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/Pass.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/Analysis/ConstantFolding.h"
22 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
23 #include "llvm/Transforms/Utils/Local.h"
24 #include "llvm/Target/TargetData.h"
25 #include "llvm/Support/CommandLine.h"
26 #include "llvm/Support/Compiler.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/ADT/SmallPtrSet.h"
31 STATISTIC(NumThreads, "Number of jumps threaded");
32 STATISTIC(NumFolds, "Number of terminators folded");
34 static cl::opt<unsigned>
35 Threshold("jump-threading-threshold",
36 cl::desc("Max block size to duplicate for jump threading"),
37 cl::init(6), cl::Hidden);
40 /// This pass performs 'jump threading', which looks at blocks that have
41 /// multiple predecessors and multiple successors. If one or more of the
42 /// predecessors of the block can be proven to always jump to one of the
43 /// successors, we forward the edge from the predecessor to the successor by
44 /// duplicating the contents of this block.
46 /// An example of when this can occur is code like this:
53 /// In this case, the unconditional branch at the end of the first if can be
54 /// revectored to the false side of the second if.
56 class VISIBILITY_HIDDEN JumpThreading : public FunctionPass {
59 static char ID; // Pass identification
60 JumpThreading() : FunctionPass(&ID) {}
62 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
63 AU.addRequired<TargetData>();
66 bool runOnFunction(Function &F);
67 bool ProcessBlock(BasicBlock *BB);
68 void ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, BasicBlock *SuccBB);
69 BasicBlock *FactorCommonPHIPreds(PHINode *PN, Constant *CstVal);
70 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
71 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
73 bool ProcessJumpOnPHI(PHINode *PN);
74 bool ProcessBranchOnLogical(Value *V, BasicBlock *BB, bool isAnd);
75 bool ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB);
77 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
81 char JumpThreading::ID = 0;
82 static RegisterPass<JumpThreading>
83 X("jump-threading", "Jump Threading");
85 // Public interface to the Jump Threading pass
86 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
88 /// runOnFunction - Top level algorithm.
90 bool JumpThreading::runOnFunction(Function &F) {
91 DOUT << "Jump threading on function '" << F.getNameStart() << "'\n";
92 TD = &getAnalysis<TargetData>();
94 bool AnotherIteration = true, EverChanged = false;
95 while (AnotherIteration) {
96 AnotherIteration = false;
98 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
100 while (ProcessBlock(BB))
105 // If the block is trivially dead, zap it. This eliminates the successor
106 // edges which simplifies the CFG.
107 if (pred_begin(BB) == pred_end(BB) &&
108 BB != &BB->getParent()->getEntryBlock()) {
109 DOUT << " JT: Deleting dead block '" << BB->getNameStart()
110 << "' with terminator: " << *BB->getTerminator();
115 AnotherIteration = Changed;
116 EverChanged |= Changed;
121 /// FactorCommonPHIPreds - If there are multiple preds with the same incoming
122 /// value for the PHI, factor them together so we get one block to thread for
124 /// This is important for things like "phi i1 [true, true, false, true, x]"
125 /// where we only need to clone the block for the true blocks once.
127 BasicBlock *JumpThreading::FactorCommonPHIPreds(PHINode *PN, Constant *CstVal) {
128 SmallVector<BasicBlock*, 16> CommonPreds;
129 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
130 if (PN->getIncomingValue(i) == CstVal)
131 CommonPreds.push_back(PN->getIncomingBlock(i));
133 if (CommonPreds.size() == 1)
134 return CommonPreds[0];
136 DOUT << " Factoring out " << CommonPreds.size()
137 << " common predecessors.\n";
138 return SplitBlockPredecessors(PN->getParent(),
139 &CommonPreds[0], CommonPreds.size(),
144 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
145 /// thread across it.
146 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
147 /// Ignore PHI nodes, these will be flattened when duplication happens.
148 BasicBlock::const_iterator I = BB->getFirstNonPHI();
150 // Sum up the cost of each instruction until we get to the terminator. Don't
151 // include the terminator because the copy won't include it.
153 for (; !isa<TerminatorInst>(I); ++I) {
154 // Debugger intrinsics don't incur code size.
155 if (isa<DbgInfoIntrinsic>(I)) continue;
157 // If this is a pointer->pointer bitcast, it is free.
158 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType()))
161 // All other instructions count for at least one unit.
164 // Calls are more expensive. If they are non-intrinsic calls, we model them
165 // as having cost of 4. If they are a non-vector intrinsic, we model them
166 // as having cost of 2 total, and if they are a vector intrinsic, we model
167 // them as having cost 1.
168 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
169 if (!isa<IntrinsicInst>(CI))
171 else if (isa<VectorType>(CI->getType()))
176 // Threading through a switch statement is particularly profitable. If this
177 // block ends in a switch, decrease its cost to make it more likely to happen.
178 if (isa<SwitchInst>(I))
179 Size = Size > 6 ? Size-6 : 0;
184 /// ProcessBlock - If there are any predecessors whose control can be threaded
185 /// through to a successor, transform them now.
186 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
187 // If this block has a single predecessor, and if that pred has a single
188 // successor, merge the blocks. This encourages recursive jump threading
189 // because now the condition in this block can be threaded through
190 // predecessors of our predecessor block.
191 if (BasicBlock *SinglePred = BB->getSinglePredecessor())
192 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
194 // Remember if SinglePred was the entry block of the function. If so, we
195 // will need to move BB back to the entry position.
196 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
197 MergeBasicBlockIntoOnlyPred(BB);
199 if (isEntry && BB != &BB->getParent()->getEntryBlock())
200 BB->moveBefore(&BB->getParent()->getEntryBlock());
204 // See if this block ends with a branch or switch. If so, see if the
205 // condition is a phi node. If so, and if an entry of the phi node is a
206 // constant, we can thread the block.
208 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
209 // Can't thread an unconditional jump.
210 if (BI->isUnconditional()) return false;
211 Condition = BI->getCondition();
212 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
213 Condition = SI->getCondition();
215 return false; // Must be an invoke.
217 // If the terminator of this block is branching on a constant, simplify the
218 // terminator to an unconditional branch. This can occur due to threading in
220 if (isa<ConstantInt>(Condition)) {
221 DOUT << " In block '" << BB->getNameStart()
222 << "' folding terminator: " << *BB->getTerminator();
224 ConstantFoldTerminator(BB);
228 // If the terminator is branching on an undef, we can pick any of the
229 // successors to branch to. Since this is arbitrary, we pick the successor
230 // with the fewest predecessors. This should reduce the in-degree of the
232 if (isa<UndefValue>(Condition)) {
233 TerminatorInst *BBTerm = BB->getTerminator();
234 unsigned MinSucc = 0;
235 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
236 // Compute the successor with the minimum number of predecessors.
237 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
238 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
239 TestBB = BBTerm->getSuccessor(i);
240 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
241 if (NumPreds < MinNumPreds)
245 // Fold the branch/switch.
246 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
247 if (i == MinSucc) continue;
248 BBTerm->getSuccessor(i)->removePredecessor(BB);
251 DOUT << " In block '" << BB->getNameStart()
252 << "' folding undef terminator: " << *BBTerm;
253 BranchInst::Create(BBTerm->getSuccessor(MinSucc), BBTerm);
254 BBTerm->eraseFromParent();
258 Instruction *CondInst = dyn_cast<Instruction>(Condition);
260 // If the condition is an instruction defined in another block, see if a
261 // predecessor has the same condition:
265 if (!Condition->hasOneUse() && // Multiple uses.
266 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition.
267 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
268 if (isa<BranchInst>(BB->getTerminator())) {
269 for (; PI != E; ++PI)
270 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
271 if (PBI->isConditional() && PBI->getCondition() == Condition &&
272 ProcessBranchOnDuplicateCond(*PI, BB))
275 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator");
276 for (; PI != E; ++PI)
277 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator()))
278 if (PSI->getCondition() == Condition &&
279 ProcessSwitchOnDuplicateCond(*PI, BB))
284 // If there is only a single predecessor of this block, nothing to fold.
285 if (BB->getSinglePredecessor())
288 // All the rest of our checks depend on the condition being an instruction.
292 // See if this is a phi node in the current block.
293 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
294 if (PN->getParent() == BB)
295 return ProcessJumpOnPHI(PN);
297 // If this is a conditional branch whose condition is and/or of a phi, try to
299 if ((CondInst->getOpcode() == Instruction::And ||
300 CondInst->getOpcode() == Instruction::Or) &&
301 isa<BranchInst>(BB->getTerminator()) &&
302 ProcessBranchOnLogical(CondInst, BB,
303 CondInst->getOpcode() == Instruction::And))
306 // If we have "br (phi != 42)" and the phi node has any constant values as
307 // operands, we can thread through this block.
308 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst))
309 if (isa<PHINode>(CondCmp->getOperand(0)) &&
310 isa<Constant>(CondCmp->getOperand(1)) &&
311 ProcessBranchOnCompare(CondCmp, BB))
314 // Check for some cases that are worth simplifying. Right now we want to look
315 // for loads that are used by a switch or by the condition for the branch. If
316 // we see one, check to see if it's partially redundant. If so, insert a PHI
317 // which can then be used to thread the values.
319 // This is particularly important because reg2mem inserts loads and stores all
320 // over the place, and this blocks jump threading if we don't zap them.
321 Value *SimplifyValue = CondInst;
322 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
323 if (isa<Constant>(CondCmp->getOperand(1)))
324 SimplifyValue = CondCmp->getOperand(0);
326 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
327 if (SimplifyPartiallyRedundantLoad(LI))
330 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
331 // "(X == 4)" thread through this block.
336 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
337 /// block that jump on exactly the same condition. This means that we almost
338 /// always know the direction of the edge in the DESTBB:
340 /// br COND, DESTBB, BBY
342 /// br COND, BBZ, BBW
344 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
345 /// in DESTBB, we have to thread over it.
346 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
348 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
350 // If both successors of PredBB go to DESTBB, we don't know anything. We can
351 // fold the branch to an unconditional one, which allows other recursive
354 if (PredBI->getSuccessor(1) != BB)
356 else if (PredBI->getSuccessor(0) != BB)
359 DOUT << " In block '" << PredBB->getNameStart()
360 << "' folding terminator: " << *PredBB->getTerminator();
362 ConstantFoldTerminator(PredBB);
366 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
368 // If the dest block has one predecessor, just fix the branch condition to a
369 // constant and fold it.
370 if (BB->getSinglePredecessor()) {
371 DOUT << " In block '" << BB->getNameStart()
372 << "' folding condition to '" << BranchDir << "': "
373 << *BB->getTerminator();
375 DestBI->setCondition(ConstantInt::get(Type::Int1Ty, BranchDir));
376 ConstantFoldTerminator(BB);
380 // Otherwise we need to thread from PredBB to DestBB's successor which
381 // involves code duplication. Check to see if it is worth it.
382 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
383 if (JumpThreadCost > Threshold) {
384 DOUT << " Not threading BB '" << BB->getNameStart()
385 << "' - Cost is too high: " << JumpThreadCost << "\n";
389 // Next, figure out which successor we are threading to.
390 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
392 // If threading to the same block as we come from, we would infinite loop.
394 DOUT << " Not threading BB '" << BB->getNameStart()
395 << "' - would thread to self!\n";
399 // And finally, do it!
400 DOUT << " Threading edge from '" << PredBB->getNameStart() << "' to '"
401 << SuccBB->getNameStart() << "' with cost: " << JumpThreadCost
402 << ", across block:\n "
405 ThreadEdge(BB, PredBB, SuccBB);
410 struct APIntUnsignedOrdering {
411 bool operator()(const APInt &LHS, const APInt &RHS) const {
416 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
417 /// block that switch on exactly the same condition. This means that we almost
418 /// always know the direction of the edge in the DESTBB:
420 /// switch COND [... DESTBB, BBY ... ]
422 /// switch COND [... BBZ, BBW ]
424 /// Optimizing switches like this is very important, because simplifycfg builds
425 /// switches out of repeated 'if' conditions.
426 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
427 BasicBlock *DestBB) {
428 // Can't thread edge to self.
429 if (PredBB == DestBB)
433 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
434 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
436 // There are a variety of optimizations that we can potentially do on these
437 // blocks: we order them from most to least preferable.
439 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
440 // directly to their destination. This does not introduce *any* code size
443 // FIXME: Thread if it just contains a PHI.
444 if (isa<SwitchInst>(DestBB->begin())) {
445 bool MadeChange = false;
446 // Ignore the default edge for now.
447 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
448 ConstantInt *DestVal = DestSI->getCaseValue(i);
449 BasicBlock *DestSucc = DestSI->getSuccessor(i);
451 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
452 // PredSI has an explicit case for it. If so, forward. If it is covered
453 // by the default case, we can't update PredSI.
454 unsigned PredCase = PredSI->findCaseValue(DestVal);
455 if (PredCase == 0) continue;
457 // If PredSI doesn't go to DestBB on this value, then it won't reach the
458 // case on this condition.
459 if (PredSI->getSuccessor(PredCase) != DestBB &&
460 DestSI->getSuccessor(i) != DestBB)
463 // Otherwise, we're safe to make the change. Make sure that the edge from
464 // DestSI to DestSucc is not critical and has no PHI nodes.
465 DOUT << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI;
466 DOUT << "THROUGH: " << *DestSI;
468 // If the destination has PHI nodes, just split the edge for updating
470 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
471 SplitCriticalEdge(DestSI, i, this);
472 DestSucc = DestSI->getSuccessor(i);
474 FoldSingleEntryPHINodes(DestSucc);
475 PredSI->setSuccessor(PredCase, DestSucc);
484 // Figure out on which of the condition allow us to get to DESTBB. If DESTBB
485 // is the default label, we compute the set of values COND is known not to be
486 // otherwise we compute the set of options that COND could be.
487 SmallVector<APInt, 16> KnownValues;
488 bool DestBBIsDefault = PredSI->getSuccessor(0) == DestBB;
490 if (DestBBIsDefault) {
491 // DestBB the default case. Collect the values where PredBB can't branch to
493 for (unsigned i = 1/*skip default*/, e = PredSI->getNumCases(); i != e; ++i)
494 if (PredSI->getSuccessor(i) != DestBB)
495 KnownValues.push_back(PredSI->getCaseValue(i)->getValue());
497 // Not the default case. Collect the values where PredBB can branch to
499 for (unsigned i = 1/*skip default*/, e = PredSI->getNumCases(); i != e; ++i)
500 if (PredSI->getSuccessor(i) == DestBB)
501 KnownValues.push_back(PredSI->getCaseValue(i)->getValue());
504 std::sort(KnownValues.begin(), KnownValues.end(), APIntUnsignedOrdering());
506 cerr << "\nFOUND THREAD BLOCKS:\n";
515 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
516 /// load instruction, eliminate it by replacing it with a PHI node. This is an
517 /// important optimization that encourages jump threading, and needs to be run
518 /// interlaced with other jump threading tasks.
519 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
520 // Don't hack volatile loads.
521 if (LI->isVolatile()) return false;
523 // If the load is defined in a block with exactly one predecessor, it can't be
524 // partially redundant.
525 BasicBlock *LoadBB = LI->getParent();
526 if (LoadBB->getSinglePredecessor())
529 Value *LoadedPtr = LI->getOperand(0);
531 // If the loaded operand is defined in the LoadBB, it can't be available.
532 // FIXME: Could do PHI translation, that would be fun :)
533 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
534 if (PtrOp->getParent() == LoadBB)
537 // Scan a few instructions up from the load, to see if it is obviously live at
538 // the entry to its block.
539 BasicBlock::iterator BBIt = LI;
541 if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB,
543 // If the value if the load is locally available within the block, just use
544 // it. This frequently occurs for reg2mem'd allocas.
545 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
547 // If the returned value is the load itself, replace with an undef. This can
548 // only happen in dead loops.
549 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
550 LI->replaceAllUsesWith(AvailableVal);
551 LI->eraseFromParent();
555 // Otherwise, if we scanned the whole block and got to the top of the block,
556 // we know the block is locally transparent to the load. If not, something
557 // might clobber its value.
558 if (BBIt != LoadBB->begin())
562 SmallPtrSet<BasicBlock*, 8> PredsScanned;
563 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
564 AvailablePredsTy AvailablePreds;
565 BasicBlock *OneUnavailablePred = 0;
567 // If we got here, the loaded value is transparent through to the start of the
568 // block. Check to see if it is available in any of the predecessor blocks.
569 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
571 BasicBlock *PredBB = *PI;
573 // If we already scanned this predecessor, skip it.
574 if (!PredsScanned.insert(PredBB))
577 // Scan the predecessor to see if the value is available in the pred.
578 BBIt = PredBB->end();
579 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
580 if (!PredAvailable) {
581 OneUnavailablePred = PredBB;
585 // If so, this load is partially redundant. Remember this info so that we
586 // can create a PHI node.
587 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
590 // If the loaded value isn't available in any predecessor, it isn't partially
592 if (AvailablePreds.empty()) return false;
594 // Okay, the loaded value is available in at least one (and maybe all!)
595 // predecessors. If the value is unavailable in more than one unique
596 // predecessor, we want to insert a merge block for those common predecessors.
597 // This ensures that we only have to insert one reload, thus not increasing
599 BasicBlock *UnavailablePred = 0;
601 // If there is exactly one predecessor where the value is unavailable, the
602 // already computed 'OneUnavailablePred' block is it. If it ends in an
603 // unconditional branch, we know that it isn't a critical edge.
604 if (PredsScanned.size() == AvailablePreds.size()+1 &&
605 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
606 UnavailablePred = OneUnavailablePred;
607 } else if (PredsScanned.size() != AvailablePreds.size()) {
608 // Otherwise, we had multiple unavailable predecessors or we had a critical
609 // edge from the one.
610 SmallVector<BasicBlock*, 8> PredsToSplit;
611 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
613 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
614 AvailablePredSet.insert(AvailablePreds[i].first);
616 // Add all the unavailable predecessors to the PredsToSplit list.
617 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
619 if (!AvailablePredSet.count(*PI))
620 PredsToSplit.push_back(*PI);
622 // Split them out to their own block.
624 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
625 "thread-split", this);
628 // If the value isn't available in all predecessors, then there will be
629 // exactly one where it isn't available. Insert a load on that edge and add
630 // it to the AvailablePreds list.
631 if (UnavailablePred) {
632 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
633 "Can't handle critical edge here!");
634 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr",
635 UnavailablePred->getTerminator());
636 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
639 // Now we know that each predecessor of this block has a value in
640 // AvailablePreds, sort them for efficient access as we're walking the preds.
641 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
643 // Create a PHI node at the start of the block for the PRE'd load value.
644 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
647 // Insert new entries into the PHI for each predecessor. A single block may
648 // have multiple entries here.
649 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
651 AvailablePredsTy::iterator I =
652 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
653 std::make_pair(*PI, (Value*)0));
655 assert(I != AvailablePreds.end() && I->first == *PI &&
656 "Didn't find entry for predecessor!");
658 PN->addIncoming(I->second, I->first);
661 //cerr << "PRE: " << *LI << *PN << "\n";
663 LI->replaceAllUsesWith(PN);
664 LI->eraseFromParent();
670 /// ProcessJumpOnPHI - We have a conditional branch of switch on a PHI node in
671 /// the current block. See if there are any simplifications we can do based on
672 /// inputs to the phi node.
674 bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) {
675 // See if the phi node has any constant values. If so, we can determine where
676 // the corresponding predecessor will branch.
677 ConstantInt *PredCst = 0;
678 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
679 if ((PredCst = dyn_cast<ConstantInt>(PN->getIncomingValue(i))))
682 // If no incoming value has a constant, we don't know the destination of any
687 // See if the cost of duplicating this block is low enough.
688 BasicBlock *BB = PN->getParent();
689 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
690 if (JumpThreadCost > Threshold) {
691 DOUT << " Not threading BB '" << BB->getNameStart()
692 << "' - Cost is too high: " << JumpThreadCost << "\n";
696 // If so, we can actually do this threading. Merge any common predecessors
697 // that will act the same.
698 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst);
700 // Next, figure out which successor we are threading to.
702 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
703 SuccBB = BI->getSuccessor(PredCst == ConstantInt::getFalse());
705 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
706 SuccBB = SI->getSuccessor(SI->findCaseValue(PredCst));
709 // If threading to the same block as we come from, we would infinite loop.
711 DOUT << " Not threading BB '" << BB->getNameStart()
712 << "' - would thread to self!\n";
716 // And finally, do it!
717 DOUT << " Threading edge from '" << PredBB->getNameStart() << "' to '"
718 << SuccBB->getNameStart() << "' with cost: " << JumpThreadCost
719 << ", across block:\n "
722 ThreadEdge(BB, PredBB, SuccBB);
727 /// ProcessJumpOnLogicalPHI - PN's basic block contains a conditional branch
728 /// whose condition is an AND/OR where one side is PN. If PN has constant
729 /// operands that permit us to evaluate the condition for some operand, thread
730 /// through the block. For example with:
731 /// br (and X, phi(Y, Z, false))
732 /// the predecessor corresponding to the 'false' will always jump to the false
733 /// destination of the branch.
735 bool JumpThreading::ProcessBranchOnLogical(Value *V, BasicBlock *BB,
737 // If this is a binary operator tree of the same AND/OR opcode, check the
739 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(V))
740 if ((isAnd && BO->getOpcode() == Instruction::And) ||
741 (!isAnd && BO->getOpcode() == Instruction::Or)) {
742 if (ProcessBranchOnLogical(BO->getOperand(0), BB, isAnd))
744 if (ProcessBranchOnLogical(BO->getOperand(1), BB, isAnd))
748 // If this isn't a PHI node, we can't handle it.
749 PHINode *PN = dyn_cast<PHINode>(V);
750 if (!PN || PN->getParent() != BB) return false;
752 // We can only do the simplification for phi nodes of 'false' with AND or
753 // 'true' with OR. See if we have any entries in the phi for this.
754 unsigned PredNo = ~0U;
755 ConstantInt *PredCst = ConstantInt::get(Type::Int1Ty, !isAnd);
756 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
757 if (PN->getIncomingValue(i) == PredCst) {
763 // If no match, bail out.
767 // See if the cost of duplicating this block is low enough.
768 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
769 if (JumpThreadCost > Threshold) {
770 DOUT << " Not threading BB '" << BB->getNameStart()
771 << "' - Cost is too high: " << JumpThreadCost << "\n";
775 // If so, we can actually do this threading. Merge any common predecessors
776 // that will act the same.
777 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst);
779 // Next, figure out which successor we are threading to. If this was an AND,
780 // the constant must be FALSE, and we must be targeting the 'false' block.
781 // If this is an OR, the constant must be TRUE, and we must be targeting the
783 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(isAnd);
785 // If threading to the same block as we come from, we would infinite loop.
787 DOUT << " Not threading BB '" << BB->getNameStart()
788 << "' - would thread to self!\n";
792 // And finally, do it!
793 DOUT << " Threading edge through bool from '" << PredBB->getNameStart()
794 << "' to '" << SuccBB->getNameStart() << "' with cost: "
795 << JumpThreadCost << ", across block:\n "
798 ThreadEdge(BB, PredBB, SuccBB);
803 /// ProcessBranchOnCompare - We found a branch on a comparison between a phi
804 /// node and a constant. If the PHI node contains any constants as inputs, we
805 /// can fold the compare for that edge and thread through it.
806 bool JumpThreading::ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB) {
807 PHINode *PN = cast<PHINode>(Cmp->getOperand(0));
808 Constant *RHS = cast<Constant>(Cmp->getOperand(1));
810 // If the phi isn't in the current block, an incoming edge to this block
811 // doesn't control the destination.
812 if (PN->getParent() != BB)
815 // We can do this simplification if any comparisons fold to true or false.
817 Constant *PredCst = 0;
818 bool TrueDirection = false;
819 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
820 PredCst = dyn_cast<Constant>(PN->getIncomingValue(i));
821 if (PredCst == 0) continue;
824 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Cmp))
825 Res = ConstantExpr::getICmp(ICI->getPredicate(), PredCst, RHS);
827 Res = ConstantExpr::getFCmp(cast<FCmpInst>(Cmp)->getPredicate(),
829 // If this folded to a constant expr, we can't do anything.
830 if (ConstantInt *ResC = dyn_cast<ConstantInt>(Res)) {
831 TrueDirection = ResC->getZExtValue();
834 // If this folded to undef, just go the false way.
835 if (isa<UndefValue>(Res)) {
836 TrueDirection = false;
840 // Otherwise, we can't fold this input.
844 // If no match, bail out.
848 // See if the cost of duplicating this block is low enough.
849 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
850 if (JumpThreadCost > Threshold) {
851 DOUT << " Not threading BB '" << BB->getNameStart()
852 << "' - Cost is too high: " << JumpThreadCost << "\n";
856 // If so, we can actually do this threading. Merge any common predecessors
857 // that will act the same.
858 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst);
860 // Next, get our successor.
861 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(!TrueDirection);
863 // If threading to the same block as we come from, we would infinite loop.
865 DOUT << " Not threading BB '" << BB->getNameStart()
866 << "' - would thread to self!\n";
871 // And finally, do it!
872 DOUT << " Threading edge through bool from '" << PredBB->getNameStart()
873 << "' to '" << SuccBB->getNameStart() << "' with cost: "
874 << JumpThreadCost << ", across block:\n "
877 ThreadEdge(BB, PredBB, SuccBB);
883 /// ThreadEdge - We have decided that it is safe and profitable to thread an
884 /// edge from PredBB to SuccBB across BB. Transform the IR to reflect this
886 void JumpThreading::ThreadEdge(BasicBlock *BB, BasicBlock *PredBB,
887 BasicBlock *SuccBB) {
889 // Jump Threading can not update SSA properties correctly if the values
890 // defined in the duplicated block are used outside of the block itself. For
891 // this reason, we spill all values that are used outside of BB to the stack.
892 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
893 if (!I->isUsedOutsideOfBlock(BB))
896 // We found a use of I outside of BB. Create a new stack slot to
897 // break this inter-block usage pattern.
898 DemoteRegToStack(*I);
901 // We are going to have to map operands from the original BB block to the new
902 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
903 // account for entry from PredBB.
904 DenseMap<Instruction*, Value*> ValueMapping;
907 BasicBlock::Create(BB->getName()+".thread", BB->getParent(), BB);
908 NewBB->moveAfter(PredBB);
910 BasicBlock::iterator BI = BB->begin();
911 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
912 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
914 // Clone the non-phi instructions of BB into NewBB, keeping track of the
915 // mapping and using it to remap operands in the cloned instructions.
916 for (; !isa<TerminatorInst>(BI); ++BI) {
917 Instruction *New = BI->clone();
918 New->setName(BI->getNameStart());
919 NewBB->getInstList().push_back(New);
920 ValueMapping[BI] = New;
922 // Remap operands to patch up intra-block references.
923 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
924 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i)))
925 if (Value *Remapped = ValueMapping[Inst])
926 New->setOperand(i, Remapped);
929 // We didn't copy the terminator from BB over to NewBB, because there is now
930 // an unconditional jump to SuccBB. Insert the unconditional jump.
931 BranchInst::Create(SuccBB, NewBB);
933 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
934 // PHI nodes for NewBB now.
935 for (BasicBlock::iterator PNI = SuccBB->begin(); isa<PHINode>(PNI); ++PNI) {
936 PHINode *PN = cast<PHINode>(PNI);
937 // Ok, we have a PHI node. Figure out what the incoming value was for the
939 Value *IV = PN->getIncomingValueForBlock(BB);
941 // Remap the value if necessary.
942 if (Instruction *Inst = dyn_cast<Instruction>(IV))
943 if (Value *MappedIV = ValueMapping[Inst])
945 PN->addIncoming(IV, NewBB);
948 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
949 // NewBB instead of BB. This eliminates predecessors from BB, which requires
950 // us to simplify any PHI nodes in BB.
951 TerminatorInst *PredTerm = PredBB->getTerminator();
952 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
953 if (PredTerm->getSuccessor(i) == BB) {
954 BB->removePredecessor(PredBB);
955 PredTerm->setSuccessor(i, NewBB);
958 // At this point, the IR is fully up to date and consistent. Do a quick scan
959 // over the new instructions and zap any that are constants or dead. This
960 // frequently happens because of phi translation.
962 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) {
963 Instruction *Inst = BI++;
964 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
965 Inst->replaceAllUsesWith(C);
966 Inst->eraseFromParent();
970 RecursivelyDeleteTriviallyDeadInstructions(Inst);