1 //===- JumpThreading.cpp - Thread control through conditional blocks ------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the Jump Threading pass.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/Scalar.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/DenseSet.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/GlobalsModRef.h"
22 #include "llvm/Analysis/CFG.h"
23 #include "llvm/Analysis/BlockFrequencyInfo.h"
24 #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
25 #include "llvm/Analysis/BranchProbabilityInfo.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/Analysis/InstructionSimplify.h"
28 #include "llvm/Analysis/LazyValueInfo.h"
29 #include "llvm/Analysis/Loads.h"
30 #include "llvm/Analysis/LoopInfo.h"
31 #include "llvm/Analysis/TargetLibraryInfo.h"
32 #include "llvm/IR/DataLayout.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/LLVMContext.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/Metadata.h"
37 #include "llvm/IR/ValueHandle.h"
38 #include "llvm/Pass.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
43 #include "llvm/Transforms/Utils/Local.h"
44 #include "llvm/Transforms/Utils/SSAUpdater.h"
49 #define DEBUG_TYPE "jump-threading"
51 STATISTIC(NumThreads, "Number of jumps threaded");
52 STATISTIC(NumFolds, "Number of terminators folded");
53 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
55 static cl::opt<unsigned>
56 BBDuplicateThreshold("jump-threading-threshold",
57 cl::desc("Max block size to duplicate for jump threading"),
58 cl::init(6), cl::Hidden);
61 // These are at global scope so static functions can use them too.
62 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo;
63 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy;
65 // This is used to keep track of what kind of constant we're currently hoping
67 enum ConstantPreference {
72 /// This pass performs 'jump threading', which looks at blocks that have
73 /// multiple predecessors and multiple successors. If one or more of the
74 /// predecessors of the block can be proven to always jump to one of the
75 /// successors, we forward the edge from the predecessor to the successor by
76 /// duplicating the contents of this block.
78 /// An example of when this can occur is code like this:
85 /// In this case, the unconditional branch at the end of the first if can be
86 /// revectored to the false side of the second if.
88 class JumpThreading : public FunctionPass {
89 TargetLibraryInfo *TLI;
91 std::unique_ptr<BlockFrequencyInfo> BFI;
92 std::unique_ptr<BranchProbabilityInfo> BPI;
95 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
97 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
99 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet;
101 unsigned BBDupThreshold;
103 // RAII helper for updating the recursion stack.
104 struct RecursionSetRemover {
105 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet;
106 std::pair<Value*, BasicBlock*> ThePair;
108 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S,
109 std::pair<Value*, BasicBlock*> P)
110 : TheSet(S), ThePair(P) { }
112 ~RecursionSetRemover() {
113 TheSet.erase(ThePair);
117 static char ID; // Pass identification
118 JumpThreading(int T = -1) : FunctionPass(ID) {
119 BBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
120 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
123 bool runOnFunction(Function &F) override;
125 void getAnalysisUsage(AnalysisUsage &AU) const override {
126 AU.addRequired<LazyValueInfo>();
127 AU.addPreserved<LazyValueInfo>();
128 AU.addPreserved<GlobalsAAWrapperPass>();
129 AU.addRequired<TargetLibraryInfoWrapperPass>();
132 void releaseMemory() override {
137 void FindLoopHeaders(Function &F);
138 bool ProcessBlock(BasicBlock *BB);
139 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
141 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
142 const SmallVectorImpl<BasicBlock *> &PredBBs);
144 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
145 PredValueInfo &Result,
146 ConstantPreference Preference,
147 Instruction *CxtI = nullptr);
148 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
149 ConstantPreference Preference,
150 Instruction *CxtI = nullptr);
152 bool ProcessBranchOnPHI(PHINode *PN);
153 bool ProcessBranchOnXOR(BinaryOperator *BO);
155 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
156 bool TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB);
159 BasicBlock *SplitBlockPreds(BasicBlock *BB, ArrayRef<BasicBlock *> Preds,
161 void UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB, BasicBlock *BB,
162 BasicBlock *NewBB, BasicBlock *SuccBB);
166 char JumpThreading::ID = 0;
167 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
168 "Jump Threading", false, false)
169 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)
170 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
171 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
172 "Jump Threading", false, false)
174 // Public interface to the Jump Threading pass
175 FunctionPass *llvm::createJumpThreadingPass(int Threshold) { return new JumpThreading(Threshold); }
177 /// runOnFunction - Top level algorithm.
179 bool JumpThreading::runOnFunction(Function &F) {
180 if (skipOptnoneFunction(F))
183 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
184 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
185 LVI = &getAnalysis<LazyValueInfo>();
188 // When profile data is available, we need to update edge weights after
189 // successful jump threading, which requires both BPI and BFI being available.
190 HasProfileData = F.getEntryCount().hasValue();
191 if (HasProfileData) {
192 LoopInfo LI{DominatorTree(F)};
193 BPI.reset(new BranchProbabilityInfo(F, LI));
194 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
197 // Remove unreachable blocks from function as they may result in infinite
198 // loop. We do threading if we found something profitable. Jump threading a
199 // branch can create other opportunities. If these opportunities form a cycle
200 // i.e. if any jump threading is undoing previous threading in the path, then
201 // we will loop forever. We take care of this issue by not jump threading for
202 // back edges. This works for normal cases but not for unreachable blocks as
203 // they may have cycle with no back edge.
204 removeUnreachableBlocks(F);
208 bool Changed, EverChanged = false;
211 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
212 BasicBlock *BB = &*I;
213 // Thread all of the branches we can over this block.
214 while (ProcessBlock(BB))
219 // If the block is trivially dead, zap it. This eliminates the successor
220 // edges which simplifies the CFG.
221 if (pred_empty(BB) &&
222 BB != &BB->getParent()->getEntryBlock()) {
223 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
224 << "' with terminator: " << *BB->getTerminator() << '\n');
225 LoopHeaders.erase(BB);
232 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
234 // Can't thread an unconditional jump, but if the block is "almost
235 // empty", we can replace uses of it with uses of the successor and make
237 if (BI && BI->isUnconditional() &&
238 BB != &BB->getParent()->getEntryBlock() &&
239 // If the terminator is the only non-phi instruction, try to nuke it.
240 BB->getFirstNonPHIOrDbg()->isTerminator()) {
241 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
242 // block, we have to make sure it isn't in the LoopHeaders set. We
243 // reinsert afterward if needed.
244 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
245 BasicBlock *Succ = BI->getSuccessor(0);
247 // FIXME: It is always conservatively correct to drop the info
248 // for a block even if it doesn't get erased. This isn't totally
249 // awesome, but it allows us to use AssertingVH to prevent nasty
250 // dangling pointer issues within LazyValueInfo.
252 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
254 // If we deleted BB and BB was the header of a loop, then the
255 // successor is now the header of the loop.
259 if (ErasedFromLoopHeaders)
260 LoopHeaders.insert(BB);
263 EverChanged |= Changed;
270 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
271 /// thread across it. Stop scanning the block when passing the threshold.
272 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
273 unsigned Threshold) {
274 /// Ignore PHI nodes, these will be flattened when duplication happens.
275 BasicBlock::const_iterator I(BB->getFirstNonPHI());
277 // FIXME: THREADING will delete values that are just used to compute the
278 // branch, so they shouldn't count against the duplication cost.
280 // Sum up the cost of each instruction until we get to the terminator. Don't
281 // include the terminator because the copy won't include it.
283 for (; !isa<TerminatorInst>(I); ++I) {
285 // Stop scanning the block if we've reached the threshold.
286 if (Size > Threshold)
289 // Debugger intrinsics don't incur code size.
290 if (isa<DbgInfoIntrinsic>(I)) continue;
292 // If this is a pointer->pointer bitcast, it is free.
293 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
296 // Bail out if this instruction gives back a token type, it is not possible
297 // to duplicate it if it is used outside this BB.
298 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
301 // All other instructions count for at least one unit.
304 // Calls are more expensive. If they are non-intrinsic calls, we model them
305 // as having cost of 4. If they are a non-vector intrinsic, we model them
306 // as having cost of 2 total, and if they are a vector intrinsic, we model
307 // them as having cost 1.
308 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
309 if (CI->cannotDuplicate() || CI->isConvergent())
310 // Blocks with NoDuplicate are modelled as having infinite cost, so they
311 // are never duplicated.
313 else if (!isa<IntrinsicInst>(CI))
315 else if (!CI->getType()->isVectorTy())
320 // Threading through a switch statement is particularly profitable. If this
321 // block ends in a switch, decrease its cost to make it more likely to happen.
322 if (isa<SwitchInst>(I))
323 Size = Size > 6 ? Size-6 : 0;
325 // The same holds for indirect branches, but slightly more so.
326 if (isa<IndirectBrInst>(I))
327 Size = Size > 8 ? Size-8 : 0;
332 /// FindLoopHeaders - We do not want jump threading to turn proper loop
333 /// structures into irreducible loops. Doing this breaks up the loop nesting
334 /// hierarchy and pessimizes later transformations. To prevent this from
335 /// happening, we first have to find the loop headers. Here we approximate this
336 /// by finding targets of backedges in the CFG.
338 /// Note that there definitely are cases when we want to allow threading of
339 /// edges across a loop header. For example, threading a jump from outside the
340 /// loop (the preheader) to an exit block of the loop is definitely profitable.
341 /// It is also almost always profitable to thread backedges from within the loop
342 /// to exit blocks, and is often profitable to thread backedges to other blocks
343 /// within the loop (forming a nested loop). This simple analysis is not rich
344 /// enough to track all of these properties and keep it up-to-date as the CFG
345 /// mutates, so we don't allow any of these transformations.
347 void JumpThreading::FindLoopHeaders(Function &F) {
348 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
349 FindFunctionBackedges(F, Edges);
351 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
352 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
355 /// getKnownConstant - Helper method to determine if we can thread over a
356 /// terminator with the given value as its condition, and if so what value to
357 /// use for that. What kind of value this is depends on whether we want an
358 /// integer or a block address, but an undef is always accepted.
359 /// Returns null if Val is null or not an appropriate constant.
360 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
364 // Undef is "known" enough.
365 if (UndefValue *U = dyn_cast<UndefValue>(Val))
368 if (Preference == WantBlockAddress)
369 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
371 return dyn_cast<ConstantInt>(Val);
374 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
375 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
376 /// in any of our predecessors. If so, return the known list of value and pred
377 /// BB in the result vector.
379 /// This returns true if there were any known values.
382 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
383 ConstantPreference Preference,
385 // This method walks up use-def chains recursively. Because of this, we could
386 // get into an infinite loop going around loops in the use-def chain. To
387 // prevent this, keep track of what (value, block) pairs we've already visited
388 // and terminate the search if we loop back to them
389 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
392 // An RAII help to remove this pair from the recursion set once the recursion
393 // stack pops back out again.
394 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
396 // If V is a constant, then it is known in all predecessors.
397 if (Constant *KC = getKnownConstant(V, Preference)) {
398 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
399 Result.push_back(std::make_pair(KC, *PI));
404 // If V is a non-instruction value, or an instruction in a different block,
405 // then it can't be derived from a PHI.
406 Instruction *I = dyn_cast<Instruction>(V);
407 if (!I || I->getParent() != BB) {
409 // Okay, if this is a live-in value, see if it has a known value at the end
410 // of any of our predecessors.
412 // FIXME: This should be an edge property, not a block end property.
413 /// TODO: Per PR2563, we could infer value range information about a
414 /// predecessor based on its terminator.
416 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
417 // "I" is a non-local compare-with-a-constant instruction. This would be
418 // able to handle value inequalities better, for example if the compare is
419 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
420 // Perhaps getConstantOnEdge should be smart enough to do this?
422 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
424 // If the value is known by LazyValueInfo to be a constant in a
425 // predecessor, use that information to try to thread this block.
426 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
427 if (Constant *KC = getKnownConstant(PredCst, Preference))
428 Result.push_back(std::make_pair(KC, P));
431 return !Result.empty();
434 /// If I is a PHI node, then we know the incoming values for any constants.
435 if (PHINode *PN = dyn_cast<PHINode>(I)) {
436 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
437 Value *InVal = PN->getIncomingValue(i);
438 if (Constant *KC = getKnownConstant(InVal, Preference)) {
439 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
441 Constant *CI = LVI->getConstantOnEdge(InVal,
442 PN->getIncomingBlock(i),
444 if (Constant *KC = getKnownConstant(CI, Preference))
445 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
449 return !Result.empty();
452 PredValueInfoTy LHSVals, RHSVals;
454 // Handle some boolean conditions.
455 if (I->getType()->getPrimitiveSizeInBits() == 1) {
456 assert(Preference == WantInteger && "One-bit non-integer type?");
458 // X & false -> false
459 if (I->getOpcode() == Instruction::Or ||
460 I->getOpcode() == Instruction::And) {
461 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
463 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
466 if (LHSVals.empty() && RHSVals.empty())
469 ConstantInt *InterestingVal;
470 if (I->getOpcode() == Instruction::Or)
471 InterestingVal = ConstantInt::getTrue(I->getContext());
473 InterestingVal = ConstantInt::getFalse(I->getContext());
475 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
477 // Scan for the sentinel. If we find an undef, force it to the
478 // interesting value: x|undef -> true and x&undef -> false.
479 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
480 if (LHSVals[i].first == InterestingVal ||
481 isa<UndefValue>(LHSVals[i].first)) {
482 Result.push_back(LHSVals[i]);
483 Result.back().first = InterestingVal;
484 LHSKnownBBs.insert(LHSVals[i].second);
486 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
487 if (RHSVals[i].first == InterestingVal ||
488 isa<UndefValue>(RHSVals[i].first)) {
489 // If we already inferred a value for this block on the LHS, don't
491 if (!LHSKnownBBs.count(RHSVals[i].second)) {
492 Result.push_back(RHSVals[i]);
493 Result.back().first = InterestingVal;
497 return !Result.empty();
500 // Handle the NOT form of XOR.
501 if (I->getOpcode() == Instruction::Xor &&
502 isa<ConstantInt>(I->getOperand(1)) &&
503 cast<ConstantInt>(I->getOperand(1))->isOne()) {
504 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
509 // Invert the known values.
510 for (unsigned i = 0, e = Result.size(); i != e; ++i)
511 Result[i].first = ConstantExpr::getNot(Result[i].first);
516 // Try to simplify some other binary operator values.
517 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
518 assert(Preference != WantBlockAddress
519 && "A binary operator creating a block address?");
520 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
521 PredValueInfoTy LHSVals;
522 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
525 // Try to use constant folding to simplify the binary operator.
526 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
527 Constant *V = LHSVals[i].first;
528 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
530 if (Constant *KC = getKnownConstant(Folded, WantInteger))
531 Result.push_back(std::make_pair(KC, LHSVals[i].second));
535 return !Result.empty();
538 // Handle compare with phi operand, where the PHI is defined in this block.
539 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
540 assert(Preference == WantInteger && "Compares only produce integers");
541 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
542 if (PN && PN->getParent() == BB) {
543 const DataLayout &DL = PN->getModule()->getDataLayout();
544 // We can do this simplification if any comparisons fold to true or false.
546 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
547 BasicBlock *PredBB = PN->getIncomingBlock(i);
548 Value *LHS = PN->getIncomingValue(i);
549 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
551 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL);
553 if (!isa<Constant>(RHS))
556 LazyValueInfo::Tristate
557 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
558 cast<Constant>(RHS), PredBB, BB,
560 if (ResT == LazyValueInfo::Unknown)
562 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
565 if (Constant *KC = getKnownConstant(Res, WantInteger))
566 Result.push_back(std::make_pair(KC, PredBB));
569 return !Result.empty();
572 // If comparing a live-in value against a constant, see if we know the
573 // live-in value on any predecessors.
574 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
575 if (!isa<Instruction>(Cmp->getOperand(0)) ||
576 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
577 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
579 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
581 // If the value is known by LazyValueInfo to be a constant in a
582 // predecessor, use that information to try to thread this block.
583 LazyValueInfo::Tristate Res =
584 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
585 RHSCst, P, BB, CxtI ? CxtI : Cmp);
586 if (Res == LazyValueInfo::Unknown)
589 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
590 Result.push_back(std::make_pair(ResC, P));
593 return !Result.empty();
596 // Try to find a constant value for the LHS of a comparison,
597 // and evaluate it statically if we can.
598 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
599 PredValueInfoTy LHSVals;
600 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
603 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
604 Constant *V = LHSVals[i].first;
605 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
607 if (Constant *KC = getKnownConstant(Folded, WantInteger))
608 Result.push_back(std::make_pair(KC, LHSVals[i].second));
611 return !Result.empty();
616 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
617 // Handle select instructions where at least one operand is a known constant
618 // and we can figure out the condition value for any predecessor block.
619 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
620 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
621 PredValueInfoTy Conds;
622 if ((TrueVal || FalseVal) &&
623 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
624 WantInteger, CxtI)) {
625 for (unsigned i = 0, e = Conds.size(); i != e; ++i) {
626 Constant *Cond = Conds[i].first;
628 // Figure out what value to use for the condition.
630 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
632 KnownCond = CI->isOne();
634 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
635 // Either operand will do, so be sure to pick the one that's a known
637 // FIXME: Do this more cleverly if both values are known constants?
638 KnownCond = (TrueVal != nullptr);
641 // See if the select has a known constant value for this predecessor.
642 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
643 Result.push_back(std::make_pair(Val, Conds[i].second));
646 return !Result.empty();
650 // If all else fails, see if LVI can figure out a constant value for us.
651 Constant *CI = LVI->getConstant(V, BB, CxtI);
652 if (Constant *KC = getKnownConstant(CI, Preference)) {
653 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
654 Result.push_back(std::make_pair(KC, *PI));
657 return !Result.empty();
662 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
663 /// in an undefined jump, decide which block is best to revector to.
665 /// Since we can pick an arbitrary destination, we pick the successor with the
666 /// fewest predecessors. This should reduce the in-degree of the others.
668 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
669 TerminatorInst *BBTerm = BB->getTerminator();
670 unsigned MinSucc = 0;
671 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
672 // Compute the successor with the minimum number of predecessors.
673 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
674 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
675 TestBB = BBTerm->getSuccessor(i);
676 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
677 if (NumPreds < MinNumPreds) {
679 MinNumPreds = NumPreds;
686 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
687 if (!BB->hasAddressTaken()) return false;
689 // If the block has its address taken, it may be a tree of dead constants
690 // hanging off of it. These shouldn't keep the block alive.
691 BlockAddress *BA = BlockAddress::get(BB);
692 BA->removeDeadConstantUsers();
693 return !BA->use_empty();
696 /// ProcessBlock - If there are any predecessors whose control can be threaded
697 /// through to a successor, transform them now.
698 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
699 // If the block is trivially dead, just return and let the caller nuke it.
700 // This simplifies other transformations.
701 if (pred_empty(BB) &&
702 BB != &BB->getParent()->getEntryBlock())
705 // If this block has a single predecessor, and if that pred has a single
706 // successor, merge the blocks. This encourages recursive jump threading
707 // because now the condition in this block can be threaded through
708 // predecessors of our predecessor block.
709 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
710 const TerminatorInst *TI = SinglePred->getTerminator();
711 if (!TI->isExceptional() && TI->getNumSuccessors() == 1 &&
712 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
713 // If SinglePred was a loop header, BB becomes one.
714 if (LoopHeaders.erase(SinglePred))
715 LoopHeaders.insert(BB);
717 LVI->eraseBlock(SinglePred);
718 MergeBasicBlockIntoOnlyPred(BB);
724 // What kind of constant we're looking for.
725 ConstantPreference Preference = WantInteger;
727 // Look to see if the terminator is a conditional branch, switch or indirect
728 // branch, if not we can't thread it.
730 Instruction *Terminator = BB->getTerminator();
731 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
732 // Can't thread an unconditional jump.
733 if (BI->isUnconditional()) return false;
734 Condition = BI->getCondition();
735 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
736 Condition = SI->getCondition();
737 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
738 // Can't thread indirect branch with no successors.
739 if (IB->getNumSuccessors() == 0) return false;
740 Condition = IB->getAddress()->stripPointerCasts();
741 Preference = WantBlockAddress;
743 return false; // Must be an invoke.
746 // Run constant folding to see if we can reduce the condition to a simple
748 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
750 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
752 I->replaceAllUsesWith(SimpleVal);
753 I->eraseFromParent();
754 Condition = SimpleVal;
758 // If the terminator is branching on an undef, we can pick any of the
759 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
760 if (isa<UndefValue>(Condition)) {
761 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
763 // Fold the branch/switch.
764 TerminatorInst *BBTerm = BB->getTerminator();
765 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
766 if (i == BestSucc) continue;
767 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
770 DEBUG(dbgs() << " In block '" << BB->getName()
771 << "' folding undef terminator: " << *BBTerm << '\n');
772 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
773 BBTerm->eraseFromParent();
777 // If the terminator of this block is branching on a constant, simplify the
778 // terminator to an unconditional branch. This can occur due to threading in
780 if (getKnownConstant(Condition, Preference)) {
781 DEBUG(dbgs() << " In block '" << BB->getName()
782 << "' folding terminator: " << *BB->getTerminator() << '\n');
784 ConstantFoldTerminator(BB, true);
788 Instruction *CondInst = dyn_cast<Instruction>(Condition);
790 // All the rest of our checks depend on the condition being an instruction.
792 // FIXME: Unify this with code below.
793 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
799 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
800 // If we're branching on a conditional, LVI might be able to determine
801 // it's value at the branch instruction. We only handle comparisons
802 // against a constant at this time.
803 // TODO: This should be extended to handle switches as well.
804 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
805 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
806 if (CondBr && CondConst && CondBr->isConditional()) {
807 LazyValueInfo::Tristate Ret =
808 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
810 if (Ret != LazyValueInfo::Unknown) {
811 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
812 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
813 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
814 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
815 CondBr->eraseFromParent();
816 if (CondCmp->use_empty())
817 CondCmp->eraseFromParent();
818 else if (CondCmp->getParent() == BB) {
819 // If the fact we just learned is true for all uses of the
820 // condition, replace it with a constant value
821 auto *CI = Ret == LazyValueInfo::True ?
822 ConstantInt::getTrue(CondCmp->getType()) :
823 ConstantInt::getFalse(CondCmp->getType());
824 CondCmp->replaceAllUsesWith(CI);
825 CondCmp->eraseFromParent();
831 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB))
835 // Check for some cases that are worth simplifying. Right now we want to look
836 // for loads that are used by a switch or by the condition for the branch. If
837 // we see one, check to see if it's partially redundant. If so, insert a PHI
838 // which can then be used to thread the values.
840 Value *SimplifyValue = CondInst;
841 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
842 if (isa<Constant>(CondCmp->getOperand(1)))
843 SimplifyValue = CondCmp->getOperand(0);
845 // TODO: There are other places where load PRE would be profitable, such as
846 // more complex comparisons.
847 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
848 if (SimplifyPartiallyRedundantLoad(LI))
852 // Handle a variety of cases where we are branching on something derived from
853 // a PHI node in the current block. If we can prove that any predecessors
854 // compute a predictable value based on a PHI node, thread those predecessors.
856 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
859 // If this is an otherwise-unfoldable branch on a phi node in the current
860 // block, see if we can simplify.
861 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
862 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
863 return ProcessBranchOnPHI(PN);
866 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
867 if (CondInst->getOpcode() == Instruction::Xor &&
868 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
869 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
872 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
873 // "(X == 4)", thread through this block.
878 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
879 /// load instruction, eliminate it by replacing it with a PHI node. This is an
880 /// important optimization that encourages jump threading, and needs to be run
881 /// interlaced with other jump threading tasks.
882 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
883 // Don't hack volatile/atomic loads.
884 if (!LI->isSimple()) return false;
886 // If the load is defined in a block with exactly one predecessor, it can't be
887 // partially redundant.
888 BasicBlock *LoadBB = LI->getParent();
889 if (LoadBB->getSinglePredecessor())
892 // If the load is defined in an EH pad, it can't be partially redundant,
893 // because the edges between the invoke and the EH pad cannot have other
894 // instructions between them.
895 if (LoadBB->isEHPad())
898 Value *LoadedPtr = LI->getOperand(0);
900 // If the loaded operand is defined in the LoadBB, it can't be available.
901 // TODO: Could do simple PHI translation, that would be fun :)
902 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
903 if (PtrOp->getParent() == LoadBB)
906 // Scan a few instructions up from the load, to see if it is obviously live at
907 // the entry to its block.
908 BasicBlock::iterator BBIt(LI);
910 if (Value *AvailableVal =
911 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, DefMaxInstsToScan)) {
912 // If the value of the load is locally available within the block, just use
913 // it. This frequently occurs for reg2mem'd allocas.
914 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
916 // If the returned value is the load itself, replace with an undef. This can
917 // only happen in dead loops.
918 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
919 if (AvailableVal->getType() != LI->getType())
921 CastInst::CreateBitOrPointerCast(AvailableVal, LI->getType(), "", LI);
922 LI->replaceAllUsesWith(AvailableVal);
923 LI->eraseFromParent();
927 // Otherwise, if we scanned the whole block and got to the top of the block,
928 // we know the block is locally transparent to the load. If not, something
929 // might clobber its value.
930 if (BBIt != LoadBB->begin())
933 // If all of the loads and stores that feed the value have the same AA tags,
934 // then we can propagate them onto any newly inserted loads.
936 LI->getAAMetadata(AATags);
938 SmallPtrSet<BasicBlock*, 8> PredsScanned;
939 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
940 AvailablePredsTy AvailablePreds;
941 BasicBlock *OneUnavailablePred = nullptr;
943 // If we got here, the loaded value is transparent through to the start of the
944 // block. Check to see if it is available in any of the predecessor blocks.
945 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
947 BasicBlock *PredBB = *PI;
949 // If we already scanned this predecessor, skip it.
950 if (!PredsScanned.insert(PredBB).second)
953 // Scan the predecessor to see if the value is available in the pred.
954 BBIt = PredBB->end();
955 AAMDNodes ThisAATags;
956 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt,
958 nullptr, &ThisAATags);
959 if (!PredAvailable) {
960 OneUnavailablePred = PredBB;
964 // If AA tags disagree or are not present, forget about them.
965 if (AATags != ThisAATags) AATags = AAMDNodes();
967 // If so, this load is partially redundant. Remember this info so that we
968 // can create a PHI node.
969 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
972 // If the loaded value isn't available in any predecessor, it isn't partially
974 if (AvailablePreds.empty()) return false;
976 // Okay, the loaded value is available in at least one (and maybe all!)
977 // predecessors. If the value is unavailable in more than one unique
978 // predecessor, we want to insert a merge block for those common predecessors.
979 // This ensures that we only have to insert one reload, thus not increasing
981 BasicBlock *UnavailablePred = nullptr;
983 // If there is exactly one predecessor where the value is unavailable, the
984 // already computed 'OneUnavailablePred' block is it. If it ends in an
985 // unconditional branch, we know that it isn't a critical edge.
986 if (PredsScanned.size() == AvailablePreds.size()+1 &&
987 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
988 UnavailablePred = OneUnavailablePred;
989 } else if (PredsScanned.size() != AvailablePreds.size()) {
990 // Otherwise, we had multiple unavailable predecessors or we had a critical
991 // edge from the one.
992 SmallVector<BasicBlock*, 8> PredsToSplit;
993 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
995 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
996 AvailablePredSet.insert(AvailablePreds[i].first);
998 // Add all the unavailable predecessors to the PredsToSplit list.
999 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
1001 BasicBlock *P = *PI;
1002 // If the predecessor is an indirect goto, we can't split the edge.
1003 if (isa<IndirectBrInst>(P->getTerminator()))
1006 if (!AvailablePredSet.count(P))
1007 PredsToSplit.push_back(P);
1010 // Split them out to their own block.
1011 UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1014 // If the value isn't available in all predecessors, then there will be
1015 // exactly one where it isn't available. Insert a load on that edge and add
1016 // it to the AvailablePreds list.
1017 if (UnavailablePred) {
1018 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1019 "Can't handle critical edge here!");
1020 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
1022 UnavailablePred->getTerminator());
1023 NewVal->setDebugLoc(LI->getDebugLoc());
1025 NewVal->setAAMetadata(AATags);
1027 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1030 // Now we know that each predecessor of this block has a value in
1031 // AvailablePreds, sort them for efficient access as we're walking the preds.
1032 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1034 // Create a PHI node at the start of the block for the PRE'd load value.
1035 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1036 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
1039 PN->setDebugLoc(LI->getDebugLoc());
1041 // Insert new entries into the PHI for each predecessor. A single block may
1042 // have multiple entries here.
1043 for (pred_iterator PI = PB; PI != PE; ++PI) {
1044 BasicBlock *P = *PI;
1045 AvailablePredsTy::iterator I =
1046 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1047 std::make_pair(P, (Value*)nullptr));
1049 assert(I != AvailablePreds.end() && I->first == P &&
1050 "Didn't find entry for predecessor!");
1052 // If we have an available predecessor but it requires casting, insert the
1053 // cast in the predecessor and use the cast. Note that we have to update the
1054 // AvailablePreds vector as we go so that all of the PHI entries for this
1055 // predecessor use the same bitcast.
1056 Value *&PredV = I->second;
1057 if (PredV->getType() != LI->getType())
1058 PredV = CastInst::CreateBitOrPointerCast(PredV, LI->getType(), "",
1059 P->getTerminator());
1061 PN->addIncoming(PredV, I->first);
1064 //cerr << "PRE: " << *LI << *PN << "\n";
1066 LI->replaceAllUsesWith(PN);
1067 LI->eraseFromParent();
1072 /// FindMostPopularDest - The specified list contains multiple possible
1073 /// threadable destinations. Pick the one that occurs the most frequently in
1076 FindMostPopularDest(BasicBlock *BB,
1077 const SmallVectorImpl<std::pair<BasicBlock*,
1078 BasicBlock*> > &PredToDestList) {
1079 assert(!PredToDestList.empty());
1081 // Determine popularity. If there are multiple possible destinations, we
1082 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1083 // blocks with known and real destinations to threading undef. We'll handle
1084 // them later if interesting.
1085 DenseMap<BasicBlock*, unsigned> DestPopularity;
1086 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1087 if (PredToDestList[i].second)
1088 DestPopularity[PredToDestList[i].second]++;
1090 // Find the most popular dest.
1091 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1092 BasicBlock *MostPopularDest = DPI->first;
1093 unsigned Popularity = DPI->second;
1094 SmallVector<BasicBlock*, 4> SamePopularity;
1096 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1097 // If the popularity of this entry isn't higher than the popularity we've
1098 // seen so far, ignore it.
1099 if (DPI->second < Popularity)
1101 else if (DPI->second == Popularity) {
1102 // If it is the same as what we've seen so far, keep track of it.
1103 SamePopularity.push_back(DPI->first);
1105 // If it is more popular, remember it.
1106 SamePopularity.clear();
1107 MostPopularDest = DPI->first;
1108 Popularity = DPI->second;
1112 // Okay, now we know the most popular destination. If there is more than one
1113 // destination, we need to determine one. This is arbitrary, but we need
1114 // to make a deterministic decision. Pick the first one that appears in the
1116 if (!SamePopularity.empty()) {
1117 SamePopularity.push_back(MostPopularDest);
1118 TerminatorInst *TI = BB->getTerminator();
1119 for (unsigned i = 0; ; ++i) {
1120 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1122 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1123 TI->getSuccessor(i)) == SamePopularity.end())
1126 MostPopularDest = TI->getSuccessor(i);
1131 // Okay, we have finally picked the most popular destination.
1132 return MostPopularDest;
1135 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1136 ConstantPreference Preference,
1137 Instruction *CxtI) {
1138 // If threading this would thread across a loop header, don't even try to
1140 if (LoopHeaders.count(BB))
1143 PredValueInfoTy PredValues;
1144 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1147 assert(!PredValues.empty() &&
1148 "ComputeValueKnownInPredecessors returned true with no values");
1150 DEBUG(dbgs() << "IN BB: " << *BB;
1151 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1152 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1153 << *PredValues[i].first
1154 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1157 // Decide what we want to thread through. Convert our list of known values to
1158 // a list of known destinations for each pred. This also discards duplicate
1159 // predecessors and keeps track of the undefined inputs (which are represented
1160 // as a null dest in the PredToDestList).
1161 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1162 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1164 BasicBlock *OnlyDest = nullptr;
1165 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1167 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1168 BasicBlock *Pred = PredValues[i].second;
1169 if (!SeenPreds.insert(Pred).second)
1170 continue; // Duplicate predecessor entry.
1172 // If the predecessor ends with an indirect goto, we can't change its
1174 if (isa<IndirectBrInst>(Pred->getTerminator()))
1177 Constant *Val = PredValues[i].first;
1180 if (isa<UndefValue>(Val))
1182 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1183 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1184 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1185 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1187 assert(isa<IndirectBrInst>(BB->getTerminator())
1188 && "Unexpected terminator");
1189 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1192 // If we have exactly one destination, remember it for efficiency below.
1193 if (PredToDestList.empty())
1195 else if (OnlyDest != DestBB)
1196 OnlyDest = MultipleDestSentinel;
1198 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1201 // If all edges were unthreadable, we fail.
1202 if (PredToDestList.empty())
1205 // Determine which is the most common successor. If we have many inputs and
1206 // this block is a switch, we want to start by threading the batch that goes
1207 // to the most popular destination first. If we only know about one
1208 // threadable destination (the common case) we can avoid this.
1209 BasicBlock *MostPopularDest = OnlyDest;
1211 if (MostPopularDest == MultipleDestSentinel)
1212 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1214 // Now that we know what the most popular destination is, factor all
1215 // predecessors that will jump to it into a single predecessor.
1216 SmallVector<BasicBlock*, 16> PredsToFactor;
1217 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1218 if (PredToDestList[i].second == MostPopularDest) {
1219 BasicBlock *Pred = PredToDestList[i].first;
1221 // This predecessor may be a switch or something else that has multiple
1222 // edges to the block. Factor each of these edges by listing them
1223 // according to # occurrences in PredsToFactor.
1224 TerminatorInst *PredTI = Pred->getTerminator();
1225 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1226 if (PredTI->getSuccessor(i) == BB)
1227 PredsToFactor.push_back(Pred);
1230 // If the threadable edges are branching on an undefined value, we get to pick
1231 // the destination that these predecessors should get to.
1232 if (!MostPopularDest)
1233 MostPopularDest = BB->getTerminator()->
1234 getSuccessor(GetBestDestForJumpOnUndef(BB));
1236 // Ok, try to thread it!
1237 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1240 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1241 /// a PHI node in the current block. See if there are any simplifications we
1242 /// can do based on inputs to the phi node.
1244 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1245 BasicBlock *BB = PN->getParent();
1247 // TODO: We could make use of this to do it once for blocks with common PHI
1249 SmallVector<BasicBlock*, 1> PredBBs;
1252 // If any of the predecessor blocks end in an unconditional branch, we can
1253 // *duplicate* the conditional branch into that block in order to further
1254 // encourage jump threading and to eliminate cases where we have branch on a
1255 // phi of an icmp (branch on icmp is much better).
1256 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1257 BasicBlock *PredBB = PN->getIncomingBlock(i);
1258 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1259 if (PredBr->isUnconditional()) {
1260 PredBBs[0] = PredBB;
1261 // Try to duplicate BB into PredBB.
1262 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1270 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1271 /// a xor instruction in the current block. See if there are any
1272 /// simplifications we can do based on inputs to the xor.
1274 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1275 BasicBlock *BB = BO->getParent();
1277 // If either the LHS or RHS of the xor is a constant, don't do this
1279 if (isa<ConstantInt>(BO->getOperand(0)) ||
1280 isa<ConstantInt>(BO->getOperand(1)))
1283 // If the first instruction in BB isn't a phi, we won't be able to infer
1284 // anything special about any particular predecessor.
1285 if (!isa<PHINode>(BB->front()))
1288 // If we have a xor as the branch input to this block, and we know that the
1289 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1290 // the condition into the predecessor and fix that value to true, saving some
1291 // logical ops on that path and encouraging other paths to simplify.
1293 // This copies something like this:
1296 // %X = phi i1 [1], [%X']
1297 // %Y = icmp eq i32 %A, %B
1298 // %Z = xor i1 %X, %Y
1303 // %Y = icmp ne i32 %A, %B
1306 PredValueInfoTy XorOpValues;
1308 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1310 assert(XorOpValues.empty());
1311 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1317 assert(!XorOpValues.empty() &&
1318 "ComputeValueKnownInPredecessors returned true with no values");
1320 // Scan the information to see which is most popular: true or false. The
1321 // predecessors can be of the set true, false, or undef.
1322 unsigned NumTrue = 0, NumFalse = 0;
1323 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1324 if (isa<UndefValue>(XorOpValues[i].first))
1325 // Ignore undefs for the count.
1327 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1333 // Determine which value to split on, true, false, or undef if neither.
1334 ConstantInt *SplitVal = nullptr;
1335 if (NumTrue > NumFalse)
1336 SplitVal = ConstantInt::getTrue(BB->getContext());
1337 else if (NumTrue != 0 || NumFalse != 0)
1338 SplitVal = ConstantInt::getFalse(BB->getContext());
1340 // Collect all of the blocks that this can be folded into so that we can
1341 // factor this once and clone it once.
1342 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1343 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1344 if (XorOpValues[i].first != SplitVal &&
1345 !isa<UndefValue>(XorOpValues[i].first))
1348 BlocksToFoldInto.push_back(XorOpValues[i].second);
1351 // If we inferred a value for all of the predecessors, then duplication won't
1352 // help us. However, we can just replace the LHS or RHS with the constant.
1353 if (BlocksToFoldInto.size() ==
1354 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1356 // If all preds provide undef, just nuke the xor, because it is undef too.
1357 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1358 BO->eraseFromParent();
1359 } else if (SplitVal->isZero()) {
1360 // If all preds provide 0, replace the xor with the other input.
1361 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1362 BO->eraseFromParent();
1364 // If all preds provide 1, set the computed value to 1.
1365 BO->setOperand(!isLHS, SplitVal);
1371 // Try to duplicate BB into PredBB.
1372 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1376 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1377 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1378 /// NewPred using the entries from OldPred (suitably mapped).
1379 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1380 BasicBlock *OldPred,
1381 BasicBlock *NewPred,
1382 DenseMap<Instruction*, Value*> &ValueMap) {
1383 for (BasicBlock::iterator PNI = PHIBB->begin();
1384 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1385 // Ok, we have a PHI node. Figure out what the incoming value was for the
1387 Value *IV = PN->getIncomingValueForBlock(OldPred);
1389 // Remap the value if necessary.
1390 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1391 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1392 if (I != ValueMap.end())
1396 PN->addIncoming(IV, NewPred);
1400 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1401 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1402 /// across BB. Transform the IR to reflect this change.
1403 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1404 const SmallVectorImpl<BasicBlock*> &PredBBs,
1405 BasicBlock *SuccBB) {
1406 // If threading to the same block as we come from, we would infinite loop.
1408 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1409 << "' - would thread to self!\n");
1413 // If threading this would thread across a loop header, don't thread the edge.
1414 // See the comments above FindLoopHeaders for justifications and caveats.
1415 if (LoopHeaders.count(BB)) {
1416 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1417 << "' to dest BB '" << SuccBB->getName()
1418 << "' - it might create an irreducible loop!\n");
1422 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, BBDupThreshold);
1423 if (JumpThreadCost > BBDupThreshold) {
1424 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1425 << "' - Cost is too high: " << JumpThreadCost << "\n");
1429 // And finally, do it! Start by factoring the predecessors if needed.
1431 if (PredBBs.size() == 1)
1432 PredBB = PredBBs[0];
1434 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1435 << " common predecessors.\n");
1436 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1439 // And finally, do it!
1440 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1441 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1442 << ", across block:\n "
1445 LVI->threadEdge(PredBB, BB, SuccBB);
1447 // We are going to have to map operands from the original BB block to the new
1448 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1449 // account for entry from PredBB.
1450 DenseMap<Instruction*, Value*> ValueMapping;
1452 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1453 BB->getName()+".thread",
1454 BB->getParent(), BB);
1455 NewBB->moveAfter(PredBB);
1457 // Set the block frequency of NewBB.
1458 if (HasProfileData) {
1460 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
1461 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
1464 BasicBlock::iterator BI = BB->begin();
1465 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1466 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1468 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1469 // mapping and using it to remap operands in the cloned instructions.
1470 for (; !isa<TerminatorInst>(BI); ++BI) {
1471 Instruction *New = BI->clone();
1472 New->setName(BI->getName());
1473 NewBB->getInstList().push_back(New);
1474 ValueMapping[&*BI] = New;
1476 // Remap operands to patch up intra-block references.
1477 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1478 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1479 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1480 if (I != ValueMapping.end())
1481 New->setOperand(i, I->second);
1485 // We didn't copy the terminator from BB over to NewBB, because there is now
1486 // an unconditional jump to SuccBB. Insert the unconditional jump.
1487 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
1488 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1490 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1491 // PHI nodes for NewBB now.
1492 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1494 // If there were values defined in BB that are used outside the block, then we
1495 // now have to update all uses of the value to use either the original value,
1496 // the cloned value, or some PHI derived value. This can require arbitrary
1497 // PHI insertion, of which we are prepared to do, clean these up now.
1498 SSAUpdater SSAUpdate;
1499 SmallVector<Use*, 16> UsesToRename;
1500 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1501 // Scan all uses of this instruction to see if it is used outside of its
1502 // block, and if so, record them in UsesToRename.
1503 for (Use &U : I->uses()) {
1504 Instruction *User = cast<Instruction>(U.getUser());
1505 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1506 if (UserPN->getIncomingBlock(U) == BB)
1508 } else if (User->getParent() == BB)
1511 UsesToRename.push_back(&U);
1514 // If there are no uses outside the block, we're done with this instruction.
1515 if (UsesToRename.empty())
1518 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1520 // We found a use of I outside of BB. Rename all uses of I that are outside
1521 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1522 // with the two values we know.
1523 SSAUpdate.Initialize(I->getType(), I->getName());
1524 SSAUpdate.AddAvailableValue(BB, &*I);
1525 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&*I]);
1527 while (!UsesToRename.empty())
1528 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1529 DEBUG(dbgs() << "\n");
1533 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1534 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1535 // us to simplify any PHI nodes in BB.
1536 TerminatorInst *PredTerm = PredBB->getTerminator();
1537 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1538 if (PredTerm->getSuccessor(i) == BB) {
1539 BB->removePredecessor(PredBB, true);
1540 PredTerm->setSuccessor(i, NewBB);
1543 // At this point, the IR is fully up to date and consistent. Do a quick scan
1544 // over the new instructions and zap any that are constants or dead. This
1545 // frequently happens because of phi translation.
1546 SimplifyInstructionsInBlock(NewBB, TLI);
1548 // Update the edge weight from BB to SuccBB, which should be less than before.
1549 UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
1551 // Threaded an edge!
1556 /// Create a new basic block that will be the predecessor of BB and successor of
1557 /// all blocks in Preds. When profile data is availble, update the frequency of
1559 BasicBlock *JumpThreading::SplitBlockPreds(BasicBlock *BB,
1560 ArrayRef<BasicBlock *> Preds,
1561 const char *Suffix) {
1562 // Collect the frequencies of all predecessors of BB, which will be used to
1563 // update the edge weight on BB->SuccBB.
1564 BlockFrequency PredBBFreq(0);
1566 for (auto Pred : Preds)
1567 PredBBFreq += BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB);
1569 BasicBlock *PredBB = SplitBlockPredecessors(BB, Preds, Suffix);
1571 // Set the block frequency of the newly created PredBB, which is the sum of
1572 // frequencies of Preds.
1574 BFI->setBlockFreq(PredBB, PredBBFreq.getFrequency());
1578 /// Update the block frequency of BB and branch weight and the metadata on the
1579 /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
1580 /// Freq(PredBB->BB) / Freq(BB->SuccBB).
1581 void JumpThreading::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
1584 BasicBlock *SuccBB) {
1585 if (!HasProfileData)
1588 assert(BFI && BPI && "BFI & BPI should have been created here");
1590 // As the edge from PredBB to BB is deleted, we have to update the block
1592 auto BBOrigFreq = BFI->getBlockFreq(BB);
1593 auto NewBBFreq = BFI->getBlockFreq(NewBB);
1594 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
1595 auto BBNewFreq = BBOrigFreq - NewBBFreq;
1596 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
1598 // Collect updated outgoing edges' frequencies from BB and use them to update
1600 SmallVector<uint64_t, 4> BBSuccFreq;
1601 for (auto I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
1602 auto SuccFreq = (*I == SuccBB)
1603 ? BB2SuccBBFreq - NewBBFreq
1604 : BBOrigFreq * BPI->getEdgeProbability(BB, *I);
1605 BBSuccFreq.push_back(SuccFreq.getFrequency());
1608 // Normalize edge weights in Weights64 so that the sum of them can fit in
1609 BranchProbability::normalizeEdgeWeights(BBSuccFreq.begin(), BBSuccFreq.end());
1611 SmallVector<uint32_t, 4> Weights;
1612 for (auto Freq : BBSuccFreq)
1613 Weights.push_back(static_cast<uint32_t>(Freq));
1615 // Update edge weights in BPI.
1616 for (int I = 0, E = Weights.size(); I < E; I++)
1617 BPI->setEdgeWeight(BB, I, Weights[I]);
1619 if (Weights.size() >= 2) {
1620 auto TI = BB->getTerminator();
1622 LLVMContext::MD_prof,
1623 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
1627 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1628 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1629 /// If we can duplicate the contents of BB up into PredBB do so now, this
1630 /// improves the odds that the branch will be on an analyzable instruction like
1632 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1633 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1634 assert(!PredBBs.empty() && "Can't handle an empty set");
1636 // If BB is a loop header, then duplicating this block outside the loop would
1637 // cause us to transform this into an irreducible loop, don't do this.
1638 // See the comments above FindLoopHeaders for justifications and caveats.
1639 if (LoopHeaders.count(BB)) {
1640 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1641 << "' into predecessor block '" << PredBBs[0]->getName()
1642 << "' - it might create an irreducible loop!\n");
1646 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, BBDupThreshold);
1647 if (DuplicationCost > BBDupThreshold) {
1648 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1649 << "' - Cost is too high: " << DuplicationCost << "\n");
1653 // And finally, do it! Start by factoring the predecessors if needed.
1655 if (PredBBs.size() == 1)
1656 PredBB = PredBBs[0];
1658 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1659 << " common predecessors.\n");
1660 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1663 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1665 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1666 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1667 << DuplicationCost << " block is:" << *BB << "\n");
1669 // Unless PredBB ends with an unconditional branch, split the edge so that we
1670 // can just clone the bits from BB into the end of the new PredBB.
1671 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1673 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
1674 PredBB = SplitEdge(PredBB, BB);
1675 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1678 // We are going to have to map operands from the original BB block into the
1679 // PredBB block. Evaluate PHI nodes in BB.
1680 DenseMap<Instruction*, Value*> ValueMapping;
1682 BasicBlock::iterator BI = BB->begin();
1683 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1684 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1685 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1686 // mapping and using it to remap operands in the cloned instructions.
1687 for (; BI != BB->end(); ++BI) {
1688 Instruction *New = BI->clone();
1690 // Remap operands to patch up intra-block references.
1691 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1692 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1693 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1694 if (I != ValueMapping.end())
1695 New->setOperand(i, I->second);
1698 // If this instruction can be simplified after the operands are updated,
1699 // just use the simplified value instead. This frequently happens due to
1702 SimplifyInstruction(New, BB->getModule()->getDataLayout())) {
1704 ValueMapping[&*BI] = IV;
1706 // Otherwise, insert the new instruction into the block.
1707 New->setName(BI->getName());
1708 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
1709 ValueMapping[&*BI] = New;
1713 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1714 // add entries to the PHI nodes for branch from PredBB now.
1715 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1716 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1718 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1721 // If there were values defined in BB that are used outside the block, then we
1722 // now have to update all uses of the value to use either the original value,
1723 // the cloned value, or some PHI derived value. This can require arbitrary
1724 // PHI insertion, of which we are prepared to do, clean these up now.
1725 SSAUpdater SSAUpdate;
1726 SmallVector<Use*, 16> UsesToRename;
1727 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1728 // Scan all uses of this instruction to see if it is used outside of its
1729 // block, and if so, record them in UsesToRename.
1730 for (Use &U : I->uses()) {
1731 Instruction *User = cast<Instruction>(U.getUser());
1732 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1733 if (UserPN->getIncomingBlock(U) == BB)
1735 } else if (User->getParent() == BB)
1738 UsesToRename.push_back(&U);
1741 // If there are no uses outside the block, we're done with this instruction.
1742 if (UsesToRename.empty())
1745 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1747 // We found a use of I outside of BB. Rename all uses of I that are outside
1748 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1749 // with the two values we know.
1750 SSAUpdate.Initialize(I->getType(), I->getName());
1751 SSAUpdate.AddAvailableValue(BB, &*I);
1752 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&*I]);
1754 while (!UsesToRename.empty())
1755 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1756 DEBUG(dbgs() << "\n");
1759 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1761 BB->removePredecessor(PredBB, true);
1763 // Remove the unconditional branch at the end of the PredBB block.
1764 OldPredBranch->eraseFromParent();
1770 /// TryToUnfoldSelect - Look for blocks of the form
1776 /// %p = phi [%a, %bb] ...
1780 /// And expand the select into a branch structure if one of its arms allows %c
1781 /// to be folded. This later enables threading from bb1 over bb2.
1782 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1783 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1784 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1785 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1787 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1788 CondLHS->getParent() != BB)
1791 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1792 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1793 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1795 // Look if one of the incoming values is a select in the corresponding
1797 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1800 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1801 if (!PredTerm || !PredTerm->isUnconditional())
1804 // Now check if one of the select values would allow us to constant fold the
1805 // terminator in BB. We don't do the transform if both sides fold, those
1806 // cases will be threaded in any case.
1807 LazyValueInfo::Tristate LHSFolds =
1808 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1809 CondRHS, Pred, BB, CondCmp);
1810 LazyValueInfo::Tristate RHSFolds =
1811 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
1812 CondRHS, Pred, BB, CondCmp);
1813 if ((LHSFolds != LazyValueInfo::Unknown ||
1814 RHSFolds != LazyValueInfo::Unknown) &&
1815 LHSFolds != RHSFolds) {
1816 // Expand the select.
1825 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
1826 BB->getParent(), BB);
1827 // Move the unconditional branch to NewBB.
1828 PredTerm->removeFromParent();
1829 NewBB->getInstList().insert(NewBB->end(), PredTerm);
1830 // Create a conditional branch and update PHI nodes.
1831 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
1832 CondLHS->setIncomingValue(I, SI->getFalseValue());
1833 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
1834 // The select is now dead.
1835 SI->eraseFromParent();
1837 // Update any other PHI nodes in BB.
1838 for (BasicBlock::iterator BI = BB->begin();
1839 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
1841 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);