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