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 have 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/Transforms/Scalar.h"
30 #include "llvm/ADT/STLExtras.h"
31 #include "llvm/ADT/SmallPtrSet.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/Analysis/GlobalsModRef.h"
34 #include "llvm/Analysis/AssumptionCache.h"
35 #include "llvm/Analysis/CodeMetrics.h"
36 #include "llvm/Analysis/InstructionSimplify.h"
37 #include "llvm/Analysis/LoopInfo.h"
38 #include "llvm/Analysis/LoopPass.h"
39 #include "llvm/Analysis/ScalarEvolution.h"
40 #include "llvm/Analysis/TargetTransformInfo.h"
41 #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
42 #include "llvm/Analysis/BlockFrequencyInfo.h"
43 #include "llvm/Analysis/BranchProbabilityInfo.h"
44 #include "llvm/Support/BranchProbability.h"
45 #include "llvm/IR/Constants.h"
46 #include "llvm/IR/DerivedTypes.h"
47 #include "llvm/IR/Dominators.h"
48 #include "llvm/IR/Function.h"
49 #include "llvm/IR/Instructions.h"
50 #include "llvm/IR/Module.h"
51 #include "llvm/IR/MDBuilder.h"
52 #include "llvm/Support/CommandLine.h"
53 #include "llvm/Support/Debug.h"
54 #include "llvm/Support/raw_ostream.h"
55 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
56 #include "llvm/Transforms/Utils/Cloning.h"
57 #include "llvm/Transforms/Utils/Local.h"
63 #define DEBUG_TYPE "loop-unswitch"
65 STATISTIC(NumBranches, "Number of branches unswitched");
66 STATISTIC(NumSwitches, "Number of switches unswitched");
67 STATISTIC(NumSelects , "Number of selects unswitched");
68 STATISTIC(NumTrivial , "Number of unswitches that are trivial");
69 STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
70 STATISTIC(TotalInsts, "Total number of instructions analyzed");
72 // The specific value of 100 here was chosen based only on intuition and a
73 // few specific examples.
74 static cl::opt<unsigned>
75 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
76 cl::init(100), cl::Hidden);
79 LoopUnswitchWithBlockFrequency("loop-unswitch-with-block-frequency",
80 cl::init(false), cl::Hidden,
81 cl::desc("Enable the use of the block frequency analysis to access PGO "
82 "heuristics to minimize code growth in cold regions."));
84 static cl::opt<unsigned>
85 ColdnessThreshold("loop-unswitch-coldness-threshold", cl::init(1), cl::Hidden,
86 cl::desc("Coldness threshold in percentage. The loop header frequency "
87 "(relative to the entry frequency) is compared with this "
88 "threshold to determine if non-trivial unswitching should be "
93 class LUAnalysisCache {
95 typedef DenseMap<const SwitchInst*, SmallPtrSet<const Value *, 8> >
98 typedef UnswitchedValsMap::iterator UnswitchedValsIt;
100 struct LoopProperties {
101 unsigned CanBeUnswitchedCount;
102 unsigned WasUnswitchedCount;
103 unsigned SizeEstimation;
104 UnswitchedValsMap UnswitchedVals;
107 // Here we use std::map instead of DenseMap, since we need to keep valid
108 // LoopProperties pointer for current loop for better performance.
109 typedef std::map<const Loop*, LoopProperties> LoopPropsMap;
110 typedef LoopPropsMap::iterator LoopPropsMapIt;
112 LoopPropsMap LoopsProperties;
113 UnswitchedValsMap *CurLoopInstructions;
114 LoopProperties *CurrentLoopProperties;
116 // A loop unswitching with an estimated cost above this threshold
117 // is not performed. MaxSize is turned into unswitching quota for
118 // the current loop, and reduced correspondingly, though note that
119 // the quota is returned by releaseMemory() when the loop has been
120 // processed, so that MaxSize will return to its previous
121 // value. So in most cases MaxSize will equal the Threshold flag
122 // when a new loop is processed. An exception to that is that
123 // MaxSize will have a smaller value while processing nested loops
124 // that were introduced due to loop unswitching of an outer loop.
126 // FIXME: The way that MaxSize works is subtle and depends on the
127 // pass manager processing loops and calling releaseMemory() in a
128 // specific order. It would be good to find a more straightforward
129 // way of doing what MaxSize does.
134 : CurLoopInstructions(nullptr), CurrentLoopProperties(nullptr),
135 MaxSize(Threshold) {}
137 // Analyze loop. Check its size, calculate is it possible to unswitch
138 // it. Returns true if we can unswitch this loop.
139 bool countLoop(const Loop *L, const TargetTransformInfo &TTI,
140 AssumptionCache *AC);
142 // Clean all data related to given loop.
143 void forgetLoop(const Loop *L);
145 // Mark case value as unswitched.
146 // Since SI instruction can be partly unswitched, in order to avoid
147 // extra unswitching in cloned loops keep track all unswitched values.
148 void setUnswitched(const SwitchInst *SI, const Value *V);
150 // Check was this case value unswitched before or not.
151 bool isUnswitched(const SwitchInst *SI, const Value *V);
153 // Returns true if another unswitching could be done within the cost
155 bool CostAllowsUnswitching();
157 // Clone all loop-unswitch related loop properties.
158 // Redistribute unswitching quotas.
159 // Note, that new loop data is stored inside the VMap.
160 void cloneData(const Loop *NewLoop, const Loop *OldLoop,
161 const ValueToValueMapTy &VMap);
164 class LoopUnswitch : public LoopPass {
165 LoopInfo *LI; // Loop information
169 // Used to check if second loop needs processing after
170 // RewriteLoopBodyWithConditionConstant rewrites first loop.
171 std::vector<Loop*> LoopProcessWorklist;
173 LUAnalysisCache BranchesInfo;
177 // BFI and ColdEntryFreq are only used when PGO and
178 // LoopUnswitchWithBlockFrequency are enabled.
179 BlockFrequencyInfo BFI;
180 BlockFrequency ColdEntryFreq;
182 bool OptimizeForSize;
187 BasicBlock *loopHeader;
188 BasicBlock *loopPreheader;
190 // LoopBlocks contains all of the basic blocks of the loop, including the
191 // preheader of the loop, the body of the loop, and the exit blocks of the
192 // loop, in that order.
193 std::vector<BasicBlock*> LoopBlocks;
194 // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
195 std::vector<BasicBlock*> NewBlocks;
198 static char ID; // Pass ID, replacement for typeid
199 explicit LoopUnswitch(bool Os = false) :
200 LoopPass(ID), OptimizeForSize(Os), redoLoop(false),
201 currentLoop(nullptr), DT(nullptr), loopHeader(nullptr),
202 loopPreheader(nullptr) {
203 initializeLoopUnswitchPass(*PassRegistry::getPassRegistry());
206 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
207 bool processCurrentLoop();
209 /// This transformation requires natural loop information & requires that
210 /// loop preheaders be inserted into the CFG.
212 void getAnalysisUsage(AnalysisUsage &AU) const override {
213 AU.addRequired<AssumptionCacheTracker>();
214 AU.addRequiredID(LoopSimplifyID);
215 AU.addPreservedID(LoopSimplifyID);
216 AU.addRequired<LoopInfoWrapperPass>();
217 AU.addPreserved<LoopInfoWrapperPass>();
218 AU.addRequiredID(LCSSAID);
219 AU.addPreservedID(LCSSAID);
220 AU.addRequired<DominatorTreeWrapperPass>();
221 AU.addPreserved<DominatorTreeWrapperPass>();
222 AU.addPreserved<ScalarEvolutionWrapperPass>();
223 AU.addRequired<TargetTransformInfoWrapperPass>();
224 AU.addPreserved<GlobalsAAWrapperPass>();
229 void releaseMemory() override {
230 BranchesInfo.forgetLoop(currentLoop);
233 void initLoopData() {
234 loopHeader = currentLoop->getHeader();
235 loopPreheader = currentLoop->getLoopPreheader();
238 /// Split all of the edges from inside the loop to their exit blocks.
239 /// Update the appropriate Phi nodes as we do so.
240 void SplitExitEdges(Loop *L,
241 const SmallVectorImpl<BasicBlock *> &ExitBlocks);
243 bool TryTrivialLoopUnswitch(bool &Changed);
245 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val,
246 TerminatorInst *TI = nullptr);
247 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
248 BasicBlock *ExitBlock, TerminatorInst *TI);
249 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L,
252 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
253 Constant *Val, bool isEqual);
255 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
256 BasicBlock *TrueDest,
257 BasicBlock *FalseDest,
258 Instruction *InsertPt,
261 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
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) {
270 LoopPropsMapIt PropsIt;
272 std::tie(PropsIt, Inserted) =
273 LoopsProperties.insert(std::make_pair(L, LoopProperties()));
275 LoopProperties &Props = PropsIt->second;
280 // Limit the number of instructions to avoid causing significant code
281 // expansion, and the number of basic blocks, to avoid loops with
282 // large numbers of branches which cause loop unswitching to go crazy.
283 // This is a very ad-hoc heuristic.
285 SmallPtrSet<const Value *, 32> EphValues;
286 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
288 // FIXME: This is overly conservative because it does not take into
289 // consideration code simplification opportunities and code that can
290 // be shared by the resultant unswitched loops.
292 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); I != E;
294 Metrics.analyzeBasicBlock(*I, TTI, EphValues);
296 Props.SizeEstimation = Metrics.NumInsts;
297 Props.CanBeUnswitchedCount = MaxSize / (Props.SizeEstimation);
298 Props.WasUnswitchedCount = 0;
299 MaxSize -= Props.SizeEstimation * Props.CanBeUnswitchedCount;
301 if (Metrics.notDuplicatable) {
302 DEBUG(dbgs() << "NOT unswitching loop %"
303 << L->getHeader()->getName() << ", contents cannot be "
309 // Be careful. This links are good only before new loop addition.
310 CurrentLoopProperties = &Props;
311 CurLoopInstructions = &Props.UnswitchedVals;
316 // Clean all data related to given loop.
317 void LUAnalysisCache::forgetLoop(const Loop *L) {
319 LoopPropsMapIt LIt = LoopsProperties.find(L);
321 if (LIt != LoopsProperties.end()) {
322 LoopProperties &Props = LIt->second;
323 MaxSize += (Props.CanBeUnswitchedCount + Props.WasUnswitchedCount) *
324 Props.SizeEstimation;
325 LoopsProperties.erase(LIt);
328 CurrentLoopProperties = nullptr;
329 CurLoopInstructions = nullptr;
332 // Mark case value as unswitched.
333 // Since SI instruction can be partly unswitched, in order to avoid
334 // extra unswitching in cloned loops keep track all unswitched values.
335 void LUAnalysisCache::setUnswitched(const SwitchInst *SI, const Value *V) {
336 (*CurLoopInstructions)[SI].insert(V);
339 // Check was this case value unswitched before or not.
340 bool LUAnalysisCache::isUnswitched(const SwitchInst *SI, const Value *V) {
341 return (*CurLoopInstructions)[SI].count(V);
344 bool LUAnalysisCache::CostAllowsUnswitching() {
345 return CurrentLoopProperties->CanBeUnswitchedCount > 0;
348 // Clone all loop-unswitch related loop properties.
349 // Redistribute unswitching quotas.
350 // Note, that new loop data is stored inside the VMap.
351 void LUAnalysisCache::cloneData(const Loop *NewLoop, const Loop *OldLoop,
352 const ValueToValueMapTy &VMap) {
354 LoopProperties &NewLoopProps = LoopsProperties[NewLoop];
355 LoopProperties &OldLoopProps = *CurrentLoopProperties;
356 UnswitchedValsMap &Insts = OldLoopProps.UnswitchedVals;
358 // Reallocate "can-be-unswitched quota"
360 --OldLoopProps.CanBeUnswitchedCount;
361 ++OldLoopProps.WasUnswitchedCount;
362 NewLoopProps.WasUnswitchedCount = 0;
363 unsigned Quota = OldLoopProps.CanBeUnswitchedCount;
364 NewLoopProps.CanBeUnswitchedCount = Quota / 2;
365 OldLoopProps.CanBeUnswitchedCount = Quota - Quota / 2;
367 NewLoopProps.SizeEstimation = OldLoopProps.SizeEstimation;
369 // Clone unswitched values info:
370 // for new loop switches we clone info about values that was
371 // already unswitched and has redundant successors.
372 for (UnswitchedValsIt I = Insts.begin(); I != Insts.end(); ++I) {
373 const SwitchInst *OldInst = I->first;
374 Value *NewI = VMap.lookup(OldInst);
375 const SwitchInst *NewInst = cast_or_null<SwitchInst>(NewI);
376 assert(NewInst && "All instructions that are in SrcBB must be in VMap.");
378 NewLoopProps.UnswitchedVals[NewInst] = OldLoopProps.UnswitchedVals[OldInst];
382 char LoopUnswitch::ID = 0;
383 INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops",
385 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
386 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
387 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
388 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
389 INITIALIZE_PASS_DEPENDENCY(LCSSA)
390 INITIALIZE_PASS_END(LoopUnswitch, "loop-unswitch", "Unswitch loops",
393 Pass *llvm::createLoopUnswitchPass(bool Os) {
394 return new LoopUnswitch(Os);
397 /// Cond is a condition that occurs in L. If it is invariant in the loop, or has
398 /// an invariant piece, return the invariant. Otherwise, return null.
399 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
401 // We started analyze new instruction, increment scanned instructions counter.
404 // We can never unswitch on vector conditions.
405 if (Cond->getType()->isVectorTy())
408 // Constants should be folded, not unswitched on!
409 if (isa<Constant>(Cond)) return nullptr;
411 // TODO: Handle: br (VARIANT|INVARIANT).
413 // Hoist simple values out.
414 if (L->makeLoopInvariant(Cond, Changed))
417 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
418 if (BO->getOpcode() == Instruction::And ||
419 BO->getOpcode() == Instruction::Or) {
420 // If either the left or right side is invariant, we can unswitch on this,
421 // which will cause the branch to go away in one loop and the condition to
422 // simplify in the other one.
423 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
425 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
432 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
433 if (skipOptnoneFunction(L))
436 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
437 *L->getHeader()->getParent());
438 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
440 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
442 Function *F = currentLoop->getHeader()->getParent();
444 EnabledPGO = F->getEntryCount().hasValue();
446 if (LoopUnswitchWithBlockFrequency && EnabledPGO) {
447 BranchProbabilityInfo BPI(*F, *LI);
448 BFI.calculate(*L->getHeader()->getParent(), BPI, *LI);
450 // Use BranchProbability to compute a minimum frequency based on
451 // function entry baseline frequency. Loops with headers below this
452 // frequency are considered as cold.
453 const BranchProbability ColdProb(ColdnessThreshold, 100);
454 ColdEntryFreq = BlockFrequency(BFI.getEntryFreq()) * ColdProb;
457 bool Changed = false;
459 assert(currentLoop->isLCSSAForm(*DT));
461 Changed |= processCurrentLoop();
464 // FIXME: Reconstruct dom info, because it is not preserved properly.
470 /// Do actual work and unswitch loop if possible and profitable.
471 bool LoopUnswitch::processCurrentLoop() {
472 bool Changed = false;
476 // If LoopSimplify was unable to form a preheader, don't do any unswitching.
480 // Loops with indirectbr cannot be cloned.
481 if (!currentLoop->isSafeToClone())
484 // Without dedicated exits, splitting the exit edge may fail.
485 if (!currentLoop->hasDedicatedExits())
488 LLVMContext &Context = loopHeader->getContext();
490 // Analyze loop cost, and stop unswitching if loop content can not be duplicated.
491 if (!BranchesInfo.countLoop(
492 currentLoop, getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
493 *currentLoop->getHeader()->getParent()),
497 // Try trivial unswitch first before loop over other basic blocks in the loop.
498 if (TryTrivialLoopUnswitch(Changed)) {
502 // Do not unswitch loops containing convergent operations, as we might be
503 // making them control dependent on the unswitch value when they were not
505 // FIXME: This could be refined to only bail if the convergent operation is
506 // not already control-dependent on the unswitch value.
507 for (const auto BB : currentLoop->blocks()) {
508 for (auto &I : *BB) {
509 auto CS = CallSite(&I);
511 if (CS.hasFnAttr(Attribute::Convergent))
516 // Do not do non-trivial unswitch while optimizing for size.
517 // FIXME: Use Function::optForSize().
518 if (OptimizeForSize ||
519 loopHeader->getParent()->hasFnAttribute(Attribute::OptimizeForSize))
522 if (LoopUnswitchWithBlockFrequency && EnabledPGO) {
523 // Compute the weighted frequency of the hottest block in the
524 // loop (loopHeader in this case since inner loops should be
525 // processed before outer loop). If it is less than ColdFrequency,
526 // we should not unswitch.
527 BlockFrequency LoopEntryFreq = BFI.getBlockFreq(loopHeader);
528 if (LoopEntryFreq < ColdEntryFreq)
532 // Loop over all of the basic blocks in the loop. If we find an interior
533 // block that is branching on a loop-invariant condition, we can unswitch this
535 for (Loop::block_iterator I = currentLoop->block_begin(),
536 E = currentLoop->block_end(); I != E; ++I) {
537 TerminatorInst *TI = (*I)->getTerminator();
538 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
539 // If this isn't branching on an invariant condition, we can't unswitch
541 if (BI->isConditional()) {
542 // See if this, or some part of it, is loop invariant. If so, we can
543 // unswitch on it if we desire.
544 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
545 currentLoop, Changed);
547 UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(Context), TI)) {
552 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
553 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
554 currentLoop, Changed);
555 unsigned NumCases = SI->getNumCases();
556 if (LoopCond && NumCases) {
557 // Find a value to unswitch on:
558 // FIXME: this should chose the most expensive case!
559 // FIXME: scan for a case with a non-critical edge?
560 Constant *UnswitchVal = nullptr;
562 // Do not process same value again and again.
563 // At this point we have some cases already unswitched and
564 // some not yet unswitched. Let's find the first not yet unswitched one.
565 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
567 Constant *UnswitchValCandidate = i.getCaseValue();
568 if (!BranchesInfo.isUnswitched(SI, UnswitchValCandidate)) {
569 UnswitchVal = UnswitchValCandidate;
577 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
584 // Scan the instructions to check for unswitchable values.
585 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
587 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
588 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
589 currentLoop, Changed);
590 if (LoopCond && UnswitchIfProfitable(LoopCond,
591 ConstantInt::getTrue(Context))) {
600 /// Check to see if all paths from BB exit the loop with no side effects
601 /// (including infinite loops).
603 /// If true, we return true and set ExitBB to the block we
606 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
608 std::set<BasicBlock*> &Visited) {
609 if (!Visited.insert(BB).second) {
610 // Already visited. Without more analysis, this could indicate an infinite
614 if (!L->contains(BB)) {
615 // Otherwise, this is a loop exit, this is fine so long as this is the
617 if (ExitBB) return false;
622 // Otherwise, this is an unvisited intra-loop node. Check all successors.
623 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
624 // Check to see if the successor is a trivial loop exit.
625 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
629 // Okay, everything after this looks good, check to make sure that this block
630 // doesn't include any side effects.
631 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
632 if (I->mayHaveSideEffects())
638 /// Return true if the specified block unconditionally leads to an exit from
639 /// the specified loop, and has no side-effects in the process. If so, return
640 /// the block that is exited to, otherwise return null.
641 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
642 std::set<BasicBlock*> Visited;
643 Visited.insert(L->getHeader()); // Branches to header make infinite loops.
644 BasicBlock *ExitBB = nullptr;
645 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
650 /// We have found that we can unswitch currentLoop when LoopCond == Val to
651 /// simplify the loop. If we decide that this is profitable,
652 /// unswitch the loop, reprocess the pieces, then return true.
653 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val,
654 TerminatorInst *TI) {
655 // Check to see if it would be profitable to unswitch current loop.
656 if (!BranchesInfo.CostAllowsUnswitching()) {
657 DEBUG(dbgs() << "NOT unswitching loop %"
658 << currentLoop->getHeader()->getName()
659 << " at non-trivial condition '" << *Val
660 << "' == " << *LoopCond << "\n"
661 << ". Cost too high.\n");
665 UnswitchNontrivialCondition(LoopCond, Val, currentLoop, TI);
669 /// Recursively clone the specified loop and all of its children,
670 /// mapping the blocks with the specified map.
671 static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM,
672 LoopInfo *LI, LPPassManager *LPM) {
673 Loop *New = new Loop();
674 LPM->insertLoop(New, PL);
676 // Add all of the blocks in L to the new loop.
677 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
679 if (LI->getLoopFor(*I) == L)
680 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI);
682 // Add all of the subloops to the new loop.
683 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
684 CloneLoop(*I, New, VM, LI, LPM);
689 static void copyMetadata(Instruction *DstInst, const Instruction *SrcInst,
691 if (!SrcInst || !SrcInst->hasMetadata())
694 SmallVector<std::pair<unsigned, MDNode *>, 4> MDs;
695 SrcInst->getAllMetadata(MDs);
696 for (auto &MD : MDs) {
700 case LLVMContext::MD_prof:
701 if (Swapped && MD.second->getNumOperands() == 3 &&
702 isa<MDString>(MD.second->getOperand(0))) {
703 MDString *MDName = cast<MDString>(MD.second->getOperand(0));
704 if (MDName->getString() == "branch_weights") {
705 auto *ValT = cast_or_null<ConstantAsMetadata>(
706 MD.second->getOperand(1))->getValue();
707 auto *ValF = cast_or_null<ConstantAsMetadata>(
708 MD.second->getOperand(2))->getValue();
709 assert(ValT && ValF && "Invalid Operands of branch_weights");
711 MDBuilder(DstInst->getParent()->getContext())
712 .createBranchWeights(cast<ConstantInt>(ValF)->getZExtValue(),
713 cast<ConstantInt>(ValT)->getZExtValue());
718 case LLVMContext::MD_make_implicit:
719 case LLVMContext::MD_dbg:
720 DstInst->setMetadata(MD.first, MD.second);
725 /// Emit a conditional branch on two values if LIC == Val, branch to TrueDst,
726 /// otherwise branch to FalseDest. Insert the code immediately before InsertPt.
727 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
728 BasicBlock *TrueDest,
729 BasicBlock *FalseDest,
730 Instruction *InsertPt,
731 TerminatorInst *TI) {
732 // Insert a conditional branch on LIC to the two preheaders. The original
733 // code is the true version and the new code is the false version.
734 Value *BranchVal = LIC;
735 bool Swapped = false;
736 if (!isa<ConstantInt>(Val) ||
737 Val->getType() != Type::getInt1Ty(LIC->getContext()))
738 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val);
739 else if (Val != ConstantInt::getTrue(Val->getContext())) {
740 // We want to enter the new loop when the condition is true.
741 std::swap(TrueDest, FalseDest);
745 // Insert the new branch.
746 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
747 copyMetadata(BI, TI, Swapped);
749 // If either edge is critical, split it. This helps preserve LoopSimplify
750 // form for enclosing loops.
751 auto Options = CriticalEdgeSplittingOptions(DT, LI).setPreserveLCSSA();
752 SplitCriticalEdge(BI, 0, Options);
753 SplitCriticalEdge(BI, 1, Options);
756 /// Given a loop that has a trivial unswitchable condition in it (a cond branch
757 /// from its header block to its latch block, where the path through the loop
758 /// that doesn't execute its body has no side-effects), unswitch it. This
759 /// doesn't involve any code duplication, just moving the conditional branch
760 /// outside of the loop and updating loop info.
761 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
762 BasicBlock *ExitBlock,
763 TerminatorInst *TI) {
764 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %"
765 << loopHeader->getName() << " [" << L->getBlocks().size()
766 << " blocks] in Function "
767 << L->getHeader()->getParent()->getName() << " on cond: " << *Val
768 << " == " << *Cond << "\n");
770 // First step, split the preheader, so that we know that there is a safe place
771 // to insert the conditional branch. We will change loopPreheader to have a
772 // conditional branch on Cond.
773 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, DT, LI);
775 // Now that we have a place to insert the conditional branch, create a place
776 // to branch to: this is the exit block out of the loop that we should
779 // Split this block now, so that the loop maintains its exit block, and so
780 // that the jump from the preheader can execute the contents of the exit block
781 // without actually branching to it (the exit block should be dominated by the
782 // loop header, not the preheader).
783 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
784 BasicBlock *NewExit = SplitBlock(ExitBlock, &ExitBlock->front(), DT, LI);
786 // Okay, now we have a position to branch from and a position to branch to,
787 // insert the new conditional branch.
788 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
789 loopPreheader->getTerminator(), TI);
790 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
791 loopPreheader->getTerminator()->eraseFromParent();
793 // We need to reprocess this loop, it could be unswitched again.
796 // Now that we know that the loop is never entered when this condition is a
797 // particular value, rewrite the loop with this info. We know that this will
798 // at least eliminate the old branch.
799 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
803 /// Check if the first non-constant condition starting from the loop header is
804 /// a trivial unswitch condition: that is, a condition controls whether or not
805 /// the loop does anything at all. If it is a trivial condition, unswitching
806 /// produces no code duplications (equivalently, it produces a simpler loop and
807 /// a new empty loop, which gets deleted). Therefore always unswitch trivial
809 bool LoopUnswitch::TryTrivialLoopUnswitch(bool &Changed) {
810 BasicBlock *CurrentBB = currentLoop->getHeader();
811 TerminatorInst *CurrentTerm = CurrentBB->getTerminator();
812 LLVMContext &Context = CurrentBB->getContext();
814 // If loop header has only one reachable successor (currently via an
815 // unconditional branch or constant foldable conditional branch, but
816 // should also consider adding constant foldable switch instruction in
817 // future), we should keep looking for trivial condition candidates in
818 // the successor as well. An alternative is to constant fold conditions
819 // and merge successors into loop header (then we only need to check header's
820 // terminator). The reason for not doing this in LoopUnswitch pass is that
821 // it could potentially break LoopPassManager's invariants. Folding dead
822 // branches could either eliminate the current loop or make other loops
823 // unreachable. LCSSA form might also not be preserved after deleting
824 // branches. The following code keeps traversing loop header's successors
825 // until it finds the trivial condition candidate (condition that is not a
826 // constant). Since unswitching generates branches with constant conditions,
827 // this scenario could be very common in practice.
828 SmallSet<BasicBlock*, 8> Visited;
831 // If we exit loop or reach a previous visited block, then
832 // we can not reach any trivial condition candidates (unfoldable
833 // branch instructions or switch instructions) and no unswitch
834 // can happen. Exit and return false.
835 if (!currentLoop->contains(CurrentBB) || !Visited.insert(CurrentBB).second)
838 // Check if this loop will execute any side-effecting instructions (e.g.
839 // stores, calls, volatile loads) in the part of the loop that the code
840 // *would* execute. Check the header first.
841 for (Instruction &I : *CurrentBB)
842 if (I.mayHaveSideEffects())
845 // FIXME: add check for constant foldable switch instructions.
846 if (BranchInst *BI = dyn_cast<BranchInst>(CurrentTerm)) {
847 if (BI->isUnconditional()) {
848 CurrentBB = BI->getSuccessor(0);
849 } else if (BI->getCondition() == ConstantInt::getTrue(Context)) {
850 CurrentBB = BI->getSuccessor(0);
851 } else if (BI->getCondition() == ConstantInt::getFalse(Context)) {
852 CurrentBB = BI->getSuccessor(1);
854 // Found a trivial condition candidate: non-foldable conditional branch.
861 CurrentTerm = CurrentBB->getTerminator();
864 // CondVal is the condition that controls the trivial condition.
865 // LoopExitBB is the BasicBlock that loop exits when meets trivial condition.
866 Constant *CondVal = nullptr;
867 BasicBlock *LoopExitBB = nullptr;
869 if (BranchInst *BI = dyn_cast<BranchInst>(CurrentTerm)) {
870 // If this isn't branching on an invariant condition, we can't unswitch it.
871 if (!BI->isConditional())
874 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
875 currentLoop, Changed);
877 // Unswitch only if the trivial condition itself is an LIV (not
878 // partial LIV which could occur in and/or)
879 if (!LoopCond || LoopCond != BI->getCondition())
882 // Check to see if a successor of the branch is guaranteed to
883 // exit through a unique exit block without having any
884 // side-effects. If so, determine the value of Cond that causes
886 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
887 BI->getSuccessor(0)))) {
888 CondVal = ConstantInt::getTrue(Context);
889 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
890 BI->getSuccessor(1)))) {
891 CondVal = ConstantInt::getFalse(Context);
894 // If we didn't find a single unique LoopExit block, or if the loop exit
895 // block contains phi nodes, this isn't trivial.
896 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
897 return false; // Can't handle this.
899 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, LoopExitBB,
903 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurrentTerm)) {
904 // If this isn't switching on an invariant condition, we can't unswitch it.
905 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
906 currentLoop, Changed);
908 // Unswitch only if the trivial condition itself is an LIV (not
909 // partial LIV which could occur in and/or)
910 if (!LoopCond || LoopCond != SI->getCondition())
913 // Check to see if a successor of the switch is guaranteed to go to the
914 // latch block or exit through a one exit block without having any
915 // side-effects. If so, determine the value of Cond that causes it to do
917 // Note that we can't trivially unswitch on the default case or
918 // on already unswitched cases.
919 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
921 BasicBlock *LoopExitCandidate;
922 if ((LoopExitCandidate = isTrivialLoopExitBlock(currentLoop,
923 i.getCaseSuccessor()))) {
924 // Okay, we found a trivial case, remember the value that is trivial.
925 ConstantInt *CaseVal = i.getCaseValue();
927 // Check that it was not unswitched before, since already unswitched
928 // trivial vals are looks trivial too.
929 if (BranchesInfo.isUnswitched(SI, CaseVal))
931 LoopExitBB = LoopExitCandidate;
937 // If we didn't find a single unique LoopExit block, or if the loop exit
938 // block contains phi nodes, this isn't trivial.
939 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
940 return false; // Can't handle this.
942 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, LoopExitBB,
950 /// Split all of the edges from inside the loop to their exit blocks.
951 /// Update the appropriate Phi nodes as we do so.
952 void LoopUnswitch::SplitExitEdges(Loop *L,
953 const SmallVectorImpl<BasicBlock *> &ExitBlocks){
955 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
956 BasicBlock *ExitBlock = ExitBlocks[i];
957 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock),
958 pred_end(ExitBlock));
960 // Although SplitBlockPredecessors doesn't preserve loop-simplify in
961 // general, if we call it on all predecessors of all exits then it does.
962 SplitBlockPredecessors(ExitBlock, Preds, ".us-lcssa", DT, LI,
963 /*PreserveLCSSA*/ true);
967 /// We determined that the loop is profitable to unswitch when LIC equal Val.
968 /// Split it into loop versions and test the condition outside of either loop.
969 /// Return the loops created as Out1/Out2.
970 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
971 Loop *L, TerminatorInst *TI) {
972 Function *F = loopHeader->getParent();
973 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %"
974 << loopHeader->getName() << " [" << L->getBlocks().size()
975 << " blocks] in Function " << F->getName()
976 << " when '" << *Val << "' == " << *LIC << "\n");
978 if (auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>())
979 SEWP->getSE().forgetLoop(L);
984 // First step, split the preheader and exit blocks, and add these blocks to
985 // the LoopBlocks list.
986 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, DT, LI);
987 LoopBlocks.push_back(NewPreheader);
989 // We want the loop to come after the preheader, but before the exit blocks.
990 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
992 SmallVector<BasicBlock*, 8> ExitBlocks;
993 L->getUniqueExitBlocks(ExitBlocks);
995 // Split all of the edges from inside the loop to their exit blocks. Update
996 // the appropriate Phi nodes as we do so.
997 SplitExitEdges(L, ExitBlocks);
999 // The exit blocks may have been changed due to edge splitting, recompute.
1001 L->getUniqueExitBlocks(ExitBlocks);
1003 // Add exit blocks to the loop blocks.
1004 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
1006 // Next step, clone all of the basic blocks that make up the loop (including
1007 // the loop preheader and exit blocks), keeping track of the mapping between
1008 // the instructions and blocks.
1009 NewBlocks.reserve(LoopBlocks.size());
1010 ValueToValueMapTy VMap;
1011 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
1012 BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F);
1014 NewBlocks.push_back(NewBB);
1015 VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping.
1016 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L);
1019 // Splice the newly inserted blocks into the function right before the
1020 // original preheader.
1021 F->getBasicBlockList().splice(NewPreheader->getIterator(),
1022 F->getBasicBlockList(),
1023 NewBlocks[0]->getIterator(), F->end());
1025 // FIXME: We could register any cloned assumptions instead of clearing the
1026 // whole function's cache.
1029 // Now we create the new Loop object for the versioned loop.
1030 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM);
1032 // Recalculate unswitching quota, inherit simplified switches info for NewBB,
1033 // Probably clone more loop-unswitch related loop properties.
1034 BranchesInfo.cloneData(NewLoop, L, VMap);
1036 Loop *ParentLoop = L->getParentLoop();
1038 // Make sure to add the cloned preheader and exit blocks to the parent loop
1040 ParentLoop->addBasicBlockToLoop(NewBlocks[0], *LI);
1043 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
1044 BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]);
1045 // The new exit block should be in the same loop as the old one.
1046 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
1047 ExitBBLoop->addBasicBlockToLoop(NewExit, *LI);
1049 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
1050 "Exit block should have been split to have one successor!");
1051 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
1053 // If the successor of the exit block had PHI nodes, add an entry for
1055 for (BasicBlock::iterator I = ExitSucc->begin();
1056 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1057 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
1058 ValueToValueMapTy::iterator It = VMap.find(V);
1059 if (It != VMap.end()) V = It->second;
1060 PN->addIncoming(V, NewExit);
1063 if (LandingPadInst *LPad = NewExit->getLandingPadInst()) {
1064 PHINode *PN = PHINode::Create(LPad->getType(), 0, "",
1065 &*ExitSucc->getFirstInsertionPt());
1067 for (pred_iterator I = pred_begin(ExitSucc), E = pred_end(ExitSucc);
1069 BasicBlock *BB = *I;
1070 LandingPadInst *LPI = BB->getLandingPadInst();
1071 LPI->replaceAllUsesWith(PN);
1072 PN->addIncoming(LPI, BB);
1077 // Rewrite the code to refer to itself.
1078 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
1079 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
1080 E = NewBlocks[i]->end(); I != E; ++I)
1081 RemapInstruction(&*I, VMap,
1082 RF_NoModuleLevelChanges | RF_IgnoreMissingEntries);
1084 // Rewrite the original preheader to select between versions of the loop.
1085 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
1086 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
1087 "Preheader splitting did not work correctly!");
1089 // Emit the new branch that selects between the two versions of this loop.
1090 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR,
1092 LPM->deleteSimpleAnalysisValue(OldBR, L);
1093 OldBR->eraseFromParent();
1095 LoopProcessWorklist.push_back(NewLoop);
1098 // Keep a WeakVH holding onto LIC. If the first call to RewriteLoopBody
1099 // deletes the instruction (for example by simplifying a PHI that feeds into
1100 // the condition that we're unswitching on), we don't rewrite the second
1102 WeakVH LICHandle(LIC);
1104 // Now we rewrite the original code to know that the condition is true and the
1105 // new code to know that the condition is false.
1106 RewriteLoopBodyWithConditionConstant(L, LIC, Val, false);
1108 // It's possible that simplifying one loop could cause the other to be
1109 // changed to another value or a constant. If its a constant, don't simplify
1111 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop &&
1112 LICHandle && !isa<Constant>(LICHandle))
1113 RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true);
1116 /// Remove all instances of I from the worklist vector specified.
1117 static void RemoveFromWorklist(Instruction *I,
1118 std::vector<Instruction*> &Worklist) {
1120 Worklist.erase(std::remove(Worklist.begin(), Worklist.end(), I),
1124 /// When we find that I really equals V, remove I from the
1125 /// program, replacing all uses with V and update the worklist.
1126 static void ReplaceUsesOfWith(Instruction *I, Value *V,
1127 std::vector<Instruction*> &Worklist,
1128 Loop *L, LPPassManager *LPM) {
1129 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I);
1131 // Add uses to the worklist, which may be dead now.
1132 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1133 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1134 Worklist.push_back(Use);
1136 // Add users to the worklist which may be simplified now.
1137 for (User *U : I->users())
1138 Worklist.push_back(cast<Instruction>(U));
1139 LPM->deleteSimpleAnalysisValue(I, L);
1140 RemoveFromWorklist(I, Worklist);
1141 I->replaceAllUsesWith(V);
1142 I->eraseFromParent();
1146 /// We know either that the value LIC has the value specified by Val in the
1147 /// specified loop, or we know it does NOT have that value.
1148 /// Rewrite any uses of LIC or of properties correlated to it.
1149 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
1152 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
1154 // FIXME: Support correlated properties, like:
1161 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
1162 // selects, switches.
1163 std::vector<Instruction*> Worklist;
1164 LLVMContext &Context = Val->getContext();
1166 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
1167 // in the loop with the appropriate one directly.
1168 if (IsEqual || (isa<ConstantInt>(Val) &&
1169 Val->getType()->isIntegerTy(1))) {
1174 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
1175 !cast<ConstantInt>(Val)->getZExtValue());
1177 for (User *U : LIC->users()) {
1178 Instruction *UI = dyn_cast<Instruction>(U);
1179 if (!UI || !L->contains(UI))
1181 Worklist.push_back(UI);
1184 for (std::vector<Instruction*>::iterator UI = Worklist.begin(),
1185 UE = Worklist.end(); UI != UE; ++UI)
1186 (*UI)->replaceUsesOfWith(LIC, Replacement);
1188 SimplifyCode(Worklist, L);
1192 // Otherwise, we don't know the precise value of LIC, but we do know that it
1193 // is certainly NOT "Val". As such, simplify any uses in the loop that we
1194 // can. This case occurs when we unswitch switch statements.
1195 for (User *U : LIC->users()) {
1196 Instruction *UI = dyn_cast<Instruction>(U);
1197 if (!UI || !L->contains(UI))
1200 Worklist.push_back(UI);
1202 // TODO: We could do other simplifications, for example, turning
1203 // 'icmp eq LIC, Val' -> false.
1205 // If we know that LIC is not Val, use this info to simplify code.
1206 SwitchInst *SI = dyn_cast<SwitchInst>(UI);
1207 if (!SI || !isa<ConstantInt>(Val)) continue;
1209 SwitchInst::CaseIt DeadCase = SI->findCaseValue(cast<ConstantInt>(Val));
1210 // Default case is live for multiple values.
1211 if (DeadCase == SI->case_default()) continue;
1213 // Found a dead case value. Don't remove PHI nodes in the
1214 // successor if they become single-entry, those PHI nodes may
1215 // be in the Users list.
1217 BasicBlock *Switch = SI->getParent();
1218 BasicBlock *SISucc = DeadCase.getCaseSuccessor();
1219 BasicBlock *Latch = L->getLoopLatch();
1221 BranchesInfo.setUnswitched(SI, Val);
1223 if (!SI->findCaseDest(SISucc)) continue; // Edge is critical.
1224 // If the DeadCase successor dominates the loop latch, then the
1225 // transformation isn't safe since it will delete the sole predecessor edge
1227 if (Latch && DT->dominates(SISucc, Latch))
1230 // FIXME: This is a hack. We need to keep the successor around
1231 // and hooked up so as to preserve the loop structure, because
1232 // trying to update it is complicated. So instead we preserve the
1233 // loop structure and put the block on a dead code path.
1234 SplitEdge(Switch, SISucc, DT, LI);
1235 // Compute the successors instead of relying on the return value
1236 // of SplitEdge, since it may have split the switch successor
1238 BasicBlock *NewSISucc = DeadCase.getCaseSuccessor();
1239 BasicBlock *OldSISucc = *succ_begin(NewSISucc);
1240 // Create an "unreachable" destination.
1241 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable",
1242 Switch->getParent(),
1244 new UnreachableInst(Context, Abort);
1245 // Force the new case destination to branch to the "unreachable"
1246 // block while maintaining a (dead) CFG edge to the old block.
1247 NewSISucc->getTerminator()->eraseFromParent();
1248 BranchInst::Create(Abort, OldSISucc,
1249 ConstantInt::getTrue(Context), NewSISucc);
1250 // Release the PHI operands for this edge.
1251 for (BasicBlock::iterator II = NewSISucc->begin();
1252 PHINode *PN = dyn_cast<PHINode>(II); ++II)
1253 PN->setIncomingValue(PN->getBasicBlockIndex(Switch),
1254 UndefValue::get(PN->getType()));
1255 // Tell the domtree about the new block. We don't fully update the
1256 // domtree here -- instead we force it to do a full recomputation
1257 // after the pass is complete -- but we do need to inform it of
1259 DT->addNewBlock(Abort, NewSISucc);
1262 SimplifyCode(Worklist, L);
1265 /// Now that we have simplified some instructions in the loop, walk over it and
1266 /// constant prop, dce, and fold control flow where possible. Note that this is
1267 /// effectively a very simple loop-structure-aware optimizer. During processing
1268 /// of this loop, L could very well be deleted, so it must not be used.
1270 /// FIXME: When the loop optimizer is more mature, separate this out to a new
1273 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
1274 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1275 while (!Worklist.empty()) {
1276 Instruction *I = Worklist.back();
1277 Worklist.pop_back();
1280 if (isInstructionTriviallyDead(I)) {
1281 DEBUG(dbgs() << "Remove dead instruction '" << *I);
1283 // Add uses to the worklist, which may be dead now.
1284 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1285 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1286 Worklist.push_back(Use);
1287 LPM->deleteSimpleAnalysisValue(I, L);
1288 RemoveFromWorklist(I, Worklist);
1289 I->eraseFromParent();
1294 // See if instruction simplification can hack this up. This is common for
1295 // things like "select false, X, Y" after unswitching made the condition be
1296 // 'false'. TODO: update the domtree properly so we can pass it here.
1297 if (Value *V = SimplifyInstruction(I, DL))
1298 if (LI->replacementPreservesLCSSAForm(I, V)) {
1299 ReplaceUsesOfWith(I, V, Worklist, L, LPM);
1303 // Special case hacks that appear commonly in unswitched code.
1304 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1305 if (BI->isUnconditional()) {
1306 // If BI's parent is the only pred of the successor, fold the two blocks
1308 BasicBlock *Pred = BI->getParent();
1309 BasicBlock *Succ = BI->getSuccessor(0);
1310 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1311 if (!SinglePred) continue; // Nothing to do.
1312 assert(SinglePred == Pred && "CFG broken");
1314 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- "
1315 << Succ->getName() << "\n");
1317 // Resolve any single entry PHI nodes in Succ.
1318 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1319 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1321 // If Succ has any successors with PHI nodes, update them to have
1322 // entries coming from Pred instead of Succ.
1323 Succ->replaceAllUsesWith(Pred);
1325 // Move all of the successor contents from Succ to Pred.
1326 Pred->getInstList().splice(BI->getIterator(), Succ->getInstList(),
1327 Succ->begin(), Succ->end());
1328 LPM->deleteSimpleAnalysisValue(BI, L);
1329 BI->eraseFromParent();
1330 RemoveFromWorklist(BI, Worklist);
1332 // Remove Succ from the loop tree.
1333 LI->removeBlock(Succ);
1334 LPM->deleteSimpleAnalysisValue(Succ, L);
1335 Succ->eraseFromParent();