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 // These are at global scope so static functions can use them too.
49 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo;
50 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy;
52 /// This pass performs 'jump threading', which looks at blocks that have
53 /// multiple predecessors and multiple successors. If one or more of the
54 /// predecessors of the block can be proven to always jump to one of the
55 /// successors, we forward the edge from the predecessor to the successor by
56 /// duplicating the contents of this block.
58 /// An example of when this can occur is code like this:
65 /// In this case, the unconditional branch at the end of the first if can be
66 /// revectored to the false side of the second if.
68 class JumpThreading : public FunctionPass {
72 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
74 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
76 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet;
78 // RAII helper for updating the recursion stack.
79 struct RecursionSetRemover {
80 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet;
81 std::pair<Value*, BasicBlock*> ThePair;
83 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S,
84 std::pair<Value*, BasicBlock*> P)
85 : TheSet(S), ThePair(P) { }
87 ~RecursionSetRemover() {
88 TheSet.erase(ThePair);
92 static char ID; // Pass identification
93 JumpThreading() : FunctionPass(ID) {
94 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
97 bool runOnFunction(Function &F);
99 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
100 AU.addRequired<LazyValueInfo>();
101 AU.addPreserved<LazyValueInfo>();
104 void FindLoopHeaders(Function &F);
105 bool ProcessBlock(BasicBlock *BB);
106 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
108 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
109 const SmallVectorImpl<BasicBlock *> &PredBBs);
111 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
112 PredValueInfo &Result);
113 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB);
116 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
117 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
119 bool ProcessBranchOnPHI(PHINode *PN);
120 bool ProcessBranchOnXOR(BinaryOperator *BO);
122 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
126 char JumpThreading::ID = 0;
127 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
128 "Jump Threading", false, false)
129 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)
130 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
131 "Jump Threading", false, false)
133 // Public interface to the Jump Threading pass
134 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
136 /// runOnFunction - Top level algorithm.
138 bool JumpThreading::runOnFunction(Function &F) {
139 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
140 TD = getAnalysisIfAvailable<TargetData>();
141 LVI = &getAnalysis<LazyValueInfo>();
145 bool Changed, EverChanged = false;
148 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
150 // Thread all of the branches we can over this block.
151 while (ProcessBlock(BB))
156 // If the block is trivially dead, zap it. This eliminates the successor
157 // edges which simplifies the CFG.
158 if (pred_begin(BB) == pred_end(BB) &&
159 BB != &BB->getParent()->getEntryBlock()) {
160 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
161 << "' with terminator: " << *BB->getTerminator() << '\n');
162 LoopHeaders.erase(BB);
166 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
167 // Can't thread an unconditional jump, but if the block is "almost
168 // empty", we can replace uses of it with uses of the successor and make
170 if (BI->isUnconditional() &&
171 BB != &BB->getParent()->getEntryBlock()) {
172 BasicBlock::iterator BBI = BB->getFirstNonPHI();
173 // Ignore dbg intrinsics.
174 while (isa<DbgInfoIntrinsic>(BBI))
176 // If the terminator is the only non-phi instruction, try to nuke it.
177 if (BBI->isTerminator()) {
178 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
179 // block, we have to make sure it isn't in the LoopHeaders set. We
180 // reinsert afterward if needed.
181 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
182 BasicBlock *Succ = BI->getSuccessor(0);
184 // FIXME: It is always conservatively correct to drop the info
185 // for a block even if it doesn't get erased. This isn't totally
186 // awesome, but it allows us to use AssertingVH to prevent nasty
187 // dangling pointer issues within LazyValueInfo.
189 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
191 // If we deleted BB and BB was the header of a loop, then the
192 // successor is now the header of the loop.
196 if (ErasedFromLoopHeaders)
197 LoopHeaders.insert(BB);
202 EverChanged |= Changed;
209 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
210 /// thread across it.
211 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
212 /// Ignore PHI nodes, these will be flattened when duplication happens.
213 BasicBlock::const_iterator I = BB->getFirstNonPHI();
215 // FIXME: THREADING will delete values that are just used to compute the
216 // branch, so they shouldn't count against the duplication cost.
219 // Sum up the cost of each instruction until we get to the terminator. Don't
220 // include the terminator because the copy won't include it.
222 for (; !isa<TerminatorInst>(I); ++I) {
223 // Debugger intrinsics don't incur code size.
224 if (isa<DbgInfoIntrinsic>(I)) continue;
226 // If this is a pointer->pointer bitcast, it is free.
227 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
230 // All other instructions count for at least one unit.
233 // Calls are more expensive. If they are non-intrinsic calls, we model them
234 // as having cost of 4. If they are a non-vector intrinsic, we model them
235 // as having cost of 2 total, and if they are a vector intrinsic, we model
236 // them as having cost 1.
237 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
238 if (!isa<IntrinsicInst>(CI))
240 else if (!CI->getType()->isVectorTy())
245 // Threading through a switch statement is particularly profitable. If this
246 // block ends in a switch, decrease its cost to make it more likely to happen.
247 if (isa<SwitchInst>(I))
248 Size = Size > 6 ? Size-6 : 0;
253 /// FindLoopHeaders - We do not want jump threading to turn proper loop
254 /// structures into irreducible loops. Doing this breaks up the loop nesting
255 /// hierarchy and pessimizes later transformations. To prevent this from
256 /// happening, we first have to find the loop headers. Here we approximate this
257 /// by finding targets of backedges in the CFG.
259 /// Note that there definitely are cases when we want to allow threading of
260 /// edges across a loop header. For example, threading a jump from outside the
261 /// loop (the preheader) to an exit block of the loop is definitely profitable.
262 /// It is also almost always profitable to thread backedges from within the loop
263 /// to exit blocks, and is often profitable to thread backedges to other blocks
264 /// within the loop (forming a nested loop). This simple analysis is not rich
265 /// enough to track all of these properties and keep it up-to-date as the CFG
266 /// mutates, so we don't allow any of these transformations.
268 void JumpThreading::FindLoopHeaders(Function &F) {
269 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
270 FindFunctionBackedges(F, Edges);
272 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
273 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
276 /// getKnownConstant - Helper method to determine if we can thread over a
277 /// terminator with the given value as its condition, and if so what value to
279 /// Returns null if Val is null or not an appropriate constant.
280 static Constant *getKnownConstant(Value *Val) {
284 // Undef is "known" enough.
285 if (UndefValue *U = dyn_cast<UndefValue>(Val))
288 return dyn_cast<ConstantInt>(Val);
291 // Helper method for ComputeValueKnownInPredecessors. If Value is a
292 // ConstantInt or undef, push it. Otherwise, do nothing.
293 static void PushKnownConstantOrUndef(PredValueInfo &Result, Constant *Value,
295 if (Constant *KC = getKnownConstant(Value))
296 Result.push_back(std::make_pair(KC, BB));
299 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
300 /// if we can infer that the value is a known ConstantInt in any of our
301 /// predecessors. If so, return the known list of value and pred BB in the
302 /// result vector. If a value is known to be undef, it is returned as null.
304 /// This returns true if there were any known values.
307 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){
308 // This method walks up use-def chains recursively. Because of this, we could
309 // get into an infinite loop going around loops in the use-def chain. To
310 // prevent this, keep track of what (value, block) pairs we've already visited
311 // and terminate the search if we loop back to them
312 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
315 // An RAII help to remove this pair from the recursion set once the recursion
316 // stack pops back out again.
317 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
319 // If V is a constant, then it is known in all predecessors.
320 if (Constant *KC = getKnownConstant(V)) {
321 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
322 Result.push_back(std::make_pair(KC, *PI));
327 // If V is a non-instruction value, or an instruction in a different block,
328 // then it can't be derived from a PHI.
329 Instruction *I = dyn_cast<Instruction>(V);
330 if (I == 0 || I->getParent() != BB) {
332 // Okay, if this is a live-in value, see if it has a known value at the end
333 // of any of our predecessors.
335 // FIXME: This should be an edge property, not a block end property.
336 /// TODO: Per PR2563, we could infer value range information about a
337 /// predecessor based on its terminator.
339 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
340 // "I" is a non-local compare-with-a-constant instruction. This would be
341 // able to handle value inequalities better, for example if the compare is
342 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
343 // Perhaps getConstantOnEdge should be smart enough to do this?
345 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
347 // If the value is known by LazyValueInfo to be a constant in a
348 // predecessor, use that information to try to thread this block.
349 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
350 if (Constant *KC = getKnownConstant(PredCst))
351 Result.push_back(std::make_pair(KC, P));
354 return !Result.empty();
357 /// If I is a PHI node, then we know the incoming values for any constants.
358 if (PHINode *PN = dyn_cast<PHINode>(I)) {
359 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
360 Value *InVal = PN->getIncomingValue(i);
361 if (Constant *KC = getKnownConstant(InVal)) {
362 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
364 Constant *CI = LVI->getConstantOnEdge(InVal,
365 PN->getIncomingBlock(i), BB);
366 // LVI returns null is no value could be determined.
368 PushKnownConstantOrUndef(Result, CI, PN->getIncomingBlock(i));
372 return !Result.empty();
375 PredValueInfoTy LHSVals, RHSVals;
377 // Handle some boolean conditions.
378 if (I->getType()->getPrimitiveSizeInBits() == 1) {
380 // X & false -> false
381 if (I->getOpcode() == Instruction::Or ||
382 I->getOpcode() == Instruction::And) {
383 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
384 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals);
386 if (LHSVals.empty() && RHSVals.empty())
389 ConstantInt *InterestingVal;
390 if (I->getOpcode() == Instruction::Or)
391 InterestingVal = ConstantInt::getTrue(I->getContext());
393 InterestingVal = ConstantInt::getFalse(I->getContext());
395 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
397 // Scan for the sentinel. If we find an undef, force it to the
398 // interesting value: x|undef -> true and x&undef -> false.
399 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
400 if (LHSVals[i].first == InterestingVal ||
401 isa<UndefValue>(LHSVals[i].first)) {
402 Result.push_back(LHSVals[i]);
403 Result.back().first = InterestingVal;
404 LHSKnownBBs.insert(LHSVals[i].second);
406 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
407 if (RHSVals[i].first == InterestingVal ||
408 isa<UndefValue>(RHSVals[i].first)) {
409 // If we already inferred a value for this block on the LHS, don't
411 if (!LHSKnownBBs.count(RHSVals[i].second)) {
412 Result.push_back(RHSVals[i]);
413 Result.back().first = InterestingVal;
417 return !Result.empty();
420 // Handle the NOT form of XOR.
421 if (I->getOpcode() == Instruction::Xor &&
422 isa<ConstantInt>(I->getOperand(1)) &&
423 cast<ConstantInt>(I->getOperand(1))->isOne()) {
424 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result);
428 // Invert the known values.
429 for (unsigned i = 0, e = Result.size(); i != e; ++i)
430 Result[i].first = ConstantExpr::getNot(Result[i].first);
435 // Try to simplify some other binary operator values.
436 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
437 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
438 PredValueInfoTy LHSVals;
439 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals);
441 // Try to use constant folding to simplify the binary operator.
442 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
443 Constant *V = LHSVals[i].first;
444 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
446 PushKnownConstantOrUndef(Result, Folded, LHSVals[i].second);
450 return !Result.empty();
453 // Handle compare with phi operand, where the PHI is defined in this block.
454 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
455 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
456 if (PN && PN->getParent() == BB) {
457 // We can do this simplification if any comparisons fold to true or false.
459 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
460 BasicBlock *PredBB = PN->getIncomingBlock(i);
461 Value *LHS = PN->getIncomingValue(i);
462 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
464 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
466 if (!isa<Constant>(RHS))
469 LazyValueInfo::Tristate
470 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
471 cast<Constant>(RHS), PredBB, BB);
472 if (ResT == LazyValueInfo::Unknown)
474 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
477 if (Constant *ConstRes = dyn_cast<Constant>(Res))
478 PushKnownConstantOrUndef(Result, ConstRes, PredBB);
481 return !Result.empty();
485 // If comparing a live-in value against a constant, see if we know the
486 // live-in value on any predecessors.
487 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
488 if (!isa<Instruction>(Cmp->getOperand(0)) ||
489 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
490 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
492 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
494 // If the value is known by LazyValueInfo to be a constant in a
495 // predecessor, use that information to try to thread this block.
496 LazyValueInfo::Tristate Res =
497 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
499 if (Res == LazyValueInfo::Unknown)
502 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
503 Result.push_back(std::make_pair(ResC, P));
506 return !Result.empty();
509 // Try to find a constant value for the LHS of a comparison,
510 // and evaluate it statically if we can.
511 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
512 PredValueInfoTy LHSVals;
513 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
515 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
516 Constant *V = LHSVals[i].first;
517 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
519 PushKnownConstantOrUndef(Result, Folded, LHSVals[i].second);
522 return !Result.empty();
527 // If all else fails, see if LVI can figure out a constant value for us.
528 Constant *CI = LVI->getConstant(V, BB);
529 if (Constant *KC = getKnownConstant(CI)) {
530 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
531 Result.push_back(std::make_pair(KC, *PI));
534 return !Result.empty();
539 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
540 /// in an undefined jump, decide which block is best to revector to.
542 /// Since we can pick an arbitrary destination, we pick the successor with the
543 /// fewest predecessors. This should reduce the in-degree of the others.
545 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
546 TerminatorInst *BBTerm = BB->getTerminator();
547 unsigned MinSucc = 0;
548 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
549 // Compute the successor with the minimum number of predecessors.
550 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
551 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
552 TestBB = BBTerm->getSuccessor(i);
553 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
554 if (NumPreds < MinNumPreds)
561 /// ProcessBlock - If there are any predecessors whose control can be threaded
562 /// through to a successor, transform them now.
563 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
564 // If the block is trivially dead, just return and let the caller nuke it.
565 // This simplifies other transformations.
566 if (pred_begin(BB) == pred_end(BB) &&
567 BB != &BB->getParent()->getEntryBlock())
570 // If this block has a single predecessor, and if that pred has a single
571 // successor, merge the blocks. This encourages recursive jump threading
572 // because now the condition in this block can be threaded through
573 // predecessors of our predecessor block.
574 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
575 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
577 // If SinglePred was a loop header, BB becomes one.
578 if (LoopHeaders.erase(SinglePred))
579 LoopHeaders.insert(BB);
581 // Remember if SinglePred was the entry block of the function. If so, we
582 // will need to move BB back to the entry position.
583 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
584 LVI->eraseBlock(SinglePred);
585 MergeBasicBlockIntoOnlyPred(BB);
587 if (isEntry && BB != &BB->getParent()->getEntryBlock())
588 BB->moveBefore(&BB->getParent()->getEntryBlock());
593 // Look to see if the terminator is a branch of switch, if not we can't thread
596 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
597 // Can't thread an unconditional jump.
598 if (BI->isUnconditional()) return false;
599 Condition = BI->getCondition();
600 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
601 Condition = SI->getCondition();
603 return false; // Must be an invoke.
605 // If the terminator is branching on an undef, we can pick any of the
606 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
607 if (isa<UndefValue>(Condition)) {
608 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
610 // Fold the branch/switch.
611 TerminatorInst *BBTerm = BB->getTerminator();
612 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
613 if (i == BestSucc) continue;
614 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
617 DEBUG(dbgs() << " In block '" << BB->getName()
618 << "' folding undef terminator: " << *BBTerm << '\n');
619 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
620 BBTerm->eraseFromParent();
624 // If the terminator of this block is branching on a constant, simplify the
625 // terminator to an unconditional branch. This can occur due to threading in
627 if (getKnownConstant(Condition)) {
628 DEBUG(dbgs() << " In block '" << BB->getName()
629 << "' folding terminator: " << *BB->getTerminator() << '\n');
631 ConstantFoldTerminator(BB);
635 Instruction *CondInst = dyn_cast<Instruction>(Condition);
637 // All the rest of our checks depend on the condition being an instruction.
639 // FIXME: Unify this with code below.
640 if (ProcessThreadableEdges(Condition, BB))
646 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
647 // For a comparison where the LHS is outside this block, it's possible
648 // that we've branched on it before. Used LVI to see if we can simplify
649 // the branch based on that.
650 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
651 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
652 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
653 if (CondBr && CondConst && CondBr->isConditional() && PI != PE &&
654 (!isa<Instruction>(CondCmp->getOperand(0)) ||
655 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
656 // For predecessor edge, determine if the comparison is true or false
657 // on that edge. If they're all true or all false, we can simplify the
659 // FIXME: We could handle mixed true/false by duplicating code.
660 LazyValueInfo::Tristate Baseline =
661 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
663 if (Baseline != LazyValueInfo::Unknown) {
664 // Check that all remaining incoming values match the first one.
666 LazyValueInfo::Tristate Ret =
667 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
668 CondCmp->getOperand(0), CondConst, *PI, BB);
669 if (Ret != Baseline) break;
672 // If we terminated early, then one of the values didn't match.
674 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
675 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
676 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
677 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
678 CondBr->eraseFromParent();
685 // Check for some cases that are worth simplifying. Right now we want to look
686 // for loads that are used by a switch or by the condition for the branch. If
687 // we see one, check to see if it's partially redundant. If so, insert a PHI
688 // which can then be used to thread the values.
690 Value *SimplifyValue = CondInst;
691 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
692 if (isa<Constant>(CondCmp->getOperand(1)))
693 SimplifyValue = CondCmp->getOperand(0);
695 // TODO: There are other places where load PRE would be profitable, such as
696 // more complex comparisons.
697 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
698 if (SimplifyPartiallyRedundantLoad(LI))
702 // Handle a variety of cases where we are branching on something derived from
703 // a PHI node in the current block. If we can prove that any predecessors
704 // compute a predictable value based on a PHI node, thread those predecessors.
706 if (ProcessThreadableEdges(CondInst, BB))
709 // If this is an otherwise-unfoldable branch on a phi node in the current
710 // block, see if we can simplify.
711 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
712 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
713 return ProcessBranchOnPHI(PN);
716 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
717 if (CondInst->getOpcode() == Instruction::Xor &&
718 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
719 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
722 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
723 // "(X == 4)", thread through this block.
728 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
729 /// block that jump on exactly the same condition. This means that we almost
730 /// always know the direction of the edge in the DESTBB:
732 /// br COND, DESTBB, BBY
734 /// br COND, BBZ, BBW
736 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
737 /// in DESTBB, we have to thread over it.
738 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
740 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
742 // If both successors of PredBB go to DESTBB, we don't know anything. We can
743 // fold the branch to an unconditional one, which allows other recursive
746 if (PredBI->getSuccessor(1) != BB)
748 else if (PredBI->getSuccessor(0) != BB)
751 DEBUG(dbgs() << " In block '" << PredBB->getName()
752 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
754 ConstantFoldTerminator(PredBB);
758 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
760 // If the dest block has one predecessor, just fix the branch condition to a
761 // constant and fold it.
762 if (BB->getSinglePredecessor()) {
763 DEBUG(dbgs() << " In block '" << BB->getName()
764 << "' folding condition to '" << BranchDir << "': "
765 << *BB->getTerminator() << '\n');
767 Value *OldCond = DestBI->getCondition();
768 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
770 // Delete dead instructions before we fold the branch. Folding the branch
771 // can eliminate edges from the CFG which can end up deleting OldCond.
772 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
773 ConstantFoldTerminator(BB);
778 // Next, figure out which successor we are threading to.
779 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
781 SmallVector<BasicBlock*, 2> Preds;
782 Preds.push_back(PredBB);
784 // Ok, try to thread it!
785 return ThreadEdge(BB, Preds, SuccBB);
788 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
789 /// block that switch on exactly the same condition. This means that we almost
790 /// always know the direction of the edge in the DESTBB:
792 /// switch COND [... DESTBB, BBY ... ]
794 /// switch COND [... BBZ, BBW ]
796 /// Optimizing switches like this is very important, because simplifycfg builds
797 /// switches out of repeated 'if' conditions.
798 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
799 BasicBlock *DestBB) {
800 // Can't thread edge to self.
801 if (PredBB == DestBB)
804 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
805 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
807 // There are a variety of optimizations that we can potentially do on these
808 // blocks: we order them from most to least preferable.
810 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
811 // directly to their destination. This does not introduce *any* code size
812 // growth. Skip debug info first.
813 BasicBlock::iterator BBI = DestBB->begin();
814 while (isa<DbgInfoIntrinsic>(BBI))
817 // FIXME: Thread if it just contains a PHI.
818 if (isa<SwitchInst>(BBI)) {
819 bool MadeChange = false;
820 // Ignore the default edge for now.
821 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
822 ConstantInt *DestVal = DestSI->getCaseValue(i);
823 BasicBlock *DestSucc = DestSI->getSuccessor(i);
825 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
826 // PredSI has an explicit case for it. If so, forward. If it is covered
827 // by the default case, we can't update PredSI.
828 unsigned PredCase = PredSI->findCaseValue(DestVal);
829 if (PredCase == 0) continue;
831 // If PredSI doesn't go to DestBB on this value, then it won't reach the
832 // case on this condition.
833 if (PredSI->getSuccessor(PredCase) != DestBB &&
834 DestSI->getSuccessor(i) != DestBB)
837 // Do not forward this if it already goes to this destination, this would
838 // be an infinite loop.
839 if (PredSI->getSuccessor(PredCase) == DestSucc)
842 // Otherwise, we're safe to make the change. Make sure that the edge from
843 // DestSI to DestSucc is not critical and has no PHI nodes.
844 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
845 DEBUG(dbgs() << "THROUGH: " << *DestSI);
847 // If the destination has PHI nodes, just split the edge for updating
849 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
850 SplitCriticalEdge(DestSI, i, this);
851 DestSucc = DestSI->getSuccessor(i);
853 FoldSingleEntryPHINodes(DestSucc);
854 PredSI->setSuccessor(PredCase, DestSucc);
866 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
867 /// load instruction, eliminate it by replacing it with a PHI node. This is an
868 /// important optimization that encourages jump threading, and needs to be run
869 /// interlaced with other jump threading tasks.
870 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
871 // Don't hack volatile loads.
872 if (LI->isVolatile()) return false;
874 // If the load is defined in a block with exactly one predecessor, it can't be
875 // partially redundant.
876 BasicBlock *LoadBB = LI->getParent();
877 if (LoadBB->getSinglePredecessor())
880 Value *LoadedPtr = LI->getOperand(0);
882 // If the loaded operand is defined in the LoadBB, it can't be available.
883 // TODO: Could do simple PHI translation, that would be fun :)
884 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
885 if (PtrOp->getParent() == LoadBB)
888 // Scan a few instructions up from the load, to see if it is obviously live at
889 // the entry to its block.
890 BasicBlock::iterator BBIt = LI;
892 if (Value *AvailableVal =
893 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
894 // If the value if the load is locally available within the block, just use
895 // it. This frequently occurs for reg2mem'd allocas.
896 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
898 // If the returned value is the load itself, replace with an undef. This can
899 // only happen in dead loops.
900 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
901 LI->replaceAllUsesWith(AvailableVal);
902 LI->eraseFromParent();
906 // Otherwise, if we scanned the whole block and got to the top of the block,
907 // we know the block is locally transparent to the load. If not, something
908 // might clobber its value.
909 if (BBIt != LoadBB->begin())
913 SmallPtrSet<BasicBlock*, 8> PredsScanned;
914 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
915 AvailablePredsTy AvailablePreds;
916 BasicBlock *OneUnavailablePred = 0;
918 // If we got here, the loaded value is transparent through to the start of the
919 // block. Check to see if it is available in any of the predecessor blocks.
920 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
922 BasicBlock *PredBB = *PI;
924 // If we already scanned this predecessor, skip it.
925 if (!PredsScanned.insert(PredBB))
928 // Scan the predecessor to see if the value is available in the pred.
929 BBIt = PredBB->end();
930 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
931 if (!PredAvailable) {
932 OneUnavailablePred = PredBB;
936 // If so, this load is partially redundant. Remember this info so that we
937 // can create a PHI node.
938 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
941 // If the loaded value isn't available in any predecessor, it isn't partially
943 if (AvailablePreds.empty()) return false;
945 // Okay, the loaded value is available in at least one (and maybe all!)
946 // predecessors. If the value is unavailable in more than one unique
947 // predecessor, we want to insert a merge block for those common predecessors.
948 // This ensures that we only have to insert one reload, thus not increasing
950 BasicBlock *UnavailablePred = 0;
952 // If there is exactly one predecessor where the value is unavailable, the
953 // already computed 'OneUnavailablePred' block is it. If it ends in an
954 // unconditional branch, we know that it isn't a critical edge.
955 if (PredsScanned.size() == AvailablePreds.size()+1 &&
956 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
957 UnavailablePred = OneUnavailablePred;
958 } else if (PredsScanned.size() != AvailablePreds.size()) {
959 // Otherwise, we had multiple unavailable predecessors or we had a critical
960 // edge from the one.
961 SmallVector<BasicBlock*, 8> PredsToSplit;
962 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
964 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
965 AvailablePredSet.insert(AvailablePreds[i].first);
967 // Add all the unavailable predecessors to the PredsToSplit list.
968 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
971 // If the predecessor is an indirect goto, we can't split the edge.
972 if (isa<IndirectBrInst>(P->getTerminator()))
975 if (!AvailablePredSet.count(P))
976 PredsToSplit.push_back(P);
979 // Split them out to their own block.
981 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
982 "thread-pre-split", this);
985 // If the value isn't available in all predecessors, then there will be
986 // exactly one where it isn't available. Insert a load on that edge and add
987 // it to the AvailablePreds list.
988 if (UnavailablePred) {
989 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
990 "Can't handle critical edge here!");
991 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
993 UnavailablePred->getTerminator());
994 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
997 // Now we know that each predecessor of this block has a value in
998 // AvailablePreds, sort them for efficient access as we're walking the preds.
999 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1001 // Create a PHI node at the start of the block for the PRE'd load value.
1002 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
1005 // Insert new entries into the PHI for each predecessor. A single block may
1006 // have multiple entries here.
1007 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
1009 BasicBlock *P = *PI;
1010 AvailablePredsTy::iterator I =
1011 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1012 std::make_pair(P, (Value*)0));
1014 assert(I != AvailablePreds.end() && I->first == P &&
1015 "Didn't find entry for predecessor!");
1017 PN->addIncoming(I->second, I->first);
1020 //cerr << "PRE: " << *LI << *PN << "\n";
1022 LI->replaceAllUsesWith(PN);
1023 LI->eraseFromParent();
1028 /// FindMostPopularDest - The specified list contains multiple possible
1029 /// threadable destinations. Pick the one that occurs the most frequently in
1032 FindMostPopularDest(BasicBlock *BB,
1033 const SmallVectorImpl<std::pair<BasicBlock*,
1034 BasicBlock*> > &PredToDestList) {
1035 assert(!PredToDestList.empty());
1037 // Determine popularity. If there are multiple possible destinations, we
1038 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1039 // blocks with known and real destinations to threading undef. We'll handle
1040 // them later if interesting.
1041 DenseMap<BasicBlock*, unsigned> DestPopularity;
1042 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1043 if (PredToDestList[i].second)
1044 DestPopularity[PredToDestList[i].second]++;
1046 // Find the most popular dest.
1047 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1048 BasicBlock *MostPopularDest = DPI->first;
1049 unsigned Popularity = DPI->second;
1050 SmallVector<BasicBlock*, 4> SamePopularity;
1052 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1053 // If the popularity of this entry isn't higher than the popularity we've
1054 // seen so far, ignore it.
1055 if (DPI->second < Popularity)
1057 else if (DPI->second == Popularity) {
1058 // If it is the same as what we've seen so far, keep track of it.
1059 SamePopularity.push_back(DPI->first);
1061 // If it is more popular, remember it.
1062 SamePopularity.clear();
1063 MostPopularDest = DPI->first;
1064 Popularity = DPI->second;
1068 // Okay, now we know the most popular destination. If there is more than
1069 // destination, we need to determine one. This is arbitrary, but we need
1070 // to make a deterministic decision. Pick the first one that appears in the
1072 if (!SamePopularity.empty()) {
1073 SamePopularity.push_back(MostPopularDest);
1074 TerminatorInst *TI = BB->getTerminator();
1075 for (unsigned i = 0; ; ++i) {
1076 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1078 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1079 TI->getSuccessor(i)) == SamePopularity.end())
1082 MostPopularDest = TI->getSuccessor(i);
1087 // Okay, we have finally picked the most popular destination.
1088 return MostPopularDest;
1091 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) {
1092 // If threading this would thread across a loop header, don't even try to
1094 if (LoopHeaders.count(BB))
1097 PredValueInfoTy PredValues;
1098 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues))
1101 assert(!PredValues.empty() &&
1102 "ComputeValueKnownInPredecessors returned true with no values");
1104 DEBUG(dbgs() << "IN BB: " << *BB;
1105 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1106 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1107 << *PredValues[i].first
1108 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1111 // Decide what we want to thread through. Convert our list of known values to
1112 // a list of known destinations for each pred. This also discards duplicate
1113 // predecessors and keeps track of the undefined inputs (which are represented
1114 // as a null dest in the PredToDestList).
1115 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1116 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1118 BasicBlock *OnlyDest = 0;
1119 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1121 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1122 BasicBlock *Pred = PredValues[i].second;
1123 if (!SeenPreds.insert(Pred))
1124 continue; // Duplicate predecessor entry.
1126 // If the predecessor ends with an indirect goto, we can't change its
1128 if (isa<IndirectBrInst>(Pred->getTerminator()))
1131 Constant *Val = PredValues[i].first;
1134 if (isa<UndefValue>(Val))
1136 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1137 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1139 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
1140 DestBB = SI->getSuccessor(SI->findCaseValue(cast<ConstantInt>(Val)));
1143 // If we have exactly one destination, remember it for efficiency below.
1146 else if (OnlyDest != DestBB)
1147 OnlyDest = MultipleDestSentinel;
1149 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1152 // If all edges were unthreadable, we fail.
1153 if (PredToDestList.empty())
1156 // Determine which is the most common successor. If we have many inputs and
1157 // this block is a switch, we want to start by threading the batch that goes
1158 // to the most popular destination first. If we only know about one
1159 // threadable destination (the common case) we can avoid this.
1160 BasicBlock *MostPopularDest = OnlyDest;
1162 if (MostPopularDest == MultipleDestSentinel)
1163 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1165 // Now that we know what the most popular destination is, factor all
1166 // predecessors that will jump to it into a single predecessor.
1167 SmallVector<BasicBlock*, 16> PredsToFactor;
1168 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1169 if (PredToDestList[i].second == MostPopularDest) {
1170 BasicBlock *Pred = PredToDestList[i].first;
1172 // This predecessor may be a switch or something else that has multiple
1173 // edges to the block. Factor each of these edges by listing them
1174 // according to # occurrences in PredsToFactor.
1175 TerminatorInst *PredTI = Pred->getTerminator();
1176 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1177 if (PredTI->getSuccessor(i) == BB)
1178 PredsToFactor.push_back(Pred);
1181 // If the threadable edges are branching on an undefined value, we get to pick
1182 // the destination that these predecessors should get to.
1183 if (MostPopularDest == 0)
1184 MostPopularDest = BB->getTerminator()->
1185 getSuccessor(GetBestDestForJumpOnUndef(BB));
1187 // Ok, try to thread it!
1188 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1191 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1192 /// a PHI node in the current block. See if there are any simplifications we
1193 /// can do based on inputs to the phi node.
1195 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1196 BasicBlock *BB = PN->getParent();
1198 // TODO: We could make use of this to do it once for blocks with common PHI
1200 SmallVector<BasicBlock*, 1> PredBBs;
1203 // If any of the predecessor blocks end in an unconditional branch, we can
1204 // *duplicate* the conditional branch into that block in order to further
1205 // encourage jump threading and to eliminate cases where we have branch on a
1206 // phi of an icmp (branch on icmp is much better).
1207 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1208 BasicBlock *PredBB = PN->getIncomingBlock(i);
1209 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1210 if (PredBr->isUnconditional()) {
1211 PredBBs[0] = PredBB;
1212 // Try to duplicate BB into PredBB.
1213 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1221 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1222 /// a xor instruction in the current block. See if there are any
1223 /// simplifications we can do based on inputs to the xor.
1225 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1226 BasicBlock *BB = BO->getParent();
1228 // If either the LHS or RHS of the xor is a constant, don't do this
1230 if (isa<ConstantInt>(BO->getOperand(0)) ||
1231 isa<ConstantInt>(BO->getOperand(1)))
1234 // If the first instruction in BB isn't a phi, we won't be able to infer
1235 // anything special about any particular predecessor.
1236 if (!isa<PHINode>(BB->front()))
1239 // If we have a xor as the branch input to this block, and we know that the
1240 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1241 // the condition into the predecessor and fix that value to true, saving some
1242 // logical ops on that path and encouraging other paths to simplify.
1244 // This copies something like this:
1247 // %X = phi i1 [1], [%X']
1248 // %Y = icmp eq i32 %A, %B
1249 // %Z = xor i1 %X, %Y
1254 // %Y = icmp ne i32 %A, %B
1257 PredValueInfoTy XorOpValues;
1259 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) {
1260 assert(XorOpValues.empty());
1261 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues))
1266 assert(!XorOpValues.empty() &&
1267 "ComputeValueKnownInPredecessors returned true with no values");
1269 // Scan the information to see which is most popular: true or false. The
1270 // predecessors can be of the set true, false, or undef.
1271 unsigned NumTrue = 0, NumFalse = 0;
1272 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1273 if (isa<UndefValue>(XorOpValues[i].first))
1274 // Ignore undefs for the count.
1276 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1282 // Determine which value to split on, true, false, or undef if neither.
1283 ConstantInt *SplitVal = 0;
1284 if (NumTrue > NumFalse)
1285 SplitVal = ConstantInt::getTrue(BB->getContext());
1286 else if (NumTrue != 0 || NumFalse != 0)
1287 SplitVal = ConstantInt::getFalse(BB->getContext());
1289 // Collect all of the blocks that this can be folded into so that we can
1290 // factor this once and clone it once.
1291 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1292 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1293 if (XorOpValues[i].first != SplitVal &&
1294 !isa<UndefValue>(XorOpValues[i].first))
1297 BlocksToFoldInto.push_back(XorOpValues[i].second);
1300 // If we inferred a value for all of the predecessors, then duplication won't
1301 // help us. However, we can just replace the LHS or RHS with the constant.
1302 if (BlocksToFoldInto.size() ==
1303 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1304 if (SplitVal == 0) {
1305 // If all preds provide undef, just nuke the xor, because it is undef too.
1306 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1307 BO->eraseFromParent();
1308 } else if (SplitVal->isZero()) {
1309 // If all preds provide 0, replace the xor with the other input.
1310 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1311 BO->eraseFromParent();
1313 // If all preds provide 1, set the computed value to 1.
1314 BO->setOperand(!isLHS, SplitVal);
1320 // Try to duplicate BB into PredBB.
1321 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1325 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1326 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1327 /// NewPred using the entries from OldPred (suitably mapped).
1328 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1329 BasicBlock *OldPred,
1330 BasicBlock *NewPred,
1331 DenseMap<Instruction*, Value*> &ValueMap) {
1332 for (BasicBlock::iterator PNI = PHIBB->begin();
1333 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1334 // Ok, we have a PHI node. Figure out what the incoming value was for the
1336 Value *IV = PN->getIncomingValueForBlock(OldPred);
1338 // Remap the value if necessary.
1339 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1340 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1341 if (I != ValueMap.end())
1345 PN->addIncoming(IV, NewPred);
1349 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1350 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1351 /// across BB. Transform the IR to reflect this change.
1352 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1353 const SmallVectorImpl<BasicBlock*> &PredBBs,
1354 BasicBlock *SuccBB) {
1355 // If threading to the same block as we come from, we would infinite loop.
1357 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1358 << "' - would thread to self!\n");
1362 // If threading this would thread across a loop header, don't thread the edge.
1363 // See the comments above FindLoopHeaders for justifications and caveats.
1364 if (LoopHeaders.count(BB)) {
1365 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1366 << "' to dest BB '" << SuccBB->getName()
1367 << "' - it might create an irreducible loop!\n");
1371 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1372 if (JumpThreadCost > Threshold) {
1373 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1374 << "' - Cost is too high: " << JumpThreadCost << "\n");
1378 // And finally, do it! Start by factoring the predecessors is needed.
1380 if (PredBBs.size() == 1)
1381 PredBB = PredBBs[0];
1383 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1384 << " common predecessors.\n");
1385 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1389 // And finally, do it!
1390 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1391 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1392 << ", across block:\n "
1395 LVI->threadEdge(PredBB, BB, SuccBB);
1397 // We are going to have to map operands from the original BB block to the new
1398 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1399 // account for entry from PredBB.
1400 DenseMap<Instruction*, Value*> ValueMapping;
1402 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1403 BB->getName()+".thread",
1404 BB->getParent(), BB);
1405 NewBB->moveAfter(PredBB);
1407 BasicBlock::iterator BI = BB->begin();
1408 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1409 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1411 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1412 // mapping and using it to remap operands in the cloned instructions.
1413 for (; !isa<TerminatorInst>(BI); ++BI) {
1414 Instruction *New = BI->clone();
1415 New->setName(BI->getName());
1416 NewBB->getInstList().push_back(New);
1417 ValueMapping[BI] = New;
1419 // Remap operands to patch up intra-block references.
1420 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1421 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1422 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1423 if (I != ValueMapping.end())
1424 New->setOperand(i, I->second);
1428 // We didn't copy the terminator from BB over to NewBB, because there is now
1429 // an unconditional jump to SuccBB. Insert the unconditional jump.
1430 BranchInst::Create(SuccBB, NewBB);
1432 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1433 // PHI nodes for NewBB now.
1434 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1436 // If there were values defined in BB that are used outside the block, then we
1437 // now have to update all uses of the value to use either the original value,
1438 // the cloned value, or some PHI derived value. This can require arbitrary
1439 // PHI insertion, of which we are prepared to do, clean these up now.
1440 SSAUpdater SSAUpdate;
1441 SmallVector<Use*, 16> UsesToRename;
1442 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1443 // Scan all uses of this instruction to see if it is used outside of its
1444 // block, and if so, record them in UsesToRename.
1445 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1447 Instruction *User = cast<Instruction>(*UI);
1448 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1449 if (UserPN->getIncomingBlock(UI) == BB)
1451 } else if (User->getParent() == BB)
1454 UsesToRename.push_back(&UI.getUse());
1457 // If there are no uses outside the block, we're done with this instruction.
1458 if (UsesToRename.empty())
1461 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1463 // We found a use of I outside of BB. Rename all uses of I that are outside
1464 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1465 // with the two values we know.
1466 SSAUpdate.Initialize(I->getType(), I->getName());
1467 SSAUpdate.AddAvailableValue(BB, I);
1468 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1470 while (!UsesToRename.empty())
1471 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1472 DEBUG(dbgs() << "\n");
1476 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1477 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1478 // us to simplify any PHI nodes in BB.
1479 TerminatorInst *PredTerm = PredBB->getTerminator();
1480 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1481 if (PredTerm->getSuccessor(i) == BB) {
1482 BB->removePredecessor(PredBB, true);
1483 PredTerm->setSuccessor(i, NewBB);
1486 // At this point, the IR is fully up to date and consistent. Do a quick scan
1487 // over the new instructions and zap any that are constants or dead. This
1488 // frequently happens because of phi translation.
1489 SimplifyInstructionsInBlock(NewBB, TD);
1491 // Threaded an edge!
1496 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1497 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1498 /// If we can duplicate the contents of BB up into PredBB do so now, this
1499 /// improves the odds that the branch will be on an analyzable instruction like
1501 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1502 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1503 assert(!PredBBs.empty() && "Can't handle an empty set");
1505 // If BB is a loop header, then duplicating this block outside the loop would
1506 // cause us to transform this into an irreducible loop, don't do this.
1507 // See the comments above FindLoopHeaders for justifications and caveats.
1508 if (LoopHeaders.count(BB)) {
1509 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1510 << "' into predecessor block '" << PredBBs[0]->getName()
1511 << "' - it might create an irreducible loop!\n");
1515 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1516 if (DuplicationCost > Threshold) {
1517 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1518 << "' - Cost is too high: " << DuplicationCost << "\n");
1522 // And finally, do it! Start by factoring the predecessors is needed.
1524 if (PredBBs.size() == 1)
1525 PredBB = PredBBs[0];
1527 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1528 << " common predecessors.\n");
1529 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1533 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1535 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1536 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1537 << DuplicationCost << " block is:" << *BB << "\n");
1539 // Unless PredBB ends with an unconditional branch, split the edge so that we
1540 // can just clone the bits from BB into the end of the new PredBB.
1541 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1543 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1544 PredBB = SplitEdge(PredBB, BB, this);
1545 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1548 // We are going to have to map operands from the original BB block into the
1549 // PredBB block. Evaluate PHI nodes in BB.
1550 DenseMap<Instruction*, Value*> ValueMapping;
1552 BasicBlock::iterator BI = BB->begin();
1553 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1554 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1556 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1557 // mapping and using it to remap operands in the cloned instructions.
1558 for (; BI != BB->end(); ++BI) {
1559 Instruction *New = BI->clone();
1561 // Remap operands to patch up intra-block references.
1562 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1563 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1564 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1565 if (I != ValueMapping.end())
1566 New->setOperand(i, I->second);
1569 // If this instruction can be simplified after the operands are updated,
1570 // just use the simplified value instead. This frequently happens due to
1572 if (Value *IV = SimplifyInstruction(New, TD)) {
1574 ValueMapping[BI] = IV;
1576 // Otherwise, insert the new instruction into the block.
1577 New->setName(BI->getName());
1578 PredBB->getInstList().insert(OldPredBranch, New);
1579 ValueMapping[BI] = New;
1583 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1584 // add entries to the PHI nodes for branch from PredBB now.
1585 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1586 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1588 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1591 // If there were values defined in BB that are used outside the block, then we
1592 // now have to update all uses of the value to use either the original value,
1593 // the cloned value, or some PHI derived value. This can require arbitrary
1594 // PHI insertion, of which we are prepared to do, clean these up now.
1595 SSAUpdater SSAUpdate;
1596 SmallVector<Use*, 16> UsesToRename;
1597 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1598 // Scan all uses of this instruction to see if it is used outside of its
1599 // block, and if so, record them in UsesToRename.
1600 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1602 Instruction *User = cast<Instruction>(*UI);
1603 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1604 if (UserPN->getIncomingBlock(UI) == BB)
1606 } else if (User->getParent() == BB)
1609 UsesToRename.push_back(&UI.getUse());
1612 // If there are no uses outside the block, we're done with this instruction.
1613 if (UsesToRename.empty())
1616 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1618 // We found a use of I outside of BB. Rename all uses of I that are outside
1619 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1620 // with the two values we know.
1621 SSAUpdate.Initialize(I->getType(), I->getName());
1622 SSAUpdate.AddAvailableValue(BB, I);
1623 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1625 while (!UsesToRename.empty())
1626 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1627 DEBUG(dbgs() << "\n");
1630 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1632 BB->removePredecessor(PredBB, true);
1634 // Remove the unconditional branch at the end of the PredBB block.
1635 OldPredBranch->eraseFromParent();