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/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, const SmallVectorImpl<BasicBlock*> &PredBBs,
77 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
80 typedef SmallVectorImpl<std::pair<ConstantInt*,
81 BasicBlock*> > PredValueInfo;
83 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
84 PredValueInfo &Result);
85 bool ProcessThreadableEdges(Instruction *CondInst, BasicBlock *BB);
88 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
89 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
91 bool ProcessJumpOnPHI(PHINode *PN);
93 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
97 char JumpThreading::ID = 0;
98 static RegisterPass<JumpThreading>
99 X("jump-threading", "Jump Threading");
101 // Public interface to the Jump Threading pass
102 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
104 /// runOnFunction - Top level algorithm.
106 bool JumpThreading::runOnFunction(Function &F) {
107 DEBUG(errs() << "Jump threading on function '" << F.getName() << "'\n");
108 TD = getAnalysisIfAvailable<TargetData>();
112 bool AnotherIteration = true, EverChanged = false;
113 while (AnotherIteration) {
114 AnotherIteration = false;
115 bool Changed = false;
116 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
118 // Thread all of the branches we can over this block.
119 while (ProcessBlock(BB))
124 // If the block is trivially dead, zap it. This eliminates the successor
125 // edges which simplifies the CFG.
126 if (pred_begin(BB) == pred_end(BB) &&
127 BB != &BB->getParent()->getEntryBlock()) {
128 DEBUG(errs() << " JT: Deleting dead block '" << BB->getName()
129 << "' with terminator: " << *BB->getTerminator() << '\n');
130 LoopHeaders.erase(BB);
133 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
134 // Can't thread an unconditional jump, but if the block is "almost
135 // empty", we can replace uses of it with uses of the successor and make
137 if (BI->isUnconditional() &&
138 BB != &BB->getParent()->getEntryBlock()) {
139 BasicBlock::iterator BBI = BB->getFirstNonPHI();
140 // Ignore dbg intrinsics.
141 while (isa<DbgInfoIntrinsic>(BBI))
143 // If the terminator is the only non-phi instruction, try to nuke it.
144 if (BBI->isTerminator()) {
145 bool Erased = LoopHeaders.erase(BB);
147 if (TryToSimplifyUncondBranchFromEmptyBlock(BB))
150 LoopHeaders.insert(BB);
155 AnotherIteration = Changed;
156 EverChanged |= Changed;
163 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
164 /// thread across it.
165 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
166 /// Ignore PHI nodes, these will be flattened when duplication happens.
167 BasicBlock::const_iterator I = BB->getFirstNonPHI();
169 // Sum up the cost of each instruction until we get to the terminator. Don't
170 // include the terminator because the copy won't include it.
172 for (; !isa<TerminatorInst>(I); ++I) {
173 // Debugger intrinsics don't incur code size.
174 if (isa<DbgInfoIntrinsic>(I)) continue;
176 // If this is a pointer->pointer bitcast, it is free.
177 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType()))
180 // All other instructions count for at least one unit.
183 // Calls are more expensive. If they are non-intrinsic calls, we model them
184 // as having cost of 4. If they are a non-vector intrinsic, we model them
185 // as having cost of 2 total, and if they are a vector intrinsic, we model
186 // them as having cost 1.
187 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
188 if (!isa<IntrinsicInst>(CI))
190 else if (!isa<VectorType>(CI->getType()))
195 // Threading through a switch statement is particularly profitable. If this
196 // block ends in a switch, decrease its cost to make it more likely to happen.
197 if (isa<SwitchInst>(I))
198 Size = Size > 6 ? Size-6 : 0;
204 //===----------------------------------------------------------------------===//
206 /// ReplaceAndSimplifyAllUses - Perform From->replaceAllUsesWith(To) and then
207 /// delete the From instruction. In addition to a basic RAUW, this does a
208 /// recursive simplification of the newly formed instructions. This catches
209 /// things where one simplification exposes other opportunities. This only
210 /// simplifies and deletes scalar operations, it does not change the CFG.
212 static void ReplaceAndSimplifyAllUses(Instruction *From, Value *To,
213 const TargetData *TD) {
214 assert(From != To && "ReplaceAndSimplifyAllUses(X,X) is not valid!");
216 // FromHandle - This keeps a weakvh on the from value so that we can know if
217 // it gets deleted out from under us in a recursive simplification.
218 WeakVH FromHandle(From);
220 while (!From->use_empty()) {
221 // Update the instruction to use the new value.
222 Use &U = From->use_begin().getUse();
223 Instruction *User = cast<Instruction>(U.getUser());
226 // See if we can simplify it.
227 if (Value *V = SimplifyInstruction(User, TD)) {
228 // Recursively simplify this.
229 ReplaceAndSimplifyAllUses(User, V, TD);
231 // If the recursive simplification ended up revisiting and deleting 'From'
237 From->eraseFromParent();
241 /// RemovePredecessorAndSimplify - Like BasicBlock::removePredecessor, this
242 /// method is called when we're about to delete Pred as a predecessor of BB. If
243 /// BB contains any PHI nodes, this drops the entries in the PHI nodes for Pred.
245 /// Unlike the removePredecessor method, this attempts to simplify uses of PHI
246 /// nodes that collapse into identity values. For example, if we have:
247 /// x = phi(1, 0, 0, 0)
250 /// .. and delete the predecessor corresponding to the '1', this will attempt to
251 /// recursively fold the and to 0.
252 static void RemovePredecessorAndSimplify(BasicBlock *BB, BasicBlock *Pred,
254 // This only adjusts blocks with PHI nodes.
255 if (!isa<PHINode>(BB->begin()))
258 // Remove the entries for Pred from the PHI nodes in BB, but do not simplify
259 // them down. This will leave us with single entry phi nodes and other phis
260 // that can be removed.
261 BB->removePredecessor(Pred, true);
263 WeakVH PhiIt = &BB->front();
264 while (PHINode *PN = dyn_cast<PHINode>(PhiIt)) {
265 PhiIt = &*++BasicBlock::iterator(cast<Instruction>(PhiIt));
267 Value *PNV = PN->hasConstantValue();
268 if (PNV == 0) continue;
270 // If we're able to simplify the phi to a single value, substitute the new
271 // value into all of its uses.
272 assert(PNV != PN && "hasConstantValue broken");
274 ReplaceAndSimplifyAllUses(PN, PNV, TD);
276 // If recursive simplification ended up deleting the next PHI node we would
277 // iterate to, then our iterator is invalid, restart scanning from the top
279 if (PhiIt == 0) PhiIt = &BB->front();
283 //===----------------------------------------------------------------------===//
286 /// FindLoopHeaders - We do not want jump threading to turn proper loop
287 /// structures into irreducible loops. Doing this breaks up the loop nesting
288 /// hierarchy and pessimizes later transformations. To prevent this from
289 /// happening, we first have to find the loop headers. Here we approximate this
290 /// by finding targets of backedges in the CFG.
292 /// Note that there definitely are cases when we want to allow threading of
293 /// edges across a loop header. For example, threading a jump from outside the
294 /// loop (the preheader) to an exit block of the loop is definitely profitable.
295 /// It is also almost always profitable to thread backedges from within the loop
296 /// to exit blocks, and is often profitable to thread backedges to other blocks
297 /// within the loop (forming a nested loop). This simple analysis is not rich
298 /// enough to track all of these properties and keep it up-to-date as the CFG
299 /// mutates, so we don't allow any of these transformations.
301 void JumpThreading::FindLoopHeaders(Function &F) {
302 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
303 FindFunctionBackedges(F, Edges);
305 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
306 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
309 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
310 /// if we can infer that the value is a known ConstantInt in any of our
311 /// predecessors. If so, return the known list of value and pred BB in the
312 /// result vector. If a value is known to be undef, it is returned as null.
314 /// The BB basic block is known to start with a PHI node.
316 /// This returns true if there were any known values.
319 /// TODO: Per PR2563, we could infer value range information about a predecessor
320 /// based on its terminator.
322 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){
323 PHINode *TheFirstPHI = cast<PHINode>(BB->begin());
325 // If V is a constantint, then it is known in all predecessors.
326 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) {
327 ConstantInt *CI = dyn_cast<ConstantInt>(V);
328 Result.resize(TheFirstPHI->getNumIncomingValues());
329 for (unsigned i = 0, e = Result.size(); i != e; ++i)
330 Result[i] = std::make_pair(CI, TheFirstPHI->getIncomingBlock(i));
334 // If V is a non-instruction value, or an instruction in a different block,
335 // then it can't be derived from a PHI.
336 Instruction *I = dyn_cast<Instruction>(V);
337 if (I == 0 || I->getParent() != BB)
340 /// If I is a PHI node, then we know the incoming values for any constants.
341 if (PHINode *PN = dyn_cast<PHINode>(I)) {
342 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
343 Value *InVal = PN->getIncomingValue(i);
344 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) {
345 ConstantInt *CI = dyn_cast<ConstantInt>(InVal);
346 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i)));
349 return !Result.empty();
352 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals;
354 // Handle some boolean conditions.
355 if (I->getType()->getPrimitiveSizeInBits() == 1) {
357 // X & false -> false
358 if (I->getOpcode() == Instruction::Or ||
359 I->getOpcode() == Instruction::And) {
360 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
361 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals);
363 if (LHSVals.empty() && RHSVals.empty())
366 ConstantInt *InterestingVal;
367 if (I->getOpcode() == Instruction::Or)
368 InterestingVal = ConstantInt::getTrue(I->getContext());
370 InterestingVal = ConstantInt::getFalse(I->getContext());
372 // Scan for the sentinel.
373 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
374 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0)
375 Result.push_back(LHSVals[i]);
376 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
377 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0)
378 Result.push_back(RHSVals[i]);
379 return !Result.empty();
382 // TODO: Should handle the NOT form of XOR.
386 // Handle compare with phi operand, where the PHI is defined in this block.
387 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
388 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
389 if (PN && PN->getParent() == BB) {
390 // We can do this simplification if any comparisons fold to true or false.
392 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
393 BasicBlock *PredBB = PN->getIncomingBlock(i);
394 Value *LHS = PN->getIncomingValue(i);
395 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
397 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS);
398 if (Res == 0) continue;
400 if (isa<UndefValue>(Res))
401 Result.push_back(std::make_pair((ConstantInt*)0, PredBB));
402 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res))
403 Result.push_back(std::make_pair(CI, PredBB));
406 return !Result.empty();
409 // TODO: We could also recurse to see if we can determine constants another
417 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
418 /// in an undefined jump, decide which block is best to revector to.
420 /// Since we can pick an arbitrary destination, we pick the successor with the
421 /// fewest predecessors. This should reduce the in-degree of the others.
423 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
424 TerminatorInst *BBTerm = BB->getTerminator();
425 unsigned MinSucc = 0;
426 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
427 // Compute the successor with the minimum number of predecessors.
428 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
429 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
430 TestBB = BBTerm->getSuccessor(i);
431 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
432 if (NumPreds < MinNumPreds)
439 /// ProcessBlock - If there are any predecessors whose control can be threaded
440 /// through to a successor, transform them now.
441 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
442 // If this block has a single predecessor, and if that pred has a single
443 // successor, merge the blocks. This encourages recursive jump threading
444 // because now the condition in this block can be threaded through
445 // predecessors of our predecessor block.
446 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
447 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
449 // If SinglePred was a loop header, BB becomes one.
450 if (LoopHeaders.erase(SinglePred))
451 LoopHeaders.insert(BB);
453 // Remember if SinglePred was the entry block of the function. If so, we
454 // will need to move BB back to the entry position.
455 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
456 MergeBasicBlockIntoOnlyPred(BB);
458 if (isEntry && BB != &BB->getParent()->getEntryBlock())
459 BB->moveBefore(&BB->getParent()->getEntryBlock());
464 // Look to see if the terminator is a branch of switch, if not we can't thread
467 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
468 // Can't thread an unconditional jump.
469 if (BI->isUnconditional()) return false;
470 Condition = BI->getCondition();
471 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
472 Condition = SI->getCondition();
474 return false; // Must be an invoke.
476 // If the terminator of this block is branching on a constant, simplify the
477 // terminator to an unconditional branch. This can occur due to threading in
479 if (isa<ConstantInt>(Condition)) {
480 DEBUG(errs() << " In block '" << BB->getName()
481 << "' folding terminator: " << *BB->getTerminator() << '\n');
483 ConstantFoldTerminator(BB);
487 // If the terminator is branching on an undef, we can pick any of the
488 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
489 if (isa<UndefValue>(Condition)) {
490 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
492 // Fold the branch/switch.
493 TerminatorInst *BBTerm = BB->getTerminator();
494 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
495 if (i == BestSucc) continue;
496 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD);
499 DEBUG(errs() << " In block '" << BB->getName()
500 << "' folding undef terminator: " << *BBTerm << '\n');
501 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
502 BBTerm->eraseFromParent();
506 Instruction *CondInst = dyn_cast<Instruction>(Condition);
508 // If the condition is an instruction defined in another block, see if a
509 // predecessor has the same condition:
513 if (!Condition->hasOneUse() && // Multiple uses.
514 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition.
515 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
516 if (isa<BranchInst>(BB->getTerminator())) {
517 for (; PI != E; ++PI)
518 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
519 if (PBI->isConditional() && PBI->getCondition() == Condition &&
520 ProcessBranchOnDuplicateCond(*PI, BB))
523 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator");
524 for (; PI != E; ++PI)
525 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator()))
526 if (PSI->getCondition() == Condition &&
527 ProcessSwitchOnDuplicateCond(*PI, BB))
532 // All the rest of our checks depend on the condition being an instruction.
536 // See if this is a phi node in the current block.
537 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
538 if (PN->getParent() == BB)
539 return ProcessJumpOnPHI(PN);
541 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
542 if (!isa<PHINode>(CondCmp->getOperand(0)) ||
543 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB) {
544 // If we have a comparison, loop over the predecessors to see if there is
545 // a condition with a lexically identical value.
546 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
547 for (; PI != E; ++PI)
548 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
549 if (PBI->isConditional() && *PI != BB) {
550 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) {
551 if (CI->getOperand(0) == CondCmp->getOperand(0) &&
552 CI->getOperand(1) == CondCmp->getOperand(1) &&
553 CI->getPredicate() == CondCmp->getPredicate()) {
554 // TODO: Could handle things like (x != 4) --> (x == 17)
555 if (ProcessBranchOnDuplicateCond(*PI, BB))
563 // Check for some cases that are worth simplifying. Right now we want to look
564 // for loads that are used by a switch or by the condition for the branch. If
565 // we see one, check to see if it's partially redundant. If so, insert a PHI
566 // which can then be used to thread the values.
568 // This is particularly important because reg2mem inserts loads and stores all
569 // over the place, and this blocks jump threading if we don't zap them.
570 Value *SimplifyValue = CondInst;
571 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
572 if (isa<Constant>(CondCmp->getOperand(1)))
573 SimplifyValue = CondCmp->getOperand(0);
575 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
576 if (SimplifyPartiallyRedundantLoad(LI))
580 // Handle a variety of cases where we are branching on something derived from
581 // a PHI node in the current block. If we can prove that any predecessors
582 // compute a predictable value based on a PHI node, thread those predecessors.
584 // We only bother doing this if the current block has a PHI node and if the
585 // conditional instruction lives in the current block. If either condition
586 // fails, this won't be a computable value anyway.
587 if (CondInst->getParent() == BB && isa<PHINode>(BB->front()))
588 if (ProcessThreadableEdges(CondInst, BB))
592 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
593 // "(X == 4)" thread through this block.
598 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
599 /// block that jump on exactly the same condition. This means that we almost
600 /// always know the direction of the edge in the DESTBB:
602 /// br COND, DESTBB, BBY
604 /// br COND, BBZ, BBW
606 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
607 /// in DESTBB, we have to thread over it.
608 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
610 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
612 // If both successors of PredBB go to DESTBB, we don't know anything. We can
613 // fold the branch to an unconditional one, which allows other recursive
616 if (PredBI->getSuccessor(1) != BB)
618 else if (PredBI->getSuccessor(0) != BB)
621 DEBUG(errs() << " In block '" << PredBB->getName()
622 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
624 ConstantFoldTerminator(PredBB);
628 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
630 // If the dest block has one predecessor, just fix the branch condition to a
631 // constant and fold it.
632 if (BB->getSinglePredecessor()) {
633 DEBUG(errs() << " In block '" << BB->getName()
634 << "' folding condition to '" << BranchDir << "': "
635 << *BB->getTerminator() << '\n');
637 Value *OldCond = DestBI->getCondition();
638 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
640 ConstantFoldTerminator(BB);
641 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
646 // Next, figure out which successor we are threading to.
647 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
649 SmallVector<BasicBlock*, 2> Preds;
650 Preds.push_back(PredBB);
652 // Ok, try to thread it!
653 return ThreadEdge(BB, Preds, SuccBB);
656 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
657 /// block that switch on exactly the same condition. This means that we almost
658 /// always know the direction of the edge in the DESTBB:
660 /// switch COND [... DESTBB, BBY ... ]
662 /// switch COND [... BBZ, BBW ]
664 /// Optimizing switches like this is very important, because simplifycfg builds
665 /// switches out of repeated 'if' conditions.
666 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
667 BasicBlock *DestBB) {
668 // Can't thread edge to self.
669 if (PredBB == DestBB)
672 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
673 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
675 // There are a variety of optimizations that we can potentially do on these
676 // blocks: we order them from most to least preferable.
678 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
679 // directly to their destination. This does not introduce *any* code size
680 // growth. Skip debug info first.
681 BasicBlock::iterator BBI = DestBB->begin();
682 while (isa<DbgInfoIntrinsic>(BBI))
685 // FIXME: Thread if it just contains a PHI.
686 if (isa<SwitchInst>(BBI)) {
687 bool MadeChange = false;
688 // Ignore the default edge for now.
689 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
690 ConstantInt *DestVal = DestSI->getCaseValue(i);
691 BasicBlock *DestSucc = DestSI->getSuccessor(i);
693 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
694 // PredSI has an explicit case for it. If so, forward. If it is covered
695 // by the default case, we can't update PredSI.
696 unsigned PredCase = PredSI->findCaseValue(DestVal);
697 if (PredCase == 0) continue;
699 // If PredSI doesn't go to DestBB on this value, then it won't reach the
700 // case on this condition.
701 if (PredSI->getSuccessor(PredCase) != DestBB &&
702 DestSI->getSuccessor(i) != DestBB)
705 // Otherwise, we're safe to make the change. Make sure that the edge from
706 // DestSI to DestSucc is not critical and has no PHI nodes.
707 DEBUG(errs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
708 DEBUG(errs() << "THROUGH: " << *DestSI);
710 // If the destination has PHI nodes, just split the edge for updating
712 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
713 SplitCriticalEdge(DestSI, i, this);
714 DestSucc = DestSI->getSuccessor(i);
716 FoldSingleEntryPHINodes(DestSucc);
717 PredSI->setSuccessor(PredCase, DestSucc);
729 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
730 /// load instruction, eliminate it by replacing it with a PHI node. This is an
731 /// important optimization that encourages jump threading, and needs to be run
732 /// interlaced with other jump threading tasks.
733 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
734 // Don't hack volatile loads.
735 if (LI->isVolatile()) return false;
737 // If the load is defined in a block with exactly one predecessor, it can't be
738 // partially redundant.
739 BasicBlock *LoadBB = LI->getParent();
740 if (LoadBB->getSinglePredecessor())
743 Value *LoadedPtr = LI->getOperand(0);
745 // If the loaded operand is defined in the LoadBB, it can't be available.
746 // FIXME: Could do PHI translation, that would be fun :)
747 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
748 if (PtrOp->getParent() == LoadBB)
751 // Scan a few instructions up from the load, to see if it is obviously live at
752 // the entry to its block.
753 BasicBlock::iterator BBIt = LI;
755 if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB,
757 // If the value if the load is locally available within the block, just use
758 // it. This frequently occurs for reg2mem'd allocas.
759 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
761 // If the returned value is the load itself, replace with an undef. This can
762 // only happen in dead loops.
763 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
764 LI->replaceAllUsesWith(AvailableVal);
765 LI->eraseFromParent();
769 // Otherwise, if we scanned the whole block and got to the top of the block,
770 // we know the block is locally transparent to the load. If not, something
771 // might clobber its value.
772 if (BBIt != LoadBB->begin())
776 SmallPtrSet<BasicBlock*, 8> PredsScanned;
777 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
778 AvailablePredsTy AvailablePreds;
779 BasicBlock *OneUnavailablePred = 0;
781 // If we got here, the loaded value is transparent through to the start of the
782 // block. Check to see if it is available in any of the predecessor blocks.
783 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
785 BasicBlock *PredBB = *PI;
787 // If we already scanned this predecessor, skip it.
788 if (!PredsScanned.insert(PredBB))
791 // Scan the predecessor to see if the value is available in the pred.
792 BBIt = PredBB->end();
793 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
794 if (!PredAvailable) {
795 OneUnavailablePred = PredBB;
799 // If so, this load is partially redundant. Remember this info so that we
800 // can create a PHI node.
801 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
804 // If the loaded value isn't available in any predecessor, it isn't partially
806 if (AvailablePreds.empty()) return false;
808 // Okay, the loaded value is available in at least one (and maybe all!)
809 // predecessors. If the value is unavailable in more than one unique
810 // predecessor, we want to insert a merge block for those common predecessors.
811 // This ensures that we only have to insert one reload, thus not increasing
813 BasicBlock *UnavailablePred = 0;
815 // If there is exactly one predecessor where the value is unavailable, the
816 // already computed 'OneUnavailablePred' block is it. If it ends in an
817 // unconditional branch, we know that it isn't a critical edge.
818 if (PredsScanned.size() == AvailablePreds.size()+1 &&
819 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
820 UnavailablePred = OneUnavailablePred;
821 } else if (PredsScanned.size() != AvailablePreds.size()) {
822 // Otherwise, we had multiple unavailable predecessors or we had a critical
823 // edge from the one.
824 SmallVector<BasicBlock*, 8> PredsToSplit;
825 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
827 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
828 AvailablePredSet.insert(AvailablePreds[i].first);
830 // Add all the unavailable predecessors to the PredsToSplit list.
831 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
833 if (!AvailablePredSet.count(*PI))
834 PredsToSplit.push_back(*PI);
836 // Split them out to their own block.
838 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
839 "thread-split", this);
842 // If the value isn't available in all predecessors, then there will be
843 // exactly one where it isn't available. Insert a load on that edge and add
844 // it to the AvailablePreds list.
845 if (UnavailablePred) {
846 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
847 "Can't handle critical edge here!");
848 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr",
849 UnavailablePred->getTerminator());
850 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
853 // Now we know that each predecessor of this block has a value in
854 // AvailablePreds, sort them for efficient access as we're walking the preds.
855 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
857 // Create a PHI node at the start of the block for the PRE'd load value.
858 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
861 // Insert new entries into the PHI for each predecessor. A single block may
862 // have multiple entries here.
863 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
865 AvailablePredsTy::iterator I =
866 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
867 std::make_pair(*PI, (Value*)0));
869 assert(I != AvailablePreds.end() && I->first == *PI &&
870 "Didn't find entry for predecessor!");
872 PN->addIncoming(I->second, I->first);
875 //cerr << "PRE: " << *LI << *PN << "\n";
877 LI->replaceAllUsesWith(PN);
878 LI->eraseFromParent();
883 /// FindMostPopularDest - The specified list contains multiple possible
884 /// threadable destinations. Pick the one that occurs the most frequently in
887 FindMostPopularDest(BasicBlock *BB,
888 const SmallVectorImpl<std::pair<BasicBlock*,
889 BasicBlock*> > &PredToDestList) {
890 assert(!PredToDestList.empty());
892 // Determine popularity. If there are multiple possible destinations, we
893 // explicitly choose to ignore 'undef' destinations. We prefer to thread
894 // blocks with known and real destinations to threading undef. We'll handle
895 // them later if interesting.
896 DenseMap<BasicBlock*, unsigned> DestPopularity;
897 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
898 if (PredToDestList[i].second)
899 DestPopularity[PredToDestList[i].second]++;
901 // Find the most popular dest.
902 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
903 BasicBlock *MostPopularDest = DPI->first;
904 unsigned Popularity = DPI->second;
905 SmallVector<BasicBlock*, 4> SamePopularity;
907 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
908 // If the popularity of this entry isn't higher than the popularity we've
909 // seen so far, ignore it.
910 if (DPI->second < Popularity)
912 else if (DPI->second == Popularity) {
913 // If it is the same as what we've seen so far, keep track of it.
914 SamePopularity.push_back(DPI->first);
916 // If it is more popular, remember it.
917 SamePopularity.clear();
918 MostPopularDest = DPI->first;
919 Popularity = DPI->second;
923 // Okay, now we know the most popular destination. If there is more than
924 // destination, we need to determine one. This is arbitrary, but we need
925 // to make a deterministic decision. Pick the first one that appears in the
927 if (!SamePopularity.empty()) {
928 SamePopularity.push_back(MostPopularDest);
929 TerminatorInst *TI = BB->getTerminator();
930 for (unsigned i = 0; ; ++i) {
931 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
933 if (std::find(SamePopularity.begin(), SamePopularity.end(),
934 TI->getSuccessor(i)) == SamePopularity.end())
937 MostPopularDest = TI->getSuccessor(i);
942 // Okay, we have finally picked the most popular destination.
943 return MostPopularDest;
946 bool JumpThreading::ProcessThreadableEdges(Instruction *CondInst,
948 // If threading this would thread across a loop header, don't even try to
950 if (LoopHeaders.count(BB))
953 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues;
954 if (!ComputeValueKnownInPredecessors(CondInst, BB, PredValues))
956 assert(!PredValues.empty() &&
957 "ComputeValueKnownInPredecessors returned true with no values");
959 DEBUG(errs() << "IN BB: " << *BB;
960 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
961 errs() << " BB '" << BB->getName() << "': FOUND condition = ";
962 if (PredValues[i].first)
963 errs() << *PredValues[i].first;
966 errs() << " for pred '" << PredValues[i].second->getName()
970 // Decide what we want to thread through. Convert our list of known values to
971 // a list of known destinations for each pred. This also discards duplicate
972 // predecessors and keeps track of the undefined inputs (which are represented
973 // as a null dest in the PredToDestList).
974 SmallPtrSet<BasicBlock*, 16> SeenPreds;
975 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
977 BasicBlock *OnlyDest = 0;
978 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
980 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
981 BasicBlock *Pred = PredValues[i].second;
982 if (!SeenPreds.insert(Pred))
983 continue; // Duplicate predecessor entry.
985 // If the predecessor ends with an indirect goto, we can't change its
987 if (isa<IndirectBrInst>(Pred->getTerminator()))
990 ConstantInt *Val = PredValues[i].first;
993 if (Val == 0) // Undef.
995 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
996 DestBB = BI->getSuccessor(Val->isZero());
998 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
999 DestBB = SI->getSuccessor(SI->findCaseValue(Val));
1002 // If we have exactly one destination, remember it for efficiency below.
1005 else if (OnlyDest != DestBB)
1006 OnlyDest = MultipleDestSentinel;
1008 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1011 // If all edges were unthreadable, we fail.
1012 if (PredToDestList.empty())
1015 // Determine which is the most common successor. If we have many inputs and
1016 // this block is a switch, we want to start by threading the batch that goes
1017 // to the most popular destination first. If we only know about one
1018 // threadable destination (the common case) we can avoid this.
1019 BasicBlock *MostPopularDest = OnlyDest;
1021 if (MostPopularDest == MultipleDestSentinel)
1022 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1024 // Now that we know what the most popular destination is, factor all
1025 // predecessors that will jump to it into a single predecessor.
1026 SmallVector<BasicBlock*, 16> PredsToFactor;
1027 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1028 if (PredToDestList[i].second == MostPopularDest) {
1029 BasicBlock *Pred = PredToDestList[i].first;
1031 // This predecessor may be a switch or something else that has multiple
1032 // edges to the block. Factor each of these edges by listing them
1033 // according to # occurrences in PredsToFactor.
1034 TerminatorInst *PredTI = Pred->getTerminator();
1035 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1036 if (PredTI->getSuccessor(i) == BB)
1037 PredsToFactor.push_back(Pred);
1040 // If the threadable edges are branching on an undefined value, we get to pick
1041 // the destination that these predecessors should get to.
1042 if (MostPopularDest == 0)
1043 MostPopularDest = BB->getTerminator()->
1044 getSuccessor(GetBestDestForJumpOnUndef(BB));
1046 // Ok, try to thread it!
1047 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1050 /// ProcessJumpOnPHI - We have a conditional branch or switch on a PHI node in
1051 /// the current block. See if there are any simplifications we can do based on
1052 /// inputs to the phi node.
1054 bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) {
1055 BasicBlock *BB = PN->getParent();
1057 // If any of the predecessor blocks end in an unconditional branch, we can
1058 // *duplicate* the jump into that block in order to further encourage jump
1059 // threading and to eliminate cases where we have branch on a phi of an icmp
1060 // (branch on icmp is much better).
1062 // We don't want to do this tranformation for switches, because we don't
1063 // really want to duplicate a switch.
1064 if (isa<SwitchInst>(BB->getTerminator()))
1067 // Look for unconditional branch predecessors.
1068 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1069 BasicBlock *PredBB = PN->getIncomingBlock(i);
1070 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1071 if (PredBr->isUnconditional() &&
1072 // Try to duplicate BB into PredBB.
1073 DuplicateCondBranchOnPHIIntoPred(BB, PredBB))
1081 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1082 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1083 /// NewPred using the entries from OldPred (suitably mapped).
1084 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1085 BasicBlock *OldPred,
1086 BasicBlock *NewPred,
1087 DenseMap<Instruction*, Value*> &ValueMap) {
1088 for (BasicBlock::iterator PNI = PHIBB->begin();
1089 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1090 // Ok, we have a PHI node. Figure out what the incoming value was for the
1092 Value *IV = PN->getIncomingValueForBlock(OldPred);
1094 // Remap the value if necessary.
1095 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1096 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1097 if (I != ValueMap.end())
1101 PN->addIncoming(IV, NewPred);
1105 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1106 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1107 /// across BB. Transform the IR to reflect this change.
1108 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1109 const SmallVectorImpl<BasicBlock*> &PredBBs,
1110 BasicBlock *SuccBB) {
1111 // If threading to the same block as we come from, we would infinite loop.
1113 DEBUG(errs() << " Not threading across BB '" << BB->getName()
1114 << "' - would thread to self!\n");
1118 // If threading this would thread across a loop header, don't thread the edge.
1119 // See the comments above FindLoopHeaders for justifications and caveats.
1120 if (LoopHeaders.count(BB)) {
1121 DEBUG(errs() << " Not threading across loop header BB '" << BB->getName()
1122 << "' to dest BB '" << SuccBB->getName()
1123 << "' - it might create an irreducible loop!\n");
1127 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1128 if (JumpThreadCost > Threshold) {
1129 DEBUG(errs() << " Not threading BB '" << BB->getName()
1130 << "' - Cost is too high: " << JumpThreadCost << "\n");
1134 // And finally, do it! Start by factoring the predecessors is needed.
1136 if (PredBBs.size() == 1)
1137 PredBB = PredBBs[0];
1139 DEBUG(errs() << " Factoring out " << PredBBs.size()
1140 << " common predecessors.\n");
1141 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1145 // And finally, do it!
1146 DEBUG(errs() << " Threading edge from '" << PredBB->getName() << "' to '"
1147 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1148 << ", across block:\n "
1151 // We are going to have to map operands from the original BB block to the new
1152 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1153 // account for entry from PredBB.
1154 DenseMap<Instruction*, Value*> ValueMapping;
1156 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1157 BB->getName()+".thread",
1158 BB->getParent(), BB);
1159 NewBB->moveAfter(PredBB);
1161 BasicBlock::iterator BI = BB->begin();
1162 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1163 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1165 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1166 // mapping and using it to remap operands in the cloned instructions.
1167 for (; !isa<TerminatorInst>(BI); ++BI) {
1168 Instruction *New = BI->clone();
1169 New->setName(BI->getName());
1170 NewBB->getInstList().push_back(New);
1171 ValueMapping[BI] = New;
1173 // Remap operands to patch up intra-block references.
1174 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1175 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1176 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1177 if (I != ValueMapping.end())
1178 New->setOperand(i, I->second);
1182 // We didn't copy the terminator from BB over to NewBB, because there is now
1183 // an unconditional jump to SuccBB. Insert the unconditional jump.
1184 BranchInst::Create(SuccBB, NewBB);
1186 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1187 // PHI nodes for NewBB now.
1188 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1190 // If there were values defined in BB that are used outside the block, then we
1191 // now have to update all uses of the value to use either the original value,
1192 // the cloned value, or some PHI derived value. This can require arbitrary
1193 // PHI insertion, of which we are prepared to do, clean these up now.
1194 SSAUpdater SSAUpdate;
1195 SmallVector<Use*, 16> UsesToRename;
1196 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1197 // Scan all uses of this instruction to see if it is used outside of its
1198 // block, and if so, record them in UsesToRename.
1199 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1201 Instruction *User = cast<Instruction>(*UI);
1202 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1203 if (UserPN->getIncomingBlock(UI) == BB)
1205 } else if (User->getParent() == BB)
1208 UsesToRename.push_back(&UI.getUse());
1211 // If there are no uses outside the block, we're done with this instruction.
1212 if (UsesToRename.empty())
1215 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n");
1217 // We found a use of I outside of BB. Rename all uses of I that are outside
1218 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1219 // with the two values we know.
1220 SSAUpdate.Initialize(I);
1221 SSAUpdate.AddAvailableValue(BB, I);
1222 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1224 while (!UsesToRename.empty())
1225 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1226 DEBUG(errs() << "\n");
1230 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1231 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1232 // us to simplify any PHI nodes in BB.
1233 TerminatorInst *PredTerm = PredBB->getTerminator();
1234 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1235 if (PredTerm->getSuccessor(i) == BB) {
1236 RemovePredecessorAndSimplify(BB, PredBB, TD);
1237 PredTerm->setSuccessor(i, NewBB);
1240 // At this point, the IR is fully up to date and consistent. Do a quick scan
1241 // over the new instructions and zap any that are constants or dead. This
1242 // frequently happens because of phi translation.
1243 BI = NewBB->begin();
1244 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) {
1245 Instruction *Inst = BI++;
1247 if (Value *V = SimplifyInstruction(Inst, TD)) {
1248 WeakVH BIHandle(BI);
1249 ReplaceAndSimplifyAllUses(Inst, V, TD);
1251 BI = NewBB->begin();
1255 RecursivelyDeleteTriviallyDeadInstructions(Inst);
1258 // Threaded an edge!
1263 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1264 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1265 /// If we can duplicate the contents of BB up into PredBB do so now, this
1266 /// improves the odds that the branch will be on an analyzable instruction like
1268 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1269 BasicBlock *PredBB) {
1270 // If BB is a loop header, then duplicating this block outside the loop would
1271 // cause us to transform this into an irreducible loop, don't do this.
1272 // See the comments above FindLoopHeaders for justifications and caveats.
1273 if (LoopHeaders.count(BB)) {
1274 DEBUG(errs() << " Not duplicating loop header '" << BB->getName()
1275 << "' into predecessor block '" << PredBB->getName()
1276 << "' - it might create an irreducible loop!\n");
1280 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1281 if (DuplicationCost > Threshold) {
1282 DEBUG(errs() << " Not duplicating BB '" << BB->getName()
1283 << "' - Cost is too high: " << DuplicationCost << "\n");
1287 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1289 DEBUG(errs() << " Duplicating block '" << BB->getName() << "' into end of '"
1290 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1291 << DuplicationCost << " block is:" << *BB << "\n");
1293 // We are going to have to map operands from the original BB block into the
1294 // PredBB block. Evaluate PHI nodes in BB.
1295 DenseMap<Instruction*, Value*> ValueMapping;
1297 BasicBlock::iterator BI = BB->begin();
1298 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1299 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1301 BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1303 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1304 // mapping and using it to remap operands in the cloned instructions.
1305 for (; BI != BB->end(); ++BI) {
1306 Instruction *New = BI->clone();
1307 New->setName(BI->getName());
1308 PredBB->getInstList().insert(OldPredBranch, New);
1309 ValueMapping[BI] = New;
1311 // Remap operands to patch up intra-block references.
1312 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1313 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1314 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1315 if (I != ValueMapping.end())
1316 New->setOperand(i, I->second);
1320 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1321 // add entries to the PHI nodes for branch from PredBB now.
1322 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1323 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1325 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1328 // If there were values defined in BB that are used outside the block, then we
1329 // now have to update all uses of the value to use either the original value,
1330 // the cloned value, or some PHI derived value. This can require arbitrary
1331 // PHI insertion, of which we are prepared to do, clean these up now.
1332 SSAUpdater SSAUpdate;
1333 SmallVector<Use*, 16> UsesToRename;
1334 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1335 // Scan all uses of this instruction to see if it is used outside of its
1336 // block, and if so, record them in UsesToRename.
1337 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1339 Instruction *User = cast<Instruction>(*UI);
1340 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1341 if (UserPN->getIncomingBlock(UI) == BB)
1343 } else if (User->getParent() == BB)
1346 UsesToRename.push_back(&UI.getUse());
1349 // If there are no uses outside the block, we're done with this instruction.
1350 if (UsesToRename.empty())
1353 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n");
1355 // We found a use of I outside of BB. Rename all uses of I that are outside
1356 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1357 // with the two values we know.
1358 SSAUpdate.Initialize(I);
1359 SSAUpdate.AddAvailableValue(BB, I);
1360 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1362 while (!UsesToRename.empty())
1363 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1364 DEBUG(errs() << "\n");
1367 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1369 RemovePredecessorAndSimplify(BB, PredBB, TD);
1371 // Remove the unconditional branch at the end of the PredBB block.
1372 OldPredBranch->eraseFromParent();