1 //===- JumpThreading.cpp - Thread control through conditional blocks ------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the Jump Threading pass.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "jump-threading"
15 #include "llvm/Transforms/Scalar.h"
16 #include "llvm/IntrinsicInst.h"
17 #include "llvm/LLVMContext.h"
18 #include "llvm/Pass.h"
19 #include "llvm/Analysis/ConstantFolding.h"
20 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
21 #include "llvm/Transforms/Utils/Local.h"
22 #include "llvm/Transforms/Utils/SSAUpdater.h"
23 #include "llvm/Target/TargetData.h"
24 #include "llvm/ADT/DenseMap.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/STLExtras.h"
27 #include "llvm/ADT/SmallPtrSet.h"
28 #include "llvm/ADT/SmallSet.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/raw_ostream.h"
34 STATISTIC(NumThreads, "Number of jumps threaded");
35 STATISTIC(NumFolds, "Number of terminators folded");
36 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
38 static cl::opt<unsigned>
39 Threshold("jump-threading-threshold",
40 cl::desc("Max block size to duplicate for jump threading"),
41 cl::init(6), cl::Hidden);
44 /// This pass performs 'jump threading', which looks at blocks that have
45 /// multiple predecessors and multiple successors. If one or more of the
46 /// predecessors of the block can be proven to always jump to one of the
47 /// successors, we forward the edge from the predecessor to the successor by
48 /// duplicating the contents of this block.
50 /// An example of when this can occur is code like this:
57 /// In this case, the unconditional branch at the end of the first if can be
58 /// revectored to the false side of the second if.
60 class JumpThreading : public FunctionPass {
63 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
65 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
68 static char ID; // Pass identification
69 JumpThreading() : FunctionPass(&ID) {}
71 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
74 bool runOnFunction(Function &F);
75 void FindLoopHeaders(Function &F);
77 bool ProcessBlock(BasicBlock *BB);
78 bool ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, BasicBlock *SuccBB);
79 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
82 BasicBlock *FactorCommonPHIPreds(PHINode *PN, Value *Val);
83 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
84 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
86 bool ProcessJumpOnPHI(PHINode *PN);
87 bool ProcessBranchOnLogical(Value *V, BasicBlock *BB, bool isAnd);
88 bool ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB);
90 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
94 char JumpThreading::ID = 0;
95 static RegisterPass<JumpThreading>
96 X("jump-threading", "Jump Threading");
98 // Public interface to the Jump Threading pass
99 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
101 /// runOnFunction - Top level algorithm.
103 bool JumpThreading::runOnFunction(Function &F) {
104 DEBUG(errs() << "Jump threading on function '" << F.getName() << "'\n");
105 TD = getAnalysisIfAvailable<TargetData>();
109 bool AnotherIteration = true, EverChanged = false;
110 while (AnotherIteration) {
111 AnotherIteration = false;
112 bool Changed = false;
113 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
115 while (ProcessBlock(BB))
120 // If the block is trivially dead, zap it. This eliminates the successor
121 // edges which simplifies the CFG.
122 if (pred_begin(BB) == pred_end(BB) &&
123 BB != &BB->getParent()->getEntryBlock()) {
124 DEBUG(errs() << " JT: Deleting dead block '" << BB->getName()
125 << "' with terminator: " << *BB->getTerminator() << '\n');
126 LoopHeaders.erase(BB);
131 AnotherIteration = Changed;
132 EverChanged |= Changed;
139 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
140 /// thread across it.
141 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
142 /// Ignore PHI nodes, these will be flattened when duplication happens.
143 BasicBlock::const_iterator I = BB->getFirstNonPHI();
145 // Sum up the cost of each instruction until we get to the terminator. Don't
146 // include the terminator because the copy won't include it.
148 for (; !isa<TerminatorInst>(I); ++I) {
149 // Debugger intrinsics don't incur code size.
150 if (isa<DbgInfoIntrinsic>(I)) continue;
152 // If this is a pointer->pointer bitcast, it is free.
153 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType()))
156 // All other instructions count for at least one unit.
159 // Calls are more expensive. If they are non-intrinsic calls, we model them
160 // as having cost of 4. If they are a non-vector intrinsic, we model them
161 // as having cost of 2 total, and if they are a vector intrinsic, we model
162 // them as having cost 1.
163 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
164 if (!isa<IntrinsicInst>(CI))
166 else if (!isa<VectorType>(CI->getType()))
171 // Threading through a switch statement is particularly profitable. If this
172 // block ends in a switch, decrease its cost to make it more likely to happen.
173 if (isa<SwitchInst>(I))
174 Size = Size > 6 ? Size-6 : 0;
181 /// FindLoopHeaders - We do not want jump threading to turn proper loop
182 /// structures into irreducible loops. Doing this breaks up the loop nesting
183 /// hierarchy and pessimizes later transformations. To prevent this from
184 /// happening, we first have to find the loop headers. Here we approximate this
185 /// by finding targets of backedges in the CFG.
187 /// Note that there definitely are cases when we want to allow threading of
188 /// edges across a loop header. For example, threading a jump from outside the
189 /// loop (the preheader) to an exit block of the loop is definitely profitable.
190 /// It is also almost always profitable to thread backedges from within the loop
191 /// to exit blocks, and is often profitable to thread backedges to other blocks
192 /// within the loop (forming a nested loop). This simple analysis is not rich
193 /// enough to track all of these properties and keep it up-to-date as the CFG
194 /// mutates, so we don't allow any of these transformations.
196 void JumpThreading::FindLoopHeaders(Function &F) {
197 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
198 FindFunctionBackedges(F, Edges);
200 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
201 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
205 /// FactorCommonPHIPreds - If there are multiple preds with the same incoming
206 /// value for the PHI, factor them together so we get one block to thread for
208 /// This is important for things like "phi i1 [true, true, false, true, x]"
209 /// where we only need to clone the block for the true blocks once.
211 BasicBlock *JumpThreading::FactorCommonPHIPreds(PHINode *PN, Value *Val) {
212 SmallVector<BasicBlock*, 16> CommonPreds;
213 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
214 if (PN->getIncomingValue(i) == Val)
215 CommonPreds.push_back(PN->getIncomingBlock(i));
217 if (CommonPreds.size() == 1)
218 return CommonPreds[0];
220 DEBUG(errs() << " Factoring out " << CommonPreds.size()
221 << " common predecessors.\n");
222 return SplitBlockPredecessors(PN->getParent(),
223 &CommonPreds[0], CommonPreds.size(),
228 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
229 /// in an undefined jump, decide which block is best to revector to.
231 /// Since we can pick an arbitrary destination, we pick the successor with the
232 /// fewest predecessors. This should reduce the in-degree of the others.
234 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
235 TerminatorInst *BBTerm = BB->getTerminator();
236 unsigned MinSucc = 0;
237 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
238 // Compute the successor with the minimum number of predecessors.
239 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
240 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
241 TestBB = BBTerm->getSuccessor(i);
242 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
243 if (NumPreds < MinNumPreds)
250 /// ProcessBlock - If there are any predecessors whose control can be threaded
251 /// through to a successor, transform them now.
252 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
253 // If this block has a single predecessor, and if that pred has a single
254 // successor, merge the blocks. This encourages recursive jump threading
255 // because now the condition in this block can be threaded through
256 // predecessors of our predecessor block.
257 if (BasicBlock *SinglePred = BB->getSinglePredecessor())
258 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
260 // If SinglePred was a loop header, BB becomes one.
261 if (LoopHeaders.erase(SinglePred))
262 LoopHeaders.insert(BB);
264 // Remember if SinglePred was the entry block of the function. If so, we
265 // will need to move BB back to the entry position.
266 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
267 MergeBasicBlockIntoOnlyPred(BB);
269 if (isEntry && BB != &BB->getParent()->getEntryBlock())
270 BB->moveBefore(&BB->getParent()->getEntryBlock());
274 // See if this block ends with a branch or switch. If so, see if the
275 // condition is a phi node. If so, and if an entry of the phi node is a
276 // constant, we can thread the block.
278 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
279 // Can't thread an unconditional jump.
280 if (BI->isUnconditional()) return false;
281 Condition = BI->getCondition();
282 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
283 Condition = SI->getCondition();
285 return false; // Must be an invoke.
287 // If the terminator of this block is branching on a constant, simplify the
288 // terminator to an unconditional branch. This can occur due to threading in
290 if (isa<ConstantInt>(Condition)) {
291 DEBUG(errs() << " In block '" << BB->getName()
292 << "' folding terminator: " << *BB->getTerminator() << '\n');
294 ConstantFoldTerminator(BB);
298 // If the terminator is branching on an undef, we can pick any of the
299 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
300 if (isa<UndefValue>(Condition)) {
301 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
303 // Fold the branch/switch.
304 TerminatorInst *BBTerm = BB->getTerminator();
305 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
306 if (i == BestSucc) continue;
307 BBTerm->getSuccessor(i)->removePredecessor(BB);
310 DEBUG(errs() << " In block '" << BB->getName()
311 << "' folding undef terminator: " << *BBTerm << '\n');
312 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
313 BBTerm->eraseFromParent();
317 Instruction *CondInst = dyn_cast<Instruction>(Condition);
319 // If the condition is an instruction defined in another block, see if a
320 // predecessor has the same condition:
324 if (!Condition->hasOneUse() && // Multiple uses.
325 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition.
326 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
327 if (isa<BranchInst>(BB->getTerminator())) {
328 for (; PI != E; ++PI)
329 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
330 if (PBI->isConditional() && PBI->getCondition() == Condition &&
331 ProcessBranchOnDuplicateCond(*PI, BB))
334 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator");
335 for (; PI != E; ++PI)
336 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator()))
337 if (PSI->getCondition() == Condition &&
338 ProcessSwitchOnDuplicateCond(*PI, BB))
343 // All the rest of our checks depend on the condition being an instruction.
347 // See if this is a phi node in the current block.
348 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
349 if (PN->getParent() == BB)
350 return ProcessJumpOnPHI(PN);
352 // If this is a conditional branch whose condition is and/or of a phi, try to
354 if ((CondInst->getOpcode() == Instruction::And ||
355 CondInst->getOpcode() == Instruction::Or) &&
356 isa<BranchInst>(BB->getTerminator()) &&
357 ProcessBranchOnLogical(CondInst, BB,
358 CondInst->getOpcode() == Instruction::And))
361 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
362 if (isa<PHINode>(CondCmp->getOperand(0))) {
363 // If we have "br (phi != 42)" and the phi node has any constant values
364 // as operands, we can thread through this block.
366 // If we have "br (cmp phi, x)" and the phi node contains x such that the
367 // comparison uniquely identifies the branch target, we can thread
368 // through this block.
370 if (ProcessBranchOnCompare(CondCmp, BB))
374 // If we have a comparison, loop over the predecessors to see if there is
375 // a condition with the same value.
376 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
377 for (; PI != E; ++PI)
378 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
379 if (PBI->isConditional() && *PI != BB) {
380 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) {
381 if (CI->getOperand(0) == CondCmp->getOperand(0) &&
382 CI->getOperand(1) == CondCmp->getOperand(1) &&
383 CI->getPredicate() == CondCmp->getPredicate()) {
384 // TODO: Could handle things like (x != 4) --> (x == 17)
385 if (ProcessBranchOnDuplicateCond(*PI, BB))
392 // Check for some cases that are worth simplifying. Right now we want to look
393 // for loads that are used by a switch or by the condition for the branch. If
394 // we see one, check to see if it's partially redundant. If so, insert a PHI
395 // which can then be used to thread the values.
397 // This is particularly important because reg2mem inserts loads and stores all
398 // over the place, and this blocks jump threading if we don't zap them.
399 Value *SimplifyValue = CondInst;
400 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
401 if (isa<Constant>(CondCmp->getOperand(1)))
402 SimplifyValue = CondCmp->getOperand(0);
404 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
405 if (SimplifyPartiallyRedundantLoad(LI))
408 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
409 // "(X == 4)" thread through this block.
414 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
415 /// block that jump on exactly the same condition. This means that we almost
416 /// always know the direction of the edge in the DESTBB:
418 /// br COND, DESTBB, BBY
420 /// br COND, BBZ, BBW
422 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
423 /// in DESTBB, we have to thread over it.
424 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
426 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
428 // If both successors of PredBB go to DESTBB, we don't know anything. We can
429 // fold the branch to an unconditional one, which allows other recursive
432 if (PredBI->getSuccessor(1) != BB)
434 else if (PredBI->getSuccessor(0) != BB)
437 DEBUG(errs() << " In block '" << PredBB->getName()
438 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
440 ConstantFoldTerminator(PredBB);
444 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
446 // If the dest block has one predecessor, just fix the branch condition to a
447 // constant and fold it.
448 if (BB->getSinglePredecessor()) {
449 DEBUG(errs() << " In block '" << BB->getName()
450 << "' folding condition to '" << BranchDir << "': "
451 << *BB->getTerminator() << '\n');
453 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
455 ConstantFoldTerminator(BB);
460 // Next, figure out which successor we are threading to.
461 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
463 // Ok, try to thread it!
464 return ThreadEdge(BB, PredBB, SuccBB);
467 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
468 /// block that switch on exactly the same condition. This means that we almost
469 /// always know the direction of the edge in the DESTBB:
471 /// switch COND [... DESTBB, BBY ... ]
473 /// switch COND [... BBZ, BBW ]
475 /// Optimizing switches like this is very important, because simplifycfg builds
476 /// switches out of repeated 'if' conditions.
477 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
478 BasicBlock *DestBB) {
479 // Can't thread edge to self.
480 if (PredBB == DestBB)
483 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
484 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
486 // There are a variety of optimizations that we can potentially do on these
487 // blocks: we order them from most to least preferable.
489 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
490 // directly to their destination. This does not introduce *any* code size
491 // growth. Skip debug info first.
492 BasicBlock::iterator BBI = DestBB->begin();
493 while (isa<DbgInfoIntrinsic>(BBI))
496 // FIXME: Thread if it just contains a PHI.
497 if (isa<SwitchInst>(BBI)) {
498 bool MadeChange = false;
499 // Ignore the default edge for now.
500 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
501 ConstantInt *DestVal = DestSI->getCaseValue(i);
502 BasicBlock *DestSucc = DestSI->getSuccessor(i);
504 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
505 // PredSI has an explicit case for it. If so, forward. If it is covered
506 // by the default case, we can't update PredSI.
507 unsigned PredCase = PredSI->findCaseValue(DestVal);
508 if (PredCase == 0) continue;
510 // If PredSI doesn't go to DestBB on this value, then it won't reach the
511 // case on this condition.
512 if (PredSI->getSuccessor(PredCase) != DestBB &&
513 DestSI->getSuccessor(i) != DestBB)
516 // Otherwise, we're safe to make the change. Make sure that the edge from
517 // DestSI to DestSucc is not critical and has no PHI nodes.
518 DEBUG(errs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
519 DEBUG(errs() << "THROUGH: " << *DestSI);
521 // If the destination has PHI nodes, just split the edge for updating
523 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
524 SplitCriticalEdge(DestSI, i, this);
525 DestSucc = DestSI->getSuccessor(i);
527 FoldSingleEntryPHINodes(DestSucc);
528 PredSI->setSuccessor(PredCase, DestSucc);
540 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
541 /// load instruction, eliminate it by replacing it with a PHI node. This is an
542 /// important optimization that encourages jump threading, and needs to be run
543 /// interlaced with other jump threading tasks.
544 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
545 // Don't hack volatile loads.
546 if (LI->isVolatile()) return false;
548 // If the load is defined in a block with exactly one predecessor, it can't be
549 // partially redundant.
550 BasicBlock *LoadBB = LI->getParent();
551 if (LoadBB->getSinglePredecessor())
554 Value *LoadedPtr = LI->getOperand(0);
556 // If the loaded operand is defined in the LoadBB, it can't be available.
557 // FIXME: Could do PHI translation, that would be fun :)
558 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
559 if (PtrOp->getParent() == LoadBB)
562 // Scan a few instructions up from the load, to see if it is obviously live at
563 // the entry to its block.
564 BasicBlock::iterator BBIt = LI;
566 if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB,
568 // If the value if the load is locally available within the block, just use
569 // it. This frequently occurs for reg2mem'd allocas.
570 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
572 // If the returned value is the load itself, replace with an undef. This can
573 // only happen in dead loops.
574 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
575 LI->replaceAllUsesWith(AvailableVal);
576 LI->eraseFromParent();
580 // Otherwise, if we scanned the whole block and got to the top of the block,
581 // we know the block is locally transparent to the load. If not, something
582 // might clobber its value.
583 if (BBIt != LoadBB->begin())
587 SmallPtrSet<BasicBlock*, 8> PredsScanned;
588 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
589 AvailablePredsTy AvailablePreds;
590 BasicBlock *OneUnavailablePred = 0;
592 // If we got here, the loaded value is transparent through to the start of the
593 // block. Check to see if it is available in any of the predecessor blocks.
594 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
596 BasicBlock *PredBB = *PI;
598 // If we already scanned this predecessor, skip it.
599 if (!PredsScanned.insert(PredBB))
602 // Scan the predecessor to see if the value is available in the pred.
603 BBIt = PredBB->end();
604 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
605 if (!PredAvailable) {
606 OneUnavailablePred = PredBB;
610 // If so, this load is partially redundant. Remember this info so that we
611 // can create a PHI node.
612 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
615 // If the loaded value isn't available in any predecessor, it isn't partially
617 if (AvailablePreds.empty()) return false;
619 // Okay, the loaded value is available in at least one (and maybe all!)
620 // predecessors. If the value is unavailable in more than one unique
621 // predecessor, we want to insert a merge block for those common predecessors.
622 // This ensures that we only have to insert one reload, thus not increasing
624 BasicBlock *UnavailablePred = 0;
626 // If there is exactly one predecessor where the value is unavailable, the
627 // already computed 'OneUnavailablePred' block is it. If it ends in an
628 // unconditional branch, we know that it isn't a critical edge.
629 if (PredsScanned.size() == AvailablePreds.size()+1 &&
630 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
631 UnavailablePred = OneUnavailablePred;
632 } else if (PredsScanned.size() != AvailablePreds.size()) {
633 // Otherwise, we had multiple unavailable predecessors or we had a critical
634 // edge from the one.
635 SmallVector<BasicBlock*, 8> PredsToSplit;
636 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
638 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
639 AvailablePredSet.insert(AvailablePreds[i].first);
641 // Add all the unavailable predecessors to the PredsToSplit list.
642 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
644 if (!AvailablePredSet.count(*PI))
645 PredsToSplit.push_back(*PI);
647 // Split them out to their own block.
649 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
650 "thread-split", this);
653 // If the value isn't available in all predecessors, then there will be
654 // exactly one where it isn't available. Insert a load on that edge and add
655 // it to the AvailablePreds list.
656 if (UnavailablePred) {
657 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
658 "Can't handle critical edge here!");
659 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr",
660 UnavailablePred->getTerminator());
661 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
664 // Now we know that each predecessor of this block has a value in
665 // AvailablePreds, sort them for efficient access as we're walking the preds.
666 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
668 // Create a PHI node at the start of the block for the PRE'd load value.
669 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
672 // Insert new entries into the PHI for each predecessor. A single block may
673 // have multiple entries here.
674 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
676 AvailablePredsTy::iterator I =
677 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
678 std::make_pair(*PI, (Value*)0));
680 assert(I != AvailablePreds.end() && I->first == *PI &&
681 "Didn't find entry for predecessor!");
683 PN->addIncoming(I->second, I->first);
686 //cerr << "PRE: " << *LI << *PN << "\n";
688 LI->replaceAllUsesWith(PN);
689 LI->eraseFromParent();
695 /// ProcessJumpOnPHI - We have a conditional branch or switch on a PHI node in
696 /// the current block. See if there are any simplifications we can do based on
697 /// inputs to the phi node.
699 bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) {
700 BasicBlock *BB = PN->getParent();
702 // See if the phi node has any constant integer or undef values. If so, we
703 // can determine where the corresponding predecessor will branch.
704 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
705 Value *PredVal = PN->getIncomingValue(i);
707 // Check to see if this input is a constant integer. If so, the direction
708 // of the branch is predictable.
709 if (ConstantInt *CI = dyn_cast<ConstantInt>(PredVal)) {
710 // Merge any common predecessors that will act the same.
711 BasicBlock *PredBB = FactorCommonPHIPreds(PN, CI);
714 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
715 SuccBB = BI->getSuccessor(CI->isZero());
717 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
718 SuccBB = SI->getSuccessor(SI->findCaseValue(CI));
721 // Ok, try to thread it!
722 return ThreadEdge(BB, PredBB, SuccBB);
725 // If the input is an undef, then it doesn't matter which way it will go.
726 // Pick an arbitrary dest and thread the edge.
727 if (UndefValue *UV = dyn_cast<UndefValue>(PredVal)) {
728 // Merge any common predecessors that will act the same.
729 BasicBlock *PredBB = FactorCommonPHIPreds(PN, UV);
731 BB->getTerminator()->getSuccessor(GetBestDestForJumpOnUndef(BB));
733 // Ok, try to thread it!
734 return ThreadEdge(BB, PredBB, SuccBB);
738 // If the incoming values are all variables, we don't know the destination of
739 // any predecessors. However, if any of the predecessor blocks end in an
740 // unconditional branch, we can *duplicate* the jump into that block in order
741 // to further encourage jump threading and to eliminate cases where we have
742 // branch on a phi of an icmp (branch on icmp is much better).
744 // We don't want to do this tranformation for switches, because we don't
745 // really want to duplicate a switch.
746 if (isa<SwitchInst>(BB->getTerminator()))
749 // Look for unconditional branch predecessors.
750 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
751 BasicBlock *PredBB = PN->getIncomingBlock(i);
752 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
753 if (PredBr->isUnconditional() &&
754 // Try to duplicate BB into PredBB.
755 DuplicateCondBranchOnPHIIntoPred(BB, PredBB))
763 /// ProcessJumpOnLogicalPHI - PN's basic block contains a conditional branch
764 /// whose condition is an AND/OR where one side is PN. If PN has constant
765 /// operands that permit us to evaluate the condition for some operand, thread
766 /// through the block. For example with:
767 /// br (and X, phi(Y, Z, false))
768 /// the predecessor corresponding to the 'false' will always jump to the false
769 /// destination of the branch.
771 bool JumpThreading::ProcessBranchOnLogical(Value *V, BasicBlock *BB,
773 // If this is a binary operator tree of the same AND/OR opcode, check the
775 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(V))
776 if ((isAnd && BO->getOpcode() == Instruction::And) ||
777 (!isAnd && BO->getOpcode() == Instruction::Or)) {
778 if (ProcessBranchOnLogical(BO->getOperand(0), BB, isAnd))
780 if (ProcessBranchOnLogical(BO->getOperand(1), BB, isAnd))
784 // If this isn't a PHI node, we can't handle it.
785 PHINode *PN = dyn_cast<PHINode>(V);
786 if (!PN || PN->getParent() != BB) return false;
788 // We can only do the simplification for phi nodes of 'false' with AND or
789 // 'true' with OR. See if we have any entries in the phi for this.
790 unsigned PredNo = ~0U;
791 ConstantInt *PredCst = ConstantInt::get(Type::getInt1Ty(BB->getContext()),
793 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
794 if (PN->getIncomingValue(i) == PredCst) {
800 // If no match, bail out.
804 // If so, we can actually do this threading. Merge any common predecessors
805 // that will act the same.
806 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst);
808 // Next, figure out which successor we are threading to. If this was an AND,
809 // the constant must be FALSE, and we must be targeting the 'false' block.
810 // If this is an OR, the constant must be TRUE, and we must be targeting the
812 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(isAnd);
814 // Ok, try to thread it!
815 return ThreadEdge(BB, PredBB, SuccBB);
818 /// GetResultOfComparison - Given an icmp/fcmp predicate and the left and right
819 /// hand sides of the compare instruction, try to determine the result. If the
820 /// result can not be determined, a null pointer is returned.
821 static Constant *GetResultOfComparison(CmpInst::Predicate pred,
822 Value *LHS, Value *RHS,
823 LLVMContext &Context) {
824 if (Constant *CLHS = dyn_cast<Constant>(LHS))
825 if (Constant *CRHS = dyn_cast<Constant>(RHS))
826 return ConstantExpr::getCompare(pred, CLHS, CRHS);
829 if (isa<IntegerType>(LHS->getType()) || isa<PointerType>(LHS->getType()))
830 return ICmpInst::isTrueWhenEqual(pred) ?
831 ConstantInt::getTrue(Context) : ConstantInt::getFalse(Context);
836 /// ProcessBranchOnCompare - We found a branch on a comparison between a phi
837 /// node and a value. If we can identify when the comparison is true between
838 /// the phi inputs and the value, we can fold the compare for that edge and
839 /// thread through it.
840 bool JumpThreading::ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB) {
841 PHINode *PN = cast<PHINode>(Cmp->getOperand(0));
842 Value *RHS = Cmp->getOperand(1);
844 // If the phi isn't in the current block, an incoming edge to this block
845 // doesn't control the destination.
846 if (PN->getParent() != BB)
849 // We can do this simplification if any comparisons fold to true or false.
852 bool TrueDirection = false;
853 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
854 PredVal = PN->getIncomingValue(i);
856 Constant *Res = GetResultOfComparison(Cmp->getPredicate(), PredVal,
857 RHS, Cmp->getContext());
863 // If this folded to a constant expr, we can't do anything.
864 if (ConstantInt *ResC = dyn_cast<ConstantInt>(Res)) {
865 TrueDirection = ResC->getZExtValue();
868 // If this folded to undef, just go the false way.
869 if (isa<UndefValue>(Res)) {
870 TrueDirection = false;
874 // Otherwise, we can't fold this input.
878 // If no match, bail out.
882 // If so, we can actually do this threading. Merge any common predecessors
883 // that will act the same.
884 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredVal);
886 // Next, get our successor.
887 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(!TrueDirection);
889 // Ok, try to thread it!
890 return ThreadEdge(BB, PredBB, SuccBB);
894 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
895 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
896 /// NewPred using the entries from OldPred (suitably mapped).
897 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
900 DenseMap<Instruction*, Value*> &ValueMap) {
901 for (BasicBlock::iterator PNI = PHIBB->begin();
902 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
903 // Ok, we have a PHI node. Figure out what the incoming value was for the
905 Value *IV = PN->getIncomingValueForBlock(OldPred);
907 // Remap the value if necessary.
908 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
909 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
910 if (I != ValueMap.end())
914 PN->addIncoming(IV, NewPred);
918 /// ThreadEdge - We have decided that it is safe and profitable to thread an
919 /// edge from PredBB to SuccBB across BB. Transform the IR to reflect this
921 bool JumpThreading::ThreadEdge(BasicBlock *BB, BasicBlock *PredBB,
922 BasicBlock *SuccBB) {
923 // If threading to the same block as we come from, we would infinite loop.
925 DEBUG(errs() << " Not threading across BB '" << BB->getName()
926 << "' - would thread to self!\n");
930 // If threading this would thread across a loop header, don't thread the edge.
931 // See the comments above FindLoopHeaders for justifications and caveats.
932 if (LoopHeaders.count(BB)) {
933 DEBUG(errs() << " Not threading from '" << PredBB->getName()
934 << "' across loop header BB '" << BB->getName()
935 << "' to dest BB '" << SuccBB->getName()
936 << "' - it might create an irreducible loop!\n");
940 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
941 if (JumpThreadCost > Threshold) {
942 DEBUG(errs() << " Not threading BB '" << BB->getName()
943 << "' - Cost is too high: " << JumpThreadCost << "\n");
947 // And finally, do it!
948 DEBUG(errs() << " Threading edge from '" << PredBB->getName() << "' to '"
949 << SuccBB->getName() << "' with cost: " << JumpThreadCost
950 << ", across block:\n "
953 // We are going to have to map operands from the original BB block to the new
954 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
955 // account for entry from PredBB.
956 DenseMap<Instruction*, Value*> ValueMapping;
958 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
959 BB->getName()+".thread",
960 BB->getParent(), BB);
961 NewBB->moveAfter(PredBB);
963 BasicBlock::iterator BI = BB->begin();
964 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
965 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
967 // Clone the non-phi instructions of BB into NewBB, keeping track of the
968 // mapping and using it to remap operands in the cloned instructions.
969 for (; !isa<TerminatorInst>(BI); ++BI) {
970 Instruction *New = BI->clone();
971 New->setName(BI->getName());
972 NewBB->getInstList().push_back(New);
973 ValueMapping[BI] = New;
975 // Remap operands to patch up intra-block references.
976 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
977 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
978 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
979 if (I != ValueMapping.end())
980 New->setOperand(i, I->second);
984 // We didn't copy the terminator from BB over to NewBB, because there is now
985 // an unconditional jump to SuccBB. Insert the unconditional jump.
986 BranchInst::Create(SuccBB, NewBB);
988 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
989 // PHI nodes for NewBB now.
990 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
992 // If there were values defined in BB that are used outside the block, then we
993 // now have to update all uses of the value to use either the original value,
994 // the cloned value, or some PHI derived value. This can require arbitrary
995 // PHI insertion, of which we are prepared to do, clean these up now.
996 SSAUpdater SSAUpdate;
997 SmallVector<Use*, 16> UsesToRename;
998 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
999 // Scan all uses of this instruction to see if it is used outside of its
1000 // block, and if so, record them in UsesToRename.
1001 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1003 Instruction *User = cast<Instruction>(*UI);
1004 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1005 if (UserPN->getIncomingBlock(UI) == BB)
1007 } else if (User->getParent() == BB)
1010 UsesToRename.push_back(&UI.getUse());
1013 // If there are no uses outside the block, we're done with this instruction.
1014 if (UsesToRename.empty())
1017 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n");
1019 // We found a use of I outside of BB. Rename all uses of I that are outside
1020 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1021 // with the two values we know.
1022 SSAUpdate.Initialize(I);
1023 SSAUpdate.AddAvailableValue(BB, I);
1024 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1026 while (!UsesToRename.empty())
1027 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1028 DEBUG(errs() << "\n");
1032 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1033 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1034 // us to simplify any PHI nodes in BB.
1035 TerminatorInst *PredTerm = PredBB->getTerminator();
1036 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1037 if (PredTerm->getSuccessor(i) == BB) {
1038 BB->removePredecessor(PredBB);
1039 PredTerm->setSuccessor(i, NewBB);
1042 // At this point, the IR is fully up to date and consistent. Do a quick scan
1043 // over the new instructions and zap any that are constants or dead. This
1044 // frequently happens because of phi translation.
1045 BI = NewBB->begin();
1046 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) {
1047 Instruction *Inst = BI++;
1048 if (Constant *C = ConstantFoldInstruction(Inst, BB->getContext(), TD)) {
1049 Inst->replaceAllUsesWith(C);
1050 Inst->eraseFromParent();
1054 RecursivelyDeleteTriviallyDeadInstructions(Inst);
1057 // Threaded an edge!
1062 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1063 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1064 /// If we can duplicate the contents of BB up into PredBB do so now, this
1065 /// improves the odds that the branch will be on an analyzable instruction like
1067 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1068 BasicBlock *PredBB) {
1069 // If BB is a loop header, then duplicating this block outside the loop would
1070 // cause us to transform this into an irreducible loop, don't do this.
1071 // See the comments above FindLoopHeaders for justifications and caveats.
1072 if (LoopHeaders.count(BB)) {
1073 DEBUG(errs() << " Not duplicating loop header '" << BB->getName()
1074 << "' into predecessor block '" << PredBB->getName()
1075 << "' - it might create an irreducible loop!\n");
1079 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1080 if (DuplicationCost > Threshold) {
1081 DEBUG(errs() << " Not duplicating BB '" << BB->getName()
1082 << "' - Cost is too high: " << DuplicationCost << "\n");
1086 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1088 DEBUG(errs() << " Duplicating block '" << BB->getName() << "' into end of '"
1089 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1090 << DuplicationCost << " block is:" << *BB << "\n");
1092 // We are going to have to map operands from the original BB block into the
1093 // PredBB block. Evaluate PHI nodes in BB.
1094 DenseMap<Instruction*, Value*> ValueMapping;
1096 BasicBlock::iterator BI = BB->begin();
1097 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1098 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1100 BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1102 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1103 // mapping and using it to remap operands in the cloned instructions.
1104 for (; BI != BB->end(); ++BI) {
1105 Instruction *New = BI->clone();
1106 New->setName(BI->getName());
1107 PredBB->getInstList().insert(OldPredBranch, New);
1108 ValueMapping[BI] = New;
1110 // Remap operands to patch up intra-block references.
1111 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1112 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1113 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1114 if (I != ValueMapping.end())
1115 New->setOperand(i, I->second);
1119 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1120 // add entries to the PHI nodes for branch from PredBB now.
1121 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1122 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1124 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1127 // If there were values defined in BB that are used outside the block, then we
1128 // now have to update all uses of the value to use either the original value,
1129 // the cloned value, or some PHI derived value. This can require arbitrary
1130 // PHI insertion, of which we are prepared to do, clean these up now.
1131 SSAUpdater SSAUpdate;
1132 SmallVector<Use*, 16> UsesToRename;
1133 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1134 // Scan all uses of this instruction to see if it is used outside of its
1135 // block, and if so, record them in UsesToRename.
1136 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1138 Instruction *User = cast<Instruction>(*UI);
1139 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1140 if (UserPN->getIncomingBlock(UI) == BB)
1142 } else if (User->getParent() == BB)
1145 UsesToRename.push_back(&UI.getUse());
1148 // If there are no uses outside the block, we're done with this instruction.
1149 if (UsesToRename.empty())
1152 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n");
1154 // We found a use of I outside of BB. Rename all uses of I that are outside
1155 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1156 // with the two values we know.
1157 SSAUpdate.Initialize(I);
1158 SSAUpdate.AddAvailableValue(BB, I);
1159 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1161 while (!UsesToRename.empty())
1162 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1163 DEBUG(errs() << "\n");
1166 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1168 BB->removePredecessor(PredBB);
1170 // Remove the unconditional branch at the end of the PredBB block.
1171 OldPredBranch->eraseFromParent();