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<BasicBlock*, 16> LoopHeaders;
105 SmallSet<AssertingVH<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);
168 BasicBlock *SplitBlockPreds(BasicBlock *BB, ArrayRef<BasicBlock *> Preds,
170 void UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB, BasicBlock *BB,
171 BasicBlock *NewBB, BasicBlock *SuccBB);
175 char JumpThreading::ID = 0;
176 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
177 "Jump Threading", false, false)
178 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)
179 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
180 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
181 "Jump Threading", false, false)
183 // Public interface to the Jump Threading pass
184 FunctionPass *llvm::createJumpThreadingPass(int Threshold) { return new JumpThreading(Threshold); }
186 /// runOnFunction - Top level algorithm.
188 bool JumpThreading::runOnFunction(Function &F) {
189 if (skipOptnoneFunction(F))
192 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
193 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
194 LVI = &getAnalysis<LazyValueInfo>();
197 // When profile data is available, we need to update edge weights after
198 // successful jump threading, which requires both BPI and BFI being available.
199 HasProfileData = F.getEntryCount().hasValue();
200 if (HasProfileData) {
201 LoopInfo LI{DominatorTree(F)};
202 BPI.reset(new BranchProbabilityInfo(F, LI));
203 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
206 // Remove unreachable blocks from function as they may result in infinite
207 // loop. We do threading if we found something profitable. Jump threading a
208 // branch can create other opportunities. If these opportunities form a cycle
209 // i.e. if any jump threading is undoing previous threading in the path, then
210 // we will loop forever. We take care of this issue by not jump threading for
211 // back edges. This works for normal cases but not for unreachable blocks as
212 // they may have cycle with no back edge.
213 removeUnreachableBlocks(F);
217 bool Changed, EverChanged = false;
220 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
221 BasicBlock *BB = &*I;
222 // Thread all of the branches we can over this block.
223 while (ProcessBlock(BB))
228 // If the block is trivially dead, zap it. This eliminates the successor
229 // edges which simplifies the CFG.
230 if (pred_empty(BB) &&
231 BB != &BB->getParent()->getEntryBlock()) {
232 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
233 << "' with terminator: " << *BB->getTerminator() << '\n');
234 LoopHeaders.erase(BB);
241 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
243 // Can't thread an unconditional jump, but if the block is "almost
244 // empty", we can replace uses of it with uses of the successor and make
246 if (BI && BI->isUnconditional() &&
247 BB != &BB->getParent()->getEntryBlock() &&
248 // If the terminator is the only non-phi instruction, try to nuke it.
249 BB->getFirstNonPHIOrDbg()->isTerminator()) {
250 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
251 // block, we have to make sure it isn't in the LoopHeaders set. We
252 // reinsert afterward if needed.
253 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
254 BasicBlock *Succ = BI->getSuccessor(0);
256 // FIXME: It is always conservatively correct to drop the info
257 // for a block even if it doesn't get erased. This isn't totally
258 // awesome, but it allows us to use AssertingVH to prevent nasty
259 // dangling pointer issues within LazyValueInfo.
261 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
263 // If we deleted BB and BB was the header of a loop, then the
264 // successor is now the header of the loop.
268 if (ErasedFromLoopHeaders)
269 LoopHeaders.insert(BB);
272 EverChanged |= Changed;
279 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
280 /// thread across it. Stop scanning the block when passing the threshold.
281 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
282 unsigned Threshold) {
283 /// Ignore PHI nodes, these will be flattened when duplication happens.
284 BasicBlock::const_iterator I(BB->getFirstNonPHI());
286 // FIXME: THREADING will delete values that are just used to compute the
287 // branch, so they shouldn't count against the duplication cost.
290 const TerminatorInst *BBTerm = BB->getTerminator();
291 // Threading through a switch statement is particularly profitable. If this
292 // block ends in a switch, decrease its cost to make it more likely to happen.
293 if (isa<SwitchInst>(BBTerm))
296 // The same holds for indirect branches, but slightly more so.
297 if (isa<IndirectBrInst>(BBTerm))
300 // Bump the threshold up so the early exit from the loop doesn't skip the
301 // terminator-based Size adjustment at the end.
304 // Sum up the cost of each instruction until we get to the terminator. Don't
305 // include the terminator because the copy won't include it.
307 for (; !isa<TerminatorInst>(I); ++I) {
309 // Stop scanning the block if we've reached the threshold.
310 if (Size > Threshold)
313 // Debugger intrinsics don't incur code size.
314 if (isa<DbgInfoIntrinsic>(I)) continue;
316 // If this is a pointer->pointer bitcast, it is free.
317 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
320 // Bail out if this instruction gives back a token type, it is not possible
321 // to duplicate it if it is used outside this BB.
322 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
325 // All other instructions count for at least one unit.
328 // Calls are more expensive. If they are non-intrinsic calls, we model them
329 // as having cost of 4. If they are a non-vector intrinsic, we model them
330 // as having cost of 2 total, and if they are a vector intrinsic, we model
331 // them as having cost 1.
332 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
333 if (CI->cannotDuplicate() || CI->isConvergent())
334 // Blocks with NoDuplicate are modelled as having infinite cost, so they
335 // are never duplicated.
337 else if (!isa<IntrinsicInst>(CI))
339 else if (!CI->getType()->isVectorTy())
344 return Size > Bonus ? Size - Bonus : 0;
347 /// FindLoopHeaders - We do not want jump threading to turn proper loop
348 /// structures into irreducible loops. Doing this breaks up the loop nesting
349 /// hierarchy and pessimizes later transformations. To prevent this from
350 /// happening, we first have to find the loop headers. Here we approximate this
351 /// by finding targets of backedges in the CFG.
353 /// Note that there definitely are cases when we want to allow threading of
354 /// edges across a loop header. For example, threading a jump from outside the
355 /// loop (the preheader) to an exit block of the loop is definitely profitable.
356 /// It is also almost always profitable to thread backedges from within the loop
357 /// to exit blocks, and is often profitable to thread backedges to other blocks
358 /// within the loop (forming a nested loop). This simple analysis is not rich
359 /// enough to track all of these properties and keep it up-to-date as the CFG
360 /// mutates, so we don't allow any of these transformations.
362 void JumpThreading::FindLoopHeaders(Function &F) {
363 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
364 FindFunctionBackedges(F, Edges);
366 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
367 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
370 /// getKnownConstant - Helper method to determine if we can thread over a
371 /// terminator with the given value as its condition, and if so what value to
372 /// use for that. What kind of value this is depends on whether we want an
373 /// integer or a block address, but an undef is always accepted.
374 /// Returns null if Val is null or not an appropriate constant.
375 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
379 // Undef is "known" enough.
380 if (UndefValue *U = dyn_cast<UndefValue>(Val))
383 if (Preference == WantBlockAddress)
384 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
386 return dyn_cast<ConstantInt>(Val);
389 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
390 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
391 /// in any of our predecessors. If so, return the known list of value and pred
392 /// BB in the result vector.
394 /// This returns true if there were any known values.
397 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
398 ConstantPreference Preference,
400 // This method walks up use-def chains recursively. Because of this, we could
401 // get into an infinite loop going around loops in the use-def chain. To
402 // prevent this, keep track of what (value, block) pairs we've already visited
403 // and terminate the search if we loop back to them
404 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
407 // An RAII help to remove this pair from the recursion set once the recursion
408 // stack pops back out again.
409 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
411 // If V is a constant, then it is known in all predecessors.
412 if (Constant *KC = getKnownConstant(V, Preference)) {
413 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
414 Result.push_back(std::make_pair(KC, *PI));
419 // If V is a non-instruction value, or an instruction in a different block,
420 // then it can't be derived from a PHI.
421 Instruction *I = dyn_cast<Instruction>(V);
422 if (!I || I->getParent() != BB) {
424 // Okay, if this is a live-in value, see if it has a known value at the end
425 // of any of our predecessors.
427 // FIXME: This should be an edge property, not a block end property.
428 /// TODO: Per PR2563, we could infer value range information about a
429 /// predecessor based on its terminator.
431 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
432 // "I" is a non-local compare-with-a-constant instruction. This would be
433 // able to handle value inequalities better, for example if the compare is
434 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
435 // Perhaps getConstantOnEdge should be smart enough to do this?
437 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
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 (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
495 if (LHSVals[i].first == InterestingVal ||
496 isa<UndefValue>(LHSVals[i].first)) {
497 Result.push_back(LHSVals[i]);
498 Result.back().first = InterestingVal;
499 LHSKnownBBs.insert(LHSVals[i].second);
501 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
502 if (RHSVals[i].first == InterestingVal ||
503 isa<UndefValue>(RHSVals[i].first)) {
504 // If we already inferred a value for this block on the LHS, don't
506 if (!LHSKnownBBs.count(RHSVals[i].second)) {
507 Result.push_back(RHSVals[i]);
508 Result.back().first = InterestingVal;
512 return !Result.empty();
515 // Handle the NOT form of XOR.
516 if (I->getOpcode() == Instruction::Xor &&
517 isa<ConstantInt>(I->getOperand(1)) &&
518 cast<ConstantInt>(I->getOperand(1))->isOne()) {
519 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
524 // Invert the known values.
525 for (unsigned i = 0, e = Result.size(); i != e; ++i)
526 Result[i].first = ConstantExpr::getNot(Result[i].first);
531 // Try to simplify some other binary operator values.
532 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
533 assert(Preference != WantBlockAddress
534 && "A binary operator creating a block address?");
535 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
536 PredValueInfoTy LHSVals;
537 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
540 // Try to use constant folding to simplify the binary operator.
541 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
542 Constant *V = LHSVals[i].first;
543 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
545 if (Constant *KC = getKnownConstant(Folded, WantInteger))
546 Result.push_back(std::make_pair(KC, LHSVals[i].second));
550 return !Result.empty();
553 // Handle compare with phi operand, where the PHI is defined in this block.
554 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
555 assert(Preference == WantInteger && "Compares only produce integers");
556 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
557 if (PN && PN->getParent() == BB) {
558 const DataLayout &DL = PN->getModule()->getDataLayout();
559 // We can do this simplification if any comparisons fold to true or false.
561 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
562 BasicBlock *PredBB = PN->getIncomingBlock(i);
563 Value *LHS = PN->getIncomingValue(i);
564 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
566 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL);
568 if (!isa<Constant>(RHS))
571 LazyValueInfo::Tristate
572 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
573 cast<Constant>(RHS), PredBB, BB,
575 if (ResT == LazyValueInfo::Unknown)
577 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
580 if (Constant *KC = getKnownConstant(Res, WantInteger))
581 Result.push_back(std::make_pair(KC, PredBB));
584 return !Result.empty();
587 // If comparing a live-in value against a constant, see if we know the
588 // live-in value on any predecessors.
589 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
590 if (!isa<Instruction>(Cmp->getOperand(0)) ||
591 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
592 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
594 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
596 // If the value is known by LazyValueInfo to be a constant in a
597 // predecessor, use that information to try to thread this block.
598 LazyValueInfo::Tristate Res =
599 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
600 RHSCst, P, BB, CxtI ? CxtI : Cmp);
601 if (Res == LazyValueInfo::Unknown)
604 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
605 Result.push_back(std::make_pair(ResC, P));
608 return !Result.empty();
611 // Try to find a constant value for the LHS of a comparison,
612 // and evaluate it statically if we can.
613 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
614 PredValueInfoTy LHSVals;
615 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
618 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
619 Constant *V = LHSVals[i].first;
620 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
622 if (Constant *KC = getKnownConstant(Folded, WantInteger))
623 Result.push_back(std::make_pair(KC, LHSVals[i].second));
626 return !Result.empty();
631 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
632 // Handle select instructions where at least one operand is a known constant
633 // and we can figure out the condition value for any predecessor block.
634 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
635 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
636 PredValueInfoTy Conds;
637 if ((TrueVal || FalseVal) &&
638 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
639 WantInteger, CxtI)) {
640 for (unsigned i = 0, e = Conds.size(); i != e; ++i) {
641 Constant *Cond = Conds[i].first;
643 // Figure out what value to use for the condition.
645 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
647 KnownCond = CI->isOne();
649 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
650 // Either operand will do, so be sure to pick the one that's a known
652 // FIXME: Do this more cleverly if both values are known constants?
653 KnownCond = (TrueVal != nullptr);
656 // See if the select has a known constant value for this predecessor.
657 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
658 Result.push_back(std::make_pair(Val, Conds[i].second));
661 return !Result.empty();
665 // If all else fails, see if LVI can figure out a constant value for us.
666 Constant *CI = LVI->getConstant(V, BB, CxtI);
667 if (Constant *KC = getKnownConstant(CI, Preference)) {
668 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
669 Result.push_back(std::make_pair(KC, *PI));
672 return !Result.empty();
677 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
678 /// in an undefined jump, decide which block is best to revector to.
680 /// Since we can pick an arbitrary destination, we pick the successor with the
681 /// fewest predecessors. This should reduce the in-degree of the others.
683 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
684 TerminatorInst *BBTerm = BB->getTerminator();
685 unsigned MinSucc = 0;
686 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
687 // Compute the successor with the minimum number of predecessors.
688 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
689 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
690 TestBB = BBTerm->getSuccessor(i);
691 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
692 if (NumPreds < MinNumPreds) {
694 MinNumPreds = NumPreds;
701 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
702 if (!BB->hasAddressTaken()) return false;
704 // If the block has its address taken, it may be a tree of dead constants
705 // hanging off of it. These shouldn't keep the block alive.
706 BlockAddress *BA = BlockAddress::get(BB);
707 BA->removeDeadConstantUsers();
708 return !BA->use_empty();
711 /// ProcessBlock - If there are any predecessors whose control can be threaded
712 /// through to a successor, transform them now.
713 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
714 // If the block is trivially dead, just return and let the caller nuke it.
715 // This simplifies other transformations.
716 if (pred_empty(BB) &&
717 BB != &BB->getParent()->getEntryBlock())
720 // If this block has a single predecessor, and if that pred has a single
721 // successor, merge the blocks. This encourages recursive jump threading
722 // because now the condition in this block can be threaded through
723 // predecessors of our predecessor block.
724 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
725 const TerminatorInst *TI = SinglePred->getTerminator();
726 if (!TI->isExceptional() && TI->getNumSuccessors() == 1 &&
727 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
728 // If SinglePred was a loop header, BB becomes one.
729 if (LoopHeaders.erase(SinglePred))
730 LoopHeaders.insert(BB);
732 LVI->eraseBlock(SinglePred);
733 MergeBasicBlockIntoOnlyPred(BB);
739 // What kind of constant we're looking for.
740 ConstantPreference Preference = WantInteger;
742 // Look to see if the terminator is a conditional branch, switch or indirect
743 // branch, if not we can't thread it.
745 Instruction *Terminator = BB->getTerminator();
746 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
747 // Can't thread an unconditional jump.
748 if (BI->isUnconditional()) return false;
749 Condition = BI->getCondition();
750 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
751 Condition = SI->getCondition();
752 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
753 // Can't thread indirect branch with no successors.
754 if (IB->getNumSuccessors() == 0) return false;
755 Condition = IB->getAddress()->stripPointerCasts();
756 Preference = WantBlockAddress;
758 return false; // Must be an invoke.
761 // Run constant folding to see if we can reduce the condition to a simple
763 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
765 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
767 I->replaceAllUsesWith(SimpleVal);
768 I->eraseFromParent();
769 Condition = SimpleVal;
773 // If the terminator is branching on an undef, we can pick any of the
774 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
775 if (isa<UndefValue>(Condition)) {
776 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
778 // Fold the branch/switch.
779 TerminatorInst *BBTerm = BB->getTerminator();
780 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
781 if (i == BestSucc) continue;
782 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
785 DEBUG(dbgs() << " In block '" << BB->getName()
786 << "' folding undef terminator: " << *BBTerm << '\n');
787 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
788 BBTerm->eraseFromParent();
792 // If the terminator of this block is branching on a constant, simplify the
793 // terminator to an unconditional branch. This can occur due to threading in
795 if (getKnownConstant(Condition, Preference)) {
796 DEBUG(dbgs() << " In block '" << BB->getName()
797 << "' folding terminator: " << *BB->getTerminator() << '\n');
799 ConstantFoldTerminator(BB, true);
803 Instruction *CondInst = dyn_cast<Instruction>(Condition);
805 // All the rest of our checks depend on the condition being an instruction.
807 // FIXME: Unify this with code below.
808 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
814 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
815 // If we're branching on a conditional, LVI might be able to determine
816 // it's value at the branch instruction. We only handle comparisons
817 // against a constant at this time.
818 // TODO: This should be extended to handle switches as well.
819 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
820 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
821 if (CondBr && CondConst && CondBr->isConditional()) {
822 LazyValueInfo::Tristate Ret =
823 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
825 if (Ret != LazyValueInfo::Unknown) {
826 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
827 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
828 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
829 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
830 CondBr->eraseFromParent();
831 if (CondCmp->use_empty())
832 CondCmp->eraseFromParent();
833 else if (CondCmp->getParent() == BB) {
834 // If the fact we just learned is true for all uses of the
835 // condition, replace it with a constant value
836 auto *CI = Ret == LazyValueInfo::True ?
837 ConstantInt::getTrue(CondCmp->getType()) :
838 ConstantInt::getFalse(CondCmp->getType());
839 CondCmp->replaceAllUsesWith(CI);
840 CondCmp->eraseFromParent();
846 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB))
850 // Check for some cases that are worth simplifying. Right now we want to look
851 // for loads that are used by a switch or by the condition for the branch. If
852 // we see one, check to see if it's partially redundant. If so, insert a PHI
853 // which can then be used to thread the values.
855 Value *SimplifyValue = CondInst;
856 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
857 if (isa<Constant>(CondCmp->getOperand(1)))
858 SimplifyValue = CondCmp->getOperand(0);
860 // TODO: There are other places where load PRE would be profitable, such as
861 // more complex comparisons.
862 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
863 if (SimplifyPartiallyRedundantLoad(LI))
867 // Handle a variety of cases where we are branching on something derived from
868 // a PHI node in the current block. If we can prove that any predecessors
869 // compute a predictable value based on a PHI node, thread those predecessors.
871 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
874 // If this is an otherwise-unfoldable branch on a phi node in the current
875 // block, see if we can simplify.
876 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
877 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
878 return ProcessBranchOnPHI(PN);
881 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
882 if (CondInst->getOpcode() == Instruction::Xor &&
883 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
884 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
886 // Search for a stronger dominating condition that can be used to simplify a
887 // conditional branch leaving BB.
888 if (ProcessImpliedCondition(BB))
894 bool JumpThreading::ProcessImpliedCondition(BasicBlock *BB) {
895 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
896 if (!BI || !BI->isConditional())
899 Value *Cond = BI->getCondition();
900 BasicBlock *CurrentBB = BB;
901 BasicBlock *CurrentPred = BB->getSinglePredecessor();
904 auto &DL = BB->getModule()->getDataLayout();
906 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
907 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
908 if (!PBI || !PBI->isConditional() || PBI->getSuccessor(0) != CurrentBB)
911 if (isImpliedCondition(PBI->getCondition(), Cond, DL)) {
912 BI->getSuccessor(1)->removePredecessor(BB);
913 BranchInst::Create(BI->getSuccessor(0), BI);
914 BI->eraseFromParent();
917 CurrentBB = CurrentPred;
918 CurrentPred = CurrentBB->getSinglePredecessor();
924 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
925 /// load instruction, eliminate it by replacing it with a PHI node. This is an
926 /// important optimization that encourages jump threading, and needs to be run
927 /// interlaced with other jump threading tasks.
928 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
929 // Don't hack volatile/atomic loads.
930 if (!LI->isSimple()) return false;
932 // If the load is defined in a block with exactly one predecessor, it can't be
933 // partially redundant.
934 BasicBlock *LoadBB = LI->getParent();
935 if (LoadBB->getSinglePredecessor())
938 // If the load is defined in an EH pad, it can't be partially redundant,
939 // because the edges between the invoke and the EH pad cannot have other
940 // instructions between them.
941 if (LoadBB->isEHPad())
944 Value *LoadedPtr = LI->getOperand(0);
946 // If the loaded operand is defined in the LoadBB, it can't be available.
947 // TODO: Could do simple PHI translation, that would be fun :)
948 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
949 if (PtrOp->getParent() == LoadBB)
952 // Scan a few instructions up from the load, to see if it is obviously live at
953 // the entry to its block.
954 BasicBlock::iterator BBIt(LI);
956 if (Value *AvailableVal =
957 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, DefMaxInstsToScan)) {
958 // If the value of the load is locally available within the block, just use
959 // it. This frequently occurs for reg2mem'd allocas.
960 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
962 // If the returned value is the load itself, replace with an undef. This can
963 // only happen in dead loops.
964 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
965 if (AvailableVal->getType() != LI->getType())
967 CastInst::CreateBitOrPointerCast(AvailableVal, LI->getType(), "", LI);
968 LI->replaceAllUsesWith(AvailableVal);
969 LI->eraseFromParent();
973 // Otherwise, if we scanned the whole block and got to the top of the block,
974 // we know the block is locally transparent to the load. If not, something
975 // might clobber its value.
976 if (BBIt != LoadBB->begin())
979 // If all of the loads and stores that feed the value have the same AA tags,
980 // then we can propagate them onto any newly inserted loads.
982 LI->getAAMetadata(AATags);
984 SmallPtrSet<BasicBlock*, 8> PredsScanned;
985 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
986 AvailablePredsTy AvailablePreds;
987 BasicBlock *OneUnavailablePred = nullptr;
989 // If we got here, the loaded value is transparent through to the start of the
990 // block. Check to see if it is available in any of the predecessor blocks.
991 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
993 BasicBlock *PredBB = *PI;
995 // If we already scanned this predecessor, skip it.
996 if (!PredsScanned.insert(PredBB).second)
999 // Scan the predecessor to see if the value is available in the pred.
1000 BBIt = PredBB->end();
1001 AAMDNodes ThisAATags;
1002 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt,
1004 nullptr, &ThisAATags);
1005 if (!PredAvailable) {
1006 OneUnavailablePred = PredBB;
1010 // If AA tags disagree or are not present, forget about them.
1011 if (AATags != ThisAATags) AATags = AAMDNodes();
1013 // If so, this load is partially redundant. Remember this info so that we
1014 // can create a PHI node.
1015 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1018 // If the loaded value isn't available in any predecessor, it isn't partially
1020 if (AvailablePreds.empty()) return false;
1022 // Okay, the loaded value is available in at least one (and maybe all!)
1023 // predecessors. If the value is unavailable in more than one unique
1024 // predecessor, we want to insert a merge block for those common predecessors.
1025 // This ensures that we only have to insert one reload, thus not increasing
1027 BasicBlock *UnavailablePred = nullptr;
1029 // If there is exactly one predecessor where the value is unavailable, the
1030 // already computed 'OneUnavailablePred' block is it. If it ends in an
1031 // unconditional branch, we know that it isn't a critical edge.
1032 if (PredsScanned.size() == AvailablePreds.size()+1 &&
1033 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1034 UnavailablePred = OneUnavailablePred;
1035 } else if (PredsScanned.size() != AvailablePreds.size()) {
1036 // Otherwise, we had multiple unavailable predecessors or we had a critical
1037 // edge from the one.
1038 SmallVector<BasicBlock*, 8> PredsToSplit;
1039 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1041 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
1042 AvailablePredSet.insert(AvailablePreds[i].first);
1044 // Add all the unavailable predecessors to the PredsToSplit list.
1045 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
1047 BasicBlock *P = *PI;
1048 // If the predecessor is an indirect goto, we can't split the edge.
1049 if (isa<IndirectBrInst>(P->getTerminator()))
1052 if (!AvailablePredSet.count(P))
1053 PredsToSplit.push_back(P);
1056 // Split them out to their own block.
1057 UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1060 // If the value isn't available in all predecessors, then there will be
1061 // exactly one where it isn't available. Insert a load on that edge and add
1062 // it to the AvailablePreds list.
1063 if (UnavailablePred) {
1064 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1065 "Can't handle critical edge here!");
1066 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
1068 UnavailablePred->getTerminator());
1069 NewVal->setDebugLoc(LI->getDebugLoc());
1071 NewVal->setAAMetadata(AATags);
1073 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1076 // Now we know that each predecessor of this block has a value in
1077 // AvailablePreds, sort them for efficient access as we're walking the preds.
1078 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1080 // Create a PHI node at the start of the block for the PRE'd load value.
1081 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1082 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
1085 PN->setDebugLoc(LI->getDebugLoc());
1087 // Insert new entries into the PHI for each predecessor. A single block may
1088 // have multiple entries here.
1089 for (pred_iterator PI = PB; PI != PE; ++PI) {
1090 BasicBlock *P = *PI;
1091 AvailablePredsTy::iterator I =
1092 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1093 std::make_pair(P, (Value*)nullptr));
1095 assert(I != AvailablePreds.end() && I->first == P &&
1096 "Didn't find entry for predecessor!");
1098 // If we have an available predecessor but it requires casting, insert the
1099 // cast in the predecessor and use the cast. Note that we have to update the
1100 // AvailablePreds vector as we go so that all of the PHI entries for this
1101 // predecessor use the same bitcast.
1102 Value *&PredV = I->second;
1103 if (PredV->getType() != LI->getType())
1104 PredV = CastInst::CreateBitOrPointerCast(PredV, LI->getType(), "",
1105 P->getTerminator());
1107 PN->addIncoming(PredV, I->first);
1110 //cerr << "PRE: " << *LI << *PN << "\n";
1112 LI->replaceAllUsesWith(PN);
1113 LI->eraseFromParent();
1118 /// FindMostPopularDest - The specified list contains multiple possible
1119 /// threadable destinations. Pick the one that occurs the most frequently in
1122 FindMostPopularDest(BasicBlock *BB,
1123 const SmallVectorImpl<std::pair<BasicBlock*,
1124 BasicBlock*> > &PredToDestList) {
1125 assert(!PredToDestList.empty());
1127 // Determine popularity. If there are multiple possible destinations, we
1128 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1129 // blocks with known and real destinations to threading undef. We'll handle
1130 // them later if interesting.
1131 DenseMap<BasicBlock*, unsigned> DestPopularity;
1132 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1133 if (PredToDestList[i].second)
1134 DestPopularity[PredToDestList[i].second]++;
1136 // Find the most popular dest.
1137 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1138 BasicBlock *MostPopularDest = DPI->first;
1139 unsigned Popularity = DPI->second;
1140 SmallVector<BasicBlock*, 4> SamePopularity;
1142 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1143 // If the popularity of this entry isn't higher than the popularity we've
1144 // seen so far, ignore it.
1145 if (DPI->second < Popularity)
1147 else if (DPI->second == Popularity) {
1148 // If it is the same as what we've seen so far, keep track of it.
1149 SamePopularity.push_back(DPI->first);
1151 // If it is more popular, remember it.
1152 SamePopularity.clear();
1153 MostPopularDest = DPI->first;
1154 Popularity = DPI->second;
1158 // Okay, now we know the most popular destination. If there is more than one
1159 // destination, we need to determine one. This is arbitrary, but we need
1160 // to make a deterministic decision. Pick the first one that appears in the
1162 if (!SamePopularity.empty()) {
1163 SamePopularity.push_back(MostPopularDest);
1164 TerminatorInst *TI = BB->getTerminator();
1165 for (unsigned i = 0; ; ++i) {
1166 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1168 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1169 TI->getSuccessor(i)) == SamePopularity.end())
1172 MostPopularDest = TI->getSuccessor(i);
1177 // Okay, we have finally picked the most popular destination.
1178 return MostPopularDest;
1181 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1182 ConstantPreference Preference,
1183 Instruction *CxtI) {
1184 // If threading this would thread across a loop header, don't even try to
1186 if (LoopHeaders.count(BB))
1189 PredValueInfoTy PredValues;
1190 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1193 assert(!PredValues.empty() &&
1194 "ComputeValueKnownInPredecessors returned true with no values");
1196 DEBUG(dbgs() << "IN BB: " << *BB;
1197 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1198 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1199 << *PredValues[i].first
1200 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1203 // Decide what we want to thread through. Convert our list of known values to
1204 // a list of known destinations for each pred. This also discards duplicate
1205 // predecessors and keeps track of the undefined inputs (which are represented
1206 // as a null dest in the PredToDestList).
1207 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1208 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1210 BasicBlock *OnlyDest = nullptr;
1211 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1213 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1214 BasicBlock *Pred = PredValues[i].second;
1215 if (!SeenPreds.insert(Pred).second)
1216 continue; // Duplicate predecessor entry.
1218 // If the predecessor ends with an indirect goto, we can't change its
1220 if (isa<IndirectBrInst>(Pred->getTerminator()))
1223 Constant *Val = PredValues[i].first;
1226 if (isa<UndefValue>(Val))
1228 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1229 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1230 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1231 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1233 assert(isa<IndirectBrInst>(BB->getTerminator())
1234 && "Unexpected terminator");
1235 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1238 // If we have exactly one destination, remember it for efficiency below.
1239 if (PredToDestList.empty())
1241 else if (OnlyDest != DestBB)
1242 OnlyDest = MultipleDestSentinel;
1244 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1247 // If all edges were unthreadable, we fail.
1248 if (PredToDestList.empty())
1251 // Determine which is the most common successor. If we have many inputs and
1252 // this block is a switch, we want to start by threading the batch that goes
1253 // to the most popular destination first. If we only know about one
1254 // threadable destination (the common case) we can avoid this.
1255 BasicBlock *MostPopularDest = OnlyDest;
1257 if (MostPopularDest == MultipleDestSentinel)
1258 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1260 // Now that we know what the most popular destination is, factor all
1261 // predecessors that will jump to it into a single predecessor.
1262 SmallVector<BasicBlock*, 16> PredsToFactor;
1263 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1264 if (PredToDestList[i].second == MostPopularDest) {
1265 BasicBlock *Pred = PredToDestList[i].first;
1267 // This predecessor may be a switch or something else that has multiple
1268 // edges to the block. Factor each of these edges by listing them
1269 // according to # occurrences in PredsToFactor.
1270 TerminatorInst *PredTI = Pred->getTerminator();
1271 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1272 if (PredTI->getSuccessor(i) == BB)
1273 PredsToFactor.push_back(Pred);
1276 // If the threadable edges are branching on an undefined value, we get to pick
1277 // the destination that these predecessors should get to.
1278 if (!MostPopularDest)
1279 MostPopularDest = BB->getTerminator()->
1280 getSuccessor(GetBestDestForJumpOnUndef(BB));
1282 // Ok, try to thread it!
1283 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1286 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1287 /// a PHI node in the current block. See if there are any simplifications we
1288 /// can do based on inputs to the phi node.
1290 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1291 BasicBlock *BB = PN->getParent();
1293 // TODO: We could make use of this to do it once for blocks with common PHI
1295 SmallVector<BasicBlock*, 1> PredBBs;
1298 // If any of the predecessor blocks end in an unconditional branch, we can
1299 // *duplicate* the conditional branch into that block in order to further
1300 // encourage jump threading and to eliminate cases where we have branch on a
1301 // phi of an icmp (branch on icmp is much better).
1302 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1303 BasicBlock *PredBB = PN->getIncomingBlock(i);
1304 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1305 if (PredBr->isUnconditional()) {
1306 PredBBs[0] = PredBB;
1307 // Try to duplicate BB into PredBB.
1308 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1316 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1317 /// a xor instruction in the current block. See if there are any
1318 /// simplifications we can do based on inputs to the xor.
1320 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1321 BasicBlock *BB = BO->getParent();
1323 // If either the LHS or RHS of the xor is a constant, don't do this
1325 if (isa<ConstantInt>(BO->getOperand(0)) ||
1326 isa<ConstantInt>(BO->getOperand(1)))
1329 // If the first instruction in BB isn't a phi, we won't be able to infer
1330 // anything special about any particular predecessor.
1331 if (!isa<PHINode>(BB->front()))
1334 // If we have a xor as the branch input to this block, and we know that the
1335 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1336 // the condition into the predecessor and fix that value to true, saving some
1337 // logical ops on that path and encouraging other paths to simplify.
1339 // This copies something like this:
1342 // %X = phi i1 [1], [%X']
1343 // %Y = icmp eq i32 %A, %B
1344 // %Z = xor i1 %X, %Y
1349 // %Y = icmp ne i32 %A, %B
1352 PredValueInfoTy XorOpValues;
1354 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1356 assert(XorOpValues.empty());
1357 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1363 assert(!XorOpValues.empty() &&
1364 "ComputeValueKnownInPredecessors returned true with no values");
1366 // Scan the information to see which is most popular: true or false. The
1367 // predecessors can be of the set true, false, or undef.
1368 unsigned NumTrue = 0, NumFalse = 0;
1369 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1370 if (isa<UndefValue>(XorOpValues[i].first))
1371 // Ignore undefs for the count.
1373 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1379 // Determine which value to split on, true, false, or undef if neither.
1380 ConstantInt *SplitVal = nullptr;
1381 if (NumTrue > NumFalse)
1382 SplitVal = ConstantInt::getTrue(BB->getContext());
1383 else if (NumTrue != 0 || NumFalse != 0)
1384 SplitVal = ConstantInt::getFalse(BB->getContext());
1386 // Collect all of the blocks that this can be folded into so that we can
1387 // factor this once and clone it once.
1388 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1389 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1390 if (XorOpValues[i].first != SplitVal &&
1391 !isa<UndefValue>(XorOpValues[i].first))
1394 BlocksToFoldInto.push_back(XorOpValues[i].second);
1397 // If we inferred a value for all of the predecessors, then duplication won't
1398 // help us. However, we can just replace the LHS or RHS with the constant.
1399 if (BlocksToFoldInto.size() ==
1400 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1402 // If all preds provide undef, just nuke the xor, because it is undef too.
1403 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1404 BO->eraseFromParent();
1405 } else if (SplitVal->isZero()) {
1406 // If all preds provide 0, replace the xor with the other input.
1407 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1408 BO->eraseFromParent();
1410 // If all preds provide 1, set the computed value to 1.
1411 BO->setOperand(!isLHS, SplitVal);
1417 // Try to duplicate BB into PredBB.
1418 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1422 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1423 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1424 /// NewPred using the entries from OldPred (suitably mapped).
1425 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1426 BasicBlock *OldPred,
1427 BasicBlock *NewPred,
1428 DenseMap<Instruction*, Value*> &ValueMap) {
1429 for (BasicBlock::iterator PNI = PHIBB->begin();
1430 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1431 // Ok, we have a PHI node. Figure out what the incoming value was for the
1433 Value *IV = PN->getIncomingValueForBlock(OldPred);
1435 // Remap the value if necessary.
1436 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1437 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1438 if (I != ValueMap.end())
1442 PN->addIncoming(IV, NewPred);
1446 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1447 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1448 /// across BB. Transform the IR to reflect this change.
1449 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1450 const SmallVectorImpl<BasicBlock*> &PredBBs,
1451 BasicBlock *SuccBB) {
1452 // If threading to the same block as we come from, we would infinite loop.
1454 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1455 << "' - would thread to self!\n");
1459 // If threading this would thread across a loop header, don't thread the edge.
1460 // See the comments above FindLoopHeaders for justifications and caveats.
1461 if (LoopHeaders.count(BB)) {
1462 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1463 << "' to dest BB '" << SuccBB->getName()
1464 << "' - it might create an irreducible loop!\n");
1468 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, BBDupThreshold);
1469 if (JumpThreadCost > BBDupThreshold) {
1470 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1471 << "' - Cost is too high: " << JumpThreadCost << "\n");
1475 // And finally, do it! Start by factoring the predecessors if needed.
1477 if (PredBBs.size() == 1)
1478 PredBB = PredBBs[0];
1480 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1481 << " common predecessors.\n");
1482 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1485 // And finally, do it!
1486 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1487 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1488 << ", across block:\n "
1491 LVI->threadEdge(PredBB, BB, SuccBB);
1493 // We are going to have to map operands from the original BB block to the new
1494 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1495 // account for entry from PredBB.
1496 DenseMap<Instruction*, Value*> ValueMapping;
1498 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1499 BB->getName()+".thread",
1500 BB->getParent(), BB);
1501 NewBB->moveAfter(PredBB);
1503 // Set the block frequency of NewBB.
1504 if (HasProfileData) {
1506 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
1507 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
1510 BasicBlock::iterator BI = BB->begin();
1511 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1512 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1514 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1515 // mapping and using it to remap operands in the cloned instructions.
1516 for (; !isa<TerminatorInst>(BI); ++BI) {
1517 Instruction *New = BI->clone();
1518 New->setName(BI->getName());
1519 NewBB->getInstList().push_back(New);
1520 ValueMapping[&*BI] = New;
1522 // Remap operands to patch up intra-block references.
1523 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1524 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1525 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1526 if (I != ValueMapping.end())
1527 New->setOperand(i, I->second);
1531 // We didn't copy the terminator from BB over to NewBB, because there is now
1532 // an unconditional jump to SuccBB. Insert the unconditional jump.
1533 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
1534 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1536 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1537 // PHI nodes for NewBB now.
1538 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1540 // If there were values defined in BB that are used outside the block, then we
1541 // now have to update all uses of the value to use either the original value,
1542 // the cloned value, or some PHI derived value. This can require arbitrary
1543 // PHI insertion, of which we are prepared to do, clean these up now.
1544 SSAUpdater SSAUpdate;
1545 SmallVector<Use*, 16> UsesToRename;
1546 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1547 // Scan all uses of this instruction to see if it is used outside of its
1548 // block, and if so, record them in UsesToRename.
1549 for (Use &U : I->uses()) {
1550 Instruction *User = cast<Instruction>(U.getUser());
1551 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1552 if (UserPN->getIncomingBlock(U) == BB)
1554 } else if (User->getParent() == BB)
1557 UsesToRename.push_back(&U);
1560 // If there are no uses outside the block, we're done with this instruction.
1561 if (UsesToRename.empty())
1564 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1566 // We found a use of I outside of BB. Rename all uses of I that are outside
1567 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1568 // with the two values we know.
1569 SSAUpdate.Initialize(I->getType(), I->getName());
1570 SSAUpdate.AddAvailableValue(BB, &*I);
1571 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&*I]);
1573 while (!UsesToRename.empty())
1574 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1575 DEBUG(dbgs() << "\n");
1579 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1580 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1581 // us to simplify any PHI nodes in BB.
1582 TerminatorInst *PredTerm = PredBB->getTerminator();
1583 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1584 if (PredTerm->getSuccessor(i) == BB) {
1585 BB->removePredecessor(PredBB, true);
1586 PredTerm->setSuccessor(i, NewBB);
1589 // At this point, the IR is fully up to date and consistent. Do a quick scan
1590 // over the new instructions and zap any that are constants or dead. This
1591 // frequently happens because of phi translation.
1592 SimplifyInstructionsInBlock(NewBB, TLI);
1594 // Update the edge weight from BB to SuccBB, which should be less than before.
1595 UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
1597 // Threaded an edge!
1602 /// Create a new basic block that will be the predecessor of BB and successor of
1603 /// all blocks in Preds. When profile data is availble, update the frequency of
1605 BasicBlock *JumpThreading::SplitBlockPreds(BasicBlock *BB,
1606 ArrayRef<BasicBlock *> Preds,
1607 const char *Suffix) {
1608 // Collect the frequencies of all predecessors of BB, which will be used to
1609 // update the edge weight on BB->SuccBB.
1610 BlockFrequency PredBBFreq(0);
1612 for (auto Pred : Preds)
1613 PredBBFreq += BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB);
1615 BasicBlock *PredBB = SplitBlockPredecessors(BB, Preds, Suffix);
1617 // Set the block frequency of the newly created PredBB, which is the sum of
1618 // frequencies of Preds.
1620 BFI->setBlockFreq(PredBB, PredBBFreq.getFrequency());
1624 /// Update the block frequency of BB and branch weight and the metadata on the
1625 /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
1626 /// Freq(PredBB->BB) / Freq(BB->SuccBB).
1627 void JumpThreading::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
1630 BasicBlock *SuccBB) {
1631 if (!HasProfileData)
1634 assert(BFI && BPI && "BFI & BPI should have been created here");
1636 // As the edge from PredBB to BB is deleted, we have to update the block
1638 auto BBOrigFreq = BFI->getBlockFreq(BB);
1639 auto NewBBFreq = BFI->getBlockFreq(NewBB);
1640 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
1641 auto BBNewFreq = BBOrigFreq - NewBBFreq;
1642 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
1644 // Collect updated outgoing edges' frequencies from BB and use them to update
1645 // edge probabilities.
1646 SmallVector<uint64_t, 4> BBSuccFreq;
1647 for (auto I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
1648 auto SuccFreq = (*I == SuccBB)
1649 ? BB2SuccBBFreq - NewBBFreq
1650 : BBOrigFreq * BPI->getEdgeProbability(BB, *I);
1651 BBSuccFreq.push_back(SuccFreq.getFrequency());
1654 uint64_t MaxBBSuccFreq =
1655 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
1657 SmallVector<BranchProbability, 4> BBSuccProbs;
1658 if (MaxBBSuccFreq == 0)
1659 BBSuccProbs.assign(BBSuccFreq.size(),
1660 {1, static_cast<unsigned>(BBSuccFreq.size())});
1662 for (uint64_t Freq : BBSuccFreq)
1663 BBSuccProbs.push_back(
1664 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
1665 // Normalize edge probabilities so that they sum up to one.
1666 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
1670 // Update edge probabilities in BPI.
1671 for (int I = 0, E = BBSuccProbs.size(); I < E; I++)
1672 BPI->setEdgeProbability(BB, I, BBSuccProbs[I]);
1674 if (BBSuccProbs.size() >= 2) {
1675 SmallVector<uint32_t, 4> Weights;
1676 for (auto Prob : BBSuccProbs)
1677 Weights.push_back(Prob.getNumerator());
1679 auto TI = BB->getTerminator();
1681 LLVMContext::MD_prof,
1682 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
1686 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1687 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1688 /// If we can duplicate the contents of BB up into PredBB do so now, this
1689 /// improves the odds that the branch will be on an analyzable instruction like
1691 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1692 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1693 assert(!PredBBs.empty() && "Can't handle an empty set");
1695 // If BB is a loop header, then duplicating this block outside the loop would
1696 // cause us to transform this into an irreducible loop, don't do this.
1697 // See the comments above FindLoopHeaders for justifications and caveats.
1698 if (LoopHeaders.count(BB)) {
1699 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1700 << "' into predecessor block '" << PredBBs[0]->getName()
1701 << "' - it might create an irreducible loop!\n");
1705 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, BBDupThreshold);
1706 if (DuplicationCost > BBDupThreshold) {
1707 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1708 << "' - Cost is too high: " << DuplicationCost << "\n");
1712 // And finally, do it! Start by factoring the predecessors if needed.
1714 if (PredBBs.size() == 1)
1715 PredBB = PredBBs[0];
1717 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1718 << " common predecessors.\n");
1719 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1722 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1724 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1725 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1726 << DuplicationCost << " block is:" << *BB << "\n");
1728 // Unless PredBB ends with an unconditional branch, split the edge so that we
1729 // can just clone the bits from BB into the end of the new PredBB.
1730 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1732 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
1733 PredBB = SplitEdge(PredBB, BB);
1734 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1737 // We are going to have to map operands from the original BB block into the
1738 // PredBB block. Evaluate PHI nodes in BB.
1739 DenseMap<Instruction*, Value*> ValueMapping;
1741 BasicBlock::iterator BI = BB->begin();
1742 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1743 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1744 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1745 // mapping and using it to remap operands in the cloned instructions.
1746 for (; BI != BB->end(); ++BI) {
1747 Instruction *New = BI->clone();
1749 // Remap operands to patch up intra-block references.
1750 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1751 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1752 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1753 if (I != ValueMapping.end())
1754 New->setOperand(i, I->second);
1757 // If this instruction can be simplified after the operands are updated,
1758 // just use the simplified value instead. This frequently happens due to
1761 SimplifyInstruction(New, BB->getModule()->getDataLayout())) {
1763 ValueMapping[&*BI] = IV;
1765 // Otherwise, insert the new instruction into the block.
1766 New->setName(BI->getName());
1767 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
1768 ValueMapping[&*BI] = New;
1772 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1773 // add entries to the PHI nodes for branch from PredBB now.
1774 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1775 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1777 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1780 // If there were values defined in BB that are used outside the block, then we
1781 // now have to update all uses of the value to use either the original value,
1782 // the cloned value, or some PHI derived value. This can require arbitrary
1783 // PHI insertion, of which we are prepared to do, clean these up now.
1784 SSAUpdater SSAUpdate;
1785 SmallVector<Use*, 16> UsesToRename;
1786 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1787 // Scan all uses of this instruction to see if it is used outside of its
1788 // block, and if so, record them in UsesToRename.
1789 for (Use &U : I->uses()) {
1790 Instruction *User = cast<Instruction>(U.getUser());
1791 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1792 if (UserPN->getIncomingBlock(U) == BB)
1794 } else if (User->getParent() == BB)
1797 UsesToRename.push_back(&U);
1800 // If there are no uses outside the block, we're done with this instruction.
1801 if (UsesToRename.empty())
1804 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1806 // We found a use of I outside of BB. Rename all uses of I that are outside
1807 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1808 // with the two values we know.
1809 SSAUpdate.Initialize(I->getType(), I->getName());
1810 SSAUpdate.AddAvailableValue(BB, &*I);
1811 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&*I]);
1813 while (!UsesToRename.empty())
1814 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1815 DEBUG(dbgs() << "\n");
1818 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1820 BB->removePredecessor(PredBB, true);
1822 // Remove the unconditional branch at the end of the PredBB block.
1823 OldPredBranch->eraseFromParent();
1829 /// TryToUnfoldSelect - Look for blocks of the form
1835 /// %p = phi [%a, %bb] ...
1839 /// And expand the select into a branch structure if one of its arms allows %c
1840 /// to be folded. This later enables threading from bb1 over bb2.
1841 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1842 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1843 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1844 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1846 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1847 CondLHS->getParent() != BB)
1850 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1851 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1852 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1854 // Look if one of the incoming values is a select in the corresponding
1856 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1859 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1860 if (!PredTerm || !PredTerm->isUnconditional())
1863 // Now check if one of the select values would allow us to constant fold the
1864 // terminator in BB. We don't do the transform if both sides fold, those
1865 // cases will be threaded in any case.
1866 LazyValueInfo::Tristate LHSFolds =
1867 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1868 CondRHS, Pred, BB, CondCmp);
1869 LazyValueInfo::Tristate RHSFolds =
1870 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
1871 CondRHS, Pred, BB, CondCmp);
1872 if ((LHSFolds != LazyValueInfo::Unknown ||
1873 RHSFolds != LazyValueInfo::Unknown) &&
1874 LHSFolds != RHSFolds) {
1875 // Expand the select.
1884 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
1885 BB->getParent(), BB);
1886 // Move the unconditional branch to NewBB.
1887 PredTerm->removeFromParent();
1888 NewBB->getInstList().insert(NewBB->end(), PredTerm);
1889 // Create a conditional branch and update PHI nodes.
1890 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
1891 CondLHS->setIncomingValue(I, SI->getFalseValue());
1892 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
1893 // The select is now dead.
1894 SI->eraseFromParent();
1896 // Update any other PHI nodes in BB.
1897 for (BasicBlock::iterator BI = BB->begin();
1898 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
1900 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);