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/DenseSet.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);
48 /// This pass performs 'jump threading', which looks at blocks that have
49 /// multiple predecessors and multiple successors. If one or more of the
50 /// predecessors of the block can be proven to always jump to one of the
51 /// successors, we forward the edge from the predecessor to the successor by
52 /// duplicating the contents of this block.
54 /// An example of when this can occur is code like this:
61 /// In this case, the unconditional branch at the end of the first if can be
62 /// revectored to the false side of the second if.
64 class JumpThreading : public FunctionPass {
68 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
70 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
72 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet;
74 // RAII helper for updating the recursion stack.
75 struct RecursionSetRemover {
76 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet;
77 std::pair<Value*, BasicBlock*> ThePair;
79 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S,
80 std::pair<Value*, BasicBlock*> P)
81 : TheSet(S), ThePair(P) { }
83 ~RecursionSetRemover() {
84 TheSet.erase(ThePair);
88 static char ID; // Pass identification
89 JumpThreading() : FunctionPass(ID) {}
91 bool runOnFunction(Function &F);
93 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
94 AU.addRequired<LazyValueInfo>();
95 AU.addPreserved<LazyValueInfo>();
98 void FindLoopHeaders(Function &F);
99 bool ProcessBlock(BasicBlock *BB);
100 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
102 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
103 const SmallVectorImpl<BasicBlock *> &PredBBs);
105 typedef SmallVectorImpl<std::pair<ConstantInt*,
106 BasicBlock*> > PredValueInfo;
108 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
109 PredValueInfo &Result);
110 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB);
113 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
114 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
116 bool ProcessBranchOnPHI(PHINode *PN);
117 bool ProcessBranchOnXOR(BinaryOperator *BO);
119 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
123 char JumpThreading::ID = 0;
124 INITIALIZE_PASS(JumpThreading, "jump-threading",
125 "Jump Threading", false, false);
127 // Public interface to the Jump Threading pass
128 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
130 /// runOnFunction - Top level algorithm.
132 bool JumpThreading::runOnFunction(Function &F) {
133 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
134 TD = getAnalysisIfAvailable<TargetData>();
135 LVI = &getAnalysis<LazyValueInfo>();
139 bool Changed, EverChanged = false;
142 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
144 // Thread all of the branches we can over this block.
145 while (ProcessBlock(BB))
150 // If the block is trivially dead, zap it. This eliminates the successor
151 // edges which simplifies the CFG.
152 if (pred_begin(BB) == pred_end(BB) &&
153 BB != &BB->getParent()->getEntryBlock()) {
154 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
155 << "' with terminator: " << *BB->getTerminator() << '\n');
156 LoopHeaders.erase(BB);
160 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
161 // Can't thread an unconditional jump, but if the block is "almost
162 // empty", we can replace uses of it with uses of the successor and make
164 if (BI->isUnconditional() &&
165 BB != &BB->getParent()->getEntryBlock()) {
166 BasicBlock::iterator BBI = BB->getFirstNonPHI();
167 // Ignore dbg intrinsics.
168 while (isa<DbgInfoIntrinsic>(BBI))
170 // If the terminator is the only non-phi instruction, try to nuke it.
171 if (BBI->isTerminator()) {
172 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
173 // block, we have to make sure it isn't in the LoopHeaders set. We
174 // reinsert afterward if needed.
175 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
176 BasicBlock *Succ = BI->getSuccessor(0);
178 // FIXME: It is always conservatively correct to drop the info
179 // for a block even if it doesn't get erased. This isn't totally
180 // awesome, but it allows us to use AssertingVH to prevent nasty
181 // dangling pointer issues within LazyValueInfo.
183 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
185 // If we deleted BB and BB was the header of a loop, then the
186 // successor is now the header of the loop.
190 if (ErasedFromLoopHeaders)
191 LoopHeaders.insert(BB);
196 EverChanged |= Changed;
203 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
204 /// thread across it.
205 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
206 /// Ignore PHI nodes, these will be flattened when duplication happens.
207 BasicBlock::const_iterator I = BB->getFirstNonPHI();
209 // FIXME: THREADING will delete values that are just used to compute the
210 // branch, so they shouldn't count against the duplication cost.
213 // Sum up the cost of each instruction until we get to the terminator. Don't
214 // include the terminator because the copy won't include it.
216 for (; !isa<TerminatorInst>(I); ++I) {
217 // Debugger intrinsics don't incur code size.
218 if (isa<DbgInfoIntrinsic>(I)) continue;
220 // If this is a pointer->pointer bitcast, it is free.
221 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
224 // All other instructions count for at least one unit.
227 // Calls are more expensive. If they are non-intrinsic calls, we model them
228 // as having cost of 4. If they are a non-vector intrinsic, we model them
229 // as having cost of 2 total, and if they are a vector intrinsic, we model
230 // them as having cost 1.
231 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
232 if (!isa<IntrinsicInst>(CI))
234 else if (!CI->getType()->isVectorTy())
239 // Threading through a switch statement is particularly profitable. If this
240 // block ends in a switch, decrease its cost to make it more likely to happen.
241 if (isa<SwitchInst>(I))
242 Size = Size > 6 ? Size-6 : 0;
247 /// FindLoopHeaders - We do not want jump threading to turn proper loop
248 /// structures into irreducible loops. Doing this breaks up the loop nesting
249 /// hierarchy and pessimizes later transformations. To prevent this from
250 /// happening, we first have to find the loop headers. Here we approximate this
251 /// by finding targets of backedges in the CFG.
253 /// Note that there definitely are cases when we want to allow threading of
254 /// edges across a loop header. For example, threading a jump from outside the
255 /// loop (the preheader) to an exit block of the loop is definitely profitable.
256 /// It is also almost always profitable to thread backedges from within the loop
257 /// to exit blocks, and is often profitable to thread backedges to other blocks
258 /// within the loop (forming a nested loop). This simple analysis is not rich
259 /// enough to track all of these properties and keep it up-to-date as the CFG
260 /// mutates, so we don't allow any of these transformations.
262 void JumpThreading::FindLoopHeaders(Function &F) {
263 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
264 FindFunctionBackedges(F, Edges);
266 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
267 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
270 // Helper method for ComputeValueKnownInPredecessors. If Value is a
271 // ConstantInt, push it. If it's an undef, push 0. Otherwise, do nothing.
272 static void PushConstantIntOrUndef(SmallVectorImpl<std::pair<ConstantInt*,
273 BasicBlock*> > &Result,
274 Constant *Value, BasicBlock* BB){
275 if (ConstantInt *FoldedCInt = dyn_cast<ConstantInt>(Value))
276 Result.push_back(std::make_pair(FoldedCInt, BB));
277 else if (isa<UndefValue>(Value))
278 Result.push_back(std::make_pair((ConstantInt*)0, BB));
281 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
282 /// if we can infer that the value is a known ConstantInt in any of our
283 /// predecessors. If so, return the known list of value and pred BB in the
284 /// result vector. If a value is known to be undef, it is returned as null.
286 /// This returns true if there were any known values.
289 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){
290 // This method walks up use-def chains recursively. Because of this, we could
291 // get into an infinite loop going around loops in the use-def chain. To
292 // prevent this, keep track of what (value, block) pairs we've already visited
293 // and terminate the search if we loop back to them
294 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
297 // An RAII help to remove this pair from the recursion set once the recursion
298 // stack pops back out again.
299 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
301 // If V is a constantint, then it is known in all predecessors.
302 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) {
303 ConstantInt *CI = dyn_cast<ConstantInt>(V);
305 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
306 Result.push_back(std::make_pair(CI, *PI));
311 // If V is a non-instruction value, or an instruction in a different block,
312 // then it can't be derived from a PHI.
313 Instruction *I = dyn_cast<Instruction>(V);
314 if (I == 0 || I->getParent() != BB) {
316 // Okay, if this is a live-in value, see if it has a known value at the end
317 // of any of our predecessors.
319 // FIXME: This should be an edge property, not a block end property.
320 /// TODO: Per PR2563, we could infer value range information about a
321 /// predecessor based on its terminator.
323 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
324 // "I" is a non-local compare-with-a-constant instruction. This would be
325 // able to handle value inequalities better, for example if the compare is
326 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
327 // Perhaps getConstantOnEdge should be smart enough to do this?
329 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
331 // If the value is known by LazyValueInfo to be a constant in a
332 // predecessor, use that information to try to thread this block.
333 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
335 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst)))
338 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), P));
341 return !Result.empty();
344 /// If I is a PHI node, then we know the incoming values for any constants.
345 if (PHINode *PN = dyn_cast<PHINode>(I)) {
346 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
347 Value *InVal = PN->getIncomingValue(i);
348 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) {
349 ConstantInt *CI = dyn_cast<ConstantInt>(InVal);
350 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i)));
352 Constant *CI = LVI->getConstantOnEdge(InVal,
353 PN->getIncomingBlock(i), BB);
354 // LVI returns null is no value could be determined.
356 PushConstantIntOrUndef(Result, CI, PN->getIncomingBlock(i));
360 return !Result.empty();
363 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals;
365 // Handle some boolean conditions.
366 if (I->getType()->getPrimitiveSizeInBits() == 1) {
368 // X & false -> false
369 if (I->getOpcode() == Instruction::Or ||
370 I->getOpcode() == Instruction::And) {
371 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
372 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals);
374 if (LHSVals.empty() && RHSVals.empty())
377 ConstantInt *InterestingVal;
378 if (I->getOpcode() == Instruction::Or)
379 InterestingVal = ConstantInt::getTrue(I->getContext());
381 InterestingVal = ConstantInt::getFalse(I->getContext());
383 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
385 // Scan for the sentinel. If we find an undef, force it to the
386 // interesting value: x|undef -> true and x&undef -> false.
387 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
388 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) {
389 Result.push_back(LHSVals[i]);
390 Result.back().first = InterestingVal;
391 LHSKnownBBs.insert(LHSVals[i].second);
393 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
394 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) {
395 // If we already inferred a value for this block on the LHS, don't
397 if (!LHSKnownBBs.count(RHSVals[i].second)) {
398 Result.push_back(RHSVals[i]);
399 Result.back().first = InterestingVal;
403 return !Result.empty();
406 // Handle the NOT form of XOR.
407 if (I->getOpcode() == Instruction::Xor &&
408 isa<ConstantInt>(I->getOperand(1)) &&
409 cast<ConstantInt>(I->getOperand(1))->isOne()) {
410 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result);
414 // Invert the known values.
415 for (unsigned i = 0, e = Result.size(); i != e; ++i)
418 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first));
423 // Try to simplify some other binary operator values.
424 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
425 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
426 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals;
427 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals);
429 // Try to use constant folding to simplify the binary operator.
430 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
431 Constant *V = LHSVals[i].first;
432 if (V == 0) V = UndefValue::get(BO->getType());
433 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
435 PushConstantIntOrUndef(Result, Folded, LHSVals[i].second);
439 return !Result.empty();
442 // Handle compare with phi operand, where the PHI is defined in this block.
443 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
444 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
445 if (PN && PN->getParent() == BB) {
446 // We can do this simplification if any comparisons fold to true or false.
448 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
449 BasicBlock *PredBB = PN->getIncomingBlock(i);
450 Value *LHS = PN->getIncomingValue(i);
451 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
453 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
455 if (!isa<Constant>(RHS))
458 LazyValueInfo::Tristate
459 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
460 cast<Constant>(RHS), PredBB, BB);
461 if (ResT == LazyValueInfo::Unknown)
463 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
466 if (Constant *ConstRes = dyn_cast<Constant>(Res))
467 PushConstantIntOrUndef(Result, ConstRes, PredBB);
470 return !Result.empty();
474 // If comparing a live-in value against a constant, see if we know the
475 // live-in value on any predecessors.
476 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
477 if (!isa<Instruction>(Cmp->getOperand(0)) ||
478 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
479 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
481 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
483 // If the value is known by LazyValueInfo to be a constant in a
484 // predecessor, use that information to try to thread this block.
485 LazyValueInfo::Tristate Res =
486 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
488 if (Res == LazyValueInfo::Unknown)
491 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
492 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), P));
495 return !Result.empty();
498 // Try to find a constant value for the LHS of a comparison,
499 // and evaluate it statically if we can.
500 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
501 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals;
502 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
504 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
505 Constant *V = LHSVals[i].first;
506 if (V == 0) V = UndefValue::get(CmpConst->getType());
507 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
509 PushConstantIntOrUndef(Result, Folded, LHSVals[i].second);
512 return !Result.empty();
517 // If all else fails, see if LVI can figure out a constant value for us.
518 Constant *CI = LVI->getConstant(V, BB);
519 ConstantInt *CInt = dyn_cast_or_null<ConstantInt>(CI);
521 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
522 Result.push_back(std::make_pair(CInt, *PI));
525 return !Result.empty();
530 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
531 /// in an undefined jump, decide which block is best to revector to.
533 /// Since we can pick an arbitrary destination, we pick the successor with the
534 /// fewest predecessors. This should reduce the in-degree of the others.
536 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
537 TerminatorInst *BBTerm = BB->getTerminator();
538 unsigned MinSucc = 0;
539 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
540 // Compute the successor with the minimum number of predecessors.
541 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
542 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
543 TestBB = BBTerm->getSuccessor(i);
544 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
545 if (NumPreds < MinNumPreds)
552 /// ProcessBlock - If there are any predecessors whose control can be threaded
553 /// through to a successor, transform them now.
554 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
555 // If the block is trivially dead, just return and let the caller nuke it.
556 // This simplifies other transformations.
557 if (pred_begin(BB) == pred_end(BB) &&
558 BB != &BB->getParent()->getEntryBlock())
561 // If this block has a single predecessor, and if that pred has a single
562 // successor, merge the blocks. This encourages recursive jump threading
563 // because now the condition in this block can be threaded through
564 // predecessors of our predecessor block.
565 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
566 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
568 // If SinglePred was a loop header, BB becomes one.
569 if (LoopHeaders.erase(SinglePred))
570 LoopHeaders.insert(BB);
572 // Remember if SinglePred was the entry block of the function. If so, we
573 // will need to move BB back to the entry position.
574 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
575 LVI->eraseBlock(SinglePred);
576 MergeBasicBlockIntoOnlyPred(BB);
578 if (isEntry && BB != &BB->getParent()->getEntryBlock())
579 BB->moveBefore(&BB->getParent()->getEntryBlock());
584 // Look to see if the terminator is a branch of switch, if not we can't thread
587 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
588 // Can't thread an unconditional jump.
589 if (BI->isUnconditional()) return false;
590 Condition = BI->getCondition();
591 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
592 Condition = SI->getCondition();
594 return false; // Must be an invoke.
596 // If the terminator of this block is branching on a constant, simplify the
597 // terminator to an unconditional branch. This can occur due to threading in
599 if (isa<ConstantInt>(Condition)) {
600 DEBUG(dbgs() << " In block '" << BB->getName()
601 << "' folding terminator: " << *BB->getTerminator() << '\n');
603 ConstantFoldTerminator(BB);
607 // If the terminator is branching on an undef, we can pick any of the
608 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
609 if (isa<UndefValue>(Condition)) {
610 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
612 // Fold the branch/switch.
613 TerminatorInst *BBTerm = BB->getTerminator();
614 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
615 if (i == BestSucc) continue;
616 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD);
619 DEBUG(dbgs() << " In block '" << BB->getName()
620 << "' folding undef terminator: " << *BBTerm << '\n');
621 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
622 BBTerm->eraseFromParent();
626 Instruction *CondInst = dyn_cast<Instruction>(Condition);
628 // All the rest of our checks depend on the condition being an instruction.
630 // FIXME: Unify this with code below.
631 if (ProcessThreadableEdges(Condition, BB))
637 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
638 // For a comparison where the LHS is outside this block, it's possible
639 // that we've branched on it before. Used LVI to see if we can simplify
640 // the branch based on that.
641 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
642 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
643 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
644 if (CondBr && CondConst && CondBr->isConditional() && PI != PE &&
645 (!isa<Instruction>(CondCmp->getOperand(0)) ||
646 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
647 // For predecessor edge, determine if the comparison is true or false
648 // on that edge. If they're all true or all false, we can simplify the
650 // FIXME: We could handle mixed true/false by duplicating code.
651 LazyValueInfo::Tristate Baseline =
652 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
654 if (Baseline != LazyValueInfo::Unknown) {
655 // Check that all remaining incoming values match the first one.
657 LazyValueInfo::Tristate Ret =
658 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
659 CondCmp->getOperand(0), CondConst, *PI, BB);
660 if (Ret != Baseline) break;
663 // If we terminated early, then one of the values didn't match.
665 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
666 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
667 RemovePredecessorAndSimplify(CondBr->getSuccessor(ToRemove), BB, TD);
668 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
669 CondBr->eraseFromParent();
676 // Check for some cases that are worth simplifying. Right now we want to look
677 // for loads that are used by a switch or by the condition for the branch. If
678 // we see one, check to see if it's partially redundant. If so, insert a PHI
679 // which can then be used to thread the values.
681 Value *SimplifyValue = CondInst;
682 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
683 if (isa<Constant>(CondCmp->getOperand(1)))
684 SimplifyValue = CondCmp->getOperand(0);
686 // TODO: There are other places where load PRE would be profitable, such as
687 // more complex comparisons.
688 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
689 if (SimplifyPartiallyRedundantLoad(LI))
693 // Handle a variety of cases where we are branching on something derived from
694 // a PHI node in the current block. If we can prove that any predecessors
695 // compute a predictable value based on a PHI node, thread those predecessors.
697 if (ProcessThreadableEdges(CondInst, BB))
700 // If this is an otherwise-unfoldable branch on a phi node in the current
701 // block, see if we can simplify.
702 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
703 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
704 return ProcessBranchOnPHI(PN);
707 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
708 if (CondInst->getOpcode() == Instruction::Xor &&
709 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
710 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
713 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
714 // "(X == 4)", thread through this block.
719 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
720 /// block that jump on exactly the same condition. This means that we almost
721 /// always know the direction of the edge in the DESTBB:
723 /// br COND, DESTBB, BBY
725 /// br COND, BBZ, BBW
727 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
728 /// in DESTBB, we have to thread over it.
729 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
731 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
733 // If both successors of PredBB go to DESTBB, we don't know anything. We can
734 // fold the branch to an unconditional one, which allows other recursive
737 if (PredBI->getSuccessor(1) != BB)
739 else if (PredBI->getSuccessor(0) != BB)
742 DEBUG(dbgs() << " In block '" << PredBB->getName()
743 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
745 ConstantFoldTerminator(PredBB);
749 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
751 // If the dest block has one predecessor, just fix the branch condition to a
752 // constant and fold it.
753 if (BB->getSinglePredecessor()) {
754 DEBUG(dbgs() << " In block '" << BB->getName()
755 << "' folding condition to '" << BranchDir << "': "
756 << *BB->getTerminator() << '\n');
758 Value *OldCond = DestBI->getCondition();
759 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
761 // Delete dead instructions before we fold the branch. Folding the branch
762 // can eliminate edges from the CFG which can end up deleting OldCond.
763 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
764 ConstantFoldTerminator(BB);
769 // Next, figure out which successor we are threading to.
770 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
772 SmallVector<BasicBlock*, 2> Preds;
773 Preds.push_back(PredBB);
775 // Ok, try to thread it!
776 return ThreadEdge(BB, Preds, SuccBB);
779 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
780 /// block that switch on exactly the same condition. This means that we almost
781 /// always know the direction of the edge in the DESTBB:
783 /// switch COND [... DESTBB, BBY ... ]
785 /// switch COND [... BBZ, BBW ]
787 /// Optimizing switches like this is very important, because simplifycfg builds
788 /// switches out of repeated 'if' conditions.
789 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
790 BasicBlock *DestBB) {
791 // Can't thread edge to self.
792 if (PredBB == DestBB)
795 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
796 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
798 // There are a variety of optimizations that we can potentially do on these
799 // blocks: we order them from most to least preferable.
801 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
802 // directly to their destination. This does not introduce *any* code size
803 // growth. Skip debug info first.
804 BasicBlock::iterator BBI = DestBB->begin();
805 while (isa<DbgInfoIntrinsic>(BBI))
808 // FIXME: Thread if it just contains a PHI.
809 if (isa<SwitchInst>(BBI)) {
810 bool MadeChange = false;
811 // Ignore the default edge for now.
812 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
813 ConstantInt *DestVal = DestSI->getCaseValue(i);
814 BasicBlock *DestSucc = DestSI->getSuccessor(i);
816 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
817 // PredSI has an explicit case for it. If so, forward. If it is covered
818 // by the default case, we can't update PredSI.
819 unsigned PredCase = PredSI->findCaseValue(DestVal);
820 if (PredCase == 0) continue;
822 // If PredSI doesn't go to DestBB on this value, then it won't reach the
823 // case on this condition.
824 if (PredSI->getSuccessor(PredCase) != DestBB &&
825 DestSI->getSuccessor(i) != DestBB)
828 // Do not forward this if it already goes to this destination, this would
829 // be an infinite loop.
830 if (PredSI->getSuccessor(PredCase) == DestSucc)
833 // Otherwise, we're safe to make the change. Make sure that the edge from
834 // DestSI to DestSucc is not critical and has no PHI nodes.
835 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
836 DEBUG(dbgs() << "THROUGH: " << *DestSI);
838 // If the destination has PHI nodes, just split the edge for updating
840 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
841 SplitCriticalEdge(DestSI, i, this);
842 DestSucc = DestSI->getSuccessor(i);
844 FoldSingleEntryPHINodes(DestSucc);
845 PredSI->setSuccessor(PredCase, DestSucc);
857 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
858 /// load instruction, eliminate it by replacing it with a PHI node. This is an
859 /// important optimization that encourages jump threading, and needs to be run
860 /// interlaced with other jump threading tasks.
861 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
862 // Don't hack volatile loads.
863 if (LI->isVolatile()) return false;
865 // If the load is defined in a block with exactly one predecessor, it can't be
866 // partially redundant.
867 BasicBlock *LoadBB = LI->getParent();
868 if (LoadBB->getSinglePredecessor())
871 Value *LoadedPtr = LI->getOperand(0);
873 // If the loaded operand is defined in the LoadBB, it can't be available.
874 // TODO: Could do simple PHI translation, that would be fun :)
875 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
876 if (PtrOp->getParent() == LoadBB)
879 // Scan a few instructions up from the load, to see if it is obviously live at
880 // the entry to its block.
881 BasicBlock::iterator BBIt = LI;
883 if (Value *AvailableVal =
884 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
885 // If the value if the load is locally available within the block, just use
886 // it. This frequently occurs for reg2mem'd allocas.
887 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
889 // If the returned value is the load itself, replace with an undef. This can
890 // only happen in dead loops.
891 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
892 LI->replaceAllUsesWith(AvailableVal);
893 LI->eraseFromParent();
897 // Otherwise, if we scanned the whole block and got to the top of the block,
898 // we know the block is locally transparent to the load. If not, something
899 // might clobber its value.
900 if (BBIt != LoadBB->begin())
904 SmallPtrSet<BasicBlock*, 8> PredsScanned;
905 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
906 AvailablePredsTy AvailablePreds;
907 BasicBlock *OneUnavailablePred = 0;
909 // If we got here, the loaded value is transparent through to the start of the
910 // block. Check to see if it is available in any of the predecessor blocks.
911 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
913 BasicBlock *PredBB = *PI;
915 // If we already scanned this predecessor, skip it.
916 if (!PredsScanned.insert(PredBB))
919 // Scan the predecessor to see if the value is available in the pred.
920 BBIt = PredBB->end();
921 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
922 if (!PredAvailable) {
923 OneUnavailablePred = PredBB;
927 // If so, this load is partially redundant. Remember this info so that we
928 // can create a PHI node.
929 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
932 // If the loaded value isn't available in any predecessor, it isn't partially
934 if (AvailablePreds.empty()) return false;
936 // Okay, the loaded value is available in at least one (and maybe all!)
937 // predecessors. If the value is unavailable in more than one unique
938 // predecessor, we want to insert a merge block for those common predecessors.
939 // This ensures that we only have to insert one reload, thus not increasing
941 BasicBlock *UnavailablePred = 0;
943 // If there is exactly one predecessor where the value is unavailable, the
944 // already computed 'OneUnavailablePred' block is it. If it ends in an
945 // unconditional branch, we know that it isn't a critical edge.
946 if (PredsScanned.size() == AvailablePreds.size()+1 &&
947 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
948 UnavailablePred = OneUnavailablePred;
949 } else if (PredsScanned.size() != AvailablePreds.size()) {
950 // Otherwise, we had multiple unavailable predecessors or we had a critical
951 // edge from the one.
952 SmallVector<BasicBlock*, 8> PredsToSplit;
953 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
955 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
956 AvailablePredSet.insert(AvailablePreds[i].first);
958 // Add all the unavailable predecessors to the PredsToSplit list.
959 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
962 // If the predecessor is an indirect goto, we can't split the edge.
963 if (isa<IndirectBrInst>(P->getTerminator()))
966 if (!AvailablePredSet.count(P))
967 PredsToSplit.push_back(P);
970 // Split them out to their own block.
972 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
973 "thread-pre-split", this);
976 // If the value isn't available in all predecessors, then there will be
977 // exactly one where it isn't available. Insert a load on that edge and add
978 // it to the AvailablePreds list.
979 if (UnavailablePred) {
980 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
981 "Can't handle critical edge here!");
982 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
984 UnavailablePred->getTerminator());
985 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
988 // Now we know that each predecessor of this block has a value in
989 // AvailablePreds, sort them for efficient access as we're walking the preds.
990 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
992 // Create a PHI node at the start of the block for the PRE'd load value.
993 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
996 // Insert new entries into the PHI for each predecessor. A single block may
997 // have multiple entries here.
998 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
1000 BasicBlock *P = *PI;
1001 AvailablePredsTy::iterator I =
1002 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1003 std::make_pair(P, (Value*)0));
1005 assert(I != AvailablePreds.end() && I->first == P &&
1006 "Didn't find entry for predecessor!");
1008 PN->addIncoming(I->second, I->first);
1011 //cerr << "PRE: " << *LI << *PN << "\n";
1013 LI->replaceAllUsesWith(PN);
1014 LI->eraseFromParent();
1019 /// FindMostPopularDest - The specified list contains multiple possible
1020 /// threadable destinations. Pick the one that occurs the most frequently in
1023 FindMostPopularDest(BasicBlock *BB,
1024 const SmallVectorImpl<std::pair<BasicBlock*,
1025 BasicBlock*> > &PredToDestList) {
1026 assert(!PredToDestList.empty());
1028 // Determine popularity. If there are multiple possible destinations, we
1029 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1030 // blocks with known and real destinations to threading undef. We'll handle
1031 // them later if interesting.
1032 DenseMap<BasicBlock*, unsigned> DestPopularity;
1033 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1034 if (PredToDestList[i].second)
1035 DestPopularity[PredToDestList[i].second]++;
1037 // Find the most popular dest.
1038 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1039 BasicBlock *MostPopularDest = DPI->first;
1040 unsigned Popularity = DPI->second;
1041 SmallVector<BasicBlock*, 4> SamePopularity;
1043 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1044 // If the popularity of this entry isn't higher than the popularity we've
1045 // seen so far, ignore it.
1046 if (DPI->second < Popularity)
1048 else if (DPI->second == Popularity) {
1049 // If it is the same as what we've seen so far, keep track of it.
1050 SamePopularity.push_back(DPI->first);
1052 // If it is more popular, remember it.
1053 SamePopularity.clear();
1054 MostPopularDest = DPI->first;
1055 Popularity = DPI->second;
1059 // Okay, now we know the most popular destination. If there is more than
1060 // destination, we need to determine one. This is arbitrary, but we need
1061 // to make a deterministic decision. Pick the first one that appears in the
1063 if (!SamePopularity.empty()) {
1064 SamePopularity.push_back(MostPopularDest);
1065 TerminatorInst *TI = BB->getTerminator();
1066 for (unsigned i = 0; ; ++i) {
1067 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1069 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1070 TI->getSuccessor(i)) == SamePopularity.end())
1073 MostPopularDest = TI->getSuccessor(i);
1078 // Okay, we have finally picked the most popular destination.
1079 return MostPopularDest;
1082 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) {
1083 // If threading this would thread across a loop header, don't even try to
1085 if (LoopHeaders.count(BB))
1088 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues;
1089 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues))
1092 assert(!PredValues.empty() &&
1093 "ComputeValueKnownInPredecessors returned true with no values");
1095 DEBUG(dbgs() << "IN BB: " << *BB;
1096 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1097 dbgs() << " BB '" << BB->getName() << "': FOUND condition = ";
1098 if (PredValues[i].first)
1099 dbgs() << *PredValues[i].first;
1102 dbgs() << " for pred '" << PredValues[i].second->getName()
1106 // Decide what we want to thread through. Convert our list of known values to
1107 // a list of known destinations for each pred. This also discards duplicate
1108 // predecessors and keeps track of the undefined inputs (which are represented
1109 // as a null dest in the PredToDestList).
1110 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1111 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1113 BasicBlock *OnlyDest = 0;
1114 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1116 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1117 BasicBlock *Pred = PredValues[i].second;
1118 if (!SeenPreds.insert(Pred))
1119 continue; // Duplicate predecessor entry.
1121 // If the predecessor ends with an indirect goto, we can't change its
1123 if (isa<IndirectBrInst>(Pred->getTerminator()))
1126 ConstantInt *Val = PredValues[i].first;
1129 if (Val == 0) // Undef.
1131 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1132 DestBB = BI->getSuccessor(Val->isZero());
1134 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
1135 DestBB = SI->getSuccessor(SI->findCaseValue(Val));
1138 // If we have exactly one destination, remember it for efficiency below.
1141 else if (OnlyDest != DestBB)
1142 OnlyDest = MultipleDestSentinel;
1144 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1147 // If all edges were unthreadable, we fail.
1148 if (PredToDestList.empty())
1151 // Determine which is the most common successor. If we have many inputs and
1152 // this block is a switch, we want to start by threading the batch that goes
1153 // to the most popular destination first. If we only know about one
1154 // threadable destination (the common case) we can avoid this.
1155 BasicBlock *MostPopularDest = OnlyDest;
1157 if (MostPopularDest == MultipleDestSentinel)
1158 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1160 // Now that we know what the most popular destination is, factor all
1161 // predecessors that will jump to it into a single predecessor.
1162 SmallVector<BasicBlock*, 16> PredsToFactor;
1163 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1164 if (PredToDestList[i].second == MostPopularDest) {
1165 BasicBlock *Pred = PredToDestList[i].first;
1167 // This predecessor may be a switch or something else that has multiple
1168 // edges to the block. Factor each of these edges by listing them
1169 // according to # occurrences in PredsToFactor.
1170 TerminatorInst *PredTI = Pred->getTerminator();
1171 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1172 if (PredTI->getSuccessor(i) == BB)
1173 PredsToFactor.push_back(Pred);
1176 // If the threadable edges are branching on an undefined value, we get to pick
1177 // the destination that these predecessors should get to.
1178 if (MostPopularDest == 0)
1179 MostPopularDest = BB->getTerminator()->
1180 getSuccessor(GetBestDestForJumpOnUndef(BB));
1182 // Ok, try to thread it!
1183 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1186 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1187 /// a PHI node in the current block. See if there are any simplifications we
1188 /// can do based on inputs to the phi node.
1190 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1191 BasicBlock *BB = PN->getParent();
1193 // TODO: We could make use of this to do it once for blocks with common PHI
1195 SmallVector<BasicBlock*, 1> PredBBs;
1198 // If any of the predecessor blocks end in an unconditional branch, we can
1199 // *duplicate* the conditional branch into that block in order to further
1200 // encourage jump threading and to eliminate cases where we have branch on a
1201 // phi of an icmp (branch on icmp is much better).
1202 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1203 BasicBlock *PredBB = PN->getIncomingBlock(i);
1204 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1205 if (PredBr->isUnconditional()) {
1206 PredBBs[0] = PredBB;
1207 // Try to duplicate BB into PredBB.
1208 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1216 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1217 /// a xor instruction in the current block. See if there are any
1218 /// simplifications we can do based on inputs to the xor.
1220 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1221 BasicBlock *BB = BO->getParent();
1223 // If either the LHS or RHS of the xor is a constant, don't do this
1225 if (isa<ConstantInt>(BO->getOperand(0)) ||
1226 isa<ConstantInt>(BO->getOperand(1)))
1229 // If the first instruction in BB isn't a phi, we won't be able to infer
1230 // anything special about any particular predecessor.
1231 if (!isa<PHINode>(BB->front()))
1234 // If we have a xor as the branch input to this block, and we know that the
1235 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1236 // the condition into the predecessor and fix that value to true, saving some
1237 // logical ops on that path and encouraging other paths to simplify.
1239 // This copies something like this:
1242 // %X = phi i1 [1], [%X']
1243 // %Y = icmp eq i32 %A, %B
1244 // %Z = xor i1 %X, %Y
1249 // %Y = icmp ne i32 %A, %B
1252 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues;
1254 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) {
1255 assert(XorOpValues.empty());
1256 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues))
1261 assert(!XorOpValues.empty() &&
1262 "ComputeValueKnownInPredecessors returned true with no values");
1264 // Scan the information to see which is most popular: true or false. The
1265 // predecessors can be of the set true, false, or undef.
1266 unsigned NumTrue = 0, NumFalse = 0;
1267 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1268 if (!XorOpValues[i].first) continue; // Ignore undefs for the count.
1269 if (XorOpValues[i].first->isZero())
1275 // Determine which value to split on, true, false, or undef if neither.
1276 ConstantInt *SplitVal = 0;
1277 if (NumTrue > NumFalse)
1278 SplitVal = ConstantInt::getTrue(BB->getContext());
1279 else if (NumTrue != 0 || NumFalse != 0)
1280 SplitVal = ConstantInt::getFalse(BB->getContext());
1282 // Collect all of the blocks that this can be folded into so that we can
1283 // factor this once and clone it once.
1284 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1285 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1286 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue;
1288 BlocksToFoldInto.push_back(XorOpValues[i].second);
1291 // If we inferred a value for all of the predecessors, then duplication won't
1292 // help us. However, we can just replace the LHS or RHS with the constant.
1293 if (BlocksToFoldInto.size() ==
1294 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1295 if (SplitVal == 0) {
1296 // If all preds provide undef, just nuke the xor, because it is undef too.
1297 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1298 BO->eraseFromParent();
1299 } else if (SplitVal->isZero()) {
1300 // If all preds provide 0, replace the xor with the other input.
1301 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1302 BO->eraseFromParent();
1304 // If all preds provide 1, set the computed value to 1.
1305 BO->setOperand(!isLHS, SplitVal);
1311 // Try to duplicate BB into PredBB.
1312 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1316 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1317 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1318 /// NewPred using the entries from OldPred (suitably mapped).
1319 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1320 BasicBlock *OldPred,
1321 BasicBlock *NewPred,
1322 DenseMap<Instruction*, Value*> &ValueMap) {
1323 for (BasicBlock::iterator PNI = PHIBB->begin();
1324 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1325 // Ok, we have a PHI node. Figure out what the incoming value was for the
1327 Value *IV = PN->getIncomingValueForBlock(OldPred);
1329 // Remap the value if necessary.
1330 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1331 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1332 if (I != ValueMap.end())
1336 PN->addIncoming(IV, NewPred);
1340 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1341 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1342 /// across BB. Transform the IR to reflect this change.
1343 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1344 const SmallVectorImpl<BasicBlock*> &PredBBs,
1345 BasicBlock *SuccBB) {
1346 // If threading to the same block as we come from, we would infinite loop.
1348 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1349 << "' - would thread to self!\n");
1353 // If threading this would thread across a loop header, don't thread the edge.
1354 // See the comments above FindLoopHeaders for justifications and caveats.
1355 if (LoopHeaders.count(BB)) {
1356 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1357 << "' to dest BB '" << SuccBB->getName()
1358 << "' - it might create an irreducible loop!\n");
1362 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1363 if (JumpThreadCost > Threshold) {
1364 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1365 << "' - Cost is too high: " << JumpThreadCost << "\n");
1369 // And finally, do it! Start by factoring the predecessors is needed.
1371 if (PredBBs.size() == 1)
1372 PredBB = PredBBs[0];
1374 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1375 << " common predecessors.\n");
1376 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1380 // And finally, do it!
1381 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1382 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1383 << ", across block:\n "
1386 LVI->threadEdge(PredBB, BB, SuccBB);
1388 // We are going to have to map operands from the original BB block to the new
1389 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1390 // account for entry from PredBB.
1391 DenseMap<Instruction*, Value*> ValueMapping;
1393 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1394 BB->getName()+".thread",
1395 BB->getParent(), BB);
1396 NewBB->moveAfter(PredBB);
1398 BasicBlock::iterator BI = BB->begin();
1399 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1400 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1402 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1403 // mapping and using it to remap operands in the cloned instructions.
1404 for (; !isa<TerminatorInst>(BI); ++BI) {
1405 Instruction *New = BI->clone();
1406 New->setName(BI->getName());
1407 NewBB->getInstList().push_back(New);
1408 ValueMapping[BI] = New;
1410 // Remap operands to patch up intra-block references.
1411 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1412 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1413 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1414 if (I != ValueMapping.end())
1415 New->setOperand(i, I->second);
1419 // We didn't copy the terminator from BB over to NewBB, because there is now
1420 // an unconditional jump to SuccBB. Insert the unconditional jump.
1421 BranchInst::Create(SuccBB, NewBB);
1423 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1424 // PHI nodes for NewBB now.
1425 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1427 // If there were values defined in BB that are used outside the block, then we
1428 // now have to update all uses of the value to use either the original value,
1429 // the cloned value, or some PHI derived value. This can require arbitrary
1430 // PHI insertion, of which we are prepared to do, clean these up now.
1431 SSAUpdater SSAUpdate;
1432 SmallVector<Use*, 16> UsesToRename;
1433 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1434 // Scan all uses of this instruction to see if it is used outside of its
1435 // block, and if so, record them in UsesToRename.
1436 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1438 Instruction *User = cast<Instruction>(*UI);
1439 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1440 if (UserPN->getIncomingBlock(UI) == BB)
1442 } else if (User->getParent() == BB)
1445 UsesToRename.push_back(&UI.getUse());
1448 // If there are no uses outside the block, we're done with this instruction.
1449 if (UsesToRename.empty())
1452 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1454 // We found a use of I outside of BB. Rename all uses of I that are outside
1455 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1456 // with the two values we know.
1457 SSAUpdate.Initialize(I->getType(), I->getName());
1458 SSAUpdate.AddAvailableValue(BB, I);
1459 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1461 while (!UsesToRename.empty())
1462 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1463 DEBUG(dbgs() << "\n");
1467 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1468 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1469 // us to simplify any PHI nodes in BB.
1470 TerminatorInst *PredTerm = PredBB->getTerminator();
1471 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1472 if (PredTerm->getSuccessor(i) == BB) {
1473 RemovePredecessorAndSimplify(BB, PredBB, TD);
1474 PredTerm->setSuccessor(i, NewBB);
1477 // At this point, the IR is fully up to date and consistent. Do a quick scan
1478 // over the new instructions and zap any that are constants or dead. This
1479 // frequently happens because of phi translation.
1480 SimplifyInstructionsInBlock(NewBB, TD);
1482 // Threaded an edge!
1487 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1488 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1489 /// If we can duplicate the contents of BB up into PredBB do so now, this
1490 /// improves the odds that the branch will be on an analyzable instruction like
1492 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1493 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1494 assert(!PredBBs.empty() && "Can't handle an empty set");
1496 // If BB is a loop header, then duplicating this block outside the loop would
1497 // cause us to transform this into an irreducible loop, don't do this.
1498 // See the comments above FindLoopHeaders for justifications and caveats.
1499 if (LoopHeaders.count(BB)) {
1500 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1501 << "' into predecessor block '" << PredBBs[0]->getName()
1502 << "' - it might create an irreducible loop!\n");
1506 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1507 if (DuplicationCost > Threshold) {
1508 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1509 << "' - Cost is too high: " << DuplicationCost << "\n");
1513 // And finally, do it! Start by factoring the predecessors is needed.
1515 if (PredBBs.size() == 1)
1516 PredBB = PredBBs[0];
1518 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1519 << " common predecessors.\n");
1520 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1524 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1526 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1527 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1528 << DuplicationCost << " block is:" << *BB << "\n");
1530 // Unless PredBB ends with an unconditional branch, split the edge so that we
1531 // can just clone the bits from BB into the end of the new PredBB.
1532 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1534 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1535 PredBB = SplitEdge(PredBB, BB, this);
1536 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1539 // We are going to have to map operands from the original BB block into the
1540 // PredBB block. Evaluate PHI nodes in BB.
1541 DenseMap<Instruction*, Value*> ValueMapping;
1543 BasicBlock::iterator BI = BB->begin();
1544 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1545 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1547 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1548 // mapping and using it to remap operands in the cloned instructions.
1549 for (; BI != BB->end(); ++BI) {
1550 Instruction *New = BI->clone();
1552 // Remap operands to patch up intra-block references.
1553 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1554 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1555 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1556 if (I != ValueMapping.end())
1557 New->setOperand(i, I->second);
1560 // If this instruction can be simplified after the operands are updated,
1561 // just use the simplified value instead. This frequently happens due to
1563 if (Value *IV = SimplifyInstruction(New, TD)) {
1565 ValueMapping[BI] = IV;
1567 // Otherwise, insert the new instruction into the block.
1568 New->setName(BI->getName());
1569 PredBB->getInstList().insert(OldPredBranch, New);
1570 ValueMapping[BI] = New;
1574 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1575 // add entries to the PHI nodes for branch from PredBB now.
1576 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1577 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1579 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1582 // If there were values defined in BB that are used outside the block, then we
1583 // now have to update all uses of the value to use either the original value,
1584 // the cloned value, or some PHI derived value. This can require arbitrary
1585 // PHI insertion, of which we are prepared to do, clean these up now.
1586 SSAUpdater SSAUpdate;
1587 SmallVector<Use*, 16> UsesToRename;
1588 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1589 // Scan all uses of this instruction to see if it is used outside of its
1590 // block, and if so, record them in UsesToRename.
1591 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1593 Instruction *User = cast<Instruction>(*UI);
1594 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1595 if (UserPN->getIncomingBlock(UI) == BB)
1597 } else if (User->getParent() == BB)
1600 UsesToRename.push_back(&UI.getUse());
1603 // If there are no uses outside the block, we're done with this instruction.
1604 if (UsesToRename.empty())
1607 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1609 // We found a use of I outside of BB. Rename all uses of I that are outside
1610 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1611 // with the two values we know.
1612 SSAUpdate.Initialize(I->getType(), I->getName());
1613 SSAUpdate.AddAvailableValue(BB, I);
1614 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1616 while (!UsesToRename.empty())
1617 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1618 DEBUG(dbgs() << "\n");
1621 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1623 RemovePredecessorAndSimplify(BB, PredBB, TD);
1625 // Remove the unconditional branch at the end of the PredBB block.
1626 OldPredBranch->eraseFromParent();