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, TD);
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();
377 // If comparing a live-in value against a constant, see if we know the
378 // live-in value on any predecessors.
379 if (LVI && isa<Constant>(Cmp->getOperand(1)) &&
380 (!isa<Instruction>(Cmp->getOperand(0)) ||
381 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB)) {
382 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
384 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
385 // If the value is known by LazyValueInfo to be a constant in a
386 // predecessor, use that information to try to thread this block.
387 Constant *PredCst = LVI->getConstant(Cmp->getOperand(0), *PI);
391 // Constant fold the compare.
392 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), PredCst, RHSCst, TD);
393 if (isa<ConstantInt>(Res) || isa<UndefValue>(Res))
394 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(Res), *PI));
397 return !Result.empty();
405 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
406 /// in an undefined jump, decide which block is best to revector to.
408 /// Since we can pick an arbitrary destination, we pick the successor with the
409 /// fewest predecessors. This should reduce the in-degree of the others.
411 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
412 TerminatorInst *BBTerm = BB->getTerminator();
413 unsigned MinSucc = 0;
414 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
415 // Compute the successor with the minimum number of predecessors.
416 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
417 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
418 TestBB = BBTerm->getSuccessor(i);
419 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
420 if (NumPreds < MinNumPreds)
427 /// ProcessBlock - If there are any predecessors whose control can be threaded
428 /// through to a successor, transform them now.
429 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
430 // If this block has a single predecessor, and if that pred has a single
431 // successor, merge the blocks. This encourages recursive jump threading
432 // because now the condition in this block can be threaded through
433 // predecessors of our predecessor block.
434 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
435 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
437 // If SinglePred was a loop header, BB becomes one.
438 if (LoopHeaders.erase(SinglePred))
439 LoopHeaders.insert(BB);
441 // Remember if SinglePred was the entry block of the function. If so, we
442 // will need to move BB back to the entry position.
443 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
444 MergeBasicBlockIntoOnlyPred(BB);
446 if (isEntry && BB != &BB->getParent()->getEntryBlock())
447 BB->moveBefore(&BB->getParent()->getEntryBlock());
452 // Look to see if the terminator is a branch of switch, if not we can't thread
455 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
456 // Can't thread an unconditional jump.
457 if (BI->isUnconditional()) return false;
458 Condition = BI->getCondition();
459 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
460 Condition = SI->getCondition();
462 return false; // Must be an invoke.
464 // If the terminator of this block is branching on a constant, simplify the
465 // terminator to an unconditional branch. This can occur due to threading in
467 if (isa<ConstantInt>(Condition)) {
468 DEBUG(errs() << " In block '" << BB->getName()
469 << "' folding terminator: " << *BB->getTerminator() << '\n');
471 ConstantFoldTerminator(BB);
475 // If the terminator is branching on an undef, we can pick any of the
476 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
477 if (isa<UndefValue>(Condition)) {
478 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
480 // Fold the branch/switch.
481 TerminatorInst *BBTerm = BB->getTerminator();
482 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
483 if (i == BestSucc) continue;
484 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD);
487 DEBUG(errs() << " In block '" << BB->getName()
488 << "' folding undef terminator: " << *BBTerm << '\n');
489 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
490 BBTerm->eraseFromParent();
494 Instruction *CondInst = dyn_cast<Instruction>(Condition);
496 // If the condition is an instruction defined in another block, see if a
497 // predecessor has the same condition:
501 if (!Condition->hasOneUse() && // Multiple uses.
502 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition.
503 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
504 if (isa<BranchInst>(BB->getTerminator())) {
505 for (; PI != E; ++PI)
506 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
507 if (PBI->isConditional() && PBI->getCondition() == Condition &&
508 ProcessBranchOnDuplicateCond(*PI, BB))
511 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator");
512 for (; PI != E; ++PI)
513 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator()))
514 if (PSI->getCondition() == Condition &&
515 ProcessSwitchOnDuplicateCond(*PI, BB))
520 // All the rest of our checks depend on the condition being an instruction.
524 // See if this is a phi node in the current block.
525 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
526 if (PN->getParent() == BB)
527 return ProcessJumpOnPHI(PN);
529 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
530 if (!isa<PHINode>(CondCmp->getOperand(0)) ||
531 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB) {
532 // If we have a comparison, loop over the predecessors to see if there is
533 // a condition with a lexically identical value.
534 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
535 for (; PI != E; ++PI)
536 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
537 if (PBI->isConditional() && *PI != BB) {
538 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) {
539 if (CI->getOperand(0) == CondCmp->getOperand(0) &&
540 CI->getOperand(1) == CondCmp->getOperand(1) &&
541 CI->getPredicate() == CondCmp->getPredicate()) {
542 // TODO: Could handle things like (x != 4) --> (x == 17)
543 if (ProcessBranchOnDuplicateCond(*PI, BB))
551 // Check for some cases that are worth simplifying. Right now we want to look
552 // for loads that are used by a switch or by the condition for the branch. If
553 // we see one, check to see if it's partially redundant. If so, insert a PHI
554 // which can then be used to thread the values.
556 // This is particularly important because reg2mem inserts loads and stores all
557 // over the place, and this blocks jump threading if we don't zap them.
558 Value *SimplifyValue = CondInst;
559 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
560 if (isa<Constant>(CondCmp->getOperand(1)))
561 SimplifyValue = CondCmp->getOperand(0);
563 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
564 if (SimplifyPartiallyRedundantLoad(LI))
568 // Handle a variety of cases where we are branching on something derived from
569 // a PHI node in the current block. If we can prove that any predecessors
570 // compute a predictable value based on a PHI node, thread those predecessors.
572 if (ProcessThreadableEdges(CondInst, BB))
576 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
577 // "(X == 4)" thread through this block.
582 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
583 /// block that jump on exactly the same condition. This means that we almost
584 /// always know the direction of the edge in the DESTBB:
586 /// br COND, DESTBB, BBY
588 /// br COND, BBZ, BBW
590 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
591 /// in DESTBB, we have to thread over it.
592 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
594 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
596 // If both successors of PredBB go to DESTBB, we don't know anything. We can
597 // fold the branch to an unconditional one, which allows other recursive
600 if (PredBI->getSuccessor(1) != BB)
602 else if (PredBI->getSuccessor(0) != BB)
605 DEBUG(errs() << " In block '" << PredBB->getName()
606 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
608 ConstantFoldTerminator(PredBB);
612 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
614 // If the dest block has one predecessor, just fix the branch condition to a
615 // constant and fold it.
616 if (BB->getSinglePredecessor()) {
617 DEBUG(errs() << " In block '" << BB->getName()
618 << "' folding condition to '" << BranchDir << "': "
619 << *BB->getTerminator() << '\n');
621 Value *OldCond = DestBI->getCondition();
622 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
624 ConstantFoldTerminator(BB);
625 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
630 // Next, figure out which successor we are threading to.
631 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
633 SmallVector<BasicBlock*, 2> Preds;
634 Preds.push_back(PredBB);
636 // Ok, try to thread it!
637 return ThreadEdge(BB, Preds, SuccBB);
640 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
641 /// block that switch on exactly the same condition. This means that we almost
642 /// always know the direction of the edge in the DESTBB:
644 /// switch COND [... DESTBB, BBY ... ]
646 /// switch COND [... BBZ, BBW ]
648 /// Optimizing switches like this is very important, because simplifycfg builds
649 /// switches out of repeated 'if' conditions.
650 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
651 BasicBlock *DestBB) {
652 // Can't thread edge to self.
653 if (PredBB == DestBB)
656 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
657 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
659 // There are a variety of optimizations that we can potentially do on these
660 // blocks: we order them from most to least preferable.
662 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
663 // directly to their destination. This does not introduce *any* code size
664 // growth. Skip debug info first.
665 BasicBlock::iterator BBI = DestBB->begin();
666 while (isa<DbgInfoIntrinsic>(BBI))
669 // FIXME: Thread if it just contains a PHI.
670 if (isa<SwitchInst>(BBI)) {
671 bool MadeChange = false;
672 // Ignore the default edge for now.
673 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
674 ConstantInt *DestVal = DestSI->getCaseValue(i);
675 BasicBlock *DestSucc = DestSI->getSuccessor(i);
677 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
678 // PredSI has an explicit case for it. If so, forward. If it is covered
679 // by the default case, we can't update PredSI.
680 unsigned PredCase = PredSI->findCaseValue(DestVal);
681 if (PredCase == 0) continue;
683 // If PredSI doesn't go to DestBB on this value, then it won't reach the
684 // case on this condition.
685 if (PredSI->getSuccessor(PredCase) != DestBB &&
686 DestSI->getSuccessor(i) != DestBB)
689 // Otherwise, we're safe to make the change. Make sure that the edge from
690 // DestSI to DestSucc is not critical and has no PHI nodes.
691 DEBUG(errs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
692 DEBUG(errs() << "THROUGH: " << *DestSI);
694 // If the destination has PHI nodes, just split the edge for updating
696 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
697 SplitCriticalEdge(DestSI, i, this);
698 DestSucc = DestSI->getSuccessor(i);
700 FoldSingleEntryPHINodes(DestSucc);
701 PredSI->setSuccessor(PredCase, DestSucc);
713 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
714 /// load instruction, eliminate it by replacing it with a PHI node. This is an
715 /// important optimization that encourages jump threading, and needs to be run
716 /// interlaced with other jump threading tasks.
717 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
718 // Don't hack volatile loads.
719 if (LI->isVolatile()) return false;
721 // If the load is defined in a block with exactly one predecessor, it can't be
722 // partially redundant.
723 BasicBlock *LoadBB = LI->getParent();
724 if (LoadBB->getSinglePredecessor())
727 Value *LoadedPtr = LI->getOperand(0);
729 // If the loaded operand is defined in the LoadBB, it can't be available.
730 // FIXME: Could do PHI translation, that would be fun :)
731 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
732 if (PtrOp->getParent() == LoadBB)
735 // Scan a few instructions up from the load, to see if it is obviously live at
736 // the entry to its block.
737 BasicBlock::iterator BBIt = LI;
739 if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB,
741 // If the value if the load is locally available within the block, just use
742 // it. This frequently occurs for reg2mem'd allocas.
743 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
745 // If the returned value is the load itself, replace with an undef. This can
746 // only happen in dead loops.
747 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
748 LI->replaceAllUsesWith(AvailableVal);
749 LI->eraseFromParent();
753 // Otherwise, if we scanned the whole block and got to the top of the block,
754 // we know the block is locally transparent to the load. If not, something
755 // might clobber its value.
756 if (BBIt != LoadBB->begin())
760 SmallPtrSet<BasicBlock*, 8> PredsScanned;
761 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
762 AvailablePredsTy AvailablePreds;
763 BasicBlock *OneUnavailablePred = 0;
765 // If we got here, the loaded value is transparent through to the start of the
766 // block. Check to see if it is available in any of the predecessor blocks.
767 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
769 BasicBlock *PredBB = *PI;
771 // If we already scanned this predecessor, skip it.
772 if (!PredsScanned.insert(PredBB))
775 // Scan the predecessor to see if the value is available in the pred.
776 BBIt = PredBB->end();
777 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
778 if (!PredAvailable) {
779 OneUnavailablePred = PredBB;
783 // If so, this load is partially redundant. Remember this info so that we
784 // can create a PHI node.
785 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
788 // If the loaded value isn't available in any predecessor, it isn't partially
790 if (AvailablePreds.empty()) return false;
792 // Okay, the loaded value is available in at least one (and maybe all!)
793 // predecessors. If the value is unavailable in more than one unique
794 // predecessor, we want to insert a merge block for those common predecessors.
795 // This ensures that we only have to insert one reload, thus not increasing
797 BasicBlock *UnavailablePred = 0;
799 // If there is exactly one predecessor where the value is unavailable, the
800 // already computed 'OneUnavailablePred' block is it. If it ends in an
801 // unconditional branch, we know that it isn't a critical edge.
802 if (PredsScanned.size() == AvailablePreds.size()+1 &&
803 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
804 UnavailablePred = OneUnavailablePred;
805 } else if (PredsScanned.size() != AvailablePreds.size()) {
806 // Otherwise, we had multiple unavailable predecessors or we had a critical
807 // edge from the one.
808 SmallVector<BasicBlock*, 8> PredsToSplit;
809 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
811 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
812 AvailablePredSet.insert(AvailablePreds[i].first);
814 // Add all the unavailable predecessors to the PredsToSplit list.
815 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
817 if (!AvailablePredSet.count(*PI))
818 PredsToSplit.push_back(*PI);
820 // Split them out to their own block.
822 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
823 "thread-split", this);
826 // If the value isn't available in all predecessors, then there will be
827 // exactly one where it isn't available. Insert a load on that edge and add
828 // it to the AvailablePreds list.
829 if (UnavailablePred) {
830 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
831 "Can't handle critical edge here!");
832 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr",
833 UnavailablePred->getTerminator());
834 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
837 // Now we know that each predecessor of this block has a value in
838 // AvailablePreds, sort them for efficient access as we're walking the preds.
839 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
841 // Create a PHI node at the start of the block for the PRE'd load value.
842 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
845 // Insert new entries into the PHI for each predecessor. A single block may
846 // have multiple entries here.
847 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
849 AvailablePredsTy::iterator I =
850 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
851 std::make_pair(*PI, (Value*)0));
853 assert(I != AvailablePreds.end() && I->first == *PI &&
854 "Didn't find entry for predecessor!");
856 PN->addIncoming(I->second, I->first);
859 //cerr << "PRE: " << *LI << *PN << "\n";
861 LI->replaceAllUsesWith(PN);
862 LI->eraseFromParent();
867 /// FindMostPopularDest - The specified list contains multiple possible
868 /// threadable destinations. Pick the one that occurs the most frequently in
871 FindMostPopularDest(BasicBlock *BB,
872 const SmallVectorImpl<std::pair<BasicBlock*,
873 BasicBlock*> > &PredToDestList) {
874 assert(!PredToDestList.empty());
876 // Determine popularity. If there are multiple possible destinations, we
877 // explicitly choose to ignore 'undef' destinations. We prefer to thread
878 // blocks with known and real destinations to threading undef. We'll handle
879 // them later if interesting.
880 DenseMap<BasicBlock*, unsigned> DestPopularity;
881 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
882 if (PredToDestList[i].second)
883 DestPopularity[PredToDestList[i].second]++;
885 // Find the most popular dest.
886 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
887 BasicBlock *MostPopularDest = DPI->first;
888 unsigned Popularity = DPI->second;
889 SmallVector<BasicBlock*, 4> SamePopularity;
891 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
892 // If the popularity of this entry isn't higher than the popularity we've
893 // seen so far, ignore it.
894 if (DPI->second < Popularity)
896 else if (DPI->second == Popularity) {
897 // If it is the same as what we've seen so far, keep track of it.
898 SamePopularity.push_back(DPI->first);
900 // If it is more popular, remember it.
901 SamePopularity.clear();
902 MostPopularDest = DPI->first;
903 Popularity = DPI->second;
907 // Okay, now we know the most popular destination. If there is more than
908 // destination, we need to determine one. This is arbitrary, but we need
909 // to make a deterministic decision. Pick the first one that appears in the
911 if (!SamePopularity.empty()) {
912 SamePopularity.push_back(MostPopularDest);
913 TerminatorInst *TI = BB->getTerminator();
914 for (unsigned i = 0; ; ++i) {
915 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
917 if (std::find(SamePopularity.begin(), SamePopularity.end(),
918 TI->getSuccessor(i)) == SamePopularity.end())
921 MostPopularDest = TI->getSuccessor(i);
926 // Okay, we have finally picked the most popular destination.
927 return MostPopularDest;
930 bool JumpThreading::ProcessThreadableEdges(Instruction *CondInst,
932 // If threading this would thread across a loop header, don't even try to
934 if (LoopHeaders.count(BB))
937 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues;
938 if (!ComputeValueKnownInPredecessors(CondInst, BB, PredValues))
940 assert(!PredValues.empty() &&
941 "ComputeValueKnownInPredecessors returned true with no values");
943 DEBUG(errs() << "IN BB: " << *BB;
944 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
945 errs() << " BB '" << BB->getName() << "': FOUND condition = ";
946 if (PredValues[i].first)
947 errs() << *PredValues[i].first;
950 errs() << " for pred '" << PredValues[i].second->getName()
954 // Decide what we want to thread through. Convert our list of known values to
955 // a list of known destinations for each pred. This also discards duplicate
956 // predecessors and keeps track of the undefined inputs (which are represented
957 // as a null dest in the PredToDestList).
958 SmallPtrSet<BasicBlock*, 16> SeenPreds;
959 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
961 BasicBlock *OnlyDest = 0;
962 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
964 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
965 BasicBlock *Pred = PredValues[i].second;
966 if (!SeenPreds.insert(Pred))
967 continue; // Duplicate predecessor entry.
969 // If the predecessor ends with an indirect goto, we can't change its
971 if (isa<IndirectBrInst>(Pred->getTerminator()))
974 ConstantInt *Val = PredValues[i].first;
977 if (Val == 0) // Undef.
979 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
980 DestBB = BI->getSuccessor(Val->isZero());
982 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
983 DestBB = SI->getSuccessor(SI->findCaseValue(Val));
986 // If we have exactly one destination, remember it for efficiency below.
989 else if (OnlyDest != DestBB)
990 OnlyDest = MultipleDestSentinel;
992 PredToDestList.push_back(std::make_pair(Pred, DestBB));
995 // If all edges were unthreadable, we fail.
996 if (PredToDestList.empty())
999 // Determine which is the most common successor. If we have many inputs and
1000 // this block is a switch, we want to start by threading the batch that goes
1001 // to the most popular destination first. If we only know about one
1002 // threadable destination (the common case) we can avoid this.
1003 BasicBlock *MostPopularDest = OnlyDest;
1005 if (MostPopularDest == MultipleDestSentinel)
1006 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1008 // Now that we know what the most popular destination is, factor all
1009 // predecessors that will jump to it into a single predecessor.
1010 SmallVector<BasicBlock*, 16> PredsToFactor;
1011 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1012 if (PredToDestList[i].second == MostPopularDest) {
1013 BasicBlock *Pred = PredToDestList[i].first;
1015 // This predecessor may be a switch or something else that has multiple
1016 // edges to the block. Factor each of these edges by listing them
1017 // according to # occurrences in PredsToFactor.
1018 TerminatorInst *PredTI = Pred->getTerminator();
1019 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1020 if (PredTI->getSuccessor(i) == BB)
1021 PredsToFactor.push_back(Pred);
1024 // If the threadable edges are branching on an undefined value, we get to pick
1025 // the destination that these predecessors should get to.
1026 if (MostPopularDest == 0)
1027 MostPopularDest = BB->getTerminator()->
1028 getSuccessor(GetBestDestForJumpOnUndef(BB));
1030 // Ok, try to thread it!
1031 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1034 /// ProcessJumpOnPHI - We have a conditional branch or switch on a PHI node in
1035 /// the current block. See if there are any simplifications we can do based on
1036 /// inputs to the phi node.
1038 bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) {
1039 BasicBlock *BB = PN->getParent();
1041 // If any of the predecessor blocks end in an unconditional branch, we can
1042 // *duplicate* the jump into that block in order to further encourage jump
1043 // threading and to eliminate cases where we have branch on a phi of an icmp
1044 // (branch on icmp is much better).
1046 // We don't want to do this tranformation for switches, because we don't
1047 // really want to duplicate a switch.
1048 if (isa<SwitchInst>(BB->getTerminator()))
1051 // Look for unconditional branch predecessors.
1052 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1053 BasicBlock *PredBB = PN->getIncomingBlock(i);
1054 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1055 if (PredBr->isUnconditional() &&
1056 // Try to duplicate BB into PredBB.
1057 DuplicateCondBranchOnPHIIntoPred(BB, PredBB))
1065 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1066 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1067 /// NewPred using the entries from OldPred (suitably mapped).
1068 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1069 BasicBlock *OldPred,
1070 BasicBlock *NewPred,
1071 DenseMap<Instruction*, Value*> &ValueMap) {
1072 for (BasicBlock::iterator PNI = PHIBB->begin();
1073 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1074 // Ok, we have a PHI node. Figure out what the incoming value was for the
1076 Value *IV = PN->getIncomingValueForBlock(OldPred);
1078 // Remap the value if necessary.
1079 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1080 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1081 if (I != ValueMap.end())
1085 PN->addIncoming(IV, NewPred);
1089 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1090 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1091 /// across BB. Transform the IR to reflect this change.
1092 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1093 const SmallVectorImpl<BasicBlock*> &PredBBs,
1094 BasicBlock *SuccBB) {
1095 // If threading to the same block as we come from, we would infinite loop.
1097 DEBUG(errs() << " Not threading across BB '" << BB->getName()
1098 << "' - would thread to self!\n");
1102 // If threading this would thread across a loop header, don't thread the edge.
1103 // See the comments above FindLoopHeaders for justifications and caveats.
1104 if (LoopHeaders.count(BB)) {
1105 DEBUG(errs() << " Not threading across loop header BB '" << BB->getName()
1106 << "' to dest BB '" << SuccBB->getName()
1107 << "' - it might create an irreducible loop!\n");
1111 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1112 if (JumpThreadCost > Threshold) {
1113 DEBUG(errs() << " Not threading BB '" << BB->getName()
1114 << "' - Cost is too high: " << JumpThreadCost << "\n");
1118 // And finally, do it! Start by factoring the predecessors is needed.
1120 if (PredBBs.size() == 1)
1121 PredBB = PredBBs[0];
1123 DEBUG(errs() << " Factoring out " << PredBBs.size()
1124 << " common predecessors.\n");
1125 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1129 // And finally, do it!
1130 DEBUG(errs() << " Threading edge from '" << PredBB->getName() << "' to '"
1131 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1132 << ", across block:\n "
1135 // We are going to have to map operands from the original BB block to the new
1136 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1137 // account for entry from PredBB.
1138 DenseMap<Instruction*, Value*> ValueMapping;
1140 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1141 BB->getName()+".thread",
1142 BB->getParent(), BB);
1143 NewBB->moveAfter(PredBB);
1145 BasicBlock::iterator BI = BB->begin();
1146 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1147 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1149 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1150 // mapping and using it to remap operands in the cloned instructions.
1151 for (; !isa<TerminatorInst>(BI); ++BI) {
1152 Instruction *New = BI->clone();
1153 New->setName(BI->getName());
1154 NewBB->getInstList().push_back(New);
1155 ValueMapping[BI] = New;
1157 // Remap operands to patch up intra-block references.
1158 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1159 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1160 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1161 if (I != ValueMapping.end())
1162 New->setOperand(i, I->second);
1166 // We didn't copy the terminator from BB over to NewBB, because there is now
1167 // an unconditional jump to SuccBB. Insert the unconditional jump.
1168 BranchInst::Create(SuccBB, NewBB);
1170 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1171 // PHI nodes for NewBB now.
1172 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1174 // If there were values defined in BB that are used outside the block, then we
1175 // now have to update all uses of the value to use either the original value,
1176 // the cloned value, or some PHI derived value. This can require arbitrary
1177 // PHI insertion, of which we are prepared to do, clean these up now.
1178 SSAUpdater SSAUpdate;
1179 SmallVector<Use*, 16> UsesToRename;
1180 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1181 // Scan all uses of this instruction to see if it is used outside of its
1182 // block, and if so, record them in UsesToRename.
1183 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1185 Instruction *User = cast<Instruction>(*UI);
1186 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1187 if (UserPN->getIncomingBlock(UI) == BB)
1189 } else if (User->getParent() == BB)
1192 UsesToRename.push_back(&UI.getUse());
1195 // If there are no uses outside the block, we're done with this instruction.
1196 if (UsesToRename.empty())
1199 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n");
1201 // We found a use of I outside of BB. Rename all uses of I that are outside
1202 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1203 // with the two values we know.
1204 SSAUpdate.Initialize(I);
1205 SSAUpdate.AddAvailableValue(BB, I);
1206 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1208 while (!UsesToRename.empty())
1209 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1210 DEBUG(errs() << "\n");
1214 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1215 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1216 // us to simplify any PHI nodes in BB.
1217 TerminatorInst *PredTerm = PredBB->getTerminator();
1218 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1219 if (PredTerm->getSuccessor(i) == BB) {
1220 RemovePredecessorAndSimplify(BB, PredBB, TD);
1221 PredTerm->setSuccessor(i, NewBB);
1224 // At this point, the IR is fully up to date and consistent. Do a quick scan
1225 // over the new instructions and zap any that are constants or dead. This
1226 // frequently happens because of phi translation.
1227 BI = NewBB->begin();
1228 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) {
1229 Instruction *Inst = BI++;
1231 if (Value *V = SimplifyInstruction(Inst, TD)) {
1232 WeakVH BIHandle(BI);
1233 ReplaceAndSimplifyAllUses(Inst, V, TD);
1235 BI = NewBB->begin();
1239 RecursivelyDeleteTriviallyDeadInstructions(Inst);
1242 // Threaded an edge!
1247 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1248 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1249 /// If we can duplicate the contents of BB up into PredBB do so now, this
1250 /// improves the odds that the branch will be on an analyzable instruction like
1252 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1253 BasicBlock *PredBB) {
1254 // If BB is a loop header, then duplicating this block outside the loop would
1255 // cause us to transform this into an irreducible loop, don't do this.
1256 // See the comments above FindLoopHeaders for justifications and caveats.
1257 if (LoopHeaders.count(BB)) {
1258 DEBUG(errs() << " Not duplicating loop header '" << BB->getName()
1259 << "' into predecessor block '" << PredBB->getName()
1260 << "' - it might create an irreducible loop!\n");
1264 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1265 if (DuplicationCost > Threshold) {
1266 DEBUG(errs() << " Not duplicating BB '" << BB->getName()
1267 << "' - Cost is too high: " << DuplicationCost << "\n");
1271 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1273 DEBUG(errs() << " Duplicating block '" << BB->getName() << "' into end of '"
1274 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1275 << DuplicationCost << " block is:" << *BB << "\n");
1277 // We are going to have to map operands from the original BB block into the
1278 // PredBB block. Evaluate PHI nodes in BB.
1279 DenseMap<Instruction*, Value*> ValueMapping;
1281 BasicBlock::iterator BI = BB->begin();
1282 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1283 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1285 BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1287 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1288 // mapping and using it to remap operands in the cloned instructions.
1289 for (; BI != BB->end(); ++BI) {
1290 Instruction *New = BI->clone();
1291 New->setName(BI->getName());
1292 PredBB->getInstList().insert(OldPredBranch, New);
1293 ValueMapping[BI] = New;
1295 // Remap operands to patch up intra-block references.
1296 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1297 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1298 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1299 if (I != ValueMapping.end())
1300 New->setOperand(i, I->second);
1304 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1305 // add entries to the PHI nodes for branch from PredBB now.
1306 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1307 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1309 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1312 // If there were values defined in BB that are used outside the block, then we
1313 // now have to update all uses of the value to use either the original value,
1314 // the cloned value, or some PHI derived value. This can require arbitrary
1315 // PHI insertion, of which we are prepared to do, clean these up now.
1316 SSAUpdater SSAUpdate;
1317 SmallVector<Use*, 16> UsesToRename;
1318 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1319 // Scan all uses of this instruction to see if it is used outside of its
1320 // block, and if so, record them in UsesToRename.
1321 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1323 Instruction *User = cast<Instruction>(*UI);
1324 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1325 if (UserPN->getIncomingBlock(UI) == BB)
1327 } else if (User->getParent() == BB)
1330 UsesToRename.push_back(&UI.getUse());
1333 // If there are no uses outside the block, we're done with this instruction.
1334 if (UsesToRename.empty())
1337 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n");
1339 // We found a use of I outside of BB. Rename all uses of I that are outside
1340 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1341 // with the two values we know.
1342 SSAUpdate.Initialize(I);
1343 SSAUpdate.AddAvailableValue(BB, I);
1344 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1346 while (!UsesToRename.empty())
1347 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1348 DEBUG(errs() << "\n");
1351 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1353 RemovePredecessorAndSimplify(BB, PredBB, TD);
1355 // Remove the unconditional branch at the end of the PredBB block.
1356 OldPredBranch->eraseFromParent();