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) && I->getType()->isPointerTy())
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 (!CI->getType()->isVectorTy())
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. If we find an undef, force it to the
340 // interesting value: x|undef -> true and x&undef -> false.
341 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
342 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) {
343 Result.push_back(LHSVals[i]);
344 Result.back().first = InterestingVal;
346 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
347 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) {
348 Result.push_back(RHSVals[i]);
349 Result.back().first = InterestingVal;
351 return !Result.empty();
354 // Handle the NOT form of XOR.
355 if (I->getOpcode() == Instruction::Xor &&
356 isa<ConstantInt>(I->getOperand(1)) &&
357 cast<ConstantInt>(I->getOperand(1))->isOne()) {
358 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result);
362 // Invert the known values.
363 for (unsigned i = 0, e = Result.size(); i != e; ++i)
366 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first));
371 // Handle compare with phi operand, where the PHI is defined in this block.
372 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
373 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
374 if (PN && PN->getParent() == BB) {
375 // We can do this simplification if any comparisons fold to true or false.
377 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
378 BasicBlock *PredBB = PN->getIncomingBlock(i);
379 Value *LHS = PN->getIncomingValue(i);
380 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
382 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
384 if (!LVI || !isa<Constant>(RHS))
387 LazyValueInfo::Tristate
388 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
389 cast<Constant>(RHS), PredBB, BB);
390 if (ResT == LazyValueInfo::Unknown)
392 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
395 if (isa<UndefValue>(Res))
396 Result.push_back(std::make_pair((ConstantInt*)0, PredBB));
397 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res))
398 Result.push_back(std::make_pair(CI, PredBB));
401 return !Result.empty();
405 // If comparing a live-in value against a constant, see if we know the
406 // live-in value on any predecessors.
407 if (LVI && isa<Constant>(Cmp->getOperand(1)) &&
408 Cmp->getType()->isIntegerTy() && // Not vector compare.
409 (!isa<Instruction>(Cmp->getOperand(0)) ||
410 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB)) {
411 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
413 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
414 // If the value is known by LazyValueInfo to be a constant in a
415 // predecessor, use that information to try to thread this block.
416 LazyValueInfo::Tristate
417 Res = LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
419 if (Res == LazyValueInfo::Unknown)
422 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
423 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), *PI));
426 return !Result.empty();
434 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
435 /// in an undefined jump, decide which block is best to revector to.
437 /// Since we can pick an arbitrary destination, we pick the successor with the
438 /// fewest predecessors. This should reduce the in-degree of the others.
440 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
441 TerminatorInst *BBTerm = BB->getTerminator();
442 unsigned MinSucc = 0;
443 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
444 // Compute the successor with the minimum number of predecessors.
445 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
446 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
447 TestBB = BBTerm->getSuccessor(i);
448 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
449 if (NumPreds < MinNumPreds)
456 /// ProcessBlock - If there are any predecessors whose control can be threaded
457 /// through to a successor, transform them now.
458 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
459 // If the block is trivially dead, just return and let the caller nuke it.
460 // This simplifies other transformations.
461 if (pred_begin(BB) == pred_end(BB) &&
462 BB != &BB->getParent()->getEntryBlock())
465 // If this block has a single predecessor, and if that pred has a single
466 // successor, merge the blocks. This encourages recursive jump threading
467 // because now the condition in this block can be threaded through
468 // predecessors of our predecessor block.
469 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
470 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
472 // If SinglePred was a loop header, BB becomes one.
473 if (LoopHeaders.erase(SinglePred))
474 LoopHeaders.insert(BB);
476 // Remember if SinglePred was the entry block of the function. If so, we
477 // will need to move BB back to the entry position.
478 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
479 MergeBasicBlockIntoOnlyPred(BB);
481 if (isEntry && BB != &BB->getParent()->getEntryBlock())
482 BB->moveBefore(&BB->getParent()->getEntryBlock());
487 // Look to see if the terminator is a branch of switch, if not we can't thread
490 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
491 // Can't thread an unconditional jump.
492 if (BI->isUnconditional()) return false;
493 Condition = BI->getCondition();
494 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
495 Condition = SI->getCondition();
497 return false; // Must be an invoke.
499 // If the terminator of this block is branching on a constant, simplify the
500 // terminator to an unconditional branch. This can occur due to threading in
502 if (isa<ConstantInt>(Condition)) {
503 DEBUG(dbgs() << " In block '" << BB->getName()
504 << "' folding terminator: " << *BB->getTerminator() << '\n');
506 ConstantFoldTerminator(BB);
510 // If the terminator is branching on an undef, we can pick any of the
511 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
512 if (isa<UndefValue>(Condition)) {
513 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
515 // Fold the branch/switch.
516 TerminatorInst *BBTerm = BB->getTerminator();
517 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
518 if (i == BestSucc) continue;
519 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD);
522 DEBUG(dbgs() << " In block '" << BB->getName()
523 << "' folding undef terminator: " << *BBTerm << '\n');
524 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
525 BBTerm->eraseFromParent();
529 Instruction *CondInst = dyn_cast<Instruction>(Condition);
531 // If the condition is an instruction defined in another block, see if a
532 // predecessor has the same condition:
537 !Condition->hasOneUse() && // Multiple uses.
538 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition.
539 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
540 if (isa<BranchInst>(BB->getTerminator())) {
541 for (; PI != E; ++PI)
542 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
543 if (PBI->isConditional() && PBI->getCondition() == Condition &&
544 ProcessBranchOnDuplicateCond(*PI, BB))
547 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator");
548 for (; PI != E; ++PI)
549 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator()))
550 if (PSI->getCondition() == Condition &&
551 ProcessSwitchOnDuplicateCond(*PI, BB))
556 // All the rest of our checks depend on the condition being an instruction.
558 // FIXME: Unify this with code below.
559 if (LVI && ProcessThreadableEdges(Condition, BB))
565 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
567 (!isa<PHINode>(CondCmp->getOperand(0)) ||
568 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) {
569 // If we have a comparison, loop over the predecessors to see if there is
570 // a condition with a lexically identical value.
571 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
572 for (; PI != E; ++PI)
573 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
574 if (PBI->isConditional() && *PI != BB) {
575 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) {
576 if (CI->getOperand(0) == CondCmp->getOperand(0) &&
577 CI->getOperand(1) == CondCmp->getOperand(1) &&
578 CI->getPredicate() == CondCmp->getPredicate()) {
579 // TODO: Could handle things like (x != 4) --> (x == 17)
580 if (ProcessBranchOnDuplicateCond(*PI, BB))
588 // Check for some cases that are worth simplifying. Right now we want to look
589 // for loads that are used by a switch or by the condition for the branch. If
590 // we see one, check to see if it's partially redundant. If so, insert a PHI
591 // which can then be used to thread the values.
593 Value *SimplifyValue = CondInst;
594 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
595 if (isa<Constant>(CondCmp->getOperand(1)))
596 SimplifyValue = CondCmp->getOperand(0);
598 // TODO: There are other places where load PRE would be profitable, such as
599 // more complex comparisons.
600 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
601 if (SimplifyPartiallyRedundantLoad(LI))
605 // Handle a variety of cases where we are branching on something derived from
606 // a PHI node in the current block. If we can prove that any predecessors
607 // compute a predictable value based on a PHI node, thread those predecessors.
609 if (ProcessThreadableEdges(CondInst, BB))
612 // If this is an otherwise-unfoldable branch on a phi node in the current
613 // block, see if we can simplify.
614 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
615 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
616 return ProcessBranchOnPHI(PN);
619 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
620 if (CondInst->getOpcode() == Instruction::Xor &&
621 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
622 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
625 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
626 // "(X == 4)", thread through this block.
631 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
632 /// block that jump on exactly the same condition. This means that we almost
633 /// always know the direction of the edge in the DESTBB:
635 /// br COND, DESTBB, BBY
637 /// br COND, BBZ, BBW
639 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
640 /// in DESTBB, we have to thread over it.
641 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
643 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
645 // If both successors of PredBB go to DESTBB, we don't know anything. We can
646 // fold the branch to an unconditional one, which allows other recursive
649 if (PredBI->getSuccessor(1) != BB)
651 else if (PredBI->getSuccessor(0) != BB)
654 DEBUG(dbgs() << " In block '" << PredBB->getName()
655 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
657 ConstantFoldTerminator(PredBB);
661 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
663 // If the dest block has one predecessor, just fix the branch condition to a
664 // constant and fold it.
665 if (BB->getSinglePredecessor()) {
666 DEBUG(dbgs() << " In block '" << BB->getName()
667 << "' folding condition to '" << BranchDir << "': "
668 << *BB->getTerminator() << '\n');
670 Value *OldCond = DestBI->getCondition();
671 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
673 // Delete dead instructions before we fold the branch. Folding the branch
674 // can eliminate edges from the CFG which can end up deleting OldCond.
675 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
676 ConstantFoldTerminator(BB);
681 // Next, figure out which successor we are threading to.
682 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
684 SmallVector<BasicBlock*, 2> Preds;
685 Preds.push_back(PredBB);
687 // Ok, try to thread it!
688 return ThreadEdge(BB, Preds, SuccBB);
691 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
692 /// block that switch on exactly the same condition. This means that we almost
693 /// always know the direction of the edge in the DESTBB:
695 /// switch COND [... DESTBB, BBY ... ]
697 /// switch COND [... BBZ, BBW ]
699 /// Optimizing switches like this is very important, because simplifycfg builds
700 /// switches out of repeated 'if' conditions.
701 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
702 BasicBlock *DestBB) {
703 // Can't thread edge to self.
704 if (PredBB == DestBB)
707 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
708 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
710 // There are a variety of optimizations that we can potentially do on these
711 // blocks: we order them from most to least preferable.
713 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
714 // directly to their destination. This does not introduce *any* code size
715 // growth. Skip debug info first.
716 BasicBlock::iterator BBI = DestBB->begin();
717 while (isa<DbgInfoIntrinsic>(BBI))
720 // FIXME: Thread if it just contains a PHI.
721 if (isa<SwitchInst>(BBI)) {
722 bool MadeChange = false;
723 // Ignore the default edge for now.
724 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
725 ConstantInt *DestVal = DestSI->getCaseValue(i);
726 BasicBlock *DestSucc = DestSI->getSuccessor(i);
728 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
729 // PredSI has an explicit case for it. If so, forward. If it is covered
730 // by the default case, we can't update PredSI.
731 unsigned PredCase = PredSI->findCaseValue(DestVal);
732 if (PredCase == 0) continue;
734 // If PredSI doesn't go to DestBB on this value, then it won't reach the
735 // case on this condition.
736 if (PredSI->getSuccessor(PredCase) != DestBB &&
737 DestSI->getSuccessor(i) != DestBB)
740 // Do not forward this if it already goes to this destination, this would
741 // be an infinite loop.
742 if (PredSI->getSuccessor(PredCase) == DestSucc)
745 // Otherwise, we're safe to make the change. Make sure that the edge from
746 // DestSI to DestSucc is not critical and has no PHI nodes.
747 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
748 DEBUG(dbgs() << "THROUGH: " << *DestSI);
750 // If the destination has PHI nodes, just split the edge for updating
752 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
753 SplitCriticalEdge(DestSI, i, this);
754 DestSucc = DestSI->getSuccessor(i);
756 FoldSingleEntryPHINodes(DestSucc);
757 PredSI->setSuccessor(PredCase, DestSucc);
769 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
770 /// load instruction, eliminate it by replacing it with a PHI node. This is an
771 /// important optimization that encourages jump threading, and needs to be run
772 /// interlaced with other jump threading tasks.
773 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
774 // Don't hack volatile loads.
775 if (LI->isVolatile()) return false;
777 // If the load is defined in a block with exactly one predecessor, it can't be
778 // partially redundant.
779 BasicBlock *LoadBB = LI->getParent();
780 if (LoadBB->getSinglePredecessor())
783 Value *LoadedPtr = LI->getOperand(0);
785 // If the loaded operand is defined in the LoadBB, it can't be available.
786 // TODO: Could do simple PHI translation, that would be fun :)
787 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
788 if (PtrOp->getParent() == LoadBB)
791 // Scan a few instructions up from the load, to see if it is obviously live at
792 // the entry to its block.
793 BasicBlock::iterator BBIt = LI;
795 if (Value *AvailableVal =
796 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
797 // If the value if the load is locally available within the block, just use
798 // it. This frequently occurs for reg2mem'd allocas.
799 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
801 // If the returned value is the load itself, replace with an undef. This can
802 // only happen in dead loops.
803 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
804 LI->replaceAllUsesWith(AvailableVal);
805 LI->eraseFromParent();
809 // Otherwise, if we scanned the whole block and got to the top of the block,
810 // we know the block is locally transparent to the load. If not, something
811 // might clobber its value.
812 if (BBIt != LoadBB->begin())
816 SmallPtrSet<BasicBlock*, 8> PredsScanned;
817 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
818 AvailablePredsTy AvailablePreds;
819 BasicBlock *OneUnavailablePred = 0;
821 // If we got here, the loaded value is transparent through to the start of the
822 // block. Check to see if it is available in any of the predecessor blocks.
823 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
825 BasicBlock *PredBB = *PI;
827 // If we already scanned this predecessor, skip it.
828 if (!PredsScanned.insert(PredBB))
831 // Scan the predecessor to see if the value is available in the pred.
832 BBIt = PredBB->end();
833 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
834 if (!PredAvailable) {
835 OneUnavailablePred = PredBB;
839 // If so, this load is partially redundant. Remember this info so that we
840 // can create a PHI node.
841 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
844 // If the loaded value isn't available in any predecessor, it isn't partially
846 if (AvailablePreds.empty()) return false;
848 // Okay, the loaded value is available in at least one (and maybe all!)
849 // predecessors. If the value is unavailable in more than one unique
850 // predecessor, we want to insert a merge block for those common predecessors.
851 // This ensures that we only have to insert one reload, thus not increasing
853 BasicBlock *UnavailablePred = 0;
855 // If there is exactly one predecessor where the value is unavailable, the
856 // already computed 'OneUnavailablePred' block is it. If it ends in an
857 // unconditional branch, we know that it isn't a critical edge.
858 if (PredsScanned.size() == AvailablePreds.size()+1 &&
859 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
860 UnavailablePred = OneUnavailablePred;
861 } else if (PredsScanned.size() != AvailablePreds.size()) {
862 // Otherwise, we had multiple unavailable predecessors or we had a critical
863 // edge from the one.
864 SmallVector<BasicBlock*, 8> PredsToSplit;
865 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
867 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
868 AvailablePredSet.insert(AvailablePreds[i].first);
870 // Add all the unavailable predecessors to the PredsToSplit list.
871 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
873 if (!AvailablePredSet.count(*PI))
874 PredsToSplit.push_back(*PI);
876 // Split them out to their own block.
878 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
879 "thread-pre-split", this);
882 // If the value isn't available in all predecessors, then there will be
883 // exactly one where it isn't available. Insert a load on that edge and add
884 // it to the AvailablePreds list.
885 if (UnavailablePred) {
886 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
887 "Can't handle critical edge here!");
888 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
890 UnavailablePred->getTerminator());
891 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
894 // Now we know that each predecessor of this block has a value in
895 // AvailablePreds, sort them for efficient access as we're walking the preds.
896 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
898 // Create a PHI node at the start of the block for the PRE'd load value.
899 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
902 // Insert new entries into the PHI for each predecessor. A single block may
903 // have multiple entries here.
904 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
906 AvailablePredsTy::iterator I =
907 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
908 std::make_pair(*PI, (Value*)0));
910 assert(I != AvailablePreds.end() && I->first == *PI &&
911 "Didn't find entry for predecessor!");
913 PN->addIncoming(I->second, I->first);
916 //cerr << "PRE: " << *LI << *PN << "\n";
918 LI->replaceAllUsesWith(PN);
919 LI->eraseFromParent();
924 /// FindMostPopularDest - The specified list contains multiple possible
925 /// threadable destinations. Pick the one that occurs the most frequently in
928 FindMostPopularDest(BasicBlock *BB,
929 const SmallVectorImpl<std::pair<BasicBlock*,
930 BasicBlock*> > &PredToDestList) {
931 assert(!PredToDestList.empty());
933 // Determine popularity. If there are multiple possible destinations, we
934 // explicitly choose to ignore 'undef' destinations. We prefer to thread
935 // blocks with known and real destinations to threading undef. We'll handle
936 // them later if interesting.
937 DenseMap<BasicBlock*, unsigned> DestPopularity;
938 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
939 if (PredToDestList[i].second)
940 DestPopularity[PredToDestList[i].second]++;
942 // Find the most popular dest.
943 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
944 BasicBlock *MostPopularDest = DPI->first;
945 unsigned Popularity = DPI->second;
946 SmallVector<BasicBlock*, 4> SamePopularity;
948 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
949 // If the popularity of this entry isn't higher than the popularity we've
950 // seen so far, ignore it.
951 if (DPI->second < Popularity)
953 else if (DPI->second == Popularity) {
954 // If it is the same as what we've seen so far, keep track of it.
955 SamePopularity.push_back(DPI->first);
957 // If it is more popular, remember it.
958 SamePopularity.clear();
959 MostPopularDest = DPI->first;
960 Popularity = DPI->second;
964 // Okay, now we know the most popular destination. If there is more than
965 // destination, we need to determine one. This is arbitrary, but we need
966 // to make a deterministic decision. Pick the first one that appears in the
968 if (!SamePopularity.empty()) {
969 SamePopularity.push_back(MostPopularDest);
970 TerminatorInst *TI = BB->getTerminator();
971 for (unsigned i = 0; ; ++i) {
972 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
974 if (std::find(SamePopularity.begin(), SamePopularity.end(),
975 TI->getSuccessor(i)) == SamePopularity.end())
978 MostPopularDest = TI->getSuccessor(i);
983 // Okay, we have finally picked the most popular destination.
984 return MostPopularDest;
987 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) {
988 // If threading this would thread across a loop header, don't even try to
990 if (LoopHeaders.count(BB))
993 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues;
994 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues))
996 assert(!PredValues.empty() &&
997 "ComputeValueKnownInPredecessors returned true with no values");
999 DEBUG(dbgs() << "IN BB: " << *BB;
1000 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1001 dbgs() << " BB '" << BB->getName() << "': FOUND condition = ";
1002 if (PredValues[i].first)
1003 dbgs() << *PredValues[i].first;
1006 dbgs() << " for pred '" << PredValues[i].second->getName()
1010 // Decide what we want to thread through. Convert our list of known values to
1011 // a list of known destinations for each pred. This also discards duplicate
1012 // predecessors and keeps track of the undefined inputs (which are represented
1013 // as a null dest in the PredToDestList).
1014 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1015 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1017 BasicBlock *OnlyDest = 0;
1018 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1020 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1021 BasicBlock *Pred = PredValues[i].second;
1022 if (!SeenPreds.insert(Pred))
1023 continue; // Duplicate predecessor entry.
1025 // If the predecessor ends with an indirect goto, we can't change its
1027 if (isa<IndirectBrInst>(Pred->getTerminator()))
1030 ConstantInt *Val = PredValues[i].first;
1033 if (Val == 0) // Undef.
1035 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1036 DestBB = BI->getSuccessor(Val->isZero());
1038 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
1039 DestBB = SI->getSuccessor(SI->findCaseValue(Val));
1042 // If we have exactly one destination, remember it for efficiency below.
1045 else if (OnlyDest != DestBB)
1046 OnlyDest = MultipleDestSentinel;
1048 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1051 // If all edges were unthreadable, we fail.
1052 if (PredToDestList.empty())
1055 // Determine which is the most common successor. If we have many inputs and
1056 // this block is a switch, we want to start by threading the batch that goes
1057 // to the most popular destination first. If we only know about one
1058 // threadable destination (the common case) we can avoid this.
1059 BasicBlock *MostPopularDest = OnlyDest;
1061 if (MostPopularDest == MultipleDestSentinel)
1062 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1064 // Now that we know what the most popular destination is, factor all
1065 // predecessors that will jump to it into a single predecessor.
1066 SmallVector<BasicBlock*, 16> PredsToFactor;
1067 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1068 if (PredToDestList[i].second == MostPopularDest) {
1069 BasicBlock *Pred = PredToDestList[i].first;
1071 // This predecessor may be a switch or something else that has multiple
1072 // edges to the block. Factor each of these edges by listing them
1073 // according to # occurrences in PredsToFactor.
1074 TerminatorInst *PredTI = Pred->getTerminator();
1075 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1076 if (PredTI->getSuccessor(i) == BB)
1077 PredsToFactor.push_back(Pred);
1080 // If the threadable edges are branching on an undefined value, we get to pick
1081 // the destination that these predecessors should get to.
1082 if (MostPopularDest == 0)
1083 MostPopularDest = BB->getTerminator()->
1084 getSuccessor(GetBestDestForJumpOnUndef(BB));
1086 // Ok, try to thread it!
1087 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1090 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1091 /// a PHI node in the current block. See if there are any simplifications we
1092 /// can do based on inputs to the phi node.
1094 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1095 BasicBlock *BB = PN->getParent();
1097 // TODO: We could make use of this to do it once for blocks with common PHI
1099 SmallVector<BasicBlock*, 1> PredBBs;
1102 // If any of the predecessor blocks end in an unconditional branch, we can
1103 // *duplicate* the conditional branch into that block in order to further
1104 // encourage jump threading and to eliminate cases where we have branch on a
1105 // phi of an icmp (branch on icmp is much better).
1106 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1107 BasicBlock *PredBB = PN->getIncomingBlock(i);
1108 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1109 if (PredBr->isUnconditional()) {
1110 PredBBs[0] = PredBB;
1111 // Try to duplicate BB into PredBB.
1112 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1120 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1121 /// a xor instruction in the current block. See if there are any
1122 /// simplifications we can do based on inputs to the xor.
1124 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1125 BasicBlock *BB = BO->getParent();
1127 // If either the LHS or RHS of the xor is a constant, don't do this
1129 if (isa<ConstantInt>(BO->getOperand(0)) ||
1130 isa<ConstantInt>(BO->getOperand(1)))
1133 // If the first instruction in BB isn't a phi, we won't be able to infer
1134 // anything special about any particular predecessor.
1135 if (!isa<PHINode>(BB->front()))
1138 // If we have a xor as the branch input to this block, and we know that the
1139 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1140 // the condition into the predecessor and fix that value to true, saving some
1141 // logical ops on that path and encouraging other paths to simplify.
1143 // This copies something like this:
1146 // %X = phi i1 [1], [%X']
1147 // %Y = icmp eq i32 %A, %B
1148 // %Z = xor i1 %X, %Y
1153 // %Y = icmp ne i32 %A, %B
1156 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues;
1158 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) {
1159 assert(XorOpValues.empty());
1160 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues))
1165 assert(!XorOpValues.empty() &&
1166 "ComputeValueKnownInPredecessors returned true with no values");
1168 // Scan the information to see which is most popular: true or false. The
1169 // predecessors can be of the set true, false, or undef.
1170 unsigned NumTrue = 0, NumFalse = 0;
1171 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1172 if (!XorOpValues[i].first) continue; // Ignore undefs for the count.
1173 if (XorOpValues[i].first->isZero())
1179 // Determine which value to split on, true, false, or undef if neither.
1180 ConstantInt *SplitVal = 0;
1181 if (NumTrue > NumFalse)
1182 SplitVal = ConstantInt::getTrue(BB->getContext());
1183 else if (NumTrue != 0 || NumFalse != 0)
1184 SplitVal = ConstantInt::getFalse(BB->getContext());
1186 // Collect all of the blocks that this can be folded into so that we can
1187 // factor this once and clone it once.
1188 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1189 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1190 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue;
1192 BlocksToFoldInto.push_back(XorOpValues[i].second);
1195 // If we inferred a value for all of the predecessors, then duplication won't
1196 // help us. However, we can just replace the LHS or RHS with the constant.
1197 if (BlocksToFoldInto.size() ==
1198 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1199 if (SplitVal == 0) {
1200 // If all preds provide undef, just nuke the xor, because it is undef too.
1201 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1202 BO->eraseFromParent();
1203 } else if (SplitVal->isZero()) {
1204 // If all preds provide 0, replace the xor with the other input.
1205 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1206 BO->eraseFromParent();
1208 // If all preds provide 1, set the computed value to 1.
1209 BO->setOperand(!isLHS, SplitVal);
1215 // Try to duplicate BB into PredBB.
1216 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1220 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1221 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1222 /// NewPred using the entries from OldPred (suitably mapped).
1223 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1224 BasicBlock *OldPred,
1225 BasicBlock *NewPred,
1226 DenseMap<Instruction*, Value*> &ValueMap) {
1227 for (BasicBlock::iterator PNI = PHIBB->begin();
1228 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1229 // Ok, we have a PHI node. Figure out what the incoming value was for the
1231 Value *IV = PN->getIncomingValueForBlock(OldPred);
1233 // Remap the value if necessary.
1234 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1235 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1236 if (I != ValueMap.end())
1240 PN->addIncoming(IV, NewPred);
1244 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1245 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1246 /// across BB. Transform the IR to reflect this change.
1247 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1248 const SmallVectorImpl<BasicBlock*> &PredBBs,
1249 BasicBlock *SuccBB) {
1250 // If threading to the same block as we come from, we would infinite loop.
1252 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1253 << "' - would thread to self!\n");
1257 // If threading this would thread across a loop header, don't thread the edge.
1258 // See the comments above FindLoopHeaders for justifications and caveats.
1259 if (LoopHeaders.count(BB)) {
1260 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1261 << "' to dest BB '" << SuccBB->getName()
1262 << "' - it might create an irreducible loop!\n");
1266 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1267 if (JumpThreadCost > Threshold) {
1268 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1269 << "' - Cost is too high: " << JumpThreadCost << "\n");
1273 // And finally, do it! Start by factoring the predecessors is needed.
1275 if (PredBBs.size() == 1)
1276 PredBB = PredBBs[0];
1278 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1279 << " common predecessors.\n");
1280 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1284 // And finally, do it!
1285 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1286 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1287 << ", across block:\n "
1290 // We are going to have to map operands from the original BB block to the new
1291 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1292 // account for entry from PredBB.
1293 DenseMap<Instruction*, Value*> ValueMapping;
1295 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1296 BB->getName()+".thread",
1297 BB->getParent(), BB);
1298 NewBB->moveAfter(PredBB);
1300 BasicBlock::iterator BI = BB->begin();
1301 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1302 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1304 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1305 // mapping and using it to remap operands in the cloned instructions.
1306 for (; !isa<TerminatorInst>(BI); ++BI) {
1307 Instruction *New = BI->clone();
1308 New->setName(BI->getName());
1309 NewBB->getInstList().push_back(New);
1310 ValueMapping[BI] = New;
1312 // Remap operands to patch up intra-block references.
1313 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1314 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1315 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1316 if (I != ValueMapping.end())
1317 New->setOperand(i, I->second);
1321 // We didn't copy the terminator from BB over to NewBB, because there is now
1322 // an unconditional jump to SuccBB. Insert the unconditional jump.
1323 BranchInst::Create(SuccBB, NewBB);
1325 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1326 // PHI nodes for NewBB now.
1327 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1329 // If there were values defined in BB that are used outside the block, then we
1330 // now have to update all uses of the value to use either the original value,
1331 // the cloned value, or some PHI derived value. This can require arbitrary
1332 // PHI insertion, of which we are prepared to do, clean these up now.
1333 SSAUpdater SSAUpdate;
1334 SmallVector<Use*, 16> UsesToRename;
1335 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1336 // Scan all uses of this instruction to see if it is used outside of its
1337 // block, and if so, record them in UsesToRename.
1338 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1340 Instruction *User = cast<Instruction>(*UI);
1341 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1342 if (UserPN->getIncomingBlock(UI) == BB)
1344 } else if (User->getParent() == BB)
1347 UsesToRename.push_back(&UI.getUse());
1350 // If there are no uses outside the block, we're done with this instruction.
1351 if (UsesToRename.empty())
1354 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1356 // We found a use of I outside of BB. Rename all uses of I that are outside
1357 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1358 // with the two values we know.
1359 SSAUpdate.Initialize(I);
1360 SSAUpdate.AddAvailableValue(BB, I);
1361 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1363 while (!UsesToRename.empty())
1364 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1365 DEBUG(dbgs() << "\n");
1369 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1370 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1371 // us to simplify any PHI nodes in BB.
1372 TerminatorInst *PredTerm = PredBB->getTerminator();
1373 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1374 if (PredTerm->getSuccessor(i) == BB) {
1375 RemovePredecessorAndSimplify(BB, PredBB, TD);
1376 PredTerm->setSuccessor(i, NewBB);
1379 // At this point, the IR is fully up to date and consistent. Do a quick scan
1380 // over the new instructions and zap any that are constants or dead. This
1381 // frequently happens because of phi translation.
1382 SimplifyInstructionsInBlock(NewBB, TD);
1384 // Threaded an edge!
1389 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1390 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1391 /// If we can duplicate the contents of BB up into PredBB do so now, this
1392 /// improves the odds that the branch will be on an analyzable instruction like
1394 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1395 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1396 assert(!PredBBs.empty() && "Can't handle an empty set");
1398 // If BB is a loop header, then duplicating this block outside the loop would
1399 // cause us to transform this into an irreducible loop, don't do this.
1400 // See the comments above FindLoopHeaders for justifications and caveats.
1401 if (LoopHeaders.count(BB)) {
1402 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1403 << "' into predecessor block '" << PredBBs[0]->getName()
1404 << "' - it might create an irreducible loop!\n");
1408 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1409 if (DuplicationCost > Threshold) {
1410 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1411 << "' - Cost is too high: " << DuplicationCost << "\n");
1415 // And finally, do it! Start by factoring the predecessors is needed.
1417 if (PredBBs.size() == 1)
1418 PredBB = PredBBs[0];
1420 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1421 << " common predecessors.\n");
1422 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1426 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1428 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1429 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1430 << DuplicationCost << " block is:" << *BB << "\n");
1432 // Unless PredBB ends with an unconditional branch, split the edge so that we
1433 // can just clone the bits from BB into the end of the new PredBB.
1434 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1436 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1437 PredBB = SplitEdge(PredBB, BB, this);
1438 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1441 // We are going to have to map operands from the original BB block into the
1442 // PredBB block. Evaluate PHI nodes in BB.
1443 DenseMap<Instruction*, Value*> ValueMapping;
1445 BasicBlock::iterator BI = BB->begin();
1446 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1447 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1449 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1450 // mapping and using it to remap operands in the cloned instructions.
1451 for (; BI != BB->end(); ++BI) {
1452 Instruction *New = BI->clone();
1454 // Remap operands to patch up intra-block references.
1455 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1456 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1457 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1458 if (I != ValueMapping.end())
1459 New->setOperand(i, I->second);
1462 // If this instruction can be simplified after the operands are updated,
1463 // just use the simplified value instead. This frequently happens due to
1465 if (Value *IV = SimplifyInstruction(New, TD)) {
1467 ValueMapping[BI] = IV;
1469 // Otherwise, insert the new instruction into the block.
1470 New->setName(BI->getName());
1471 PredBB->getInstList().insert(OldPredBranch, New);
1472 ValueMapping[BI] = New;
1476 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1477 // add entries to the PHI nodes for branch from PredBB now.
1478 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1479 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1481 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1484 // If there were values defined in BB that are used outside the block, then we
1485 // now have to update all uses of the value to use either the original value,
1486 // the cloned value, or some PHI derived value. This can require arbitrary
1487 // PHI insertion, of which we are prepared to do, clean these up now.
1488 SSAUpdater SSAUpdate;
1489 SmallVector<Use*, 16> UsesToRename;
1490 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1491 // Scan all uses of this instruction to see if it is used outside of its
1492 // block, and if so, record them in UsesToRename.
1493 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1495 Instruction *User = cast<Instruction>(*UI);
1496 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1497 if (UserPN->getIncomingBlock(UI) == BB)
1499 } else if (User->getParent() == BB)
1502 UsesToRename.push_back(&UI.getUse());
1505 // If there are no uses outside the block, we're done with this instruction.
1506 if (UsesToRename.empty())
1509 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1511 // We found a use of I outside of BB. Rename all uses of I that are outside
1512 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1513 // with the two values we know.
1514 SSAUpdate.Initialize(I);
1515 SSAUpdate.AddAvailableValue(BB, I);
1516 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1518 while (!UsesToRename.empty())
1519 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1520 DEBUG(dbgs() << "\n");
1523 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1525 RemovePredecessorAndSimplify(BB, PredBB, TD);
1527 // Remove the unconditional branch at the end of the PredBB block.
1528 OldPredBranch->eraseFromParent();