1 //===- InductiveRangeCheckElimination.cpp - -------------------------------===// 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 InductiveRangeCheckElimination pass splits a loop's iteration space into 10 // three disjoint ranges. It does that in a way such that the loop running in 11 // the middle loop provably does not need range checks. As an example, it will 12 // convert 13 // 14 // len = < known positive > 15 // for (i = 0; i < n; i++) { 16 // if (0 <= i && i < len) { 17 // do_something(); 18 // } else { 19 // throw_out_of_bounds(); 20 // } 21 // } 22 // 23 // to 24 // 25 // len = < known positive > 26 // limit = smin(n, len) 27 // // no first segment 28 // for (i = 0; i < limit; i++) { 29 // if (0 <= i && i < len) { // this check is fully redundant 30 // do_something(); 31 // } else { 32 // throw_out_of_bounds(); 33 // } 34 // } 35 // for (i = limit; i < n; i++) { 36 // if (0 <= i && i < len) { 37 // do_something(); 38 // } else { 39 // throw_out_of_bounds(); 40 // } 41 // } 42 // 43 //===----------------------------------------------------------------------===// 44 45 #include "llvm/Transforms/Scalar/InductiveRangeCheckElimination.h" 46 #include "llvm/ADT/APInt.h" 47 #include "llvm/ADT/ArrayRef.h" 48 #include "llvm/ADT/PriorityWorklist.h" 49 #include "llvm/ADT/SmallPtrSet.h" 50 #include "llvm/ADT/SmallVector.h" 51 #include "llvm/ADT/StringRef.h" 52 #include "llvm/ADT/Twine.h" 53 #include "llvm/Analysis/BlockFrequencyInfo.h" 54 #include "llvm/Analysis/BranchProbabilityInfo.h" 55 #include "llvm/Analysis/LoopAnalysisManager.h" 56 #include "llvm/Analysis/LoopInfo.h" 57 #include "llvm/Analysis/ScalarEvolution.h" 58 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 59 #include "llvm/IR/BasicBlock.h" 60 #include "llvm/IR/CFG.h" 61 #include "llvm/IR/Constants.h" 62 #include "llvm/IR/DerivedTypes.h" 63 #include "llvm/IR/Dominators.h" 64 #include "llvm/IR/Function.h" 65 #include "llvm/IR/IRBuilder.h" 66 #include "llvm/IR/InstrTypes.h" 67 #include "llvm/IR/Instructions.h" 68 #include "llvm/IR/Metadata.h" 69 #include "llvm/IR/Module.h" 70 #include "llvm/IR/PatternMatch.h" 71 #include "llvm/IR/Type.h" 72 #include "llvm/IR/Use.h" 73 #include "llvm/IR/User.h" 74 #include "llvm/IR/Value.h" 75 #include "llvm/Support/BranchProbability.h" 76 #include "llvm/Support/Casting.h" 77 #include "llvm/Support/CommandLine.h" 78 #include "llvm/Support/Compiler.h" 79 #include "llvm/Support/Debug.h" 80 #include "llvm/Support/ErrorHandling.h" 81 #include "llvm/Support/raw_ostream.h" 82 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 83 #include "llvm/Transforms/Utils/Cloning.h" 84 #include "llvm/Transforms/Utils/LoopSimplify.h" 85 #include "llvm/Transforms/Utils/LoopUtils.h" 86 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" 87 #include "llvm/Transforms/Utils/ValueMapper.h" 88 #include <algorithm> 89 #include <cassert> 90 #include <iterator> 91 #include <limits> 92 #include <optional> 93 #include <utility> 94 #include <vector> 95 96 using namespace llvm; 97 using namespace llvm::PatternMatch; 98 99 static cl::opt<unsigned> LoopSizeCutoff("irce-loop-size-cutoff", cl::Hidden, 100 cl::init(64)); 101 102 static cl::opt<bool> PrintChangedLoops("irce-print-changed-loops", cl::Hidden, 103 cl::init(false)); 104 105 static cl::opt<bool> PrintRangeChecks("irce-print-range-checks", cl::Hidden, 106 cl::init(false)); 107 108 static cl::opt<bool> SkipProfitabilityChecks("irce-skip-profitability-checks", 109 cl::Hidden, cl::init(false)); 110 111 static cl::opt<unsigned> MinRuntimeIterations("irce-min-runtime-iterations", 112 cl::Hidden, cl::init(10)); 113 114 static cl::opt<bool> AllowUnsignedLatchCondition("irce-allow-unsigned-latch", 115 cl::Hidden, cl::init(true)); 116 117 static cl::opt<bool> AllowNarrowLatchCondition( 118 "irce-allow-narrow-latch", cl::Hidden, cl::init(true), 119 cl::desc("If set to true, IRCE may eliminate wide range checks in loops " 120 "with narrow latch condition.")); 121 122 static cl::opt<unsigned> MaxTypeSizeForOverflowCheck( 123 "irce-max-type-size-for-overflow-check", cl::Hidden, cl::init(32), 124 cl::desc( 125 "Maximum size of range check type for which can be produced runtime " 126 "overflow check of its limit's computation")); 127 128 static cl::opt<bool> 129 PrintScaledBoundaryRangeChecks("irce-print-scaled-boundary-range-checks", 130 cl::Hidden, cl::init(false)); 131 132 static const char *ClonedLoopTag = "irce.loop.clone"; 133 134 #define DEBUG_TYPE "irce" 135 136 namespace { 137 138 /// An inductive range check is conditional branch in a loop with 139 /// 140 /// 1. a very cold successor (i.e. the branch jumps to that successor very 141 /// rarely) 142 /// 143 /// and 144 /// 145 /// 2. a condition that is provably true for some contiguous range of values 146 /// taken by the containing loop's induction variable. 147 /// 148 class InductiveRangeCheck { 149 150 const SCEV *Begin = nullptr; 151 const SCEV *Step = nullptr; 152 const SCEV *End = nullptr; 153 Use *CheckUse = nullptr; 154 155 static bool parseRangeCheckICmp(Loop *L, ICmpInst *ICI, ScalarEvolution &SE, 156 const SCEVAddRecExpr *&Index, 157 const SCEV *&End); 158 159 static void 160 extractRangeChecksFromCond(Loop *L, ScalarEvolution &SE, Use &ConditionUse, 161 SmallVectorImpl<InductiveRangeCheck> &Checks, 162 SmallPtrSetImpl<Value *> &Visited); 163 164 static bool parseIvAgaisntLimit(Loop *L, Value *LHS, Value *RHS, 165 ICmpInst::Predicate Pred, ScalarEvolution &SE, 166 const SCEVAddRecExpr *&Index, 167 const SCEV *&End); 168 169 static bool reassociateSubLHS(Loop *L, Value *VariantLHS, Value *InvariantRHS, 170 ICmpInst::Predicate Pred, ScalarEvolution &SE, 171 const SCEVAddRecExpr *&Index, const SCEV *&End); 172 173 public: 174 const SCEV *getBegin() const { return Begin; } 175 const SCEV *getStep() const { return Step; } 176 const SCEV *getEnd() const { return End; } 177 178 void print(raw_ostream &OS) const { 179 OS << "InductiveRangeCheck:\n"; 180 OS << " Begin: "; 181 Begin->print(OS); 182 OS << " Step: "; 183 Step->print(OS); 184 OS << " End: "; 185 End->print(OS); 186 OS << "\n CheckUse: "; 187 getCheckUse()->getUser()->print(OS); 188 OS << " Operand: " << getCheckUse()->getOperandNo() << "\n"; 189 } 190 191 LLVM_DUMP_METHOD 192 void dump() { 193 print(dbgs()); 194 } 195 196 Use *getCheckUse() const { return CheckUse; } 197 198 /// Represents an signed integer range [Range.getBegin(), Range.getEnd()). If 199 /// R.getEnd() le R.getBegin(), then R denotes the empty range. 200 201 class Range { 202 const SCEV *Begin; 203 const SCEV *End; 204 205 public: 206 Range(const SCEV *Begin, const SCEV *End) : Begin(Begin), End(End) { 207 assert(Begin->getType() == End->getType() && "ill-typed range!"); 208 } 209 210 Type *getType() const { return Begin->getType(); } 211 const SCEV *getBegin() const { return Begin; } 212 const SCEV *getEnd() const { return End; } 213 bool isEmpty(ScalarEvolution &SE, bool IsSigned) const { 214 if (Begin == End) 215 return true; 216 if (IsSigned) 217 return SE.isKnownPredicate(ICmpInst::ICMP_SGE, Begin, End); 218 else 219 return SE.isKnownPredicate(ICmpInst::ICMP_UGE, Begin, End); 220 } 221 }; 222 223 /// This is the value the condition of the branch needs to evaluate to for the 224 /// branch to take the hot successor (see (1) above). 225 bool getPassingDirection() { return true; } 226 227 /// Computes a range for the induction variable (IndVar) in which the range 228 /// check is redundant and can be constant-folded away. The induction 229 /// variable is not required to be the canonical {0,+,1} induction variable. 230 std::optional<Range> computeSafeIterationSpace(ScalarEvolution &SE, 231 const SCEVAddRecExpr *IndVar, 232 bool IsLatchSigned) const; 233 234 /// Parse out a set of inductive range checks from \p BI and append them to \p 235 /// Checks. 236 /// 237 /// NB! There may be conditions feeding into \p BI that aren't inductive range 238 /// checks, and hence don't end up in \p Checks. 239 static void extractRangeChecksFromBranch( 240 BranchInst *BI, Loop *L, ScalarEvolution &SE, BranchProbabilityInfo *BPI, 241 SmallVectorImpl<InductiveRangeCheck> &Checks, bool &Changed); 242 }; 243 244 struct LoopStructure; 245 246 class InductiveRangeCheckElimination { 247 ScalarEvolution &SE; 248 BranchProbabilityInfo *BPI; 249 DominatorTree &DT; 250 LoopInfo &LI; 251 252 using GetBFIFunc = 253 std::optional<llvm::function_ref<llvm::BlockFrequencyInfo &()>>; 254 GetBFIFunc GetBFI; 255 256 // Returns true if it is profitable to do a transform basing on estimation of 257 // number of iterations. 258 bool isProfitableToTransform(const Loop &L, LoopStructure &LS); 259 260 public: 261 InductiveRangeCheckElimination(ScalarEvolution &SE, 262 BranchProbabilityInfo *BPI, DominatorTree &DT, 263 LoopInfo &LI, GetBFIFunc GetBFI = std::nullopt) 264 : SE(SE), BPI(BPI), DT(DT), LI(LI), GetBFI(GetBFI) {} 265 266 bool run(Loop *L, function_ref<void(Loop *, bool)> LPMAddNewLoop); 267 }; 268 269 } // end anonymous namespace 270 271 /// Parse a single ICmp instruction, `ICI`, into a range check. If `ICI` cannot 272 /// be interpreted as a range check, return false. Otherwise set `Index` to the 273 /// SCEV being range checked, and set `End` to the upper or lower limit `Index` 274 /// is being range checked. 275 bool InductiveRangeCheck::parseRangeCheckICmp(Loop *L, ICmpInst *ICI, 276 ScalarEvolution &SE, 277 const SCEVAddRecExpr *&Index, 278 const SCEV *&End) { 279 auto IsLoopInvariant = [&SE, L](Value *V) { 280 return SE.isLoopInvariant(SE.getSCEV(V), L); 281 }; 282 283 ICmpInst::Predicate Pred = ICI->getPredicate(); 284 Value *LHS = ICI->getOperand(0); 285 Value *RHS = ICI->getOperand(1); 286 287 // Canonicalize to the `Index Pred Invariant` comparison 288 if (IsLoopInvariant(LHS)) { 289 std::swap(LHS, RHS); 290 Pred = CmpInst::getSwappedPredicate(Pred); 291 } else if (!IsLoopInvariant(RHS)) 292 // Both LHS and RHS are loop variant 293 return false; 294 295 if (parseIvAgaisntLimit(L, LHS, RHS, Pred, SE, Index, End)) 296 return true; 297 298 if (reassociateSubLHS(L, LHS, RHS, Pred, SE, Index, End)) 299 return true; 300 301 // TODO: support ReassociateAddLHS 302 return false; 303 } 304 305 // Try to parse range check in the form of "IV vs Limit" 306 bool InductiveRangeCheck::parseIvAgaisntLimit(Loop *L, Value *LHS, Value *RHS, 307 ICmpInst::Predicate Pred, 308 ScalarEvolution &SE, 309 const SCEVAddRecExpr *&Index, 310 const SCEV *&End) { 311 312 auto SIntMaxSCEV = [&](Type *T) { 313 unsigned BitWidth = cast<IntegerType>(T)->getBitWidth(); 314 return SE.getConstant(APInt::getSignedMaxValue(BitWidth)); 315 }; 316 317 const auto *AddRec = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(LHS)); 318 if (!AddRec) 319 return false; 320 321 // We strengthen "0 <= I" to "0 <= I < INT_SMAX" and "I < L" to "0 <= I < L". 322 // We can potentially do much better here. 323 // If we want to adjust upper bound for the unsigned range check as we do it 324 // for signed one, we will need to pick Unsigned max 325 switch (Pred) { 326 default: 327 return false; 328 329 case ICmpInst::ICMP_SGE: 330 if (match(RHS, m_ConstantInt<0>())) { 331 Index = AddRec; 332 End = SIntMaxSCEV(Index->getType()); 333 return true; 334 } 335 return false; 336 337 case ICmpInst::ICMP_SGT: 338 if (match(RHS, m_ConstantInt<-1>())) { 339 Index = AddRec; 340 End = SIntMaxSCEV(Index->getType()); 341 return true; 342 } 343 return false; 344 345 case ICmpInst::ICMP_SLT: 346 case ICmpInst::ICMP_ULT: 347 Index = AddRec; 348 End = SE.getSCEV(RHS); 349 return true; 350 351 case ICmpInst::ICMP_SLE: 352 case ICmpInst::ICMP_ULE: 353 const SCEV *One = SE.getOne(RHS->getType()); 354 const SCEV *RHSS = SE.getSCEV(RHS); 355 bool Signed = Pred == ICmpInst::ICMP_SLE; 356 if (SE.willNotOverflow(Instruction::BinaryOps::Add, Signed, RHSS, One)) { 357 Index = AddRec; 358 End = SE.getAddExpr(RHSS, One); 359 return true; 360 } 361 return false; 362 } 363 364 llvm_unreachable("default clause returns!"); 365 } 366 367 // Try to parse range check in the form of "IV - Offset vs Limit" or "Offset - 368 // IV vs Limit" 369 bool InductiveRangeCheck::reassociateSubLHS( 370 Loop *L, Value *VariantLHS, Value *InvariantRHS, ICmpInst::Predicate Pred, 371 ScalarEvolution &SE, const SCEVAddRecExpr *&Index, const SCEV *&End) { 372 Value *LHS, *RHS; 373 if (!match(VariantLHS, m_Sub(m_Value(LHS), m_Value(RHS)))) 374 return false; 375 376 const SCEV *IV = SE.getSCEV(LHS); 377 const SCEV *Offset = SE.getSCEV(RHS); 378 const SCEV *Limit = SE.getSCEV(InvariantRHS); 379 380 bool OffsetSubtracted = false; 381 if (SE.isLoopInvariant(IV, L)) 382 // "Offset - IV vs Limit" 383 std::swap(IV, Offset); 384 else if (SE.isLoopInvariant(Offset, L)) 385 // "IV - Offset vs Limit" 386 OffsetSubtracted = true; 387 else 388 return false; 389 390 const auto *AddRec = dyn_cast<SCEVAddRecExpr>(IV); 391 if (!AddRec) 392 return false; 393 394 // In order to turn "IV - Offset < Limit" into "IV < Limit + Offset", we need 395 // to be able to freely move values from left side of inequality to right side 396 // (just as in normal linear arithmetics). Overflows make things much more 397 // complicated, so we want to avoid this. 398 // 399 // Let's prove that the initial subtraction doesn't overflow with all IV's 400 // values from the safe range constructed for that check. 401 // 402 // [Case 1] IV - Offset < Limit 403 // It doesn't overflow if: 404 // SINT_MIN <= IV - Offset <= SINT_MAX 405 // In terms of scaled SINT we need to prove: 406 // SINT_MIN + Offset <= IV <= SINT_MAX + Offset 407 // Safe range will be constructed: 408 // 0 <= IV < Limit + Offset 409 // It means that 'IV - Offset' doesn't underflow, because: 410 // SINT_MIN + Offset < 0 <= IV 411 // and doesn't overflow: 412 // IV < Limit + Offset <= SINT_MAX + Offset 413 // 414 // [Case 2] Offset - IV > Limit 415 // It doesn't overflow if: 416 // SINT_MIN <= Offset - IV <= SINT_MAX 417 // In terms of scaled SINT we need to prove: 418 // -SINT_MIN >= IV - Offset >= -SINT_MAX 419 // Offset - SINT_MIN >= IV >= Offset - SINT_MAX 420 // Safe range will be constructed: 421 // 0 <= IV < Offset - Limit 422 // It means that 'Offset - IV' doesn't underflow, because 423 // Offset - SINT_MAX < 0 <= IV 424 // and doesn't overflow: 425 // IV < Offset - Limit <= Offset - SINT_MIN 426 // 427 // For the computed upper boundary of the IV's range (Offset +/- Limit) we 428 // don't know exactly whether it overflows or not. So if we can't prove this 429 // fact at compile time, we scale boundary computations to a wider type with 430 // the intention to add runtime overflow check. 431 432 auto getExprScaledIfOverflow = [&](Instruction::BinaryOps BinOp, 433 const SCEV *LHS, 434 const SCEV *RHS) -> const SCEV * { 435 const SCEV *(ScalarEvolution::*Operation)(const SCEV *, const SCEV *, 436 SCEV::NoWrapFlags, unsigned); 437 switch (BinOp) { 438 default: 439 llvm_unreachable("Unsupported binary op"); 440 case Instruction::Add: 441 Operation = &ScalarEvolution::getAddExpr; 442 break; 443 case Instruction::Sub: 444 Operation = &ScalarEvolution::getMinusSCEV; 445 break; 446 } 447 448 if (SE.willNotOverflow(BinOp, ICmpInst::isSigned(Pred), LHS, RHS, 449 cast<Instruction>(VariantLHS))) 450 return (SE.*Operation)(LHS, RHS, SCEV::FlagAnyWrap, 0); 451 452 // We couldn't prove that the expression does not overflow. 453 // Than scale it to a wider type to check overflow at runtime. 454 auto *Ty = cast<IntegerType>(LHS->getType()); 455 if (Ty->getBitWidth() > MaxTypeSizeForOverflowCheck) 456 return nullptr; 457 458 auto WideTy = IntegerType::get(Ty->getContext(), Ty->getBitWidth() * 2); 459 return (SE.*Operation)(SE.getSignExtendExpr(LHS, WideTy), 460 SE.getSignExtendExpr(RHS, WideTy), SCEV::FlagAnyWrap, 461 0); 462 }; 463 464 if (OffsetSubtracted) 465 // "IV - Offset < Limit" -> "IV" < Offset + Limit 466 Limit = getExprScaledIfOverflow(Instruction::BinaryOps::Add, Offset, Limit); 467 else { 468 // "Offset - IV > Limit" -> "IV" < Offset - Limit 469 Limit = getExprScaledIfOverflow(Instruction::BinaryOps::Sub, Offset, Limit); 470 Pred = ICmpInst::getSwappedPredicate(Pred); 471 } 472 473 if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) { 474 // "Expr <= Limit" -> "Expr < Limit + 1" 475 if (Pred == ICmpInst::ICMP_SLE && Limit) 476 Limit = getExprScaledIfOverflow(Instruction::BinaryOps::Add, Limit, 477 SE.getOne(Limit->getType())); 478 if (Limit) { 479 Index = AddRec; 480 End = Limit; 481 return true; 482 } 483 } 484 return false; 485 } 486 487 void InductiveRangeCheck::extractRangeChecksFromCond( 488 Loop *L, ScalarEvolution &SE, Use &ConditionUse, 489 SmallVectorImpl<InductiveRangeCheck> &Checks, 490 SmallPtrSetImpl<Value *> &Visited) { 491 Value *Condition = ConditionUse.get(); 492 if (!Visited.insert(Condition).second) 493 return; 494 495 // TODO: Do the same for OR, XOR, NOT etc? 496 if (match(Condition, m_LogicalAnd(m_Value(), m_Value()))) { 497 extractRangeChecksFromCond(L, SE, cast<User>(Condition)->getOperandUse(0), 498 Checks, Visited); 499 extractRangeChecksFromCond(L, SE, cast<User>(Condition)->getOperandUse(1), 500 Checks, Visited); 501 return; 502 } 503 504 ICmpInst *ICI = dyn_cast<ICmpInst>(Condition); 505 if (!ICI) 506 return; 507 508 const SCEV *End = nullptr; 509 const SCEVAddRecExpr *IndexAddRec = nullptr; 510 if (!parseRangeCheckICmp(L, ICI, SE, IndexAddRec, End)) 511 return; 512 513 assert(IndexAddRec && "IndexAddRec was not computed"); 514 assert(End && "End was not computed"); 515 516 if ((IndexAddRec->getLoop() != L) || !IndexAddRec->isAffine()) 517 return; 518 519 InductiveRangeCheck IRC; 520 IRC.End = End; 521 IRC.Begin = IndexAddRec->getStart(); 522 IRC.Step = IndexAddRec->getStepRecurrence(SE); 523 IRC.CheckUse = &ConditionUse; 524 Checks.push_back(IRC); 525 } 526 527 void InductiveRangeCheck::extractRangeChecksFromBranch( 528 BranchInst *BI, Loop *L, ScalarEvolution &SE, BranchProbabilityInfo *BPI, 529 SmallVectorImpl<InductiveRangeCheck> &Checks, bool &Changed) { 530 if (BI->isUnconditional() || BI->getParent() == L->getLoopLatch()) 531 return; 532 533 unsigned IndexLoopSucc = L->contains(BI->getSuccessor(0)) ? 0 : 1; 534 assert(L->contains(BI->getSuccessor(IndexLoopSucc)) && 535 "No edges coming to loop?"); 536 BranchProbability LikelyTaken(15, 16); 537 538 if (!SkipProfitabilityChecks && BPI && 539 BPI->getEdgeProbability(BI->getParent(), IndexLoopSucc) < LikelyTaken) 540 return; 541 542 // IRCE expects branch's true edge comes to loop. Invert branch for opposite 543 // case. 544 if (IndexLoopSucc != 0) { 545 IRBuilder<> Builder(BI); 546 InvertBranch(BI, Builder); 547 if (BPI) 548 BPI->swapSuccEdgesProbabilities(BI->getParent()); 549 Changed = true; 550 } 551 552 SmallPtrSet<Value *, 8> Visited; 553 InductiveRangeCheck::extractRangeChecksFromCond(L, SE, BI->getOperandUse(0), 554 Checks, Visited); 555 } 556 557 // Add metadata to the loop L to disable loop optimizations. Callers need to 558 // confirm that optimizing loop L is not beneficial. 559 static void DisableAllLoopOptsOnLoop(Loop &L) { 560 // We do not care about any existing loopID related metadata for L, since we 561 // are setting all loop metadata to false. 562 LLVMContext &Context = L.getHeader()->getContext(); 563 // Reserve first location for self reference to the LoopID metadata node. 564 MDNode *Dummy = MDNode::get(Context, {}); 565 MDNode *DisableUnroll = MDNode::get( 566 Context, {MDString::get(Context, "llvm.loop.unroll.disable")}); 567 Metadata *FalseVal = 568 ConstantAsMetadata::get(ConstantInt::get(Type::getInt1Ty(Context), 0)); 569 MDNode *DisableVectorize = MDNode::get( 570 Context, 571 {MDString::get(Context, "llvm.loop.vectorize.enable"), FalseVal}); 572 MDNode *DisableLICMVersioning = MDNode::get( 573 Context, {MDString::get(Context, "llvm.loop.licm_versioning.disable")}); 574 MDNode *DisableDistribution= MDNode::get( 575 Context, 576 {MDString::get(Context, "llvm.loop.distribute.enable"), FalseVal}); 577 MDNode *NewLoopID = 578 MDNode::get(Context, {Dummy, DisableUnroll, DisableVectorize, 579 DisableLICMVersioning, DisableDistribution}); 580 // Set operand 0 to refer to the loop id itself. 581 NewLoopID->replaceOperandWith(0, NewLoopID); 582 L.setLoopID(NewLoopID); 583 } 584 585 namespace { 586 587 // Keeps track of the structure of a loop. This is similar to llvm::Loop, 588 // except that it is more lightweight and can track the state of a loop through 589 // changing and potentially invalid IR. This structure also formalizes the 590 // kinds of loops we can deal with -- ones that have a single latch that is also 591 // an exiting block *and* have a canonical induction variable. 592 struct LoopStructure { 593 const char *Tag = ""; 594 595 BasicBlock *Header = nullptr; 596 BasicBlock *Latch = nullptr; 597 598 // `Latch's terminator instruction is `LatchBr', and it's `LatchBrExitIdx'th 599 // successor is `LatchExit', the exit block of the loop. 600 BranchInst *LatchBr = nullptr; 601 BasicBlock *LatchExit = nullptr; 602 unsigned LatchBrExitIdx = std::numeric_limits<unsigned>::max(); 603 604 // The loop represented by this instance of LoopStructure is semantically 605 // equivalent to: 606 // 607 // intN_ty inc = IndVarIncreasing ? 1 : -1; 608 // pred_ty predicate = IndVarIncreasing ? ICMP_SLT : ICMP_SGT; 609 // 610 // for (intN_ty iv = IndVarStart; predicate(iv, LoopExitAt); iv = IndVarBase) 611 // ... body ... 612 613 Value *IndVarBase = nullptr; 614 Value *IndVarStart = nullptr; 615 Value *IndVarStep = nullptr; 616 Value *LoopExitAt = nullptr; 617 bool IndVarIncreasing = false; 618 bool IsSignedPredicate = true; 619 620 LoopStructure() = default; 621 622 template <typename M> LoopStructure map(M Map) const { 623 LoopStructure Result; 624 Result.Tag = Tag; 625 Result.Header = cast<BasicBlock>(Map(Header)); 626 Result.Latch = cast<BasicBlock>(Map(Latch)); 627 Result.LatchBr = cast<BranchInst>(Map(LatchBr)); 628 Result.LatchExit = cast<BasicBlock>(Map(LatchExit)); 629 Result.LatchBrExitIdx = LatchBrExitIdx; 630 Result.IndVarBase = Map(IndVarBase); 631 Result.IndVarStart = Map(IndVarStart); 632 Result.IndVarStep = Map(IndVarStep); 633 Result.LoopExitAt = Map(LoopExitAt); 634 Result.IndVarIncreasing = IndVarIncreasing; 635 Result.IsSignedPredicate = IsSignedPredicate; 636 return Result; 637 } 638 639 static std::optional<LoopStructure> parseLoopStructure(ScalarEvolution &, 640 Loop &, const char *&); 641 }; 642 643 /// This class is used to constrain loops to run within a given iteration space. 644 /// The algorithm this class implements is given a Loop and a range [Begin, 645 /// End). The algorithm then tries to break out a "main loop" out of the loop 646 /// it is given in a way that the "main loop" runs with the induction variable 647 /// in a subset of [Begin, End). The algorithm emits appropriate pre and post 648 /// loops to run any remaining iterations. The pre loop runs any iterations in 649 /// which the induction variable is < Begin, and the post loop runs any 650 /// iterations in which the induction variable is >= End. 651 class LoopConstrainer { 652 // The representation of a clone of the original loop we started out with. 653 struct ClonedLoop { 654 // The cloned blocks 655 std::vector<BasicBlock *> Blocks; 656 657 // `Map` maps values in the clonee into values in the cloned version 658 ValueToValueMapTy Map; 659 660 // An instance of `LoopStructure` for the cloned loop 661 LoopStructure Structure; 662 }; 663 664 // Result of rewriting the range of a loop. See changeIterationSpaceEnd for 665 // more details on what these fields mean. 666 struct RewrittenRangeInfo { 667 BasicBlock *PseudoExit = nullptr; 668 BasicBlock *ExitSelector = nullptr; 669 std::vector<PHINode *> PHIValuesAtPseudoExit; 670 PHINode *IndVarEnd = nullptr; 671 672 RewrittenRangeInfo() = default; 673 }; 674 675 // Calculated subranges we restrict the iteration space of the main loop to. 676 // See the implementation of `calculateSubRanges' for more details on how 677 // these fields are computed. `LowLimit` is std::nullopt if there is no 678 // restriction on low end of the restricted iteration space of the main loop. 679 // `HighLimit` is std::nullopt if there is no restriction on high end of the 680 // restricted iteration space of the main loop. 681 682 struct SubRanges { 683 std::optional<const SCEV *> LowLimit; 684 std::optional<const SCEV *> HighLimit; 685 }; 686 687 // Compute a safe set of limits for the main loop to run in -- effectively the 688 // intersection of `Range' and the iteration space of the original loop. 689 // Return std::nullopt if unable to compute the set of subranges. 690 std::optional<SubRanges> calculateSubRanges(bool IsSignedPredicate) const; 691 692 // Clone `OriginalLoop' and return the result in CLResult. The IR after 693 // running `cloneLoop' is well formed except for the PHI nodes in CLResult -- 694 // the PHI nodes say that there is an incoming edge from `OriginalPreheader` 695 // but there is no such edge. 696 void cloneLoop(ClonedLoop &CLResult, const char *Tag) const; 697 698 // Create the appropriate loop structure needed to describe a cloned copy of 699 // `Original`. The clone is described by `VM`. 700 Loop *createClonedLoopStructure(Loop *Original, Loop *Parent, 701 ValueToValueMapTy &VM, bool IsSubloop); 702 703 // Rewrite the iteration space of the loop denoted by (LS, Preheader). The 704 // iteration space of the rewritten loop ends at ExitLoopAt. The start of the 705 // iteration space is not changed. `ExitLoopAt' is assumed to be slt 706 // `OriginalHeaderCount'. 707 // 708 // If there are iterations left to execute, control is made to jump to 709 // `ContinuationBlock', otherwise they take the normal loop exit. The 710 // returned `RewrittenRangeInfo' object is populated as follows: 711 // 712 // .PseudoExit is a basic block that unconditionally branches to 713 // `ContinuationBlock'. 714 // 715 // .ExitSelector is a basic block that decides, on exit from the loop, 716 // whether to branch to the "true" exit or to `PseudoExit'. 717 // 718 // .PHIValuesAtPseudoExit are PHINodes in `PseudoExit' that compute the value 719 // for each PHINode in the loop header on taking the pseudo exit. 720 // 721 // After changeIterationSpaceEnd, `Preheader' is no longer a legitimate 722 // preheader because it is made to branch to the loop header only 723 // conditionally. 724 RewrittenRangeInfo 725 changeIterationSpaceEnd(const LoopStructure &LS, BasicBlock *Preheader, 726 Value *ExitLoopAt, 727 BasicBlock *ContinuationBlock) const; 728 729 // The loop denoted by `LS' has `OldPreheader' as its preheader. This 730 // function creates a new preheader for `LS' and returns it. 731 BasicBlock *createPreheader(const LoopStructure &LS, BasicBlock *OldPreheader, 732 const char *Tag) const; 733 734 // `ContinuationBlockAndPreheader' was the continuation block for some call to 735 // `changeIterationSpaceEnd' and is the preheader to the loop denoted by `LS'. 736 // This function rewrites the PHI nodes in `LS.Header' to start with the 737 // correct value. 738 void rewriteIncomingValuesForPHIs( 739 LoopStructure &LS, BasicBlock *ContinuationBlockAndPreheader, 740 const LoopConstrainer::RewrittenRangeInfo &RRI) const; 741 742 // Even though we do not preserve any passes at this time, we at least need to 743 // keep the parent loop structure consistent. The `LPPassManager' seems to 744 // verify this after running a loop pass. This function adds the list of 745 // blocks denoted by BBs to this loops parent loop if required. 746 void addToParentLoopIfNeeded(ArrayRef<BasicBlock *> BBs); 747 748 // Some global state. 749 Function &F; 750 LLVMContext &Ctx; 751 ScalarEvolution &SE; 752 DominatorTree &DT; 753 LoopInfo &LI; 754 function_ref<void(Loop *, bool)> LPMAddNewLoop; 755 756 // Information about the original loop we started out with. 757 Loop &OriginalLoop; 758 759 const IntegerType *ExitCountTy = nullptr; 760 BasicBlock *OriginalPreheader = nullptr; 761 762 // The preheader of the main loop. This may or may not be different from 763 // `OriginalPreheader'. 764 BasicBlock *MainLoopPreheader = nullptr; 765 766 // The range we need to run the main loop in. 767 InductiveRangeCheck::Range Range; 768 769 // The structure of the main loop (see comment at the beginning of this class 770 // for a definition) 771 LoopStructure MainLoopStructure; 772 773 public: 774 LoopConstrainer(Loop &L, LoopInfo &LI, 775 function_ref<void(Loop *, bool)> LPMAddNewLoop, 776 const LoopStructure &LS, ScalarEvolution &SE, 777 DominatorTree &DT, InductiveRangeCheck::Range R) 778 : F(*L.getHeader()->getParent()), Ctx(L.getHeader()->getContext()), 779 SE(SE), DT(DT), LI(LI), LPMAddNewLoop(LPMAddNewLoop), OriginalLoop(L), 780 Range(R), MainLoopStructure(LS) {} 781 782 // Entry point for the algorithm. Returns true on success. 783 bool run(); 784 }; 785 786 } // end anonymous namespace 787 788 /// Given a loop with an deccreasing induction variable, is it possible to 789 /// safely calculate the bounds of a new loop using the given Predicate. 790 static bool isSafeDecreasingBound(const SCEV *Start, 791 const SCEV *BoundSCEV, const SCEV *Step, 792 ICmpInst::Predicate Pred, 793 unsigned LatchBrExitIdx, 794 Loop *L, ScalarEvolution &SE) { 795 if (Pred != ICmpInst::ICMP_SLT && Pred != ICmpInst::ICMP_SGT && 796 Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_UGT) 797 return false; 798 799 if (!SE.isAvailableAtLoopEntry(BoundSCEV, L)) 800 return false; 801 802 assert(SE.isKnownNegative(Step) && "expecting negative step"); 803 804 LLVM_DEBUG(dbgs() << "irce: isSafeDecreasingBound with:\n"); 805 LLVM_DEBUG(dbgs() << "irce: Start: " << *Start << "\n"); 806 LLVM_DEBUG(dbgs() << "irce: Step: " << *Step << "\n"); 807 LLVM_DEBUG(dbgs() << "irce: BoundSCEV: " << *BoundSCEV << "\n"); 808 LLVM_DEBUG(dbgs() << "irce: Pred: " << Pred << "\n"); 809 LLVM_DEBUG(dbgs() << "irce: LatchExitBrIdx: " << LatchBrExitIdx << "\n"); 810 811 bool IsSigned = ICmpInst::isSigned(Pred); 812 // The predicate that we need to check that the induction variable lies 813 // within bounds. 814 ICmpInst::Predicate BoundPred = 815 IsSigned ? CmpInst::ICMP_SGT : CmpInst::ICMP_UGT; 816 817 if (LatchBrExitIdx == 1) 818 return SE.isLoopEntryGuardedByCond(L, BoundPred, Start, BoundSCEV); 819 820 assert(LatchBrExitIdx == 0 && 821 "LatchBrExitIdx should be either 0 or 1"); 822 823 const SCEV *StepPlusOne = SE.getAddExpr(Step, SE.getOne(Step->getType())); 824 unsigned BitWidth = cast<IntegerType>(BoundSCEV->getType())->getBitWidth(); 825 APInt Min = IsSigned ? APInt::getSignedMinValue(BitWidth) : 826 APInt::getMinValue(BitWidth); 827 const SCEV *Limit = SE.getMinusSCEV(SE.getConstant(Min), StepPlusOne); 828 829 const SCEV *MinusOne = 830 SE.getMinusSCEV(BoundSCEV, SE.getOne(BoundSCEV->getType())); 831 832 return SE.isLoopEntryGuardedByCond(L, BoundPred, Start, MinusOne) && 833 SE.isLoopEntryGuardedByCond(L, BoundPred, BoundSCEV, Limit); 834 835 } 836 837 /// Given a loop with an increasing induction variable, is it possible to 838 /// safely calculate the bounds of a new loop using the given Predicate. 839 static bool isSafeIncreasingBound(const SCEV *Start, 840 const SCEV *BoundSCEV, const SCEV *Step, 841 ICmpInst::Predicate Pred, 842 unsigned LatchBrExitIdx, 843 Loop *L, ScalarEvolution &SE) { 844 if (Pred != ICmpInst::ICMP_SLT && Pred != ICmpInst::ICMP_SGT && 845 Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_UGT) 846 return false; 847 848 if (!SE.isAvailableAtLoopEntry(BoundSCEV, L)) 849 return false; 850 851 LLVM_DEBUG(dbgs() << "irce: isSafeIncreasingBound with:\n"); 852 LLVM_DEBUG(dbgs() << "irce: Start: " << *Start << "\n"); 853 LLVM_DEBUG(dbgs() << "irce: Step: " << *Step << "\n"); 854 LLVM_DEBUG(dbgs() << "irce: BoundSCEV: " << *BoundSCEV << "\n"); 855 LLVM_DEBUG(dbgs() << "irce: Pred: " << Pred << "\n"); 856 LLVM_DEBUG(dbgs() << "irce: LatchExitBrIdx: " << LatchBrExitIdx << "\n"); 857 858 bool IsSigned = ICmpInst::isSigned(Pred); 859 // The predicate that we need to check that the induction variable lies 860 // within bounds. 861 ICmpInst::Predicate BoundPred = 862 IsSigned ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT; 863 864 if (LatchBrExitIdx == 1) 865 return SE.isLoopEntryGuardedByCond(L, BoundPred, Start, BoundSCEV); 866 867 assert(LatchBrExitIdx == 0 && "LatchBrExitIdx should be 0 or 1"); 868 869 const SCEV *StepMinusOne = 870 SE.getMinusSCEV(Step, SE.getOne(Step->getType())); 871 unsigned BitWidth = cast<IntegerType>(BoundSCEV->getType())->getBitWidth(); 872 APInt Max = IsSigned ? APInt::getSignedMaxValue(BitWidth) : 873 APInt::getMaxValue(BitWidth); 874 const SCEV *Limit = SE.getMinusSCEV(SE.getConstant(Max), StepMinusOne); 875 876 return (SE.isLoopEntryGuardedByCond(L, BoundPred, Start, 877 SE.getAddExpr(BoundSCEV, Step)) && 878 SE.isLoopEntryGuardedByCond(L, BoundPred, BoundSCEV, Limit)); 879 } 880 881 /// Returns estimate for max latch taken count of the loop of the narrowest 882 /// available type. If the latch block has such estimate, it is returned. 883 /// Otherwise, we use max exit count of whole loop (that is potentially of wider 884 /// type than latch check itself), which is still better than no estimate. 885 static const SCEV *getNarrowestLatchMaxTakenCountEstimate(ScalarEvolution &SE, 886 const Loop &L) { 887 const SCEV *FromBlock = 888 SE.getExitCount(&L, L.getLoopLatch(), ScalarEvolution::SymbolicMaximum); 889 if (isa<SCEVCouldNotCompute>(FromBlock)) 890 return SE.getSymbolicMaxBackedgeTakenCount(&L); 891 return FromBlock; 892 } 893 894 std::optional<LoopStructure> 895 LoopStructure::parseLoopStructure(ScalarEvolution &SE, Loop &L, 896 const char *&FailureReason) { 897 if (!L.isLoopSimplifyForm()) { 898 FailureReason = "loop not in LoopSimplify form"; 899 return std::nullopt; 900 } 901 902 BasicBlock *Latch = L.getLoopLatch(); 903 assert(Latch && "Simplified loops only have one latch!"); 904 905 if (Latch->getTerminator()->getMetadata(ClonedLoopTag)) { 906 FailureReason = "loop has already been cloned"; 907 return std::nullopt; 908 } 909 910 if (!L.isLoopExiting(Latch)) { 911 FailureReason = "no loop latch"; 912 return std::nullopt; 913 } 914 915 BasicBlock *Header = L.getHeader(); 916 BasicBlock *Preheader = L.getLoopPreheader(); 917 if (!Preheader) { 918 FailureReason = "no preheader"; 919 return std::nullopt; 920 } 921 922 BranchInst *LatchBr = dyn_cast<BranchInst>(Latch->getTerminator()); 923 if (!LatchBr || LatchBr->isUnconditional()) { 924 FailureReason = "latch terminator not conditional branch"; 925 return std::nullopt; 926 } 927 928 unsigned LatchBrExitIdx = LatchBr->getSuccessor(0) == Header ? 1 : 0; 929 930 ICmpInst *ICI = dyn_cast<ICmpInst>(LatchBr->getCondition()); 931 if (!ICI || !isa<IntegerType>(ICI->getOperand(0)->getType())) { 932 FailureReason = "latch terminator branch not conditional on integral icmp"; 933 return std::nullopt; 934 } 935 936 const SCEV *MaxBETakenCount = getNarrowestLatchMaxTakenCountEstimate(SE, L); 937 if (isa<SCEVCouldNotCompute>(MaxBETakenCount)) { 938 FailureReason = "could not compute latch count"; 939 return std::nullopt; 940 } 941 assert(SE.getLoopDisposition(MaxBETakenCount, &L) == 942 ScalarEvolution::LoopInvariant && 943 "loop variant exit count doesn't make sense!"); 944 945 ICmpInst::Predicate Pred = ICI->getPredicate(); 946 Value *LeftValue = ICI->getOperand(0); 947 const SCEV *LeftSCEV = SE.getSCEV(LeftValue); 948 IntegerType *IndVarTy = cast<IntegerType>(LeftValue->getType()); 949 950 Value *RightValue = ICI->getOperand(1); 951 const SCEV *RightSCEV = SE.getSCEV(RightValue); 952 953 // We canonicalize `ICI` such that `LeftSCEV` is an add recurrence. 954 if (!isa<SCEVAddRecExpr>(LeftSCEV)) { 955 if (isa<SCEVAddRecExpr>(RightSCEV)) { 956 std::swap(LeftSCEV, RightSCEV); 957 std::swap(LeftValue, RightValue); 958 Pred = ICmpInst::getSwappedPredicate(Pred); 959 } else { 960 FailureReason = "no add recurrences in the icmp"; 961 return std::nullopt; 962 } 963 } 964 965 auto HasNoSignedWrap = [&](const SCEVAddRecExpr *AR) { 966 if (AR->getNoWrapFlags(SCEV::FlagNSW)) 967 return true; 968 969 IntegerType *Ty = cast<IntegerType>(AR->getType()); 970 IntegerType *WideTy = 971 IntegerType::get(Ty->getContext(), Ty->getBitWidth() * 2); 972 973 const SCEVAddRecExpr *ExtendAfterOp = 974 dyn_cast<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy)); 975 if (ExtendAfterOp) { 976 const SCEV *ExtendedStart = SE.getSignExtendExpr(AR->getStart(), WideTy); 977 const SCEV *ExtendedStep = 978 SE.getSignExtendExpr(AR->getStepRecurrence(SE), WideTy); 979 980 bool NoSignedWrap = ExtendAfterOp->getStart() == ExtendedStart && 981 ExtendAfterOp->getStepRecurrence(SE) == ExtendedStep; 982 983 if (NoSignedWrap) 984 return true; 985 } 986 987 // We may have proved this when computing the sign extension above. 988 return AR->getNoWrapFlags(SCEV::FlagNSW) != SCEV::FlagAnyWrap; 989 }; 990 991 // `ICI` is interpreted as taking the backedge if the *next* value of the 992 // induction variable satisfies some constraint. 993 994 const SCEVAddRecExpr *IndVarBase = cast<SCEVAddRecExpr>(LeftSCEV); 995 if (IndVarBase->getLoop() != &L) { 996 FailureReason = "LHS in cmp is not an AddRec for this loop"; 997 return std::nullopt; 998 } 999 if (!IndVarBase->isAffine()) { 1000 FailureReason = "LHS in icmp not induction variable"; 1001 return std::nullopt; 1002 } 1003 const SCEV* StepRec = IndVarBase->getStepRecurrence(SE); 1004 if (!isa<SCEVConstant>(StepRec)) { 1005 FailureReason = "LHS in icmp not induction variable"; 1006 return std::nullopt; 1007 } 1008 ConstantInt *StepCI = cast<SCEVConstant>(StepRec)->getValue(); 1009 1010 if (ICI->isEquality() && !HasNoSignedWrap(IndVarBase)) { 1011 FailureReason = "LHS in icmp needs nsw for equality predicates"; 1012 return std::nullopt; 1013 } 1014 1015 assert(!StepCI->isZero() && "Zero step?"); 1016 bool IsIncreasing = !StepCI->isNegative(); 1017 bool IsSignedPredicate; 1018 const SCEV *StartNext = IndVarBase->getStart(); 1019 const SCEV *Addend = SE.getNegativeSCEV(IndVarBase->getStepRecurrence(SE)); 1020 const SCEV *IndVarStart = SE.getAddExpr(StartNext, Addend); 1021 const SCEV *Step = SE.getSCEV(StepCI); 1022 1023 const SCEV *FixedRightSCEV = nullptr; 1024 1025 // If RightValue resides within loop (but still being loop invariant), 1026 // regenerate it as preheader. 1027 if (auto *I = dyn_cast<Instruction>(RightValue)) 1028 if (L.contains(I->getParent())) 1029 FixedRightSCEV = RightSCEV; 1030 1031 if (IsIncreasing) { 1032 bool DecreasedRightValueByOne = false; 1033 if (StepCI->isOne()) { 1034 // Try to turn eq/ne predicates to those we can work with. 1035 if (Pred == ICmpInst::ICMP_NE && LatchBrExitIdx == 1) 1036 // while (++i != len) { while (++i < len) { 1037 // ... ---> ... 1038 // } } 1039 // If both parts are known non-negative, it is profitable to use 1040 // unsigned comparison in increasing loop. This allows us to make the 1041 // comparison check against "RightSCEV + 1" more optimistic. 1042 if (isKnownNonNegativeInLoop(IndVarStart, &L, SE) && 1043 isKnownNonNegativeInLoop(RightSCEV, &L, SE)) 1044 Pred = ICmpInst::ICMP_ULT; 1045 else 1046 Pred = ICmpInst::ICMP_SLT; 1047 else if (Pred == ICmpInst::ICMP_EQ && LatchBrExitIdx == 0) { 1048 // while (true) { while (true) { 1049 // if (++i == len) ---> if (++i > len - 1) 1050 // break; break; 1051 // ... ... 1052 // } } 1053 if (IndVarBase->getNoWrapFlags(SCEV::FlagNUW) && 1054 cannotBeMinInLoop(RightSCEV, &L, SE, /*Signed*/false)) { 1055 Pred = ICmpInst::ICMP_UGT; 1056 RightSCEV = SE.getMinusSCEV(RightSCEV, 1057 SE.getOne(RightSCEV->getType())); 1058 DecreasedRightValueByOne = true; 1059 } else if (cannotBeMinInLoop(RightSCEV, &L, SE, /*Signed*/true)) { 1060 Pred = ICmpInst::ICMP_SGT; 1061 RightSCEV = SE.getMinusSCEV(RightSCEV, 1062 SE.getOne(RightSCEV->getType())); 1063 DecreasedRightValueByOne = true; 1064 } 1065 } 1066 } 1067 1068 bool LTPred = (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_ULT); 1069 bool GTPred = (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_UGT); 1070 bool FoundExpectedPred = 1071 (LTPred && LatchBrExitIdx == 1) || (GTPred && LatchBrExitIdx == 0); 1072 1073 if (!FoundExpectedPred) { 1074 FailureReason = "expected icmp slt semantically, found something else"; 1075 return std::nullopt; 1076 } 1077 1078 IsSignedPredicate = ICmpInst::isSigned(Pred); 1079 if (!IsSignedPredicate && !AllowUnsignedLatchCondition) { 1080 FailureReason = "unsigned latch conditions are explicitly prohibited"; 1081 return std::nullopt; 1082 } 1083 1084 if (!isSafeIncreasingBound(IndVarStart, RightSCEV, Step, Pred, 1085 LatchBrExitIdx, &L, SE)) { 1086 FailureReason = "Unsafe loop bounds"; 1087 return std::nullopt; 1088 } 1089 if (LatchBrExitIdx == 0) { 1090 // We need to increase the right value unless we have already decreased 1091 // it virtually when we replaced EQ with SGT. 1092 if (!DecreasedRightValueByOne) 1093 FixedRightSCEV = 1094 SE.getAddExpr(RightSCEV, SE.getOne(RightSCEV->getType())); 1095 } else { 1096 assert(!DecreasedRightValueByOne && 1097 "Right value can be decreased only for LatchBrExitIdx == 0!"); 1098 } 1099 } else { 1100 bool IncreasedRightValueByOne = false; 1101 if (StepCI->isMinusOne()) { 1102 // Try to turn eq/ne predicates to those we can work with. 1103 if (Pred == ICmpInst::ICMP_NE && LatchBrExitIdx == 1) 1104 // while (--i != len) { while (--i > len) { 1105 // ... ---> ... 1106 // } } 1107 // We intentionally don't turn the predicate into UGT even if we know 1108 // that both operands are non-negative, because it will only pessimize 1109 // our check against "RightSCEV - 1". 1110 Pred = ICmpInst::ICMP_SGT; 1111 else if (Pred == ICmpInst::ICMP_EQ && LatchBrExitIdx == 0) { 1112 // while (true) { while (true) { 1113 // if (--i == len) ---> if (--i < len + 1) 1114 // break; break; 1115 // ... ... 1116 // } } 1117 if (IndVarBase->getNoWrapFlags(SCEV::FlagNUW) && 1118 cannotBeMaxInLoop(RightSCEV, &L, SE, /* Signed */ false)) { 1119 Pred = ICmpInst::ICMP_ULT; 1120 RightSCEV = SE.getAddExpr(RightSCEV, SE.getOne(RightSCEV->getType())); 1121 IncreasedRightValueByOne = true; 1122 } else if (cannotBeMaxInLoop(RightSCEV, &L, SE, /* Signed */ true)) { 1123 Pred = ICmpInst::ICMP_SLT; 1124 RightSCEV = SE.getAddExpr(RightSCEV, SE.getOne(RightSCEV->getType())); 1125 IncreasedRightValueByOne = true; 1126 } 1127 } 1128 } 1129 1130 bool LTPred = (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_ULT); 1131 bool GTPred = (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_UGT); 1132 1133 bool FoundExpectedPred = 1134 (GTPred && LatchBrExitIdx == 1) || (LTPred && LatchBrExitIdx == 0); 1135 1136 if (!FoundExpectedPred) { 1137 FailureReason = "expected icmp sgt semantically, found something else"; 1138 return std::nullopt; 1139 } 1140 1141 IsSignedPredicate = 1142 Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGT; 1143 1144 if (!IsSignedPredicate && !AllowUnsignedLatchCondition) { 1145 FailureReason = "unsigned latch conditions are explicitly prohibited"; 1146 return std::nullopt; 1147 } 1148 1149 if (!isSafeDecreasingBound(IndVarStart, RightSCEV, Step, Pred, 1150 LatchBrExitIdx, &L, SE)) { 1151 FailureReason = "Unsafe bounds"; 1152 return std::nullopt; 1153 } 1154 1155 if (LatchBrExitIdx == 0) { 1156 // We need to decrease the right value unless we have already increased 1157 // it virtually when we replaced EQ with SLT. 1158 if (!IncreasedRightValueByOne) 1159 FixedRightSCEV = 1160 SE.getMinusSCEV(RightSCEV, SE.getOne(RightSCEV->getType())); 1161 } else { 1162 assert(!IncreasedRightValueByOne && 1163 "Right value can be increased only for LatchBrExitIdx == 0!"); 1164 } 1165 } 1166 BasicBlock *LatchExit = LatchBr->getSuccessor(LatchBrExitIdx); 1167 1168 assert(!L.contains(LatchExit) && "expected an exit block!"); 1169 const DataLayout &DL = Preheader->getModule()->getDataLayout(); 1170 SCEVExpander Expander(SE, DL, "irce"); 1171 Instruction *Ins = Preheader->getTerminator(); 1172 1173 if (FixedRightSCEV) 1174 RightValue = 1175 Expander.expandCodeFor(FixedRightSCEV, FixedRightSCEV->getType(), Ins); 1176 1177 Value *IndVarStartV = Expander.expandCodeFor(IndVarStart, IndVarTy, Ins); 1178 IndVarStartV->setName("indvar.start"); 1179 1180 LoopStructure Result; 1181 1182 Result.Tag = "main"; 1183 Result.Header = Header; 1184 Result.Latch = Latch; 1185 Result.LatchBr = LatchBr; 1186 Result.LatchExit = LatchExit; 1187 Result.LatchBrExitIdx = LatchBrExitIdx; 1188 Result.IndVarStart = IndVarStartV; 1189 Result.IndVarStep = StepCI; 1190 Result.IndVarBase = LeftValue; 1191 Result.IndVarIncreasing = IsIncreasing; 1192 Result.LoopExitAt = RightValue; 1193 Result.IsSignedPredicate = IsSignedPredicate; 1194 1195 FailureReason = nullptr; 1196 1197 return Result; 1198 } 1199 1200 /// If the type of \p S matches with \p Ty, return \p S. Otherwise, return 1201 /// signed or unsigned extension of \p S to type \p Ty. 1202 static const SCEV *NoopOrExtend(const SCEV *S, Type *Ty, ScalarEvolution &SE, 1203 bool Signed) { 1204 return Signed ? SE.getNoopOrSignExtend(S, Ty) : SE.getNoopOrZeroExtend(S, Ty); 1205 } 1206 1207 std::optional<LoopConstrainer::SubRanges> 1208 LoopConstrainer::calculateSubRanges(bool IsSignedPredicate) const { 1209 auto *RTy = cast<IntegerType>(Range.getType()); 1210 // We only support wide range checks and narrow latches. 1211 if (!AllowNarrowLatchCondition && RTy != ExitCountTy) 1212 return std::nullopt; 1213 if (RTy->getBitWidth() < ExitCountTy->getBitWidth()) 1214 return std::nullopt; 1215 1216 LoopConstrainer::SubRanges Result; 1217 1218 // I think we can be more aggressive here and make this nuw / nsw if the 1219 // addition that feeds into the icmp for the latch's terminating branch is nuw 1220 // / nsw. In any case, a wrapping 2's complement addition is safe. 1221 const SCEV *Start = NoopOrExtend(SE.getSCEV(MainLoopStructure.IndVarStart), 1222 RTy, SE, IsSignedPredicate); 1223 const SCEV *End = NoopOrExtend(SE.getSCEV(MainLoopStructure.LoopExitAt), RTy, 1224 SE, IsSignedPredicate); 1225 1226 bool Increasing = MainLoopStructure.IndVarIncreasing; 1227 1228 // We compute `Smallest` and `Greatest` such that [Smallest, Greatest), or 1229 // [Smallest, GreatestSeen] is the range of values the induction variable 1230 // takes. 1231 1232 const SCEV *Smallest = nullptr, *Greatest = nullptr, *GreatestSeen = nullptr; 1233 1234 const SCEV *One = SE.getOne(RTy); 1235 if (Increasing) { 1236 Smallest = Start; 1237 Greatest = End; 1238 // No overflow, because the range [Smallest, GreatestSeen] is not empty. 1239 GreatestSeen = SE.getMinusSCEV(End, One); 1240 } else { 1241 // These two computations may sign-overflow. Here is why that is okay: 1242 // 1243 // We know that the induction variable does not sign-overflow on any 1244 // iteration except the last one, and it starts at `Start` and ends at 1245 // `End`, decrementing by one every time. 1246 // 1247 // * if `Smallest` sign-overflows we know `End` is `INT_SMAX`. Since the 1248 // induction variable is decreasing we know that that the smallest value 1249 // the loop body is actually executed with is `INT_SMIN` == `Smallest`. 1250 // 1251 // * if `Greatest` sign-overflows, we know it can only be `INT_SMIN`. In 1252 // that case, `Clamp` will always return `Smallest` and 1253 // [`Result.LowLimit`, `Result.HighLimit`) = [`Smallest`, `Smallest`) 1254 // will be an empty range. Returning an empty range is always safe. 1255 1256 Smallest = SE.getAddExpr(End, One); 1257 Greatest = SE.getAddExpr(Start, One); 1258 GreatestSeen = Start; 1259 } 1260 1261 auto Clamp = [this, Smallest, Greatest, IsSignedPredicate](const SCEV *S) { 1262 return IsSignedPredicate 1263 ? SE.getSMaxExpr(Smallest, SE.getSMinExpr(Greatest, S)) 1264 : SE.getUMaxExpr(Smallest, SE.getUMinExpr(Greatest, S)); 1265 }; 1266 1267 // In some cases we can prove that we don't need a pre or post loop. 1268 ICmpInst::Predicate PredLE = 1269 IsSignedPredicate ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1270 ICmpInst::Predicate PredLT = 1271 IsSignedPredicate ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1272 1273 bool ProvablyNoPreloop = 1274 SE.isKnownPredicate(PredLE, Range.getBegin(), Smallest); 1275 if (!ProvablyNoPreloop) 1276 Result.LowLimit = Clamp(Range.getBegin()); 1277 1278 bool ProvablyNoPostLoop = 1279 SE.isKnownPredicate(PredLT, GreatestSeen, Range.getEnd()); 1280 if (!ProvablyNoPostLoop) 1281 Result.HighLimit = Clamp(Range.getEnd()); 1282 1283 return Result; 1284 } 1285 1286 void LoopConstrainer::cloneLoop(LoopConstrainer::ClonedLoop &Result, 1287 const char *Tag) const { 1288 for (BasicBlock *BB : OriginalLoop.getBlocks()) { 1289 BasicBlock *Clone = CloneBasicBlock(BB, Result.Map, Twine(".") + Tag, &F); 1290 Result.Blocks.push_back(Clone); 1291 Result.Map[BB] = Clone; 1292 } 1293 1294 auto GetClonedValue = [&Result](Value *V) { 1295 assert(V && "null values not in domain!"); 1296 auto It = Result.Map.find(V); 1297 if (It == Result.Map.end()) 1298 return V; 1299 return static_cast<Value *>(It->second); 1300 }; 1301 1302 auto *ClonedLatch = 1303 cast<BasicBlock>(GetClonedValue(OriginalLoop.getLoopLatch())); 1304 ClonedLatch->getTerminator()->setMetadata(ClonedLoopTag, 1305 MDNode::get(Ctx, {})); 1306 1307 Result.Structure = MainLoopStructure.map(GetClonedValue); 1308 Result.Structure.Tag = Tag; 1309 1310 for (unsigned i = 0, e = Result.Blocks.size(); i != e; ++i) { 1311 BasicBlock *ClonedBB = Result.Blocks[i]; 1312 BasicBlock *OriginalBB = OriginalLoop.getBlocks()[i]; 1313 1314 assert(Result.Map[OriginalBB] == ClonedBB && "invariant!"); 1315 1316 for (Instruction &I : *ClonedBB) 1317 RemapInstruction(&I, Result.Map, 1318 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals); 1319 1320 // Exit blocks will now have one more predecessor and their PHI nodes need 1321 // to be edited to reflect that. No phi nodes need to be introduced because 1322 // the loop is in LCSSA. 1323 1324 for (auto *SBB : successors(OriginalBB)) { 1325 if (OriginalLoop.contains(SBB)) 1326 continue; // not an exit block 1327 1328 for (PHINode &PN : SBB->phis()) { 1329 Value *OldIncoming = PN.getIncomingValueForBlock(OriginalBB); 1330 PN.addIncoming(GetClonedValue(OldIncoming), ClonedBB); 1331 SE.forgetValue(&PN); 1332 } 1333 } 1334 } 1335 } 1336 1337 LoopConstrainer::RewrittenRangeInfo LoopConstrainer::changeIterationSpaceEnd( 1338 const LoopStructure &LS, BasicBlock *Preheader, Value *ExitSubloopAt, 1339 BasicBlock *ContinuationBlock) const { 1340 // We start with a loop with a single latch: 1341 // 1342 // +--------------------+ 1343 // | | 1344 // | preheader | 1345 // | | 1346 // +--------+-----------+ 1347 // | ----------------\ 1348 // | / | 1349 // +--------v----v------+ | 1350 // | | | 1351 // | header | | 1352 // | | | 1353 // +--------------------+ | 1354 // | 1355 // ..... | 1356 // | 1357 // +--------------------+ | 1358 // | | | 1359 // | latch >----------/ 1360 // | | 1361 // +-------v------------+ 1362 // | 1363 // | 1364 // | +--------------------+ 1365 // | | | 1366 // +---> original exit | 1367 // | | 1368 // +--------------------+ 1369 // 1370 // We change the control flow to look like 1371 // 1372 // 1373 // +--------------------+ 1374 // | | 1375 // | preheader >-------------------------+ 1376 // | | | 1377 // +--------v-----------+ | 1378 // | /-------------+ | 1379 // | / | | 1380 // +--------v--v--------+ | | 1381 // | | | | 1382 // | header | | +--------+ | 1383 // | | | | | | 1384 // +--------------------+ | | +-----v-----v-----------+ 1385 // | | | | 1386 // | | | .pseudo.exit | 1387 // | | | | 1388 // | | +-----------v-----------+ 1389 // | | | 1390 // ..... | | | 1391 // | | +--------v-------------+ 1392 // +--------------------+ | | | | 1393 // | | | | | ContinuationBlock | 1394 // | latch >------+ | | | 1395 // | | | +----------------------+ 1396 // +---------v----------+ | 1397 // | | 1398 // | | 1399 // | +---------------^-----+ 1400 // | | | 1401 // +-----> .exit.selector | 1402 // | | 1403 // +----------v----------+ 1404 // | 1405 // +--------------------+ | 1406 // | | | 1407 // | original exit <----+ 1408 // | | 1409 // +--------------------+ 1410 1411 RewrittenRangeInfo RRI; 1412 1413 BasicBlock *BBInsertLocation = LS.Latch->getNextNode(); 1414 RRI.ExitSelector = BasicBlock::Create(Ctx, Twine(LS.Tag) + ".exit.selector", 1415 &F, BBInsertLocation); 1416 RRI.PseudoExit = BasicBlock::Create(Ctx, Twine(LS.Tag) + ".pseudo.exit", &F, 1417 BBInsertLocation); 1418 1419 BranchInst *PreheaderJump = cast<BranchInst>(Preheader->getTerminator()); 1420 bool Increasing = LS.IndVarIncreasing; 1421 bool IsSignedPredicate = LS.IsSignedPredicate; 1422 1423 IRBuilder<> B(PreheaderJump); 1424 auto *RangeTy = Range.getBegin()->getType(); 1425 auto NoopOrExt = [&](Value *V) { 1426 if (V->getType() == RangeTy) 1427 return V; 1428 return IsSignedPredicate ? B.CreateSExt(V, RangeTy, "wide." + V->getName()) 1429 : B.CreateZExt(V, RangeTy, "wide." + V->getName()); 1430 }; 1431 1432 // EnterLoopCond - is it okay to start executing this `LS'? 1433 Value *EnterLoopCond = nullptr; 1434 auto Pred = 1435 Increasing 1436 ? (IsSignedPredicate ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT) 1437 : (IsSignedPredicate ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT); 1438 Value *IndVarStart = NoopOrExt(LS.IndVarStart); 1439 EnterLoopCond = B.CreateICmp(Pred, IndVarStart, ExitSubloopAt); 1440 1441 B.CreateCondBr(EnterLoopCond, LS.Header, RRI.PseudoExit); 1442 PreheaderJump->eraseFromParent(); 1443 1444 LS.LatchBr->setSuccessor(LS.LatchBrExitIdx, RRI.ExitSelector); 1445 B.SetInsertPoint(LS.LatchBr); 1446 Value *IndVarBase = NoopOrExt(LS.IndVarBase); 1447 Value *TakeBackedgeLoopCond = B.CreateICmp(Pred, IndVarBase, ExitSubloopAt); 1448 1449 Value *CondForBranch = LS.LatchBrExitIdx == 1 1450 ? TakeBackedgeLoopCond 1451 : B.CreateNot(TakeBackedgeLoopCond); 1452 1453 LS.LatchBr->setCondition(CondForBranch); 1454 1455 B.SetInsertPoint(RRI.ExitSelector); 1456 1457 // IterationsLeft - are there any more iterations left, given the original 1458 // upper bound on the induction variable? If not, we branch to the "real" 1459 // exit. 1460 Value *LoopExitAt = NoopOrExt(LS.LoopExitAt); 1461 Value *IterationsLeft = B.CreateICmp(Pred, IndVarBase, LoopExitAt); 1462 B.CreateCondBr(IterationsLeft, RRI.PseudoExit, LS.LatchExit); 1463 1464 BranchInst *BranchToContinuation = 1465 BranchInst::Create(ContinuationBlock, RRI.PseudoExit); 1466 1467 // We emit PHI nodes into `RRI.PseudoExit' that compute the "latest" value of 1468 // each of the PHI nodes in the loop header. This feeds into the initial 1469 // value of the same PHI nodes if/when we continue execution. 1470 for (PHINode &PN : LS.Header->phis()) { 1471 PHINode *NewPHI = PHINode::Create(PN.getType(), 2, PN.getName() + ".copy", 1472 BranchToContinuation); 1473 1474 NewPHI->addIncoming(PN.getIncomingValueForBlock(Preheader), Preheader); 1475 NewPHI->addIncoming(PN.getIncomingValueForBlock(LS.Latch), 1476 RRI.ExitSelector); 1477 RRI.PHIValuesAtPseudoExit.push_back(NewPHI); 1478 } 1479 1480 RRI.IndVarEnd = PHINode::Create(IndVarBase->getType(), 2, "indvar.end", 1481 BranchToContinuation); 1482 RRI.IndVarEnd->addIncoming(IndVarStart, Preheader); 1483 RRI.IndVarEnd->addIncoming(IndVarBase, RRI.ExitSelector); 1484 1485 // The latch exit now has a branch from `RRI.ExitSelector' instead of 1486 // `LS.Latch'. The PHI nodes need to be updated to reflect that. 1487 LS.LatchExit->replacePhiUsesWith(LS.Latch, RRI.ExitSelector); 1488 1489 return RRI; 1490 } 1491 1492 void LoopConstrainer::rewriteIncomingValuesForPHIs( 1493 LoopStructure &LS, BasicBlock *ContinuationBlock, 1494 const LoopConstrainer::RewrittenRangeInfo &RRI) const { 1495 unsigned PHIIndex = 0; 1496 for (PHINode &PN : LS.Header->phis()) 1497 PN.setIncomingValueForBlock(ContinuationBlock, 1498 RRI.PHIValuesAtPseudoExit[PHIIndex++]); 1499 1500 LS.IndVarStart = RRI.IndVarEnd; 1501 } 1502 1503 BasicBlock *LoopConstrainer::createPreheader(const LoopStructure &LS, 1504 BasicBlock *OldPreheader, 1505 const char *Tag) const { 1506 BasicBlock *Preheader = BasicBlock::Create(Ctx, Tag, &F, LS.Header); 1507 BranchInst::Create(LS.Header, Preheader); 1508 1509 LS.Header->replacePhiUsesWith(OldPreheader, Preheader); 1510 1511 return Preheader; 1512 } 1513 1514 void LoopConstrainer::addToParentLoopIfNeeded(ArrayRef<BasicBlock *> BBs) { 1515 Loop *ParentLoop = OriginalLoop.getParentLoop(); 1516 if (!ParentLoop) 1517 return; 1518 1519 for (BasicBlock *BB : BBs) 1520 ParentLoop->addBasicBlockToLoop(BB, LI); 1521 } 1522 1523 Loop *LoopConstrainer::createClonedLoopStructure(Loop *Original, Loop *Parent, 1524 ValueToValueMapTy &VM, 1525 bool IsSubloop) { 1526 Loop &New = *LI.AllocateLoop(); 1527 if (Parent) 1528 Parent->addChildLoop(&New); 1529 else 1530 LI.addTopLevelLoop(&New); 1531 LPMAddNewLoop(&New, IsSubloop); 1532 1533 // Add all of the blocks in Original to the new loop. 1534 for (auto *BB : Original->blocks()) 1535 if (LI.getLoopFor(BB) == Original) 1536 New.addBasicBlockToLoop(cast<BasicBlock>(VM[BB]), LI); 1537 1538 // Add all of the subloops to the new loop. 1539 for (Loop *SubLoop : *Original) 1540 createClonedLoopStructure(SubLoop, &New, VM, /* IsSubloop */ true); 1541 1542 return &New; 1543 } 1544 1545 bool LoopConstrainer::run() { 1546 BasicBlock *Preheader = nullptr; 1547 const SCEV *MaxBETakenCount = 1548 getNarrowestLatchMaxTakenCountEstimate(SE, OriginalLoop); 1549 Preheader = OriginalLoop.getLoopPreheader(); 1550 assert(!isa<SCEVCouldNotCompute>(MaxBETakenCount) && Preheader != nullptr && 1551 "preconditions!"); 1552 ExitCountTy = cast<IntegerType>(MaxBETakenCount->getType()); 1553 1554 OriginalPreheader = Preheader; 1555 MainLoopPreheader = Preheader; 1556 1557 bool IsSignedPredicate = MainLoopStructure.IsSignedPredicate; 1558 std::optional<SubRanges> MaybeSR = calculateSubRanges(IsSignedPredicate); 1559 if (!MaybeSR) { 1560 LLVM_DEBUG(dbgs() << "irce: could not compute subranges\n"); 1561 return false; 1562 } 1563 1564 SubRanges SR = *MaybeSR; 1565 bool Increasing = MainLoopStructure.IndVarIncreasing; 1566 IntegerType *IVTy = 1567 cast<IntegerType>(Range.getBegin()->getType()); 1568 1569 SCEVExpander Expander(SE, F.getParent()->getDataLayout(), "irce"); 1570 Instruction *InsertPt = OriginalPreheader->getTerminator(); 1571 1572 // It would have been better to make `PreLoop' and `PostLoop' 1573 // `std::optional<ClonedLoop>'s, but `ValueToValueMapTy' does not have a copy 1574 // constructor. 1575 ClonedLoop PreLoop, PostLoop; 1576 bool NeedsPreLoop = 1577 Increasing ? SR.LowLimit.has_value() : SR.HighLimit.has_value(); 1578 bool NeedsPostLoop = 1579 Increasing ? SR.HighLimit.has_value() : SR.LowLimit.has_value(); 1580 1581 Value *ExitPreLoopAt = nullptr; 1582 Value *ExitMainLoopAt = nullptr; 1583 const SCEVConstant *MinusOneS = 1584 cast<SCEVConstant>(SE.getConstant(IVTy, -1, true /* isSigned */)); 1585 1586 if (NeedsPreLoop) { 1587 const SCEV *ExitPreLoopAtSCEV = nullptr; 1588 1589 if (Increasing) 1590 ExitPreLoopAtSCEV = *SR.LowLimit; 1591 else if (cannotBeMinInLoop(*SR.HighLimit, &OriginalLoop, SE, 1592 IsSignedPredicate)) 1593 ExitPreLoopAtSCEV = SE.getAddExpr(*SR.HighLimit, MinusOneS); 1594 else { 1595 LLVM_DEBUG(dbgs() << "irce: could not prove no-overflow when computing " 1596 << "preloop exit limit. HighLimit = " 1597 << *(*SR.HighLimit) << "\n"); 1598 return false; 1599 } 1600 1601 if (!Expander.isSafeToExpandAt(ExitPreLoopAtSCEV, InsertPt)) { 1602 LLVM_DEBUG(dbgs() << "irce: could not prove that it is safe to expand the" 1603 << " preloop exit limit " << *ExitPreLoopAtSCEV 1604 << " at block " << InsertPt->getParent()->getName() 1605 << "\n"); 1606 return false; 1607 } 1608 1609 ExitPreLoopAt = Expander.expandCodeFor(ExitPreLoopAtSCEV, IVTy, InsertPt); 1610 ExitPreLoopAt->setName("exit.preloop.at"); 1611 } 1612 1613 if (NeedsPostLoop) { 1614 const SCEV *ExitMainLoopAtSCEV = nullptr; 1615 1616 if (Increasing) 1617 ExitMainLoopAtSCEV = *SR.HighLimit; 1618 else if (cannotBeMinInLoop(*SR.LowLimit, &OriginalLoop, SE, 1619 IsSignedPredicate)) 1620 ExitMainLoopAtSCEV = SE.getAddExpr(*SR.LowLimit, MinusOneS); 1621 else { 1622 LLVM_DEBUG(dbgs() << "irce: could not prove no-overflow when computing " 1623 << "mainloop exit limit. LowLimit = " 1624 << *(*SR.LowLimit) << "\n"); 1625 return false; 1626 } 1627 1628 if (!Expander.isSafeToExpandAt(ExitMainLoopAtSCEV, InsertPt)) { 1629 LLVM_DEBUG(dbgs() << "irce: could not prove that it is safe to expand the" 1630 << " main loop exit limit " << *ExitMainLoopAtSCEV 1631 << " at block " << InsertPt->getParent()->getName() 1632 << "\n"); 1633 return false; 1634 } 1635 1636 ExitMainLoopAt = Expander.expandCodeFor(ExitMainLoopAtSCEV, IVTy, InsertPt); 1637 ExitMainLoopAt->setName("exit.mainloop.at"); 1638 } 1639 1640 // We clone these ahead of time so that we don't have to deal with changing 1641 // and temporarily invalid IR as we transform the loops. 1642 if (NeedsPreLoop) 1643 cloneLoop(PreLoop, "preloop"); 1644 if (NeedsPostLoop) 1645 cloneLoop(PostLoop, "postloop"); 1646 1647 RewrittenRangeInfo PreLoopRRI; 1648 1649 if (NeedsPreLoop) { 1650 Preheader->getTerminator()->replaceUsesOfWith(MainLoopStructure.Header, 1651 PreLoop.Structure.Header); 1652 1653 MainLoopPreheader = 1654 createPreheader(MainLoopStructure, Preheader, "mainloop"); 1655 PreLoopRRI = changeIterationSpaceEnd(PreLoop.Structure, Preheader, 1656 ExitPreLoopAt, MainLoopPreheader); 1657 rewriteIncomingValuesForPHIs(MainLoopStructure, MainLoopPreheader, 1658 PreLoopRRI); 1659 } 1660 1661 BasicBlock *PostLoopPreheader = nullptr; 1662 RewrittenRangeInfo PostLoopRRI; 1663 1664 if (NeedsPostLoop) { 1665 PostLoopPreheader = 1666 createPreheader(PostLoop.Structure, Preheader, "postloop"); 1667 PostLoopRRI = changeIterationSpaceEnd(MainLoopStructure, MainLoopPreheader, 1668 ExitMainLoopAt, PostLoopPreheader); 1669 rewriteIncomingValuesForPHIs(PostLoop.Structure, PostLoopPreheader, 1670 PostLoopRRI); 1671 } 1672 1673 BasicBlock *NewMainLoopPreheader = 1674 MainLoopPreheader != Preheader ? MainLoopPreheader : nullptr; 1675 BasicBlock *NewBlocks[] = {PostLoopPreheader, PreLoopRRI.PseudoExit, 1676 PreLoopRRI.ExitSelector, PostLoopRRI.PseudoExit, 1677 PostLoopRRI.ExitSelector, NewMainLoopPreheader}; 1678 1679 // Some of the above may be nullptr, filter them out before passing to 1680 // addToParentLoopIfNeeded. 1681 auto NewBlocksEnd = 1682 std::remove(std::begin(NewBlocks), std::end(NewBlocks), nullptr); 1683 1684 addToParentLoopIfNeeded(ArrayRef(std::begin(NewBlocks), NewBlocksEnd)); 1685 1686 DT.recalculate(F); 1687 1688 // We need to first add all the pre and post loop blocks into the loop 1689 // structures (as part of createClonedLoopStructure), and then update the 1690 // LCSSA form and LoopSimplifyForm. This is necessary for correctly updating 1691 // LI when LoopSimplifyForm is generated. 1692 Loop *PreL = nullptr, *PostL = nullptr; 1693 if (!PreLoop.Blocks.empty()) { 1694 PreL = createClonedLoopStructure(&OriginalLoop, 1695 OriginalLoop.getParentLoop(), PreLoop.Map, 1696 /* IsSubLoop */ false); 1697 } 1698 1699 if (!PostLoop.Blocks.empty()) { 1700 PostL = 1701 createClonedLoopStructure(&OriginalLoop, OriginalLoop.getParentLoop(), 1702 PostLoop.Map, /* IsSubLoop */ false); 1703 } 1704 1705 // This function canonicalizes the loop into Loop-Simplify and LCSSA forms. 1706 auto CanonicalizeLoop = [&] (Loop *L, bool IsOriginalLoop) { 1707 formLCSSARecursively(*L, DT, &LI, &SE); 1708 simplifyLoop(L, &DT, &LI, &SE, nullptr, nullptr, true); 1709 // Pre/post loops are slow paths, we do not need to perform any loop 1710 // optimizations on them. 1711 if (!IsOriginalLoop) 1712 DisableAllLoopOptsOnLoop(*L); 1713 }; 1714 if (PreL) 1715 CanonicalizeLoop(PreL, false); 1716 if (PostL) 1717 CanonicalizeLoop(PostL, false); 1718 CanonicalizeLoop(&OriginalLoop, true); 1719 1720 /// At this point: 1721 /// - We've broken a "main loop" out of the loop in a way that the "main loop" 1722 /// runs with the induction variable in a subset of [Begin, End). 1723 /// - There is no overflow when computing "main loop" exit limit. 1724 /// - Max latch taken count of the loop is limited. 1725 /// It guarantees that induction variable will not overflow iterating in the 1726 /// "main loop". 1727 if (auto BO = dyn_cast<BinaryOperator>(MainLoopStructure.IndVarBase)) 1728 if (IsSignedPredicate) 1729 BO->setHasNoSignedWrap(true); 1730 /// TODO: support unsigned predicate. 1731 /// To add NUW flag we need to prove that both operands of BO are 1732 /// non-negative. E.g: 1733 /// ... 1734 /// %iv.next = add nsw i32 %iv, -1 1735 /// %cmp = icmp ult i32 %iv.next, %n 1736 /// br i1 %cmp, label %loopexit, label %loop 1737 /// 1738 /// -1 is MAX_UINT in terms of unsigned int. Adding anything but zero will 1739 /// overflow, therefore NUW flag is not legal here. 1740 1741 return true; 1742 } 1743 1744 /// Computes and returns a range of values for the induction variable (IndVar) 1745 /// in which the range check can be safely elided. If it cannot compute such a 1746 /// range, returns std::nullopt. 1747 std::optional<InductiveRangeCheck::Range> 1748 InductiveRangeCheck::computeSafeIterationSpace(ScalarEvolution &SE, 1749 const SCEVAddRecExpr *IndVar, 1750 bool IsLatchSigned) const { 1751 // We can deal when types of latch check and range checks don't match in case 1752 // if latch check is more narrow. 1753 auto *IVType = dyn_cast<IntegerType>(IndVar->getType()); 1754 auto *RCType = dyn_cast<IntegerType>(getBegin()->getType()); 1755 auto *EndType = dyn_cast<IntegerType>(getEnd()->getType()); 1756 // Do not work with pointer types. 1757 if (!IVType || !RCType) 1758 return std::nullopt; 1759 if (IVType->getBitWidth() > RCType->getBitWidth()) 1760 return std::nullopt; 1761 1762 // IndVar is of the form "A + B * I" (where "I" is the canonical induction 1763 // variable, that may or may not exist as a real llvm::Value in the loop) and 1764 // this inductive range check is a range check on the "C + D * I" ("C" is 1765 // getBegin() and "D" is getStep()). We rewrite the value being range 1766 // checked to "M + N * IndVar" where "N" = "D * B^(-1)" and "M" = "C - NA". 1767 // 1768 // The actual inequalities we solve are of the form 1769 // 1770 // 0 <= M + 1 * IndVar < L given L >= 0 (i.e. N == 1) 1771 // 1772 // Here L stands for upper limit of the safe iteration space. 1773 // The inequality is satisfied by (0 - M) <= IndVar < (L - M). To avoid 1774 // overflows when calculating (0 - M) and (L - M) we, depending on type of 1775 // IV's iteration space, limit the calculations by borders of the iteration 1776 // space. For example, if IndVar is unsigned, (0 - M) overflows for any M > 0. 1777 // If we figured out that "anything greater than (-M) is safe", we strengthen 1778 // this to "everything greater than 0 is safe", assuming that values between 1779 // -M and 0 just do not exist in unsigned iteration space, and we don't want 1780 // to deal with overflown values. 1781 1782 if (!IndVar->isAffine()) 1783 return std::nullopt; 1784 1785 const SCEV *A = NoopOrExtend(IndVar->getStart(), RCType, SE, IsLatchSigned); 1786 const SCEVConstant *B = dyn_cast<SCEVConstant>( 1787 NoopOrExtend(IndVar->getStepRecurrence(SE), RCType, SE, IsLatchSigned)); 1788 if (!B) 1789 return std::nullopt; 1790 assert(!B->isZero() && "Recurrence with zero step?"); 1791 1792 const SCEV *C = getBegin(); 1793 const SCEVConstant *D = dyn_cast<SCEVConstant>(getStep()); 1794 if (D != B) 1795 return std::nullopt; 1796 1797 assert(!D->getValue()->isZero() && "Recurrence with zero step?"); 1798 unsigned BitWidth = RCType->getBitWidth(); 1799 const SCEV *SIntMax = SE.getConstant(APInt::getSignedMaxValue(BitWidth)); 1800 const SCEV *SIntMin = SE.getConstant(APInt::getSignedMinValue(BitWidth)); 1801 1802 // Subtract Y from X so that it does not go through border of the IV 1803 // iteration space. Mathematically, it is equivalent to: 1804 // 1805 // ClampedSubtract(X, Y) = min(max(X - Y, INT_MIN), INT_MAX). [1] 1806 // 1807 // In [1], 'X - Y' is a mathematical subtraction (result is not bounded to 1808 // any width of bit grid). But after we take min/max, the result is 1809 // guaranteed to be within [INT_MIN, INT_MAX]. 1810 // 1811 // In [1], INT_MAX and INT_MIN are respectively signed and unsigned max/min 1812 // values, depending on type of latch condition that defines IV iteration 1813 // space. 1814 auto ClampedSubtract = [&](const SCEV *X, const SCEV *Y) { 1815 // FIXME: The current implementation assumes that X is in [0, SINT_MAX]. 1816 // This is required to ensure that SINT_MAX - X does not overflow signed and 1817 // that X - Y does not overflow unsigned if Y is negative. Can we lift this 1818 // restriction and make it work for negative X either? 1819 if (IsLatchSigned) { 1820 // X is a number from signed range, Y is interpreted as signed. 1821 // Even if Y is SINT_MAX, (X - Y) does not reach SINT_MIN. So the only 1822 // thing we should care about is that we didn't cross SINT_MAX. 1823 // So, if Y is positive, we subtract Y safely. 1824 // Rule 1: Y > 0 ---> Y. 1825 // If 0 <= -Y <= (SINT_MAX - X), we subtract Y safely. 1826 // Rule 2: Y >=s (X - SINT_MAX) ---> Y. 1827 // If 0 <= (SINT_MAX - X) < -Y, we can only subtract (X - SINT_MAX). 1828 // Rule 3: Y <s (X - SINT_MAX) ---> (X - SINT_MAX). 1829 // It gives us smax(Y, X - SINT_MAX) to subtract in all cases. 1830 const SCEV *XMinusSIntMax = SE.getMinusSCEV(X, SIntMax); 1831 return SE.getMinusSCEV(X, SE.getSMaxExpr(Y, XMinusSIntMax), 1832 SCEV::FlagNSW); 1833 } else 1834 // X is a number from unsigned range, Y is interpreted as signed. 1835 // Even if Y is SINT_MIN, (X - Y) does not reach UINT_MAX. So the only 1836 // thing we should care about is that we didn't cross zero. 1837 // So, if Y is negative, we subtract Y safely. 1838 // Rule 1: Y <s 0 ---> Y. 1839 // If 0 <= Y <= X, we subtract Y safely. 1840 // Rule 2: Y <=s X ---> Y. 1841 // If 0 <= X < Y, we should stop at 0 and can only subtract X. 1842 // Rule 3: Y >s X ---> X. 1843 // It gives us smin(X, Y) to subtract in all cases. 1844 return SE.getMinusSCEV(X, SE.getSMinExpr(X, Y), SCEV::FlagNUW); 1845 }; 1846 const SCEV *M = SE.getMinusSCEV(C, A); 1847 const SCEV *Zero = SE.getZero(M->getType()); 1848 1849 // This function returns SCEV equal to 1 if X is non-negative 0 otherwise. 1850 auto SCEVCheckNonNegative = [&](const SCEV *X) { 1851 const Loop *L = IndVar->getLoop(); 1852 const SCEV *Zero = SE.getZero(X->getType()); 1853 const SCEV *One = SE.getOne(X->getType()); 1854 // Can we trivially prove that X is a non-negative or negative value? 1855 if (isKnownNonNegativeInLoop(X, L, SE)) 1856 return One; 1857 else if (isKnownNegativeInLoop(X, L, SE)) 1858 return Zero; 1859 // If not, we will have to figure it out during the execution. 1860 // Function smax(smin(X, 0), -1) + 1 equals to 1 if X >= 0 and 0 if X < 0. 1861 const SCEV *NegOne = SE.getNegativeSCEV(One); 1862 return SE.getAddExpr(SE.getSMaxExpr(SE.getSMinExpr(X, Zero), NegOne), One); 1863 }; 1864 1865 // This function returns SCEV equal to 1 if X will not overflow in terms of 1866 // range check type, 0 otherwise. 1867 auto SCEVCheckWillNotOverflow = [&](const SCEV *X) { 1868 // X doesn't overflow if SINT_MAX >= X. 1869 // Then if (SINT_MAX - X) >= 0, X doesn't overflow 1870 const SCEV *SIntMaxExt = SE.getSignExtendExpr(SIntMax, X->getType()); 1871 const SCEV *OverflowCheck = 1872 SCEVCheckNonNegative(SE.getMinusSCEV(SIntMaxExt, X)); 1873 1874 // X doesn't underflow if X >= SINT_MIN. 1875 // Then if (X - SINT_MIN) >= 0, X doesn't underflow 1876 const SCEV *SIntMinExt = SE.getSignExtendExpr(SIntMin, X->getType()); 1877 const SCEV *UnderflowCheck = 1878 SCEVCheckNonNegative(SE.getMinusSCEV(X, SIntMinExt)); 1879 1880 return SE.getMulExpr(OverflowCheck, UnderflowCheck); 1881 }; 1882 1883 // FIXME: Current implementation of ClampedSubtract implicitly assumes that 1884 // X is non-negative (in sense of a signed value). We need to re-implement 1885 // this function in a way that it will correctly handle negative X as well. 1886 // We use it twice: for X = 0 everything is fine, but for X = getEnd() we can 1887 // end up with a negative X and produce wrong results. So currently we ensure 1888 // that if getEnd() is negative then both ends of the safe range are zero. 1889 // Note that this may pessimize elimination of unsigned range checks against 1890 // negative values. 1891 const SCEV *REnd = getEnd(); 1892 const SCEV *EndWillNotOverflow = SE.getOne(RCType); 1893 1894 auto PrintRangeCheck = [&](raw_ostream &OS) { 1895 auto L = IndVar->getLoop(); 1896 OS << "irce: in function "; 1897 OS << L->getHeader()->getParent()->getName(); 1898 OS << ", in "; 1899 L->print(OS); 1900 OS << "there is range check with scaled boundary:\n"; 1901 print(OS); 1902 }; 1903 1904 if (EndType->getBitWidth() > RCType->getBitWidth()) { 1905 assert(EndType->getBitWidth() == RCType->getBitWidth() * 2); 1906 if (PrintScaledBoundaryRangeChecks) 1907 PrintRangeCheck(errs()); 1908 // End is computed with extended type but will be truncated to a narrow one 1909 // type of range check. Therefore we need a check that the result will not 1910 // overflow in terms of narrow type. 1911 EndWillNotOverflow = 1912 SE.getTruncateExpr(SCEVCheckWillNotOverflow(REnd), RCType); 1913 REnd = SE.getTruncateExpr(REnd, RCType); 1914 } 1915 1916 const SCEV *RuntimeChecks = 1917 SE.getMulExpr(SCEVCheckNonNegative(REnd), EndWillNotOverflow); 1918 const SCEV *Begin = SE.getMulExpr(ClampedSubtract(Zero, M), RuntimeChecks); 1919 const SCEV *End = SE.getMulExpr(ClampedSubtract(REnd, M), RuntimeChecks); 1920 1921 return InductiveRangeCheck::Range(Begin, End); 1922 } 1923 1924 static std::optional<InductiveRangeCheck::Range> 1925 IntersectSignedRange(ScalarEvolution &SE, 1926 const std::optional<InductiveRangeCheck::Range> &R1, 1927 const InductiveRangeCheck::Range &R2) { 1928 if (R2.isEmpty(SE, /* IsSigned */ true)) 1929 return std::nullopt; 1930 if (!R1) 1931 return R2; 1932 auto &R1Value = *R1; 1933 // We never return empty ranges from this function, and R1 is supposed to be 1934 // a result of intersection. Thus, R1 is never empty. 1935 assert(!R1Value.isEmpty(SE, /* IsSigned */ true) && 1936 "We should never have empty R1!"); 1937 1938 // TODO: we could widen the smaller range and have this work; but for now we 1939 // bail out to keep things simple. 1940 if (R1Value.getType() != R2.getType()) 1941 return std::nullopt; 1942 1943 const SCEV *NewBegin = SE.getSMaxExpr(R1Value.getBegin(), R2.getBegin()); 1944 const SCEV *NewEnd = SE.getSMinExpr(R1Value.getEnd(), R2.getEnd()); 1945 1946 // If the resulting range is empty, just return std::nullopt. 1947 auto Ret = InductiveRangeCheck::Range(NewBegin, NewEnd); 1948 if (Ret.isEmpty(SE, /* IsSigned */ true)) 1949 return std::nullopt; 1950 return Ret; 1951 } 1952 1953 static std::optional<InductiveRangeCheck::Range> 1954 IntersectUnsignedRange(ScalarEvolution &SE, 1955 const std::optional<InductiveRangeCheck::Range> &R1, 1956 const InductiveRangeCheck::Range &R2) { 1957 if (R2.isEmpty(SE, /* IsSigned */ false)) 1958 return std::nullopt; 1959 if (!R1) 1960 return R2; 1961 auto &R1Value = *R1; 1962 // We never return empty ranges from this function, and R1 is supposed to be 1963 // a result of intersection. Thus, R1 is never empty. 1964 assert(!R1Value.isEmpty(SE, /* IsSigned */ false) && 1965 "We should never have empty R1!"); 1966 1967 // TODO: we could widen the smaller range and have this work; but for now we 1968 // bail out to keep things simple. 1969 if (R1Value.getType() != R2.getType()) 1970 return std::nullopt; 1971 1972 const SCEV *NewBegin = SE.getUMaxExpr(R1Value.getBegin(), R2.getBegin()); 1973 const SCEV *NewEnd = SE.getUMinExpr(R1Value.getEnd(), R2.getEnd()); 1974 1975 // If the resulting range is empty, just return std::nullopt. 1976 auto Ret = InductiveRangeCheck::Range(NewBegin, NewEnd); 1977 if (Ret.isEmpty(SE, /* IsSigned */ false)) 1978 return std::nullopt; 1979 return Ret; 1980 } 1981 1982 PreservedAnalyses IRCEPass::run(Function &F, FunctionAnalysisManager &AM) { 1983 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 1984 LoopInfo &LI = AM.getResult<LoopAnalysis>(F); 1985 // There are no loops in the function. Return before computing other expensive 1986 // analyses. 1987 if (LI.empty()) 1988 return PreservedAnalyses::all(); 1989 auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F); 1990 auto &BPI = AM.getResult<BranchProbabilityAnalysis>(F); 1991 1992 // Get BFI analysis result on demand. Please note that modification of 1993 // CFG invalidates this analysis and we should handle it. 1994 auto getBFI = [&F, &AM ]()->BlockFrequencyInfo & { 1995 return AM.getResult<BlockFrequencyAnalysis>(F); 1996 }; 1997 InductiveRangeCheckElimination IRCE(SE, &BPI, DT, LI, { getBFI }); 1998 1999 bool Changed = false; 2000 { 2001 bool CFGChanged = false; 2002 for (const auto &L : LI) { 2003 CFGChanged |= simplifyLoop(L, &DT, &LI, &SE, nullptr, nullptr, 2004 /*PreserveLCSSA=*/false); 2005 Changed |= formLCSSARecursively(*L, DT, &LI, &SE); 2006 } 2007 Changed |= CFGChanged; 2008 2009 if (CFGChanged && !SkipProfitabilityChecks) { 2010 PreservedAnalyses PA = PreservedAnalyses::all(); 2011 PA.abandon<BlockFrequencyAnalysis>(); 2012 AM.invalidate(F, PA); 2013 } 2014 } 2015 2016 SmallPriorityWorklist<Loop *, 4> Worklist; 2017 appendLoopsToWorklist(LI, Worklist); 2018 auto LPMAddNewLoop = [&Worklist](Loop *NL, bool IsSubloop) { 2019 if (!IsSubloop) 2020 appendLoopsToWorklist(*NL, Worklist); 2021 }; 2022 2023 while (!Worklist.empty()) { 2024 Loop *L = Worklist.pop_back_val(); 2025 if (IRCE.run(L, LPMAddNewLoop)) { 2026 Changed = true; 2027 if (!SkipProfitabilityChecks) { 2028 PreservedAnalyses PA = PreservedAnalyses::all(); 2029 PA.abandon<BlockFrequencyAnalysis>(); 2030 AM.invalidate(F, PA); 2031 } 2032 } 2033 } 2034 2035 if (!Changed) 2036 return PreservedAnalyses::all(); 2037 return getLoopPassPreservedAnalyses(); 2038 } 2039 2040 bool 2041 InductiveRangeCheckElimination::isProfitableToTransform(const Loop &L, 2042 LoopStructure &LS) { 2043 if (SkipProfitabilityChecks) 2044 return true; 2045 if (GetBFI) { 2046 BlockFrequencyInfo &BFI = (*GetBFI)(); 2047 uint64_t hFreq = BFI.getBlockFreq(LS.Header).getFrequency(); 2048 uint64_t phFreq = BFI.getBlockFreq(L.getLoopPreheader()).getFrequency(); 2049 if (phFreq != 0 && hFreq != 0 && (hFreq / phFreq < MinRuntimeIterations)) { 2050 LLVM_DEBUG(dbgs() << "irce: could not prove profitability: " 2051 << "the estimated number of iterations basing on " 2052 "frequency info is " << (hFreq / phFreq) << "\n";); 2053 return false; 2054 } 2055 return true; 2056 } 2057 2058 if (!BPI) 2059 return true; 2060 BranchProbability ExitProbability = 2061 BPI->getEdgeProbability(LS.Latch, LS.LatchBrExitIdx); 2062 if (ExitProbability > BranchProbability(1, MinRuntimeIterations)) { 2063 LLVM_DEBUG(dbgs() << "irce: could not prove profitability: " 2064 << "the exit probability is too big " << ExitProbability 2065 << "\n";); 2066 return false; 2067 } 2068 return true; 2069 } 2070 2071 bool InductiveRangeCheckElimination::run( 2072 Loop *L, function_ref<void(Loop *, bool)> LPMAddNewLoop) { 2073 if (L->getBlocks().size() >= LoopSizeCutoff) { 2074 LLVM_DEBUG(dbgs() << "irce: giving up constraining loop, too large\n"); 2075 return false; 2076 } 2077 2078 BasicBlock *Preheader = L->getLoopPreheader(); 2079 if (!Preheader) { 2080 LLVM_DEBUG(dbgs() << "irce: loop has no preheader, leaving\n"); 2081 return false; 2082 } 2083 2084 LLVMContext &Context = Preheader->getContext(); 2085 SmallVector<InductiveRangeCheck, 16> RangeChecks; 2086 bool Changed = false; 2087 2088 for (auto *BBI : L->getBlocks()) 2089 if (BranchInst *TBI = dyn_cast<BranchInst>(BBI->getTerminator())) 2090 InductiveRangeCheck::extractRangeChecksFromBranch(TBI, L, SE, BPI, 2091 RangeChecks, Changed); 2092 2093 if (RangeChecks.empty()) 2094 return Changed; 2095 2096 auto PrintRecognizedRangeChecks = [&](raw_ostream &OS) { 2097 OS << "irce: looking at loop "; L->print(OS); 2098 OS << "irce: loop has " << RangeChecks.size() 2099 << " inductive range checks: \n"; 2100 for (InductiveRangeCheck &IRC : RangeChecks) 2101 IRC.print(OS); 2102 }; 2103 2104 LLVM_DEBUG(PrintRecognizedRangeChecks(dbgs())); 2105 2106 if (PrintRangeChecks) 2107 PrintRecognizedRangeChecks(errs()); 2108 2109 const char *FailureReason = nullptr; 2110 std::optional<LoopStructure> MaybeLoopStructure = 2111 LoopStructure::parseLoopStructure(SE, *L, FailureReason); 2112 if (!MaybeLoopStructure) { 2113 LLVM_DEBUG(dbgs() << "irce: could not parse loop structure: " 2114 << FailureReason << "\n";); 2115 return Changed; 2116 } 2117 LoopStructure LS = *MaybeLoopStructure; 2118 if (!isProfitableToTransform(*L, LS)) 2119 return Changed; 2120 const SCEVAddRecExpr *IndVar = 2121 cast<SCEVAddRecExpr>(SE.getMinusSCEV(SE.getSCEV(LS.IndVarBase), SE.getSCEV(LS.IndVarStep))); 2122 2123 std::optional<InductiveRangeCheck::Range> SafeIterRange; 2124 2125 SmallVector<InductiveRangeCheck, 4> RangeChecksToEliminate; 2126 // Basing on the type of latch predicate, we interpret the IV iteration range 2127 // as signed or unsigned range. We use different min/max functions (signed or 2128 // unsigned) when intersecting this range with safe iteration ranges implied 2129 // by range checks. 2130 auto IntersectRange = 2131 LS.IsSignedPredicate ? IntersectSignedRange : IntersectUnsignedRange; 2132 2133 for (InductiveRangeCheck &IRC : RangeChecks) { 2134 auto Result = IRC.computeSafeIterationSpace(SE, IndVar, 2135 LS.IsSignedPredicate); 2136 if (Result) { 2137 auto MaybeSafeIterRange = IntersectRange(SE, SafeIterRange, *Result); 2138 if (MaybeSafeIterRange) { 2139 assert(!MaybeSafeIterRange->isEmpty(SE, LS.IsSignedPredicate) && 2140 "We should never return empty ranges!"); 2141 RangeChecksToEliminate.push_back(IRC); 2142 SafeIterRange = *MaybeSafeIterRange; 2143 } 2144 } 2145 } 2146 2147 if (!SafeIterRange) 2148 return Changed; 2149 2150 LoopConstrainer LC(*L, LI, LPMAddNewLoop, LS, SE, DT, *SafeIterRange); 2151 2152 if (LC.run()) { 2153 Changed = true; 2154 2155 auto PrintConstrainedLoopInfo = [L]() { 2156 dbgs() << "irce: in function "; 2157 dbgs() << L->getHeader()->getParent()->getName() << ": "; 2158 dbgs() << "constrained "; 2159 L->print(dbgs()); 2160 }; 2161 2162 LLVM_DEBUG(PrintConstrainedLoopInfo()); 2163 2164 if (PrintChangedLoops) 2165 PrintConstrainedLoopInfo(); 2166 2167 // Optimize away the now-redundant range checks. 2168 2169 for (InductiveRangeCheck &IRC : RangeChecksToEliminate) { 2170 ConstantInt *FoldedRangeCheck = IRC.getPassingDirection() 2171 ? ConstantInt::getTrue(Context) 2172 : ConstantInt::getFalse(Context); 2173 IRC.getCheckUse()->set(FoldedRangeCheck); 2174 } 2175 } 2176 2177 return Changed; 2178 } 2179