1 //===- InductiveRangeCheckElimination.cpp - -------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // The InductiveRangeCheckElimination pass splits a loop's iteration space into
11 // three disjoint ranges. It does that in a way such that the loop running in
12 // the middle loop provably does not need range checks. As an example, it will
15 // len = < known positive >
16 // for (i = 0; i < n; i++) {
17 // if (0 <= i && i < len) {
20 // throw_out_of_bounds();
26 // len = < known positive >
27 // limit = smin(n, len)
28 // // no first segment
29 // for (i = 0; i < limit; i++) {
30 // if (0 <= i && i < len) { // this check is fully redundant
33 // throw_out_of_bounds();
36 // for (i = limit; i < n; i++) {
37 // if (0 <= i && i < len) {
40 // throw_out_of_bounds();
44 //===----------------------------------------------------------------------===//
46 #include "llvm/Transforms/Scalar/InductiveRangeCheckElimination.h"
47 #include "llvm/ADT/APInt.h"
48 #include "llvm/ADT/ArrayRef.h"
49 #include "llvm/ADT/None.h"
50 #include "llvm/ADT/Optional.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/BranchProbabilityInfo.h"
56 #include "llvm/Analysis/LoopAnalysisManager.h"
57 #include "llvm/Analysis/LoopInfo.h"
58 #include "llvm/Analysis/LoopPass.h"
59 #include "llvm/Analysis/ScalarEvolution.h"
60 #include "llvm/Analysis/ScalarEvolutionExpander.h"
61 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
62 #include "llvm/IR/BasicBlock.h"
63 #include "llvm/IR/CFG.h"
64 #include "llvm/IR/Constants.h"
65 #include "llvm/IR/DerivedTypes.h"
66 #include "llvm/IR/Dominators.h"
67 #include "llvm/IR/Function.h"
68 #include "llvm/IR/IRBuilder.h"
69 #include "llvm/IR/InstrTypes.h"
70 #include "llvm/IR/Instructions.h"
71 #include "llvm/IR/Metadata.h"
72 #include "llvm/IR/Module.h"
73 #include "llvm/IR/PatternMatch.h"
74 #include "llvm/IR/Type.h"
75 #include "llvm/IR/Use.h"
76 #include "llvm/IR/User.h"
77 #include "llvm/IR/Value.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/ValueMapper.h"
99 using namespace llvm::PatternMatch;
101 static cl::opt<unsigned> LoopSizeCutoff("irce-loop-size-cutoff", cl::Hidden,
104 static cl::opt<bool> PrintChangedLoops("irce-print-changed-loops", cl::Hidden,
107 static cl::opt<bool> PrintRangeChecks("irce-print-range-checks", cl::Hidden,
110 static cl::opt<int> MaxExitProbReciprocal("irce-max-exit-prob-reciprocal",
111 cl::Hidden, cl::init(10));
113 static cl::opt<bool> SkipProfitabilityChecks("irce-skip-profitability-checks",
114 cl::Hidden, cl::init(false));
116 static cl::opt<bool> AllowUnsignedLatchCondition("irce-allow-unsigned-latch",
117 cl::Hidden, cl::init(true));
119 static const char *ClonedLoopTag = "irce.loop.clone";
121 #define DEBUG_TYPE "irce"
125 /// An inductive range check is conditional branch in a loop with
127 /// 1. a very cold successor (i.e. the branch jumps to that successor very
132 /// 2. a condition that is provably true for some contiguous range of values
133 /// taken by the containing loop's induction variable.
135 class InductiveRangeCheck {
136 // Classifies a range check
137 enum RangeCheckKind : unsigned {
138 // Range check of the form "0 <= I".
139 RANGE_CHECK_LOWER = 1,
141 // Range check of the form "I < L" where L is known positive.
142 RANGE_CHECK_UPPER = 2,
144 // The logical and of the RANGE_CHECK_LOWER and RANGE_CHECK_UPPER
146 RANGE_CHECK_BOTH = RANGE_CHECK_LOWER | RANGE_CHECK_UPPER,
148 // Unrecognized range check condition.
149 RANGE_CHECK_UNKNOWN = (unsigned)-1
152 static StringRef rangeCheckKindToStr(RangeCheckKind);
154 const SCEV *Begin = nullptr;
155 const SCEV *Step = nullptr;
156 const SCEV *End = nullptr;
157 Use *CheckUse = nullptr;
158 RangeCheckKind Kind = RANGE_CHECK_UNKNOWN;
159 bool IsSigned = true;
161 static RangeCheckKind parseRangeCheckICmp(Loop *L, ICmpInst *ICI,
162 ScalarEvolution &SE, Value *&Index,
163 Value *&Length, bool &IsSigned);
166 extractRangeChecksFromCond(Loop *L, ScalarEvolution &SE, Use &ConditionUse,
167 SmallVectorImpl<InductiveRangeCheck> &Checks,
168 SmallPtrSetImpl<Value *> &Visited);
171 const SCEV *getBegin() const { return Begin; }
172 const SCEV *getStep() const { return Step; }
173 const SCEV *getEnd() const { return End; }
174 bool isSigned() const { return IsSigned; }
176 void print(raw_ostream &OS) const {
177 OS << "InductiveRangeCheck:\n";
178 OS << " Kind: " << rangeCheckKindToStr(Kind) << "\n";
185 OS << "\n CheckUse: ";
186 getCheckUse()->getUser()->print(OS);
187 OS << " Operand: " << getCheckUse()->getOperandNo() << "\n";
195 Use *getCheckUse() const { return CheckUse; }
197 /// Represents an signed integer range [Range.getBegin(), Range.getEnd()). If
198 /// R.getEnd() le R.getBegin(), then R denotes the empty range.
205 Range(const SCEV *Begin, const SCEV *End) : Begin(Begin), End(End) {
206 assert(Begin->getType() == End->getType() && "ill-typed range!");
209 Type *getType() const { return Begin->getType(); }
210 const SCEV *getBegin() const { return Begin; }
211 const SCEV *getEnd() const { return End; }
212 bool isEmpty(ScalarEvolution &SE, bool IsSigned) const {
216 return SE.isKnownPredicate(ICmpInst::ICMP_SGE, Begin, End);
218 return SE.isKnownPredicate(ICmpInst::ICMP_UGE, Begin, End);
222 /// This is the value the condition of the branch needs to evaluate to for the
223 /// branch to take the hot successor (see (1) above).
224 bool getPassingDirection() { return true; }
226 /// Computes a range for the induction variable (IndVar) in which the range
227 /// check is redundant and can be constant-folded away. The induction
228 /// variable is not required to be the canonical {0,+,1} induction variable.
229 Optional<Range> computeSafeIterationSpace(ScalarEvolution &SE,
230 const SCEVAddRecExpr *IndVar,
231 bool IsLatchSigned) const;
233 /// Parse out a set of inductive range checks from \p BI and append them to \p
236 /// NB! There may be conditions feeding into \p BI that aren't inductive range
237 /// checks, and hence don't end up in \p Checks.
239 extractRangeChecksFromBranch(BranchInst *BI, Loop *L, ScalarEvolution &SE,
240 BranchProbabilityInfo *BPI,
241 SmallVectorImpl<InductiveRangeCheck> &Checks);
244 class InductiveRangeCheckElimination {
246 BranchProbabilityInfo *BPI;
251 InductiveRangeCheckElimination(ScalarEvolution &SE,
252 BranchProbabilityInfo *BPI, DominatorTree &DT,
254 : SE(SE), BPI(BPI), DT(DT), LI(LI) {}
256 bool run(Loop *L, function_ref<void(Loop *, bool)> LPMAddNewLoop);
259 class IRCELegacyPass : public LoopPass {
263 IRCELegacyPass() : LoopPass(ID) {
264 initializeIRCELegacyPassPass(*PassRegistry::getPassRegistry());
267 void getAnalysisUsage(AnalysisUsage &AU) const override {
268 AU.addRequired<BranchProbabilityInfoWrapperPass>();
269 getLoopAnalysisUsage(AU);
272 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
275 } // end anonymous namespace
277 char IRCELegacyPass::ID = 0;
279 INITIALIZE_PASS_BEGIN(IRCELegacyPass, "irce",
280 "Inductive range check elimination", false, false)
281 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)
282 INITIALIZE_PASS_DEPENDENCY(LoopPass)
283 INITIALIZE_PASS_END(IRCELegacyPass, "irce", "Inductive range check elimination",
286 StringRef InductiveRangeCheck::rangeCheckKindToStr(
287 InductiveRangeCheck::RangeCheckKind RCK) {
289 case InductiveRangeCheck::RANGE_CHECK_UNKNOWN:
290 return "RANGE_CHECK_UNKNOWN";
292 case InductiveRangeCheck::RANGE_CHECK_UPPER:
293 return "RANGE_CHECK_UPPER";
295 case InductiveRangeCheck::RANGE_CHECK_LOWER:
296 return "RANGE_CHECK_LOWER";
298 case InductiveRangeCheck::RANGE_CHECK_BOTH:
299 return "RANGE_CHECK_BOTH";
302 llvm_unreachable("unknown range check type!");
305 /// Parse a single ICmp instruction, `ICI`, into a range check. If `ICI` cannot
306 /// be interpreted as a range check, return `RANGE_CHECK_UNKNOWN` and set
307 /// `Index` and `Length` to `nullptr`. Otherwise set `Index` to the value being
308 /// range checked, and set `Length` to the upper limit `Index` is being range
309 /// checked with if (and only if) the range check type is stronger or equal to
310 /// RANGE_CHECK_UPPER.
311 InductiveRangeCheck::RangeCheckKind
312 InductiveRangeCheck::parseRangeCheckICmp(Loop *L, ICmpInst *ICI,
313 ScalarEvolution &SE, Value *&Index,
314 Value *&Length, bool &IsSigned) {
315 auto IsLoopInvariant = [&SE, L](Value *V) {
316 return SE.isLoopInvariant(SE.getSCEV(V), L);
319 ICmpInst::Predicate Pred = ICI->getPredicate();
320 Value *LHS = ICI->getOperand(0);
321 Value *RHS = ICI->getOperand(1);
325 return RANGE_CHECK_UNKNOWN;
327 case ICmpInst::ICMP_SLE:
330 case ICmpInst::ICMP_SGE:
332 if (match(RHS, m_ConstantInt<0>())) {
334 return RANGE_CHECK_LOWER;
336 return RANGE_CHECK_UNKNOWN;
338 case ICmpInst::ICMP_SLT:
341 case ICmpInst::ICMP_SGT:
343 if (match(RHS, m_ConstantInt<-1>())) {
345 return RANGE_CHECK_LOWER;
348 if (IsLoopInvariant(LHS)) {
351 return RANGE_CHECK_UPPER;
353 return RANGE_CHECK_UNKNOWN;
355 case ICmpInst::ICMP_ULT:
358 case ICmpInst::ICMP_UGT:
360 if (IsLoopInvariant(LHS)) {
363 return RANGE_CHECK_BOTH;
365 return RANGE_CHECK_UNKNOWN;
368 llvm_unreachable("default clause returns!");
371 void InductiveRangeCheck::extractRangeChecksFromCond(
372 Loop *L, ScalarEvolution &SE, Use &ConditionUse,
373 SmallVectorImpl<InductiveRangeCheck> &Checks,
374 SmallPtrSetImpl<Value *> &Visited) {
375 Value *Condition = ConditionUse.get();
376 if (!Visited.insert(Condition).second)
379 // TODO: Do the same for OR, XOR, NOT etc?
380 if (match(Condition, m_And(m_Value(), m_Value()))) {
381 extractRangeChecksFromCond(L, SE, cast<User>(Condition)->getOperandUse(0),
383 extractRangeChecksFromCond(L, SE, cast<User>(Condition)->getOperandUse(1),
388 ICmpInst *ICI = dyn_cast<ICmpInst>(Condition);
392 Value *Length = nullptr, *Index;
394 auto RCKind = parseRangeCheckICmp(L, ICI, SE, Index, Length, IsSigned);
395 if (RCKind == InductiveRangeCheck::RANGE_CHECK_UNKNOWN)
398 const auto *IndexAddRec = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Index));
400 IndexAddRec && (IndexAddRec->getLoop() == L) && IndexAddRec->isAffine();
405 const SCEV *End = nullptr;
406 // We strengthen "0 <= I" to "0 <= I < INT_SMAX" and "I < L" to "0 <= I < L".
407 // We can potentially do much better here.
409 End = SE.getSCEV(Length);
411 assert(RCKind == InductiveRangeCheck::RANGE_CHECK_LOWER && "invariant!");
412 // So far we can only reach this point for Signed range check. This may
413 // change in future. In this case we will need to pick Unsigned max for the
414 // unsigned range check.
415 unsigned BitWidth = cast<IntegerType>(IndexAddRec->getType())->getBitWidth();
416 const SCEV *SIntMax = SE.getConstant(APInt::getSignedMaxValue(BitWidth));
420 InductiveRangeCheck IRC;
422 IRC.Begin = IndexAddRec->getStart();
423 IRC.Step = IndexAddRec->getStepRecurrence(SE);
424 IRC.CheckUse = &ConditionUse;
426 IRC.IsSigned = IsSigned;
427 Checks.push_back(IRC);
430 void InductiveRangeCheck::extractRangeChecksFromBranch(
431 BranchInst *BI, Loop *L, ScalarEvolution &SE, BranchProbabilityInfo *BPI,
432 SmallVectorImpl<InductiveRangeCheck> &Checks) {
433 if (BI->isUnconditional() || BI->getParent() == L->getLoopLatch())
436 BranchProbability LikelyTaken(15, 16);
438 if (!SkipProfitabilityChecks && BPI &&
439 BPI->getEdgeProbability(BI->getParent(), (unsigned)0) < LikelyTaken)
442 SmallPtrSet<Value *, 8> Visited;
443 InductiveRangeCheck::extractRangeChecksFromCond(L, SE, BI->getOperandUse(0),
447 // Add metadata to the loop L to disable loop optimizations. Callers need to
448 // confirm that optimizing loop L is not beneficial.
449 static void DisableAllLoopOptsOnLoop(Loop &L) {
450 // We do not care about any existing loopID related metadata for L, since we
451 // are setting all loop metadata to false.
452 LLVMContext &Context = L.getHeader()->getContext();
453 // Reserve first location for self reference to the LoopID metadata node.
454 MDNode *Dummy = MDNode::get(Context, {});
455 MDNode *DisableUnroll = MDNode::get(
456 Context, {MDString::get(Context, "llvm.loop.unroll.disable")});
458 ConstantAsMetadata::get(ConstantInt::get(Type::getInt1Ty(Context), 0));
459 MDNode *DisableVectorize = MDNode::get(
461 {MDString::get(Context, "llvm.loop.vectorize.enable"), FalseVal});
462 MDNode *DisableLICMVersioning = MDNode::get(
463 Context, {MDString::get(Context, "llvm.loop.licm_versioning.disable")});
464 MDNode *DisableDistribution= MDNode::get(
466 {MDString::get(Context, "llvm.loop.distribute.enable"), FalseVal});
468 MDNode::get(Context, {Dummy, DisableUnroll, DisableVectorize,
469 DisableLICMVersioning, DisableDistribution});
470 // Set operand 0 to refer to the loop id itself.
471 NewLoopID->replaceOperandWith(0, NewLoopID);
472 L.setLoopID(NewLoopID);
477 // Keeps track of the structure of a loop. This is similar to llvm::Loop,
478 // except that it is more lightweight and can track the state of a loop through
479 // changing and potentially invalid IR. This structure also formalizes the
480 // kinds of loops we can deal with -- ones that have a single latch that is also
481 // an exiting block *and* have a canonical induction variable.
482 struct LoopStructure {
483 const char *Tag = "";
485 BasicBlock *Header = nullptr;
486 BasicBlock *Latch = nullptr;
488 // `Latch's terminator instruction is `LatchBr', and it's `LatchBrExitIdx'th
489 // successor is `LatchExit', the exit block of the loop.
490 BranchInst *LatchBr = nullptr;
491 BasicBlock *LatchExit = nullptr;
492 unsigned LatchBrExitIdx = std::numeric_limits<unsigned>::max();
494 // The loop represented by this instance of LoopStructure is semantically
497 // intN_ty inc = IndVarIncreasing ? 1 : -1;
498 // pred_ty predicate = IndVarIncreasing ? ICMP_SLT : ICMP_SGT;
500 // for (intN_ty iv = IndVarStart; predicate(iv, LoopExitAt); iv = IndVarBase)
503 Value *IndVarBase = nullptr;
504 Value *IndVarStart = nullptr;
505 Value *IndVarStep = nullptr;
506 Value *LoopExitAt = nullptr;
507 bool IndVarIncreasing = false;
508 bool IsSignedPredicate = true;
510 LoopStructure() = default;
512 template <typename M> LoopStructure map(M Map) const {
513 LoopStructure Result;
515 Result.Header = cast<BasicBlock>(Map(Header));
516 Result.Latch = cast<BasicBlock>(Map(Latch));
517 Result.LatchBr = cast<BranchInst>(Map(LatchBr));
518 Result.LatchExit = cast<BasicBlock>(Map(LatchExit));
519 Result.LatchBrExitIdx = LatchBrExitIdx;
520 Result.IndVarBase = Map(IndVarBase);
521 Result.IndVarStart = Map(IndVarStart);
522 Result.IndVarStep = Map(IndVarStep);
523 Result.LoopExitAt = Map(LoopExitAt);
524 Result.IndVarIncreasing = IndVarIncreasing;
525 Result.IsSignedPredicate = IsSignedPredicate;
529 static Optional<LoopStructure> parseLoopStructure(ScalarEvolution &,
530 BranchProbabilityInfo *BPI,
531 Loop &, const char *&);
534 /// This class is used to constrain loops to run within a given iteration space.
535 /// The algorithm this class implements is given a Loop and a range [Begin,
536 /// End). The algorithm then tries to break out a "main loop" out of the loop
537 /// it is given in a way that the "main loop" runs with the induction variable
538 /// in a subset of [Begin, End). The algorithm emits appropriate pre and post
539 /// loops to run any remaining iterations. The pre loop runs any iterations in
540 /// which the induction variable is < Begin, and the post loop runs any
541 /// iterations in which the induction variable is >= End.
542 class LoopConstrainer {
543 // The representation of a clone of the original loop we started out with.
546 std::vector<BasicBlock *> Blocks;
548 // `Map` maps values in the clonee into values in the cloned version
549 ValueToValueMapTy Map;
551 // An instance of `LoopStructure` for the cloned loop
552 LoopStructure Structure;
555 // Result of rewriting the range of a loop. See changeIterationSpaceEnd for
556 // more details on what these fields mean.
557 struct RewrittenRangeInfo {
558 BasicBlock *PseudoExit = nullptr;
559 BasicBlock *ExitSelector = nullptr;
560 std::vector<PHINode *> PHIValuesAtPseudoExit;
561 PHINode *IndVarEnd = nullptr;
563 RewrittenRangeInfo() = default;
566 // Calculated subranges we restrict the iteration space of the main loop to.
567 // See the implementation of `calculateSubRanges' for more details on how
568 // these fields are computed. `LowLimit` is None if there is no restriction
569 // on low end of the restricted iteration space of the main loop. `HighLimit`
570 // is None if there is no restriction on high end of the restricted iteration
571 // space of the main loop.
574 Optional<const SCEV *> LowLimit;
575 Optional<const SCEV *> HighLimit;
578 // A utility function that does a `replaceUsesOfWith' on the incoming block
579 // set of a `PHINode' -- replaces instances of `Block' in the `PHINode's
580 // incoming block list with `ReplaceBy'.
581 static void replacePHIBlock(PHINode *PN, BasicBlock *Block,
582 BasicBlock *ReplaceBy);
584 // Compute a safe set of limits for the main loop to run in -- effectively the
585 // intersection of `Range' and the iteration space of the original loop.
586 // Return None if unable to compute the set of subranges.
587 Optional<SubRanges> calculateSubRanges(bool IsSignedPredicate) const;
589 // Clone `OriginalLoop' and return the result in CLResult. The IR after
590 // running `cloneLoop' is well formed except for the PHI nodes in CLResult --
591 // the PHI nodes say that there is an incoming edge from `OriginalPreheader`
592 // but there is no such edge.
593 void cloneLoop(ClonedLoop &CLResult, const char *Tag) const;
595 // Create the appropriate loop structure needed to describe a cloned copy of
596 // `Original`. The clone is described by `VM`.
597 Loop *createClonedLoopStructure(Loop *Original, Loop *Parent,
598 ValueToValueMapTy &VM, bool IsSubloop);
600 // Rewrite the iteration space of the loop denoted by (LS, Preheader). The
601 // iteration space of the rewritten loop ends at ExitLoopAt. The start of the
602 // iteration space is not changed. `ExitLoopAt' is assumed to be slt
603 // `OriginalHeaderCount'.
605 // If there are iterations left to execute, control is made to jump to
606 // `ContinuationBlock', otherwise they take the normal loop exit. The
607 // returned `RewrittenRangeInfo' object is populated as follows:
609 // .PseudoExit is a basic block that unconditionally branches to
610 // `ContinuationBlock'.
612 // .ExitSelector is a basic block that decides, on exit from the loop,
613 // whether to branch to the "true" exit or to `PseudoExit'.
615 // .PHIValuesAtPseudoExit are PHINodes in `PseudoExit' that compute the value
616 // for each PHINode in the loop header on taking the pseudo exit.
618 // After changeIterationSpaceEnd, `Preheader' is no longer a legitimate
619 // preheader because it is made to branch to the loop header only
622 changeIterationSpaceEnd(const LoopStructure &LS, BasicBlock *Preheader,
624 BasicBlock *ContinuationBlock) const;
626 // The loop denoted by `LS' has `OldPreheader' as its preheader. This
627 // function creates a new preheader for `LS' and returns it.
628 BasicBlock *createPreheader(const LoopStructure &LS, BasicBlock *OldPreheader,
629 const char *Tag) const;
631 // `ContinuationBlockAndPreheader' was the continuation block for some call to
632 // `changeIterationSpaceEnd' and is the preheader to the loop denoted by `LS'.
633 // This function rewrites the PHI nodes in `LS.Header' to start with the
635 void rewriteIncomingValuesForPHIs(
636 LoopStructure &LS, BasicBlock *ContinuationBlockAndPreheader,
637 const LoopConstrainer::RewrittenRangeInfo &RRI) const;
639 // Even though we do not preserve any passes at this time, we at least need to
640 // keep the parent loop structure consistent. The `LPPassManager' seems to
641 // verify this after running a loop pass. This function adds the list of
642 // blocks denoted by BBs to this loops parent loop if required.
643 void addToParentLoopIfNeeded(ArrayRef<BasicBlock *> BBs);
645 // Some global state.
651 function_ref<void(Loop *, bool)> LPMAddNewLoop;
653 // Information about the original loop we started out with.
656 const SCEV *LatchTakenCount = nullptr;
657 BasicBlock *OriginalPreheader = nullptr;
659 // The preheader of the main loop. This may or may not be different from
660 // `OriginalPreheader'.
661 BasicBlock *MainLoopPreheader = nullptr;
663 // The range we need to run the main loop in.
664 InductiveRangeCheck::Range Range;
666 // The structure of the main loop (see comment at the beginning of this class
668 LoopStructure MainLoopStructure;
671 LoopConstrainer(Loop &L, LoopInfo &LI,
672 function_ref<void(Loop *, bool)> LPMAddNewLoop,
673 const LoopStructure &LS, ScalarEvolution &SE,
674 DominatorTree &DT, InductiveRangeCheck::Range R)
675 : F(*L.getHeader()->getParent()), Ctx(L.getHeader()->getContext()),
676 SE(SE), DT(DT), LI(LI), LPMAddNewLoop(LPMAddNewLoop), OriginalLoop(L),
677 Range(R), MainLoopStructure(LS) {}
679 // Entry point for the algorithm. Returns true on success.
683 } // end anonymous namespace
685 void LoopConstrainer::replacePHIBlock(PHINode *PN, BasicBlock *Block,
686 BasicBlock *ReplaceBy) {
687 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
688 if (PN->getIncomingBlock(i) == Block)
689 PN->setIncomingBlock(i, ReplaceBy);
692 static bool CannotBeMaxInLoop(const SCEV *BoundSCEV, Loop *L,
693 ScalarEvolution &SE, bool Signed) {
694 unsigned BitWidth = cast<IntegerType>(BoundSCEV->getType())->getBitWidth();
695 APInt Max = Signed ? APInt::getSignedMaxValue(BitWidth) :
696 APInt::getMaxValue(BitWidth);
697 auto Predicate = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
698 return SE.isAvailableAtLoopEntry(BoundSCEV, L) &&
699 SE.isLoopEntryGuardedByCond(L, Predicate, BoundSCEV,
700 SE.getConstant(Max));
703 /// Given a loop with an deccreasing induction variable, is it possible to
704 /// safely calculate the bounds of a new loop using the given Predicate.
705 static bool isSafeDecreasingBound(const SCEV *Start,
706 const SCEV *BoundSCEV, const SCEV *Step,
707 ICmpInst::Predicate Pred,
708 unsigned LatchBrExitIdx,
709 Loop *L, ScalarEvolution &SE) {
710 if (Pred != ICmpInst::ICMP_SLT && Pred != ICmpInst::ICMP_SGT &&
711 Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_UGT)
714 if (!SE.isAvailableAtLoopEntry(BoundSCEV, L))
717 assert(SE.isKnownNegative(Step) && "expecting negative step");
719 LLVM_DEBUG(dbgs() << "irce: isSafeDecreasingBound 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)
725 LLVM_DEBUG(dbgs() << "irce: LatchExitBrIdx: " << LatchBrExitIdx << "\n");
727 bool IsSigned = ICmpInst::isSigned(Pred);
728 // The predicate that we need to check that the induction variable lies
730 ICmpInst::Predicate BoundPred =
731 IsSigned ? CmpInst::ICMP_SGT : CmpInst::ICMP_UGT;
733 if (LatchBrExitIdx == 1)
734 return SE.isLoopEntryGuardedByCond(L, BoundPred, Start, BoundSCEV);
736 assert(LatchBrExitIdx == 0 &&
737 "LatchBrExitIdx should be either 0 or 1");
739 const SCEV *StepPlusOne = SE.getAddExpr(Step, SE.getOne(Step->getType()));
740 unsigned BitWidth = cast<IntegerType>(BoundSCEV->getType())->getBitWidth();
741 APInt Min = IsSigned ? APInt::getSignedMinValue(BitWidth) :
742 APInt::getMinValue(BitWidth);
743 const SCEV *Limit = SE.getMinusSCEV(SE.getConstant(Min), StepPlusOne);
745 const SCEV *MinusOne =
746 SE.getMinusSCEV(BoundSCEV, SE.getOne(BoundSCEV->getType()));
748 return SE.isLoopEntryGuardedByCond(L, BoundPred, Start, MinusOne) &&
749 SE.isLoopEntryGuardedByCond(L, BoundPred, BoundSCEV, Limit);
753 /// Given a loop with an increasing induction variable, is it possible to
754 /// safely calculate the bounds of a new loop using the given Predicate.
755 static bool isSafeIncreasingBound(const SCEV *Start,
756 const SCEV *BoundSCEV, const SCEV *Step,
757 ICmpInst::Predicate Pred,
758 unsigned LatchBrExitIdx,
759 Loop *L, ScalarEvolution &SE) {
760 if (Pred != ICmpInst::ICMP_SLT && Pred != ICmpInst::ICMP_SGT &&
761 Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_UGT)
764 if (!SE.isAvailableAtLoopEntry(BoundSCEV, L))
767 LLVM_DEBUG(dbgs() << "irce: isSafeIncreasingBound with:\n");
768 LLVM_DEBUG(dbgs() << "irce: Start: " << *Start << "\n");
769 LLVM_DEBUG(dbgs() << "irce: Step: " << *Step << "\n");
770 LLVM_DEBUG(dbgs() << "irce: BoundSCEV: " << *BoundSCEV << "\n");
771 LLVM_DEBUG(dbgs() << "irce: Pred: " << ICmpInst::getPredicateName(Pred)
773 LLVM_DEBUG(dbgs() << "irce: LatchExitBrIdx: " << LatchBrExitIdx << "\n");
775 bool IsSigned = ICmpInst::isSigned(Pred);
776 // The predicate that we need to check that the induction variable lies
778 ICmpInst::Predicate BoundPred =
779 IsSigned ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
781 if (LatchBrExitIdx == 1)
782 return SE.isLoopEntryGuardedByCond(L, BoundPred, Start, BoundSCEV);
784 assert(LatchBrExitIdx == 0 && "LatchBrExitIdx should be 0 or 1");
786 const SCEV *StepMinusOne =
787 SE.getMinusSCEV(Step, SE.getOne(Step->getType()));
788 unsigned BitWidth = cast<IntegerType>(BoundSCEV->getType())->getBitWidth();
789 APInt Max = IsSigned ? APInt::getSignedMaxValue(BitWidth) :
790 APInt::getMaxValue(BitWidth);
791 const SCEV *Limit = SE.getMinusSCEV(SE.getConstant(Max), StepMinusOne);
793 return (SE.isLoopEntryGuardedByCond(L, BoundPred, Start,
794 SE.getAddExpr(BoundSCEV, Step)) &&
795 SE.isLoopEntryGuardedByCond(L, BoundPred, BoundSCEV, Limit));
798 static bool CannotBeMinInLoop(const SCEV *BoundSCEV, Loop *L,
799 ScalarEvolution &SE, bool Signed) {
800 unsigned BitWidth = cast<IntegerType>(BoundSCEV->getType())->getBitWidth();
801 APInt Min = Signed ? APInt::getSignedMinValue(BitWidth) :
802 APInt::getMinValue(BitWidth);
803 auto Predicate = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
804 return SE.isAvailableAtLoopEntry(BoundSCEV, L) &&
805 SE.isLoopEntryGuardedByCond(L, Predicate, BoundSCEV,
806 SE.getConstant(Min));
809 static bool isKnownNonNegativeInLoop(const SCEV *BoundSCEV, const Loop *L,
810 ScalarEvolution &SE) {
811 const SCEV *Zero = SE.getZero(BoundSCEV->getType());
812 return SE.isAvailableAtLoopEntry(BoundSCEV, L) &&
813 SE.isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SGE, BoundSCEV, Zero);
816 static bool isKnownNegativeInLoop(const SCEV *BoundSCEV, const Loop *L,
817 ScalarEvolution &SE) {
818 const SCEV *Zero = SE.getZero(BoundSCEV->getType());
819 return SE.isAvailableAtLoopEntry(BoundSCEV, L) &&
820 SE.isLoopEntryGuardedByCond(L, ICmpInst::ICMP_SLT, BoundSCEV, Zero);
823 Optional<LoopStructure>
824 LoopStructure::parseLoopStructure(ScalarEvolution &SE,
825 BranchProbabilityInfo *BPI, Loop &L,
826 const char *&FailureReason) {
827 if (!L.isLoopSimplifyForm()) {
828 FailureReason = "loop not in LoopSimplify form";
832 BasicBlock *Latch = L.getLoopLatch();
833 assert(Latch && "Simplified loops only have one latch!");
835 if (Latch->getTerminator()->getMetadata(ClonedLoopTag)) {
836 FailureReason = "loop has already been cloned";
840 if (!L.isLoopExiting(Latch)) {
841 FailureReason = "no loop latch";
845 BasicBlock *Header = L.getHeader();
846 BasicBlock *Preheader = L.getLoopPreheader();
848 FailureReason = "no preheader";
852 BranchInst *LatchBr = dyn_cast<BranchInst>(Latch->getTerminator());
853 if (!LatchBr || LatchBr->isUnconditional()) {
854 FailureReason = "latch terminator not conditional branch";
858 unsigned LatchBrExitIdx = LatchBr->getSuccessor(0) == Header ? 1 : 0;
860 BranchProbability ExitProbability =
861 BPI ? BPI->getEdgeProbability(LatchBr->getParent(), LatchBrExitIdx)
862 : BranchProbability::getZero();
864 if (!SkipProfitabilityChecks &&
865 ExitProbability > BranchProbability(1, MaxExitProbReciprocal)) {
866 FailureReason = "short running loop, not profitable";
870 ICmpInst *ICI = dyn_cast<ICmpInst>(LatchBr->getCondition());
871 if (!ICI || !isa<IntegerType>(ICI->getOperand(0)->getType())) {
872 FailureReason = "latch terminator branch not conditional on integral icmp";
876 const SCEV *LatchCount = SE.getExitCount(&L, Latch);
877 if (isa<SCEVCouldNotCompute>(LatchCount)) {
878 FailureReason = "could not compute latch count";
882 ICmpInst::Predicate Pred = ICI->getPredicate();
883 Value *LeftValue = ICI->getOperand(0);
884 const SCEV *LeftSCEV = SE.getSCEV(LeftValue);
885 IntegerType *IndVarTy = cast<IntegerType>(LeftValue->getType());
887 Value *RightValue = ICI->getOperand(1);
888 const SCEV *RightSCEV = SE.getSCEV(RightValue);
890 // We canonicalize `ICI` such that `LeftSCEV` is an add recurrence.
891 if (!isa<SCEVAddRecExpr>(LeftSCEV)) {
892 if (isa<SCEVAddRecExpr>(RightSCEV)) {
893 std::swap(LeftSCEV, RightSCEV);
894 std::swap(LeftValue, RightValue);
895 Pred = ICmpInst::getSwappedPredicate(Pred);
897 FailureReason = "no add recurrences in the icmp";
902 auto HasNoSignedWrap = [&](const SCEVAddRecExpr *AR) {
903 if (AR->getNoWrapFlags(SCEV::FlagNSW))
906 IntegerType *Ty = cast<IntegerType>(AR->getType());
907 IntegerType *WideTy =
908 IntegerType::get(Ty->getContext(), Ty->getBitWidth() * 2);
910 const SCEVAddRecExpr *ExtendAfterOp =
911 dyn_cast<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
913 const SCEV *ExtendedStart = SE.getSignExtendExpr(AR->getStart(), WideTy);
914 const SCEV *ExtendedStep =
915 SE.getSignExtendExpr(AR->getStepRecurrence(SE), WideTy);
917 bool NoSignedWrap = ExtendAfterOp->getStart() == ExtendedStart &&
918 ExtendAfterOp->getStepRecurrence(SE) == ExtendedStep;
924 // We may have proved this when computing the sign extension above.
925 return AR->getNoWrapFlags(SCEV::FlagNSW) != SCEV::FlagAnyWrap;
928 // `ICI` is interpreted as taking the backedge if the *next* value of the
929 // induction variable satisfies some constraint.
931 const SCEVAddRecExpr *IndVarBase = cast<SCEVAddRecExpr>(LeftSCEV);
932 if (!IndVarBase->isAffine()) {
933 FailureReason = "LHS in icmp not induction variable";
936 const SCEV* StepRec = IndVarBase->getStepRecurrence(SE);
937 if (!isa<SCEVConstant>(StepRec)) {
938 FailureReason = "LHS in icmp not induction variable";
941 ConstantInt *StepCI = cast<SCEVConstant>(StepRec)->getValue();
943 if (ICI->isEquality() && !HasNoSignedWrap(IndVarBase)) {
944 FailureReason = "LHS in icmp needs nsw for equality predicates";
948 assert(!StepCI->isZero() && "Zero step?");
949 bool IsIncreasing = !StepCI->isNegative();
950 bool IsSignedPredicate = ICmpInst::isSigned(Pred);
951 const SCEV *StartNext = IndVarBase->getStart();
952 const SCEV *Addend = SE.getNegativeSCEV(IndVarBase->getStepRecurrence(SE));
953 const SCEV *IndVarStart = SE.getAddExpr(StartNext, Addend);
954 const SCEV *Step = SE.getSCEV(StepCI);
956 ConstantInt *One = ConstantInt::get(IndVarTy, 1);
958 bool DecreasedRightValueByOne = false;
959 if (StepCI->isOne()) {
960 // Try to turn eq/ne predicates to those we can work with.
961 if (Pred == ICmpInst::ICMP_NE && LatchBrExitIdx == 1)
962 // while (++i != len) { while (++i < len) {
965 // If both parts are known non-negative, it is profitable to use
966 // unsigned comparison in increasing loop. This allows us to make the
967 // comparison check against "RightSCEV + 1" more optimistic.
968 if (isKnownNonNegativeInLoop(IndVarStart, &L, SE) &&
969 isKnownNonNegativeInLoop(RightSCEV, &L, SE))
970 Pred = ICmpInst::ICMP_ULT;
972 Pred = ICmpInst::ICMP_SLT;
973 else if (Pred == ICmpInst::ICMP_EQ && LatchBrExitIdx == 0) {
974 // while (true) { while (true) {
975 // if (++i == len) ---> if (++i > len - 1)
979 if (IndVarBase->getNoWrapFlags(SCEV::FlagNUW) &&
980 CannotBeMinInLoop(RightSCEV, &L, SE, /*Signed*/false)) {
981 Pred = ICmpInst::ICMP_UGT;
982 RightSCEV = SE.getMinusSCEV(RightSCEV,
983 SE.getOne(RightSCEV->getType()));
984 DecreasedRightValueByOne = true;
985 } else if (CannotBeMinInLoop(RightSCEV, &L, SE, /*Signed*/true)) {
986 Pred = ICmpInst::ICMP_SGT;
987 RightSCEV = SE.getMinusSCEV(RightSCEV,
988 SE.getOne(RightSCEV->getType()));
989 DecreasedRightValueByOne = true;
994 bool LTPred = (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_ULT);
995 bool GTPred = (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_UGT);
996 bool FoundExpectedPred =
997 (LTPred && LatchBrExitIdx == 1) || (GTPred && LatchBrExitIdx == 0);
999 if (!FoundExpectedPred) {
1000 FailureReason = "expected icmp slt semantically, found something else";
1004 IsSignedPredicate = ICmpInst::isSigned(Pred);
1005 if (!IsSignedPredicate && !AllowUnsignedLatchCondition) {
1006 FailureReason = "unsigned latch conditions are explicitly prohibited";
1010 if (!isSafeIncreasingBound(IndVarStart, RightSCEV, Step, Pred,
1011 LatchBrExitIdx, &L, SE)) {
1012 FailureReason = "Unsafe loop bounds";
1015 if (LatchBrExitIdx == 0) {
1016 // We need to increase the right value unless we have already decreased
1017 // it virtually when we replaced EQ with SGT.
1018 if (!DecreasedRightValueByOne) {
1019 IRBuilder<> B(Preheader->getTerminator());
1020 RightValue = B.CreateAdd(RightValue, One);
1023 assert(!DecreasedRightValueByOne &&
1024 "Right value can be decreased only for LatchBrExitIdx == 0!");
1027 bool IncreasedRightValueByOne = false;
1028 if (StepCI->isMinusOne()) {
1029 // Try to turn eq/ne predicates to those we can work with.
1030 if (Pred == ICmpInst::ICMP_NE && LatchBrExitIdx == 1)
1031 // while (--i != len) { while (--i > len) {
1034 // We intentionally don't turn the predicate into UGT even if we know
1035 // that both operands are non-negative, because it will only pessimize
1036 // our check against "RightSCEV - 1".
1037 Pred = ICmpInst::ICMP_SGT;
1038 else if (Pred == ICmpInst::ICMP_EQ && LatchBrExitIdx == 0) {
1039 // while (true) { while (true) {
1040 // if (--i == len) ---> if (--i < len + 1)
1044 if (IndVarBase->getNoWrapFlags(SCEV::FlagNUW) &&
1045 CannotBeMaxInLoop(RightSCEV, &L, SE, /* Signed */ false)) {
1046 Pred = ICmpInst::ICMP_ULT;
1047 RightSCEV = SE.getAddExpr(RightSCEV, SE.getOne(RightSCEV->getType()));
1048 IncreasedRightValueByOne = true;
1049 } else if (CannotBeMaxInLoop(RightSCEV, &L, SE, /* Signed */ true)) {
1050 Pred = ICmpInst::ICMP_SLT;
1051 RightSCEV = SE.getAddExpr(RightSCEV, SE.getOne(RightSCEV->getType()));
1052 IncreasedRightValueByOne = true;
1057 bool LTPred = (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_ULT);
1058 bool GTPred = (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_UGT);
1060 bool FoundExpectedPred =
1061 (GTPred && LatchBrExitIdx == 1) || (LTPred && LatchBrExitIdx == 0);
1063 if (!FoundExpectedPred) {
1064 FailureReason = "expected icmp sgt semantically, found something else";
1069 Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGT;
1071 if (!IsSignedPredicate && !AllowUnsignedLatchCondition) {
1072 FailureReason = "unsigned latch conditions are explicitly prohibited";
1076 if (!isSafeDecreasingBound(IndVarStart, RightSCEV, Step, Pred,
1077 LatchBrExitIdx, &L, SE)) {
1078 FailureReason = "Unsafe bounds";
1082 if (LatchBrExitIdx == 0) {
1083 // We need to decrease the right value unless we have already increased
1084 // it virtually when we replaced EQ with SLT.
1085 if (!IncreasedRightValueByOne) {
1086 IRBuilder<> B(Preheader->getTerminator());
1087 RightValue = B.CreateSub(RightValue, One);
1090 assert(!IncreasedRightValueByOne &&
1091 "Right value can be increased only for LatchBrExitIdx == 0!");
1094 BasicBlock *LatchExit = LatchBr->getSuccessor(LatchBrExitIdx);
1096 assert(SE.getLoopDisposition(LatchCount, &L) ==
1097 ScalarEvolution::LoopInvariant &&
1098 "loop variant exit count doesn't make sense!");
1100 assert(!L.contains(LatchExit) && "expected an exit block!");
1101 const DataLayout &DL = Preheader->getModule()->getDataLayout();
1102 Value *IndVarStartV =
1103 SCEVExpander(SE, DL, "irce")
1104 .expandCodeFor(IndVarStart, IndVarTy, Preheader->getTerminator());
1105 IndVarStartV->setName("indvar.start");
1107 LoopStructure Result;
1109 Result.Tag = "main";
1110 Result.Header = Header;
1111 Result.Latch = Latch;
1112 Result.LatchBr = LatchBr;
1113 Result.LatchExit = LatchExit;
1114 Result.LatchBrExitIdx = LatchBrExitIdx;
1115 Result.IndVarStart = IndVarStartV;
1116 Result.IndVarStep = StepCI;
1117 Result.IndVarBase = LeftValue;
1118 Result.IndVarIncreasing = IsIncreasing;
1119 Result.LoopExitAt = RightValue;
1120 Result.IsSignedPredicate = IsSignedPredicate;
1122 FailureReason = nullptr;
1127 Optional<LoopConstrainer::SubRanges>
1128 LoopConstrainer::calculateSubRanges(bool IsSignedPredicate) const {
1129 IntegerType *Ty = cast<IntegerType>(LatchTakenCount->getType());
1131 if (Range.getType() != Ty)
1134 LoopConstrainer::SubRanges Result;
1136 // I think we can be more aggressive here and make this nuw / nsw if the
1137 // addition that feeds into the icmp for the latch's terminating branch is nuw
1138 // / nsw. In any case, a wrapping 2's complement addition is safe.
1139 const SCEV *Start = SE.getSCEV(MainLoopStructure.IndVarStart);
1140 const SCEV *End = SE.getSCEV(MainLoopStructure.LoopExitAt);
1142 bool Increasing = MainLoopStructure.IndVarIncreasing;
1144 // We compute `Smallest` and `Greatest` such that [Smallest, Greatest), or
1145 // [Smallest, GreatestSeen] is the range of values the induction variable
1148 const SCEV *Smallest = nullptr, *Greatest = nullptr, *GreatestSeen = nullptr;
1150 const SCEV *One = SE.getOne(Ty);
1154 // No overflow, because the range [Smallest, GreatestSeen] is not empty.
1155 GreatestSeen = SE.getMinusSCEV(End, One);
1157 // These two computations may sign-overflow. Here is why that is okay:
1159 // We know that the induction variable does not sign-overflow on any
1160 // iteration except the last one, and it starts at `Start` and ends at
1161 // `End`, decrementing by one every time.
1163 // * if `Smallest` sign-overflows we know `End` is `INT_SMAX`. Since the
1164 // induction variable is decreasing we know that that the smallest value
1165 // the loop body is actually executed with is `INT_SMIN` == `Smallest`.
1167 // * if `Greatest` sign-overflows, we know it can only be `INT_SMIN`. In
1168 // that case, `Clamp` will always return `Smallest` and
1169 // [`Result.LowLimit`, `Result.HighLimit`) = [`Smallest`, `Smallest`)
1170 // will be an empty range. Returning an empty range is always safe.
1172 Smallest = SE.getAddExpr(End, One);
1173 Greatest = SE.getAddExpr(Start, One);
1174 GreatestSeen = Start;
1177 auto Clamp = [this, Smallest, Greatest, IsSignedPredicate](const SCEV *S) {
1178 return IsSignedPredicate
1179 ? SE.getSMaxExpr(Smallest, SE.getSMinExpr(Greatest, S))
1180 : SE.getUMaxExpr(Smallest, SE.getUMinExpr(Greatest, S));
1183 // In some cases we can prove that we don't need a pre or post loop.
1184 ICmpInst::Predicate PredLE =
1185 IsSignedPredicate ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1186 ICmpInst::Predicate PredLT =
1187 IsSignedPredicate ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1189 bool ProvablyNoPreloop =
1190 SE.isKnownPredicate(PredLE, Range.getBegin(), Smallest);
1191 if (!ProvablyNoPreloop)
1192 Result.LowLimit = Clamp(Range.getBegin());
1194 bool ProvablyNoPostLoop =
1195 SE.isKnownPredicate(PredLT, GreatestSeen, Range.getEnd());
1196 if (!ProvablyNoPostLoop)
1197 Result.HighLimit = Clamp(Range.getEnd());
1202 void LoopConstrainer::cloneLoop(LoopConstrainer::ClonedLoop &Result,
1203 const char *Tag) const {
1204 for (BasicBlock *BB : OriginalLoop.getBlocks()) {
1205 BasicBlock *Clone = CloneBasicBlock(BB, Result.Map, Twine(".") + Tag, &F);
1206 Result.Blocks.push_back(Clone);
1207 Result.Map[BB] = Clone;
1210 auto GetClonedValue = [&Result](Value *V) {
1211 assert(V && "null values not in domain!");
1212 auto It = Result.Map.find(V);
1213 if (It == Result.Map.end())
1215 return static_cast<Value *>(It->second);
1219 cast<BasicBlock>(GetClonedValue(OriginalLoop.getLoopLatch()));
1220 ClonedLatch->getTerminator()->setMetadata(ClonedLoopTag,
1221 MDNode::get(Ctx, {}));
1223 Result.Structure = MainLoopStructure.map(GetClonedValue);
1224 Result.Structure.Tag = Tag;
1226 for (unsigned i = 0, e = Result.Blocks.size(); i != e; ++i) {
1227 BasicBlock *ClonedBB = Result.Blocks[i];
1228 BasicBlock *OriginalBB = OriginalLoop.getBlocks()[i];
1230 assert(Result.Map[OriginalBB] == ClonedBB && "invariant!");
1232 for (Instruction &I : *ClonedBB)
1233 RemapInstruction(&I, Result.Map,
1234 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
1236 // Exit blocks will now have one more predecessor and their PHI nodes need
1237 // to be edited to reflect that. No phi nodes need to be introduced because
1238 // the loop is in LCSSA.
1240 for (auto *SBB : successors(OriginalBB)) {
1241 if (OriginalLoop.contains(SBB))
1242 continue; // not an exit block
1244 for (PHINode &PN : SBB->phis()) {
1245 Value *OldIncoming = PN.getIncomingValueForBlock(OriginalBB);
1246 PN.addIncoming(GetClonedValue(OldIncoming), ClonedBB);
1252 LoopConstrainer::RewrittenRangeInfo LoopConstrainer::changeIterationSpaceEnd(
1253 const LoopStructure &LS, BasicBlock *Preheader, Value *ExitSubloopAt,
1254 BasicBlock *ContinuationBlock) const {
1255 // We start with a loop with a single latch:
1257 // +--------------------+
1261 // +--------+-----------+
1262 // | ----------------\
1264 // +--------v----v------+ |
1268 // +--------------------+ |
1272 // +--------------------+ |
1274 // | latch >----------/
1276 // +-------v------------+
1279 // | +--------------------+
1281 // +---> original exit |
1283 // +--------------------+
1285 // We change the control flow to look like
1288 // +--------------------+
1290 // | preheader >-------------------------+
1292 // +--------v-----------+ |
1293 // | /-------------+ |
1295 // +--------v--v--------+ | |
1297 // | header | | +--------+ |
1299 // +--------------------+ | | +-----v-----v-----------+
1301 // | | | .pseudo.exit |
1303 // | | +-----------v-----------+
1306 // | | +--------v-------------+
1307 // +--------------------+ | | | |
1308 // | | | | | ContinuationBlock |
1309 // | latch >------+ | | |
1310 // | | | +----------------------+
1311 // +---------v----------+ |
1314 // | +---------------^-----+
1316 // +-----> .exit.selector |
1318 // +----------v----------+
1320 // +--------------------+ |
1322 // | original exit <----+
1324 // +--------------------+
1326 RewrittenRangeInfo RRI;
1328 BasicBlock *BBInsertLocation = LS.Latch->getNextNode();
1329 RRI.ExitSelector = BasicBlock::Create(Ctx, Twine(LS.Tag) + ".exit.selector",
1330 &F, BBInsertLocation);
1331 RRI.PseudoExit = BasicBlock::Create(Ctx, Twine(LS.Tag) + ".pseudo.exit", &F,
1334 BranchInst *PreheaderJump = cast<BranchInst>(Preheader->getTerminator());
1335 bool Increasing = LS.IndVarIncreasing;
1336 bool IsSignedPredicate = LS.IsSignedPredicate;
1338 IRBuilder<> B(PreheaderJump);
1340 // EnterLoopCond - is it okay to start executing this `LS'?
1341 Value *EnterLoopCond = nullptr;
1343 EnterLoopCond = IsSignedPredicate
1344 ? B.CreateICmpSLT(LS.IndVarStart, ExitSubloopAt)
1345 : B.CreateICmpULT(LS.IndVarStart, ExitSubloopAt);
1347 EnterLoopCond = IsSignedPredicate
1348 ? B.CreateICmpSGT(LS.IndVarStart, ExitSubloopAt)
1349 : B.CreateICmpUGT(LS.IndVarStart, ExitSubloopAt);
1351 B.CreateCondBr(EnterLoopCond, LS.Header, RRI.PseudoExit);
1352 PreheaderJump->eraseFromParent();
1354 LS.LatchBr->setSuccessor(LS.LatchBrExitIdx, RRI.ExitSelector);
1355 B.SetInsertPoint(LS.LatchBr);
1356 Value *TakeBackedgeLoopCond = nullptr;
1358 TakeBackedgeLoopCond = IsSignedPredicate
1359 ? B.CreateICmpSLT(LS.IndVarBase, ExitSubloopAt)
1360 : B.CreateICmpULT(LS.IndVarBase, ExitSubloopAt);
1362 TakeBackedgeLoopCond = IsSignedPredicate
1363 ? B.CreateICmpSGT(LS.IndVarBase, ExitSubloopAt)
1364 : B.CreateICmpUGT(LS.IndVarBase, ExitSubloopAt);
1365 Value *CondForBranch = LS.LatchBrExitIdx == 1
1366 ? TakeBackedgeLoopCond
1367 : B.CreateNot(TakeBackedgeLoopCond);
1369 LS.LatchBr->setCondition(CondForBranch);
1371 B.SetInsertPoint(RRI.ExitSelector);
1373 // IterationsLeft - are there any more iterations left, given the original
1374 // upper bound on the induction variable? If not, we branch to the "real"
1376 Value *IterationsLeft = nullptr;
1378 IterationsLeft = IsSignedPredicate
1379 ? B.CreateICmpSLT(LS.IndVarBase, LS.LoopExitAt)
1380 : B.CreateICmpULT(LS.IndVarBase, LS.LoopExitAt);
1382 IterationsLeft = IsSignedPredicate
1383 ? B.CreateICmpSGT(LS.IndVarBase, LS.LoopExitAt)
1384 : B.CreateICmpUGT(LS.IndVarBase, LS.LoopExitAt);
1385 B.CreateCondBr(IterationsLeft, RRI.PseudoExit, LS.LatchExit);
1387 BranchInst *BranchToContinuation =
1388 BranchInst::Create(ContinuationBlock, RRI.PseudoExit);
1390 // We emit PHI nodes into `RRI.PseudoExit' that compute the "latest" value of
1391 // each of the PHI nodes in the loop header. This feeds into the initial
1392 // value of the same PHI nodes if/when we continue execution.
1393 for (PHINode &PN : LS.Header->phis()) {
1394 PHINode *NewPHI = PHINode::Create(PN.getType(), 2, PN.getName() + ".copy",
1395 BranchToContinuation);
1397 NewPHI->addIncoming(PN.getIncomingValueForBlock(Preheader), Preheader);
1398 NewPHI->addIncoming(PN.getIncomingValueForBlock(LS.Latch),
1400 RRI.PHIValuesAtPseudoExit.push_back(NewPHI);
1403 RRI.IndVarEnd = PHINode::Create(LS.IndVarBase->getType(), 2, "indvar.end",
1404 BranchToContinuation);
1405 RRI.IndVarEnd->addIncoming(LS.IndVarStart, Preheader);
1406 RRI.IndVarEnd->addIncoming(LS.IndVarBase, RRI.ExitSelector);
1408 // The latch exit now has a branch from `RRI.ExitSelector' instead of
1409 // `LS.Latch'. The PHI nodes need to be updated to reflect that.
1410 for (PHINode &PN : LS.LatchExit->phis())
1411 replacePHIBlock(&PN, LS.Latch, RRI.ExitSelector);
1416 void LoopConstrainer::rewriteIncomingValuesForPHIs(
1417 LoopStructure &LS, BasicBlock *ContinuationBlock,
1418 const LoopConstrainer::RewrittenRangeInfo &RRI) const {
1419 unsigned PHIIndex = 0;
1420 for (PHINode &PN : LS.Header->phis())
1421 for (unsigned i = 0, e = PN.getNumIncomingValues(); i < e; ++i)
1422 if (PN.getIncomingBlock(i) == ContinuationBlock)
1423 PN.setIncomingValue(i, RRI.PHIValuesAtPseudoExit[PHIIndex++]);
1425 LS.IndVarStart = RRI.IndVarEnd;
1428 BasicBlock *LoopConstrainer::createPreheader(const LoopStructure &LS,
1429 BasicBlock *OldPreheader,
1430 const char *Tag) const {
1431 BasicBlock *Preheader = BasicBlock::Create(Ctx, Tag, &F, LS.Header);
1432 BranchInst::Create(LS.Header, Preheader);
1434 for (PHINode &PN : LS.Header->phis())
1435 for (unsigned i = 0, e = PN.getNumIncomingValues(); i < e; ++i)
1436 replacePHIBlock(&PN, OldPreheader, Preheader);
1441 void LoopConstrainer::addToParentLoopIfNeeded(ArrayRef<BasicBlock *> BBs) {
1442 Loop *ParentLoop = OriginalLoop.getParentLoop();
1446 for (BasicBlock *BB : BBs)
1447 ParentLoop->addBasicBlockToLoop(BB, LI);
1450 Loop *LoopConstrainer::createClonedLoopStructure(Loop *Original, Loop *Parent,
1451 ValueToValueMapTy &VM,
1453 Loop &New = *LI.AllocateLoop();
1455 Parent->addChildLoop(&New);
1457 LI.addTopLevelLoop(&New);
1458 LPMAddNewLoop(&New, IsSubloop);
1460 // Add all of the blocks in Original to the new loop.
1461 for (auto *BB : Original->blocks())
1462 if (LI.getLoopFor(BB) == Original)
1463 New.addBasicBlockToLoop(cast<BasicBlock>(VM[BB]), LI);
1465 // Add all of the subloops to the new loop.
1466 for (Loop *SubLoop : *Original)
1467 createClonedLoopStructure(SubLoop, &New, VM, /* IsSubloop */ true);
1472 bool LoopConstrainer::run() {
1473 BasicBlock *Preheader = nullptr;
1474 LatchTakenCount = SE.getExitCount(&OriginalLoop, MainLoopStructure.Latch);
1475 Preheader = OriginalLoop.getLoopPreheader();
1476 assert(!isa<SCEVCouldNotCompute>(LatchTakenCount) && Preheader != nullptr &&
1479 OriginalPreheader = Preheader;
1480 MainLoopPreheader = Preheader;
1482 bool IsSignedPredicate = MainLoopStructure.IsSignedPredicate;
1483 Optional<SubRanges> MaybeSR = calculateSubRanges(IsSignedPredicate);
1484 if (!MaybeSR.hasValue()) {
1485 LLVM_DEBUG(dbgs() << "irce: could not compute subranges\n");
1489 SubRanges SR = MaybeSR.getValue();
1490 bool Increasing = MainLoopStructure.IndVarIncreasing;
1492 cast<IntegerType>(MainLoopStructure.IndVarBase->getType());
1494 SCEVExpander Expander(SE, F.getParent()->getDataLayout(), "irce");
1495 Instruction *InsertPt = OriginalPreheader->getTerminator();
1497 // It would have been better to make `PreLoop' and `PostLoop'
1498 // `Optional<ClonedLoop>'s, but `ValueToValueMapTy' does not have a copy
1500 ClonedLoop PreLoop, PostLoop;
1502 Increasing ? SR.LowLimit.hasValue() : SR.HighLimit.hasValue();
1503 bool NeedsPostLoop =
1504 Increasing ? SR.HighLimit.hasValue() : SR.LowLimit.hasValue();
1506 Value *ExitPreLoopAt = nullptr;
1507 Value *ExitMainLoopAt = nullptr;
1508 const SCEVConstant *MinusOneS =
1509 cast<SCEVConstant>(SE.getConstant(IVTy, -1, true /* isSigned */));
1512 const SCEV *ExitPreLoopAtSCEV = nullptr;
1515 ExitPreLoopAtSCEV = *SR.LowLimit;
1517 if (CannotBeMinInLoop(*SR.HighLimit, &OriginalLoop, SE,
1519 ExitPreLoopAtSCEV = SE.getAddExpr(*SR.HighLimit, MinusOneS);
1521 LLVM_DEBUG(dbgs() << "irce: could not prove no-overflow when computing "
1522 << "preloop exit limit. HighLimit = "
1523 << *(*SR.HighLimit) << "\n");
1528 if (!isSafeToExpandAt(ExitPreLoopAtSCEV, InsertPt, SE)) {
1529 LLVM_DEBUG(dbgs() << "irce: could not prove that it is safe to expand the"
1530 << " preloop exit limit " << *ExitPreLoopAtSCEV
1531 << " at block " << InsertPt->getParent()->getName()
1536 ExitPreLoopAt = Expander.expandCodeFor(ExitPreLoopAtSCEV, IVTy, InsertPt);
1537 ExitPreLoopAt->setName("exit.preloop.at");
1540 if (NeedsPostLoop) {
1541 const SCEV *ExitMainLoopAtSCEV = nullptr;
1544 ExitMainLoopAtSCEV = *SR.HighLimit;
1546 if (CannotBeMinInLoop(*SR.LowLimit, &OriginalLoop, SE,
1548 ExitMainLoopAtSCEV = SE.getAddExpr(*SR.LowLimit, MinusOneS);
1550 LLVM_DEBUG(dbgs() << "irce: could not prove no-overflow when computing "
1551 << "mainloop exit limit. LowLimit = "
1552 << *(*SR.LowLimit) << "\n");
1557 if (!isSafeToExpandAt(ExitMainLoopAtSCEV, InsertPt, SE)) {
1558 LLVM_DEBUG(dbgs() << "irce: could not prove that it is safe to expand the"
1559 << " main loop exit limit " << *ExitMainLoopAtSCEV
1560 << " at block " << InsertPt->getParent()->getName()
1565 ExitMainLoopAt = Expander.expandCodeFor(ExitMainLoopAtSCEV, IVTy, InsertPt);
1566 ExitMainLoopAt->setName("exit.mainloop.at");
1569 // We clone these ahead of time so that we don't have to deal with changing
1570 // and temporarily invalid IR as we transform the loops.
1572 cloneLoop(PreLoop, "preloop");
1574 cloneLoop(PostLoop, "postloop");
1576 RewrittenRangeInfo PreLoopRRI;
1579 Preheader->getTerminator()->replaceUsesOfWith(MainLoopStructure.Header,
1580 PreLoop.Structure.Header);
1583 createPreheader(MainLoopStructure, Preheader, "mainloop");
1584 PreLoopRRI = changeIterationSpaceEnd(PreLoop.Structure, Preheader,
1585 ExitPreLoopAt, MainLoopPreheader);
1586 rewriteIncomingValuesForPHIs(MainLoopStructure, MainLoopPreheader,
1590 BasicBlock *PostLoopPreheader = nullptr;
1591 RewrittenRangeInfo PostLoopRRI;
1593 if (NeedsPostLoop) {
1595 createPreheader(PostLoop.Structure, Preheader, "postloop");
1596 PostLoopRRI = changeIterationSpaceEnd(MainLoopStructure, MainLoopPreheader,
1597 ExitMainLoopAt, PostLoopPreheader);
1598 rewriteIncomingValuesForPHIs(PostLoop.Structure, PostLoopPreheader,
1602 BasicBlock *NewMainLoopPreheader =
1603 MainLoopPreheader != Preheader ? MainLoopPreheader : nullptr;
1604 BasicBlock *NewBlocks[] = {PostLoopPreheader, PreLoopRRI.PseudoExit,
1605 PreLoopRRI.ExitSelector, PostLoopRRI.PseudoExit,
1606 PostLoopRRI.ExitSelector, NewMainLoopPreheader};
1608 // Some of the above may be nullptr, filter them out before passing to
1609 // addToParentLoopIfNeeded.
1611 std::remove(std::begin(NewBlocks), std::end(NewBlocks), nullptr);
1613 addToParentLoopIfNeeded(makeArrayRef(std::begin(NewBlocks), NewBlocksEnd));
1617 // We need to first add all the pre and post loop blocks into the loop
1618 // structures (as part of createClonedLoopStructure), and then update the
1619 // LCSSA form and LoopSimplifyForm. This is necessary for correctly updating
1620 // LI when LoopSimplifyForm is generated.
1621 Loop *PreL = nullptr, *PostL = nullptr;
1622 if (!PreLoop.Blocks.empty()) {
1623 PreL = createClonedLoopStructure(&OriginalLoop,
1624 OriginalLoop.getParentLoop(), PreLoop.Map,
1625 /* IsSubLoop */ false);
1628 if (!PostLoop.Blocks.empty()) {
1630 createClonedLoopStructure(&OriginalLoop, OriginalLoop.getParentLoop(),
1631 PostLoop.Map, /* IsSubLoop */ false);
1634 // This function canonicalizes the loop into Loop-Simplify and LCSSA forms.
1635 auto CanonicalizeLoop = [&] (Loop *L, bool IsOriginalLoop) {
1636 formLCSSARecursively(*L, DT, &LI, &SE);
1637 simplifyLoop(L, &DT, &LI, &SE, nullptr, true);
1638 // Pre/post loops are slow paths, we do not need to perform any loop
1639 // optimizations on them.
1640 if (!IsOriginalLoop)
1641 DisableAllLoopOptsOnLoop(*L);
1644 CanonicalizeLoop(PreL, false);
1646 CanonicalizeLoop(PostL, false);
1647 CanonicalizeLoop(&OriginalLoop, true);
1652 /// Computes and returns a range of values for the induction variable (IndVar)
1653 /// in which the range check can be safely elided. If it cannot compute such a
1654 /// range, returns None.
1655 Optional<InductiveRangeCheck::Range>
1656 InductiveRangeCheck::computeSafeIterationSpace(
1657 ScalarEvolution &SE, const SCEVAddRecExpr *IndVar,
1658 bool IsLatchSigned) const {
1659 // IndVar is of the form "A + B * I" (where "I" is the canonical induction
1660 // variable, that may or may not exist as a real llvm::Value in the loop) and
1661 // this inductive range check is a range check on the "C + D * I" ("C" is
1662 // getBegin() and "D" is getStep()). We rewrite the value being range
1663 // checked to "M + N * IndVar" where "N" = "D * B^(-1)" and "M" = "C - NA".
1665 // The actual inequalities we solve are of the form
1667 // 0 <= M + 1 * IndVar < L given L >= 0 (i.e. N == 1)
1669 // Here L stands for upper limit of the safe iteration space.
1670 // The inequality is satisfied by (0 - M) <= IndVar < (L - M). To avoid
1671 // overflows when calculating (0 - M) and (L - M) we, depending on type of
1672 // IV's iteration space, limit the calculations by borders of the iteration
1673 // space. For example, if IndVar is unsigned, (0 - M) overflows for any M > 0.
1674 // If we figured out that "anything greater than (-M) is safe", we strengthen
1675 // this to "everything greater than 0 is safe", assuming that values between
1676 // -M and 0 just do not exist in unsigned iteration space, and we don't want
1677 // to deal with overflown values.
1679 if (!IndVar->isAffine())
1682 const SCEV *A = IndVar->getStart();
1683 const SCEVConstant *B = dyn_cast<SCEVConstant>(IndVar->getStepRecurrence(SE));
1686 assert(!B->isZero() && "Recurrence with zero step?");
1688 const SCEV *C = getBegin();
1689 const SCEVConstant *D = dyn_cast<SCEVConstant>(getStep());
1693 assert(!D->getValue()->isZero() && "Recurrence with zero step?");
1694 unsigned BitWidth = cast<IntegerType>(IndVar->getType())->getBitWidth();
1695 const SCEV *SIntMax = SE.getConstant(APInt::getSignedMaxValue(BitWidth));
1697 // Subtract Y from X so that it does not go through border of the IV
1698 // iteration space. Mathematically, it is equivalent to:
1700 // ClampedSubtract(X, Y) = min(max(X - Y, INT_MIN), INT_MAX). [1]
1702 // In [1], 'X - Y' is a mathematical subtraction (result is not bounded to
1703 // any width of bit grid). But after we take min/max, the result is
1704 // guaranteed to be within [INT_MIN, INT_MAX].
1706 // In [1], INT_MAX and INT_MIN are respectively signed and unsigned max/min
1707 // values, depending on type of latch condition that defines IV iteration
1709 auto ClampedSubtract = [&](const SCEV *X, const SCEV *Y) {
1710 // FIXME: The current implementation assumes that X is in [0, SINT_MAX].
1711 // This is required to ensure that SINT_MAX - X does not overflow signed and
1712 // that X - Y does not overflow unsigned if Y is negative. Can we lift this
1713 // restriction and make it work for negative X either?
1714 if (IsLatchSigned) {
1715 // X is a number from signed range, Y is interpreted as signed.
1716 // Even if Y is SINT_MAX, (X - Y) does not reach SINT_MIN. So the only
1717 // thing we should care about is that we didn't cross SINT_MAX.
1718 // So, if Y is positive, we subtract Y safely.
1719 // Rule 1: Y > 0 ---> Y.
1720 // If 0 <= -Y <= (SINT_MAX - X), we subtract Y safely.
1721 // Rule 2: Y >=s (X - SINT_MAX) ---> Y.
1722 // If 0 <= (SINT_MAX - X) < -Y, we can only subtract (X - SINT_MAX).
1723 // Rule 3: Y <s (X - SINT_MAX) ---> (X - SINT_MAX).
1724 // It gives us smax(Y, X - SINT_MAX) to subtract in all cases.
1725 const SCEV *XMinusSIntMax = SE.getMinusSCEV(X, SIntMax);
1726 return SE.getMinusSCEV(X, SE.getSMaxExpr(Y, XMinusSIntMax),
1729 // X is a number from unsigned range, Y is interpreted as signed.
1730 // Even if Y is SINT_MIN, (X - Y) does not reach UINT_MAX. So the only
1731 // thing we should care about is that we didn't cross zero.
1732 // So, if Y is negative, we subtract Y safely.
1733 // Rule 1: Y <s 0 ---> Y.
1734 // If 0 <= Y <= X, we subtract Y safely.
1735 // Rule 2: Y <=s X ---> Y.
1736 // If 0 <= X < Y, we should stop at 0 and can only subtract X.
1737 // Rule 3: Y >s X ---> X.
1738 // It gives us smin(X, Y) to subtract in all cases.
1739 return SE.getMinusSCEV(X, SE.getSMinExpr(X, Y), SCEV::FlagNUW);
1741 const SCEV *M = SE.getMinusSCEV(C, A);
1742 const SCEV *Zero = SE.getZero(M->getType());
1744 // This function returns SCEV equal to 1 if X is non-negative 0 otherwise.
1745 auto SCEVCheckNonNegative = [&](const SCEV *X) {
1746 const Loop *L = IndVar->getLoop();
1747 const SCEV *One = SE.getOne(X->getType());
1748 // Can we trivially prove that X is a non-negative or negative value?
1749 if (isKnownNonNegativeInLoop(X, L, SE))
1751 else if (isKnownNegativeInLoop(X, L, SE))
1753 // If not, we will have to figure it out during the execution.
1754 // Function smax(smin(X, 0), -1) + 1 equals to 1 if X >= 0 and 0 if X < 0.
1755 const SCEV *NegOne = SE.getNegativeSCEV(One);
1756 return SE.getAddExpr(SE.getSMaxExpr(SE.getSMinExpr(X, Zero), NegOne), One);
1758 // FIXME: Current implementation of ClampedSubtract implicitly assumes that
1759 // X is non-negative (in sense of a signed value). We need to re-implement
1760 // this function in a way that it will correctly handle negative X as well.
1761 // We use it twice: for X = 0 everything is fine, but for X = getEnd() we can
1762 // end up with a negative X and produce wrong results. So currently we ensure
1763 // that if getEnd() is negative then both ends of the safe range are zero.
1764 // Note that this may pessimize elimination of unsigned range checks against
1766 const SCEV *REnd = getEnd();
1767 const SCEV *EndIsNonNegative = SCEVCheckNonNegative(REnd);
1769 const SCEV *Begin = SE.getMulExpr(ClampedSubtract(Zero, M), EndIsNonNegative);
1770 const SCEV *End = SE.getMulExpr(ClampedSubtract(REnd, M), EndIsNonNegative);
1771 return InductiveRangeCheck::Range(Begin, End);
1774 static Optional<InductiveRangeCheck::Range>
1775 IntersectSignedRange(ScalarEvolution &SE,
1776 const Optional<InductiveRangeCheck::Range> &R1,
1777 const InductiveRangeCheck::Range &R2) {
1778 if (R2.isEmpty(SE, /* IsSigned */ true))
1782 auto &R1Value = R1.getValue();
1783 // We never return empty ranges from this function, and R1 is supposed to be
1784 // a result of intersection. Thus, R1 is never empty.
1785 assert(!R1Value.isEmpty(SE, /* IsSigned */ true) &&
1786 "We should never have empty R1!");
1788 // TODO: we could widen the smaller range and have this work; but for now we
1789 // bail out to keep things simple.
1790 if (R1Value.getType() != R2.getType())
1793 const SCEV *NewBegin = SE.getSMaxExpr(R1Value.getBegin(), R2.getBegin());
1794 const SCEV *NewEnd = SE.getSMinExpr(R1Value.getEnd(), R2.getEnd());
1796 // If the resulting range is empty, just return None.
1797 auto Ret = InductiveRangeCheck::Range(NewBegin, NewEnd);
1798 if (Ret.isEmpty(SE, /* IsSigned */ true))
1803 static Optional<InductiveRangeCheck::Range>
1804 IntersectUnsignedRange(ScalarEvolution &SE,
1805 const Optional<InductiveRangeCheck::Range> &R1,
1806 const InductiveRangeCheck::Range &R2) {
1807 if (R2.isEmpty(SE, /* IsSigned */ false))
1811 auto &R1Value = R1.getValue();
1812 // We never return empty ranges from this function, and R1 is supposed to be
1813 // a result of intersection. Thus, R1 is never empty.
1814 assert(!R1Value.isEmpty(SE, /* IsSigned */ false) &&
1815 "We should never have empty R1!");
1817 // TODO: we could widen the smaller range and have this work; but for now we
1818 // bail out to keep things simple.
1819 if (R1Value.getType() != R2.getType())
1822 const SCEV *NewBegin = SE.getUMaxExpr(R1Value.getBegin(), R2.getBegin());
1823 const SCEV *NewEnd = SE.getUMinExpr(R1Value.getEnd(), R2.getEnd());
1825 // If the resulting range is empty, just return None.
1826 auto Ret = InductiveRangeCheck::Range(NewBegin, NewEnd);
1827 if (Ret.isEmpty(SE, /* IsSigned */ false))
1832 PreservedAnalyses IRCEPass::run(Loop &L, LoopAnalysisManager &AM,
1833 LoopStandardAnalysisResults &AR,
1835 Function *F = L.getHeader()->getParent();
1837 AM.getResult<FunctionAnalysisManagerLoopProxy>(L, AR).getManager();
1838 auto *BPI = FAM.getCachedResult<BranchProbabilityAnalysis>(*F);
1839 InductiveRangeCheckElimination IRCE(AR.SE, BPI, AR.DT, AR.LI);
1840 auto LPMAddNewLoop = [&U](Loop *NL, bool IsSubloop) {
1842 U.addSiblingLoops(NL);
1844 bool Changed = IRCE.run(&L, LPMAddNewLoop);
1846 return PreservedAnalyses::all();
1848 return getLoopPassPreservedAnalyses();
1851 bool IRCELegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) {
1855 ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1856 BranchProbabilityInfo &BPI =
1857 getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI();
1858 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1859 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1860 InductiveRangeCheckElimination IRCE(SE, &BPI, DT, LI);
1861 auto LPMAddNewLoop = [&LPM](Loop *NL, bool /* IsSubLoop */) {
1864 return IRCE.run(L, LPMAddNewLoop);
1867 bool InductiveRangeCheckElimination::run(
1868 Loop *L, function_ref<void(Loop *, bool)> LPMAddNewLoop) {
1869 if (L->getBlocks().size() >= LoopSizeCutoff) {
1870 LLVM_DEBUG(dbgs() << "irce: giving up constraining loop, too large\n");
1874 BasicBlock *Preheader = L->getLoopPreheader();
1876 LLVM_DEBUG(dbgs() << "irce: loop has no preheader, leaving\n");
1880 LLVMContext &Context = Preheader->getContext();
1881 SmallVector<InductiveRangeCheck, 16> RangeChecks;
1883 for (auto BBI : L->getBlocks())
1884 if (BranchInst *TBI = dyn_cast<BranchInst>(BBI->getTerminator()))
1885 InductiveRangeCheck::extractRangeChecksFromBranch(TBI, L, SE, BPI,
1888 if (RangeChecks.empty())
1891 auto PrintRecognizedRangeChecks = [&](raw_ostream &OS) {
1892 OS << "irce: looking at loop "; L->print(OS);
1893 OS << "irce: loop has " << RangeChecks.size()
1894 << " inductive range checks: \n";
1895 for (InductiveRangeCheck &IRC : RangeChecks)
1899 LLVM_DEBUG(PrintRecognizedRangeChecks(dbgs()));
1901 if (PrintRangeChecks)
1902 PrintRecognizedRangeChecks(errs());
1904 const char *FailureReason = nullptr;
1905 Optional<LoopStructure> MaybeLoopStructure =
1906 LoopStructure::parseLoopStructure(SE, BPI, *L, FailureReason);
1907 if (!MaybeLoopStructure.hasValue()) {
1908 LLVM_DEBUG(dbgs() << "irce: could not parse loop structure: "
1909 << FailureReason << "\n";);
1912 LoopStructure LS = MaybeLoopStructure.getValue();
1913 const SCEVAddRecExpr *IndVar =
1914 cast<SCEVAddRecExpr>(SE.getMinusSCEV(SE.getSCEV(LS.IndVarBase), SE.getSCEV(LS.IndVarStep)));
1916 Optional<InductiveRangeCheck::Range> SafeIterRange;
1917 Instruction *ExprInsertPt = Preheader->getTerminator();
1919 SmallVector<InductiveRangeCheck, 4> RangeChecksToEliminate;
1920 // Basing on the type of latch predicate, we interpret the IV iteration range
1921 // as signed or unsigned range. We use different min/max functions (signed or
1922 // unsigned) when intersecting this range with safe iteration ranges implied
1924 auto IntersectRange =
1925 LS.IsSignedPredicate ? IntersectSignedRange : IntersectUnsignedRange;
1927 IRBuilder<> B(ExprInsertPt);
1928 for (InductiveRangeCheck &IRC : RangeChecks) {
1929 auto Result = IRC.computeSafeIterationSpace(SE, IndVar,
1930 LS.IsSignedPredicate);
1931 if (Result.hasValue()) {
1932 auto MaybeSafeIterRange =
1933 IntersectRange(SE, SafeIterRange, Result.getValue());
1934 if (MaybeSafeIterRange.hasValue()) {
1936 !MaybeSafeIterRange.getValue().isEmpty(SE, LS.IsSignedPredicate) &&
1937 "We should never return empty ranges!");
1938 RangeChecksToEliminate.push_back(IRC);
1939 SafeIterRange = MaybeSafeIterRange.getValue();
1944 if (!SafeIterRange.hasValue())
1947 LoopConstrainer LC(*L, LI, LPMAddNewLoop, LS, SE, DT,
1948 SafeIterRange.getValue());
1949 bool Changed = LC.run();
1952 auto PrintConstrainedLoopInfo = [L]() {
1953 dbgs() << "irce: in function ";
1954 dbgs() << L->getHeader()->getParent()->getName() << ": ";
1955 dbgs() << "constrained ";
1959 LLVM_DEBUG(PrintConstrainedLoopInfo());
1961 if (PrintChangedLoops)
1962 PrintConstrainedLoopInfo();
1964 // Optimize away the now-redundant range checks.
1966 for (InductiveRangeCheck &IRC : RangeChecksToEliminate) {
1967 ConstantInt *FoldedRangeCheck = IRC.getPassingDirection()
1968 ? ConstantInt::getTrue(Context)
1969 : ConstantInt::getFalse(Context);
1970 IRC.getCheckUse()->set(FoldedRangeCheck);
1977 Pass *llvm::createInductiveRangeCheckEliminationPass() {
1978 return new IRCELegacyPass();