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/IR/Constants.h"
42 #include "llvm/IR/DerivedTypes.h"
43 #include "llvm/IR/Dominators.h"
44 #include "llvm/IR/Function.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/Module.h"
47 #include "llvm/IR/MDBuilder.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/raw_ostream.h"
51 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
52 #include "llvm/Transforms/Utils/Cloning.h"
53 #include "llvm/Transforms/Utils/Local.h"
59 #define DEBUG_TYPE "loop-unswitch"
61 STATISTIC(NumBranches, "Number of branches unswitched");
62 STATISTIC(NumSwitches, "Number of switches unswitched");
63 STATISTIC(NumSelects , "Number of selects unswitched");
64 STATISTIC(NumTrivial , "Number of unswitches that are trivial");
65 STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
66 STATISTIC(TotalInsts, "Total number of instructions analyzed");
68 // The specific value of 100 here was chosen based only on intuition and a
69 // few specific examples.
70 static cl::opt<unsigned>
71 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
72 cl::init(100), cl::Hidden);
76 class LUAnalysisCache {
78 typedef DenseMap<const SwitchInst*, SmallPtrSet<const Value *, 8> >
81 typedef UnswitchedValsMap::iterator UnswitchedValsIt;
83 struct LoopProperties {
84 unsigned CanBeUnswitchedCount;
85 unsigned WasUnswitchedCount;
86 unsigned SizeEstimation;
87 UnswitchedValsMap UnswitchedVals;
90 // Here we use std::map instead of DenseMap, since we need to keep valid
91 // LoopProperties pointer for current loop for better performance.
92 typedef std::map<const Loop*, LoopProperties> LoopPropsMap;
93 typedef LoopPropsMap::iterator LoopPropsMapIt;
95 LoopPropsMap LoopsProperties;
96 UnswitchedValsMap *CurLoopInstructions;
97 LoopProperties *CurrentLoopProperties;
99 // A loop unswitching with an estimated cost above this threshold
100 // is not performed. MaxSize is turned into unswitching quota for
101 // the current loop, and reduced correspondingly, though note that
102 // the quota is returned by releaseMemory() when the loop has been
103 // processed, so that MaxSize will return to its previous
104 // value. So in most cases MaxSize will equal the Threshold flag
105 // when a new loop is processed. An exception to that is that
106 // MaxSize will have a smaller value while processing nested loops
107 // that were introduced due to loop unswitching of an outer loop.
109 // FIXME: The way that MaxSize works is subtle and depends on the
110 // pass manager processing loops and calling releaseMemory() in a
111 // specific order. It would be good to find a more straightforward
112 // way of doing what MaxSize does.
117 : CurLoopInstructions(nullptr), CurrentLoopProperties(nullptr),
118 MaxSize(Threshold) {}
120 // Analyze loop. Check its size, calculate is it possible to unswitch
121 // it. Returns true if we can unswitch this loop.
122 bool countLoop(const Loop *L, const TargetTransformInfo &TTI,
123 AssumptionCache *AC);
125 // Clean all data related to given loop.
126 void forgetLoop(const Loop *L);
128 // Mark case value as unswitched.
129 // Since SI instruction can be partly unswitched, in order to avoid
130 // extra unswitching in cloned loops keep track all unswitched values.
131 void setUnswitched(const SwitchInst *SI, const Value *V);
133 // Check was this case value unswitched before or not.
134 bool isUnswitched(const SwitchInst *SI, const Value *V);
136 // Returns true if another unswitching could be done within the cost
138 bool CostAllowsUnswitching();
140 // Clone all loop-unswitch related loop properties.
141 // Redistribute unswitching quotas.
142 // Note, that new loop data is stored inside the VMap.
143 void cloneData(const Loop *NewLoop, const Loop *OldLoop,
144 const ValueToValueMapTy &VMap);
147 class LoopUnswitch : public LoopPass {
148 LoopInfo *LI; // Loop information
152 // Used to check if second loop needs processing after
153 // RewriteLoopBodyWithConditionConstant rewrites first loop.
154 std::vector<Loop*> LoopProcessWorklist;
156 LUAnalysisCache BranchesInfo;
158 bool OptimizeForSize;
163 BasicBlock *loopHeader;
164 BasicBlock *loopPreheader;
166 // LoopBlocks contains all of the basic blocks of the loop, including the
167 // preheader of the loop, the body of the loop, and the exit blocks of the
168 // loop, in that order.
169 std::vector<BasicBlock*> LoopBlocks;
170 // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
171 std::vector<BasicBlock*> NewBlocks;
174 static char ID; // Pass ID, replacement for typeid
175 explicit LoopUnswitch(bool Os = false) :
176 LoopPass(ID), OptimizeForSize(Os), redoLoop(false),
177 currentLoop(nullptr), DT(nullptr), loopHeader(nullptr),
178 loopPreheader(nullptr) {
179 initializeLoopUnswitchPass(*PassRegistry::getPassRegistry());
182 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
183 bool processCurrentLoop();
185 /// This transformation requires natural loop information & requires that
186 /// loop preheaders be inserted into the CFG.
188 void getAnalysisUsage(AnalysisUsage &AU) const override {
189 AU.addRequired<AssumptionCacheTracker>();
190 AU.addRequiredID(LoopSimplifyID);
191 AU.addPreservedID(LoopSimplifyID);
192 AU.addRequired<LoopInfoWrapperPass>();
193 AU.addPreserved<LoopInfoWrapperPass>();
194 AU.addRequiredID(LCSSAID);
195 AU.addPreservedID(LCSSAID);
196 AU.addRequired<DominatorTreeWrapperPass>();
197 AU.addPreserved<DominatorTreeWrapperPass>();
198 AU.addPreserved<ScalarEvolutionWrapperPass>();
199 AU.addRequired<TargetTransformInfoWrapperPass>();
200 AU.addPreserved<GlobalsAAWrapperPass>();
205 void releaseMemory() override {
206 BranchesInfo.forgetLoop(currentLoop);
209 void initLoopData() {
210 loopHeader = currentLoop->getHeader();
211 loopPreheader = currentLoop->getLoopPreheader();
214 /// Split all of the edges from inside the loop to their exit blocks.
215 /// Update the appropriate Phi nodes as we do so.
216 void SplitExitEdges(Loop *L,
217 const SmallVectorImpl<BasicBlock *> &ExitBlocks);
219 bool TryTrivialLoopUnswitch(bool &Changed);
221 bool UnswitchIfProfitable(Value *LoopCond, Constant *Val,
222 TerminatorInst *TI = nullptr);
223 void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
224 BasicBlock *ExitBlock, TerminatorInst *TI);
225 void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L,
228 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
229 Constant *Val, bool isEqual);
231 void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
232 BasicBlock *TrueDest,
233 BasicBlock *FalseDest,
234 Instruction *InsertPt,
237 void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
241 // Analyze loop. Check its size, calculate is it possible to unswitch
242 // it. Returns true if we can unswitch this loop.
243 bool LUAnalysisCache::countLoop(const Loop *L, const TargetTransformInfo &TTI,
244 AssumptionCache *AC) {
246 LoopPropsMapIt PropsIt;
248 std::tie(PropsIt, Inserted) =
249 LoopsProperties.insert(std::make_pair(L, LoopProperties()));
251 LoopProperties &Props = PropsIt->second;
256 // Limit the number of instructions to avoid causing significant code
257 // expansion, and the number of basic blocks, to avoid loops with
258 // large numbers of branches which cause loop unswitching to go crazy.
259 // This is a very ad-hoc heuristic.
261 SmallPtrSet<const Value *, 32> EphValues;
262 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
264 // FIXME: This is overly conservative because it does not take into
265 // consideration code simplification opportunities and code that can
266 // be shared by the resultant unswitched loops.
268 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); I != E;
270 Metrics.analyzeBasicBlock(*I, TTI, EphValues);
272 Props.SizeEstimation = Metrics.NumInsts;
273 Props.CanBeUnswitchedCount = MaxSize / (Props.SizeEstimation);
274 Props.WasUnswitchedCount = 0;
275 MaxSize -= Props.SizeEstimation * Props.CanBeUnswitchedCount;
277 if (Metrics.notDuplicatable) {
278 DEBUG(dbgs() << "NOT unswitching loop %"
279 << L->getHeader()->getName() << ", contents cannot be "
285 // Be careful. This links are good only before new loop addition.
286 CurrentLoopProperties = &Props;
287 CurLoopInstructions = &Props.UnswitchedVals;
292 // Clean all data related to given loop.
293 void LUAnalysisCache::forgetLoop(const Loop *L) {
295 LoopPropsMapIt LIt = LoopsProperties.find(L);
297 if (LIt != LoopsProperties.end()) {
298 LoopProperties &Props = LIt->second;
299 MaxSize += (Props.CanBeUnswitchedCount + Props.WasUnswitchedCount) *
300 Props.SizeEstimation;
301 LoopsProperties.erase(LIt);
304 CurrentLoopProperties = nullptr;
305 CurLoopInstructions = nullptr;
308 // Mark case value as unswitched.
309 // Since SI instruction can be partly unswitched, in order to avoid
310 // extra unswitching in cloned loops keep track all unswitched values.
311 void LUAnalysisCache::setUnswitched(const SwitchInst *SI, const Value *V) {
312 (*CurLoopInstructions)[SI].insert(V);
315 // Check was this case value unswitched before or not.
316 bool LUAnalysisCache::isUnswitched(const SwitchInst *SI, const Value *V) {
317 return (*CurLoopInstructions)[SI].count(V);
320 bool LUAnalysisCache::CostAllowsUnswitching() {
321 return CurrentLoopProperties->CanBeUnswitchedCount > 0;
324 // Clone all loop-unswitch related loop properties.
325 // Redistribute unswitching quotas.
326 // Note, that new loop data is stored inside the VMap.
327 void LUAnalysisCache::cloneData(const Loop *NewLoop, const Loop *OldLoop,
328 const ValueToValueMapTy &VMap) {
330 LoopProperties &NewLoopProps = LoopsProperties[NewLoop];
331 LoopProperties &OldLoopProps = *CurrentLoopProperties;
332 UnswitchedValsMap &Insts = OldLoopProps.UnswitchedVals;
334 // Reallocate "can-be-unswitched quota"
336 --OldLoopProps.CanBeUnswitchedCount;
337 ++OldLoopProps.WasUnswitchedCount;
338 NewLoopProps.WasUnswitchedCount = 0;
339 unsigned Quota = OldLoopProps.CanBeUnswitchedCount;
340 NewLoopProps.CanBeUnswitchedCount = Quota / 2;
341 OldLoopProps.CanBeUnswitchedCount = Quota - Quota / 2;
343 NewLoopProps.SizeEstimation = OldLoopProps.SizeEstimation;
345 // Clone unswitched values info:
346 // for new loop switches we clone info about values that was
347 // already unswitched and has redundant successors.
348 for (UnswitchedValsIt I = Insts.begin(); I != Insts.end(); ++I) {
349 const SwitchInst *OldInst = I->first;
350 Value *NewI = VMap.lookup(OldInst);
351 const SwitchInst *NewInst = cast_or_null<SwitchInst>(NewI);
352 assert(NewInst && "All instructions that are in SrcBB must be in VMap.");
354 NewLoopProps.UnswitchedVals[NewInst] = OldLoopProps.UnswitchedVals[OldInst];
358 char LoopUnswitch::ID = 0;
359 INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops",
361 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
362 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
363 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
364 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
365 INITIALIZE_PASS_DEPENDENCY(LCSSA)
366 INITIALIZE_PASS_END(LoopUnswitch, "loop-unswitch", "Unswitch loops",
369 Pass *llvm::createLoopUnswitchPass(bool Os) {
370 return new LoopUnswitch(Os);
373 /// Cond is a condition that occurs in L. If it is invariant in the loop, or has
374 /// an invariant piece, return the invariant. Otherwise, return null.
375 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
377 // We started analyze new instruction, increment scanned instructions counter.
380 // We can never unswitch on vector conditions.
381 if (Cond->getType()->isVectorTy())
384 // Constants should be folded, not unswitched on!
385 if (isa<Constant>(Cond)) return nullptr;
387 // TODO: Handle: br (VARIANT|INVARIANT).
389 // Hoist simple values out.
390 if (L->makeLoopInvariant(Cond, Changed))
393 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
394 if (BO->getOpcode() == Instruction::And ||
395 BO->getOpcode() == Instruction::Or) {
396 // If either the left or right side is invariant, we can unswitch on this,
397 // which will cause the branch to go away in one loop and the condition to
398 // simplify in the other one.
399 if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
401 if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
408 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
409 if (skipOptnoneFunction(L))
412 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
413 *L->getHeader()->getParent());
414 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
416 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
418 Function *F = currentLoop->getHeader()->getParent();
419 bool Changed = false;
421 assert(currentLoop->isLCSSAForm(*DT));
423 Changed |= processCurrentLoop();
426 // FIXME: Reconstruct dom info, because it is not preserved properly.
432 /// Do actual work and unswitch loop if possible and profitable.
433 bool LoopUnswitch::processCurrentLoop() {
434 bool Changed = false;
438 // If LoopSimplify was unable to form a preheader, don't do any unswitching.
442 // Loops with indirectbr cannot be cloned.
443 if (!currentLoop->isSafeToClone())
446 // Without dedicated exits, splitting the exit edge may fail.
447 if (!currentLoop->hasDedicatedExits())
450 LLVMContext &Context = loopHeader->getContext();
452 // Probably we reach the quota of branches for this loop. If so
454 if (!BranchesInfo.countLoop(
455 currentLoop, getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
456 *currentLoop->getHeader()->getParent()),
460 // Try trivial unswitch first before loop over other basic blocks in the loop.
461 if (TryTrivialLoopUnswitch(Changed)) {
465 // Do not do non-trivial unswitch while optimizing for size.
466 // FIXME: Use Function::optForSize().
467 if (OptimizeForSize ||
468 loopHeader->getParent()->hasFnAttribute(Attribute::OptimizeForSize))
471 // Loop over all of the basic blocks in the loop. If we find an interior
472 // block that is branching on a loop-invariant condition, we can unswitch this
474 for (Loop::block_iterator I = currentLoop->block_begin(),
475 E = currentLoop->block_end(); I != E; ++I) {
476 TerminatorInst *TI = (*I)->getTerminator();
477 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
478 // If this isn't branching on an invariant condition, we can't unswitch
480 if (BI->isConditional()) {
481 // See if this, or some part of it, is loop invariant. If so, we can
482 // unswitch on it if we desire.
483 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
484 currentLoop, Changed);
486 UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(Context), TI)) {
491 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
492 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
493 currentLoop, Changed);
494 unsigned NumCases = SI->getNumCases();
495 if (LoopCond && NumCases) {
496 // Find a value to unswitch on:
497 // FIXME: this should chose the most expensive case!
498 // FIXME: scan for a case with a non-critical edge?
499 Constant *UnswitchVal = nullptr;
501 // Do not process same value again and again.
502 // At this point we have some cases already unswitched and
503 // some not yet unswitched. Let's find the first not yet unswitched one.
504 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
506 Constant *UnswitchValCandidate = i.getCaseValue();
507 if (!BranchesInfo.isUnswitched(SI, UnswitchValCandidate)) {
508 UnswitchVal = UnswitchValCandidate;
516 if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
523 // Scan the instructions to check for unswitchable values.
524 for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
526 if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
527 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
528 currentLoop, Changed);
529 if (LoopCond && UnswitchIfProfitable(LoopCond,
530 ConstantInt::getTrue(Context))) {
539 /// Check to see if all paths from BB exit the loop with no side effects
540 /// (including infinite loops).
542 /// If true, we return true and set ExitBB to the block we
545 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
547 std::set<BasicBlock*> &Visited) {
548 if (!Visited.insert(BB).second) {
549 // Already visited. Without more analysis, this could indicate an infinite
553 if (!L->contains(BB)) {
554 // Otherwise, this is a loop exit, this is fine so long as this is the
556 if (ExitBB) return false;
561 // Otherwise, this is an unvisited intra-loop node. Check all successors.
562 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
563 // Check to see if the successor is a trivial loop exit.
564 if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
568 // Okay, everything after this looks good, check to make sure that this block
569 // doesn't include any side effects.
570 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
571 if (I->mayHaveSideEffects())
577 /// Return true if the specified block unconditionally leads to an exit from
578 /// the specified loop, and has no side-effects in the process. If so, return
579 /// the block that is exited to, otherwise return null.
580 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
581 std::set<BasicBlock*> Visited;
582 Visited.insert(L->getHeader()); // Branches to header make infinite loops.
583 BasicBlock *ExitBB = nullptr;
584 if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
589 /// We have found that we can unswitch currentLoop when LoopCond == Val to
590 /// simplify the loop. If we decide that this is profitable,
591 /// unswitch the loop, reprocess the pieces, then return true.
592 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val,
593 TerminatorInst *TI) {
594 // Check to see if it would be profitable to unswitch current loop.
595 if (!BranchesInfo.CostAllowsUnswitching()) {
596 DEBUG(dbgs() << "NOT unswitching loop %"
597 << currentLoop->getHeader()->getName()
598 << " at non-trivial condition '" << *Val
599 << "' == " << *LoopCond << "\n"
600 << ". Cost too high.\n");
604 UnswitchNontrivialCondition(LoopCond, Val, currentLoop, TI);
608 /// Recursively clone the specified loop and all of its children,
609 /// mapping the blocks with the specified map.
610 static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM,
611 LoopInfo *LI, LPPassManager *LPM) {
612 Loop *New = new Loop();
613 LPM->insertLoop(New, PL);
615 // Add all of the blocks in L to the new loop.
616 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
618 if (LI->getLoopFor(*I) == L)
619 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI);
621 // Add all of the subloops to the new loop.
622 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
623 CloneLoop(*I, New, VM, LI, LPM);
628 static void copyMetadata(Instruction *DstInst, const Instruction *SrcInst,
630 if (!SrcInst || !SrcInst->hasMetadata())
633 SmallVector<std::pair<unsigned, MDNode *>, 4> MDs;
634 SrcInst->getAllMetadata(MDs);
635 for (auto &MD : MDs) {
639 case LLVMContext::MD_prof:
640 if (Swapped && MD.second->getNumOperands() == 3 &&
641 isa<MDString>(MD.second->getOperand(0))) {
642 MDString *MDName = cast<MDString>(MD.second->getOperand(0));
643 if (MDName->getString() == "branch_weights") {
644 auto *ValT = cast_or_null<ConstantAsMetadata>(
645 MD.second->getOperand(1))->getValue();
646 auto *ValF = cast_or_null<ConstantAsMetadata>(
647 MD.second->getOperand(2))->getValue();
648 assert(ValT && ValF && "Invalid Operands of branch_weights");
650 MDBuilder(DstInst->getParent()->getContext())
651 .createBranchWeights(cast<ConstantInt>(ValF)->getZExtValue(),
652 cast<ConstantInt>(ValT)->getZExtValue());
657 case LLVMContext::MD_make_implicit:
658 case LLVMContext::MD_dbg:
659 DstInst->setMetadata(MD.first, MD.second);
664 /// Emit a conditional branch on two values if LIC == Val, branch to TrueDst,
665 /// otherwise branch to FalseDest. Insert the code immediately before InsertPt.
666 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
667 BasicBlock *TrueDest,
668 BasicBlock *FalseDest,
669 Instruction *InsertPt,
670 TerminatorInst *TI) {
671 // Insert a conditional branch on LIC to the two preheaders. The original
672 // code is the true version and the new code is the false version.
673 Value *BranchVal = LIC;
674 bool Swapped = false;
675 if (!isa<ConstantInt>(Val) ||
676 Val->getType() != Type::getInt1Ty(LIC->getContext()))
677 BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val);
678 else if (Val != ConstantInt::getTrue(Val->getContext())) {
679 // We want to enter the new loop when the condition is true.
680 std::swap(TrueDest, FalseDest);
684 // Insert the new branch.
685 BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
686 copyMetadata(BI, TI, Swapped);
688 // If either edge is critical, split it. This helps preserve LoopSimplify
689 // form for enclosing loops.
690 auto Options = CriticalEdgeSplittingOptions(DT, LI).setPreserveLCSSA();
691 SplitCriticalEdge(BI, 0, Options);
692 SplitCriticalEdge(BI, 1, Options);
695 /// Given a loop that has a trivial unswitchable condition in it (a cond branch
696 /// from its header block to its latch block, where the path through the loop
697 /// that doesn't execute its body has no side-effects), unswitch it. This
698 /// doesn't involve any code duplication, just moving the conditional branch
699 /// outside of the loop and updating loop info.
700 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
701 BasicBlock *ExitBlock,
702 TerminatorInst *TI) {
703 DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %"
704 << loopHeader->getName() << " [" << L->getBlocks().size()
705 << " blocks] in Function "
706 << L->getHeader()->getParent()->getName() << " on cond: " << *Val
707 << " == " << *Cond << "\n");
709 // First step, split the preheader, so that we know that there is a safe place
710 // to insert the conditional branch. We will change loopPreheader to have a
711 // conditional branch on Cond.
712 BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, DT, LI);
714 // Now that we have a place to insert the conditional branch, create a place
715 // to branch to: this is the exit block out of the loop that we should
718 // Split this block now, so that the loop maintains its exit block, and so
719 // that the jump from the preheader can execute the contents of the exit block
720 // without actually branching to it (the exit block should be dominated by the
721 // loop header, not the preheader).
722 assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
723 BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), DT, LI);
725 // Okay, now we have a position to branch from and a position to branch to,
726 // insert the new conditional branch.
727 EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
728 loopPreheader->getTerminator(), TI);
729 LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
730 loopPreheader->getTerminator()->eraseFromParent();
732 // We need to reprocess this loop, it could be unswitched again.
735 // Now that we know that the loop is never entered when this condition is a
736 // particular value, rewrite the loop with this info. We know that this will
737 // at least eliminate the old branch.
738 RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
742 /// Check if the first non-constant condition starting from the loop header is
743 /// a trivial unswitch condition: that is, a condition controls whether or not
744 /// the loop does anything at all. If it is a trivial condition, unswitching
745 /// produces no code duplications (equivalently, it produces a simpler loop and
746 /// a new empty loop, which gets deleted). Therefore always unswitch trivial
748 bool LoopUnswitch::TryTrivialLoopUnswitch(bool &Changed) {
749 BasicBlock *CurrentBB = currentLoop->getHeader();
750 TerminatorInst *CurrentTerm = CurrentBB->getTerminator();
751 LLVMContext &Context = CurrentBB->getContext();
753 // If loop header has only one reachable successor (currently via an
754 // unconditional branch or constant foldable conditional branch, but
755 // should also consider adding constant foldable switch instruction in
756 // future), we should keep looking for trivial condition candidates in
757 // the successor as well. An alternative is to constant fold conditions
758 // and merge successors into loop header (then we only need to check header's
759 // terminator). The reason for not doing this in LoopUnswitch pass is that
760 // it could potentially break LoopPassManager's invariants. Folding dead
761 // branches could either eliminate the current loop or make other loops
762 // unreachable. LCSSA form might also not be preserved after deleting
763 // branches. The following code keeps traversing loop header's successors
764 // until it finds the trivial condition candidate (condition that is not a
765 // constant). Since unswitching generates branches with constant conditions,
766 // this scenario could be very common in practice.
767 SmallSet<BasicBlock*, 8> Visited;
770 // If we exit loop or reach a previous visited block, then
771 // we can not reach any trivial condition candidates (unfoldable
772 // branch instructions or switch instructions) and no unswitch
773 // can happen. Exit and return false.
774 if (!currentLoop->contains(CurrentBB) || !Visited.insert(CurrentBB).second)
777 // Check if this loop will execute any side-effecting instructions (e.g.
778 // stores, calls, volatile loads) in the part of the loop that the code
779 // *would* execute. Check the header first.
780 for (BasicBlock::iterator I : *CurrentBB)
781 if (I->mayHaveSideEffects())
784 // FIXME: add check for constant foldable switch instructions.
785 if (BranchInst *BI = dyn_cast<BranchInst>(CurrentTerm)) {
786 if (BI->isUnconditional()) {
787 CurrentBB = BI->getSuccessor(0);
788 } else if (BI->getCondition() == ConstantInt::getTrue(Context)) {
789 CurrentBB = BI->getSuccessor(0);
790 } else if (BI->getCondition() == ConstantInt::getFalse(Context)) {
791 CurrentBB = BI->getSuccessor(1);
793 // Found a trivial condition candidate: non-foldable conditional branch.
800 CurrentTerm = CurrentBB->getTerminator();
803 // CondVal is the condition that controls the trivial condition.
804 // LoopExitBB is the BasicBlock that loop exits when meets trivial condition.
805 Constant *CondVal = nullptr;
806 BasicBlock *LoopExitBB = nullptr;
808 if (BranchInst *BI = dyn_cast<BranchInst>(CurrentTerm)) {
809 // If this isn't branching on an invariant condition, we can't unswitch it.
810 if (!BI->isConditional())
813 Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
814 currentLoop, Changed);
816 // Unswitch only if the trivial condition itself is an LIV (not
817 // partial LIV which could occur in and/or)
818 if (!LoopCond || LoopCond != BI->getCondition())
821 // Check to see if a successor of the branch is guaranteed to
822 // exit through a unique exit block without having any
823 // side-effects. If so, determine the value of Cond that causes
825 if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
826 BI->getSuccessor(0)))) {
827 CondVal = ConstantInt::getTrue(Context);
828 } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
829 BI->getSuccessor(1)))) {
830 CondVal = ConstantInt::getFalse(Context);
833 // If we didn't find a single unique LoopExit block, or if the loop exit
834 // block contains phi nodes, this isn't trivial.
835 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
836 return false; // Can't handle this.
838 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, LoopExitBB,
842 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(CurrentTerm)) {
843 // If this isn't switching on an invariant condition, we can't unswitch it.
844 Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
845 currentLoop, Changed);
847 // Unswitch only if the trivial condition itself is an LIV (not
848 // partial LIV which could occur in and/or)
849 if (!LoopCond || LoopCond != SI->getCondition())
852 // Check to see if a successor of the switch is guaranteed to go to the
853 // latch block or exit through a one exit block without having any
854 // side-effects. If so, determine the value of Cond that causes it to do
856 // Note that we can't trivially unswitch on the default case or
857 // on already unswitched cases.
858 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
860 BasicBlock *LoopExitCandidate;
861 if ((LoopExitCandidate = isTrivialLoopExitBlock(currentLoop,
862 i.getCaseSuccessor()))) {
863 // Okay, we found a trivial case, remember the value that is trivial.
864 ConstantInt *CaseVal = i.getCaseValue();
866 // Check that it was not unswitched before, since already unswitched
867 // trivial vals are looks trivial too.
868 if (BranchesInfo.isUnswitched(SI, CaseVal))
870 LoopExitBB = LoopExitCandidate;
876 // If we didn't find a single unique LoopExit block, or if the loop exit
877 // block contains phi nodes, this isn't trivial.
878 if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
879 return false; // Can't handle this.
881 UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, LoopExitBB,
889 /// Split all of the edges from inside the loop to their exit blocks.
890 /// Update the appropriate Phi nodes as we do so.
891 void LoopUnswitch::SplitExitEdges(Loop *L,
892 const SmallVectorImpl<BasicBlock *> &ExitBlocks){
894 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
895 BasicBlock *ExitBlock = ExitBlocks[i];
896 SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock),
897 pred_end(ExitBlock));
899 // Although SplitBlockPredecessors doesn't preserve loop-simplify in
900 // general, if we call it on all predecessors of all exits then it does.
901 SplitBlockPredecessors(ExitBlock, Preds, ".us-lcssa", DT, LI,
902 /*PreserveLCSSA*/ true);
906 /// We determined that the loop is profitable to unswitch when LIC equal Val.
907 /// Split it into loop versions and test the condition outside of either loop.
908 /// Return the loops created as Out1/Out2.
909 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
910 Loop *L, TerminatorInst *TI) {
911 Function *F = loopHeader->getParent();
912 DEBUG(dbgs() << "loop-unswitch: Unswitching loop %"
913 << loopHeader->getName() << " [" << L->getBlocks().size()
914 << " blocks] in Function " << F->getName()
915 << " when '" << *Val << "' == " << *LIC << "\n");
917 if (auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>())
918 SEWP->getSE().forgetLoop(L);
923 // First step, split the preheader and exit blocks, and add these blocks to
924 // the LoopBlocks list.
925 BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, DT, LI);
926 LoopBlocks.push_back(NewPreheader);
928 // We want the loop to come after the preheader, but before the exit blocks.
929 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
931 SmallVector<BasicBlock*, 8> ExitBlocks;
932 L->getUniqueExitBlocks(ExitBlocks);
934 // Split all of the edges from inside the loop to their exit blocks. Update
935 // the appropriate Phi nodes as we do so.
936 SplitExitEdges(L, ExitBlocks);
938 // The exit blocks may have been changed due to edge splitting, recompute.
940 L->getUniqueExitBlocks(ExitBlocks);
942 // Add exit blocks to the loop blocks.
943 LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
945 // Next step, clone all of the basic blocks that make up the loop (including
946 // the loop preheader and exit blocks), keeping track of the mapping between
947 // the instructions and blocks.
948 NewBlocks.reserve(LoopBlocks.size());
949 ValueToValueMapTy VMap;
950 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
951 BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F);
953 NewBlocks.push_back(NewBB);
954 VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping.
955 LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L);
958 // Splice the newly inserted blocks into the function right before the
959 // original preheader.
960 F->getBasicBlockList().splice(NewPreheader, F->getBasicBlockList(),
961 NewBlocks[0], F->end());
963 // FIXME: We could register any cloned assumptions instead of clearing the
964 // whole function's cache.
967 // Now we create the new Loop object for the versioned loop.
968 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM);
970 // Recalculate unswitching quota, inherit simplified switches info for NewBB,
971 // Probably clone more loop-unswitch related loop properties.
972 BranchesInfo.cloneData(NewLoop, L, VMap);
974 Loop *ParentLoop = L->getParentLoop();
976 // Make sure to add the cloned preheader and exit blocks to the parent loop
978 ParentLoop->addBasicBlockToLoop(NewBlocks[0], *LI);
981 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
982 BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]);
983 // The new exit block should be in the same loop as the old one.
984 if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
985 ExitBBLoop->addBasicBlockToLoop(NewExit, *LI);
987 assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
988 "Exit block should have been split to have one successor!");
989 BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
991 // If the successor of the exit block had PHI nodes, add an entry for
993 for (BasicBlock::iterator I = ExitSucc->begin();
994 PHINode *PN = dyn_cast<PHINode>(I); ++I) {
995 Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
996 ValueToValueMapTy::iterator It = VMap.find(V);
997 if (It != VMap.end()) V = It->second;
998 PN->addIncoming(V, NewExit);
1001 if (LandingPadInst *LPad = NewExit->getLandingPadInst()) {
1002 PHINode *PN = PHINode::Create(LPad->getType(), 0, "",
1003 ExitSucc->getFirstInsertionPt());
1005 for (pred_iterator I = pred_begin(ExitSucc), E = pred_end(ExitSucc);
1007 BasicBlock *BB = *I;
1008 LandingPadInst *LPI = BB->getLandingPadInst();
1009 LPI->replaceAllUsesWith(PN);
1010 PN->addIncoming(LPI, BB);
1015 // Rewrite the code to refer to itself.
1016 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
1017 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
1018 E = NewBlocks[i]->end(); I != E; ++I)
1019 RemapInstruction(I, VMap,RF_NoModuleLevelChanges|RF_IgnoreMissingEntries);
1021 // Rewrite the original preheader to select between versions of the loop.
1022 BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
1023 assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
1024 "Preheader splitting did not work correctly!");
1026 // Emit the new branch that selects between the two versions of this loop.
1027 EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR,
1029 LPM->deleteSimpleAnalysisValue(OldBR, L);
1030 OldBR->eraseFromParent();
1032 LoopProcessWorklist.push_back(NewLoop);
1035 // Keep a WeakVH holding onto LIC. If the first call to RewriteLoopBody
1036 // deletes the instruction (for example by simplifying a PHI that feeds into
1037 // the condition that we're unswitching on), we don't rewrite the second
1039 WeakVH LICHandle(LIC);
1041 // Now we rewrite the original code to know that the condition is true and the
1042 // new code to know that the condition is false.
1043 RewriteLoopBodyWithConditionConstant(L, LIC, Val, false);
1045 // It's possible that simplifying one loop could cause the other to be
1046 // changed to another value or a constant. If its a constant, don't simplify
1048 if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop &&
1049 LICHandle && !isa<Constant>(LICHandle))
1050 RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true);
1053 /// Remove all instances of I from the worklist vector specified.
1054 static void RemoveFromWorklist(Instruction *I,
1055 std::vector<Instruction*> &Worklist) {
1057 Worklist.erase(std::remove(Worklist.begin(), Worklist.end(), I),
1061 /// When we find that I really equals V, remove I from the
1062 /// program, replacing all uses with V and update the worklist.
1063 static void ReplaceUsesOfWith(Instruction *I, Value *V,
1064 std::vector<Instruction*> &Worklist,
1065 Loop *L, LPPassManager *LPM) {
1066 DEBUG(dbgs() << "Replace with '" << *V << "': " << *I);
1068 // Add uses to the worklist, which may be dead now.
1069 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1070 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1071 Worklist.push_back(Use);
1073 // Add users to the worklist which may be simplified now.
1074 for (User *U : I->users())
1075 Worklist.push_back(cast<Instruction>(U));
1076 LPM->deleteSimpleAnalysisValue(I, L);
1077 RemoveFromWorklist(I, Worklist);
1078 I->replaceAllUsesWith(V);
1079 I->eraseFromParent();
1083 /// We know either that the value LIC has the value specified by Val in the
1084 /// specified loop, or we know it does NOT have that value.
1085 /// Rewrite any uses of LIC or of properties correlated to it.
1086 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
1089 assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
1091 // FIXME: Support correlated properties, like:
1098 // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches,
1099 // selects, switches.
1100 std::vector<Instruction*> Worklist;
1101 LLVMContext &Context = Val->getContext();
1103 // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
1104 // in the loop with the appropriate one directly.
1105 if (IsEqual || (isa<ConstantInt>(Val) &&
1106 Val->getType()->isIntegerTy(1))) {
1111 Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
1112 !cast<ConstantInt>(Val)->getZExtValue());
1114 for (User *U : LIC->users()) {
1115 Instruction *UI = dyn_cast<Instruction>(U);
1116 if (!UI || !L->contains(UI))
1118 Worklist.push_back(UI);
1121 for (std::vector<Instruction*>::iterator UI = Worklist.begin(),
1122 UE = Worklist.end(); UI != UE; ++UI)
1123 (*UI)->replaceUsesOfWith(LIC, Replacement);
1125 SimplifyCode(Worklist, L);
1129 // Otherwise, we don't know the precise value of LIC, but we do know that it
1130 // is certainly NOT "Val". As such, simplify any uses in the loop that we
1131 // can. This case occurs when we unswitch switch statements.
1132 for (User *U : LIC->users()) {
1133 Instruction *UI = dyn_cast<Instruction>(U);
1134 if (!UI || !L->contains(UI))
1137 Worklist.push_back(UI);
1139 // TODO: We could do other simplifications, for example, turning
1140 // 'icmp eq LIC, Val' -> false.
1142 // If we know that LIC is not Val, use this info to simplify code.
1143 SwitchInst *SI = dyn_cast<SwitchInst>(UI);
1144 if (!SI || !isa<ConstantInt>(Val)) continue;
1146 SwitchInst::CaseIt DeadCase = SI->findCaseValue(cast<ConstantInt>(Val));
1147 // Default case is live for multiple values.
1148 if (DeadCase == SI->case_default()) continue;
1150 // Found a dead case value. Don't remove PHI nodes in the
1151 // successor if they become single-entry, those PHI nodes may
1152 // be in the Users list.
1154 BasicBlock *Switch = SI->getParent();
1155 BasicBlock *SISucc = DeadCase.getCaseSuccessor();
1156 BasicBlock *Latch = L->getLoopLatch();
1158 BranchesInfo.setUnswitched(SI, Val);
1160 if (!SI->findCaseDest(SISucc)) continue; // Edge is critical.
1161 // If the DeadCase successor dominates the loop latch, then the
1162 // transformation isn't safe since it will delete the sole predecessor edge
1164 if (Latch && DT->dominates(SISucc, Latch))
1167 // FIXME: This is a hack. We need to keep the successor around
1168 // and hooked up so as to preserve the loop structure, because
1169 // trying to update it is complicated. So instead we preserve the
1170 // loop structure and put the block on a dead code path.
1171 SplitEdge(Switch, SISucc, DT, LI);
1172 // Compute the successors instead of relying on the return value
1173 // of SplitEdge, since it may have split the switch successor
1175 BasicBlock *NewSISucc = DeadCase.getCaseSuccessor();
1176 BasicBlock *OldSISucc = *succ_begin(NewSISucc);
1177 // Create an "unreachable" destination.
1178 BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable",
1179 Switch->getParent(),
1181 new UnreachableInst(Context, Abort);
1182 // Force the new case destination to branch to the "unreachable"
1183 // block while maintaining a (dead) CFG edge to the old block.
1184 NewSISucc->getTerminator()->eraseFromParent();
1185 BranchInst::Create(Abort, OldSISucc,
1186 ConstantInt::getTrue(Context), NewSISucc);
1187 // Release the PHI operands for this edge.
1188 for (BasicBlock::iterator II = NewSISucc->begin();
1189 PHINode *PN = dyn_cast<PHINode>(II); ++II)
1190 PN->setIncomingValue(PN->getBasicBlockIndex(Switch),
1191 UndefValue::get(PN->getType()));
1192 // Tell the domtree about the new block. We don't fully update the
1193 // domtree here -- instead we force it to do a full recomputation
1194 // after the pass is complete -- but we do need to inform it of
1196 DT->addNewBlock(Abort, NewSISucc);
1199 SimplifyCode(Worklist, L);
1202 /// Now that we have simplified some instructions in the loop, walk over it and
1203 /// constant prop, dce, and fold control flow where possible. Note that this is
1204 /// effectively a very simple loop-structure-aware optimizer. During processing
1205 /// of this loop, L could very well be deleted, so it must not be used.
1207 /// FIXME: When the loop optimizer is more mature, separate this out to a new
1210 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
1211 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
1212 while (!Worklist.empty()) {
1213 Instruction *I = Worklist.back();
1214 Worklist.pop_back();
1217 if (isInstructionTriviallyDead(I)) {
1218 DEBUG(dbgs() << "Remove dead instruction '" << *I);
1220 // Add uses to the worklist, which may be dead now.
1221 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
1222 if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
1223 Worklist.push_back(Use);
1224 LPM->deleteSimpleAnalysisValue(I, L);
1225 RemoveFromWorklist(I, Worklist);
1226 I->eraseFromParent();
1231 // See if instruction simplification can hack this up. This is common for
1232 // things like "select false, X, Y" after unswitching made the condition be
1233 // 'false'. TODO: update the domtree properly so we can pass it here.
1234 if (Value *V = SimplifyInstruction(I, DL))
1235 if (LI->replacementPreservesLCSSAForm(I, V)) {
1236 ReplaceUsesOfWith(I, V, Worklist, L, LPM);
1240 // Special case hacks that appear commonly in unswitched code.
1241 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1242 if (BI->isUnconditional()) {
1243 // If BI's parent is the only pred of the successor, fold the two blocks
1245 BasicBlock *Pred = BI->getParent();
1246 BasicBlock *Succ = BI->getSuccessor(0);
1247 BasicBlock *SinglePred = Succ->getSinglePredecessor();
1248 if (!SinglePred) continue; // Nothing to do.
1249 assert(SinglePred == Pred && "CFG broken");
1251 DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- "
1252 << Succ->getName() << "\n");
1254 // Resolve any single entry PHI nodes in Succ.
1255 while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
1256 ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
1258 // If Succ has any successors with PHI nodes, update them to have
1259 // entries coming from Pred instead of Succ.
1260 Succ->replaceAllUsesWith(Pred);
1262 // Move all of the successor contents from Succ to Pred.
1263 Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
1265 LPM->deleteSimpleAnalysisValue(BI, L);
1266 BI->eraseFromParent();
1267 RemoveFromWorklist(BI, Worklist);
1269 // Remove Succ from the loop tree.
1270 LI->removeBlock(Succ);
1271 LPM->deleteSimpleAnalysisValue(Succ, L);
1272 Succ->eraseFromParent();