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/Transforms/Utils/BasicBlockUtils.h"
22 #include "llvm/Transforms/Utils/Local.h"
23 #include "llvm/Transforms/Utils/SSAUpdater.h"
24 #include "llvm/Target/TargetData.h"
25 #include "llvm/ADT/DenseMap.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/ADT/STLExtras.h"
28 #include "llvm/ADT/SmallPtrSet.h"
29 #include "llvm/ADT/SmallSet.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/raw_ostream.h"
35 STATISTIC(NumThreads, "Number of jumps threaded");
36 STATISTIC(NumFolds, "Number of terminators folded");
37 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
39 static cl::opt<unsigned>
40 Threshold("jump-threading-threshold",
41 cl::desc("Max block size to duplicate for jump threading"),
42 cl::init(6), cl::Hidden);
44 // Turn on use of LazyValueInfo.
46 EnableLVI("enable-jump-threading-lvi", cl::ReallyHidden);
51 /// This pass performs 'jump threading', which looks at blocks that have
52 /// multiple predecessors and multiple successors. If one or more of the
53 /// predecessors of the block can be proven to always jump to one of the
54 /// successors, we forward the edge from the predecessor to the successor by
55 /// duplicating the contents of this block.
57 /// An example of when this can occur is code like this:
64 /// In this case, the unconditional branch at the end of the first if can be
65 /// revectored to the false side of the second if.
67 class JumpThreading : public FunctionPass {
71 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
73 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
76 static char ID; // Pass identification
77 JumpThreading() : FunctionPass(&ID) {}
79 bool runOnFunction(Function &F);
81 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
83 AU.addRequired<LazyValueInfo>();
86 void FindLoopHeaders(Function &F);
87 bool ProcessBlock(BasicBlock *BB);
88 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
90 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
93 typedef SmallVectorImpl<std::pair<ConstantInt*,
94 BasicBlock*> > PredValueInfo;
96 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
97 PredValueInfo &Result);
98 bool ProcessThreadableEdges(Instruction *CondInst, BasicBlock *BB);
101 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
102 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
104 bool ProcessJumpOnPHI(PHINode *PN);
106 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
110 char JumpThreading::ID = 0;
111 static RegisterPass<JumpThreading>
112 X("jump-threading", "Jump Threading");
114 // Public interface to the Jump Threading pass
115 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
117 /// runOnFunction - Top level algorithm.
119 bool JumpThreading::runOnFunction(Function &F) {
120 DEBUG(errs() << "Jump threading on function '" << F.getName() << "'\n");
121 TD = getAnalysisIfAvailable<TargetData>();
122 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0;
126 bool AnotherIteration = true, EverChanged = false;
127 while (AnotherIteration) {
128 AnotherIteration = false;
129 bool Changed = false;
130 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
132 // Thread all of the branches we can over this block.
133 while (ProcessBlock(BB))
138 // If the block is trivially dead, zap it. This eliminates the successor
139 // edges which simplifies the CFG.
140 if (pred_begin(BB) == pred_end(BB) &&
141 BB != &BB->getParent()->getEntryBlock()) {
142 DEBUG(errs() << " JT: Deleting dead block '" << BB->getName()
143 << "' with terminator: " << *BB->getTerminator() << '\n');
144 LoopHeaders.erase(BB);
147 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
148 // Can't thread an unconditional jump, but if the block is "almost
149 // empty", we can replace uses of it with uses of the successor and make
151 if (BI->isUnconditional() &&
152 BB != &BB->getParent()->getEntryBlock()) {
153 BasicBlock::iterator BBI = BB->getFirstNonPHI();
154 // Ignore dbg intrinsics.
155 while (isa<DbgInfoIntrinsic>(BBI))
157 // If the terminator is the only non-phi instruction, try to nuke it.
158 if (BBI->isTerminator()) {
159 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
160 // block, we have to make sure it isn't in the LoopHeaders set. We
161 // reinsert afterward in the rare case when the block isn't deleted.
162 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
164 if (TryToSimplifyUncondBranchFromEmptyBlock(BB))
166 else if (ErasedFromLoopHeaders)
167 LoopHeaders.insert(BB);
172 AnotherIteration = Changed;
173 EverChanged |= Changed;
180 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
181 /// thread across it.
182 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
183 /// Ignore PHI nodes, these will be flattened when duplication happens.
184 BasicBlock::const_iterator I = BB->getFirstNonPHI();
186 // FIXME: THREADING will delete values that are just used to compute the
187 // branch, so they shouldn't count against the duplication cost.
190 // Sum up the cost of each instruction until we get to the terminator. Don't
191 // include the terminator because the copy won't include it.
193 for (; !isa<TerminatorInst>(I); ++I) {
194 // Debugger intrinsics don't incur code size.
195 if (isa<DbgInfoIntrinsic>(I)) continue;
197 // If this is a pointer->pointer bitcast, it is free.
198 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType()))
201 // All other instructions count for at least one unit.
204 // Calls are more expensive. If they are non-intrinsic calls, we model them
205 // as having cost of 4. If they are a non-vector intrinsic, we model them
206 // as having cost of 2 total, and if they are a vector intrinsic, we model
207 // them as having cost 1.
208 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
209 if (!isa<IntrinsicInst>(CI))
211 else if (!isa<VectorType>(CI->getType()))
216 // Threading through a switch statement is particularly profitable. If this
217 // block ends in a switch, decrease its cost to make it more likely to happen.
218 if (isa<SwitchInst>(I))
219 Size = Size > 6 ? Size-6 : 0;
224 /// FindLoopHeaders - We do not want jump threading to turn proper loop
225 /// structures into irreducible loops. Doing this breaks up the loop nesting
226 /// hierarchy and pessimizes later transformations. To prevent this from
227 /// happening, we first have to find the loop headers. Here we approximate this
228 /// by finding targets of backedges in the CFG.
230 /// Note that there definitely are cases when we want to allow threading of
231 /// edges across a loop header. For example, threading a jump from outside the
232 /// loop (the preheader) to an exit block of the loop is definitely profitable.
233 /// It is also almost always profitable to thread backedges from within the loop
234 /// to exit blocks, and is often profitable to thread backedges to other blocks
235 /// within the loop (forming a nested loop). This simple analysis is not rich
236 /// enough to track all of these properties and keep it up-to-date as the CFG
237 /// mutates, so we don't allow any of these transformations.
239 void JumpThreading::FindLoopHeaders(Function &F) {
240 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
241 FindFunctionBackedges(F, Edges);
243 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
244 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
247 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
248 /// if we can infer that the value is a known ConstantInt in any of our
249 /// predecessors. If so, return the known list of value and pred BB in the
250 /// result vector. If a value is known to be undef, it is returned as null.
252 /// This returns true if there were any known values.
255 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){
256 // If V is a constantint, then it is known in all predecessors.
257 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) {
258 ConstantInt *CI = dyn_cast<ConstantInt>(V);
260 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
261 Result.push_back(std::make_pair(CI, *PI));
265 // If V is a non-instruction value, or an instruction in a different block,
266 // then it can't be derived from a PHI.
267 Instruction *I = dyn_cast<Instruction>(V);
268 if (I == 0 || I->getParent() != BB) {
270 // Okay, if this is a live-in value, see if it has a known value at the end
271 // of any of our predecessors.
273 // FIXME: This should be an edge property, not a block end property.
274 /// TODO: Per PR2563, we could infer value range information about a
275 /// predecessor based on its terminator.
278 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
279 // If the value is known by LazyValueInfo to be a constant in a
280 // predecessor, use that information to try to thread this block.
281 Constant *PredCst = LVI->getConstant(V, *PI);
283 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst)))
286 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), *PI));
289 return !Result.empty();
295 /// If I is a PHI node, then we know the incoming values for any constants.
296 if (PHINode *PN = dyn_cast<PHINode>(I)) {
297 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
298 Value *InVal = PN->getIncomingValue(i);
299 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) {
300 ConstantInt *CI = dyn_cast<ConstantInt>(InVal);
301 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i)));
304 return !Result.empty();
307 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals;
309 // Handle some boolean conditions.
310 if (I->getType()->getPrimitiveSizeInBits() == 1) {
312 // X & false -> false
313 if (I->getOpcode() == Instruction::Or ||
314 I->getOpcode() == Instruction::And) {
315 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
316 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals);
318 if (LHSVals.empty() && RHSVals.empty())
321 ConstantInt *InterestingVal;
322 if (I->getOpcode() == Instruction::Or)
323 InterestingVal = ConstantInt::getTrue(I->getContext());
325 InterestingVal = ConstantInt::getFalse(I->getContext());
327 // Scan for the sentinel.
328 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
329 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0)
330 Result.push_back(LHSVals[i]);
331 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
332 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0)
333 Result.push_back(RHSVals[i]);
334 return !Result.empty();
337 // Handle the NOT form of XOR.
338 if (I->getOpcode() == Instruction::Xor &&
339 isa<ConstantInt>(I->getOperand(1)) &&
340 cast<ConstantInt>(I->getOperand(1))->isOne()) {
341 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result);
345 // Invert the known values.
346 for (unsigned i = 0, e = Result.size(); i != e; ++i)
348 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first));
353 // Handle compare with phi operand, where the PHI is defined in this block.
354 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
355 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
356 if (PN && PN->getParent() == BB) {
357 // We can do this simplification if any comparisons fold to true or false.
359 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
360 BasicBlock *PredBB = PN->getIncomingBlock(i);
361 Value *LHS = PN->getIncomingValue(i);
362 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
364 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS);
365 if (Res == 0) continue;
367 if (isa<UndefValue>(Res))
368 Result.push_back(std::make_pair((ConstantInt*)0, PredBB));
369 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res))
370 Result.push_back(std::make_pair(CI, PredBB));
373 return !Result.empty();
376 // TODO: We could also recurse to see if we can determine constants another
384 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
385 /// in an undefined jump, decide which block is best to revector to.
387 /// Since we can pick an arbitrary destination, we pick the successor with the
388 /// fewest predecessors. This should reduce the in-degree of the others.
390 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
391 TerminatorInst *BBTerm = BB->getTerminator();
392 unsigned MinSucc = 0;
393 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
394 // Compute the successor with the minimum number of predecessors.
395 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
396 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
397 TestBB = BBTerm->getSuccessor(i);
398 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
399 if (NumPreds < MinNumPreds)
406 /// ProcessBlock - If there are any predecessors whose control can be threaded
407 /// through to a successor, transform them now.
408 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
409 // If this block has a single predecessor, and if that pred has a single
410 // successor, merge the blocks. This encourages recursive jump threading
411 // because now the condition in this block can be threaded through
412 // predecessors of our predecessor block.
413 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
414 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
416 // If SinglePred was a loop header, BB becomes one.
417 if (LoopHeaders.erase(SinglePred))
418 LoopHeaders.insert(BB);
420 // Remember if SinglePred was the entry block of the function. If so, we
421 // will need to move BB back to the entry position.
422 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
423 MergeBasicBlockIntoOnlyPred(BB);
425 if (isEntry && BB != &BB->getParent()->getEntryBlock())
426 BB->moveBefore(&BB->getParent()->getEntryBlock());
431 // Look to see if the terminator is a branch of switch, if not we can't thread
434 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
435 // Can't thread an unconditional jump.
436 if (BI->isUnconditional()) return false;
437 Condition = BI->getCondition();
438 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
439 Condition = SI->getCondition();
441 return false; // Must be an invoke.
443 // If the terminator of this block is branching on a constant, simplify the
444 // terminator to an unconditional branch. This can occur due to threading in
446 if (isa<ConstantInt>(Condition)) {
447 DEBUG(errs() << " In block '" << BB->getName()
448 << "' folding terminator: " << *BB->getTerminator() << '\n');
450 ConstantFoldTerminator(BB);
454 // If the terminator is branching on an undef, we can pick any of the
455 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
456 if (isa<UndefValue>(Condition)) {
457 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
459 // Fold the branch/switch.
460 TerminatorInst *BBTerm = BB->getTerminator();
461 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
462 if (i == BestSucc) continue;
463 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD);
466 DEBUG(errs() << " In block '" << BB->getName()
467 << "' folding undef terminator: " << *BBTerm << '\n');
468 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
469 BBTerm->eraseFromParent();
473 Instruction *CondInst = dyn_cast<Instruction>(Condition);
475 // If the condition is an instruction defined in another block, see if a
476 // predecessor has the same condition:
480 if (!Condition->hasOneUse() && // Multiple uses.
481 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition.
482 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
483 if (isa<BranchInst>(BB->getTerminator())) {
484 for (; PI != E; ++PI)
485 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
486 if (PBI->isConditional() && PBI->getCondition() == Condition &&
487 ProcessBranchOnDuplicateCond(*PI, BB))
490 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator");
491 for (; PI != E; ++PI)
492 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator()))
493 if (PSI->getCondition() == Condition &&
494 ProcessSwitchOnDuplicateCond(*PI, BB))
499 // All the rest of our checks depend on the condition being an instruction.
503 // See if this is a phi node in the current block.
504 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
505 if (PN->getParent() == BB)
506 return ProcessJumpOnPHI(PN);
508 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
509 if (!isa<PHINode>(CondCmp->getOperand(0)) ||
510 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB) {
511 // If we have a comparison, loop over the predecessors to see if there is
512 // a condition with a lexically identical value.
513 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
514 for (; PI != E; ++PI)
515 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
516 if (PBI->isConditional() && *PI != BB) {
517 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) {
518 if (CI->getOperand(0) == CondCmp->getOperand(0) &&
519 CI->getOperand(1) == CondCmp->getOperand(1) &&
520 CI->getPredicate() == CondCmp->getPredicate()) {
521 // TODO: Could handle things like (x != 4) --> (x == 17)
522 if (ProcessBranchOnDuplicateCond(*PI, BB))
530 // Check for some cases that are worth simplifying. Right now we want to look
531 // for loads that are used by a switch or by the condition for the branch. If
532 // we see one, check to see if it's partially redundant. If so, insert a PHI
533 // which can then be used to thread the values.
535 // This is particularly important because reg2mem inserts loads and stores all
536 // over the place, and this blocks jump threading if we don't zap them.
537 Value *SimplifyValue = CondInst;
538 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
539 if (isa<Constant>(CondCmp->getOperand(1)))
540 SimplifyValue = CondCmp->getOperand(0);
542 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
543 if (SimplifyPartiallyRedundantLoad(LI))
547 // Handle a variety of cases where we are branching on something derived from
548 // a PHI node in the current block. If we can prove that any predecessors
549 // compute a predictable value based on a PHI node, thread those predecessors.
551 if (ProcessThreadableEdges(CondInst, BB))
555 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
556 // "(X == 4)" thread through this block.
561 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
562 /// block that jump on exactly the same condition. This means that we almost
563 /// always know the direction of the edge in the DESTBB:
565 /// br COND, DESTBB, BBY
567 /// br COND, BBZ, BBW
569 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
570 /// in DESTBB, we have to thread over it.
571 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
573 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
575 // If both successors of PredBB go to DESTBB, we don't know anything. We can
576 // fold the branch to an unconditional one, which allows other recursive
579 if (PredBI->getSuccessor(1) != BB)
581 else if (PredBI->getSuccessor(0) != BB)
584 DEBUG(errs() << " In block '" << PredBB->getName()
585 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
587 ConstantFoldTerminator(PredBB);
591 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
593 // If the dest block has one predecessor, just fix the branch condition to a
594 // constant and fold it.
595 if (BB->getSinglePredecessor()) {
596 DEBUG(errs() << " In block '" << BB->getName()
597 << "' folding condition to '" << BranchDir << "': "
598 << *BB->getTerminator() << '\n');
600 Value *OldCond = DestBI->getCondition();
601 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
603 ConstantFoldTerminator(BB);
604 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
609 // Next, figure out which successor we are threading to.
610 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
612 SmallVector<BasicBlock*, 2> Preds;
613 Preds.push_back(PredBB);
615 // Ok, try to thread it!
616 return ThreadEdge(BB, Preds, SuccBB);
619 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
620 /// block that switch on exactly the same condition. This means that we almost
621 /// always know the direction of the edge in the DESTBB:
623 /// switch COND [... DESTBB, BBY ... ]
625 /// switch COND [... BBZ, BBW ]
627 /// Optimizing switches like this is very important, because simplifycfg builds
628 /// switches out of repeated 'if' conditions.
629 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
630 BasicBlock *DestBB) {
631 // Can't thread edge to self.
632 if (PredBB == DestBB)
635 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
636 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
638 // There are a variety of optimizations that we can potentially do on these
639 // blocks: we order them from most to least preferable.
641 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
642 // directly to their destination. This does not introduce *any* code size
643 // growth. Skip debug info first.
644 BasicBlock::iterator BBI = DestBB->begin();
645 while (isa<DbgInfoIntrinsic>(BBI))
648 // FIXME: Thread if it just contains a PHI.
649 if (isa<SwitchInst>(BBI)) {
650 bool MadeChange = false;
651 // Ignore the default edge for now.
652 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
653 ConstantInt *DestVal = DestSI->getCaseValue(i);
654 BasicBlock *DestSucc = DestSI->getSuccessor(i);
656 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
657 // PredSI has an explicit case for it. If so, forward. If it is covered
658 // by the default case, we can't update PredSI.
659 unsigned PredCase = PredSI->findCaseValue(DestVal);
660 if (PredCase == 0) continue;
662 // If PredSI doesn't go to DestBB on this value, then it won't reach the
663 // case on this condition.
664 if (PredSI->getSuccessor(PredCase) != DestBB &&
665 DestSI->getSuccessor(i) != DestBB)
668 // Otherwise, we're safe to make the change. Make sure that the edge from
669 // DestSI to DestSucc is not critical and has no PHI nodes.
670 DEBUG(errs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
671 DEBUG(errs() << "THROUGH: " << *DestSI);
673 // If the destination has PHI nodes, just split the edge for updating
675 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
676 SplitCriticalEdge(DestSI, i, this);
677 DestSucc = DestSI->getSuccessor(i);
679 FoldSingleEntryPHINodes(DestSucc);
680 PredSI->setSuccessor(PredCase, DestSucc);
692 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
693 /// load instruction, eliminate it by replacing it with a PHI node. This is an
694 /// important optimization that encourages jump threading, and needs to be run
695 /// interlaced with other jump threading tasks.
696 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
697 // Don't hack volatile loads.
698 if (LI->isVolatile()) return false;
700 // If the load is defined in a block with exactly one predecessor, it can't be
701 // partially redundant.
702 BasicBlock *LoadBB = LI->getParent();
703 if (LoadBB->getSinglePredecessor())
706 Value *LoadedPtr = LI->getOperand(0);
708 // If the loaded operand is defined in the LoadBB, it can't be available.
709 // FIXME: Could do PHI translation, that would be fun :)
710 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
711 if (PtrOp->getParent() == LoadBB)
714 // Scan a few instructions up from the load, to see if it is obviously live at
715 // the entry to its block.
716 BasicBlock::iterator BBIt = LI;
718 if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB,
720 // If the value if the load is locally available within the block, just use
721 // it. This frequently occurs for reg2mem'd allocas.
722 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
724 // If the returned value is the load itself, replace with an undef. This can
725 // only happen in dead loops.
726 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
727 LI->replaceAllUsesWith(AvailableVal);
728 LI->eraseFromParent();
732 // Otherwise, if we scanned the whole block and got to the top of the block,
733 // we know the block is locally transparent to the load. If not, something
734 // might clobber its value.
735 if (BBIt != LoadBB->begin())
739 SmallPtrSet<BasicBlock*, 8> PredsScanned;
740 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
741 AvailablePredsTy AvailablePreds;
742 BasicBlock *OneUnavailablePred = 0;
744 // If we got here, the loaded value is transparent through to the start of the
745 // block. Check to see if it is available in any of the predecessor blocks.
746 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
748 BasicBlock *PredBB = *PI;
750 // If we already scanned this predecessor, skip it.
751 if (!PredsScanned.insert(PredBB))
754 // Scan the predecessor to see if the value is available in the pred.
755 BBIt = PredBB->end();
756 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
757 if (!PredAvailable) {
758 OneUnavailablePred = PredBB;
762 // If so, this load is partially redundant. Remember this info so that we
763 // can create a PHI node.
764 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
767 // If the loaded value isn't available in any predecessor, it isn't partially
769 if (AvailablePreds.empty()) return false;
771 // Okay, the loaded value is available in at least one (and maybe all!)
772 // predecessors. If the value is unavailable in more than one unique
773 // predecessor, we want to insert a merge block for those common predecessors.
774 // This ensures that we only have to insert one reload, thus not increasing
776 BasicBlock *UnavailablePred = 0;
778 // If there is exactly one predecessor where the value is unavailable, the
779 // already computed 'OneUnavailablePred' block is it. If it ends in an
780 // unconditional branch, we know that it isn't a critical edge.
781 if (PredsScanned.size() == AvailablePreds.size()+1 &&
782 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
783 UnavailablePred = OneUnavailablePred;
784 } else if (PredsScanned.size() != AvailablePreds.size()) {
785 // Otherwise, we had multiple unavailable predecessors or we had a critical
786 // edge from the one.
787 SmallVector<BasicBlock*, 8> PredsToSplit;
788 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
790 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
791 AvailablePredSet.insert(AvailablePreds[i].first);
793 // Add all the unavailable predecessors to the PredsToSplit list.
794 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
796 if (!AvailablePredSet.count(*PI))
797 PredsToSplit.push_back(*PI);
799 // Split them out to their own block.
801 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
802 "thread-split", this);
805 // If the value isn't available in all predecessors, then there will be
806 // exactly one where it isn't available. Insert a load on that edge and add
807 // it to the AvailablePreds list.
808 if (UnavailablePred) {
809 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
810 "Can't handle critical edge here!");
811 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr",
812 UnavailablePred->getTerminator());
813 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
816 // Now we know that each predecessor of this block has a value in
817 // AvailablePreds, sort them for efficient access as we're walking the preds.
818 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
820 // Create a PHI node at the start of the block for the PRE'd load value.
821 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
824 // Insert new entries into the PHI for each predecessor. A single block may
825 // have multiple entries here.
826 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
828 AvailablePredsTy::iterator I =
829 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
830 std::make_pair(*PI, (Value*)0));
832 assert(I != AvailablePreds.end() && I->first == *PI &&
833 "Didn't find entry for predecessor!");
835 PN->addIncoming(I->second, I->first);
838 //cerr << "PRE: " << *LI << *PN << "\n";
840 LI->replaceAllUsesWith(PN);
841 LI->eraseFromParent();
846 /// FindMostPopularDest - The specified list contains multiple possible
847 /// threadable destinations. Pick the one that occurs the most frequently in
850 FindMostPopularDest(BasicBlock *BB,
851 const SmallVectorImpl<std::pair<BasicBlock*,
852 BasicBlock*> > &PredToDestList) {
853 assert(!PredToDestList.empty());
855 // Determine popularity. If there are multiple possible destinations, we
856 // explicitly choose to ignore 'undef' destinations. We prefer to thread
857 // blocks with known and real destinations to threading undef. We'll handle
858 // them later if interesting.
859 DenseMap<BasicBlock*, unsigned> DestPopularity;
860 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
861 if (PredToDestList[i].second)
862 DestPopularity[PredToDestList[i].second]++;
864 // Find the most popular dest.
865 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
866 BasicBlock *MostPopularDest = DPI->first;
867 unsigned Popularity = DPI->second;
868 SmallVector<BasicBlock*, 4> SamePopularity;
870 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
871 // If the popularity of this entry isn't higher than the popularity we've
872 // seen so far, ignore it.
873 if (DPI->second < Popularity)
875 else if (DPI->second == Popularity) {
876 // If it is the same as what we've seen so far, keep track of it.
877 SamePopularity.push_back(DPI->first);
879 // If it is more popular, remember it.
880 SamePopularity.clear();
881 MostPopularDest = DPI->first;
882 Popularity = DPI->second;
886 // Okay, now we know the most popular destination. If there is more than
887 // destination, we need to determine one. This is arbitrary, but we need
888 // to make a deterministic decision. Pick the first one that appears in the
890 if (!SamePopularity.empty()) {
891 SamePopularity.push_back(MostPopularDest);
892 TerminatorInst *TI = BB->getTerminator();
893 for (unsigned i = 0; ; ++i) {
894 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
896 if (std::find(SamePopularity.begin(), SamePopularity.end(),
897 TI->getSuccessor(i)) == SamePopularity.end())
900 MostPopularDest = TI->getSuccessor(i);
905 // Okay, we have finally picked the most popular destination.
906 return MostPopularDest;
909 bool JumpThreading::ProcessThreadableEdges(Instruction *CondInst,
911 // If threading this would thread across a loop header, don't even try to
913 if (LoopHeaders.count(BB))
916 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues;
917 if (!ComputeValueKnownInPredecessors(CondInst, BB, PredValues))
919 assert(!PredValues.empty() &&
920 "ComputeValueKnownInPredecessors returned true with no values");
922 DEBUG(errs() << "IN BB: " << *BB;
923 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
924 errs() << " BB '" << BB->getName() << "': FOUND condition = ";
925 if (PredValues[i].first)
926 errs() << *PredValues[i].first;
929 errs() << " for pred '" << PredValues[i].second->getName()
933 // Decide what we want to thread through. Convert our list of known values to
934 // a list of known destinations for each pred. This also discards duplicate
935 // predecessors and keeps track of the undefined inputs (which are represented
936 // as a null dest in the PredToDestList).
937 SmallPtrSet<BasicBlock*, 16> SeenPreds;
938 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
940 BasicBlock *OnlyDest = 0;
941 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
943 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
944 BasicBlock *Pred = PredValues[i].second;
945 if (!SeenPreds.insert(Pred))
946 continue; // Duplicate predecessor entry.
948 // If the predecessor ends with an indirect goto, we can't change its
950 if (isa<IndirectBrInst>(Pred->getTerminator()))
953 ConstantInt *Val = PredValues[i].first;
956 if (Val == 0) // Undef.
958 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
959 DestBB = BI->getSuccessor(Val->isZero());
961 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
962 DestBB = SI->getSuccessor(SI->findCaseValue(Val));
965 // If we have exactly one destination, remember it for efficiency below.
968 else if (OnlyDest != DestBB)
969 OnlyDest = MultipleDestSentinel;
971 PredToDestList.push_back(std::make_pair(Pred, DestBB));
974 // If all edges were unthreadable, we fail.
975 if (PredToDestList.empty())
978 // Determine which is the most common successor. If we have many inputs and
979 // this block is a switch, we want to start by threading the batch that goes
980 // to the most popular destination first. If we only know about one
981 // threadable destination (the common case) we can avoid this.
982 BasicBlock *MostPopularDest = OnlyDest;
984 if (MostPopularDest == MultipleDestSentinel)
985 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
987 // Now that we know what the most popular destination is, factor all
988 // predecessors that will jump to it into a single predecessor.
989 SmallVector<BasicBlock*, 16> PredsToFactor;
990 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
991 if (PredToDestList[i].second == MostPopularDest) {
992 BasicBlock *Pred = PredToDestList[i].first;
994 // This predecessor may be a switch or something else that has multiple
995 // edges to the block. Factor each of these edges by listing them
996 // according to # occurrences in PredsToFactor.
997 TerminatorInst *PredTI = Pred->getTerminator();
998 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
999 if (PredTI->getSuccessor(i) == BB)
1000 PredsToFactor.push_back(Pred);
1003 // If the threadable edges are branching on an undefined value, we get to pick
1004 // the destination that these predecessors should get to.
1005 if (MostPopularDest == 0)
1006 MostPopularDest = BB->getTerminator()->
1007 getSuccessor(GetBestDestForJumpOnUndef(BB));
1009 // Ok, try to thread it!
1010 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1013 /// ProcessJumpOnPHI - We have a conditional branch or switch on a PHI node in
1014 /// the current block. See if there are any simplifications we can do based on
1015 /// inputs to the phi node.
1017 bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) {
1018 BasicBlock *BB = PN->getParent();
1020 // If any of the predecessor blocks end in an unconditional branch, we can
1021 // *duplicate* the jump into that block in order to further encourage jump
1022 // threading and to eliminate cases where we have branch on a phi of an icmp
1023 // (branch on icmp is much better).
1025 // We don't want to do this tranformation for switches, because we don't
1026 // really want to duplicate a switch.
1027 if (isa<SwitchInst>(BB->getTerminator()))
1030 // Look for unconditional branch predecessors.
1031 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1032 BasicBlock *PredBB = PN->getIncomingBlock(i);
1033 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1034 if (PredBr->isUnconditional() &&
1035 // Try to duplicate BB into PredBB.
1036 DuplicateCondBranchOnPHIIntoPred(BB, PredBB))
1044 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1045 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1046 /// NewPred using the entries from OldPred (suitably mapped).
1047 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1048 BasicBlock *OldPred,
1049 BasicBlock *NewPred,
1050 DenseMap<Instruction*, Value*> &ValueMap) {
1051 for (BasicBlock::iterator PNI = PHIBB->begin();
1052 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1053 // Ok, we have a PHI node. Figure out what the incoming value was for the
1055 Value *IV = PN->getIncomingValueForBlock(OldPred);
1057 // Remap the value if necessary.
1058 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1059 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1060 if (I != ValueMap.end())
1064 PN->addIncoming(IV, NewPred);
1068 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1069 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1070 /// across BB. Transform the IR to reflect this change.
1071 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1072 const SmallVectorImpl<BasicBlock*> &PredBBs,
1073 BasicBlock *SuccBB) {
1074 // If threading to the same block as we come from, we would infinite loop.
1076 DEBUG(errs() << " Not threading across BB '" << BB->getName()
1077 << "' - would thread to self!\n");
1081 // If threading this would thread across a loop header, don't thread the edge.
1082 // See the comments above FindLoopHeaders for justifications and caveats.
1083 if (LoopHeaders.count(BB)) {
1084 DEBUG(errs() << " Not threading across loop header BB '" << BB->getName()
1085 << "' to dest BB '" << SuccBB->getName()
1086 << "' - it might create an irreducible loop!\n");
1090 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1091 if (JumpThreadCost > Threshold) {
1092 DEBUG(errs() << " Not threading BB '" << BB->getName()
1093 << "' - Cost is too high: " << JumpThreadCost << "\n");
1097 // And finally, do it! Start by factoring the predecessors is needed.
1099 if (PredBBs.size() == 1)
1100 PredBB = PredBBs[0];
1102 DEBUG(errs() << " Factoring out " << PredBBs.size()
1103 << " common predecessors.\n");
1104 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1108 // And finally, do it!
1109 DEBUG(errs() << " Threading edge from '" << PredBB->getName() << "' to '"
1110 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1111 << ", across block:\n "
1114 // We are going to have to map operands from the original BB block to the new
1115 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1116 // account for entry from PredBB.
1117 DenseMap<Instruction*, Value*> ValueMapping;
1119 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1120 BB->getName()+".thread",
1121 BB->getParent(), BB);
1122 NewBB->moveAfter(PredBB);
1124 BasicBlock::iterator BI = BB->begin();
1125 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1126 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1128 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1129 // mapping and using it to remap operands in the cloned instructions.
1130 for (; !isa<TerminatorInst>(BI); ++BI) {
1131 Instruction *New = BI->clone();
1132 New->setName(BI->getName());
1133 NewBB->getInstList().push_back(New);
1134 ValueMapping[BI] = New;
1136 // Remap operands to patch up intra-block references.
1137 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1138 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1139 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1140 if (I != ValueMapping.end())
1141 New->setOperand(i, I->second);
1145 // We didn't copy the terminator from BB over to NewBB, because there is now
1146 // an unconditional jump to SuccBB. Insert the unconditional jump.
1147 BranchInst::Create(SuccBB, NewBB);
1149 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1150 // PHI nodes for NewBB now.
1151 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1153 // If there were values defined in BB that are used outside the block, then we
1154 // now have to update all uses of the value to use either the original value,
1155 // the cloned value, or some PHI derived value. This can require arbitrary
1156 // PHI insertion, of which we are prepared to do, clean these up now.
1157 SSAUpdater SSAUpdate;
1158 SmallVector<Use*, 16> UsesToRename;
1159 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1160 // Scan all uses of this instruction to see if it is used outside of its
1161 // block, and if so, record them in UsesToRename.
1162 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1164 Instruction *User = cast<Instruction>(*UI);
1165 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1166 if (UserPN->getIncomingBlock(UI) == BB)
1168 } else if (User->getParent() == BB)
1171 UsesToRename.push_back(&UI.getUse());
1174 // If there are no uses outside the block, we're done with this instruction.
1175 if (UsesToRename.empty())
1178 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n");
1180 // We found a use of I outside of BB. Rename all uses of I that are outside
1181 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1182 // with the two values we know.
1183 SSAUpdate.Initialize(I);
1184 SSAUpdate.AddAvailableValue(BB, I);
1185 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1187 while (!UsesToRename.empty())
1188 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1189 DEBUG(errs() << "\n");
1193 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1194 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1195 // us to simplify any PHI nodes in BB.
1196 TerminatorInst *PredTerm = PredBB->getTerminator();
1197 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1198 if (PredTerm->getSuccessor(i) == BB) {
1199 RemovePredecessorAndSimplify(BB, PredBB, TD);
1200 PredTerm->setSuccessor(i, NewBB);
1203 // At this point, the IR is fully up to date and consistent. Do a quick scan
1204 // over the new instructions and zap any that are constants or dead. This
1205 // frequently happens because of phi translation.
1206 BI = NewBB->begin();
1207 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) {
1208 Instruction *Inst = BI++;
1210 if (Value *V = SimplifyInstruction(Inst, TD)) {
1211 WeakVH BIHandle(BI);
1212 ReplaceAndSimplifyAllUses(Inst, V, TD);
1214 BI = NewBB->begin();
1218 RecursivelyDeleteTriviallyDeadInstructions(Inst);
1221 // Threaded an edge!
1226 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1227 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1228 /// If we can duplicate the contents of BB up into PredBB do so now, this
1229 /// improves the odds that the branch will be on an analyzable instruction like
1231 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1232 BasicBlock *PredBB) {
1233 // If BB is a loop header, then duplicating this block outside the loop would
1234 // cause us to transform this into an irreducible loop, don't do this.
1235 // See the comments above FindLoopHeaders for justifications and caveats.
1236 if (LoopHeaders.count(BB)) {
1237 DEBUG(errs() << " Not duplicating loop header '" << BB->getName()
1238 << "' into predecessor block '" << PredBB->getName()
1239 << "' - it might create an irreducible loop!\n");
1243 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1244 if (DuplicationCost > Threshold) {
1245 DEBUG(errs() << " Not duplicating BB '" << BB->getName()
1246 << "' - Cost is too high: " << DuplicationCost << "\n");
1250 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1252 DEBUG(errs() << " Duplicating block '" << BB->getName() << "' into end of '"
1253 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1254 << DuplicationCost << " block is:" << *BB << "\n");
1256 // We are going to have to map operands from the original BB block into the
1257 // PredBB block. Evaluate PHI nodes in BB.
1258 DenseMap<Instruction*, Value*> ValueMapping;
1260 BasicBlock::iterator BI = BB->begin();
1261 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1262 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1264 BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1266 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1267 // mapping and using it to remap operands in the cloned instructions.
1268 for (; BI != BB->end(); ++BI) {
1269 Instruction *New = BI->clone();
1270 New->setName(BI->getName());
1271 PredBB->getInstList().insert(OldPredBranch, New);
1272 ValueMapping[BI] = New;
1274 // Remap operands to patch up intra-block references.
1275 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1276 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1277 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1278 if (I != ValueMapping.end())
1279 New->setOperand(i, I->second);
1283 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1284 // add entries to the PHI nodes for branch from PredBB now.
1285 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1286 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1288 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1291 // If there were values defined in BB that are used outside the block, then we
1292 // now have to update all uses of the value to use either the original value,
1293 // the cloned value, or some PHI derived value. This can require arbitrary
1294 // PHI insertion, of which we are prepared to do, clean these up now.
1295 SSAUpdater SSAUpdate;
1296 SmallVector<Use*, 16> UsesToRename;
1297 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1298 // Scan all uses of this instruction to see if it is used outside of its
1299 // block, and if so, record them in UsesToRename.
1300 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1302 Instruction *User = cast<Instruction>(*UI);
1303 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1304 if (UserPN->getIncomingBlock(UI) == BB)
1306 } else if (User->getParent() == BB)
1309 UsesToRename.push_back(&UI.getUse());
1312 // If there are no uses outside the block, we're done with this instruction.
1313 if (UsesToRename.empty())
1316 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n");
1318 // We found a use of I outside of BB. Rename all uses of I that are outside
1319 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1320 // with the two values we know.
1321 SSAUpdate.Initialize(I);
1322 SSAUpdate.AddAvailableValue(BB, I);
1323 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1325 while (!UsesToRename.empty())
1326 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1327 DEBUG(errs() << "\n");
1330 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1332 RemovePredecessorAndSimplify(BB, PredBB, TD);
1334 // Remove the unconditional branch at the end of the PredBB block.
1335 OldPredBranch->eraseFromParent();