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);
174 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
176 // Can't thread an unconditional jump, but if the block is "almost
177 // empty", we can replace uses of it with uses of the successor and make
179 if (BI && BI->isUnconditional() &&
180 BB != &BB->getParent()->getEntryBlock() &&
181 // If the terminator is the only non-phi instruction, try to nuke it.
182 BB->getFirstNonPHIOrDbg()->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);
205 EverChanged |= Changed;
212 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
213 /// thread across it.
214 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
215 /// Ignore PHI nodes, these will be flattened when duplication happens.
216 BasicBlock::const_iterator I = BB->getFirstNonPHI();
218 // FIXME: THREADING will delete values that are just used to compute the
219 // branch, so they shouldn't count against the duplication cost.
222 // Sum up the cost of each instruction until we get to the terminator. Don't
223 // include the terminator because the copy won't include it.
225 for (; !isa<TerminatorInst>(I); ++I) {
226 // Debugger intrinsics don't incur code size.
227 if (isa<DbgInfoIntrinsic>(I)) continue;
229 // If this is a pointer->pointer bitcast, it is free.
230 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
233 // All other instructions count for at least one unit.
236 // Calls are more expensive. If they are non-intrinsic calls, we model them
237 // as having cost of 4. If they are a non-vector intrinsic, we model them
238 // as having cost of 2 total, and if they are a vector intrinsic, we model
239 // them as having cost 1.
240 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
241 if (!isa<IntrinsicInst>(CI))
243 else if (!CI->getType()->isVectorTy())
248 // Threading through a switch statement is particularly profitable. If this
249 // block ends in a switch, decrease its cost to make it more likely to happen.
250 if (isa<SwitchInst>(I))
251 Size = Size > 6 ? Size-6 : 0;
253 // The same holds for indirect branches, but slightly more so.
254 if (isa<IndirectBrInst>(I))
255 Size = Size > 8 ? Size-8 : 0;
260 /// FindLoopHeaders - We do not want jump threading to turn proper loop
261 /// structures into irreducible loops. Doing this breaks up the loop nesting
262 /// hierarchy and pessimizes later transformations. To prevent this from
263 /// happening, we first have to find the loop headers. Here we approximate this
264 /// by finding targets of backedges in the CFG.
266 /// Note that there definitely are cases when we want to allow threading of
267 /// edges across a loop header. For example, threading a jump from outside the
268 /// loop (the preheader) to an exit block of the loop is definitely profitable.
269 /// It is also almost always profitable to thread backedges from within the loop
270 /// to exit blocks, and is often profitable to thread backedges to other blocks
271 /// within the loop (forming a nested loop). This simple analysis is not rich
272 /// enough to track all of these properties and keep it up-to-date as the CFG
273 /// mutates, so we don't allow any of these transformations.
275 void JumpThreading::FindLoopHeaders(Function &F) {
276 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
277 FindFunctionBackedges(F, Edges);
279 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
280 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
283 /// getKnownConstant - Helper method to determine if we can thread over a
284 /// terminator with the given value as its condition, and if so what value to
285 /// use for that. What kind of value this is depends on whether we want an
286 /// integer or a block address, but an undef is always accepted.
287 /// Returns null if Val is null or not an appropriate constant.
288 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
292 // Undef is "known" enough.
293 if (UndefValue *U = dyn_cast<UndefValue>(Val))
296 if (Preference == WantBlockAddress)
297 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
299 return dyn_cast<ConstantInt>(Val);
302 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
303 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
304 /// in any of our predecessors. If so, return the known list of value and pred
305 /// BB in the result vector.
307 /// This returns true if there were any known values.
310 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
311 ConstantPreference Preference) {
312 // This method walks up use-def chains recursively. Because of this, we could
313 // get into an infinite loop going around loops in the use-def chain. To
314 // prevent this, keep track of what (value, block) pairs we've already visited
315 // and terminate the search if we loop back to them
316 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
319 // An RAII help to remove this pair from the recursion set once the recursion
320 // stack pops back out again.
321 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
323 // If V is a constant, then it is known in all predecessors.
324 if (Constant *KC = getKnownConstant(V, Preference)) {
325 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
326 Result.push_back(std::make_pair(KC, *PI));
331 // If V is a non-instruction value, or an instruction in a different block,
332 // then it can't be derived from a PHI.
333 Instruction *I = dyn_cast<Instruction>(V);
334 if (I == 0 || I->getParent() != BB) {
336 // Okay, if this is a live-in value, see if it has a known value at the end
337 // of any of our predecessors.
339 // FIXME: This should be an edge property, not a block end property.
340 /// TODO: Per PR2563, we could infer value range information about a
341 /// predecessor based on its terminator.
343 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
344 // "I" is a non-local compare-with-a-constant instruction. This would be
345 // able to handle value inequalities better, for example if the compare is
346 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
347 // Perhaps getConstantOnEdge should be smart enough to do this?
349 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
351 // If the value is known by LazyValueInfo to be a constant in a
352 // predecessor, use that information to try to thread this block.
353 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
354 if (Constant *KC = getKnownConstant(PredCst, Preference))
355 Result.push_back(std::make_pair(KC, P));
358 return !Result.empty();
361 /// If I is a PHI node, then we know the incoming values for any constants.
362 if (PHINode *PN = dyn_cast<PHINode>(I)) {
363 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
364 Value *InVal = PN->getIncomingValue(i);
365 if (Constant *KC = getKnownConstant(InVal, Preference)) {
366 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
368 Constant *CI = LVI->getConstantOnEdge(InVal,
369 PN->getIncomingBlock(i), BB);
370 if (Constant *KC = getKnownConstant(CI, Preference))
371 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
375 return !Result.empty();
378 PredValueInfoTy LHSVals, RHSVals;
380 // Handle some boolean conditions.
381 if (I->getType()->getPrimitiveSizeInBits() == 1) {
382 assert(Preference == WantInteger && "One-bit non-integer type?");
384 // X & false -> false
385 if (I->getOpcode() == Instruction::Or ||
386 I->getOpcode() == Instruction::And) {
387 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
389 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
392 if (LHSVals.empty() && RHSVals.empty())
395 ConstantInt *InterestingVal;
396 if (I->getOpcode() == Instruction::Or)
397 InterestingVal = ConstantInt::getTrue(I->getContext());
399 InterestingVal = ConstantInt::getFalse(I->getContext());
401 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
403 // Scan for the sentinel. If we find an undef, force it to the
404 // interesting value: x|undef -> true and x&undef -> false.
405 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
406 if (LHSVals[i].first == InterestingVal ||
407 isa<UndefValue>(LHSVals[i].first)) {
408 Result.push_back(LHSVals[i]);
409 Result.back().first = InterestingVal;
410 LHSKnownBBs.insert(LHSVals[i].second);
412 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
413 if (RHSVals[i].first == InterestingVal ||
414 isa<UndefValue>(RHSVals[i].first)) {
415 // If we already inferred a value for this block on the LHS, don't
417 if (!LHSKnownBBs.count(RHSVals[i].second)) {
418 Result.push_back(RHSVals[i]);
419 Result.back().first = InterestingVal;
423 return !Result.empty();
426 // Handle the NOT form of XOR.
427 if (I->getOpcode() == Instruction::Xor &&
428 isa<ConstantInt>(I->getOperand(1)) &&
429 cast<ConstantInt>(I->getOperand(1))->isOne()) {
430 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
435 // Invert the known values.
436 for (unsigned i = 0, e = Result.size(); i != e; ++i)
437 Result[i].first = ConstantExpr::getNot(Result[i].first);
442 // Try to simplify some other binary operator values.
443 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
444 assert(Preference != WantBlockAddress
445 && "A binary operator creating a block address?");
446 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
447 PredValueInfoTy LHSVals;
448 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
451 // Try to use constant folding to simplify the binary operator.
452 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
453 Constant *V = LHSVals[i].first;
454 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
456 if (Constant *KC = getKnownConstant(Folded, WantInteger))
457 Result.push_back(std::make_pair(KC, LHSVals[i].second));
461 return !Result.empty();
464 // Handle compare with phi operand, where the PHI is defined in this block.
465 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
466 assert(Preference == WantInteger && "Compares only produce integers");
467 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
468 if (PN && PN->getParent() == BB) {
469 // We can do this simplification if any comparisons fold to true or false.
471 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
472 BasicBlock *PredBB = PN->getIncomingBlock(i);
473 Value *LHS = PN->getIncomingValue(i);
474 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
476 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
478 if (!isa<Constant>(RHS))
481 LazyValueInfo::Tristate
482 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
483 cast<Constant>(RHS), PredBB, BB);
484 if (ResT == LazyValueInfo::Unknown)
486 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
489 if (Constant *KC = getKnownConstant(Res, WantInteger))
490 Result.push_back(std::make_pair(KC, PredBB));
493 return !Result.empty();
497 // If comparing a live-in value against a constant, see if we know the
498 // live-in value on any predecessors.
499 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
500 if (!isa<Instruction>(Cmp->getOperand(0)) ||
501 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
502 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
504 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
506 // If the value is known by LazyValueInfo to be a constant in a
507 // predecessor, use that information to try to thread this block.
508 LazyValueInfo::Tristate Res =
509 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
511 if (Res == LazyValueInfo::Unknown)
514 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
515 Result.push_back(std::make_pair(ResC, P));
518 return !Result.empty();
521 // Try to find a constant value for the LHS of a comparison,
522 // and evaluate it statically if we can.
523 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
524 PredValueInfoTy LHSVals;
525 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
528 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
529 Constant *V = LHSVals[i].first;
530 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
532 if (Constant *KC = getKnownConstant(Folded, WantInteger))
533 Result.push_back(std::make_pair(KC, LHSVals[i].second));
536 return !Result.empty();
541 // If all else fails, see if LVI can figure out a constant value for us.
542 Constant *CI = LVI->getConstant(V, BB);
543 if (Constant *KC = getKnownConstant(CI, Preference)) {
544 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
545 Result.push_back(std::make_pair(KC, *PI));
548 return !Result.empty();
553 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
554 /// in an undefined jump, decide which block is best to revector to.
556 /// Since we can pick an arbitrary destination, we pick the successor with the
557 /// fewest predecessors. This should reduce the in-degree of the others.
559 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
560 TerminatorInst *BBTerm = BB->getTerminator();
561 unsigned MinSucc = 0;
562 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
563 // Compute the successor with the minimum number of predecessors.
564 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
565 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
566 TestBB = BBTerm->getSuccessor(i);
567 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
568 if (NumPreds < MinNumPreds)
575 /// ProcessBlock - If there are any predecessors whose control can be threaded
576 /// through to a successor, transform them now.
577 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
578 // If the block is trivially dead, just return and let the caller nuke it.
579 // This simplifies other transformations.
580 if (pred_begin(BB) == pred_end(BB) &&
581 BB != &BB->getParent()->getEntryBlock())
584 // If this block has a single predecessor, and if that pred has a single
585 // successor, merge the blocks. This encourages recursive jump threading
586 // because now the condition in this block can be threaded through
587 // predecessors of our predecessor block.
588 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
589 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
591 // If SinglePred was a loop header, BB becomes one.
592 if (LoopHeaders.erase(SinglePred))
593 LoopHeaders.insert(BB);
595 // Remember if SinglePred was the entry block of the function. If so, we
596 // will need to move BB back to the entry position.
597 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
598 LVI->eraseBlock(SinglePred);
599 MergeBasicBlockIntoOnlyPred(BB);
601 if (isEntry && BB != &BB->getParent()->getEntryBlock())
602 BB->moveBefore(&BB->getParent()->getEntryBlock());
607 // What kind of constant we're looking for.
608 ConstantPreference Preference = WantInteger;
610 // Look to see if the terminator is a conditional branch, switch or indirect
611 // branch, if not we can't thread it.
613 Instruction *Terminator = BB->getTerminator();
614 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
615 // Can't thread an unconditional jump.
616 if (BI->isUnconditional()) return false;
617 Condition = BI->getCondition();
618 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
619 Condition = SI->getCondition();
620 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
621 Condition = IB->getAddress()->stripPointerCasts();
622 Preference = WantBlockAddress;
624 return false; // Must be an invoke.
627 // If the terminator is branching on an undef, we can pick any of the
628 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
629 if (isa<UndefValue>(Condition)) {
630 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
632 // Fold the branch/switch.
633 TerminatorInst *BBTerm = BB->getTerminator();
634 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
635 if (i == BestSucc) continue;
636 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
639 DEBUG(dbgs() << " In block '" << BB->getName()
640 << "' folding undef terminator: " << *BBTerm << '\n');
641 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
642 BBTerm->eraseFromParent();
646 // If the terminator of this block is branching on a constant, simplify the
647 // terminator to an unconditional branch. This can occur due to threading in
649 if (getKnownConstant(Condition, Preference)) {
650 DEBUG(dbgs() << " In block '" << BB->getName()
651 << "' folding terminator: " << *BB->getTerminator() << '\n');
653 ConstantFoldTerminator(BB);
657 Instruction *CondInst = dyn_cast<Instruction>(Condition);
659 // All the rest of our checks depend on the condition being an instruction.
661 // FIXME: Unify this with code below.
662 if (ProcessThreadableEdges(Condition, BB, Preference))
668 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
669 // For a comparison where the LHS is outside this block, it's possible
670 // that we've branched on it before. Used LVI to see if we can simplify
671 // the branch based on that.
672 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
673 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
674 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
675 if (CondBr && CondConst && CondBr->isConditional() && PI != PE &&
676 (!isa<Instruction>(CondCmp->getOperand(0)) ||
677 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
678 // For predecessor edge, determine if the comparison is true or false
679 // on that edge. If they're all true or all false, we can simplify the
681 // FIXME: We could handle mixed true/false by duplicating code.
682 LazyValueInfo::Tristate Baseline =
683 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
685 if (Baseline != LazyValueInfo::Unknown) {
686 // Check that all remaining incoming values match the first one.
688 LazyValueInfo::Tristate Ret =
689 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
690 CondCmp->getOperand(0), CondConst, *PI, BB);
691 if (Ret != Baseline) break;
694 // If we terminated early, then one of the values didn't match.
696 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
697 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
698 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
699 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
700 CondBr->eraseFromParent();
707 // Check for some cases that are worth simplifying. Right now we want to look
708 // for loads that are used by a switch or by the condition for the branch. If
709 // we see one, check to see if it's partially redundant. If so, insert a PHI
710 // which can then be used to thread the values.
712 Value *SimplifyValue = CondInst;
713 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
714 if (isa<Constant>(CondCmp->getOperand(1)))
715 SimplifyValue = CondCmp->getOperand(0);
717 // TODO: There are other places where load PRE would be profitable, such as
718 // more complex comparisons.
719 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
720 if (SimplifyPartiallyRedundantLoad(LI))
724 // Handle a variety of cases where we are branching on something derived from
725 // a PHI node in the current block. If we can prove that any predecessors
726 // compute a predictable value based on a PHI node, thread those predecessors.
728 if (ProcessThreadableEdges(CondInst, BB, Preference))
731 // If this is an otherwise-unfoldable branch on a phi node in the current
732 // block, see if we can simplify.
733 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
734 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
735 return ProcessBranchOnPHI(PN);
738 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
739 if (CondInst->getOpcode() == Instruction::Xor &&
740 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
741 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
744 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
745 // "(X == 4)", thread through this block.
751 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
752 /// load instruction, eliminate it by replacing it with a PHI node. This is an
753 /// important optimization that encourages jump threading, and needs to be run
754 /// interlaced with other jump threading tasks.
755 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
756 // Don't hack volatile loads.
757 if (LI->isVolatile()) return false;
759 // If the load is defined in a block with exactly one predecessor, it can't be
760 // partially redundant.
761 BasicBlock *LoadBB = LI->getParent();
762 if (LoadBB->getSinglePredecessor())
765 Value *LoadedPtr = LI->getOperand(0);
767 // If the loaded operand is defined in the LoadBB, it can't be available.
768 // TODO: Could do simple PHI translation, that would be fun :)
769 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
770 if (PtrOp->getParent() == LoadBB)
773 // Scan a few instructions up from the load, to see if it is obviously live at
774 // the entry to its block.
775 BasicBlock::iterator BBIt = LI;
777 if (Value *AvailableVal =
778 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
779 // If the value if the load is locally available within the block, just use
780 // it. This frequently occurs for reg2mem'd allocas.
781 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
783 // If the returned value is the load itself, replace with an undef. This can
784 // only happen in dead loops.
785 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
786 LI->replaceAllUsesWith(AvailableVal);
787 LI->eraseFromParent();
791 // Otherwise, if we scanned the whole block and got to the top of the block,
792 // we know the block is locally transparent to the load. If not, something
793 // might clobber its value.
794 if (BBIt != LoadBB->begin())
798 SmallPtrSet<BasicBlock*, 8> PredsScanned;
799 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
800 AvailablePredsTy AvailablePreds;
801 BasicBlock *OneUnavailablePred = 0;
803 // If we got here, the loaded value is transparent through to the start of the
804 // block. Check to see if it is available in any of the predecessor blocks.
805 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
807 BasicBlock *PredBB = *PI;
809 // If we already scanned this predecessor, skip it.
810 if (!PredsScanned.insert(PredBB))
813 // Scan the predecessor to see if the value is available in the pred.
814 BBIt = PredBB->end();
815 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
816 if (!PredAvailable) {
817 OneUnavailablePred = PredBB;
821 // If so, this load is partially redundant. Remember this info so that we
822 // can create a PHI node.
823 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
826 // If the loaded value isn't available in any predecessor, it isn't partially
828 if (AvailablePreds.empty()) return false;
830 // Okay, the loaded value is available in at least one (and maybe all!)
831 // predecessors. If the value is unavailable in more than one unique
832 // predecessor, we want to insert a merge block for those common predecessors.
833 // This ensures that we only have to insert one reload, thus not increasing
835 BasicBlock *UnavailablePred = 0;
837 // If there is exactly one predecessor where the value is unavailable, the
838 // already computed 'OneUnavailablePred' block is it. If it ends in an
839 // unconditional branch, we know that it isn't a critical edge.
840 if (PredsScanned.size() == AvailablePreds.size()+1 &&
841 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
842 UnavailablePred = OneUnavailablePred;
843 } else if (PredsScanned.size() != AvailablePreds.size()) {
844 // Otherwise, we had multiple unavailable predecessors or we had a critical
845 // edge from the one.
846 SmallVector<BasicBlock*, 8> PredsToSplit;
847 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
849 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
850 AvailablePredSet.insert(AvailablePreds[i].first);
852 // Add all the unavailable predecessors to the PredsToSplit list.
853 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
856 // If the predecessor is an indirect goto, we can't split the edge.
857 if (isa<IndirectBrInst>(P->getTerminator()))
860 if (!AvailablePredSet.count(P))
861 PredsToSplit.push_back(P);
864 // Split them out to their own block.
866 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
867 "thread-pre-split", this);
870 // If the value isn't available in all predecessors, then there will be
871 // exactly one where it isn't available. Insert a load on that edge and add
872 // it to the AvailablePreds list.
873 if (UnavailablePred) {
874 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
875 "Can't handle critical edge here!");
876 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
878 UnavailablePred->getTerminator());
879 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
882 // Now we know that each predecessor of this block has a value in
883 // AvailablePreds, sort them for efficient access as we're walking the preds.
884 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
886 // Create a PHI node at the start of the block for the PRE'd load value.
887 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
890 // Insert new entries into the PHI for each predecessor. A single block may
891 // have multiple entries here.
892 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
895 AvailablePredsTy::iterator I =
896 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
897 std::make_pair(P, (Value*)0));
899 assert(I != AvailablePreds.end() && I->first == P &&
900 "Didn't find entry for predecessor!");
902 PN->addIncoming(I->second, I->first);
905 //cerr << "PRE: " << *LI << *PN << "\n";
907 LI->replaceAllUsesWith(PN);
908 LI->eraseFromParent();
913 /// FindMostPopularDest - The specified list contains multiple possible
914 /// threadable destinations. Pick the one that occurs the most frequently in
917 FindMostPopularDest(BasicBlock *BB,
918 const SmallVectorImpl<std::pair<BasicBlock*,
919 BasicBlock*> > &PredToDestList) {
920 assert(!PredToDestList.empty());
922 // Determine popularity. If there are multiple possible destinations, we
923 // explicitly choose to ignore 'undef' destinations. We prefer to thread
924 // blocks with known and real destinations to threading undef. We'll handle
925 // them later if interesting.
926 DenseMap<BasicBlock*, unsigned> DestPopularity;
927 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
928 if (PredToDestList[i].second)
929 DestPopularity[PredToDestList[i].second]++;
931 // Find the most popular dest.
932 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
933 BasicBlock *MostPopularDest = DPI->first;
934 unsigned Popularity = DPI->second;
935 SmallVector<BasicBlock*, 4> SamePopularity;
937 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
938 // If the popularity of this entry isn't higher than the popularity we've
939 // seen so far, ignore it.
940 if (DPI->second < Popularity)
942 else if (DPI->second == Popularity) {
943 // If it is the same as what we've seen so far, keep track of it.
944 SamePopularity.push_back(DPI->first);
946 // If it is more popular, remember it.
947 SamePopularity.clear();
948 MostPopularDest = DPI->first;
949 Popularity = DPI->second;
953 // Okay, now we know the most popular destination. If there is more than
954 // destination, we need to determine one. This is arbitrary, but we need
955 // to make a deterministic decision. Pick the first one that appears in the
957 if (!SamePopularity.empty()) {
958 SamePopularity.push_back(MostPopularDest);
959 TerminatorInst *TI = BB->getTerminator();
960 for (unsigned i = 0; ; ++i) {
961 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
963 if (std::find(SamePopularity.begin(), SamePopularity.end(),
964 TI->getSuccessor(i)) == SamePopularity.end())
967 MostPopularDest = TI->getSuccessor(i);
972 // Okay, we have finally picked the most popular destination.
973 return MostPopularDest;
976 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
977 ConstantPreference Preference) {
978 // If threading this would thread across a loop header, don't even try to
980 if (LoopHeaders.count(BB))
983 PredValueInfoTy PredValues;
984 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference))
987 assert(!PredValues.empty() &&
988 "ComputeValueKnownInPredecessors returned true with no values");
990 DEBUG(dbgs() << "IN BB: " << *BB;
991 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
992 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
993 << *PredValues[i].first
994 << " for pred '" << PredValues[i].second->getName() << "'.\n";
997 // Decide what we want to thread through. Convert our list of known values to
998 // a list of known destinations for each pred. This also discards duplicate
999 // predecessors and keeps track of the undefined inputs (which are represented
1000 // as a null dest in the PredToDestList).
1001 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1002 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1004 BasicBlock *OnlyDest = 0;
1005 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1007 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1008 BasicBlock *Pred = PredValues[i].second;
1009 if (!SeenPreds.insert(Pred))
1010 continue; // Duplicate predecessor entry.
1012 // If the predecessor ends with an indirect goto, we can't change its
1014 if (isa<IndirectBrInst>(Pred->getTerminator()))
1017 Constant *Val = PredValues[i].first;
1020 if (isa<UndefValue>(Val))
1022 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1023 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1024 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
1025 DestBB = SI->getSuccessor(SI->findCaseValue(cast<ConstantInt>(Val)));
1027 assert(isa<IndirectBrInst>(BB->getTerminator())
1028 && "Unexpected terminator");
1029 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1032 // If we have exactly one destination, remember it for efficiency below.
1035 else if (OnlyDest != DestBB)
1036 OnlyDest = MultipleDestSentinel;
1038 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1041 // If all edges were unthreadable, we fail.
1042 if (PredToDestList.empty())
1045 // Determine which is the most common successor. If we have many inputs and
1046 // this block is a switch, we want to start by threading the batch that goes
1047 // to the most popular destination first. If we only know about one
1048 // threadable destination (the common case) we can avoid this.
1049 BasicBlock *MostPopularDest = OnlyDest;
1051 if (MostPopularDest == MultipleDestSentinel)
1052 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1054 // Now that we know what the most popular destination is, factor all
1055 // predecessors that will jump to it into a single predecessor.
1056 SmallVector<BasicBlock*, 16> PredsToFactor;
1057 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1058 if (PredToDestList[i].second == MostPopularDest) {
1059 BasicBlock *Pred = PredToDestList[i].first;
1061 // This predecessor may be a switch or something else that has multiple
1062 // edges to the block. Factor each of these edges by listing them
1063 // according to # occurrences in PredsToFactor.
1064 TerminatorInst *PredTI = Pred->getTerminator();
1065 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1066 if (PredTI->getSuccessor(i) == BB)
1067 PredsToFactor.push_back(Pred);
1070 // If the threadable edges are branching on an undefined value, we get to pick
1071 // the destination that these predecessors should get to.
1072 if (MostPopularDest == 0)
1073 MostPopularDest = BB->getTerminator()->
1074 getSuccessor(GetBestDestForJumpOnUndef(BB));
1076 // Ok, try to thread it!
1077 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1080 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1081 /// a PHI node in the current block. See if there are any simplifications we
1082 /// can do based on inputs to the phi node.
1084 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1085 BasicBlock *BB = PN->getParent();
1087 // TODO: We could make use of this to do it once for blocks with common PHI
1089 SmallVector<BasicBlock*, 1> PredBBs;
1092 // If any of the predecessor blocks end in an unconditional branch, we can
1093 // *duplicate* the conditional branch into that block in order to further
1094 // encourage jump threading and to eliminate cases where we have branch on a
1095 // phi of an icmp (branch on icmp is much better).
1096 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1097 BasicBlock *PredBB = PN->getIncomingBlock(i);
1098 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1099 if (PredBr->isUnconditional()) {
1100 PredBBs[0] = PredBB;
1101 // Try to duplicate BB into PredBB.
1102 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1110 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1111 /// a xor instruction in the current block. See if there are any
1112 /// simplifications we can do based on inputs to the xor.
1114 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1115 BasicBlock *BB = BO->getParent();
1117 // If either the LHS or RHS of the xor is a constant, don't do this
1119 if (isa<ConstantInt>(BO->getOperand(0)) ||
1120 isa<ConstantInt>(BO->getOperand(1)))
1123 // If the first instruction in BB isn't a phi, we won't be able to infer
1124 // anything special about any particular predecessor.
1125 if (!isa<PHINode>(BB->front()))
1128 // If we have a xor as the branch input to this block, and we know that the
1129 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1130 // the condition into the predecessor and fix that value to true, saving some
1131 // logical ops on that path and encouraging other paths to simplify.
1133 // This copies something like this:
1136 // %X = phi i1 [1], [%X']
1137 // %Y = icmp eq i32 %A, %B
1138 // %Z = xor i1 %X, %Y
1143 // %Y = icmp ne i32 %A, %B
1146 PredValueInfoTy XorOpValues;
1148 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1150 assert(XorOpValues.empty());
1151 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1157 assert(!XorOpValues.empty() &&
1158 "ComputeValueKnownInPredecessors returned true with no values");
1160 // Scan the information to see which is most popular: true or false. The
1161 // predecessors can be of the set true, false, or undef.
1162 unsigned NumTrue = 0, NumFalse = 0;
1163 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1164 if (isa<UndefValue>(XorOpValues[i].first))
1165 // Ignore undefs for the count.
1167 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1173 // Determine which value to split on, true, false, or undef if neither.
1174 ConstantInt *SplitVal = 0;
1175 if (NumTrue > NumFalse)
1176 SplitVal = ConstantInt::getTrue(BB->getContext());
1177 else if (NumTrue != 0 || NumFalse != 0)
1178 SplitVal = ConstantInt::getFalse(BB->getContext());
1180 // Collect all of the blocks that this can be folded into so that we can
1181 // factor this once and clone it once.
1182 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1183 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1184 if (XorOpValues[i].first != SplitVal &&
1185 !isa<UndefValue>(XorOpValues[i].first))
1188 BlocksToFoldInto.push_back(XorOpValues[i].second);
1191 // If we inferred a value for all of the predecessors, then duplication won't
1192 // help us. However, we can just replace the LHS or RHS with the constant.
1193 if (BlocksToFoldInto.size() ==
1194 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1195 if (SplitVal == 0) {
1196 // If all preds provide undef, just nuke the xor, because it is undef too.
1197 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1198 BO->eraseFromParent();
1199 } else if (SplitVal->isZero()) {
1200 // If all preds provide 0, replace the xor with the other input.
1201 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1202 BO->eraseFromParent();
1204 // If all preds provide 1, set the computed value to 1.
1205 BO->setOperand(!isLHS, SplitVal);
1211 // Try to duplicate BB into PredBB.
1212 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1216 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1217 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1218 /// NewPred using the entries from OldPred (suitably mapped).
1219 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1220 BasicBlock *OldPred,
1221 BasicBlock *NewPred,
1222 DenseMap<Instruction*, Value*> &ValueMap) {
1223 for (BasicBlock::iterator PNI = PHIBB->begin();
1224 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1225 // Ok, we have a PHI node. Figure out what the incoming value was for the
1227 Value *IV = PN->getIncomingValueForBlock(OldPred);
1229 // Remap the value if necessary.
1230 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1231 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1232 if (I != ValueMap.end())
1236 PN->addIncoming(IV, NewPred);
1240 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1241 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1242 /// across BB. Transform the IR to reflect this change.
1243 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1244 const SmallVectorImpl<BasicBlock*> &PredBBs,
1245 BasicBlock *SuccBB) {
1246 // If threading to the same block as we come from, we would infinite loop.
1248 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1249 << "' - would thread to self!\n");
1253 // If threading this would thread across a loop header, don't thread the edge.
1254 // See the comments above FindLoopHeaders for justifications and caveats.
1255 if (LoopHeaders.count(BB)) {
1256 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1257 << "' to dest BB '" << SuccBB->getName()
1258 << "' - it might create an irreducible loop!\n");
1262 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1263 if (JumpThreadCost > Threshold) {
1264 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1265 << "' - Cost is too high: " << JumpThreadCost << "\n");
1269 // And finally, do it! Start by factoring the predecessors is needed.
1271 if (PredBBs.size() == 1)
1272 PredBB = PredBBs[0];
1274 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1275 << " common predecessors.\n");
1276 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1280 // And finally, do it!
1281 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1282 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1283 << ", across block:\n "
1286 LVI->threadEdge(PredBB, BB, SuccBB);
1288 // We are going to have to map operands from the original BB block to the new
1289 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1290 // account for entry from PredBB.
1291 DenseMap<Instruction*, Value*> ValueMapping;
1293 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1294 BB->getName()+".thread",
1295 BB->getParent(), BB);
1296 NewBB->moveAfter(PredBB);
1298 BasicBlock::iterator BI = BB->begin();
1299 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1300 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1302 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1303 // mapping and using it to remap operands in the cloned instructions.
1304 for (; !isa<TerminatorInst>(BI); ++BI) {
1305 Instruction *New = BI->clone();
1306 New->setName(BI->getName());
1307 NewBB->getInstList().push_back(New);
1308 ValueMapping[BI] = New;
1310 // Remap operands to patch up intra-block references.
1311 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1312 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1313 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1314 if (I != ValueMapping.end())
1315 New->setOperand(i, I->second);
1319 // We didn't copy the terminator from BB over to NewBB, because there is now
1320 // an unconditional jump to SuccBB. Insert the unconditional jump.
1321 BranchInst::Create(SuccBB, NewBB);
1323 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1324 // PHI nodes for NewBB now.
1325 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1327 // If there were values defined in BB that are used outside the block, then we
1328 // now have to update all uses of the value to use either the original value,
1329 // the cloned value, or some PHI derived value. This can require arbitrary
1330 // PHI insertion, of which we are prepared to do, clean these up now.
1331 SSAUpdater SSAUpdate;
1332 SmallVector<Use*, 16> UsesToRename;
1333 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1334 // Scan all uses of this instruction to see if it is used outside of its
1335 // block, and if so, record them in UsesToRename.
1336 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1338 Instruction *User = cast<Instruction>(*UI);
1339 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1340 if (UserPN->getIncomingBlock(UI) == BB)
1342 } else if (User->getParent() == BB)
1345 UsesToRename.push_back(&UI.getUse());
1348 // If there are no uses outside the block, we're done with this instruction.
1349 if (UsesToRename.empty())
1352 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1354 // We found a use of I outside of BB. Rename all uses of I that are outside
1355 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1356 // with the two values we know.
1357 SSAUpdate.Initialize(I->getType(), I->getName());
1358 SSAUpdate.AddAvailableValue(BB, I);
1359 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1361 while (!UsesToRename.empty())
1362 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1363 DEBUG(dbgs() << "\n");
1367 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1368 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1369 // us to simplify any PHI nodes in BB.
1370 TerminatorInst *PredTerm = PredBB->getTerminator();
1371 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1372 if (PredTerm->getSuccessor(i) == BB) {
1373 BB->removePredecessor(PredBB, true);
1374 PredTerm->setSuccessor(i, NewBB);
1377 // At this point, the IR is fully up to date and consistent. Do a quick scan
1378 // over the new instructions and zap any that are constants or dead. This
1379 // frequently happens because of phi translation.
1380 SimplifyInstructionsInBlock(NewBB, TD);
1382 // Threaded an edge!
1387 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1388 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1389 /// If we can duplicate the contents of BB up into PredBB do so now, this
1390 /// improves the odds that the branch will be on an analyzable instruction like
1392 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1393 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1394 assert(!PredBBs.empty() && "Can't handle an empty set");
1396 // If BB is a loop header, then duplicating this block outside the loop would
1397 // cause us to transform this into an irreducible loop, don't do this.
1398 // See the comments above FindLoopHeaders for justifications and caveats.
1399 if (LoopHeaders.count(BB)) {
1400 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1401 << "' into predecessor block '" << PredBBs[0]->getName()
1402 << "' - it might create an irreducible loop!\n");
1406 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1407 if (DuplicationCost > Threshold) {
1408 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1409 << "' - Cost is too high: " << DuplicationCost << "\n");
1413 // And finally, do it! Start by factoring the predecessors is needed.
1415 if (PredBBs.size() == 1)
1416 PredBB = PredBBs[0];
1418 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1419 << " common predecessors.\n");
1420 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1424 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1426 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1427 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1428 << DuplicationCost << " block is:" << *BB << "\n");
1430 // Unless PredBB ends with an unconditional branch, split the edge so that we
1431 // can just clone the bits from BB into the end of the new PredBB.
1432 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1434 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1435 PredBB = SplitEdge(PredBB, BB, this);
1436 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1439 // We are going to have to map operands from the original BB block into the
1440 // PredBB block. Evaluate PHI nodes in BB.
1441 DenseMap<Instruction*, Value*> ValueMapping;
1443 BasicBlock::iterator BI = BB->begin();
1444 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1445 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1447 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1448 // mapping and using it to remap operands in the cloned instructions.
1449 for (; BI != BB->end(); ++BI) {
1450 Instruction *New = BI->clone();
1452 // Remap operands to patch up intra-block references.
1453 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1454 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1455 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1456 if (I != ValueMapping.end())
1457 New->setOperand(i, I->second);
1460 // If this instruction can be simplified after the operands are updated,
1461 // just use the simplified value instead. This frequently happens due to
1463 if (Value *IV = SimplifyInstruction(New, TD)) {
1465 ValueMapping[BI] = IV;
1467 // Otherwise, insert the new instruction into the block.
1468 New->setName(BI->getName());
1469 PredBB->getInstList().insert(OldPredBranch, New);
1470 ValueMapping[BI] = New;
1474 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1475 // add entries to the PHI nodes for branch from PredBB now.
1476 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1477 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1479 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1482 // If there were values defined in BB that are used outside the block, then we
1483 // now have to update all uses of the value to use either the original value,
1484 // the cloned value, or some PHI derived value. This can require arbitrary
1485 // PHI insertion, of which we are prepared to do, clean these up now.
1486 SSAUpdater SSAUpdate;
1487 SmallVector<Use*, 16> UsesToRename;
1488 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1489 // Scan all uses of this instruction to see if it is used outside of its
1490 // block, and if so, record them in UsesToRename.
1491 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1493 Instruction *User = cast<Instruction>(*UI);
1494 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1495 if (UserPN->getIncomingBlock(UI) == BB)
1497 } else if (User->getParent() == BB)
1500 UsesToRename.push_back(&UI.getUse());
1503 // If there are no uses outside the block, we're done with this instruction.
1504 if (UsesToRename.empty())
1507 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1509 // We found a use of I outside of BB. Rename all uses of I that are outside
1510 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1511 // with the two values we know.
1512 SSAUpdate.Initialize(I->getType(), I->getName());
1513 SSAUpdate.AddAvailableValue(BB, I);
1514 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1516 while (!UsesToRename.empty())
1517 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1518 DEBUG(dbgs() << "\n");
1521 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1523 BB->removePredecessor(PredBB, true);
1525 // Remove the unconditional branch at the end of the PredBB block.
1526 OldPredBranch->eraseFromParent();