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/ConstantFolding.h"
20 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
21 #include "llvm/Transforms/Utils/Local.h"
22 #include "llvm/Transforms/Utils/SSAUpdater.h"
23 #include "llvm/Target/TargetData.h"
24 #include "llvm/ADT/DenseMap.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/STLExtras.h"
27 #include "llvm/ADT/SmallPtrSet.h"
28 #include "llvm/ADT/SmallSet.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/raw_ostream.h"
34 STATISTIC(NumThreads, "Number of jumps threaded");
35 STATISTIC(NumFolds, "Number of terminators folded");
36 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
38 static cl::opt<unsigned>
39 Threshold("jump-threading-threshold",
40 cl::desc("Max block size to duplicate for jump threading"),
41 cl::init(6), cl::Hidden);
44 /// This pass performs 'jump threading', which looks at blocks that have
45 /// multiple predecessors and multiple successors. If one or more of the
46 /// predecessors of the block can be proven to always jump to one of the
47 /// successors, we forward the edge from the predecessor to the successor by
48 /// duplicating the contents of this block.
50 /// An example of when this can occur is code like this:
57 /// In this case, the unconditional branch at the end of the first if can be
58 /// revectored to the false side of the second if.
60 class JumpThreading : public FunctionPass {
63 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
65 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
68 static char ID; // Pass identification
69 JumpThreading() : FunctionPass(&ID) {}
71 bool runOnFunction(Function &F);
72 void FindLoopHeaders(Function &F);
74 bool ProcessBlock(BasicBlock *BB);
75 bool ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, BasicBlock *SuccBB);
76 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
79 typedef SmallVectorImpl<std::pair<ConstantInt*,
80 BasicBlock*> > PredValueInfo;
82 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
83 PredValueInfo &Result);
84 bool ProcessThreadableEdges(Instruction *CondInst, BasicBlock *BB);
87 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
88 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
90 bool ProcessJumpOnPHI(PHINode *PN);
92 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
96 char JumpThreading::ID = 0;
97 static RegisterPass<JumpThreading>
98 X("jump-threading", "Jump Threading");
100 // Public interface to the Jump Threading pass
101 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
103 /// runOnFunction - Top level algorithm.
105 bool JumpThreading::runOnFunction(Function &F) {
106 DEBUG(errs() << "Jump threading on function '" << F.getName() << "'\n");
107 TD = getAnalysisIfAvailable<TargetData>();
111 bool AnotherIteration = true, EverChanged = false;
112 while (AnotherIteration) {
113 AnotherIteration = false;
114 bool Changed = false;
115 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
117 while (ProcessBlock(BB))
122 // If the block is trivially dead, zap it. This eliminates the successor
123 // edges which simplifies the CFG.
124 if (pred_begin(BB) == pred_end(BB) &&
125 BB != &BB->getParent()->getEntryBlock()) {
126 DEBUG(errs() << " JT: Deleting dead block '" << BB->getName()
127 << "' with terminator: " << *BB->getTerminator() << '\n');
128 LoopHeaders.erase(BB);
133 AnotherIteration = Changed;
134 EverChanged |= Changed;
141 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
142 /// thread across it.
143 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
144 /// Ignore PHI nodes, these will be flattened when duplication happens.
145 BasicBlock::const_iterator I = BB->getFirstNonPHI();
147 // Sum up the cost of each instruction until we get to the terminator. Don't
148 // include the terminator because the copy won't include it.
150 for (; !isa<TerminatorInst>(I); ++I) {
151 // Debugger intrinsics don't incur code size.
152 if (isa<DbgInfoIntrinsic>(I)) continue;
154 // If this is a pointer->pointer bitcast, it is free.
155 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType()))
158 // All other instructions count for at least one unit.
161 // Calls are more expensive. If they are non-intrinsic calls, we model them
162 // as having cost of 4. If they are a non-vector intrinsic, we model them
163 // as having cost of 2 total, and if they are a vector intrinsic, we model
164 // them as having cost 1.
165 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
166 if (!isa<IntrinsicInst>(CI))
168 else if (!isa<VectorType>(CI->getType()))
173 // Threading through a switch statement is particularly profitable. If this
174 // block ends in a switch, decrease its cost to make it more likely to happen.
175 if (isa<SwitchInst>(I))
176 Size = Size > 6 ? Size-6 : 0;
183 /// FindLoopHeaders - We do not want jump threading to turn proper loop
184 /// structures into irreducible loops. Doing this breaks up the loop nesting
185 /// hierarchy and pessimizes later transformations. To prevent this from
186 /// happening, we first have to find the loop headers. Here we approximate this
187 /// by finding targets of backedges in the CFG.
189 /// Note that there definitely are cases when we want to allow threading of
190 /// edges across a loop header. For example, threading a jump from outside the
191 /// loop (the preheader) to an exit block of the loop is definitely profitable.
192 /// It is also almost always profitable to thread backedges from within the loop
193 /// to exit blocks, and is often profitable to thread backedges to other blocks
194 /// within the loop (forming a nested loop). This simple analysis is not rich
195 /// enough to track all of these properties and keep it up-to-date as the CFG
196 /// mutates, so we don't allow any of these transformations.
198 void JumpThreading::FindLoopHeaders(Function &F) {
199 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
200 FindFunctionBackedges(F, Edges);
202 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
203 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
206 /// GetResultOfComparison - Given an icmp/fcmp predicate and the left and right
207 /// hand sides of the compare instruction, try to determine the result. If the
208 /// result can not be determined, a null pointer is returned.
209 static Constant *GetResultOfComparison(CmpInst::Predicate pred,
210 Value *LHS, Value *RHS) {
211 if (Constant *CLHS = dyn_cast<Constant>(LHS))
212 if (Constant *CRHS = dyn_cast<Constant>(RHS))
213 return ConstantExpr::getCompare(pred, CLHS, CRHS);
216 if (isa<IntegerType>(LHS->getType()) || isa<PointerType>(LHS->getType()))
217 if (ICmpInst::isTrueWhenEqual(pred))
218 return ConstantInt::getTrue(LHS->getContext());
220 return ConstantInt::getFalse(LHS->getContext());
225 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
226 /// if we can infer that the value is a known ConstantInt in any of our
227 /// predecessors. If so, return the known the list of value and pred BB in the
228 /// result vector. If a value is known to be undef, it is returned as null.
230 /// The BB basic block is known to start with a PHI node.
232 /// This returns true if there were any known values.
235 /// TODO: Per PR2563, we could infer value range information about a predecessor
236 /// based on its terminator.
238 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){
239 PHINode *TheFirstPHI = cast<PHINode>(BB->begin());
241 // If V is a constantint, then it is known in all predecessors.
242 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) {
243 ConstantInt *CI = dyn_cast<ConstantInt>(V);
244 Result.resize(TheFirstPHI->getNumIncomingValues());
245 for (unsigned i = 0, e = Result.size(); i != e; ++i)
246 Result[i] = std::make_pair(CI, TheFirstPHI->getIncomingBlock(i));
250 // If V is a non-instruction value, or an instruction in a different block,
251 // then it can't be derived from a PHI.
252 Instruction *I = dyn_cast<Instruction>(V);
253 if (I == 0 || I->getParent() != BB)
256 /// If I is a PHI node, then we know the incoming values for any constants.
257 if (PHINode *PN = dyn_cast<PHINode>(I)) {
258 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
259 Value *InVal = PN->getIncomingValue(i);
260 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) {
261 ConstantInt *CI = dyn_cast<ConstantInt>(InVal);
262 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i)));
265 return !Result.empty();
268 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals;
270 // Handle some boolean conditions.
271 if (I->getType()->getPrimitiveSizeInBits() == 1) {
273 // X & false -> false
274 if (I->getOpcode() == Instruction::Or ||
275 I->getOpcode() == Instruction::And) {
276 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
277 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals);
279 if (LHSVals.empty() && RHSVals.empty())
282 ConstantInt *InterestingVal;
283 if (I->getOpcode() == Instruction::Or)
284 InterestingVal = ConstantInt::getTrue(I->getContext());
286 InterestingVal = ConstantInt::getFalse(I->getContext());
288 // Scan for the sentinel.
289 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
290 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0)
291 Result.push_back(LHSVals[i]);
292 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
293 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0)
294 Result.push_back(RHSVals[i]);
295 return !Result.empty();
298 // TODO: Should handle the NOT form of XOR.
302 // Handle compare with phi operand, where the PHI is defined in this block.
303 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
304 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
305 if (PN && PN->getParent() == BB) {
306 // We can do this simplification if any comparisons fold to true or false.
308 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
309 BasicBlock *PredBB = PN->getIncomingBlock(i);
310 Value *LHS = PN->getIncomingValue(i);
311 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
313 Constant *Res = GetResultOfComparison(Cmp->getPredicate(), LHS, RHS);
314 if (Res == 0) continue;
316 if (isa<UndefValue>(Res))
317 Result.push_back(std::make_pair((ConstantInt*)0, PredBB));
318 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res))
319 Result.push_back(std::make_pair(CI, PredBB));
322 return !Result.empty();
325 // TODO: We could also recurse to see if we can determine constants another
333 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
334 /// in an undefined jump, decide which block is best to revector to.
336 /// Since we can pick an arbitrary destination, we pick the successor with the
337 /// fewest predecessors. This should reduce the in-degree of the others.
339 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
340 TerminatorInst *BBTerm = BB->getTerminator();
341 unsigned MinSucc = 0;
342 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
343 // Compute the successor with the minimum number of predecessors.
344 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
345 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
346 TestBB = BBTerm->getSuccessor(i);
347 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
348 if (NumPreds < MinNumPreds)
355 /// ProcessBlock - If there are any predecessors whose control can be threaded
356 /// through to a successor, transform them now.
357 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
358 // If this block has a single predecessor, and if that pred has a single
359 // successor, merge the blocks. This encourages recursive jump threading
360 // because now the condition in this block can be threaded through
361 // predecessors of our predecessor block.
362 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
363 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
365 // If SinglePred was a loop header, BB becomes one.
366 if (LoopHeaders.erase(SinglePred))
367 LoopHeaders.insert(BB);
369 // Remember if SinglePred was the entry block of the function. If so, we
370 // will need to move BB back to the entry position.
371 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
372 MergeBasicBlockIntoOnlyPred(BB);
374 if (isEntry && BB != &BB->getParent()->getEntryBlock())
375 BB->moveBefore(&BB->getParent()->getEntryBlock());
380 // Look to see if the terminator is a branch of switch, if not we can't thread
383 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
384 // Can't thread an unconditional jump.
385 if (BI->isUnconditional()) return false;
386 Condition = BI->getCondition();
387 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
388 Condition = SI->getCondition();
390 return false; // Must be an invoke.
392 // If the terminator of this block is branching on a constant, simplify the
393 // terminator to an unconditional branch. This can occur due to threading in
395 if (isa<ConstantInt>(Condition)) {
396 DEBUG(errs() << " In block '" << BB->getName()
397 << "' folding terminator: " << *BB->getTerminator() << '\n');
399 ConstantFoldTerminator(BB);
403 // If the terminator is branching on an undef, we can pick any of the
404 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
405 if (isa<UndefValue>(Condition)) {
406 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
408 // Fold the branch/switch.
409 TerminatorInst *BBTerm = BB->getTerminator();
410 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
411 if (i == BestSucc) continue;
412 BBTerm->getSuccessor(i)->removePredecessor(BB);
415 DEBUG(errs() << " In block '" << BB->getName()
416 << "' folding undef terminator: " << *BBTerm << '\n');
417 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
418 BBTerm->eraseFromParent();
422 Instruction *CondInst = dyn_cast<Instruction>(Condition);
424 // If the condition is an instruction defined in another block, see if a
425 // predecessor has the same condition:
429 if (!Condition->hasOneUse() && // Multiple uses.
430 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition.
431 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
432 if (isa<BranchInst>(BB->getTerminator())) {
433 for (; PI != E; ++PI)
434 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
435 if (PBI->isConditional() && PBI->getCondition() == Condition &&
436 ProcessBranchOnDuplicateCond(*PI, BB))
439 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator");
440 for (; PI != E; ++PI)
441 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator()))
442 if (PSI->getCondition() == Condition &&
443 ProcessSwitchOnDuplicateCond(*PI, BB))
448 // All the rest of our checks depend on the condition being an instruction.
452 // See if this is a phi node in the current block.
453 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
454 if (PN->getParent() == BB)
455 return ProcessJumpOnPHI(PN);
457 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
458 if (!isa<PHINode>(CondCmp->getOperand(0)) ||
459 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB) {
460 // If we have a comparison, loop over the predecessors to see if there is
461 // a condition with a lexically identical value.
462 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
463 for (; PI != E; ++PI)
464 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
465 if (PBI->isConditional() && *PI != BB) {
466 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) {
467 if (CI->getOperand(0) == CondCmp->getOperand(0) &&
468 CI->getOperand(1) == CondCmp->getOperand(1) &&
469 CI->getPredicate() == CondCmp->getPredicate()) {
470 // TODO: Could handle things like (x != 4) --> (x == 17)
471 if (ProcessBranchOnDuplicateCond(*PI, BB))
479 // Check for some cases that are worth simplifying. Right now we want to look
480 // for loads that are used by a switch or by the condition for the branch. If
481 // we see one, check to see if it's partially redundant. If so, insert a PHI
482 // which can then be used to thread the values.
484 // This is particularly important because reg2mem inserts loads and stores all
485 // over the place, and this blocks jump threading if we don't zap them.
486 Value *SimplifyValue = CondInst;
487 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
488 if (isa<Constant>(CondCmp->getOperand(1)))
489 SimplifyValue = CondCmp->getOperand(0);
491 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
492 if (SimplifyPartiallyRedundantLoad(LI))
496 // Handle a variety of cases where we are branching on something derived from
497 // a PHI node in the current block. If we can prove that any predecessors
498 // compute a predictable value based on a PHI node, thread those predecessors.
500 // We only bother doing this if the current block has a PHI node and if the
501 // conditional instruction lives in the current block. If either condition
502 // fail, this won't be a computable value anyway.
503 if (CondInst->getParent() == BB && isa<PHINode>(BB->front()))
504 if (ProcessThreadableEdges(CondInst, BB))
508 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
509 // "(X == 4)" thread through this block.
514 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
515 /// block that jump on exactly the same condition. This means that we almost
516 /// always know the direction of the edge in the DESTBB:
518 /// br COND, DESTBB, BBY
520 /// br COND, BBZ, BBW
522 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
523 /// in DESTBB, we have to thread over it.
524 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
526 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
528 // If both successors of PredBB go to DESTBB, we don't know anything. We can
529 // fold the branch to an unconditional one, which allows other recursive
532 if (PredBI->getSuccessor(1) != BB)
534 else if (PredBI->getSuccessor(0) != BB)
537 DEBUG(errs() << " In block '" << PredBB->getName()
538 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
540 ConstantFoldTerminator(PredBB);
544 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
546 // If the dest block has one predecessor, just fix the branch condition to a
547 // constant and fold it.
548 if (BB->getSinglePredecessor()) {
549 DEBUG(errs() << " In block '" << BB->getName()
550 << "' folding condition to '" << BranchDir << "': "
551 << *BB->getTerminator() << '\n');
553 Value *OldCond = DestBI->getCondition();
554 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
556 ConstantFoldTerminator(BB);
557 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
562 // Next, figure out which successor we are threading to.
563 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
565 // Ok, try to thread it!
566 return ThreadEdge(BB, PredBB, SuccBB);
569 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
570 /// block that switch on exactly the same condition. This means that we almost
571 /// always know the direction of the edge in the DESTBB:
573 /// switch COND [... DESTBB, BBY ... ]
575 /// switch COND [... BBZ, BBW ]
577 /// Optimizing switches like this is very important, because simplifycfg builds
578 /// switches out of repeated 'if' conditions.
579 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
580 BasicBlock *DestBB) {
581 // Can't thread edge to self.
582 if (PredBB == DestBB)
585 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
586 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
588 // There are a variety of optimizations that we can potentially do on these
589 // blocks: we order them from most to least preferable.
591 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
592 // directly to their destination. This does not introduce *any* code size
593 // growth. Skip debug info first.
594 BasicBlock::iterator BBI = DestBB->begin();
595 while (isa<DbgInfoIntrinsic>(BBI))
598 // FIXME: Thread if it just contains a PHI.
599 if (isa<SwitchInst>(BBI)) {
600 bool MadeChange = false;
601 // Ignore the default edge for now.
602 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
603 ConstantInt *DestVal = DestSI->getCaseValue(i);
604 BasicBlock *DestSucc = DestSI->getSuccessor(i);
606 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
607 // PredSI has an explicit case for it. If so, forward. If it is covered
608 // by the default case, we can't update PredSI.
609 unsigned PredCase = PredSI->findCaseValue(DestVal);
610 if (PredCase == 0) continue;
612 // If PredSI doesn't go to DestBB on this value, then it won't reach the
613 // case on this condition.
614 if (PredSI->getSuccessor(PredCase) != DestBB &&
615 DestSI->getSuccessor(i) != DestBB)
618 // Otherwise, we're safe to make the change. Make sure that the edge from
619 // DestSI to DestSucc is not critical and has no PHI nodes.
620 DEBUG(errs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
621 DEBUG(errs() << "THROUGH: " << *DestSI);
623 // If the destination has PHI nodes, just split the edge for updating
625 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
626 SplitCriticalEdge(DestSI, i, this);
627 DestSucc = DestSI->getSuccessor(i);
629 FoldSingleEntryPHINodes(DestSucc);
630 PredSI->setSuccessor(PredCase, DestSucc);
642 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
643 /// load instruction, eliminate it by replacing it with a PHI node. This is an
644 /// important optimization that encourages jump threading, and needs to be run
645 /// interlaced with other jump threading tasks.
646 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
647 // Don't hack volatile loads.
648 if (LI->isVolatile()) return false;
650 // If the load is defined in a block with exactly one predecessor, it can't be
651 // partially redundant.
652 BasicBlock *LoadBB = LI->getParent();
653 if (LoadBB->getSinglePredecessor())
656 Value *LoadedPtr = LI->getOperand(0);
658 // If the loaded operand is defined in the LoadBB, it can't be available.
659 // FIXME: Could do PHI translation, that would be fun :)
660 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
661 if (PtrOp->getParent() == LoadBB)
664 // Scan a few instructions up from the load, to see if it is obviously live at
665 // the entry to its block.
666 BasicBlock::iterator BBIt = LI;
668 if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB,
670 // If the value if the load is locally available within the block, just use
671 // it. This frequently occurs for reg2mem'd allocas.
672 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
674 // If the returned value is the load itself, replace with an undef. This can
675 // only happen in dead loops.
676 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
677 LI->replaceAllUsesWith(AvailableVal);
678 LI->eraseFromParent();
682 // Otherwise, if we scanned the whole block and got to the top of the block,
683 // we know the block is locally transparent to the load. If not, something
684 // might clobber its value.
685 if (BBIt != LoadBB->begin())
689 SmallPtrSet<BasicBlock*, 8> PredsScanned;
690 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
691 AvailablePredsTy AvailablePreds;
692 BasicBlock *OneUnavailablePred = 0;
694 // If we got here, the loaded value is transparent through to the start of the
695 // block. Check to see if it is available in any of the predecessor blocks.
696 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
698 BasicBlock *PredBB = *PI;
700 // If we already scanned this predecessor, skip it.
701 if (!PredsScanned.insert(PredBB))
704 // Scan the predecessor to see if the value is available in the pred.
705 BBIt = PredBB->end();
706 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
707 if (!PredAvailable) {
708 OneUnavailablePred = PredBB;
712 // If so, this load is partially redundant. Remember this info so that we
713 // can create a PHI node.
714 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
717 // If the loaded value isn't available in any predecessor, it isn't partially
719 if (AvailablePreds.empty()) return false;
721 // Okay, the loaded value is available in at least one (and maybe all!)
722 // predecessors. If the value is unavailable in more than one unique
723 // predecessor, we want to insert a merge block for those common predecessors.
724 // This ensures that we only have to insert one reload, thus not increasing
726 BasicBlock *UnavailablePred = 0;
728 // If there is exactly one predecessor where the value is unavailable, the
729 // already computed 'OneUnavailablePred' block is it. If it ends in an
730 // unconditional branch, we know that it isn't a critical edge.
731 if (PredsScanned.size() == AvailablePreds.size()+1 &&
732 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
733 UnavailablePred = OneUnavailablePred;
734 } else if (PredsScanned.size() != AvailablePreds.size()) {
735 // Otherwise, we had multiple unavailable predecessors or we had a critical
736 // edge from the one.
737 SmallVector<BasicBlock*, 8> PredsToSplit;
738 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
740 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
741 AvailablePredSet.insert(AvailablePreds[i].first);
743 // Add all the unavailable predecessors to the PredsToSplit list.
744 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
746 if (!AvailablePredSet.count(*PI))
747 PredsToSplit.push_back(*PI);
749 // Split them out to their own block.
751 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
752 "thread-split", this);
755 // If the value isn't available in all predecessors, then there will be
756 // exactly one where it isn't available. Insert a load on that edge and add
757 // it to the AvailablePreds list.
758 if (UnavailablePred) {
759 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
760 "Can't handle critical edge here!");
761 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr",
762 UnavailablePred->getTerminator());
763 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
766 // Now we know that each predecessor of this block has a value in
767 // AvailablePreds, sort them for efficient access as we're walking the preds.
768 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
770 // Create a PHI node at the start of the block for the PRE'd load value.
771 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
774 // Insert new entries into the PHI for each predecessor. A single block may
775 // have multiple entries here.
776 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
778 AvailablePredsTy::iterator I =
779 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
780 std::make_pair(*PI, (Value*)0));
782 assert(I != AvailablePreds.end() && I->first == *PI &&
783 "Didn't find entry for predecessor!");
785 PN->addIncoming(I->second, I->first);
788 //cerr << "PRE: " << *LI << *PN << "\n";
790 LI->replaceAllUsesWith(PN);
791 LI->eraseFromParent();
796 /// FindMostPopularDest - The specified list contains multiple possible
797 /// threadable destinations. Pick the one that occurs the most frequently in
800 FindMostPopularDest(BasicBlock *BB,
801 const SmallVectorImpl<std::pair<BasicBlock*,
802 BasicBlock*> > &PredToDestList) {
803 assert(!PredToDestList.empty());
805 // Determine popularity. If there are multiple possible destinations, we
806 // explicitly choose to ignore 'undef' destinations. We prefer to thread
807 // blocks with known and real destinations to threading undef. We'll handle
808 // them later if interesting.
809 DenseMap<BasicBlock*, unsigned> DestPopularity;
810 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
811 if (PredToDestList[i].second)
812 DestPopularity[PredToDestList[i].second]++;
814 // Find the most popular dest.
815 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
816 BasicBlock *MostPopularDest = DPI->first;
817 unsigned Popularity = DPI->second;
818 SmallVector<BasicBlock*, 4> SamePopularity;
820 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
821 // If the popularity of this entry isn't higher than the popularity we've
822 // seen so far, ignore it.
823 if (DPI->second < Popularity)
825 else if (DPI->second == Popularity) {
826 // If it is the same as what we've seen so far, keep track of it.
827 SamePopularity.push_back(DPI->first);
829 // If it is more popular, remember it.
830 SamePopularity.clear();
831 MostPopularDest = DPI->first;
832 Popularity = DPI->second;
836 // Okay, now we know the most popular destination. If there is more than
837 // destination, we need to determine one. This is arbitrary, but we need
838 // to make a deterministic decision. Pick the first one that appears in the
840 if (!SamePopularity.empty()) {
841 SamePopularity.push_back(MostPopularDest);
842 TerminatorInst *TI = BB->getTerminator();
843 for (unsigned i = 0; ; ++i) {
844 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
846 if (std::find(SamePopularity.begin(), SamePopularity.end(),
847 TI->getSuccessor(i)) == SamePopularity.end())
850 MostPopularDest = TI->getSuccessor(i);
855 // Okay, we have finally picked the most popular destination.
856 return MostPopularDest;
859 bool JumpThreading::ProcessThreadableEdges(Instruction *CondInst,
861 // If threading this would thread across a loop header, don't even try to
863 if (LoopHeaders.count(BB))
868 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues;
869 if (!ComputeValueKnownInPredecessors(CondInst, BB, PredValues))
871 assert(!PredValues.empty() &&
872 "ComputeValueKnownInPredecessors returned true with no values");
874 DEBUG(errs() << "IN BB: " << *BB;
875 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
876 errs() << " BB '" << BB->getName() << "': FOUND condition = ";
877 if (PredValues[i].first)
878 errs() << *PredValues[i].first;
881 errs() << " for pred '" << PredValues[i].second->getName()
885 // Decide what we want to thread through. Convert our list of known values to
886 // a list of known destinations for each pred. This also discards duplicate
887 // predecessors and keeps track of the undefined inputs (which are represented
888 // as a null dest in the PredToDestList.
889 SmallPtrSet<BasicBlock*, 16> SeenPreds;
890 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
892 BasicBlock *OnlyDest = 0;
893 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
895 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
896 BasicBlock *Pred = PredValues[i].second;
897 if (!SeenPreds.insert(Pred))
898 continue; // Duplicate predecessor entry.
900 // If the predecessor ends with an indirect goto, we can't change its
902 if (isa<IndirectBrInst>(Pred->getTerminator()))
905 ConstantInt *Val = PredValues[i].first;
908 if (Val == 0) // Undef.
910 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
911 DestBB = BI->getSuccessor(Val->isZero());
913 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
914 DestBB = SI->getSuccessor(SI->findCaseValue(Val));
917 // If we have exactly one destination, remember it for efficiency below.
920 else if (OnlyDest != DestBB)
921 OnlyDest = MultipleDestSentinel;
923 PredToDestList.push_back(std::make_pair(Pred, DestBB));
926 // If all edges were unthreadable, we fail.
927 if (PredToDestList.empty())
930 // Determine which is the most common successor. If we have many inputs and
931 // this block is a switch, we want to start by threading the batch that goes
932 // to the most popular destination first. If we only know about one
933 // threadable destination (the common case) we can avoid this.
934 BasicBlock *MostPopularDest = OnlyDest;
936 if (MostPopularDest == MultipleDestSentinel)
937 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
939 // Now that we know what the most popular destination is, factor all
940 // predecessors that will jump to it into a single predecessor.
941 SmallVector<BasicBlock*, 16> PredsToFactor;
942 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
943 if (PredToDestList[i].second == MostPopularDest) {
944 BasicBlock *Pred = PredToDestList[i].first;
946 // This predecessor may be a switch or something else that has multiple
947 // edges to the block. Factor each of these edges by listing them
948 // according to # occurrences in PredsToFactor.
949 TerminatorInst *PredTI = Pred->getTerminator();
950 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
951 if (PredTI->getSuccessor(i) == BB)
952 PredsToFactor.push_back(Pred);
955 BasicBlock *PredToThread;
956 if (PredsToFactor.size() == 1)
957 PredToThread = PredsToFactor[0];
959 DEBUG(errs() << " Factoring out " << PredsToFactor.size()
960 << " common predecessors.\n");
961 PredToThread = SplitBlockPredecessors(BB, &PredsToFactor[0],
962 PredsToFactor.size(),
966 // If the threadable edges are branching on an undefined value, we get to pick
967 // the destination that these predecessors should get to.
968 if (MostPopularDest == 0)
969 MostPopularDest = BB->getTerminator()->
970 getSuccessor(GetBestDestForJumpOnUndef(BB));
972 // Ok, try to thread it!
973 return ThreadEdge(BB, PredToThread, MostPopularDest);
976 /// ProcessJumpOnPHI - We have a conditional branch or switch on a PHI node in
977 /// the current block. See if there are any simplifications we can do based on
978 /// inputs to the phi node.
980 bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) {
981 BasicBlock *BB = PN->getParent();
983 // If any of the predecessor blocks end in an unconditional branch, we can
984 // *duplicate* the jump into that block in order to further encourage jump
985 // threading and to eliminate cases where we have branch on a phi of an icmp
986 // (branch on icmp is much better).
988 // We don't want to do this tranformation for switches, because we don't
989 // really want to duplicate a switch.
990 if (isa<SwitchInst>(BB->getTerminator()))
993 // Look for unconditional branch predecessors.
994 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
995 BasicBlock *PredBB = PN->getIncomingBlock(i);
996 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
997 if (PredBr->isUnconditional() &&
998 // Try to duplicate BB into PredBB.
999 DuplicateCondBranchOnPHIIntoPred(BB, PredBB))
1007 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1008 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1009 /// NewPred using the entries from OldPred (suitably mapped).
1010 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1011 BasicBlock *OldPred,
1012 BasicBlock *NewPred,
1013 DenseMap<Instruction*, Value*> &ValueMap) {
1014 for (BasicBlock::iterator PNI = PHIBB->begin();
1015 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1016 // Ok, we have a PHI node. Figure out what the incoming value was for the
1018 Value *IV = PN->getIncomingValueForBlock(OldPred);
1020 // Remap the value if necessary.
1021 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1022 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1023 if (I != ValueMap.end())
1027 PN->addIncoming(IV, NewPred);
1031 /// ThreadEdge - We have decided that it is safe and profitable to thread an
1032 /// edge from PredBB to SuccBB across BB. Transform the IR to reflect this
1034 bool JumpThreading::ThreadEdge(BasicBlock *BB, BasicBlock *PredBB,
1035 BasicBlock *SuccBB) {
1036 // If threading to the same block as we come from, we would infinite loop.
1038 DEBUG(errs() << " Not threading across BB '" << BB->getName()
1039 << "' - would thread to self!\n");
1043 // If threading this would thread across a loop header, don't thread the edge.
1044 // See the comments above FindLoopHeaders for justifications and caveats.
1045 if (LoopHeaders.count(BB)) {
1046 DEBUG(errs() << " Not threading from '" << PredBB->getName()
1047 << "' across loop header BB '" << BB->getName()
1048 << "' to dest BB '" << SuccBB->getName()
1049 << "' - it might create an irreducible loop!\n");
1053 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1054 if (JumpThreadCost > Threshold) {
1055 DEBUG(errs() << " Not threading BB '" << BB->getName()
1056 << "' - Cost is too high: " << JumpThreadCost << "\n");
1060 // And finally, do it!
1061 DEBUG(errs() << " Threading edge from '" << PredBB->getName() << "' to '"
1062 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1063 << ", across block:\n "
1066 // We are going to have to map operands from the original BB block to the new
1067 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1068 // account for entry from PredBB.
1069 DenseMap<Instruction*, Value*> ValueMapping;
1071 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1072 BB->getName()+".thread",
1073 BB->getParent(), BB);
1074 NewBB->moveAfter(PredBB);
1076 BasicBlock::iterator BI = BB->begin();
1077 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1078 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1080 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1081 // mapping and using it to remap operands in the cloned instructions.
1082 for (; !isa<TerminatorInst>(BI); ++BI) {
1083 Instruction *New = BI->clone();
1084 New->setName(BI->getName());
1085 NewBB->getInstList().push_back(New);
1086 ValueMapping[BI] = New;
1088 // Remap operands to patch up intra-block references.
1089 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1090 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1091 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1092 if (I != ValueMapping.end())
1093 New->setOperand(i, I->second);
1097 // We didn't copy the terminator from BB over to NewBB, because there is now
1098 // an unconditional jump to SuccBB. Insert the unconditional jump.
1099 BranchInst::Create(SuccBB, NewBB);
1101 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1102 // PHI nodes for NewBB now.
1103 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1105 // If there were values defined in BB that are used outside the block, then we
1106 // now have to update all uses of the value to use either the original value,
1107 // the cloned value, or some PHI derived value. This can require arbitrary
1108 // PHI insertion, of which we are prepared to do, clean these up now.
1109 SSAUpdater SSAUpdate;
1110 SmallVector<Use*, 16> UsesToRename;
1111 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1112 // Scan all uses of this instruction to see if it is used outside of its
1113 // block, and if so, record them in UsesToRename.
1114 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1116 Instruction *User = cast<Instruction>(*UI);
1117 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1118 if (UserPN->getIncomingBlock(UI) == BB)
1120 } else if (User->getParent() == BB)
1123 UsesToRename.push_back(&UI.getUse());
1126 // If there are no uses outside the block, we're done with this instruction.
1127 if (UsesToRename.empty())
1130 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n");
1132 // We found a use of I outside of BB. Rename all uses of I that are outside
1133 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1134 // with the two values we know.
1135 SSAUpdate.Initialize(I);
1136 SSAUpdate.AddAvailableValue(BB, I);
1137 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1139 while (!UsesToRename.empty())
1140 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1141 DEBUG(errs() << "\n");
1145 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1146 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1147 // us to simplify any PHI nodes in BB.
1148 TerminatorInst *PredTerm = PredBB->getTerminator();
1149 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1150 if (PredTerm->getSuccessor(i) == BB) {
1151 BB->removePredecessor(PredBB);
1152 PredTerm->setSuccessor(i, NewBB);
1155 // At this point, the IR is fully up to date and consistent. Do a quick scan
1156 // over the new instructions and zap any that are constants or dead. This
1157 // frequently happens because of phi translation.
1158 BI = NewBB->begin();
1159 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) {
1160 Instruction *Inst = BI++;
1161 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
1162 Inst->replaceAllUsesWith(C);
1163 Inst->eraseFromParent();
1167 RecursivelyDeleteTriviallyDeadInstructions(Inst);
1170 // Threaded an edge!
1175 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1176 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1177 /// If we can duplicate the contents of BB up into PredBB do so now, this
1178 /// improves the odds that the branch will be on an analyzable instruction like
1180 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1181 BasicBlock *PredBB) {
1182 // If BB is a loop header, then duplicating this block outside the loop would
1183 // cause us to transform this into an irreducible loop, don't do this.
1184 // See the comments above FindLoopHeaders for justifications and caveats.
1185 if (LoopHeaders.count(BB)) {
1186 DEBUG(errs() << " Not duplicating loop header '" << BB->getName()
1187 << "' into predecessor block '" << PredBB->getName()
1188 << "' - it might create an irreducible loop!\n");
1192 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1193 if (DuplicationCost > Threshold) {
1194 DEBUG(errs() << " Not duplicating BB '" << BB->getName()
1195 << "' - Cost is too high: " << DuplicationCost << "\n");
1199 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1201 DEBUG(errs() << " Duplicating block '" << BB->getName() << "' into end of '"
1202 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1203 << DuplicationCost << " block is:" << *BB << "\n");
1205 // We are going to have to map operands from the original BB block into the
1206 // PredBB block. Evaluate PHI nodes in BB.
1207 DenseMap<Instruction*, Value*> ValueMapping;
1209 BasicBlock::iterator BI = BB->begin();
1210 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1211 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1213 BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1215 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1216 // mapping and using it to remap operands in the cloned instructions.
1217 for (; BI != BB->end(); ++BI) {
1218 Instruction *New = BI->clone();
1219 New->setName(BI->getName());
1220 PredBB->getInstList().insert(OldPredBranch, New);
1221 ValueMapping[BI] = New;
1223 // Remap operands to patch up intra-block references.
1224 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1225 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1226 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1227 if (I != ValueMapping.end())
1228 New->setOperand(i, I->second);
1232 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1233 // add entries to the PHI nodes for branch from PredBB now.
1234 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1235 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1237 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1240 // If there were values defined in BB that are used outside the block, then we
1241 // now have to update all uses of the value to use either the original value,
1242 // the cloned value, or some PHI derived value. This can require arbitrary
1243 // PHI insertion, of which we are prepared to do, clean these up now.
1244 SSAUpdater SSAUpdate;
1245 SmallVector<Use*, 16> UsesToRename;
1246 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1247 // Scan all uses of this instruction to see if it is used outside of its
1248 // block, and if so, record them in UsesToRename.
1249 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1251 Instruction *User = cast<Instruction>(*UI);
1252 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1253 if (UserPN->getIncomingBlock(UI) == BB)
1255 } else if (User->getParent() == BB)
1258 UsesToRename.push_back(&UI.getUse());
1261 // If there are no uses outside the block, we're done with this instruction.
1262 if (UsesToRename.empty())
1265 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n");
1267 // We found a use of I outside of BB. Rename all uses of I that are outside
1268 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1269 // with the two values we know.
1270 SSAUpdate.Initialize(I);
1271 SSAUpdate.AddAvailableValue(BB, I);
1272 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1274 while (!UsesToRename.empty())
1275 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1276 DEBUG(errs() << "\n");
1279 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1281 BB->removePredecessor(PredBB);
1283 // Remove the unconditional branch at the end of the PredBB block.
1284 OldPredBranch->eraseFromParent();