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/ConstantFolding.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/LazyValueInfo.h"
22 #include "llvm/Analysis/Loads.h"
23 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
24 #include "llvm/Transforms/Utils/Local.h"
25 #include "llvm/Transforms/Utils/SSAUpdater.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/ADT/DenseMap.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/ADT/STLExtras.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/SmallSet.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ValueHandle.h"
35 #include "llvm/Support/raw_ostream.h"
38 STATISTIC(NumThreads, "Number of jumps threaded");
39 STATISTIC(NumFolds, "Number of terminators folded");
40 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
42 static cl::opt<unsigned>
43 Threshold("jump-threading-threshold",
44 cl::desc("Max block size to duplicate for jump threading"),
45 cl::init(6), cl::Hidden);
47 // Turn on use of LazyValueInfo.
49 EnableLVI("enable-jump-threading-lvi",
50 cl::desc("Use LVI for jump threading"),
57 /// This pass performs 'jump threading', which looks at blocks that have
58 /// multiple predecessors and multiple successors. If one or more of the
59 /// predecessors of the block can be proven to always jump to one of the
60 /// successors, we forward the edge from the predecessor to the successor by
61 /// duplicating the contents of this block.
63 /// An example of when this can occur is code like this:
70 /// In this case, the unconditional branch at the end of the first if can be
71 /// revectored to the false side of the second if.
73 class JumpThreading : public FunctionPass {
77 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
79 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
82 static char ID; // Pass identification
83 JumpThreading() : FunctionPass(ID) {}
85 bool runOnFunction(Function &F);
87 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
89 AU.addRequired<LazyValueInfo>();
92 void FindLoopHeaders(Function &F);
93 bool ProcessBlock(BasicBlock *BB);
94 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
96 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
97 const SmallVectorImpl<BasicBlock *> &PredBBs);
99 typedef SmallVectorImpl<std::pair<ConstantInt*,
100 BasicBlock*> > PredValueInfo;
102 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
103 PredValueInfo &Result);
104 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB);
107 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
108 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
110 bool ProcessBranchOnPHI(PHINode *PN);
111 bool ProcessBranchOnXOR(BinaryOperator *BO);
113 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
117 char JumpThreading::ID = 0;
118 INITIALIZE_PASS(JumpThreading, "jump-threading",
119 "Jump Threading", false, false);
121 // Public interface to the Jump Threading pass
122 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
124 /// runOnFunction - Top level algorithm.
126 bool JumpThreading::runOnFunction(Function &F) {
127 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
128 TD = getAnalysisIfAvailable<TargetData>();
129 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0;
133 bool Changed, EverChanged = false;
136 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
138 // Thread all of the branches we can over this block.
139 while (ProcessBlock(BB))
144 // If the block is trivially dead, zap it. This eliminates the successor
145 // edges which simplifies the CFG.
146 if (pred_begin(BB) == pred_end(BB) &&
147 BB != &BB->getParent()->getEntryBlock()) {
148 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
149 << "' with terminator: " << *BB->getTerminator() << '\n');
150 LoopHeaders.erase(BB);
151 if (LVI) LVI->eraseBlock(BB);
154 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
155 // Can't thread an unconditional jump, but if the block is "almost
156 // empty", we can replace uses of it with uses of the successor and make
158 if (BI->isUnconditional() &&
159 BB != &BB->getParent()->getEntryBlock()) {
160 BasicBlock::iterator BBI = BB->getFirstNonPHI();
161 // Ignore dbg intrinsics.
162 while (isa<DbgInfoIntrinsic>(BBI))
164 // If the terminator is the only non-phi instruction, try to nuke it.
165 if (BBI->isTerminator()) {
166 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
167 // block, we have to make sure it isn't in the LoopHeaders set. We
168 // reinsert afterward if needed.
169 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
170 BasicBlock *Succ = BI->getSuccessor(0);
172 // FIXME: It is always conservatively correct to drop the info
173 // for a block even if it doesn't get erased. This isn't totally
174 // awesome, but it allows us to use AssertingVH to prevent nasty
175 // dangling pointer issues within LazyValueInfo.
176 if (LVI) LVI->eraseBlock(BB);
177 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
179 // If we deleted BB and BB was the header of a loop, then the
180 // successor is now the header of the loop.
184 if (ErasedFromLoopHeaders)
185 LoopHeaders.insert(BB);
190 EverChanged |= Changed;
197 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
198 /// thread across it.
199 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
200 /// Ignore PHI nodes, these will be flattened when duplication happens.
201 BasicBlock::const_iterator I = BB->getFirstNonPHI();
203 // FIXME: THREADING will delete values that are just used to compute the
204 // branch, so they shouldn't count against the duplication cost.
207 // Sum up the cost of each instruction until we get to the terminator. Don't
208 // include the terminator because the copy won't include it.
210 for (; !isa<TerminatorInst>(I); ++I) {
211 // Debugger intrinsics don't incur code size.
212 if (isa<DbgInfoIntrinsic>(I)) continue;
214 // If this is a pointer->pointer bitcast, it is free.
215 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
218 // All other instructions count for at least one unit.
221 // Calls are more expensive. If they are non-intrinsic calls, we model them
222 // as having cost of 4. If they are a non-vector intrinsic, we model them
223 // as having cost of 2 total, and if they are a vector intrinsic, we model
224 // them as having cost 1.
225 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
226 if (!isa<IntrinsicInst>(CI))
228 else if (!CI->getType()->isVectorTy())
233 // Threading through a switch statement is particularly profitable. If this
234 // block ends in a switch, decrease its cost to make it more likely to happen.
235 if (isa<SwitchInst>(I))
236 Size = Size > 6 ? Size-6 : 0;
241 /// FindLoopHeaders - We do not want jump threading to turn proper loop
242 /// structures into irreducible loops. Doing this breaks up the loop nesting
243 /// hierarchy and pessimizes later transformations. To prevent this from
244 /// happening, we first have to find the loop headers. Here we approximate this
245 /// by finding targets of backedges in the CFG.
247 /// Note that there definitely are cases when we want to allow threading of
248 /// edges across a loop header. For example, threading a jump from outside the
249 /// loop (the preheader) to an exit block of the loop is definitely profitable.
250 /// It is also almost always profitable to thread backedges from within the loop
251 /// to exit blocks, and is often profitable to thread backedges to other blocks
252 /// within the loop (forming a nested loop). This simple analysis is not rich
253 /// enough to track all of these properties and keep it up-to-date as the CFG
254 /// mutates, so we don't allow any of these transformations.
256 void JumpThreading::FindLoopHeaders(Function &F) {
257 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
258 FindFunctionBackedges(F, Edges);
260 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
261 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
264 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
265 /// if we can infer that the value is a known ConstantInt in any of our
266 /// predecessors. If so, return the known list of value and pred BB in the
267 /// result vector. If a value is known to be undef, it is returned as null.
269 /// This returns true if there were any known values.
272 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){
273 // If V is a constantint, then it is known in all predecessors.
274 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) {
275 ConstantInt *CI = dyn_cast<ConstantInt>(V);
277 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
278 Result.push_back(std::make_pair(CI, *PI));
282 // If V is a non-instruction value, or an instruction in a different block,
283 // then it can't be derived from a PHI.
284 Instruction *I = dyn_cast<Instruction>(V);
285 if (I == 0 || I->getParent() != BB) {
287 // Okay, if this is a live-in value, see if it has a known value at the end
288 // of any of our predecessors.
290 // FIXME: This should be an edge property, not a block end property.
291 /// TODO: Per PR2563, we could infer value range information about a
292 /// predecessor based on its terminator.
295 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
296 // "I" is a non-local compare-with-a-constant instruction. This would be
297 // able to handle value inequalities better, for example if the compare is
298 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
299 // Perhaps getConstantOnEdge should be smart enough to do this?
301 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
303 // If the value is known by LazyValueInfo to be a constant in a
304 // predecessor, use that information to try to thread this block.
305 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
307 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst)))
310 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), P));
313 return !Result.empty();
319 /// If I is a PHI node, then we know the incoming values for any constants.
320 if (PHINode *PN = dyn_cast<PHINode>(I)) {
321 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
322 Value *InVal = PN->getIncomingValue(i);
323 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) {
324 ConstantInt *CI = dyn_cast<ConstantInt>(InVal);
325 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i)));
327 Constant *CI = LVI->getConstantOnEdge(InVal,
328 PN->getIncomingBlock(i), BB);
329 // LVI returns null is no value could be determined.
331 ConstantInt *CInt = dyn_cast<ConstantInt>(CI);
332 Result.push_back(std::make_pair(CInt, PN->getIncomingBlock(i)));
335 return !Result.empty();
338 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals;
340 // Handle some boolean conditions.
341 if (I->getType()->getPrimitiveSizeInBits() == 1) {
343 // X & false -> false
344 if (I->getOpcode() == Instruction::Or ||
345 I->getOpcode() == Instruction::And) {
346 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
347 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals);
349 if (LHSVals.empty() && RHSVals.empty())
352 ConstantInt *InterestingVal;
353 if (I->getOpcode() == Instruction::Or)
354 InterestingVal = ConstantInt::getTrue(I->getContext());
356 InterestingVal = ConstantInt::getFalse(I->getContext());
358 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
360 // Scan for the sentinel. If we find an undef, force it to the
361 // interesting value: x|undef -> true and x&undef -> false.
362 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
363 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) {
364 Result.push_back(LHSVals[i]);
365 Result.back().first = InterestingVal;
366 LHSKnownBBs.insert(LHSVals[i].second);
368 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
369 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) {
370 // If we already inferred a value for this block on the LHS, don't
372 if (!LHSKnownBBs.count(RHSVals[i].second)) {
373 Result.push_back(RHSVals[i]);
374 Result.back().first = InterestingVal;
377 return !Result.empty();
380 // Handle the NOT form of XOR.
381 if (I->getOpcode() == Instruction::Xor &&
382 isa<ConstantInt>(I->getOperand(1)) &&
383 cast<ConstantInt>(I->getOperand(1))->isOne()) {
384 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result);
388 // Invert the known values.
389 for (unsigned i = 0, e = Result.size(); i != e; ++i)
392 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first));
396 // Try to simplify some other binary operator values.
397 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
398 // AND or OR of a value with itself is that value.
399 ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1));
400 if (CI && (BO->getOpcode() == Instruction::And ||
401 BO->getOpcode() == Instruction::Or)) {
402 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals;
403 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals);
404 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
405 if (LHSVals[i].first == 0) {
407 cast<ConstantInt>(ConstantInt::get(BO->getType(), 0));
408 Result.push_back(std::make_pair(Zero, LHSVals[i].second));
409 } else if (Constant *Folded = ConstantExpr::get(BO->getOpcode(),
410 LHSVals[i].first, CI)) {
411 Result.push_back(std::make_pair(cast<ConstantInt>(Folded),
415 return !Result.empty();
419 // Handle compare with phi operand, where the PHI is defined in this block.
420 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
421 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
422 if (PN && PN->getParent() == BB) {
423 // We can do this simplification if any comparisons fold to true or false.
425 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
426 BasicBlock *PredBB = PN->getIncomingBlock(i);
427 Value *LHS = PN->getIncomingValue(i);
428 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
430 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
432 if (!LVI || !isa<Constant>(RHS))
435 LazyValueInfo::Tristate
436 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
437 cast<Constant>(RHS), PredBB, BB);
438 if (ResT == LazyValueInfo::Unknown)
440 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
443 if (isa<UndefValue>(Res))
444 Result.push_back(std::make_pair((ConstantInt*)0, PredBB));
445 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res))
446 Result.push_back(std::make_pair(CI, PredBB));
449 return !Result.empty();
453 // If comparing a live-in value against a constant, see if we know the
454 // live-in value on any predecessors.
455 if (LVI && isa<Constant>(Cmp->getOperand(1)) &&
456 Cmp->getType()->isIntegerTy()) {
457 if (!isa<Instruction>(Cmp->getOperand(0)) ||
458 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
459 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
461 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
463 // If the value is known by LazyValueInfo to be a constant in a
464 // predecessor, use that information to try to thread this block.
465 LazyValueInfo::Tristate Res =
466 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
468 if (Res == LazyValueInfo::Unknown)
471 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
472 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), P));
475 return !Result.empty();
478 // Try to find a constant value for the LHS of an equality comparison,
479 // and evaluate it statically if we can.
480 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
481 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals;
482 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
484 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
485 if (LHSVals[i].first == 0)
486 Result.push_back(std::make_pair((ConstantInt*)0,
488 else if (Constant *Folded = ConstantExpr::getCompare(
489 Cmp->getPredicate(), LHSVals[i].first, CmpConst))
490 Result.push_back(std::make_pair(cast<ConstantInt>(Folded),
494 return !Result.empty();
500 // If all else fails, see if LVI can figure out a constant value for us.
501 Constant *CI = LVI->getConstant(V, BB);
502 ConstantInt *CInt = dyn_cast_or_null<ConstantInt>(CI);
504 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
505 Result.push_back(std::make_pair(CInt, *PI));
508 return !Result.empty();
516 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
517 /// in an undefined jump, decide which block is best to revector to.
519 /// Since we can pick an arbitrary destination, we pick the successor with the
520 /// fewest predecessors. This should reduce the in-degree of the others.
522 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
523 TerminatorInst *BBTerm = BB->getTerminator();
524 unsigned MinSucc = 0;
525 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
526 // Compute the successor with the minimum number of predecessors.
527 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
528 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
529 TestBB = BBTerm->getSuccessor(i);
530 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
531 if (NumPreds < MinNumPreds)
538 /// ProcessBlock - If there are any predecessors whose control can be threaded
539 /// through to a successor, transform them now.
540 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
541 // If the block is trivially dead, just return and let the caller nuke it.
542 // This simplifies other transformations.
543 if (pred_begin(BB) == pred_end(BB) &&
544 BB != &BB->getParent()->getEntryBlock())
547 // If this block has a single predecessor, and if that pred has a single
548 // successor, merge the blocks. This encourages recursive jump threading
549 // because now the condition in this block can be threaded through
550 // predecessors of our predecessor block.
551 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
552 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
554 // If SinglePred was a loop header, BB becomes one.
555 if (LoopHeaders.erase(SinglePred))
556 LoopHeaders.insert(BB);
558 // Remember if SinglePred was the entry block of the function. If so, we
559 // will need to move BB back to the entry position.
560 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
561 if (LVI) LVI->eraseBlock(SinglePred);
562 MergeBasicBlockIntoOnlyPred(BB);
564 if (isEntry && BB != &BB->getParent()->getEntryBlock())
565 BB->moveBefore(&BB->getParent()->getEntryBlock());
570 // Look to see if the terminator is a branch of switch, if not we can't thread
573 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
574 // Can't thread an unconditional jump.
575 if (BI->isUnconditional()) return false;
576 Condition = BI->getCondition();
577 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
578 Condition = SI->getCondition();
580 return false; // Must be an invoke.
582 // If the terminator of this block is branching on a constant, simplify the
583 // terminator to an unconditional branch. This can occur due to threading in
585 if (isa<ConstantInt>(Condition)) {
586 DEBUG(dbgs() << " In block '" << BB->getName()
587 << "' folding terminator: " << *BB->getTerminator() << '\n');
589 ConstantFoldTerminator(BB);
593 // If the terminator is branching on an undef, we can pick any of the
594 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
595 if (isa<UndefValue>(Condition)) {
596 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
598 // Fold the branch/switch.
599 TerminatorInst *BBTerm = BB->getTerminator();
600 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
601 if (i == BestSucc) continue;
602 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD);
605 DEBUG(dbgs() << " In block '" << BB->getName()
606 << "' folding undef terminator: " << *BBTerm << '\n');
607 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
608 BBTerm->eraseFromParent();
612 Instruction *CondInst = dyn_cast<Instruction>(Condition);
614 // If the condition is an instruction defined in another block, see if a
615 // predecessor has the same condition:
620 !Condition->hasOneUse() && // Multiple uses.
621 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition.
622 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
623 if (isa<BranchInst>(BB->getTerminator())) {
624 for (; PI != E; ++PI) {
626 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator()))
627 if (PBI->isConditional() && PBI->getCondition() == Condition &&
628 ProcessBranchOnDuplicateCond(P, BB))
632 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator");
633 for (; PI != E; ++PI) {
635 if (SwitchInst *PSI = dyn_cast<SwitchInst>(P->getTerminator()))
636 if (PSI->getCondition() == Condition &&
637 ProcessSwitchOnDuplicateCond(P, BB))
643 // All the rest of our checks depend on the condition being an instruction.
645 // FIXME: Unify this with code below.
646 if (LVI && ProcessThreadableEdges(Condition, BB))
652 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
654 (!isa<PHINode>(CondCmp->getOperand(0)) ||
655 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) {
656 // If we have a comparison, loop over the predecessors to see if there is
657 // a condition with a lexically identical value.
658 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
659 for (; PI != E; ++PI) {
661 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator()))
662 if (PBI->isConditional() && P != BB) {
663 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) {
664 if (CI->getOperand(0) == CondCmp->getOperand(0) &&
665 CI->getOperand(1) == CondCmp->getOperand(1) &&
666 CI->getPredicate() == CondCmp->getPredicate()) {
667 // TODO: Could handle things like (x != 4) --> (x == 17)
668 if (ProcessBranchOnDuplicateCond(P, BB))
676 // For a comparison where the LHS is outside this block, it's possible
677 // that we've branched on it before. Used LVI to see if we can simplify
678 // the branch based on that.
679 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
680 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
681 if (LVI && CondBr && CondConst && CondBr->isConditional() &&
682 (!isa<Instruction>(CondCmp->getOperand(0)) ||
683 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
684 // For predecessor edge, determine if the comparison is true or false
685 // on that edge. If they're all true or all false, we can simplify the
687 // FIXME: We could handle mixed true/false by duplicating code.
688 unsigned Trues = 0, Falses = 0, predcount = 0;
689 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB);PI != PE; ++PI){
691 LazyValueInfo::Tristate Ret =
692 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
693 CondCmp->getOperand(0), CondConst, *PI, BB);
694 if (Ret == LazyValueInfo::True)
696 else if (Ret == LazyValueInfo::False)
700 // If we can determine the branch direction statically, convert
701 // the conditional branch to an unconditional one.
702 if (Trues && Trues == predcount) {
703 RemovePredecessorAndSimplify(CondBr->getSuccessor(1), BB, TD);
704 BranchInst::Create(CondBr->getSuccessor(0), CondBr);
705 CondBr->eraseFromParent();
707 } else if (Falses && Falses == predcount) {
708 RemovePredecessorAndSimplify(CondBr->getSuccessor(0), BB, TD);
709 BranchInst::Create(CondBr->getSuccessor(1), CondBr);
710 CondBr->eraseFromParent();
716 // Check for some cases that are worth simplifying. Right now we want to look
717 // for loads that are used by a switch or by the condition for the branch. If
718 // we see one, check to see if it's partially redundant. If so, insert a PHI
719 // which can then be used to thread the values.
721 Value *SimplifyValue = CondInst;
722 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
723 if (isa<Constant>(CondCmp->getOperand(1)))
724 SimplifyValue = CondCmp->getOperand(0);
726 // TODO: There are other places where load PRE would be profitable, such as
727 // more complex comparisons.
728 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
729 if (SimplifyPartiallyRedundantLoad(LI))
733 // Handle a variety of cases where we are branching on something derived from
734 // a PHI node in the current block. If we can prove that any predecessors
735 // compute a predictable value based on a PHI node, thread those predecessors.
737 if (ProcessThreadableEdges(CondInst, BB))
740 // If this is an otherwise-unfoldable branch on a phi node in the current
741 // block, see if we can simplify.
742 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
743 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
744 return ProcessBranchOnPHI(PN);
747 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
748 if (CondInst->getOpcode() == Instruction::Xor &&
749 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
750 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
753 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
754 // "(X == 4)", thread through this block.
759 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
760 /// block that jump on exactly the same condition. This means that we almost
761 /// always know the direction of the edge in the DESTBB:
763 /// br COND, DESTBB, BBY
765 /// br COND, BBZ, BBW
767 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
768 /// in DESTBB, we have to thread over it.
769 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
771 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
773 // If both successors of PredBB go to DESTBB, we don't know anything. We can
774 // fold the branch to an unconditional one, which allows other recursive
777 if (PredBI->getSuccessor(1) != BB)
779 else if (PredBI->getSuccessor(0) != BB)
782 DEBUG(dbgs() << " In block '" << PredBB->getName()
783 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
785 ConstantFoldTerminator(PredBB);
789 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
791 // If the dest block has one predecessor, just fix the branch condition to a
792 // constant and fold it.
793 if (BB->getSinglePredecessor()) {
794 DEBUG(dbgs() << " In block '" << BB->getName()
795 << "' folding condition to '" << BranchDir << "': "
796 << *BB->getTerminator() << '\n');
798 Value *OldCond = DestBI->getCondition();
799 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
801 // Delete dead instructions before we fold the branch. Folding the branch
802 // can eliminate edges from the CFG which can end up deleting OldCond.
803 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
804 ConstantFoldTerminator(BB);
809 // Next, figure out which successor we are threading to.
810 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
812 SmallVector<BasicBlock*, 2> Preds;
813 Preds.push_back(PredBB);
815 // Ok, try to thread it!
816 return ThreadEdge(BB, Preds, SuccBB);
819 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
820 /// block that switch on exactly the same condition. This means that we almost
821 /// always know the direction of the edge in the DESTBB:
823 /// switch COND [... DESTBB, BBY ... ]
825 /// switch COND [... BBZ, BBW ]
827 /// Optimizing switches like this is very important, because simplifycfg builds
828 /// switches out of repeated 'if' conditions.
829 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
830 BasicBlock *DestBB) {
831 // Can't thread edge to self.
832 if (PredBB == DestBB)
835 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
836 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
838 // There are a variety of optimizations that we can potentially do on these
839 // blocks: we order them from most to least preferable.
841 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
842 // directly to their destination. This does not introduce *any* code size
843 // growth. Skip debug info first.
844 BasicBlock::iterator BBI = DestBB->begin();
845 while (isa<DbgInfoIntrinsic>(BBI))
848 // FIXME: Thread if it just contains a PHI.
849 if (isa<SwitchInst>(BBI)) {
850 bool MadeChange = false;
851 // Ignore the default edge for now.
852 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
853 ConstantInt *DestVal = DestSI->getCaseValue(i);
854 BasicBlock *DestSucc = DestSI->getSuccessor(i);
856 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
857 // PredSI has an explicit case for it. If so, forward. If it is covered
858 // by the default case, we can't update PredSI.
859 unsigned PredCase = PredSI->findCaseValue(DestVal);
860 if (PredCase == 0) continue;
862 // If PredSI doesn't go to DestBB on this value, then it won't reach the
863 // case on this condition.
864 if (PredSI->getSuccessor(PredCase) != DestBB &&
865 DestSI->getSuccessor(i) != DestBB)
868 // Do not forward this if it already goes to this destination, this would
869 // be an infinite loop.
870 if (PredSI->getSuccessor(PredCase) == DestSucc)
873 // Otherwise, we're safe to make the change. Make sure that the edge from
874 // DestSI to DestSucc is not critical and has no PHI nodes.
875 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
876 DEBUG(dbgs() << "THROUGH: " << *DestSI);
878 // If the destination has PHI nodes, just split the edge for updating
880 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
881 SplitCriticalEdge(DestSI, i, this);
882 DestSucc = DestSI->getSuccessor(i);
884 FoldSingleEntryPHINodes(DestSucc);
885 PredSI->setSuccessor(PredCase, DestSucc);
897 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
898 /// load instruction, eliminate it by replacing it with a PHI node. This is an
899 /// important optimization that encourages jump threading, and needs to be run
900 /// interlaced with other jump threading tasks.
901 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
902 // Don't hack volatile loads.
903 if (LI->isVolatile()) return false;
905 // If the load is defined in a block with exactly one predecessor, it can't be
906 // partially redundant.
907 BasicBlock *LoadBB = LI->getParent();
908 if (LoadBB->getSinglePredecessor())
911 Value *LoadedPtr = LI->getOperand(0);
913 // If the loaded operand is defined in the LoadBB, it can't be available.
914 // TODO: Could do simple PHI translation, that would be fun :)
915 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
916 if (PtrOp->getParent() == LoadBB)
919 // Scan a few instructions up from the load, to see if it is obviously live at
920 // the entry to its block.
921 BasicBlock::iterator BBIt = LI;
923 if (Value *AvailableVal =
924 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
925 // If the value if the load is locally available within the block, just use
926 // it. This frequently occurs for reg2mem'd allocas.
927 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
929 // If the returned value is the load itself, replace with an undef. This can
930 // only happen in dead loops.
931 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
932 LI->replaceAllUsesWith(AvailableVal);
933 LI->eraseFromParent();
937 // Otherwise, if we scanned the whole block and got to the top of the block,
938 // we know the block is locally transparent to the load. If not, something
939 // might clobber its value.
940 if (BBIt != LoadBB->begin())
944 SmallPtrSet<BasicBlock*, 8> PredsScanned;
945 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
946 AvailablePredsTy AvailablePreds;
947 BasicBlock *OneUnavailablePred = 0;
949 // If we got here, the loaded value is transparent through to the start of the
950 // block. Check to see if it is available in any of the predecessor blocks.
951 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
953 BasicBlock *PredBB = *PI;
955 // If we already scanned this predecessor, skip it.
956 if (!PredsScanned.insert(PredBB))
959 // Scan the predecessor to see if the value is available in the pred.
960 BBIt = PredBB->end();
961 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
962 if (!PredAvailable) {
963 OneUnavailablePred = PredBB;
967 // If so, this load is partially redundant. Remember this info so that we
968 // can create a PHI node.
969 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
972 // If the loaded value isn't available in any predecessor, it isn't partially
974 if (AvailablePreds.empty()) return false;
976 // Okay, the loaded value is available in at least one (and maybe all!)
977 // predecessors. If the value is unavailable in more than one unique
978 // predecessor, we want to insert a merge block for those common predecessors.
979 // This ensures that we only have to insert one reload, thus not increasing
981 BasicBlock *UnavailablePred = 0;
983 // If there is exactly one predecessor where the value is unavailable, the
984 // already computed 'OneUnavailablePred' block is it. If it ends in an
985 // unconditional branch, we know that it isn't a critical edge.
986 if (PredsScanned.size() == AvailablePreds.size()+1 &&
987 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
988 UnavailablePred = OneUnavailablePred;
989 } else if (PredsScanned.size() != AvailablePreds.size()) {
990 // Otherwise, we had multiple unavailable predecessors or we had a critical
991 // edge from the one.
992 SmallVector<BasicBlock*, 8> PredsToSplit;
993 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
995 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
996 AvailablePredSet.insert(AvailablePreds[i].first);
998 // Add all the unavailable predecessors to the PredsToSplit list.
999 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
1001 BasicBlock *P = *PI;
1002 // If the predecessor is an indirect goto, we can't split the edge.
1003 if (isa<IndirectBrInst>(P->getTerminator()))
1006 if (!AvailablePredSet.count(P))
1007 PredsToSplit.push_back(P);
1010 // Split them out to their own block.
1012 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
1013 "thread-pre-split", this);
1016 // If the value isn't available in all predecessors, then there will be
1017 // exactly one where it isn't available. Insert a load on that edge and add
1018 // it to the AvailablePreds list.
1019 if (UnavailablePred) {
1020 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1021 "Can't handle critical edge here!");
1022 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
1024 UnavailablePred->getTerminator());
1025 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1028 // Now we know that each predecessor of this block has a value in
1029 // AvailablePreds, sort them for efficient access as we're walking the preds.
1030 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1032 // Create a PHI node at the start of the block for the PRE'd load value.
1033 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
1036 // Insert new entries into the PHI for each predecessor. A single block may
1037 // have multiple entries here.
1038 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
1040 BasicBlock *P = *PI;
1041 AvailablePredsTy::iterator I =
1042 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1043 std::make_pair(P, (Value*)0));
1045 assert(I != AvailablePreds.end() && I->first == P &&
1046 "Didn't find entry for predecessor!");
1048 PN->addIncoming(I->second, I->first);
1051 //cerr << "PRE: " << *LI << *PN << "\n";
1053 LI->replaceAllUsesWith(PN);
1054 LI->eraseFromParent();
1059 /// FindMostPopularDest - The specified list contains multiple possible
1060 /// threadable destinations. Pick the one that occurs the most frequently in
1063 FindMostPopularDest(BasicBlock *BB,
1064 const SmallVectorImpl<std::pair<BasicBlock*,
1065 BasicBlock*> > &PredToDestList) {
1066 assert(!PredToDestList.empty());
1068 // Determine popularity. If there are multiple possible destinations, we
1069 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1070 // blocks with known and real destinations to threading undef. We'll handle
1071 // them later if interesting.
1072 DenseMap<BasicBlock*, unsigned> DestPopularity;
1073 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1074 if (PredToDestList[i].second)
1075 DestPopularity[PredToDestList[i].second]++;
1077 // Find the most popular dest.
1078 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1079 BasicBlock *MostPopularDest = DPI->first;
1080 unsigned Popularity = DPI->second;
1081 SmallVector<BasicBlock*, 4> SamePopularity;
1083 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1084 // If the popularity of this entry isn't higher than the popularity we've
1085 // seen so far, ignore it.
1086 if (DPI->second < Popularity)
1088 else if (DPI->second == Popularity) {
1089 // If it is the same as what we've seen so far, keep track of it.
1090 SamePopularity.push_back(DPI->first);
1092 // If it is more popular, remember it.
1093 SamePopularity.clear();
1094 MostPopularDest = DPI->first;
1095 Popularity = DPI->second;
1099 // Okay, now we know the most popular destination. If there is more than
1100 // destination, we need to determine one. This is arbitrary, but we need
1101 // to make a deterministic decision. Pick the first one that appears in the
1103 if (!SamePopularity.empty()) {
1104 SamePopularity.push_back(MostPopularDest);
1105 TerminatorInst *TI = BB->getTerminator();
1106 for (unsigned i = 0; ; ++i) {
1107 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1109 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1110 TI->getSuccessor(i)) == SamePopularity.end())
1113 MostPopularDest = TI->getSuccessor(i);
1118 // Okay, we have finally picked the most popular destination.
1119 return MostPopularDest;
1122 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) {
1123 // If threading this would thread across a loop header, don't even try to
1125 if (LoopHeaders.count(BB))
1128 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues;
1129 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues))
1131 assert(!PredValues.empty() &&
1132 "ComputeValueKnownInPredecessors returned true with no values");
1134 DEBUG(dbgs() << "IN BB: " << *BB;
1135 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1136 dbgs() << " BB '" << BB->getName() << "': FOUND condition = ";
1137 if (PredValues[i].first)
1138 dbgs() << *PredValues[i].first;
1141 dbgs() << " for pred '" << PredValues[i].second->getName()
1145 // Decide what we want to thread through. Convert our list of known values to
1146 // a list of known destinations for each pred. This also discards duplicate
1147 // predecessors and keeps track of the undefined inputs (which are represented
1148 // as a null dest in the PredToDestList).
1149 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1150 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1152 BasicBlock *OnlyDest = 0;
1153 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1155 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1156 BasicBlock *Pred = PredValues[i].second;
1157 if (!SeenPreds.insert(Pred))
1158 continue; // Duplicate predecessor entry.
1160 // If the predecessor ends with an indirect goto, we can't change its
1162 if (isa<IndirectBrInst>(Pred->getTerminator()))
1165 ConstantInt *Val = PredValues[i].first;
1168 if (Val == 0) // Undef.
1170 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1171 DestBB = BI->getSuccessor(Val->isZero());
1173 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
1174 DestBB = SI->getSuccessor(SI->findCaseValue(Val));
1177 // If we have exactly one destination, remember it for efficiency below.
1180 else if (OnlyDest != DestBB)
1181 OnlyDest = MultipleDestSentinel;
1183 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1186 // If all edges were unthreadable, we fail.
1187 if (PredToDestList.empty())
1190 // Determine which is the most common successor. If we have many inputs and
1191 // this block is a switch, we want to start by threading the batch that goes
1192 // to the most popular destination first. If we only know about one
1193 // threadable destination (the common case) we can avoid this.
1194 BasicBlock *MostPopularDest = OnlyDest;
1196 if (MostPopularDest == MultipleDestSentinel)
1197 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1199 // Now that we know what the most popular destination is, factor all
1200 // predecessors that will jump to it into a single predecessor.
1201 SmallVector<BasicBlock*, 16> PredsToFactor;
1202 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1203 if (PredToDestList[i].second == MostPopularDest) {
1204 BasicBlock *Pred = PredToDestList[i].first;
1206 // This predecessor may be a switch or something else that has multiple
1207 // edges to the block. Factor each of these edges by listing them
1208 // according to # occurrences in PredsToFactor.
1209 TerminatorInst *PredTI = Pred->getTerminator();
1210 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1211 if (PredTI->getSuccessor(i) == BB)
1212 PredsToFactor.push_back(Pred);
1215 // If the threadable edges are branching on an undefined value, we get to pick
1216 // the destination that these predecessors should get to.
1217 if (MostPopularDest == 0)
1218 MostPopularDest = BB->getTerminator()->
1219 getSuccessor(GetBestDestForJumpOnUndef(BB));
1221 // Ok, try to thread it!
1222 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1225 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1226 /// a PHI node in the current block. See if there are any simplifications we
1227 /// can do based on inputs to the phi node.
1229 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1230 BasicBlock *BB = PN->getParent();
1232 // TODO: We could make use of this to do it once for blocks with common PHI
1234 SmallVector<BasicBlock*, 1> PredBBs;
1237 // If any of the predecessor blocks end in an unconditional branch, we can
1238 // *duplicate* the conditional branch into that block in order to further
1239 // encourage jump threading and to eliminate cases where we have branch on a
1240 // phi of an icmp (branch on icmp is much better).
1241 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1242 BasicBlock *PredBB = PN->getIncomingBlock(i);
1243 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1244 if (PredBr->isUnconditional()) {
1245 PredBBs[0] = PredBB;
1246 // Try to duplicate BB into PredBB.
1247 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1255 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1256 /// a xor instruction in the current block. See if there are any
1257 /// simplifications we can do based on inputs to the xor.
1259 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1260 BasicBlock *BB = BO->getParent();
1262 // If either the LHS or RHS of the xor is a constant, don't do this
1264 if (isa<ConstantInt>(BO->getOperand(0)) ||
1265 isa<ConstantInt>(BO->getOperand(1)))
1268 // If the first instruction in BB isn't a phi, we won't be able to infer
1269 // anything special about any particular predecessor.
1270 if (!isa<PHINode>(BB->front()))
1273 // If we have a xor as the branch input to this block, and we know that the
1274 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1275 // the condition into the predecessor and fix that value to true, saving some
1276 // logical ops on that path and encouraging other paths to simplify.
1278 // This copies something like this:
1281 // %X = phi i1 [1], [%X']
1282 // %Y = icmp eq i32 %A, %B
1283 // %Z = xor i1 %X, %Y
1288 // %Y = icmp ne i32 %A, %B
1291 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues;
1293 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) {
1294 assert(XorOpValues.empty());
1295 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues))
1300 assert(!XorOpValues.empty() &&
1301 "ComputeValueKnownInPredecessors returned true with no values");
1303 // Scan the information to see which is most popular: true or false. The
1304 // predecessors can be of the set true, false, or undef.
1305 unsigned NumTrue = 0, NumFalse = 0;
1306 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1307 if (!XorOpValues[i].first) continue; // Ignore undefs for the count.
1308 if (XorOpValues[i].first->isZero())
1314 // Determine which value to split on, true, false, or undef if neither.
1315 ConstantInt *SplitVal = 0;
1316 if (NumTrue > NumFalse)
1317 SplitVal = ConstantInt::getTrue(BB->getContext());
1318 else if (NumTrue != 0 || NumFalse != 0)
1319 SplitVal = ConstantInt::getFalse(BB->getContext());
1321 // Collect all of the blocks that this can be folded into so that we can
1322 // factor this once and clone it once.
1323 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1324 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1325 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue;
1327 BlocksToFoldInto.push_back(XorOpValues[i].second);
1330 // If we inferred a value for all of the predecessors, then duplication won't
1331 // help us. However, we can just replace the LHS or RHS with the constant.
1332 if (BlocksToFoldInto.size() ==
1333 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1334 if (SplitVal == 0) {
1335 // If all preds provide undef, just nuke the xor, because it is undef too.
1336 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1337 BO->eraseFromParent();
1338 } else if (SplitVal->isZero()) {
1339 // If all preds provide 0, replace the xor with the other input.
1340 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1341 BO->eraseFromParent();
1343 // If all preds provide 1, set the computed value to 1.
1344 BO->setOperand(!isLHS, SplitVal);
1350 // Try to duplicate BB into PredBB.
1351 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1355 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1356 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1357 /// NewPred using the entries from OldPred (suitably mapped).
1358 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1359 BasicBlock *OldPred,
1360 BasicBlock *NewPred,
1361 DenseMap<Instruction*, Value*> &ValueMap) {
1362 for (BasicBlock::iterator PNI = PHIBB->begin();
1363 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1364 // Ok, we have a PHI node. Figure out what the incoming value was for the
1366 Value *IV = PN->getIncomingValueForBlock(OldPred);
1368 // Remap the value if necessary.
1369 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1370 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1371 if (I != ValueMap.end())
1375 PN->addIncoming(IV, NewPred);
1379 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1380 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1381 /// across BB. Transform the IR to reflect this change.
1382 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1383 const SmallVectorImpl<BasicBlock*> &PredBBs,
1384 BasicBlock *SuccBB) {
1385 // If threading to the same block as we come from, we would infinite loop.
1387 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1388 << "' - would thread to self!\n");
1392 // If threading this would thread across a loop header, don't thread the edge.
1393 // See the comments above FindLoopHeaders for justifications and caveats.
1394 if (LoopHeaders.count(BB)) {
1395 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1396 << "' to dest BB '" << SuccBB->getName()
1397 << "' - it might create an irreducible loop!\n");
1401 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1402 if (JumpThreadCost > Threshold) {
1403 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1404 << "' - Cost is too high: " << JumpThreadCost << "\n");
1408 // And finally, do it! Start by factoring the predecessors is needed.
1410 if (PredBBs.size() == 1)
1411 PredBB = PredBBs[0];
1413 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1414 << " common predecessors.\n");
1415 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1419 // And finally, do it!
1420 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1421 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1422 << ", across block:\n "
1426 LVI->threadEdge(PredBB, BB, SuccBB);
1428 // We are going to have to map operands from the original BB block to the new
1429 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1430 // account for entry from PredBB.
1431 DenseMap<Instruction*, Value*> ValueMapping;
1433 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1434 BB->getName()+".thread",
1435 BB->getParent(), BB);
1436 NewBB->moveAfter(PredBB);
1438 BasicBlock::iterator BI = BB->begin();
1439 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1440 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1442 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1443 // mapping and using it to remap operands in the cloned instructions.
1444 for (; !isa<TerminatorInst>(BI); ++BI) {
1445 Instruction *New = BI->clone();
1446 New->setName(BI->getName());
1447 NewBB->getInstList().push_back(New);
1448 ValueMapping[BI] = New;
1450 // Remap operands to patch up intra-block references.
1451 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1452 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1453 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1454 if (I != ValueMapping.end())
1455 New->setOperand(i, I->second);
1459 // We didn't copy the terminator from BB over to NewBB, because there is now
1460 // an unconditional jump to SuccBB. Insert the unconditional jump.
1461 BranchInst::Create(SuccBB, NewBB);
1463 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1464 // PHI nodes for NewBB now.
1465 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1467 // If there were values defined in BB that are used outside the block, then we
1468 // now have to update all uses of the value to use either the original value,
1469 // the cloned value, or some PHI derived value. This can require arbitrary
1470 // PHI insertion, of which we are prepared to do, clean these up now.
1471 SSAUpdater SSAUpdate;
1472 SmallVector<Use*, 16> UsesToRename;
1473 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1474 // Scan all uses of this instruction to see if it is used outside of its
1475 // block, and if so, record them in UsesToRename.
1476 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1478 Instruction *User = cast<Instruction>(*UI);
1479 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1480 if (UserPN->getIncomingBlock(UI) == BB)
1482 } else if (User->getParent() == BB)
1485 UsesToRename.push_back(&UI.getUse());
1488 // If there are no uses outside the block, we're done with this instruction.
1489 if (UsesToRename.empty())
1492 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1494 // We found a use of I outside of BB. Rename all uses of I that are outside
1495 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1496 // with the two values we know.
1497 SSAUpdate.Initialize(I);
1498 SSAUpdate.AddAvailableValue(BB, I);
1499 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1501 while (!UsesToRename.empty())
1502 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1503 DEBUG(dbgs() << "\n");
1507 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1508 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1509 // us to simplify any PHI nodes in BB.
1510 TerminatorInst *PredTerm = PredBB->getTerminator();
1511 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1512 if (PredTerm->getSuccessor(i) == BB) {
1513 RemovePredecessorAndSimplify(BB, PredBB, TD);
1514 PredTerm->setSuccessor(i, NewBB);
1517 // At this point, the IR is fully up to date and consistent. Do a quick scan
1518 // over the new instructions and zap any that are constants or dead. This
1519 // frequently happens because of phi translation.
1520 SimplifyInstructionsInBlock(NewBB, TD);
1522 // Threaded an edge!
1527 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1528 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1529 /// If we can duplicate the contents of BB up into PredBB do so now, this
1530 /// improves the odds that the branch will be on an analyzable instruction like
1532 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1533 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1534 assert(!PredBBs.empty() && "Can't handle an empty set");
1536 // If BB is a loop header, then duplicating this block outside the loop would
1537 // cause us to transform this into an irreducible loop, don't do this.
1538 // See the comments above FindLoopHeaders for justifications and caveats.
1539 if (LoopHeaders.count(BB)) {
1540 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1541 << "' into predecessor block '" << PredBBs[0]->getName()
1542 << "' - it might create an irreducible loop!\n");
1546 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1547 if (DuplicationCost > Threshold) {
1548 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1549 << "' - Cost is too high: " << DuplicationCost << "\n");
1553 // And finally, do it! Start by factoring the predecessors is needed.
1555 if (PredBBs.size() == 1)
1556 PredBB = PredBBs[0];
1558 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1559 << " common predecessors.\n");
1560 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1564 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1566 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1567 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1568 << DuplicationCost << " block is:" << *BB << "\n");
1570 // Unless PredBB ends with an unconditional branch, split the edge so that we
1571 // can just clone the bits from BB into the end of the new PredBB.
1572 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1574 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1575 PredBB = SplitEdge(PredBB, BB, this);
1576 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1579 // We are going to have to map operands from the original BB block into the
1580 // PredBB block. Evaluate PHI nodes in BB.
1581 DenseMap<Instruction*, Value*> ValueMapping;
1583 BasicBlock::iterator BI = BB->begin();
1584 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1585 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1587 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1588 // mapping and using it to remap operands in the cloned instructions.
1589 for (; BI != BB->end(); ++BI) {
1590 Instruction *New = BI->clone();
1592 // Remap operands to patch up intra-block references.
1593 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1594 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1595 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1596 if (I != ValueMapping.end())
1597 New->setOperand(i, I->second);
1600 // If this instruction can be simplified after the operands are updated,
1601 // just use the simplified value instead. This frequently happens due to
1603 if (Value *IV = SimplifyInstruction(New, TD)) {
1605 ValueMapping[BI] = IV;
1607 // Otherwise, insert the new instruction into the block.
1608 New->setName(BI->getName());
1609 PredBB->getInstList().insert(OldPredBranch, New);
1610 ValueMapping[BI] = New;
1614 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1615 // add entries to the PHI nodes for branch from PredBB now.
1616 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1617 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1619 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1622 // If there were values defined in BB that are used outside the block, then we
1623 // now have to update all uses of the value to use either the original value,
1624 // the cloned value, or some PHI derived value. This can require arbitrary
1625 // PHI insertion, of which we are prepared to do, clean these up now.
1626 SSAUpdater SSAUpdate;
1627 SmallVector<Use*, 16> UsesToRename;
1628 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1629 // Scan all uses of this instruction to see if it is used outside of its
1630 // block, and if so, record them in UsesToRename.
1631 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1633 Instruction *User = cast<Instruction>(*UI);
1634 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1635 if (UserPN->getIncomingBlock(UI) == BB)
1637 } else if (User->getParent() == BB)
1640 UsesToRename.push_back(&UI.getUse());
1643 // If there are no uses outside the block, we're done with this instruction.
1644 if (UsesToRename.empty())
1647 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1649 // We found a use of I outside of BB. Rename all uses of I that are outside
1650 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1651 // with the two values we know.
1652 SSAUpdate.Initialize(I);
1653 SSAUpdate.AddAvailableValue(BB, I);
1654 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1656 while (!UsesToRename.empty())
1657 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1658 DEBUG(dbgs() << "\n");
1661 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1663 RemovePredecessorAndSimplify(BB, PredBB, TD);
1665 // Remove the unconditional branch at the end of the PredBB block.
1666 OldPredBranch->eraseFromParent();