1 //===- GuardWidening.cpp - ---- Guard widening ----------------------------===//
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 // This file implements the guard widening pass. The semantics of the
11 // @llvm.experimental.guard intrinsic lets LLVM transform it so that it fails
12 // more often that it did before the transform. This optimization is called
13 // "widening" and can be used hoist and common runtime checks in situations like
16 // %cmp0 = 7 u< Length
17 // call @llvm.experimental.guard(i1 %cmp0) [ "deopt"(...) ]
18 // call @unknown_side_effects()
19 // %cmp1 = 9 u< Length
20 // call @llvm.experimental.guard(i1 %cmp1) [ "deopt"(...) ]
25 // %cmp0 = 9 u< Length
26 // call @llvm.experimental.guard(i1 %cmp0) [ "deopt"(...) ]
27 // call @unknown_side_effects()
30 // If %cmp0 is false, @llvm.experimental.guard will "deoptimize" back to a
31 // generic implementation of the same function, which will have the correct
32 // semantics from that point onward. It is always _legal_ to deoptimize (so
33 // replacing %cmp0 with false is "correct"), though it may not always be
34 // profitable to do so.
36 // NB! This pass is a work in progress. It hasn't been tuned to be "production
37 // ready" yet. It is known to have quadriatic running time and will not scale
38 // to large numbers of guards
40 //===----------------------------------------------------------------------===//
42 #include "llvm/Transforms/Scalar/GuardWidening.h"
44 #include "llvm/ADT/DenseMap.h"
45 #include "llvm/ADT/DepthFirstIterator.h"
46 #include "llvm/ADT/Statistic.h"
47 #include "llvm/Analysis/BranchProbabilityInfo.h"
48 #include "llvm/Analysis/GuardUtils.h"
49 #include "llvm/Analysis/LoopInfo.h"
50 #include "llvm/Analysis/LoopPass.h"
51 #include "llvm/Analysis/PostDominators.h"
52 #include "llvm/Analysis/ValueTracking.h"
53 #include "llvm/IR/ConstantRange.h"
54 #include "llvm/IR/Dominators.h"
55 #include "llvm/IR/IntrinsicInst.h"
56 #include "llvm/IR/PatternMatch.h"
57 #include "llvm/Pass.h"
58 #include "llvm/Support/Debug.h"
59 #include "llvm/Support/KnownBits.h"
60 #include "llvm/Transforms/Scalar.h"
61 #include "llvm/Transforms/Utils/LoopUtils.h"
65 #define DEBUG_TYPE "guard-widening"
67 STATISTIC(GuardsEliminated, "Number of eliminated guards");
68 STATISTIC(CondBranchEliminated, "Number of eliminated conditional branches");
70 static cl::opt<bool> WidenFrequentBranches(
71 "guard-widening-widen-frequent-branches", cl::Hidden,
72 cl::desc("Widen conditions of explicit branches into dominating guards in "
73 "case if their taken frequency exceeds threshold set by "
74 "guard-widening-frequent-branch-threshold option"),
77 static cl::opt<unsigned> FrequentBranchThreshold(
78 "guard-widening-frequent-branch-threshold", cl::Hidden,
79 cl::desc("When WidenFrequentBranches is set to true, this option is used "
80 "to determine which branches are frequently taken. The criteria "
81 "that a branch is taken more often than "
82 "((FrequentBranchThreshold - 1) / FrequentBranchThreshold), then "
83 "it is considered frequently taken"),
89 // Get the condition of \p I. It can either be a guard or a conditional branch.
90 static Value *getCondition(Instruction *I) {
91 if (IntrinsicInst *GI = dyn_cast<IntrinsicInst>(I)) {
92 assert(GI->getIntrinsicID() == Intrinsic::experimental_guard &&
93 "Bad guard intrinsic?");
94 return GI->getArgOperand(0);
96 return cast<BranchInst>(I)->getCondition();
99 // Set the condition for \p I to \p NewCond. \p I can either be a guard or a
100 // conditional branch.
101 static void setCondition(Instruction *I, Value *NewCond) {
102 if (IntrinsicInst *GI = dyn_cast<IntrinsicInst>(I)) {
103 assert(GI->getIntrinsicID() == Intrinsic::experimental_guard &&
104 "Bad guard intrinsic?");
105 GI->setArgOperand(0, NewCond);
108 cast<BranchInst>(I)->setCondition(NewCond);
111 // Eliminates the guard instruction properly.
112 static void eliminateGuard(Instruction *GuardInst) {
113 GuardInst->eraseFromParent();
117 class GuardWideningImpl {
119 PostDominatorTree *PDT;
121 BranchProbabilityInfo *BPI;
123 /// Together, these describe the region of interest. This might be all of
124 /// the blocks within a function, or only a given loop's blocks and preheader.
126 std::function<bool(BasicBlock*)> BlockFilter;
128 /// The set of guards and conditional branches whose conditions have been
129 /// widened into dominating guards.
130 SmallVector<Instruction *, 16> EliminatedGuardsAndBranches;
132 /// The set of guards which have been widened to include conditions to other
134 DenseSet<Instruction *> WidenedGuards;
136 /// Try to eliminate guard \p Guard by widening it into an earlier dominating
137 /// guard. \p DFSI is the DFS iterator on the dominator tree that is
138 /// currently visiting the block containing \p Guard, and \p GuardsPerBlock
139 /// maps BasicBlocks to the set of guards seen in that block.
140 bool eliminateGuardViaWidening(
141 Instruction *Guard, const df_iterator<DomTreeNode *> &DFSI,
142 const DenseMap<BasicBlock *, SmallVector<Instruction *, 8>> &
143 GuardsPerBlock, bool InvertCondition = false);
145 /// Used to keep track of which widening potential is more effective.
148 WS_IllegalOrNegative,
150 /// Widening is performance neutral as far as the cycles spent in check
151 /// conditions goes (but can still help, e.g., code layout, having less
155 /// Widening is profitable.
158 /// Widening is very profitable. Not significantly different from \c
159 /// WS_Positive, except by the order.
163 static StringRef scoreTypeToString(WideningScore WS);
165 /// Compute the score for widening the condition in \p DominatedGuard
166 /// (contained in \p DominatedGuardLoop) into \p DominatingGuard (contained in
167 /// \p DominatingGuardLoop). If \p InvertCond is set, then we widen the
168 /// inverted condition of the dominating guard.
169 WideningScore computeWideningScore(Instruction *DominatedGuard,
170 Loop *DominatedGuardLoop,
171 Instruction *DominatingGuard,
172 Loop *DominatingGuardLoop,
175 /// Helper to check if \p V can be hoisted to \p InsertPos.
176 bool isAvailableAt(Value *V, Instruction *InsertPos) {
177 SmallPtrSet<Instruction *, 8> Visited;
178 return isAvailableAt(V, InsertPos, Visited);
181 bool isAvailableAt(Value *V, Instruction *InsertPos,
182 SmallPtrSetImpl<Instruction *> &Visited);
184 /// Helper to hoist \p V to \p InsertPos. Guaranteed to succeed if \c
185 /// isAvailableAt returned true.
186 void makeAvailableAt(Value *V, Instruction *InsertPos);
188 /// Common helper used by \c widenGuard and \c isWideningCondProfitable. Try
189 /// to generate an expression computing the logical AND of \p Cond0 and (\p
190 /// Cond1 XOR \p InvertCondition).
191 /// Return true if the expression computing the AND is only as
192 /// expensive as computing one of the two. If \p InsertPt is true then
193 /// actually generate the resulting expression, make it available at \p
194 /// InsertPt and return it in \p Result (else no change to the IR is made).
195 bool widenCondCommon(Value *Cond0, Value *Cond1, Instruction *InsertPt,
196 Value *&Result, bool InvertCondition);
198 /// Represents a range check of the form \c Base + \c Offset u< \c Length,
199 /// with the constraint that \c Length is not negative. \c CheckInst is the
200 /// pre-existing instruction in the IR that computes the result of this range
209 explicit RangeCheck(Value *Base, ConstantInt *Offset, Value *Length,
211 : Base(Base), Offset(Offset), Length(Length), CheckInst(CheckInst) {}
213 void setBase(Value *NewBase) { Base = NewBase; }
214 void setOffset(ConstantInt *NewOffset) { Offset = NewOffset; }
216 Value *getBase() const { return Base; }
217 ConstantInt *getOffset() const { return Offset; }
218 const APInt &getOffsetValue() const { return getOffset()->getValue(); }
219 Value *getLength() const { return Length; };
220 ICmpInst *getCheckInst() const { return CheckInst; }
222 void print(raw_ostream &OS, bool PrintTypes = false) {
224 Base->printAsOperand(OS, PrintTypes);
226 Offset->printAsOperand(OS, PrintTypes);
228 Length->printAsOperand(OS, PrintTypes);
231 LLVM_DUMP_METHOD void dump() {
237 /// Parse \p CheckCond into a conjunction (logical-and) of range checks; and
238 /// append them to \p Checks. Returns true on success, may clobber \c Checks
240 bool parseRangeChecks(Value *CheckCond, SmallVectorImpl<RangeCheck> &Checks) {
241 SmallPtrSet<Value *, 8> Visited;
242 return parseRangeChecks(CheckCond, Checks, Visited);
245 bool parseRangeChecks(Value *CheckCond, SmallVectorImpl<RangeCheck> &Checks,
246 SmallPtrSetImpl<Value *> &Visited);
248 /// Combine the checks in \p Checks into a smaller set of checks and append
249 /// them into \p CombinedChecks. Return true on success (i.e. all of checks
250 /// in \p Checks were combined into \p CombinedChecks). Clobbers \p Checks
251 /// and \p CombinedChecks on success and on failure.
252 bool combineRangeChecks(SmallVectorImpl<RangeCheck> &Checks,
253 SmallVectorImpl<RangeCheck> &CombinedChecks);
255 /// Can we compute the logical AND of \p Cond0 and \p Cond1 for the price of
256 /// computing only one of the two expressions?
257 bool isWideningCondProfitable(Value *Cond0, Value *Cond1, bool InvertCond) {
259 return widenCondCommon(Cond0, Cond1, /*InsertPt=*/nullptr, ResultUnused,
263 /// If \p InvertCondition is false, Widen \p ToWiden to fail if
264 /// \p NewCondition is false, otherwise make it fail if \p NewCondition is
265 /// true (in addition to whatever it is already checking).
266 void widenGuard(Instruction *ToWiden, Value *NewCondition,
267 bool InvertCondition) {
269 widenCondCommon(ToWiden->getOperand(0), NewCondition, ToWiden, Result,
271 setCondition(ToWiden, Result);
276 explicit GuardWideningImpl(DominatorTree &DT, PostDominatorTree *PDT,
277 LoopInfo &LI, BranchProbabilityInfo *BPI,
279 std::function<bool(BasicBlock*)> BlockFilter)
280 : DT(DT), PDT(PDT), LI(LI), BPI(BPI), Root(Root), BlockFilter(BlockFilter)
283 /// The entry point for this pass.
288 bool GuardWideningImpl::run() {
289 DenseMap<BasicBlock *, SmallVector<Instruction *, 8>> GuardsInBlock;
290 bool Changed = false;
291 Optional<BranchProbability> LikelyTaken = None;
292 if (WidenFrequentBranches && BPI) {
293 unsigned Threshold = FrequentBranchThreshold;
294 assert(Threshold > 0 && "Zero threshold makes no sense!");
295 LikelyTaken = BranchProbability(Threshold - 1, Threshold);
298 for (auto DFI = df_begin(Root), DFE = df_end(Root);
300 auto *BB = (*DFI)->getBlock();
301 if (!BlockFilter(BB))
304 auto &CurrentList = GuardsInBlock[BB];
308 CurrentList.push_back(cast<Instruction>(&I));
310 for (auto *II : CurrentList)
311 Changed |= eliminateGuardViaWidening(II, DFI, GuardsInBlock);
312 if (WidenFrequentBranches && BPI)
313 if (auto *BI = dyn_cast<BranchInst>(BB->getTerminator()))
314 if (BI->isConditional()) {
315 // If one of branches of a conditional is likely taken, try to
317 if (BPI->getEdgeProbability(BB, 0U) >= *LikelyTaken)
318 Changed |= eliminateGuardViaWidening(BI, DFI, GuardsInBlock);
319 else if (BPI->getEdgeProbability(BB, 1U) >= *LikelyTaken)
320 Changed |= eliminateGuardViaWidening(BI, DFI, GuardsInBlock,
321 /*InvertCondition*/true);
325 assert(EliminatedGuardsAndBranches.empty() || Changed);
326 for (auto *I : EliminatedGuardsAndBranches)
327 if (!WidenedGuards.count(I)) {
328 assert(isa<ConstantInt>(getCondition(I)) && "Should be!");
332 assert(isa<BranchInst>(I) &&
333 "Eliminated something other than guard or branch?");
334 ++CondBranchEliminated;
341 bool GuardWideningImpl::eliminateGuardViaWidening(
342 Instruction *GuardInst, const df_iterator<DomTreeNode *> &DFSI,
343 const DenseMap<BasicBlock *, SmallVector<Instruction *, 8>> &
344 GuardsInBlock, bool InvertCondition) {
345 // Ignore trivial true or false conditions. These instructions will be
346 // trivially eliminated by any cleanup pass. Do not erase them because other
347 // guards can possibly be widened into them.
348 if (isa<ConstantInt>(getCondition(GuardInst)))
351 Instruction *BestSoFar = nullptr;
352 auto BestScoreSoFar = WS_IllegalOrNegative;
353 auto *GuardInstLoop = LI.getLoopFor(GuardInst->getParent());
355 // In the set of dominating guards, find the one we can merge GuardInst with
356 // for the most profit.
357 for (unsigned i = 0, e = DFSI.getPathLength(); i != e; ++i) {
358 auto *CurBB = DFSI.getPath(i)->getBlock();
359 if (!BlockFilter(CurBB))
361 auto *CurLoop = LI.getLoopFor(CurBB);
362 assert(GuardsInBlock.count(CurBB) && "Must have been populated by now!");
363 const auto &GuardsInCurBB = GuardsInBlock.find(CurBB)->second;
365 auto I = GuardsInCurBB.begin();
366 auto E = GuardsInCurBB.end();
371 for (auto &I : *CurBB) {
372 if (Index == GuardsInCurBB.size())
374 if (GuardsInCurBB[Index] == &I)
377 assert(Index == GuardsInCurBB.size() &&
378 "Guards expected to be in order!");
382 assert((i == (e - 1)) == (GuardInst->getParent() == CurBB) && "Bad DFS?");
384 if (i == (e - 1) && CurBB->getTerminator() != GuardInst) {
385 // Corner case: make sure we're only looking at guards strictly dominating
386 // GuardInst when visiting GuardInst->getParent().
387 auto NewEnd = std::find(I, E, GuardInst);
388 assert(NewEnd != E && "GuardInst not in its own block?");
392 for (auto *Candidate : make_range(I, E)) {
394 computeWideningScore(GuardInst, GuardInstLoop, Candidate, CurLoop,
396 LLVM_DEBUG(dbgs() << "Score between " << *getCondition(GuardInst)
397 << " and " << *getCondition(Candidate) << " is "
398 << scoreTypeToString(Score) << "\n");
399 if (Score > BestScoreSoFar) {
400 BestScoreSoFar = Score;
401 BestSoFar = Candidate;
406 if (BestScoreSoFar == WS_IllegalOrNegative) {
407 LLVM_DEBUG(dbgs() << "Did not eliminate guard " << *GuardInst << "\n");
411 assert(BestSoFar != GuardInst && "Should have never visited same guard!");
412 assert(DT.dominates(BestSoFar, GuardInst) && "Should be!");
414 LLVM_DEBUG(dbgs() << "Widening " << *GuardInst << " into " << *BestSoFar
415 << " with score " << scoreTypeToString(BestScoreSoFar)
417 widenGuard(BestSoFar, getCondition(GuardInst), InvertCondition);
418 auto NewGuardCondition = InvertCondition
419 ? ConstantInt::getFalse(GuardInst->getContext())
420 : ConstantInt::getTrue(GuardInst->getContext());
421 setCondition(GuardInst, NewGuardCondition);
422 EliminatedGuardsAndBranches.push_back(GuardInst);
423 WidenedGuards.insert(BestSoFar);
427 GuardWideningImpl::WideningScore GuardWideningImpl::computeWideningScore(
428 Instruction *DominatedGuard, Loop *DominatedGuardLoop,
429 Instruction *DominatingGuard, Loop *DominatingGuardLoop, bool InvertCond) {
430 bool HoistingOutOfLoop = false;
432 if (DominatingGuardLoop != DominatedGuardLoop) {
433 // Be conservative and don't widen into a sibling loop. TODO: If the
434 // sibling is colder, we should consider allowing this.
435 if (DominatingGuardLoop &&
436 !DominatingGuardLoop->contains(DominatedGuardLoop))
437 return WS_IllegalOrNegative;
439 HoistingOutOfLoop = true;
442 if (!isAvailableAt(getCondition(DominatedGuard), DominatingGuard))
443 return WS_IllegalOrNegative;
445 // If the guard was conditional executed, it may never be reached
446 // dynamically. There are two potential downsides to hoisting it out of the
447 // conditionally executed region: 1) we may spuriously deopt without need and
448 // 2) we have the extra cost of computing the guard condition in the common
449 // case. At the moment, we really only consider the second in our heuristic
450 // here. TODO: evaluate cost model for spurious deopt
451 // NOTE: As written, this also lets us hoist right over another guard which
452 // is essentially just another spelling for control flow.
453 if (isWideningCondProfitable(getCondition(DominatedGuard),
454 getCondition(DominatingGuard), InvertCond))
455 return HoistingOutOfLoop ? WS_VeryPositive : WS_Positive;
457 if (HoistingOutOfLoop)
460 // Returns true if we might be hoisting above explicit control flow. Note
461 // that this completely ignores implicit control flow (guards, calls which
462 // throw, etc...). That choice appears arbitrary.
463 auto MaybeHoistingOutOfIf = [&]() {
464 auto *DominatingBlock = DominatingGuard->getParent();
465 auto *DominatedBlock = DominatedGuard->getParent();
468 if (DominatedBlock == DominatingBlock)
470 // Obvious successor (common loop header/preheader case)
471 if (DominatedBlock == DominatingBlock->getUniqueSuccessor())
473 // TODO: diamond, triangle cases
474 if (!PDT) return true;
475 return !PDT->dominates(DominatedBlock, DominatingBlock);
478 return MaybeHoistingOutOfIf() ? WS_IllegalOrNegative : WS_Neutral;
481 bool GuardWideningImpl::isAvailableAt(Value *V, Instruction *Loc,
482 SmallPtrSetImpl<Instruction *> &Visited) {
483 auto *Inst = dyn_cast<Instruction>(V);
484 if (!Inst || DT.dominates(Inst, Loc) || Visited.count(Inst))
487 if (!isSafeToSpeculativelyExecute(Inst, Loc, &DT) ||
488 Inst->mayReadFromMemory())
491 Visited.insert(Inst);
493 // We only want to go _up_ the dominance chain when recursing.
494 assert(!isa<PHINode>(Loc) &&
495 "PHIs should return false for isSafeToSpeculativelyExecute");
496 assert(DT.isReachableFromEntry(Inst->getParent()) &&
497 "We did a DFS from the block entry!");
498 return all_of(Inst->operands(),
499 [&](Value *Op) { return isAvailableAt(Op, Loc, Visited); });
502 void GuardWideningImpl::makeAvailableAt(Value *V, Instruction *Loc) {
503 auto *Inst = dyn_cast<Instruction>(V);
504 if (!Inst || DT.dominates(Inst, Loc))
507 assert(isSafeToSpeculativelyExecute(Inst, Loc, &DT) &&
508 !Inst->mayReadFromMemory() && "Should've checked with isAvailableAt!");
510 for (Value *Op : Inst->operands())
511 makeAvailableAt(Op, Loc);
513 Inst->moveBefore(Loc);
516 bool GuardWideningImpl::widenCondCommon(Value *Cond0, Value *Cond1,
517 Instruction *InsertPt, Value *&Result,
518 bool InvertCondition) {
519 using namespace llvm::PatternMatch;
522 // L >u C0 && L >u C1 -> L >u max(C0, C1)
523 ConstantInt *RHS0, *RHS1;
525 ICmpInst::Predicate Pred0, Pred1;
526 if (match(Cond0, m_ICmp(Pred0, m_Value(LHS), m_ConstantInt(RHS0))) &&
527 match(Cond1, m_ICmp(Pred1, m_Specific(LHS), m_ConstantInt(RHS1)))) {
529 Pred1 = ICmpInst::getInversePredicate(Pred1);
532 ConstantRange::makeExactICmpRegion(Pred0, RHS0->getValue());
534 ConstantRange::makeExactICmpRegion(Pred1, RHS1->getValue());
536 // SubsetIntersect is a subset of the actual mathematical intersection of
537 // CR0 and CR1, while SupersetIntersect is a superset of the actual
538 // mathematical intersection. If these two ConstantRanges are equal, then
539 // we know we were able to represent the actual mathematical intersection
540 // of CR0 and CR1, and can use the same to generate an icmp instruction.
542 // Given what we're doing here and the semantics of guards, it would
543 // actually be correct to just use SubsetIntersect, but that may be too
544 // aggressive in cases we care about.
545 auto SubsetIntersect = CR0.inverse().unionWith(CR1.inverse()).inverse();
546 auto SupersetIntersect = CR0.intersectWith(CR1);
549 CmpInst::Predicate Pred;
550 if (SubsetIntersect == SupersetIntersect &&
551 SubsetIntersect.getEquivalentICmp(Pred, NewRHSAP)) {
553 ConstantInt *NewRHS = ConstantInt::get(Cond0->getContext(), NewRHSAP);
554 Result = new ICmpInst(InsertPt, Pred, LHS, NewRHS, "wide.chk");
562 SmallVector<GuardWideningImpl::RangeCheck, 4> Checks, CombinedChecks;
563 // TODO: Support InvertCondition case?
564 if (!InvertCondition &&
565 parseRangeChecks(Cond0, Checks) && parseRangeChecks(Cond1, Checks) &&
566 combineRangeChecks(Checks, CombinedChecks)) {
569 for (auto &RC : CombinedChecks) {
570 makeAvailableAt(RC.getCheckInst(), InsertPt);
572 Result = BinaryOperator::CreateAnd(RC.getCheckInst(), Result, "",
575 Result = RC.getCheckInst();
578 Result->setName("wide.chk");
584 // Base case -- just logical-and the two conditions together.
587 makeAvailableAt(Cond0, InsertPt);
588 makeAvailableAt(Cond1, InsertPt);
590 Cond1 = BinaryOperator::CreateNot(Cond1, "inverted", InsertPt);
591 Result = BinaryOperator::CreateAnd(Cond0, Cond1, "wide.chk", InsertPt);
594 // We were not able to compute Cond0 AND Cond1 for the price of one.
598 bool GuardWideningImpl::parseRangeChecks(
599 Value *CheckCond, SmallVectorImpl<GuardWideningImpl::RangeCheck> &Checks,
600 SmallPtrSetImpl<Value *> &Visited) {
601 if (!Visited.insert(CheckCond).second)
604 using namespace llvm::PatternMatch;
607 Value *AndLHS, *AndRHS;
608 if (match(CheckCond, m_And(m_Value(AndLHS), m_Value(AndRHS))))
609 return parseRangeChecks(AndLHS, Checks) &&
610 parseRangeChecks(AndRHS, Checks);
613 auto *IC = dyn_cast<ICmpInst>(CheckCond);
614 if (!IC || !IC->getOperand(0)->getType()->isIntegerTy() ||
615 (IC->getPredicate() != ICmpInst::ICMP_ULT &&
616 IC->getPredicate() != ICmpInst::ICMP_UGT))
619 Value *CmpLHS = IC->getOperand(0), *CmpRHS = IC->getOperand(1);
620 if (IC->getPredicate() == ICmpInst::ICMP_UGT)
621 std::swap(CmpLHS, CmpRHS);
623 auto &DL = IC->getModule()->getDataLayout();
625 GuardWideningImpl::RangeCheck Check(
626 CmpLHS, cast<ConstantInt>(ConstantInt::getNullValue(CmpRHS->getType())),
629 if (!isKnownNonNegative(Check.getLength(), DL))
632 // What we have in \c Check now is a correct interpretation of \p CheckCond.
633 // Try to see if we can move some constant offsets into the \c Offset field.
636 auto &Ctx = CheckCond->getContext();
644 auto *BaseInst = dyn_cast<Instruction>(Check.getBase());
645 assert((!BaseInst || DT.isReachableFromEntry(BaseInst->getParent())) &&
646 "Unreachable instruction?");
649 if (match(Check.getBase(), m_Add(m_Value(OpLHS), m_ConstantInt(OpRHS)))) {
650 Check.setBase(OpLHS);
651 APInt NewOffset = Check.getOffsetValue() + OpRHS->getValue();
652 Check.setOffset(ConstantInt::get(Ctx, NewOffset));
654 } else if (match(Check.getBase(),
655 m_Or(m_Value(OpLHS), m_ConstantInt(OpRHS)))) {
656 KnownBits Known = computeKnownBits(OpLHS, DL);
657 if ((OpRHS->getValue() & Known.Zero) == OpRHS->getValue()) {
658 Check.setBase(OpLHS);
659 APInt NewOffset = Check.getOffsetValue() + OpRHS->getValue();
660 Check.setOffset(ConstantInt::get(Ctx, NewOffset));
666 Checks.push_back(Check);
670 bool GuardWideningImpl::combineRangeChecks(
671 SmallVectorImpl<GuardWideningImpl::RangeCheck> &Checks,
672 SmallVectorImpl<GuardWideningImpl::RangeCheck> &RangeChecksOut) {
673 unsigned OldCount = Checks.size();
674 while (!Checks.empty()) {
675 // Pick all of the range checks with a specific base and length, and try to
677 Value *CurrentBase = Checks.front().getBase();
678 Value *CurrentLength = Checks.front().getLength();
680 SmallVector<GuardWideningImpl::RangeCheck, 3> CurrentChecks;
682 auto IsCurrentCheck = [&](GuardWideningImpl::RangeCheck &RC) {
683 return RC.getBase() == CurrentBase && RC.getLength() == CurrentLength;
686 copy_if(Checks, std::back_inserter(CurrentChecks), IsCurrentCheck);
687 Checks.erase(remove_if(Checks, IsCurrentCheck), Checks.end());
689 assert(CurrentChecks.size() != 0 && "We know we have at least one!");
691 if (CurrentChecks.size() < 3) {
692 RangeChecksOut.insert(RangeChecksOut.end(), CurrentChecks.begin(),
693 CurrentChecks.end());
697 // CurrentChecks.size() will typically be 3 here, but so far there has been
698 // no need to hard-code that fact.
700 llvm::sort(CurrentChecks, [&](const GuardWideningImpl::RangeCheck &LHS,
701 const GuardWideningImpl::RangeCheck &RHS) {
702 return LHS.getOffsetValue().slt(RHS.getOffsetValue());
705 // Note: std::sort should not invalidate the ChecksStart iterator.
707 ConstantInt *MinOffset = CurrentChecks.front().getOffset(),
708 *MaxOffset = CurrentChecks.back().getOffset();
710 unsigned BitWidth = MaxOffset->getValue().getBitWidth();
711 if ((MaxOffset->getValue() - MinOffset->getValue())
712 .ugt(APInt::getSignedMinValue(BitWidth)))
715 APInt MaxDiff = MaxOffset->getValue() - MinOffset->getValue();
716 const APInt &HighOffset = MaxOffset->getValue();
717 auto OffsetOK = [&](const GuardWideningImpl::RangeCheck &RC) {
718 return (HighOffset - RC.getOffsetValue()).ult(MaxDiff);
721 if (MaxDiff.isMinValue() ||
722 !std::all_of(std::next(CurrentChecks.begin()), CurrentChecks.end(),
726 // We have a series of f+1 checks as:
728 // I+k_0 u< L ... Chk_0
729 // I+k_1 u< L ... Chk_1
731 // I+k_f u< L ... Chk_f
733 // with forall i in [0,f]: k_f-k_i u< k_f-k_0 ... Precond_0
734 // k_f-k_0 u< INT_MIN+k_f ... Precond_1
735 // k_f != k_0 ... Precond_2
738 // Chk_0 AND Chk_f implies all the other checks
740 // Informal proof sketch:
742 // We will show that the integer range [I+k_0,I+k_f] does not unsigned-wrap
743 // (i.e. going from I+k_0 to I+k_f does not cross the -1,0 boundary) and
744 // thus I+k_f is the greatest unsigned value in that range.
746 // This combined with Ckh_(f+1) shows that everything in that range is u< L.
747 // Via Precond_0 we know that all of the indices in Chk_0 through Chk_(f+1)
748 // lie in [I+k_0,I+k_f], this proving our claim.
750 // To see that [I+k_0,I+k_f] is not a wrapping range, note that there are
751 // two possibilities: I+k_0 u< I+k_f or I+k_0 >u I+k_f (they can't be equal
752 // since k_0 != k_f). In the former case, [I+k_0,I+k_f] is not a wrapping
753 // range by definition, and the latter case is impossible:
755 // 0-----I+k_f---I+k_0----L---INT_MAX,INT_MIN------------------(-1)
756 // xxxxxx xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
758 // For Chk_0 to succeed, we'd have to have k_f-k_0 (the range highlighted
759 // with 'x' above) to be at least >u INT_MIN.
761 RangeChecksOut.emplace_back(CurrentChecks.front());
762 RangeChecksOut.emplace_back(CurrentChecks.back());
765 assert(RangeChecksOut.size() <= OldCount && "We pessimized!");
766 return RangeChecksOut.size() != OldCount;
770 StringRef GuardWideningImpl::scoreTypeToString(WideningScore WS) {
772 case WS_IllegalOrNegative:
773 return "IllegalOrNegative";
778 case WS_VeryPositive:
779 return "VeryPositive";
782 llvm_unreachable("Fully covered switch above!");
786 PreservedAnalyses GuardWideningPass::run(Function &F,
787 FunctionAnalysisManager &AM) {
788 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
789 auto &LI = AM.getResult<LoopAnalysis>(F);
790 auto &PDT = AM.getResult<PostDominatorTreeAnalysis>(F);
791 BranchProbabilityInfo *BPI = nullptr;
792 if (WidenFrequentBranches)
793 BPI = AM.getCachedResult<BranchProbabilityAnalysis>(F);
794 if (!GuardWideningImpl(DT, &PDT, LI, BPI, DT.getRootNode(),
795 [](BasicBlock*) { return true; } ).run())
796 return PreservedAnalyses::all();
798 PreservedAnalyses PA;
799 PA.preserveSet<CFGAnalyses>();
804 struct GuardWideningLegacyPass : public FunctionPass {
807 GuardWideningLegacyPass() : FunctionPass(ID) {
808 initializeGuardWideningLegacyPassPass(*PassRegistry::getPassRegistry());
811 bool runOnFunction(Function &F) override {
814 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
815 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
816 auto &PDT = getAnalysis<PostDominatorTreeWrapperPass>().getPostDomTree();
817 BranchProbabilityInfo *BPI = nullptr;
818 if (WidenFrequentBranches)
819 BPI = &getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI();
820 return GuardWideningImpl(DT, &PDT, LI, BPI, DT.getRootNode(),
821 [](BasicBlock*) { return true; } ).run();
824 void getAnalysisUsage(AnalysisUsage &AU) const override {
825 AU.setPreservesCFG();
826 AU.addRequired<DominatorTreeWrapperPass>();
827 AU.addRequired<PostDominatorTreeWrapperPass>();
828 AU.addRequired<LoopInfoWrapperPass>();
829 if (WidenFrequentBranches)
830 AU.addRequired<BranchProbabilityInfoWrapperPass>();
834 /// Same as above, but restricted to a single loop at a time. Can be
835 /// scheduled with other loop passes w/o breaking out of LPM
836 struct LoopGuardWideningLegacyPass : public LoopPass {
839 LoopGuardWideningLegacyPass() : LoopPass(ID) {
840 initializeLoopGuardWideningLegacyPassPass(*PassRegistry::getPassRegistry());
843 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
846 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
847 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
848 auto *PDTWP = getAnalysisIfAvailable<PostDominatorTreeWrapperPass>();
849 auto *PDT = PDTWP ? &PDTWP->getPostDomTree() : nullptr;
850 BasicBlock *RootBB = L->getLoopPredecessor();
852 RootBB = L->getHeader();
853 auto BlockFilter = [&](BasicBlock *BB) {
854 return BB == RootBB || L->contains(BB);
856 BranchProbabilityInfo *BPI = nullptr;
857 if (WidenFrequentBranches)
858 BPI = &getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI();
859 return GuardWideningImpl(DT, PDT, LI, BPI,
860 DT.getNode(RootBB), BlockFilter).run();
863 void getAnalysisUsage(AnalysisUsage &AU) const override {
864 if (WidenFrequentBranches)
865 AU.addRequired<BranchProbabilityInfoWrapperPass>();
866 AU.setPreservesCFG();
867 getLoopAnalysisUsage(AU);
868 AU.addPreserved<PostDominatorTreeWrapperPass>();
873 char GuardWideningLegacyPass::ID = 0;
874 char LoopGuardWideningLegacyPass::ID = 0;
876 INITIALIZE_PASS_BEGIN(GuardWideningLegacyPass, "guard-widening", "Widen guards",
878 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
879 INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass)
880 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
881 if (WidenFrequentBranches)
882 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)
883 INITIALIZE_PASS_END(GuardWideningLegacyPass, "guard-widening", "Widen guards",
886 INITIALIZE_PASS_BEGIN(LoopGuardWideningLegacyPass, "loop-guard-widening",
887 "Widen guards (within a single loop, as a loop pass)",
889 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
890 INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass)
891 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
892 if (WidenFrequentBranches)
893 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)
894 INITIALIZE_PASS_END(LoopGuardWideningLegacyPass, "loop-guard-widening",
895 "Widen guards (within a single loop, as a loop pass)",
898 FunctionPass *llvm::createGuardWideningPass() {
899 return new GuardWideningLegacyPass();
902 Pass *llvm::createLoopGuardWideningPass() {
903 return new LoopGuardWideningLegacyPass();