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 bool EverChanged = false;
215 EverChanged |= removeUnreachableBlocks(F, LVI);
222 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
223 BasicBlock *BB = &*I;
224 // Thread all of the branches we can over this block.
225 while (ProcessBlock(BB))
230 // If the block is trivially dead, zap it. This eliminates the successor
231 // edges which simplifies the CFG.
232 if (pred_empty(BB) &&
233 BB != &BB->getParent()->getEntryBlock()) {
234 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
235 << "' with terminator: " << *BB->getTerminator() << '\n');
236 LoopHeaders.erase(BB);
243 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
245 // Can't thread an unconditional jump, but if the block is "almost
246 // empty", we can replace uses of it with uses of the successor and make
248 if (BI && BI->isUnconditional() &&
249 BB != &BB->getParent()->getEntryBlock() &&
250 // If the terminator is the only non-phi instruction, try to nuke it.
251 BB->getFirstNonPHIOrDbg()->isTerminator()) {
252 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
253 // block, we have to make sure it isn't in the LoopHeaders set. We
254 // reinsert afterward if needed.
255 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
256 BasicBlock *Succ = BI->getSuccessor(0);
258 // FIXME: It is always conservatively correct to drop the info
259 // for a block even if it doesn't get erased. This isn't totally
260 // awesome, but it allows us to use AssertingVH to prevent nasty
261 // dangling pointer issues within LazyValueInfo.
263 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
265 // If we deleted BB and BB was the header of a loop, then the
266 // successor is now the header of the loop.
270 if (ErasedFromLoopHeaders)
271 LoopHeaders.insert(BB);
274 EverChanged |= Changed;
281 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
282 /// thread across it. Stop scanning the block when passing the threshold.
283 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
284 unsigned Threshold) {
285 /// Ignore PHI nodes, these will be flattened when duplication happens.
286 BasicBlock::const_iterator I(BB->getFirstNonPHI());
288 // FIXME: THREADING will delete values that are just used to compute the
289 // branch, so they shouldn't count against the duplication cost.
292 const TerminatorInst *BBTerm = BB->getTerminator();
293 // Threading through a switch statement is particularly profitable. If this
294 // block ends in a switch, decrease its cost to make it more likely to happen.
295 if (isa<SwitchInst>(BBTerm))
298 // The same holds for indirect branches, but slightly more so.
299 if (isa<IndirectBrInst>(BBTerm))
302 // Bump the threshold up so the early exit from the loop doesn't skip the
303 // terminator-based Size adjustment at the end.
306 // Sum up the cost of each instruction until we get to the terminator. Don't
307 // include the terminator because the copy won't include it.
309 for (; !isa<TerminatorInst>(I); ++I) {
311 // Stop scanning the block if we've reached the threshold.
312 if (Size > Threshold)
315 // Debugger intrinsics don't incur code size.
316 if (isa<DbgInfoIntrinsic>(I)) continue;
318 // If this is a pointer->pointer bitcast, it is free.
319 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
322 // Bail out if this instruction gives back a token type, it is not possible
323 // to duplicate it if it is used outside this BB.
324 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
327 // All other instructions count for at least one unit.
330 // Calls are more expensive. If they are non-intrinsic calls, we model them
331 // as having cost of 4. If they are a non-vector intrinsic, we model them
332 // as having cost of 2 total, and if they are a vector intrinsic, we model
333 // them as having cost 1.
334 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
335 if (CI->cannotDuplicate() || CI->isConvergent())
336 // Blocks with NoDuplicate are modelled as having infinite cost, so they
337 // are never duplicated.
339 else if (!isa<IntrinsicInst>(CI))
341 else if (!CI->getType()->isVectorTy())
346 return Size > Bonus ? Size - Bonus : 0;
349 /// FindLoopHeaders - We do not want jump threading to turn proper loop
350 /// structures into irreducible loops. Doing this breaks up the loop nesting
351 /// hierarchy and pessimizes later transformations. To prevent this from
352 /// happening, we first have to find the loop headers. Here we approximate this
353 /// by finding targets of backedges in the CFG.
355 /// Note that there definitely are cases when we want to allow threading of
356 /// edges across a loop header. For example, threading a jump from outside the
357 /// loop (the preheader) to an exit block of the loop is definitely profitable.
358 /// It is also almost always profitable to thread backedges from within the loop
359 /// to exit blocks, and is often profitable to thread backedges to other blocks
360 /// within the loop (forming a nested loop). This simple analysis is not rich
361 /// enough to track all of these properties and keep it up-to-date as the CFG
362 /// mutates, so we don't allow any of these transformations.
364 void JumpThreading::FindLoopHeaders(Function &F) {
365 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
366 FindFunctionBackedges(F, Edges);
368 for (const auto &Edge : Edges)
369 LoopHeaders.insert(Edge.second);
372 /// getKnownConstant - Helper method to determine if we can thread over a
373 /// terminator with the given value as its condition, and if so what value to
374 /// use for that. What kind of value this is depends on whether we want an
375 /// integer or a block address, but an undef is always accepted.
376 /// Returns null if Val is null or not an appropriate constant.
377 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
381 // Undef is "known" enough.
382 if (UndefValue *U = dyn_cast<UndefValue>(Val))
385 if (Preference == WantBlockAddress)
386 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
388 return dyn_cast<ConstantInt>(Val);
391 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
392 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
393 /// in any of our predecessors. If so, return the known list of value and pred
394 /// BB in the result vector.
396 /// This returns true if there were any known values.
399 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
400 ConstantPreference Preference,
402 // This method walks up use-def chains recursively. Because of this, we could
403 // get into an infinite loop going around loops in the use-def chain. To
404 // prevent this, keep track of what (value, block) pairs we've already visited
405 // and terminate the search if we loop back to them
406 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
409 // An RAII help to remove this pair from the recursion set once the recursion
410 // stack pops back out again.
411 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
413 // If V is a constant, then it is known in all predecessors.
414 if (Constant *KC = getKnownConstant(V, Preference)) {
415 for (BasicBlock *Pred : predecessors(BB))
416 Result.push_back(std::make_pair(KC, Pred));
421 // If V is a non-instruction value, or an instruction in a different block,
422 // then it can't be derived from a PHI.
423 Instruction *I = dyn_cast<Instruction>(V);
424 if (!I || I->getParent() != BB) {
426 // Okay, if this is a live-in value, see if it has a known value at the end
427 // of any of our predecessors.
429 // FIXME: This should be an edge property, not a block end property.
430 /// TODO: Per PR2563, we could infer value range information about a
431 /// predecessor based on its terminator.
433 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
434 // "I" is a non-local compare-with-a-constant instruction. This would be
435 // able to handle value inequalities better, for example if the compare is
436 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
437 // Perhaps getConstantOnEdge should be smart enough to do this?
439 for (BasicBlock *P : predecessors(BB)) {
440 // If the value is known by LazyValueInfo to be a constant in a
441 // predecessor, use that information to try to thread this block.
442 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
443 if (Constant *KC = getKnownConstant(PredCst, Preference))
444 Result.push_back(std::make_pair(KC, P));
447 return !Result.empty();
450 /// If I is a PHI node, then we know the incoming values for any constants.
451 if (PHINode *PN = dyn_cast<PHINode>(I)) {
452 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
453 Value *InVal = PN->getIncomingValue(i);
454 if (Constant *KC = getKnownConstant(InVal, Preference)) {
455 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
457 Constant *CI = LVI->getConstantOnEdge(InVal,
458 PN->getIncomingBlock(i),
460 if (Constant *KC = getKnownConstant(CI, Preference))
461 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
465 return !Result.empty();
468 PredValueInfoTy LHSVals, RHSVals;
470 // Handle some boolean conditions.
471 if (I->getType()->getPrimitiveSizeInBits() == 1) {
472 assert(Preference == WantInteger && "One-bit non-integer type?");
474 // X & false -> false
475 if (I->getOpcode() == Instruction::Or ||
476 I->getOpcode() == Instruction::And) {
477 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
479 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
482 if (LHSVals.empty() && RHSVals.empty())
485 ConstantInt *InterestingVal;
486 if (I->getOpcode() == Instruction::Or)
487 InterestingVal = ConstantInt::getTrue(I->getContext());
489 InterestingVal = ConstantInt::getFalse(I->getContext());
491 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
493 // Scan for the sentinel. If we find an undef, force it to the
494 // interesting value: x|undef -> true and x&undef -> false.
495 for (const auto &LHSVal : LHSVals)
496 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
497 Result.emplace_back(InterestingVal, LHSVal.second);
498 LHSKnownBBs.insert(LHSVal.second);
500 for (const auto &RHSVal : RHSVals)
501 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
502 // If we already inferred a value for this block on the LHS, don't
504 if (!LHSKnownBBs.count(RHSVal.second))
505 Result.emplace_back(InterestingVal, RHSVal.second);
508 return !Result.empty();
511 // Handle the NOT form of XOR.
512 if (I->getOpcode() == Instruction::Xor &&
513 isa<ConstantInt>(I->getOperand(1)) &&
514 cast<ConstantInt>(I->getOperand(1))->isOne()) {
515 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
520 // Invert the known values.
521 for (auto &R : Result)
522 R.first = ConstantExpr::getNot(R.first);
527 // Try to simplify some other binary operator values.
528 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
529 assert(Preference != WantBlockAddress
530 && "A binary operator creating a block address?");
531 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
532 PredValueInfoTy LHSVals;
533 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
536 // Try to use constant folding to simplify the binary operator.
537 for (const auto &LHSVal : LHSVals) {
538 Constant *V = LHSVal.first;
539 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
541 if (Constant *KC = getKnownConstant(Folded, WantInteger))
542 Result.push_back(std::make_pair(KC, LHSVal.second));
546 return !Result.empty();
549 // Handle compare with phi operand, where the PHI is defined in this block.
550 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
551 assert(Preference == WantInteger && "Compares only produce integers");
552 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
553 if (PN && PN->getParent() == BB) {
554 const DataLayout &DL = PN->getModule()->getDataLayout();
555 // We can do this simplification if any comparisons fold to true or false.
557 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
558 BasicBlock *PredBB = PN->getIncomingBlock(i);
559 Value *LHS = PN->getIncomingValue(i);
560 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
562 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL);
564 if (!isa<Constant>(RHS))
567 LazyValueInfo::Tristate
568 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
569 cast<Constant>(RHS), PredBB, BB,
571 if (ResT == LazyValueInfo::Unknown)
573 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
576 if (Constant *KC = getKnownConstant(Res, WantInteger))
577 Result.push_back(std::make_pair(KC, PredBB));
580 return !Result.empty();
583 // If comparing a live-in value against a constant, see if we know the
584 // live-in value on any predecessors.
585 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
586 if (!isa<Instruction>(Cmp->getOperand(0)) ||
587 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
588 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
590 for (BasicBlock *P : predecessors(BB)) {
591 // If the value is known by LazyValueInfo to be a constant in a
592 // predecessor, use that information to try to thread this block.
593 LazyValueInfo::Tristate Res =
594 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
595 RHSCst, P, BB, CxtI ? CxtI : Cmp);
596 if (Res == LazyValueInfo::Unknown)
599 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
600 Result.push_back(std::make_pair(ResC, P));
603 return !Result.empty();
606 // Try to find a constant value for the LHS of a comparison,
607 // and evaluate it statically if we can.
608 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
609 PredValueInfoTy LHSVals;
610 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
613 for (const auto &LHSVal : LHSVals) {
614 Constant *V = LHSVal.first;
615 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
617 if (Constant *KC = getKnownConstant(Folded, WantInteger))
618 Result.push_back(std::make_pair(KC, LHSVal.second));
621 return !Result.empty();
626 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
627 // Handle select instructions where at least one operand is a known constant
628 // and we can figure out the condition value for any predecessor block.
629 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
630 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
631 PredValueInfoTy Conds;
632 if ((TrueVal || FalseVal) &&
633 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
634 WantInteger, CxtI)) {
635 for (auto &C : Conds) {
636 Constant *Cond = C.first;
638 // Figure out what value to use for the condition.
640 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
642 KnownCond = CI->isOne();
644 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
645 // Either operand will do, so be sure to pick the one that's a known
647 // FIXME: Do this more cleverly if both values are known constants?
648 KnownCond = (TrueVal != nullptr);
651 // See if the select has a known constant value for this predecessor.
652 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
653 Result.push_back(std::make_pair(Val, C.second));
656 return !Result.empty();
660 // If all else fails, see if LVI can figure out a constant value for us.
661 Constant *CI = LVI->getConstant(V, BB, CxtI);
662 if (Constant *KC = getKnownConstant(CI, Preference)) {
663 for (BasicBlock *Pred : predecessors(BB))
664 Result.push_back(std::make_pair(KC, Pred));
667 return !Result.empty();
672 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
673 /// in an undefined jump, decide which block is best to revector to.
675 /// Since we can pick an arbitrary destination, we pick the successor with the
676 /// fewest predecessors. This should reduce the in-degree of the others.
678 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
679 TerminatorInst *BBTerm = BB->getTerminator();
680 unsigned MinSucc = 0;
681 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
682 // Compute the successor with the minimum number of predecessors.
683 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
684 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
685 TestBB = BBTerm->getSuccessor(i);
686 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
687 if (NumPreds < MinNumPreds) {
689 MinNumPreds = NumPreds;
696 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
697 if (!BB->hasAddressTaken()) return false;
699 // If the block has its address taken, it may be a tree of dead constants
700 // hanging off of it. These shouldn't keep the block alive.
701 BlockAddress *BA = BlockAddress::get(BB);
702 BA->removeDeadConstantUsers();
703 return !BA->use_empty();
706 /// ProcessBlock - If there are any predecessors whose control can be threaded
707 /// through to a successor, transform them now.
708 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
709 // If the block is trivially dead, just return and let the caller nuke it.
710 // This simplifies other transformations.
711 if (pred_empty(BB) &&
712 BB != &BB->getParent()->getEntryBlock())
715 // If this block has a single predecessor, and if that pred has a single
716 // successor, merge the blocks. This encourages recursive jump threading
717 // because now the condition in this block can be threaded through
718 // predecessors of our predecessor block.
719 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
720 const TerminatorInst *TI = SinglePred->getTerminator();
721 if (!TI->isExceptional() && TI->getNumSuccessors() == 1 &&
722 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
723 // If SinglePred was a loop header, BB becomes one.
724 if (LoopHeaders.erase(SinglePred))
725 LoopHeaders.insert(BB);
727 LVI->eraseBlock(SinglePred);
728 MergeBasicBlockIntoOnlyPred(BB);
734 if (TryToUnfoldSelectInCurrBB(BB))
737 // What kind of constant we're looking for.
738 ConstantPreference Preference = WantInteger;
740 // Look to see if the terminator is a conditional branch, switch or indirect
741 // branch, if not we can't thread it.
743 Instruction *Terminator = BB->getTerminator();
744 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
745 // Can't thread an unconditional jump.
746 if (BI->isUnconditional()) return false;
747 Condition = BI->getCondition();
748 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
749 Condition = SI->getCondition();
750 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
751 // Can't thread indirect branch with no successors.
752 if (IB->getNumSuccessors() == 0) return false;
753 Condition = IB->getAddress()->stripPointerCasts();
754 Preference = WantBlockAddress;
756 return false; // Must be an invoke.
759 // Run constant folding to see if we can reduce the condition to a simple
761 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
763 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
765 I->replaceAllUsesWith(SimpleVal);
766 I->eraseFromParent();
767 Condition = SimpleVal;
771 // If the terminator is branching on an undef, we can pick any of the
772 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
773 if (isa<UndefValue>(Condition)) {
774 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
776 // Fold the branch/switch.
777 TerminatorInst *BBTerm = BB->getTerminator();
778 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
779 if (i == BestSucc) continue;
780 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
783 DEBUG(dbgs() << " In block '" << BB->getName()
784 << "' folding undef terminator: " << *BBTerm << '\n');
785 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
786 BBTerm->eraseFromParent();
790 // If the terminator of this block is branching on a constant, simplify the
791 // terminator to an unconditional branch. This can occur due to threading in
793 if (getKnownConstant(Condition, Preference)) {
794 DEBUG(dbgs() << " In block '" << BB->getName()
795 << "' folding terminator: " << *BB->getTerminator() << '\n');
797 ConstantFoldTerminator(BB, true);
801 Instruction *CondInst = dyn_cast<Instruction>(Condition);
803 // All the rest of our checks depend on the condition being an instruction.
805 // FIXME: Unify this with code below.
806 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
812 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
813 // If we're branching on a conditional, LVI might be able to determine
814 // it's value at the branch instruction. We only handle comparisons
815 // against a constant at this time.
816 // TODO: This should be extended to handle switches as well.
817 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
818 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
819 if (CondBr && CondConst && CondBr->isConditional()) {
820 LazyValueInfo::Tristate Ret =
821 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
823 if (Ret != LazyValueInfo::Unknown) {
824 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
825 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
826 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
827 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
828 CondBr->eraseFromParent();
829 if (CondCmp->use_empty())
830 CondCmp->eraseFromParent();
831 else if (CondCmp->getParent() == BB) {
832 // If the fact we just learned is true for all uses of the
833 // condition, replace it with a constant value
834 auto *CI = Ret == LazyValueInfo::True ?
835 ConstantInt::getTrue(CondCmp->getType()) :
836 ConstantInt::getFalse(CondCmp->getType());
837 CondCmp->replaceAllUsesWith(CI);
838 CondCmp->eraseFromParent();
844 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB))
848 // Check for some cases that are worth simplifying. Right now we want to look
849 // for loads that are used by a switch or by the condition for the branch. If
850 // we see one, check to see if it's partially redundant. If so, insert a PHI
851 // which can then be used to thread the values.
853 Value *SimplifyValue = CondInst;
854 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
855 if (isa<Constant>(CondCmp->getOperand(1)))
856 SimplifyValue = CondCmp->getOperand(0);
858 // TODO: There are other places where load PRE would be profitable, such as
859 // more complex comparisons.
860 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
861 if (SimplifyPartiallyRedundantLoad(LI))
865 // Handle a variety of cases where we are branching on something derived from
866 // a PHI node in the current block. If we can prove that any predecessors
867 // compute a predictable value based on a PHI node, thread those predecessors.
869 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
872 // If this is an otherwise-unfoldable branch on a phi node in the current
873 // block, see if we can simplify.
874 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
875 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
876 return ProcessBranchOnPHI(PN);
879 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
880 if (CondInst->getOpcode() == Instruction::Xor &&
881 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
882 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
884 // Search for a stronger dominating condition that can be used to simplify a
885 // conditional branch leaving BB.
886 if (ProcessImpliedCondition(BB))
892 bool JumpThreading::ProcessImpliedCondition(BasicBlock *BB) {
893 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
894 if (!BI || !BI->isConditional())
897 Value *Cond = BI->getCondition();
898 BasicBlock *CurrentBB = BB;
899 BasicBlock *CurrentPred = BB->getSinglePredecessor();
902 auto &DL = BB->getModule()->getDataLayout();
904 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
905 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
906 if (!PBI || !PBI->isConditional() || PBI->getSuccessor(0) != CurrentBB)
909 if (isImpliedCondition(PBI->getCondition(), Cond, DL)) {
910 BI->getSuccessor(1)->removePredecessor(BB);
911 BranchInst::Create(BI->getSuccessor(0), BI);
912 BI->eraseFromParent();
915 CurrentBB = CurrentPred;
916 CurrentPred = CurrentBB->getSinglePredecessor();
922 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
923 /// load instruction, eliminate it by replacing it with a PHI node. This is an
924 /// important optimization that encourages jump threading, and needs to be run
925 /// interlaced with other jump threading tasks.
926 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
927 // Don't hack volatile/atomic loads.
928 if (!LI->isSimple()) return false;
930 // If the load is defined in a block with exactly one predecessor, it can't be
931 // partially redundant.
932 BasicBlock *LoadBB = LI->getParent();
933 if (LoadBB->getSinglePredecessor())
936 // If the load is defined in an EH pad, it can't be partially redundant,
937 // because the edges between the invoke and the EH pad cannot have other
938 // instructions between them.
939 if (LoadBB->isEHPad())
942 Value *LoadedPtr = LI->getOperand(0);
944 // If the loaded operand is defined in the LoadBB, it can't be available.
945 // TODO: Could do simple PHI translation, that would be fun :)
946 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
947 if (PtrOp->getParent() == LoadBB)
950 // Scan a few instructions up from the load, to see if it is obviously live at
951 // the entry to its block.
952 BasicBlock::iterator BBIt(LI);
954 if (Value *AvailableVal =
955 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, DefMaxInstsToScan)) {
956 // If the value of the load is locally available within the block, just use
957 // it. This frequently occurs for reg2mem'd allocas.
958 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
960 // If the returned value is the load itself, replace with an undef. This can
961 // only happen in dead loops.
962 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
963 if (AvailableVal->getType() != LI->getType())
965 CastInst::CreateBitOrPointerCast(AvailableVal, LI->getType(), "", LI);
966 LI->replaceAllUsesWith(AvailableVal);
967 LI->eraseFromParent();
971 // Otherwise, if we scanned the whole block and got to the top of the block,
972 // we know the block is locally transparent to the load. If not, something
973 // might clobber its value.
974 if (BBIt != LoadBB->begin())
977 // If all of the loads and stores that feed the value have the same AA tags,
978 // then we can propagate them onto any newly inserted loads.
980 LI->getAAMetadata(AATags);
982 SmallPtrSet<BasicBlock*, 8> PredsScanned;
983 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
984 AvailablePredsTy AvailablePreds;
985 BasicBlock *OneUnavailablePred = nullptr;
987 // If we got here, the loaded value is transparent through to the start of the
988 // block. Check to see if it is available in any of the predecessor blocks.
989 for (BasicBlock *PredBB : predecessors(LoadBB)) {
990 // If we already scanned this predecessor, skip it.
991 if (!PredsScanned.insert(PredBB).second)
994 // Scan the predecessor to see if the value is available in the pred.
995 BBIt = PredBB->end();
996 AAMDNodes ThisAATags;
997 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt,
999 nullptr, &ThisAATags);
1000 if (!PredAvailable) {
1001 OneUnavailablePred = PredBB;
1005 // If AA tags disagree or are not present, forget about them.
1006 if (AATags != ThisAATags) AATags = AAMDNodes();
1008 // If so, this load is partially redundant. Remember this info so that we
1009 // can create a PHI node.
1010 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1013 // If the loaded value isn't available in any predecessor, it isn't partially
1015 if (AvailablePreds.empty()) return false;
1017 // Okay, the loaded value is available in at least one (and maybe all!)
1018 // predecessors. If the value is unavailable in more than one unique
1019 // predecessor, we want to insert a merge block for those common predecessors.
1020 // This ensures that we only have to insert one reload, thus not increasing
1022 BasicBlock *UnavailablePred = nullptr;
1024 // If there is exactly one predecessor where the value is unavailable, the
1025 // already computed 'OneUnavailablePred' block is it. If it ends in an
1026 // unconditional branch, we know that it isn't a critical edge.
1027 if (PredsScanned.size() == AvailablePreds.size()+1 &&
1028 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1029 UnavailablePred = OneUnavailablePred;
1030 } else if (PredsScanned.size() != AvailablePreds.size()) {
1031 // Otherwise, we had multiple unavailable predecessors or we had a critical
1032 // edge from the one.
1033 SmallVector<BasicBlock*, 8> PredsToSplit;
1034 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1036 for (const auto &AvailablePred : AvailablePreds)
1037 AvailablePredSet.insert(AvailablePred.first);
1039 // Add all the unavailable predecessors to the PredsToSplit list.
1040 for (BasicBlock *P : predecessors(LoadBB)) {
1041 // If the predecessor is an indirect goto, we can't split the edge.
1042 if (isa<IndirectBrInst>(P->getTerminator()))
1045 if (!AvailablePredSet.count(P))
1046 PredsToSplit.push_back(P);
1049 // Split them out to their own block.
1050 UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1053 // If the value isn't available in all predecessors, then there will be
1054 // exactly one where it isn't available. Insert a load on that edge and add
1055 // it to the AvailablePreds list.
1056 if (UnavailablePred) {
1057 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1058 "Can't handle critical edge here!");
1059 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
1061 UnavailablePred->getTerminator());
1062 NewVal->setDebugLoc(LI->getDebugLoc());
1064 NewVal->setAAMetadata(AATags);
1066 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1069 // Now we know that each predecessor of this block has a value in
1070 // AvailablePreds, sort them for efficient access as we're walking the preds.
1071 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1073 // Create a PHI node at the start of the block for the PRE'd load value.
1074 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1075 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
1078 PN->setDebugLoc(LI->getDebugLoc());
1080 // Insert new entries into the PHI for each predecessor. A single block may
1081 // have multiple entries here.
1082 for (pred_iterator PI = PB; PI != PE; ++PI) {
1083 BasicBlock *P = *PI;
1084 AvailablePredsTy::iterator I =
1085 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1086 std::make_pair(P, (Value*)nullptr));
1088 assert(I != AvailablePreds.end() && I->first == P &&
1089 "Didn't find entry for predecessor!");
1091 // If we have an available predecessor but it requires casting, insert the
1092 // cast in the predecessor and use the cast. Note that we have to update the
1093 // AvailablePreds vector as we go so that all of the PHI entries for this
1094 // predecessor use the same bitcast.
1095 Value *&PredV = I->second;
1096 if (PredV->getType() != LI->getType())
1097 PredV = CastInst::CreateBitOrPointerCast(PredV, LI->getType(), "",
1098 P->getTerminator());
1100 PN->addIncoming(PredV, I->first);
1103 //cerr << "PRE: " << *LI << *PN << "\n";
1105 LI->replaceAllUsesWith(PN);
1106 LI->eraseFromParent();
1111 /// FindMostPopularDest - The specified list contains multiple possible
1112 /// threadable destinations. Pick the one that occurs the most frequently in
1115 FindMostPopularDest(BasicBlock *BB,
1116 const SmallVectorImpl<std::pair<BasicBlock*,
1117 BasicBlock*> > &PredToDestList) {
1118 assert(!PredToDestList.empty());
1120 // Determine popularity. If there are multiple possible destinations, we
1121 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1122 // blocks with known and real destinations to threading undef. We'll handle
1123 // them later if interesting.
1124 DenseMap<BasicBlock*, unsigned> DestPopularity;
1125 for (const auto &PredToDest : PredToDestList)
1126 if (PredToDest.second)
1127 DestPopularity[PredToDest.second]++;
1129 // Find the most popular dest.
1130 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1131 BasicBlock *MostPopularDest = DPI->first;
1132 unsigned Popularity = DPI->second;
1133 SmallVector<BasicBlock*, 4> SamePopularity;
1135 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1136 // If the popularity of this entry isn't higher than the popularity we've
1137 // seen so far, ignore it.
1138 if (DPI->second < Popularity)
1140 else if (DPI->second == Popularity) {
1141 // If it is the same as what we've seen so far, keep track of it.
1142 SamePopularity.push_back(DPI->first);
1144 // If it is more popular, remember it.
1145 SamePopularity.clear();
1146 MostPopularDest = DPI->first;
1147 Popularity = DPI->second;
1151 // Okay, now we know the most popular destination. If there is more than one
1152 // destination, we need to determine one. This is arbitrary, but we need
1153 // to make a deterministic decision. Pick the first one that appears in the
1155 if (!SamePopularity.empty()) {
1156 SamePopularity.push_back(MostPopularDest);
1157 TerminatorInst *TI = BB->getTerminator();
1158 for (unsigned i = 0; ; ++i) {
1159 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1161 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1162 TI->getSuccessor(i)) == SamePopularity.end())
1165 MostPopularDest = TI->getSuccessor(i);
1170 // Okay, we have finally picked the most popular destination.
1171 return MostPopularDest;
1174 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1175 ConstantPreference Preference,
1176 Instruction *CxtI) {
1177 // If threading this would thread across a loop header, don't even try to
1179 if (LoopHeaders.count(BB))
1182 PredValueInfoTy PredValues;
1183 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1186 assert(!PredValues.empty() &&
1187 "ComputeValueKnownInPredecessors returned true with no values");
1189 DEBUG(dbgs() << "IN BB: " << *BB;
1190 for (const auto &PredValue : PredValues) {
1191 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1193 << " for pred '" << PredValue.second->getName() << "'.\n";
1196 // Decide what we want to thread through. Convert our list of known values to
1197 // a list of known destinations for each pred. This also discards duplicate
1198 // predecessors and keeps track of the undefined inputs (which are represented
1199 // as a null dest in the PredToDestList).
1200 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1201 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1203 BasicBlock *OnlyDest = nullptr;
1204 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1206 for (const auto &PredValue : PredValues) {
1207 BasicBlock *Pred = PredValue.second;
1208 if (!SeenPreds.insert(Pred).second)
1209 continue; // Duplicate predecessor entry.
1211 // If the predecessor ends with an indirect goto, we can't change its
1213 if (isa<IndirectBrInst>(Pred->getTerminator()))
1216 Constant *Val = PredValue.first;
1219 if (isa<UndefValue>(Val))
1221 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1222 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1223 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1224 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1226 assert(isa<IndirectBrInst>(BB->getTerminator())
1227 && "Unexpected terminator");
1228 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1231 // If we have exactly one destination, remember it for efficiency below.
1232 if (PredToDestList.empty())
1234 else if (OnlyDest != DestBB)
1235 OnlyDest = MultipleDestSentinel;
1237 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1240 // If all edges were unthreadable, we fail.
1241 if (PredToDestList.empty())
1244 // Determine which is the most common successor. If we have many inputs and
1245 // this block is a switch, we want to start by threading the batch that goes
1246 // to the most popular destination first. If we only know about one
1247 // threadable destination (the common case) we can avoid this.
1248 BasicBlock *MostPopularDest = OnlyDest;
1250 if (MostPopularDest == MultipleDestSentinel)
1251 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1253 // Now that we know what the most popular destination is, factor all
1254 // predecessors that will jump to it into a single predecessor.
1255 SmallVector<BasicBlock*, 16> PredsToFactor;
1256 for (const auto &PredToDest : PredToDestList)
1257 if (PredToDest.second == MostPopularDest) {
1258 BasicBlock *Pred = PredToDest.first;
1260 // This predecessor may be a switch or something else that has multiple
1261 // edges to the block. Factor each of these edges by listing them
1262 // according to # occurrences in PredsToFactor.
1263 for (BasicBlock *Succ : successors(Pred))
1265 PredsToFactor.push_back(Pred);
1268 // If the threadable edges are branching on an undefined value, we get to pick
1269 // the destination that these predecessors should get to.
1270 if (!MostPopularDest)
1271 MostPopularDest = BB->getTerminator()->
1272 getSuccessor(GetBestDestForJumpOnUndef(BB));
1274 // Ok, try to thread it!
1275 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1278 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1279 /// a PHI node in the current block. See if there are any simplifications we
1280 /// can do based on inputs to the phi node.
1282 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1283 BasicBlock *BB = PN->getParent();
1285 // TODO: We could make use of this to do it once for blocks with common PHI
1287 SmallVector<BasicBlock*, 1> PredBBs;
1290 // If any of the predecessor blocks end in an unconditional branch, we can
1291 // *duplicate* the conditional branch into that block in order to further
1292 // encourage jump threading and to eliminate cases where we have branch on a
1293 // phi of an icmp (branch on icmp is much better).
1294 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1295 BasicBlock *PredBB = PN->getIncomingBlock(i);
1296 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1297 if (PredBr->isUnconditional()) {
1298 PredBBs[0] = PredBB;
1299 // Try to duplicate BB into PredBB.
1300 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1308 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1309 /// a xor instruction in the current block. See if there are any
1310 /// simplifications we can do based on inputs to the xor.
1312 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1313 BasicBlock *BB = BO->getParent();
1315 // If either the LHS or RHS of the xor is a constant, don't do this
1317 if (isa<ConstantInt>(BO->getOperand(0)) ||
1318 isa<ConstantInt>(BO->getOperand(1)))
1321 // If the first instruction in BB isn't a phi, we won't be able to infer
1322 // anything special about any particular predecessor.
1323 if (!isa<PHINode>(BB->front()))
1326 // If we have a xor as the branch input to this block, and we know that the
1327 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1328 // the condition into the predecessor and fix that value to true, saving some
1329 // logical ops on that path and encouraging other paths to simplify.
1331 // This copies something like this:
1334 // %X = phi i1 [1], [%X']
1335 // %Y = icmp eq i32 %A, %B
1336 // %Z = xor i1 %X, %Y
1341 // %Y = icmp ne i32 %A, %B
1344 PredValueInfoTy XorOpValues;
1346 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1348 assert(XorOpValues.empty());
1349 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1355 assert(!XorOpValues.empty() &&
1356 "ComputeValueKnownInPredecessors returned true with no values");
1358 // Scan the information to see which is most popular: true or false. The
1359 // predecessors can be of the set true, false, or undef.
1360 unsigned NumTrue = 0, NumFalse = 0;
1361 for (const auto &XorOpValue : XorOpValues) {
1362 if (isa<UndefValue>(XorOpValue.first))
1363 // Ignore undefs for the count.
1365 if (cast<ConstantInt>(XorOpValue.first)->isZero())
1371 // Determine which value to split on, true, false, or undef if neither.
1372 ConstantInt *SplitVal = nullptr;
1373 if (NumTrue > NumFalse)
1374 SplitVal = ConstantInt::getTrue(BB->getContext());
1375 else if (NumTrue != 0 || NumFalse != 0)
1376 SplitVal = ConstantInt::getFalse(BB->getContext());
1378 // Collect all of the blocks that this can be folded into so that we can
1379 // factor this once and clone it once.
1380 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1381 for (const auto &XorOpValue : XorOpValues) {
1382 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1385 BlocksToFoldInto.push_back(XorOpValue.second);
1388 // If we inferred a value for all of the predecessors, then duplication won't
1389 // help us. However, we can just replace the LHS or RHS with the constant.
1390 if (BlocksToFoldInto.size() ==
1391 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1393 // If all preds provide undef, just nuke the xor, because it is undef too.
1394 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1395 BO->eraseFromParent();
1396 } else if (SplitVal->isZero()) {
1397 // If all preds provide 0, replace the xor with the other input.
1398 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1399 BO->eraseFromParent();
1401 // If all preds provide 1, set the computed value to 1.
1402 BO->setOperand(!isLHS, SplitVal);
1408 // Try to duplicate BB into PredBB.
1409 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1413 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1414 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1415 /// NewPred using the entries from OldPred (suitably mapped).
1416 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1417 BasicBlock *OldPred,
1418 BasicBlock *NewPred,
1419 DenseMap<Instruction*, Value*> &ValueMap) {
1420 for (BasicBlock::iterator PNI = PHIBB->begin();
1421 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1422 // Ok, we have a PHI node. Figure out what the incoming value was for the
1424 Value *IV = PN->getIncomingValueForBlock(OldPred);
1426 // Remap the value if necessary.
1427 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1428 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1429 if (I != ValueMap.end())
1433 PN->addIncoming(IV, NewPred);
1437 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1438 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1439 /// across BB. Transform the IR to reflect this change.
1440 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1441 const SmallVectorImpl<BasicBlock*> &PredBBs,
1442 BasicBlock *SuccBB) {
1443 // If threading to the same block as we come from, we would infinite loop.
1445 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1446 << "' - would thread to self!\n");
1450 // If threading this would thread across a loop header, don't thread the edge.
1451 // See the comments above FindLoopHeaders for justifications and caveats.
1452 if (LoopHeaders.count(BB)) {
1453 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1454 << "' to dest BB '" << SuccBB->getName()
1455 << "' - it might create an irreducible loop!\n");
1459 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, BBDupThreshold);
1460 if (JumpThreadCost > BBDupThreshold) {
1461 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1462 << "' - Cost is too high: " << JumpThreadCost << "\n");
1466 // And finally, do it! Start by factoring the predecessors if needed.
1468 if (PredBBs.size() == 1)
1469 PredBB = PredBBs[0];
1471 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1472 << " common predecessors.\n");
1473 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1476 // And finally, do it!
1477 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1478 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1479 << ", across block:\n "
1482 LVI->threadEdge(PredBB, BB, SuccBB);
1484 // We are going to have to map operands from the original BB block to the new
1485 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1486 // account for entry from PredBB.
1487 DenseMap<Instruction*, Value*> ValueMapping;
1489 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1490 BB->getName()+".thread",
1491 BB->getParent(), BB);
1492 NewBB->moveAfter(PredBB);
1494 // Set the block frequency of NewBB.
1495 if (HasProfileData) {
1497 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
1498 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
1501 BasicBlock::iterator BI = BB->begin();
1502 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1503 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1505 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1506 // mapping and using it to remap operands in the cloned instructions.
1507 for (; !isa<TerminatorInst>(BI); ++BI) {
1508 Instruction *New = BI->clone();
1509 New->setName(BI->getName());
1510 NewBB->getInstList().push_back(New);
1511 ValueMapping[&*BI] = New;
1513 // Remap operands to patch up intra-block references.
1514 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1515 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1516 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1517 if (I != ValueMapping.end())
1518 New->setOperand(i, I->second);
1522 // We didn't copy the terminator from BB over to NewBB, because there is now
1523 // an unconditional jump to SuccBB. Insert the unconditional jump.
1524 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
1525 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1527 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1528 // PHI nodes for NewBB now.
1529 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1531 // If there were values defined in BB that are used outside the block, then we
1532 // now have to update all uses of the value to use either the original value,
1533 // the cloned value, or some PHI derived value. This can require arbitrary
1534 // PHI insertion, of which we are prepared to do, clean these up now.
1535 SSAUpdater SSAUpdate;
1536 SmallVector<Use*, 16> UsesToRename;
1537 for (Instruction &I : *BB) {
1538 // Scan all uses of this instruction to see if it is used outside of its
1539 // block, and if so, record them in UsesToRename.
1540 for (Use &U : I.uses()) {
1541 Instruction *User = cast<Instruction>(U.getUser());
1542 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1543 if (UserPN->getIncomingBlock(U) == BB)
1545 } else if (User->getParent() == BB)
1548 UsesToRename.push_back(&U);
1551 // If there are no uses outside the block, we're done with this instruction.
1552 if (UsesToRename.empty())
1555 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
1557 // We found a use of I outside of BB. Rename all uses of I that are outside
1558 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1559 // with the two values we know.
1560 SSAUpdate.Initialize(I.getType(), I.getName());
1561 SSAUpdate.AddAvailableValue(BB, &I);
1562 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
1564 while (!UsesToRename.empty())
1565 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1566 DEBUG(dbgs() << "\n");
1570 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1571 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1572 // us to simplify any PHI nodes in BB.
1573 TerminatorInst *PredTerm = PredBB->getTerminator();
1574 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1575 if (PredTerm->getSuccessor(i) == BB) {
1576 BB->removePredecessor(PredBB, true);
1577 PredTerm->setSuccessor(i, NewBB);
1580 // At this point, the IR is fully up to date and consistent. Do a quick scan
1581 // over the new instructions and zap any that are constants or dead. This
1582 // frequently happens because of phi translation.
1583 SimplifyInstructionsInBlock(NewBB, TLI);
1585 // Update the edge weight from BB to SuccBB, which should be less than before.
1586 UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
1588 // Threaded an edge!
1593 /// Create a new basic block that will be the predecessor of BB and successor of
1594 /// all blocks in Preds. When profile data is availble, update the frequency of
1596 BasicBlock *JumpThreading::SplitBlockPreds(BasicBlock *BB,
1597 ArrayRef<BasicBlock *> Preds,
1598 const char *Suffix) {
1599 // Collect the frequencies of all predecessors of BB, which will be used to
1600 // update the edge weight on BB->SuccBB.
1601 BlockFrequency PredBBFreq(0);
1603 for (auto Pred : Preds)
1604 PredBBFreq += BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB);
1606 BasicBlock *PredBB = SplitBlockPredecessors(BB, Preds, Suffix);
1608 // Set the block frequency of the newly created PredBB, which is the sum of
1609 // frequencies of Preds.
1611 BFI->setBlockFreq(PredBB, PredBBFreq.getFrequency());
1615 /// Update the block frequency of BB and branch weight and the metadata on the
1616 /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
1617 /// Freq(PredBB->BB) / Freq(BB->SuccBB).
1618 void JumpThreading::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
1621 BasicBlock *SuccBB) {
1622 if (!HasProfileData)
1625 assert(BFI && BPI && "BFI & BPI should have been created here");
1627 // As the edge from PredBB to BB is deleted, we have to update the block
1629 auto BBOrigFreq = BFI->getBlockFreq(BB);
1630 auto NewBBFreq = BFI->getBlockFreq(NewBB);
1631 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
1632 auto BBNewFreq = BBOrigFreq - NewBBFreq;
1633 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
1635 // Collect updated outgoing edges' frequencies from BB and use them to update
1636 // edge probabilities.
1637 SmallVector<uint64_t, 4> BBSuccFreq;
1638 for (BasicBlock *Succ : successors(BB)) {
1639 auto SuccFreq = (Succ == SuccBB)
1640 ? BB2SuccBBFreq - NewBBFreq
1641 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
1642 BBSuccFreq.push_back(SuccFreq.getFrequency());
1645 uint64_t MaxBBSuccFreq =
1646 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
1648 SmallVector<BranchProbability, 4> BBSuccProbs;
1649 if (MaxBBSuccFreq == 0)
1650 BBSuccProbs.assign(BBSuccFreq.size(),
1651 {1, static_cast<unsigned>(BBSuccFreq.size())});
1653 for (uint64_t Freq : BBSuccFreq)
1654 BBSuccProbs.push_back(
1655 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
1656 // Normalize edge probabilities so that they sum up to one.
1657 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
1661 // Update edge probabilities in BPI.
1662 for (int I = 0, E = BBSuccProbs.size(); I < E; I++)
1663 BPI->setEdgeProbability(BB, I, BBSuccProbs[I]);
1665 if (BBSuccProbs.size() >= 2) {
1666 SmallVector<uint32_t, 4> Weights;
1667 for (auto Prob : BBSuccProbs)
1668 Weights.push_back(Prob.getNumerator());
1670 auto TI = BB->getTerminator();
1672 LLVMContext::MD_prof,
1673 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
1677 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1678 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1679 /// If we can duplicate the contents of BB up into PredBB do so now, this
1680 /// improves the odds that the branch will be on an analyzable instruction like
1682 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1683 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1684 assert(!PredBBs.empty() && "Can't handle an empty set");
1686 // If BB is a loop header, then duplicating this block outside the loop would
1687 // cause us to transform this into an irreducible loop, don't do this.
1688 // See the comments above FindLoopHeaders for justifications and caveats.
1689 if (LoopHeaders.count(BB)) {
1690 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1691 << "' into predecessor block '" << PredBBs[0]->getName()
1692 << "' - it might create an irreducible loop!\n");
1696 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, BBDupThreshold);
1697 if (DuplicationCost > BBDupThreshold) {
1698 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1699 << "' - Cost is too high: " << DuplicationCost << "\n");
1703 // And finally, do it! Start by factoring the predecessors if needed.
1705 if (PredBBs.size() == 1)
1706 PredBB = PredBBs[0];
1708 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1709 << " common predecessors.\n");
1710 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1713 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1715 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1716 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1717 << DuplicationCost << " block is:" << *BB << "\n");
1719 // Unless PredBB ends with an unconditional branch, split the edge so that we
1720 // can just clone the bits from BB into the end of the new PredBB.
1721 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1723 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
1724 PredBB = SplitEdge(PredBB, BB);
1725 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1728 // We are going to have to map operands from the original BB block into the
1729 // PredBB block. Evaluate PHI nodes in BB.
1730 DenseMap<Instruction*, Value*> ValueMapping;
1732 BasicBlock::iterator BI = BB->begin();
1733 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1734 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1735 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1736 // mapping and using it to remap operands in the cloned instructions.
1737 for (; BI != BB->end(); ++BI) {
1738 Instruction *New = BI->clone();
1740 // Remap operands to patch up intra-block references.
1741 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1742 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1743 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1744 if (I != ValueMapping.end())
1745 New->setOperand(i, I->second);
1748 // If this instruction can be simplified after the operands are updated,
1749 // just use the simplified value instead. This frequently happens due to
1752 SimplifyInstruction(New, BB->getModule()->getDataLayout())) {
1754 ValueMapping[&*BI] = IV;
1756 // Otherwise, insert the new instruction into the block.
1757 New->setName(BI->getName());
1758 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
1759 ValueMapping[&*BI] = New;
1763 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1764 // add entries to the PHI nodes for branch from PredBB now.
1765 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1766 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1768 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1771 // If there were values defined in BB that are used outside the block, then we
1772 // now have to update all uses of the value to use either the original value,
1773 // the cloned value, or some PHI derived value. This can require arbitrary
1774 // PHI insertion, of which we are prepared to do, clean these up now.
1775 SSAUpdater SSAUpdate;
1776 SmallVector<Use*, 16> UsesToRename;
1777 for (Instruction &I : *BB) {
1778 // Scan all uses of this instruction to see if it is used outside of its
1779 // block, and if so, record them in UsesToRename.
1780 for (Use &U : I.uses()) {
1781 Instruction *User = cast<Instruction>(U.getUser());
1782 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1783 if (UserPN->getIncomingBlock(U) == BB)
1785 } else if (User->getParent() == BB)
1788 UsesToRename.push_back(&U);
1791 // If there are no uses outside the block, we're done with this instruction.
1792 if (UsesToRename.empty())
1795 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
1797 // We found a use of I outside of BB. Rename all uses of I that are outside
1798 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1799 // with the two values we know.
1800 SSAUpdate.Initialize(I.getType(), I.getName());
1801 SSAUpdate.AddAvailableValue(BB, &I);
1802 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&I]);
1804 while (!UsesToRename.empty())
1805 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1806 DEBUG(dbgs() << "\n");
1809 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1811 BB->removePredecessor(PredBB, true);
1813 // Remove the unconditional branch at the end of the PredBB block.
1814 OldPredBranch->eraseFromParent();
1820 /// TryToUnfoldSelect - Look for blocks of the form
1826 /// %p = phi [%a, %bb] ...
1830 /// And expand the select into a branch structure if one of its arms allows %c
1831 /// to be folded. This later enables threading from bb1 over bb2.
1832 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1833 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1834 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1835 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1837 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1838 CondLHS->getParent() != BB)
1841 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1842 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1843 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1845 // Look if one of the incoming values is a select in the corresponding
1847 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1850 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1851 if (!PredTerm || !PredTerm->isUnconditional())
1854 // Now check if one of the select values would allow us to constant fold the
1855 // terminator in BB. We don't do the transform if both sides fold, those
1856 // cases will be threaded in any case.
1857 LazyValueInfo::Tristate LHSFolds =
1858 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1859 CondRHS, Pred, BB, CondCmp);
1860 LazyValueInfo::Tristate RHSFolds =
1861 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
1862 CondRHS, Pred, BB, CondCmp);
1863 if ((LHSFolds != LazyValueInfo::Unknown ||
1864 RHSFolds != LazyValueInfo::Unknown) &&
1865 LHSFolds != RHSFolds) {
1866 // Expand the select.
1875 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
1876 BB->getParent(), BB);
1877 // Move the unconditional branch to NewBB.
1878 PredTerm->removeFromParent();
1879 NewBB->getInstList().insert(NewBB->end(), PredTerm);
1880 // Create a conditional branch and update PHI nodes.
1881 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
1882 CondLHS->setIncomingValue(I, SI->getFalseValue());
1883 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
1884 // The select is now dead.
1885 SI->eraseFromParent();
1887 // Update any other PHI nodes in BB.
1888 for (BasicBlock::iterator BI = BB->begin();
1889 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
1891 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
1898 /// TryToUnfoldSelectInCurrBB - Look for PHI/Select in the same BB of the form
1900 /// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
1901 /// %s = select p, trueval, falseval
1903 /// And expand the select into a branch structure. This later enables
1904 /// jump-threading over bb in this pass.
1906 /// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
1907 /// select if the associated PHI has at least one constant. If the unfolded
1908 /// select is not jump-threaded, it will be folded again in the later
1910 bool JumpThreading::TryToUnfoldSelectInCurrBB(BasicBlock *BB) {
1911 // If threading this would thread across a loop header, don't thread the edge.
1912 // See the comments above FindLoopHeaders for justifications and caveats.
1913 if (LoopHeaders.count(BB))
1916 // Look for a Phi/Select pair in the same basic block. The Phi feeds the
1917 // condition of the Select and at least one of the incoming values is a
1919 for (BasicBlock::iterator BI = BB->begin();
1920 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
1921 unsigned NumPHIValues = PN->getNumIncomingValues();
1922 if (NumPHIValues == 0 || !PN->hasOneUse())
1925 SelectInst *SI = dyn_cast<SelectInst>(PN->user_back());
1926 if (!SI || SI->getParent() != BB)
1929 Value *Cond = SI->getCondition();
1930 if (!Cond || Cond != PN || !Cond->getType()->isIntegerTy(1))
1933 bool HasConst = false;
1934 for (unsigned i = 0; i != NumPHIValues; ++i) {
1935 if (PN->getIncomingBlock(i) == BB)
1937 if (isa<ConstantInt>(PN->getIncomingValue(i)))
1942 // Expand the select.
1943 TerminatorInst *Term =
1944 SplitBlockAndInsertIfThen(SI->getCondition(), SI, false);
1945 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
1946 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
1947 NewPN->addIncoming(SI->getFalseValue(), BB);
1948 SI->replaceAllUsesWith(NewPN);
1949 SI->eraseFromParent();