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 #define DEBUG_TYPE "jump-threading"
15 #include "llvm/Transforms/Scalar.h"
16 #include "llvm/IntrinsicInst.h"
17 #include "llvm/LLVMContext.h"
18 #include "llvm/Pass.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/Analysis/LazyValueInfo.h"
21 #include "llvm/Analysis/Loads.h"
22 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
23 #include "llvm/Transforms/Utils/Local.h"
24 #include "llvm/Transforms/Utils/SSAUpdater.h"
25 #include "llvm/Target/TargetData.h"
26 #include "llvm/ADT/DenseMap.h"
27 #include "llvm/ADT/DenseSet.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/ADT/STLExtras.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/SmallSet.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ValueHandle.h"
35 #include "llvm/Support/raw_ostream.h"
38 STATISTIC(NumThreads, "Number of jumps threaded");
39 STATISTIC(NumFolds, "Number of terminators folded");
40 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
42 static cl::opt<unsigned>
43 Threshold("jump-threading-threshold",
44 cl::desc("Max block size to duplicate for jump threading"),
45 cl::init(6), cl::Hidden);
48 // These are at global scope so static functions can use them too.
49 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo;
50 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy;
52 // This is used to keep track of what kind of constant we're currently hoping
54 enum ConstantPreference {
59 /// This pass performs 'jump threading', which looks at blocks that have
60 /// multiple predecessors and multiple successors. If one or more of the
61 /// predecessors of the block can be proven to always jump to one of the
62 /// successors, we forward the edge from the predecessor to the successor by
63 /// duplicating the contents of this block.
65 /// An example of when this can occur is code like this:
72 /// In this case, the unconditional branch at the end of the first if can be
73 /// revectored to the false side of the second if.
75 class JumpThreading : public FunctionPass {
79 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
81 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
83 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet;
85 // RAII helper for updating the recursion stack.
86 struct RecursionSetRemover {
87 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet;
88 std::pair<Value*, BasicBlock*> ThePair;
90 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S,
91 std::pair<Value*, BasicBlock*> P)
92 : TheSet(S), ThePair(P) { }
94 ~RecursionSetRemover() {
95 TheSet.erase(ThePair);
99 static char ID; // Pass identification
100 JumpThreading() : FunctionPass(ID) {
101 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
104 bool runOnFunction(Function &F);
106 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
107 AU.addRequired<LazyValueInfo>();
108 AU.addPreserved<LazyValueInfo>();
111 void FindLoopHeaders(Function &F);
112 bool ProcessBlock(BasicBlock *BB);
113 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
115 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
116 const SmallVectorImpl<BasicBlock *> &PredBBs);
118 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
119 PredValueInfo &Result,
120 ConstantPreference Preference);
121 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
122 ConstantPreference Preference);
125 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
126 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
128 bool ProcessBranchOnPHI(PHINode *PN);
129 bool ProcessBranchOnXOR(BinaryOperator *BO);
131 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
135 char JumpThreading::ID = 0;
136 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
137 "Jump Threading", false, false)
138 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)
139 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
140 "Jump Threading", false, false)
142 // Public interface to the Jump Threading pass
143 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
145 /// runOnFunction - Top level algorithm.
147 bool JumpThreading::runOnFunction(Function &F) {
148 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
149 TD = getAnalysisIfAvailable<TargetData>();
150 LVI = &getAnalysis<LazyValueInfo>();
154 bool Changed, EverChanged = false;
157 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
159 // Thread all of the branches we can over this block.
160 while (ProcessBlock(BB))
165 // If the block is trivially dead, zap it. This eliminates the successor
166 // edges which simplifies the CFG.
167 if (pred_begin(BB) == pred_end(BB) &&
168 BB != &BB->getParent()->getEntryBlock()) {
169 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
170 << "' with terminator: " << *BB->getTerminator() << '\n');
171 LoopHeaders.erase(BB);
175 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
176 // Can't thread an unconditional jump, but if the block is "almost
177 // empty", we can replace uses of it with uses of the successor and make
179 if (BI->isUnconditional() &&
180 BB != &BB->getParent()->getEntryBlock()) {
181 BasicBlock::iterator BBI = BB->getFirstNonPHI();
182 // Ignore dbg intrinsics.
183 while (isa<DbgInfoIntrinsic>(BBI))
185 // If the terminator is the only non-phi instruction, try to nuke it.
186 if (BBI->isTerminator()) {
187 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
188 // block, we have to make sure it isn't in the LoopHeaders set. We
189 // reinsert afterward if needed.
190 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
191 BasicBlock *Succ = BI->getSuccessor(0);
193 // FIXME: It is always conservatively correct to drop the info
194 // for a block even if it doesn't get erased. This isn't totally
195 // awesome, but it allows us to use AssertingVH to prevent nasty
196 // dangling pointer issues within LazyValueInfo.
198 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
200 // If we deleted BB and BB was the header of a loop, then the
201 // successor is now the header of the loop.
205 if (ErasedFromLoopHeaders)
206 LoopHeaders.insert(BB);
211 EverChanged |= Changed;
218 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
219 /// thread across it.
220 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
221 /// Ignore PHI nodes, these will be flattened when duplication happens.
222 BasicBlock::const_iterator I = BB->getFirstNonPHI();
224 // FIXME: THREADING will delete values that are just used to compute the
225 // branch, so they shouldn't count against the duplication cost.
228 // Sum up the cost of each instruction until we get to the terminator. Don't
229 // include the terminator because the copy won't include it.
231 for (; !isa<TerminatorInst>(I); ++I) {
232 // Debugger intrinsics don't incur code size.
233 if (isa<DbgInfoIntrinsic>(I)) continue;
235 // If this is a pointer->pointer bitcast, it is free.
236 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
239 // All other instructions count for at least one unit.
242 // Calls are more expensive. If they are non-intrinsic calls, we model them
243 // as having cost of 4. If they are a non-vector intrinsic, we model them
244 // as having cost of 2 total, and if they are a vector intrinsic, we model
245 // them as having cost 1.
246 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
247 if (!isa<IntrinsicInst>(CI))
249 else if (!CI->getType()->isVectorTy())
254 // Threading through a switch statement is particularly profitable. If this
255 // block ends in a switch, decrease its cost to make it more likely to happen.
256 if (isa<SwitchInst>(I))
257 Size = Size > 6 ? Size-6 : 0;
259 // The same holds for indirect branches, but slightly more so.
260 if (isa<IndirectBrInst>(I))
261 Size = Size > 8 ? Size-8 : 0;
266 /// FindLoopHeaders - We do not want jump threading to turn proper loop
267 /// structures into irreducible loops. Doing this breaks up the loop nesting
268 /// hierarchy and pessimizes later transformations. To prevent this from
269 /// happening, we first have to find the loop headers. Here we approximate this
270 /// by finding targets of backedges in the CFG.
272 /// Note that there definitely are cases when we want to allow threading of
273 /// edges across a loop header. For example, threading a jump from outside the
274 /// loop (the preheader) to an exit block of the loop is definitely profitable.
275 /// It is also almost always profitable to thread backedges from within the loop
276 /// to exit blocks, and is often profitable to thread backedges to other blocks
277 /// within the loop (forming a nested loop). This simple analysis is not rich
278 /// enough to track all of these properties and keep it up-to-date as the CFG
279 /// mutates, so we don't allow any of these transformations.
281 void JumpThreading::FindLoopHeaders(Function &F) {
282 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
283 FindFunctionBackedges(F, Edges);
285 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
286 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
289 /// getKnownConstant - Helper method to determine if we can thread over a
290 /// terminator with the given value as its condition, and if so what value to
291 /// use for that. What kind of value this is depends on whether we want an
292 /// integer or a block address, but an undef is always accepted.
293 /// Returns null if Val is null or not an appropriate constant.
294 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
298 // Undef is "known" enough.
299 if (UndefValue *U = dyn_cast<UndefValue>(Val))
302 if (Preference == WantBlockAddress)
303 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
305 return dyn_cast<ConstantInt>(Val);
308 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
309 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
310 /// in any of our predecessors. If so, return the known list of value and pred
311 /// BB in the result vector.
313 /// This returns true if there were any known values.
316 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
317 ConstantPreference Preference) {
318 // This method walks up use-def chains recursively. Because of this, we could
319 // get into an infinite loop going around loops in the use-def chain. To
320 // prevent this, keep track of what (value, block) pairs we've already visited
321 // and terminate the search if we loop back to them
322 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
325 // An RAII help to remove this pair from the recursion set once the recursion
326 // stack pops back out again.
327 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
329 // If V is a constant, then it is known in all predecessors.
330 if (Constant *KC = getKnownConstant(V, Preference)) {
331 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
332 Result.push_back(std::make_pair(KC, *PI));
337 // If V is a non-instruction value, or an instruction in a different block,
338 // then it can't be derived from a PHI.
339 Instruction *I = dyn_cast<Instruction>(V);
340 if (I == 0 || I->getParent() != BB) {
342 // Okay, if this is a live-in value, see if it has a known value at the end
343 // of any of our predecessors.
345 // FIXME: This should be an edge property, not a block end property.
346 /// TODO: Per PR2563, we could infer value range information about a
347 /// predecessor based on its terminator.
349 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
350 // "I" is a non-local compare-with-a-constant instruction. This would be
351 // able to handle value inequalities better, for example if the compare is
352 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
353 // Perhaps getConstantOnEdge should be smart enough to do this?
355 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
357 // If the value is known by LazyValueInfo to be a constant in a
358 // predecessor, use that information to try to thread this block.
359 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
360 if (Constant *KC = getKnownConstant(PredCst, Preference))
361 Result.push_back(std::make_pair(KC, P));
364 return !Result.empty();
367 /// If I is a PHI node, then we know the incoming values for any constants.
368 if (PHINode *PN = dyn_cast<PHINode>(I)) {
369 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
370 Value *InVal = PN->getIncomingValue(i);
371 if (Constant *KC = getKnownConstant(InVal, Preference)) {
372 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
374 Constant *CI = LVI->getConstantOnEdge(InVal,
375 PN->getIncomingBlock(i), BB);
376 if (Constant *KC = getKnownConstant(CI, Preference))
377 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
381 return !Result.empty();
384 PredValueInfoTy LHSVals, RHSVals;
386 // Handle some boolean conditions.
387 if (I->getType()->getPrimitiveSizeInBits() == 1) {
388 assert(Preference == WantInteger && "One-bit non-integer type?");
390 // X & false -> false
391 if (I->getOpcode() == Instruction::Or ||
392 I->getOpcode() == Instruction::And) {
393 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
395 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
398 if (LHSVals.empty() && RHSVals.empty())
401 ConstantInt *InterestingVal;
402 if (I->getOpcode() == Instruction::Or)
403 InterestingVal = ConstantInt::getTrue(I->getContext());
405 InterestingVal = ConstantInt::getFalse(I->getContext());
407 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
409 // Scan for the sentinel. If we find an undef, force it to the
410 // interesting value: x|undef -> true and x&undef -> false.
411 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
412 if (LHSVals[i].first == InterestingVal ||
413 isa<UndefValue>(LHSVals[i].first)) {
414 Result.push_back(LHSVals[i]);
415 Result.back().first = InterestingVal;
416 LHSKnownBBs.insert(LHSVals[i].second);
418 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
419 if (RHSVals[i].first == InterestingVal ||
420 isa<UndefValue>(RHSVals[i].first)) {
421 // If we already inferred a value for this block on the LHS, don't
423 if (!LHSKnownBBs.count(RHSVals[i].second)) {
424 Result.push_back(RHSVals[i]);
425 Result.back().first = InterestingVal;
429 return !Result.empty();
432 // Handle the NOT form of XOR.
433 if (I->getOpcode() == Instruction::Xor &&
434 isa<ConstantInt>(I->getOperand(1)) &&
435 cast<ConstantInt>(I->getOperand(1))->isOne()) {
436 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
441 // Invert the known values.
442 for (unsigned i = 0, e = Result.size(); i != e; ++i)
443 Result[i].first = ConstantExpr::getNot(Result[i].first);
448 // Try to simplify some other binary operator values.
449 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
450 assert(Preference != WantBlockAddress
451 && "A binary operator creating a block address?");
452 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
453 PredValueInfoTy LHSVals;
454 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
457 // Try to use constant folding to simplify the binary operator.
458 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
459 Constant *V = LHSVals[i].first;
460 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
462 if (Constant *KC = getKnownConstant(Folded, WantInteger))
463 Result.push_back(std::make_pair(KC, LHSVals[i].second));
467 return !Result.empty();
470 // Handle compare with phi operand, where the PHI is defined in this block.
471 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
472 assert(Preference == WantInteger && "Compares only produce integers");
473 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
474 if (PN && PN->getParent() == BB) {
475 // We can do this simplification if any comparisons fold to true or false.
477 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
478 BasicBlock *PredBB = PN->getIncomingBlock(i);
479 Value *LHS = PN->getIncomingValue(i);
480 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
482 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
484 if (!isa<Constant>(RHS))
487 LazyValueInfo::Tristate
488 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
489 cast<Constant>(RHS), PredBB, BB);
490 if (ResT == LazyValueInfo::Unknown)
492 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
495 if (Constant *KC = getKnownConstant(Res, WantInteger))
496 Result.push_back(std::make_pair(KC, PredBB));
499 return !Result.empty();
503 // If comparing a live-in value against a constant, see if we know the
504 // live-in value on any predecessors.
505 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
506 if (!isa<Instruction>(Cmp->getOperand(0)) ||
507 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
508 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
510 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
512 // If the value is known by LazyValueInfo to be a constant in a
513 // predecessor, use that information to try to thread this block.
514 LazyValueInfo::Tristate Res =
515 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
517 if (Res == LazyValueInfo::Unknown)
520 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
521 Result.push_back(std::make_pair(ResC, P));
524 return !Result.empty();
527 // Try to find a constant value for the LHS of a comparison,
528 // and evaluate it statically if we can.
529 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
530 PredValueInfoTy LHSVals;
531 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
534 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
535 Constant *V = LHSVals[i].first;
536 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
538 if (Constant *KC = getKnownConstant(Folded, WantInteger))
539 Result.push_back(std::make_pair(KC, LHSVals[i].second));
542 return !Result.empty();
547 // If all else fails, see if LVI can figure out a constant value for us.
548 Constant *CI = LVI->getConstant(V, BB);
549 if (Constant *KC = getKnownConstant(CI, Preference)) {
550 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
551 Result.push_back(std::make_pair(KC, *PI));
554 return !Result.empty();
559 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
560 /// in an undefined jump, decide which block is best to revector to.
562 /// Since we can pick an arbitrary destination, we pick the successor with the
563 /// fewest predecessors. This should reduce the in-degree of the others.
565 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
566 TerminatorInst *BBTerm = BB->getTerminator();
567 unsigned MinSucc = 0;
568 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
569 // Compute the successor with the minimum number of predecessors.
570 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
571 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
572 TestBB = BBTerm->getSuccessor(i);
573 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
574 if (NumPreds < MinNumPreds)
581 /// ProcessBlock - If there are any predecessors whose control can be threaded
582 /// through to a successor, transform them now.
583 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
584 // If the block is trivially dead, just return and let the caller nuke it.
585 // This simplifies other transformations.
586 if (pred_begin(BB) == pred_end(BB) &&
587 BB != &BB->getParent()->getEntryBlock())
590 // If this block has a single predecessor, and if that pred has a single
591 // successor, merge the blocks. This encourages recursive jump threading
592 // because now the condition in this block can be threaded through
593 // predecessors of our predecessor block.
594 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
595 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
597 // If SinglePred was a loop header, BB becomes one.
598 if (LoopHeaders.erase(SinglePred))
599 LoopHeaders.insert(BB);
601 // Remember if SinglePred was the entry block of the function. If so, we
602 // will need to move BB back to the entry position.
603 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
604 LVI->eraseBlock(SinglePred);
605 MergeBasicBlockIntoOnlyPred(BB);
607 if (isEntry && BB != &BB->getParent()->getEntryBlock())
608 BB->moveBefore(&BB->getParent()->getEntryBlock());
613 // What kind of constant we're looking for.
614 ConstantPreference Preference = WantInteger;
616 // Look to see if the terminator is a conditional branch, switch or indirect
617 // branch, if not we can't thread it.
619 Instruction *Terminator = BB->getTerminator();
620 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
621 // Can't thread an unconditional jump.
622 if (BI->isUnconditional()) return false;
623 Condition = BI->getCondition();
624 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
625 Condition = SI->getCondition();
626 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
627 Condition = IB->getAddress()->stripPointerCasts();
628 Preference = WantBlockAddress;
630 return false; // Must be an invoke.
633 // If the terminator is branching on an undef, we can pick any of the
634 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
635 if (isa<UndefValue>(Condition)) {
636 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
638 // Fold the branch/switch.
639 TerminatorInst *BBTerm = BB->getTerminator();
640 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
641 if (i == BestSucc) continue;
642 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
645 DEBUG(dbgs() << " In block '" << BB->getName()
646 << "' folding undef terminator: " << *BBTerm << '\n');
647 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
648 BBTerm->eraseFromParent();
652 // If the terminator of this block is branching on a constant, simplify the
653 // terminator to an unconditional branch. This can occur due to threading in
655 if (getKnownConstant(Condition, Preference)) {
656 DEBUG(dbgs() << " In block '" << BB->getName()
657 << "' folding terminator: " << *BB->getTerminator() << '\n');
659 ConstantFoldTerminator(BB);
663 Instruction *CondInst = dyn_cast<Instruction>(Condition);
665 // All the rest of our checks depend on the condition being an instruction.
667 // FIXME: Unify this with code below.
668 if (ProcessThreadableEdges(Condition, BB, Preference))
674 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
675 // For a comparison where the LHS is outside this block, it's possible
676 // that we've branched on it before. Used LVI to see if we can simplify
677 // the branch based on that.
678 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
679 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
680 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
681 if (CondBr && CondConst && CondBr->isConditional() && PI != PE &&
682 (!isa<Instruction>(CondCmp->getOperand(0)) ||
683 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
684 // For predecessor edge, determine if the comparison is true or false
685 // on that edge. If they're all true or all false, we can simplify the
687 // FIXME: We could handle mixed true/false by duplicating code.
688 LazyValueInfo::Tristate Baseline =
689 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
691 if (Baseline != LazyValueInfo::Unknown) {
692 // Check that all remaining incoming values match the first one.
694 LazyValueInfo::Tristate Ret =
695 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
696 CondCmp->getOperand(0), CondConst, *PI, BB);
697 if (Ret != Baseline) break;
700 // If we terminated early, then one of the values didn't match.
702 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
703 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
704 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
705 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
706 CondBr->eraseFromParent();
713 // Check for some cases that are worth simplifying. Right now we want to look
714 // for loads that are used by a switch or by the condition for the branch. If
715 // we see one, check to see if it's partially redundant. If so, insert a PHI
716 // which can then be used to thread the values.
718 Value *SimplifyValue = CondInst;
719 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
720 if (isa<Constant>(CondCmp->getOperand(1)))
721 SimplifyValue = CondCmp->getOperand(0);
723 // TODO: There are other places where load PRE would be profitable, such as
724 // more complex comparisons.
725 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
726 if (SimplifyPartiallyRedundantLoad(LI))
730 // Handle a variety of cases where we are branching on something derived from
731 // a PHI node in the current block. If we can prove that any predecessors
732 // compute a predictable value based on a PHI node, thread those predecessors.
734 if (ProcessThreadableEdges(CondInst, BB, Preference))
737 // If this is an otherwise-unfoldable branch on a phi node in the current
738 // block, see if we can simplify.
739 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
740 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
741 return ProcessBranchOnPHI(PN);
744 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
745 if (CondInst->getOpcode() == Instruction::Xor &&
746 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
747 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
750 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
751 // "(X == 4)", thread through this block.
756 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
757 /// block that jump on exactly the same condition. This means that we almost
758 /// always know the direction of the edge in the DESTBB:
760 /// br COND, DESTBB, BBY
762 /// br COND, BBZ, BBW
764 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
765 /// in DESTBB, we have to thread over it.
766 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
768 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
770 // If both successors of PredBB go to DESTBB, we don't know anything. We can
771 // fold the branch to an unconditional one, which allows other recursive
774 if (PredBI->getSuccessor(1) != BB)
776 else if (PredBI->getSuccessor(0) != BB)
779 DEBUG(dbgs() << " In block '" << PredBB->getName()
780 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
782 ConstantFoldTerminator(PredBB);
786 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
788 // If the dest block has one predecessor, just fix the branch condition to a
789 // constant and fold it.
790 if (BB->getSinglePredecessor()) {
791 DEBUG(dbgs() << " In block '" << BB->getName()
792 << "' folding condition to '" << BranchDir << "': "
793 << *BB->getTerminator() << '\n');
795 Value *OldCond = DestBI->getCondition();
796 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
798 // Delete dead instructions before we fold the branch. Folding the branch
799 // can eliminate edges from the CFG which can end up deleting OldCond.
800 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
801 ConstantFoldTerminator(BB);
806 // Next, figure out which successor we are threading to.
807 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
809 SmallVector<BasicBlock*, 2> Preds;
810 Preds.push_back(PredBB);
812 // Ok, try to thread it!
813 return ThreadEdge(BB, Preds, SuccBB);
816 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
817 /// block that switch on exactly the same condition. This means that we almost
818 /// always know the direction of the edge in the DESTBB:
820 /// switch COND [... DESTBB, BBY ... ]
822 /// switch COND [... BBZ, BBW ]
824 /// Optimizing switches like this is very important, because simplifycfg builds
825 /// switches out of repeated 'if' conditions.
826 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
827 BasicBlock *DestBB) {
828 // Can't thread edge to self.
829 if (PredBB == DestBB)
832 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
833 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
835 // There are a variety of optimizations that we can potentially do on these
836 // blocks: we order them from most to least preferable.
838 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
839 // directly to their destination. This does not introduce *any* code size
840 // growth. Skip debug info first.
841 BasicBlock::iterator BBI = DestBB->begin();
842 while (isa<DbgInfoIntrinsic>(BBI))
845 // FIXME: Thread if it just contains a PHI.
846 if (isa<SwitchInst>(BBI)) {
847 bool MadeChange = false;
848 // Ignore the default edge for now.
849 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
850 ConstantInt *DestVal = DestSI->getCaseValue(i);
851 BasicBlock *DestSucc = DestSI->getSuccessor(i);
853 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
854 // PredSI has an explicit case for it. If so, forward. If it is covered
855 // by the default case, we can't update PredSI.
856 unsigned PredCase = PredSI->findCaseValue(DestVal);
857 if (PredCase == 0) continue;
859 // If PredSI doesn't go to DestBB on this value, then it won't reach the
860 // case on this condition.
861 if (PredSI->getSuccessor(PredCase) != DestBB &&
862 DestSI->getSuccessor(i) != DestBB)
865 // Do not forward this if it already goes to this destination, this would
866 // be an infinite loop.
867 if (PredSI->getSuccessor(PredCase) == DestSucc)
870 // Otherwise, we're safe to make the change. Make sure that the edge from
871 // DestSI to DestSucc is not critical and has no PHI nodes.
872 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
873 DEBUG(dbgs() << "THROUGH: " << *DestSI);
875 // If the destination has PHI nodes, just split the edge for updating
877 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
878 SplitCriticalEdge(DestSI, i, this);
879 DestSucc = DestSI->getSuccessor(i);
881 FoldSingleEntryPHINodes(DestSucc);
882 PredSI->setSuccessor(PredCase, DestSucc);
894 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
895 /// load instruction, eliminate it by replacing it with a PHI node. This is an
896 /// important optimization that encourages jump threading, and needs to be run
897 /// interlaced with other jump threading tasks.
898 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
899 // Don't hack volatile loads.
900 if (LI->isVolatile()) return false;
902 // If the load is defined in a block with exactly one predecessor, it can't be
903 // partially redundant.
904 BasicBlock *LoadBB = LI->getParent();
905 if (LoadBB->getSinglePredecessor())
908 Value *LoadedPtr = LI->getOperand(0);
910 // If the loaded operand is defined in the LoadBB, it can't be available.
911 // TODO: Could do simple PHI translation, that would be fun :)
912 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
913 if (PtrOp->getParent() == LoadBB)
916 // Scan a few instructions up from the load, to see if it is obviously live at
917 // the entry to its block.
918 BasicBlock::iterator BBIt = LI;
920 if (Value *AvailableVal =
921 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
922 // If the value if the load is locally available within the block, just use
923 // it. This frequently occurs for reg2mem'd allocas.
924 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
926 // If the returned value is the load itself, replace with an undef. This can
927 // only happen in dead loops.
928 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
929 LI->replaceAllUsesWith(AvailableVal);
930 LI->eraseFromParent();
934 // Otherwise, if we scanned the whole block and got to the top of the block,
935 // we know the block is locally transparent to the load. If not, something
936 // might clobber its value.
937 if (BBIt != LoadBB->begin())
941 SmallPtrSet<BasicBlock*, 8> PredsScanned;
942 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
943 AvailablePredsTy AvailablePreds;
944 BasicBlock *OneUnavailablePred = 0;
946 // If we got here, the loaded value is transparent through to the start of the
947 // block. Check to see if it is available in any of the predecessor blocks.
948 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
950 BasicBlock *PredBB = *PI;
952 // If we already scanned this predecessor, skip it.
953 if (!PredsScanned.insert(PredBB))
956 // Scan the predecessor to see if the value is available in the pred.
957 BBIt = PredBB->end();
958 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
959 if (!PredAvailable) {
960 OneUnavailablePred = PredBB;
964 // If so, this load is partially redundant. Remember this info so that we
965 // can create a PHI node.
966 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
969 // If the loaded value isn't available in any predecessor, it isn't partially
971 if (AvailablePreds.empty()) return false;
973 // Okay, the loaded value is available in at least one (and maybe all!)
974 // predecessors. If the value is unavailable in more than one unique
975 // predecessor, we want to insert a merge block for those common predecessors.
976 // This ensures that we only have to insert one reload, thus not increasing
978 BasicBlock *UnavailablePred = 0;
980 // If there is exactly one predecessor where the value is unavailable, the
981 // already computed 'OneUnavailablePred' block is it. If it ends in an
982 // unconditional branch, we know that it isn't a critical edge.
983 if (PredsScanned.size() == AvailablePreds.size()+1 &&
984 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
985 UnavailablePred = OneUnavailablePred;
986 } else if (PredsScanned.size() != AvailablePreds.size()) {
987 // Otherwise, we had multiple unavailable predecessors or we had a critical
988 // edge from the one.
989 SmallVector<BasicBlock*, 8> PredsToSplit;
990 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
992 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
993 AvailablePredSet.insert(AvailablePreds[i].first);
995 // Add all the unavailable predecessors to the PredsToSplit list.
996 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
999 // If the predecessor is an indirect goto, we can't split the edge.
1000 if (isa<IndirectBrInst>(P->getTerminator()))
1003 if (!AvailablePredSet.count(P))
1004 PredsToSplit.push_back(P);
1007 // Split them out to their own block.
1009 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
1010 "thread-pre-split", this);
1013 // If the value isn't available in all predecessors, then there will be
1014 // exactly one where it isn't available. Insert a load on that edge and add
1015 // it to the AvailablePreds list.
1016 if (UnavailablePred) {
1017 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1018 "Can't handle critical edge here!");
1019 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
1021 UnavailablePred->getTerminator());
1022 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1025 // Now we know that each predecessor of this block has a value in
1026 // AvailablePreds, sort them for efficient access as we're walking the preds.
1027 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1029 // Create a PHI node at the start of the block for the PRE'd load value.
1030 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
1033 // Insert new entries into the PHI for each predecessor. A single block may
1034 // have multiple entries here.
1035 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
1037 BasicBlock *P = *PI;
1038 AvailablePredsTy::iterator I =
1039 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1040 std::make_pair(P, (Value*)0));
1042 assert(I != AvailablePreds.end() && I->first == P &&
1043 "Didn't find entry for predecessor!");
1045 PN->addIncoming(I->second, I->first);
1048 //cerr << "PRE: " << *LI << *PN << "\n";
1050 LI->replaceAllUsesWith(PN);
1051 LI->eraseFromParent();
1056 /// FindMostPopularDest - The specified list contains multiple possible
1057 /// threadable destinations. Pick the one that occurs the most frequently in
1060 FindMostPopularDest(BasicBlock *BB,
1061 const SmallVectorImpl<std::pair<BasicBlock*,
1062 BasicBlock*> > &PredToDestList) {
1063 assert(!PredToDestList.empty());
1065 // Determine popularity. If there are multiple possible destinations, we
1066 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1067 // blocks with known and real destinations to threading undef. We'll handle
1068 // them later if interesting.
1069 DenseMap<BasicBlock*, unsigned> DestPopularity;
1070 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1071 if (PredToDestList[i].second)
1072 DestPopularity[PredToDestList[i].second]++;
1074 // Find the most popular dest.
1075 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1076 BasicBlock *MostPopularDest = DPI->first;
1077 unsigned Popularity = DPI->second;
1078 SmallVector<BasicBlock*, 4> SamePopularity;
1080 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1081 // If the popularity of this entry isn't higher than the popularity we've
1082 // seen so far, ignore it.
1083 if (DPI->second < Popularity)
1085 else if (DPI->second == Popularity) {
1086 // If it is the same as what we've seen so far, keep track of it.
1087 SamePopularity.push_back(DPI->first);
1089 // If it is more popular, remember it.
1090 SamePopularity.clear();
1091 MostPopularDest = DPI->first;
1092 Popularity = DPI->second;
1096 // Okay, now we know the most popular destination. If there is more than
1097 // destination, we need to determine one. This is arbitrary, but we need
1098 // to make a deterministic decision. Pick the first one that appears in the
1100 if (!SamePopularity.empty()) {
1101 SamePopularity.push_back(MostPopularDest);
1102 TerminatorInst *TI = BB->getTerminator();
1103 for (unsigned i = 0; ; ++i) {
1104 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1106 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1107 TI->getSuccessor(i)) == SamePopularity.end())
1110 MostPopularDest = TI->getSuccessor(i);
1115 // Okay, we have finally picked the most popular destination.
1116 return MostPopularDest;
1119 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1120 ConstantPreference Preference) {
1121 // If threading this would thread across a loop header, don't even try to
1123 if (LoopHeaders.count(BB))
1126 PredValueInfoTy PredValues;
1127 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference))
1130 assert(!PredValues.empty() &&
1131 "ComputeValueKnownInPredecessors returned true with no values");
1133 DEBUG(dbgs() << "IN BB: " << *BB;
1134 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1135 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1136 << *PredValues[i].first
1137 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1140 // Decide what we want to thread through. Convert our list of known values to
1141 // a list of known destinations for each pred. This also discards duplicate
1142 // predecessors and keeps track of the undefined inputs (which are represented
1143 // as a null dest in the PredToDestList).
1144 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1145 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1147 BasicBlock *OnlyDest = 0;
1148 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1150 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1151 BasicBlock *Pred = PredValues[i].second;
1152 if (!SeenPreds.insert(Pred))
1153 continue; // Duplicate predecessor entry.
1155 // If the predecessor ends with an indirect goto, we can't change its
1157 if (isa<IndirectBrInst>(Pred->getTerminator()))
1160 Constant *Val = PredValues[i].first;
1163 if (isa<UndefValue>(Val))
1165 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1166 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1167 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
1168 DestBB = SI->getSuccessor(SI->findCaseValue(cast<ConstantInt>(Val)));
1170 assert(isa<IndirectBrInst>(BB->getTerminator())
1171 && "Unexpected terminator");
1172 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1175 // If we have exactly one destination, remember it for efficiency below.
1178 else if (OnlyDest != DestBB)
1179 OnlyDest = MultipleDestSentinel;
1181 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1184 // If all edges were unthreadable, we fail.
1185 if (PredToDestList.empty())
1188 // Determine which is the most common successor. If we have many inputs and
1189 // this block is a switch, we want to start by threading the batch that goes
1190 // to the most popular destination first. If we only know about one
1191 // threadable destination (the common case) we can avoid this.
1192 BasicBlock *MostPopularDest = OnlyDest;
1194 if (MostPopularDest == MultipleDestSentinel)
1195 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1197 // Now that we know what the most popular destination is, factor all
1198 // predecessors that will jump to it into a single predecessor.
1199 SmallVector<BasicBlock*, 16> PredsToFactor;
1200 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1201 if (PredToDestList[i].second == MostPopularDest) {
1202 BasicBlock *Pred = PredToDestList[i].first;
1204 // This predecessor may be a switch or something else that has multiple
1205 // edges to the block. Factor each of these edges by listing them
1206 // according to # occurrences in PredsToFactor.
1207 TerminatorInst *PredTI = Pred->getTerminator();
1208 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1209 if (PredTI->getSuccessor(i) == BB)
1210 PredsToFactor.push_back(Pred);
1213 // If the threadable edges are branching on an undefined value, we get to pick
1214 // the destination that these predecessors should get to.
1215 if (MostPopularDest == 0)
1216 MostPopularDest = BB->getTerminator()->
1217 getSuccessor(GetBestDestForJumpOnUndef(BB));
1219 // Ok, try to thread it!
1220 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1223 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1224 /// a PHI node in the current block. See if there are any simplifications we
1225 /// can do based on inputs to the phi node.
1227 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1228 BasicBlock *BB = PN->getParent();
1230 // TODO: We could make use of this to do it once for blocks with common PHI
1232 SmallVector<BasicBlock*, 1> PredBBs;
1235 // If any of the predecessor blocks end in an unconditional branch, we can
1236 // *duplicate* the conditional branch into that block in order to further
1237 // encourage jump threading and to eliminate cases where we have branch on a
1238 // phi of an icmp (branch on icmp is much better).
1239 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1240 BasicBlock *PredBB = PN->getIncomingBlock(i);
1241 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1242 if (PredBr->isUnconditional()) {
1243 PredBBs[0] = PredBB;
1244 // Try to duplicate BB into PredBB.
1245 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1253 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1254 /// a xor instruction in the current block. See if there are any
1255 /// simplifications we can do based on inputs to the xor.
1257 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1258 BasicBlock *BB = BO->getParent();
1260 // If either the LHS or RHS of the xor is a constant, don't do this
1262 if (isa<ConstantInt>(BO->getOperand(0)) ||
1263 isa<ConstantInt>(BO->getOperand(1)))
1266 // If the first instruction in BB isn't a phi, we won't be able to infer
1267 // anything special about any particular predecessor.
1268 if (!isa<PHINode>(BB->front()))
1271 // If we have a xor as the branch input to this block, and we know that the
1272 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1273 // the condition into the predecessor and fix that value to true, saving some
1274 // logical ops on that path and encouraging other paths to simplify.
1276 // This copies something like this:
1279 // %X = phi i1 [1], [%X']
1280 // %Y = icmp eq i32 %A, %B
1281 // %Z = xor i1 %X, %Y
1286 // %Y = icmp ne i32 %A, %B
1289 PredValueInfoTy XorOpValues;
1291 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1293 assert(XorOpValues.empty());
1294 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1300 assert(!XorOpValues.empty() &&
1301 "ComputeValueKnownInPredecessors returned true with no values");
1303 // Scan the information to see which is most popular: true or false. The
1304 // predecessors can be of the set true, false, or undef.
1305 unsigned NumTrue = 0, NumFalse = 0;
1306 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1307 if (isa<UndefValue>(XorOpValues[i].first))
1308 // Ignore undefs for the count.
1310 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1316 // Determine which value to split on, true, false, or undef if neither.
1317 ConstantInt *SplitVal = 0;
1318 if (NumTrue > NumFalse)
1319 SplitVal = ConstantInt::getTrue(BB->getContext());
1320 else if (NumTrue != 0 || NumFalse != 0)
1321 SplitVal = ConstantInt::getFalse(BB->getContext());
1323 // Collect all of the blocks that this can be folded into so that we can
1324 // factor this once and clone it once.
1325 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1326 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1327 if (XorOpValues[i].first != SplitVal &&
1328 !isa<UndefValue>(XorOpValues[i].first))
1331 BlocksToFoldInto.push_back(XorOpValues[i].second);
1334 // If we inferred a value for all of the predecessors, then duplication won't
1335 // help us. However, we can just replace the LHS or RHS with the constant.
1336 if (BlocksToFoldInto.size() ==
1337 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1338 if (SplitVal == 0) {
1339 // If all preds provide undef, just nuke the xor, because it is undef too.
1340 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1341 BO->eraseFromParent();
1342 } else if (SplitVal->isZero()) {
1343 // If all preds provide 0, replace the xor with the other input.
1344 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1345 BO->eraseFromParent();
1347 // If all preds provide 1, set the computed value to 1.
1348 BO->setOperand(!isLHS, SplitVal);
1354 // Try to duplicate BB into PredBB.
1355 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1359 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1360 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1361 /// NewPred using the entries from OldPred (suitably mapped).
1362 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1363 BasicBlock *OldPred,
1364 BasicBlock *NewPred,
1365 DenseMap<Instruction*, Value*> &ValueMap) {
1366 for (BasicBlock::iterator PNI = PHIBB->begin();
1367 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1368 // Ok, we have a PHI node. Figure out what the incoming value was for the
1370 Value *IV = PN->getIncomingValueForBlock(OldPred);
1372 // Remap the value if necessary.
1373 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1374 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1375 if (I != ValueMap.end())
1379 PN->addIncoming(IV, NewPred);
1383 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1384 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1385 /// across BB. Transform the IR to reflect this change.
1386 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1387 const SmallVectorImpl<BasicBlock*> &PredBBs,
1388 BasicBlock *SuccBB) {
1389 // If threading to the same block as we come from, we would infinite loop.
1391 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1392 << "' - would thread to self!\n");
1396 // If threading this would thread across a loop header, don't thread the edge.
1397 // See the comments above FindLoopHeaders for justifications and caveats.
1398 if (LoopHeaders.count(BB)) {
1399 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1400 << "' to dest BB '" << SuccBB->getName()
1401 << "' - it might create an irreducible loop!\n");
1405 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1406 if (JumpThreadCost > Threshold) {
1407 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1408 << "' - Cost is too high: " << JumpThreadCost << "\n");
1412 // And finally, do it! Start by factoring the predecessors is needed.
1414 if (PredBBs.size() == 1)
1415 PredBB = PredBBs[0];
1417 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1418 << " common predecessors.\n");
1419 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1423 // And finally, do it!
1424 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1425 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1426 << ", across block:\n "
1429 LVI->threadEdge(PredBB, BB, SuccBB);
1431 // We are going to have to map operands from the original BB block to the new
1432 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1433 // account for entry from PredBB.
1434 DenseMap<Instruction*, Value*> ValueMapping;
1436 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1437 BB->getName()+".thread",
1438 BB->getParent(), BB);
1439 NewBB->moveAfter(PredBB);
1441 BasicBlock::iterator BI = BB->begin();
1442 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1443 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1445 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1446 // mapping and using it to remap operands in the cloned instructions.
1447 for (; !isa<TerminatorInst>(BI); ++BI) {
1448 Instruction *New = BI->clone();
1449 New->setName(BI->getName());
1450 NewBB->getInstList().push_back(New);
1451 ValueMapping[BI] = New;
1453 // Remap operands to patch up intra-block references.
1454 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1455 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1456 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1457 if (I != ValueMapping.end())
1458 New->setOperand(i, I->second);
1462 // We didn't copy the terminator from BB over to NewBB, because there is now
1463 // an unconditional jump to SuccBB. Insert the unconditional jump.
1464 BranchInst::Create(SuccBB, NewBB);
1466 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1467 // PHI nodes for NewBB now.
1468 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1470 // If there were values defined in BB that are used outside the block, then we
1471 // now have to update all uses of the value to use either the original value,
1472 // the cloned value, or some PHI derived value. This can require arbitrary
1473 // PHI insertion, of which we are prepared to do, clean these up now.
1474 SSAUpdater SSAUpdate;
1475 SmallVector<Use*, 16> UsesToRename;
1476 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1477 // Scan all uses of this instruction to see if it is used outside of its
1478 // block, and if so, record them in UsesToRename.
1479 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1481 Instruction *User = cast<Instruction>(*UI);
1482 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1483 if (UserPN->getIncomingBlock(UI) == BB)
1485 } else if (User->getParent() == BB)
1488 UsesToRename.push_back(&UI.getUse());
1491 // If there are no uses outside the block, we're done with this instruction.
1492 if (UsesToRename.empty())
1495 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1497 // We found a use of I outside of BB. Rename all uses of I that are outside
1498 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1499 // with the two values we know.
1500 SSAUpdate.Initialize(I->getType(), I->getName());
1501 SSAUpdate.AddAvailableValue(BB, I);
1502 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1504 while (!UsesToRename.empty())
1505 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1506 DEBUG(dbgs() << "\n");
1510 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1511 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1512 // us to simplify any PHI nodes in BB.
1513 TerminatorInst *PredTerm = PredBB->getTerminator();
1514 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1515 if (PredTerm->getSuccessor(i) == BB) {
1516 BB->removePredecessor(PredBB, true);
1517 PredTerm->setSuccessor(i, NewBB);
1520 // At this point, the IR is fully up to date and consistent. Do a quick scan
1521 // over the new instructions and zap any that are constants or dead. This
1522 // frequently happens because of phi translation.
1523 SimplifyInstructionsInBlock(NewBB, TD);
1525 // Threaded an edge!
1530 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1531 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1532 /// If we can duplicate the contents of BB up into PredBB do so now, this
1533 /// improves the odds that the branch will be on an analyzable instruction like
1535 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1536 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1537 assert(!PredBBs.empty() && "Can't handle an empty set");
1539 // If BB is a loop header, then duplicating this block outside the loop would
1540 // cause us to transform this into an irreducible loop, don't do this.
1541 // See the comments above FindLoopHeaders for justifications and caveats.
1542 if (LoopHeaders.count(BB)) {
1543 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1544 << "' into predecessor block '" << PredBBs[0]->getName()
1545 << "' - it might create an irreducible loop!\n");
1549 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1550 if (DuplicationCost > Threshold) {
1551 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1552 << "' - Cost is too high: " << DuplicationCost << "\n");
1556 // And finally, do it! Start by factoring the predecessors is needed.
1558 if (PredBBs.size() == 1)
1559 PredBB = PredBBs[0];
1561 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1562 << " common predecessors.\n");
1563 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1567 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1569 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1570 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1571 << DuplicationCost << " block is:" << *BB << "\n");
1573 // Unless PredBB ends with an unconditional branch, split the edge so that we
1574 // can just clone the bits from BB into the end of the new PredBB.
1575 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1577 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1578 PredBB = SplitEdge(PredBB, BB, this);
1579 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1582 // We are going to have to map operands from the original BB block into the
1583 // PredBB block. Evaluate PHI nodes in BB.
1584 DenseMap<Instruction*, Value*> ValueMapping;
1586 BasicBlock::iterator BI = BB->begin();
1587 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1588 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1590 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1591 // mapping and using it to remap operands in the cloned instructions.
1592 for (; BI != BB->end(); ++BI) {
1593 Instruction *New = BI->clone();
1595 // Remap operands to patch up intra-block references.
1596 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1597 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1598 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1599 if (I != ValueMapping.end())
1600 New->setOperand(i, I->second);
1603 // If this instruction can be simplified after the operands are updated,
1604 // just use the simplified value instead. This frequently happens due to
1606 if (Value *IV = SimplifyInstruction(New, TD)) {
1608 ValueMapping[BI] = IV;
1610 // Otherwise, insert the new instruction into the block.
1611 New->setName(BI->getName());
1612 PredBB->getInstList().insert(OldPredBranch, New);
1613 ValueMapping[BI] = New;
1617 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1618 // add entries to the PHI nodes for branch from PredBB now.
1619 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1620 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1622 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1625 // If there were values defined in BB that are used outside the block, then we
1626 // now have to update all uses of the value to use either the original value,
1627 // the cloned value, or some PHI derived value. This can require arbitrary
1628 // PHI insertion, of which we are prepared to do, clean these up now.
1629 SSAUpdater SSAUpdate;
1630 SmallVector<Use*, 16> UsesToRename;
1631 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1632 // Scan all uses of this instruction to see if it is used outside of its
1633 // block, and if so, record them in UsesToRename.
1634 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1636 Instruction *User = cast<Instruction>(*UI);
1637 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1638 if (UserPN->getIncomingBlock(UI) == BB)
1640 } else if (User->getParent() == BB)
1643 UsesToRename.push_back(&UI.getUse());
1646 // If there are no uses outside the block, we're done with this instruction.
1647 if (UsesToRename.empty())
1650 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1652 // We found a use of I outside of BB. Rename all uses of I that are outside
1653 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1654 // with the two values we know.
1655 SSAUpdate.Initialize(I->getType(), I->getName());
1656 SSAUpdate.AddAvailableValue(BB, I);
1657 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1659 while (!UsesToRename.empty())
1660 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1661 DEBUG(dbgs() << "\n");
1664 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1666 BB->removePredecessor(PredBB, true);
1668 // Remove the unconditional branch at the end of the PredBB block.
1669 OldPredBranch->eraseFromParent();