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/ConstantFolding.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/LazyValueInfo.h"
25 #include "llvm/Analysis/Loads.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/Pass.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ValueHandle.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetLibraryInfo.h"
35 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
36 #include "llvm/Transforms/Utils/Local.h"
37 #include "llvm/Transforms/Utils/SSAUpdater.h"
40 STATISTIC(NumThreads, "Number of jumps threaded");
41 STATISTIC(NumFolds, "Number of terminators folded");
42 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
44 static cl::opt<unsigned>
45 Threshold("jump-threading-threshold",
46 cl::desc("Max block size to duplicate for jump threading"),
47 cl::init(6), cl::Hidden);
50 // These are at global scope so static functions can use them too.
51 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo;
52 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy;
54 // This is used to keep track of what kind of constant we're currently hoping
56 enum ConstantPreference {
61 /// This pass performs 'jump threading', which looks at blocks that have
62 /// multiple predecessors and multiple successors. If one or more of the
63 /// predecessors of the block can be proven to always jump to one of the
64 /// successors, we forward the edge from the predecessor to the successor by
65 /// duplicating the contents of this block.
67 /// An example of when this can occur is code like this:
74 /// In this case, the unconditional branch at the end of the first if can be
75 /// revectored to the false side of the second if.
77 class JumpThreading : public FunctionPass {
79 TargetLibraryInfo *TLI;
82 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
84 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
86 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet;
88 // RAII helper for updating the recursion stack.
89 struct RecursionSetRemover {
90 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet;
91 std::pair<Value*, BasicBlock*> ThePair;
93 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S,
94 std::pair<Value*, BasicBlock*> P)
95 : TheSet(S), ThePair(P) { }
97 ~RecursionSetRemover() {
98 TheSet.erase(ThePair);
102 static char ID; // Pass identification
103 JumpThreading() : FunctionPass(ID) {
104 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
107 bool runOnFunction(Function &F);
109 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
110 AU.addRequired<LazyValueInfo>();
111 AU.addPreserved<LazyValueInfo>();
112 AU.addRequired<TargetLibraryInfo>();
115 void FindLoopHeaders(Function &F);
116 bool ProcessBlock(BasicBlock *BB);
117 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
119 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
120 const SmallVectorImpl<BasicBlock *> &PredBBs);
122 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
123 PredValueInfo &Result,
124 ConstantPreference Preference);
125 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
126 ConstantPreference Preference);
128 bool ProcessBranchOnPHI(PHINode *PN);
129 bool ProcessBranchOnXOR(BinaryOperator *BO);
131 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
135 char JumpThreading::ID = 0;
136 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
137 "Jump Threading", false, false)
138 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)
139 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
140 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
141 "Jump Threading", false, false)
143 // Public interface to the Jump Threading pass
144 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
146 /// runOnFunction - Top level algorithm.
148 bool JumpThreading::runOnFunction(Function &F) {
149 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
150 TD = getAnalysisIfAvailable<DataLayout>();
151 TLI = &getAnalysis<TargetLibraryInfo>();
152 LVI = &getAnalysis<LazyValueInfo>();
156 bool Changed, EverChanged = false;
159 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
161 // Thread all of the branches we can over this block.
162 while (ProcessBlock(BB))
167 // If the block is trivially dead, zap it. This eliminates the successor
168 // edges which simplifies the CFG.
169 if (pred_begin(BB) == pred_end(BB) &&
170 BB != &BB->getParent()->getEntryBlock()) {
171 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
172 << "' with terminator: " << *BB->getTerminator() << '\n');
173 LoopHeaders.erase(BB);
180 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
182 // Can't thread an unconditional jump, but if the block is "almost
183 // empty", we can replace uses of it with uses of the successor and make
185 if (BI && BI->isUnconditional() &&
186 BB != &BB->getParent()->getEntryBlock() &&
187 // If the terminator is the only non-phi instruction, try to nuke it.
188 BB->getFirstNonPHIOrDbg()->isTerminator()) {
189 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
190 // block, we have to make sure it isn't in the LoopHeaders set. We
191 // reinsert afterward if needed.
192 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
193 BasicBlock *Succ = BI->getSuccessor(0);
195 // FIXME: It is always conservatively correct to drop the info
196 // for a block even if it doesn't get erased. This isn't totally
197 // awesome, but it allows us to use AssertingVH to prevent nasty
198 // dangling pointer issues within LazyValueInfo.
200 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
202 // If we deleted BB and BB was the header of a loop, then the
203 // successor is now the header of the loop.
207 if (ErasedFromLoopHeaders)
208 LoopHeaders.insert(BB);
211 EverChanged |= Changed;
218 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
219 /// thread across it. Stop scanning the block when passing the threshold.
220 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
221 unsigned Threshold) {
222 /// Ignore PHI nodes, these will be flattened when duplication happens.
223 BasicBlock::const_iterator I = BB->getFirstNonPHI();
225 // FIXME: THREADING will delete values that are just used to compute the
226 // branch, so they shouldn't count against the duplication cost.
228 // Sum up the cost of each instruction until we get to the terminator. Don't
229 // include the terminator because the copy won't include it.
231 for (; !isa<TerminatorInst>(I); ++I) {
233 // Stop scanning the block if we've reached the threshold.
234 if (Size > Threshold)
237 // Debugger intrinsics don't incur code size.
238 if (isa<DbgInfoIntrinsic>(I)) continue;
240 // If this is a pointer->pointer bitcast, it is free.
241 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
244 // All other instructions count for at least one unit.
247 // Calls are more expensive. If they are non-intrinsic calls, we model them
248 // as having cost of 4. If they are a non-vector intrinsic, we model them
249 // as having cost of 2 total, and if they are a vector intrinsic, we model
250 // them as having cost 1.
251 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
252 if (CI->hasFnAttr(Attribute::NoDuplicate))
253 // Blocks with NoDuplicate are modelled as having infinite cost, so they
254 // are never duplicated.
256 else if (!isa<IntrinsicInst>(CI))
258 else if (!CI->getType()->isVectorTy())
263 // Threading through a switch statement is particularly profitable. If this
264 // block ends in a switch, decrease its cost to make it more likely to happen.
265 if (isa<SwitchInst>(I))
266 Size = Size > 6 ? Size-6 : 0;
268 // The same holds for indirect branches, but slightly more so.
269 if (isa<IndirectBrInst>(I))
270 Size = Size > 8 ? Size-8 : 0;
275 /// FindLoopHeaders - We do not want jump threading to turn proper loop
276 /// structures into irreducible loops. Doing this breaks up the loop nesting
277 /// hierarchy and pessimizes later transformations. To prevent this from
278 /// happening, we first have to find the loop headers. Here we approximate this
279 /// by finding targets of backedges in the CFG.
281 /// Note that there definitely are cases when we want to allow threading of
282 /// edges across a loop header. For example, threading a jump from outside the
283 /// loop (the preheader) to an exit block of the loop is definitely profitable.
284 /// It is also almost always profitable to thread backedges from within the loop
285 /// to exit blocks, and is often profitable to thread backedges to other blocks
286 /// within the loop (forming a nested loop). This simple analysis is not rich
287 /// enough to track all of these properties and keep it up-to-date as the CFG
288 /// mutates, so we don't allow any of these transformations.
290 void JumpThreading::FindLoopHeaders(Function &F) {
291 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
292 FindFunctionBackedges(F, Edges);
294 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
295 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
298 /// getKnownConstant - Helper method to determine if we can thread over a
299 /// terminator with the given value as its condition, and if so what value to
300 /// use for that. What kind of value this is depends on whether we want an
301 /// integer or a block address, but an undef is always accepted.
302 /// Returns null if Val is null or not an appropriate constant.
303 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
307 // Undef is "known" enough.
308 if (UndefValue *U = dyn_cast<UndefValue>(Val))
311 if (Preference == WantBlockAddress)
312 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
314 return dyn_cast<ConstantInt>(Val);
317 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
318 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
319 /// in any of our predecessors. If so, return the known list of value and pred
320 /// BB in the result vector.
322 /// This returns true if there were any known values.
325 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
326 ConstantPreference Preference) {
327 // This method walks up use-def chains recursively. Because of this, we could
328 // get into an infinite loop going around loops in the use-def chain. To
329 // prevent this, keep track of what (value, block) pairs we've already visited
330 // and terminate the search if we loop back to them
331 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
334 // An RAII help to remove this pair from the recursion set once the recursion
335 // stack pops back out again.
336 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
338 // If V is a constant, then it is known in all predecessors.
339 if (Constant *KC = getKnownConstant(V, Preference)) {
340 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
341 Result.push_back(std::make_pair(KC, *PI));
346 // If V is a non-instruction value, or an instruction in a different block,
347 // then it can't be derived from a PHI.
348 Instruction *I = dyn_cast<Instruction>(V);
349 if (I == 0 || I->getParent() != BB) {
351 // Okay, if this is a live-in value, see if it has a known value at the end
352 // of any of our predecessors.
354 // FIXME: This should be an edge property, not a block end property.
355 /// TODO: Per PR2563, we could infer value range information about a
356 /// predecessor based on its terminator.
358 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
359 // "I" is a non-local compare-with-a-constant instruction. This would be
360 // able to handle value inequalities better, for example if the compare is
361 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
362 // Perhaps getConstantOnEdge should be smart enough to do this?
364 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
366 // If the value is known by LazyValueInfo to be a constant in a
367 // predecessor, use that information to try to thread this block.
368 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
369 if (Constant *KC = getKnownConstant(PredCst, Preference))
370 Result.push_back(std::make_pair(KC, P));
373 return !Result.empty();
376 /// If I is a PHI node, then we know the incoming values for any constants.
377 if (PHINode *PN = dyn_cast<PHINode>(I)) {
378 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
379 Value *InVal = PN->getIncomingValue(i);
380 if (Constant *KC = getKnownConstant(InVal, Preference)) {
381 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
383 Constant *CI = LVI->getConstantOnEdge(InVal,
384 PN->getIncomingBlock(i), BB);
385 if (Constant *KC = getKnownConstant(CI, Preference))
386 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
390 return !Result.empty();
393 PredValueInfoTy LHSVals, RHSVals;
395 // Handle some boolean conditions.
396 if (I->getType()->getPrimitiveSizeInBits() == 1) {
397 assert(Preference == WantInteger && "One-bit non-integer type?");
399 // X & false -> false
400 if (I->getOpcode() == Instruction::Or ||
401 I->getOpcode() == Instruction::And) {
402 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
404 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
407 if (LHSVals.empty() && RHSVals.empty())
410 ConstantInt *InterestingVal;
411 if (I->getOpcode() == Instruction::Or)
412 InterestingVal = ConstantInt::getTrue(I->getContext());
414 InterestingVal = ConstantInt::getFalse(I->getContext());
416 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
418 // Scan for the sentinel. If we find an undef, force it to the
419 // interesting value: x|undef -> true and x&undef -> false.
420 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
421 if (LHSVals[i].first == InterestingVal ||
422 isa<UndefValue>(LHSVals[i].first)) {
423 Result.push_back(LHSVals[i]);
424 Result.back().first = InterestingVal;
425 LHSKnownBBs.insert(LHSVals[i].second);
427 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
428 if (RHSVals[i].first == InterestingVal ||
429 isa<UndefValue>(RHSVals[i].first)) {
430 // If we already inferred a value for this block on the LHS, don't
432 if (!LHSKnownBBs.count(RHSVals[i].second)) {
433 Result.push_back(RHSVals[i]);
434 Result.back().first = InterestingVal;
438 return !Result.empty();
441 // Handle the NOT form of XOR.
442 if (I->getOpcode() == Instruction::Xor &&
443 isa<ConstantInt>(I->getOperand(1)) &&
444 cast<ConstantInt>(I->getOperand(1))->isOne()) {
445 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
450 // Invert the known values.
451 for (unsigned i = 0, e = Result.size(); i != e; ++i)
452 Result[i].first = ConstantExpr::getNot(Result[i].first);
457 // Try to simplify some other binary operator values.
458 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
459 assert(Preference != WantBlockAddress
460 && "A binary operator creating a block address?");
461 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
462 PredValueInfoTy LHSVals;
463 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
466 // Try to use constant folding to simplify the binary operator.
467 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
468 Constant *V = LHSVals[i].first;
469 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
471 if (Constant *KC = getKnownConstant(Folded, WantInteger))
472 Result.push_back(std::make_pair(KC, LHSVals[i].second));
476 return !Result.empty();
479 // Handle compare with phi operand, where the PHI is defined in this block.
480 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
481 assert(Preference == WantInteger && "Compares only produce integers");
482 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
483 if (PN && PN->getParent() == BB) {
484 // We can do this simplification if any comparisons fold to true or false.
486 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
487 BasicBlock *PredBB = PN->getIncomingBlock(i);
488 Value *LHS = PN->getIncomingValue(i);
489 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
491 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
493 if (!isa<Constant>(RHS))
496 LazyValueInfo::Tristate
497 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
498 cast<Constant>(RHS), PredBB, BB);
499 if (ResT == LazyValueInfo::Unknown)
501 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
504 if (Constant *KC = getKnownConstant(Res, WantInteger))
505 Result.push_back(std::make_pair(KC, PredBB));
508 return !Result.empty();
512 // If comparing a live-in value against a constant, see if we know the
513 // live-in value on any predecessors.
514 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
515 if (!isa<Instruction>(Cmp->getOperand(0)) ||
516 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
517 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
519 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
521 // If the value is known by LazyValueInfo to be a constant in a
522 // predecessor, use that information to try to thread this block.
523 LazyValueInfo::Tristate Res =
524 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
526 if (Res == LazyValueInfo::Unknown)
529 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
530 Result.push_back(std::make_pair(ResC, P));
533 return !Result.empty();
536 // Try to find a constant value for the LHS of a comparison,
537 // and evaluate it statically if we can.
538 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
539 PredValueInfoTy LHSVals;
540 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
543 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
544 Constant *V = LHSVals[i].first;
545 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
547 if (Constant *KC = getKnownConstant(Folded, WantInteger))
548 Result.push_back(std::make_pair(KC, LHSVals[i].second));
551 return !Result.empty();
556 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
557 // Handle select instructions where at least one operand is a known constant
558 // and we can figure out the condition value for any predecessor block.
559 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
560 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
561 PredValueInfoTy Conds;
562 if ((TrueVal || FalseVal) &&
563 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
565 for (unsigned i = 0, e = Conds.size(); i != e; ++i) {
566 Constant *Cond = Conds[i].first;
568 // Figure out what value to use for the condition.
570 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
572 KnownCond = CI->isOne();
574 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
575 // Either operand will do, so be sure to pick the one that's a known
577 // FIXME: Do this more cleverly if both values are known constants?
578 KnownCond = (TrueVal != 0);
581 // See if the select has a known constant value for this predecessor.
582 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
583 Result.push_back(std::make_pair(Val, Conds[i].second));
586 return !Result.empty();
590 // If all else fails, see if LVI can figure out a constant value for us.
591 Constant *CI = LVI->getConstant(V, BB);
592 if (Constant *KC = getKnownConstant(CI, Preference)) {
593 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
594 Result.push_back(std::make_pair(KC, *PI));
597 return !Result.empty();
602 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
603 /// in an undefined jump, decide which block is best to revector to.
605 /// Since we can pick an arbitrary destination, we pick the successor with the
606 /// fewest predecessors. This should reduce the in-degree of the others.
608 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
609 TerminatorInst *BBTerm = BB->getTerminator();
610 unsigned MinSucc = 0;
611 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
612 // Compute the successor with the minimum number of predecessors.
613 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
614 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
615 TestBB = BBTerm->getSuccessor(i);
616 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
617 if (NumPreds < MinNumPreds) {
619 MinNumPreds = NumPreds;
626 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
627 if (!BB->hasAddressTaken()) return false;
629 // If the block has its address taken, it may be a tree of dead constants
630 // hanging off of it. These shouldn't keep the block alive.
631 BlockAddress *BA = BlockAddress::get(BB);
632 BA->removeDeadConstantUsers();
633 return !BA->use_empty();
636 /// ProcessBlock - If there are any predecessors whose control can be threaded
637 /// through to a successor, transform them now.
638 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
639 // If the block is trivially dead, just return and let the caller nuke it.
640 // This simplifies other transformations.
641 if (pred_begin(BB) == pred_end(BB) &&
642 BB != &BB->getParent()->getEntryBlock())
645 // If this block has a single predecessor, and if that pred has a single
646 // successor, merge the blocks. This encourages recursive jump threading
647 // because now the condition in this block can be threaded through
648 // predecessors of our predecessor block.
649 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
650 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
651 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
652 // If SinglePred was a loop header, BB becomes one.
653 if (LoopHeaders.erase(SinglePred))
654 LoopHeaders.insert(BB);
656 // Remember if SinglePred was the entry block of the function. If so, we
657 // will need to move BB back to the entry position.
658 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
659 LVI->eraseBlock(SinglePred);
660 MergeBasicBlockIntoOnlyPred(BB);
662 if (isEntry && BB != &BB->getParent()->getEntryBlock())
663 BB->moveBefore(&BB->getParent()->getEntryBlock());
668 // What kind of constant we're looking for.
669 ConstantPreference Preference = WantInteger;
671 // Look to see if the terminator is a conditional branch, switch or indirect
672 // branch, if not we can't thread it.
674 Instruction *Terminator = BB->getTerminator();
675 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
676 // Can't thread an unconditional jump.
677 if (BI->isUnconditional()) return false;
678 Condition = BI->getCondition();
679 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
680 Condition = SI->getCondition();
681 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
682 // Can't thread indirect branch with no successors.
683 if (IB->getNumSuccessors() == 0) return false;
684 Condition = IB->getAddress()->stripPointerCasts();
685 Preference = WantBlockAddress;
687 return false; // Must be an invoke.
690 // Run constant folding to see if we can reduce the condition to a simple
692 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
693 Value *SimpleVal = ConstantFoldInstruction(I, TD, TLI);
695 I->replaceAllUsesWith(SimpleVal);
696 I->eraseFromParent();
697 Condition = SimpleVal;
701 // If the terminator is branching on an undef, we can pick any of the
702 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
703 if (isa<UndefValue>(Condition)) {
704 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
706 // Fold the branch/switch.
707 TerminatorInst *BBTerm = BB->getTerminator();
708 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
709 if (i == BestSucc) continue;
710 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
713 DEBUG(dbgs() << " In block '" << BB->getName()
714 << "' folding undef terminator: " << *BBTerm << '\n');
715 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
716 BBTerm->eraseFromParent();
720 // If the terminator of this block is branching on a constant, simplify the
721 // terminator to an unconditional branch. This can occur due to threading in
723 if (getKnownConstant(Condition, Preference)) {
724 DEBUG(dbgs() << " In block '" << BB->getName()
725 << "' folding terminator: " << *BB->getTerminator() << '\n');
727 ConstantFoldTerminator(BB, true);
731 Instruction *CondInst = dyn_cast<Instruction>(Condition);
733 // All the rest of our checks depend on the condition being an instruction.
735 // FIXME: Unify this with code below.
736 if (ProcessThreadableEdges(Condition, BB, Preference))
742 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
743 // For a comparison where the LHS is outside this block, it's possible
744 // that we've branched on it before. Used LVI to see if we can simplify
745 // the branch based on that.
746 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
747 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
748 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
749 if (CondBr && CondConst && CondBr->isConditional() && PI != PE &&
750 (!isa<Instruction>(CondCmp->getOperand(0)) ||
751 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
752 // For predecessor edge, determine if the comparison is true or false
753 // on that edge. If they're all true or all false, we can simplify the
755 // FIXME: We could handle mixed true/false by duplicating code.
756 LazyValueInfo::Tristate Baseline =
757 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
759 if (Baseline != LazyValueInfo::Unknown) {
760 // Check that all remaining incoming values match the first one.
762 LazyValueInfo::Tristate Ret =
763 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
764 CondCmp->getOperand(0), CondConst, *PI, BB);
765 if (Ret != Baseline) break;
768 // If we terminated early, then one of the values didn't match.
770 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
771 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
772 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
773 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
774 CondBr->eraseFromParent();
781 // Check for some cases that are worth simplifying. Right now we want to look
782 // for loads that are used by a switch or by the condition for the branch. If
783 // we see one, check to see if it's partially redundant. If so, insert a PHI
784 // which can then be used to thread the values.
786 Value *SimplifyValue = CondInst;
787 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
788 if (isa<Constant>(CondCmp->getOperand(1)))
789 SimplifyValue = CondCmp->getOperand(0);
791 // TODO: There are other places where load PRE would be profitable, such as
792 // more complex comparisons.
793 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
794 if (SimplifyPartiallyRedundantLoad(LI))
798 // Handle a variety of cases where we are branching on something derived from
799 // a PHI node in the current block. If we can prove that any predecessors
800 // compute a predictable value based on a PHI node, thread those predecessors.
802 if (ProcessThreadableEdges(CondInst, BB, Preference))
805 // If this is an otherwise-unfoldable branch on a phi node in the current
806 // block, see if we can simplify.
807 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
808 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
809 return ProcessBranchOnPHI(PN);
812 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
813 if (CondInst->getOpcode() == Instruction::Xor &&
814 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
815 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
818 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
819 // "(X == 4)", thread through this block.
825 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
826 /// load instruction, eliminate it by replacing it with a PHI node. This is an
827 /// important optimization that encourages jump threading, and needs to be run
828 /// interlaced with other jump threading tasks.
829 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
830 // Don't hack volatile/atomic loads.
831 if (!LI->isSimple()) return false;
833 // If the load is defined in a block with exactly one predecessor, it can't be
834 // partially redundant.
835 BasicBlock *LoadBB = LI->getParent();
836 if (LoadBB->getSinglePredecessor())
839 Value *LoadedPtr = LI->getOperand(0);
841 // If the loaded operand is defined in the LoadBB, it can't be available.
842 // TODO: Could do simple PHI translation, that would be fun :)
843 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
844 if (PtrOp->getParent() == LoadBB)
847 // Scan a few instructions up from the load, to see if it is obviously live at
848 // the entry to its block.
849 BasicBlock::iterator BBIt = LI;
851 if (Value *AvailableVal =
852 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
853 // If the value if the load is locally available within the block, just use
854 // it. This frequently occurs for reg2mem'd allocas.
855 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
857 // If the returned value is the load itself, replace with an undef. This can
858 // only happen in dead loops.
859 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
860 LI->replaceAllUsesWith(AvailableVal);
861 LI->eraseFromParent();
865 // Otherwise, if we scanned the whole block and got to the top of the block,
866 // we know the block is locally transparent to the load. If not, something
867 // might clobber its value.
868 if (BBIt != LoadBB->begin())
871 // If all of the loads and stores that feed the value have the same TBAA tag,
872 // then we can propagate it onto any newly inserted loads.
873 MDNode *TBAATag = LI->getMetadata(LLVMContext::MD_tbaa);
875 SmallPtrSet<BasicBlock*, 8> PredsScanned;
876 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
877 AvailablePredsTy AvailablePreds;
878 BasicBlock *OneUnavailablePred = 0;
880 // If we got here, the loaded value is transparent through to the start of the
881 // block. Check to see if it is available in any of the predecessor blocks.
882 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
884 BasicBlock *PredBB = *PI;
886 // If we already scanned this predecessor, skip it.
887 if (!PredsScanned.insert(PredBB))
890 // Scan the predecessor to see if the value is available in the pred.
891 BBIt = PredBB->end();
892 MDNode *ThisTBAATag = 0;
893 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6,
895 if (!PredAvailable) {
896 OneUnavailablePred = PredBB;
900 // If tbaa tags disagree or are not present, forget about them.
901 if (TBAATag != ThisTBAATag) TBAATag = 0;
903 // If so, this load is partially redundant. Remember this info so that we
904 // can create a PHI node.
905 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
908 // If the loaded value isn't available in any predecessor, it isn't partially
910 if (AvailablePreds.empty()) return false;
912 // Okay, the loaded value is available in at least one (and maybe all!)
913 // predecessors. If the value is unavailable in more than one unique
914 // predecessor, we want to insert a merge block for those common predecessors.
915 // This ensures that we only have to insert one reload, thus not increasing
917 BasicBlock *UnavailablePred = 0;
919 // If there is exactly one predecessor where the value is unavailable, the
920 // already computed 'OneUnavailablePred' block is it. If it ends in an
921 // unconditional branch, we know that it isn't a critical edge.
922 if (PredsScanned.size() == AvailablePreds.size()+1 &&
923 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
924 UnavailablePred = OneUnavailablePred;
925 } else if (PredsScanned.size() != AvailablePreds.size()) {
926 // Otherwise, we had multiple unavailable predecessors or we had a critical
927 // edge from the one.
928 SmallVector<BasicBlock*, 8> PredsToSplit;
929 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
931 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
932 AvailablePredSet.insert(AvailablePreds[i].first);
934 // Add all the unavailable predecessors to the PredsToSplit list.
935 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
938 // If the predecessor is an indirect goto, we can't split the edge.
939 if (isa<IndirectBrInst>(P->getTerminator()))
942 if (!AvailablePredSet.count(P))
943 PredsToSplit.push_back(P);
946 // Split them out to their own block.
948 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split", this);
951 // If the value isn't available in all predecessors, then there will be
952 // exactly one where it isn't available. Insert a load on that edge and add
953 // it to the AvailablePreds list.
954 if (UnavailablePred) {
955 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
956 "Can't handle critical edge here!");
957 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
959 UnavailablePred->getTerminator());
960 NewVal->setDebugLoc(LI->getDebugLoc());
962 NewVal->setMetadata(LLVMContext::MD_tbaa, TBAATag);
964 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
967 // Now we know that each predecessor of this block has a value in
968 // AvailablePreds, sort them for efficient access as we're walking the preds.
969 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
971 // Create a PHI node at the start of the block for the PRE'd load value.
972 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
973 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
976 PN->setDebugLoc(LI->getDebugLoc());
978 // Insert new entries into the PHI for each predecessor. A single block may
979 // have multiple entries here.
980 for (pred_iterator PI = PB; PI != PE; ++PI) {
982 AvailablePredsTy::iterator I =
983 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
984 std::make_pair(P, (Value*)0));
986 assert(I != AvailablePreds.end() && I->first == P &&
987 "Didn't find entry for predecessor!");
989 PN->addIncoming(I->second, I->first);
992 //cerr << "PRE: " << *LI << *PN << "\n";
994 LI->replaceAllUsesWith(PN);
995 LI->eraseFromParent();
1000 /// FindMostPopularDest - The specified list contains multiple possible
1001 /// threadable destinations. Pick the one that occurs the most frequently in
1004 FindMostPopularDest(BasicBlock *BB,
1005 const SmallVectorImpl<std::pair<BasicBlock*,
1006 BasicBlock*> > &PredToDestList) {
1007 assert(!PredToDestList.empty());
1009 // Determine popularity. If there are multiple possible destinations, we
1010 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1011 // blocks with known and real destinations to threading undef. We'll handle
1012 // them later if interesting.
1013 DenseMap<BasicBlock*, unsigned> DestPopularity;
1014 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1015 if (PredToDestList[i].second)
1016 DestPopularity[PredToDestList[i].second]++;
1018 // Find the most popular dest.
1019 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1020 BasicBlock *MostPopularDest = DPI->first;
1021 unsigned Popularity = DPI->second;
1022 SmallVector<BasicBlock*, 4> SamePopularity;
1024 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1025 // If the popularity of this entry isn't higher than the popularity we've
1026 // seen so far, ignore it.
1027 if (DPI->second < Popularity)
1029 else if (DPI->second == Popularity) {
1030 // If it is the same as what we've seen so far, keep track of it.
1031 SamePopularity.push_back(DPI->first);
1033 // If it is more popular, remember it.
1034 SamePopularity.clear();
1035 MostPopularDest = DPI->first;
1036 Popularity = DPI->second;
1040 // Okay, now we know the most popular destination. If there is more than one
1041 // destination, we need to determine one. This is arbitrary, but we need
1042 // to make a deterministic decision. Pick the first one that appears in the
1044 if (!SamePopularity.empty()) {
1045 SamePopularity.push_back(MostPopularDest);
1046 TerminatorInst *TI = BB->getTerminator();
1047 for (unsigned i = 0; ; ++i) {
1048 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1050 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1051 TI->getSuccessor(i)) == SamePopularity.end())
1054 MostPopularDest = TI->getSuccessor(i);
1059 // Okay, we have finally picked the most popular destination.
1060 return MostPopularDest;
1063 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1064 ConstantPreference Preference) {
1065 // If threading this would thread across a loop header, don't even try to
1067 if (LoopHeaders.count(BB))
1070 PredValueInfoTy PredValues;
1071 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference))
1074 assert(!PredValues.empty() &&
1075 "ComputeValueKnownInPredecessors returned true with no values");
1077 DEBUG(dbgs() << "IN BB: " << *BB;
1078 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1079 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1080 << *PredValues[i].first
1081 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1084 // Decide what we want to thread through. Convert our list of known values to
1085 // a list of known destinations for each pred. This also discards duplicate
1086 // predecessors and keeps track of the undefined inputs (which are represented
1087 // as a null dest in the PredToDestList).
1088 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1089 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1091 BasicBlock *OnlyDest = 0;
1092 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1094 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1095 BasicBlock *Pred = PredValues[i].second;
1096 if (!SeenPreds.insert(Pred))
1097 continue; // Duplicate predecessor entry.
1099 // If the predecessor ends with an indirect goto, we can't change its
1101 if (isa<IndirectBrInst>(Pred->getTerminator()))
1104 Constant *Val = PredValues[i].first;
1107 if (isa<UndefValue>(Val))
1109 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1110 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1111 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1112 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1114 assert(isa<IndirectBrInst>(BB->getTerminator())
1115 && "Unexpected terminator");
1116 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1119 // If we have exactly one destination, remember it for efficiency below.
1120 if (PredToDestList.empty())
1122 else if (OnlyDest != DestBB)
1123 OnlyDest = MultipleDestSentinel;
1125 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1128 // If all edges were unthreadable, we fail.
1129 if (PredToDestList.empty())
1132 // Determine which is the most common successor. If we have many inputs and
1133 // this block is a switch, we want to start by threading the batch that goes
1134 // to the most popular destination first. If we only know about one
1135 // threadable destination (the common case) we can avoid this.
1136 BasicBlock *MostPopularDest = OnlyDest;
1138 if (MostPopularDest == MultipleDestSentinel)
1139 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1141 // Now that we know what the most popular destination is, factor all
1142 // predecessors that will jump to it into a single predecessor.
1143 SmallVector<BasicBlock*, 16> PredsToFactor;
1144 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1145 if (PredToDestList[i].second == MostPopularDest) {
1146 BasicBlock *Pred = PredToDestList[i].first;
1148 // This predecessor may be a switch or something else that has multiple
1149 // edges to the block. Factor each of these edges by listing them
1150 // according to # occurrences in PredsToFactor.
1151 TerminatorInst *PredTI = Pred->getTerminator();
1152 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1153 if (PredTI->getSuccessor(i) == BB)
1154 PredsToFactor.push_back(Pred);
1157 // If the threadable edges are branching on an undefined value, we get to pick
1158 // the destination that these predecessors should get to.
1159 if (MostPopularDest == 0)
1160 MostPopularDest = BB->getTerminator()->
1161 getSuccessor(GetBestDestForJumpOnUndef(BB));
1163 // Ok, try to thread it!
1164 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1167 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1168 /// a PHI node in the current block. See if there are any simplifications we
1169 /// can do based on inputs to the phi node.
1171 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1172 BasicBlock *BB = PN->getParent();
1174 // TODO: We could make use of this to do it once for blocks with common PHI
1176 SmallVector<BasicBlock*, 1> PredBBs;
1179 // If any of the predecessor blocks end in an unconditional branch, we can
1180 // *duplicate* the conditional branch into that block in order to further
1181 // encourage jump threading and to eliminate cases where we have branch on a
1182 // phi of an icmp (branch on icmp is much better).
1183 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1184 BasicBlock *PredBB = PN->getIncomingBlock(i);
1185 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1186 if (PredBr->isUnconditional()) {
1187 PredBBs[0] = PredBB;
1188 // Try to duplicate BB into PredBB.
1189 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1197 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1198 /// a xor instruction in the current block. See if there are any
1199 /// simplifications we can do based on inputs to the xor.
1201 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1202 BasicBlock *BB = BO->getParent();
1204 // If either the LHS or RHS of the xor is a constant, don't do this
1206 if (isa<ConstantInt>(BO->getOperand(0)) ||
1207 isa<ConstantInt>(BO->getOperand(1)))
1210 // If the first instruction in BB isn't a phi, we won't be able to infer
1211 // anything special about any particular predecessor.
1212 if (!isa<PHINode>(BB->front()))
1215 // If we have a xor as the branch input to this block, and we know that the
1216 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1217 // the condition into the predecessor and fix that value to true, saving some
1218 // logical ops on that path and encouraging other paths to simplify.
1220 // This copies something like this:
1223 // %X = phi i1 [1], [%X']
1224 // %Y = icmp eq i32 %A, %B
1225 // %Z = xor i1 %X, %Y
1230 // %Y = icmp ne i32 %A, %B
1233 PredValueInfoTy XorOpValues;
1235 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1237 assert(XorOpValues.empty());
1238 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1244 assert(!XorOpValues.empty() &&
1245 "ComputeValueKnownInPredecessors returned true with no values");
1247 // Scan the information to see which is most popular: true or false. The
1248 // predecessors can be of the set true, false, or undef.
1249 unsigned NumTrue = 0, NumFalse = 0;
1250 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1251 if (isa<UndefValue>(XorOpValues[i].first))
1252 // Ignore undefs for the count.
1254 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1260 // Determine which value to split on, true, false, or undef if neither.
1261 ConstantInt *SplitVal = 0;
1262 if (NumTrue > NumFalse)
1263 SplitVal = ConstantInt::getTrue(BB->getContext());
1264 else if (NumTrue != 0 || NumFalse != 0)
1265 SplitVal = ConstantInt::getFalse(BB->getContext());
1267 // Collect all of the blocks that this can be folded into so that we can
1268 // factor this once and clone it once.
1269 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1270 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1271 if (XorOpValues[i].first != SplitVal &&
1272 !isa<UndefValue>(XorOpValues[i].first))
1275 BlocksToFoldInto.push_back(XorOpValues[i].second);
1278 // If we inferred a value for all of the predecessors, then duplication won't
1279 // help us. However, we can just replace the LHS or RHS with the constant.
1280 if (BlocksToFoldInto.size() ==
1281 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1282 if (SplitVal == 0) {
1283 // If all preds provide undef, just nuke the xor, because it is undef too.
1284 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1285 BO->eraseFromParent();
1286 } else if (SplitVal->isZero()) {
1287 // If all preds provide 0, replace the xor with the other input.
1288 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1289 BO->eraseFromParent();
1291 // If all preds provide 1, set the computed value to 1.
1292 BO->setOperand(!isLHS, SplitVal);
1298 // Try to duplicate BB into PredBB.
1299 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1303 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1304 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1305 /// NewPred using the entries from OldPred (suitably mapped).
1306 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1307 BasicBlock *OldPred,
1308 BasicBlock *NewPred,
1309 DenseMap<Instruction*, Value*> &ValueMap) {
1310 for (BasicBlock::iterator PNI = PHIBB->begin();
1311 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1312 // Ok, we have a PHI node. Figure out what the incoming value was for the
1314 Value *IV = PN->getIncomingValueForBlock(OldPred);
1316 // Remap the value if necessary.
1317 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1318 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1319 if (I != ValueMap.end())
1323 PN->addIncoming(IV, NewPred);
1327 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1328 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1329 /// across BB. Transform the IR to reflect this change.
1330 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1331 const SmallVectorImpl<BasicBlock*> &PredBBs,
1332 BasicBlock *SuccBB) {
1333 // If threading to the same block as we come from, we would infinite loop.
1335 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1336 << "' - would thread to self!\n");
1340 // If threading this would thread across a loop header, don't thread the edge.
1341 // See the comments above FindLoopHeaders for justifications and caveats.
1342 if (LoopHeaders.count(BB)) {
1343 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1344 << "' to dest BB '" << SuccBB->getName()
1345 << "' - it might create an irreducible loop!\n");
1349 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, Threshold);
1350 if (JumpThreadCost > Threshold) {
1351 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1352 << "' - Cost is too high: " << JumpThreadCost << "\n");
1356 // And finally, do it! Start by factoring the predecessors is needed.
1358 if (PredBBs.size() == 1)
1359 PredBB = PredBBs[0];
1361 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1362 << " common predecessors.\n");
1363 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1366 // And finally, do it!
1367 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1368 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1369 << ", across block:\n "
1372 LVI->threadEdge(PredBB, BB, SuccBB);
1374 // We are going to have to map operands from the original BB block to the new
1375 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1376 // account for entry from PredBB.
1377 DenseMap<Instruction*, Value*> ValueMapping;
1379 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1380 BB->getName()+".thread",
1381 BB->getParent(), BB);
1382 NewBB->moveAfter(PredBB);
1384 BasicBlock::iterator BI = BB->begin();
1385 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1386 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1388 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1389 // mapping and using it to remap operands in the cloned instructions.
1390 for (; !isa<TerminatorInst>(BI); ++BI) {
1391 Instruction *New = BI->clone();
1392 New->setName(BI->getName());
1393 NewBB->getInstList().push_back(New);
1394 ValueMapping[BI] = New;
1396 // Remap operands to patch up intra-block references.
1397 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1398 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1399 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1400 if (I != ValueMapping.end())
1401 New->setOperand(i, I->second);
1405 // We didn't copy the terminator from BB over to NewBB, because there is now
1406 // an unconditional jump to SuccBB. Insert the unconditional jump.
1407 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB);
1408 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1410 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1411 // PHI nodes for NewBB now.
1412 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1414 // If there were values defined in BB that are used outside the block, then we
1415 // now have to update all uses of the value to use either the original value,
1416 // the cloned value, or some PHI derived value. This can require arbitrary
1417 // PHI insertion, of which we are prepared to do, clean these up now.
1418 SSAUpdater SSAUpdate;
1419 SmallVector<Use*, 16> UsesToRename;
1420 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1421 // Scan all uses of this instruction to see if it is used outside of its
1422 // block, and if so, record them in UsesToRename.
1423 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1425 Instruction *User = cast<Instruction>(*UI);
1426 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1427 if (UserPN->getIncomingBlock(UI) == BB)
1429 } else if (User->getParent() == BB)
1432 UsesToRename.push_back(&UI.getUse());
1435 // If there are no uses outside the block, we're done with this instruction.
1436 if (UsesToRename.empty())
1439 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1441 // We found a use of I outside of BB. Rename all uses of I that are outside
1442 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1443 // with the two values we know.
1444 SSAUpdate.Initialize(I->getType(), I->getName());
1445 SSAUpdate.AddAvailableValue(BB, I);
1446 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1448 while (!UsesToRename.empty())
1449 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1450 DEBUG(dbgs() << "\n");
1454 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1455 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1456 // us to simplify any PHI nodes in BB.
1457 TerminatorInst *PredTerm = PredBB->getTerminator();
1458 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1459 if (PredTerm->getSuccessor(i) == BB) {
1460 BB->removePredecessor(PredBB, true);
1461 PredTerm->setSuccessor(i, NewBB);
1464 // At this point, the IR is fully up to date and consistent. Do a quick scan
1465 // over the new instructions and zap any that are constants or dead. This
1466 // frequently happens because of phi translation.
1467 SimplifyInstructionsInBlock(NewBB, TD, TLI);
1469 // Threaded an edge!
1474 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1475 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1476 /// If we can duplicate the contents of BB up into PredBB do so now, this
1477 /// improves the odds that the branch will be on an analyzable instruction like
1479 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1480 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1481 assert(!PredBBs.empty() && "Can't handle an empty set");
1483 // If BB is a loop header, then duplicating this block outside the loop would
1484 // cause us to transform this into an irreducible loop, don't do this.
1485 // See the comments above FindLoopHeaders for justifications and caveats.
1486 if (LoopHeaders.count(BB)) {
1487 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1488 << "' into predecessor block '" << PredBBs[0]->getName()
1489 << "' - it might create an irreducible loop!\n");
1493 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, Threshold);
1494 if (DuplicationCost > Threshold) {
1495 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1496 << "' - Cost is too high: " << DuplicationCost << "\n");
1500 // And finally, do it! Start by factoring the predecessors is needed.
1502 if (PredBBs.size() == 1)
1503 PredBB = PredBBs[0];
1505 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1506 << " common predecessors.\n");
1507 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1510 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1512 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1513 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1514 << DuplicationCost << " block is:" << *BB << "\n");
1516 // Unless PredBB ends with an unconditional branch, split the edge so that we
1517 // can just clone the bits from BB into the end of the new PredBB.
1518 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1520 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1521 PredBB = SplitEdge(PredBB, BB, this);
1522 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1525 // We are going to have to map operands from the original BB block into the
1526 // PredBB block. Evaluate PHI nodes in BB.
1527 DenseMap<Instruction*, Value*> ValueMapping;
1529 BasicBlock::iterator BI = BB->begin();
1530 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1531 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1533 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1534 // mapping and using it to remap operands in the cloned instructions.
1535 for (; BI != BB->end(); ++BI) {
1536 Instruction *New = BI->clone();
1538 // Remap operands to patch up intra-block references.
1539 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1540 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1541 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1542 if (I != ValueMapping.end())
1543 New->setOperand(i, I->second);
1546 // If this instruction can be simplified after the operands are updated,
1547 // just use the simplified value instead. This frequently happens due to
1549 if (Value *IV = SimplifyInstruction(New, TD)) {
1551 ValueMapping[BI] = IV;
1553 // Otherwise, insert the new instruction into the block.
1554 New->setName(BI->getName());
1555 PredBB->getInstList().insert(OldPredBranch, New);
1556 ValueMapping[BI] = New;
1560 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1561 // add entries to the PHI nodes for branch from PredBB now.
1562 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1563 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1565 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1568 // If there were values defined in BB that are used outside the block, then we
1569 // now have to update all uses of the value to use either the original value,
1570 // the cloned value, or some PHI derived value. This can require arbitrary
1571 // PHI insertion, of which we are prepared to do, clean these up now.
1572 SSAUpdater SSAUpdate;
1573 SmallVector<Use*, 16> UsesToRename;
1574 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1575 // Scan all uses of this instruction to see if it is used outside of its
1576 // block, and if so, record them in UsesToRename.
1577 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1579 Instruction *User = cast<Instruction>(*UI);
1580 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1581 if (UserPN->getIncomingBlock(UI) == BB)
1583 } else if (User->getParent() == BB)
1586 UsesToRename.push_back(&UI.getUse());
1589 // If there are no uses outside the block, we're done with this instruction.
1590 if (UsesToRename.empty())
1593 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1595 // We found a use of I outside of BB. Rename all uses of I that are outside
1596 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1597 // with the two values we know.
1598 SSAUpdate.Initialize(I->getType(), I->getName());
1599 SSAUpdate.AddAvailableValue(BB, I);
1600 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1602 while (!UsesToRename.empty())
1603 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1604 DEBUG(dbgs() << "\n");
1607 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1609 BB->removePredecessor(PredBB, true);
1611 // Remove the unconditional branch at the end of the PredBB block.
1612 OldPredBranch->eraseFromParent();