1 //===- LoopUnswitch.cpp - Hoist loop-invariant conditionals in loop -------===//
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 pass transforms loops that contain branches on loop-invariant conditions
11 // to multiple loops. For example, it turns the left into the right code:
20 // This can increase the size of the code exponentially (doubling it every time
21 // a loop is unswitched) so we only unswitch if the resultant code will be
22 // smaller than a threshold.
24 // This pass expects LICM to be run before it to hoist invariant conditions out
25 // of the loop, to make the unswitching opportunity obvious.
27 //===----------------------------------------------------------------------===//
29 #include "llvm/ADT/DenseMap.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/SmallSet.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/Analysis/AssumptionCache.h"
35 #include "llvm/Analysis/CodeMetrics.h"
36 #include "llvm/Analysis/DivergenceAnalysis.h"
37 #include "llvm/Analysis/InstructionSimplify.h"
38 #include "llvm/Analysis/LoopInfo.h"
39 #include "llvm/Analysis/LoopPass.h"
40 #include "llvm/Analysis/ScalarEvolution.h"
41 #include "llvm/Analysis/TargetTransformInfo.h"
42 #include "llvm/IR/Attributes.h"
43 #include "llvm/IR/BasicBlock.h"
44 #include "llvm/IR/CallSite.h"
45 #include "llvm/IR/Constant.h"
46 #include "llvm/IR/Constants.h"
47 #include "llvm/IR/DerivedTypes.h"
48 #include "llvm/IR/Dominators.h"
49 #include "llvm/IR/Function.h"
50 #include "llvm/IR/IRBuilder.h"
51 #include "llvm/IR/InstrTypes.h"
52 #include "llvm/IR/Instruction.h"
53 #include "llvm/IR/Instructions.h"
54 #include "llvm/IR/IntrinsicInst.h"
55 #include "llvm/IR/Intrinsics.h"
56 #include "llvm/IR/Module.h"
57 #include "llvm/IR/Type.h"
58 #include "llvm/IR/User.h"
59 #include "llvm/IR/Value.h"
60 #include "llvm/IR/ValueHandle.h"
61 #include "llvm/Pass.h"
62 #include "llvm/Support/Casting.h"
63 #include "llvm/Support/CommandLine.h"
64 #include "llvm/Support/Debug.h"
65 #include "llvm/Support/raw_ostream.h"
66 #include "llvm/Transforms/Scalar.h"
67 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
68 #include "llvm/Transforms/Utils/Cloning.h"
69 #include "llvm/Transforms/Utils/Local.h"
70 #include "llvm/Transforms/Utils/LoopUtils.h"
71 #include "llvm/Transforms/Utils/ValueMapper.h"
82 #define DEBUG_TYPE "loop-unswitch"
84 STATISTIC(NumBranches, "Number of branches unswitched");
85 STATISTIC(NumSwitches, "Number of switches unswitched");
86 STATISTIC(NumGuards, "Number of guards unswitched");
87 STATISTIC(NumSelects , "Number of selects unswitched");
88 STATISTIC(NumTrivial , "Number of unswitches that are trivial");
89 STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
90 STATISTIC(TotalInsts, "Total number of instructions analyzed");
92 // The specific value of 100 here was chosen based only on intuition and a
93 // few specific examples.
94 static cl::opt<unsigned>
95 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
96 cl::init(100), cl::Hidden);
100 class LUAnalysisCache {
101 using UnswitchedValsMap =
102 DenseMap<const SwitchInst *, SmallPtrSet<const Value *, 8>>;
103 using UnswitchedValsIt = UnswitchedValsMap::iterator;
105 struct LoopProperties {
106 unsigned CanBeUnswitchedCount;
107 unsigned WasUnswitchedCount;
108 unsigned SizeEstimation;
109 UnswitchedValsMap UnswitchedVals;
112 // Here we use std::map instead of DenseMap, since we need to keep valid
113 // LoopProperties pointer for current loop for better performance.
114 using LoopPropsMap = std::map<const Loop *, LoopProperties>;
115 using LoopPropsMapIt = LoopPropsMap::iterator;
117 LoopPropsMap LoopsProperties;
118 UnswitchedValsMap *CurLoopInstructions = nullptr;
119 LoopProperties *CurrentLoopProperties = nullptr;
121 // A loop unswitching with an estimated cost above this threshold
122 // is not performed. MaxSize is turned into unswitching quota for
123 // the current loop, and reduced correspondingly, though note that
124 // the quota is returned by releaseMemory() when the loop has been
125 // processed, so that MaxSize will return to its previous
126 // value. So in most cases MaxSize will equal the Threshold flag
127 // when a new loop is processed. An exception to that is that
128 // MaxSize will have a smaller value while processing nested loops
129 // that were introduced due to loop unswitching of an outer loop.
131 // FIXME: The way that MaxSize works is subtle and depends on the
132 // pass manager processing loops and calling releaseMemory() in a
133 // specific order. It would be good to find a more straightforward
134 // way of doing what MaxSize does.
138 LUAnalysisCache() : MaxSize(Threshold) {}
140 // Analyze loop. Check its size, calculate is it possible to unswitch
141 // it. Returns true if we can unswitch this loop.
142 bool countLoop(const Loop *L, const TargetTransformInfo &TTI,
143 AssumptionCache *AC);
145 // Clean all data related to given loop.
146 void forgetLoop(const Loop *L);
148 // Mark case value as unswitched.
149 // Since SI instruction can be partly unswitched, in order to avoid
150 // extra unswitching in cloned loops keep track all unswitched values.
151 void setUnswitched(const SwitchInst *SI, const Value *V);
153 // Check was this case value unswitched before or not.
154 bool isUnswitched(const SwitchInst *SI, const Value *V);
156 // Returns true if another unswitching could be done within the cost
158 bool CostAllowsUnswitching();
160 // Clone all loop-unswitch related loop properties.
161 // Redistribute unswitching quotas.
162 // Note, that new loop data is stored inside the VMap.
163 void cloneData(const Loop *NewLoop, const Loop *OldLoop,
164 const ValueToValueMapTy &VMap);
167 class LoopUnswitch : public LoopPass {
168 LoopInfo *LI; // Loop information
172 // Used to check if second loop needs processing after
173 // RewriteLoopBodyWithConditionConstant rewrites first loop.
174 std::vector<Loop*> LoopProcessWorklist;
176 LUAnalysisCache BranchesInfo;
178 bool OptimizeForSize;
179 bool redoLoop = false;
181 Loop *currentLoop = nullptr;
182 DominatorTree *DT = nullptr;
183 BasicBlock *loopHeader = nullptr;
184 BasicBlock *loopPreheader = nullptr;
187 LoopSafetyInfo SafetyInfo;
189 // LoopBlocks contains all of the basic blocks of the loop, including the
190 // preheader of the loop, the body of the loop, and the exit blocks of the
191 // loop, in that order.
192 std::vector<BasicBlock*> LoopBlocks;
193 // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
194 std::vector<BasicBlock*> NewBlocks;
196 bool hasBranchDivergence;
199 static char ID; // Pass ID, replacement for typeid
201 explicit LoopUnswitch(bool Os = false, bool hasBranchDivergence = false)
202 : LoopPass(ID), OptimizeForSize(Os),
203 hasBranchDivergence(hasBranchDivergence) {
204 initializeLoopUnswitchPass(*PassRegistry::getPassRegistry());
207 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
208 bool processCurrentLoop();
209 bool isUnreachableDueToPreviousUnswitching(BasicBlock *);
211 /// This transformation requires natural loop information & requires that
212 /// loop preheaders be inserted into the CFG.
214 void getAnalysisUsage(AnalysisUsage &AU) const override {
215 AU.addRequired<AssumptionCacheTracker>();
216 AU.addRequired<TargetTransformInfoWrapperPass>();
217 if (hasBranchDivergence)
218 AU.addRequired<DivergenceAnalysis>();
219 getLoopAnalysisUsage(AU);
223 void releaseMemory() override {
224 BranchesInfo.forgetLoop(currentLoop);
227 void initLoopData() {
228 loopHeader = currentLoop->getHeader();
229 loopPreheader = currentLoop->getLoopPreheader();
232 /// Split all of the edges from inside the loop to their exit blocks.
233 /// Update the appropriate Phi nodes as we do so.
234 void SplitExitEdges(Loop *L,
235 const SmallVectorImpl<BasicBlock *> &ExitBlocks);
237 bool TryTrivialLoopUnswitch(bool &Changed);
239 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val,
240 TerminatorInst *TI = nullptr);
241 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
242 BasicBlock *ExitBlock, TerminatorInst *TI);
243 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L,
246 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
247 Constant *Val, bool isEqual);
249 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
250 BasicBlock *TrueDest,
251 BasicBlock *FalseDest,
252 BranchInst *OldBranch,
255 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
257 /// Given that the Invariant is not equal to Val. Simplify instructions
259 Value *SimplifyInstructionWithNotEqual(Instruction *Inst, Value *Invariant,
263 } // end anonymous namespace
265 // Analyze loop. Check its size, calculate is it possible to unswitch
266 // it. Returns true if we can unswitch this loop.
267 bool LUAnalysisCache::countLoop(const Loop *L, const TargetTransformInfo &TTI,
268 AssumptionCache *AC) {
269 LoopPropsMapIt PropsIt;
271 std::tie(PropsIt, Inserted) =
272 LoopsProperties.insert(std::make_pair(L, LoopProperties()));
274 LoopProperties &Props = PropsIt->second;
279 // Limit the number of instructions to avoid causing significant code
280 // expansion, and the number of basic blocks, to avoid loops with
281 // large numbers of branches which cause loop unswitching to go crazy.
282 // This is a very ad-hoc heuristic.
284 SmallPtrSet<const Value *, 32> EphValues;
285 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
287 // FIXME: This is overly conservative because it does not take into
288 // consideration code simplification opportunities and code that can
289 // be shared by the resultant unswitched loops.
291 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); I != E;
293 Metrics.analyzeBasicBlock(*I, TTI, EphValues);
295 Props.SizeEstimation = Metrics.NumInsts;
296 Props.CanBeUnswitchedCount = MaxSize / (Props.SizeEstimation);
297 Props.WasUnswitchedCount = 0;
298 MaxSize -= Props.SizeEstimation * Props.CanBeUnswitchedCount;
300 if (Metrics.notDuplicatable) {
301 DEBUG(dbgs() << "NOT unswitching loop %"
302 << L->getHeader()->getName() << ", contents cannot be "
308 // Be careful. This links are good only before new loop addition.
309 CurrentLoopProperties = &Props;
310 CurLoopInstructions = &Props.UnswitchedVals;
315 // Clean all data related to given loop.
316 void LUAnalysisCache::forgetLoop(const Loop *L) {
317 LoopPropsMapIt LIt = LoopsProperties.find(L);
319 if (LIt != LoopsProperties.end()) {
320 LoopProperties &Props = LIt->second;
321 MaxSize += (Props.CanBeUnswitchedCount + Props.WasUnswitchedCount) *
322 Props.SizeEstimation;
323 LoopsProperties.erase(LIt);
326 CurrentLoopProperties = nullptr;
327 CurLoopInstructions = nullptr;
330 // Mark case value as unswitched.
331 // Since SI instruction can be partly unswitched, in order to avoid
332 // extra unswitching in cloned loops keep track all unswitched values.
333 void LUAnalysisCache::setUnswitched(const SwitchInst *SI, const Value *V) {
334 (*CurLoopInstructions)[SI].insert(V);
337 // Check was this case value unswitched before or not.
338 bool LUAnalysisCache::isUnswitched(const SwitchInst *SI, const Value *V) {
339 return (*CurLoopInstructions)[SI].count(V);
342 bool LUAnalysisCache::CostAllowsUnswitching() {
343 return CurrentLoopProperties->CanBeUnswitchedCount > 0;
346 // Clone all loop-unswitch related loop properties.
347 // Redistribute unswitching quotas.
348 // Note, that new loop data is stored inside the VMap.
349 void LUAnalysisCache::cloneData(const Loop *NewLoop, const Loop *OldLoop,
350 const ValueToValueMapTy &VMap) {
351 LoopProperties &NewLoopProps = LoopsProperties[NewLoop];
352 LoopProperties &OldLoopProps = *CurrentLoopProperties;
353 UnswitchedValsMap &Insts = OldLoopProps.UnswitchedVals;
355 // Reallocate "can-be-unswitched quota"
357 --OldLoopProps.CanBeUnswitchedCount;
358 ++OldLoopProps.WasUnswitchedCount;
359 NewLoopProps.WasUnswitchedCount = 0;
360 unsigned Quota = OldLoopProps.CanBeUnswitchedCount;
361 NewLoopProps.CanBeUnswitchedCount = Quota / 2;
362 OldLoopProps.CanBeUnswitchedCount = Quota - Quota / 2;
364 NewLoopProps.SizeEstimation = OldLoopProps.SizeEstimation;
366 // Clone unswitched values info:
367 // for new loop switches we clone info about values that was
368 // already unswitched and has redundant successors.
369 for (UnswitchedValsIt I = Insts.begin(); I != Insts.end(); ++I) {
370 const SwitchInst *OldInst = I->first;
371 Value *NewI = VMap.lookup(OldInst);
372 const SwitchInst *NewInst = cast_or_null<SwitchInst>(NewI);
373 assert(NewInst && "All instructions that are in SrcBB must be in VMap.");
375 NewLoopProps.UnswitchedVals[NewInst] = OldLoopProps.UnswitchedVals[OldInst];
379 char LoopUnswitch::ID = 0;
381 INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops",
383 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
384 INITIALIZE_PASS_DEPENDENCY(LoopPass)
385 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
386 INITIALIZE_PASS_DEPENDENCY(DivergenceAnalysis)
387 INITIALIZE_PASS_END(LoopUnswitch, "loop-unswitch", "Unswitch loops",
390 Pass *llvm::createLoopUnswitchPass(bool Os, bool hasBranchDivergence) {
391 return new LoopUnswitch(Os, hasBranchDivergence);
394 /// Operator chain lattice.
396 OC_OpChainNone, ///< There is no operator.
397 OC_OpChainOr, ///< There are only ORs.
398 OC_OpChainAnd, ///< There are only ANDs.
399 OC_OpChainMixed ///< There are ANDs and ORs.
402 /// Cond is a condition that occurs in L. If it is invariant in the loop, or has
403 /// an invariant piece, return the invariant. Otherwise, return null.
405 /// NOTE: FindLIVLoopCondition will not return a partial LIV by walking up a
406 /// mixed operator chain, as we can not reliably find a value which will simplify
407 /// the operator chain. If the chain is AND-only or OR-only, we can use 0 or ~0
408 /// to simplify the chain.
410 /// NOTE: In case a partial LIV and a mixed operator chain, we may be able to
411 /// simplify the condition itself to a loop variant condition, but at the
412 /// cost of creating an entirely new loop.
413 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed,
414 OperatorChain &ParentChain,
415 DenseMap<Value *, Value *> &Cache) {
416 auto CacheIt = Cache.find(Cond);
417 if (CacheIt != Cache.end())
418 return CacheIt->second;
420 // We started analyze new instruction, increment scanned instructions counter.
423 // We can never unswitch on vector conditions.
424 if (Cond->getType()->isVectorTy())
427 // Constants should be folded, not unswitched on!
428 if (isa<Constant>(Cond)) return nullptr;
430 // TODO: Handle: br (VARIANT|INVARIANT).
432 // Hoist simple values out.
433 if (L->makeLoopInvariant(Cond, Changed)) {
438 // Walk up the operator chain to find partial invariant conditions.
439 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
440 if (BO->getOpcode() == Instruction::And ||
441 BO->getOpcode() == Instruction::Or) {
442 // Given the previous operator, compute the current operator chain status.
443 OperatorChain NewChain;
444 switch (ParentChain) {
446 NewChain = BO->getOpcode() == Instruction::And ? OC_OpChainAnd :
450 NewChain = BO->getOpcode() == Instruction::Or ? OC_OpChainOr :
454 NewChain = BO->getOpcode() == Instruction::And ? OC_OpChainAnd :
457 case OC_OpChainMixed:
458 NewChain = OC_OpChainMixed;
462 // If we reach a Mixed state, we do not want to keep walking up as we can not
463 // reliably find a value that will simplify the chain. With this check, we
464 // will return null on the first sight of mixed chain and the caller will
465 // either backtrack to find partial LIV in other operand or return null.
466 if (NewChain != OC_OpChainMixed) {
467 // Update the current operator chain type before we search up the chain.
468 ParentChain = NewChain;
469 // If either the left or right side is invariant, we can unswitch on this,
470 // which will cause the branch to go away in one loop and the condition to
471 // simplify in the other one.
472 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed,
473 ParentChain, Cache)) {
477 // We did not manage to find a partial LIV in operand(0). Backtrack and try
479 ParentChain = NewChain;
480 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed,
481 ParentChain, Cache)) {
488 Cache[Cond] = nullptr;
492 /// Cond is a condition that occurs in L. If it is invariant in the loop, or has
493 /// an invariant piece, return the invariant along with the operator chain type.
494 /// Otherwise, return null.
495 static std::pair<Value *, OperatorChain> FindLIVLoopCondition(Value *Cond,
498 DenseMap<Value *, Value *> Cache;
499 OperatorChain OpChain = OC_OpChainNone;
500 Value *FCond = FindLIVLoopCondition(Cond, L, Changed, OpChain, Cache);
502 // In case we do find a LIV, it can not be obtained by walking up a mixed
504 assert((!FCond || OpChain != OC_OpChainMixed) &&
505 "Do not expect a partial LIV with mixed operator chain");
506 return {FCond, OpChain};
509 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
513 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
514 *L->getHeader()->getParent());
515 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
517 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
519 Function *F = currentLoop->getHeader()->getParent();
521 SanitizeMemory = F->hasFnAttribute(Attribute::SanitizeMemory);
523 computeLoopSafetyInfo(&SafetyInfo, L);
525 bool Changed = false;
527 assert(currentLoop->isLCSSAForm(*DT));
529 Changed |= processCurrentLoop();
535 // Return true if the BasicBlock BB is unreachable from the loop header.
536 // Return false, otherwise.
537 bool LoopUnswitch::isUnreachableDueToPreviousUnswitching(BasicBlock *BB) {
538 auto *Node = DT->getNode(BB)->getIDom();
539 BasicBlock *DomBB = Node->getBlock();
540 while (currentLoop->contains(DomBB)) {
541 BranchInst *BInst = dyn_cast<BranchInst>(DomBB->getTerminator());
543 Node = DT->getNode(DomBB)->getIDom();
544 DomBB = Node->getBlock();
546 if (!BInst || !BInst->isConditional())
549 Value *Cond = BInst->getCondition();
550 if (!isa<ConstantInt>(Cond))
553 BasicBlock *UnreachableSucc =
554 Cond == ConstantInt::getTrue(Cond->getContext())
555 ? BInst->getSuccessor(1)
556 : BInst->getSuccessor(0);
558 if (DT->dominates(UnreachableSucc, BB))
564 /// FIXME: Remove this workaround when freeze related patches are done.
565 /// LoopUnswitch and Equality propagation in GVN have discrepancy about
566 /// whether branch on undef/poison has undefine behavior. Here it is to
567 /// rule out some common cases that we found such discrepancy already
568 /// causing problems. Detail could be found in PR31652. Note if the
569 /// func returns true, it is unsafe. But if it is false, it doesn't mean
570 /// it is necessarily safe.
571 static bool EqualityPropUnSafe(Value &LoopCond) {
572 ICmpInst *CI = dyn_cast<ICmpInst>(&LoopCond);
573 if (!CI || !CI->isEquality())
576 Value *LHS = CI->getOperand(0);
577 Value *RHS = CI->getOperand(1);
578 if (isa<UndefValue>(LHS) || isa<UndefValue>(RHS))
581 auto hasUndefInPHI = [](PHINode &PN) {
582 for (Value *Opd : PN.incoming_values()) {
583 if (isa<UndefValue>(Opd))
588 PHINode *LPHI = dyn_cast<PHINode>(LHS);
589 PHINode *RPHI = dyn_cast<PHINode>(RHS);
590 if ((LPHI && hasUndefInPHI(*LPHI)) || (RPHI && hasUndefInPHI(*RPHI)))
593 auto hasUndefInSelect = [](SelectInst &SI) {
594 if (isa<UndefValue>(SI.getTrueValue()) ||
595 isa<UndefValue>(SI.getFalseValue()))
599 SelectInst *LSI = dyn_cast<SelectInst>(LHS);
600 SelectInst *RSI = dyn_cast<SelectInst>(RHS);
601 if ((LSI && hasUndefInSelect(*LSI)) || (RSI && hasUndefInSelect(*RSI)))
606 /// Do actual work and unswitch loop if possible and profitable.
607 bool LoopUnswitch::processCurrentLoop() {
608 bool Changed = false;
612 // If LoopSimplify was unable to form a preheader, don't do any unswitching.
616 // Loops with indirectbr cannot be cloned.
617 if (!currentLoop->isSafeToClone())
620 // Without dedicated exits, splitting the exit edge may fail.
621 if (!currentLoop->hasDedicatedExits())
624 LLVMContext &Context = loopHeader->getContext();
626 // Analyze loop cost, and stop unswitching if loop content can not be duplicated.
627 if (!BranchesInfo.countLoop(
628 currentLoop, getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
629 *currentLoop->getHeader()->getParent()),
633 // Try trivial unswitch first before loop over other basic blocks in the loop.
634 if (TryTrivialLoopUnswitch(Changed)) {
638 // Run through the instructions in the loop, keeping track of three things:
640 // - That we do not unswitch loops containing convergent operations, as we
641 // might be making them control dependent on the unswitch value when they
643 // FIXME: This could be refined to only bail if the convergent operation is
644 // not already control-dependent on the unswitch value.
646 // - That basic blocks in the loop contain invokes whose predecessor edges we
649 // - The set of guard intrinsics encountered (these are non terminator
650 // instructions that are also profitable to be unswitched).
652 SmallVector<IntrinsicInst *, 4> Guards;
654 for (const auto BB : currentLoop->blocks()) {
655 for (auto &I : *BB) {
656 auto CS = CallSite(&I);
658 if (CS.hasFnAttr(Attribute::Convergent))
660 if (auto *II = dyn_cast<InvokeInst>(&I))
661 if (!II->getUnwindDest()->canSplitPredecessors())
663 if (auto *II = dyn_cast<IntrinsicInst>(&I))
664 if (II->getIntrinsicID() == Intrinsic::experimental_guard)
665 Guards.push_back(II);
669 // Do not do non-trivial unswitch while optimizing for size.
670 // FIXME: Use Function::optForSize().
671 if (OptimizeForSize ||
672 loopHeader->getParent()->hasFnAttribute(Attribute::OptimizeForSize))
675 for (IntrinsicInst *Guard : Guards) {
677 FindLIVLoopCondition(Guard->getOperand(0), currentLoop, Changed).first;
679 UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(Context))) {
680 // NB! Unswitching (if successful) could have erased some of the
681 // instructions in Guards leaving dangling pointers there. This is fine
682 // because we're returning now, and won't look at Guards again.
688 // Loop over all of the basic blocks in the loop. If we find an interior
689 // block that is branching on a loop-invariant condition, we can unswitch this
691 for (Loop::block_iterator I = currentLoop->block_begin(),
692 E = currentLoop->block_end(); I != E; ++I) {
693 TerminatorInst *TI = (*I)->getTerminator();
695 // Unswitching on a potentially uninitialized predicate is not
696 // MSan-friendly. Limit this to the cases when the original predicate is
697 // guaranteed to execute, to avoid creating a use-of-uninitialized-value
698 // in the code that did not have one.
699 // This is a workaround for the discrepancy between LLVM IR and MSan
700 // semantics. See PR28054 for more details.
701 if (SanitizeMemory &&
702 !isGuaranteedToExecute(*TI, DT, currentLoop, &SafetyInfo))
705 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
706 // Some branches may be rendered unreachable because of previous
708 // Unswitch only those branches that are reachable.
709 if (isUnreachableDueToPreviousUnswitching(*I))
712 // If this isn't branching on an invariant condition, we can't unswitch
714 if (BI->isConditional()) {
715 // See if this, or some part of it, is loop invariant. If so, we can
716 // unswitch on it if we desire.
717 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
718 currentLoop, Changed).first;
719 if (LoopCond && !EqualityPropUnSafe(*LoopCond) &&
720 UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(Context), TI)) {
725 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
726 Value *SC = SI->getCondition();
728 OperatorChain OpChain;
729 std::tie(LoopCond, OpChain) =
730 FindLIVLoopCondition(SC, currentLoop, Changed);
732 unsigned NumCases = SI->getNumCases();
733 if (LoopCond && NumCases) {
734 // Find a value to unswitch on:
735 // FIXME: this should chose the most expensive case!
736 // FIXME: scan for a case with a non-critical edge?
737 Constant *UnswitchVal = nullptr;
738 // Find a case value such that at least one case value is unswitched
740 if (OpChain == OC_OpChainAnd) {
741 // If the chain only has ANDs and the switch has a case value of 0.
742 // Dropping in a 0 to the chain will unswitch out the 0-casevalue.
743 auto *AllZero = cast<ConstantInt>(Constant::getNullValue(SC->getType()));
744 if (BranchesInfo.isUnswitched(SI, AllZero))
746 // We are unswitching 0 out.
747 UnswitchVal = AllZero;
748 } else if (OpChain == OC_OpChainOr) {
749 // If the chain only has ORs and the switch has a case value of ~0.
750 // Dropping in a ~0 to the chain will unswitch out the ~0-casevalue.
751 auto *AllOne = cast<ConstantInt>(Constant::getAllOnesValue(SC->getType()));
752 if (BranchesInfo.isUnswitched(SI, AllOne))
754 // We are unswitching ~0 out.
755 UnswitchVal = AllOne;
757 assert(OpChain == OC_OpChainNone &&
758 "Expect to unswitch on trivial chain");
759 // Do not process same value again and again.
760 // At this point we have some cases already unswitched and
761 // some not yet unswitched. Let's find the first not yet unswitched one.
762 for (auto Case : SI->cases()) {
763 Constant *UnswitchValCandidate = Case.getCaseValue();
764 if (!BranchesInfo.isUnswitched(SI, UnswitchValCandidate)) {
765 UnswitchVal = UnswitchValCandidate;
774 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
776 // In case of a full LIV, UnswitchVal is the value we unswitched out.
777 // In case of a partial LIV, we only unswitch when its an AND-chain
778 // or OR-chain. In both cases switch input value simplifies to
780 BranchesInfo.setUnswitched(SI, UnswitchVal);
786 // Scan the instructions to check for unswitchable values.
787 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
789 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
790 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
791 currentLoop, Changed).first;
792 if (LoopCond && UnswitchIfProfitable(LoopCond,
793 ConstantInt::getTrue(Context))) {
802 /// Check to see if all paths from BB exit the loop with no side effects
803 /// (including infinite loops).
805 /// If true, we return true and set ExitBB to the block we
808 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
810 std::set<BasicBlock*> &Visited) {
811 if (!Visited.insert(BB).second) {
812 // Already visited. Without more analysis, this could indicate an infinite
816 if (!L->contains(BB)) {
817 // Otherwise, this is a loop exit, this is fine so long as this is the
819 if (ExitBB) return false;
824 // Otherwise, this is an unvisited intra-loop node. Check all successors.
825 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
826 // Check to see if the successor is a trivial loop exit.
827 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
831 // Okay, everything after this looks good, check to make sure that this block
832 // doesn't include any side effects.
833 for (Instruction &I : *BB)
834 if (I.mayHaveSideEffects())
840 /// Return true if the specified block unconditionally leads to an exit from
841 /// the specified loop, and has no side-effects in the process. If so, return
842 /// the block that is exited to, otherwise return null.
843 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
844 std::set<BasicBlock*> Visited;
845 Visited.insert(L->getHeader()); // Branches to header make infinite loops.
846 BasicBlock *ExitBB = nullptr;
847 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
852 /// We have found that we can unswitch currentLoop when LoopCond == Val to
853 /// simplify the loop. If we decide that this is profitable,
854 /// unswitch the loop, reprocess the pieces, then return true.
855 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val,
856 TerminatorInst *TI) {
857 // Check to see if it would be profitable to unswitch current loop.
858 if (!BranchesInfo.CostAllowsUnswitching()) {
859 DEBUG(dbgs() << "NOT unswitching loop %"
860 << currentLoop->getHeader()->getName()
861 << " at non-trivial condition '" << *Val
862 << "' == " << *LoopCond << "\n"
863 << ". Cost too high.\n");
866 if (hasBranchDivergence &&
867 getAnalysis<DivergenceAnalysis>().isDivergent(LoopCond)) {
868 DEBUG(dbgs() << "NOT unswitching loop %"
869 << currentLoop->getHeader()->getName()
870 << " at non-trivial condition '" << *Val
871 << "' == " << *LoopCond << "\n"
872 << ". Condition is divergent.\n");
876 UnswitchNontrivialCondition(LoopCond, Val, currentLoop, TI);
880 /// Recursively clone the specified loop and all of its children,
881 /// mapping the blocks with the specified map.
882 static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM,
883 LoopInfo *LI, LPPassManager *LPM) {
884 Loop &New = *LI->AllocateLoop();
886 PL->addChildLoop(&New);
888 LI->addTopLevelLoop(&New);
891 // Add all of the blocks in L to the new loop.
892 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
894 if (LI->getLoopFor(*I) == L)
895 New.addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI);
897 // Add all of the subloops to the new loop.
899 CloneLoop(I, &New, VM, LI, LPM);
904 /// Emit a conditional branch on two values if LIC == Val, branch to TrueDst,
905 /// otherwise branch to FalseDest. Insert the code immediately before OldBranch
906 /// and remove (but not erase!) it from the function.
907 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
908 BasicBlock *TrueDest,
909 BasicBlock *FalseDest,
910 BranchInst *OldBranch,
911 TerminatorInst *TI) {
912 assert(OldBranch->isUnconditional() && "Preheader is not split correctly");
913 // Insert a conditional branch on LIC to the two preheaders. The original
914 // code is the true version and the new code is the false version.
915 Value *BranchVal = LIC;
916 bool Swapped = false;
917 if (!isa<ConstantInt>(Val) ||
918 Val->getType() != Type::getInt1Ty(LIC->getContext()))
919 BranchVal = new ICmpInst(OldBranch, ICmpInst::ICMP_EQ, LIC, Val);
920 else if (Val != ConstantInt::getTrue(Val->getContext())) {
921 // We want to enter the new loop when the condition is true.
922 std::swap(TrueDest, FalseDest);
926 // Old branch will be removed, so save its parent and successor to update the
928 auto *OldBranchSucc = OldBranch->getSuccessor(0);
929 auto *OldBranchParent = OldBranch->getParent();
931 // Insert the new branch.
933 IRBuilder<>(OldBranch).CreateCondBr(BranchVal, TrueDest, FalseDest, TI);
935 BI->swapProfMetadata();
937 // Remove the old branch so there is only one branch at the end. This is
938 // needed to perform DomTree's internal DFS walk on the function's CFG.
939 OldBranch->removeFromParent();
941 // Inform the DT about the new branch.
943 // First, add both successors.
944 SmallVector<DominatorTree::UpdateType, 3> Updates;
945 if (TrueDest != OldBranchParent)
946 Updates.push_back({DominatorTree::Insert, OldBranchParent, TrueDest});
947 if (FalseDest != OldBranchParent)
948 Updates.push_back({DominatorTree::Insert, OldBranchParent, FalseDest});
949 // If both of the new successors are different from the old one, inform the
950 // DT that the edge was deleted.
951 if (OldBranchSucc != TrueDest && OldBranchSucc != FalseDest) {
952 Updates.push_back({DominatorTree::Delete, OldBranchParent, OldBranchSucc});
955 DT->applyUpdates(Updates);
958 // If either edge is critical, split it. This helps preserve LoopSimplify
959 // form for enclosing loops.
960 auto Options = CriticalEdgeSplittingOptions(DT, LI).setPreserveLCSSA();
961 SplitCriticalEdge(BI, 0, Options);
962 SplitCriticalEdge(BI, 1, Options);
965 /// Given a loop that has a trivial unswitchable condition in it (a cond branch
966 /// from its header block to its latch block, where the path through the loop
967 /// that doesn't execute its body has no side-effects), unswitch it. This
968 /// doesn't involve any code duplication, just moving the conditional branch
969 /// outside of the loop and updating loop info.
970 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
971 BasicBlock *ExitBlock,
972 TerminatorInst *TI) {
973 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %"
974 << loopHeader->getName() << " [" << L->getBlocks().size()
975 << " blocks] in Function "
976 << L->getHeader()->getParent()->getName() << " on cond: " << *Val
977 << " == " << *Cond << "\n");
979 // First step, split the preheader, so that we know that there is a safe place
980 // to insert the conditional branch. We will change loopPreheader to have a
981 // conditional branch on Cond.
982 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, DT, LI);
984 // Now that we have a place to insert the conditional branch, create a place
985 // to branch to: this is the exit block out of the loop that we should
988 // Split this block now, so that the loop maintains its exit block, and so
989 // that the jump from the preheader can execute the contents of the exit block
990 // without actually branching to it (the exit block should be dominated by the
991 // loop header, not the preheader).
992 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
993 BasicBlock *NewExit = SplitBlock(ExitBlock, &ExitBlock->front(), DT, LI);
995 // Okay, now we have a position to branch from and a position to branch to,
996 // insert the new conditional branch.
997 auto *OldBranch = dyn_cast<BranchInst>(loopPreheader->getTerminator());
998 assert(OldBranch && "Failed to split the preheader");
999 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, OldBranch, TI);
1000 LPM->deleteSimpleAnalysisValue(OldBranch, L);
1002 // EmitPreheaderBranchOnCondition removed the OldBranch from the function.
1003 // Delete it, as it is no longer needed.
1006 // We need to reprocess this loop, it could be unswitched again.
1009 // Now that we know that the loop is never entered when this condition is a
1010 // particular value, rewrite the loop with this info. We know that this will
1011 // at least eliminate the old branch.
1012 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
1016 /// Check if the first non-constant condition starting from the loop header is
1017 /// a trivial unswitch condition: that is, a condition controls whether or not
1018 /// the loop does anything at all. If it is a trivial condition, unswitching
1019 /// produces no code duplications (equivalently, it produces a simpler loop and
1020 /// a new empty loop, which gets deleted). Therefore always unswitch trivial
1022 bool LoopUnswitch::TryTrivialLoopUnswitch(bool &Changed) {
1023 BasicBlock *CurrentBB = currentLoop->getHeader();
1024 TerminatorInst *CurrentTerm = CurrentBB->getTerminator();
1025 LLVMContext &Context = CurrentBB->getContext();
1027 // If loop header has only one reachable successor (currently via an
1028 // unconditional branch or constant foldable conditional branch, but
1029 // should also consider adding constant foldable switch instruction in
1030 // future), we should keep looking for trivial condition candidates in
1031 // the successor as well. An alternative is to constant fold conditions
1032 // and merge successors into loop header (then we only need to check header's
1033 // terminator). The reason for not doing this in LoopUnswitch pass is that
1034 // it could potentially break LoopPassManager's invariants. Folding dead
1035 // branches could either eliminate the current loop or make other loops
1036 // unreachable. LCSSA form might also not be preserved after deleting
1037 // branches. The following code keeps traversing loop header's successors
1038 // until it finds the trivial condition candidate (condition that is not a
1039 // constant). Since unswitching generates branches with constant conditions,
1040 // this scenario could be very common in practice.
1041 SmallSet<BasicBlock*, 8> Visited;
1044 // If we exit loop or reach a previous visited block, then
1045 // we can not reach any trivial condition candidates (unfoldable
1046 // branch instructions or switch instructions) and no unswitch
1047 // can happen. Exit and return false.
1048 if (!currentLoop->contains(CurrentBB) || !Visited.insert(CurrentBB).second)
1051 // Check if this loop will execute any side-effecting instructions (e.g.
1052 // stores, calls, volatile loads) in the part of the loop that the code
1053 // *would* execute. Check the header first.
1054 for (Instruction &I : *CurrentBB)
1055 if (I.mayHaveSideEffects())
1058 if (BranchInst *BI = dyn_cast<BranchInst>(CurrentTerm)) {
1059 if (BI->isUnconditional()) {
1060 CurrentBB = BI->getSuccessor(0);
1061 } else if (BI->getCondition() == ConstantInt::getTrue(Context)) {
1062 CurrentBB = BI->getSuccessor(0);
1063 } else if (BI->getCondition() == ConstantInt::getFalse(Context)) {
1064 CurrentBB = BI->getSuccessor(1);
1066 // Found a trivial condition candidate: non-foldable conditional branch.
1069 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurrentTerm)) {
1070 // At this point, any constant-foldable instructions should have probably
1072 ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition());
1075 // Find the target block we are definitely going to.
1076 CurrentBB = SI->findCaseValue(Cond)->getCaseSuccessor();
1078 // We do not understand these terminator instructions.
1082 CurrentTerm = CurrentBB->getTerminator();
1085 // CondVal is the condition that controls the trivial condition.
1086 // LoopExitBB is the BasicBlock that loop exits when meets trivial condition.
1087 Constant *CondVal = nullptr;
1088 BasicBlock *LoopExitBB = nullptr;
1090 if (BranchInst *BI = dyn_cast<BranchInst>(CurrentTerm)) {
1091 // If this isn't branching on an invariant condition, we can't unswitch it.
1092 if (!BI->isConditional())
1095 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
1096 currentLoop, Changed).first;
1098 // Unswitch only if the trivial condition itself is an LIV (not
1099 // partial LIV which could occur in and/or)
1100 if (!LoopCond || LoopCond != BI->getCondition())
1103 // Check to see if a successor of the branch is guaranteed to
1104 // exit through a unique exit block without having any
1105 // side-effects. If so, determine the value of Cond that causes
1107 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
1108 BI->getSuccessor(0)))) {
1109 CondVal = ConstantInt::getTrue(Context);
1110 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
1111 BI->getSuccessor(1)))) {
1112 CondVal = ConstantInt::getFalse(Context);
1115 // If we didn't find a single unique LoopExit block, or if the loop exit
1116 // block contains phi nodes, this isn't trivial.
1117 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
1118 return false; // Can't handle this.
1120 if (EqualityPropUnSafe(*LoopCond))
1123 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, LoopExitBB,
1127 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurrentTerm)) {
1128 // If this isn't switching on an invariant condition, we can't unswitch it.
1129 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
1130 currentLoop, Changed).first;
1132 // Unswitch only if the trivial condition itself is an LIV (not
1133 // partial LIV which could occur in and/or)
1134 if (!LoopCond || LoopCond != SI->getCondition())
1137 // Check to see if a successor of the switch is guaranteed to go to the
1138 // latch block or exit through a one exit block without having any
1139 // side-effects. If so, determine the value of Cond that causes it to do
1141 // Note that we can't trivially unswitch on the default case or
1142 // on already unswitched cases.
1143 for (auto Case : SI->cases()) {
1144 BasicBlock *LoopExitCandidate;
1145 if ((LoopExitCandidate =
1146 isTrivialLoopExitBlock(currentLoop, Case.getCaseSuccessor()))) {
1147 // Okay, we found a trivial case, remember the value that is trivial.
1148 ConstantInt *CaseVal = Case.getCaseValue();
1150 // Check that it was not unswitched before, since already unswitched
1151 // trivial vals are looks trivial too.
1152 if (BranchesInfo.isUnswitched(SI, CaseVal))
1154 LoopExitBB = LoopExitCandidate;
1160 // If we didn't find a single unique LoopExit block, or if the loop exit
1161 // block contains phi nodes, this isn't trivial.
1162 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
1163 return false; // Can't handle this.
1165 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, LoopExitBB,
1168 // We are only unswitching full LIV.
1169 BranchesInfo.setUnswitched(SI, CondVal);
1176 /// Split all of the edges from inside the loop to their exit blocks.
1177 /// Update the appropriate Phi nodes as we do so.
1178 void LoopUnswitch::SplitExitEdges(Loop *L,
1179 const SmallVectorImpl<BasicBlock *> &ExitBlocks){
1181 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
1182 BasicBlock *ExitBlock = ExitBlocks[i];
1183 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock),
1184 pred_end(ExitBlock));
1186 // Although SplitBlockPredecessors doesn't preserve loop-simplify in
1187 // general, if we call it on all predecessors of all exits then it does.
1188 SplitBlockPredecessors(ExitBlock, Preds, ".us-lcssa", DT, LI,
1189 /*PreserveLCSSA*/ true);
1193 /// We determined that the loop is profitable to unswitch when LIC equal Val.
1194 /// Split it into loop versions and test the condition outside of either loop.
1195 /// Return the loops created as Out1/Out2.
1196 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
1197 Loop *L, TerminatorInst *TI) {
1198 Function *F = loopHeader->getParent();
1199 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %"
1200 << loopHeader->getName() << " [" << L->getBlocks().size()
1201 << " blocks] in Function " << F->getName()
1202 << " when '" << *Val << "' == " << *LIC << "\n");
1204 if (auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>())
1205 SEWP->getSE().forgetLoop(L);
1210 // First step, split the preheader and exit blocks, and add these blocks to
1211 // the LoopBlocks list.
1212 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, DT, LI);
1213 LoopBlocks.push_back(NewPreheader);
1215 // We want the loop to come after the preheader, but before the exit blocks.
1216 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
1218 SmallVector<BasicBlock*, 8> ExitBlocks;
1219 L->getUniqueExitBlocks(ExitBlocks);
1221 // Split all of the edges from inside the loop to their exit blocks. Update
1222 // the appropriate Phi nodes as we do so.
1223 SplitExitEdges(L, ExitBlocks);
1225 // The exit blocks may have been changed due to edge splitting, recompute.
1227 L->getUniqueExitBlocks(ExitBlocks);
1229 // Add exit blocks to the loop blocks.
1230 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
1232 // Next step, clone all of the basic blocks that make up the loop (including
1233 // the loop preheader and exit blocks), keeping track of the mapping between
1234 // the instructions and blocks.
1235 NewBlocks.reserve(LoopBlocks.size());
1236 ValueToValueMapTy VMap;
1237 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
1238 BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F);
1240 NewBlocks.push_back(NewBB);
1241 VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping.
1242 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L);
1245 // Splice the newly inserted blocks into the function right before the
1246 // original preheader.
1247 F->getBasicBlockList().splice(NewPreheader->getIterator(),
1248 F->getBasicBlockList(),
1249 NewBlocks[0]->getIterator(), F->end());
1251 // Now we create the new Loop object for the versioned loop.
1252 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM);
1254 // Recalculate unswitching quota, inherit simplified switches info for NewBB,
1255 // Probably clone more loop-unswitch related loop properties.
1256 BranchesInfo.cloneData(NewLoop, L, VMap);
1258 Loop *ParentLoop = L->getParentLoop();
1260 // Make sure to add the cloned preheader and exit blocks to the parent loop
1262 ParentLoop->addBasicBlockToLoop(NewBlocks[0], *LI);
1265 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
1266 BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]);
1267 // The new exit block should be in the same loop as the old one.
1268 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
1269 ExitBBLoop->addBasicBlockToLoop(NewExit, *LI);
1271 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
1272 "Exit block should have been split to have one successor!");
1273 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
1275 // If the successor of the exit block had PHI nodes, add an entry for
1277 for (PHINode &PN : ExitSucc->phis()) {
1278 Value *V = PN.getIncomingValueForBlock(ExitBlocks[i]);
1279 ValueToValueMapTy::iterator It = VMap.find(V);
1280 if (It != VMap.end()) V = It->second;
1281 PN.addIncoming(V, NewExit);
1284 if (LandingPadInst *LPad = NewExit->getLandingPadInst()) {
1285 PHINode *PN = PHINode::Create(LPad->getType(), 0, "",
1286 &*ExitSucc->getFirstInsertionPt());
1288 for (pred_iterator I = pred_begin(ExitSucc), E = pred_end(ExitSucc);
1290 BasicBlock *BB = *I;
1291 LandingPadInst *LPI = BB->getLandingPadInst();
1292 LPI->replaceAllUsesWith(PN);
1293 PN->addIncoming(LPI, BB);
1298 // Rewrite the code to refer to itself.
1299 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) {
1300 for (Instruction &I : *NewBlocks[i]) {
1301 RemapInstruction(&I, VMap,
1302 RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
1303 if (auto *II = dyn_cast<IntrinsicInst>(&I))
1304 if (II->getIntrinsicID() == Intrinsic::assume)
1305 AC->registerAssumption(II);
1309 // Rewrite the original preheader to select between versions of the loop.
1310 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
1311 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
1312 "Preheader splitting did not work correctly!");
1314 // Emit the new branch that selects between the two versions of this loop.
1315 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR,
1317 LPM->deleteSimpleAnalysisValue(OldBR, L);
1319 // The OldBr was replaced by a new one and removed (but not erased) by
1320 // EmitPreheaderBranchOnCondition. It is no longer needed, so delete it.
1323 LoopProcessWorklist.push_back(NewLoop);
1326 // Keep a WeakTrackingVH holding onto LIC. If the first call to
1328 // deletes the instruction (for example by simplifying a PHI that feeds into
1329 // the condition that we're unswitching on), we don't rewrite the second
1331 WeakTrackingVH LICHandle(LIC);
1333 // Now we rewrite the original code to know that the condition is true and the
1334 // new code to know that the condition is false.
1335 RewriteLoopBodyWithConditionConstant(L, LIC, Val, false);
1337 // It's possible that simplifying one loop could cause the other to be
1338 // changed to another value or a constant. If its a constant, don't simplify
1340 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop &&
1341 LICHandle && !isa<Constant>(LICHandle))
1342 RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true);
1345 /// Remove all instances of I from the worklist vector specified.
1346 static void RemoveFromWorklist(Instruction *I,
1347 std::vector<Instruction*> &Worklist) {
1349 Worklist.erase(std::remove(Worklist.begin(), Worklist.end(), I),
1353 /// When we find that I really equals V, remove I from the
1354 /// program, replacing all uses with V and update the worklist.
1355 static void ReplaceUsesOfWith(Instruction *I, Value *V,
1356 std::vector<Instruction*> &Worklist,
1357 Loop *L, LPPassManager *LPM) {
1358 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I << "\n");
1360 // Add uses to the worklist, which may be dead now.
1361 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1362 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1363 Worklist.push_back(Use);
1365 // Add users to the worklist which may be simplified now.
1366 for (User *U : I->users())
1367 Worklist.push_back(cast<Instruction>(U));
1368 LPM->deleteSimpleAnalysisValue(I, L);
1369 RemoveFromWorklist(I, Worklist);
1370 I->replaceAllUsesWith(V);
1371 if (!I->mayHaveSideEffects())
1372 I->eraseFromParent();
1376 /// We know either that the value LIC has the value specified by Val in the
1377 /// specified loop, or we know it does NOT have that value.
1378 /// Rewrite any uses of LIC or of properties correlated to it.
1379 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
1382 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
1384 // FIXME: Support correlated properties, like:
1391 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
1392 // selects, switches.
1393 std::vector<Instruction*> Worklist;
1394 LLVMContext &Context = Val->getContext();
1396 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
1397 // in the loop with the appropriate one directly.
1398 if (IsEqual || (isa<ConstantInt>(Val) &&
1399 Val->getType()->isIntegerTy(1))) {
1404 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
1405 !cast<ConstantInt>(Val)->getZExtValue());
1407 for (User *U : LIC->users()) {
1408 Instruction *UI = dyn_cast<Instruction>(U);
1409 if (!UI || !L->contains(UI))
1411 Worklist.push_back(UI);
1414 for (Instruction *UI : Worklist)
1415 UI->replaceUsesOfWith(LIC, Replacement);
1417 SimplifyCode(Worklist, L);
1421 // Otherwise, we don't know the precise value of LIC, but we do know that it
1422 // is certainly NOT "Val". As such, simplify any uses in the loop that we
1423 // can. This case occurs when we unswitch switch statements.
1424 for (User *U : LIC->users()) {
1425 Instruction *UI = dyn_cast<Instruction>(U);
1426 if (!UI || !L->contains(UI))
1429 // At this point, we know LIC is definitely not Val. Try to use some simple
1430 // logic to simplify the user w.r.t. to the context.
1431 if (Value *Replacement = SimplifyInstructionWithNotEqual(UI, LIC, Val)) {
1432 if (LI->replacementPreservesLCSSAForm(UI, Replacement)) {
1433 // This in-loop instruction has been simplified w.r.t. its context,
1434 // i.e. LIC != Val, make sure we propagate its replacement value to
1437 // We can not yet delete UI, the LIC user, yet, because that would invalidate
1438 // the LIC->users() iterator !. However, we can make this instruction
1439 // dead by replacing all its users and push it onto the worklist so that
1440 // it can be properly deleted and its operands simplified.
1441 UI->replaceAllUsesWith(Replacement);
1445 // This is a LIC user, push it into the worklist so that SimplifyCode can
1446 // attempt to simplify it.
1447 Worklist.push_back(UI);
1449 // If we know that LIC is not Val, use this info to simplify code.
1450 SwitchInst *SI = dyn_cast<SwitchInst>(UI);
1451 if (!SI || !isa<ConstantInt>(Val)) continue;
1453 // NOTE: if a case value for the switch is unswitched out, we record it
1454 // after the unswitch finishes. We can not record it here as the switch
1455 // is not a direct user of the partial LIV.
1456 SwitchInst::CaseHandle DeadCase =
1457 *SI->findCaseValue(cast<ConstantInt>(Val));
1458 // Default case is live for multiple values.
1459 if (DeadCase == *SI->case_default())
1462 // Found a dead case value. Don't remove PHI nodes in the
1463 // successor if they become single-entry, those PHI nodes may
1464 // be in the Users list.
1466 BasicBlock *Switch = SI->getParent();
1467 BasicBlock *SISucc = DeadCase.getCaseSuccessor();
1468 BasicBlock *Latch = L->getLoopLatch();
1470 if (!SI->findCaseDest(SISucc)) continue; // Edge is critical.
1471 // If the DeadCase successor dominates the loop latch, then the
1472 // transformation isn't safe since it will delete the sole predecessor edge
1474 if (Latch && DT->dominates(SISucc, Latch))
1477 // FIXME: This is a hack. We need to keep the successor around
1478 // and hooked up so as to preserve the loop structure, because
1479 // trying to update it is complicated. So instead we preserve the
1480 // loop structure and put the block on a dead code path.
1481 SplitEdge(Switch, SISucc, DT, LI);
1482 // Compute the successors instead of relying on the return value
1483 // of SplitEdge, since it may have split the switch successor
1485 BasicBlock *NewSISucc = DeadCase.getCaseSuccessor();
1486 BasicBlock *OldSISucc = *succ_begin(NewSISucc);
1487 // Create an "unreachable" destination.
1488 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable",
1489 Switch->getParent(),
1491 new UnreachableInst(Context, Abort);
1492 // Force the new case destination to branch to the "unreachable"
1493 // block while maintaining a (dead) CFG edge to the old block.
1494 NewSISucc->getTerminator()->eraseFromParent();
1495 BranchInst::Create(Abort, OldSISucc,
1496 ConstantInt::getTrue(Context), NewSISucc);
1497 // Release the PHI operands for this edge.
1498 for (PHINode &PN : NewSISucc->phis())
1499 PN.setIncomingValue(PN.getBasicBlockIndex(Switch),
1500 UndefValue::get(PN.getType()));
1501 // Tell the domtree about the new block. We don't fully update the
1502 // domtree here -- instead we force it to do a full recomputation
1503 // after the pass is complete -- but we do need to inform it of
1505 DT->addNewBlock(Abort, NewSISucc);
1508 SimplifyCode(Worklist, L);
1511 /// Now that we have simplified some instructions in the loop, walk over it and
1512 /// constant prop, dce, and fold control flow where possible. Note that this is
1513 /// effectively a very simple loop-structure-aware optimizer. During processing
1514 /// of this loop, L could very well be deleted, so it must not be used.
1516 /// FIXME: When the loop optimizer is more mature, separate this out to a new
1519 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
1520 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1521 while (!Worklist.empty()) {
1522 Instruction *I = Worklist.back();
1523 Worklist.pop_back();
1526 if (isInstructionTriviallyDead(I)) {
1527 DEBUG(dbgs() << "Remove dead instruction '" << *I << "\n");
1529 // Add uses to the worklist, which may be dead now.
1530 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1531 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1532 Worklist.push_back(Use);
1533 LPM->deleteSimpleAnalysisValue(I, L);
1534 RemoveFromWorklist(I, Worklist);
1535 I->eraseFromParent();
1540 // See if instruction simplification can hack this up. This is common for
1541 // things like "select false, X, Y" after unswitching made the condition be
1542 // 'false'. TODO: update the domtree properly so we can pass it here.
1543 if (Value *V = SimplifyInstruction(I, DL))
1544 if (LI->replacementPreservesLCSSAForm(I, V)) {
1545 ReplaceUsesOfWith(I, V, Worklist, L, LPM);
1549 // Special case hacks that appear commonly in unswitched code.
1550 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1551 if (BI->isUnconditional()) {
1552 // If BI's parent is the only pred of the successor, fold the two blocks
1554 BasicBlock *Pred = BI->getParent();
1555 BasicBlock *Succ = BI->getSuccessor(0);
1556 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1557 if (!SinglePred) continue; // Nothing to do.
1558 assert(SinglePred == Pred && "CFG broken");
1560 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- "
1561 << Succ->getName() << "\n");
1563 // Resolve any single entry PHI nodes in Succ.
1564 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1565 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1567 // If Succ has any successors with PHI nodes, update them to have
1568 // entries coming from Pred instead of Succ.
1569 Succ->replaceAllUsesWith(Pred);
1571 // Move all of the successor contents from Succ to Pred.
1572 Pred->getInstList().splice(BI->getIterator(), Succ->getInstList(),
1573 Succ->begin(), Succ->end());
1574 LPM->deleteSimpleAnalysisValue(BI, L);
1575 RemoveFromWorklist(BI, Worklist);
1576 BI->eraseFromParent();
1578 // Remove Succ from the loop tree.
1579 LI->removeBlock(Succ);
1580 LPM->deleteSimpleAnalysisValue(Succ, L);
1581 Succ->eraseFromParent();
1591 /// Simple simplifications we can do given the information that Cond is
1592 /// definitely not equal to Val.
1593 Value *LoopUnswitch::SimplifyInstructionWithNotEqual(Instruction *Inst,
1596 // icmp eq cond, val -> false
1597 ICmpInst *CI = dyn_cast<ICmpInst>(Inst);
1598 if (CI && CI->isEquality()) {
1599 Value *Op0 = CI->getOperand(0);
1600 Value *Op1 = CI->getOperand(1);
1601 if ((Op0 == Invariant && Op1 == Val) || (Op0 == Val && Op1 == Invariant)) {
1602 LLVMContext &Ctx = Inst->getContext();
1603 if (CI->getPredicate() == CmpInst::ICMP_EQ)
1604 return ConstantInt::getFalse(Ctx);
1606 return ConstantInt::getTrue(Ctx);
1610 // FIXME: there may be other opportunities, e.g. comparison with floating
1611 // point, or Invariant - Val != 0, etc.