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/CodeGen/MachineBranchProbabilityInfo.h"
33 #include "llvm/IR/DataLayout.h"
34 #include "llvm/IR/IntrinsicInst.h"
35 #include "llvm/IR/LLVMContext.h"
36 #include "llvm/IR/MDBuilder.h"
37 #include "llvm/IR/Metadata.h"
38 #include "llvm/IR/ValueHandle.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
44 #include "llvm/Transforms/Utils/Local.h"
45 #include "llvm/Transforms/Utils/SSAUpdater.h"
50 #define DEBUG_TYPE "jump-threading"
52 STATISTIC(NumThreads, "Number of jumps threaded");
53 STATISTIC(NumFolds, "Number of terminators folded");
54 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
56 static cl::opt<unsigned>
57 BBDuplicateThreshold("jump-threading-threshold",
58 cl::desc("Max block size to duplicate for jump threading"),
59 cl::init(6), cl::Hidden);
62 // These are at global scope so static functions can use them too.
63 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo;
64 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy;
66 // This is used to keep track of what kind of constant we're currently hoping
68 enum ConstantPreference {
73 /// This pass performs 'jump threading', which looks at blocks that have
74 /// multiple predecessors and multiple successors. If one or more of the
75 /// predecessors of the block can be proven to always jump to one of the
76 /// successors, we forward the edge from the predecessor to the successor by
77 /// duplicating the contents of this block.
79 /// An example of when this can occur is code like this:
86 /// In this case, the unconditional branch at the end of the first if can be
87 /// revectored to the false side of the second if.
89 class JumpThreading : public FunctionPass {
90 TargetLibraryInfo *TLI;
92 std::unique_ptr<BlockFrequencyInfo> BFI;
93 std::unique_ptr<BranchProbabilityInfo> BPI;
96 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
98 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
100 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet;
102 unsigned BBDupThreshold;
104 // RAII helper for updating the recursion stack.
105 struct RecursionSetRemover {
106 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet;
107 std::pair<Value*, BasicBlock*> ThePair;
109 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S,
110 std::pair<Value*, BasicBlock*> P)
111 : TheSet(S), ThePair(P) { }
113 ~RecursionSetRemover() {
114 TheSet.erase(ThePair);
118 static char ID; // Pass identification
119 JumpThreading(int T = -1) : FunctionPass(ID) {
120 BBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
121 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
124 bool runOnFunction(Function &F) override;
126 void getAnalysisUsage(AnalysisUsage &AU) const override {
127 AU.addRequired<LazyValueInfo>();
128 AU.addPreserved<LazyValueInfo>();
129 AU.addPreserved<GlobalsAAWrapperPass>();
130 AU.addRequired<TargetLibraryInfoWrapperPass>();
133 void releaseMemory() override {
138 void FindLoopHeaders(Function &F);
139 bool ProcessBlock(BasicBlock *BB);
140 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
142 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
143 const SmallVectorImpl<BasicBlock *> &PredBBs);
145 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
146 PredValueInfo &Result,
147 ConstantPreference Preference,
148 Instruction *CxtI = nullptr);
149 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
150 ConstantPreference Preference,
151 Instruction *CxtI = nullptr);
153 bool ProcessBranchOnPHI(PHINode *PN);
154 bool ProcessBranchOnXOR(BinaryOperator *BO);
156 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
157 bool TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB);
160 BasicBlock *SplitBlockPreds(BasicBlock *BB, ArrayRef<BasicBlock *> Preds,
162 void UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB, BasicBlock *BB,
163 BasicBlock *NewBB, BasicBlock *SuccBB);
167 char JumpThreading::ID = 0;
168 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
169 "Jump Threading", false, false)
170 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)
171 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
172 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
173 "Jump Threading", false, false)
175 // Public interface to the Jump Threading pass
176 FunctionPass *llvm::createJumpThreadingPass(int Threshold) { return new JumpThreading(Threshold); }
178 /// runOnFunction - Top level algorithm.
180 bool JumpThreading::runOnFunction(Function &F) {
181 if (skipOptnoneFunction(F))
184 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
185 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
186 LVI = &getAnalysis<LazyValueInfo>();
189 // When profile data is available, we need to update edge weights after
190 // successful jump threading, which requires both BPI and BFI being available.
191 HasProfileData = F.getEntryCount().hasValue();
192 if (HasProfileData) {
193 LoopInfo LI{DominatorTree(F)};
194 BPI.reset(new BranchProbabilityInfo(F, LI));
195 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
198 // Remove unreachable blocks from function as they may result in infinite
199 // loop. We do threading if we found something profitable. Jump threading a
200 // branch can create other opportunities. If these opportunities form a cycle
201 // i.e. if any jump threading is undoing previous threading in the path, then
202 // we will loop forever. We take care of this issue by not jump threading for
203 // back edges. This works for normal cases but not for unreachable blocks as
204 // they may have cycle with no back edge.
205 removeUnreachableBlocks(F);
209 bool Changed, EverChanged = false;
212 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
213 BasicBlock *BB = &*I;
214 // Thread all of the branches we can over this block.
215 while (ProcessBlock(BB))
220 // If the block is trivially dead, zap it. This eliminates the successor
221 // edges which simplifies the CFG.
222 if (pred_empty(BB) &&
223 BB != &BB->getParent()->getEntryBlock()) {
224 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
225 << "' with terminator: " << *BB->getTerminator() << '\n');
226 LoopHeaders.erase(BB);
233 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
235 // Can't thread an unconditional jump, but if the block is "almost
236 // empty", we can replace uses of it with uses of the successor and make
238 if (BI && BI->isUnconditional() &&
239 BB != &BB->getParent()->getEntryBlock() &&
240 // If the terminator is the only non-phi instruction, try to nuke it.
241 BB->getFirstNonPHIOrDbg()->isTerminator()) {
242 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
243 // block, we have to make sure it isn't in the LoopHeaders set. We
244 // reinsert afterward if needed.
245 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
246 BasicBlock *Succ = BI->getSuccessor(0);
248 // FIXME: It is always conservatively correct to drop the info
249 // for a block even if it doesn't get erased. This isn't totally
250 // awesome, but it allows us to use AssertingVH to prevent nasty
251 // dangling pointer issues within LazyValueInfo.
253 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
255 // If we deleted BB and BB was the header of a loop, then the
256 // successor is now the header of the loop.
260 if (ErasedFromLoopHeaders)
261 LoopHeaders.insert(BB);
264 EverChanged |= Changed;
271 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
272 /// thread across it. Stop scanning the block when passing the threshold.
273 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
274 unsigned Threshold) {
275 /// Ignore PHI nodes, these will be flattened when duplication happens.
276 BasicBlock::const_iterator I(BB->getFirstNonPHI());
278 // FIXME: THREADING will delete values that are just used to compute the
279 // branch, so they shouldn't count against the duplication cost.
281 // Sum up the cost of each instruction until we get to the terminator. Don't
282 // include the terminator because the copy won't include it.
284 for (; !isa<TerminatorInst>(I); ++I) {
286 // Stop scanning the block if we've reached the threshold.
287 if (Size > Threshold)
290 // Debugger intrinsics don't incur code size.
291 if (isa<DbgInfoIntrinsic>(I)) continue;
293 // If this is a pointer->pointer bitcast, it is free.
294 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
297 // Bail out if this instruction gives back a token type, it is not possible
298 // to duplicate it if it is used outside this BB.
299 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
302 // All other instructions count for at least one unit.
305 // Calls are more expensive. If they are non-intrinsic calls, we model them
306 // as having cost of 4. If they are a non-vector intrinsic, we model them
307 // as having cost of 2 total, and if they are a vector intrinsic, we model
308 // them as having cost 1.
309 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
310 if (CI->cannotDuplicate() || CI->isConvergent())
311 // Blocks with NoDuplicate are modelled as having infinite cost, so they
312 // are never duplicated.
314 else if (!isa<IntrinsicInst>(CI))
316 else if (!CI->getType()->isVectorTy())
321 // Threading through a switch statement is particularly profitable. If this
322 // block ends in a switch, decrease its cost to make it more likely to happen.
323 if (isa<SwitchInst>(I))
324 Size = Size > 6 ? Size-6 : 0;
326 // The same holds for indirect branches, but slightly more so.
327 if (isa<IndirectBrInst>(I))
328 Size = Size > 8 ? Size-8 : 0;
333 /// FindLoopHeaders - We do not want jump threading to turn proper loop
334 /// structures into irreducible loops. Doing this breaks up the loop nesting
335 /// hierarchy and pessimizes later transformations. To prevent this from
336 /// happening, we first have to find the loop headers. Here we approximate this
337 /// by finding targets of backedges in the CFG.
339 /// Note that there definitely are cases when we want to allow threading of
340 /// edges across a loop header. For example, threading a jump from outside the
341 /// loop (the preheader) to an exit block of the loop is definitely profitable.
342 /// It is also almost always profitable to thread backedges from within the loop
343 /// to exit blocks, and is often profitable to thread backedges to other blocks
344 /// within the loop (forming a nested loop). This simple analysis is not rich
345 /// enough to track all of these properties and keep it up-to-date as the CFG
346 /// mutates, so we don't allow any of these transformations.
348 void JumpThreading::FindLoopHeaders(Function &F) {
349 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
350 FindFunctionBackedges(F, Edges);
352 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
353 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
356 /// getKnownConstant - Helper method to determine if we can thread over a
357 /// terminator with the given value as its condition, and if so what value to
358 /// use for that. What kind of value this is depends on whether we want an
359 /// integer or a block address, but an undef is always accepted.
360 /// Returns null if Val is null or not an appropriate constant.
361 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
365 // Undef is "known" enough.
366 if (UndefValue *U = dyn_cast<UndefValue>(Val))
369 if (Preference == WantBlockAddress)
370 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
372 return dyn_cast<ConstantInt>(Val);
375 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
376 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
377 /// in any of our predecessors. If so, return the known list of value and pred
378 /// BB in the result vector.
380 /// This returns true if there were any known values.
383 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
384 ConstantPreference Preference,
386 // This method walks up use-def chains recursively. Because of this, we could
387 // get into an infinite loop going around loops in the use-def chain. To
388 // prevent this, keep track of what (value, block) pairs we've already visited
389 // and terminate the search if we loop back to them
390 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
393 // An RAII help to remove this pair from the recursion set once the recursion
394 // stack pops back out again.
395 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
397 // If V is a constant, then it is known in all predecessors.
398 if (Constant *KC = getKnownConstant(V, Preference)) {
399 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
400 Result.push_back(std::make_pair(KC, *PI));
405 // If V is a non-instruction value, or an instruction in a different block,
406 // then it can't be derived from a PHI.
407 Instruction *I = dyn_cast<Instruction>(V);
408 if (!I || I->getParent() != BB) {
410 // Okay, if this is a live-in value, see if it has a known value at the end
411 // of any of our predecessors.
413 // FIXME: This should be an edge property, not a block end property.
414 /// TODO: Per PR2563, we could infer value range information about a
415 /// predecessor based on its terminator.
417 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
418 // "I" is a non-local compare-with-a-constant instruction. This would be
419 // able to handle value inequalities better, for example if the compare is
420 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
421 // Perhaps getConstantOnEdge should be smart enough to do this?
423 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
425 // If the value is known by LazyValueInfo to be a constant in a
426 // predecessor, use that information to try to thread this block.
427 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
428 if (Constant *KC = getKnownConstant(PredCst, Preference))
429 Result.push_back(std::make_pair(KC, P));
432 return !Result.empty();
435 /// If I is a PHI node, then we know the incoming values for any constants.
436 if (PHINode *PN = dyn_cast<PHINode>(I)) {
437 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
438 Value *InVal = PN->getIncomingValue(i);
439 if (Constant *KC = getKnownConstant(InVal, Preference)) {
440 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
442 Constant *CI = LVI->getConstantOnEdge(InVal,
443 PN->getIncomingBlock(i),
445 if (Constant *KC = getKnownConstant(CI, Preference))
446 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
450 return !Result.empty();
453 PredValueInfoTy LHSVals, RHSVals;
455 // Handle some boolean conditions.
456 if (I->getType()->getPrimitiveSizeInBits() == 1) {
457 assert(Preference == WantInteger && "One-bit non-integer type?");
459 // X & false -> false
460 if (I->getOpcode() == Instruction::Or ||
461 I->getOpcode() == Instruction::And) {
462 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
464 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
467 if (LHSVals.empty() && RHSVals.empty())
470 ConstantInt *InterestingVal;
471 if (I->getOpcode() == Instruction::Or)
472 InterestingVal = ConstantInt::getTrue(I->getContext());
474 InterestingVal = ConstantInt::getFalse(I->getContext());
476 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
478 // Scan for the sentinel. If we find an undef, force it to the
479 // interesting value: x|undef -> true and x&undef -> false.
480 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
481 if (LHSVals[i].first == InterestingVal ||
482 isa<UndefValue>(LHSVals[i].first)) {
483 Result.push_back(LHSVals[i]);
484 Result.back().first = InterestingVal;
485 LHSKnownBBs.insert(LHSVals[i].second);
487 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
488 if (RHSVals[i].first == InterestingVal ||
489 isa<UndefValue>(RHSVals[i].first)) {
490 // If we already inferred a value for this block on the LHS, don't
492 if (!LHSKnownBBs.count(RHSVals[i].second)) {
493 Result.push_back(RHSVals[i]);
494 Result.back().first = InterestingVal;
498 return !Result.empty();
501 // Handle the NOT form of XOR.
502 if (I->getOpcode() == Instruction::Xor &&
503 isa<ConstantInt>(I->getOperand(1)) &&
504 cast<ConstantInt>(I->getOperand(1))->isOne()) {
505 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
510 // Invert the known values.
511 for (unsigned i = 0, e = Result.size(); i != e; ++i)
512 Result[i].first = ConstantExpr::getNot(Result[i].first);
517 // Try to simplify some other binary operator values.
518 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
519 assert(Preference != WantBlockAddress
520 && "A binary operator creating a block address?");
521 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
522 PredValueInfoTy LHSVals;
523 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
526 // Try to use constant folding to simplify the binary operator.
527 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
528 Constant *V = LHSVals[i].first;
529 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
531 if (Constant *KC = getKnownConstant(Folded, WantInteger))
532 Result.push_back(std::make_pair(KC, LHSVals[i].second));
536 return !Result.empty();
539 // Handle compare with phi operand, where the PHI is defined in this block.
540 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
541 assert(Preference == WantInteger && "Compares only produce integers");
542 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
543 if (PN && PN->getParent() == BB) {
544 const DataLayout &DL = PN->getModule()->getDataLayout();
545 // We can do this simplification if any comparisons fold to true or false.
547 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
548 BasicBlock *PredBB = PN->getIncomingBlock(i);
549 Value *LHS = PN->getIncomingValue(i);
550 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
552 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL);
554 if (!isa<Constant>(RHS))
557 LazyValueInfo::Tristate
558 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
559 cast<Constant>(RHS), PredBB, BB,
561 if (ResT == LazyValueInfo::Unknown)
563 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
566 if (Constant *KC = getKnownConstant(Res, WantInteger))
567 Result.push_back(std::make_pair(KC, PredBB));
570 return !Result.empty();
573 // If comparing a live-in value against a constant, see if we know the
574 // live-in value on any predecessors.
575 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
576 if (!isa<Instruction>(Cmp->getOperand(0)) ||
577 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
578 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
580 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
582 // If the value is known by LazyValueInfo to be a constant in a
583 // predecessor, use that information to try to thread this block.
584 LazyValueInfo::Tristate Res =
585 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
586 RHSCst, P, BB, CxtI ? CxtI : Cmp);
587 if (Res == LazyValueInfo::Unknown)
590 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
591 Result.push_back(std::make_pair(ResC, P));
594 return !Result.empty();
597 // Try to find a constant value for the LHS of a comparison,
598 // and evaluate it statically if we can.
599 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
600 PredValueInfoTy LHSVals;
601 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
604 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
605 Constant *V = LHSVals[i].first;
606 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
608 if (Constant *KC = getKnownConstant(Folded, WantInteger))
609 Result.push_back(std::make_pair(KC, LHSVals[i].second));
612 return !Result.empty();
617 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
618 // Handle select instructions where at least one operand is a known constant
619 // and we can figure out the condition value for any predecessor block.
620 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
621 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
622 PredValueInfoTy Conds;
623 if ((TrueVal || FalseVal) &&
624 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
625 WantInteger, CxtI)) {
626 for (unsigned i = 0, e = Conds.size(); i != e; ++i) {
627 Constant *Cond = Conds[i].first;
629 // Figure out what value to use for the condition.
631 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
633 KnownCond = CI->isOne();
635 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
636 // Either operand will do, so be sure to pick the one that's a known
638 // FIXME: Do this more cleverly if both values are known constants?
639 KnownCond = (TrueVal != nullptr);
642 // See if the select has a known constant value for this predecessor.
643 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
644 Result.push_back(std::make_pair(Val, Conds[i].second));
647 return !Result.empty();
651 // If all else fails, see if LVI can figure out a constant value for us.
652 Constant *CI = LVI->getConstant(V, BB, CxtI);
653 if (Constant *KC = getKnownConstant(CI, Preference)) {
654 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
655 Result.push_back(std::make_pair(KC, *PI));
658 return !Result.empty();
663 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
664 /// in an undefined jump, decide which block is best to revector to.
666 /// Since we can pick an arbitrary destination, we pick the successor with the
667 /// fewest predecessors. This should reduce the in-degree of the others.
669 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
670 TerminatorInst *BBTerm = BB->getTerminator();
671 unsigned MinSucc = 0;
672 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
673 // Compute the successor with the minimum number of predecessors.
674 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
675 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
676 TestBB = BBTerm->getSuccessor(i);
677 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
678 if (NumPreds < MinNumPreds) {
680 MinNumPreds = NumPreds;
687 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
688 if (!BB->hasAddressTaken()) return false;
690 // If the block has its address taken, it may be a tree of dead constants
691 // hanging off of it. These shouldn't keep the block alive.
692 BlockAddress *BA = BlockAddress::get(BB);
693 BA->removeDeadConstantUsers();
694 return !BA->use_empty();
697 /// ProcessBlock - If there are any predecessors whose control can be threaded
698 /// through to a successor, transform them now.
699 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
700 // If the block is trivially dead, just return and let the caller nuke it.
701 // This simplifies other transformations.
702 if (pred_empty(BB) &&
703 BB != &BB->getParent()->getEntryBlock())
706 // If this block has a single predecessor, and if that pred has a single
707 // successor, merge the blocks. This encourages recursive jump threading
708 // because now the condition in this block can be threaded through
709 // predecessors of our predecessor block.
710 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
711 const TerminatorInst *TI = SinglePred->getTerminator();
712 if (!TI->isExceptional() && TI->getNumSuccessors() == 1 &&
713 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
714 // If SinglePred was a loop header, BB becomes one.
715 if (LoopHeaders.erase(SinglePred))
716 LoopHeaders.insert(BB);
718 LVI->eraseBlock(SinglePred);
719 MergeBasicBlockIntoOnlyPred(BB);
725 // What kind of constant we're looking for.
726 ConstantPreference Preference = WantInteger;
728 // Look to see if the terminator is a conditional branch, switch or indirect
729 // branch, if not we can't thread it.
731 Instruction *Terminator = BB->getTerminator();
732 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
733 // Can't thread an unconditional jump.
734 if (BI->isUnconditional()) return false;
735 Condition = BI->getCondition();
736 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
737 Condition = SI->getCondition();
738 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
739 // Can't thread indirect branch with no successors.
740 if (IB->getNumSuccessors() == 0) return false;
741 Condition = IB->getAddress()->stripPointerCasts();
742 Preference = WantBlockAddress;
744 return false; // Must be an invoke.
747 // Run constant folding to see if we can reduce the condition to a simple
749 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
751 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
753 I->replaceAllUsesWith(SimpleVal);
754 I->eraseFromParent();
755 Condition = SimpleVal;
759 // If the terminator is branching on an undef, we can pick any of the
760 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
761 if (isa<UndefValue>(Condition)) {
762 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
764 // Fold the branch/switch.
765 TerminatorInst *BBTerm = BB->getTerminator();
766 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
767 if (i == BestSucc) continue;
768 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
771 DEBUG(dbgs() << " In block '" << BB->getName()
772 << "' folding undef terminator: " << *BBTerm << '\n');
773 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
774 BBTerm->eraseFromParent();
778 // If the terminator of this block is branching on a constant, simplify the
779 // terminator to an unconditional branch. This can occur due to threading in
781 if (getKnownConstant(Condition, Preference)) {
782 DEBUG(dbgs() << " In block '" << BB->getName()
783 << "' folding terminator: " << *BB->getTerminator() << '\n');
785 ConstantFoldTerminator(BB, true);
789 Instruction *CondInst = dyn_cast<Instruction>(Condition);
791 // All the rest of our checks depend on the condition being an instruction.
793 // FIXME: Unify this with code below.
794 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
800 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
801 // If we're branching on a conditional, LVI might be able to determine
802 // it's value at the branch instruction. We only handle comparisons
803 // against a constant at this time.
804 // TODO: This should be extended to handle switches as well.
805 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
806 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
807 if (CondBr && CondConst && CondBr->isConditional()) {
808 LazyValueInfo::Tristate Ret =
809 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
811 if (Ret != LazyValueInfo::Unknown) {
812 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
813 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
814 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
815 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
816 CondBr->eraseFromParent();
817 if (CondCmp->use_empty())
818 CondCmp->eraseFromParent();
819 else if (CondCmp->getParent() == BB) {
820 // If the fact we just learned is true for all uses of the
821 // condition, replace it with a constant value
822 auto *CI = Ret == LazyValueInfo::True ?
823 ConstantInt::getTrue(CondCmp->getType()) :
824 ConstantInt::getFalse(CondCmp->getType());
825 CondCmp->replaceAllUsesWith(CI);
826 CondCmp->eraseFromParent();
832 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB))
836 // Check for some cases that are worth simplifying. Right now we want to look
837 // for loads that are used by a switch or by the condition for the branch. If
838 // we see one, check to see if it's partially redundant. If so, insert a PHI
839 // which can then be used to thread the values.
841 Value *SimplifyValue = CondInst;
842 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
843 if (isa<Constant>(CondCmp->getOperand(1)))
844 SimplifyValue = CondCmp->getOperand(0);
846 // TODO: There are other places where load PRE would be profitable, such as
847 // more complex comparisons.
848 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
849 if (SimplifyPartiallyRedundantLoad(LI))
853 // Handle a variety of cases where we are branching on something derived from
854 // a PHI node in the current block. If we can prove that any predecessors
855 // compute a predictable value based on a PHI node, thread those predecessors.
857 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
860 // If this is an otherwise-unfoldable branch on a phi node in the current
861 // block, see if we can simplify.
862 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
863 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
864 return ProcessBranchOnPHI(PN);
867 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
868 if (CondInst->getOpcode() == Instruction::Xor &&
869 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
870 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
873 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
874 // "(X == 4)", thread through this block.
879 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
880 /// load instruction, eliminate it by replacing it with a PHI node. This is an
881 /// important optimization that encourages jump threading, and needs to be run
882 /// interlaced with other jump threading tasks.
883 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
884 // Don't hack volatile/atomic loads.
885 if (!LI->isSimple()) return false;
887 // If the load is defined in a block with exactly one predecessor, it can't be
888 // partially redundant.
889 BasicBlock *LoadBB = LI->getParent();
890 if (LoadBB->getSinglePredecessor())
893 // If the load is defined in an EH pad, it can't be partially redundant,
894 // because the edges between the invoke and the EH pad cannot have other
895 // instructions between them.
896 if (LoadBB->isEHPad())
899 Value *LoadedPtr = LI->getOperand(0);
901 // If the loaded operand is defined in the LoadBB, it can't be available.
902 // TODO: Could do simple PHI translation, that would be fun :)
903 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
904 if (PtrOp->getParent() == LoadBB)
907 // Scan a few instructions up from the load, to see if it is obviously live at
908 // the entry to its block.
909 BasicBlock::iterator BBIt(LI);
911 if (Value *AvailableVal =
912 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, DefMaxInstsToScan)) {
913 // If the value of the load is locally available within the block, just use
914 // it. This frequently occurs for reg2mem'd allocas.
915 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
917 // If the returned value is the load itself, replace with an undef. This can
918 // only happen in dead loops.
919 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
920 if (AvailableVal->getType() != LI->getType())
922 CastInst::CreateBitOrPointerCast(AvailableVal, LI->getType(), "", LI);
923 LI->replaceAllUsesWith(AvailableVal);
924 LI->eraseFromParent();
928 // Otherwise, if we scanned the whole block and got to the top of the block,
929 // we know the block is locally transparent to the load. If not, something
930 // might clobber its value.
931 if (BBIt != LoadBB->begin())
934 // If all of the loads and stores that feed the value have the same AA tags,
935 // then we can propagate them onto any newly inserted loads.
937 LI->getAAMetadata(AATags);
939 SmallPtrSet<BasicBlock*, 8> PredsScanned;
940 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
941 AvailablePredsTy AvailablePreds;
942 BasicBlock *OneUnavailablePred = nullptr;
944 // If we got here, the loaded value is transparent through to the start of the
945 // block. Check to see if it is available in any of the predecessor blocks.
946 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
948 BasicBlock *PredBB = *PI;
950 // If we already scanned this predecessor, skip it.
951 if (!PredsScanned.insert(PredBB).second)
954 // Scan the predecessor to see if the value is available in the pred.
955 BBIt = PredBB->end();
956 AAMDNodes ThisAATags;
957 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt,
959 nullptr, &ThisAATags);
960 if (!PredAvailable) {
961 OneUnavailablePred = PredBB;
965 // If AA tags disagree or are not present, forget about them.
966 if (AATags != ThisAATags) AATags = AAMDNodes();
968 // If so, this load is partially redundant. Remember this info so that we
969 // can create a PHI node.
970 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
973 // If the loaded value isn't available in any predecessor, it isn't partially
975 if (AvailablePreds.empty()) return false;
977 // Okay, the loaded value is available in at least one (and maybe all!)
978 // predecessors. If the value is unavailable in more than one unique
979 // predecessor, we want to insert a merge block for those common predecessors.
980 // This ensures that we only have to insert one reload, thus not increasing
982 BasicBlock *UnavailablePred = nullptr;
984 // If there is exactly one predecessor where the value is unavailable, the
985 // already computed 'OneUnavailablePred' block is it. If it ends in an
986 // unconditional branch, we know that it isn't a critical edge.
987 if (PredsScanned.size() == AvailablePreds.size()+1 &&
988 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
989 UnavailablePred = OneUnavailablePred;
990 } else if (PredsScanned.size() != AvailablePreds.size()) {
991 // Otherwise, we had multiple unavailable predecessors or we had a critical
992 // edge from the one.
993 SmallVector<BasicBlock*, 8> PredsToSplit;
994 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
996 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
997 AvailablePredSet.insert(AvailablePreds[i].first);
999 // Add all the unavailable predecessors to the PredsToSplit list.
1000 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
1002 BasicBlock *P = *PI;
1003 // If the predecessor is an indirect goto, we can't split the edge.
1004 if (isa<IndirectBrInst>(P->getTerminator()))
1007 if (!AvailablePredSet.count(P))
1008 PredsToSplit.push_back(P);
1011 // Split them out to their own block.
1012 UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1015 // If the value isn't available in all predecessors, then there will be
1016 // exactly one where it isn't available. Insert a load on that edge and add
1017 // it to the AvailablePreds list.
1018 if (UnavailablePred) {
1019 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1020 "Can't handle critical edge here!");
1021 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
1023 UnavailablePred->getTerminator());
1024 NewVal->setDebugLoc(LI->getDebugLoc());
1026 NewVal->setAAMetadata(AATags);
1028 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1031 // Now we know that each predecessor of this block has a value in
1032 // AvailablePreds, sort them for efficient access as we're walking the preds.
1033 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1035 // Create a PHI node at the start of the block for the PRE'd load value.
1036 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1037 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
1040 PN->setDebugLoc(LI->getDebugLoc());
1042 // Insert new entries into the PHI for each predecessor. A single block may
1043 // have multiple entries here.
1044 for (pred_iterator PI = PB; PI != PE; ++PI) {
1045 BasicBlock *P = *PI;
1046 AvailablePredsTy::iterator I =
1047 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1048 std::make_pair(P, (Value*)nullptr));
1050 assert(I != AvailablePreds.end() && I->first == P &&
1051 "Didn't find entry for predecessor!");
1053 // If we have an available predecessor but it requires casting, insert the
1054 // cast in the predecessor and use the cast. Note that we have to update the
1055 // AvailablePreds vector as we go so that all of the PHI entries for this
1056 // predecessor use the same bitcast.
1057 Value *&PredV = I->second;
1058 if (PredV->getType() != LI->getType())
1059 PredV = CastInst::CreateBitOrPointerCast(PredV, LI->getType(), "",
1060 P->getTerminator());
1062 PN->addIncoming(PredV, I->first);
1065 //cerr << "PRE: " << *LI << *PN << "\n";
1067 LI->replaceAllUsesWith(PN);
1068 LI->eraseFromParent();
1073 /// FindMostPopularDest - The specified list contains multiple possible
1074 /// threadable destinations. Pick the one that occurs the most frequently in
1077 FindMostPopularDest(BasicBlock *BB,
1078 const SmallVectorImpl<std::pair<BasicBlock*,
1079 BasicBlock*> > &PredToDestList) {
1080 assert(!PredToDestList.empty());
1082 // Determine popularity. If there are multiple possible destinations, we
1083 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1084 // blocks with known and real destinations to threading undef. We'll handle
1085 // them later if interesting.
1086 DenseMap<BasicBlock*, unsigned> DestPopularity;
1087 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1088 if (PredToDestList[i].second)
1089 DestPopularity[PredToDestList[i].second]++;
1091 // Find the most popular dest.
1092 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1093 BasicBlock *MostPopularDest = DPI->first;
1094 unsigned Popularity = DPI->second;
1095 SmallVector<BasicBlock*, 4> SamePopularity;
1097 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1098 // If the popularity of this entry isn't higher than the popularity we've
1099 // seen so far, ignore it.
1100 if (DPI->second < Popularity)
1102 else if (DPI->second == Popularity) {
1103 // If it is the same as what we've seen so far, keep track of it.
1104 SamePopularity.push_back(DPI->first);
1106 // If it is more popular, remember it.
1107 SamePopularity.clear();
1108 MostPopularDest = DPI->first;
1109 Popularity = DPI->second;
1113 // Okay, now we know the most popular destination. If there is more than one
1114 // destination, we need to determine one. This is arbitrary, but we need
1115 // to make a deterministic decision. Pick the first one that appears in the
1117 if (!SamePopularity.empty()) {
1118 SamePopularity.push_back(MostPopularDest);
1119 TerminatorInst *TI = BB->getTerminator();
1120 for (unsigned i = 0; ; ++i) {
1121 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1123 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1124 TI->getSuccessor(i)) == SamePopularity.end())
1127 MostPopularDest = TI->getSuccessor(i);
1132 // Okay, we have finally picked the most popular destination.
1133 return MostPopularDest;
1136 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1137 ConstantPreference Preference,
1138 Instruction *CxtI) {
1139 // If threading this would thread across a loop header, don't even try to
1141 if (LoopHeaders.count(BB))
1144 PredValueInfoTy PredValues;
1145 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1148 assert(!PredValues.empty() &&
1149 "ComputeValueKnownInPredecessors returned true with no values");
1151 DEBUG(dbgs() << "IN BB: " << *BB;
1152 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1153 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1154 << *PredValues[i].first
1155 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1158 // Decide what we want to thread through. Convert our list of known values to
1159 // a list of known destinations for each pred. This also discards duplicate
1160 // predecessors and keeps track of the undefined inputs (which are represented
1161 // as a null dest in the PredToDestList).
1162 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1163 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1165 BasicBlock *OnlyDest = nullptr;
1166 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1168 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1169 BasicBlock *Pred = PredValues[i].second;
1170 if (!SeenPreds.insert(Pred).second)
1171 continue; // Duplicate predecessor entry.
1173 // If the predecessor ends with an indirect goto, we can't change its
1175 if (isa<IndirectBrInst>(Pred->getTerminator()))
1178 Constant *Val = PredValues[i].first;
1181 if (isa<UndefValue>(Val))
1183 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1184 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1185 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1186 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1188 assert(isa<IndirectBrInst>(BB->getTerminator())
1189 && "Unexpected terminator");
1190 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1193 // If we have exactly one destination, remember it for efficiency below.
1194 if (PredToDestList.empty())
1196 else if (OnlyDest != DestBB)
1197 OnlyDest = MultipleDestSentinel;
1199 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1202 // If all edges were unthreadable, we fail.
1203 if (PredToDestList.empty())
1206 // Determine which is the most common successor. If we have many inputs and
1207 // this block is a switch, we want to start by threading the batch that goes
1208 // to the most popular destination first. If we only know about one
1209 // threadable destination (the common case) we can avoid this.
1210 BasicBlock *MostPopularDest = OnlyDest;
1212 if (MostPopularDest == MultipleDestSentinel)
1213 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1215 // Now that we know what the most popular destination is, factor all
1216 // predecessors that will jump to it into a single predecessor.
1217 SmallVector<BasicBlock*, 16> PredsToFactor;
1218 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1219 if (PredToDestList[i].second == MostPopularDest) {
1220 BasicBlock *Pred = PredToDestList[i].first;
1222 // This predecessor may be a switch or something else that has multiple
1223 // edges to the block. Factor each of these edges by listing them
1224 // according to # occurrences in PredsToFactor.
1225 TerminatorInst *PredTI = Pred->getTerminator();
1226 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1227 if (PredTI->getSuccessor(i) == BB)
1228 PredsToFactor.push_back(Pred);
1231 // If the threadable edges are branching on an undefined value, we get to pick
1232 // the destination that these predecessors should get to.
1233 if (!MostPopularDest)
1234 MostPopularDest = BB->getTerminator()->
1235 getSuccessor(GetBestDestForJumpOnUndef(BB));
1237 // Ok, try to thread it!
1238 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1241 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1242 /// a PHI node in the current block. See if there are any simplifications we
1243 /// can do based on inputs to the phi node.
1245 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1246 BasicBlock *BB = PN->getParent();
1248 // TODO: We could make use of this to do it once for blocks with common PHI
1250 SmallVector<BasicBlock*, 1> PredBBs;
1253 // If any of the predecessor blocks end in an unconditional branch, we can
1254 // *duplicate* the conditional branch into that block in order to further
1255 // encourage jump threading and to eliminate cases where we have branch on a
1256 // phi of an icmp (branch on icmp is much better).
1257 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1258 BasicBlock *PredBB = PN->getIncomingBlock(i);
1259 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1260 if (PredBr->isUnconditional()) {
1261 PredBBs[0] = PredBB;
1262 // Try to duplicate BB into PredBB.
1263 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1271 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1272 /// a xor instruction in the current block. See if there are any
1273 /// simplifications we can do based on inputs to the xor.
1275 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1276 BasicBlock *BB = BO->getParent();
1278 // If either the LHS or RHS of the xor is a constant, don't do this
1280 if (isa<ConstantInt>(BO->getOperand(0)) ||
1281 isa<ConstantInt>(BO->getOperand(1)))
1284 // If the first instruction in BB isn't a phi, we won't be able to infer
1285 // anything special about any particular predecessor.
1286 if (!isa<PHINode>(BB->front()))
1289 // If we have a xor as the branch input to this block, and we know that the
1290 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1291 // the condition into the predecessor and fix that value to true, saving some
1292 // logical ops on that path and encouraging other paths to simplify.
1294 // This copies something like this:
1297 // %X = phi i1 [1], [%X']
1298 // %Y = icmp eq i32 %A, %B
1299 // %Z = xor i1 %X, %Y
1304 // %Y = icmp ne i32 %A, %B
1307 PredValueInfoTy XorOpValues;
1309 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1311 assert(XorOpValues.empty());
1312 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1318 assert(!XorOpValues.empty() &&
1319 "ComputeValueKnownInPredecessors returned true with no values");
1321 // Scan the information to see which is most popular: true or false. The
1322 // predecessors can be of the set true, false, or undef.
1323 unsigned NumTrue = 0, NumFalse = 0;
1324 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1325 if (isa<UndefValue>(XorOpValues[i].first))
1326 // Ignore undefs for the count.
1328 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1334 // Determine which value to split on, true, false, or undef if neither.
1335 ConstantInt *SplitVal = nullptr;
1336 if (NumTrue > NumFalse)
1337 SplitVal = ConstantInt::getTrue(BB->getContext());
1338 else if (NumTrue != 0 || NumFalse != 0)
1339 SplitVal = ConstantInt::getFalse(BB->getContext());
1341 // Collect all of the blocks that this can be folded into so that we can
1342 // factor this once and clone it once.
1343 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1344 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1345 if (XorOpValues[i].first != SplitVal &&
1346 !isa<UndefValue>(XorOpValues[i].first))
1349 BlocksToFoldInto.push_back(XorOpValues[i].second);
1352 // If we inferred a value for all of the predecessors, then duplication won't
1353 // help us. However, we can just replace the LHS or RHS with the constant.
1354 if (BlocksToFoldInto.size() ==
1355 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1357 // If all preds provide undef, just nuke the xor, because it is undef too.
1358 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1359 BO->eraseFromParent();
1360 } else if (SplitVal->isZero()) {
1361 // If all preds provide 0, replace the xor with the other input.
1362 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1363 BO->eraseFromParent();
1365 // If all preds provide 1, set the computed value to 1.
1366 BO->setOperand(!isLHS, SplitVal);
1372 // Try to duplicate BB into PredBB.
1373 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1377 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1378 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1379 /// NewPred using the entries from OldPred (suitably mapped).
1380 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1381 BasicBlock *OldPred,
1382 BasicBlock *NewPred,
1383 DenseMap<Instruction*, Value*> &ValueMap) {
1384 for (BasicBlock::iterator PNI = PHIBB->begin();
1385 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1386 // Ok, we have a PHI node. Figure out what the incoming value was for the
1388 Value *IV = PN->getIncomingValueForBlock(OldPred);
1390 // Remap the value if necessary.
1391 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1392 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1393 if (I != ValueMap.end())
1397 PN->addIncoming(IV, NewPred);
1401 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1402 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1403 /// across BB. Transform the IR to reflect this change.
1404 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1405 const SmallVectorImpl<BasicBlock*> &PredBBs,
1406 BasicBlock *SuccBB) {
1407 // If threading to the same block as we come from, we would infinite loop.
1409 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1410 << "' - would thread to self!\n");
1414 // If threading this would thread across a loop header, don't thread the edge.
1415 // See the comments above FindLoopHeaders for justifications and caveats.
1416 if (LoopHeaders.count(BB)) {
1417 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1418 << "' to dest BB '" << SuccBB->getName()
1419 << "' - it might create an irreducible loop!\n");
1423 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, BBDupThreshold);
1424 if (JumpThreadCost > BBDupThreshold) {
1425 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1426 << "' - Cost is too high: " << JumpThreadCost << "\n");
1430 // And finally, do it! Start by factoring the predecessors if needed.
1432 if (PredBBs.size() == 1)
1433 PredBB = PredBBs[0];
1435 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1436 << " common predecessors.\n");
1437 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1440 // And finally, do it!
1441 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1442 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1443 << ", across block:\n "
1446 LVI->threadEdge(PredBB, BB, SuccBB);
1448 // We are going to have to map operands from the original BB block to the new
1449 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1450 // account for entry from PredBB.
1451 DenseMap<Instruction*, Value*> ValueMapping;
1453 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1454 BB->getName()+".thread",
1455 BB->getParent(), BB);
1456 NewBB->moveAfter(PredBB);
1458 // Set the block frequency of NewBB.
1459 if (HasProfileData) {
1461 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
1462 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
1465 BasicBlock::iterator BI = BB->begin();
1466 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1467 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1469 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1470 // mapping and using it to remap operands in the cloned instructions.
1471 for (; !isa<TerminatorInst>(BI); ++BI) {
1472 Instruction *New = BI->clone();
1473 New->setName(BI->getName());
1474 NewBB->getInstList().push_back(New);
1475 ValueMapping[&*BI] = New;
1477 // Remap operands to patch up intra-block references.
1478 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1479 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1480 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1481 if (I != ValueMapping.end())
1482 New->setOperand(i, I->second);
1486 // We didn't copy the terminator from BB over to NewBB, because there is now
1487 // an unconditional jump to SuccBB. Insert the unconditional jump.
1488 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
1489 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1491 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1492 // PHI nodes for NewBB now.
1493 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1495 // If there were values defined in BB that are used outside the block, then we
1496 // now have to update all uses of the value to use either the original value,
1497 // the cloned value, or some PHI derived value. This can require arbitrary
1498 // PHI insertion, of which we are prepared to do, clean these up now.
1499 SSAUpdater SSAUpdate;
1500 SmallVector<Use*, 16> UsesToRename;
1501 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1502 // Scan all uses of this instruction to see if it is used outside of its
1503 // block, and if so, record them in UsesToRename.
1504 for (Use &U : I->uses()) {
1505 Instruction *User = cast<Instruction>(U.getUser());
1506 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1507 if (UserPN->getIncomingBlock(U) == BB)
1509 } else if (User->getParent() == BB)
1512 UsesToRename.push_back(&U);
1515 // If there are no uses outside the block, we're done with this instruction.
1516 if (UsesToRename.empty())
1519 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1521 // We found a use of I outside of BB. Rename all uses of I that are outside
1522 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1523 // with the two values we know.
1524 SSAUpdate.Initialize(I->getType(), I->getName());
1525 SSAUpdate.AddAvailableValue(BB, &*I);
1526 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&*I]);
1528 while (!UsesToRename.empty())
1529 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1530 DEBUG(dbgs() << "\n");
1534 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1535 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1536 // us to simplify any PHI nodes in BB.
1537 TerminatorInst *PredTerm = PredBB->getTerminator();
1538 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1539 if (PredTerm->getSuccessor(i) == BB) {
1540 BB->removePredecessor(PredBB, true);
1541 PredTerm->setSuccessor(i, NewBB);
1544 // At this point, the IR is fully up to date and consistent. Do a quick scan
1545 // over the new instructions and zap any that are constants or dead. This
1546 // frequently happens because of phi translation.
1547 SimplifyInstructionsInBlock(NewBB, TLI);
1549 // Update the edge weight from BB to SuccBB, which should be less than before.
1550 UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
1552 // Threaded an edge!
1557 /// Create a new basic block that will be the predecessor of BB and successor of
1558 /// all blocks in Preds. When profile data is availble, update the frequency of
1560 BasicBlock *JumpThreading::SplitBlockPreds(BasicBlock *BB,
1561 ArrayRef<BasicBlock *> Preds,
1562 const char *Suffix) {
1563 // Collect the frequencies of all predecessors of BB, which will be used to
1564 // update the edge weight on BB->SuccBB.
1565 BlockFrequency PredBBFreq(0);
1567 for (auto Pred : Preds)
1568 PredBBFreq += BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB);
1570 BasicBlock *PredBB = SplitBlockPredecessors(BB, Preds, Suffix);
1572 // Set the block frequency of the newly created PredBB, which is the sum of
1573 // frequencies of Preds.
1575 BFI->setBlockFreq(PredBB, PredBBFreq.getFrequency());
1579 /// Update the block frequency of BB and branch weight and the metadata on the
1580 /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
1581 /// Freq(PredBB->BB) / Freq(BB->SuccBB).
1582 void JumpThreading::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
1585 BasicBlock *SuccBB) {
1586 if (!HasProfileData)
1589 assert(BFI && BPI && "BFI & BPI should have been created here");
1591 // As the edge from PredBB to BB is deleted, we have to update the block
1593 auto BBOrigFreq = BFI->getBlockFreq(BB);
1594 auto NewBBFreq = BFI->getBlockFreq(NewBB);
1595 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
1596 auto BBNewFreq = BBOrigFreq - NewBBFreq;
1597 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
1599 // Collect updated outgoing edges' frequencies from BB and use them to update
1601 SmallVector<uint64_t, 4> BBSuccFreq;
1602 for (auto I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
1603 auto SuccFreq = (*I == SuccBB)
1604 ? BB2SuccBBFreq - NewBBFreq
1605 : BBOrigFreq * BPI->getEdgeProbability(BB, *I);
1606 BBSuccFreq.push_back(SuccFreq.getFrequency());
1609 // Normalize edge weights in Weights64 so that the sum of them can fit in
1610 MachineBranchProbabilityInfo::normalizeEdgeWeights(BBSuccFreq.begin(),
1613 SmallVector<uint32_t, 4> Weights;
1614 for (auto Freq : BBSuccFreq)
1615 Weights.push_back(static_cast<uint32_t>(Freq));
1617 // Update edge weights in BPI.
1618 for (int I = 0, E = Weights.size(); I < E; I++)
1619 BPI->setEdgeWeight(BB, I, Weights[I]);
1621 if (Weights.size() >= 2) {
1622 auto TI = BB->getTerminator();
1624 LLVMContext::MD_prof,
1625 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
1629 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1630 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1631 /// If we can duplicate the contents of BB up into PredBB do so now, this
1632 /// improves the odds that the branch will be on an analyzable instruction like
1634 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1635 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1636 assert(!PredBBs.empty() && "Can't handle an empty set");
1638 // If BB is a loop header, then duplicating this block outside the loop would
1639 // cause us to transform this into an irreducible loop, don't do this.
1640 // See the comments above FindLoopHeaders for justifications and caveats.
1641 if (LoopHeaders.count(BB)) {
1642 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1643 << "' into predecessor block '" << PredBBs[0]->getName()
1644 << "' - it might create an irreducible loop!\n");
1648 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, BBDupThreshold);
1649 if (DuplicationCost > BBDupThreshold) {
1650 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1651 << "' - Cost is too high: " << DuplicationCost << "\n");
1655 // And finally, do it! Start by factoring the predecessors if needed.
1657 if (PredBBs.size() == 1)
1658 PredBB = PredBBs[0];
1660 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1661 << " common predecessors.\n");
1662 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1665 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1667 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1668 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1669 << DuplicationCost << " block is:" << *BB << "\n");
1671 // Unless PredBB ends with an unconditional branch, split the edge so that we
1672 // can just clone the bits from BB into the end of the new PredBB.
1673 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1675 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
1676 PredBB = SplitEdge(PredBB, BB);
1677 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1680 // We are going to have to map operands from the original BB block into the
1681 // PredBB block. Evaluate PHI nodes in BB.
1682 DenseMap<Instruction*, Value*> ValueMapping;
1684 BasicBlock::iterator BI = BB->begin();
1685 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1686 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1687 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1688 // mapping and using it to remap operands in the cloned instructions.
1689 for (; BI != BB->end(); ++BI) {
1690 Instruction *New = BI->clone();
1692 // Remap operands to patch up intra-block references.
1693 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1694 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1695 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1696 if (I != ValueMapping.end())
1697 New->setOperand(i, I->second);
1700 // If this instruction can be simplified after the operands are updated,
1701 // just use the simplified value instead. This frequently happens due to
1704 SimplifyInstruction(New, BB->getModule()->getDataLayout())) {
1706 ValueMapping[&*BI] = IV;
1708 // Otherwise, insert the new instruction into the block.
1709 New->setName(BI->getName());
1710 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
1711 ValueMapping[&*BI] = New;
1715 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1716 // add entries to the PHI nodes for branch from PredBB now.
1717 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1718 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1720 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1723 // If there were values defined in BB that are used outside the block, then we
1724 // now have to update all uses of the value to use either the original value,
1725 // the cloned value, or some PHI derived value. This can require arbitrary
1726 // PHI insertion, of which we are prepared to do, clean these up now.
1727 SSAUpdater SSAUpdate;
1728 SmallVector<Use*, 16> UsesToRename;
1729 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1730 // Scan all uses of this instruction to see if it is used outside of its
1731 // block, and if so, record them in UsesToRename.
1732 for (Use &U : I->uses()) {
1733 Instruction *User = cast<Instruction>(U.getUser());
1734 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1735 if (UserPN->getIncomingBlock(U) == BB)
1737 } else if (User->getParent() == BB)
1740 UsesToRename.push_back(&U);
1743 // If there are no uses outside the block, we're done with this instruction.
1744 if (UsesToRename.empty())
1747 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1749 // We found a use of I outside of BB. Rename all uses of I that are outside
1750 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1751 // with the two values we know.
1752 SSAUpdate.Initialize(I->getType(), I->getName());
1753 SSAUpdate.AddAvailableValue(BB, &*I);
1754 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&*I]);
1756 while (!UsesToRename.empty())
1757 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1758 DEBUG(dbgs() << "\n");
1761 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1763 BB->removePredecessor(PredBB, true);
1765 // Remove the unconditional branch at the end of the PredBB block.
1766 OldPredBranch->eraseFromParent();
1772 /// TryToUnfoldSelect - Look for blocks of the form
1778 /// %p = phi [%a, %bb] ...
1782 /// And expand the select into a branch structure if one of its arms allows %c
1783 /// to be folded. This later enables threading from bb1 over bb2.
1784 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1785 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1786 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1787 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1789 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1790 CondLHS->getParent() != BB)
1793 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1794 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1795 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1797 // Look if one of the incoming values is a select in the corresponding
1799 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1802 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1803 if (!PredTerm || !PredTerm->isUnconditional())
1806 // Now check if one of the select values would allow us to constant fold the
1807 // terminator in BB. We don't do the transform if both sides fold, those
1808 // cases will be threaded in any case.
1809 LazyValueInfo::Tristate LHSFolds =
1810 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1811 CondRHS, Pred, BB, CondCmp);
1812 LazyValueInfo::Tristate RHSFolds =
1813 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
1814 CondRHS, Pred, BB, CondCmp);
1815 if ((LHSFolds != LazyValueInfo::Unknown ||
1816 RHSFolds != LazyValueInfo::Unknown) &&
1817 LHSFolds != RHSFolds) {
1818 // Expand the select.
1827 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
1828 BB->getParent(), BB);
1829 // Move the unconditional branch to NewBB.
1830 PredTerm->removeFromParent();
1831 NewBB->getInstList().insert(NewBB->end(), PredTerm);
1832 // Create a conditional branch and update PHI nodes.
1833 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
1834 CondLHS->setIncomingValue(I, SI->getFalseValue());
1835 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
1836 // The select is now dead.
1837 SI->eraseFromParent();
1839 // Update any other PHI nodes in BB.
1840 for (BasicBlock::iterator BI = BB->begin();
1841 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
1843 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);