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 is used to keep track of what kind of constant we're currently hoping
54 enum ConstantPreference {
59 /// This pass performs 'jump threading', which looks at blocks that have
60 /// multiple predecessors and multiple successors. If one or more of the
61 /// predecessors of the block can be proven to always jump to one of the
62 /// successors, we forward the edge from the predecessor to the successor by
63 /// duplicating the contents of this block.
65 /// An example of when this can occur is code like this:
72 /// In this case, the unconditional branch at the end of the first if can be
73 /// revectored to the false side of the second if.
75 class JumpThreading : public FunctionPass {
79 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
81 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
83 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet;
85 // RAII helper for updating the recursion stack.
86 struct RecursionSetRemover {
87 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet;
88 std::pair<Value*, BasicBlock*> ThePair;
90 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S,
91 std::pair<Value*, BasicBlock*> P)
92 : TheSet(S), ThePair(P) { }
94 ~RecursionSetRemover() {
95 TheSet.erase(ThePair);
99 static char ID; // Pass identification
100 JumpThreading() : FunctionPass(ID) {
101 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
104 bool runOnFunction(Function &F);
106 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
107 AU.addRequired<LazyValueInfo>();
108 AU.addPreserved<LazyValueInfo>();
111 void FindLoopHeaders(Function &F);
112 bool ProcessBlock(BasicBlock *BB);
113 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
115 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
116 const SmallVectorImpl<BasicBlock *> &PredBBs);
118 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
119 PredValueInfo &Result,
120 ConstantPreference Preference);
121 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
122 ConstantPreference Preference);
124 bool ProcessBranchOnPHI(PHINode *PN);
125 bool ProcessBranchOnXOR(BinaryOperator *BO);
127 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
131 char JumpThreading::ID = 0;
132 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
133 "Jump Threading", false, false)
134 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)
135 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
136 "Jump Threading", false, false)
138 // Public interface to the Jump Threading pass
139 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
141 /// runOnFunction - Top level algorithm.
143 bool JumpThreading::runOnFunction(Function &F) {
144 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
145 TD = getAnalysisIfAvailable<TargetData>();
146 LVI = &getAnalysis<LazyValueInfo>();
150 bool Changed, EverChanged = false;
153 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
155 // Thread all of the branches we can over this block.
156 while (ProcessBlock(BB))
161 // If the block is trivially dead, zap it. This eliminates the successor
162 // edges which simplifies the CFG.
163 if (pred_begin(BB) == pred_end(BB) &&
164 BB != &BB->getParent()->getEntryBlock()) {
165 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
166 << "' with terminator: " << *BB->getTerminator() << '\n');
167 LoopHeaders.erase(BB);
171 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
172 // Can't thread an unconditional jump, but if the block is "almost
173 // empty", we can replace uses of it with uses of the successor and make
175 if (BI->isUnconditional() &&
176 BB != &BB->getParent()->getEntryBlock()) {
177 BasicBlock::iterator BBI = BB->getFirstNonPHI();
178 // Ignore dbg intrinsics.
179 while (isa<DbgInfoIntrinsic>(BBI))
181 // If the terminator is the only non-phi instruction, try to nuke it.
182 if (BBI->isTerminator()) {
183 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
184 // block, we have to make sure it isn't in the LoopHeaders set. We
185 // reinsert afterward if needed.
186 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
187 BasicBlock *Succ = BI->getSuccessor(0);
189 // FIXME: It is always conservatively correct to drop the info
190 // for a block even if it doesn't get erased. This isn't totally
191 // awesome, but it allows us to use AssertingVH to prevent nasty
192 // dangling pointer issues within LazyValueInfo.
194 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
196 // If we deleted BB and BB was the header of a loop, then the
197 // successor is now the header of the loop.
201 if (ErasedFromLoopHeaders)
202 LoopHeaders.insert(BB);
207 EverChanged |= Changed;
214 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
215 /// thread across it.
216 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
217 /// Ignore PHI nodes, these will be flattened when duplication happens.
218 BasicBlock::const_iterator I = BB->getFirstNonPHI();
220 // FIXME: THREADING will delete values that are just used to compute the
221 // branch, so they shouldn't count against the duplication cost.
224 // Sum up the cost of each instruction until we get to the terminator. Don't
225 // include the terminator because the copy won't include it.
227 for (; !isa<TerminatorInst>(I); ++I) {
228 // Debugger intrinsics don't incur code size.
229 if (isa<DbgInfoIntrinsic>(I)) continue;
231 // If this is a pointer->pointer bitcast, it is free.
232 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
235 // All other instructions count for at least one unit.
238 // Calls are more expensive. If they are non-intrinsic calls, we model them
239 // as having cost of 4. If they are a non-vector intrinsic, we model them
240 // as having cost of 2 total, and if they are a vector intrinsic, we model
241 // them as having cost 1.
242 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
243 if (!isa<IntrinsicInst>(CI))
245 else if (!CI->getType()->isVectorTy())
250 // Threading through a switch statement is particularly profitable. If this
251 // block ends in a switch, decrease its cost to make it more likely to happen.
252 if (isa<SwitchInst>(I))
253 Size = Size > 6 ? Size-6 : 0;
255 // The same holds for indirect branches, but slightly more so.
256 if (isa<IndirectBrInst>(I))
257 Size = Size > 8 ? Size-8 : 0;
262 /// FindLoopHeaders - We do not want jump threading to turn proper loop
263 /// structures into irreducible loops. Doing this breaks up the loop nesting
264 /// hierarchy and pessimizes later transformations. To prevent this from
265 /// happening, we first have to find the loop headers. Here we approximate this
266 /// by finding targets of backedges in the CFG.
268 /// Note that there definitely are cases when we want to allow threading of
269 /// edges across a loop header. For example, threading a jump from outside the
270 /// loop (the preheader) to an exit block of the loop is definitely profitable.
271 /// It is also almost always profitable to thread backedges from within the loop
272 /// to exit blocks, and is often profitable to thread backedges to other blocks
273 /// within the loop (forming a nested loop). This simple analysis is not rich
274 /// enough to track all of these properties and keep it up-to-date as the CFG
275 /// mutates, so we don't allow any of these transformations.
277 void JumpThreading::FindLoopHeaders(Function &F) {
278 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
279 FindFunctionBackedges(F, Edges);
281 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
282 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
285 /// getKnownConstant - Helper method to determine if we can thread over a
286 /// terminator with the given value as its condition, and if so what value to
287 /// use for that. What kind of value this is depends on whether we want an
288 /// integer or a block address, but an undef is always accepted.
289 /// Returns null if Val is null or not an appropriate constant.
290 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
294 // Undef is "known" enough.
295 if (UndefValue *U = dyn_cast<UndefValue>(Val))
298 if (Preference == WantBlockAddress)
299 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
301 return dyn_cast<ConstantInt>(Val);
304 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
305 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
306 /// in any of our predecessors. If so, return the known list of value and pred
307 /// BB in the result vector.
309 /// This returns true if there were any known values.
312 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
313 ConstantPreference Preference) {
314 // This method walks up use-def chains recursively. Because of this, we could
315 // get into an infinite loop going around loops in the use-def chain. To
316 // prevent this, keep track of what (value, block) pairs we've already visited
317 // and terminate the search if we loop back to them
318 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
321 // An RAII help to remove this pair from the recursion set once the recursion
322 // stack pops back out again.
323 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
325 // If V is a constant, then it is known in all predecessors.
326 if (Constant *KC = getKnownConstant(V, Preference)) {
327 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
328 Result.push_back(std::make_pair(KC, *PI));
333 // If V is a non-instruction value, or an instruction in a different block,
334 // then it can't be derived from a PHI.
335 Instruction *I = dyn_cast<Instruction>(V);
336 if (I == 0 || I->getParent() != BB) {
338 // Okay, if this is a live-in value, see if it has a known value at the end
339 // of any of our predecessors.
341 // FIXME: This should be an edge property, not a block end property.
342 /// TODO: Per PR2563, we could infer value range information about a
343 /// predecessor based on its terminator.
345 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
346 // "I" is a non-local compare-with-a-constant instruction. This would be
347 // able to handle value inequalities better, for example if the compare is
348 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
349 // Perhaps getConstantOnEdge should be smart enough to do this?
351 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
353 // If the value is known by LazyValueInfo to be a constant in a
354 // predecessor, use that information to try to thread this block.
355 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
356 if (Constant *KC = getKnownConstant(PredCst, Preference))
357 Result.push_back(std::make_pair(KC, P));
360 return !Result.empty();
363 /// If I is a PHI node, then we know the incoming values for any constants.
364 if (PHINode *PN = dyn_cast<PHINode>(I)) {
365 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
366 Value *InVal = PN->getIncomingValue(i);
367 if (Constant *KC = getKnownConstant(InVal, Preference)) {
368 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
370 Constant *CI = LVI->getConstantOnEdge(InVal,
371 PN->getIncomingBlock(i), BB);
372 if (Constant *KC = getKnownConstant(CI, Preference))
373 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
377 return !Result.empty();
380 PredValueInfoTy LHSVals, RHSVals;
382 // Handle some boolean conditions.
383 if (I->getType()->getPrimitiveSizeInBits() == 1) {
384 assert(Preference == WantInteger && "One-bit non-integer type?");
386 // X & false -> false
387 if (I->getOpcode() == Instruction::Or ||
388 I->getOpcode() == Instruction::And) {
389 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
391 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
394 if (LHSVals.empty() && RHSVals.empty())
397 ConstantInt *InterestingVal;
398 if (I->getOpcode() == Instruction::Or)
399 InterestingVal = ConstantInt::getTrue(I->getContext());
401 InterestingVal = ConstantInt::getFalse(I->getContext());
403 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
405 // Scan for the sentinel. If we find an undef, force it to the
406 // interesting value: x|undef -> true and x&undef -> false.
407 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
408 if (LHSVals[i].first == InterestingVal ||
409 isa<UndefValue>(LHSVals[i].first)) {
410 Result.push_back(LHSVals[i]);
411 Result.back().first = InterestingVal;
412 LHSKnownBBs.insert(LHSVals[i].second);
414 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
415 if (RHSVals[i].first == InterestingVal ||
416 isa<UndefValue>(RHSVals[i].first)) {
417 // If we already inferred a value for this block on the LHS, don't
419 if (!LHSKnownBBs.count(RHSVals[i].second)) {
420 Result.push_back(RHSVals[i]);
421 Result.back().first = InterestingVal;
425 return !Result.empty();
428 // Handle the NOT form of XOR.
429 if (I->getOpcode() == Instruction::Xor &&
430 isa<ConstantInt>(I->getOperand(1)) &&
431 cast<ConstantInt>(I->getOperand(1))->isOne()) {
432 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
437 // Invert the known values.
438 for (unsigned i = 0, e = Result.size(); i != e; ++i)
439 Result[i].first = ConstantExpr::getNot(Result[i].first);
444 // Try to simplify some other binary operator values.
445 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
446 assert(Preference != WantBlockAddress
447 && "A binary operator creating a block address?");
448 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
449 PredValueInfoTy LHSVals;
450 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
453 // Try to use constant folding to simplify the binary operator.
454 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
455 Constant *V = LHSVals[i].first;
456 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
458 if (Constant *KC = getKnownConstant(Folded, WantInteger))
459 Result.push_back(std::make_pair(KC, LHSVals[i].second));
463 return !Result.empty();
466 // Handle compare with phi operand, where the PHI is defined in this block.
467 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
468 assert(Preference == WantInteger && "Compares only produce integers");
469 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
470 if (PN && PN->getParent() == BB) {
471 // We can do this simplification if any comparisons fold to true or false.
473 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
474 BasicBlock *PredBB = PN->getIncomingBlock(i);
475 Value *LHS = PN->getIncomingValue(i);
476 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
478 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
480 if (!isa<Constant>(RHS))
483 LazyValueInfo::Tristate
484 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
485 cast<Constant>(RHS), PredBB, BB);
486 if (ResT == LazyValueInfo::Unknown)
488 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
491 if (Constant *KC = getKnownConstant(Res, WantInteger))
492 Result.push_back(std::make_pair(KC, PredBB));
495 return !Result.empty();
499 // If comparing a live-in value against a constant, see if we know the
500 // live-in value on any predecessors.
501 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
502 if (!isa<Instruction>(Cmp->getOperand(0)) ||
503 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
504 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
506 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
508 // If the value is known by LazyValueInfo to be a constant in a
509 // predecessor, use that information to try to thread this block.
510 LazyValueInfo::Tristate Res =
511 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
513 if (Res == LazyValueInfo::Unknown)
516 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
517 Result.push_back(std::make_pair(ResC, P));
520 return !Result.empty();
523 // Try to find a constant value for the LHS of a comparison,
524 // and evaluate it statically if we can.
525 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
526 PredValueInfoTy LHSVals;
527 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
530 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
531 Constant *V = LHSVals[i].first;
532 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
534 if (Constant *KC = getKnownConstant(Folded, WantInteger))
535 Result.push_back(std::make_pair(KC, LHSVals[i].second));
538 return !Result.empty();
543 // If all else fails, see if LVI can figure out a constant value for us.
544 Constant *CI = LVI->getConstant(V, BB);
545 if (Constant *KC = getKnownConstant(CI, Preference)) {
546 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
547 Result.push_back(std::make_pair(KC, *PI));
550 return !Result.empty();
555 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
556 /// in an undefined jump, decide which block is best to revector to.
558 /// Since we can pick an arbitrary destination, we pick the successor with the
559 /// fewest predecessors. This should reduce the in-degree of the others.
561 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
562 TerminatorInst *BBTerm = BB->getTerminator();
563 unsigned MinSucc = 0;
564 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
565 // Compute the successor with the minimum number of predecessors.
566 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
567 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
568 TestBB = BBTerm->getSuccessor(i);
569 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
570 if (NumPreds < MinNumPreds)
577 /// ProcessBlock - If there are any predecessors whose control can be threaded
578 /// through to a successor, transform them now.
579 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
580 // If the block is trivially dead, just return and let the caller nuke it.
581 // This simplifies other transformations.
582 if (pred_begin(BB) == pred_end(BB) &&
583 BB != &BB->getParent()->getEntryBlock())
586 // If this block has a single predecessor, and if that pred has a single
587 // successor, merge the blocks. This encourages recursive jump threading
588 // because now the condition in this block can be threaded through
589 // predecessors of our predecessor block.
590 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
591 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
593 // If SinglePred was a loop header, BB becomes one.
594 if (LoopHeaders.erase(SinglePred))
595 LoopHeaders.insert(BB);
597 // Remember if SinglePred was the entry block of the function. If so, we
598 // will need to move BB back to the entry position.
599 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
600 LVI->eraseBlock(SinglePred);
601 MergeBasicBlockIntoOnlyPred(BB);
603 if (isEntry && BB != &BB->getParent()->getEntryBlock())
604 BB->moveBefore(&BB->getParent()->getEntryBlock());
609 // What kind of constant we're looking for.
610 ConstantPreference Preference = WantInteger;
612 // Look to see if the terminator is a conditional branch, switch or indirect
613 // branch, if not we can't thread it.
615 Instruction *Terminator = BB->getTerminator();
616 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
617 // Can't thread an unconditional jump.
618 if (BI->isUnconditional()) return false;
619 Condition = BI->getCondition();
620 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
621 Condition = SI->getCondition();
622 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
623 Condition = IB->getAddress()->stripPointerCasts();
624 Preference = WantBlockAddress;
626 return false; // Must be an invoke.
629 // If the terminator is branching on an undef, we can pick any of the
630 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
631 if (isa<UndefValue>(Condition)) {
632 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
634 // Fold the branch/switch.
635 TerminatorInst *BBTerm = BB->getTerminator();
636 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
637 if (i == BestSucc) continue;
638 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
641 DEBUG(dbgs() << " In block '" << BB->getName()
642 << "' folding undef terminator: " << *BBTerm << '\n');
643 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
644 BBTerm->eraseFromParent();
648 // If the terminator of this block is branching on a constant, simplify the
649 // terminator to an unconditional branch. This can occur due to threading in
651 if (getKnownConstant(Condition, Preference)) {
652 DEBUG(dbgs() << " In block '" << BB->getName()
653 << "' folding terminator: " << *BB->getTerminator() << '\n');
655 ConstantFoldTerminator(BB);
659 Instruction *CondInst = dyn_cast<Instruction>(Condition);
661 // All the rest of our checks depend on the condition being an instruction.
663 // FIXME: Unify this with code below.
664 if (ProcessThreadableEdges(Condition, BB, Preference))
670 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
671 // For a comparison where the LHS is outside this block, it's possible
672 // that we've branched on it before. Used LVI to see if we can simplify
673 // the branch based on that.
674 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
675 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
676 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
677 if (CondBr && CondConst && CondBr->isConditional() && PI != PE &&
678 (!isa<Instruction>(CondCmp->getOperand(0)) ||
679 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
680 // For predecessor edge, determine if the comparison is true or false
681 // on that edge. If they're all true or all false, we can simplify the
683 // FIXME: We could handle mixed true/false by duplicating code.
684 LazyValueInfo::Tristate Baseline =
685 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
687 if (Baseline != LazyValueInfo::Unknown) {
688 // Check that all remaining incoming values match the first one.
690 LazyValueInfo::Tristate Ret =
691 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
692 CondCmp->getOperand(0), CondConst, *PI, BB);
693 if (Ret != Baseline) break;
696 // If we terminated early, then one of the values didn't match.
698 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
699 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
700 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
701 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
702 CondBr->eraseFromParent();
709 // Check for some cases that are worth simplifying. Right now we want to look
710 // for loads that are used by a switch or by the condition for the branch. If
711 // we see one, check to see if it's partially redundant. If so, insert a PHI
712 // which can then be used to thread the values.
714 Value *SimplifyValue = CondInst;
715 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
716 if (isa<Constant>(CondCmp->getOperand(1)))
717 SimplifyValue = CondCmp->getOperand(0);
719 // TODO: There are other places where load PRE would be profitable, such as
720 // more complex comparisons.
721 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
722 if (SimplifyPartiallyRedundantLoad(LI))
726 // Handle a variety of cases where we are branching on something derived from
727 // a PHI node in the current block. If we can prove that any predecessors
728 // compute a predictable value based on a PHI node, thread those predecessors.
730 if (ProcessThreadableEdges(CondInst, BB, Preference))
733 // If this is an otherwise-unfoldable branch on a phi node in the current
734 // block, see if we can simplify.
735 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
736 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
737 return ProcessBranchOnPHI(PN);
740 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
741 if (CondInst->getOpcode() == Instruction::Xor &&
742 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
743 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
746 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
747 // "(X == 4)", thread through this block.
753 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
754 /// load instruction, eliminate it by replacing it with a PHI node. This is an
755 /// important optimization that encourages jump threading, and needs to be run
756 /// interlaced with other jump threading tasks.
757 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
758 // Don't hack volatile loads.
759 if (LI->isVolatile()) return false;
761 // If the load is defined in a block with exactly one predecessor, it can't be
762 // partially redundant.
763 BasicBlock *LoadBB = LI->getParent();
764 if (LoadBB->getSinglePredecessor())
767 Value *LoadedPtr = LI->getOperand(0);
769 // If the loaded operand is defined in the LoadBB, it can't be available.
770 // TODO: Could do simple PHI translation, that would be fun :)
771 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
772 if (PtrOp->getParent() == LoadBB)
775 // Scan a few instructions up from the load, to see if it is obviously live at
776 // the entry to its block.
777 BasicBlock::iterator BBIt = LI;
779 if (Value *AvailableVal =
780 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
781 // If the value if the load is locally available within the block, just use
782 // it. This frequently occurs for reg2mem'd allocas.
783 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
785 // If the returned value is the load itself, replace with an undef. This can
786 // only happen in dead loops.
787 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
788 LI->replaceAllUsesWith(AvailableVal);
789 LI->eraseFromParent();
793 // Otherwise, if we scanned the whole block and got to the top of the block,
794 // we know the block is locally transparent to the load. If not, something
795 // might clobber its value.
796 if (BBIt != LoadBB->begin())
800 SmallPtrSet<BasicBlock*, 8> PredsScanned;
801 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
802 AvailablePredsTy AvailablePreds;
803 BasicBlock *OneUnavailablePred = 0;
805 // If we got here, the loaded value is transparent through to the start of the
806 // block. Check to see if it is available in any of the predecessor blocks.
807 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
809 BasicBlock *PredBB = *PI;
811 // If we already scanned this predecessor, skip it.
812 if (!PredsScanned.insert(PredBB))
815 // Scan the predecessor to see if the value is available in the pred.
816 BBIt = PredBB->end();
817 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
818 if (!PredAvailable) {
819 OneUnavailablePred = PredBB;
823 // If so, this load is partially redundant. Remember this info so that we
824 // can create a PHI node.
825 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
828 // If the loaded value isn't available in any predecessor, it isn't partially
830 if (AvailablePreds.empty()) return false;
832 // Okay, the loaded value is available in at least one (and maybe all!)
833 // predecessors. If the value is unavailable in more than one unique
834 // predecessor, we want to insert a merge block for those common predecessors.
835 // This ensures that we only have to insert one reload, thus not increasing
837 BasicBlock *UnavailablePred = 0;
839 // If there is exactly one predecessor where the value is unavailable, the
840 // already computed 'OneUnavailablePred' block is it. If it ends in an
841 // unconditional branch, we know that it isn't a critical edge.
842 if (PredsScanned.size() == AvailablePreds.size()+1 &&
843 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
844 UnavailablePred = OneUnavailablePred;
845 } else if (PredsScanned.size() != AvailablePreds.size()) {
846 // Otherwise, we had multiple unavailable predecessors or we had a critical
847 // edge from the one.
848 SmallVector<BasicBlock*, 8> PredsToSplit;
849 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
851 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
852 AvailablePredSet.insert(AvailablePreds[i].first);
854 // Add all the unavailable predecessors to the PredsToSplit list.
855 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
858 // If the predecessor is an indirect goto, we can't split the edge.
859 if (isa<IndirectBrInst>(P->getTerminator()))
862 if (!AvailablePredSet.count(P))
863 PredsToSplit.push_back(P);
866 // Split them out to their own block.
868 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
869 "thread-pre-split", this);
872 // If the value isn't available in all predecessors, then there will be
873 // exactly one where it isn't available. Insert a load on that edge and add
874 // it to the AvailablePreds list.
875 if (UnavailablePred) {
876 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
877 "Can't handle critical edge here!");
878 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
880 UnavailablePred->getTerminator());
881 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
884 // Now we know that each predecessor of this block has a value in
885 // AvailablePreds, sort them for efficient access as we're walking the preds.
886 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
888 // Create a PHI node at the start of the block for the PRE'd load value.
889 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
892 // Insert new entries into the PHI for each predecessor. A single block may
893 // have multiple entries here.
894 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
897 AvailablePredsTy::iterator I =
898 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
899 std::make_pair(P, (Value*)0));
901 assert(I != AvailablePreds.end() && I->first == P &&
902 "Didn't find entry for predecessor!");
904 PN->addIncoming(I->second, I->first);
907 //cerr << "PRE: " << *LI << *PN << "\n";
909 LI->replaceAllUsesWith(PN);
910 LI->eraseFromParent();
915 /// FindMostPopularDest - The specified list contains multiple possible
916 /// threadable destinations. Pick the one that occurs the most frequently in
919 FindMostPopularDest(BasicBlock *BB,
920 const SmallVectorImpl<std::pair<BasicBlock*,
921 BasicBlock*> > &PredToDestList) {
922 assert(!PredToDestList.empty());
924 // Determine popularity. If there are multiple possible destinations, we
925 // explicitly choose to ignore 'undef' destinations. We prefer to thread
926 // blocks with known and real destinations to threading undef. We'll handle
927 // them later if interesting.
928 DenseMap<BasicBlock*, unsigned> DestPopularity;
929 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
930 if (PredToDestList[i].second)
931 DestPopularity[PredToDestList[i].second]++;
933 // Find the most popular dest.
934 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
935 BasicBlock *MostPopularDest = DPI->first;
936 unsigned Popularity = DPI->second;
937 SmallVector<BasicBlock*, 4> SamePopularity;
939 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
940 // If the popularity of this entry isn't higher than the popularity we've
941 // seen so far, ignore it.
942 if (DPI->second < Popularity)
944 else if (DPI->second == Popularity) {
945 // If it is the same as what we've seen so far, keep track of it.
946 SamePopularity.push_back(DPI->first);
948 // If it is more popular, remember it.
949 SamePopularity.clear();
950 MostPopularDest = DPI->first;
951 Popularity = DPI->second;
955 // Okay, now we know the most popular destination. If there is more than
956 // destination, we need to determine one. This is arbitrary, but we need
957 // to make a deterministic decision. Pick the first one that appears in the
959 if (!SamePopularity.empty()) {
960 SamePopularity.push_back(MostPopularDest);
961 TerminatorInst *TI = BB->getTerminator();
962 for (unsigned i = 0; ; ++i) {
963 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
965 if (std::find(SamePopularity.begin(), SamePopularity.end(),
966 TI->getSuccessor(i)) == SamePopularity.end())
969 MostPopularDest = TI->getSuccessor(i);
974 // Okay, we have finally picked the most popular destination.
975 return MostPopularDest;
978 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
979 ConstantPreference Preference) {
980 // If threading this would thread across a loop header, don't even try to
982 if (LoopHeaders.count(BB))
985 PredValueInfoTy PredValues;
986 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference))
989 assert(!PredValues.empty() &&
990 "ComputeValueKnownInPredecessors returned true with no values");
992 DEBUG(dbgs() << "IN BB: " << *BB;
993 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
994 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
995 << *PredValues[i].first
996 << " for pred '" << PredValues[i].second->getName() << "'.\n";
999 // Decide what we want to thread through. Convert our list of known values to
1000 // a list of known destinations for each pred. This also discards duplicate
1001 // predecessors and keeps track of the undefined inputs (which are represented
1002 // as a null dest in the PredToDestList).
1003 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1004 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1006 BasicBlock *OnlyDest = 0;
1007 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1009 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1010 BasicBlock *Pred = PredValues[i].second;
1011 if (!SeenPreds.insert(Pred))
1012 continue; // Duplicate predecessor entry.
1014 // If the predecessor ends with an indirect goto, we can't change its
1016 if (isa<IndirectBrInst>(Pred->getTerminator()))
1019 Constant *Val = PredValues[i].first;
1022 if (isa<UndefValue>(Val))
1024 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1025 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1026 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
1027 DestBB = SI->getSuccessor(SI->findCaseValue(cast<ConstantInt>(Val)));
1029 assert(isa<IndirectBrInst>(BB->getTerminator())
1030 && "Unexpected terminator");
1031 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1034 // If we have exactly one destination, remember it for efficiency below.
1037 else if (OnlyDest != DestBB)
1038 OnlyDest = MultipleDestSentinel;
1040 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1043 // If all edges were unthreadable, we fail.
1044 if (PredToDestList.empty())
1047 // Determine which is the most common successor. If we have many inputs and
1048 // this block is a switch, we want to start by threading the batch that goes
1049 // to the most popular destination first. If we only know about one
1050 // threadable destination (the common case) we can avoid this.
1051 BasicBlock *MostPopularDest = OnlyDest;
1053 if (MostPopularDest == MultipleDestSentinel)
1054 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1056 // Now that we know what the most popular destination is, factor all
1057 // predecessors that will jump to it into a single predecessor.
1058 SmallVector<BasicBlock*, 16> PredsToFactor;
1059 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1060 if (PredToDestList[i].second == MostPopularDest) {
1061 BasicBlock *Pred = PredToDestList[i].first;
1063 // This predecessor may be a switch or something else that has multiple
1064 // edges to the block. Factor each of these edges by listing them
1065 // according to # occurrences in PredsToFactor.
1066 TerminatorInst *PredTI = Pred->getTerminator();
1067 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1068 if (PredTI->getSuccessor(i) == BB)
1069 PredsToFactor.push_back(Pred);
1072 // If the threadable edges are branching on an undefined value, we get to pick
1073 // the destination that these predecessors should get to.
1074 if (MostPopularDest == 0)
1075 MostPopularDest = BB->getTerminator()->
1076 getSuccessor(GetBestDestForJumpOnUndef(BB));
1078 // Ok, try to thread it!
1079 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1082 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1083 /// a PHI node in the current block. See if there are any simplifications we
1084 /// can do based on inputs to the phi node.
1086 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1087 BasicBlock *BB = PN->getParent();
1089 // TODO: We could make use of this to do it once for blocks with common PHI
1091 SmallVector<BasicBlock*, 1> PredBBs;
1094 // If any of the predecessor blocks end in an unconditional branch, we can
1095 // *duplicate* the conditional branch into that block in order to further
1096 // encourage jump threading and to eliminate cases where we have branch on a
1097 // phi of an icmp (branch on icmp is much better).
1098 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1099 BasicBlock *PredBB = PN->getIncomingBlock(i);
1100 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1101 if (PredBr->isUnconditional()) {
1102 PredBBs[0] = PredBB;
1103 // Try to duplicate BB into PredBB.
1104 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1112 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1113 /// a xor instruction in the current block. See if there are any
1114 /// simplifications we can do based on inputs to the xor.
1116 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1117 BasicBlock *BB = BO->getParent();
1119 // If either the LHS or RHS of the xor is a constant, don't do this
1121 if (isa<ConstantInt>(BO->getOperand(0)) ||
1122 isa<ConstantInt>(BO->getOperand(1)))
1125 // If the first instruction in BB isn't a phi, we won't be able to infer
1126 // anything special about any particular predecessor.
1127 if (!isa<PHINode>(BB->front()))
1130 // If we have a xor as the branch input to this block, and we know that the
1131 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1132 // the condition into the predecessor and fix that value to true, saving some
1133 // logical ops on that path and encouraging other paths to simplify.
1135 // This copies something like this:
1138 // %X = phi i1 [1], [%X']
1139 // %Y = icmp eq i32 %A, %B
1140 // %Z = xor i1 %X, %Y
1145 // %Y = icmp ne i32 %A, %B
1148 PredValueInfoTy XorOpValues;
1150 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1152 assert(XorOpValues.empty());
1153 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1159 assert(!XorOpValues.empty() &&
1160 "ComputeValueKnownInPredecessors returned true with no values");
1162 // Scan the information to see which is most popular: true or false. The
1163 // predecessors can be of the set true, false, or undef.
1164 unsigned NumTrue = 0, NumFalse = 0;
1165 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1166 if (isa<UndefValue>(XorOpValues[i].first))
1167 // Ignore undefs for the count.
1169 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1175 // Determine which value to split on, true, false, or undef if neither.
1176 ConstantInt *SplitVal = 0;
1177 if (NumTrue > NumFalse)
1178 SplitVal = ConstantInt::getTrue(BB->getContext());
1179 else if (NumTrue != 0 || NumFalse != 0)
1180 SplitVal = ConstantInt::getFalse(BB->getContext());
1182 // Collect all of the blocks that this can be folded into so that we can
1183 // factor this once and clone it once.
1184 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1185 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1186 if (XorOpValues[i].first != SplitVal &&
1187 !isa<UndefValue>(XorOpValues[i].first))
1190 BlocksToFoldInto.push_back(XorOpValues[i].second);
1193 // If we inferred a value for all of the predecessors, then duplication won't
1194 // help us. However, we can just replace the LHS or RHS with the constant.
1195 if (BlocksToFoldInto.size() ==
1196 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1197 if (SplitVal == 0) {
1198 // If all preds provide undef, just nuke the xor, because it is undef too.
1199 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1200 BO->eraseFromParent();
1201 } else if (SplitVal->isZero()) {
1202 // If all preds provide 0, replace the xor with the other input.
1203 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1204 BO->eraseFromParent();
1206 // If all preds provide 1, set the computed value to 1.
1207 BO->setOperand(!isLHS, SplitVal);
1213 // Try to duplicate BB into PredBB.
1214 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1218 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1219 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1220 /// NewPred using the entries from OldPred (suitably mapped).
1221 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1222 BasicBlock *OldPred,
1223 BasicBlock *NewPred,
1224 DenseMap<Instruction*, Value*> &ValueMap) {
1225 for (BasicBlock::iterator PNI = PHIBB->begin();
1226 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1227 // Ok, we have a PHI node. Figure out what the incoming value was for the
1229 Value *IV = PN->getIncomingValueForBlock(OldPred);
1231 // Remap the value if necessary.
1232 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1233 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1234 if (I != ValueMap.end())
1238 PN->addIncoming(IV, NewPred);
1242 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1243 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1244 /// across BB. Transform the IR to reflect this change.
1245 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1246 const SmallVectorImpl<BasicBlock*> &PredBBs,
1247 BasicBlock *SuccBB) {
1248 // If threading to the same block as we come from, we would infinite loop.
1250 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1251 << "' - would thread to self!\n");
1255 // If threading this would thread across a loop header, don't thread the edge.
1256 // See the comments above FindLoopHeaders for justifications and caveats.
1257 if (LoopHeaders.count(BB)) {
1258 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1259 << "' to dest BB '" << SuccBB->getName()
1260 << "' - it might create an irreducible loop!\n");
1264 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1265 if (JumpThreadCost > Threshold) {
1266 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1267 << "' - Cost is too high: " << JumpThreadCost << "\n");
1271 // And finally, do it! Start by factoring the predecessors is needed.
1273 if (PredBBs.size() == 1)
1274 PredBB = PredBBs[0];
1276 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1277 << " common predecessors.\n");
1278 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1282 // And finally, do it!
1283 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1284 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1285 << ", across block:\n "
1288 LVI->threadEdge(PredBB, BB, SuccBB);
1290 // We are going to have to map operands from the original BB block to the new
1291 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1292 // account for entry from PredBB.
1293 DenseMap<Instruction*, Value*> ValueMapping;
1295 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1296 BB->getName()+".thread",
1297 BB->getParent(), BB);
1298 NewBB->moveAfter(PredBB);
1300 BasicBlock::iterator BI = BB->begin();
1301 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1302 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1304 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1305 // mapping and using it to remap operands in the cloned instructions.
1306 for (; !isa<TerminatorInst>(BI); ++BI) {
1307 Instruction *New = BI->clone();
1308 New->setName(BI->getName());
1309 NewBB->getInstList().push_back(New);
1310 ValueMapping[BI] = New;
1312 // Remap operands to patch up intra-block references.
1313 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1314 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1315 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1316 if (I != ValueMapping.end())
1317 New->setOperand(i, I->second);
1321 // We didn't copy the terminator from BB over to NewBB, because there is now
1322 // an unconditional jump to SuccBB. Insert the unconditional jump.
1323 BranchInst::Create(SuccBB, NewBB);
1325 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1326 // PHI nodes for NewBB now.
1327 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1329 // If there were values defined in BB that are used outside the block, then we
1330 // now have to update all uses of the value to use either the original value,
1331 // the cloned value, or some PHI derived value. This can require arbitrary
1332 // PHI insertion, of which we are prepared to do, clean these up now.
1333 SSAUpdater SSAUpdate;
1334 SmallVector<Use*, 16> UsesToRename;
1335 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1336 // Scan all uses of this instruction to see if it is used outside of its
1337 // block, and if so, record them in UsesToRename.
1338 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1340 Instruction *User = cast<Instruction>(*UI);
1341 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1342 if (UserPN->getIncomingBlock(UI) == BB)
1344 } else if (User->getParent() == BB)
1347 UsesToRename.push_back(&UI.getUse());
1350 // If there are no uses outside the block, we're done with this instruction.
1351 if (UsesToRename.empty())
1354 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1356 // We found a use of I outside of BB. Rename all uses of I that are outside
1357 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1358 // with the two values we know.
1359 SSAUpdate.Initialize(I->getType(), I->getName());
1360 SSAUpdate.AddAvailableValue(BB, I);
1361 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1363 while (!UsesToRename.empty())
1364 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1365 DEBUG(dbgs() << "\n");
1369 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1370 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1371 // us to simplify any PHI nodes in BB.
1372 TerminatorInst *PredTerm = PredBB->getTerminator();
1373 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1374 if (PredTerm->getSuccessor(i) == BB) {
1375 BB->removePredecessor(PredBB, true);
1376 PredTerm->setSuccessor(i, NewBB);
1379 // At this point, the IR is fully up to date and consistent. Do a quick scan
1380 // over the new instructions and zap any that are constants or dead. This
1381 // frequently happens because of phi translation.
1382 SimplifyInstructionsInBlock(NewBB, TD);
1384 // Threaded an edge!
1389 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1390 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1391 /// If we can duplicate the contents of BB up into PredBB do so now, this
1392 /// improves the odds that the branch will be on an analyzable instruction like
1394 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1395 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1396 assert(!PredBBs.empty() && "Can't handle an empty set");
1398 // If BB is a loop header, then duplicating this block outside the loop would
1399 // cause us to transform this into an irreducible loop, don't do this.
1400 // See the comments above FindLoopHeaders for justifications and caveats.
1401 if (LoopHeaders.count(BB)) {
1402 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1403 << "' into predecessor block '" << PredBBs[0]->getName()
1404 << "' - it might create an irreducible loop!\n");
1408 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1409 if (DuplicationCost > Threshold) {
1410 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1411 << "' - Cost is too high: " << DuplicationCost << "\n");
1415 // And finally, do it! Start by factoring the predecessors is needed.
1417 if (PredBBs.size() == 1)
1418 PredBB = PredBBs[0];
1420 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1421 << " common predecessors.\n");
1422 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1426 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1428 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1429 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1430 << DuplicationCost << " block is:" << *BB << "\n");
1432 // Unless PredBB ends with an unconditional branch, split the edge so that we
1433 // can just clone the bits from BB into the end of the new PredBB.
1434 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1436 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1437 PredBB = SplitEdge(PredBB, BB, this);
1438 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1441 // We are going to have to map operands from the original BB block into the
1442 // PredBB block. Evaluate PHI nodes in BB.
1443 DenseMap<Instruction*, Value*> ValueMapping;
1445 BasicBlock::iterator BI = BB->begin();
1446 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1447 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1449 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1450 // mapping and using it to remap operands in the cloned instructions.
1451 for (; BI != BB->end(); ++BI) {
1452 Instruction *New = BI->clone();
1454 // Remap operands to patch up intra-block references.
1455 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1456 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1457 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1458 if (I != ValueMapping.end())
1459 New->setOperand(i, I->second);
1462 // If this instruction can be simplified after the operands are updated,
1463 // just use the simplified value instead. This frequently happens due to
1465 if (Value *IV = SimplifyInstruction(New, TD)) {
1467 ValueMapping[BI] = IV;
1469 // Otherwise, insert the new instruction into the block.
1470 New->setName(BI->getName());
1471 PredBB->getInstList().insert(OldPredBranch, New);
1472 ValueMapping[BI] = New;
1476 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1477 // add entries to the PHI nodes for branch from PredBB now.
1478 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1479 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1481 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1484 // If there were values defined in BB that are used outside the block, then we
1485 // now have to update all uses of the value to use either the original value,
1486 // the cloned value, or some PHI derived value. This can require arbitrary
1487 // PHI insertion, of which we are prepared to do, clean these up now.
1488 SSAUpdater SSAUpdate;
1489 SmallVector<Use*, 16> UsesToRename;
1490 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1491 // Scan all uses of this instruction to see if it is used outside of its
1492 // block, and if so, record them in UsesToRename.
1493 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1495 Instruction *User = cast<Instruction>(*UI);
1496 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1497 if (UserPN->getIncomingBlock(UI) == BB)
1499 } else if (User->getParent() == BB)
1502 UsesToRename.push_back(&UI.getUse());
1505 // If there are no uses outside the block, we're done with this instruction.
1506 if (UsesToRename.empty())
1509 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1511 // We found a use of I outside of BB. Rename all uses of I that are outside
1512 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1513 // with the two values we know.
1514 SSAUpdate.Initialize(I->getType(), I->getName());
1515 SSAUpdate.AddAvailableValue(BB, I);
1516 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1518 while (!UsesToRename.empty())
1519 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1520 DEBUG(dbgs() << "\n");
1523 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1525 BB->removePredecessor(PredBB, true);
1527 // Remove the unconditional branch at the end of the PredBB block.
1528 OldPredBranch->eraseFromParent();