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/ADT/DenseMap.h"
17 #include "llvm/ADT/DenseSet.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallSet.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/CFG.h"
23 #include "llvm/Analysis/ConstantFolding.h"
24 #include "llvm/Analysis/InstructionSimplify.h"
25 #include "llvm/Analysis/LazyValueInfo.h"
26 #include "llvm/Analysis/Loads.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/IntrinsicInst.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/Pass.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ValueHandle.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetLibraryInfo.h"
36 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
37 #include "llvm/Transforms/Utils/Local.h"
38 #include "llvm/Transforms/Utils/SSAUpdater.h"
41 STATISTIC(NumThreads, "Number of jumps threaded");
42 STATISTIC(NumFolds, "Number of terminators folded");
43 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
45 static cl::opt<unsigned>
46 Threshold("jump-threading-threshold",
47 cl::desc("Max block size to duplicate for jump threading"),
48 cl::init(6), cl::Hidden);
51 // These are at global scope so static functions can use them too.
52 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo;
53 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy;
55 // This is used to keep track of what kind of constant we're currently hoping
57 enum ConstantPreference {
62 /// This pass performs 'jump threading', which looks at blocks that have
63 /// multiple predecessors and multiple successors. If one or more of the
64 /// predecessors of the block can be proven to always jump to one of the
65 /// successors, we forward the edge from the predecessor to the successor by
66 /// duplicating the contents of this block.
68 /// An example of when this can occur is code like this:
75 /// In this case, the unconditional branch at the end of the first if can be
76 /// revectored to the false side of the second if.
78 class JumpThreading : public FunctionPass {
80 TargetLibraryInfo *TLI;
83 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
85 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
87 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet;
89 // RAII helper for updating the recursion stack.
90 struct RecursionSetRemover {
91 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet;
92 std::pair<Value*, BasicBlock*> ThePair;
94 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S,
95 std::pair<Value*, BasicBlock*> P)
96 : TheSet(S), ThePair(P) { }
98 ~RecursionSetRemover() {
99 TheSet.erase(ThePair);
103 static char ID; // Pass identification
104 JumpThreading() : FunctionPass(ID) {
105 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
108 bool runOnFunction(Function &F);
110 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
111 AU.addRequired<LazyValueInfo>();
112 AU.addPreserved<LazyValueInfo>();
113 AU.addRequired<TargetLibraryInfo>();
116 void FindLoopHeaders(Function &F);
117 bool ProcessBlock(BasicBlock *BB);
118 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
120 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
121 const SmallVectorImpl<BasicBlock *> &PredBBs);
123 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
124 PredValueInfo &Result,
125 ConstantPreference Preference);
126 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
127 ConstantPreference Preference);
129 bool ProcessBranchOnPHI(PHINode *PN);
130 bool ProcessBranchOnXOR(BinaryOperator *BO);
132 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
133 bool TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB);
137 char JumpThreading::ID = 0;
138 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
139 "Jump Threading", false, false)
140 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)
141 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
142 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
143 "Jump Threading", false, false)
145 // Public interface to the Jump Threading pass
146 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
148 /// runOnFunction - Top level algorithm.
150 bool JumpThreading::runOnFunction(Function &F) {
151 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
152 TD = getAnalysisIfAvailable<DataLayout>();
153 TLI = &getAnalysis<TargetLibraryInfo>();
154 LVI = &getAnalysis<LazyValueInfo>();
158 bool Changed, EverChanged = false;
161 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
163 // Thread all of the branches we can over this block.
164 while (ProcessBlock(BB))
169 // If the block is trivially dead, zap it. This eliminates the successor
170 // edges which simplifies the CFG.
171 if (pred_begin(BB) == pred_end(BB) &&
172 BB != &BB->getParent()->getEntryBlock()) {
173 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
174 << "' with terminator: " << *BB->getTerminator() << '\n');
175 LoopHeaders.erase(BB);
182 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
184 // Can't thread an unconditional jump, but if the block is "almost
185 // empty", we can replace uses of it with uses of the successor and make
187 if (BI && BI->isUnconditional() &&
188 BB != &BB->getParent()->getEntryBlock() &&
189 // If the terminator is the only non-phi instruction, try to nuke it.
190 BB->getFirstNonPHIOrDbg()->isTerminator()) {
191 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
192 // block, we have to make sure it isn't in the LoopHeaders set. We
193 // reinsert afterward if needed.
194 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
195 BasicBlock *Succ = BI->getSuccessor(0);
197 // FIXME: It is always conservatively correct to drop the info
198 // for a block even if it doesn't get erased. This isn't totally
199 // awesome, but it allows us to use AssertingVH to prevent nasty
200 // dangling pointer issues within LazyValueInfo.
202 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
204 // If we deleted BB and BB was the header of a loop, then the
205 // successor is now the header of the loop.
209 if (ErasedFromLoopHeaders)
210 LoopHeaders.insert(BB);
213 EverChanged |= Changed;
220 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
221 /// thread across it. Stop scanning the block when passing the threshold.
222 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
223 unsigned Threshold) {
224 /// Ignore PHI nodes, these will be flattened when duplication happens.
225 BasicBlock::const_iterator I = BB->getFirstNonPHI();
227 // FIXME: THREADING will delete values that are just used to compute the
228 // branch, so they shouldn't count against the duplication cost.
230 // Sum up the cost of each instruction until we get to the terminator. Don't
231 // include the terminator because the copy won't include it.
233 for (; !isa<TerminatorInst>(I); ++I) {
235 // Stop scanning the block if we've reached the threshold.
236 if (Size > Threshold)
239 // Debugger intrinsics don't incur code size.
240 if (isa<DbgInfoIntrinsic>(I)) continue;
242 // If this is a pointer->pointer bitcast, it is free.
243 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
246 // All other instructions count for at least one unit.
249 // Calls are more expensive. If they are non-intrinsic calls, we model them
250 // as having cost of 4. If they are a non-vector intrinsic, we model them
251 // as having cost of 2 total, and if they are a vector intrinsic, we model
252 // them as having cost 1.
253 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
254 if (CI->hasFnAttr(Attribute::NoDuplicate))
255 // Blocks with NoDuplicate are modelled as having infinite cost, so they
256 // are never duplicated.
258 else if (!isa<IntrinsicInst>(CI))
260 else if (!CI->getType()->isVectorTy())
265 // Threading through a switch statement is particularly profitable. If this
266 // block ends in a switch, decrease its cost to make it more likely to happen.
267 if (isa<SwitchInst>(I))
268 Size = Size > 6 ? Size-6 : 0;
270 // The same holds for indirect branches, but slightly more so.
271 if (isa<IndirectBrInst>(I))
272 Size = Size > 8 ? Size-8 : 0;
277 /// FindLoopHeaders - We do not want jump threading to turn proper loop
278 /// structures into irreducible loops. Doing this breaks up the loop nesting
279 /// hierarchy and pessimizes later transformations. To prevent this from
280 /// happening, we first have to find the loop headers. Here we approximate this
281 /// by finding targets of backedges in the CFG.
283 /// Note that there definitely are cases when we want to allow threading of
284 /// edges across a loop header. For example, threading a jump from outside the
285 /// loop (the preheader) to an exit block of the loop is definitely profitable.
286 /// It is also almost always profitable to thread backedges from within the loop
287 /// to exit blocks, and is often profitable to thread backedges to other blocks
288 /// within the loop (forming a nested loop). This simple analysis is not rich
289 /// enough to track all of these properties and keep it up-to-date as the CFG
290 /// mutates, so we don't allow any of these transformations.
292 void JumpThreading::FindLoopHeaders(Function &F) {
293 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
294 FindFunctionBackedges(F, Edges);
296 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
297 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
300 /// getKnownConstant - Helper method to determine if we can thread over a
301 /// terminator with the given value as its condition, and if so what value to
302 /// use for that. What kind of value this is depends on whether we want an
303 /// integer or a block address, but an undef is always accepted.
304 /// Returns null if Val is null or not an appropriate constant.
305 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
309 // Undef is "known" enough.
310 if (UndefValue *U = dyn_cast<UndefValue>(Val))
313 if (Preference == WantBlockAddress)
314 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
316 return dyn_cast<ConstantInt>(Val);
319 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
320 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
321 /// in any of our predecessors. If so, return the known list of value and pred
322 /// BB in the result vector.
324 /// This returns true if there were any known values.
327 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
328 ConstantPreference Preference) {
329 // This method walks up use-def chains recursively. Because of this, we could
330 // get into an infinite loop going around loops in the use-def chain. To
331 // prevent this, keep track of what (value, block) pairs we've already visited
332 // and terminate the search if we loop back to them
333 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
336 // An RAII help to remove this pair from the recursion set once the recursion
337 // stack pops back out again.
338 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
340 // If V is a constant, then it is known in all predecessors.
341 if (Constant *KC = getKnownConstant(V, Preference)) {
342 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
343 Result.push_back(std::make_pair(KC, *PI));
348 // If V is a non-instruction value, or an instruction in a different block,
349 // then it can't be derived from a PHI.
350 Instruction *I = dyn_cast<Instruction>(V);
351 if (I == 0 || I->getParent() != BB) {
353 // Okay, if this is a live-in value, see if it has a known value at the end
354 // of any of our predecessors.
356 // FIXME: This should be an edge property, not a block end property.
357 /// TODO: Per PR2563, we could infer value range information about a
358 /// predecessor based on its terminator.
360 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
361 // "I" is a non-local compare-with-a-constant instruction. This would be
362 // able to handle value inequalities better, for example if the compare is
363 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
364 // Perhaps getConstantOnEdge should be smart enough to do this?
366 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
368 // If the value is known by LazyValueInfo to be a constant in a
369 // predecessor, use that information to try to thread this block.
370 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
371 if (Constant *KC = getKnownConstant(PredCst, Preference))
372 Result.push_back(std::make_pair(KC, P));
375 return !Result.empty();
378 /// If I is a PHI node, then we know the incoming values for any constants.
379 if (PHINode *PN = dyn_cast<PHINode>(I)) {
380 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
381 Value *InVal = PN->getIncomingValue(i);
382 if (Constant *KC = getKnownConstant(InVal, Preference)) {
383 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
385 Constant *CI = LVI->getConstantOnEdge(InVal,
386 PN->getIncomingBlock(i), BB);
387 if (Constant *KC = getKnownConstant(CI, Preference))
388 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
392 return !Result.empty();
395 PredValueInfoTy LHSVals, RHSVals;
397 // Handle some boolean conditions.
398 if (I->getType()->getPrimitiveSizeInBits() == 1) {
399 assert(Preference == WantInteger && "One-bit non-integer type?");
401 // X & false -> false
402 if (I->getOpcode() == Instruction::Or ||
403 I->getOpcode() == Instruction::And) {
404 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
406 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
409 if (LHSVals.empty() && RHSVals.empty())
412 ConstantInt *InterestingVal;
413 if (I->getOpcode() == Instruction::Or)
414 InterestingVal = ConstantInt::getTrue(I->getContext());
416 InterestingVal = ConstantInt::getFalse(I->getContext());
418 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
420 // Scan for the sentinel. If we find an undef, force it to the
421 // interesting value: x|undef -> true and x&undef -> false.
422 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
423 if (LHSVals[i].first == InterestingVal ||
424 isa<UndefValue>(LHSVals[i].first)) {
425 Result.push_back(LHSVals[i]);
426 Result.back().first = InterestingVal;
427 LHSKnownBBs.insert(LHSVals[i].second);
429 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
430 if (RHSVals[i].first == InterestingVal ||
431 isa<UndefValue>(RHSVals[i].first)) {
432 // If we already inferred a value for this block on the LHS, don't
434 if (!LHSKnownBBs.count(RHSVals[i].second)) {
435 Result.push_back(RHSVals[i]);
436 Result.back().first = InterestingVal;
440 return !Result.empty();
443 // Handle the NOT form of XOR.
444 if (I->getOpcode() == Instruction::Xor &&
445 isa<ConstantInt>(I->getOperand(1)) &&
446 cast<ConstantInt>(I->getOperand(1))->isOne()) {
447 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
452 // Invert the known values.
453 for (unsigned i = 0, e = Result.size(); i != e; ++i)
454 Result[i].first = ConstantExpr::getNot(Result[i].first);
459 // Try to simplify some other binary operator values.
460 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
461 assert(Preference != WantBlockAddress
462 && "A binary operator creating a block address?");
463 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
464 PredValueInfoTy LHSVals;
465 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
468 // Try to use constant folding to simplify the binary operator.
469 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
470 Constant *V = LHSVals[i].first;
471 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
473 if (Constant *KC = getKnownConstant(Folded, WantInteger))
474 Result.push_back(std::make_pair(KC, LHSVals[i].second));
478 return !Result.empty();
481 // Handle compare with phi operand, where the PHI is defined in this block.
482 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
483 assert(Preference == WantInteger && "Compares only produce integers");
484 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
485 if (PN && PN->getParent() == BB) {
486 // We can do this simplification if any comparisons fold to true or false.
488 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
489 BasicBlock *PredBB = PN->getIncomingBlock(i);
490 Value *LHS = PN->getIncomingValue(i);
491 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
493 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
495 if (!isa<Constant>(RHS))
498 LazyValueInfo::Tristate
499 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
500 cast<Constant>(RHS), PredBB, BB);
501 if (ResT == LazyValueInfo::Unknown)
503 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
506 if (Constant *KC = getKnownConstant(Res, WantInteger))
507 Result.push_back(std::make_pair(KC, PredBB));
510 return !Result.empty();
514 // If comparing a live-in value against a constant, see if we know the
515 // live-in value on any predecessors.
516 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
517 if (!isa<Instruction>(Cmp->getOperand(0)) ||
518 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
519 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
521 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
523 // If the value is known by LazyValueInfo to be a constant in a
524 // predecessor, use that information to try to thread this block.
525 LazyValueInfo::Tristate Res =
526 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
528 if (Res == LazyValueInfo::Unknown)
531 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
532 Result.push_back(std::make_pair(ResC, P));
535 return !Result.empty();
538 // Try to find a constant value for the LHS of a comparison,
539 // and evaluate it statically if we can.
540 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
541 PredValueInfoTy LHSVals;
542 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
545 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
546 Constant *V = LHSVals[i].first;
547 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
549 if (Constant *KC = getKnownConstant(Folded, WantInteger))
550 Result.push_back(std::make_pair(KC, LHSVals[i].second));
553 return !Result.empty();
558 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
559 // Handle select instructions where at least one operand is a known constant
560 // and we can figure out the condition value for any predecessor block.
561 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
562 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
563 PredValueInfoTy Conds;
564 if ((TrueVal || FalseVal) &&
565 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
567 for (unsigned i = 0, e = Conds.size(); i != e; ++i) {
568 Constant *Cond = Conds[i].first;
570 // Figure out what value to use for the condition.
572 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
574 KnownCond = CI->isOne();
576 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
577 // Either operand will do, so be sure to pick the one that's a known
579 // FIXME: Do this more cleverly if both values are known constants?
580 KnownCond = (TrueVal != 0);
583 // See if the select has a known constant value for this predecessor.
584 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
585 Result.push_back(std::make_pair(Val, Conds[i].second));
588 return !Result.empty();
592 // If all else fails, see if LVI can figure out a constant value for us.
593 Constant *CI = LVI->getConstant(V, BB);
594 if (Constant *KC = getKnownConstant(CI, Preference)) {
595 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
596 Result.push_back(std::make_pair(KC, *PI));
599 return !Result.empty();
604 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
605 /// in an undefined jump, decide which block is best to revector to.
607 /// Since we can pick an arbitrary destination, we pick the successor with the
608 /// fewest predecessors. This should reduce the in-degree of the others.
610 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
611 TerminatorInst *BBTerm = BB->getTerminator();
612 unsigned MinSucc = 0;
613 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
614 // Compute the successor with the minimum number of predecessors.
615 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
616 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
617 TestBB = BBTerm->getSuccessor(i);
618 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
619 if (NumPreds < MinNumPreds) {
621 MinNumPreds = NumPreds;
628 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
629 if (!BB->hasAddressTaken()) return false;
631 // If the block has its address taken, it may be a tree of dead constants
632 // hanging off of it. These shouldn't keep the block alive.
633 BlockAddress *BA = BlockAddress::get(BB);
634 BA->removeDeadConstantUsers();
635 return !BA->use_empty();
638 /// ProcessBlock - If there are any predecessors whose control can be threaded
639 /// through to a successor, transform them now.
640 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
641 // If the block is trivially dead, just return and let the caller nuke it.
642 // This simplifies other transformations.
643 if (pred_begin(BB) == pred_end(BB) &&
644 BB != &BB->getParent()->getEntryBlock())
647 // If this block has a single predecessor, and if that pred has a single
648 // successor, merge the blocks. This encourages recursive jump threading
649 // because now the condition in this block can be threaded through
650 // predecessors of our predecessor block.
651 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
652 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
653 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
654 // If SinglePred was a loop header, BB becomes one.
655 if (LoopHeaders.erase(SinglePred))
656 LoopHeaders.insert(BB);
658 // Remember if SinglePred was the entry block of the function. If so, we
659 // will need to move BB back to the entry position.
660 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
661 LVI->eraseBlock(SinglePred);
662 MergeBasicBlockIntoOnlyPred(BB);
664 if (isEntry && BB != &BB->getParent()->getEntryBlock())
665 BB->moveBefore(&BB->getParent()->getEntryBlock());
670 // What kind of constant we're looking for.
671 ConstantPreference Preference = WantInteger;
673 // Look to see if the terminator is a conditional branch, switch or indirect
674 // branch, if not we can't thread it.
676 Instruction *Terminator = BB->getTerminator();
677 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
678 // Can't thread an unconditional jump.
679 if (BI->isUnconditional()) return false;
680 Condition = BI->getCondition();
681 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
682 Condition = SI->getCondition();
683 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
684 // Can't thread indirect branch with no successors.
685 if (IB->getNumSuccessors() == 0) return false;
686 Condition = IB->getAddress()->stripPointerCasts();
687 Preference = WantBlockAddress;
689 return false; // Must be an invoke.
692 // Run constant folding to see if we can reduce the condition to a simple
694 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
695 Value *SimpleVal = ConstantFoldInstruction(I, TD, TLI);
697 I->replaceAllUsesWith(SimpleVal);
698 I->eraseFromParent();
699 Condition = SimpleVal;
703 // If the terminator is branching on an undef, we can pick any of the
704 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
705 if (isa<UndefValue>(Condition)) {
706 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
708 // Fold the branch/switch.
709 TerminatorInst *BBTerm = BB->getTerminator();
710 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
711 if (i == BestSucc) continue;
712 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
715 DEBUG(dbgs() << " In block '" << BB->getName()
716 << "' folding undef terminator: " << *BBTerm << '\n');
717 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
718 BBTerm->eraseFromParent();
722 // If the terminator of this block is branching on a constant, simplify the
723 // terminator to an unconditional branch. This can occur due to threading in
725 if (getKnownConstant(Condition, Preference)) {
726 DEBUG(dbgs() << " In block '" << BB->getName()
727 << "' folding terminator: " << *BB->getTerminator() << '\n');
729 ConstantFoldTerminator(BB, true);
733 Instruction *CondInst = dyn_cast<Instruction>(Condition);
735 // All the rest of our checks depend on the condition being an instruction.
737 // FIXME: Unify this with code below.
738 if (ProcessThreadableEdges(Condition, BB, Preference))
744 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
745 // For a comparison where the LHS is outside this block, it's possible
746 // that we've branched on it before. Used LVI to see if we can simplify
747 // the branch based on that.
748 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
749 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
750 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
751 if (CondBr && CondConst && CondBr->isConditional() && PI != PE &&
752 (!isa<Instruction>(CondCmp->getOperand(0)) ||
753 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
754 // For predecessor edge, determine if the comparison is true or false
755 // on that edge. If they're all true or all false, we can simplify the
757 // FIXME: We could handle mixed true/false by duplicating code.
758 LazyValueInfo::Tristate Baseline =
759 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
761 if (Baseline != LazyValueInfo::Unknown) {
762 // Check that all remaining incoming values match the first one.
764 LazyValueInfo::Tristate Ret =
765 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
766 CondCmp->getOperand(0), CondConst, *PI, BB);
767 if (Ret != Baseline) break;
770 // If we terminated early, then one of the values didn't match.
772 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
773 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
774 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
775 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
776 CondBr->eraseFromParent();
783 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB))
787 // Check for some cases that are worth simplifying. Right now we want to look
788 // for loads that are used by a switch or by the condition for the branch. If
789 // we see one, check to see if it's partially redundant. If so, insert a PHI
790 // which can then be used to thread the values.
792 Value *SimplifyValue = CondInst;
793 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
794 if (isa<Constant>(CondCmp->getOperand(1)))
795 SimplifyValue = CondCmp->getOperand(0);
797 // TODO: There are other places where load PRE would be profitable, such as
798 // more complex comparisons.
799 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
800 if (SimplifyPartiallyRedundantLoad(LI))
804 // Handle a variety of cases where we are branching on something derived from
805 // a PHI node in the current block. If we can prove that any predecessors
806 // compute a predictable value based on a PHI node, thread those predecessors.
808 if (ProcessThreadableEdges(CondInst, BB, Preference))
811 // If this is an otherwise-unfoldable branch on a phi node in the current
812 // block, see if we can simplify.
813 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
814 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
815 return ProcessBranchOnPHI(PN);
818 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
819 if (CondInst->getOpcode() == Instruction::Xor &&
820 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
821 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
824 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
825 // "(X == 4)", thread through this block.
830 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
831 /// load instruction, eliminate it by replacing it with a PHI node. This is an
832 /// important optimization that encourages jump threading, and needs to be run
833 /// interlaced with other jump threading tasks.
834 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
835 // Don't hack volatile/atomic loads.
836 if (!LI->isSimple()) return false;
838 // If the load is defined in a block with exactly one predecessor, it can't be
839 // partially redundant.
840 BasicBlock *LoadBB = LI->getParent();
841 if (LoadBB->getSinglePredecessor())
844 // If the load is defined in a landing pad, it can't be partially redundant,
845 // because the edges between the invoke and the landing pad cannot have other
846 // instructions between them.
847 if (LoadBB->isLandingPad())
850 Value *LoadedPtr = LI->getOperand(0);
852 // If the loaded operand is defined in the LoadBB, it can't be available.
853 // TODO: Could do simple PHI translation, that would be fun :)
854 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
855 if (PtrOp->getParent() == LoadBB)
858 // Scan a few instructions up from the load, to see if it is obviously live at
859 // the entry to its block.
860 BasicBlock::iterator BBIt = LI;
862 if (Value *AvailableVal =
863 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
864 // If the value if the load is locally available within the block, just use
865 // it. This frequently occurs for reg2mem'd allocas.
866 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
868 // If the returned value is the load itself, replace with an undef. This can
869 // only happen in dead loops.
870 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
871 LI->replaceAllUsesWith(AvailableVal);
872 LI->eraseFromParent();
876 // Otherwise, if we scanned the whole block and got to the top of the block,
877 // we know the block is locally transparent to the load. If not, something
878 // might clobber its value.
879 if (BBIt != LoadBB->begin())
882 // If all of the loads and stores that feed the value have the same TBAA tag,
883 // then we can propagate it onto any newly inserted loads.
884 MDNode *TBAATag = LI->getMetadata(LLVMContext::MD_tbaa);
886 SmallPtrSet<BasicBlock*, 8> PredsScanned;
887 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
888 AvailablePredsTy AvailablePreds;
889 BasicBlock *OneUnavailablePred = 0;
891 // If we got here, the loaded value is transparent through to the start of the
892 // block. Check to see if it is available in any of the predecessor blocks.
893 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
895 BasicBlock *PredBB = *PI;
897 // If we already scanned this predecessor, skip it.
898 if (!PredsScanned.insert(PredBB))
901 // Scan the predecessor to see if the value is available in the pred.
902 BBIt = PredBB->end();
903 MDNode *ThisTBAATag = 0;
904 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6,
906 if (!PredAvailable) {
907 OneUnavailablePred = PredBB;
911 // If tbaa tags disagree or are not present, forget about them.
912 if (TBAATag != ThisTBAATag) TBAATag = 0;
914 // If so, this load is partially redundant. Remember this info so that we
915 // can create a PHI node.
916 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
919 // If the loaded value isn't available in any predecessor, it isn't partially
921 if (AvailablePreds.empty()) return false;
923 // Okay, the loaded value is available in at least one (and maybe all!)
924 // predecessors. If the value is unavailable in more than one unique
925 // predecessor, we want to insert a merge block for those common predecessors.
926 // This ensures that we only have to insert one reload, thus not increasing
928 BasicBlock *UnavailablePred = 0;
930 // If there is exactly one predecessor where the value is unavailable, the
931 // already computed 'OneUnavailablePred' block is it. If it ends in an
932 // unconditional branch, we know that it isn't a critical edge.
933 if (PredsScanned.size() == AvailablePreds.size()+1 &&
934 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
935 UnavailablePred = OneUnavailablePred;
936 } else if (PredsScanned.size() != AvailablePreds.size()) {
937 // Otherwise, we had multiple unavailable predecessors or we had a critical
938 // edge from the one.
939 SmallVector<BasicBlock*, 8> PredsToSplit;
940 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
942 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
943 AvailablePredSet.insert(AvailablePreds[i].first);
945 // Add all the unavailable predecessors to the PredsToSplit list.
946 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
949 // If the predecessor is an indirect goto, we can't split the edge.
950 if (isa<IndirectBrInst>(P->getTerminator()))
953 if (!AvailablePredSet.count(P))
954 PredsToSplit.push_back(P);
957 // Split them out to their own block.
959 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split", this);
962 // If the value isn't available in all predecessors, then there will be
963 // exactly one where it isn't available. Insert a load on that edge and add
964 // it to the AvailablePreds list.
965 if (UnavailablePred) {
966 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
967 "Can't handle critical edge here!");
968 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
970 UnavailablePred->getTerminator());
971 NewVal->setDebugLoc(LI->getDebugLoc());
973 NewVal->setMetadata(LLVMContext::MD_tbaa, TBAATag);
975 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
978 // Now we know that each predecessor of this block has a value in
979 // AvailablePreds, sort them for efficient access as we're walking the preds.
980 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
982 // Create a PHI node at the start of the block for the PRE'd load value.
983 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
984 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
987 PN->setDebugLoc(LI->getDebugLoc());
989 // Insert new entries into the PHI for each predecessor. A single block may
990 // have multiple entries here.
991 for (pred_iterator PI = PB; PI != PE; ++PI) {
993 AvailablePredsTy::iterator I =
994 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
995 std::make_pair(P, (Value*)0));
997 assert(I != AvailablePreds.end() && I->first == P &&
998 "Didn't find entry for predecessor!");
1000 PN->addIncoming(I->second, I->first);
1003 //cerr << "PRE: " << *LI << *PN << "\n";
1005 LI->replaceAllUsesWith(PN);
1006 LI->eraseFromParent();
1011 /// FindMostPopularDest - The specified list contains multiple possible
1012 /// threadable destinations. Pick the one that occurs the most frequently in
1015 FindMostPopularDest(BasicBlock *BB,
1016 const SmallVectorImpl<std::pair<BasicBlock*,
1017 BasicBlock*> > &PredToDestList) {
1018 assert(!PredToDestList.empty());
1020 // Determine popularity. If there are multiple possible destinations, we
1021 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1022 // blocks with known and real destinations to threading undef. We'll handle
1023 // them later if interesting.
1024 DenseMap<BasicBlock*, unsigned> DestPopularity;
1025 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1026 if (PredToDestList[i].second)
1027 DestPopularity[PredToDestList[i].second]++;
1029 // Find the most popular dest.
1030 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1031 BasicBlock *MostPopularDest = DPI->first;
1032 unsigned Popularity = DPI->second;
1033 SmallVector<BasicBlock*, 4> SamePopularity;
1035 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1036 // If the popularity of this entry isn't higher than the popularity we've
1037 // seen so far, ignore it.
1038 if (DPI->second < Popularity)
1040 else if (DPI->second == Popularity) {
1041 // If it is the same as what we've seen so far, keep track of it.
1042 SamePopularity.push_back(DPI->first);
1044 // If it is more popular, remember it.
1045 SamePopularity.clear();
1046 MostPopularDest = DPI->first;
1047 Popularity = DPI->second;
1051 // Okay, now we know the most popular destination. If there is more than one
1052 // destination, we need to determine one. This is arbitrary, but we need
1053 // to make a deterministic decision. Pick the first one that appears in the
1055 if (!SamePopularity.empty()) {
1056 SamePopularity.push_back(MostPopularDest);
1057 TerminatorInst *TI = BB->getTerminator();
1058 for (unsigned i = 0; ; ++i) {
1059 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1061 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1062 TI->getSuccessor(i)) == SamePopularity.end())
1065 MostPopularDest = TI->getSuccessor(i);
1070 // Okay, we have finally picked the most popular destination.
1071 return MostPopularDest;
1074 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1075 ConstantPreference Preference) {
1076 // If threading this would thread across a loop header, don't even try to
1078 if (LoopHeaders.count(BB))
1081 PredValueInfoTy PredValues;
1082 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference))
1085 assert(!PredValues.empty() &&
1086 "ComputeValueKnownInPredecessors returned true with no values");
1088 DEBUG(dbgs() << "IN BB: " << *BB;
1089 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1090 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1091 << *PredValues[i].first
1092 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1095 // Decide what we want to thread through. Convert our list of known values to
1096 // a list of known destinations for each pred. This also discards duplicate
1097 // predecessors and keeps track of the undefined inputs (which are represented
1098 // as a null dest in the PredToDestList).
1099 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1100 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1102 BasicBlock *OnlyDest = 0;
1103 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1105 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1106 BasicBlock *Pred = PredValues[i].second;
1107 if (!SeenPreds.insert(Pred))
1108 continue; // Duplicate predecessor entry.
1110 // If the predecessor ends with an indirect goto, we can't change its
1112 if (isa<IndirectBrInst>(Pred->getTerminator()))
1115 Constant *Val = PredValues[i].first;
1118 if (isa<UndefValue>(Val))
1120 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1121 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1122 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1123 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1125 assert(isa<IndirectBrInst>(BB->getTerminator())
1126 && "Unexpected terminator");
1127 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1130 // If we have exactly one destination, remember it for efficiency below.
1131 if (PredToDestList.empty())
1133 else if (OnlyDest != DestBB)
1134 OnlyDest = MultipleDestSentinel;
1136 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1139 // If all edges were unthreadable, we fail.
1140 if (PredToDestList.empty())
1143 // Determine which is the most common successor. If we have many inputs and
1144 // this block is a switch, we want to start by threading the batch that goes
1145 // to the most popular destination first. If we only know about one
1146 // threadable destination (the common case) we can avoid this.
1147 BasicBlock *MostPopularDest = OnlyDest;
1149 if (MostPopularDest == MultipleDestSentinel)
1150 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1152 // Now that we know what the most popular destination is, factor all
1153 // predecessors that will jump to it into a single predecessor.
1154 SmallVector<BasicBlock*, 16> PredsToFactor;
1155 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1156 if (PredToDestList[i].second == MostPopularDest) {
1157 BasicBlock *Pred = PredToDestList[i].first;
1159 // This predecessor may be a switch or something else that has multiple
1160 // edges to the block. Factor each of these edges by listing them
1161 // according to # occurrences in PredsToFactor.
1162 TerminatorInst *PredTI = Pred->getTerminator();
1163 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1164 if (PredTI->getSuccessor(i) == BB)
1165 PredsToFactor.push_back(Pred);
1168 // If the threadable edges are branching on an undefined value, we get to pick
1169 // the destination that these predecessors should get to.
1170 if (MostPopularDest == 0)
1171 MostPopularDest = BB->getTerminator()->
1172 getSuccessor(GetBestDestForJumpOnUndef(BB));
1174 // Ok, try to thread it!
1175 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1178 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1179 /// a PHI node in the current block. See if there are any simplifications we
1180 /// can do based on inputs to the phi node.
1182 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1183 BasicBlock *BB = PN->getParent();
1185 // TODO: We could make use of this to do it once for blocks with common PHI
1187 SmallVector<BasicBlock*, 1> PredBBs;
1190 // If any of the predecessor blocks end in an unconditional branch, we can
1191 // *duplicate* the conditional branch into that block in order to further
1192 // encourage jump threading and to eliminate cases where we have branch on a
1193 // phi of an icmp (branch on icmp is much better).
1194 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1195 BasicBlock *PredBB = PN->getIncomingBlock(i);
1196 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1197 if (PredBr->isUnconditional()) {
1198 PredBBs[0] = PredBB;
1199 // Try to duplicate BB into PredBB.
1200 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1208 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1209 /// a xor instruction in the current block. See if there are any
1210 /// simplifications we can do based on inputs to the xor.
1212 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1213 BasicBlock *BB = BO->getParent();
1215 // If either the LHS or RHS of the xor is a constant, don't do this
1217 if (isa<ConstantInt>(BO->getOperand(0)) ||
1218 isa<ConstantInt>(BO->getOperand(1)))
1221 // If the first instruction in BB isn't a phi, we won't be able to infer
1222 // anything special about any particular predecessor.
1223 if (!isa<PHINode>(BB->front()))
1226 // If we have a xor as the branch input to this block, and we know that the
1227 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1228 // the condition into the predecessor and fix that value to true, saving some
1229 // logical ops on that path and encouraging other paths to simplify.
1231 // This copies something like this:
1234 // %X = phi i1 [1], [%X']
1235 // %Y = icmp eq i32 %A, %B
1236 // %Z = xor i1 %X, %Y
1241 // %Y = icmp ne i32 %A, %B
1244 PredValueInfoTy XorOpValues;
1246 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1248 assert(XorOpValues.empty());
1249 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1255 assert(!XorOpValues.empty() &&
1256 "ComputeValueKnownInPredecessors returned true with no values");
1258 // Scan the information to see which is most popular: true or false. The
1259 // predecessors can be of the set true, false, or undef.
1260 unsigned NumTrue = 0, NumFalse = 0;
1261 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1262 if (isa<UndefValue>(XorOpValues[i].first))
1263 // Ignore undefs for the count.
1265 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1271 // Determine which value to split on, true, false, or undef if neither.
1272 ConstantInt *SplitVal = 0;
1273 if (NumTrue > NumFalse)
1274 SplitVal = ConstantInt::getTrue(BB->getContext());
1275 else if (NumTrue != 0 || NumFalse != 0)
1276 SplitVal = ConstantInt::getFalse(BB->getContext());
1278 // Collect all of the blocks that this can be folded into so that we can
1279 // factor this once and clone it once.
1280 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1281 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1282 if (XorOpValues[i].first != SplitVal &&
1283 !isa<UndefValue>(XorOpValues[i].first))
1286 BlocksToFoldInto.push_back(XorOpValues[i].second);
1289 // If we inferred a value for all of the predecessors, then duplication won't
1290 // help us. However, we can just replace the LHS or RHS with the constant.
1291 if (BlocksToFoldInto.size() ==
1292 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1293 if (SplitVal == 0) {
1294 // If all preds provide undef, just nuke the xor, because it is undef too.
1295 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1296 BO->eraseFromParent();
1297 } else if (SplitVal->isZero()) {
1298 // If all preds provide 0, replace the xor with the other input.
1299 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1300 BO->eraseFromParent();
1302 // If all preds provide 1, set the computed value to 1.
1303 BO->setOperand(!isLHS, SplitVal);
1309 // Try to duplicate BB into PredBB.
1310 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1314 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1315 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1316 /// NewPred using the entries from OldPred (suitably mapped).
1317 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1318 BasicBlock *OldPred,
1319 BasicBlock *NewPred,
1320 DenseMap<Instruction*, Value*> &ValueMap) {
1321 for (BasicBlock::iterator PNI = PHIBB->begin();
1322 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1323 // Ok, we have a PHI node. Figure out what the incoming value was for the
1325 Value *IV = PN->getIncomingValueForBlock(OldPred);
1327 // Remap the value if necessary.
1328 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1329 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1330 if (I != ValueMap.end())
1334 PN->addIncoming(IV, NewPred);
1338 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1339 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1340 /// across BB. Transform the IR to reflect this change.
1341 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1342 const SmallVectorImpl<BasicBlock*> &PredBBs,
1343 BasicBlock *SuccBB) {
1344 // If threading to the same block as we come from, we would infinite loop.
1346 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1347 << "' - would thread to self!\n");
1351 // If threading this would thread across a loop header, don't thread the edge.
1352 // See the comments above FindLoopHeaders for justifications and caveats.
1353 if (LoopHeaders.count(BB)) {
1354 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1355 << "' to dest BB '" << SuccBB->getName()
1356 << "' - it might create an irreducible loop!\n");
1360 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, Threshold);
1361 if (JumpThreadCost > Threshold) {
1362 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1363 << "' - Cost is too high: " << JumpThreadCost << "\n");
1367 // And finally, do it! Start by factoring the predecessors is needed.
1369 if (PredBBs.size() == 1)
1370 PredBB = PredBBs[0];
1372 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1373 << " common predecessors.\n");
1374 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1377 // And finally, do it!
1378 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1379 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1380 << ", across block:\n "
1383 LVI->threadEdge(PredBB, BB, SuccBB);
1385 // We are going to have to map operands from the original BB block to the new
1386 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1387 // account for entry from PredBB.
1388 DenseMap<Instruction*, Value*> ValueMapping;
1390 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1391 BB->getName()+".thread",
1392 BB->getParent(), BB);
1393 NewBB->moveAfter(PredBB);
1395 BasicBlock::iterator BI = BB->begin();
1396 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1397 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1399 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1400 // mapping and using it to remap operands in the cloned instructions.
1401 for (; !isa<TerminatorInst>(BI); ++BI) {
1402 Instruction *New = BI->clone();
1403 New->setName(BI->getName());
1404 NewBB->getInstList().push_back(New);
1405 ValueMapping[BI] = New;
1407 // Remap operands to patch up intra-block references.
1408 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1409 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1410 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1411 if (I != ValueMapping.end())
1412 New->setOperand(i, I->second);
1416 // We didn't copy the terminator from BB over to NewBB, because there is now
1417 // an unconditional jump to SuccBB. Insert the unconditional jump.
1418 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB);
1419 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1421 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1422 // PHI nodes for NewBB now.
1423 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1425 // If there were values defined in BB that are used outside the block, then we
1426 // now have to update all uses of the value to use either the original value,
1427 // the cloned value, or some PHI derived value. This can require arbitrary
1428 // PHI insertion, of which we are prepared to do, clean these up now.
1429 SSAUpdater SSAUpdate;
1430 SmallVector<Use*, 16> UsesToRename;
1431 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1432 // Scan all uses of this instruction to see if it is used outside of its
1433 // block, and if so, record them in UsesToRename.
1434 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1436 Instruction *User = cast<Instruction>(*UI);
1437 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1438 if (UserPN->getIncomingBlock(UI) == BB)
1440 } else if (User->getParent() == BB)
1443 UsesToRename.push_back(&UI.getUse());
1446 // If there are no uses outside the block, we're done with this instruction.
1447 if (UsesToRename.empty())
1450 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1452 // We found a use of I outside of BB. Rename all uses of I that are outside
1453 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1454 // with the two values we know.
1455 SSAUpdate.Initialize(I->getType(), I->getName());
1456 SSAUpdate.AddAvailableValue(BB, I);
1457 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1459 while (!UsesToRename.empty())
1460 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1461 DEBUG(dbgs() << "\n");
1465 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1466 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1467 // us to simplify any PHI nodes in BB.
1468 TerminatorInst *PredTerm = PredBB->getTerminator();
1469 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1470 if (PredTerm->getSuccessor(i) == BB) {
1471 BB->removePredecessor(PredBB, true);
1472 PredTerm->setSuccessor(i, NewBB);
1475 // At this point, the IR is fully up to date and consistent. Do a quick scan
1476 // over the new instructions and zap any that are constants or dead. This
1477 // frequently happens because of phi translation.
1478 SimplifyInstructionsInBlock(NewBB, TD, TLI);
1480 // Threaded an edge!
1485 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1486 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1487 /// If we can duplicate the contents of BB up into PredBB do so now, this
1488 /// improves the odds that the branch will be on an analyzable instruction like
1490 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1491 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1492 assert(!PredBBs.empty() && "Can't handle an empty set");
1494 // If BB is a loop header, then duplicating this block outside the loop would
1495 // cause us to transform this into an irreducible loop, don't do this.
1496 // See the comments above FindLoopHeaders for justifications and caveats.
1497 if (LoopHeaders.count(BB)) {
1498 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1499 << "' into predecessor block '" << PredBBs[0]->getName()
1500 << "' - it might create an irreducible loop!\n");
1504 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, Threshold);
1505 if (DuplicationCost > Threshold) {
1506 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1507 << "' - Cost is too high: " << DuplicationCost << "\n");
1511 // And finally, do it! Start by factoring the predecessors is needed.
1513 if (PredBBs.size() == 1)
1514 PredBB = PredBBs[0];
1516 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1517 << " common predecessors.\n");
1518 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1521 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1523 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1524 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1525 << DuplicationCost << " block is:" << *BB << "\n");
1527 // Unless PredBB ends with an unconditional branch, split the edge so that we
1528 // can just clone the bits from BB into the end of the new PredBB.
1529 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1531 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1532 PredBB = SplitEdge(PredBB, BB, this);
1533 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1536 // We are going to have to map operands from the original BB block into the
1537 // PredBB block. Evaluate PHI nodes in BB.
1538 DenseMap<Instruction*, Value*> ValueMapping;
1540 BasicBlock::iterator BI = BB->begin();
1541 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1542 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1544 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1545 // mapping and using it to remap operands in the cloned instructions.
1546 for (; BI != BB->end(); ++BI) {
1547 Instruction *New = BI->clone();
1549 // Remap operands to patch up intra-block references.
1550 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1551 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1552 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1553 if (I != ValueMapping.end())
1554 New->setOperand(i, I->second);
1557 // If this instruction can be simplified after the operands are updated,
1558 // just use the simplified value instead. This frequently happens due to
1560 if (Value *IV = SimplifyInstruction(New, TD)) {
1562 ValueMapping[BI] = IV;
1564 // Otherwise, insert the new instruction into the block.
1565 New->setName(BI->getName());
1566 PredBB->getInstList().insert(OldPredBranch, New);
1567 ValueMapping[BI] = New;
1571 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1572 // add entries to the PHI nodes for branch from PredBB now.
1573 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1574 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1576 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1579 // If there were values defined in BB that are used outside the block, then we
1580 // now have to update all uses of the value to use either the original value,
1581 // the cloned value, or some PHI derived value. This can require arbitrary
1582 // PHI insertion, of which we are prepared to do, clean these up now.
1583 SSAUpdater SSAUpdate;
1584 SmallVector<Use*, 16> UsesToRename;
1585 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1586 // Scan all uses of this instruction to see if it is used outside of its
1587 // block, and if so, record them in UsesToRename.
1588 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1590 Instruction *User = cast<Instruction>(*UI);
1591 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1592 if (UserPN->getIncomingBlock(UI) == BB)
1594 } else if (User->getParent() == BB)
1597 UsesToRename.push_back(&UI.getUse());
1600 // If there are no uses outside the block, we're done with this instruction.
1601 if (UsesToRename.empty())
1604 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1606 // We found a use of I outside of BB. Rename all uses of I that are outside
1607 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1608 // with the two values we know.
1609 SSAUpdate.Initialize(I->getType(), I->getName());
1610 SSAUpdate.AddAvailableValue(BB, I);
1611 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1613 while (!UsesToRename.empty())
1614 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1615 DEBUG(dbgs() << "\n");
1618 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1620 BB->removePredecessor(PredBB, true);
1622 // Remove the unconditional branch at the end of the PredBB block.
1623 OldPredBranch->eraseFromParent();
1629 /// TryToUnfoldSelect - Look for blocks of the form
1635 /// %p = phi [%a, %bb] ...
1639 /// And expand the select into a branch structure if one of its arms allows %c
1640 /// to be folded. This later enables threading from bb1 over bb2.
1641 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1642 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1643 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1644 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1646 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1647 CondLHS->getParent() != BB)
1650 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1651 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1652 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1654 // Look if one of the incoming values is a select in the corresponding
1656 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1659 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1660 if (!PredTerm || !PredTerm->isUnconditional())
1663 // Now check if one of the select values would allow us to constant fold the
1664 // terminator in BB. We don't do the transform if both sides fold, those
1665 // cases will be threaded in any case.
1666 LazyValueInfo::Tristate LHSFolds =
1667 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1669 LazyValueInfo::Tristate RHSFolds =
1670 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
1672 if ((LHSFolds != LazyValueInfo::Unknown ||
1673 RHSFolds != LazyValueInfo::Unknown) &&
1674 LHSFolds != RHSFolds) {
1675 // Expand the select.
1684 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
1685 BB->getParent(), BB);
1686 // Move the unconditional branch to NewBB.
1687 PredTerm->removeFromParent();
1688 NewBB->getInstList().insert(NewBB->end(), PredTerm);
1689 // Create a conditional branch and update PHI nodes.
1690 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
1691 CondLHS->setIncomingValue(I, SI->getFalseValue());
1692 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
1693 // The select is now dead.
1694 SI->eraseFromParent();
1696 // Update any other PHI nodes in BB.
1697 for (BasicBlock::iterator BI = BB->begin();
1698 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
1700 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);