1 //===-- LoopPredication.cpp - Guard based loop predication pass -----------===// 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 // The LoopPredication pass tries to convert loop variant range checks to loop 10 // invariant by widening checks across loop iterations. For example, it will 11 // convert 12 // 13 // for (i = 0; i < n; i++) { 14 // guard(i < len); 15 // ... 16 // } 17 // 18 // to 19 // 20 // for (i = 0; i < n; i++) { 21 // guard(n - 1 < len); 22 // ... 23 // } 24 // 25 // After this transformation the condition of the guard is loop invariant, so 26 // loop-unswitch can later unswitch the loop by this condition which basically 27 // predicates the loop by the widened condition: 28 // 29 // if (n - 1 < len) 30 // for (i = 0; i < n; i++) { 31 // ... 32 // } 33 // else 34 // deoptimize 35 // 36 // It's tempting to rely on SCEV here, but it has proven to be problematic. 37 // Generally the facts SCEV provides about the increment step of add 38 // recurrences are true if the backedge of the loop is taken, which implicitly 39 // assumes that the guard doesn't fail. Using these facts to optimize the 40 // guard results in a circular logic where the guard is optimized under the 41 // assumption that it never fails. 42 // 43 // For example, in the loop below the induction variable will be marked as nuw 44 // basing on the guard. Basing on nuw the guard predicate will be considered 45 // monotonic. Given a monotonic condition it's tempting to replace the induction 46 // variable in the condition with its value on the last iteration. But this 47 // transformation is not correct, e.g. e = 4, b = 5 breaks the loop. 48 // 49 // for (int i = b; i != e; i++) 50 // guard(i u< len) 51 // 52 // One of the ways to reason about this problem is to use an inductive proof 53 // approach. Given the loop: 54 // 55 // if (B(0)) { 56 // do { 57 // I = PHI(0, I.INC) 58 // I.INC = I + Step 59 // guard(G(I)); 60 // } while (B(I)); 61 // } 62 // 63 // where B(x) and G(x) are predicates that map integers to booleans, we want a 64 // loop invariant expression M such the following program has the same semantics 65 // as the above: 66 // 67 // if (B(0)) { 68 // do { 69 // I = PHI(0, I.INC) 70 // I.INC = I + Step 71 // guard(G(0) && M); 72 // } while (B(I)); 73 // } 74 // 75 // One solution for M is M = forall X . (G(X) && B(X)) => G(X + Step) 76 // 77 // Informal proof that the transformation above is correct: 78 // 79 // By the definition of guards we can rewrite the guard condition to: 80 // G(I) && G(0) && M 81 // 82 // Let's prove that for each iteration of the loop: 83 // G(0) && M => G(I) 84 // And the condition above can be simplified to G(Start) && M. 85 // 86 // Induction base. 87 // G(0) && M => G(0) 88 // 89 // Induction step. Assuming G(0) && M => G(I) on the subsequent 90 // iteration: 91 // 92 // B(I) is true because it's the backedge condition. 93 // G(I) is true because the backedge is guarded by this condition. 94 // 95 // So M = forall X . (G(X) && B(X)) => G(X + Step) implies G(I + Step). 96 // 97 // Note that we can use anything stronger than M, i.e. any condition which 98 // implies M. 99 // 100 // When S = 1 (i.e. forward iterating loop), the transformation is supported 101 // when: 102 // * The loop has a single latch with the condition of the form: 103 // B(X) = latchStart + X <pred> latchLimit, 104 // where <pred> is u<, u<=, s<, or s<=. 105 // * The guard condition is of the form 106 // G(X) = guardStart + X u< guardLimit 107 // 108 // For the ult latch comparison case M is: 109 // forall X . guardStart + X u< guardLimit && latchStart + X <u latchLimit => 110 // guardStart + X + 1 u< guardLimit 111 // 112 // The only way the antecedent can be true and the consequent can be false is 113 // if 114 // X == guardLimit - 1 - guardStart 115 // (and guardLimit is non-zero, but we won't use this latter fact). 116 // If X == guardLimit - 1 - guardStart then the second half of the antecedent is 117 // latchStart + guardLimit - 1 - guardStart u< latchLimit 118 // and its negation is 119 // latchStart + guardLimit - 1 - guardStart u>= latchLimit 120 // 121 // In other words, if 122 // latchLimit u<= latchStart + guardLimit - 1 - guardStart 123 // then: 124 // (the ranges below are written in ConstantRange notation, where [A, B) is the 125 // set for (I = A; I != B; I++ /*maywrap*/) yield(I);) 126 // 127 // forall X . guardStart + X u< guardLimit && 128 // latchStart + X u< latchLimit => 129 // guardStart + X + 1 u< guardLimit 130 // == forall X . guardStart + X u< guardLimit && 131 // latchStart + X u< latchStart + guardLimit - 1 - guardStart => 132 // guardStart + X + 1 u< guardLimit 133 // == forall X . (guardStart + X) in [0, guardLimit) && 134 // (latchStart + X) in [0, latchStart + guardLimit - 1 - guardStart) => 135 // (guardStart + X + 1) in [0, guardLimit) 136 // == forall X . X in [-guardStart, guardLimit - guardStart) && 137 // X in [-latchStart, guardLimit - 1 - guardStart) => 138 // X in [-guardStart - 1, guardLimit - guardStart - 1) 139 // == true 140 // 141 // So the widened condition is: 142 // guardStart u< guardLimit && 143 // latchStart + guardLimit - 1 - guardStart u>= latchLimit 144 // Similarly for ule condition the widened condition is: 145 // guardStart u< guardLimit && 146 // latchStart + guardLimit - 1 - guardStart u> latchLimit 147 // For slt condition the widened condition is: 148 // guardStart u< guardLimit && 149 // latchStart + guardLimit - 1 - guardStart s>= latchLimit 150 // For sle condition the widened condition is: 151 // guardStart u< guardLimit && 152 // latchStart + guardLimit - 1 - guardStart s> latchLimit 153 // 154 // When S = -1 (i.e. reverse iterating loop), the transformation is supported 155 // when: 156 // * The loop has a single latch with the condition of the form: 157 // B(X) = X <pred> latchLimit, where <pred> is u>, u>=, s>, or s>=. 158 // * The guard condition is of the form 159 // G(X) = X - 1 u< guardLimit 160 // 161 // For the ugt latch comparison case M is: 162 // forall X. X-1 u< guardLimit and X u> latchLimit => X-2 u< guardLimit 163 // 164 // The only way the antecedent can be true and the consequent can be false is if 165 // X == 1. 166 // If X == 1 then the second half of the antecedent is 167 // 1 u> latchLimit, and its negation is latchLimit u>= 1. 168 // 169 // So the widened condition is: 170 // guardStart u< guardLimit && latchLimit u>= 1. 171 // Similarly for sgt condition the widened condition is: 172 // guardStart u< guardLimit && latchLimit s>= 1. 173 // For uge condition the widened condition is: 174 // guardStart u< guardLimit && latchLimit u> 1. 175 // For sge condition the widened condition is: 176 // guardStart u< guardLimit && latchLimit s> 1. 177 //===----------------------------------------------------------------------===// 178 179 #include "llvm/Transforms/Scalar/LoopPredication.h" 180 #include "llvm/ADT/Statistic.h" 181 #include "llvm/Analysis/AliasAnalysis.h" 182 #include "llvm/Analysis/BranchProbabilityInfo.h" 183 #include "llvm/Analysis/GuardUtils.h" 184 #include "llvm/Analysis/LoopInfo.h" 185 #include "llvm/Analysis/LoopPass.h" 186 #include "llvm/Analysis/ScalarEvolution.h" 187 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 188 #include "llvm/IR/Function.h" 189 #include "llvm/IR/GlobalValue.h" 190 #include "llvm/IR/IntrinsicInst.h" 191 #include "llvm/IR/Module.h" 192 #include "llvm/IR/PatternMatch.h" 193 #include "llvm/InitializePasses.h" 194 #include "llvm/Pass.h" 195 #include "llvm/Support/CommandLine.h" 196 #include "llvm/Support/Debug.h" 197 #include "llvm/Transforms/Scalar.h" 198 #include "llvm/Transforms/Utils/GuardUtils.h" 199 #include "llvm/Transforms/Utils/Local.h" 200 #include "llvm/Transforms/Utils/LoopUtils.h" 201 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" 202 203 #define DEBUG_TYPE "loop-predication" 204 205 STATISTIC(TotalConsidered, "Number of guards considered"); 206 STATISTIC(TotalWidened, "Number of checks widened"); 207 208 using namespace llvm; 209 210 static cl::opt<bool> EnableIVTruncation("loop-predication-enable-iv-truncation", 211 cl::Hidden, cl::init(true)); 212 213 static cl::opt<bool> EnableCountDownLoop("loop-predication-enable-count-down-loop", 214 cl::Hidden, cl::init(true)); 215 216 static cl::opt<bool> 217 SkipProfitabilityChecks("loop-predication-skip-profitability-checks", 218 cl::Hidden, cl::init(false)); 219 220 // This is the scale factor for the latch probability. We use this during 221 // profitability analysis to find other exiting blocks that have a much higher 222 // probability of exiting the loop instead of loop exiting via latch. 223 // This value should be greater than 1 for a sane profitability check. 224 static cl::opt<float> LatchExitProbabilityScale( 225 "loop-predication-latch-probability-scale", cl::Hidden, cl::init(2.0), 226 cl::desc("scale factor for the latch probability. Value should be greater " 227 "than 1. Lower values are ignored")); 228 229 static cl::opt<bool> PredicateWidenableBranchGuards( 230 "loop-predication-predicate-widenable-branches-to-deopt", cl::Hidden, 231 cl::desc("Whether or not we should predicate guards " 232 "expressed as widenable branches to deoptimize blocks"), 233 cl::init(true)); 234 235 namespace { 236 /// Represents an induction variable check: 237 /// icmp Pred, <induction variable>, <loop invariant limit> 238 struct LoopICmp { 239 ICmpInst::Predicate Pred; 240 const SCEVAddRecExpr *IV; 241 const SCEV *Limit; 242 LoopICmp(ICmpInst::Predicate Pred, const SCEVAddRecExpr *IV, 243 const SCEV *Limit) 244 : Pred(Pred), IV(IV), Limit(Limit) {} 245 LoopICmp() {} 246 void dump() { 247 dbgs() << "LoopICmp Pred = " << Pred << ", IV = " << *IV 248 << ", Limit = " << *Limit << "\n"; 249 } 250 }; 251 252 class LoopPredication { 253 AliasAnalysis *AA; 254 DominatorTree *DT; 255 ScalarEvolution *SE; 256 LoopInfo *LI; 257 BranchProbabilityInfo *BPI; 258 259 Loop *L; 260 const DataLayout *DL; 261 BasicBlock *Preheader; 262 LoopICmp LatchCheck; 263 264 bool isSupportedStep(const SCEV* Step); 265 Optional<LoopICmp> parseLoopICmp(ICmpInst *ICI); 266 Optional<LoopICmp> parseLoopLatchICmp(); 267 268 /// Return an insertion point suitable for inserting a safe to speculate 269 /// instruction whose only user will be 'User' which has operands 'Ops'. A 270 /// trivial result would be the at the User itself, but we try to return a 271 /// loop invariant location if possible. 272 Instruction *findInsertPt(Instruction *User, ArrayRef<Value*> Ops); 273 /// Same as above, *except* that this uses the SCEV definition of invariant 274 /// which is that an expression *can be made* invariant via SCEVExpander. 275 /// Thus, this version is only suitable for finding an insert point to be be 276 /// passed to SCEVExpander! 277 Instruction *findInsertPt(Instruction *User, ArrayRef<const SCEV*> Ops); 278 279 /// Return true if the value is known to produce a single fixed value across 280 /// all iterations on which it executes. Note that this does not imply 281 /// speculation safety. That must be established separately. 282 bool isLoopInvariantValue(const SCEV* S); 283 284 Value *expandCheck(SCEVExpander &Expander, Instruction *Guard, 285 ICmpInst::Predicate Pred, const SCEV *LHS, 286 const SCEV *RHS); 287 288 Optional<Value *> widenICmpRangeCheck(ICmpInst *ICI, SCEVExpander &Expander, 289 Instruction *Guard); 290 Optional<Value *> widenICmpRangeCheckIncrementingLoop(LoopICmp LatchCheck, 291 LoopICmp RangeCheck, 292 SCEVExpander &Expander, 293 Instruction *Guard); 294 Optional<Value *> widenICmpRangeCheckDecrementingLoop(LoopICmp LatchCheck, 295 LoopICmp RangeCheck, 296 SCEVExpander &Expander, 297 Instruction *Guard); 298 unsigned collectChecks(SmallVectorImpl<Value *> &Checks, Value *Condition, 299 SCEVExpander &Expander, Instruction *Guard); 300 bool widenGuardConditions(IntrinsicInst *II, SCEVExpander &Expander); 301 bool widenWidenableBranchGuardConditions(BranchInst *Guard, SCEVExpander &Expander); 302 // If the loop always exits through another block in the loop, we should not 303 // predicate based on the latch check. For example, the latch check can be a 304 // very coarse grained check and there can be more fine grained exit checks 305 // within the loop. We identify such unprofitable loops through BPI. 306 bool isLoopProfitableToPredicate(); 307 308 bool predicateLoopExits(Loop *L, SCEVExpander &Rewriter); 309 310 public: 311 LoopPredication(AliasAnalysis *AA, DominatorTree *DT, 312 ScalarEvolution *SE, LoopInfo *LI, 313 BranchProbabilityInfo *BPI) 314 : AA(AA), DT(DT), SE(SE), LI(LI), BPI(BPI) {}; 315 bool runOnLoop(Loop *L); 316 }; 317 318 class LoopPredicationLegacyPass : public LoopPass { 319 public: 320 static char ID; 321 LoopPredicationLegacyPass() : LoopPass(ID) { 322 initializeLoopPredicationLegacyPassPass(*PassRegistry::getPassRegistry()); 323 } 324 325 void getAnalysisUsage(AnalysisUsage &AU) const override { 326 AU.addRequired<BranchProbabilityInfoWrapperPass>(); 327 getLoopAnalysisUsage(AU); 328 } 329 330 bool runOnLoop(Loop *L, LPPassManager &LPM) override { 331 if (skipLoop(L)) 332 return false; 333 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 334 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 335 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 336 BranchProbabilityInfo &BPI = 337 getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI(); 338 auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 339 LoopPredication LP(AA, DT, SE, LI, &BPI); 340 return LP.runOnLoop(L); 341 } 342 }; 343 344 char LoopPredicationLegacyPass::ID = 0; 345 } // end namespace 346 347 INITIALIZE_PASS_BEGIN(LoopPredicationLegacyPass, "loop-predication", 348 "Loop predication", false, false) 349 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass) 350 INITIALIZE_PASS_DEPENDENCY(LoopPass) 351 INITIALIZE_PASS_END(LoopPredicationLegacyPass, "loop-predication", 352 "Loop predication", false, false) 353 354 Pass *llvm::createLoopPredicationPass() { 355 return new LoopPredicationLegacyPass(); 356 } 357 358 PreservedAnalyses LoopPredicationPass::run(Loop &L, LoopAnalysisManager &AM, 359 LoopStandardAnalysisResults &AR, 360 LPMUpdater &U) { 361 Function *F = L.getHeader()->getParent(); 362 // For the new PM, we also can't use BranchProbabilityInfo as an analysis 363 // pass. Function analyses need to be preserved across loop transformations 364 // but BPI is not preserved, hence a newly built one is needed. 365 BranchProbabilityInfo BPI(*F, AR.LI, &AR.TLI, &AR.DT, nullptr); 366 LoopPredication LP(&AR.AA, &AR.DT, &AR.SE, &AR.LI, &BPI); 367 if (!LP.runOnLoop(&L)) 368 return PreservedAnalyses::all(); 369 370 return getLoopPassPreservedAnalyses(); 371 } 372 373 Optional<LoopICmp> 374 LoopPredication::parseLoopICmp(ICmpInst *ICI) { 375 auto Pred = ICI->getPredicate(); 376 auto *LHS = ICI->getOperand(0); 377 auto *RHS = ICI->getOperand(1); 378 379 const SCEV *LHSS = SE->getSCEV(LHS); 380 if (isa<SCEVCouldNotCompute>(LHSS)) 381 return None; 382 const SCEV *RHSS = SE->getSCEV(RHS); 383 if (isa<SCEVCouldNotCompute>(RHSS)) 384 return None; 385 386 // Canonicalize RHS to be loop invariant bound, LHS - a loop computable IV 387 if (SE->isLoopInvariant(LHSS, L)) { 388 std::swap(LHS, RHS); 389 std::swap(LHSS, RHSS); 390 Pred = ICmpInst::getSwappedPredicate(Pred); 391 } 392 393 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHSS); 394 if (!AR || AR->getLoop() != L) 395 return None; 396 397 return LoopICmp(Pred, AR, RHSS); 398 } 399 400 Value *LoopPredication::expandCheck(SCEVExpander &Expander, 401 Instruction *Guard, 402 ICmpInst::Predicate Pred, const SCEV *LHS, 403 const SCEV *RHS) { 404 Type *Ty = LHS->getType(); 405 assert(Ty == RHS->getType() && "expandCheck operands have different types?"); 406 407 if (SE->isLoopInvariant(LHS, L) && SE->isLoopInvariant(RHS, L)) { 408 IRBuilder<> Builder(Guard); 409 if (SE->isLoopEntryGuardedByCond(L, Pred, LHS, RHS)) 410 return Builder.getTrue(); 411 if (SE->isLoopEntryGuardedByCond(L, ICmpInst::getInversePredicate(Pred), 412 LHS, RHS)) 413 return Builder.getFalse(); 414 } 415 416 Value *LHSV = Expander.expandCodeFor(LHS, Ty, findInsertPt(Guard, {LHS})); 417 Value *RHSV = Expander.expandCodeFor(RHS, Ty, findInsertPt(Guard, {RHS})); 418 IRBuilder<> Builder(findInsertPt(Guard, {LHSV, RHSV})); 419 return Builder.CreateICmp(Pred, LHSV, RHSV); 420 } 421 422 423 // Returns true if its safe to truncate the IV to RangeCheckType. 424 // When the IV type is wider than the range operand type, we can still do loop 425 // predication, by generating SCEVs for the range and latch that are of the 426 // same type. We achieve this by generating a SCEV truncate expression for the 427 // latch IV. This is done iff truncation of the IV is a safe operation, 428 // without loss of information. 429 // Another way to achieve this is by generating a wider type SCEV for the 430 // range check operand, however, this needs a more involved check that 431 // operands do not overflow. This can lead to loss of information when the 432 // range operand is of the form: add i32 %offset, %iv. We need to prove that 433 // sext(x + y) is same as sext(x) + sext(y). 434 // This function returns true if we can safely represent the IV type in 435 // the RangeCheckType without loss of information. 436 static bool isSafeToTruncateWideIVType(const DataLayout &DL, 437 ScalarEvolution &SE, 438 const LoopICmp LatchCheck, 439 Type *RangeCheckType) { 440 if (!EnableIVTruncation) 441 return false; 442 assert(DL.getTypeSizeInBits(LatchCheck.IV->getType()).getFixedSize() > 443 DL.getTypeSizeInBits(RangeCheckType).getFixedSize() && 444 "Expected latch check IV type to be larger than range check operand " 445 "type!"); 446 // The start and end values of the IV should be known. This is to guarantee 447 // that truncating the wide type will not lose information. 448 auto *Limit = dyn_cast<SCEVConstant>(LatchCheck.Limit); 449 auto *Start = dyn_cast<SCEVConstant>(LatchCheck.IV->getStart()); 450 if (!Limit || !Start) 451 return false; 452 // This check makes sure that the IV does not change sign during loop 453 // iterations. Consider latchType = i64, LatchStart = 5, Pred = ICMP_SGE, 454 // LatchEnd = 2, rangeCheckType = i32. If it's not a monotonic predicate, the 455 // IV wraps around, and the truncation of the IV would lose the range of 456 // iterations between 2^32 and 2^64. 457 if (!SE.getMonotonicPredicateType(LatchCheck.IV, LatchCheck.Pred)) 458 return false; 459 // The active bits should be less than the bits in the RangeCheckType. This 460 // guarantees that truncating the latch check to RangeCheckType is a safe 461 // operation. 462 auto RangeCheckTypeBitSize = 463 DL.getTypeSizeInBits(RangeCheckType).getFixedSize(); 464 return Start->getAPInt().getActiveBits() < RangeCheckTypeBitSize && 465 Limit->getAPInt().getActiveBits() < RangeCheckTypeBitSize; 466 } 467 468 469 // Return an LoopICmp describing a latch check equivlent to LatchCheck but with 470 // the requested type if safe to do so. May involve the use of a new IV. 471 static Optional<LoopICmp> generateLoopLatchCheck(const DataLayout &DL, 472 ScalarEvolution &SE, 473 const LoopICmp LatchCheck, 474 Type *RangeCheckType) { 475 476 auto *LatchType = LatchCheck.IV->getType(); 477 if (RangeCheckType == LatchType) 478 return LatchCheck; 479 // For now, bail out if latch type is narrower than range type. 480 if (DL.getTypeSizeInBits(LatchType).getFixedSize() < 481 DL.getTypeSizeInBits(RangeCheckType).getFixedSize()) 482 return None; 483 if (!isSafeToTruncateWideIVType(DL, SE, LatchCheck, RangeCheckType)) 484 return None; 485 // We can now safely identify the truncated version of the IV and limit for 486 // RangeCheckType. 487 LoopICmp NewLatchCheck; 488 NewLatchCheck.Pred = LatchCheck.Pred; 489 NewLatchCheck.IV = dyn_cast<SCEVAddRecExpr>( 490 SE.getTruncateExpr(LatchCheck.IV, RangeCheckType)); 491 if (!NewLatchCheck.IV) 492 return None; 493 NewLatchCheck.Limit = SE.getTruncateExpr(LatchCheck.Limit, RangeCheckType); 494 LLVM_DEBUG(dbgs() << "IV of type: " << *LatchType 495 << "can be represented as range check type:" 496 << *RangeCheckType << "\n"); 497 LLVM_DEBUG(dbgs() << "LatchCheck.IV: " << *NewLatchCheck.IV << "\n"); 498 LLVM_DEBUG(dbgs() << "LatchCheck.Limit: " << *NewLatchCheck.Limit << "\n"); 499 return NewLatchCheck; 500 } 501 502 bool LoopPredication::isSupportedStep(const SCEV* Step) { 503 return Step->isOne() || (Step->isAllOnesValue() && EnableCountDownLoop); 504 } 505 506 Instruction *LoopPredication::findInsertPt(Instruction *Use, 507 ArrayRef<Value*> Ops) { 508 for (Value *Op : Ops) 509 if (!L->isLoopInvariant(Op)) 510 return Use; 511 return Preheader->getTerminator(); 512 } 513 514 Instruction *LoopPredication::findInsertPt(Instruction *Use, 515 ArrayRef<const SCEV*> Ops) { 516 // Subtlety: SCEV considers things to be invariant if the value produced is 517 // the same across iterations. This is not the same as being able to 518 // evaluate outside the loop, which is what we actually need here. 519 for (const SCEV *Op : Ops) 520 if (!SE->isLoopInvariant(Op, L) || 521 !isSafeToExpandAt(Op, Preheader->getTerminator(), *SE)) 522 return Use; 523 return Preheader->getTerminator(); 524 } 525 526 bool LoopPredication::isLoopInvariantValue(const SCEV* S) { 527 // Handling expressions which produce invariant results, but *haven't* yet 528 // been removed from the loop serves two important purposes. 529 // 1) Most importantly, it resolves a pass ordering cycle which would 530 // otherwise need us to iteration licm, loop-predication, and either 531 // loop-unswitch or loop-peeling to make progress on examples with lots of 532 // predicable range checks in a row. (Since, in the general case, we can't 533 // hoist the length checks until the dominating checks have been discharged 534 // as we can't prove doing so is safe.) 535 // 2) As a nice side effect, this exposes the value of peeling or unswitching 536 // much more obviously in the IR. Otherwise, the cost modeling for other 537 // transforms would end up needing to duplicate all of this logic to model a 538 // check which becomes predictable based on a modeled peel or unswitch. 539 // 540 // The cost of doing so in the worst case is an extra fill from the stack in 541 // the loop to materialize the loop invariant test value instead of checking 542 // against the original IV which is presumable in a register inside the loop. 543 // Such cases are presumably rare, and hint at missing oppurtunities for 544 // other passes. 545 546 if (SE->isLoopInvariant(S, L)) 547 // Note: This the SCEV variant, so the original Value* may be within the 548 // loop even though SCEV has proven it is loop invariant. 549 return true; 550 551 // Handle a particular important case which SCEV doesn't yet know about which 552 // shows up in range checks on arrays with immutable lengths. 553 // TODO: This should be sunk inside SCEV. 554 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) 555 if (const auto *LI = dyn_cast<LoadInst>(U->getValue())) 556 if (LI->isUnordered() && L->hasLoopInvariantOperands(LI)) 557 if (AA->pointsToConstantMemory(LI->getOperand(0)) || 558 LI->hasMetadata(LLVMContext::MD_invariant_load)) 559 return true; 560 return false; 561 } 562 563 Optional<Value *> LoopPredication::widenICmpRangeCheckIncrementingLoop( 564 LoopICmp LatchCheck, LoopICmp RangeCheck, 565 SCEVExpander &Expander, Instruction *Guard) { 566 auto *Ty = RangeCheck.IV->getType(); 567 // Generate the widened condition for the forward loop: 568 // guardStart u< guardLimit && 569 // latchLimit <pred> guardLimit - 1 - guardStart + latchStart 570 // where <pred> depends on the latch condition predicate. See the file 571 // header comment for the reasoning. 572 // guardLimit - guardStart + latchStart - 1 573 const SCEV *GuardStart = RangeCheck.IV->getStart(); 574 const SCEV *GuardLimit = RangeCheck.Limit; 575 const SCEV *LatchStart = LatchCheck.IV->getStart(); 576 const SCEV *LatchLimit = LatchCheck.Limit; 577 // Subtlety: We need all the values to be *invariant* across all iterations, 578 // but we only need to check expansion safety for those which *aren't* 579 // already guaranteed to dominate the guard. 580 if (!isLoopInvariantValue(GuardStart) || 581 !isLoopInvariantValue(GuardLimit) || 582 !isLoopInvariantValue(LatchStart) || 583 !isLoopInvariantValue(LatchLimit)) { 584 LLVM_DEBUG(dbgs() << "Can't expand limit check!\n"); 585 return None; 586 } 587 if (!isSafeToExpandAt(LatchStart, Guard, *SE) || 588 !isSafeToExpandAt(LatchLimit, Guard, *SE)) { 589 LLVM_DEBUG(dbgs() << "Can't expand limit check!\n"); 590 return None; 591 } 592 593 // guardLimit - guardStart + latchStart - 1 594 const SCEV *RHS = 595 SE->getAddExpr(SE->getMinusSCEV(GuardLimit, GuardStart), 596 SE->getMinusSCEV(LatchStart, SE->getOne(Ty))); 597 auto LimitCheckPred = 598 ICmpInst::getFlippedStrictnessPredicate(LatchCheck.Pred); 599 600 LLVM_DEBUG(dbgs() << "LHS: " << *LatchLimit << "\n"); 601 LLVM_DEBUG(dbgs() << "RHS: " << *RHS << "\n"); 602 LLVM_DEBUG(dbgs() << "Pred: " << LimitCheckPred << "\n"); 603 604 auto *LimitCheck = 605 expandCheck(Expander, Guard, LimitCheckPred, LatchLimit, RHS); 606 auto *FirstIterationCheck = expandCheck(Expander, Guard, RangeCheck.Pred, 607 GuardStart, GuardLimit); 608 IRBuilder<> Builder(findInsertPt(Guard, {FirstIterationCheck, LimitCheck})); 609 return Builder.CreateAnd(FirstIterationCheck, LimitCheck); 610 } 611 612 Optional<Value *> LoopPredication::widenICmpRangeCheckDecrementingLoop( 613 LoopICmp LatchCheck, LoopICmp RangeCheck, 614 SCEVExpander &Expander, Instruction *Guard) { 615 auto *Ty = RangeCheck.IV->getType(); 616 const SCEV *GuardStart = RangeCheck.IV->getStart(); 617 const SCEV *GuardLimit = RangeCheck.Limit; 618 const SCEV *LatchStart = LatchCheck.IV->getStart(); 619 const SCEV *LatchLimit = LatchCheck.Limit; 620 // Subtlety: We need all the values to be *invariant* across all iterations, 621 // but we only need to check expansion safety for those which *aren't* 622 // already guaranteed to dominate the guard. 623 if (!isLoopInvariantValue(GuardStart) || 624 !isLoopInvariantValue(GuardLimit) || 625 !isLoopInvariantValue(LatchStart) || 626 !isLoopInvariantValue(LatchLimit)) { 627 LLVM_DEBUG(dbgs() << "Can't expand limit check!\n"); 628 return None; 629 } 630 if (!isSafeToExpandAt(LatchStart, Guard, *SE) || 631 !isSafeToExpandAt(LatchLimit, Guard, *SE)) { 632 LLVM_DEBUG(dbgs() << "Can't expand limit check!\n"); 633 return None; 634 } 635 // The decrement of the latch check IV should be the same as the 636 // rangeCheckIV. 637 auto *PostDecLatchCheckIV = LatchCheck.IV->getPostIncExpr(*SE); 638 if (RangeCheck.IV != PostDecLatchCheckIV) { 639 LLVM_DEBUG(dbgs() << "Not the same. PostDecLatchCheckIV: " 640 << *PostDecLatchCheckIV 641 << " and RangeCheckIV: " << *RangeCheck.IV << "\n"); 642 return None; 643 } 644 645 // Generate the widened condition for CountDownLoop: 646 // guardStart u< guardLimit && 647 // latchLimit <pred> 1. 648 // See the header comment for reasoning of the checks. 649 auto LimitCheckPred = 650 ICmpInst::getFlippedStrictnessPredicate(LatchCheck.Pred); 651 auto *FirstIterationCheck = expandCheck(Expander, Guard, 652 ICmpInst::ICMP_ULT, 653 GuardStart, GuardLimit); 654 auto *LimitCheck = expandCheck(Expander, Guard, LimitCheckPred, LatchLimit, 655 SE->getOne(Ty)); 656 IRBuilder<> Builder(findInsertPt(Guard, {FirstIterationCheck, LimitCheck})); 657 return Builder.CreateAnd(FirstIterationCheck, LimitCheck); 658 } 659 660 static void normalizePredicate(ScalarEvolution *SE, Loop *L, 661 LoopICmp& RC) { 662 // LFTR canonicalizes checks to the ICMP_NE/EQ form; normalize back to the 663 // ULT/UGE form for ease of handling by our caller. 664 if (ICmpInst::isEquality(RC.Pred) && 665 RC.IV->getStepRecurrence(*SE)->isOne() && 666 SE->isKnownPredicate(ICmpInst::ICMP_ULE, RC.IV->getStart(), RC.Limit)) 667 RC.Pred = RC.Pred == ICmpInst::ICMP_NE ? 668 ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE; 669 } 670 671 672 /// If ICI can be widened to a loop invariant condition emits the loop 673 /// invariant condition in the loop preheader and return it, otherwise 674 /// returns None. 675 Optional<Value *> LoopPredication::widenICmpRangeCheck(ICmpInst *ICI, 676 SCEVExpander &Expander, 677 Instruction *Guard) { 678 LLVM_DEBUG(dbgs() << "Analyzing ICmpInst condition:\n"); 679 LLVM_DEBUG(ICI->dump()); 680 681 // parseLoopStructure guarantees that the latch condition is: 682 // ++i <pred> latchLimit, where <pred> is u<, u<=, s<, or s<=. 683 // We are looking for the range checks of the form: 684 // i u< guardLimit 685 auto RangeCheck = parseLoopICmp(ICI); 686 if (!RangeCheck) { 687 LLVM_DEBUG(dbgs() << "Failed to parse the loop latch condition!\n"); 688 return None; 689 } 690 LLVM_DEBUG(dbgs() << "Guard check:\n"); 691 LLVM_DEBUG(RangeCheck->dump()); 692 if (RangeCheck->Pred != ICmpInst::ICMP_ULT) { 693 LLVM_DEBUG(dbgs() << "Unsupported range check predicate(" 694 << RangeCheck->Pred << ")!\n"); 695 return None; 696 } 697 auto *RangeCheckIV = RangeCheck->IV; 698 if (!RangeCheckIV->isAffine()) { 699 LLVM_DEBUG(dbgs() << "Range check IV is not affine!\n"); 700 return None; 701 } 702 auto *Step = RangeCheckIV->getStepRecurrence(*SE); 703 // We cannot just compare with latch IV step because the latch and range IVs 704 // may have different types. 705 if (!isSupportedStep(Step)) { 706 LLVM_DEBUG(dbgs() << "Range check and latch have IVs different steps!\n"); 707 return None; 708 } 709 auto *Ty = RangeCheckIV->getType(); 710 auto CurrLatchCheckOpt = generateLoopLatchCheck(*DL, *SE, LatchCheck, Ty); 711 if (!CurrLatchCheckOpt) { 712 LLVM_DEBUG(dbgs() << "Failed to generate a loop latch check " 713 "corresponding to range type: " 714 << *Ty << "\n"); 715 return None; 716 } 717 718 LoopICmp CurrLatchCheck = *CurrLatchCheckOpt; 719 // At this point, the range and latch step should have the same type, but need 720 // not have the same value (we support both 1 and -1 steps). 721 assert(Step->getType() == 722 CurrLatchCheck.IV->getStepRecurrence(*SE)->getType() && 723 "Range and latch steps should be of same type!"); 724 if (Step != CurrLatchCheck.IV->getStepRecurrence(*SE)) { 725 LLVM_DEBUG(dbgs() << "Range and latch have different step values!\n"); 726 return None; 727 } 728 729 if (Step->isOne()) 730 return widenICmpRangeCheckIncrementingLoop(CurrLatchCheck, *RangeCheck, 731 Expander, Guard); 732 else { 733 assert(Step->isAllOnesValue() && "Step should be -1!"); 734 return widenICmpRangeCheckDecrementingLoop(CurrLatchCheck, *RangeCheck, 735 Expander, Guard); 736 } 737 } 738 739 unsigned LoopPredication::collectChecks(SmallVectorImpl<Value *> &Checks, 740 Value *Condition, 741 SCEVExpander &Expander, 742 Instruction *Guard) { 743 unsigned NumWidened = 0; 744 // The guard condition is expected to be in form of: 745 // cond1 && cond2 && cond3 ... 746 // Iterate over subconditions looking for icmp conditions which can be 747 // widened across loop iterations. Widening these conditions remember the 748 // resulting list of subconditions in Checks vector. 749 SmallVector<Value *, 4> Worklist(1, Condition); 750 SmallPtrSet<Value *, 4> Visited; 751 Value *WideableCond = nullptr; 752 do { 753 Value *Condition = Worklist.pop_back_val(); 754 if (!Visited.insert(Condition).second) 755 continue; 756 757 Value *LHS, *RHS; 758 using namespace llvm::PatternMatch; 759 if (match(Condition, m_And(m_Value(LHS), m_Value(RHS)))) { 760 Worklist.push_back(LHS); 761 Worklist.push_back(RHS); 762 continue; 763 } 764 765 if (match(Condition, 766 m_Intrinsic<Intrinsic::experimental_widenable_condition>())) { 767 // Pick any, we don't care which 768 WideableCond = Condition; 769 continue; 770 } 771 772 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Condition)) { 773 if (auto NewRangeCheck = widenICmpRangeCheck(ICI, Expander, 774 Guard)) { 775 Checks.push_back(NewRangeCheck.getValue()); 776 NumWidened++; 777 continue; 778 } 779 } 780 781 // Save the condition as is if we can't widen it 782 Checks.push_back(Condition); 783 } while (!Worklist.empty()); 784 // At the moment, our matching logic for wideable conditions implicitly 785 // assumes we preserve the form: (br (and Cond, WC())). FIXME 786 // Note that if there were multiple calls to wideable condition in the 787 // traversal, we only need to keep one, and which one is arbitrary. 788 if (WideableCond) 789 Checks.push_back(WideableCond); 790 return NumWidened; 791 } 792 793 bool LoopPredication::widenGuardConditions(IntrinsicInst *Guard, 794 SCEVExpander &Expander) { 795 LLVM_DEBUG(dbgs() << "Processing guard:\n"); 796 LLVM_DEBUG(Guard->dump()); 797 798 TotalConsidered++; 799 SmallVector<Value *, 4> Checks; 800 unsigned NumWidened = collectChecks(Checks, Guard->getOperand(0), Expander, 801 Guard); 802 if (NumWidened == 0) 803 return false; 804 805 TotalWidened += NumWidened; 806 807 // Emit the new guard condition 808 IRBuilder<> Builder(findInsertPt(Guard, Checks)); 809 Value *AllChecks = Builder.CreateAnd(Checks); 810 auto *OldCond = Guard->getOperand(0); 811 Guard->setOperand(0, AllChecks); 812 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 813 814 LLVM_DEBUG(dbgs() << "Widened checks = " << NumWidened << "\n"); 815 return true; 816 } 817 818 bool LoopPredication::widenWidenableBranchGuardConditions( 819 BranchInst *BI, SCEVExpander &Expander) { 820 assert(isGuardAsWidenableBranch(BI) && "Must be!"); 821 LLVM_DEBUG(dbgs() << "Processing guard:\n"); 822 LLVM_DEBUG(BI->dump()); 823 824 TotalConsidered++; 825 SmallVector<Value *, 4> Checks; 826 unsigned NumWidened = collectChecks(Checks, BI->getCondition(), 827 Expander, BI); 828 if (NumWidened == 0) 829 return false; 830 831 TotalWidened += NumWidened; 832 833 // Emit the new guard condition 834 IRBuilder<> Builder(findInsertPt(BI, Checks)); 835 Value *AllChecks = Builder.CreateAnd(Checks); 836 auto *OldCond = BI->getCondition(); 837 BI->setCondition(AllChecks); 838 RecursivelyDeleteTriviallyDeadInstructions(OldCond); 839 assert(isGuardAsWidenableBranch(BI) && 840 "Stopped being a guard after transform?"); 841 842 LLVM_DEBUG(dbgs() << "Widened checks = " << NumWidened << "\n"); 843 return true; 844 } 845 846 Optional<LoopICmp> LoopPredication::parseLoopLatchICmp() { 847 using namespace PatternMatch; 848 849 BasicBlock *LoopLatch = L->getLoopLatch(); 850 if (!LoopLatch) { 851 LLVM_DEBUG(dbgs() << "The loop doesn't have a single latch!\n"); 852 return None; 853 } 854 855 auto *BI = dyn_cast<BranchInst>(LoopLatch->getTerminator()); 856 if (!BI || !BI->isConditional()) { 857 LLVM_DEBUG(dbgs() << "Failed to match the latch terminator!\n"); 858 return None; 859 } 860 BasicBlock *TrueDest = BI->getSuccessor(0); 861 assert( 862 (TrueDest == L->getHeader() || BI->getSuccessor(1) == L->getHeader()) && 863 "One of the latch's destinations must be the header"); 864 865 auto *ICI = dyn_cast<ICmpInst>(BI->getCondition()); 866 if (!ICI) { 867 LLVM_DEBUG(dbgs() << "Failed to match the latch condition!\n"); 868 return None; 869 } 870 auto Result = parseLoopICmp(ICI); 871 if (!Result) { 872 LLVM_DEBUG(dbgs() << "Failed to parse the loop latch condition!\n"); 873 return None; 874 } 875 876 if (TrueDest != L->getHeader()) 877 Result->Pred = ICmpInst::getInversePredicate(Result->Pred); 878 879 // Check affine first, so if it's not we don't try to compute the step 880 // recurrence. 881 if (!Result->IV->isAffine()) { 882 LLVM_DEBUG(dbgs() << "The induction variable is not affine!\n"); 883 return None; 884 } 885 886 auto *Step = Result->IV->getStepRecurrence(*SE); 887 if (!isSupportedStep(Step)) { 888 LLVM_DEBUG(dbgs() << "Unsupported loop stride(" << *Step << ")!\n"); 889 return None; 890 } 891 892 auto IsUnsupportedPredicate = [](const SCEV *Step, ICmpInst::Predicate Pred) { 893 if (Step->isOne()) { 894 return Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_SLT && 895 Pred != ICmpInst::ICMP_ULE && Pred != ICmpInst::ICMP_SLE; 896 } else { 897 assert(Step->isAllOnesValue() && "Step should be -1!"); 898 return Pred != ICmpInst::ICMP_UGT && Pred != ICmpInst::ICMP_SGT && 899 Pred != ICmpInst::ICMP_UGE && Pred != ICmpInst::ICMP_SGE; 900 } 901 }; 902 903 normalizePredicate(SE, L, *Result); 904 if (IsUnsupportedPredicate(Step, Result->Pred)) { 905 LLVM_DEBUG(dbgs() << "Unsupported loop latch predicate(" << Result->Pred 906 << ")!\n"); 907 return None; 908 } 909 910 return Result; 911 } 912 913 914 bool LoopPredication::isLoopProfitableToPredicate() { 915 if (SkipProfitabilityChecks || !BPI) 916 return true; 917 918 SmallVector<std::pair<BasicBlock *, BasicBlock *>, 8> ExitEdges; 919 L->getExitEdges(ExitEdges); 920 // If there is only one exiting edge in the loop, it is always profitable to 921 // predicate the loop. 922 if (ExitEdges.size() == 1) 923 return true; 924 925 // Calculate the exiting probabilities of all exiting edges from the loop, 926 // starting with the LatchExitProbability. 927 // Heuristic for profitability: If any of the exiting blocks' probability of 928 // exiting the loop is larger than exiting through the latch block, it's not 929 // profitable to predicate the loop. 930 auto *LatchBlock = L->getLoopLatch(); 931 assert(LatchBlock && "Should have a single latch at this point!"); 932 auto *LatchTerm = LatchBlock->getTerminator(); 933 assert(LatchTerm->getNumSuccessors() == 2 && 934 "expected to be an exiting block with 2 succs!"); 935 unsigned LatchBrExitIdx = 936 LatchTerm->getSuccessor(0) == L->getHeader() ? 1 : 0; 937 BranchProbability LatchExitProbability = 938 BPI->getEdgeProbability(LatchBlock, LatchBrExitIdx); 939 940 // Protect against degenerate inputs provided by the user. Providing a value 941 // less than one, can invert the definition of profitable loop predication. 942 float ScaleFactor = LatchExitProbabilityScale; 943 if (ScaleFactor < 1) { 944 LLVM_DEBUG( 945 dbgs() 946 << "Ignored user setting for loop-predication-latch-probability-scale: " 947 << LatchExitProbabilityScale << "\n"); 948 LLVM_DEBUG(dbgs() << "The value is set to 1.0\n"); 949 ScaleFactor = 1.0; 950 } 951 const auto LatchProbabilityThreshold = 952 LatchExitProbability * ScaleFactor; 953 954 for (const auto &ExitEdge : ExitEdges) { 955 BranchProbability ExitingBlockProbability = 956 BPI->getEdgeProbability(ExitEdge.first, ExitEdge.second); 957 // Some exiting edge has higher probability than the latch exiting edge. 958 // No longer profitable to predicate. 959 if (ExitingBlockProbability > LatchProbabilityThreshold) 960 return false; 961 } 962 // Using BPI, we have concluded that the most probable way to exit from the 963 // loop is through the latch (or there's no profile information and all 964 // exits are equally likely). 965 return true; 966 } 967 968 /// If we can (cheaply) find a widenable branch which controls entry into the 969 /// loop, return it. 970 static BranchInst *FindWidenableTerminatorAboveLoop(Loop *L, LoopInfo &LI) { 971 // Walk back through any unconditional executed blocks and see if we can find 972 // a widenable condition which seems to control execution of this loop. Note 973 // that we predict that maythrow calls are likely untaken and thus that it's 974 // profitable to widen a branch before a maythrow call with a condition 975 // afterwards even though that may cause the slow path to run in a case where 976 // it wouldn't have otherwise. 977 BasicBlock *BB = L->getLoopPreheader(); 978 if (!BB) 979 return nullptr; 980 do { 981 if (BasicBlock *Pred = BB->getSinglePredecessor()) 982 if (BB == Pred->getSingleSuccessor()) { 983 BB = Pred; 984 continue; 985 } 986 break; 987 } while (true); 988 989 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 990 auto *Term = Pred->getTerminator(); 991 992 Value *Cond, *WC; 993 BasicBlock *IfTrueBB, *IfFalseBB; 994 if (parseWidenableBranch(Term, Cond, WC, IfTrueBB, IfFalseBB) && 995 IfTrueBB == BB) 996 return cast<BranchInst>(Term); 997 } 998 return nullptr; 999 } 1000 1001 /// Return the minimum of all analyzeable exit counts. This is an upper bound 1002 /// on the actual exit count. If there are not at least two analyzeable exits, 1003 /// returns SCEVCouldNotCompute. 1004 static const SCEV *getMinAnalyzeableBackedgeTakenCount(ScalarEvolution &SE, 1005 DominatorTree &DT, 1006 Loop *L) { 1007 SmallVector<BasicBlock *, 16> ExitingBlocks; 1008 L->getExitingBlocks(ExitingBlocks); 1009 1010 SmallVector<const SCEV *, 4> ExitCounts; 1011 for (BasicBlock *ExitingBB : ExitingBlocks) { 1012 const SCEV *ExitCount = SE.getExitCount(L, ExitingBB); 1013 if (isa<SCEVCouldNotCompute>(ExitCount)) 1014 continue; 1015 assert(DT.dominates(ExitingBB, L->getLoopLatch()) && 1016 "We should only have known counts for exiting blocks that " 1017 "dominate latch!"); 1018 ExitCounts.push_back(ExitCount); 1019 } 1020 if (ExitCounts.size() < 2) 1021 return SE.getCouldNotCompute(); 1022 return SE.getUMinFromMismatchedTypes(ExitCounts); 1023 } 1024 1025 /// This implements an analogous, but entirely distinct transform from the main 1026 /// loop predication transform. This one is phrased in terms of using a 1027 /// widenable branch *outside* the loop to allow us to simplify loop exits in a 1028 /// following loop. This is close in spirit to the IndVarSimplify transform 1029 /// of the same name, but is materially different widening loosens legality 1030 /// sharply. 1031 bool LoopPredication::predicateLoopExits(Loop *L, SCEVExpander &Rewriter) { 1032 // The transformation performed here aims to widen a widenable condition 1033 // above the loop such that all analyzeable exit leading to deopt are dead. 1034 // It assumes that the latch is the dominant exit for profitability and that 1035 // exits branching to deoptimizing blocks are rarely taken. It relies on the 1036 // semantics of widenable expressions for legality. (i.e. being able to fall 1037 // down the widenable path spuriously allows us to ignore exit order, 1038 // unanalyzeable exits, side effects, exceptional exits, and other challenges 1039 // which restrict the applicability of the non-WC based version of this 1040 // transform in IndVarSimplify.) 1041 // 1042 // NOTE ON POISON/UNDEF - We're hoisting an expression above guards which may 1043 // imply flags on the expression being hoisted and inserting new uses (flags 1044 // are only correct for current uses). The result is that we may be 1045 // inserting a branch on the value which can be either poison or undef. In 1046 // this case, the branch can legally go either way; we just need to avoid 1047 // introducing UB. This is achieved through the use of the freeze 1048 // instruction. 1049 1050 SmallVector<BasicBlock *, 16> ExitingBlocks; 1051 L->getExitingBlocks(ExitingBlocks); 1052 1053 if (ExitingBlocks.empty()) 1054 return false; // Nothing to do. 1055 1056 auto *Latch = L->getLoopLatch(); 1057 if (!Latch) 1058 return false; 1059 1060 auto *WidenableBR = FindWidenableTerminatorAboveLoop(L, *LI); 1061 if (!WidenableBR) 1062 return false; 1063 1064 const SCEV *LatchEC = SE->getExitCount(L, Latch); 1065 if (isa<SCEVCouldNotCompute>(LatchEC)) 1066 return false; // profitability - want hot exit in analyzeable set 1067 1068 // At this point, we have found an analyzeable latch, and a widenable 1069 // condition above the loop. If we have a widenable exit within the loop 1070 // (for which we can't compute exit counts), drop the ability to further 1071 // widen so that we gain ability to analyze it's exit count and perform this 1072 // transform. TODO: It'd be nice to know for sure the exit became 1073 // analyzeable after dropping widenability. 1074 { 1075 bool Invalidate = false; 1076 1077 for (auto *ExitingBB : ExitingBlocks) { 1078 if (LI->getLoopFor(ExitingBB) != L) 1079 continue; 1080 1081 auto *BI = dyn_cast<BranchInst>(ExitingBB->getTerminator()); 1082 if (!BI) 1083 continue; 1084 1085 Use *Cond, *WC; 1086 BasicBlock *IfTrueBB, *IfFalseBB; 1087 if (parseWidenableBranch(BI, Cond, WC, IfTrueBB, IfFalseBB) && 1088 L->contains(IfTrueBB)) { 1089 WC->set(ConstantInt::getTrue(IfTrueBB->getContext())); 1090 Invalidate = true; 1091 } 1092 } 1093 if (Invalidate) 1094 SE->forgetLoop(L); 1095 } 1096 1097 // The use of umin(all analyzeable exits) instead of latch is subtle, but 1098 // important for profitability. We may have a loop which hasn't been fully 1099 // canonicalized just yet. If the exit we chose to widen is provably never 1100 // taken, we want the widened form to *also* be provably never taken. We 1101 // can't guarantee this as a current unanalyzeable exit may later become 1102 // analyzeable, but we can at least avoid the obvious cases. 1103 const SCEV *MinEC = getMinAnalyzeableBackedgeTakenCount(*SE, *DT, L); 1104 if (isa<SCEVCouldNotCompute>(MinEC) || MinEC->getType()->isPointerTy() || 1105 !SE->isLoopInvariant(MinEC, L) || 1106 !isSafeToExpandAt(MinEC, WidenableBR, *SE)) 1107 return false; 1108 1109 // Subtlety: We need to avoid inserting additional uses of the WC. We know 1110 // that it can only have one transitive use at the moment, and thus moving 1111 // that use to just before the branch and inserting code before it and then 1112 // modifying the operand is legal. 1113 auto *IP = cast<Instruction>(WidenableBR->getCondition()); 1114 IP->moveBefore(WidenableBR); 1115 Rewriter.setInsertPoint(IP); 1116 IRBuilder<> B(IP); 1117 1118 bool Changed = false; 1119 Value *MinECV = nullptr; // lazily generated if needed 1120 for (BasicBlock *ExitingBB : ExitingBlocks) { 1121 // If our exiting block exits multiple loops, we can only rewrite the 1122 // innermost one. Otherwise, we're changing how many times the innermost 1123 // loop runs before it exits. 1124 if (LI->getLoopFor(ExitingBB) != L) 1125 continue; 1126 1127 // Can't rewrite non-branch yet. 1128 auto *BI = dyn_cast<BranchInst>(ExitingBB->getTerminator()); 1129 if (!BI) 1130 continue; 1131 1132 // If already constant, nothing to do. 1133 if (isa<Constant>(BI->getCondition())) 1134 continue; 1135 1136 const SCEV *ExitCount = SE->getExitCount(L, ExitingBB); 1137 if (isa<SCEVCouldNotCompute>(ExitCount) || 1138 ExitCount->getType()->isPointerTy() || 1139 !isSafeToExpandAt(ExitCount, WidenableBR, *SE)) 1140 continue; 1141 1142 const bool ExitIfTrue = !L->contains(*succ_begin(ExitingBB)); 1143 BasicBlock *ExitBB = BI->getSuccessor(ExitIfTrue ? 0 : 1); 1144 if (!ExitBB->getPostdominatingDeoptimizeCall()) 1145 continue; 1146 1147 /// Here we can be fairly sure that executing this exit will most likely 1148 /// lead to executing llvm.experimental.deoptimize. 1149 /// This is a profitability heuristic, not a legality constraint. 1150 1151 // If we found a widenable exit condition, do two things: 1152 // 1) fold the widened exit test into the widenable condition 1153 // 2) fold the branch to untaken - avoids infinite looping 1154 1155 Value *ECV = Rewriter.expandCodeFor(ExitCount); 1156 if (!MinECV) 1157 MinECV = Rewriter.expandCodeFor(MinEC); 1158 Value *RHS = MinECV; 1159 if (ECV->getType() != RHS->getType()) { 1160 Type *WiderTy = SE->getWiderType(ECV->getType(), RHS->getType()); 1161 ECV = B.CreateZExt(ECV, WiderTy); 1162 RHS = B.CreateZExt(RHS, WiderTy); 1163 } 1164 assert(!Latch || DT->dominates(ExitingBB, Latch)); 1165 Value *NewCond = B.CreateICmp(ICmpInst::ICMP_UGT, ECV, RHS); 1166 // Freeze poison or undef to an arbitrary bit pattern to ensure we can 1167 // branch without introducing UB. See NOTE ON POISON/UNDEF above for 1168 // context. 1169 NewCond = B.CreateFreeze(NewCond); 1170 1171 widenWidenableBranch(WidenableBR, NewCond); 1172 1173 Value *OldCond = BI->getCondition(); 1174 BI->setCondition(ConstantInt::get(OldCond->getType(), !ExitIfTrue)); 1175 Changed = true; 1176 } 1177 1178 if (Changed) 1179 // We just mutated a bunch of loop exits changing there exit counts 1180 // widely. We need to force recomputation of the exit counts given these 1181 // changes. Note that all of the inserted exits are never taken, and 1182 // should be removed next time the CFG is modified. 1183 SE->forgetLoop(L); 1184 return Changed; 1185 } 1186 1187 bool LoopPredication::runOnLoop(Loop *Loop) { 1188 L = Loop; 1189 1190 LLVM_DEBUG(dbgs() << "Analyzing "); 1191 LLVM_DEBUG(L->dump()); 1192 1193 Module *M = L->getHeader()->getModule(); 1194 1195 // There is nothing to do if the module doesn't use guards 1196 auto *GuardDecl = 1197 M->getFunction(Intrinsic::getName(Intrinsic::experimental_guard)); 1198 bool HasIntrinsicGuards = GuardDecl && !GuardDecl->use_empty(); 1199 auto *WCDecl = M->getFunction( 1200 Intrinsic::getName(Intrinsic::experimental_widenable_condition)); 1201 bool HasWidenableConditions = 1202 PredicateWidenableBranchGuards && WCDecl && !WCDecl->use_empty(); 1203 if (!HasIntrinsicGuards && !HasWidenableConditions) 1204 return false; 1205 1206 DL = &M->getDataLayout(); 1207 1208 Preheader = L->getLoopPreheader(); 1209 if (!Preheader) 1210 return false; 1211 1212 auto LatchCheckOpt = parseLoopLatchICmp(); 1213 if (!LatchCheckOpt) 1214 return false; 1215 LatchCheck = *LatchCheckOpt; 1216 1217 LLVM_DEBUG(dbgs() << "Latch check:\n"); 1218 LLVM_DEBUG(LatchCheck.dump()); 1219 1220 if (!isLoopProfitableToPredicate()) { 1221 LLVM_DEBUG(dbgs() << "Loop not profitable to predicate!\n"); 1222 return false; 1223 } 1224 // Collect all the guards into a vector and process later, so as not 1225 // to invalidate the instruction iterator. 1226 SmallVector<IntrinsicInst *, 4> Guards; 1227 SmallVector<BranchInst *, 4> GuardsAsWidenableBranches; 1228 for (const auto BB : L->blocks()) { 1229 for (auto &I : *BB) 1230 if (isGuard(&I)) 1231 Guards.push_back(cast<IntrinsicInst>(&I)); 1232 if (PredicateWidenableBranchGuards && 1233 isGuardAsWidenableBranch(BB->getTerminator())) 1234 GuardsAsWidenableBranches.push_back( 1235 cast<BranchInst>(BB->getTerminator())); 1236 } 1237 1238 SCEVExpander Expander(*SE, *DL, "loop-predication"); 1239 bool Changed = false; 1240 for (auto *Guard : Guards) 1241 Changed |= widenGuardConditions(Guard, Expander); 1242 for (auto *Guard : GuardsAsWidenableBranches) 1243 Changed |= widenWidenableBranchGuardConditions(Guard, Expander); 1244 Changed |= predicateLoopExits(L, Expander); 1245 return Changed; 1246 } 1247