[InstCombine] Signed saturation patterns
[llvm-complete.git] / lib / Transforms / Scalar / GuardWidening.cpp
blob2697d78095681abe289250ddd30abe8315e10afe
1 //===- GuardWidening.cpp - ---- Guard widening ----------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the guard widening pass. The semantics of the
10 // @llvm.experimental.guard intrinsic lets LLVM transform it so that it fails
11 // more often that it did before the transform. This optimization is called
12 // "widening" and can be used hoist and common runtime checks in situations like
13 // these:
15 // %cmp0 = 7 u< Length
16 // call @llvm.experimental.guard(i1 %cmp0) [ "deopt"(...) ]
17 // call @unknown_side_effects()
18 // %cmp1 = 9 u< Length
19 // call @llvm.experimental.guard(i1 %cmp1) [ "deopt"(...) ]
20 // ...
22 // =>
24 // %cmp0 = 9 u< Length
25 // call @llvm.experimental.guard(i1 %cmp0) [ "deopt"(...) ]
26 // call @unknown_side_effects()
27 // ...
29 // If %cmp0 is false, @llvm.experimental.guard will "deoptimize" back to a
30 // generic implementation of the same function, which will have the correct
31 // semantics from that point onward. It is always _legal_ to deoptimize (so
32 // replacing %cmp0 with false is "correct"), though it may not always be
33 // profitable to do so.
35 // NB! This pass is a work in progress. It hasn't been tuned to be "production
36 // ready" yet. It is known to have quadriatic running time and will not scale
37 // to large numbers of guards
39 //===----------------------------------------------------------------------===//
41 #include "llvm/Transforms/Scalar/GuardWidening.h"
42 #include <functional>
43 #include "llvm/ADT/DenseMap.h"
44 #include "llvm/ADT/DepthFirstIterator.h"
45 #include "llvm/ADT/Statistic.h"
46 #include "llvm/Analysis/BranchProbabilityInfo.h"
47 #include "llvm/Analysis/GuardUtils.h"
48 #include "llvm/Analysis/LoopInfo.h"
49 #include "llvm/Analysis/LoopPass.h"
50 #include "llvm/Analysis/PostDominators.h"
51 #include "llvm/Analysis/ValueTracking.h"
52 #include "llvm/IR/ConstantRange.h"
53 #include "llvm/IR/Dominators.h"
54 #include "llvm/IR/IntrinsicInst.h"
55 #include "llvm/IR/PatternMatch.h"
56 #include "llvm/Pass.h"
57 #include "llvm/Support/Debug.h"
58 #include "llvm/Support/KnownBits.h"
59 #include "llvm/Transforms/Scalar.h"
60 #include "llvm/Transforms/Utils/LoopUtils.h"
62 using namespace llvm;
64 #define DEBUG_TYPE "guard-widening"
66 STATISTIC(GuardsEliminated, "Number of eliminated guards");
67 STATISTIC(CondBranchEliminated, "Number of eliminated conditional branches");
69 static cl::opt<bool> WidenFrequentBranches(
70 "guard-widening-widen-frequent-branches", cl::Hidden,
71 cl::desc("Widen conditions of explicit branches into dominating guards in "
72 "case if their taken frequency exceeds threshold set by "
73 "guard-widening-frequent-branch-threshold option"),
74 cl::init(false));
76 static cl::opt<unsigned> FrequentBranchThreshold(
77 "guard-widening-frequent-branch-threshold", cl::Hidden,
78 cl::desc("When WidenFrequentBranches is set to true, this option is used "
79 "to determine which branches are frequently taken. The criteria "
80 "that a branch is taken more often than "
81 "((FrequentBranchThreshold - 1) / FrequentBranchThreshold), then "
82 "it is considered frequently taken"),
83 cl::init(1000));
85 static cl::opt<bool>
86 WidenBranchGuards("guard-widening-widen-branch-guards", cl::Hidden,
87 cl::desc("Whether or not we should widen guards "
88 "expressed as branches by widenable conditions"),
89 cl::init(true));
91 namespace {
93 // Get the condition of \p I. It can either be a guard or a conditional branch.
94 static Value *getCondition(Instruction *I) {
95 if (IntrinsicInst *GI = dyn_cast<IntrinsicInst>(I)) {
96 assert(GI->getIntrinsicID() == Intrinsic::experimental_guard &&
97 "Bad guard intrinsic?");
98 return GI->getArgOperand(0);
100 if (isGuardAsWidenableBranch(I)) {
101 auto *Cond = cast<BranchInst>(I)->getCondition();
102 return cast<BinaryOperator>(Cond)->getOperand(0);
104 return cast<BranchInst>(I)->getCondition();
107 // Set the condition for \p I to \p NewCond. \p I can either be a guard or a
108 // conditional branch.
109 static void setCondition(Instruction *I, Value *NewCond) {
110 if (IntrinsicInst *GI = dyn_cast<IntrinsicInst>(I)) {
111 assert(GI->getIntrinsicID() == Intrinsic::experimental_guard &&
112 "Bad guard intrinsic?");
113 GI->setArgOperand(0, NewCond);
114 return;
116 cast<BranchInst>(I)->setCondition(NewCond);
119 // Eliminates the guard instruction properly.
120 static void eliminateGuard(Instruction *GuardInst) {
121 GuardInst->eraseFromParent();
122 ++GuardsEliminated;
125 class GuardWideningImpl {
126 DominatorTree &DT;
127 PostDominatorTree *PDT;
128 LoopInfo &LI;
129 BranchProbabilityInfo *BPI;
131 /// Together, these describe the region of interest. This might be all of
132 /// the blocks within a function, or only a given loop's blocks and preheader.
133 DomTreeNode *Root;
134 std::function<bool(BasicBlock*)> BlockFilter;
136 /// The set of guards and conditional branches whose conditions have been
137 /// widened into dominating guards.
138 SmallVector<Instruction *, 16> EliminatedGuardsAndBranches;
140 /// The set of guards which have been widened to include conditions to other
141 /// guards.
142 DenseSet<Instruction *> WidenedGuards;
144 /// Try to eliminate instruction \p Instr by widening it into an earlier
145 /// dominating guard. \p DFSI is the DFS iterator on the dominator tree that
146 /// is currently visiting the block containing \p Guard, and \p GuardsPerBlock
147 /// maps BasicBlocks to the set of guards seen in that block.
148 bool eliminateInstrViaWidening(
149 Instruction *Instr, const df_iterator<DomTreeNode *> &DFSI,
150 const DenseMap<BasicBlock *, SmallVector<Instruction *, 8>> &
151 GuardsPerBlock, bool InvertCondition = false);
153 /// Used to keep track of which widening potential is more effective.
154 enum WideningScore {
155 /// Don't widen.
156 WS_IllegalOrNegative,
158 /// Widening is performance neutral as far as the cycles spent in check
159 /// conditions goes (but can still help, e.g., code layout, having less
160 /// deopt state).
161 WS_Neutral,
163 /// Widening is profitable.
164 WS_Positive,
166 /// Widening is very profitable. Not significantly different from \c
167 /// WS_Positive, except by the order.
168 WS_VeryPositive
171 static StringRef scoreTypeToString(WideningScore WS);
173 /// Compute the score for widening the condition in \p DominatedInstr
174 /// into \p DominatingGuard. If \p InvertCond is set, then we widen the
175 /// inverted condition of the dominating guard.
176 WideningScore computeWideningScore(Instruction *DominatedInstr,
177 Instruction *DominatingGuard,
178 bool InvertCond);
180 /// Helper to check if \p V can be hoisted to \p InsertPos.
181 bool isAvailableAt(const Value *V, const Instruction *InsertPos) const {
182 SmallPtrSet<const Instruction *, 8> Visited;
183 return isAvailableAt(V, InsertPos, Visited);
186 bool isAvailableAt(const Value *V, const Instruction *InsertPos,
187 SmallPtrSetImpl<const Instruction *> &Visited) const;
189 /// Helper to hoist \p V to \p InsertPos. Guaranteed to succeed if \c
190 /// isAvailableAt returned true.
191 void makeAvailableAt(Value *V, Instruction *InsertPos) const;
193 /// Common helper used by \c widenGuard and \c isWideningCondProfitable. Try
194 /// to generate an expression computing the logical AND of \p Cond0 and (\p
195 /// Cond1 XOR \p InvertCondition).
196 /// Return true if the expression computing the AND is only as
197 /// expensive as computing one of the two. If \p InsertPt is true then
198 /// actually generate the resulting expression, make it available at \p
199 /// InsertPt and return it in \p Result (else no change to the IR is made).
200 bool widenCondCommon(Value *Cond0, Value *Cond1, Instruction *InsertPt,
201 Value *&Result, bool InvertCondition);
203 /// Represents a range check of the form \c Base + \c Offset u< \c Length,
204 /// with the constraint that \c Length is not negative. \c CheckInst is the
205 /// pre-existing instruction in the IR that computes the result of this range
206 /// check.
207 class RangeCheck {
208 const Value *Base;
209 const ConstantInt *Offset;
210 const Value *Length;
211 ICmpInst *CheckInst;
213 public:
214 explicit RangeCheck(const Value *Base, const ConstantInt *Offset,
215 const Value *Length, ICmpInst *CheckInst)
216 : Base(Base), Offset(Offset), Length(Length), CheckInst(CheckInst) {}
218 void setBase(const Value *NewBase) { Base = NewBase; }
219 void setOffset(const ConstantInt *NewOffset) { Offset = NewOffset; }
221 const Value *getBase() const { return Base; }
222 const ConstantInt *getOffset() const { return Offset; }
223 const APInt &getOffsetValue() const { return getOffset()->getValue(); }
224 const Value *getLength() const { return Length; };
225 ICmpInst *getCheckInst() const { return CheckInst; }
227 void print(raw_ostream &OS, bool PrintTypes = false) {
228 OS << "Base: ";
229 Base->printAsOperand(OS, PrintTypes);
230 OS << " Offset: ";
231 Offset->printAsOperand(OS, PrintTypes);
232 OS << " Length: ";
233 Length->printAsOperand(OS, PrintTypes);
236 LLVM_DUMP_METHOD void dump() {
237 print(dbgs());
238 dbgs() << "\n";
242 /// Parse \p CheckCond into a conjunction (logical-and) of range checks; and
243 /// append them to \p Checks. Returns true on success, may clobber \c Checks
244 /// on failure.
245 bool parseRangeChecks(Value *CheckCond, SmallVectorImpl<RangeCheck> &Checks) {
246 SmallPtrSet<const Value *, 8> Visited;
247 return parseRangeChecks(CheckCond, Checks, Visited);
250 bool parseRangeChecks(Value *CheckCond, SmallVectorImpl<RangeCheck> &Checks,
251 SmallPtrSetImpl<const Value *> &Visited);
253 /// Combine the checks in \p Checks into a smaller set of checks and append
254 /// them into \p CombinedChecks. Return true on success (i.e. all of checks
255 /// in \p Checks were combined into \p CombinedChecks). Clobbers \p Checks
256 /// and \p CombinedChecks on success and on failure.
257 bool combineRangeChecks(SmallVectorImpl<RangeCheck> &Checks,
258 SmallVectorImpl<RangeCheck> &CombinedChecks) const;
260 /// Can we compute the logical AND of \p Cond0 and \p Cond1 for the price of
261 /// computing only one of the two expressions?
262 bool isWideningCondProfitable(Value *Cond0, Value *Cond1, bool InvertCond) {
263 Value *ResultUnused;
264 return widenCondCommon(Cond0, Cond1, /*InsertPt=*/nullptr, ResultUnused,
265 InvertCond);
268 /// If \p InvertCondition is false, Widen \p ToWiden to fail if
269 /// \p NewCondition is false, otherwise make it fail if \p NewCondition is
270 /// true (in addition to whatever it is already checking).
271 void widenGuard(Instruction *ToWiden, Value *NewCondition,
272 bool InvertCondition) {
273 Value *Result;
274 widenCondCommon(getCondition(ToWiden), NewCondition, ToWiden, Result,
275 InvertCondition);
276 Value *WidenableCondition = nullptr;
277 if (isGuardAsWidenableBranch(ToWiden)) {
278 auto *Cond = cast<BranchInst>(ToWiden)->getCondition();
279 WidenableCondition = cast<BinaryOperator>(Cond)->getOperand(1);
281 if (WidenableCondition)
282 Result = BinaryOperator::CreateAnd(Result, WidenableCondition,
283 "guard.chk", ToWiden);
284 setCondition(ToWiden, Result);
287 public:
289 explicit GuardWideningImpl(DominatorTree &DT, PostDominatorTree *PDT,
290 LoopInfo &LI, BranchProbabilityInfo *BPI,
291 DomTreeNode *Root,
292 std::function<bool(BasicBlock*)> BlockFilter)
293 : DT(DT), PDT(PDT), LI(LI), BPI(BPI), Root(Root), BlockFilter(BlockFilter)
296 /// The entry point for this pass.
297 bool run();
301 static bool isSupportedGuardInstruction(const Instruction *Insn) {
302 if (isGuard(Insn))
303 return true;
304 if (WidenBranchGuards && isGuardAsWidenableBranch(Insn))
305 return true;
306 return false;
309 bool GuardWideningImpl::run() {
310 DenseMap<BasicBlock *, SmallVector<Instruction *, 8>> GuardsInBlock;
311 bool Changed = false;
312 Optional<BranchProbability> LikelyTaken = None;
313 if (WidenFrequentBranches && BPI) {
314 unsigned Threshold = FrequentBranchThreshold;
315 assert(Threshold > 0 && "Zero threshold makes no sense!");
316 LikelyTaken = BranchProbability(Threshold - 1, Threshold);
319 for (auto DFI = df_begin(Root), DFE = df_end(Root);
320 DFI != DFE; ++DFI) {
321 auto *BB = (*DFI)->getBlock();
322 if (!BlockFilter(BB))
323 continue;
325 auto &CurrentList = GuardsInBlock[BB];
327 for (auto &I : *BB)
328 if (isSupportedGuardInstruction(&I))
329 CurrentList.push_back(cast<Instruction>(&I));
331 for (auto *II : CurrentList)
332 Changed |= eliminateInstrViaWidening(II, DFI, GuardsInBlock);
333 if (WidenFrequentBranches && BPI)
334 if (auto *BI = dyn_cast<BranchInst>(BB->getTerminator()))
335 if (BI->isConditional()) {
336 // If one of branches of a conditional is likely taken, try to
337 // eliminate it.
338 if (BPI->getEdgeProbability(BB, 0U) >= *LikelyTaken)
339 Changed |= eliminateInstrViaWidening(BI, DFI, GuardsInBlock);
340 else if (BPI->getEdgeProbability(BB, 1U) >= *LikelyTaken)
341 Changed |= eliminateInstrViaWidening(BI, DFI, GuardsInBlock,
342 /*InvertCondition*/true);
346 assert(EliminatedGuardsAndBranches.empty() || Changed);
347 for (auto *I : EliminatedGuardsAndBranches)
348 if (!WidenedGuards.count(I)) {
349 assert(isa<ConstantInt>(getCondition(I)) && "Should be!");
350 if (isSupportedGuardInstruction(I))
351 eliminateGuard(I);
352 else {
353 assert(isa<BranchInst>(I) &&
354 "Eliminated something other than guard or branch?");
355 ++CondBranchEliminated;
359 return Changed;
362 bool GuardWideningImpl::eliminateInstrViaWidening(
363 Instruction *Instr, const df_iterator<DomTreeNode *> &DFSI,
364 const DenseMap<BasicBlock *, SmallVector<Instruction *, 8>> &
365 GuardsInBlock, bool InvertCondition) {
366 // Ignore trivial true or false conditions. These instructions will be
367 // trivially eliminated by any cleanup pass. Do not erase them because other
368 // guards can possibly be widened into them.
369 if (isa<ConstantInt>(getCondition(Instr)))
370 return false;
372 Instruction *BestSoFar = nullptr;
373 auto BestScoreSoFar = WS_IllegalOrNegative;
375 // In the set of dominating guards, find the one we can merge GuardInst with
376 // for the most profit.
377 for (unsigned i = 0, e = DFSI.getPathLength(); i != e; ++i) {
378 auto *CurBB = DFSI.getPath(i)->getBlock();
379 if (!BlockFilter(CurBB))
380 break;
381 assert(GuardsInBlock.count(CurBB) && "Must have been populated by now!");
382 const auto &GuardsInCurBB = GuardsInBlock.find(CurBB)->second;
384 auto I = GuardsInCurBB.begin();
385 auto E = Instr->getParent() == CurBB
386 ? std::find(GuardsInCurBB.begin(), GuardsInCurBB.end(), Instr)
387 : GuardsInCurBB.end();
389 #ifndef NDEBUG
391 unsigned Index = 0;
392 for (auto &I : *CurBB) {
393 if (Index == GuardsInCurBB.size())
394 break;
395 if (GuardsInCurBB[Index] == &I)
396 Index++;
398 assert(Index == GuardsInCurBB.size() &&
399 "Guards expected to be in order!");
401 #endif
403 assert((i == (e - 1)) == (Instr->getParent() == CurBB) && "Bad DFS?");
405 for (auto *Candidate : make_range(I, E)) {
406 auto Score = computeWideningScore(Instr, Candidate, InvertCondition);
407 LLVM_DEBUG(dbgs() << "Score between " << *getCondition(Instr)
408 << " and " << *getCondition(Candidate) << " is "
409 << scoreTypeToString(Score) << "\n");
410 if (Score > BestScoreSoFar) {
411 BestScoreSoFar = Score;
412 BestSoFar = Candidate;
417 if (BestScoreSoFar == WS_IllegalOrNegative) {
418 LLVM_DEBUG(dbgs() << "Did not eliminate guard " << *Instr << "\n");
419 return false;
422 assert(BestSoFar != Instr && "Should have never visited same guard!");
423 assert(DT.dominates(BestSoFar, Instr) && "Should be!");
425 LLVM_DEBUG(dbgs() << "Widening " << *Instr << " into " << *BestSoFar
426 << " with score " << scoreTypeToString(BestScoreSoFar)
427 << "\n");
428 widenGuard(BestSoFar, getCondition(Instr), InvertCondition);
429 auto NewGuardCondition = InvertCondition
430 ? ConstantInt::getFalse(Instr->getContext())
431 : ConstantInt::getTrue(Instr->getContext());
432 setCondition(Instr, NewGuardCondition);
433 EliminatedGuardsAndBranches.push_back(Instr);
434 WidenedGuards.insert(BestSoFar);
435 return true;
438 GuardWideningImpl::WideningScore
439 GuardWideningImpl::computeWideningScore(Instruction *DominatedInstr,
440 Instruction *DominatingGuard,
441 bool InvertCond) {
442 Loop *DominatedInstrLoop = LI.getLoopFor(DominatedInstr->getParent());
443 Loop *DominatingGuardLoop = LI.getLoopFor(DominatingGuard->getParent());
444 bool HoistingOutOfLoop = false;
446 if (DominatingGuardLoop != DominatedInstrLoop) {
447 // Be conservative and don't widen into a sibling loop. TODO: If the
448 // sibling is colder, we should consider allowing this.
449 if (DominatingGuardLoop &&
450 !DominatingGuardLoop->contains(DominatedInstrLoop))
451 return WS_IllegalOrNegative;
453 HoistingOutOfLoop = true;
456 if (!isAvailableAt(getCondition(DominatedInstr), DominatingGuard))
457 return WS_IllegalOrNegative;
459 // If the guard was conditional executed, it may never be reached
460 // dynamically. There are two potential downsides to hoisting it out of the
461 // conditionally executed region: 1) we may spuriously deopt without need and
462 // 2) we have the extra cost of computing the guard condition in the common
463 // case. At the moment, we really only consider the second in our heuristic
464 // here. TODO: evaluate cost model for spurious deopt
465 // NOTE: As written, this also lets us hoist right over another guard which
466 // is essentially just another spelling for control flow.
467 if (isWideningCondProfitable(getCondition(DominatedInstr),
468 getCondition(DominatingGuard), InvertCond))
469 return HoistingOutOfLoop ? WS_VeryPositive : WS_Positive;
471 if (HoistingOutOfLoop)
472 return WS_Positive;
474 // Returns true if we might be hoisting above explicit control flow. Note
475 // that this completely ignores implicit control flow (guards, calls which
476 // throw, etc...). That choice appears arbitrary.
477 auto MaybeHoistingOutOfIf = [&]() {
478 auto *DominatingBlock = DominatingGuard->getParent();
479 auto *DominatedBlock = DominatedInstr->getParent();
480 if (isGuardAsWidenableBranch(DominatingGuard))
481 DominatingBlock = cast<BranchInst>(DominatingGuard)->getSuccessor(0);
483 // Same Block?
484 if (DominatedBlock == DominatingBlock)
485 return false;
486 // Obvious successor (common loop header/preheader case)
487 if (DominatedBlock == DominatingBlock->getUniqueSuccessor())
488 return false;
489 // TODO: diamond, triangle cases
490 if (!PDT) return true;
491 return !PDT->dominates(DominatedBlock, DominatingBlock);
494 return MaybeHoistingOutOfIf() ? WS_IllegalOrNegative : WS_Neutral;
497 bool GuardWideningImpl::isAvailableAt(
498 const Value *V, const Instruction *Loc,
499 SmallPtrSetImpl<const Instruction *> &Visited) const {
500 auto *Inst = dyn_cast<Instruction>(V);
501 if (!Inst || DT.dominates(Inst, Loc) || Visited.count(Inst))
502 return true;
504 if (!isSafeToSpeculativelyExecute(Inst, Loc, &DT) ||
505 Inst->mayReadFromMemory())
506 return false;
508 Visited.insert(Inst);
510 // We only want to go _up_ the dominance chain when recursing.
511 assert(!isa<PHINode>(Loc) &&
512 "PHIs should return false for isSafeToSpeculativelyExecute");
513 assert(DT.isReachableFromEntry(Inst->getParent()) &&
514 "We did a DFS from the block entry!");
515 return all_of(Inst->operands(),
516 [&](Value *Op) { return isAvailableAt(Op, Loc, Visited); });
519 void GuardWideningImpl::makeAvailableAt(Value *V, Instruction *Loc) const {
520 auto *Inst = dyn_cast<Instruction>(V);
521 if (!Inst || DT.dominates(Inst, Loc))
522 return;
524 assert(isSafeToSpeculativelyExecute(Inst, Loc, &DT) &&
525 !Inst->mayReadFromMemory() && "Should've checked with isAvailableAt!");
527 for (Value *Op : Inst->operands())
528 makeAvailableAt(Op, Loc);
530 Inst->moveBefore(Loc);
533 bool GuardWideningImpl::widenCondCommon(Value *Cond0, Value *Cond1,
534 Instruction *InsertPt, Value *&Result,
535 bool InvertCondition) {
536 using namespace llvm::PatternMatch;
539 // L >u C0 && L >u C1 -> L >u max(C0, C1)
540 ConstantInt *RHS0, *RHS1;
541 Value *LHS;
542 ICmpInst::Predicate Pred0, Pred1;
543 if (match(Cond0, m_ICmp(Pred0, m_Value(LHS), m_ConstantInt(RHS0))) &&
544 match(Cond1, m_ICmp(Pred1, m_Specific(LHS), m_ConstantInt(RHS1)))) {
545 if (InvertCondition)
546 Pred1 = ICmpInst::getInversePredicate(Pred1);
548 ConstantRange CR0 =
549 ConstantRange::makeExactICmpRegion(Pred0, RHS0->getValue());
550 ConstantRange CR1 =
551 ConstantRange::makeExactICmpRegion(Pred1, RHS1->getValue());
553 // SubsetIntersect is a subset of the actual mathematical intersection of
554 // CR0 and CR1, while SupersetIntersect is a superset of the actual
555 // mathematical intersection. If these two ConstantRanges are equal, then
556 // we know we were able to represent the actual mathematical intersection
557 // of CR0 and CR1, and can use the same to generate an icmp instruction.
559 // Given what we're doing here and the semantics of guards, it would
560 // actually be correct to just use SubsetIntersect, but that may be too
561 // aggressive in cases we care about.
562 auto SubsetIntersect = CR0.inverse().unionWith(CR1.inverse()).inverse();
563 auto SupersetIntersect = CR0.intersectWith(CR1);
565 APInt NewRHSAP;
566 CmpInst::Predicate Pred;
567 if (SubsetIntersect == SupersetIntersect &&
568 SubsetIntersect.getEquivalentICmp(Pred, NewRHSAP)) {
569 if (InsertPt) {
570 ConstantInt *NewRHS = ConstantInt::get(Cond0->getContext(), NewRHSAP);
571 Result = new ICmpInst(InsertPt, Pred, LHS, NewRHS, "wide.chk");
573 return true;
579 SmallVector<GuardWideningImpl::RangeCheck, 4> Checks, CombinedChecks;
580 // TODO: Support InvertCondition case?
581 if (!InvertCondition &&
582 parseRangeChecks(Cond0, Checks) && parseRangeChecks(Cond1, Checks) &&
583 combineRangeChecks(Checks, CombinedChecks)) {
584 if (InsertPt) {
585 Result = nullptr;
586 for (auto &RC : CombinedChecks) {
587 makeAvailableAt(RC.getCheckInst(), InsertPt);
588 if (Result)
589 Result = BinaryOperator::CreateAnd(RC.getCheckInst(), Result, "",
590 InsertPt);
591 else
592 Result = RC.getCheckInst();
594 assert(Result && "Failed to find result value");
595 Result->setName("wide.chk");
597 return true;
601 // Base case -- just logical-and the two conditions together.
603 if (InsertPt) {
604 makeAvailableAt(Cond0, InsertPt);
605 makeAvailableAt(Cond1, InsertPt);
606 if (InvertCondition)
607 Cond1 = BinaryOperator::CreateNot(Cond1, "inverted", InsertPt);
608 Result = BinaryOperator::CreateAnd(Cond0, Cond1, "wide.chk", InsertPt);
611 // We were not able to compute Cond0 AND Cond1 for the price of one.
612 return false;
615 bool GuardWideningImpl::parseRangeChecks(
616 Value *CheckCond, SmallVectorImpl<GuardWideningImpl::RangeCheck> &Checks,
617 SmallPtrSetImpl<const Value *> &Visited) {
618 if (!Visited.insert(CheckCond).second)
619 return true;
621 using namespace llvm::PatternMatch;
624 Value *AndLHS, *AndRHS;
625 if (match(CheckCond, m_And(m_Value(AndLHS), m_Value(AndRHS))))
626 return parseRangeChecks(AndLHS, Checks) &&
627 parseRangeChecks(AndRHS, Checks);
630 auto *IC = dyn_cast<ICmpInst>(CheckCond);
631 if (!IC || !IC->getOperand(0)->getType()->isIntegerTy() ||
632 (IC->getPredicate() != ICmpInst::ICMP_ULT &&
633 IC->getPredicate() != ICmpInst::ICMP_UGT))
634 return false;
636 const Value *CmpLHS = IC->getOperand(0), *CmpRHS = IC->getOperand(1);
637 if (IC->getPredicate() == ICmpInst::ICMP_UGT)
638 std::swap(CmpLHS, CmpRHS);
640 auto &DL = IC->getModule()->getDataLayout();
642 GuardWideningImpl::RangeCheck Check(
643 CmpLHS, cast<ConstantInt>(ConstantInt::getNullValue(CmpRHS->getType())),
644 CmpRHS, IC);
646 if (!isKnownNonNegative(Check.getLength(), DL))
647 return false;
649 // What we have in \c Check now is a correct interpretation of \p CheckCond.
650 // Try to see if we can move some constant offsets into the \c Offset field.
652 bool Changed;
653 auto &Ctx = CheckCond->getContext();
655 do {
656 Value *OpLHS;
657 ConstantInt *OpRHS;
658 Changed = false;
660 #ifndef NDEBUG
661 auto *BaseInst = dyn_cast<Instruction>(Check.getBase());
662 assert((!BaseInst || DT.isReachableFromEntry(BaseInst->getParent())) &&
663 "Unreachable instruction?");
664 #endif
666 if (match(Check.getBase(), m_Add(m_Value(OpLHS), m_ConstantInt(OpRHS)))) {
667 Check.setBase(OpLHS);
668 APInt NewOffset = Check.getOffsetValue() + OpRHS->getValue();
669 Check.setOffset(ConstantInt::get(Ctx, NewOffset));
670 Changed = true;
671 } else if (match(Check.getBase(),
672 m_Or(m_Value(OpLHS), m_ConstantInt(OpRHS)))) {
673 KnownBits Known = computeKnownBits(OpLHS, DL);
674 if ((OpRHS->getValue() & Known.Zero) == OpRHS->getValue()) {
675 Check.setBase(OpLHS);
676 APInt NewOffset = Check.getOffsetValue() + OpRHS->getValue();
677 Check.setOffset(ConstantInt::get(Ctx, NewOffset));
678 Changed = true;
681 } while (Changed);
683 Checks.push_back(Check);
684 return true;
687 bool GuardWideningImpl::combineRangeChecks(
688 SmallVectorImpl<GuardWideningImpl::RangeCheck> &Checks,
689 SmallVectorImpl<GuardWideningImpl::RangeCheck> &RangeChecksOut) const {
690 unsigned OldCount = Checks.size();
691 while (!Checks.empty()) {
692 // Pick all of the range checks with a specific base and length, and try to
693 // merge them.
694 const Value *CurrentBase = Checks.front().getBase();
695 const Value *CurrentLength = Checks.front().getLength();
697 SmallVector<GuardWideningImpl::RangeCheck, 3> CurrentChecks;
699 auto IsCurrentCheck = [&](GuardWideningImpl::RangeCheck &RC) {
700 return RC.getBase() == CurrentBase && RC.getLength() == CurrentLength;
703 copy_if(Checks, std::back_inserter(CurrentChecks), IsCurrentCheck);
704 Checks.erase(remove_if(Checks, IsCurrentCheck), Checks.end());
706 assert(CurrentChecks.size() != 0 && "We know we have at least one!");
708 if (CurrentChecks.size() < 3) {
709 RangeChecksOut.insert(RangeChecksOut.end(), CurrentChecks.begin(),
710 CurrentChecks.end());
711 continue;
714 // CurrentChecks.size() will typically be 3 here, but so far there has been
715 // no need to hard-code that fact.
717 llvm::sort(CurrentChecks, [&](const GuardWideningImpl::RangeCheck &LHS,
718 const GuardWideningImpl::RangeCheck &RHS) {
719 return LHS.getOffsetValue().slt(RHS.getOffsetValue());
722 // Note: std::sort should not invalidate the ChecksStart iterator.
724 const ConstantInt *MinOffset = CurrentChecks.front().getOffset();
725 const ConstantInt *MaxOffset = CurrentChecks.back().getOffset();
727 unsigned BitWidth = MaxOffset->getValue().getBitWidth();
728 if ((MaxOffset->getValue() - MinOffset->getValue())
729 .ugt(APInt::getSignedMinValue(BitWidth)))
730 return false;
732 APInt MaxDiff = MaxOffset->getValue() - MinOffset->getValue();
733 const APInt &HighOffset = MaxOffset->getValue();
734 auto OffsetOK = [&](const GuardWideningImpl::RangeCheck &RC) {
735 return (HighOffset - RC.getOffsetValue()).ult(MaxDiff);
738 if (MaxDiff.isMinValue() ||
739 !std::all_of(std::next(CurrentChecks.begin()), CurrentChecks.end(),
740 OffsetOK))
741 return false;
743 // We have a series of f+1 checks as:
745 // I+k_0 u< L ... Chk_0
746 // I+k_1 u< L ... Chk_1
747 // ...
748 // I+k_f u< L ... Chk_f
750 // with forall i in [0,f]: k_f-k_i u< k_f-k_0 ... Precond_0
751 // k_f-k_0 u< INT_MIN+k_f ... Precond_1
752 // k_f != k_0 ... Precond_2
754 // Claim:
755 // Chk_0 AND Chk_f implies all the other checks
757 // Informal proof sketch:
759 // We will show that the integer range [I+k_0,I+k_f] does not unsigned-wrap
760 // (i.e. going from I+k_0 to I+k_f does not cross the -1,0 boundary) and
761 // thus I+k_f is the greatest unsigned value in that range.
763 // This combined with Ckh_(f+1) shows that everything in that range is u< L.
764 // Via Precond_0 we know that all of the indices in Chk_0 through Chk_(f+1)
765 // lie in [I+k_0,I+k_f], this proving our claim.
767 // To see that [I+k_0,I+k_f] is not a wrapping range, note that there are
768 // two possibilities: I+k_0 u< I+k_f or I+k_0 >u I+k_f (they can't be equal
769 // since k_0 != k_f). In the former case, [I+k_0,I+k_f] is not a wrapping
770 // range by definition, and the latter case is impossible:
772 // 0-----I+k_f---I+k_0----L---INT_MAX,INT_MIN------------------(-1)
773 // xxxxxx xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
775 // For Chk_0 to succeed, we'd have to have k_f-k_0 (the range highlighted
776 // with 'x' above) to be at least >u INT_MIN.
778 RangeChecksOut.emplace_back(CurrentChecks.front());
779 RangeChecksOut.emplace_back(CurrentChecks.back());
782 assert(RangeChecksOut.size() <= OldCount && "We pessimized!");
783 return RangeChecksOut.size() != OldCount;
786 #ifndef NDEBUG
787 StringRef GuardWideningImpl::scoreTypeToString(WideningScore WS) {
788 switch (WS) {
789 case WS_IllegalOrNegative:
790 return "IllegalOrNegative";
791 case WS_Neutral:
792 return "Neutral";
793 case WS_Positive:
794 return "Positive";
795 case WS_VeryPositive:
796 return "VeryPositive";
799 llvm_unreachable("Fully covered switch above!");
801 #endif
803 PreservedAnalyses GuardWideningPass::run(Function &F,
804 FunctionAnalysisManager &AM) {
805 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
806 auto &LI = AM.getResult<LoopAnalysis>(F);
807 auto &PDT = AM.getResult<PostDominatorTreeAnalysis>(F);
808 BranchProbabilityInfo *BPI = nullptr;
809 if (WidenFrequentBranches)
810 BPI = AM.getCachedResult<BranchProbabilityAnalysis>(F);
811 if (!GuardWideningImpl(DT, &PDT, LI, BPI, DT.getRootNode(),
812 [](BasicBlock*) { return true; } ).run())
813 return PreservedAnalyses::all();
815 PreservedAnalyses PA;
816 PA.preserveSet<CFGAnalyses>();
817 return PA;
820 PreservedAnalyses GuardWideningPass::run(Loop &L, LoopAnalysisManager &AM,
821 LoopStandardAnalysisResults &AR,
822 LPMUpdater &U) {
824 const auto &FAM =
825 AM.getResult<FunctionAnalysisManagerLoopProxy>(L, AR).getManager();
826 Function &F = *L.getHeader()->getParent();
827 BranchProbabilityInfo *BPI = nullptr;
828 if (WidenFrequentBranches)
829 BPI = FAM.getCachedResult<BranchProbabilityAnalysis>(F);
831 BasicBlock *RootBB = L.getLoopPredecessor();
832 if (!RootBB)
833 RootBB = L.getHeader();
834 auto BlockFilter = [&](BasicBlock *BB) {
835 return BB == RootBB || L.contains(BB);
837 if (!GuardWideningImpl(AR.DT, nullptr, AR.LI, BPI,
838 AR.DT.getNode(RootBB),
839 BlockFilter).run())
840 return PreservedAnalyses::all();
842 return getLoopPassPreservedAnalyses();
845 namespace {
846 struct GuardWideningLegacyPass : public FunctionPass {
847 static char ID;
849 GuardWideningLegacyPass() : FunctionPass(ID) {
850 initializeGuardWideningLegacyPassPass(*PassRegistry::getPassRegistry());
853 bool runOnFunction(Function &F) override {
854 if (skipFunction(F))
855 return false;
856 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
857 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
858 auto &PDT = getAnalysis<PostDominatorTreeWrapperPass>().getPostDomTree();
859 BranchProbabilityInfo *BPI = nullptr;
860 if (WidenFrequentBranches)
861 BPI = &getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI();
862 return GuardWideningImpl(DT, &PDT, LI, BPI, DT.getRootNode(),
863 [](BasicBlock*) { return true; } ).run();
866 void getAnalysisUsage(AnalysisUsage &AU) const override {
867 AU.setPreservesCFG();
868 AU.addRequired<DominatorTreeWrapperPass>();
869 AU.addRequired<PostDominatorTreeWrapperPass>();
870 AU.addRequired<LoopInfoWrapperPass>();
871 if (WidenFrequentBranches)
872 AU.addRequired<BranchProbabilityInfoWrapperPass>();
876 /// Same as above, but restricted to a single loop at a time. Can be
877 /// scheduled with other loop passes w/o breaking out of LPM
878 struct LoopGuardWideningLegacyPass : public LoopPass {
879 static char ID;
881 LoopGuardWideningLegacyPass() : LoopPass(ID) {
882 initializeLoopGuardWideningLegacyPassPass(*PassRegistry::getPassRegistry());
885 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
886 if (skipLoop(L))
887 return false;
888 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
889 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
890 auto *PDTWP = getAnalysisIfAvailable<PostDominatorTreeWrapperPass>();
891 auto *PDT = PDTWP ? &PDTWP->getPostDomTree() : nullptr;
892 BasicBlock *RootBB = L->getLoopPredecessor();
893 if (!RootBB)
894 RootBB = L->getHeader();
895 auto BlockFilter = [&](BasicBlock *BB) {
896 return BB == RootBB || L->contains(BB);
898 BranchProbabilityInfo *BPI = nullptr;
899 if (WidenFrequentBranches)
900 BPI = &getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI();
901 return GuardWideningImpl(DT, PDT, LI, BPI,
902 DT.getNode(RootBB), BlockFilter).run();
905 void getAnalysisUsage(AnalysisUsage &AU) const override {
906 if (WidenFrequentBranches)
907 AU.addRequired<BranchProbabilityInfoWrapperPass>();
908 AU.setPreservesCFG();
909 getLoopAnalysisUsage(AU);
910 AU.addPreserved<PostDominatorTreeWrapperPass>();
915 char GuardWideningLegacyPass::ID = 0;
916 char LoopGuardWideningLegacyPass::ID = 0;
918 INITIALIZE_PASS_BEGIN(GuardWideningLegacyPass, "guard-widening", "Widen guards",
919 false, false)
920 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
921 INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass)
922 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
923 if (WidenFrequentBranches)
924 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)
925 INITIALIZE_PASS_END(GuardWideningLegacyPass, "guard-widening", "Widen guards",
926 false, false)
928 INITIALIZE_PASS_BEGIN(LoopGuardWideningLegacyPass, "loop-guard-widening",
929 "Widen guards (within a single loop, as a loop pass)",
930 false, false)
931 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
932 INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass)
933 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
934 if (WidenFrequentBranches)
935 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)
936 INITIALIZE_PASS_END(LoopGuardWideningLegacyPass, "loop-guard-widening",
937 "Widen guards (within a single loop, as a loop pass)",
938 false, false)
940 FunctionPass *llvm::createGuardWideningPass() {
941 return new GuardWideningLegacyPass();
944 Pass *llvm::createLoopGuardWideningPass() {
945 return new LoopGuardWideningLegacyPass();