1 //===- JumpThreading.cpp - Thread control through conditional blocks ------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the Jump Threading pass.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "jump-threading"
15 #include "llvm/Transforms/Scalar.h"
16 #include "llvm/IntrinsicInst.h"
17 #include "llvm/LLVMContext.h"
18 #include "llvm/Pass.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/Analysis/LazyValueInfo.h"
21 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
22 #include "llvm/Transforms/Utils/Local.h"
23 #include "llvm/Transforms/Utils/SSAUpdater.h"
24 #include "llvm/Target/TargetData.h"
25 #include "llvm/ADT/DenseMap.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/ADT/STLExtras.h"
28 #include "llvm/ADT/SmallPtrSet.h"
29 #include "llvm/ADT/SmallSet.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/raw_ostream.h"
35 STATISTIC(NumThreads, "Number of jumps threaded");
36 STATISTIC(NumFolds, "Number of terminators folded");
37 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
39 static cl::opt<unsigned>
40 Threshold("jump-threading-threshold",
41 cl::desc("Max block size to duplicate for jump threading"),
42 cl::init(6), cl::Hidden);
44 // Turn on use of LazyValueInfo.
46 EnableLVI("enable-jump-threading-lvi", cl::ReallyHidden);
51 /// This pass performs 'jump threading', which looks at blocks that have
52 /// multiple predecessors and multiple successors. If one or more of the
53 /// predecessors of the block can be proven to always jump to one of the
54 /// successors, we forward the edge from the predecessor to the successor by
55 /// duplicating the contents of this block.
57 /// An example of when this can occur is code like this:
64 /// In this case, the unconditional branch at the end of the first if can be
65 /// revectored to the false side of the second if.
67 class JumpThreading : public FunctionPass {
71 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
73 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
76 static char ID; // Pass identification
77 JumpThreading() : FunctionPass(&ID) {}
79 bool runOnFunction(Function &F);
81 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
83 AU.addRequired<LazyValueInfo>();
86 void FindLoopHeaders(Function &F);
87 bool ProcessBlock(BasicBlock *BB);
88 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
90 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
93 typedef SmallVectorImpl<std::pair<ConstantInt*,
94 BasicBlock*> > PredValueInfo;
96 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
97 PredValueInfo &Result);
98 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB);
101 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
102 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
104 bool ProcessJumpOnPHI(PHINode *PN);
106 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
110 char JumpThreading::ID = 0;
111 static RegisterPass<JumpThreading>
112 X("jump-threading", "Jump Threading");
114 // Public interface to the Jump Threading pass
115 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
117 /// runOnFunction - Top level algorithm.
119 bool JumpThreading::runOnFunction(Function &F) {
120 DEBUG(errs() << "Jump threading on function '" << F.getName() << "'\n");
121 TD = getAnalysisIfAvailable<TargetData>();
122 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0;
126 bool AnotherIteration = true, EverChanged = false;
127 while (AnotherIteration) {
128 AnotherIteration = false;
129 bool Changed = false;
130 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
132 // Thread all of the branches we can over this block.
133 while (ProcessBlock(BB))
138 // If the block is trivially dead, zap it. This eliminates the successor
139 // edges which simplifies the CFG.
140 if (pred_begin(BB) == pred_end(BB) &&
141 BB != &BB->getParent()->getEntryBlock()) {
142 DEBUG(errs() << " JT: Deleting dead block '" << BB->getName()
143 << "' with terminator: " << *BB->getTerminator() << '\n');
144 LoopHeaders.erase(BB);
147 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
148 // Can't thread an unconditional jump, but if the block is "almost
149 // empty", we can replace uses of it with uses of the successor and make
151 if (BI->isUnconditional() &&
152 BB != &BB->getParent()->getEntryBlock()) {
153 BasicBlock::iterator BBI = BB->getFirstNonPHI();
154 // Ignore dbg intrinsics.
155 while (isa<DbgInfoIntrinsic>(BBI))
157 // If the terminator is the only non-phi instruction, try to nuke it.
158 if (BBI->isTerminator()) {
159 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
160 // block, we have to make sure it isn't in the LoopHeaders set. We
161 // reinsert afterward if needed.
162 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
163 BasicBlock *Succ = BI->getSuccessor(0);
165 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
167 // If we deleted BB and BB was the header of a loop, then the
168 // successor is now the header of the loop.
172 if (ErasedFromLoopHeaders)
173 LoopHeaders.insert(BB);
178 AnotherIteration = Changed;
179 EverChanged |= Changed;
186 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
187 /// thread across it.
188 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
189 /// Ignore PHI nodes, these will be flattened when duplication happens.
190 BasicBlock::const_iterator I = BB->getFirstNonPHI();
192 // FIXME: THREADING will delete values that are just used to compute the
193 // branch, so they shouldn't count against the duplication cost.
196 // Sum up the cost of each instruction until we get to the terminator. Don't
197 // include the terminator because the copy won't include it.
199 for (; !isa<TerminatorInst>(I); ++I) {
200 // Debugger intrinsics don't incur code size.
201 if (isa<DbgInfoIntrinsic>(I)) continue;
203 // If this is a pointer->pointer bitcast, it is free.
204 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType()))
207 // All other instructions count for at least one unit.
210 // Calls are more expensive. If they are non-intrinsic calls, we model them
211 // as having cost of 4. If they are a non-vector intrinsic, we model them
212 // as having cost of 2 total, and if they are a vector intrinsic, we model
213 // them as having cost 1.
214 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
215 if (!isa<IntrinsicInst>(CI))
217 else if (!isa<VectorType>(CI->getType()))
222 // Threading through a switch statement is particularly profitable. If this
223 // block ends in a switch, decrease its cost to make it more likely to happen.
224 if (isa<SwitchInst>(I))
225 Size = Size > 6 ? Size-6 : 0;
230 /// FindLoopHeaders - We do not want jump threading to turn proper loop
231 /// structures into irreducible loops. Doing this breaks up the loop nesting
232 /// hierarchy and pessimizes later transformations. To prevent this from
233 /// happening, we first have to find the loop headers. Here we approximate this
234 /// by finding targets of backedges in the CFG.
236 /// Note that there definitely are cases when we want to allow threading of
237 /// edges across a loop header. For example, threading a jump from outside the
238 /// loop (the preheader) to an exit block of the loop is definitely profitable.
239 /// It is also almost always profitable to thread backedges from within the loop
240 /// to exit blocks, and is often profitable to thread backedges to other blocks
241 /// within the loop (forming a nested loop). This simple analysis is not rich
242 /// enough to track all of these properties and keep it up-to-date as the CFG
243 /// mutates, so we don't allow any of these transformations.
245 void JumpThreading::FindLoopHeaders(Function &F) {
246 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
247 FindFunctionBackedges(F, Edges);
249 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
250 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
253 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
254 /// if we can infer that the value is a known ConstantInt in any of our
255 /// predecessors. If so, return the known list of value and pred BB in the
256 /// result vector. If a value is known to be undef, it is returned as null.
258 /// This returns true if there were any known values.
261 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){
262 // If V is a constantint, then it is known in all predecessors.
263 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) {
264 ConstantInt *CI = dyn_cast<ConstantInt>(V);
266 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
267 Result.push_back(std::make_pair(CI, *PI));
271 // If V is a non-instruction value, or an instruction in a different block,
272 // then it can't be derived from a PHI.
273 Instruction *I = dyn_cast<Instruction>(V);
274 if (I == 0 || I->getParent() != BB) {
276 // Okay, if this is a live-in value, see if it has a known value at the end
277 // of any of our predecessors.
279 // FIXME: This should be an edge property, not a block end property.
280 /// TODO: Per PR2563, we could infer value range information about a
281 /// predecessor based on its terminator.
284 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
285 // "I" is a non-local compare-with-a-constant instruction. This would be
286 // able to handle value inequalities better, for example if the compare is
287 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
288 // Perhaps getConstantOnEdge should be smart enough to do this?
290 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
291 // If the value is known by LazyValueInfo to be a constant in a
292 // predecessor, use that information to try to thread this block.
293 Constant *PredCst = LVI->getConstantOnEdge(V, *PI, BB);
295 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst)))
298 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), *PI));
301 return !Result.empty();
307 /// If I is a PHI node, then we know the incoming values for any constants.
308 if (PHINode *PN = dyn_cast<PHINode>(I)) {
309 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
310 Value *InVal = PN->getIncomingValue(i);
311 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) {
312 ConstantInt *CI = dyn_cast<ConstantInt>(InVal);
313 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i)));
316 return !Result.empty();
319 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals;
321 // Handle some boolean conditions.
322 if (I->getType()->getPrimitiveSizeInBits() == 1) {
324 // X & false -> false
325 if (I->getOpcode() == Instruction::Or ||
326 I->getOpcode() == Instruction::And) {
327 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
328 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals);
330 if (LHSVals.empty() && RHSVals.empty())
333 ConstantInt *InterestingVal;
334 if (I->getOpcode() == Instruction::Or)
335 InterestingVal = ConstantInt::getTrue(I->getContext());
337 InterestingVal = ConstantInt::getFalse(I->getContext());
339 // Scan for the sentinel.
340 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
341 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0)
342 Result.push_back(LHSVals[i]);
343 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
344 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0)
345 Result.push_back(RHSVals[i]);
346 return !Result.empty();
349 // Handle the NOT form of XOR.
350 if (I->getOpcode() == Instruction::Xor &&
351 isa<ConstantInt>(I->getOperand(1)) &&
352 cast<ConstantInt>(I->getOperand(1))->isOne()) {
353 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result);
357 // Invert the known values.
358 for (unsigned i = 0, e = Result.size(); i != e; ++i)
361 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first));
366 // Handle compare with phi operand, where the PHI is defined in this block.
367 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
368 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
369 if (PN && PN->getParent() == BB) {
370 // We can do this simplification if any comparisons fold to true or false.
372 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
373 BasicBlock *PredBB = PN->getIncomingBlock(i);
374 Value *LHS = PN->getIncomingValue(i);
375 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
377 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
379 if (!LVI || !isa<Constant>(RHS))
382 LazyValueInfo::Tristate
383 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
384 cast<Constant>(RHS), PredBB, BB);
385 if (ResT == LazyValueInfo::Unknown)
387 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
390 if (isa<UndefValue>(Res))
391 Result.push_back(std::make_pair((ConstantInt*)0, PredBB));
392 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res))
393 Result.push_back(std::make_pair(CI, PredBB));
396 return !Result.empty();
400 // If comparing a live-in value against a constant, see if we know the
401 // live-in value on any predecessors.
402 if (LVI && isa<Constant>(Cmp->getOperand(1)) &&
403 Cmp->getType()->isInteger() && // Not vector compare.
404 (!isa<Instruction>(Cmp->getOperand(0)) ||
405 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB)) {
406 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
408 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
409 // If the value is known by LazyValueInfo to be a constant in a
410 // predecessor, use that information to try to thread this block.
411 LazyValueInfo::Tristate
412 Res = LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
414 if (Res == LazyValueInfo::Unknown)
417 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
418 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), *PI));
421 return !Result.empty();
429 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
430 /// in an undefined jump, decide which block is best to revector to.
432 /// Since we can pick an arbitrary destination, we pick the successor with the
433 /// fewest predecessors. This should reduce the in-degree of the others.
435 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
436 TerminatorInst *BBTerm = BB->getTerminator();
437 unsigned MinSucc = 0;
438 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
439 // Compute the successor with the minimum number of predecessors.
440 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
441 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
442 TestBB = BBTerm->getSuccessor(i);
443 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
444 if (NumPreds < MinNumPreds)
451 /// ProcessBlock - If there are any predecessors whose control can be threaded
452 /// through to a successor, transform them now.
453 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
454 // If this block has a single predecessor, and if that pred has a single
455 // successor, merge the blocks. This encourages recursive jump threading
456 // because now the condition in this block can be threaded through
457 // predecessors of our predecessor block.
458 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
459 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
461 // If SinglePred was a loop header, BB becomes one.
462 if (LoopHeaders.erase(SinglePred))
463 LoopHeaders.insert(BB);
465 // Remember if SinglePred was the entry block of the function. If so, we
466 // will need to move BB back to the entry position.
467 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
468 MergeBasicBlockIntoOnlyPred(BB);
470 if (isEntry && BB != &BB->getParent()->getEntryBlock())
471 BB->moveBefore(&BB->getParent()->getEntryBlock());
476 // Look to see if the terminator is a branch of switch, if not we can't thread
479 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
480 // Can't thread an unconditional jump.
481 if (BI->isUnconditional()) return false;
482 Condition = BI->getCondition();
483 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
484 Condition = SI->getCondition();
486 return false; // Must be an invoke.
488 // If the terminator of this block is branching on a constant, simplify the
489 // terminator to an unconditional branch. This can occur due to threading in
491 if (isa<ConstantInt>(Condition)) {
492 DEBUG(errs() << " In block '" << BB->getName()
493 << "' folding terminator: " << *BB->getTerminator() << '\n');
495 ConstantFoldTerminator(BB);
499 // If the terminator is branching on an undef, we can pick any of the
500 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
501 if (isa<UndefValue>(Condition)) {
502 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
504 // Fold the branch/switch.
505 TerminatorInst *BBTerm = BB->getTerminator();
506 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
507 if (i == BestSucc) continue;
508 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD);
511 DEBUG(errs() << " In block '" << BB->getName()
512 << "' folding undef terminator: " << *BBTerm << '\n');
513 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
514 BBTerm->eraseFromParent();
518 Instruction *CondInst = dyn_cast<Instruction>(Condition);
520 // If the condition is an instruction defined in another block, see if a
521 // predecessor has the same condition:
526 !Condition->hasOneUse() && // Multiple uses.
527 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition.
528 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
529 if (isa<BranchInst>(BB->getTerminator())) {
530 for (; PI != E; ++PI)
531 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
532 if (PBI->isConditional() && PBI->getCondition() == Condition &&
533 ProcessBranchOnDuplicateCond(*PI, BB))
536 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator");
537 for (; PI != E; ++PI)
538 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator()))
539 if (PSI->getCondition() == Condition &&
540 ProcessSwitchOnDuplicateCond(*PI, BB))
545 // All the rest of our checks depend on the condition being an instruction.
547 // FIXME: Unify this with code below.
548 if (LVI && ProcessThreadableEdges(Condition, BB))
554 // See if this is a phi node in the current block.
555 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
556 if (PN->getParent() == BB)
557 return ProcessJumpOnPHI(PN);
559 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
561 (!isa<PHINode>(CondCmp->getOperand(0)) ||
562 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) {
563 // If we have a comparison, loop over the predecessors to see if there is
564 // a condition with a lexically identical value.
565 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
566 for (; PI != E; ++PI)
567 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
568 if (PBI->isConditional() && *PI != BB) {
569 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) {
570 if (CI->getOperand(0) == CondCmp->getOperand(0) &&
571 CI->getOperand(1) == CondCmp->getOperand(1) &&
572 CI->getPredicate() == CondCmp->getPredicate()) {
573 // TODO: Could handle things like (x != 4) --> (x == 17)
574 if (ProcessBranchOnDuplicateCond(*PI, BB))
582 // Check for some cases that are worth simplifying. Right now we want to look
583 // for loads that are used by a switch or by the condition for the branch. If
584 // we see one, check to see if it's partially redundant. If so, insert a PHI
585 // which can then be used to thread the values.
587 // This is particularly important because reg2mem inserts loads and stores all
588 // over the place, and this blocks jump threading if we don't zap them.
589 Value *SimplifyValue = CondInst;
590 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
591 if (isa<Constant>(CondCmp->getOperand(1)))
592 SimplifyValue = CondCmp->getOperand(0);
594 // TODO: There are other places where load PRE would be profitable, such as
595 // more complex comparisons.
596 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
597 if (SimplifyPartiallyRedundantLoad(LI))
601 // Handle a variety of cases where we are branching on something derived from
602 // a PHI node in the current block. If we can prove that any predecessors
603 // compute a predictable value based on a PHI node, thread those predecessors.
605 if (ProcessThreadableEdges(CondInst, BB))
609 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
610 // "(X == 4)" thread through this block.
615 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
616 /// block that jump on exactly the same condition. This means that we almost
617 /// always know the direction of the edge in the DESTBB:
619 /// br COND, DESTBB, BBY
621 /// br COND, BBZ, BBW
623 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
624 /// in DESTBB, we have to thread over it.
625 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
627 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
629 // If both successors of PredBB go to DESTBB, we don't know anything. We can
630 // fold the branch to an unconditional one, which allows other recursive
633 if (PredBI->getSuccessor(1) != BB)
635 else if (PredBI->getSuccessor(0) != BB)
638 DEBUG(errs() << " In block '" << PredBB->getName()
639 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
641 ConstantFoldTerminator(PredBB);
645 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
647 // If the dest block has one predecessor, just fix the branch condition to a
648 // constant and fold it.
649 if (BB->getSinglePredecessor()) {
650 DEBUG(errs() << " In block '" << BB->getName()
651 << "' folding condition to '" << BranchDir << "': "
652 << *BB->getTerminator() << '\n');
654 Value *OldCond = DestBI->getCondition();
655 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
657 ConstantFoldTerminator(BB);
658 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
663 // Next, figure out which successor we are threading to.
664 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
666 SmallVector<BasicBlock*, 2> Preds;
667 Preds.push_back(PredBB);
669 // Ok, try to thread it!
670 return ThreadEdge(BB, Preds, SuccBB);
673 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
674 /// block that switch on exactly the same condition. This means that we almost
675 /// always know the direction of the edge in the DESTBB:
677 /// switch COND [... DESTBB, BBY ... ]
679 /// switch COND [... BBZ, BBW ]
681 /// Optimizing switches like this is very important, because simplifycfg builds
682 /// switches out of repeated 'if' conditions.
683 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
684 BasicBlock *DestBB) {
685 // Can't thread edge to self.
686 if (PredBB == DestBB)
689 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
690 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
692 // There are a variety of optimizations that we can potentially do on these
693 // blocks: we order them from most to least preferable.
695 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
696 // directly to their destination. This does not introduce *any* code size
697 // growth. Skip debug info first.
698 BasicBlock::iterator BBI = DestBB->begin();
699 while (isa<DbgInfoIntrinsic>(BBI))
702 // FIXME: Thread if it just contains a PHI.
703 if (isa<SwitchInst>(BBI)) {
704 bool MadeChange = false;
705 // Ignore the default edge for now.
706 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
707 ConstantInt *DestVal = DestSI->getCaseValue(i);
708 BasicBlock *DestSucc = DestSI->getSuccessor(i);
710 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
711 // PredSI has an explicit case for it. If so, forward. If it is covered
712 // by the default case, we can't update PredSI.
713 unsigned PredCase = PredSI->findCaseValue(DestVal);
714 if (PredCase == 0) continue;
716 // If PredSI doesn't go to DestBB on this value, then it won't reach the
717 // case on this condition.
718 if (PredSI->getSuccessor(PredCase) != DestBB &&
719 DestSI->getSuccessor(i) != DestBB)
722 // Do not forward this if it already goes to this destination, this would
723 // be an infinite loop.
724 if (PredSI->getSuccessor(PredCase) == DestSucc)
727 // Otherwise, we're safe to make the change. Make sure that the edge from
728 // DestSI to DestSucc is not critical and has no PHI nodes.
729 DEBUG(errs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
730 DEBUG(errs() << "THROUGH: " << *DestSI);
732 // If the destination has PHI nodes, just split the edge for updating
734 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
735 SplitCriticalEdge(DestSI, i, this);
736 DestSucc = DestSI->getSuccessor(i);
738 FoldSingleEntryPHINodes(DestSucc);
739 PredSI->setSuccessor(PredCase, DestSucc);
751 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
752 /// load instruction, eliminate it by replacing it with a PHI node. This is an
753 /// important optimization that encourages jump threading, and needs to be run
754 /// interlaced with other jump threading tasks.
755 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
756 // Don't hack volatile loads.
757 if (LI->isVolatile()) return false;
759 // If the load is defined in a block with exactly one predecessor, it can't be
760 // partially redundant.
761 BasicBlock *LoadBB = LI->getParent();
762 if (LoadBB->getSinglePredecessor())
765 Value *LoadedPtr = LI->getOperand(0);
767 // If the loaded operand is defined in the LoadBB, it can't be available.
768 // TODO: Could do simple PHI translation, that would be fun :)
769 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
770 if (PtrOp->getParent() == LoadBB)
773 // Scan a few instructions up from the load, to see if it is obviously live at
774 // the entry to its block.
775 BasicBlock::iterator BBIt = LI;
777 if (Value *AvailableVal =
778 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
779 // If the value if the load is locally available within the block, just use
780 // it. This frequently occurs for reg2mem'd allocas.
781 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
783 // If the returned value is the load itself, replace with an undef. This can
784 // only happen in dead loops.
785 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
786 LI->replaceAllUsesWith(AvailableVal);
787 LI->eraseFromParent();
791 // Otherwise, if we scanned the whole block and got to the top of the block,
792 // we know the block is locally transparent to the load. If not, something
793 // might clobber its value.
794 if (BBIt != LoadBB->begin())
798 SmallPtrSet<BasicBlock*, 8> PredsScanned;
799 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
800 AvailablePredsTy AvailablePreds;
801 BasicBlock *OneUnavailablePred = 0;
803 // If we got here, the loaded value is transparent through to the start of the
804 // block. Check to see if it is available in any of the predecessor blocks.
805 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
807 BasicBlock *PredBB = *PI;
809 // If we already scanned this predecessor, skip it.
810 if (!PredsScanned.insert(PredBB))
813 // Scan the predecessor to see if the value is available in the pred.
814 BBIt = PredBB->end();
815 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
816 if (!PredAvailable) {
817 OneUnavailablePred = PredBB;
821 // If so, this load is partially redundant. Remember this info so that we
822 // can create a PHI node.
823 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
826 // If the loaded value isn't available in any predecessor, it isn't partially
828 if (AvailablePreds.empty()) return false;
830 // Okay, the loaded value is available in at least one (and maybe all!)
831 // predecessors. If the value is unavailable in more than one unique
832 // predecessor, we want to insert a merge block for those common predecessors.
833 // This ensures that we only have to insert one reload, thus not increasing
835 BasicBlock *UnavailablePred = 0;
837 // If there is exactly one predecessor where the value is unavailable, the
838 // already computed 'OneUnavailablePred' block is it. If it ends in an
839 // unconditional branch, we know that it isn't a critical edge.
840 if (PredsScanned.size() == AvailablePreds.size()+1 &&
841 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
842 UnavailablePred = OneUnavailablePred;
843 } else if (PredsScanned.size() != AvailablePreds.size()) {
844 // Otherwise, we had multiple unavailable predecessors or we had a critical
845 // edge from the one.
846 SmallVector<BasicBlock*, 8> PredsToSplit;
847 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
849 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
850 AvailablePredSet.insert(AvailablePreds[i].first);
852 // Add all the unavailable predecessors to the PredsToSplit list.
853 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
855 if (!AvailablePredSet.count(*PI))
856 PredsToSplit.push_back(*PI);
858 // Split them out to their own block.
860 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
861 "thread-pre-split", this);
864 // If the value isn't available in all predecessors, then there will be
865 // exactly one where it isn't available. Insert a load on that edge and add
866 // it to the AvailablePreds list.
867 if (UnavailablePred) {
868 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
869 "Can't handle critical edge here!");
870 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
872 UnavailablePred->getTerminator());
873 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
876 // Now we know that each predecessor of this block has a value in
877 // AvailablePreds, sort them for efficient access as we're walking the preds.
878 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
880 // Create a PHI node at the start of the block for the PRE'd load value.
881 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
884 // Insert new entries into the PHI for each predecessor. A single block may
885 // have multiple entries here.
886 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
888 AvailablePredsTy::iterator I =
889 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
890 std::make_pair(*PI, (Value*)0));
892 assert(I != AvailablePreds.end() && I->first == *PI &&
893 "Didn't find entry for predecessor!");
895 PN->addIncoming(I->second, I->first);
898 //cerr << "PRE: " << *LI << *PN << "\n";
900 LI->replaceAllUsesWith(PN);
901 LI->eraseFromParent();
906 /// FindMostPopularDest - The specified list contains multiple possible
907 /// threadable destinations. Pick the one that occurs the most frequently in
910 FindMostPopularDest(BasicBlock *BB,
911 const SmallVectorImpl<std::pair<BasicBlock*,
912 BasicBlock*> > &PredToDestList) {
913 assert(!PredToDestList.empty());
915 // Determine popularity. If there are multiple possible destinations, we
916 // explicitly choose to ignore 'undef' destinations. We prefer to thread
917 // blocks with known and real destinations to threading undef. We'll handle
918 // them later if interesting.
919 DenseMap<BasicBlock*, unsigned> DestPopularity;
920 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
921 if (PredToDestList[i].second)
922 DestPopularity[PredToDestList[i].second]++;
924 // Find the most popular dest.
925 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
926 BasicBlock *MostPopularDest = DPI->first;
927 unsigned Popularity = DPI->second;
928 SmallVector<BasicBlock*, 4> SamePopularity;
930 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
931 // If the popularity of this entry isn't higher than the popularity we've
932 // seen so far, ignore it.
933 if (DPI->second < Popularity)
935 else if (DPI->second == Popularity) {
936 // If it is the same as what we've seen so far, keep track of it.
937 SamePopularity.push_back(DPI->first);
939 // If it is more popular, remember it.
940 SamePopularity.clear();
941 MostPopularDest = DPI->first;
942 Popularity = DPI->second;
946 // Okay, now we know the most popular destination. If there is more than
947 // destination, we need to determine one. This is arbitrary, but we need
948 // to make a deterministic decision. Pick the first one that appears in the
950 if (!SamePopularity.empty()) {
951 SamePopularity.push_back(MostPopularDest);
952 TerminatorInst *TI = BB->getTerminator();
953 for (unsigned i = 0; ; ++i) {
954 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
956 if (std::find(SamePopularity.begin(), SamePopularity.end(),
957 TI->getSuccessor(i)) == SamePopularity.end())
960 MostPopularDest = TI->getSuccessor(i);
965 // Okay, we have finally picked the most popular destination.
966 return MostPopularDest;
969 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) {
970 // If threading this would thread across a loop header, don't even try to
972 if (LoopHeaders.count(BB))
975 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues;
976 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues))
978 assert(!PredValues.empty() &&
979 "ComputeValueKnownInPredecessors returned true with no values");
981 DEBUG(errs() << "IN BB: " << *BB;
982 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
983 errs() << " BB '" << BB->getName() << "': FOUND condition = ";
984 if (PredValues[i].first)
985 errs() << *PredValues[i].first;
988 errs() << " for pred '" << PredValues[i].second->getName()
992 // Decide what we want to thread through. Convert our list of known values to
993 // a list of known destinations for each pred. This also discards duplicate
994 // predecessors and keeps track of the undefined inputs (which are represented
995 // as a null dest in the PredToDestList).
996 SmallPtrSet<BasicBlock*, 16> SeenPreds;
997 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
999 BasicBlock *OnlyDest = 0;
1000 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1002 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1003 BasicBlock *Pred = PredValues[i].second;
1004 if (!SeenPreds.insert(Pred))
1005 continue; // Duplicate predecessor entry.
1007 // If the predecessor ends with an indirect goto, we can't change its
1009 if (isa<IndirectBrInst>(Pred->getTerminator()))
1012 ConstantInt *Val = PredValues[i].first;
1015 if (Val == 0) // Undef.
1017 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1018 DestBB = BI->getSuccessor(Val->isZero());
1020 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
1021 DestBB = SI->getSuccessor(SI->findCaseValue(Val));
1024 // If we have exactly one destination, remember it for efficiency below.
1027 else if (OnlyDest != DestBB)
1028 OnlyDest = MultipleDestSentinel;
1030 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1033 // If all edges were unthreadable, we fail.
1034 if (PredToDestList.empty())
1037 // Determine which is the most common successor. If we have many inputs and
1038 // this block is a switch, we want to start by threading the batch that goes
1039 // to the most popular destination first. If we only know about one
1040 // threadable destination (the common case) we can avoid this.
1041 BasicBlock *MostPopularDest = OnlyDest;
1043 if (MostPopularDest == MultipleDestSentinel)
1044 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1046 // Now that we know what the most popular destination is, factor all
1047 // predecessors that will jump to it into a single predecessor.
1048 SmallVector<BasicBlock*, 16> PredsToFactor;
1049 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1050 if (PredToDestList[i].second == MostPopularDest) {
1051 BasicBlock *Pred = PredToDestList[i].first;
1053 // This predecessor may be a switch or something else that has multiple
1054 // edges to the block. Factor each of these edges by listing them
1055 // according to # occurrences in PredsToFactor.
1056 TerminatorInst *PredTI = Pred->getTerminator();
1057 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1058 if (PredTI->getSuccessor(i) == BB)
1059 PredsToFactor.push_back(Pred);
1062 // If the threadable edges are branching on an undefined value, we get to pick
1063 // the destination that these predecessors should get to.
1064 if (MostPopularDest == 0)
1065 MostPopularDest = BB->getTerminator()->
1066 getSuccessor(GetBestDestForJumpOnUndef(BB));
1068 // Ok, try to thread it!
1069 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1072 /// ProcessJumpOnPHI - We have a conditional branch or switch on a PHI node in
1073 /// the current block. See if there are any simplifications we can do based on
1074 /// inputs to the phi node.
1076 bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) {
1077 BasicBlock *BB = PN->getParent();
1079 // If any of the predecessor blocks end in an unconditional branch, we can
1080 // *duplicate* the jump into that block in order to further encourage jump
1081 // threading and to eliminate cases where we have branch on a phi of an icmp
1082 // (branch on icmp is much better).
1084 // We don't want to do this tranformation for switches, because we don't
1085 // really want to duplicate a switch.
1086 if (isa<SwitchInst>(BB->getTerminator()))
1089 // Look for unconditional branch predecessors.
1090 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1091 BasicBlock *PredBB = PN->getIncomingBlock(i);
1092 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1093 if (PredBr->isUnconditional() &&
1094 // Try to duplicate BB into PredBB.
1095 DuplicateCondBranchOnPHIIntoPred(BB, PredBB))
1103 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1104 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1105 /// NewPred using the entries from OldPred (suitably mapped).
1106 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1107 BasicBlock *OldPred,
1108 BasicBlock *NewPred,
1109 DenseMap<Instruction*, Value*> &ValueMap) {
1110 for (BasicBlock::iterator PNI = PHIBB->begin();
1111 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1112 // Ok, we have a PHI node. Figure out what the incoming value was for the
1114 Value *IV = PN->getIncomingValueForBlock(OldPred);
1116 // Remap the value if necessary.
1117 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1118 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1119 if (I != ValueMap.end())
1123 PN->addIncoming(IV, NewPred);
1127 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1128 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1129 /// across BB. Transform the IR to reflect this change.
1130 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1131 const SmallVectorImpl<BasicBlock*> &PredBBs,
1132 BasicBlock *SuccBB) {
1133 // If threading to the same block as we come from, we would infinite loop.
1135 DEBUG(errs() << " Not threading across BB '" << BB->getName()
1136 << "' - would thread to self!\n");
1140 // If threading this would thread across a loop header, don't thread the edge.
1141 // See the comments above FindLoopHeaders for justifications and caveats.
1142 if (LoopHeaders.count(BB)) {
1143 DEBUG(errs() << " Not threading across loop header BB '" << BB->getName()
1144 << "' to dest BB '" << SuccBB->getName()
1145 << "' - it might create an irreducible loop!\n");
1149 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1150 if (JumpThreadCost > Threshold) {
1151 DEBUG(errs() << " Not threading BB '" << BB->getName()
1152 << "' - Cost is too high: " << JumpThreadCost << "\n");
1156 // And finally, do it! Start by factoring the predecessors is needed.
1158 if (PredBBs.size() == 1)
1159 PredBB = PredBBs[0];
1161 DEBUG(errs() << " Factoring out " << PredBBs.size()
1162 << " common predecessors.\n");
1163 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1167 // And finally, do it!
1168 DEBUG(errs() << " Threading edge from '" << PredBB->getName() << "' to '"
1169 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1170 << ", across block:\n "
1173 // We are going to have to map operands from the original BB block to the new
1174 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1175 // account for entry from PredBB.
1176 DenseMap<Instruction*, Value*> ValueMapping;
1178 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1179 BB->getName()+".thread",
1180 BB->getParent(), BB);
1181 NewBB->moveAfter(PredBB);
1183 BasicBlock::iterator BI = BB->begin();
1184 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1185 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1187 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1188 // mapping and using it to remap operands in the cloned instructions.
1189 for (; !isa<TerminatorInst>(BI); ++BI) {
1190 Instruction *New = BI->clone();
1191 New->setName(BI->getName());
1192 NewBB->getInstList().push_back(New);
1193 ValueMapping[BI] = New;
1195 // Remap operands to patch up intra-block references.
1196 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1197 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1198 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1199 if (I != ValueMapping.end())
1200 New->setOperand(i, I->second);
1204 // We didn't copy the terminator from BB over to NewBB, because there is now
1205 // an unconditional jump to SuccBB. Insert the unconditional jump.
1206 BranchInst::Create(SuccBB, NewBB);
1208 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1209 // PHI nodes for NewBB now.
1210 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1212 // If there were values defined in BB that are used outside the block, then we
1213 // now have to update all uses of the value to use either the original value,
1214 // the cloned value, or some PHI derived value. This can require arbitrary
1215 // PHI insertion, of which we are prepared to do, clean these up now.
1216 SSAUpdater SSAUpdate;
1217 SmallVector<Use*, 16> UsesToRename;
1218 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1219 // Scan all uses of this instruction to see if it is used outside of its
1220 // block, and if so, record them in UsesToRename.
1221 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1223 Instruction *User = cast<Instruction>(*UI);
1224 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1225 if (UserPN->getIncomingBlock(UI) == BB)
1227 } else if (User->getParent() == BB)
1230 UsesToRename.push_back(&UI.getUse());
1233 // If there are no uses outside the block, we're done with this instruction.
1234 if (UsesToRename.empty())
1237 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n");
1239 // We found a use of I outside of BB. Rename all uses of I that are outside
1240 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1241 // with the two values we know.
1242 SSAUpdate.Initialize(I);
1243 SSAUpdate.AddAvailableValue(BB, I);
1244 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1246 while (!UsesToRename.empty())
1247 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1248 DEBUG(errs() << "\n");
1252 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1253 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1254 // us to simplify any PHI nodes in BB.
1255 TerminatorInst *PredTerm = PredBB->getTerminator();
1256 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1257 if (PredTerm->getSuccessor(i) == BB) {
1258 RemovePredecessorAndSimplify(BB, PredBB, TD);
1259 PredTerm->setSuccessor(i, NewBB);
1262 // At this point, the IR is fully up to date and consistent. Do a quick scan
1263 // over the new instructions and zap any that are constants or dead. This
1264 // frequently happens because of phi translation.
1265 BI = NewBB->begin();
1266 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) {
1267 Instruction *Inst = BI++;
1269 if (Value *V = SimplifyInstruction(Inst, TD)) {
1270 WeakVH BIHandle(BI);
1271 ReplaceAndSimplifyAllUses(Inst, V, TD);
1273 BI = NewBB->begin();
1277 RecursivelyDeleteTriviallyDeadInstructions(Inst);
1280 // Threaded an edge!
1285 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1286 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1287 /// If we can duplicate the contents of BB up into PredBB do so now, this
1288 /// improves the odds that the branch will be on an analyzable instruction like
1290 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1291 BasicBlock *PredBB) {
1292 // If BB is a loop header, then duplicating this block outside the loop would
1293 // cause us to transform this into an irreducible loop, don't do this.
1294 // See the comments above FindLoopHeaders for justifications and caveats.
1295 if (LoopHeaders.count(BB)) {
1296 DEBUG(errs() << " Not duplicating loop header '" << BB->getName()
1297 << "' into predecessor block '" << PredBB->getName()
1298 << "' - it might create an irreducible loop!\n");
1302 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1303 if (DuplicationCost > Threshold) {
1304 DEBUG(errs() << " Not duplicating BB '" << BB->getName()
1305 << "' - Cost is too high: " << DuplicationCost << "\n");
1309 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1311 DEBUG(errs() << " Duplicating block '" << BB->getName() << "' into end of '"
1312 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1313 << DuplicationCost << " block is:" << *BB << "\n");
1315 // We are going to have to map operands from the original BB block into the
1316 // PredBB block. Evaluate PHI nodes in BB.
1317 DenseMap<Instruction*, Value*> ValueMapping;
1319 BasicBlock::iterator BI = BB->begin();
1320 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1321 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1323 BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1325 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1326 // mapping and using it to remap operands in the cloned instructions.
1327 for (; BI != BB->end(); ++BI) {
1328 Instruction *New = BI->clone();
1329 New->setName(BI->getName());
1330 PredBB->getInstList().insert(OldPredBranch, New);
1331 ValueMapping[BI] = New;
1333 // Remap operands to patch up intra-block references.
1334 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1335 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1336 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1337 if (I != ValueMapping.end())
1338 New->setOperand(i, I->second);
1342 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1343 // add entries to the PHI nodes for branch from PredBB now.
1344 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1345 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1347 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1350 // If there were values defined in BB that are used outside the block, then we
1351 // now have to update all uses of the value to use either the original value,
1352 // the cloned value, or some PHI derived value. This can require arbitrary
1353 // PHI insertion, of which we are prepared to do, clean these up now.
1354 SSAUpdater SSAUpdate;
1355 SmallVector<Use*, 16> UsesToRename;
1356 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1357 // Scan all uses of this instruction to see if it is used outside of its
1358 // block, and if so, record them in UsesToRename.
1359 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1361 Instruction *User = cast<Instruction>(*UI);
1362 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1363 if (UserPN->getIncomingBlock(UI) == BB)
1365 } else if (User->getParent() == BB)
1368 UsesToRename.push_back(&UI.getUse());
1371 // If there are no uses outside the block, we're done with this instruction.
1372 if (UsesToRename.empty())
1375 DEBUG(errs() << "JT: Renaming non-local uses of: " << *I << "\n");
1377 // We found a use of I outside of BB. Rename all uses of I that are outside
1378 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1379 // with the two values we know.
1380 SSAUpdate.Initialize(I);
1381 SSAUpdate.AddAvailableValue(BB, I);
1382 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1384 while (!UsesToRename.empty())
1385 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1386 DEBUG(errs() << "\n");
1389 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1391 RemovePredecessorAndSimplify(BB, PredBB, TD);
1393 // Remove the unconditional branch at the end of the PredBB block.
1394 OldPredBranch->eraseFromParent();