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 //===----------------------------------------------------------------------===//
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
14 // len = < known positive >
15 // for (i = 0; i < n; i++) {
16 // if (0 <= i && i < len) {
19 // throw_out_of_bounds();
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
32 // throw_out_of_bounds();
35 // for (i = limit; i < n; i++) {
36 // if (0 <= i && i < len) {
39 // throw_out_of_bounds();
42 //===----------------------------------------------------------------------===//
44 #include "llvm/ADT/Optional.h"
45 #include "llvm/Analysis/BranchProbabilityInfo.h"
46 #include "llvm/Analysis/LoopInfo.h"
47 #include "llvm/Analysis/LoopPass.h"
48 #include "llvm/Analysis/ScalarEvolution.h"
49 #include "llvm/Analysis/ScalarEvolutionExpander.h"
50 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
51 #include "llvm/IR/Dominators.h"
52 #include "llvm/IR/Function.h"
53 #include "llvm/IR/IRBuilder.h"
54 #include "llvm/IR/Instructions.h"
55 #include "llvm/IR/PatternMatch.h"
56 #include "llvm/Pass.h"
57 #include "llvm/Support/Debug.h"
58 #include "llvm/Support/raw_ostream.h"
59 #include "llvm/Transforms/Scalar.h"
60 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
61 #include "llvm/Transforms/Utils/Cloning.h"
62 #include "llvm/Transforms/Utils/LoopSimplify.h"
63 #include "llvm/Transforms/Utils/LoopUtils.h"
67 static cl::opt<unsigned> LoopSizeCutoff("irce-loop-size-cutoff", cl::Hidden,
70 static cl::opt<bool> PrintChangedLoops("irce-print-changed-loops", cl::Hidden,
73 static cl::opt<bool> PrintRangeChecks("irce-print-range-checks", cl::Hidden,
76 static cl::opt<int> MaxExitProbReciprocal("irce-max-exit-prob-reciprocal",
77 cl::Hidden, cl::init(10));
79 static cl::opt<bool> SkipProfitabilityChecks("irce-skip-profitability-checks",
80 cl::Hidden, cl::init(false));
82 static const char *ClonedLoopTag = "irce.loop.clone";
84 #define DEBUG_TYPE "irce"
88 /// An inductive range check is conditional branch in a loop with
90 /// 1. a very cold successor (i.e. the branch jumps to that successor very
95 /// 2. a condition that is provably true for some contiguous range of values
96 /// taken by the containing loop's induction variable.
98 class InductiveRangeCheck {
99 // Classifies a range check
100 enum RangeCheckKind : unsigned {
101 // Range check of the form "0 <= I".
102 RANGE_CHECK_LOWER = 1,
104 // Range check of the form "I < L" where L is known positive.
105 RANGE_CHECK_UPPER = 2,
107 // The logical and of the RANGE_CHECK_LOWER and RANGE_CHECK_UPPER
109 RANGE_CHECK_BOTH = RANGE_CHECK_LOWER | RANGE_CHECK_UPPER,
111 // Unrecognized range check condition.
112 RANGE_CHECK_UNKNOWN = (unsigned)-1
115 static StringRef rangeCheckKindToStr(RangeCheckKind);
117 const SCEV *Offset = nullptr;
118 const SCEV *Scale = nullptr;
119 Value *Length = nullptr;
120 Use *CheckUse = nullptr;
121 RangeCheckKind Kind = RANGE_CHECK_UNKNOWN;
123 static RangeCheckKind parseRangeCheckICmp(Loop *L, ICmpInst *ICI,
124 ScalarEvolution &SE, Value *&Index,
128 extractRangeChecksFromCond(Loop *L, ScalarEvolution &SE, Use &ConditionUse,
129 SmallVectorImpl<InductiveRangeCheck> &Checks,
130 SmallPtrSetImpl<Value *> &Visited);
133 const SCEV *getOffset() const { return Offset; }
134 const SCEV *getScale() const { return Scale; }
135 Value *getLength() const { return Length; }
137 void print(raw_ostream &OS) const {
138 OS << "InductiveRangeCheck:\n";
139 OS << " Kind: " << rangeCheckKindToStr(Kind) << "\n";
149 OS << "\n CheckUse: ";
150 getCheckUse()->getUser()->print(OS);
151 OS << " Operand: " << getCheckUse()->getOperandNo() << "\n";
159 Use *getCheckUse() const { return CheckUse; }
161 /// Represents an signed integer range [Range.getBegin(), Range.getEnd()). If
162 /// R.getEnd() sle R.getBegin(), then R denotes the empty range.
169 Range(const SCEV *Begin, const SCEV *End) : Begin(Begin), End(End) {
170 assert(Begin->getType() == End->getType() && "ill-typed range!");
173 Type *getType() const { return Begin->getType(); }
174 const SCEV *getBegin() const { return Begin; }
175 const SCEV *getEnd() const { return End; }
178 /// This is the value the condition of the branch needs to evaluate to for the
179 /// branch to take the hot successor (see (1) above).
180 bool getPassingDirection() { return true; }
182 /// Computes a range for the induction variable (IndVar) in which the range
183 /// check is redundant and can be constant-folded away. The induction
184 /// variable is not required to be the canonical {0,+,1} induction variable.
185 Optional<Range> computeSafeIterationSpace(ScalarEvolution &SE,
186 const SCEVAddRecExpr *IndVar) const;
188 /// Parse out a set of inductive range checks from \p BI and append them to \p
191 /// NB! There may be conditions feeding into \p BI that aren't inductive range
192 /// checks, and hence don't end up in \p Checks.
194 extractRangeChecksFromBranch(BranchInst *BI, Loop *L, ScalarEvolution &SE,
195 BranchProbabilityInfo &BPI,
196 SmallVectorImpl<InductiveRangeCheck> &Checks);
199 class InductiveRangeCheckElimination : public LoopPass {
202 InductiveRangeCheckElimination() : LoopPass(ID) {
203 initializeInductiveRangeCheckEliminationPass(
204 *PassRegistry::getPassRegistry());
207 void getAnalysisUsage(AnalysisUsage &AU) const override {
208 AU.addRequired<BranchProbabilityInfoWrapperPass>();
209 getLoopAnalysisUsage(AU);
212 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
215 char InductiveRangeCheckElimination::ID = 0;
218 INITIALIZE_PASS_BEGIN(InductiveRangeCheckElimination, "irce",
219 "Inductive range check elimination", false, false)
220 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)
221 INITIALIZE_PASS_DEPENDENCY(LoopPass)
222 INITIALIZE_PASS_END(InductiveRangeCheckElimination, "irce",
223 "Inductive range check elimination", false, false)
225 StringRef InductiveRangeCheck::rangeCheckKindToStr(
226 InductiveRangeCheck::RangeCheckKind RCK) {
228 case InductiveRangeCheck::RANGE_CHECK_UNKNOWN:
229 return "RANGE_CHECK_UNKNOWN";
231 case InductiveRangeCheck::RANGE_CHECK_UPPER:
232 return "RANGE_CHECK_UPPER";
234 case InductiveRangeCheck::RANGE_CHECK_LOWER:
235 return "RANGE_CHECK_LOWER";
237 case InductiveRangeCheck::RANGE_CHECK_BOTH:
238 return "RANGE_CHECK_BOTH";
241 llvm_unreachable("unknown range check type!");
244 /// Parse a single ICmp instruction, `ICI`, into a range check. If `ICI` cannot
245 /// be interpreted as a range check, return `RANGE_CHECK_UNKNOWN` and set
246 /// `Index` and `Length` to `nullptr`. Otherwise set `Index` to the value being
247 /// range checked, and set `Length` to the upper limit `Index` is being range
248 /// checked with if (and only if) the range check type is stronger or equal to
249 /// RANGE_CHECK_UPPER.
251 InductiveRangeCheck::RangeCheckKind
252 InductiveRangeCheck::parseRangeCheckICmp(Loop *L, ICmpInst *ICI,
253 ScalarEvolution &SE, Value *&Index,
256 auto IsNonNegativeAndNotLoopVarying = [&SE, L](Value *V) {
257 const SCEV *S = SE.getSCEV(V);
258 if (isa<SCEVCouldNotCompute>(S))
261 return SE.getLoopDisposition(S, L) == ScalarEvolution::LoopInvariant &&
262 SE.isKnownNonNegative(S);
265 using namespace llvm::PatternMatch;
267 ICmpInst::Predicate Pred = ICI->getPredicate();
268 Value *LHS = ICI->getOperand(0);
269 Value *RHS = ICI->getOperand(1);
273 return RANGE_CHECK_UNKNOWN;
275 case ICmpInst::ICMP_SLE:
278 case ICmpInst::ICMP_SGE:
279 if (match(RHS, m_ConstantInt<0>())) {
281 return RANGE_CHECK_LOWER;
283 return RANGE_CHECK_UNKNOWN;
285 case ICmpInst::ICMP_SLT:
288 case ICmpInst::ICMP_SGT:
289 if (match(RHS, m_ConstantInt<-1>())) {
291 return RANGE_CHECK_LOWER;
294 if (IsNonNegativeAndNotLoopVarying(LHS)) {
297 return RANGE_CHECK_UPPER;
299 return RANGE_CHECK_UNKNOWN;
301 case ICmpInst::ICMP_ULT:
304 case ICmpInst::ICMP_UGT:
305 if (IsNonNegativeAndNotLoopVarying(LHS)) {
308 return RANGE_CHECK_BOTH;
310 return RANGE_CHECK_UNKNOWN;
313 llvm_unreachable("default clause returns!");
316 void InductiveRangeCheck::extractRangeChecksFromCond(
317 Loop *L, ScalarEvolution &SE, Use &ConditionUse,
318 SmallVectorImpl<InductiveRangeCheck> &Checks,
319 SmallPtrSetImpl<Value *> &Visited) {
320 using namespace llvm::PatternMatch;
322 Value *Condition = ConditionUse.get();
323 if (!Visited.insert(Condition).second)
326 if (match(Condition, m_And(m_Value(), m_Value()))) {
327 SmallVector<InductiveRangeCheck, 8> SubChecks;
328 extractRangeChecksFromCond(L, SE, cast<User>(Condition)->getOperandUse(0),
330 extractRangeChecksFromCond(L, SE, cast<User>(Condition)->getOperandUse(1),
333 if (SubChecks.size() == 2) {
334 // Handle a special case where we know how to merge two checks separately
335 // checking the upper and lower bounds into a full range check.
336 const auto &RChkA = SubChecks[0];
337 const auto &RChkB = SubChecks[1];
338 if ((RChkA.Length == RChkB.Length || !RChkA.Length || !RChkB.Length) &&
339 RChkA.Offset == RChkB.Offset && RChkA.Scale == RChkB.Scale) {
341 // If RChkA.Kind == RChkB.Kind then we just found two identical checks.
342 // But if one of them is a RANGE_CHECK_LOWER and the other is a
343 // RANGE_CHECK_UPPER (only possibility if they're different) then
344 // together they form a RANGE_CHECK_BOTH.
346 (InductiveRangeCheck::RangeCheckKind)(RChkA.Kind | RChkB.Kind);
347 SubChecks[0].Length = RChkA.Length ? RChkA.Length : RChkB.Length;
348 SubChecks[0].CheckUse = &ConditionUse;
350 // We updated one of the checks in place, now erase the other.
351 SubChecks.pop_back();
355 Checks.insert(Checks.end(), SubChecks.begin(), SubChecks.end());
359 ICmpInst *ICI = dyn_cast<ICmpInst>(Condition);
363 Value *Length = nullptr, *Index;
364 auto RCKind = parseRangeCheckICmp(L, ICI, SE, Index, Length);
365 if (RCKind == InductiveRangeCheck::RANGE_CHECK_UNKNOWN)
368 const auto *IndexAddRec = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Index));
370 IndexAddRec && (IndexAddRec->getLoop() == L) && IndexAddRec->isAffine();
375 InductiveRangeCheck IRC;
377 IRC.Offset = IndexAddRec->getStart();
378 IRC.Scale = IndexAddRec->getStepRecurrence(SE);
379 IRC.CheckUse = &ConditionUse;
381 Checks.push_back(IRC);
384 void InductiveRangeCheck::extractRangeChecksFromBranch(
385 BranchInst *BI, Loop *L, ScalarEvolution &SE, BranchProbabilityInfo &BPI,
386 SmallVectorImpl<InductiveRangeCheck> &Checks) {
388 if (BI->isUnconditional() || BI->getParent() == L->getLoopLatch())
391 BranchProbability LikelyTaken(15, 16);
393 if (!SkipProfitabilityChecks &&
394 BPI.getEdgeProbability(BI->getParent(), (unsigned)0) < LikelyTaken)
397 SmallPtrSet<Value *, 8> Visited;
398 InductiveRangeCheck::extractRangeChecksFromCond(L, SE, BI->getOperandUse(0),
402 // Add metadata to the loop L to disable loop optimizations. Callers need to
403 // confirm that optimizing loop L is not beneficial.
404 static void DisableAllLoopOptsOnLoop(Loop &L) {
405 // We do not care about any existing loopID related metadata for L, since we
406 // are setting all loop metadata to false.
407 LLVMContext &Context = L.getHeader()->getContext();
408 // Reserve first location for self reference to the LoopID metadata node.
409 MDNode *Dummy = MDNode::get(Context, {});
410 MDNode *DisableUnroll = MDNode::get(
411 Context, {MDString::get(Context, "llvm.loop.unroll.disable")});
413 ConstantAsMetadata::get(ConstantInt::get(Type::getInt1Ty(Context), 0));
414 MDNode *DisableVectorize = MDNode::get(
416 {MDString::get(Context, "llvm.loop.vectorize.enable"), FalseVal});
417 MDNode *DisableLICMVersioning = MDNode::get(
418 Context, {MDString::get(Context, "llvm.loop.licm_versioning.disable")});
419 MDNode *DisableDistribution= MDNode::get(
421 {MDString::get(Context, "llvm.loop.distribute.enable"), FalseVal});
423 MDNode::get(Context, {Dummy, DisableUnroll, DisableVectorize,
424 DisableLICMVersioning, DisableDistribution});
425 // Set operand 0 to refer to the loop id itself.
426 NewLoopID->replaceOperandWith(0, NewLoopID);
427 L.setLoopID(NewLoopID);
432 // Keeps track of the structure of a loop. This is similar to llvm::Loop,
433 // except that it is more lightweight and can track the state of a loop through
434 // changing and potentially invalid IR. This structure also formalizes the
435 // kinds of loops we can deal with -- ones that have a single latch that is also
436 // an exiting block *and* have a canonical induction variable.
437 struct LoopStructure {
443 // `Latch's terminator instruction is `LatchBr', and it's `LatchBrExitIdx'th
444 // successor is `LatchExit', the exit block of the loop.
446 BasicBlock *LatchExit;
447 unsigned LatchBrExitIdx;
449 // The loop represented by this instance of LoopStructure is semantically
452 // intN_ty inc = IndVarIncreasing ? 1 : -1;
453 // pred_ty predicate = IndVarIncreasing
454 // ? IsSignedPredicate ? ICMP_SLT : ICMP_ULT
455 // : IsSignedPredicate ? ICMP_SGT : ICMP_UGT;
458 // for (intN_ty iv = IndVarStart; predicate(IndVarBase, LoopExitAt);
462 // Here IndVarBase is either current or next value of the induction variable.
463 // in the former case, IsIndVarNext = false and IndVarBase points to the
464 // Phi node of the induction variable. Otherwise, IsIndVarNext = true and
465 // IndVarBase points to IV increment instruction.
472 bool IndVarIncreasing;
473 bool IsSignedPredicate;
477 : Tag(""), Header(nullptr), Latch(nullptr), LatchBr(nullptr),
478 LatchExit(nullptr), LatchBrExitIdx(-1), IndVarBase(nullptr),
479 IndVarStart(nullptr), IndVarStep(nullptr), LoopExitAt(nullptr),
480 IndVarIncreasing(false), IsSignedPredicate(true), IsIndVarNext(false) {}
482 template <typename M> LoopStructure map(M Map) const {
483 LoopStructure Result;
485 Result.Header = cast<BasicBlock>(Map(Header));
486 Result.Latch = cast<BasicBlock>(Map(Latch));
487 Result.LatchBr = cast<BranchInst>(Map(LatchBr));
488 Result.LatchExit = cast<BasicBlock>(Map(LatchExit));
489 Result.LatchBrExitIdx = LatchBrExitIdx;
490 Result.IndVarBase = Map(IndVarBase);
491 Result.IndVarStart = Map(IndVarStart);
492 Result.IndVarStep = Map(IndVarStep);
493 Result.LoopExitAt = Map(LoopExitAt);
494 Result.IndVarIncreasing = IndVarIncreasing;
495 Result.IsSignedPredicate = IsSignedPredicate;
496 Result.IsIndVarNext = IsIndVarNext;
500 static Optional<LoopStructure> parseLoopStructure(ScalarEvolution &,
501 BranchProbabilityInfo &BPI,
506 /// This class is used to constrain loops to run within a given iteration space.
507 /// The algorithm this class implements is given a Loop and a range [Begin,
508 /// End). The algorithm then tries to break out a "main loop" out of the loop
509 /// it is given in a way that the "main loop" runs with the induction variable
510 /// in a subset of [Begin, End). The algorithm emits appropriate pre and post
511 /// loops to run any remaining iterations. The pre loop runs any iterations in
512 /// which the induction variable is < Begin, and the post loop runs any
513 /// iterations in which the induction variable is >= End.
515 class LoopConstrainer {
516 // The representation of a clone of the original loop we started out with.
519 std::vector<BasicBlock *> Blocks;
521 // `Map` maps values in the clonee into values in the cloned version
522 ValueToValueMapTy Map;
524 // An instance of `LoopStructure` for the cloned loop
525 LoopStructure Structure;
528 // Result of rewriting the range of a loop. See changeIterationSpaceEnd for
529 // more details on what these fields mean.
530 struct RewrittenRangeInfo {
531 BasicBlock *PseudoExit;
532 BasicBlock *ExitSelector;
533 std::vector<PHINode *> PHIValuesAtPseudoExit;
537 : PseudoExit(nullptr), ExitSelector(nullptr), IndVarEnd(nullptr) {}
540 // Calculated subranges we restrict the iteration space of the main loop to.
541 // See the implementation of `calculateSubRanges' for more details on how
542 // these fields are computed. `LowLimit` is None if there is no restriction
543 // on low end of the restricted iteration space of the main loop. `HighLimit`
544 // is None if there is no restriction on high end of the restricted iteration
545 // space of the main loop.
548 Optional<const SCEV *> LowLimit;
549 Optional<const SCEV *> HighLimit;
552 // A utility function that does a `replaceUsesOfWith' on the incoming block
553 // set of a `PHINode' -- replaces instances of `Block' in the `PHINode's
554 // incoming block list with `ReplaceBy'.
555 static void replacePHIBlock(PHINode *PN, BasicBlock *Block,
556 BasicBlock *ReplaceBy);
558 // Compute a safe set of limits for the main loop to run in -- effectively the
559 // intersection of `Range' and the iteration space of the original loop.
560 // Return None if unable to compute the set of subranges.
562 Optional<SubRanges> calculateSubRanges(bool IsSignedPredicate) const;
564 // Clone `OriginalLoop' and return the result in CLResult. The IR after
565 // running `cloneLoop' is well formed except for the PHI nodes in CLResult --
566 // the PHI nodes say that there is an incoming edge from `OriginalPreheader`
567 // but there is no such edge.
569 void cloneLoop(ClonedLoop &CLResult, const char *Tag) const;
571 // Create the appropriate loop structure needed to describe a cloned copy of
572 // `Original`. The clone is described by `VM`.
573 Loop *createClonedLoopStructure(Loop *Original, Loop *Parent,
574 ValueToValueMapTy &VM);
576 // Rewrite the iteration space of the loop denoted by (LS, Preheader). The
577 // iteration space of the rewritten loop ends at ExitLoopAt. The start of the
578 // iteration space is not changed. `ExitLoopAt' is assumed to be slt
579 // `OriginalHeaderCount'.
581 // If there are iterations left to execute, control is made to jump to
582 // `ContinuationBlock', otherwise they take the normal loop exit. The
583 // returned `RewrittenRangeInfo' object is populated as follows:
585 // .PseudoExit is a basic block that unconditionally branches to
586 // `ContinuationBlock'.
588 // .ExitSelector is a basic block that decides, on exit from the loop,
589 // whether to branch to the "true" exit or to `PseudoExit'.
591 // .PHIValuesAtPseudoExit are PHINodes in `PseudoExit' that compute the value
592 // for each PHINode in the loop header on taking the pseudo exit.
594 // After changeIterationSpaceEnd, `Preheader' is no longer a legitimate
595 // preheader because it is made to branch to the loop header only
599 changeIterationSpaceEnd(const LoopStructure &LS, BasicBlock *Preheader,
601 BasicBlock *ContinuationBlock) const;
603 // The loop denoted by `LS' has `OldPreheader' as its preheader. This
604 // function creates a new preheader for `LS' and returns it.
606 BasicBlock *createPreheader(const LoopStructure &LS, BasicBlock *OldPreheader,
607 const char *Tag) const;
609 // `ContinuationBlockAndPreheader' was the continuation block for some call to
610 // `changeIterationSpaceEnd' and is the preheader to the loop denoted by `LS'.
611 // This function rewrites the PHI nodes in `LS.Header' to start with the
613 void rewriteIncomingValuesForPHIs(
614 LoopStructure &LS, BasicBlock *ContinuationBlockAndPreheader,
615 const LoopConstrainer::RewrittenRangeInfo &RRI) const;
617 // Even though we do not preserve any passes at this time, we at least need to
618 // keep the parent loop structure consistent. The `LPPassManager' seems to
619 // verify this after running a loop pass. This function adds the list of
620 // blocks denoted by BBs to this loops parent loop if required.
621 void addToParentLoopIfNeeded(ArrayRef<BasicBlock *> BBs);
623 // Some global state.
631 // Information about the original loop we started out with.
633 const SCEV *LatchTakenCount;
634 BasicBlock *OriginalPreheader;
636 // The preheader of the main loop. This may or may not be different from
637 // `OriginalPreheader'.
638 BasicBlock *MainLoopPreheader;
640 // The range we need to run the main loop in.
641 InductiveRangeCheck::Range Range;
643 // The structure of the main loop (see comment at the beginning of this class
645 LoopStructure MainLoopStructure;
648 LoopConstrainer(Loop &L, LoopInfo &LI, LPPassManager &LPM,
649 const LoopStructure &LS, ScalarEvolution &SE,
650 DominatorTree &DT, InductiveRangeCheck::Range R)
651 : F(*L.getHeader()->getParent()), Ctx(L.getHeader()->getContext()),
652 SE(SE), DT(DT), LPM(LPM), LI(LI), OriginalLoop(L),
653 LatchTakenCount(nullptr), OriginalPreheader(nullptr),
654 MainLoopPreheader(nullptr), Range(R), MainLoopStructure(LS) {}
656 // Entry point for the algorithm. Returns true on success.
662 void LoopConstrainer::replacePHIBlock(PHINode *PN, BasicBlock *Block,
663 BasicBlock *ReplaceBy) {
664 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
665 if (PN->getIncomingBlock(i) == Block)
666 PN->setIncomingBlock(i, ReplaceBy);
669 static bool CanBeMax(ScalarEvolution &SE, const SCEV *S, bool Signed) {
671 APInt::getSignedMaxValue(cast<IntegerType>(S->getType())->getBitWidth()) :
672 APInt::getMaxValue(cast<IntegerType>(S->getType())->getBitWidth());
673 return SE.getSignedRange(S).contains(Max) &&
674 SE.getUnsignedRange(S).contains(Max);
677 static bool SumCanReachMax(ScalarEvolution &SE, const SCEV *S1, const SCEV *S2,
679 // S1 < INT_MAX - S2 ===> S1 + S2 < INT_MAX.
680 assert(SE.isKnownNonNegative(S2) &&
681 "We expected the 2nd arg to be non-negative!");
682 const SCEV *Max = SE.getConstant(
683 Signed ? APInt::getSignedMaxValue(
684 cast<IntegerType>(S1->getType())->getBitWidth())
685 : APInt::getMaxValue(
686 cast<IntegerType>(S1->getType())->getBitWidth()));
687 const SCEV *CapForS1 = SE.getMinusSCEV(Max, S2);
688 return !SE.isKnownPredicate(Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
692 static bool CanBeMin(ScalarEvolution &SE, const SCEV *S, bool Signed) {
694 APInt::getSignedMinValue(cast<IntegerType>(S->getType())->getBitWidth()) :
695 APInt::getMinValue(cast<IntegerType>(S->getType())->getBitWidth());
696 return SE.getSignedRange(S).contains(Min) &&
697 SE.getUnsignedRange(S).contains(Min);
700 static bool SumCanReachMin(ScalarEvolution &SE, const SCEV *S1, const SCEV *S2,
702 // S1 > INT_MIN - S2 ===> S1 + S2 > INT_MIN.
703 assert(SE.isKnownNonPositive(S2) &&
704 "We expected the 2nd arg to be non-positive!");
705 const SCEV *Max = SE.getConstant(
706 Signed ? APInt::getSignedMinValue(
707 cast<IntegerType>(S1->getType())->getBitWidth())
708 : APInt::getMinValue(
709 cast<IntegerType>(S1->getType())->getBitWidth()));
710 const SCEV *CapForS1 = SE.getMinusSCEV(Max, S2);
711 return !SE.isKnownPredicate(Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT,
715 Optional<LoopStructure>
716 LoopStructure::parseLoopStructure(ScalarEvolution &SE,
717 BranchProbabilityInfo &BPI,
718 Loop &L, const char *&FailureReason) {
719 if (!L.isLoopSimplifyForm()) {
720 FailureReason = "loop not in LoopSimplify form";
724 BasicBlock *Latch = L.getLoopLatch();
725 assert(Latch && "Simplified loops only have one latch!");
727 if (Latch->getTerminator()->getMetadata(ClonedLoopTag)) {
728 FailureReason = "loop has already been cloned";
732 if (!L.isLoopExiting(Latch)) {
733 FailureReason = "no loop latch";
737 BasicBlock *Header = L.getHeader();
738 BasicBlock *Preheader = L.getLoopPreheader();
740 FailureReason = "no preheader";
744 BranchInst *LatchBr = dyn_cast<BranchInst>(Latch->getTerminator());
745 if (!LatchBr || LatchBr->isUnconditional()) {
746 FailureReason = "latch terminator not conditional branch";
750 unsigned LatchBrExitIdx = LatchBr->getSuccessor(0) == Header ? 1 : 0;
752 BranchProbability ExitProbability =
753 BPI.getEdgeProbability(LatchBr->getParent(), LatchBrExitIdx);
755 if (!SkipProfitabilityChecks &&
756 ExitProbability > BranchProbability(1, MaxExitProbReciprocal)) {
757 FailureReason = "short running loop, not profitable";
761 ICmpInst *ICI = dyn_cast<ICmpInst>(LatchBr->getCondition());
762 if (!ICI || !isa<IntegerType>(ICI->getOperand(0)->getType())) {
763 FailureReason = "latch terminator branch not conditional on integral icmp";
767 const SCEV *LatchCount = SE.getExitCount(&L, Latch);
768 if (isa<SCEVCouldNotCompute>(LatchCount)) {
769 FailureReason = "could not compute latch count";
773 ICmpInst::Predicate Pred = ICI->getPredicate();
774 Value *LeftValue = ICI->getOperand(0);
775 const SCEV *LeftSCEV = SE.getSCEV(LeftValue);
776 IntegerType *IndVarTy = cast<IntegerType>(LeftValue->getType());
778 Value *RightValue = ICI->getOperand(1);
779 const SCEV *RightSCEV = SE.getSCEV(RightValue);
781 // We canonicalize `ICI` such that `LeftSCEV` is an add recurrence.
782 if (!isa<SCEVAddRecExpr>(LeftSCEV)) {
783 if (isa<SCEVAddRecExpr>(RightSCEV)) {
784 std::swap(LeftSCEV, RightSCEV);
785 std::swap(LeftValue, RightValue);
786 Pred = ICmpInst::getSwappedPredicate(Pred);
788 FailureReason = "no add recurrences in the icmp";
793 auto HasNoSignedWrap = [&](const SCEVAddRecExpr *AR) {
794 if (AR->getNoWrapFlags(SCEV::FlagNSW))
797 IntegerType *Ty = cast<IntegerType>(AR->getType());
798 IntegerType *WideTy =
799 IntegerType::get(Ty->getContext(), Ty->getBitWidth() * 2);
801 const SCEVAddRecExpr *ExtendAfterOp =
802 dyn_cast<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
804 const SCEV *ExtendedStart = SE.getSignExtendExpr(AR->getStart(), WideTy);
805 const SCEV *ExtendedStep =
806 SE.getSignExtendExpr(AR->getStepRecurrence(SE), WideTy);
808 bool NoSignedWrap = ExtendAfterOp->getStart() == ExtendedStart &&
809 ExtendAfterOp->getStepRecurrence(SE) == ExtendedStep;
815 // We may have proved this when computing the sign extension above.
816 return AR->getNoWrapFlags(SCEV::FlagNSW) != SCEV::FlagAnyWrap;
819 // Here we check whether the suggested AddRec is an induction variable that
820 // can be handled (i.e. with known constant step), and if yes, calculate its
821 // step and identify whether it is increasing or decreasing.
822 auto IsInductionVar = [&](const SCEVAddRecExpr *AR, bool &IsIncreasing,
823 ConstantInt *&StepCI) {
827 // Currently we only work with induction variables that have been proved to
828 // not wrap. This restriction can potentially be lifted in the future.
830 if (!HasNoSignedWrap(AR))
833 if (const SCEVConstant *StepExpr =
834 dyn_cast<SCEVConstant>(AR->getStepRecurrence(SE))) {
835 StepCI = StepExpr->getValue();
836 assert(!StepCI->isZero() && "Zero step?");
837 IsIncreasing = !StepCI->isNegative();
844 // `ICI` can either be a comparison against IV or a comparison of IV.next.
845 // Depending on the interpretation, we calculate the start value differently.
847 // Pair {IndVarBase; IsIndVarNext} semantically designates whether the latch
848 // comparisons happens against the IV before or after its value is
849 // incremented. Two valid combinations for them are:
851 // 1) { phi [ iv.start, preheader ], [ iv.next, latch ]; false },
852 // 2) { iv.next; true }.
854 // The latch comparison happens against IndVarBase which can be either current
855 // or next value of the induction variable.
856 const SCEVAddRecExpr *IndVarBase = cast<SCEVAddRecExpr>(LeftSCEV);
857 bool IsIncreasing = false;
858 bool IsSignedPredicate = true;
859 bool IsIndVarNext = false;
861 if (!IsInductionVar(IndVarBase, IsIncreasing, StepCI)) {
862 FailureReason = "LHS in icmp not induction variable";
866 const SCEV *IndVarStart = nullptr;
867 // TODO: Currently we only handle comparison against IV, but we can extend
868 // this analysis to be able to deal with comparison against sext(iv) and such.
869 if (isa<PHINode>(LeftValue) &&
870 cast<PHINode>(LeftValue)->getParent() == Header)
871 // The comparison is made against current IV value.
872 IndVarStart = IndVarBase->getStart();
874 // Assume that the comparison is made against next IV value.
875 const SCEV *StartNext = IndVarBase->getStart();
876 const SCEV *Addend = SE.getNegativeSCEV(IndVarBase->getStepRecurrence(SE));
877 IndVarStart = SE.getAddExpr(StartNext, Addend);
880 const SCEV *Step = SE.getSCEV(StepCI);
882 ConstantInt *One = ConstantInt::get(IndVarTy, 1);
884 bool DecreasedRightValueByOne = false;
885 if (StepCI->isOne()) {
886 // Try to turn eq/ne predicates to those we can work with.
887 if (Pred == ICmpInst::ICMP_NE && LatchBrExitIdx == 1)
888 // while (++i != len) { while (++i < len) {
891 // If both parts are known non-negative, it is profitable to use
892 // unsigned comparison in increasing loop. This allows us to make the
893 // comparison check against "RightSCEV + 1" more optimistic.
894 if (SE.isKnownNonNegative(IndVarStart) &&
895 SE.isKnownNonNegative(RightSCEV))
896 Pred = ICmpInst::ICMP_ULT;
898 Pred = ICmpInst::ICMP_SLT;
899 else if (Pred == ICmpInst::ICMP_EQ && LatchBrExitIdx == 0 &&
900 !CanBeMin(SE, RightSCEV, /* IsSignedPredicate */ true)) {
901 // while (true) { while (true) {
902 // if (++i == len) ---> if (++i > len - 1)
906 // TODO: Insert ICMP_UGT if both are non-negative?
907 Pred = ICmpInst::ICMP_SGT;
908 RightSCEV = SE.getMinusSCEV(RightSCEV, SE.getOne(RightSCEV->getType()));
909 DecreasedRightValueByOne = true;
913 bool LTPred = (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_ULT);
914 bool GTPred = (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_UGT);
915 bool FoundExpectedPred =
916 (LTPred && LatchBrExitIdx == 1) || (GTPred && LatchBrExitIdx == 0);
918 if (!FoundExpectedPred) {
919 FailureReason = "expected icmp slt semantically, found something else";
924 Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGT;
925 // The predicate that we need to check that the induction variable lies
927 ICmpInst::Predicate BoundPred =
928 IsSignedPredicate ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
930 if (LatchBrExitIdx == 0) {
931 const SCEV *StepMinusOne = SE.getMinusSCEV(Step,
932 SE.getOne(Step->getType()));
933 if (SumCanReachMax(SE, RightSCEV, StepMinusOne, IsSignedPredicate)) {
934 // TODO: this restriction is easily removable -- we just have to
935 // remember that the icmp was an slt and not an sle.
936 FailureReason = "limit may overflow when coercing le to lt";
940 if (!SE.isLoopEntryGuardedByCond(
941 &L, BoundPred, IndVarStart,
942 SE.getAddExpr(RightSCEV, Step))) {
943 FailureReason = "Induction variable start not bounded by upper limit";
947 // We need to increase the right value unless we have already decreased
948 // it virtually when we replaced EQ with SGT.
949 if (!DecreasedRightValueByOne) {
950 IRBuilder<> B(Preheader->getTerminator());
951 RightValue = B.CreateAdd(RightValue, One);
954 if (!SE.isLoopEntryGuardedByCond(&L, BoundPred, IndVarStart, RightSCEV)) {
955 FailureReason = "Induction variable start not bounded by upper limit";
958 assert(!DecreasedRightValueByOne &&
959 "Right value can be decreased only for LatchBrExitIdx == 0!");
962 bool IncreasedRightValueByOne = false;
963 if (StepCI->isMinusOne()) {
964 // Try to turn eq/ne predicates to those we can work with.
965 if (Pred == ICmpInst::ICMP_NE && LatchBrExitIdx == 1)
966 // while (--i != len) { while (--i > len) {
969 // We intentionally don't turn the predicate into UGT even if we know
970 // that both operands are non-negative, because it will only pessimize
971 // our check against "RightSCEV - 1".
972 Pred = ICmpInst::ICMP_SGT;
973 else if (Pred == ICmpInst::ICMP_EQ && LatchBrExitIdx == 0 &&
974 !CanBeMax(SE, RightSCEV, /* IsSignedPredicate */ true)) {
975 // while (true) { while (true) {
976 // if (--i == len) ---> if (--i < len + 1)
980 // TODO: Insert ICMP_ULT if both are non-negative?
981 Pred = ICmpInst::ICMP_SLT;
982 RightSCEV = SE.getAddExpr(RightSCEV, SE.getOne(RightSCEV->getType()));
983 IncreasedRightValueByOne = true;
987 bool LTPred = (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_ULT);
988 bool GTPred = (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_UGT);
990 bool FoundExpectedPred =
991 (GTPred && LatchBrExitIdx == 1) || (LTPred && LatchBrExitIdx == 0);
993 if (!FoundExpectedPred) {
994 FailureReason = "expected icmp sgt semantically, found something else";
999 Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGT;
1000 // The predicate that we need to check that the induction variable lies
1002 ICmpInst::Predicate BoundPred =
1003 IsSignedPredicate ? CmpInst::ICMP_SGT : CmpInst::ICMP_UGT;
1005 if (LatchBrExitIdx == 0) {
1006 const SCEV *StepPlusOne = SE.getAddExpr(Step, SE.getOne(Step->getType()));
1007 if (SumCanReachMin(SE, RightSCEV, StepPlusOne, IsSignedPredicate)) {
1008 // TODO: this restriction is easily removable -- we just have to
1009 // remember that the icmp was an sgt and not an sge.
1010 FailureReason = "limit may overflow when coercing ge to gt";
1014 if (!SE.isLoopEntryGuardedByCond(
1015 &L, BoundPred, IndVarStart,
1016 SE.getMinusSCEV(RightSCEV, SE.getOne(RightSCEV->getType())))) {
1017 FailureReason = "Induction variable start not bounded by lower limit";
1021 // We need to decrease the right value unless we have already increased
1022 // it virtually when we replaced EQ with SLT.
1023 if (!IncreasedRightValueByOne) {
1024 IRBuilder<> B(Preheader->getTerminator());
1025 RightValue = B.CreateSub(RightValue, One);
1028 if (!SE.isLoopEntryGuardedByCond(&L, BoundPred, IndVarStart, RightSCEV)) {
1029 FailureReason = "Induction variable start not bounded by lower limit";
1032 assert(!IncreasedRightValueByOne &&
1033 "Right value can be increased only for LatchBrExitIdx == 0!");
1036 BasicBlock *LatchExit = LatchBr->getSuccessor(LatchBrExitIdx);
1038 assert(SE.getLoopDisposition(LatchCount, &L) ==
1039 ScalarEvolution::LoopInvariant &&
1040 "loop variant exit count doesn't make sense!");
1042 assert(!L.contains(LatchExit) && "expected an exit block!");
1043 const DataLayout &DL = Preheader->getModule()->getDataLayout();
1044 Value *IndVarStartV =
1045 SCEVExpander(SE, DL, "irce")
1046 .expandCodeFor(IndVarStart, IndVarTy, Preheader->getTerminator());
1047 IndVarStartV->setName("indvar.start");
1049 LoopStructure Result;
1051 Result.Tag = "main";
1052 Result.Header = Header;
1053 Result.Latch = Latch;
1054 Result.LatchBr = LatchBr;
1055 Result.LatchExit = LatchExit;
1056 Result.LatchBrExitIdx = LatchBrExitIdx;
1057 Result.IndVarStart = IndVarStartV;
1058 Result.IndVarStep = StepCI;
1059 Result.IndVarBase = LeftValue;
1060 Result.IndVarIncreasing = IsIncreasing;
1061 Result.LoopExitAt = RightValue;
1062 Result.IsSignedPredicate = IsSignedPredicate;
1063 Result.IsIndVarNext = IsIndVarNext;
1065 FailureReason = nullptr;
1070 Optional<LoopConstrainer::SubRanges>
1071 LoopConstrainer::calculateSubRanges(bool IsSignedPredicate) const {
1072 IntegerType *Ty = cast<IntegerType>(LatchTakenCount->getType());
1074 if (Range.getType() != Ty)
1077 LoopConstrainer::SubRanges Result;
1079 // I think we can be more aggressive here and make this nuw / nsw if the
1080 // addition that feeds into the icmp for the latch's terminating branch is nuw
1081 // / nsw. In any case, a wrapping 2's complement addition is safe.
1082 const SCEV *Start = SE.getSCEV(MainLoopStructure.IndVarStart);
1083 const SCEV *End = SE.getSCEV(MainLoopStructure.LoopExitAt);
1085 bool Increasing = MainLoopStructure.IndVarIncreasing;
1087 // We compute `Smallest` and `Greatest` such that [Smallest, Greatest), or
1088 // [Smallest, GreatestSeen] is the range of values the induction variable
1091 const SCEV *Smallest = nullptr, *Greatest = nullptr, *GreatestSeen = nullptr;
1093 const SCEV *One = SE.getOne(Ty);
1097 // No overflow, because the range [Smallest, GreatestSeen] is not empty.
1098 GreatestSeen = SE.getMinusSCEV(End, One);
1100 // These two computations may sign-overflow. Here is why that is okay:
1102 // We know that the induction variable does not sign-overflow on any
1103 // iteration except the last one, and it starts at `Start` and ends at
1104 // `End`, decrementing by one every time.
1106 // * if `Smallest` sign-overflows we know `End` is `INT_SMAX`. Since the
1107 // induction variable is decreasing we know that that the smallest value
1108 // the loop body is actually executed with is `INT_SMIN` == `Smallest`.
1110 // * if `Greatest` sign-overflows, we know it can only be `INT_SMIN`. In
1111 // that case, `Clamp` will always return `Smallest` and
1112 // [`Result.LowLimit`, `Result.HighLimit`) = [`Smallest`, `Smallest`)
1113 // will be an empty range. Returning an empty range is always safe.
1116 Smallest = SE.getAddExpr(End, One);
1117 Greatest = SE.getAddExpr(Start, One);
1118 GreatestSeen = Start;
1121 auto Clamp = [this, Smallest, Greatest, IsSignedPredicate](const SCEV *S) {
1122 bool MaybeNegativeValues = IsSignedPredicate || !SE.isKnownNonNegative(S);
1123 return MaybeNegativeValues
1124 ? SE.getSMaxExpr(Smallest, SE.getSMinExpr(Greatest, S))
1125 : SE.getUMaxExpr(Smallest, SE.getUMinExpr(Greatest, S));
1128 // In some cases we can prove that we don't need a pre or post loop.
1129 ICmpInst::Predicate PredLE =
1130 IsSignedPredicate ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1131 ICmpInst::Predicate PredLT =
1132 IsSignedPredicate ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1134 bool ProvablyNoPreloop =
1135 SE.isKnownPredicate(PredLE, Range.getBegin(), Smallest);
1136 if (!ProvablyNoPreloop)
1137 Result.LowLimit = Clamp(Range.getBegin());
1139 bool ProvablyNoPostLoop =
1140 SE.isKnownPredicate(PredLT, GreatestSeen, Range.getEnd());
1141 if (!ProvablyNoPostLoop)
1142 Result.HighLimit = Clamp(Range.getEnd());
1147 void LoopConstrainer::cloneLoop(LoopConstrainer::ClonedLoop &Result,
1148 const char *Tag) const {
1149 for (BasicBlock *BB : OriginalLoop.getBlocks()) {
1150 BasicBlock *Clone = CloneBasicBlock(BB, Result.Map, Twine(".") + Tag, &F);
1151 Result.Blocks.push_back(Clone);
1152 Result.Map[BB] = Clone;
1155 auto GetClonedValue = [&Result](Value *V) {
1156 assert(V && "null values not in domain!");
1157 auto It = Result.Map.find(V);
1158 if (It == Result.Map.end())
1160 return static_cast<Value *>(It->second);
1164 cast<BasicBlock>(GetClonedValue(OriginalLoop.getLoopLatch()));
1165 ClonedLatch->getTerminator()->setMetadata(ClonedLoopTag,
1166 MDNode::get(Ctx, {}));
1168 Result.Structure = MainLoopStructure.map(GetClonedValue);
1169 Result.Structure.Tag = Tag;
1171 for (unsigned i = 0, e = Result.Blocks.size(); i != e; ++i) {
1172 BasicBlock *ClonedBB = Result.Blocks[i];
1173 BasicBlock *OriginalBB = OriginalLoop.getBlocks()[i];
1175 assert(Result.Map[OriginalBB] == ClonedBB && "invariant!");
1177 for (Instruction &I : *ClonedBB)
1178 RemapInstruction(&I, Result.Map,
1179 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
1181 // Exit blocks will now have one more predecessor and their PHI nodes need
1182 // to be edited to reflect that. No phi nodes need to be introduced because
1183 // the loop is in LCSSA.
1185 for (auto *SBB : successors(OriginalBB)) {
1186 if (OriginalLoop.contains(SBB))
1187 continue; // not an exit block
1189 for (Instruction &I : *SBB) {
1190 auto *PN = dyn_cast<PHINode>(&I);
1194 Value *OldIncoming = PN->getIncomingValueForBlock(OriginalBB);
1195 PN->addIncoming(GetClonedValue(OldIncoming), ClonedBB);
1201 LoopConstrainer::RewrittenRangeInfo LoopConstrainer::changeIterationSpaceEnd(
1202 const LoopStructure &LS, BasicBlock *Preheader, Value *ExitSubloopAt,
1203 BasicBlock *ContinuationBlock) const {
1205 // We start with a loop with a single latch:
1207 // +--------------------+
1211 // +--------+-----------+
1212 // | ----------------\
1214 // +--------v----v------+ |
1218 // +--------------------+ |
1222 // +--------------------+ |
1224 // | latch >----------/
1226 // +-------v------------+
1229 // | +--------------------+
1231 // +---> original exit |
1233 // +--------------------+
1235 // We change the control flow to look like
1238 // +--------------------+
1240 // | preheader >-------------------------+
1242 // +--------v-----------+ |
1243 // | /-------------+ |
1245 // +--------v--v--------+ | |
1247 // | header | | +--------+ |
1249 // +--------------------+ | | +-----v-----v-----------+
1251 // | | | .pseudo.exit |
1253 // | | +-----------v-----------+
1256 // | | +--------v-------------+
1257 // +--------------------+ | | | |
1258 // | | | | | ContinuationBlock |
1259 // | latch >------+ | | |
1260 // | | | +----------------------+
1261 // +---------v----------+ |
1264 // | +---------------^-----+
1266 // +-----> .exit.selector |
1268 // +----------v----------+
1270 // +--------------------+ |
1272 // | original exit <----+
1274 // +--------------------+
1277 RewrittenRangeInfo RRI;
1279 BasicBlock *BBInsertLocation = LS.Latch->getNextNode();
1280 RRI.ExitSelector = BasicBlock::Create(Ctx, Twine(LS.Tag) + ".exit.selector",
1281 &F, BBInsertLocation);
1282 RRI.PseudoExit = BasicBlock::Create(Ctx, Twine(LS.Tag) + ".pseudo.exit", &F,
1285 BranchInst *PreheaderJump = cast<BranchInst>(Preheader->getTerminator());
1286 bool Increasing = LS.IndVarIncreasing;
1287 bool IsSignedPredicate = LS.IsSignedPredicate;
1289 IRBuilder<> B(PreheaderJump);
1291 // EnterLoopCond - is it okay to start executing this `LS'?
1292 Value *EnterLoopCond = nullptr;
1294 EnterLoopCond = IsSignedPredicate
1295 ? B.CreateICmpSLT(LS.IndVarStart, ExitSubloopAt)
1296 : B.CreateICmpULT(LS.IndVarStart, ExitSubloopAt);
1298 EnterLoopCond = IsSignedPredicate
1299 ? B.CreateICmpSGT(LS.IndVarStart, ExitSubloopAt)
1300 : B.CreateICmpUGT(LS.IndVarStart, ExitSubloopAt);
1302 B.CreateCondBr(EnterLoopCond, LS.Header, RRI.PseudoExit);
1303 PreheaderJump->eraseFromParent();
1305 LS.LatchBr->setSuccessor(LS.LatchBrExitIdx, RRI.ExitSelector);
1306 B.SetInsertPoint(LS.LatchBr);
1307 Value *TakeBackedgeLoopCond = nullptr;
1309 TakeBackedgeLoopCond = IsSignedPredicate
1310 ? B.CreateICmpSLT(LS.IndVarBase, ExitSubloopAt)
1311 : B.CreateICmpULT(LS.IndVarBase, ExitSubloopAt);
1313 TakeBackedgeLoopCond = IsSignedPredicate
1314 ? B.CreateICmpSGT(LS.IndVarBase, ExitSubloopAt)
1315 : B.CreateICmpUGT(LS.IndVarBase, ExitSubloopAt);
1316 Value *CondForBranch = LS.LatchBrExitIdx == 1
1317 ? TakeBackedgeLoopCond
1318 : B.CreateNot(TakeBackedgeLoopCond);
1320 LS.LatchBr->setCondition(CondForBranch);
1322 B.SetInsertPoint(RRI.ExitSelector);
1324 // IterationsLeft - are there any more iterations left, given the original
1325 // upper bound on the induction variable? If not, we branch to the "real"
1327 Value *IterationsLeft = nullptr;
1329 IterationsLeft = IsSignedPredicate
1330 ? B.CreateICmpSLT(LS.IndVarBase, LS.LoopExitAt)
1331 : B.CreateICmpULT(LS.IndVarBase, LS.LoopExitAt);
1333 IterationsLeft = IsSignedPredicate
1334 ? B.CreateICmpSGT(LS.IndVarBase, LS.LoopExitAt)
1335 : B.CreateICmpUGT(LS.IndVarBase, LS.LoopExitAt);
1336 B.CreateCondBr(IterationsLeft, RRI.PseudoExit, LS.LatchExit);
1338 BranchInst *BranchToContinuation =
1339 BranchInst::Create(ContinuationBlock, RRI.PseudoExit);
1341 // We emit PHI nodes into `RRI.PseudoExit' that compute the "latest" value of
1342 // each of the PHI nodes in the loop header. This feeds into the initial
1343 // value of the same PHI nodes if/when we continue execution.
1344 for (Instruction &I : *LS.Header) {
1345 auto *PN = dyn_cast<PHINode>(&I);
1349 PHINode *NewPHI = PHINode::Create(PN->getType(), 2, PN->getName() + ".copy",
1350 BranchToContinuation);
1352 NewPHI->addIncoming(PN->getIncomingValueForBlock(Preheader), Preheader);
1354 LS.IsIndVarNext ? PN->getIncomingValueForBlock(LS.Latch) : PN;
1355 NewPHI->addIncoming(FixupValue, RRI.ExitSelector);
1356 RRI.PHIValuesAtPseudoExit.push_back(NewPHI);
1359 RRI.IndVarEnd = PHINode::Create(LS.IndVarBase->getType(), 2, "indvar.end",
1360 BranchToContinuation);
1361 RRI.IndVarEnd->addIncoming(LS.IndVarStart, Preheader);
1362 RRI.IndVarEnd->addIncoming(LS.IndVarBase, RRI.ExitSelector);
1364 // The latch exit now has a branch from `RRI.ExitSelector' instead of
1365 // `LS.Latch'. The PHI nodes need to be updated to reflect that.
1366 for (Instruction &I : *LS.LatchExit) {
1367 if (PHINode *PN = dyn_cast<PHINode>(&I))
1368 replacePHIBlock(PN, LS.Latch, RRI.ExitSelector);
1376 void LoopConstrainer::rewriteIncomingValuesForPHIs(
1377 LoopStructure &LS, BasicBlock *ContinuationBlock,
1378 const LoopConstrainer::RewrittenRangeInfo &RRI) const {
1380 unsigned PHIIndex = 0;
1381 for (Instruction &I : *LS.Header) {
1382 auto *PN = dyn_cast<PHINode>(&I);
1386 for (unsigned i = 0, e = PN->getNumIncomingValues(); i < e; ++i)
1387 if (PN->getIncomingBlock(i) == ContinuationBlock)
1388 PN->setIncomingValue(i, RRI.PHIValuesAtPseudoExit[PHIIndex++]);
1391 LS.IndVarStart = RRI.IndVarEnd;
1394 BasicBlock *LoopConstrainer::createPreheader(const LoopStructure &LS,
1395 BasicBlock *OldPreheader,
1396 const char *Tag) const {
1398 BasicBlock *Preheader = BasicBlock::Create(Ctx, Tag, &F, LS.Header);
1399 BranchInst::Create(LS.Header, Preheader);
1401 for (Instruction &I : *LS.Header) {
1402 auto *PN = dyn_cast<PHINode>(&I);
1406 for (unsigned i = 0, e = PN->getNumIncomingValues(); i < e; ++i)
1407 replacePHIBlock(PN, OldPreheader, Preheader);
1413 void LoopConstrainer::addToParentLoopIfNeeded(ArrayRef<BasicBlock *> BBs) {
1414 Loop *ParentLoop = OriginalLoop.getParentLoop();
1418 for (BasicBlock *BB : BBs)
1419 ParentLoop->addBasicBlockToLoop(BB, LI);
1422 Loop *LoopConstrainer::createClonedLoopStructure(Loop *Original, Loop *Parent,
1423 ValueToValueMapTy &VM) {
1424 Loop &New = *new Loop();
1426 Parent->addChildLoop(&New);
1428 LI.addTopLevelLoop(&New);
1431 // Add all of the blocks in Original to the new loop.
1432 for (auto *BB : Original->blocks())
1433 if (LI.getLoopFor(BB) == Original)
1434 New.addBasicBlockToLoop(cast<BasicBlock>(VM[BB]), LI);
1436 // Add all of the subloops to the new loop.
1437 for (Loop *SubLoop : *Original)
1438 createClonedLoopStructure(SubLoop, &New, VM);
1443 bool LoopConstrainer::run() {
1444 BasicBlock *Preheader = nullptr;
1445 LatchTakenCount = SE.getExitCount(&OriginalLoop, MainLoopStructure.Latch);
1446 Preheader = OriginalLoop.getLoopPreheader();
1447 assert(!isa<SCEVCouldNotCompute>(LatchTakenCount) && Preheader != nullptr &&
1450 OriginalPreheader = Preheader;
1451 MainLoopPreheader = Preheader;
1453 bool IsSignedPredicate = MainLoopStructure.IsSignedPredicate;
1454 Optional<SubRanges> MaybeSR = calculateSubRanges(IsSignedPredicate);
1455 if (!MaybeSR.hasValue()) {
1456 DEBUG(dbgs() << "irce: could not compute subranges\n");
1460 SubRanges SR = MaybeSR.getValue();
1461 bool Increasing = MainLoopStructure.IndVarIncreasing;
1463 cast<IntegerType>(MainLoopStructure.IndVarBase->getType());
1465 SCEVExpander Expander(SE, F.getParent()->getDataLayout(), "irce");
1466 Instruction *InsertPt = OriginalPreheader->getTerminator();
1468 // It would have been better to make `PreLoop' and `PostLoop'
1469 // `Optional<ClonedLoop>'s, but `ValueToValueMapTy' does not have a copy
1471 ClonedLoop PreLoop, PostLoop;
1473 Increasing ? SR.LowLimit.hasValue() : SR.HighLimit.hasValue();
1474 bool NeedsPostLoop =
1475 Increasing ? SR.HighLimit.hasValue() : SR.LowLimit.hasValue();
1477 Value *ExitPreLoopAt = nullptr;
1478 Value *ExitMainLoopAt = nullptr;
1479 const SCEVConstant *MinusOneS =
1480 cast<SCEVConstant>(SE.getConstant(IVTy, -1, true /* isSigned */));
1483 const SCEV *ExitPreLoopAtSCEV = nullptr;
1486 ExitPreLoopAtSCEV = *SR.LowLimit;
1488 if (CanBeMin(SE, *SR.HighLimit, IsSignedPredicate)) {
1489 DEBUG(dbgs() << "irce: could not prove no-overflow when computing "
1490 << "preloop exit limit. HighLimit = " << *(*SR.HighLimit)
1494 ExitPreLoopAtSCEV = SE.getAddExpr(*SR.HighLimit, MinusOneS);
1497 ExitPreLoopAt = Expander.expandCodeFor(ExitPreLoopAtSCEV, IVTy, InsertPt);
1498 ExitPreLoopAt->setName("exit.preloop.at");
1501 if (NeedsPostLoop) {
1502 const SCEV *ExitMainLoopAtSCEV = nullptr;
1505 ExitMainLoopAtSCEV = *SR.HighLimit;
1507 if (CanBeMin(SE, *SR.LowLimit, IsSignedPredicate)) {
1508 DEBUG(dbgs() << "irce: could not prove no-overflow when computing "
1509 << "mainloop exit limit. LowLimit = " << *(*SR.LowLimit)
1513 ExitMainLoopAtSCEV = SE.getAddExpr(*SR.LowLimit, MinusOneS);
1516 ExitMainLoopAt = Expander.expandCodeFor(ExitMainLoopAtSCEV, IVTy, InsertPt);
1517 ExitMainLoopAt->setName("exit.mainloop.at");
1520 // We clone these ahead of time so that we don't have to deal with changing
1521 // and temporarily invalid IR as we transform the loops.
1523 cloneLoop(PreLoop, "preloop");
1525 cloneLoop(PostLoop, "postloop");
1527 RewrittenRangeInfo PreLoopRRI;
1530 Preheader->getTerminator()->replaceUsesOfWith(MainLoopStructure.Header,
1531 PreLoop.Structure.Header);
1534 createPreheader(MainLoopStructure, Preheader, "mainloop");
1535 PreLoopRRI = changeIterationSpaceEnd(PreLoop.Structure, Preheader,
1536 ExitPreLoopAt, MainLoopPreheader);
1537 rewriteIncomingValuesForPHIs(MainLoopStructure, MainLoopPreheader,
1541 BasicBlock *PostLoopPreheader = nullptr;
1542 RewrittenRangeInfo PostLoopRRI;
1544 if (NeedsPostLoop) {
1546 createPreheader(PostLoop.Structure, Preheader, "postloop");
1547 PostLoopRRI = changeIterationSpaceEnd(MainLoopStructure, MainLoopPreheader,
1548 ExitMainLoopAt, PostLoopPreheader);
1549 rewriteIncomingValuesForPHIs(PostLoop.Structure, PostLoopPreheader,
1553 BasicBlock *NewMainLoopPreheader =
1554 MainLoopPreheader != Preheader ? MainLoopPreheader : nullptr;
1555 BasicBlock *NewBlocks[] = {PostLoopPreheader, PreLoopRRI.PseudoExit,
1556 PreLoopRRI.ExitSelector, PostLoopRRI.PseudoExit,
1557 PostLoopRRI.ExitSelector, NewMainLoopPreheader};
1559 // Some of the above may be nullptr, filter them out before passing to
1560 // addToParentLoopIfNeeded.
1562 std::remove(std::begin(NewBlocks), std::end(NewBlocks), nullptr);
1564 addToParentLoopIfNeeded(makeArrayRef(std::begin(NewBlocks), NewBlocksEnd));
1568 // We need to first add all the pre and post loop blocks into the loop
1569 // structures (as part of createClonedLoopStructure), and then update the
1570 // LCSSA form and LoopSimplifyForm. This is necessary for correctly updating
1571 // LI when LoopSimplifyForm is generated.
1572 Loop *PreL = nullptr, *PostL = nullptr;
1573 if (!PreLoop.Blocks.empty()) {
1574 PreL = createClonedLoopStructure(
1575 &OriginalLoop, OriginalLoop.getParentLoop(), PreLoop.Map);
1578 if (!PostLoop.Blocks.empty()) {
1579 PostL = createClonedLoopStructure(
1580 &OriginalLoop, OriginalLoop.getParentLoop(), PostLoop.Map);
1583 // This function canonicalizes the loop into Loop-Simplify and LCSSA forms.
1584 auto CanonicalizeLoop = [&] (Loop *L, bool IsOriginalLoop) {
1585 formLCSSARecursively(*L, DT, &LI, &SE);
1586 simplifyLoop(L, &DT, &LI, &SE, nullptr, true);
1587 // Pre/post loops are slow paths, we do not need to perform any loop
1588 // optimizations on them.
1589 if (!IsOriginalLoop)
1590 DisableAllLoopOptsOnLoop(*L);
1593 CanonicalizeLoop(PreL, false);
1595 CanonicalizeLoop(PostL, false);
1596 CanonicalizeLoop(&OriginalLoop, true);
1601 /// Computes and returns a range of values for the induction variable (IndVar)
1602 /// in which the range check can be safely elided. If it cannot compute such a
1603 /// range, returns None.
1604 Optional<InductiveRangeCheck::Range>
1605 InductiveRangeCheck::computeSafeIterationSpace(
1606 ScalarEvolution &SE, const SCEVAddRecExpr *IndVar) const {
1607 // IndVar is of the form "A + B * I" (where "I" is the canonical induction
1608 // variable, that may or may not exist as a real llvm::Value in the loop) and
1609 // this inductive range check is a range check on the "C + D * I" ("C" is
1610 // getOffset() and "D" is getScale()). We rewrite the value being range
1611 // checked to "M + N * IndVar" where "N" = "D * B^(-1)" and "M" = "C - NA".
1613 // The actual inequalities we solve are of the form
1615 // 0 <= M + 1 * IndVar < L given L >= 0 (i.e. N == 1)
1617 // The inequality is satisfied by -M <= IndVar < (L - M) [^1]. All additions
1618 // and subtractions are twos-complement wrapping and comparisons are signed.
1622 // If there exists IndVar such that -M <= IndVar < (L - M) then it follows
1623 // that -M <= (-M + L) [== Eq. 1]. Since L >= 0, if (-M + L) sign-overflows
1624 // then (-M + L) < (-M). Hence by [Eq. 1], (-M + L) could not have
1627 // This means IndVar = t + (-M) for t in [0, L). Hence (IndVar + M) = t.
1628 // Hence 0 <= (IndVar + M) < L
1630 // [^1]: Note that the solution does _not_ apply if L < 0; consider values M =
1631 // 127, IndVar = 126 and L = -2 in an i8 world.
1633 if (!IndVar->isAffine())
1636 const SCEV *A = IndVar->getStart();
1637 const SCEVConstant *B = dyn_cast<SCEVConstant>(IndVar->getStepRecurrence(SE));
1640 assert(!B->isZero() && "Recurrence with zero step?");
1642 const SCEV *C = getOffset();
1643 const SCEVConstant *D = dyn_cast<SCEVConstant>(getScale());
1647 assert(!D->getValue()->isZero() && "Recurrence with zero step?");
1649 const SCEV *M = SE.getMinusSCEV(C, A);
1650 const SCEV *Begin = SE.getNegativeSCEV(M);
1651 const SCEV *UpperLimit = nullptr;
1653 // We strengthen "0 <= I" to "0 <= I < INT_SMAX" and "I < L" to "0 <= I < L".
1654 // We can potentially do much better here.
1655 if (Value *V = getLength()) {
1656 UpperLimit = SE.getSCEV(V);
1658 assert(Kind == InductiveRangeCheck::RANGE_CHECK_LOWER && "invariant!");
1659 unsigned BitWidth = cast<IntegerType>(IndVar->getType())->getBitWidth();
1660 UpperLimit = SE.getConstant(APInt::getSignedMaxValue(BitWidth));
1663 const SCEV *End = SE.getMinusSCEV(UpperLimit, M);
1664 return InductiveRangeCheck::Range(Begin, End);
1667 static Optional<InductiveRangeCheck::Range>
1668 IntersectRange(ScalarEvolution &SE,
1669 const Optional<InductiveRangeCheck::Range> &R1,
1670 const InductiveRangeCheck::Range &R2) {
1673 auto &R1Value = R1.getValue();
1675 // TODO: we could widen the smaller range and have this work; but for now we
1676 // bail out to keep things simple.
1677 if (R1Value.getType() != R2.getType())
1680 const SCEV *NewBegin = SE.getSMaxExpr(R1Value.getBegin(), R2.getBegin());
1681 const SCEV *NewEnd = SE.getSMinExpr(R1Value.getEnd(), R2.getEnd());
1683 return InductiveRangeCheck::Range(NewBegin, NewEnd);
1686 bool InductiveRangeCheckElimination::runOnLoop(Loop *L, LPPassManager &LPM) {
1690 if (L->getBlocks().size() >= LoopSizeCutoff) {
1691 DEBUG(dbgs() << "irce: giving up constraining loop, too large\n";);
1695 BasicBlock *Preheader = L->getLoopPreheader();
1697 DEBUG(dbgs() << "irce: loop has no preheader, leaving\n");
1701 LLVMContext &Context = Preheader->getContext();
1702 SmallVector<InductiveRangeCheck, 16> RangeChecks;
1703 ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1704 BranchProbabilityInfo &BPI =
1705 getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI();
1707 for (auto BBI : L->getBlocks())
1708 if (BranchInst *TBI = dyn_cast<BranchInst>(BBI->getTerminator()))
1709 InductiveRangeCheck::extractRangeChecksFromBranch(TBI, L, SE, BPI,
1712 if (RangeChecks.empty())
1715 auto PrintRecognizedRangeChecks = [&](raw_ostream &OS) {
1716 OS << "irce: looking at loop "; L->print(OS);
1717 OS << "irce: loop has " << RangeChecks.size()
1718 << " inductive range checks: \n";
1719 for (InductiveRangeCheck &IRC : RangeChecks)
1723 DEBUG(PrintRecognizedRangeChecks(dbgs()));
1725 if (PrintRangeChecks)
1726 PrintRecognizedRangeChecks(errs());
1728 const char *FailureReason = nullptr;
1729 Optional<LoopStructure> MaybeLoopStructure =
1730 LoopStructure::parseLoopStructure(SE, BPI, *L, FailureReason);
1731 if (!MaybeLoopStructure.hasValue()) {
1732 DEBUG(dbgs() << "irce: could not parse loop structure: " << FailureReason
1736 LoopStructure LS = MaybeLoopStructure.getValue();
1737 const SCEVAddRecExpr *IndVar =
1738 cast<SCEVAddRecExpr>(SE.getSCEV(LS.IndVarBase));
1739 if (LS.IsIndVarNext)
1740 IndVar = cast<SCEVAddRecExpr>(SE.getMinusSCEV(IndVar,
1741 SE.getSCEV(LS.IndVarStep)));
1743 Optional<InductiveRangeCheck::Range> SafeIterRange;
1744 Instruction *ExprInsertPt = Preheader->getTerminator();
1746 SmallVector<InductiveRangeCheck, 4> RangeChecksToEliminate;
1748 IRBuilder<> B(ExprInsertPt);
1749 for (InductiveRangeCheck &IRC : RangeChecks) {
1750 auto Result = IRC.computeSafeIterationSpace(SE, IndVar);
1751 if (Result.hasValue()) {
1752 auto MaybeSafeIterRange =
1753 IntersectRange(SE, SafeIterRange, Result.getValue());
1754 if (MaybeSafeIterRange.hasValue()) {
1755 RangeChecksToEliminate.push_back(IRC);
1756 SafeIterRange = MaybeSafeIterRange.getValue();
1761 if (!SafeIterRange.hasValue())
1764 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1765 LoopConstrainer LC(*L, getAnalysis<LoopInfoWrapperPass>().getLoopInfo(), LPM,
1766 LS, SE, DT, SafeIterRange.getValue());
1767 bool Changed = LC.run();
1770 auto PrintConstrainedLoopInfo = [L]() {
1771 dbgs() << "irce: in function ";
1772 dbgs() << L->getHeader()->getParent()->getName() << ": ";
1773 dbgs() << "constrained ";
1777 DEBUG(PrintConstrainedLoopInfo());
1779 if (PrintChangedLoops)
1780 PrintConstrainedLoopInfo();
1782 // Optimize away the now-redundant range checks.
1784 for (InductiveRangeCheck &IRC : RangeChecksToEliminate) {
1785 ConstantInt *FoldedRangeCheck = IRC.getPassingDirection()
1786 ? ConstantInt::getTrue(Context)
1787 : ConstantInt::getFalse(Context);
1788 IRC.getCheckUse()->set(FoldedRangeCheck);
1795 Pass *llvm::createInductiveRangeCheckEliminationPass() {
1796 return new InductiveRangeCheckElimination;