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/ValueHandle.h"
33 #include "llvm/Support/raw_ostream.h"
36 STATISTIC(NumThreads, "Number of jumps threaded");
37 STATISTIC(NumFolds, "Number of terminators folded");
38 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
40 static cl::opt<unsigned>
41 Threshold("jump-threading-threshold",
42 cl::desc("Max block size to duplicate for jump threading"),
43 cl::init(6), cl::Hidden);
45 // Turn on use of LazyValueInfo.
47 EnableLVI("enable-jump-threading-lvi", cl::ReallyHidden);
52 /// This pass performs 'jump threading', which looks at blocks that have
53 /// multiple predecessors and multiple successors. If one or more of the
54 /// predecessors of the block can be proven to always jump to one of the
55 /// successors, we forward the edge from the predecessor to the successor by
56 /// duplicating the contents of this block.
58 /// An example of when this can occur is code like this:
65 /// In this case, the unconditional branch at the end of the first if can be
66 /// revectored to the false side of the second if.
68 class JumpThreading : public FunctionPass {
72 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
74 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
77 static char ID; // Pass identification
78 JumpThreading() : FunctionPass(&ID) {}
80 bool runOnFunction(Function &F);
82 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
84 AU.addRequired<LazyValueInfo>();
87 void FindLoopHeaders(Function &F);
88 bool ProcessBlock(BasicBlock *BB);
89 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
91 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
92 const SmallVectorImpl<BasicBlock *> &PredBBs);
94 typedef SmallVectorImpl<std::pair<ConstantInt*,
95 BasicBlock*> > PredValueInfo;
97 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
98 PredValueInfo &Result);
99 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB);
102 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
103 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
105 bool ProcessBranchOnPHI(PHINode *PN);
106 bool ProcessBranchOnXOR(BinaryOperator *BO);
108 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
112 char JumpThreading::ID = 0;
113 static RegisterPass<JumpThreading>
114 X("jump-threading", "Jump Threading");
116 // Public interface to the Jump Threading pass
117 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
119 /// runOnFunction - Top level algorithm.
121 bool JumpThreading::runOnFunction(Function &F) {
122 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
123 TD = getAnalysisIfAvailable<TargetData>();
124 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0;
128 bool Changed, EverChanged = false;
131 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
133 // Thread all of the branches we can over this block.
134 while (ProcessBlock(BB))
139 // If the block is trivially dead, zap it. This eliminates the successor
140 // edges which simplifies the CFG.
141 if (pred_begin(BB) == pred_end(BB) &&
142 BB != &BB->getParent()->getEntryBlock()) {
143 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
144 << "' with terminator: " << *BB->getTerminator() << '\n');
145 LoopHeaders.erase(BB);
148 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
149 // Can't thread an unconditional jump, but if the block is "almost
150 // empty", we can replace uses of it with uses of the successor and make
152 if (BI->isUnconditional() &&
153 BB != &BB->getParent()->getEntryBlock()) {
154 BasicBlock::iterator BBI = BB->getFirstNonPHI();
155 // Ignore dbg intrinsics.
156 while (isa<DbgInfoIntrinsic>(BBI))
158 // If the terminator is the only non-phi instruction, try to nuke it.
159 if (BBI->isTerminator()) {
160 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
161 // block, we have to make sure it isn't in the LoopHeaders set. We
162 // reinsert afterward if needed.
163 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
164 BasicBlock *Succ = BI->getSuccessor(0);
166 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
168 // If we deleted BB and BB was the header of a loop, then the
169 // successor is now the header of the loop.
173 if (ErasedFromLoopHeaders)
174 LoopHeaders.insert(BB);
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(dbgs() << " 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(dbgs() << " 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 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
556 (!isa<PHINode>(CondCmp->getOperand(0)) ||
557 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) {
558 // If we have a comparison, loop over the predecessors to see if there is
559 // a condition with a lexically identical value.
560 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
561 for (; PI != E; ++PI)
562 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
563 if (PBI->isConditional() && *PI != BB) {
564 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) {
565 if (CI->getOperand(0) == CondCmp->getOperand(0) &&
566 CI->getOperand(1) == CondCmp->getOperand(1) &&
567 CI->getPredicate() == CondCmp->getPredicate()) {
568 // TODO: Could handle things like (x != 4) --> (x == 17)
569 if (ProcessBranchOnDuplicateCond(*PI, BB))
577 // Check for some cases that are worth simplifying. Right now we want to look
578 // for loads that are used by a switch or by the condition for the branch. If
579 // we see one, check to see if it's partially redundant. If so, insert a PHI
580 // which can then be used to thread the values.
582 Value *SimplifyValue = CondInst;
583 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
584 if (isa<Constant>(CondCmp->getOperand(1)))
585 SimplifyValue = CondCmp->getOperand(0);
587 // TODO: There are other places where load PRE would be profitable, such as
588 // more complex comparisons.
589 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
590 if (SimplifyPartiallyRedundantLoad(LI))
594 // Handle a variety of cases where we are branching on something derived from
595 // a PHI node in the current block. If we can prove that any predecessors
596 // compute a predictable value based on a PHI node, thread those predecessors.
598 if (ProcessThreadableEdges(CondInst, BB))
601 // If this is an otherwise-unfoldable branch on a phi node in the current
602 // block, see if we can simplify.
603 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
604 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
605 return ProcessBranchOnPHI(PN);
608 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
609 if (CondInst->getOpcode() == Instruction::Xor &&
610 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
611 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
614 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
615 // "(X == 4)", thread through this block.
620 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
621 /// block that jump on exactly the same condition. This means that we almost
622 /// always know the direction of the edge in the DESTBB:
624 /// br COND, DESTBB, BBY
626 /// br COND, BBZ, BBW
628 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
629 /// in DESTBB, we have to thread over it.
630 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
632 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
634 // If both successors of PredBB go to DESTBB, we don't know anything. We can
635 // fold the branch to an unconditional one, which allows other recursive
638 if (PredBI->getSuccessor(1) != BB)
640 else if (PredBI->getSuccessor(0) != BB)
643 DEBUG(dbgs() << " In block '" << PredBB->getName()
644 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
646 ConstantFoldTerminator(PredBB);
650 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
652 // If the dest block has one predecessor, just fix the branch condition to a
653 // constant and fold it.
654 if (BB->getSinglePredecessor()) {
655 DEBUG(dbgs() << " In block '" << BB->getName()
656 << "' folding condition to '" << BranchDir << "': "
657 << *BB->getTerminator() << '\n');
659 Value *OldCond = DestBI->getCondition();
660 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
662 ConstantFoldTerminator(BB);
663 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
668 // Next, figure out which successor we are threading to.
669 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
671 SmallVector<BasicBlock*, 2> Preds;
672 Preds.push_back(PredBB);
674 // Ok, try to thread it!
675 return ThreadEdge(BB, Preds, SuccBB);
678 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
679 /// block that switch on exactly the same condition. This means that we almost
680 /// always know the direction of the edge in the DESTBB:
682 /// switch COND [... DESTBB, BBY ... ]
684 /// switch COND [... BBZ, BBW ]
686 /// Optimizing switches like this is very important, because simplifycfg builds
687 /// switches out of repeated 'if' conditions.
688 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
689 BasicBlock *DestBB) {
690 // Can't thread edge to self.
691 if (PredBB == DestBB)
694 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
695 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
697 // There are a variety of optimizations that we can potentially do on these
698 // blocks: we order them from most to least preferable.
700 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
701 // directly to their destination. This does not introduce *any* code size
702 // growth. Skip debug info first.
703 BasicBlock::iterator BBI = DestBB->begin();
704 while (isa<DbgInfoIntrinsic>(BBI))
707 // FIXME: Thread if it just contains a PHI.
708 if (isa<SwitchInst>(BBI)) {
709 bool MadeChange = false;
710 // Ignore the default edge for now.
711 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
712 ConstantInt *DestVal = DestSI->getCaseValue(i);
713 BasicBlock *DestSucc = DestSI->getSuccessor(i);
715 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
716 // PredSI has an explicit case for it. If so, forward. If it is covered
717 // by the default case, we can't update PredSI.
718 unsigned PredCase = PredSI->findCaseValue(DestVal);
719 if (PredCase == 0) continue;
721 // If PredSI doesn't go to DestBB on this value, then it won't reach the
722 // case on this condition.
723 if (PredSI->getSuccessor(PredCase) != DestBB &&
724 DestSI->getSuccessor(i) != DestBB)
727 // Do not forward this if it already goes to this destination, this would
728 // be an infinite loop.
729 if (PredSI->getSuccessor(PredCase) == DestSucc)
732 // Otherwise, we're safe to make the change. Make sure that the edge from
733 // DestSI to DestSucc is not critical and has no PHI nodes.
734 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
735 DEBUG(dbgs() << "THROUGH: " << *DestSI);
737 // If the destination has PHI nodes, just split the edge for updating
739 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
740 SplitCriticalEdge(DestSI, i, this);
741 DestSucc = DestSI->getSuccessor(i);
743 FoldSingleEntryPHINodes(DestSucc);
744 PredSI->setSuccessor(PredCase, DestSucc);
756 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
757 /// load instruction, eliminate it by replacing it with a PHI node. This is an
758 /// important optimization that encourages jump threading, and needs to be run
759 /// interlaced with other jump threading tasks.
760 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
761 // Don't hack volatile loads.
762 if (LI->isVolatile()) return false;
764 // If the load is defined in a block with exactly one predecessor, it can't be
765 // partially redundant.
766 BasicBlock *LoadBB = LI->getParent();
767 if (LoadBB->getSinglePredecessor())
770 Value *LoadedPtr = LI->getOperand(0);
772 // If the loaded operand is defined in the LoadBB, it can't be available.
773 // TODO: Could do simple PHI translation, that would be fun :)
774 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
775 if (PtrOp->getParent() == LoadBB)
778 // Scan a few instructions up from the load, to see if it is obviously live at
779 // the entry to its block.
780 BasicBlock::iterator BBIt = LI;
782 if (Value *AvailableVal =
783 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
784 // If the value if the load is locally available within the block, just use
785 // it. This frequently occurs for reg2mem'd allocas.
786 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
788 // If the returned value is the load itself, replace with an undef. This can
789 // only happen in dead loops.
790 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
791 LI->replaceAllUsesWith(AvailableVal);
792 LI->eraseFromParent();
796 // Otherwise, if we scanned the whole block and got to the top of the block,
797 // we know the block is locally transparent to the load. If not, something
798 // might clobber its value.
799 if (BBIt != LoadBB->begin())
803 SmallPtrSet<BasicBlock*, 8> PredsScanned;
804 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
805 AvailablePredsTy AvailablePreds;
806 BasicBlock *OneUnavailablePred = 0;
808 // If we got here, the loaded value is transparent through to the start of the
809 // block. Check to see if it is available in any of the predecessor blocks.
810 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
812 BasicBlock *PredBB = *PI;
814 // If we already scanned this predecessor, skip it.
815 if (!PredsScanned.insert(PredBB))
818 // Scan the predecessor to see if the value is available in the pred.
819 BBIt = PredBB->end();
820 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
821 if (!PredAvailable) {
822 OneUnavailablePred = PredBB;
826 // If so, this load is partially redundant. Remember this info so that we
827 // can create a PHI node.
828 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
831 // If the loaded value isn't available in any predecessor, it isn't partially
833 if (AvailablePreds.empty()) return false;
835 // Okay, the loaded value is available in at least one (and maybe all!)
836 // predecessors. If the value is unavailable in more than one unique
837 // predecessor, we want to insert a merge block for those common predecessors.
838 // This ensures that we only have to insert one reload, thus not increasing
840 BasicBlock *UnavailablePred = 0;
842 // If there is exactly one predecessor where the value is unavailable, the
843 // already computed 'OneUnavailablePred' block is it. If it ends in an
844 // unconditional branch, we know that it isn't a critical edge.
845 if (PredsScanned.size() == AvailablePreds.size()+1 &&
846 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
847 UnavailablePred = OneUnavailablePred;
848 } else if (PredsScanned.size() != AvailablePreds.size()) {
849 // Otherwise, we had multiple unavailable predecessors or we had a critical
850 // edge from the one.
851 SmallVector<BasicBlock*, 8> PredsToSplit;
852 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
854 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
855 AvailablePredSet.insert(AvailablePreds[i].first);
857 // Add all the unavailable predecessors to the PredsToSplit list.
858 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
860 if (!AvailablePredSet.count(*PI))
861 PredsToSplit.push_back(*PI);
863 // Split them out to their own block.
865 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
866 "thread-pre-split", this);
869 // If the value isn't available in all predecessors, then there will be
870 // exactly one where it isn't available. Insert a load on that edge and add
871 // it to the AvailablePreds list.
872 if (UnavailablePred) {
873 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
874 "Can't handle critical edge here!");
875 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
877 UnavailablePred->getTerminator());
878 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
881 // Now we know that each predecessor of this block has a value in
882 // AvailablePreds, sort them for efficient access as we're walking the preds.
883 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
885 // Create a PHI node at the start of the block for the PRE'd load value.
886 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
889 // Insert new entries into the PHI for each predecessor. A single block may
890 // have multiple entries here.
891 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
893 AvailablePredsTy::iterator I =
894 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
895 std::make_pair(*PI, (Value*)0));
897 assert(I != AvailablePreds.end() && I->first == *PI &&
898 "Didn't find entry for predecessor!");
900 PN->addIncoming(I->second, I->first);
903 //cerr << "PRE: " << *LI << *PN << "\n";
905 LI->replaceAllUsesWith(PN);
906 LI->eraseFromParent();
911 /// FindMostPopularDest - The specified list contains multiple possible
912 /// threadable destinations. Pick the one that occurs the most frequently in
915 FindMostPopularDest(BasicBlock *BB,
916 const SmallVectorImpl<std::pair<BasicBlock*,
917 BasicBlock*> > &PredToDestList) {
918 assert(!PredToDestList.empty());
920 // Determine popularity. If there are multiple possible destinations, we
921 // explicitly choose to ignore 'undef' destinations. We prefer to thread
922 // blocks with known and real destinations to threading undef. We'll handle
923 // them later if interesting.
924 DenseMap<BasicBlock*, unsigned> DestPopularity;
925 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
926 if (PredToDestList[i].second)
927 DestPopularity[PredToDestList[i].second]++;
929 // Find the most popular dest.
930 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
931 BasicBlock *MostPopularDest = DPI->first;
932 unsigned Popularity = DPI->second;
933 SmallVector<BasicBlock*, 4> SamePopularity;
935 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
936 // If the popularity of this entry isn't higher than the popularity we've
937 // seen so far, ignore it.
938 if (DPI->second < Popularity)
940 else if (DPI->second == Popularity) {
941 // If it is the same as what we've seen so far, keep track of it.
942 SamePopularity.push_back(DPI->first);
944 // If it is more popular, remember it.
945 SamePopularity.clear();
946 MostPopularDest = DPI->first;
947 Popularity = DPI->second;
951 // Okay, now we know the most popular destination. If there is more than
952 // destination, we need to determine one. This is arbitrary, but we need
953 // to make a deterministic decision. Pick the first one that appears in the
955 if (!SamePopularity.empty()) {
956 SamePopularity.push_back(MostPopularDest);
957 TerminatorInst *TI = BB->getTerminator();
958 for (unsigned i = 0; ; ++i) {
959 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
961 if (std::find(SamePopularity.begin(), SamePopularity.end(),
962 TI->getSuccessor(i)) == SamePopularity.end())
965 MostPopularDest = TI->getSuccessor(i);
970 // Okay, we have finally picked the most popular destination.
971 return MostPopularDest;
974 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) {
975 // If threading this would thread across a loop header, don't even try to
977 if (LoopHeaders.count(BB))
980 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues;
981 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues))
983 assert(!PredValues.empty() &&
984 "ComputeValueKnownInPredecessors returned true with no values");
986 DEBUG(dbgs() << "IN BB: " << *BB;
987 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
988 dbgs() << " BB '" << BB->getName() << "': FOUND condition = ";
989 if (PredValues[i].first)
990 dbgs() << *PredValues[i].first;
993 dbgs() << " for pred '" << PredValues[i].second->getName()
997 // Decide what we want to thread through. Convert our list of known values to
998 // a list of known destinations for each pred. This also discards duplicate
999 // predecessors and keeps track of the undefined inputs (which are represented
1000 // as a null dest in the PredToDestList).
1001 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1002 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1004 BasicBlock *OnlyDest = 0;
1005 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1007 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1008 BasicBlock *Pred = PredValues[i].second;
1009 if (!SeenPreds.insert(Pred))
1010 continue; // Duplicate predecessor entry.
1012 // If the predecessor ends with an indirect goto, we can't change its
1014 if (isa<IndirectBrInst>(Pred->getTerminator()))
1017 ConstantInt *Val = PredValues[i].first;
1020 if (Val == 0) // Undef.
1022 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1023 DestBB = BI->getSuccessor(Val->isZero());
1025 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
1026 DestBB = SI->getSuccessor(SI->findCaseValue(Val));
1029 // If we have exactly one destination, remember it for efficiency below.
1032 else if (OnlyDest != DestBB)
1033 OnlyDest = MultipleDestSentinel;
1035 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1038 // If all edges were unthreadable, we fail.
1039 if (PredToDestList.empty())
1042 // Determine which is the most common successor. If we have many inputs and
1043 // this block is a switch, we want to start by threading the batch that goes
1044 // to the most popular destination first. If we only know about one
1045 // threadable destination (the common case) we can avoid this.
1046 BasicBlock *MostPopularDest = OnlyDest;
1048 if (MostPopularDest == MultipleDestSentinel)
1049 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1051 // Now that we know what the most popular destination is, factor all
1052 // predecessors that will jump to it into a single predecessor.
1053 SmallVector<BasicBlock*, 16> PredsToFactor;
1054 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1055 if (PredToDestList[i].second == MostPopularDest) {
1056 BasicBlock *Pred = PredToDestList[i].first;
1058 // This predecessor may be a switch or something else that has multiple
1059 // edges to the block. Factor each of these edges by listing them
1060 // according to # occurrences in PredsToFactor.
1061 TerminatorInst *PredTI = Pred->getTerminator();
1062 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1063 if (PredTI->getSuccessor(i) == BB)
1064 PredsToFactor.push_back(Pred);
1067 // If the threadable edges are branching on an undefined value, we get to pick
1068 // the destination that these predecessors should get to.
1069 if (MostPopularDest == 0)
1070 MostPopularDest = BB->getTerminator()->
1071 getSuccessor(GetBestDestForJumpOnUndef(BB));
1073 // Ok, try to thread it!
1074 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1077 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1078 /// a PHI node in the current block. See if there are any simplifications we
1079 /// can do based on inputs to the phi node.
1081 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1082 BasicBlock *BB = PN->getParent();
1084 // TODO: We could make use of this to do it once for blocks with common PHI
1086 SmallVector<BasicBlock*, 1> PredBBs;
1089 // If any of the predecessor blocks end in an unconditional branch, we can
1090 // *duplicate* the conditional branch into that block in order to further
1091 // encourage jump threading and to eliminate cases where we have branch on a
1092 // phi of an icmp (branch on icmp is much better).
1093 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1094 BasicBlock *PredBB = PN->getIncomingBlock(i);
1095 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1096 if (PredBr->isUnconditional()) {
1097 PredBBs[0] = PredBB;
1098 // Try to duplicate BB into PredBB.
1099 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1107 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1108 /// a xor instruction in the current block. See if there are any
1109 /// simplifications we can do based on inputs to the xor.
1111 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1112 BasicBlock *BB = BO->getParent();
1114 // If either the LHS or RHS of the xor is a constant, don't do this
1116 if (isa<ConstantInt>(BO->getOperand(0)) ||
1117 isa<ConstantInt>(BO->getOperand(1)))
1120 // If we have a xor as the branch input to this block, and we know that the
1121 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1122 // the condition into the predecessor and fix that value to true, saving some
1123 // logical ops on that path and encouraging other paths to simplify.
1125 // This copies something like this:
1128 // %X = phi i1 [1], [%X']
1129 // %Y = icmp eq i32 %A, %B
1130 // %Z = xor i1 %X, %Y
1135 // %Y = icmp ne i32 %A, %B
1138 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues;
1140 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) {
1141 assert(XorOpValues.empty());
1142 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues))
1147 assert(!XorOpValues.empty() &&
1148 "ComputeValueKnownInPredecessors returned true with no values");
1150 // Scan the information to see which is most popular: true or false. The
1151 // predecessors can be of the set true, false, or undef.
1152 unsigned NumTrue = 0, NumFalse = 0;
1153 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1154 if (!XorOpValues[i].first) continue; // Ignore undefs for the count.
1155 if (XorOpValues[i].first->isZero())
1161 // Determine which value to split on, true, false, or undef if neither.
1162 ConstantInt *SplitVal = 0;
1163 if (NumTrue > NumFalse)
1164 SplitVal = ConstantInt::getTrue(BB->getContext());
1165 else if (NumTrue != 0 || NumFalse != 0)
1166 SplitVal = ConstantInt::getFalse(BB->getContext());
1168 // Collect all of the blocks that this can be folded into so that we can
1169 // factor this once and clone it once.
1170 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1171 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1172 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue;
1174 BlocksToFoldInto.push_back(XorOpValues[i].second);
1177 // Try to duplicate BB into PredBB.
1178 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1182 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1183 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1184 /// NewPred using the entries from OldPred (suitably mapped).
1185 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1186 BasicBlock *OldPred,
1187 BasicBlock *NewPred,
1188 DenseMap<Instruction*, Value*> &ValueMap) {
1189 for (BasicBlock::iterator PNI = PHIBB->begin();
1190 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1191 // Ok, we have a PHI node. Figure out what the incoming value was for the
1193 Value *IV = PN->getIncomingValueForBlock(OldPred);
1195 // Remap the value if necessary.
1196 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1197 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1198 if (I != ValueMap.end())
1202 PN->addIncoming(IV, NewPred);
1206 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1207 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1208 /// across BB. Transform the IR to reflect this change.
1209 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1210 const SmallVectorImpl<BasicBlock*> &PredBBs,
1211 BasicBlock *SuccBB) {
1212 // If threading to the same block as we come from, we would infinite loop.
1214 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1215 << "' - would thread to self!\n");
1219 // If threading this would thread across a loop header, don't thread the edge.
1220 // See the comments above FindLoopHeaders for justifications and caveats.
1221 if (LoopHeaders.count(BB)) {
1222 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1223 << "' to dest BB '" << SuccBB->getName()
1224 << "' - it might create an irreducible loop!\n");
1228 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1229 if (JumpThreadCost > Threshold) {
1230 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1231 << "' - Cost is too high: " << JumpThreadCost << "\n");
1235 // And finally, do it! Start by factoring the predecessors is needed.
1237 if (PredBBs.size() == 1)
1238 PredBB = PredBBs[0];
1240 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1241 << " common predecessors.\n");
1242 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1246 // And finally, do it!
1247 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1248 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1249 << ", across block:\n "
1252 // We are going to have to map operands from the original BB block to the new
1253 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1254 // account for entry from PredBB.
1255 DenseMap<Instruction*, Value*> ValueMapping;
1257 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1258 BB->getName()+".thread",
1259 BB->getParent(), BB);
1260 NewBB->moveAfter(PredBB);
1262 BasicBlock::iterator BI = BB->begin();
1263 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1264 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1266 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1267 // mapping and using it to remap operands in the cloned instructions.
1268 for (; !isa<TerminatorInst>(BI); ++BI) {
1269 Instruction *New = BI->clone();
1270 New->setName(BI->getName());
1271 NewBB->getInstList().push_back(New);
1272 ValueMapping[BI] = New;
1274 // Remap operands to patch up intra-block references.
1275 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1276 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1277 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1278 if (I != ValueMapping.end())
1279 New->setOperand(i, I->second);
1283 // We didn't copy the terminator from BB over to NewBB, because there is now
1284 // an unconditional jump to SuccBB. Insert the unconditional jump.
1285 BranchInst::Create(SuccBB, NewBB);
1287 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1288 // PHI nodes for NewBB now.
1289 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1291 // If there were values defined in BB that are used outside the block, then we
1292 // now have to update all uses of the value to use either the original value,
1293 // the cloned value, or some PHI derived value. This can require arbitrary
1294 // PHI insertion, of which we are prepared to do, clean these up now.
1295 SSAUpdater SSAUpdate;
1296 SmallVector<Use*, 16> UsesToRename;
1297 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1298 // Scan all uses of this instruction to see if it is used outside of its
1299 // block, and if so, record them in UsesToRename.
1300 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1302 Instruction *User = cast<Instruction>(*UI);
1303 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1304 if (UserPN->getIncomingBlock(UI) == BB)
1306 } else if (User->getParent() == BB)
1309 UsesToRename.push_back(&UI.getUse());
1312 // If there are no uses outside the block, we're done with this instruction.
1313 if (UsesToRename.empty())
1316 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1318 // We found a use of I outside of BB. Rename all uses of I that are outside
1319 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1320 // with the two values we know.
1321 SSAUpdate.Initialize(I);
1322 SSAUpdate.AddAvailableValue(BB, I);
1323 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1325 while (!UsesToRename.empty())
1326 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1327 DEBUG(dbgs() << "\n");
1331 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1332 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1333 // us to simplify any PHI nodes in BB.
1334 TerminatorInst *PredTerm = PredBB->getTerminator();
1335 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1336 if (PredTerm->getSuccessor(i) == BB) {
1337 RemovePredecessorAndSimplify(BB, PredBB, TD);
1338 PredTerm->setSuccessor(i, NewBB);
1341 // At this point, the IR is fully up to date and consistent. Do a quick scan
1342 // over the new instructions and zap any that are constants or dead. This
1343 // frequently happens because of phi translation.
1344 SimplifyInstructionsInBlock(NewBB, TD);
1346 // Threaded an edge!
1351 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1352 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1353 /// If we can duplicate the contents of BB up into PredBB do so now, this
1354 /// improves the odds that the branch will be on an analyzable instruction like
1356 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1357 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1358 assert(!PredBBs.empty() && "Can't handle an empty set");
1360 // If BB is a loop header, then duplicating this block outside the loop would
1361 // cause us to transform this into an irreducible loop, don't do this.
1362 // See the comments above FindLoopHeaders for justifications and caveats.
1363 if (LoopHeaders.count(BB)) {
1364 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1365 << "' into predecessor block '" << PredBBs[0]->getName()
1366 << "' - it might create an irreducible loop!\n");
1370 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1371 if (DuplicationCost > Threshold) {
1372 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1373 << "' - Cost is too high: " << DuplicationCost << "\n");
1377 // And finally, do it! Start by factoring the predecessors is needed.
1379 if (PredBBs.size() == 1)
1380 PredBB = PredBBs[0];
1382 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1383 << " common predecessors.\n");
1384 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1388 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1390 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1391 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1392 << DuplicationCost << " block is:" << *BB << "\n");
1394 // Unless PredBB ends with an unconditional branch, split the edge so that we
1395 // can just clone the bits from BB into the end of the new PredBB.
1396 BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1398 if (!OldPredBranch->isUnconditional()) {
1399 PredBB = SplitEdge(PredBB, BB, this);
1400 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1403 // We are going to have to map operands from the original BB block into the
1404 // PredBB block. Evaluate PHI nodes in BB.
1405 DenseMap<Instruction*, Value*> ValueMapping;
1407 BasicBlock::iterator BI = BB->begin();
1408 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1409 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1411 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1412 // mapping and using it to remap operands in the cloned instructions.
1413 for (; BI != BB->end(); ++BI) {
1414 Instruction *New = BI->clone();
1416 // Remap operands to patch up intra-block references.
1417 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1418 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1419 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1420 if (I != ValueMapping.end())
1421 New->setOperand(i, I->second);
1424 // If this instruction can be simplified after the operands are updated,
1425 // just use the simplified value instead. This frequently happens due to
1427 if (Value *IV = SimplifyInstruction(New, TD)) {
1429 ValueMapping[BI] = IV;
1431 // Otherwise, insert the new instruction into the block.
1432 New->setName(BI->getName());
1433 PredBB->getInstList().insert(OldPredBranch, New);
1434 ValueMapping[BI] = New;
1438 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1439 // add entries to the PHI nodes for branch from PredBB now.
1440 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1441 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1443 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1446 // If there were values defined in BB that are used outside the block, then we
1447 // now have to update all uses of the value to use either the original value,
1448 // the cloned value, or some PHI derived value. This can require arbitrary
1449 // PHI insertion, of which we are prepared to do, clean these up now.
1450 SSAUpdater SSAUpdate;
1451 SmallVector<Use*, 16> UsesToRename;
1452 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1453 // Scan all uses of this instruction to see if it is used outside of its
1454 // block, and if so, record them in UsesToRename.
1455 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1457 Instruction *User = cast<Instruction>(*UI);
1458 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1459 if (UserPN->getIncomingBlock(UI) == BB)
1461 } else if (User->getParent() == BB)
1464 UsesToRename.push_back(&UI.getUse());
1467 // If there are no uses outside the block, we're done with this instruction.
1468 if (UsesToRename.empty())
1471 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1473 // We found a use of I outside of BB. Rename all uses of I that are outside
1474 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1475 // with the two values we know.
1476 SSAUpdate.Initialize(I);
1477 SSAUpdate.AddAvailableValue(BB, I);
1478 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1480 while (!UsesToRename.empty())
1481 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1482 DEBUG(dbgs() << "\n");
1485 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1487 RemovePredecessorAndSimplify(BB, PredBB, TD);
1489 // Remove the unconditional branch at the end of the PredBB block.
1490 OldPredBranch->eraseFromParent();