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/Analysis/Loads.h"
22 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
23 #include "llvm/Transforms/Utils/Local.h"
24 #include "llvm/Transforms/Utils/SSAUpdater.h"
25 #include "llvm/Target/TargetData.h"
26 #include "llvm/ADT/DenseMap.h"
27 #include "llvm/ADT/Statistic.h"
28 #include "llvm/ADT/STLExtras.h"
29 #include "llvm/ADT/SmallPtrSet.h"
30 #include "llvm/ADT/SmallSet.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ValueHandle.h"
34 #include "llvm/Support/raw_ostream.h"
37 STATISTIC(NumThreads, "Number of jumps threaded");
38 STATISTIC(NumFolds, "Number of terminators folded");
39 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
41 static cl::opt<unsigned>
42 Threshold("jump-threading-threshold",
43 cl::desc("Max block size to duplicate for jump threading"),
44 cl::init(6), cl::Hidden);
46 // Turn on use of LazyValueInfo.
48 EnableLVI("enable-jump-threading-lvi", cl::ReallyHidden);
53 /// This pass performs 'jump threading', which looks at blocks that have
54 /// multiple predecessors and multiple successors. If one or more of the
55 /// predecessors of the block can be proven to always jump to one of the
56 /// successors, we forward the edge from the predecessor to the successor by
57 /// duplicating the contents of this block.
59 /// An example of when this can occur is code like this:
66 /// In this case, the unconditional branch at the end of the first if can be
67 /// revectored to the false side of the second if.
69 class JumpThreading : public FunctionPass {
73 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
75 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
78 static char ID; // Pass identification
79 JumpThreading() : FunctionPass(&ID) {}
81 bool runOnFunction(Function &F);
83 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
85 AU.addRequired<LazyValueInfo>();
88 void FindLoopHeaders(Function &F);
89 bool ProcessBlock(BasicBlock *BB);
90 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
92 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
93 const SmallVectorImpl<BasicBlock *> &PredBBs);
95 typedef SmallVectorImpl<std::pair<ConstantInt*,
96 BasicBlock*> > PredValueInfo;
98 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
99 PredValueInfo &Result);
100 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB);
103 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
104 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
106 bool ProcessBranchOnPHI(PHINode *PN);
107 bool ProcessBranchOnXOR(BinaryOperator *BO);
109 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
113 char JumpThreading::ID = 0;
114 static RegisterPass<JumpThreading>
115 X("jump-threading", "Jump Threading");
117 // Public interface to the Jump Threading pass
118 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
120 /// runOnFunction - Top level algorithm.
122 bool JumpThreading::runOnFunction(Function &F) {
123 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
124 TD = getAnalysisIfAvailable<TargetData>();
125 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0;
129 bool Changed, EverChanged = false;
132 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
134 // Thread all of the branches we can over this block.
135 while (ProcessBlock(BB))
140 // If the block is trivially dead, zap it. This eliminates the successor
141 // edges which simplifies the CFG.
142 if (pred_begin(BB) == pred_end(BB) &&
143 BB != &BB->getParent()->getEntryBlock()) {
144 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
145 << "' with terminator: " << *BB->getTerminator() << '\n');
146 LoopHeaders.erase(BB);
149 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
150 // Can't thread an unconditional jump, but if the block is "almost
151 // empty", we can replace uses of it with uses of the successor and make
153 if (BI->isUnconditional() &&
154 BB != &BB->getParent()->getEntryBlock()) {
155 BasicBlock::iterator BBI = BB->getFirstNonPHI();
156 // Ignore dbg intrinsics.
157 while (isa<DbgInfoIntrinsic>(BBI))
159 // If the terminator is the only non-phi instruction, try to nuke it.
160 if (BBI->isTerminator()) {
161 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
162 // block, we have to make sure it isn't in the LoopHeaders set. We
163 // reinsert afterward if needed.
164 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
165 BasicBlock *Succ = BI->getSuccessor(0);
167 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
169 // If we deleted BB and BB was the header of a loop, then the
170 // successor is now the header of the loop.
174 if (ErasedFromLoopHeaders)
175 LoopHeaders.insert(BB);
180 EverChanged |= Changed;
187 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
188 /// thread across it.
189 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
190 /// Ignore PHI nodes, these will be flattened when duplication happens.
191 BasicBlock::const_iterator I = BB->getFirstNonPHI();
193 // FIXME: THREADING will delete values that are just used to compute the
194 // branch, so they shouldn't count against the duplication cost.
197 // Sum up the cost of each instruction until we get to the terminator. Don't
198 // include the terminator because the copy won't include it.
200 for (; !isa<TerminatorInst>(I); ++I) {
201 // Debugger intrinsics don't incur code size.
202 if (isa<DbgInfoIntrinsic>(I)) continue;
204 // If this is a pointer->pointer bitcast, it is free.
205 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
208 // All other instructions count for at least one unit.
211 // Calls are more expensive. If they are non-intrinsic calls, we model them
212 // as having cost of 4. If they are a non-vector intrinsic, we model them
213 // as having cost of 2 total, and if they are a vector intrinsic, we model
214 // them as having cost 1.
215 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
216 if (!isa<IntrinsicInst>(CI))
218 else if (!CI->getType()->isVectorTy())
223 // Threading through a switch statement is particularly profitable. If this
224 // block ends in a switch, decrease its cost to make it more likely to happen.
225 if (isa<SwitchInst>(I))
226 Size = Size > 6 ? Size-6 : 0;
231 /// FindLoopHeaders - We do not want jump threading to turn proper loop
232 /// structures into irreducible loops. Doing this breaks up the loop nesting
233 /// hierarchy and pessimizes later transformations. To prevent this from
234 /// happening, we first have to find the loop headers. Here we approximate this
235 /// by finding targets of backedges in the CFG.
237 /// Note that there definitely are cases when we want to allow threading of
238 /// edges across a loop header. For example, threading a jump from outside the
239 /// loop (the preheader) to an exit block of the loop is definitely profitable.
240 /// It is also almost always profitable to thread backedges from within the loop
241 /// to exit blocks, and is often profitable to thread backedges to other blocks
242 /// within the loop (forming a nested loop). This simple analysis is not rich
243 /// enough to track all of these properties and keep it up-to-date as the CFG
244 /// mutates, so we don't allow any of these transformations.
246 void JumpThreading::FindLoopHeaders(Function &F) {
247 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
248 FindFunctionBackedges(F, Edges);
250 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
251 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
254 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
255 /// if we can infer that the value is a known ConstantInt in any of our
256 /// predecessors. If so, return the known list of value and pred BB in the
257 /// result vector. If a value is known to be undef, it is returned as null.
259 /// This returns true if there were any known values.
262 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){
263 // If V is a constantint, then it is known in all predecessors.
264 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) {
265 ConstantInt *CI = dyn_cast<ConstantInt>(V);
267 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
268 Result.push_back(std::make_pair(CI, *PI));
272 // If V is a non-instruction value, or an instruction in a different block,
273 // then it can't be derived from a PHI.
274 Instruction *I = dyn_cast<Instruction>(V);
275 if (I == 0 || I->getParent() != BB) {
277 // Okay, if this is a live-in value, see if it has a known value at the end
278 // of any of our predecessors.
280 // FIXME: This should be an edge property, not a block end property.
281 /// TODO: Per PR2563, we could infer value range information about a
282 /// predecessor based on its terminator.
285 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
286 // "I" is a non-local compare-with-a-constant instruction. This would be
287 // able to handle value inequalities better, for example if the compare is
288 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
289 // Perhaps getConstantOnEdge should be smart enough to do this?
291 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
292 // If the value is known by LazyValueInfo to be a constant in a
293 // predecessor, use that information to try to thread this block.
294 Constant *PredCst = LVI->getConstantOnEdge(V, *PI, BB);
296 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst)))
299 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), *PI));
302 return !Result.empty();
308 /// If I is a PHI node, then we know the incoming values for any constants.
309 if (PHINode *PN = dyn_cast<PHINode>(I)) {
310 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
311 Value *InVal = PN->getIncomingValue(i);
312 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) {
313 ConstantInt *CI = dyn_cast<ConstantInt>(InVal);
314 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i)));
317 return !Result.empty();
320 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals;
322 // Handle some boolean conditions.
323 if (I->getType()->getPrimitiveSizeInBits() == 1) {
325 // X & false -> false
326 if (I->getOpcode() == Instruction::Or ||
327 I->getOpcode() == Instruction::And) {
328 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
329 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals);
331 if (LHSVals.empty() && RHSVals.empty())
334 ConstantInt *InterestingVal;
335 if (I->getOpcode() == Instruction::Or)
336 InterestingVal = ConstantInt::getTrue(I->getContext());
338 InterestingVal = ConstantInt::getFalse(I->getContext());
340 // Scan for the sentinel. If we find an undef, force it to the
341 // interesting value: x|undef -> true and x&undef -> false.
342 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
343 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) {
344 Result.push_back(LHSVals[i]);
345 Result.back().first = InterestingVal;
347 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
348 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) {
349 Result.push_back(RHSVals[i]);
350 Result.back().first = InterestingVal;
352 return !Result.empty();
355 // Handle the NOT form of XOR.
356 if (I->getOpcode() == Instruction::Xor &&
357 isa<ConstantInt>(I->getOperand(1)) &&
358 cast<ConstantInt>(I->getOperand(1))->isOne()) {
359 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result);
363 // Invert the known values.
364 for (unsigned i = 0, e = Result.size(); i != e; ++i)
367 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first));
372 // Handle compare with phi operand, where the PHI is defined in this block.
373 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
374 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
375 if (PN && PN->getParent() == BB) {
376 // We can do this simplification if any comparisons fold to true or false.
378 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
379 BasicBlock *PredBB = PN->getIncomingBlock(i);
380 Value *LHS = PN->getIncomingValue(i);
381 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
383 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
385 if (!LVI || !isa<Constant>(RHS))
388 LazyValueInfo::Tristate
389 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
390 cast<Constant>(RHS), PredBB, BB);
391 if (ResT == LazyValueInfo::Unknown)
393 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
396 if (isa<UndefValue>(Res))
397 Result.push_back(std::make_pair((ConstantInt*)0, PredBB));
398 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res))
399 Result.push_back(std::make_pair(CI, PredBB));
402 return !Result.empty();
406 // If comparing a live-in value against a constant, see if we know the
407 // live-in value on any predecessors.
408 if (LVI && isa<Constant>(Cmp->getOperand(1)) &&
409 Cmp->getType()->isIntegerTy() && // Not vector compare.
410 (!isa<Instruction>(Cmp->getOperand(0)) ||
411 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB)) {
412 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
414 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
415 // If the value is known by LazyValueInfo to be a constant in a
416 // predecessor, use that information to try to thread this block.
417 LazyValueInfo::Tristate
418 Res = LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
420 if (Res == LazyValueInfo::Unknown)
423 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
424 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), *PI));
427 return !Result.empty();
435 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
436 /// in an undefined jump, decide which block is best to revector to.
438 /// Since we can pick an arbitrary destination, we pick the successor with the
439 /// fewest predecessors. This should reduce the in-degree of the others.
441 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
442 TerminatorInst *BBTerm = BB->getTerminator();
443 unsigned MinSucc = 0;
444 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
445 // Compute the successor with the minimum number of predecessors.
446 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
447 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
448 TestBB = BBTerm->getSuccessor(i);
449 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
450 if (NumPreds < MinNumPreds)
457 /// ProcessBlock - If there are any predecessors whose control can be threaded
458 /// through to a successor, transform them now.
459 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
460 // If the block is trivially dead, just return and let the caller nuke it.
461 // This simplifies other transformations.
462 if (pred_begin(BB) == pred_end(BB) &&
463 BB != &BB->getParent()->getEntryBlock())
466 // If this block has a single predecessor, and if that pred has a single
467 // successor, merge the blocks. This encourages recursive jump threading
468 // because now the condition in this block can be threaded through
469 // predecessors of our predecessor block.
470 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
471 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
473 // If SinglePred was a loop header, BB becomes one.
474 if (LoopHeaders.erase(SinglePred))
475 LoopHeaders.insert(BB);
477 // Remember if SinglePred was the entry block of the function. If so, we
478 // will need to move BB back to the entry position.
479 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
480 MergeBasicBlockIntoOnlyPred(BB);
482 if (isEntry && BB != &BB->getParent()->getEntryBlock())
483 BB->moveBefore(&BB->getParent()->getEntryBlock());
488 // Look to see if the terminator is a branch of switch, if not we can't thread
491 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
492 // Can't thread an unconditional jump.
493 if (BI->isUnconditional()) return false;
494 Condition = BI->getCondition();
495 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
496 Condition = SI->getCondition();
498 return false; // Must be an invoke.
500 // If the terminator of this block is branching on a constant, simplify the
501 // terminator to an unconditional branch. This can occur due to threading in
503 if (isa<ConstantInt>(Condition)) {
504 DEBUG(dbgs() << " In block '" << BB->getName()
505 << "' folding terminator: " << *BB->getTerminator() << '\n');
507 ConstantFoldTerminator(BB);
511 // If the terminator is branching on an undef, we can pick any of the
512 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
513 if (isa<UndefValue>(Condition)) {
514 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
516 // Fold the branch/switch.
517 TerminatorInst *BBTerm = BB->getTerminator();
518 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
519 if (i == BestSucc) continue;
520 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD);
523 DEBUG(dbgs() << " In block '" << BB->getName()
524 << "' folding undef terminator: " << *BBTerm << '\n');
525 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
526 BBTerm->eraseFromParent();
530 Instruction *CondInst = dyn_cast<Instruction>(Condition);
532 // If the condition is an instruction defined in another block, see if a
533 // predecessor has the same condition:
538 !Condition->hasOneUse() && // Multiple uses.
539 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition.
540 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
541 if (isa<BranchInst>(BB->getTerminator())) {
542 for (; PI != E; ++PI)
543 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
544 if (PBI->isConditional() && PBI->getCondition() == Condition &&
545 ProcessBranchOnDuplicateCond(*PI, BB))
548 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator");
549 for (; PI != E; ++PI)
550 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator()))
551 if (PSI->getCondition() == Condition &&
552 ProcessSwitchOnDuplicateCond(*PI, BB))
557 // All the rest of our checks depend on the condition being an instruction.
559 // FIXME: Unify this with code below.
560 if (LVI && ProcessThreadableEdges(Condition, BB))
566 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
568 (!isa<PHINode>(CondCmp->getOperand(0)) ||
569 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) {
570 // If we have a comparison, loop over the predecessors to see if there is
571 // a condition with a lexically identical value.
572 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
573 for (; PI != E; ++PI)
574 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
575 if (PBI->isConditional() && *PI != BB) {
576 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) {
577 if (CI->getOperand(0) == CondCmp->getOperand(0) &&
578 CI->getOperand(1) == CondCmp->getOperand(1) &&
579 CI->getPredicate() == CondCmp->getPredicate()) {
580 // TODO: Could handle things like (x != 4) --> (x == 17)
581 if (ProcessBranchOnDuplicateCond(*PI, BB))
589 // Check for some cases that are worth simplifying. Right now we want to look
590 // for loads that are used by a switch or by the condition for the branch. If
591 // we see one, check to see if it's partially redundant. If so, insert a PHI
592 // which can then be used to thread the values.
594 Value *SimplifyValue = CondInst;
595 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
596 if (isa<Constant>(CondCmp->getOperand(1)))
597 SimplifyValue = CondCmp->getOperand(0);
599 // TODO: There are other places where load PRE would be profitable, such as
600 // more complex comparisons.
601 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
602 if (SimplifyPartiallyRedundantLoad(LI))
606 // Handle a variety of cases where we are branching on something derived from
607 // a PHI node in the current block. If we can prove that any predecessors
608 // compute a predictable value based on a PHI node, thread those predecessors.
610 if (ProcessThreadableEdges(CondInst, BB))
613 // If this is an otherwise-unfoldable branch on a phi node in the current
614 // block, see if we can simplify.
615 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
616 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
617 return ProcessBranchOnPHI(PN);
620 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
621 if (CondInst->getOpcode() == Instruction::Xor &&
622 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
623 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
626 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
627 // "(X == 4)", thread through this block.
632 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
633 /// block that jump on exactly the same condition. This means that we almost
634 /// always know the direction of the edge in the DESTBB:
636 /// br COND, DESTBB, BBY
638 /// br COND, BBZ, BBW
640 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
641 /// in DESTBB, we have to thread over it.
642 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
644 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
646 // If both successors of PredBB go to DESTBB, we don't know anything. We can
647 // fold the branch to an unconditional one, which allows other recursive
650 if (PredBI->getSuccessor(1) != BB)
652 else if (PredBI->getSuccessor(0) != BB)
655 DEBUG(dbgs() << " In block '" << PredBB->getName()
656 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
658 ConstantFoldTerminator(PredBB);
662 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
664 // If the dest block has one predecessor, just fix the branch condition to a
665 // constant and fold it.
666 if (BB->getSinglePredecessor()) {
667 DEBUG(dbgs() << " In block '" << BB->getName()
668 << "' folding condition to '" << BranchDir << "': "
669 << *BB->getTerminator() << '\n');
671 Value *OldCond = DestBI->getCondition();
672 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
674 // Delete dead instructions before we fold the branch. Folding the branch
675 // can eliminate edges from the CFG which can end up deleting OldCond.
676 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
677 ConstantFoldTerminator(BB);
682 // Next, figure out which successor we are threading to.
683 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
685 SmallVector<BasicBlock*, 2> Preds;
686 Preds.push_back(PredBB);
688 // Ok, try to thread it!
689 return ThreadEdge(BB, Preds, SuccBB);
692 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
693 /// block that switch on exactly the same condition. This means that we almost
694 /// always know the direction of the edge in the DESTBB:
696 /// switch COND [... DESTBB, BBY ... ]
698 /// switch COND [... BBZ, BBW ]
700 /// Optimizing switches like this is very important, because simplifycfg builds
701 /// switches out of repeated 'if' conditions.
702 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
703 BasicBlock *DestBB) {
704 // Can't thread edge to self.
705 if (PredBB == DestBB)
708 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
709 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
711 // There are a variety of optimizations that we can potentially do on these
712 // blocks: we order them from most to least preferable.
714 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
715 // directly to their destination. This does not introduce *any* code size
716 // growth. Skip debug info first.
717 BasicBlock::iterator BBI = DestBB->begin();
718 while (isa<DbgInfoIntrinsic>(BBI))
721 // FIXME: Thread if it just contains a PHI.
722 if (isa<SwitchInst>(BBI)) {
723 bool MadeChange = false;
724 // Ignore the default edge for now.
725 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
726 ConstantInt *DestVal = DestSI->getCaseValue(i);
727 BasicBlock *DestSucc = DestSI->getSuccessor(i);
729 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
730 // PredSI has an explicit case for it. If so, forward. If it is covered
731 // by the default case, we can't update PredSI.
732 unsigned PredCase = PredSI->findCaseValue(DestVal);
733 if (PredCase == 0) continue;
735 // If PredSI doesn't go to DestBB on this value, then it won't reach the
736 // case on this condition.
737 if (PredSI->getSuccessor(PredCase) != DestBB &&
738 DestSI->getSuccessor(i) != DestBB)
741 // Do not forward this if it already goes to this destination, this would
742 // be an infinite loop.
743 if (PredSI->getSuccessor(PredCase) == DestSucc)
746 // Otherwise, we're safe to make the change. Make sure that the edge from
747 // DestSI to DestSucc is not critical and has no PHI nodes.
748 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
749 DEBUG(dbgs() << "THROUGH: " << *DestSI);
751 // If the destination has PHI nodes, just split the edge for updating
753 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
754 SplitCriticalEdge(DestSI, i, this);
755 DestSucc = DestSI->getSuccessor(i);
757 FoldSingleEntryPHINodes(DestSucc);
758 PredSI->setSuccessor(PredCase, DestSucc);
770 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
771 /// load instruction, eliminate it by replacing it with a PHI node. This is an
772 /// important optimization that encourages jump threading, and needs to be run
773 /// interlaced with other jump threading tasks.
774 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
775 // Don't hack volatile loads.
776 if (LI->isVolatile()) return false;
778 // If the load is defined in a block with exactly one predecessor, it can't be
779 // partially redundant.
780 BasicBlock *LoadBB = LI->getParent();
781 if (LoadBB->getSinglePredecessor())
784 Value *LoadedPtr = LI->getOperand(0);
786 // If the loaded operand is defined in the LoadBB, it can't be available.
787 // TODO: Could do simple PHI translation, that would be fun :)
788 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
789 if (PtrOp->getParent() == LoadBB)
792 // Scan a few instructions up from the load, to see if it is obviously live at
793 // the entry to its block.
794 BasicBlock::iterator BBIt = LI;
796 if (Value *AvailableVal =
797 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
798 // If the value if the load is locally available within the block, just use
799 // it. This frequently occurs for reg2mem'd allocas.
800 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
802 // If the returned value is the load itself, replace with an undef. This can
803 // only happen in dead loops.
804 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
805 LI->replaceAllUsesWith(AvailableVal);
806 LI->eraseFromParent();
810 // Otherwise, if we scanned the whole block and got to the top of the block,
811 // we know the block is locally transparent to the load. If not, something
812 // might clobber its value.
813 if (BBIt != LoadBB->begin())
817 SmallPtrSet<BasicBlock*, 8> PredsScanned;
818 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
819 AvailablePredsTy AvailablePreds;
820 BasicBlock *OneUnavailablePred = 0;
822 // If we got here, the loaded value is transparent through to the start of the
823 // block. Check to see if it is available in any of the predecessor blocks.
824 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
826 BasicBlock *PredBB = *PI;
828 // If we already scanned this predecessor, skip it.
829 if (!PredsScanned.insert(PredBB))
832 // Scan the predecessor to see if the value is available in the pred.
833 BBIt = PredBB->end();
834 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
835 if (!PredAvailable) {
836 OneUnavailablePred = PredBB;
840 // If so, this load is partially redundant. Remember this info so that we
841 // can create a PHI node.
842 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
845 // If the loaded value isn't available in any predecessor, it isn't partially
847 if (AvailablePreds.empty()) return false;
849 // Okay, the loaded value is available in at least one (and maybe all!)
850 // predecessors. If the value is unavailable in more than one unique
851 // predecessor, we want to insert a merge block for those common predecessors.
852 // This ensures that we only have to insert one reload, thus not increasing
854 BasicBlock *UnavailablePred = 0;
856 // If there is exactly one predecessor where the value is unavailable, the
857 // already computed 'OneUnavailablePred' block is it. If it ends in an
858 // unconditional branch, we know that it isn't a critical edge.
859 if (PredsScanned.size() == AvailablePreds.size()+1 &&
860 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
861 UnavailablePred = OneUnavailablePred;
862 } else if (PredsScanned.size() != AvailablePreds.size()) {
863 // Otherwise, we had multiple unavailable predecessors or we had a critical
864 // edge from the one.
865 SmallVector<BasicBlock*, 8> PredsToSplit;
866 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
868 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
869 AvailablePredSet.insert(AvailablePreds[i].first);
871 // Add all the unavailable predecessors to the PredsToSplit list.
872 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
874 if (!AvailablePredSet.count(*PI))
875 PredsToSplit.push_back(*PI);
877 // Split them out to their own block.
879 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
880 "thread-pre-split", this);
883 // If the value isn't available in all predecessors, then there will be
884 // exactly one where it isn't available. Insert a load on that edge and add
885 // it to the AvailablePreds list.
886 if (UnavailablePred) {
887 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
888 "Can't handle critical edge here!");
889 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
891 UnavailablePred->getTerminator());
892 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
895 // Now we know that each predecessor of this block has a value in
896 // AvailablePreds, sort them for efficient access as we're walking the preds.
897 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
899 // Create a PHI node at the start of the block for the PRE'd load value.
900 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
903 // Insert new entries into the PHI for each predecessor. A single block may
904 // have multiple entries here.
905 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
907 AvailablePredsTy::iterator I =
908 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
909 std::make_pair(*PI, (Value*)0));
911 assert(I != AvailablePreds.end() && I->first == *PI &&
912 "Didn't find entry for predecessor!");
914 PN->addIncoming(I->second, I->first);
917 //cerr << "PRE: " << *LI << *PN << "\n";
919 LI->replaceAllUsesWith(PN);
920 LI->eraseFromParent();
925 /// FindMostPopularDest - The specified list contains multiple possible
926 /// threadable destinations. Pick the one that occurs the most frequently in
929 FindMostPopularDest(BasicBlock *BB,
930 const SmallVectorImpl<std::pair<BasicBlock*,
931 BasicBlock*> > &PredToDestList) {
932 assert(!PredToDestList.empty());
934 // Determine popularity. If there are multiple possible destinations, we
935 // explicitly choose to ignore 'undef' destinations. We prefer to thread
936 // blocks with known and real destinations to threading undef. We'll handle
937 // them later if interesting.
938 DenseMap<BasicBlock*, unsigned> DestPopularity;
939 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
940 if (PredToDestList[i].second)
941 DestPopularity[PredToDestList[i].second]++;
943 // Find the most popular dest.
944 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
945 BasicBlock *MostPopularDest = DPI->first;
946 unsigned Popularity = DPI->second;
947 SmallVector<BasicBlock*, 4> SamePopularity;
949 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
950 // If the popularity of this entry isn't higher than the popularity we've
951 // seen so far, ignore it.
952 if (DPI->second < Popularity)
954 else if (DPI->second == Popularity) {
955 // If it is the same as what we've seen so far, keep track of it.
956 SamePopularity.push_back(DPI->first);
958 // If it is more popular, remember it.
959 SamePopularity.clear();
960 MostPopularDest = DPI->first;
961 Popularity = DPI->second;
965 // Okay, now we know the most popular destination. If there is more than
966 // destination, we need to determine one. This is arbitrary, but we need
967 // to make a deterministic decision. Pick the first one that appears in the
969 if (!SamePopularity.empty()) {
970 SamePopularity.push_back(MostPopularDest);
971 TerminatorInst *TI = BB->getTerminator();
972 for (unsigned i = 0; ; ++i) {
973 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
975 if (std::find(SamePopularity.begin(), SamePopularity.end(),
976 TI->getSuccessor(i)) == SamePopularity.end())
979 MostPopularDest = TI->getSuccessor(i);
984 // Okay, we have finally picked the most popular destination.
985 return MostPopularDest;
988 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) {
989 // If threading this would thread across a loop header, don't even try to
991 if (LoopHeaders.count(BB))
994 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues;
995 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues))
997 assert(!PredValues.empty() &&
998 "ComputeValueKnownInPredecessors returned true with no values");
1000 DEBUG(dbgs() << "IN BB: " << *BB;
1001 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1002 dbgs() << " BB '" << BB->getName() << "': FOUND condition = ";
1003 if (PredValues[i].first)
1004 dbgs() << *PredValues[i].first;
1007 dbgs() << " for pred '" << PredValues[i].second->getName()
1011 // Decide what we want to thread through. Convert our list of known values to
1012 // a list of known destinations for each pred. This also discards duplicate
1013 // predecessors and keeps track of the undefined inputs (which are represented
1014 // as a null dest in the PredToDestList).
1015 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1016 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1018 BasicBlock *OnlyDest = 0;
1019 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1021 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1022 BasicBlock *Pred = PredValues[i].second;
1023 if (!SeenPreds.insert(Pred))
1024 continue; // Duplicate predecessor entry.
1026 // If the predecessor ends with an indirect goto, we can't change its
1028 if (isa<IndirectBrInst>(Pred->getTerminator()))
1031 ConstantInt *Val = PredValues[i].first;
1034 if (Val == 0) // Undef.
1036 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1037 DestBB = BI->getSuccessor(Val->isZero());
1039 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
1040 DestBB = SI->getSuccessor(SI->findCaseValue(Val));
1043 // If we have exactly one destination, remember it for efficiency below.
1046 else if (OnlyDest != DestBB)
1047 OnlyDest = MultipleDestSentinel;
1049 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1052 // If all edges were unthreadable, we fail.
1053 if (PredToDestList.empty())
1056 // Determine which is the most common successor. If we have many inputs and
1057 // this block is a switch, we want to start by threading the batch that goes
1058 // to the most popular destination first. If we only know about one
1059 // threadable destination (the common case) we can avoid this.
1060 BasicBlock *MostPopularDest = OnlyDest;
1062 if (MostPopularDest == MultipleDestSentinel)
1063 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1065 // Now that we know what the most popular destination is, factor all
1066 // predecessors that will jump to it into a single predecessor.
1067 SmallVector<BasicBlock*, 16> PredsToFactor;
1068 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1069 if (PredToDestList[i].second == MostPopularDest) {
1070 BasicBlock *Pred = PredToDestList[i].first;
1072 // This predecessor may be a switch or something else that has multiple
1073 // edges to the block. Factor each of these edges by listing them
1074 // according to # occurrences in PredsToFactor.
1075 TerminatorInst *PredTI = Pred->getTerminator();
1076 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1077 if (PredTI->getSuccessor(i) == BB)
1078 PredsToFactor.push_back(Pred);
1081 // If the threadable edges are branching on an undefined value, we get to pick
1082 // the destination that these predecessors should get to.
1083 if (MostPopularDest == 0)
1084 MostPopularDest = BB->getTerminator()->
1085 getSuccessor(GetBestDestForJumpOnUndef(BB));
1087 // Ok, try to thread it!
1088 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1091 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1092 /// a PHI node in the current block. See if there are any simplifications we
1093 /// can do based on inputs to the phi node.
1095 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1096 BasicBlock *BB = PN->getParent();
1098 // TODO: We could make use of this to do it once for blocks with common PHI
1100 SmallVector<BasicBlock*, 1> PredBBs;
1103 // If any of the predecessor blocks end in an unconditional branch, we can
1104 // *duplicate* the conditional branch into that block in order to further
1105 // encourage jump threading and to eliminate cases where we have branch on a
1106 // phi of an icmp (branch on icmp is much better).
1107 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1108 BasicBlock *PredBB = PN->getIncomingBlock(i);
1109 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1110 if (PredBr->isUnconditional()) {
1111 PredBBs[0] = PredBB;
1112 // Try to duplicate BB into PredBB.
1113 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1121 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1122 /// a xor instruction in the current block. See if there are any
1123 /// simplifications we can do based on inputs to the xor.
1125 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1126 BasicBlock *BB = BO->getParent();
1128 // If either the LHS or RHS of the xor is a constant, don't do this
1130 if (isa<ConstantInt>(BO->getOperand(0)) ||
1131 isa<ConstantInt>(BO->getOperand(1)))
1134 // If the first instruction in BB isn't a phi, we won't be able to infer
1135 // anything special about any particular predecessor.
1136 if (!isa<PHINode>(BB->front()))
1139 // If we have a xor as the branch input to this block, and we know that the
1140 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1141 // the condition into the predecessor and fix that value to true, saving some
1142 // logical ops on that path and encouraging other paths to simplify.
1144 // This copies something like this:
1147 // %X = phi i1 [1], [%X']
1148 // %Y = icmp eq i32 %A, %B
1149 // %Z = xor i1 %X, %Y
1154 // %Y = icmp ne i32 %A, %B
1157 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues;
1159 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) {
1160 assert(XorOpValues.empty());
1161 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues))
1166 assert(!XorOpValues.empty() &&
1167 "ComputeValueKnownInPredecessors returned true with no values");
1169 // Scan the information to see which is most popular: true or false. The
1170 // predecessors can be of the set true, false, or undef.
1171 unsigned NumTrue = 0, NumFalse = 0;
1172 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1173 if (!XorOpValues[i].first) continue; // Ignore undefs for the count.
1174 if (XorOpValues[i].first->isZero())
1180 // Determine which value to split on, true, false, or undef if neither.
1181 ConstantInt *SplitVal = 0;
1182 if (NumTrue > NumFalse)
1183 SplitVal = ConstantInt::getTrue(BB->getContext());
1184 else if (NumTrue != 0 || NumFalse != 0)
1185 SplitVal = ConstantInt::getFalse(BB->getContext());
1187 // Collect all of the blocks that this can be folded into so that we can
1188 // factor this once and clone it once.
1189 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1190 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1191 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue;
1193 BlocksToFoldInto.push_back(XorOpValues[i].second);
1196 // If we inferred a value for all of the predecessors, then duplication won't
1197 // help us. However, we can just replace the LHS or RHS with the constant.
1198 if (BlocksToFoldInto.size() ==
1199 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1200 if (SplitVal == 0) {
1201 // If all preds provide undef, just nuke the xor, because it is undef too.
1202 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1203 BO->eraseFromParent();
1204 } else if (SplitVal->isZero()) {
1205 // If all preds provide 0, replace the xor with the other input.
1206 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1207 BO->eraseFromParent();
1209 // If all preds provide 1, set the computed value to 1.
1210 BO->setOperand(!isLHS, SplitVal);
1216 // Try to duplicate BB into PredBB.
1217 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1221 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1222 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1223 /// NewPred using the entries from OldPred (suitably mapped).
1224 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1225 BasicBlock *OldPred,
1226 BasicBlock *NewPred,
1227 DenseMap<Instruction*, Value*> &ValueMap) {
1228 for (BasicBlock::iterator PNI = PHIBB->begin();
1229 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1230 // Ok, we have a PHI node. Figure out what the incoming value was for the
1232 Value *IV = PN->getIncomingValueForBlock(OldPred);
1234 // Remap the value if necessary.
1235 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1236 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1237 if (I != ValueMap.end())
1241 PN->addIncoming(IV, NewPred);
1245 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1246 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1247 /// across BB. Transform the IR to reflect this change.
1248 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1249 const SmallVectorImpl<BasicBlock*> &PredBBs,
1250 BasicBlock *SuccBB) {
1251 // If threading to the same block as we come from, we would infinite loop.
1253 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1254 << "' - would thread to self!\n");
1258 // If threading this would thread across a loop header, don't thread the edge.
1259 // See the comments above FindLoopHeaders for justifications and caveats.
1260 if (LoopHeaders.count(BB)) {
1261 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1262 << "' to dest BB '" << SuccBB->getName()
1263 << "' - it might create an irreducible loop!\n");
1267 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1268 if (JumpThreadCost > Threshold) {
1269 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1270 << "' - Cost is too high: " << JumpThreadCost << "\n");
1274 // And finally, do it! Start by factoring the predecessors is needed.
1276 if (PredBBs.size() == 1)
1277 PredBB = PredBBs[0];
1279 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1280 << " common predecessors.\n");
1281 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1285 // And finally, do it!
1286 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1287 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1288 << ", across block:\n "
1291 // We are going to have to map operands from the original BB block to the new
1292 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1293 // account for entry from PredBB.
1294 DenseMap<Instruction*, Value*> ValueMapping;
1296 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1297 BB->getName()+".thread",
1298 BB->getParent(), BB);
1299 NewBB->moveAfter(PredBB);
1301 BasicBlock::iterator BI = BB->begin();
1302 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1303 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1305 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1306 // mapping and using it to remap operands in the cloned instructions.
1307 for (; !isa<TerminatorInst>(BI); ++BI) {
1308 Instruction *New = BI->clone();
1309 New->setName(BI->getName());
1310 NewBB->getInstList().push_back(New);
1311 ValueMapping[BI] = New;
1313 // Remap operands to patch up intra-block references.
1314 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1315 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1316 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1317 if (I != ValueMapping.end())
1318 New->setOperand(i, I->second);
1322 // We didn't copy the terminator from BB over to NewBB, because there is now
1323 // an unconditional jump to SuccBB. Insert the unconditional jump.
1324 BranchInst::Create(SuccBB, NewBB);
1326 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1327 // PHI nodes for NewBB now.
1328 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1330 // If there were values defined in BB that are used outside the block, then we
1331 // now have to update all uses of the value to use either the original value,
1332 // the cloned value, or some PHI derived value. This can require arbitrary
1333 // PHI insertion, of which we are prepared to do, clean these up now.
1334 SSAUpdater SSAUpdate;
1335 SmallVector<Use*, 16> UsesToRename;
1336 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1337 // Scan all uses of this instruction to see if it is used outside of its
1338 // block, and if so, record them in UsesToRename.
1339 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1341 Instruction *User = cast<Instruction>(*UI);
1342 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1343 if (UserPN->getIncomingBlock(UI) == BB)
1345 } else if (User->getParent() == BB)
1348 UsesToRename.push_back(&UI.getUse());
1351 // If there are no uses outside the block, we're done with this instruction.
1352 if (UsesToRename.empty())
1355 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1357 // We found a use of I outside of BB. Rename all uses of I that are outside
1358 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1359 // with the two values we know.
1360 SSAUpdate.Initialize(I);
1361 SSAUpdate.AddAvailableValue(BB, I);
1362 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1364 while (!UsesToRename.empty())
1365 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1366 DEBUG(dbgs() << "\n");
1370 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1371 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1372 // us to simplify any PHI nodes in BB.
1373 TerminatorInst *PredTerm = PredBB->getTerminator();
1374 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1375 if (PredTerm->getSuccessor(i) == BB) {
1376 RemovePredecessorAndSimplify(BB, PredBB, TD);
1377 PredTerm->setSuccessor(i, NewBB);
1380 // At this point, the IR is fully up to date and consistent. Do a quick scan
1381 // over the new instructions and zap any that are constants or dead. This
1382 // frequently happens because of phi translation.
1383 SimplifyInstructionsInBlock(NewBB, TD);
1385 // Threaded an edge!
1390 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1391 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1392 /// If we can duplicate the contents of BB up into PredBB do so now, this
1393 /// improves the odds that the branch will be on an analyzable instruction like
1395 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1396 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1397 assert(!PredBBs.empty() && "Can't handle an empty set");
1399 // If BB is a loop header, then duplicating this block outside the loop would
1400 // cause us to transform this into an irreducible loop, don't do this.
1401 // See the comments above FindLoopHeaders for justifications and caveats.
1402 if (LoopHeaders.count(BB)) {
1403 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1404 << "' into predecessor block '" << PredBBs[0]->getName()
1405 << "' - it might create an irreducible loop!\n");
1409 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1410 if (DuplicationCost > Threshold) {
1411 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1412 << "' - Cost is too high: " << DuplicationCost << "\n");
1416 // And finally, do it! Start by factoring the predecessors is needed.
1418 if (PredBBs.size() == 1)
1419 PredBB = PredBBs[0];
1421 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1422 << " common predecessors.\n");
1423 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1427 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1429 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1430 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1431 << DuplicationCost << " block is:" << *BB << "\n");
1433 // Unless PredBB ends with an unconditional branch, split the edge so that we
1434 // can just clone the bits from BB into the end of the new PredBB.
1435 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1437 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1438 PredBB = SplitEdge(PredBB, BB, this);
1439 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1442 // We are going to have to map operands from the original BB block into the
1443 // PredBB block. Evaluate PHI nodes in BB.
1444 DenseMap<Instruction*, Value*> ValueMapping;
1446 BasicBlock::iterator BI = BB->begin();
1447 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1448 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1450 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1451 // mapping and using it to remap operands in the cloned instructions.
1452 for (; BI != BB->end(); ++BI) {
1453 Instruction *New = BI->clone();
1455 // Remap operands to patch up intra-block references.
1456 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1457 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1458 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1459 if (I != ValueMapping.end())
1460 New->setOperand(i, I->second);
1463 // If this instruction can be simplified after the operands are updated,
1464 // just use the simplified value instead. This frequently happens due to
1466 if (Value *IV = SimplifyInstruction(New, TD)) {
1468 ValueMapping[BI] = IV;
1470 // Otherwise, insert the new instruction into the block.
1471 New->setName(BI->getName());
1472 PredBB->getInstList().insert(OldPredBranch, New);
1473 ValueMapping[BI] = New;
1477 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1478 // add entries to the PHI nodes for branch from PredBB now.
1479 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1480 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1482 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1485 // If there were values defined in BB that are used outside the block, then we
1486 // now have to update all uses of the value to use either the original value,
1487 // the cloned value, or some PHI derived value. This can require arbitrary
1488 // PHI insertion, of which we are prepared to do, clean these up now.
1489 SSAUpdater SSAUpdate;
1490 SmallVector<Use*, 16> UsesToRename;
1491 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1492 // Scan all uses of this instruction to see if it is used outside of its
1493 // block, and if so, record them in UsesToRename.
1494 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1496 Instruction *User = cast<Instruction>(*UI);
1497 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1498 if (UserPN->getIncomingBlock(UI) == BB)
1500 } else if (User->getParent() == BB)
1503 UsesToRename.push_back(&UI.getUse());
1506 // If there are no uses outside the block, we're done with this instruction.
1507 if (UsesToRename.empty())
1510 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1512 // We found a use of I outside of BB. Rename all uses of I that are outside
1513 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1514 // with the two values we know.
1515 SSAUpdate.Initialize(I);
1516 SSAUpdate.AddAvailableValue(BB, I);
1517 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1519 while (!UsesToRename.empty())
1520 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1521 DEBUG(dbgs() << "\n");
1524 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1526 RemovePredecessorAndSimplify(BB, PredBB, TD);
1528 // Remove the unconditional branch at the end of the PredBB block.
1529 OldPredBranch->eraseFromParent();