependent ranges for post increment IV uses. The key is 1087 // a pair of IV def and a use of this def denoting the context. The value is 1088 // a ConstantRange representing possible values of the def at the given 1089 // context. 1090 DenseMap<DefUserPair, ConstantRange> PostIncRangeInfos; 1091 1092 std::optional<ConstantRange> getPostIncRangeInfo(Value *Def, 1093 Instruction *UseI) { 1094 DefUserPair Key(Def, UseI); 1095 auto It = PostIncRangeInfos.find(Key); 1096 return It == PostIncRangeInfos.end() 1097 ? std::optional<ConstantRange>(std::nullopt) 1098 : std::optional<ConstantRange>(It->second); 1099 } 1100 1101 void calculatePostIncRanges(PHINode *OrigPhi); 1102 void calculatePostIncRange(Instruction *NarrowDef, Instruction *NarrowUser); 1103 1104 void updatePostIncRangeInfo(Value *Def, Instruction *UseI, ConstantRange R) { 1105 DefUserPair Key(Def, UseI); 1106 auto It = PostIncRangeInfos.find(Key); 1107 if (It == PostIncRangeInfos.end()) 1108 PostIncRangeInfos.insert({Key, R}); 1109 else 1110 It->second = R.intersectWith(It->second); 1111 } 1112 1113 public: 1114 /// Record a link in the Narrow IV def-use chain along with the WideIV that 1115 /// computes the same value as the Narrow IV def. This avoids caching Use* 1116 /// pointers. 1117 struct NarrowIVDefUse { 1118 Instruction *NarrowDef = nullptr; 1119 Instruction *NarrowUse = nullptr; 1120 Instruction *WideDef = nullptr; 1121 1122 // True if the narrow def is never negative. Tracking this information lets 1123 // us use a sign extension instead of a zero extension or vice versa, when 1124 // profitable and legal. 1125 bool NeverNegative = false; 1126 1127 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD, 1128 bool NeverNegative) 1129 : NarrowDef(ND), NarrowUse(NU), WideDef(WD), 1130 NeverNegative(NeverNegative) {} 1131 }; 1132 1133 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo, ScalarEvolution *SEv, 1134 DominatorTree *DTree, SmallVectorImpl<WeakTrackingVH> &DI, 1135 bool HasGuards, bool UsePostIncrementRanges = true); 1136 1137 PHINode *createWideIV(SCEVExpander &Rewriter); 1138 1139 unsigned getNumElimExt() { return NumElimExt; }; 1140 unsigned getNumWidened() { return NumWidened; }; 1141 1142 protected: 1143 Value *createExtendInst(Value *NarrowOper, Type *WideType, bool IsSigned, 1144 Instruction *Use); 1145 1146 Instruction *cloneIVUser(NarrowIVDefUse DU, const SCEVAddRecExpr *WideAR); 1147 Instruction *cloneArithmeticIVUser(NarrowIVDefUse DU, 1148 const SCEVAddRecExpr *WideAR); 1149 Instruction *cloneBitwiseIVUser(NarrowIVDefUse DU); 1150 1151 ExtendKind getExtendKind(Instruction *I); 1152 1153 using WidenedRecTy = std::pair<const SCEVAddRecExpr *, ExtendKind>; 1154 1155 WidenedRecTy getWideRecurrence(NarrowIVDefUse DU); 1156 1157 WidenedRecTy getExtendedOperandRecurrence(NarrowIVDefUse DU); 1158 1159 const SCEV *getSCEVByOpCode(const SCEV *LHS, const SCEV *RHS, 1160 unsigned OpCode) const; 1161 1162 Instruction *widenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter, 1163 PHINode *OrigPhi, PHINode *WidePhi); 1164 void truncateIVUse(NarrowIVDefUse DU); 1165 1166 bool widenLoopCompare(NarrowIVDefUse DU); 1167 bool widenWithVariantUse(NarrowIVDefUse DU); 1168 1169 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef); 1170 1171 private: 1172 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers; 1173 }; 1174 } // namespace 1175 1176 /// Determine the insertion point for this user. By default, insert immediately 1177 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the 1178 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest 1179 /// common dominator for the incoming blocks. A nullptr can be returned if no 1180 /// viable location is found: it may happen if User is a PHI and Def only comes 1181 /// to this PHI from unreachable blocks. 1182 static Instruction *getInsertPointForUses(Instruction *User, Value *Def, 1183 DominatorTree *DT, LoopInfo *LI) { 1184 PHINode *PHI = dyn_cast<PHINode>(User); 1185 if (!PHI) 1186 return User; 1187 1188 Instruction *InsertPt = nullptr; 1189 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) { 1190 if (PHI->getIncomingValue(i) != Def) 1191 continue; 1192 1193 BasicBlock *InsertBB = PHI->getIncomingBlock(i); 1194 1195 if (!DT->isReachableFromEntry(InsertBB)) 1196 continue; 1197 1198 if (!InsertPt) { 1199 InsertPt = InsertBB->getTerminator(); 1200 continue; 1201 } 1202 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB); 1203 InsertPt = InsertBB->getTerminator(); 1204 } 1205 1206 // If we have skipped all inputs, it means that Def only comes to Phi from 1207 // unreachable blocks. 1208 if (!InsertPt) 1209 return nullptr; 1210 1211 auto *DefI = dyn_cast<Instruction>(Def); 1212 if (!DefI) 1213 return InsertPt; 1214 1215 assert(DT->dominates(DefI, InsertPt) && "def does not dominate all uses"); 1216 1217 auto *L = LI->getLoopFor(DefI->getParent()); 1218 assert(!L || L->contains(LI->getLoopFor(InsertPt->getParent()))); 1219 1220 for (auto *DTN = (*DT)[InsertPt->getParent()]; DTN; DTN = DTN->getIDom()) 1221 if (LI->getLoopFor(DTN->getBlock()) == L) 1222 return DTN->getBlock()->getTerminator(); 1223 1224 llvm_unreachable("DefI dominates InsertPt!"); 1225 } 1226 1227 WidenIV::WidenIV(const WideIVInfo &WI, LoopInfo *LInfo, ScalarEvolution *SEv, 1228 DominatorTree *DTree, SmallVectorImpl<WeakTrackingVH> &DI, 1229 bool HasGuards, bool UsePostIncrementRanges) 1230 : OrigPhi(WI.NarrowIV), WideType(WI.WidestNativeType), LI(LInfo), 1231 L(LI->getLoopFor(OrigPhi->getParent())), SE(SEv), DT(DTree), 1232 HasGuards(HasGuards), UsePostIncrementRanges(UsePostIncrementRanges), 1233 DeadInsts(DI) { 1234 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV"); 1235 ExtendKindMap[OrigPhi] = WI.IsSigned ? ExtendKind::Sign : ExtendKind::Zero; 1236 } 1237 1238 Value *WidenIV::createExtendInst(Value *NarrowOper, Type *WideType, 1239 bool IsSigned, Instruction *Use) { 1240 // Set the debug location and conservative insertion point. 1241 IRBuilder<> Builder(Use); 1242 // Hoist the insertion point into loop preheaders as far as possible. 1243 for (const Loop *L = LI->getLoopFor(Use->getParent()); 1244 L && L->getLoopPreheader() && L->isLoopInvariant(NarrowOper); 1245 L = L->getParentLoop()) 1246 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator()); 1247 1248 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) : 1249 Builder.CreateZExt(NarrowOper, WideType); 1250 } 1251 1252 /// Instantiate a wide operation to replace a narrow operation. This only needs 1253 /// to handle operations that can evaluation to SCEVAddRec. It can safely return 1254 /// 0 for any operation we decide not to clone. 1255 Instruction *WidenIV::cloneIVUser(WidenIV::NarrowIVDefUse DU, 1256 const SCEVAddRecExpr *WideAR) { 1257 unsigned Opcode = DU.NarrowUse->getOpcode(); 1258 switch (Opcode) { 1259 default: 1260 return nullptr; 1261 case Instruction::Add: 1262 case Instruction::Mul: 1263 case Instruction::UDiv: 1264 case Instruction::Sub: 1265 return cloneArithmeticIVUser(DU, WideAR); 1266 1267 case Instruction::And: 1268 case Instruction::Or: 1269 case Instruction::Xor: 1270 case Instruction::Shl: 1271 case Instruction::LShr: 1272 case Instruction::AShr: 1273 return cloneBitwiseIVUser(DU); 1274 } 1275 } 1276 1277 Instruction *WidenIV::cloneBitwiseIVUser(WidenIV::NarrowIVDefUse DU) { 1278 Instruction *NarrowUse = DU.NarrowUse; 1279 Instruction *NarrowDef = DU.NarrowDef; 1280 Instruction *WideDef = DU.WideDef; 1281 1282 LLVM_DEBUG(dbgs() << "Cloning bitwise IVUser: " << *NarrowUse << "\n"); 1283 1284 // Replace NarrowDef operands with WideDef. Otherwise, we don't know anything 1285 // about the narrow operand yet so must insert a [sz]ext. It is probably loop 1286 // invariant and will be folded or hoisted. If it actually comes from a 1287 // widened IV, it should be removed during a future call to widenIVUse. 1288 bool IsSigned = getExtendKind(NarrowDef) == ExtendKind::Sign; 1289 Value *LHS = (NarrowUse->getOperand(0) == NarrowDef) 1290 ? WideDef 1291 : createExtendInst(NarrowUse->getOperand(0), WideType, 1292 IsSigned, NarrowUse); 1293 Value *RHS = (NarrowUse->getOperand(1) == NarrowDef) 1294 ? WideDef 1295 : createExtendInst(NarrowUse->getOperand(1), WideType, 1296 IsSigned, NarrowUse); 1297 1298 auto *NarrowBO = cast<BinaryOperator>(NarrowUse); 1299 auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS, 1300 NarrowBO->getName()); 1301 IRBuilder<> Builder(NarrowUse); 1302 Builder.Insert(WideBO); 1303 WideBO->copyIRFlags(NarrowBO); 1304 return WideBO; 1305 } 1306 1307 Instruction *WidenIV::cloneArithmeticIVUser(WidenIV::NarrowIVDefUse DU, 1308 const SCEVAddRecExpr *WideAR) { 1309 Instruction *NarrowUse = DU.NarrowUse; 1310 Instruction *NarrowDef = DU.NarrowDef; 1311 Instruction *WideDef = DU.WideDef; 1312 1313 LLVM_DEBUG(dbgs() << "Cloning arithmetic IVUser: " << *NarrowUse << "\n"); 1314 1315 unsigned IVOpIdx = (NarrowUse->getOperand(0) == NarrowDef) ? 0 : 1; 1316 1317 // We're trying to find X such that 1318 // 1319 // Widen(NarrowDef `op` NonIVNarrowDef) == WideAR == WideDef `op.wide` X 1320 // 1321 // We guess two solutions to X, sext(NonIVNarrowDef) and zext(NonIVNarrowDef), 1322 // and check using SCEV if any of them are correct. 1323 1324 // Returns true if extending NonIVNarrowDef according to `SignExt` is a 1325 // correct solution to X. 1326 auto GuessNonIVOperand = [&](bool SignExt) { 1327 const SCEV *WideLHS; 1328 const SCEV *WideRHS; 1329 1330 auto GetExtend = [this, SignExt](const SCEV *S, Type *Ty) { 1331 if (SignExt) 1332 return SE->getSignExtendExpr(S, Ty); 1333 return SE->getZeroExtendExpr(S, Ty); 1334 }; 1335 1336 if (IVOpIdx == 0) { 1337 WideLHS = SE->getSCEV(WideDef); 1338 const SCEV *NarrowRHS = SE->getSCEV(NarrowUse->getOperand(1)); 1339 WideRHS = GetExtend(NarrowRHS, WideType); 1340 } else { 1341 const SCEV *NarrowLHS = SE->getSCEV(NarrowUse->getOperand(0)); 1342 WideLHS = GetExtend(NarrowLHS, WideType); 1343 WideRHS = SE->getSCEV(WideDef); 1344 } 1345 1346 // WideUse is "WideDef `op.wide` X" as described in the comment. 1347 const SCEV *WideUse = 1348 getSCEVByOpCode(WideLHS, WideRHS, NarrowUse->getOpcode()); 1349 1350 return WideUse == WideAR; 1351 }; 1352 1353 bool SignExtend = getExtendKind(NarrowDef) == ExtendKind::Sign; 1354 if (!GuessNonIVOperand(SignExtend)) { 1355 SignExtend = !SignExtend; 1356 if (!GuessNonIVOperand(SignExtend)) 1357 return nullptr; 1358 } 1359 1360 Value *LHS = (NarrowUse->getOperand(0) == NarrowDef) 1361 ? WideDef 1362 : createExtendInst(NarrowUse->getOperand(0), WideType, 1363 SignExtend, NarrowUse); 1364 Value *RHS = (NarrowUse->getOperand(1) == NarrowDef) 1365 ? WideDef 1366 : createExtendInst(NarrowUse->getOperand(1), WideType, 1367 SignExtend, NarrowUse); 1368 1369 auto *NarrowBO = cast<BinaryOperator>(NarrowUse); 1370 auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS, 1371 NarrowBO->getName()); 1372 1373 IRBuilder<> Builder(NarrowUse); 1374 Builder.Insert(WideBO); 1375 WideBO->copyIRFlags(NarrowBO); 1376 return WideBO; 1377 } 1378 1379 WidenIV::ExtendKind WidenIV::getExtendKind(Instruction *I) { 1380 auto It = ExtendKindMap.find(I); 1381 assert(It != ExtendKindMap.end() && "Instruction not yet extended!"); 1382 return It->second; 1383 } 1384 1385 const SCEV *WidenIV::getSCEVByOpCode(const SCEV *LHS, const SCEV *RHS, 1386 unsigned OpCode) const { 1387 switch (OpCode) { 1388 case Instruction::Add: 1389 return SE->getAddExpr(LHS, RHS); 1390 case Instruction::Sub: 1391 return SE->getMinusSCEV(LHS, RHS); 1392 case Instruction::Mul: 1393 return SE->getMulExpr(LHS, RHS); 1394 case Instruction::UDiv: 1395 return SE->getUDivExpr(LHS, RHS); 1396 default: 1397 llvm_unreachable("Unsupported opcode."); 1398 }; 1399 } 1400 1401 namespace { 1402 1403 // Represents a interesting integer binary operation for 1404 // getExtendedOperandRecurrence. This may be a shl that is being treated as a 1405 // multiply or a 'or disjoint' that is being treated as 'add nsw nuw'. 1406 struct BinaryOp { 1407 unsigned Opcode; 1408 std::array<Value *, 2> Operands; 1409 bool IsNSW = false; 1410 bool IsNUW = false; 1411 1412 explicit BinaryOp(Instruction *Op) 1413 : Opcode(Op->getOpcode()), 1414 Operands({Op->getOperand(0), Op->getOperand(1)}) { 1415 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(Op)) { 1416 IsNSW = OBO->hasNoSignedWrap(); 1417 IsNUW = OBO->hasNoUnsignedWrap(); 1418 } 1419 } 1420 1421 explicit BinaryOp(Instruction::BinaryOps Opcode, Value *LHS, Value *RHS, 1422 bool IsNSW = false, bool IsNUW = false) 1423 : Opcode(Opcode), Operands({LHS, RHS}), IsNSW(IsNSW), IsNUW(IsNUW) {} 1424 }; 1425 1426 } // end anonymous namespace 1427 1428 static std::optional<BinaryOp> matchBinaryOp(Instruction *Op) { 1429 switch (Op->getOpcode()) { 1430 case Instruction::Add: 1431 case Instruction::Sub: 1432 case Instruction::Mul: 1433 return BinaryOp(Op); 1434 case Instruction::Or: { 1435 // Convert or disjoint into add nuw nsw. 1436 if (cast<PossiblyDisjointInst>(Op)->isDisjoint()) 1437 return BinaryOp(Instruction::Add, Op->getOperand(0), Op->getOperand(1), 1438 /*IsNSW=*/true, /*IsNUW=*/true); 1439 break; 1440 } 1441 case Instruction::Shl: { 1442 if (ConstantInt *SA = dyn_cast<ConstantInt>(Op->getOperand(1))) { 1443 unsigned BitWidth = cast<IntegerType>(SA->getType())->getBitWidth(); 1444 1445 // If the shift count is not less than the bitwidth, the result of 1446 // the shift is undefined. Don't try to analyze it, because the 1447 // resolution chosen here may differ from the resolution chosen in 1448 // other parts of the compiler. 1449 if (SA->getValue().ult(BitWidth)) { 1450 // We can safely preserve the nuw flag in all cases. It's also safe to 1451 // turn a nuw nsw shl into a nuw nsw mul. However, nsw in isolation 1452 // requires special handling. It can be preserved as long as we're not 1453 // left shifting by bitwidth - 1. 1454 bool IsNUW = Op->hasNoUnsignedWrap(); 1455 bool IsNSW = Op->hasNoSignedWrap() && 1456 (IsNUW || SA->getValue().ult(BitWidth - 1)); 1457 1458 ConstantInt *X = 1459 ConstantInt::get(Op->getContext(), 1460 APInt::getOneBitSet(BitWidth, SA->getZExtValue())); 1461 return BinaryOp(Instruction::Mul, Op->getOperand(0), X, IsNSW, IsNUW); 1462 } 1463 } 1464 1465 break; 1466 } 1467 } 1468 1469 return std::nullopt; 1470 } 1471 1472 /// No-wrap operations can transfer sign extension of their result to their 1473 /// operands. Generate the SCEV value for the widened operation without 1474 /// actually modifying the IR yet. If the expression after extending the 1475 /// operands is an AddRec for this loop, return the AddRec and the kind of 1476 /// extension used. 1477 WidenIV::WidenedRecTy 1478 WidenIV::getExtendedOperandRecurrence(WidenIV::NarrowIVDefUse DU) { 1479 auto Op = matchBinaryOp(DU.NarrowUse); 1480 if (!Op) 1481 return {nullptr, ExtendKind::Unknown}; 1482 1483 assert((Op->Opcode == Instruction::Add || Op->Opcode == Instruction::Sub || 1484 Op->Opcode == Instruction::Mul) && 1485 "Unexpected opcode"); 1486 1487 // One operand (NarrowDef) has already been extended to WideDef. Now determine 1488 // if extending the other will lead to a recurrence. 1489 const unsigned ExtendOperIdx = Op->Operands[0] == DU.NarrowDef ? 1 : 0; 1490 assert(Op->Operands[1 - ExtendOperIdx] == DU.NarrowDef && "bad DU"); 1491 1492 ExtendKind ExtKind = getExtendKind(DU.NarrowDef); 1493 if (!(ExtKind == ExtendKind::Sign && Op->IsNSW) && 1494 !(ExtKind == ExtendKind::Zero && Op->IsNUW)) { 1495 ExtKind = ExtendKind::Unknown; 1496 1497 // For a non-negative NarrowDef, we can choose either type of 1498 // extension. We want to use the current extend kind if legal 1499 // (see above), and we only hit this code if we need to check 1500 // the opposite case. 1501 if (DU.NeverNegative) { 1502 if (Op->IsNSW) { 1503 ExtKind = ExtendKind::Sign; 1504 } else if (Op->IsNUW) { 1505 ExtKind = ExtendKind::Zero; 1506 } 1507 } 1508 } 1509 1510 const SCEV *ExtendOperExpr = SE->getSCEV(Op->Operands[ExtendOperIdx]); 1511 if (ExtKind == ExtendKind::Sign) 1512 ExtendOperExpr = SE->getSignExtendExpr(ExtendOperExpr, WideType); 1513 else if (ExtKind == ExtendKind::Zero) 1514 ExtendOperExpr = SE->getZeroExtendExpr(ExtendOperExpr, WideType); 1515 else 1516 return {nullptr, ExtendKind::Unknown}; 1517 1518 // When creating this SCEV expr, don't apply the current operations NSW or NUW 1519 // flags. This instruction may be guarded by control flow that the no-wrap 1520 // behavior depends on. Non-control-equivalent instructions can be mapped to 1521 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW 1522 // semantics to those operations. 1523 const SCEV *lhs = SE->getSCEV(DU.WideDef); 1524 const SCEV *rhs = ExtendOperExpr; 1525 1526 // Let's swap operands to the initial order for the case of non-commutative 1527 // operations, like SUB. See PR21014. 1528 if (ExtendOperIdx == 0) 1529 std::swap(lhs, rhs); 1530 const SCEVAddRecExpr *AddRec = 1531 dyn_cast<SCEVAddRecExpr>(getSCEVByOpCode(lhs, rhs, Op->Opcode)); 1532 1533 if (!AddRec || AddRec->getLoop() != L) 1534 return {nullptr, ExtendKind::Unknown}; 1535 1536 return {AddRec, ExtKind}; 1537 } 1538 1539 /// Is this instruction potentially interesting for further simplification after 1540 /// widening it's type? In other words, can the extend be safely hoisted out of 1541 /// the loop with SCEV reducing the value to a recurrence on the same loop. If 1542 /// so, return the extended recurrence and the kind of extension used. Otherwise 1543 /// return {nullptr, ExtendKind::Unknown}. 1544 WidenIV::WidenedRecTy WidenIV::getWideRecurrence(WidenIV::NarrowIVDefUse DU) { 1545 if (!DU.NarrowUse->getType()->isIntegerTy()) 1546 return {nullptr, ExtendKind::Unknown}; 1547 1548 const SCEV *NarrowExpr = SE->getSCEV(DU.NarrowUse); 1549 if (SE->getTypeSizeInBits(NarrowExpr->getType()) >= 1550 SE->getTypeSizeInBits(WideType)) { 1551 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow 1552 // index. So don't follow this use. 1553 return {nullptr, ExtendKind::Unknown}; 1554 } 1555 1556 const SCEV *WideExpr; 1557 ExtendKind ExtKind; 1558 if (DU.NeverNegative) { 1559 WideExpr = SE->getSignExtendExpr(NarrowExpr, WideType); 1560 if (isa<SCEVAddRecExpr>(WideExpr)) 1561 ExtKind = ExtendKind::Sign; 1562 else { 1563 WideExpr = SE->getZeroExtendExpr(NarrowExpr, WideType); 1564 ExtKind = ExtendKind::Zero; 1565 } 1566 } else if (getExtendKind(DU.NarrowDef) == ExtendKind::Sign) { 1567 WideExpr = SE->getSignExtendExpr(NarrowExpr, WideType); 1568 ExtKind = ExtendKind::Sign; 1569 } else { 1570 WideExpr = SE->getZeroExtendExpr(NarrowExpr, WideType); 1571 ExtKind = ExtendKind::Zero; 1572 } 1573 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr); 1574 if (!AddRec || AddRec->getLoop() != L) 1575 return {nullptr, ExtendKind::Unknown}; 1576 return {AddRec, ExtKind}; 1577 } 1578 1579 /// This IV user cannot be widened. Replace this use of the original narrow IV 1580 /// with a truncation of the new wide IV to isolate and eliminate the narrow IV. 1581 void WidenIV::truncateIVUse(NarrowIVDefUse DU) { 1582 auto *InsertPt = getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT, LI); 1583 if (!InsertPt) 1584 return; 1585 LLVM_DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef << " for user " 1586 << *DU.NarrowUse << "\n"); 1587 ExtendKind ExtKind = getExtendKind(DU.NarrowDef); 1588 IRBuilder<> Builder(InsertPt); 1589 Value *Trunc = 1590 Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType(), "", 1591 DU.NeverNegative || ExtKind == ExtendKind::Zero, 1592 DU.NeverNegative || ExtKind == ExtendKind::Sign); 1593 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc); 1594 } 1595 1596 /// If the narrow use is a compare instruction, then widen the compare 1597 // (and possibly the other operand). The extend operation is hoisted into the 1598 // loop preheader as far as possible. 1599 bool WidenIV::widenLoopCompare(WidenIV::NarrowIVDefUse DU) { 1600 ICmpInst *Cmp = dyn_cast<ICmpInst>(DU.NarrowUse); 1601 if (!Cmp) 1602 return false; 1603 1604 // We can legally widen the comparison in the following two cases: 1605 // 1606 // - The signedness of the IV extension and comparison match 1607 // 1608 // - The narrow IV is always positive (and thus its sign extension is equal 1609 // to its zero extension). For instance, let's say we're zero extending 1610 // %narrow for the following use 1611 // 1612 // icmp slt i32 %narrow, %val ... (A) 1613 // 1614 // and %narrow is always positive. Then 1615 // 1616 // (A) == icmp slt i32 sext(%narrow), sext(%val) 1617 // == icmp slt i32 zext(%narrow), sext(%val) 1618 bool IsSigned = getExtendKind(DU.NarrowDef) == ExtendKind::Sign; 1619 if (!(DU.NeverNegative || IsSigned == Cmp->isSigned())) 1620 return false; 1621 1622 Value *Op = Cmp->getOperand(Cmp->getOperand(0) == DU.NarrowDef ? 1 : 0); 1623 unsigned CastWidth = SE->getTypeSizeInBits(Op->getType()); 1624 unsigned IVWidth = SE->getTypeSizeInBits(WideType); 1625 assert(CastWidth <= IVWidth && "Unexpected width while widening compare."); 1626 1627 // Widen the compare instruction. 1628 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef); 1629 1630 // Widen the other operand of the compare, if necessary. 1631 if (CastWidth < IVWidth) { 1632 Value *ExtOp = createExtendInst(Op, WideType, Cmp->isSigned(), Cmp); 1633 DU.NarrowUse->replaceUsesOfWith(Op, ExtOp); 1634 } 1635 return true; 1636 } 1637 1638 // The widenIVUse avoids generating trunc by evaluating the use as AddRec, this 1639 // will not work when: 1640 // 1) SCEV traces back to an instruction inside the loop that SCEV can not 1641 // expand, eg. add %indvar, (load %addr) 1642 // 2) SCEV finds a loop variant, eg. add %indvar, %loopvariant 1643 // While SCEV fails to avoid trunc, we can still try to use instruction 1644 // combining approach to prove trunc is not required. This can be further 1645 // extended with other instruction combining checks, but for now we handle the 1646 // following case (sub can be "add" and "mul", "nsw + sext" can be "nus + zext") 1647 // 1648 // Src: 1649 // %c = sub nsw %b, %indvar 1650 // %d = sext %c to i64 1651 // Dst: 1652 // %indvar.ext1 = sext %indvar to i64 1653 // %m = sext %b to i64 1654 // %d = sub nsw i64 %m, %indvar.ext1 1655 // Therefore, as long as the result of add/sub/mul is extended to wide type, no 1656 // trunc is required regardless of how %b is generated. This pattern is common 1657 // when calculating address in 64 bit architecture 1658 bool WidenIV::widenWithVariantUse(WidenIV::NarrowIVDefUse DU) { 1659 Instruction *NarrowUse = DU.NarrowUse; 1660 Instruction *NarrowDef = DU.NarrowDef; 1661 Instruction *WideDef = DU.WideDef; 1662 1663 // Handle the common case of add<nsw/nuw> 1664 const unsigned OpCode = NarrowUse->getOpcode(); 1665 // Only Add/Sub/Mul instructions are supported. 1666 if (OpCode != Instruction::Add && OpCode != Instruction::Sub && 1667 OpCode != Instruction::Mul) 1668 return false; 1669 1670 // The operand that is not defined by NarrowDef of DU. Let's call it the 1671 // other operand. 1672 assert((NarrowUse->getOperand(0) == NarrowDef || 1673 NarrowUse->getOperand(1) == NarrowDef) && 1674 "bad DU"); 1675 1676 const OverflowingBinaryOperator *OBO = 1677 cast<OverflowingBinaryOperator>(NarrowUse); 1678 ExtendKind ExtKind = getExtendKind(NarrowDef); 1679 bool CanSignExtend = ExtKind == ExtendKind::Sign && OBO->hasNoSignedWrap(); 1680 bool CanZeroExtend = ExtKind == ExtendKind::Zero && OBO->hasNoUnsignedWrap(); 1681 auto AnotherOpExtKind = ExtKind; 1682 1683 // Check that all uses are either: 1684 // - narrow def (in case of we are widening the IV increment); 1685 // - single-input LCSSA Phis; 1686 // - comparison of the chosen type; 1687 // - extend of the chosen type (raison d'etre). 1688 SmallVector<Instruction *, 4> ExtUsers; 1689 SmallVector<PHINode *, 4> LCSSAPhiUsers; 1690 SmallVector<ICmpInst *, 4> ICmpUsers; 1691 for (Use &U : NarrowUse->uses()) { 1692 Instruction *User = cast<Instruction>(U.getUser()); 1693 if (User == NarrowDef) 1694 continue; 1695 if (!L->contains(User)) { 1696 auto *LCSSAPhi = cast<PHINode>(User); 1697 // Make sure there is only 1 input, so that we don't have to split 1698 // critical edges. 1699 if (LCSSAPhi->getNumOperands() != 1) 1700 return false; 1701 LCSSAPhiUsers.push_back(LCSSAPhi); 1702 continue; 1703 } 1704 if (auto *ICmp = dyn_cast<ICmpInst>(User)) { 1705 auto Pred = ICmp->getPredicate(); 1706 // We have 3 types of predicates: signed, unsigned and equality 1707 // predicates. For equality, it's legal to widen icmp for either sign and 1708 // zero extend. For sign extend, we can also do so for signed predicates, 1709 // likeweise for zero extend we can widen icmp for unsigned predicates. 1710 if (ExtKind == ExtendKind::Zero && ICmpInst::isSigned(Pred)) 1711 return false; 1712 if (ExtKind == ExtendKind::Sign && ICmpInst::isUnsigned(Pred)) 1713 return false; 1714 ICmpUsers.push_back(ICmp); 1715 continue; 1716 } 1717 if (ExtKind == ExtendKind::Sign) 1718 User = dyn_cast<SExtInst>(User); 1719 else 1720 User = dyn_cast<ZExtInst>(User); 1721 if (!User || User->getType() != WideType) 1722 return false; 1723 ExtUsers.push_back(User); 1724 } 1725 if (ExtUsers.empty()) { 1726 DeadInsts.emplace_back(NarrowUse); 1727 return true; 1728 } 1729 1730 // We'll prove some facts that should be true in the context of ext users. If 1731 // there is no users, we are done now. If there are some, pick their common 1732 // dominator as context. 1733 const Instruction *CtxI = findCommonDominator(ExtUsers, *DT); 1734 1735 if (!CanSignExtend && !CanZeroExtend) { 1736 // Because InstCombine turns 'sub nuw' to 'add' losing the no-wrap flag, we 1737 // will most likely not see it. Let's try to prove it. 1738 if (OpCode != Instruction::Add) 1739 return false; 1740 if (ExtKind != ExtendKind::Zero) 1741 return false; 1742 const SCEV *LHS = SE->getSCEV(OBO->getOperand(0)); 1743 const SCEV *RHS = SE->getSCEV(OBO->getOperand(1)); 1744 // TODO: Support case for NarrowDef = NarrowUse->getOperand(1). 1745 if (NarrowUse->getOperand(0) != NarrowDef) 1746 return false; 1747 if (!SE->isKnownNegative(RHS)) 1748 return false; 1749 bool ProvedSubNUW = SE->isKnownPredicateAt(ICmpInst::ICMP_UGE, LHS, 1750 SE->getNegativeSCEV(RHS), CtxI); 1751 if (!ProvedSubNUW) 1752 return false; 1753 // In fact, our 'add' is 'sub nuw'. We will need to widen the 2nd operand as 1754 // neg(zext(neg(op))), which is basically sext(op). 1755 AnotherOpExtKind = ExtendKind::Sign; 1756 } 1757 1758 // Verifying that Defining operand is an AddRec 1759 const SCEV *Op1 = SE->getSCEV(WideDef); 1760 const SCEVAddRecExpr *AddRecOp1 = dyn_cast<SCEVAddRecExpr>(Op1); 1761 if (!AddRecOp1 || AddRecOp1->getLoop() != L) 1762 return false; 1763 1764 LLVM_DEBUG(dbgs() << "Cloning arithmetic IVUser: " << *NarrowUse << "\n"); 1765 1766 // Generating a widening use instruction. 1767 Value *LHS = 1768 (NarrowUse->getOperand(0) == NarrowDef) 1769 ? WideDef 1770 : createExtendInst(NarrowUse->getOperand(0), WideType, 1771 AnotherOpExtKind == ExtendKind::Sign, NarrowUse); 1772 Value *RHS = 1773 (NarrowUse->getOperand(1) == NarrowDef) 1774 ? WideDef 1775 : createExtendInst(NarrowUse->getOperand(1), WideType, 1776 AnotherOpExtKind == ExtendKind::Sign, NarrowUse); 1777 1778 auto *NarrowBO = cast<BinaryOperator>(NarrowUse); 1779 auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS, 1780 NarrowBO->getName()); 1781 IRBuilder<> Builder(NarrowUse); 1782 Builder.Insert(WideBO); 1783 WideBO->copyIRFlags(NarrowBO); 1784 ExtendKindMap[NarrowUse] = ExtKind; 1785 1786 for (Instruction *User : ExtUsers) { 1787 assert(User->getType() == WideType && "Checked before!"); 1788 LLVM_DEBUG(dbgs() << "INDVARS: eliminating " << *User << " replaced by " 1789 << *WideBO << "\n"); 1790 ++NumElimExt; 1791 User->replaceAllUsesWith(WideBO); 1792 DeadInsts.emplace_back(User); 1793 } 1794 1795 for (PHINode *User : LCSSAPhiUsers) { 1796 assert(User->getNumOperands() == 1 && "Checked before!"); 1797 Builder.SetInsertPoint(User); 1798 auto *WidePN = 1799 Builder.CreatePHI(WideBO->getType(), 1, User->getName() + ".wide"); 1800 BasicBlock *LoopExitingBlock = User->getParent()->getSinglePredecessor(); 1801 assert(LoopExitingBlock && L->contains(LoopExitingBlock) && 1802 "Not a LCSSA Phi?"); 1803 WidePN->addIncoming(WideBO, LoopExitingBlock); 1804 Builder.SetInsertPoint(User->getParent(), 1805 User->getParent()->getFirstInsertionPt()); 1806 auto *TruncPN = Builder.CreateTrunc(WidePN, User->getType()); 1807 User->replaceAllUsesWith(TruncPN); 1808 DeadInsts.emplace_back(User); 1809 } 1810 1811 for (ICmpInst *User : ICmpUsers) { 1812 Builder.SetInsertPoint(User); 1813 auto ExtendedOp = [&](Value * V)->Value * { 1814 if (V == NarrowUse) 1815 return WideBO; 1816 if (ExtKind == ExtendKind::Zero) 1817 return Builder.CreateZExt(V, WideBO->getType()); 1818 else 1819 return Builder.CreateSExt(V, WideBO->getType()); 1820 }; 1821 auto Pred = User->getPredicate(); 1822 auto *LHS = ExtendedOp(User->getOperand(0)); 1823 auto *RHS = ExtendedOp(User->getOperand(1)); 1824 auto *WideCmp = 1825 Builder.CreateICmp(Pred, LHS, RHS, User->getName() + ".wide"); 1826 User->replaceAllUsesWith(WideCmp); 1827 DeadInsts.emplace_back(User); 1828 } 1829 1830 return true; 1831 } 1832 1833 /// Determine whether an individual user of the narrow IV can be widened. If so, 1834 /// return the wide clone of the user. 1835 Instruction *WidenIV::widenIVUse(WidenIV::NarrowIVDefUse DU, 1836 SCEVExpander &Rewriter, PHINode *OrigPhi, 1837 PHINode *WidePhi) { 1838 assert(ExtendKindMap.count(DU.NarrowDef) && 1839 "Should already know the kind of extension used to widen NarrowDef"); 1840 1841 // This narrow use can be widened by a sext if it's non-negative or its narrow 1842 // def was widened by a sext. Same for zext. 1843 bool CanWidenBySExt = 1844 DU.NeverNegative || getExtendKind(DU.NarrowDef) == ExtendKind::Sign; 1845 bool CanWidenByZExt = 1846 DU.NeverNegative || getExtendKind(DU.NarrowDef) == ExtendKind::Zero; 1847 1848 // Stop traversing the def-use chain at inner-loop phis or post-loop phis. 1849 if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) { 1850 if (LI->getLoopFor(UsePhi->getParent()) != L) { 1851 // For LCSSA phis, sink the truncate outside the loop. 1852 // After SimplifyCFG most loop exit targets have a single predecessor. 1853 // Otherwise fall back to a truncate within the loop. 1854 if (UsePhi->getNumOperands() != 1) 1855 truncateIVUse(DU); 1856 else { 1857 // Widening the PHI requires us to insert a trunc. The logical place 1858 // for this trunc is in the same BB as the PHI. This is not possible if 1859 // the BB is terminated by a catchswitch. 1860 if (isa<CatchSwitchInst>(UsePhi->getParent()->getTerminator())) 1861 return nullptr; 1862 1863 PHINode *WidePhi = 1864 PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide", 1865 UsePhi->getIterator()); 1866 WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0)); 1867 BasicBlock *WidePhiBB = WidePhi->getParent(); 1868 IRBuilder<> Builder(WidePhiBB, WidePhiBB->getFirstInsertionPt()); 1869 Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType(), "", 1870 CanWidenByZExt, CanWidenBySExt); 1871 UsePhi->replaceAllUsesWith(Trunc); 1872 DeadInsts.emplace_back(UsePhi); 1873 LLVM_DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi << " to " 1874 << *WidePhi << "\n"); 1875 } 1876 return nullptr; 1877 } 1878 } 1879 1880 // Our raison d'etre! Eliminate sign and zero extension. 1881 if ((match(DU.NarrowUse, m_SExtLike(m_Value())) && CanWidenBySExt) || 1882 (isa<ZExtInst>(DU.NarrowUse) && CanWidenByZExt)) { 1883 Value *NewDef = DU.WideDef; 1884 if (DU.NarrowUse->getType() != WideType) { 1885 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType()); 1886 unsigned IVWidth = SE->getTypeSizeInBits(WideType); 1887 if (CastWidth < IVWidth) { 1888 // The cast isn't as wide as the IV, so insert a Trunc. 1889 IRBuilder<> Builder(DU.NarrowUse); 1890 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType(), "", 1891 CanWidenByZExt, CanWidenBySExt); 1892 } 1893 else { 1894 // A wider extend was hidden behind a narrower one. This may induce 1895 // another round of IV widening in which the intermediate IV becomes 1896 // dead. It should be very rare. 1897 LLVM_DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi 1898 << " not wide enough to subsume " << *DU.NarrowUse 1899 << "\n"); 1900 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef); 1901 NewDef = DU.NarrowUse; 1902 } 1903 } 1904 if (NewDef != DU.NarrowUse) { 1905 LLVM_DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse 1906 << " replaced by " << *DU.WideDef << "\n"); 1907 ++NumElimExt; 1908 DU.NarrowUse->replaceAllUsesWith(NewDef); 1909 DeadInsts.emplace_back(DU.NarrowUse); 1910 } 1911 // Now that the extend is gone, we want to expose it's uses for potential 1912 // further simplification. We don't need to directly inform SimplifyIVUsers 1913 // of the new users, because their parent IV will be processed later as a 1914 // new loop phi. If we preserved IVUsers analysis, we would also want to 1915 // push the uses of WideDef here. 1916 1917 // No further widening is needed. The deceased [sz]ext had done it for us. 1918 return nullptr; 1919 } 1920 1921 auto tryAddRecExpansion = [&]() -> Instruction* { 1922 // Does this user itself evaluate to a recurrence after widening? 1923 WidenedRecTy WideAddRec = getExtendedOperandRecurrence(DU); 1924 if (!WideAddRec.first) 1925 WideAddRec = getWideRecurrence(DU); 1926 assert((WideAddRec.first == nullptr) == 1927 (WideAddRec.second == ExtendKind::Unknown)); 1928 if (!WideAddRec.first) 1929 return nullptr; 1930 1931 auto CanUseWideInc = [&]() { 1932 if (!WideInc) 1933 return false; 1934 // Reuse the IV increment that SCEVExpander created. Recompute flags, 1935 // unless the flags for both increments agree and it is safe to use the 1936 // ones from the original inc. In that case, the new use of the wide 1937 // increment won't be more poisonous. 1938 bool NeedToRecomputeFlags = 1939 !SCEVExpander::canReuseFlagsFromOriginalIVInc( 1940 OrigPhi, WidePhi, DU.NarrowUse, WideInc) || 1941 DU.NarrowUse->hasNoUnsignedWrap() != WideInc->hasNoUnsignedWrap() || 1942 DU.NarrowUse->hasNoSignedWrap() != WideInc->hasNoSignedWrap(); 1943 return WideAddRec.first == WideIncExpr && 1944 Rewriter.hoistIVInc(WideInc, DU.NarrowUse, NeedToRecomputeFlags); 1945 }; 1946 1947 Instruction *WideUse = nullptr; 1948 if (CanUseWideInc()) 1949 WideUse = WideInc; 1950 else { 1951 WideUse = cloneIVUser(DU, WideAddRec.first); 1952 if (!WideUse) 1953 return nullptr; 1954 } 1955 // Evaluation of WideAddRec ensured that the narrow expression could be 1956 // extended outside the loop without overflow. This suggests that the wide use 1957 // evaluates to the same expression as the extended narrow use, but doesn't 1958 // absolutely guarantee it. Hence the following failsafe check. In rare cases 1959 // where it fails, we simply throw away the newly created wide use. 1960 if (WideAddRec.first != SE->getSCEV(WideUse)) { 1961 LLVM_DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse << ": " 1962 << *SE->getSCEV(WideUse) << " != " << *WideAddRec.first 1963 << "\n"); 1964 DeadInsts.emplace_back(WideUse); 1965 return nullptr; 1966 }; 1967 1968 // if we reached this point then we are going to replace 1969 // DU.NarrowUse with WideUse. Reattach DbgValue then. 1970 replaceAllDbgUsesWith(*DU.NarrowUse, *WideUse, *WideUse, *DT); 1971 1972 ExtendKindMap[DU.NarrowUse] = WideAddRec.second; 1973 // Returning WideUse pushes it on the worklist. 1974 return WideUse; 1975 }; 1976 1977 if (auto *I = tryAddRecExpansion()) 1978 return I; 1979 1980 // If use is a loop condition, try to promote the condition instead of 1981 // truncating the IV first. 1982 if (widenLoopCompare(DU)) 1983 return nullptr; 1984 1985 // We are here about to generate a truncate instruction that may hurt 1986 // performance because the scalar evolution expression computed earlier 1987 // in WideAddRec.first does not indicate a polynomial induction expression. 1988 // In that case, look at the operands of the use instruction to determine 1989 // if we can still widen the use instead of truncating its operand. 1990 if (widenWithVariantUse(DU)) 1991 return nullptr; 1992 1993 // This user does not evaluate to a recurrence after widening, so don't 1994 // follow it. Instead insert a Trunc to kill off the original use, 1995 // eventually isolating the original narrow IV so it can be removed. 1996 truncateIVUse(DU); 1997 return nullptr; 1998 } 1999 2000 /// Add eligible users of NarrowDef to NarrowIVUsers. 2001 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) { 2002 const SCEV *NarrowSCEV = SE->getSCEV(NarrowDef); 2003 bool NonNegativeDef = 2004 SE->isKnownPredicate(ICmpInst::ICMP_SGE, NarrowSCEV, 2005 SE->getZero(NarrowSCEV->getType())); 2006 for (User *U : NarrowDef->users()) { 2007 Instruction *NarrowUser = cast<Instruction>(U); 2008 2009 // Handle data flow merges and bizarre phi cycles. 2010 if (!Widened.insert(NarrowUser).second) 2011 continue; 2012 2013 bool NonNegativeUse = false; 2014 if (!NonNegativeDef) { 2015 // We might have a control-dependent range information for this context. 2016 if (auto RangeInfo = getPostIncRangeInfo(NarrowDef, NarrowUser)) 2017 NonNegativeUse = RangeInfo->getSignedMin().isNonNegative(); 2018 } 2019 2020 NarrowIVUsers.emplace_back(NarrowDef, NarrowUser, WideDef, 2021 NonNegativeDef || NonNegativeUse); 2022 } 2023 } 2024 2025 /// Process a single induction variable. First use the SCEVExpander to create a 2026 /// wide induction variable that evaluates to the same recurrence as the 2027 /// original narrow IV. Then use a worklist to forward traverse the narrow IV's 2028 /// def-use chain. After widenIVUse has processed all interesting IV users, the 2029 /// narrow IV will be isolated for removal by DeleteDeadPHIs. 2030 /// 2031 /// It would be simpler to delete uses as they are processed, but we must avoid 2032 /// invalidating SCEV expressions. 2033 PHINode *WidenIV::createWideIV(SCEVExpander &Rewriter) { 2034 // Is this phi an induction variable? 2035 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi)); 2036 if (!AddRec) 2037 return nullptr; 2038 2039 // Widen the induction variable expression. 2040 const SCEV *WideIVExpr = getExtendKind(OrigPhi) == ExtendKind::Sign 2041 ? SE->getSignExtendExpr(AddRec, WideType) 2042 : SE->getZeroExtendExpr(AddRec, WideType); 2043 2044 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType && 2045 "Expect the new IV expression to preserve its type"); 2046 2047 // Can the IV be extended outside the loop without overflow? 2048 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr); 2049 if (!AddRec || AddRec->getLoop() != L) 2050 return nullptr; 2051 2052 // An AddRec must have loop-invariant operands. Since this AddRec is 2053 // materialized by a loop header phi, the expression cannot have any post-loop 2054 // operands, so they must dominate the loop header. 2055 assert( 2056 SE->properlyDominates(AddRec->getStart(), L->getHeader()) && 2057 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader()) && 2058 "Loop header phi recurrence inputs do not dominate the loop"); 2059 2060 // Iterate over IV uses (including transitive ones) looking for IV increments 2061 // of the form 'add nsw %iv, <const>'. For each increment and each use of 2062 // the increment calculate control-dependent range information basing on 2063 // dominating conditions inside of the loop (e.g. a range check inside of the 2064 // loop). Calculated ranges are stored in PostIncRangeInfos map. 2065 // 2066 // Control-dependent range information is later used to prove that a narrow 2067 // definition is not negative (see pushNarrowIVUsers). It's difficult to do 2068 // this on demand because when pushNarrowIVUsers needs this information some 2069 // of the dominating conditions might be already widened. 2070 if (UsePostIncrementRanges) 2071 calculatePostIncRanges(OrigPhi); 2072 2073 // The rewriter provides a value for the desired IV expression. This may 2074 // either find an existing phi or materialize a new one. Either way, we 2075 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part 2076 // of the phi-SCC dominates the loop entry. 2077 Instruction *InsertPt = &*L->getHeader()->getFirstInsertionPt(); 2078 Value *ExpandInst = Rewriter.expandCodeFor(AddRec, WideType, InsertPt); 2079 // If the wide phi is not a phi node, for example a cast node, like bitcast, 2080 // inttoptr, ptrtoint, just skip for now. 2081 if (!(WidePhi = dyn_cast<PHINode>(ExpandInst))) { 2082 // if the cast node is an inserted instruction without any user, we should 2083 // remove it to make sure the pass don't touch the function as we can not 2084 // wide the phi. 2085 if (ExpandInst->hasNUses(0) && 2086 Rewriter.isInsertedInstruction(cast<Instruction>(ExpandInst))) 2087 DeadInsts.emplace_back(ExpandInst); 2088 return nullptr; 2089 } 2090 2091 // Remembering the WideIV increment generated by SCEVExpander allows 2092 // widenIVUse to reuse it when widening the narrow IV's increment. We don't 2093 // employ a general reuse mechanism because the call above is the only call to 2094 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses. 2095 if (BasicBlock *LatchBlock = L->getLoopLatch()) { 2096 WideInc = 2097 dyn_cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock)); 2098 if (WideInc) { 2099 WideIncExpr = SE->getSCEV(WideInc); 2100 // Propagate the debug location associated with the original loop 2101 // increment to the new (widened) increment. 2102 auto *OrigInc = 2103 cast<Instruction>(OrigPhi->getIncomingValueForBlock(LatchBlock)); 2104 2105 WideInc->setDebugLoc(OrigInc->getDebugLoc()); 2106 // We are replacing a narrow IV increment with a wider IV increment. If 2107 // the original (narrow) increment did not wrap, the wider increment one 2108 // should not wrap either. Set the flags to be the union of both wide 2109 // increment and original increment; this ensures we preserve flags SCEV 2110 // could infer for the wider increment. Limit this only to cases where 2111 // both increments directly increment the corresponding PHI nodes and have 2112 // the same opcode. It is not safe to re-use the flags from the original 2113 // increment, if it is more complex and SCEV expansion may have yielded a 2114 // more simplified wider increment. 2115 if (SCEVExpander::canReuseFlagsFromOriginalIVInc(OrigPhi, WidePhi, 2116 OrigInc, WideInc) && 2117 isa<OverflowingBinaryOperator>(OrigInc) && 2118 isa<OverflowingBinaryOperator>(WideInc)) { 2119 WideInc->setHasNoUnsignedWrap(WideInc->hasNoUnsignedWrap() || 2120 OrigInc->hasNoUnsignedWrap()); 2121 WideInc->setHasNoSignedWrap(WideInc->hasNoSignedWrap() || 2122 OrigInc->hasNoSignedWrap()); 2123 } 2124 } 2125 } 2126 2127 LLVM_DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n"); 2128 ++NumWidened; 2129 2130 // Traverse the def-use chain using a worklist starting at the original IV. 2131 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" ); 2132 2133 Widened.insert(OrigPhi); 2134 pushNarrowIVUsers(OrigPhi, WidePhi); 2135 2136 while (!NarrowIVUsers.empty()) { 2137 WidenIV::NarrowIVDefUse DU = NarrowIVUsers.pop_back_val(); 2138 2139 // Process a def-use edge. This may replace the use, so don't hold a 2140 // use_iterator across it. 2141 Instruction *WideUse = widenIVUse(DU, Rewriter, OrigPhi, WidePhi); 2142 2143 // Follow all def-use edges from the previous narrow use. 2144 if (WideUse) 2145 pushNarrowIVUsers(DU.NarrowUse, WideUse); 2146 2147 // widenIVUse may have removed the def-use edge. 2148 if (DU.NarrowDef->use_empty()) 2149 DeadInsts.emplace_back(DU.NarrowDef); 2150 } 2151 2152 // Attach any debug information to the new PHI. 2153 replaceAllDbgUsesWith(*OrigPhi, *WidePhi, *WidePhi, *DT); 2154 2155 return WidePhi; 2156 } 2157 2158 /// Calculates control-dependent range for the given def at the given context 2159 /// by looking at dominating conditions inside of the loop 2160 void WidenIV::calculatePostIncRange(Instruction *NarrowDef, 2161 Instruction *NarrowUser) { 2162 Value *NarrowDefLHS; 2163 const APInt *NarrowDefRHS; 2164 if (!match(NarrowDef, m_NSWAdd(m_Value(NarrowDefLHS), 2165 m_APInt(NarrowDefRHS))) || 2166 !NarrowDefRHS->isNonNegative()) 2167 return; 2168 2169 auto UpdateRangeFromCondition = [&] (Value *Condition, 2170 bool TrueDest) { 2171 CmpInst::Predicate Pred; 2172 Value *CmpRHS; 2173 if (!match(Condition, m_ICmp(Pred, m_Specific(NarrowDefLHS), 2174 m_Value(CmpRHS)))) 2175 return; 2176 2177 CmpInst::Predicate P = 2178 TrueDest ? Pred : CmpInst::getInversePredicate(Pred); 2179 2180 auto CmpRHSRange = SE->getSignedRange(SE->getSCEV(CmpRHS)); 2181 auto CmpConstrainedLHSRange = 2182 ConstantRange::makeAllowedICmpRegion(P, CmpRHSRange); 2183 auto NarrowDefRange = CmpConstrainedLHSRange.addWithNoWrap( 2184 *NarrowDefRHS, OverflowingBinaryOperator::NoSignedWrap); 2185 2186 updatePostIncRangeInfo(NarrowDef, NarrowUser, NarrowDefRange); 2187 }; 2188 2189 auto UpdateRangeFromGuards = [&](Instruction *Ctx) { 2190 if (!HasGuards) 2191 return; 2192 2193 for (Instruction &I : make_range(Ctx->getIterator().getReverse(), 2194 Ctx->getParent()->rend())) { 2195 Value *C = nullptr; 2196 if (match(&I, m_Intrinsic<Intrinsic::experimental_guard>(m_Value(C)))) 2197 UpdateRangeFromCondition(C, /*TrueDest=*/true); 2198 } 2199 }; 2200 2201 UpdateRangeFromGuards(NarrowUser); 2202 2203 BasicBlock *NarrowUserBB = NarrowUser->getParent(); 2204 // If NarrowUserBB is statically unreachable asking dominator queries may 2205 // yield surprising results. (e.g. the block may not have a dom tree node) 2206 if (!DT->isReachableFromEntry(NarrowUserBB)) 2207 return; 2208 2209 for (auto *DTB = (*DT)[NarrowUserBB]->getIDom(); 2210 L->contains(DTB->getBlock()); 2211 DTB = DTB->getIDom()) { 2212 auto *BB = DTB->getBlock(); 2213 auto *TI = BB->getTerminator(); 2214 UpdateRangeFromGuards(TI); 2215 2216 auto *BI = dyn_cast<BranchInst>(TI); 2217 if (!BI || !BI->isConditional()) 2218 continue; 2219 2220 auto *TrueSuccessor = BI->getSuccessor(0); 2221 auto *FalseSuccessor = BI->getSuccessor(1); 2222 2223 auto DominatesNarrowUser = [this, NarrowUser] (BasicBlockEdge BBE) { 2224 return BBE.isSingleEdge() && 2225 DT->dominates(BBE, NarrowUser->getParent()); 2226 }; 2227 2228 if (DominatesNarrowUser(BasicBlockEdge(BB, TrueSuccessor))) 2229 UpdateRangeFromCondition(BI->getCondition(), /*TrueDest=*/true); 2230 2231 if (DominatesNarrowUser(BasicBlockEdge(BB, FalseSuccessor))) 2232 UpdateRangeFromCondition(BI->getCondition(), /*TrueDest=*/false); 2233 } 2234 } 2235 2236 /// Calculates PostIncRangeInfos map for the given IV 2237 void WidenIV::calculatePostIncRanges(PHINode *OrigPhi) { 2238 SmallPtrSet<Instruction *, 16> Visited; 2239 SmallVector<Instruction *, 6> Worklist; 2240 Worklist.push_back(OrigPhi); 2241 Visited.insert(OrigPhi); 2242 2243 while (!Worklist.empty()) { 2244 Instruction *NarrowDef = Worklist.pop_back_val(); 2245 2246 for (Use &U : NarrowDef->uses()) { 2247 auto *NarrowUser = cast<Instruction>(U.getUser()); 2248 2249 // Don't go looking outside the current loop. 2250 auto *NarrowUserLoop = (*LI)[NarrowUser->getParent()]; 2251 if (!NarrowUserLoop || !L->contains(NarrowUserLoop)) 2252 continue; 2253 2254 if (!Visited.insert(NarrowUser).second) 2255 continue; 2256 2257 Worklist.push_back(NarrowUser); 2258 2259 calculatePostIncRange(NarrowDef, NarrowUser); 2260 } 2261 } 2262 } 2263 2264 PHINode *llvm::createWideIV(const WideIVInfo &WI, 2265 LoopInfo *LI, ScalarEvolution *SE, SCEVExpander &Rewriter, 2266 DominatorTree *DT, SmallVectorImpl<WeakTrackingVH> &DeadInsts, 2267 unsigned &NumElimExt, unsigned &NumWidened, 2268 bool HasGuards, bool UsePostIncrementRanges) { 2269 WidenIV Widener(WI, LI, SE, DT, DeadInsts, HasGuards, UsePostIncrementRanges); 2270 PHINode *WidePHI = Widener.createWideIV(Rewriter); 2271 NumElimExt = Widener.getNumElimExt(); 2272 NumWidened = Widener.getNumWidened(); 2273 return WidePHI; 2274 } 2275