1 //===-- LoopPredication.cpp - Guard based loop predication pass -----------===//
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 LoopPredication pass tries to convert loop variant range checks to loop
11 // invariant by widening checks across loop iterations. For example, it will
14 // for (i = 0; i < n; i++) {
21 // for (i = 0; i < n; i++) {
22 // guard(n - 1 < len);
26 // After this transformation the condition of the guard is loop invariant, so
27 // loop-unswitch can later unswitch the loop by this condition which basically
28 // predicates the loop by the widened condition:
31 // for (i = 0; i < n; i++) {
37 // It's tempting to rely on SCEV here, but it has proven to be problematic.
38 // Generally the facts SCEV provides about the increment step of add
39 // recurrences are true if the backedge of the loop is taken, which implicitly
40 // assumes that the guard doesn't fail. Using these facts to optimize the
41 // guard results in a circular logic where the guard is optimized under the
42 // assumption that it never fails.
44 // For example, in the loop below the induction variable will be marked as nuw
45 // basing on the guard. Basing on nuw the guard predicate will be considered
46 // monotonic. Given a monotonic condition it's tempting to replace the induction
47 // variable in the condition with its value on the last iteration. But this
48 // transformation is not correct, e.g. e = 4, b = 5 breaks the loop.
50 // for (int i = b; i != e; i++)
53 // One of the ways to reason about this problem is to use an inductive proof
54 // approach. Given the loop:
64 // where B(x) and G(x) are predicates that map integers to booleans, we want a
65 // loop invariant expression M such the following program has the same semantics
76 // One solution for M is M = forall X . (G(X) && B(X)) => G(X + Step)
78 // Informal proof that the transformation above is correct:
80 // By the definition of guards we can rewrite the guard condition to:
83 // Let's prove that for each iteration of the loop:
85 // And the condition above can be simplified to G(Start) && M.
90 // Induction step. Assuming G(0) && M => G(I) on the subsequent
93 // B(I) is true because it's the backedge condition.
94 // G(I) is true because the backedge is guarded by this condition.
96 // So M = forall X . (G(X) && B(X)) => G(X + Step) implies G(I + Step).
98 // Note that we can use anything stronger than M, i.e. any condition which
101 // When S = 1 (i.e. forward iterating loop), the transformation is supported
103 // * The loop has a single latch with the condition of the form:
104 // B(X) = latchStart + X <pred> latchLimit,
105 // where <pred> is u<, u<=, s<, or s<=.
106 // * The guard condition is of the form
107 // G(X) = guardStart + X u< guardLimit
109 // For the ult latch comparison case M is:
110 // forall X . guardStart + X u< guardLimit && latchStart + X <u latchLimit =>
111 // guardStart + X + 1 u< guardLimit
113 // The only way the antecedent can be true and the consequent can be false is
115 // X == guardLimit - 1 - guardStart
116 // (and guardLimit is non-zero, but we won't use this latter fact).
117 // If X == guardLimit - 1 - guardStart then the second half of the antecedent is
118 // latchStart + guardLimit - 1 - guardStart u< latchLimit
119 // and its negation is
120 // latchStart + guardLimit - 1 - guardStart u>= latchLimit
122 // In other words, if
123 // latchLimit u<= latchStart + guardLimit - 1 - guardStart
125 // (the ranges below are written in ConstantRange notation, where [A, B) is the
126 // set for (I = A; I != B; I++ /*maywrap*/) yield(I);)
128 // forall X . guardStart + X u< guardLimit &&
129 // latchStart + X u< latchLimit =>
130 // guardStart + X + 1 u< guardLimit
131 // == forall X . guardStart + X u< guardLimit &&
132 // latchStart + X u< latchStart + guardLimit - 1 - guardStart =>
133 // guardStart + X + 1 u< guardLimit
134 // == forall X . (guardStart + X) in [0, guardLimit) &&
135 // (latchStart + X) in [0, latchStart + guardLimit - 1 - guardStart) =>
136 // (guardStart + X + 1) in [0, guardLimit)
137 // == forall X . X in [-guardStart, guardLimit - guardStart) &&
138 // X in [-latchStart, guardLimit - 1 - guardStart) =>
139 // X in [-guardStart - 1, guardLimit - guardStart - 1)
142 // So the widened condition is:
143 // guardStart u< guardLimit &&
144 // latchStart + guardLimit - 1 - guardStart u>= latchLimit
145 // Similarly for ule condition the widened condition is:
146 // guardStart u< guardLimit &&
147 // latchStart + guardLimit - 1 - guardStart u> latchLimit
148 // For slt condition the widened condition is:
149 // guardStart u< guardLimit &&
150 // latchStart + guardLimit - 1 - guardStart s>= latchLimit
151 // For sle condition the widened condition is:
152 // guardStart u< guardLimit &&
153 // latchStart + guardLimit - 1 - guardStart s> latchLimit
155 // When S = -1 (i.e. reverse iterating loop), the transformation is supported
157 // * The loop has a single latch with the condition of the form:
158 // B(X) = X <pred> latchLimit, where <pred> is u> or s>.
159 // * The guard condition is of the form
160 // G(X) = X - 1 u< guardLimit
162 // For the ugt latch comparison case M is:
163 // forall X. X-1 u< guardLimit and X u> latchLimit => X-2 u< guardLimit
165 // The only way the antecedent can be true and the consequent can be false is if
167 // If X == 1 then the second half of the antecedent is
168 // 1 u> latchLimit, and its negation is latchLimit u>= 1.
170 // So the widened condition is:
171 // guardStart u< guardLimit && latchLimit u>= 1.
172 // Similarly for sgt condition the widened condition is:
173 // guardStart u< guardLimit && latchLimit s>= 1.
174 //===----------------------------------------------------------------------===//
176 #include "llvm/Transforms/Scalar/LoopPredication.h"
177 #include "llvm/Analysis/LoopInfo.h"
178 #include "llvm/Analysis/LoopPass.h"
179 #include "llvm/Analysis/ScalarEvolution.h"
180 #include "llvm/Analysis/ScalarEvolutionExpander.h"
181 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
182 #include "llvm/IR/Function.h"
183 #include "llvm/IR/GlobalValue.h"
184 #include "llvm/IR/IntrinsicInst.h"
185 #include "llvm/IR/Module.h"
186 #include "llvm/IR/PatternMatch.h"
187 #include "llvm/Pass.h"
188 #include "llvm/Support/Debug.h"
189 #include "llvm/Transforms/Scalar.h"
190 #include "llvm/Transforms/Utils/LoopUtils.h"
192 #define DEBUG_TYPE "loop-predication"
194 using namespace llvm;
196 static cl::opt<bool> EnableIVTruncation("loop-predication-enable-iv-truncation",
197 cl::Hidden, cl::init(true));
199 static cl::opt<bool> EnableCountDownLoop("loop-predication-enable-count-down-loop",
200 cl::Hidden, cl::init(true));
202 class LoopPredication {
203 /// Represents an induction variable check:
204 /// icmp Pred, <induction variable>, <loop invariant limit>
206 ICmpInst::Predicate Pred;
207 const SCEVAddRecExpr *IV;
209 LoopICmp(ICmpInst::Predicate Pred, const SCEVAddRecExpr *IV,
211 : Pred(Pred), IV(IV), Limit(Limit) {}
214 dbgs() << "LoopICmp Pred = " << Pred << ", IV = " << *IV
215 << ", Limit = " << *Limit << "\n";
222 const DataLayout *DL;
223 BasicBlock *Preheader;
226 bool isSupportedStep(const SCEV* Step);
227 Optional<LoopICmp> parseLoopICmp(ICmpInst *ICI) {
228 return parseLoopICmp(ICI->getPredicate(), ICI->getOperand(0),
231 Optional<LoopICmp> parseLoopICmp(ICmpInst::Predicate Pred, Value *LHS,
234 Optional<LoopICmp> parseLoopLatchICmp();
236 bool CanExpand(const SCEV* S);
237 Value *expandCheck(SCEVExpander &Expander, IRBuilder<> &Builder,
238 ICmpInst::Predicate Pred, const SCEV *LHS, const SCEV *RHS,
239 Instruction *InsertAt);
241 Optional<Value *> widenICmpRangeCheck(ICmpInst *ICI, SCEVExpander &Expander,
242 IRBuilder<> &Builder);
243 Optional<Value *> widenICmpRangeCheckIncrementingLoop(LoopICmp LatchCheck,
245 SCEVExpander &Expander,
246 IRBuilder<> &Builder);
247 Optional<Value *> widenICmpRangeCheckDecrementingLoop(LoopICmp LatchCheck,
249 SCEVExpander &Expander,
250 IRBuilder<> &Builder);
251 bool widenGuardConditions(IntrinsicInst *II, SCEVExpander &Expander);
253 // When the IV type is wider than the range operand type, we can still do loop
254 // predication, by generating SCEVs for the range and latch that are of the
255 // same type. We achieve this by generating a SCEV truncate expression for the
256 // latch IV. This is done iff truncation of the IV is a safe operation,
257 // without loss of information.
258 // Another way to achieve this is by generating a wider type SCEV for the
259 // range check operand, however, this needs a more involved check that
260 // operands do not overflow. This can lead to loss of information when the
261 // range operand is of the form: add i32 %offset, %iv. We need to prove that
262 // sext(x + y) is same as sext(x) + sext(y).
263 // This function returns true if we can safely represent the IV type in
264 // the RangeCheckType without loss of information.
265 bool isSafeToTruncateWideIVType(Type *RangeCheckType);
266 // Return the loopLatchCheck corresponding to the RangeCheckType if safe to do
268 Optional<LoopICmp> generateLoopLatchCheck(Type *RangeCheckType);
270 LoopPredication(ScalarEvolution *SE) : SE(SE){};
271 bool runOnLoop(Loop *L);
274 class LoopPredicationLegacyPass : public LoopPass {
277 LoopPredicationLegacyPass() : LoopPass(ID) {
278 initializeLoopPredicationLegacyPassPass(*PassRegistry::getPassRegistry());
281 void getAnalysisUsage(AnalysisUsage &AU) const override {
282 getLoopAnalysisUsage(AU);
285 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
288 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
289 LoopPredication LP(SE);
290 return LP.runOnLoop(L);
294 char LoopPredicationLegacyPass::ID = 0;
295 } // end namespace llvm
297 INITIALIZE_PASS_BEGIN(LoopPredicationLegacyPass, "loop-predication",
298 "Loop predication", false, false)
299 INITIALIZE_PASS_DEPENDENCY(LoopPass)
300 INITIALIZE_PASS_END(LoopPredicationLegacyPass, "loop-predication",
301 "Loop predication", false, false)
303 Pass *llvm::createLoopPredicationPass() {
304 return new LoopPredicationLegacyPass();
307 PreservedAnalyses LoopPredicationPass::run(Loop &L, LoopAnalysisManager &AM,
308 LoopStandardAnalysisResults &AR,
310 LoopPredication LP(&AR.SE);
311 if (!LP.runOnLoop(&L))
312 return PreservedAnalyses::all();
314 return getLoopPassPreservedAnalyses();
317 Optional<LoopPredication::LoopICmp>
318 LoopPredication::parseLoopICmp(ICmpInst::Predicate Pred, Value *LHS,
320 const SCEV *LHSS = SE->getSCEV(LHS);
321 if (isa<SCEVCouldNotCompute>(LHSS))
323 const SCEV *RHSS = SE->getSCEV(RHS);
324 if (isa<SCEVCouldNotCompute>(RHSS))
327 // Canonicalize RHS to be loop invariant bound, LHS - a loop computable IV
328 if (SE->isLoopInvariant(LHSS, L)) {
330 std::swap(LHSS, RHSS);
331 Pred = ICmpInst::getSwappedPredicate(Pred);
334 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHSS);
335 if (!AR || AR->getLoop() != L)
338 return LoopICmp(Pred, AR, RHSS);
341 Value *LoopPredication::expandCheck(SCEVExpander &Expander,
342 IRBuilder<> &Builder,
343 ICmpInst::Predicate Pred, const SCEV *LHS,
344 const SCEV *RHS, Instruction *InsertAt) {
345 // TODO: we can check isLoopEntryGuardedByCond before emitting the check
347 Type *Ty = LHS->getType();
348 assert(Ty == RHS->getType() && "expandCheck operands have different types?");
350 if (SE->isLoopEntryGuardedByCond(L, Pred, LHS, RHS))
351 return Builder.getTrue();
353 Value *LHSV = Expander.expandCodeFor(LHS, Ty, InsertAt);
354 Value *RHSV = Expander.expandCodeFor(RHS, Ty, InsertAt);
355 return Builder.CreateICmp(Pred, LHSV, RHSV);
358 Optional<LoopPredication::LoopICmp>
359 LoopPredication::generateLoopLatchCheck(Type *RangeCheckType) {
361 auto *LatchType = LatchCheck.IV->getType();
362 if (RangeCheckType == LatchType)
364 // For now, bail out if latch type is narrower than range type.
365 if (DL->getTypeSizeInBits(LatchType) < DL->getTypeSizeInBits(RangeCheckType))
367 if (!isSafeToTruncateWideIVType(RangeCheckType))
369 // We can now safely identify the truncated version of the IV and limit for
371 LoopICmp NewLatchCheck;
372 NewLatchCheck.Pred = LatchCheck.Pred;
373 NewLatchCheck.IV = dyn_cast<SCEVAddRecExpr>(
374 SE->getTruncateExpr(LatchCheck.IV, RangeCheckType));
375 if (!NewLatchCheck.IV)
377 NewLatchCheck.Limit = SE->getTruncateExpr(LatchCheck.Limit, RangeCheckType);
378 DEBUG(dbgs() << "IV of type: " << *LatchType
379 << "can be represented as range check type:" << *RangeCheckType
381 DEBUG(dbgs() << "LatchCheck.IV: " << *NewLatchCheck.IV << "\n");
382 DEBUG(dbgs() << "LatchCheck.Limit: " << *NewLatchCheck.Limit << "\n");
383 return NewLatchCheck;
386 bool LoopPredication::isSupportedStep(const SCEV* Step) {
387 return Step->isOne() || (Step->isAllOnesValue() && EnableCountDownLoop);
390 bool LoopPredication::CanExpand(const SCEV* S) {
391 return SE->isLoopInvariant(S, L) && isSafeToExpand(S, *SE);
394 Optional<Value *> LoopPredication::widenICmpRangeCheckIncrementingLoop(
395 LoopPredication::LoopICmp LatchCheck, LoopPredication::LoopICmp RangeCheck,
396 SCEVExpander &Expander, IRBuilder<> &Builder) {
397 auto *Ty = RangeCheck.IV->getType();
398 // Generate the widened condition for the forward loop:
399 // guardStart u< guardLimit &&
400 // latchLimit <pred> guardLimit - 1 - guardStart + latchStart
401 // where <pred> depends on the latch condition predicate. See the file
402 // header comment for the reasoning.
403 // guardLimit - guardStart + latchStart - 1
404 const SCEV *GuardStart = RangeCheck.IV->getStart();
405 const SCEV *GuardLimit = RangeCheck.Limit;
406 const SCEV *LatchStart = LatchCheck.IV->getStart();
407 const SCEV *LatchLimit = LatchCheck.Limit;
409 // guardLimit - guardStart + latchStart - 1
411 SE->getAddExpr(SE->getMinusSCEV(GuardLimit, GuardStart),
412 SE->getMinusSCEV(LatchStart, SE->getOne(Ty)));
413 if (!CanExpand(GuardStart) || !CanExpand(GuardLimit) ||
414 !CanExpand(LatchLimit) || !CanExpand(RHS)) {
415 DEBUG(dbgs() << "Can't expand limit check!\n");
418 ICmpInst::Predicate LimitCheckPred;
419 switch (LatchCheck.Pred) {
420 case ICmpInst::ICMP_ULT:
421 LimitCheckPred = ICmpInst::ICMP_ULE;
423 case ICmpInst::ICMP_ULE:
424 LimitCheckPred = ICmpInst::ICMP_ULT;
426 case ICmpInst::ICMP_SLT:
427 LimitCheckPred = ICmpInst::ICMP_SLE;
429 case ICmpInst::ICMP_SLE:
430 LimitCheckPred = ICmpInst::ICMP_SLT;
433 llvm_unreachable("Unsupported loop latch!");
436 DEBUG(dbgs() << "LHS: " << *LatchLimit << "\n");
437 DEBUG(dbgs() << "RHS: " << *RHS << "\n");
438 DEBUG(dbgs() << "Pred: " << LimitCheckPred << "\n");
440 Instruction *InsertAt = Preheader->getTerminator();
442 expandCheck(Expander, Builder, LimitCheckPred, LatchLimit, RHS, InsertAt);
443 auto *FirstIterationCheck = expandCheck(Expander, Builder, RangeCheck.Pred,
444 GuardStart, GuardLimit, InsertAt);
445 return Builder.CreateAnd(FirstIterationCheck, LimitCheck);
448 Optional<Value *> LoopPredication::widenICmpRangeCheckDecrementingLoop(
449 LoopPredication::LoopICmp LatchCheck, LoopPredication::LoopICmp RangeCheck,
450 SCEVExpander &Expander, IRBuilder<> &Builder) {
451 auto *Ty = RangeCheck.IV->getType();
452 const SCEV *GuardStart = RangeCheck.IV->getStart();
453 const SCEV *GuardLimit = RangeCheck.Limit;
454 const SCEV *LatchLimit = LatchCheck.Limit;
455 if (!CanExpand(GuardStart) || !CanExpand(GuardLimit) ||
456 !CanExpand(LatchLimit)) {
457 DEBUG(dbgs() << "Can't expand limit check!\n");
460 // The decrement of the latch check IV should be the same as the
462 auto *PostDecLatchCheckIV = LatchCheck.IV->getPostIncExpr(*SE);
463 if (RangeCheck.IV != PostDecLatchCheckIV) {
464 DEBUG(dbgs() << "Not the same. PostDecLatchCheckIV: "
465 << *PostDecLatchCheckIV
466 << " and RangeCheckIV: " << *RangeCheck.IV << "\n");
470 // Generate the widened condition for CountDownLoop:
471 // guardStart u< guardLimit &&
472 // latchLimit <pred> 1.
473 // See the header comment for reasoning of the checks.
474 Instruction *InsertAt = Preheader->getTerminator();
475 auto LimitCheckPred = ICmpInst::isSigned(LatchCheck.Pred)
477 : ICmpInst::ICMP_UGE;
478 auto *FirstIterationCheck = expandCheck(Expander, Builder, ICmpInst::ICMP_ULT,
479 GuardStart, GuardLimit, InsertAt);
480 auto *LimitCheck = expandCheck(Expander, Builder, LimitCheckPred, LatchLimit,
481 SE->getOne(Ty), InsertAt);
482 return Builder.CreateAnd(FirstIterationCheck, LimitCheck);
485 /// If ICI can be widened to a loop invariant condition emits the loop
486 /// invariant condition in the loop preheader and return it, otherwise
488 Optional<Value *> LoopPredication::widenICmpRangeCheck(ICmpInst *ICI,
489 SCEVExpander &Expander,
490 IRBuilder<> &Builder) {
491 DEBUG(dbgs() << "Analyzing ICmpInst condition:\n");
494 // parseLoopStructure guarantees that the latch condition is:
495 // ++i <pred> latchLimit, where <pred> is u<, u<=, s<, or s<=.
496 // We are looking for the range checks of the form:
498 auto RangeCheck = parseLoopICmp(ICI);
500 DEBUG(dbgs() << "Failed to parse the loop latch condition!\n");
503 DEBUG(dbgs() << "Guard check:\n");
504 DEBUG(RangeCheck->dump());
505 if (RangeCheck->Pred != ICmpInst::ICMP_ULT) {
506 DEBUG(dbgs() << "Unsupported range check predicate(" << RangeCheck->Pred
510 auto *RangeCheckIV = RangeCheck->IV;
511 if (!RangeCheckIV->isAffine()) {
512 DEBUG(dbgs() << "Range check IV is not affine!\n");
515 auto *Step = RangeCheckIV->getStepRecurrence(*SE);
516 // We cannot just compare with latch IV step because the latch and range IVs
517 // may have different types.
518 if (!isSupportedStep(Step)) {
519 DEBUG(dbgs() << "Range check and latch have IVs different steps!\n");
522 auto *Ty = RangeCheckIV->getType();
523 auto CurrLatchCheckOpt = generateLoopLatchCheck(Ty);
524 if (!CurrLatchCheckOpt) {
525 DEBUG(dbgs() << "Failed to generate a loop latch check "
526 "corresponding to range type: "
531 LoopICmp CurrLatchCheck = *CurrLatchCheckOpt;
532 // At this point, the range and latch step should have the same type, but need
533 // not have the same value (we support both 1 and -1 steps).
534 assert(Step->getType() ==
535 CurrLatchCheck.IV->getStepRecurrence(*SE)->getType() &&
536 "Range and latch steps should be of same type!");
537 if (Step != CurrLatchCheck.IV->getStepRecurrence(*SE)) {
538 DEBUG(dbgs() << "Range and latch have different step values!\n");
543 return widenICmpRangeCheckIncrementingLoop(CurrLatchCheck, *RangeCheck,
546 assert(Step->isAllOnesValue() && "Step should be -1!");
547 return widenICmpRangeCheckDecrementingLoop(CurrLatchCheck, *RangeCheck,
552 bool LoopPredication::widenGuardConditions(IntrinsicInst *Guard,
553 SCEVExpander &Expander) {
554 DEBUG(dbgs() << "Processing guard:\n");
555 DEBUG(Guard->dump());
557 IRBuilder<> Builder(cast<Instruction>(Preheader->getTerminator()));
559 // The guard condition is expected to be in form of:
560 // cond1 && cond2 && cond3 ...
561 // Iterate over subconditions looking for for icmp conditions which can be
562 // widened across loop iterations. Widening these conditions remember the
563 // resulting list of subconditions in Checks vector.
564 SmallVector<Value *, 4> Worklist(1, Guard->getOperand(0));
565 SmallPtrSet<Value *, 4> Visited;
567 SmallVector<Value *, 4> Checks;
569 unsigned NumWidened = 0;
571 Value *Condition = Worklist.pop_back_val();
572 if (!Visited.insert(Condition).second)
576 using namespace llvm::PatternMatch;
577 if (match(Condition, m_And(m_Value(LHS), m_Value(RHS)))) {
578 Worklist.push_back(LHS);
579 Worklist.push_back(RHS);
583 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Condition)) {
584 if (auto NewRangeCheck = widenICmpRangeCheck(ICI, Expander, Builder)) {
585 Checks.push_back(NewRangeCheck.getValue());
591 // Save the condition as is if we can't widen it
592 Checks.push_back(Condition);
593 } while (Worklist.size() != 0);
598 // Emit the new guard condition
599 Builder.SetInsertPoint(Guard);
600 Value *LastCheck = nullptr;
601 for (auto *Check : Checks)
605 LastCheck = Builder.CreateAnd(LastCheck, Check);
606 Guard->setOperand(0, LastCheck);
608 DEBUG(dbgs() << "Widened checks = " << NumWidened << "\n");
612 Optional<LoopPredication::LoopICmp> LoopPredication::parseLoopLatchICmp() {
613 using namespace PatternMatch;
615 BasicBlock *LoopLatch = L->getLoopLatch();
617 DEBUG(dbgs() << "The loop doesn't have a single latch!\n");
621 ICmpInst::Predicate Pred;
623 BasicBlock *TrueDest, *FalseDest;
625 if (!match(LoopLatch->getTerminator(),
626 m_Br(m_ICmp(Pred, m_Value(LHS), m_Value(RHS)), TrueDest,
628 DEBUG(dbgs() << "Failed to match the latch terminator!\n");
631 assert((TrueDest == L->getHeader() || FalseDest == L->getHeader()) &&
632 "One of the latch's destinations must be the header");
633 if (TrueDest != L->getHeader())
634 Pred = ICmpInst::getInversePredicate(Pred);
636 auto Result = parseLoopICmp(Pred, LHS, RHS);
638 DEBUG(dbgs() << "Failed to parse the loop latch condition!\n");
642 // Check affine first, so if it's not we don't try to compute the step
644 if (!Result->IV->isAffine()) {
645 DEBUG(dbgs() << "The induction variable is not affine!\n");
649 auto *Step = Result->IV->getStepRecurrence(*SE);
650 if (!isSupportedStep(Step)) {
651 DEBUG(dbgs() << "Unsupported loop stride(" << *Step << ")!\n");
655 auto IsUnsupportedPredicate = [](const SCEV *Step, ICmpInst::Predicate Pred) {
657 return Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_SLT &&
658 Pred != ICmpInst::ICMP_ULE && Pred != ICmpInst::ICMP_SLE;
660 assert(Step->isAllOnesValue() && "Step should be -1!");
661 return Pred != ICmpInst::ICMP_UGT && Pred != ICmpInst::ICMP_SGT;
665 if (IsUnsupportedPredicate(Step, Result->Pred)) {
666 DEBUG(dbgs() << "Unsupported loop latch predicate(" << Result->Pred
673 // Returns true if its safe to truncate the IV to RangeCheckType.
674 bool LoopPredication::isSafeToTruncateWideIVType(Type *RangeCheckType) {
675 if (!EnableIVTruncation)
677 assert(DL->getTypeSizeInBits(LatchCheck.IV->getType()) >
678 DL->getTypeSizeInBits(RangeCheckType) &&
679 "Expected latch check IV type to be larger than range check operand "
681 // The start and end values of the IV should be known. This is to guarantee
682 // that truncating the wide type will not lose information.
683 auto *Limit = dyn_cast<SCEVConstant>(LatchCheck.Limit);
684 auto *Start = dyn_cast<SCEVConstant>(LatchCheck.IV->getStart());
685 if (!Limit || !Start)
687 // This check makes sure that the IV does not change sign during loop
688 // iterations. Consider latchType = i64, LatchStart = 5, Pred = ICMP_SGE,
689 // LatchEnd = 2, rangeCheckType = i32. If it's not a monotonic predicate, the
690 // IV wraps around, and the truncation of the IV would lose the range of
691 // iterations between 2^32 and 2^64.
693 if (!SE->isMonotonicPredicate(LatchCheck.IV, LatchCheck.Pred, Increasing))
695 // The active bits should be less than the bits in the RangeCheckType. This
696 // guarantees that truncating the latch check to RangeCheckType is a safe
698 auto RangeCheckTypeBitSize = DL->getTypeSizeInBits(RangeCheckType);
699 return Start->getAPInt().getActiveBits() < RangeCheckTypeBitSize &&
700 Limit->getAPInt().getActiveBits() < RangeCheckTypeBitSize;
703 bool LoopPredication::runOnLoop(Loop *Loop) {
706 DEBUG(dbgs() << "Analyzing ");
709 Module *M = L->getHeader()->getModule();
711 // There is nothing to do if the module doesn't use guards
713 M->getFunction(Intrinsic::getName(Intrinsic::experimental_guard));
714 if (!GuardDecl || GuardDecl->use_empty())
717 DL = &M->getDataLayout();
719 Preheader = L->getLoopPreheader();
723 auto LatchCheckOpt = parseLoopLatchICmp();
726 LatchCheck = *LatchCheckOpt;
728 DEBUG(dbgs() << "Latch check:\n");
729 DEBUG(LatchCheck.dump());
731 // Collect all the guards into a vector and process later, so as not
732 // to invalidate the instruction iterator.
733 SmallVector<IntrinsicInst *, 4> Guards;
734 for (const auto BB : L->blocks())
736 if (auto *II = dyn_cast<IntrinsicInst>(&I))
737 if (II->getIntrinsicID() == Intrinsic::experimental_guard)
738 Guards.push_back(II);
743 SCEVExpander Expander(*SE, *DL, "loop-predication");
745 bool Changed = false;
746 for (auto *Guard : Guards)
747 Changed |= widenGuardConditions(Guard, Expander);