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/Analysis/ValueTracking.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);
61 static cl::opt<unsigned>
62 ImplicationSearchThreshold(
63 "jump-threading-implication-search-threshold",
64 cl::desc("The number of predecessors to search for a stronger "
65 "condition to use to thread over a weaker condition"),
66 cl::init(3), cl::Hidden);
69 // These are at global scope so static functions can use them too.
70 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo;
71 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy;
73 // This is used to keep track of what kind of constant we're currently hoping
75 enum ConstantPreference {
80 /// This pass performs 'jump threading', which looks at blocks that have
81 /// multiple predecessors and multiple successors. If one or more of the
82 /// predecessors of the block can be proven to always jump to one of the
83 /// successors, we forward the edge from the predecessor to the successor by
84 /// duplicating the contents of this block.
86 /// An example of when this can occur is code like this:
93 /// In this case, the unconditional branch at the end of the first if can be
94 /// revectored to the false side of the second if.
96 class JumpThreading : public FunctionPass {
97 TargetLibraryInfo *TLI;
99 std::unique_ptr<BlockFrequencyInfo> BFI;
100 std::unique_ptr<BranchProbabilityInfo> BPI;
103 SmallPtrSet<const BasicBlock *, 16> LoopHeaders;
105 SmallSet<AssertingVH<const BasicBlock>, 16> LoopHeaders;
107 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet;
109 unsigned BBDupThreshold;
111 // RAII helper for updating the recursion stack.
112 struct RecursionSetRemover {
113 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet;
114 std::pair<Value*, BasicBlock*> ThePair;
116 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S,
117 std::pair<Value*, BasicBlock*> P)
118 : TheSet(S), ThePair(P) { }
120 ~RecursionSetRemover() {
121 TheSet.erase(ThePair);
125 static char ID; // Pass identification
126 JumpThreading(int T = -1) : FunctionPass(ID) {
127 BBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
128 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
131 bool runOnFunction(Function &F) override;
133 void getAnalysisUsage(AnalysisUsage &AU) const override {
134 AU.addRequired<LazyValueInfo>();
135 AU.addPreserved<LazyValueInfo>();
136 AU.addPreserved<GlobalsAAWrapperPass>();
137 AU.addRequired<TargetLibraryInfoWrapperPass>();
140 void releaseMemory() override {
145 void FindLoopHeaders(Function &F);
146 bool ProcessBlock(BasicBlock *BB);
147 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
149 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
150 const SmallVectorImpl<BasicBlock *> &PredBBs);
152 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
153 PredValueInfo &Result,
154 ConstantPreference Preference,
155 Instruction *CxtI = nullptr);
156 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
157 ConstantPreference Preference,
158 Instruction *CxtI = nullptr);
160 bool ProcessBranchOnPHI(PHINode *PN);
161 bool ProcessBranchOnXOR(BinaryOperator *BO);
162 bool ProcessImpliedCondition(BasicBlock *BB);
164 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
165 bool TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB);
166 bool TryToUnfoldSelectInCurrBB(BasicBlock *BB);
169 BasicBlock *SplitBlockPreds(BasicBlock *BB, ArrayRef<BasicBlock *> Preds,
171 void UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB, BasicBlock *BB,
172 BasicBlock *NewBB, BasicBlock *SuccBB);
176 char JumpThreading::ID = 0;
177 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
178 "Jump Threading", false, false)
179 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)
180 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
181 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
182 "Jump Threading", false, false)
184 // Public interface to the Jump Threading pass
185 FunctionPass *llvm::createJumpThreadingPass(int Threshold) { return new JumpThreading(Threshold); }
187 /// runOnFunction - Top level algorithm.
189 bool JumpThreading::runOnFunction(Function &F) {
190 if (skipOptnoneFunction(F))
193 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
194 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
195 LVI = &getAnalysis<LazyValueInfo>();
198 // When profile data is available, we need to update edge weights after
199 // successful jump threading, which requires both BPI and BFI being available.
200 HasProfileData = F.getEntryCount().hasValue();
201 if (HasProfileData) {
202 LoopInfo LI{DominatorTree(F)};
203 BPI.reset(new BranchProbabilityInfo(F, LI));
204 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
207 // Remove unreachable blocks from function as they may result in infinite
208 // loop. We do threading if we found something profitable. Jump threading a
209 // branch can create other opportunities. If these opportunities form a cycle
210 // i.e. if any jump threading is undoing previous threading in the path, then
211 // we will loop forever. We take care of this issue by not jump threading for
212 // back edges. This works for normal cases but not for unreachable blocks as
213 // they may have cycle with no back edge.
214 removeUnreachableBlocks(F);
218 bool Changed, EverChanged = false;
221 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
222 BasicBlock *BB = &*I;
223 // Thread all of the branches we can over this block.
224 while (ProcessBlock(BB))
229 // If the block is trivially dead, zap it. This eliminates the successor
230 // edges which simplifies the CFG.
231 if (pred_empty(BB) &&
232 BB != &BB->getParent()->getEntryBlock()) {
233 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
234 << "' with terminator: " << *BB->getTerminator() << '\n');
235 LoopHeaders.erase(BB);
242 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
244 // Can't thread an unconditional jump, but if the block is "almost
245 // empty", we can replace uses of it with uses of the successor and make
247 if (BI && BI->isUnconditional() &&
248 BB != &BB->getParent()->getEntryBlock() &&
249 // If the terminator is the only non-phi instruction, try to nuke it.
250 BB->getFirstNonPHIOrDbg()->isTerminator()) {
251 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
252 // block, we have to make sure it isn't in the LoopHeaders set. We
253 // reinsert afterward if needed.
254 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
255 BasicBlock *Succ = BI->getSuccessor(0);
257 // FIXME: It is always conservatively correct to drop the info
258 // for a block even if it doesn't get erased. This isn't totally
259 // awesome, but it allows us to use AssertingVH to prevent nasty
260 // dangling pointer issues within LazyValueInfo.
262 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
264 // If we deleted BB and BB was the header of a loop, then the
265 // successor is now the header of the loop.
269 if (ErasedFromLoopHeaders)
270 LoopHeaders.insert(BB);
273 EverChanged |= Changed;
280 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
281 /// thread across it. Stop scanning the block when passing the threshold.
282 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
283 unsigned Threshold) {
284 /// Ignore PHI nodes, these will be flattened when duplication happens.
285 BasicBlock::const_iterator I(BB->getFirstNonPHI());
287 // FIXME: THREADING will delete values that are just used to compute the
288 // branch, so they shouldn't count against the duplication cost.
291 const TerminatorInst *BBTerm = BB->getTerminator();
292 // Threading through a switch statement is particularly profitable. If this
293 // block ends in a switch, decrease its cost to make it more likely to happen.
294 if (isa<SwitchInst>(BBTerm))
297 // The same holds for indirect branches, but slightly more so.
298 if (isa<IndirectBrInst>(BBTerm))
301 // Bump the threshold up so the early exit from the loop doesn't skip the
302 // terminator-based Size adjustment at the end.
305 // Sum up the cost of each instruction until we get to the terminator. Don't
306 // include the terminator because the copy won't include it.
308 for (; !isa<TerminatorInst>(I); ++I) {
310 // Stop scanning the block if we've reached the threshold.
311 if (Size > Threshold)
314 // Debugger intrinsics don't incur code size.
315 if (isa<DbgInfoIntrinsic>(I)) continue;
317 // If this is a pointer->pointer bitcast, it is free.
318 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
321 // Bail out if this instruction gives back a token type, it is not possible
322 // to duplicate it if it is used outside this BB.
323 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
326 // All other instructions count for at least one unit.
329 // Calls are more expensive. If they are non-intrinsic calls, we model them
330 // as having cost of 4. If they are a non-vector intrinsic, we model them
331 // as having cost of 2 total, and if they are a vector intrinsic, we model
332 // them as having cost 1.
333 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
334 if (CI->cannotDuplicate() || CI->isConvergent())
335 // Blocks with NoDuplicate are modelled as having infinite cost, so they
336 // are never duplicated.
338 else if (!isa<IntrinsicInst>(CI))
340 else if (!CI->getType()->isVectorTy())
345 return Size > Bonus ? Size - Bonus : 0;
348 /// FindLoopHeaders - We do not want jump threading to turn proper loop
349 /// structures into irreducible loops. Doing this breaks up the loop nesting
350 /// hierarchy and pessimizes later transformations. To prevent this from
351 /// happening, we first have to find the loop headers. Here we approximate this
352 /// by finding targets of backedges in the CFG.
354 /// Note that there definitely are cases when we want to allow threading of
355 /// edges across a loop header. For example, threading a jump from outside the
356 /// loop (the preheader) to an exit block of the loop is definitely profitable.
357 /// It is also almost always profitable to thread backedges from within the loop
358 /// to exit blocks, and is often profitable to thread backedges to other blocks
359 /// within the loop (forming a nested loop). This simple analysis is not rich
360 /// enough to track all of these properties and keep it up-to-date as the CFG
361 /// mutates, so we don't allow any of these transformations.
363 void JumpThreading::FindLoopHeaders(Function &F) {
364 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
365 FindFunctionBackedges(F, Edges);
367 for (const auto &Edge : Edges)
368 LoopHeaders.insert(Edge.second);
371 /// getKnownConstant - Helper method to determine if we can thread over a
372 /// terminator with the given value as its condition, and if so what value to
373 /// use for that. What kind of value this is depends on whether we want an
374 /// integer or a block address, but an undef is always accepted.
375 /// Returns null if Val is null or not an appropriate constant.
376 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
380 // Undef is "known" enough.
381 if (UndefValue *U = dyn_cast<UndefValue>(Val))
384 if (Preference == WantBlockAddress)
385 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
387 return dyn_cast<ConstantInt>(Val);
390 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
391 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
392 /// in any of our predecessors. If so, return the known list of value and pred
393 /// BB in the result vector.
395 /// This returns true if there were any known values.
398 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
399 ConstantPreference Preference,
401 // This method walks up use-def chains recursively. Because of this, we could
402 // get into an infinite loop going around loops in the use-def chain. To
403 // prevent this, keep track of what (value, block) pairs we've already visited
404 // and terminate the search if we loop back to them
405 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
408 // An RAII help to remove this pair from the recursion set once the recursion
409 // stack pops back out again.
410 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
412 // If V is a constant, then it is known in all predecessors.
413 if (Constant *KC = getKnownConstant(V, Preference)) {
414 for (BasicBlock *Pred : predecessors(BB))
415 Result.push_back(std::make_pair(KC, Pred));
420 // If V is a non-instruction value, or an instruction in a different block,
421 // then it can't be derived from a PHI.
422 Instruction *I = dyn_cast<Instruction>(V);
423 if (!I || I->getParent() != BB) {
425 // Okay, if this is a live-in value, see if it has a known value at the end
426 // of any of our predecessors.
428 // FIXME: This should be an edge property, not a block end property.
429 /// TODO: Per PR2563, we could infer value range information about a
430 /// predecessor based on its terminator.
432 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
433 // "I" is a non-local compare-with-a-constant instruction. This would be
434 // able to handle value inequalities better, for example if the compare is
435 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
436 // Perhaps getConstantOnEdge should be smart enough to do this?
438 for (BasicBlock *P : predecessors(BB)) {
439 // If the value is known by LazyValueInfo to be a constant in a
440 // predecessor, use that information to try to thread this block.
441 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
442 if (Constant *KC = getKnownConstant(PredCst, Preference))
443 Result.push_back(std::make_pair(KC, P));
446 return !Result.empty();
449 /// If I is a PHI node, then we know the incoming values for any constants.
450 if (PHINode *PN = dyn_cast<PHINode>(I)) {
451 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
452 Value *InVal = PN->getIncomingValue(i);
453 if (Constant *KC = getKnownConstant(InVal, Preference)) {
454 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
456 Constant *CI = LVI->getConstantOnEdge(InVal,
457 PN->getIncomingBlock(i),
459 if (Constant *KC = getKnownConstant(CI, Preference))
460 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
464 return !Result.empty();
467 PredValueInfoTy LHSVals, RHSVals;
469 // Handle some boolean conditions.
470 if (I->getType()->getPrimitiveSizeInBits() == 1) {
471 assert(Preference == WantInteger && "One-bit non-integer type?");
473 // X & false -> false
474 if (I->getOpcode() == Instruction::Or ||
475 I->getOpcode() == Instruction::And) {
476 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
478 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
481 if (LHSVals.empty() && RHSVals.empty())
484 ConstantInt *InterestingVal;
485 if (I->getOpcode() == Instruction::Or)
486 InterestingVal = ConstantInt::getTrue(I->getContext());
488 InterestingVal = ConstantInt::getFalse(I->getContext());
490 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
492 // Scan for the sentinel. If we find an undef, force it to the
493 // interesting value: x|undef -> true and x&undef -> false.
494 for (const auto &LHSVal : LHSVals)
495 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
496 Result.emplace_back(InterestingVal, LHSVal.second);
497 LHSKnownBBs.insert(LHSVal.second);
499 for (const auto &RHSVal : RHSVals)
500 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
501 // If we already inferred a value for this block on the LHS, don't
503 if (!LHSKnownBBs.count(RHSVal.second))
504 Result.emplace_back(InterestingVal, RHSVal.second);
507 return !Result.empty();
510 // Handle the NOT form of XOR.
511 if (I->getOpcode() == Instruction::Xor &&
512 isa<ConstantInt>(I->getOperand(1)) &&
513 cast<ConstantInt>(I->getOperand(1))->isOne()) {
514 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
519 // Invert the known values.
520 for (auto &R : Result)
521 R.first = ConstantExpr::getNot(R.first);
526 // Try to simplify some other binary operator values.
527 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
528 assert(Preference != WantBlockAddress
529 && "A binary operator creating a block address?");
530 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
531 PredValueInfoTy LHSVals;
532 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
535 // Try to use constant folding to simplify the binary operator.
536 for (const auto &LHSVal : LHSVals) {
537 Constant *V = LHSVal.first;
538 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
540 if (Constant *KC = getKnownConstant(Folded, WantInteger))
541 Result.push_back(std::make_pair(KC, LHSVal.second));
545 return !Result.empty();
548 // Handle compare with phi operand, where the PHI is defined in this block.
549 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
550 assert(Preference == WantInteger && "Compares only produce integers");
551 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
552 if (PN && PN->getParent() == BB) {
553 const DataLayout &DL = PN->getModule()->getDataLayout();
554 // We can do this simplification if any comparisons fold to true or false.
556 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
557 BasicBlock *PredBB = PN->getIncomingBlock(i);
558 Value *LHS = PN->getIncomingValue(i);
559 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
561 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL);
563 if (!isa<Constant>(RHS))
566 LazyValueInfo::Tristate
567 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
568 cast<Constant>(RHS), PredBB, BB,
570 if (ResT == LazyValueInfo::Unknown)
572 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
575 if (Constant *KC = getKnownConstant(Res, WantInteger))
576 Result.push_back(std::make_pair(KC, PredBB));
579 return !Result.empty();
582 // If comparing a live-in value against a constant, see if we know the
583 // live-in value on any predecessors.
584 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
585 if (!isa<Instruction>(Cmp->getOperand(0)) ||
586 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
587 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
589 for (BasicBlock *P : predecessors(BB)) {
590 // If the value is known by LazyValueInfo to be a constant in a
591 // predecessor, use that information to try to thread this block.
592 LazyValueInfo::Tristate Res =
593 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
594 RHSCst, P, BB, CxtI ? CxtI : Cmp);
595 if (Res == LazyValueInfo::Unknown)
598 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
599 Result.push_back(std::make_pair(ResC, P));
602 return !Result.empty();
605 // Try to find a constant value for the LHS of a comparison,
606 // and evaluate it statically if we can.
607 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
608 PredValueInfoTy LHSVals;
609 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
612 for (const auto &LHSVal : LHSVals) {
613 Constant *V = LHSVal.first;
614 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
616 if (Constant *KC = getKnownConstant(Folded, WantInteger))
617 Result.push_back(std::make_pair(KC, LHSVal.second));
620 return !Result.empty();
625 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
626 // Handle select instructions where at least one operand is a known constant
627 // and we can figure out the condition value for any predecessor block.
628 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
629 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
630 PredValueInfoTy Conds;
631 if ((TrueVal || FalseVal) &&
632 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
633 WantInteger, CxtI)) {
634 for (auto &C : Conds) {
635 Constant *Cond = C.first;
637 // Figure out what value to use for the condition.
639 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
641 KnownCond = CI->isOne();
643 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
644 // Either operand will do, so be sure to pick the one that's a known
646 // FIXME: Do this more cleverly if both values are known constants?
647 KnownCond = (TrueVal != nullptr);
650 // See if the select has a known constant value for this predecessor.
651 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
652 Result.push_back(std::make_pair(Val, C.second));
655 return !Result.empty();
659 // If all else fails, see if LVI can figure out a constant value for us.
660 Constant *CI = LVI->getConstant(V, BB, CxtI);
661 if (Constant *KC = getKnownConstant(CI, Preference)) {
662 for (BasicBlock *Pred : predecessors(BB))
663 Result.push_back(std::make_pair(KC, Pred));
666 return !Result.empty();
671 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
672 /// in an undefined jump, decide which block is best to revector to.
674 /// Since we can pick an arbitrary destination, we pick the successor with the
675 /// fewest predecessors. This should reduce the in-degree of the others.
677 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
678 TerminatorInst *BBTerm = BB->getTerminator();
679 unsigned MinSucc = 0;
680 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
681 // Compute the successor with the minimum number of predecessors.
682 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
683 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
684 TestBB = BBTerm->getSuccessor(i);
685 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
686 if (NumPreds < MinNumPreds) {
688 MinNumPreds = NumPreds;
695 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
696 if (!BB->hasAddressTaken()) return false;
698 // If the block has its address taken, it may be a tree of dead constants
699 // hanging off of it. These shouldn't keep the block alive.
700 BlockAddress *BA = BlockAddress::get(BB);
701 BA->removeDeadConstantUsers();
702 return !BA->use_empty();
705 /// ProcessBlock - If there are any predecessors whose control can be threaded
706 /// through to a successor, transform them now.
707 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
708 // If the block is trivially dead, just return and let the caller nuke it.
709 // This simplifies other transformations.
710 if (pred_empty(BB) &&
711 BB != &BB->getParent()->getEntryBlock())
714 // If this block has a single predecessor, and if that pred has a single
715 // successor, merge the blocks. This encourages recursive jump threading
716 // because now the condition in this block can be threaded through
717 // predecessors of our predecessor block.
718 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
719 const TerminatorInst *TI = SinglePred->getTerminator();
720 if (!TI->isExceptional() && TI->getNumSuccessors() == 1 &&
721 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
722 // If SinglePred was a loop header, BB becomes one.
723 if (LoopHeaders.erase(SinglePred))
724 LoopHeaders.insert(BB);
726 LVI->eraseBlock(SinglePred);
727 MergeBasicBlockIntoOnlyPred(BB);
733 if (TryToUnfoldSelectInCurrBB(BB))
736 // What kind of constant we're looking for.
737 ConstantPreference Preference = WantInteger;
739 // Look to see if the terminator is a conditional branch, switch or indirect
740 // branch, if not we can't thread it.
742 Instruction *Terminator = BB->getTerminator();
743 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
744 // Can't thread an unconditional jump.
745 if (BI->isUnconditional()) return false;
746 Condition = BI->getCondition();
747 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
748 Condition = SI->getCondition();
749 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
750 // Can't thread indirect branch with no successors.
751 if (IB->getNumSuccessors() == 0) return false;
752 Condition = IB->getAddress()->stripPointerCasts();
753 Preference = WantBlockAddress;
755 return false; // Must be an invoke.
758 // Run constant folding to see if we can reduce the condition to a simple
760 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
762 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
764 I->replaceAllUsesWith(SimpleVal);
765 I->eraseFromParent();
766 Condition = SimpleVal;
770 // If the terminator is branching on an undef, we can pick any of the
771 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
772 if (isa<UndefValue>(Condition)) {
773 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
775 // Fold the branch/switch.
776 TerminatorInst *BBTerm = BB->getTerminator();
777 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
778 if (i == BestSucc) continue;
779 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
782 DEBUG(dbgs() << " In block '" << BB->getName()
783 << "' folding undef terminator: " << *BBTerm << '\n');
784 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
785 BBTerm->eraseFromParent();
789 // If the terminator of this block is branching on a constant, simplify the
790 // terminator to an unconditional branch. This can occur due to threading in
792 if (getKnownConstant(Condition, Preference)) {
793 DEBUG(dbgs() << " In block '" << BB->getName()
794 << "' folding terminator: " << *BB->getTerminator() << '\n');
796 ConstantFoldTerminator(BB, true);
800 Instruction *CondInst = dyn_cast<Instruction>(Condition);
802 // All the rest of our checks depend on the condition being an instruction.
804 // FIXME: Unify this with code below.
805 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
811 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
812 // If we're branching on a conditional, LVI might be able to determine
813 // it's value at the branch instruction. We only handle comparisons
814 // against a constant at this time.
815 // TODO: This should be extended to handle switches as well.
816 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
817 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
818 if (CondBr && CondConst && CondBr->isConditional()) {
819 LazyValueInfo::Tristate Ret =
820 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
822 if (Ret != LazyValueInfo::Unknown) {
823 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
824 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
825 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
826 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
827 CondBr->eraseFromParent();
828 if (CondCmp->use_empty())
829 CondCmp->eraseFromParent();
830 else if (CondCmp->getParent() == BB) {
831 // If the fact we just learned is true for all uses of the
832 // condition, replace it with a constant value
833 auto *CI = Ret == LazyValueInfo::True ?
834 ConstantInt::getTrue(CondCmp->getType()) :
835 ConstantInt::getFalse(CondCmp->getType());
836 CondCmp->replaceAllUsesWith(CI);
837 CondCmp->eraseFromParent();
843 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB))
847 // Check for some cases that are worth simplifying. Right now we want to look
848 // for loads that are used by a switch or by the condition for the branch. If
849 // we see one, check to see if it's partially redundant. If so, insert a PHI
850 // which can then be used to thread the values.
852 Value *SimplifyValue = CondInst;
853 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
854 if (isa<Constant>(CondCmp->getOperand(1)))
855 SimplifyValue = CondCmp->getOperand(0);
857 // TODO: There are other places where load PRE would be profitable, such as
858 // more complex comparisons.
859 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
860 if (SimplifyPartiallyRedundantLoad(LI))
864 // Handle a variety of cases where we are branching on something derived from
865 // a PHI node in the current block. If we can prove that any predecessors
866 // compute a predictable value based on a PHI node, thread those predecessors.
868 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
871 // If this is an otherwise-unfoldable branch on a phi node in the current
872 // block, see if we can simplify.
873 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
874 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
875 return ProcessBranchOnPHI(PN);
878 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
879 if (CondInst->getOpcode() == Instruction::Xor &&
880 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
881 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
883 // Search for a stronger dominating condition that can be used to simplify a
884 // conditional branch leaving BB.
885 if (ProcessImpliedCondition(BB))
891 bool JumpThreading::ProcessImpliedCondition(BasicBlock *BB) {
892 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
893 if (!BI || !BI->isConditional())
896 Value *Cond = BI->getCondition();
897 BasicBlock *CurrentBB = BB;
898 BasicBlock *CurrentPred = BB->getSinglePredecessor();
901 auto &DL = BB->getModule()->getDataLayout();
903 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
904 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
905 if (!PBI || !PBI->isConditional() || PBI->getSuccessor(0) != CurrentBB)
908 if (isImpliedCondition(PBI->getCondition(), Cond, DL)) {
909 BI->getSuccessor(1)->removePredecessor(BB);
910 BranchInst::Create(BI->getSuccessor(0), BI);
911 BI->eraseFromParent();
914 CurrentBB = CurrentPred;
915 CurrentPred = CurrentBB->getSinglePredecessor();
921 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
922 /// load instruction, eliminate it by replacing it with a PHI node. This is an
923 /// important optimization that encourages jump threading, and needs to be run
924 /// interlaced with other jump threading tasks.
925 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
926 // Don't hack volatile/atomic loads.
927 if (!LI->isSimple()) return false;
929 // If the load is defined in a block with exactly one predecessor, it can't be
930 // partially redundant.
931 BasicBlock *LoadBB = LI->getParent();
932 if (LoadBB->getSinglePredecessor())
935 // If the load is defined in an EH pad, it can't be partially redundant,
936 // because the edges between the invoke and the EH pad cannot have other
937 // instructions between them.
938 if (LoadBB->isEHPad())
941 Value *LoadedPtr = LI->getOperand(0);
943 // If the loaded operand is defined in the LoadBB, it can't be available.
944 // TODO: Could do simple PHI translation, that would be fun :)
945 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
946 if (PtrOp->getParent() == LoadBB)
949 // Scan a few instructions up from the load, to see if it is obviously live at
950 // the entry to its block.
951 BasicBlock::iterator BBIt(LI);
953 if (Value *AvailableVal =
954 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, DefMaxInstsToScan)) {
955 // If the value of the load is locally available within the block, just use
956 // it. This frequently occurs for reg2mem'd allocas.
957 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
959 // If the returned value is the load itself, replace with an undef. This can
960 // only happen in dead loops.
961 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
962 if (AvailableVal->getType() != LI->getType())
964 CastInst::CreateBitOrPointerCast(AvailableVal, LI->getType(), "", LI);
965 LI->replaceAllUsesWith(AvailableVal);
966 LI->eraseFromParent();
970 // Otherwise, if we scanned the whole block and got to the top of the block,
971 // we know the block is locally transparent to the load. If not, something
972 // might clobber its value.
973 if (BBIt != LoadBB->begin())
976 // If all of the loads and stores that feed the value have the same AA tags,
977 // then we can propagate them onto any newly inserted loads.
979 LI->getAAMetadata(AATags);
981 SmallPtrSet<BasicBlock*, 8> PredsScanned;
982 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
983 AvailablePredsTy AvailablePreds;
984 BasicBlock *OneUnavailablePred = nullptr;
986 // If we got here, the loaded value is transparent through to the start of the
987 // block. Check to see if it is available in any of the predecessor blocks.
988 for (BasicBlock *PredBB : predecessors(LoadBB)) {
989 // If we already scanned this predecessor, skip it.
990 if (!PredsScanned.insert(PredBB).second)
993 // Scan the predecessor to see if the value is available in the pred.
994 BBIt = PredBB->end();
995 AAMDNodes ThisAATags;
996 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt,
998 nullptr, &ThisAATags);
999 if (!PredAvailable) {
1000 OneUnavailablePred = PredBB;
1004 // If AA tags disagree or are not present, forget about them.
1005 if (AATags != ThisAATags) AATags = AAMDNodes();
1007 // If so, this load is partially redundant. Remember this info so that we
1008 // can create a PHI node.
1009 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1012 // If the loaded value isn't available in any predecessor, it isn't partially
1014 if (AvailablePreds.empty()) return false;
1016 // Okay, the loaded value is available in at least one (and maybe all!)
1017 // predecessors. If the value is unavailable in more than one unique
1018 // predecessor, we want to insert a merge block for those common predecessors.
1019 // This ensures that we only have to insert one reload, thus not increasing
1021 BasicBlock *UnavailablePred = nullptr;
1023 // If there is exactly one predecessor where the value is unavailable, the
1024 // already computed 'OneUnavailablePred' block is it. If it ends in an
1025 // unconditional branch, we know that it isn't a critical edge.
1026 if (PredsScanned.size() == AvailablePreds.size()+1 &&
1027 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1028 UnavailablePred = OneUnavailablePred;
1029 } else if (PredsScanned.size() != AvailablePreds.size()) {
1030 // Otherwise, we had multiple unavailable predecessors or we had a critical
1031 // edge from the one.
1032 SmallVector<BasicBlock*, 8> PredsToSplit;
1033 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1035 for (const auto &AvailablePred : AvailablePreds)
1036 AvailablePredSet.insert(AvailablePred.first);
1038 // Add all the unavailable predecessors to the PredsToSplit list.
1039 for (BasicBlock *P : predecessors(LoadBB)) {
1040 // If the predecessor is an indirect goto, we can't split the edge.
1041 if (isa<IndirectBrInst>(P->getTerminator()))
1044 if (!AvailablePredSet.count(P))
1045 PredsToSplit.push_back(P);
1048 // Split them out to their own block.
1049 UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1052 // If the value isn't available in all predecessors, then there will be
1053 // exactly one where it isn't available. Insert a load on that edge and add
1054 // it to the AvailablePreds list.
1055 if (UnavailablePred) {
1056 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1057 "Can't handle critical edge here!");
1058 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
1060 UnavailablePred->getTerminator());
1061 NewVal->setDebugLoc(LI->getDebugLoc());
1063 NewVal->setAAMetadata(AATags);
1065 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1068 // Now we know that each predecessor of this block has a value in
1069 // AvailablePreds, sort them for efficient access as we're walking the preds.
1070 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1072 // Create a PHI node at the start of the block for the PRE'd load value.
1073 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1074 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
1077 PN->setDebugLoc(LI->getDebugLoc());
1079 // Insert new entries into the PHI for each predecessor. A single block may
1080 // have multiple entries here.
1081 for (pred_iterator PI = PB; PI != PE; ++PI) {
1082 BasicBlock *P = *PI;
1083 AvailablePredsTy::iterator I =
1084 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1085 std::make_pair(P, (Value*)nullptr));
1087 assert(I != AvailablePreds.end() && I->first == P &&
1088 "Didn't find entry for predecessor!");
1090 // If we have an available predecessor but it requires casting, insert the
1091 // cast in the predecessor and use the cast. Note that we have to update the
1092 // AvailablePreds vector as we go so that all of the PHI entries for this
1093 // predecessor use the same bitcast.
1094 Value *&PredV = I->second;
1095 if (PredV->getType() != LI->getType())
1096 PredV = CastInst::CreateBitOrPointerCast(PredV, LI->getType(), "",
1097 P->getTerminator());
1099 PN->addIncoming(PredV, I->first);
1102 //cerr << "PRE: " << *LI << *PN << "\n";
1104 LI->replaceAllUsesWith(PN);
1105 LI->eraseFromParent();
1110 /// FindMostPopularDest - The specified list contains multiple possible
1111 /// threadable destinations. Pick the one that occurs the most frequently in
1114 FindMostPopularDest(BasicBlock *BB,
1115 const SmallVectorImpl<std::pair<BasicBlock*,
1116 BasicBlock*> > &PredToDestList) {
1117 assert(!PredToDestList.empty());
1119 // Determine popularity. If there are multiple possible destinations, we
1120 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1121 // blocks with known and real destinations to threading undef. We'll handle
1122 // them later if interesting.
1123 DenseMap<BasicBlock*, unsigned> DestPopularity;
1124 for (const auto &PredToDest : PredToDestList)
1125 if (PredToDest.second)
1126 DestPopularity[PredToDest.second]++;
1128 // Find the most popular dest.
1129 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1130 BasicBlock *MostPopularDest = DPI->first;
1131 unsigned Popularity = DPI->second;
1132 SmallVector<BasicBlock*, 4> SamePopularity;
1134 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1135 // If the popularity of this entry isn't higher than the popularity we've
1136 // seen so far, ignore it.
1137 if (DPI->second < Popularity)
1139 else if (DPI->second == Popularity) {
1140 // If it is the same as what we've seen so far, keep track of it.
1141 SamePopularity.push_back(DPI->first);
1143 // If it is more popular, remember it.
1144 SamePopularity.clear();
1145 MostPopularDest = DPI->first;
1146 Popularity = DPI->second;
1150 // Okay, now we know the most popular destination. If there is more than one
1151 // destination, we need to determine one. This is arbitrary, but we need
1152 // to make a deterministic decision. Pick the first one that appears in the
1154 if (!SamePopularity.empty()) {
1155 SamePopularity.push_back(MostPopularDest);
1156 TerminatorInst *TI = BB->getTerminator();
1157 for (unsigned i = 0; ; ++i) {
1158 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1160 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1161 TI->getSuccessor(i)) == SamePopularity.end())
1164 MostPopularDest = TI->getSuccessor(i);
1169 // Okay, we have finally picked the most popular destination.
1170 return MostPopularDest;
1173 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1174 ConstantPreference Preference,
1175 Instruction *CxtI) {
1176 // If threading this would thread across a loop header, don't even try to
1178 if (LoopHeaders.count(BB))
1181 PredValueInfoTy PredValues;
1182 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1185 assert(!PredValues.empty() &&
1186 "ComputeValueKnownInPredecessors returned true with no values");
1188 DEBUG(dbgs() << "IN BB: " << *BB;
1189 for (const auto &PredValue : PredValues) {
1190 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1192 << " for pred '" << PredValue.second->getName() << "'.\n";
1195 // Decide what we want to thread through. Convert our list of known values to
1196 // a list of known destinations for each pred. This also discards duplicate
1197 // predecessors and keeps track of the undefined inputs (which are represented
1198 // as a null dest in the PredToDestList).
1199 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1200 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1202 BasicBlock *OnlyDest = nullptr;
1203 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1205 for (const auto &PredValue : PredValues) {
1206 BasicBlock *Pred = PredValue.second;
1207 if (!SeenPreds.insert(Pred).second)
1208 continue; // Duplicate predecessor entry.
1210 // If the predecessor ends with an indirect goto, we can't change its
1212 if (isa<IndirectBrInst>(Pred->getTerminator()))
1215 Constant *Val = PredValue.first;
1218 if (isa<UndefValue>(Val))
1220 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1221 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1222 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1223 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1225 assert(isa<IndirectBrInst>(BB->getTerminator())
1226 && "Unexpected terminator");
1227 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1230 // If we have exactly one destination, remember it for efficiency below.
1231 if (PredToDestList.empty())
1233 else if (OnlyDest != DestBB)
1234 OnlyDest = MultipleDestSentinel;
1236 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1239 // If all edges were unthreadable, we fail.
1240 if (PredToDestList.empty())
1243 // Determine which is the most common successor. If we have many inputs and
1244 // this block is a switch, we want to start by threading the batch that goes
1245 // to the most popular destination first. If we only know about one
1246 // threadable destination (the common case) we can avoid this.
1247 BasicBlock *MostPopularDest = OnlyDest;
1249 if (MostPopularDest == MultipleDestSentinel)
1250 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1252 // Now that we know what the most popular destination is, factor all
1253 // predecessors that will jump to it into a single predecessor.
1254 SmallVector<BasicBlock*, 16> PredsToFactor;
1255 for (const auto &PredToDest : PredToDestList)
1256 if (PredToDest.second == MostPopularDest) {
1257 BasicBlock *Pred = PredToDest.first;
1259 // This predecessor may be a switch or something else that has multiple
1260 // edges to the block. Factor each of these edges by listing them
1261 // according to # occurrences in PredsToFactor.
1262 for (BasicBlock *Succ : successors(Pred))
1264 PredsToFactor.push_back(Pred);
1267 // If the threadable edges are branching on an undefined value, we get to pick
1268 // the destination that these predecessors should get to.
1269 if (!MostPopularDest)
1270 MostPopularDest = BB->getTerminator()->
1271 getSuccessor(GetBestDestForJumpOnUndef(BB));
1273 // Ok, try to thread it!
1274 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1277 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1278 /// a PHI node in the current block. See if there are any simplifications we
1279 /// can do based on inputs to the phi node.
1281 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1282 BasicBlock *BB = PN->getParent();
1284 // TODO: We could make use of this to do it once for blocks with common PHI
1286 SmallVector<BasicBlock*, 1> PredBBs;
1289 // If any of the predecessor blocks end in an unconditional branch, we can
1290 // *duplicate* the conditional branch into that block in order to further
1291 // encourage jump threading and to eliminate cases where we have branch on a
1292 // phi of an icmp (branch on icmp is much better).
1293 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1294 BasicBlock *PredBB = PN->getIncomingBlock(i);
1295 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1296 if (PredBr->isUnconditional()) {
1297 PredBBs[0] = PredBB;
1298 // Try to duplicate BB into PredBB.
1299 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1307 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1308 /// a xor instruction in the current block. See if there are any
1309 /// simplifications we can do based on inputs to the xor.
1311 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1312 BasicBlock *BB = BO->getParent();
1314 // If either the LHS or RHS of the xor is a constant, don't do this
1316 if (isa<ConstantInt>(BO->getOperand(0)) ||
1317 isa<ConstantInt>(BO->getOperand(1)))
1320 // If the first instruction in BB isn't a phi, we won't be able to infer
1321 // anything special about any particular predecessor.
1322 if (!isa<PHINode>(BB->front()))
1325 // If we have a xor as the branch input to this block, and we know that the
1326 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1327 // the condition into the predecessor and fix that value to true, saving some
1328 // logical ops on that path and encouraging other paths to simplify.
1330 // This copies something like this:
1333 // %X = phi i1 [1], [%X']
1334 // %Y = icmp eq i32 %A, %B
1335 // %Z = xor i1 %X, %Y
1340 // %Y = icmp ne i32 %A, %B
1343 PredValueInfoTy XorOpValues;
1345 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1347 assert(XorOpValues.empty());
1348 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1354 assert(!XorOpValues.empty() &&
1355 "ComputeValueKnownInPredecessors returned true with no values");
1357 // Scan the information to see which is most popular: true or false. The
1358 // predecessors can be of the set true, false, or undef.
1359 unsigned NumTrue = 0, NumFalse = 0;
1360 for (const auto &XorOpValue : XorOpValues) {
1361 if (isa<UndefValue>(XorOpValue.first))
1362 // Ignore undefs for the count.
1364 if (cast<ConstantInt>(XorOpValue.first)->isZero())
1370 // Determine which value to split on, true, false, or undef if neither.
1371 ConstantInt *SplitVal = nullptr;
1372 if (NumTrue > NumFalse)
1373 SplitVal = ConstantInt::getTrue(BB->getContext());
1374 else if (NumTrue != 0 || NumFalse != 0)
1375 SplitVal = ConstantInt::getFalse(BB->getContext());
1377 // Collect all of the blocks that this can be folded into so that we can
1378 // factor this once and clone it once.
1379 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1380 for (const auto &XorOpValue : XorOpValues) {
1381 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1384 BlocksToFoldInto.push_back(XorOpValue.second);
1387 // If we inferred a value for all of the predecessors, then duplication won't
1388 // help us. However, we can just replace the LHS or RHS with the constant.
1389 if (BlocksToFoldInto.size() ==
1390 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1392 // If all preds provide undef, just nuke the xor, because it is undef too.
1393 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1394 BO->eraseFromParent();
1395 } else if (SplitVal->isZero()) {
1396 // If all preds provide 0, replace the xor with the other input.
1397 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1398 BO->eraseFromParent();
1400 // If all preds provide 1, set the computed value to 1.
1401 BO->setOperand(!isLHS, SplitVal);
1407 // Try to duplicate BB into PredBB.
1408 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1412 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1413 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1414 /// NewPred using the entries from OldPred (suitably mapped).
1415 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1416 BasicBlock *OldPred,
1417 BasicBlock *NewPred,
1418 DenseMap<Instruction*, Value*> &ValueMap) {
1419 for (BasicBlock::iterator PNI = PHIBB->begin();
1420 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1421 // Ok, we have a PHI node. Figure out what the incoming value was for the
1423 Value *IV = PN->getIncomingValueForBlock(OldPred);
1425 // Remap the value if necessary.
1426 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1427 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1428 if (I != ValueMap.end())
1432 PN->addIncoming(IV, NewPred);
1436 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1437 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1438 /// across BB. Transform the IR to reflect this change.
1439 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1440 const SmallVectorImpl<BasicBlock*> &PredBBs,
1441 BasicBlock *SuccBB) {
1442 // If threading to the same block as we come from, we would infinite loop.
1444 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1445 << "' - would thread to self!\n");
1449 // If threading this would thread across a loop header, don't thread the edge.
1450 // See the comments above FindLoopHeaders for justifications and caveats.
1451 if (LoopHeaders.count(BB)) {
1452 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1453 << "' to dest BB '" << SuccBB->getName()
1454 << "' - it might create an irreducible loop!\n");
1458 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, BBDupThreshold);
1459 if (JumpThreadCost > BBDupThreshold) {
1460 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1461 << "' - Cost is too high: " << JumpThreadCost << "\n");
1465 // And finally, do it! Start by factoring the predecessors if needed.
1467 if (PredBBs.size() == 1)
1468 PredBB = PredBBs[0];
1470 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1471 << " common predecessors.\n");
1472 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1475 // And finally, do it!
1476 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1477 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1478 << ", across block:\n "
1481 LVI->threadEdge(PredBB, BB, SuccBB);
1483 // We are going to have to map operands from the original BB block to the new
1484 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1485 // account for entry from PredBB.
1486 DenseMap<Instruction*, Value*> ValueMapping;
1488 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1489 BB->getName()+".thread",
1490 BB->getParent(), BB);
1491 NewBB->moveAfter(PredBB);
1493 // Set the block frequency of NewBB.
1494 if (HasProfileData) {
1496 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
1497 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
1500 BasicBlock::iterator BI = BB->begin();
1501 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1502 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1504 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1505 // mapping and using it to remap operands in the cloned instructions.
1506 for (; !isa<TerminatorInst>(BI); ++BI) {
1507 Instruction *New = BI->clone();
1508 New->setName(BI->getName());
1509 NewBB->getInstList().push_back(New);
1510 ValueMapping[&*BI] = New;
1512 // Remap operands to patch up intra-block references.
1513 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1514 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1515 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1516 if (I != ValueMapping.end())
1517 New->setOperand(i, I->second);
1521 // We didn't copy the terminator from BB over to NewBB, because there is now
1522 // an unconditional jump to SuccBB. Insert the unconditional jump.
1523 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
1524 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1526 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1527 // PHI nodes for NewBB now.
1528 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1530 // If there were values defined in BB that are used outside the block, then we
1531 // now have to update all uses of the value to use either the original value,
1532 // the cloned value, or some PHI derived value. This can require arbitrary
1533 // PHI insertion, of which we are prepared to do, clean these up now.
1534 SSAUpdater SSAUpdate;
1535 SmallVector<Use*, 16> UsesToRename;
1536 for (Instruction &I : *BB) {
1537 // Scan all uses of this instruction to see if it is used outside of its
1538 // block, and if so, record them in UsesToRename.
1539 for (Use &U : I.uses()) {
1540 Instruction *User = cast<Instruction>(U.getUser());
1541 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1542 if (UserPN->getIncomingBlock(U) == BB)
1544 } else if (User->getParent() == BB)
1547 UsesToRename.push_back(&U);
1550 // If there are no uses outside the block, we're done with this instruction.
1551 if (UsesToRename.empty())
1554 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
1556 // We found a use of I outside of BB. Rename all uses of I that are outside
1557 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1558 // with the two values we know.
1559 SSAUpdate.Initialize(I.getType(), I.getName());
1560 SSAUpdate.AddAvailableValue(BB, &I);
1561 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
1563 while (!UsesToRename.empty())
1564 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1565 DEBUG(dbgs() << "\n");
1569 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1570 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1571 // us to simplify any PHI nodes in BB.
1572 TerminatorInst *PredTerm = PredBB->getTerminator();
1573 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1574 if (PredTerm->getSuccessor(i) == BB) {
1575 BB->removePredecessor(PredBB, true);
1576 PredTerm->setSuccessor(i, NewBB);
1579 // At this point, the IR is fully up to date and consistent. Do a quick scan
1580 // over the new instructions and zap any that are constants or dead. This
1581 // frequently happens because of phi translation.
1582 SimplifyInstructionsInBlock(NewBB, TLI);
1584 // Update the edge weight from BB to SuccBB, which should be less than before.
1585 UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
1587 // Threaded an edge!
1592 /// Create a new basic block that will be the predecessor of BB and successor of
1593 /// all blocks in Preds. When profile data is availble, update the frequency of
1595 BasicBlock *JumpThreading::SplitBlockPreds(BasicBlock *BB,
1596 ArrayRef<BasicBlock *> Preds,
1597 const char *Suffix) {
1598 // Collect the frequencies of all predecessors of BB, which will be used to
1599 // update the edge weight on BB->SuccBB.
1600 BlockFrequency PredBBFreq(0);
1602 for (auto Pred : Preds)
1603 PredBBFreq += BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB);
1605 BasicBlock *PredBB = SplitBlockPredecessors(BB, Preds, Suffix);
1607 // Set the block frequency of the newly created PredBB, which is the sum of
1608 // frequencies of Preds.
1610 BFI->setBlockFreq(PredBB, PredBBFreq.getFrequency());
1614 /// Update the block frequency of BB and branch weight and the metadata on the
1615 /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
1616 /// Freq(PredBB->BB) / Freq(BB->SuccBB).
1617 void JumpThreading::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
1620 BasicBlock *SuccBB) {
1621 if (!HasProfileData)
1624 assert(BFI && BPI && "BFI & BPI should have been created here");
1626 // As the edge from PredBB to BB is deleted, we have to update the block
1628 auto BBOrigFreq = BFI->getBlockFreq(BB);
1629 auto NewBBFreq = BFI->getBlockFreq(NewBB);
1630 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
1631 auto BBNewFreq = BBOrigFreq - NewBBFreq;
1632 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
1634 // Collect updated outgoing edges' frequencies from BB and use them to update
1635 // edge probabilities.
1636 SmallVector<uint64_t, 4> BBSuccFreq;
1637 for (BasicBlock *Succ : successors(BB)) {
1638 auto SuccFreq = (Succ == SuccBB)
1639 ? BB2SuccBBFreq - NewBBFreq
1640 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
1641 BBSuccFreq.push_back(SuccFreq.getFrequency());
1644 uint64_t MaxBBSuccFreq =
1645 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
1647 SmallVector<BranchProbability, 4> BBSuccProbs;
1648 if (MaxBBSuccFreq == 0)
1649 BBSuccProbs.assign(BBSuccFreq.size(),
1650 {1, static_cast<unsigned>(BBSuccFreq.size())});
1652 for (uint64_t Freq : BBSuccFreq)
1653 BBSuccProbs.push_back(
1654 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
1655 // Normalize edge probabilities so that they sum up to one.
1656 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
1660 // Update edge probabilities in BPI.
1661 for (int I = 0, E = BBSuccProbs.size(); I < E; I++)
1662 BPI->setEdgeProbability(BB, I, BBSuccProbs[I]);
1664 if (BBSuccProbs.size() >= 2) {
1665 SmallVector<uint32_t, 4> Weights;
1666 for (auto Prob : BBSuccProbs)
1667 Weights.push_back(Prob.getNumerator());
1669 auto TI = BB->getTerminator();
1671 LLVMContext::MD_prof,
1672 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
1676 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1677 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1678 /// If we can duplicate the contents of BB up into PredBB do so now, this
1679 /// improves the odds that the branch will be on an analyzable instruction like
1681 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1682 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1683 assert(!PredBBs.empty() && "Can't handle an empty set");
1685 // If BB is a loop header, then duplicating this block outside the loop would
1686 // cause us to transform this into an irreducible loop, don't do this.
1687 // See the comments above FindLoopHeaders for justifications and caveats.
1688 if (LoopHeaders.count(BB)) {
1689 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1690 << "' into predecessor block '" << PredBBs[0]->getName()
1691 << "' - it might create an irreducible loop!\n");
1695 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, BBDupThreshold);
1696 if (DuplicationCost > BBDupThreshold) {
1697 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1698 << "' - Cost is too high: " << DuplicationCost << "\n");
1702 // And finally, do it! Start by factoring the predecessors if needed.
1704 if (PredBBs.size() == 1)
1705 PredBB = PredBBs[0];
1707 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1708 << " common predecessors.\n");
1709 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1712 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1714 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1715 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1716 << DuplicationCost << " block is:" << *BB << "\n");
1718 // Unless PredBB ends with an unconditional branch, split the edge so that we
1719 // can just clone the bits from BB into the end of the new PredBB.
1720 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1722 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
1723 PredBB = SplitEdge(PredBB, BB);
1724 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1727 // We are going to have to map operands from the original BB block into the
1728 // PredBB block. Evaluate PHI nodes in BB.
1729 DenseMap<Instruction*, Value*> ValueMapping;
1731 BasicBlock::iterator BI = BB->begin();
1732 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1733 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1734 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1735 // mapping and using it to remap operands in the cloned instructions.
1736 for (; BI != BB->end(); ++BI) {
1737 Instruction *New = BI->clone();
1739 // Remap operands to patch up intra-block references.
1740 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1741 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1742 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1743 if (I != ValueMapping.end())
1744 New->setOperand(i, I->second);
1747 // If this instruction can be simplified after the operands are updated,
1748 // just use the simplified value instead. This frequently happens due to
1751 SimplifyInstruction(New, BB->getModule()->getDataLayout())) {
1753 ValueMapping[&*BI] = IV;
1755 // Otherwise, insert the new instruction into the block.
1756 New->setName(BI->getName());
1757 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
1758 ValueMapping[&*BI] = New;
1762 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1763 // add entries to the PHI nodes for branch from PredBB now.
1764 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1765 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1767 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1770 // If there were values defined in BB that are used outside the block, then we
1771 // now have to update all uses of the value to use either the original value,
1772 // the cloned value, or some PHI derived value. This can require arbitrary
1773 // PHI insertion, of which we are prepared to do, clean these up now.
1774 SSAUpdater SSAUpdate;
1775 SmallVector<Use*, 16> UsesToRename;
1776 for (Instruction &I : *BB) {
1777 // Scan all uses of this instruction to see if it is used outside of its
1778 // block, and if so, record them in UsesToRename.
1779 for (Use &U : I.uses()) {
1780 Instruction *User = cast<Instruction>(U.getUser());
1781 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1782 if (UserPN->getIncomingBlock(U) == BB)
1784 } else if (User->getParent() == BB)
1787 UsesToRename.push_back(&U);
1790 // If there are no uses outside the block, we're done with this instruction.
1791 if (UsesToRename.empty())
1794 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
1796 // We found a use of I outside of BB. Rename all uses of I that are outside
1797 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1798 // with the two values we know.
1799 SSAUpdate.Initialize(I.getType(), I.getName());
1800 SSAUpdate.AddAvailableValue(BB, &I);
1801 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&I]);
1803 while (!UsesToRename.empty())
1804 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1805 DEBUG(dbgs() << "\n");
1808 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1810 BB->removePredecessor(PredBB, true);
1812 // Remove the unconditional branch at the end of the PredBB block.
1813 OldPredBranch->eraseFromParent();
1819 /// TryToUnfoldSelect - Look for blocks of the form
1825 /// %p = phi [%a, %bb] ...
1829 /// And expand the select into a branch structure if one of its arms allows %c
1830 /// to be folded. This later enables threading from bb1 over bb2.
1831 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1832 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1833 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1834 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1836 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1837 CondLHS->getParent() != BB)
1840 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1841 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1842 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1844 // Look if one of the incoming values is a select in the corresponding
1846 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1849 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1850 if (!PredTerm || !PredTerm->isUnconditional())
1853 // Now check if one of the select values would allow us to constant fold the
1854 // terminator in BB. We don't do the transform if both sides fold, those
1855 // cases will be threaded in any case.
1856 LazyValueInfo::Tristate LHSFolds =
1857 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1858 CondRHS, Pred, BB, CondCmp);
1859 LazyValueInfo::Tristate RHSFolds =
1860 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
1861 CondRHS, Pred, BB, CondCmp);
1862 if ((LHSFolds != LazyValueInfo::Unknown ||
1863 RHSFolds != LazyValueInfo::Unknown) &&
1864 LHSFolds != RHSFolds) {
1865 // Expand the select.
1874 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
1875 BB->getParent(), BB);
1876 // Move the unconditional branch to NewBB.
1877 PredTerm->removeFromParent();
1878 NewBB->getInstList().insert(NewBB->end(), PredTerm);
1879 // Create a conditional branch and update PHI nodes.
1880 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
1881 CondLHS->setIncomingValue(I, SI->getFalseValue());
1882 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
1883 // The select is now dead.
1884 SI->eraseFromParent();
1886 // Update any other PHI nodes in BB.
1887 for (BasicBlock::iterator BI = BB->begin();
1888 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
1890 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
1897 /// TryToUnfoldSelectInCurrBB - Look for PHI/Select in the same BB of the form
1899 /// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
1900 /// %s = select p, trueval, falseval
1902 /// And expand the select into a branch structure. This later enables
1903 /// jump-threading over bb in this pass.
1905 /// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
1906 /// select if the associated PHI has at least one constant. If the unfolded
1907 /// select is not jump-threaded, it will be folded again in the later
1909 bool JumpThreading::TryToUnfoldSelectInCurrBB(BasicBlock *BB) {
1910 // If threading this would thread across a loop header, don't thread the edge.
1911 // See the comments above FindLoopHeaders for justifications and caveats.
1912 if (LoopHeaders.count(BB))
1915 // Look for a Phi/Select pair in the same basic block. The Phi feeds the
1916 // condition of the Select and at least one of the incoming values is a
1918 for (BasicBlock::iterator BI = BB->begin();
1919 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
1920 unsigned NumPHIValues = PN->getNumIncomingValues();
1921 if (NumPHIValues == 0 || !PN->hasOneUse())
1924 SelectInst *SI = dyn_cast<SelectInst>(PN->user_back());
1925 if (!SI || SI->getParent() != BB)
1928 Value *Cond = SI->getCondition();
1929 if (!Cond || Cond != PN || !Cond->getType()->isIntegerTy(1))
1932 bool HasConst = false;
1933 for (unsigned i = 0; i != NumPHIValues; ++i) {
1934 if (PN->getIncomingBlock(i) == BB)
1936 if (isa<ConstantInt>(PN->getIncomingValue(i)))
1941 // Expand the select.
1942 TerminatorInst *Term =
1943 SplitBlockAndInsertIfThen(SI->getCondition(), SI, false);
1944 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
1945 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
1946 NewPN->addIncoming(SI->getFalseValue(), BB);
1947 SI->replaceAllUsesWith(NewPN);
1948 SI->eraseFromParent();