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 if (skipOptnoneFunction(F))
154 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
155 TD = getAnalysisIfAvailable<DataLayout>();
156 TLI = &getAnalysis<TargetLibraryInfo>();
157 LVI = &getAnalysis<LazyValueInfo>();
161 bool Changed, EverChanged = false;
164 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
166 // Thread all of the branches we can over this block.
167 while (ProcessBlock(BB))
172 // If the block is trivially dead, zap it. This eliminates the successor
173 // edges which simplifies the CFG.
174 if (pred_begin(BB) == pred_end(BB) &&
175 BB != &BB->getParent()->getEntryBlock()) {
176 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
177 << "' with terminator: " << *BB->getTerminator() << '\n');
178 LoopHeaders.erase(BB);
185 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
187 // Can't thread an unconditional jump, but if the block is "almost
188 // empty", we can replace uses of it with uses of the successor and make
190 if (BI && BI->isUnconditional() &&
191 BB != &BB->getParent()->getEntryBlock() &&
192 // If the terminator is the only non-phi instruction, try to nuke it.
193 BB->getFirstNonPHIOrDbg()->isTerminator()) {
194 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
195 // block, we have to make sure it isn't in the LoopHeaders set. We
196 // reinsert afterward if needed.
197 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
198 BasicBlock *Succ = BI->getSuccessor(0);
200 // FIXME: It is always conservatively correct to drop the info
201 // for a block even if it doesn't get erased. This isn't totally
202 // awesome, but it allows us to use AssertingVH to prevent nasty
203 // dangling pointer issues within LazyValueInfo.
205 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
207 // If we deleted BB and BB was the header of a loop, then the
208 // successor is now the header of the loop.
212 if (ErasedFromLoopHeaders)
213 LoopHeaders.insert(BB);
216 EverChanged |= Changed;
223 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
224 /// thread across it. Stop scanning the block when passing the threshold.
225 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
226 unsigned Threshold) {
227 /// Ignore PHI nodes, these will be flattened when duplication happens.
228 BasicBlock::const_iterator I = BB->getFirstNonPHI();
230 // FIXME: THREADING will delete values that are just used to compute the
231 // branch, so they shouldn't count against the duplication cost.
233 // Sum up the cost of each instruction until we get to the terminator. Don't
234 // include the terminator because the copy won't include it.
236 for (; !isa<TerminatorInst>(I); ++I) {
238 // Stop scanning the block if we've reached the threshold.
239 if (Size > Threshold)
242 // Debugger intrinsics don't incur code size.
243 if (isa<DbgInfoIntrinsic>(I)) continue;
245 // If this is a pointer->pointer bitcast, it is free.
246 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
249 // All other instructions count for at least one unit.
252 // Calls are more expensive. If they are non-intrinsic calls, we model them
253 // as having cost of 4. If they are a non-vector intrinsic, we model them
254 // as having cost of 2 total, and if they are a vector intrinsic, we model
255 // them as having cost 1.
256 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
257 if (CI->hasFnAttr(Attribute::NoDuplicate))
258 // Blocks with NoDuplicate are modelled as having infinite cost, so they
259 // are never duplicated.
261 else if (!isa<IntrinsicInst>(CI))
263 else if (!CI->getType()->isVectorTy())
268 // Threading through a switch statement is particularly profitable. If this
269 // block ends in a switch, decrease its cost to make it more likely to happen.
270 if (isa<SwitchInst>(I))
271 Size = Size > 6 ? Size-6 : 0;
273 // The same holds for indirect branches, but slightly more so.
274 if (isa<IndirectBrInst>(I))
275 Size = Size > 8 ? Size-8 : 0;
280 /// FindLoopHeaders - We do not want jump threading to turn proper loop
281 /// structures into irreducible loops. Doing this breaks up the loop nesting
282 /// hierarchy and pessimizes later transformations. To prevent this from
283 /// happening, we first have to find the loop headers. Here we approximate this
284 /// by finding targets of backedges in the CFG.
286 /// Note that there definitely are cases when we want to allow threading of
287 /// edges across a loop header. For example, threading a jump from outside the
288 /// loop (the preheader) to an exit block of the loop is definitely profitable.
289 /// It is also almost always profitable to thread backedges from within the loop
290 /// to exit blocks, and is often profitable to thread backedges to other blocks
291 /// within the loop (forming a nested loop). This simple analysis is not rich
292 /// enough to track all of these properties and keep it up-to-date as the CFG
293 /// mutates, so we don't allow any of these transformations.
295 void JumpThreading::FindLoopHeaders(Function &F) {
296 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
297 FindFunctionBackedges(F, Edges);
299 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
300 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
303 /// getKnownConstant - Helper method to determine if we can thread over a
304 /// terminator with the given value as its condition, and if so what value to
305 /// use for that. What kind of value this is depends on whether we want an
306 /// integer or a block address, but an undef is always accepted.
307 /// Returns null if Val is null or not an appropriate constant.
308 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
312 // Undef is "known" enough.
313 if (UndefValue *U = dyn_cast<UndefValue>(Val))
316 if (Preference == WantBlockAddress)
317 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
319 return dyn_cast<ConstantInt>(Val);
322 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
323 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
324 /// in any of our predecessors. If so, return the known list of value and pred
325 /// BB in the result vector.
327 /// This returns true if there were any known values.
330 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
331 ConstantPreference Preference) {
332 // This method walks up use-def chains recursively. Because of this, we could
333 // get into an infinite loop going around loops in the use-def chain. To
334 // prevent this, keep track of what (value, block) pairs we've already visited
335 // and terminate the search if we loop back to them
336 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
339 // An RAII help to remove this pair from the recursion set once the recursion
340 // stack pops back out again.
341 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
343 // If V is a constant, then it is known in all predecessors.
344 if (Constant *KC = getKnownConstant(V, Preference)) {
345 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
346 Result.push_back(std::make_pair(KC, *PI));
351 // If V is a non-instruction value, or an instruction in a different block,
352 // then it can't be derived from a PHI.
353 Instruction *I = dyn_cast<Instruction>(V);
354 if (I == 0 || I->getParent() != BB) {
356 // Okay, if this is a live-in value, see if it has a known value at the end
357 // of any of our predecessors.
359 // FIXME: This should be an edge property, not a block end property.
360 /// TODO: Per PR2563, we could infer value range information about a
361 /// predecessor based on its terminator.
363 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
364 // "I" is a non-local compare-with-a-constant instruction. This would be
365 // able to handle value inequalities better, for example if the compare is
366 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
367 // Perhaps getConstantOnEdge should be smart enough to do this?
369 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
371 // If the value is known by LazyValueInfo to be a constant in a
372 // predecessor, use that information to try to thread this block.
373 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
374 if (Constant *KC = getKnownConstant(PredCst, Preference))
375 Result.push_back(std::make_pair(KC, P));
378 return !Result.empty();
381 /// If I is a PHI node, then we know the incoming values for any constants.
382 if (PHINode *PN = dyn_cast<PHINode>(I)) {
383 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
384 Value *InVal = PN->getIncomingValue(i);
385 if (Constant *KC = getKnownConstant(InVal, Preference)) {
386 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
388 Constant *CI = LVI->getConstantOnEdge(InVal,
389 PN->getIncomingBlock(i), BB);
390 if (Constant *KC = getKnownConstant(CI, Preference))
391 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
395 return !Result.empty();
398 PredValueInfoTy LHSVals, RHSVals;
400 // Handle some boolean conditions.
401 if (I->getType()->getPrimitiveSizeInBits() == 1) {
402 assert(Preference == WantInteger && "One-bit non-integer type?");
404 // X & false -> false
405 if (I->getOpcode() == Instruction::Or ||
406 I->getOpcode() == Instruction::And) {
407 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
409 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
412 if (LHSVals.empty() && RHSVals.empty())
415 ConstantInt *InterestingVal;
416 if (I->getOpcode() == Instruction::Or)
417 InterestingVal = ConstantInt::getTrue(I->getContext());
419 InterestingVal = ConstantInt::getFalse(I->getContext());
421 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
423 // Scan for the sentinel. If we find an undef, force it to the
424 // interesting value: x|undef -> true and x&undef -> false.
425 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
426 if (LHSVals[i].first == InterestingVal ||
427 isa<UndefValue>(LHSVals[i].first)) {
428 Result.push_back(LHSVals[i]);
429 Result.back().first = InterestingVal;
430 LHSKnownBBs.insert(LHSVals[i].second);
432 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
433 if (RHSVals[i].first == InterestingVal ||
434 isa<UndefValue>(RHSVals[i].first)) {
435 // If we already inferred a value for this block on the LHS, don't
437 if (!LHSKnownBBs.count(RHSVals[i].second)) {
438 Result.push_back(RHSVals[i]);
439 Result.back().first = InterestingVal;
443 return !Result.empty();
446 // Handle the NOT form of XOR.
447 if (I->getOpcode() == Instruction::Xor &&
448 isa<ConstantInt>(I->getOperand(1)) &&
449 cast<ConstantInt>(I->getOperand(1))->isOne()) {
450 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
455 // Invert the known values.
456 for (unsigned i = 0, e = Result.size(); i != e; ++i)
457 Result[i].first = ConstantExpr::getNot(Result[i].first);
462 // Try to simplify some other binary operator values.
463 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
464 assert(Preference != WantBlockAddress
465 && "A binary operator creating a block address?");
466 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
467 PredValueInfoTy LHSVals;
468 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
471 // Try to use constant folding to simplify the binary operator.
472 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
473 Constant *V = LHSVals[i].first;
474 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
476 if (Constant *KC = getKnownConstant(Folded, WantInteger))
477 Result.push_back(std::make_pair(KC, LHSVals[i].second));
481 return !Result.empty();
484 // Handle compare with phi operand, where the PHI is defined in this block.
485 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
486 assert(Preference == WantInteger && "Compares only produce integers");
487 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
488 if (PN && PN->getParent() == BB) {
489 // We can do this simplification if any comparisons fold to true or false.
491 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
492 BasicBlock *PredBB = PN->getIncomingBlock(i);
493 Value *LHS = PN->getIncomingValue(i);
494 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
496 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
498 if (!isa<Constant>(RHS))
501 LazyValueInfo::Tristate
502 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
503 cast<Constant>(RHS), PredBB, BB);
504 if (ResT == LazyValueInfo::Unknown)
506 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
509 if (Constant *KC = getKnownConstant(Res, WantInteger))
510 Result.push_back(std::make_pair(KC, PredBB));
513 return !Result.empty();
517 // If comparing a live-in value against a constant, see if we know the
518 // live-in value on any predecessors.
519 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
520 if (!isa<Instruction>(Cmp->getOperand(0)) ||
521 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
522 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
524 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
526 // If the value is known by LazyValueInfo to be a constant in a
527 // predecessor, use that information to try to thread this block.
528 LazyValueInfo::Tristate Res =
529 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
531 if (Res == LazyValueInfo::Unknown)
534 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
535 Result.push_back(std::make_pair(ResC, P));
538 return !Result.empty();
541 // Try to find a constant value for the LHS of a comparison,
542 // and evaluate it statically if we can.
543 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
544 PredValueInfoTy LHSVals;
545 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
548 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
549 Constant *V = LHSVals[i].first;
550 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
552 if (Constant *KC = getKnownConstant(Folded, WantInteger))
553 Result.push_back(std::make_pair(KC, LHSVals[i].second));
556 return !Result.empty();
561 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
562 // Handle select instructions where at least one operand is a known constant
563 // and we can figure out the condition value for any predecessor block.
564 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
565 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
566 PredValueInfoTy Conds;
567 if ((TrueVal || FalseVal) &&
568 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
570 for (unsigned i = 0, e = Conds.size(); i != e; ++i) {
571 Constant *Cond = Conds[i].first;
573 // Figure out what value to use for the condition.
575 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
577 KnownCond = CI->isOne();
579 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
580 // Either operand will do, so be sure to pick the one that's a known
582 // FIXME: Do this more cleverly if both values are known constants?
583 KnownCond = (TrueVal != 0);
586 // See if the select has a known constant value for this predecessor.
587 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
588 Result.push_back(std::make_pair(Val, Conds[i].second));
591 return !Result.empty();
595 // If all else fails, see if LVI can figure out a constant value for us.
596 Constant *CI = LVI->getConstant(V, BB);
597 if (Constant *KC = getKnownConstant(CI, Preference)) {
598 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
599 Result.push_back(std::make_pair(KC, *PI));
602 return !Result.empty();
607 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
608 /// in an undefined jump, decide which block is best to revector to.
610 /// Since we can pick an arbitrary destination, we pick the successor with the
611 /// fewest predecessors. This should reduce the in-degree of the others.
613 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
614 TerminatorInst *BBTerm = BB->getTerminator();
615 unsigned MinSucc = 0;
616 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
617 // Compute the successor with the minimum number of predecessors.
618 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
619 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
620 TestBB = BBTerm->getSuccessor(i);
621 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
622 if (NumPreds < MinNumPreds) {
624 MinNumPreds = NumPreds;
631 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
632 if (!BB->hasAddressTaken()) return false;
634 // If the block has its address taken, it may be a tree of dead constants
635 // hanging off of it. These shouldn't keep the block alive.
636 BlockAddress *BA = BlockAddress::get(BB);
637 BA->removeDeadConstantUsers();
638 return !BA->use_empty();
641 /// ProcessBlock - If there are any predecessors whose control can be threaded
642 /// through to a successor, transform them now.
643 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
644 // If the block is trivially dead, just return and let the caller nuke it.
645 // This simplifies other transformations.
646 if (pred_begin(BB) == pred_end(BB) &&
647 BB != &BB->getParent()->getEntryBlock())
650 // If this block has a single predecessor, and if that pred has a single
651 // successor, merge the blocks. This encourages recursive jump threading
652 // because now the condition in this block can be threaded through
653 // predecessors of our predecessor block.
654 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
655 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
656 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
657 // If SinglePred was a loop header, BB becomes one.
658 if (LoopHeaders.erase(SinglePred))
659 LoopHeaders.insert(BB);
661 // Remember if SinglePred was the entry block of the function. If so, we
662 // will need to move BB back to the entry position.
663 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
664 LVI->eraseBlock(SinglePred);
665 MergeBasicBlockIntoOnlyPred(BB);
667 if (isEntry && BB != &BB->getParent()->getEntryBlock())
668 BB->moveBefore(&BB->getParent()->getEntryBlock());
673 // What kind of constant we're looking for.
674 ConstantPreference Preference = WantInteger;
676 // Look to see if the terminator is a conditional branch, switch or indirect
677 // branch, if not we can't thread it.
679 Instruction *Terminator = BB->getTerminator();
680 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
681 // Can't thread an unconditional jump.
682 if (BI->isUnconditional()) return false;
683 Condition = BI->getCondition();
684 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
685 Condition = SI->getCondition();
686 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
687 // Can't thread indirect branch with no successors.
688 if (IB->getNumSuccessors() == 0) return false;
689 Condition = IB->getAddress()->stripPointerCasts();
690 Preference = WantBlockAddress;
692 return false; // Must be an invoke.
695 // Run constant folding to see if we can reduce the condition to a simple
697 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
698 Value *SimpleVal = ConstantFoldInstruction(I, TD, TLI);
700 I->replaceAllUsesWith(SimpleVal);
701 I->eraseFromParent();
702 Condition = SimpleVal;
706 // If the terminator is branching on an undef, we can pick any of the
707 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
708 if (isa<UndefValue>(Condition)) {
709 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
711 // Fold the branch/switch.
712 TerminatorInst *BBTerm = BB->getTerminator();
713 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
714 if (i == BestSucc) continue;
715 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
718 DEBUG(dbgs() << " In block '" << BB->getName()
719 << "' folding undef terminator: " << *BBTerm << '\n');
720 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
721 BBTerm->eraseFromParent();
725 // If the terminator of this block is branching on a constant, simplify the
726 // terminator to an unconditional branch. This can occur due to threading in
728 if (getKnownConstant(Condition, Preference)) {
729 DEBUG(dbgs() << " In block '" << BB->getName()
730 << "' folding terminator: " << *BB->getTerminator() << '\n');
732 ConstantFoldTerminator(BB, true);
736 Instruction *CondInst = dyn_cast<Instruction>(Condition);
738 // All the rest of our checks depend on the condition being an instruction.
740 // FIXME: Unify this with code below.
741 if (ProcessThreadableEdges(Condition, BB, Preference))
747 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
748 // For a comparison where the LHS is outside this block, it's possible
749 // that we've branched on it before. Used LVI to see if we can simplify
750 // the branch based on that.
751 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
752 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
753 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
754 if (CondBr && CondConst && CondBr->isConditional() && PI != PE &&
755 (!isa<Instruction>(CondCmp->getOperand(0)) ||
756 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
757 // For predecessor edge, determine if the comparison is true or false
758 // on that edge. If they're all true or all false, we can simplify the
760 // FIXME: We could handle mixed true/false by duplicating code.
761 LazyValueInfo::Tristate Baseline =
762 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
764 if (Baseline != LazyValueInfo::Unknown) {
765 // Check that all remaining incoming values match the first one.
767 LazyValueInfo::Tristate Ret =
768 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
769 CondCmp->getOperand(0), CondConst, *PI, BB);
770 if (Ret != Baseline) break;
773 // If we terminated early, then one of the values didn't match.
775 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
776 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
777 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
778 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
779 CondBr->eraseFromParent();
786 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB))
790 // Check for some cases that are worth simplifying. Right now we want to look
791 // for loads that are used by a switch or by the condition for the branch. If
792 // we see one, check to see if it's partially redundant. If so, insert a PHI
793 // which can then be used to thread the values.
795 Value *SimplifyValue = CondInst;
796 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
797 if (isa<Constant>(CondCmp->getOperand(1)))
798 SimplifyValue = CondCmp->getOperand(0);
800 // TODO: There are other places where load PRE would be profitable, such as
801 // more complex comparisons.
802 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
803 if (SimplifyPartiallyRedundantLoad(LI))
807 // Handle a variety of cases where we are branching on something derived from
808 // a PHI node in the current block. If we can prove that any predecessors
809 // compute a predictable value based on a PHI node, thread those predecessors.
811 if (ProcessThreadableEdges(CondInst, BB, Preference))
814 // If this is an otherwise-unfoldable branch on a phi node in the current
815 // block, see if we can simplify.
816 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
817 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
818 return ProcessBranchOnPHI(PN);
821 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
822 if (CondInst->getOpcode() == Instruction::Xor &&
823 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
824 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
827 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
828 // "(X == 4)", thread through this block.
833 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
834 /// load instruction, eliminate it by replacing it with a PHI node. This is an
835 /// important optimization that encourages jump threading, and needs to be run
836 /// interlaced with other jump threading tasks.
837 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
838 // Don't hack volatile/atomic loads.
839 if (!LI->isSimple()) return false;
841 // If the load is defined in a block with exactly one predecessor, it can't be
842 // partially redundant.
843 BasicBlock *LoadBB = LI->getParent();
844 if (LoadBB->getSinglePredecessor())
847 // If the load is defined in a landing pad, it can't be partially redundant,
848 // because the edges between the invoke and the landing pad cannot have other
849 // instructions between them.
850 if (LoadBB->isLandingPad())
853 Value *LoadedPtr = LI->getOperand(0);
855 // If the loaded operand is defined in the LoadBB, it can't be available.
856 // TODO: Could do simple PHI translation, that would be fun :)
857 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
858 if (PtrOp->getParent() == LoadBB)
861 // Scan a few instructions up from the load, to see if it is obviously live at
862 // the entry to its block.
863 BasicBlock::iterator BBIt = LI;
865 if (Value *AvailableVal =
866 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
867 // If the value if the load is locally available within the block, just use
868 // it. This frequently occurs for reg2mem'd allocas.
869 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
871 // If the returned value is the load itself, replace with an undef. This can
872 // only happen in dead loops.
873 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
874 LI->replaceAllUsesWith(AvailableVal);
875 LI->eraseFromParent();
879 // Otherwise, if we scanned the whole block and got to the top of the block,
880 // we know the block is locally transparent to the load. If not, something
881 // might clobber its value.
882 if (BBIt != LoadBB->begin())
885 // If all of the loads and stores that feed the value have the same TBAA tag,
886 // then we can propagate it onto any newly inserted loads.
887 MDNode *TBAATag = LI->getMetadata(LLVMContext::MD_tbaa);
889 SmallPtrSet<BasicBlock*, 8> PredsScanned;
890 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
891 AvailablePredsTy AvailablePreds;
892 BasicBlock *OneUnavailablePred = 0;
894 // If we got here, the loaded value is transparent through to the start of the
895 // block. Check to see if it is available in any of the predecessor blocks.
896 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
898 BasicBlock *PredBB = *PI;
900 // If we already scanned this predecessor, skip it.
901 if (!PredsScanned.insert(PredBB))
904 // Scan the predecessor to see if the value is available in the pred.
905 BBIt = PredBB->end();
906 MDNode *ThisTBAATag = 0;
907 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6,
909 if (!PredAvailable) {
910 OneUnavailablePred = PredBB;
914 // If tbaa tags disagree or are not present, forget about them.
915 if (TBAATag != ThisTBAATag) TBAATag = 0;
917 // If so, this load is partially redundant. Remember this info so that we
918 // can create a PHI node.
919 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
922 // If the loaded value isn't available in any predecessor, it isn't partially
924 if (AvailablePreds.empty()) return false;
926 // Okay, the loaded value is available in at least one (and maybe all!)
927 // predecessors. If the value is unavailable in more than one unique
928 // predecessor, we want to insert a merge block for those common predecessors.
929 // This ensures that we only have to insert one reload, thus not increasing
931 BasicBlock *UnavailablePred = 0;
933 // If there is exactly one predecessor where the value is unavailable, the
934 // already computed 'OneUnavailablePred' block is it. If it ends in an
935 // unconditional branch, we know that it isn't a critical edge.
936 if (PredsScanned.size() == AvailablePreds.size()+1 &&
937 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
938 UnavailablePred = OneUnavailablePred;
939 } else if (PredsScanned.size() != AvailablePreds.size()) {
940 // Otherwise, we had multiple unavailable predecessors or we had a critical
941 // edge from the one.
942 SmallVector<BasicBlock*, 8> PredsToSplit;
943 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
945 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
946 AvailablePredSet.insert(AvailablePreds[i].first);
948 // Add all the unavailable predecessors to the PredsToSplit list.
949 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
952 // If the predecessor is an indirect goto, we can't split the edge.
953 if (isa<IndirectBrInst>(P->getTerminator()))
956 if (!AvailablePredSet.count(P))
957 PredsToSplit.push_back(P);
960 // Split them out to their own block.
962 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split", this);
965 // If the value isn't available in all predecessors, then there will be
966 // exactly one where it isn't available. Insert a load on that edge and add
967 // it to the AvailablePreds list.
968 if (UnavailablePred) {
969 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
970 "Can't handle critical edge here!");
971 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
973 UnavailablePred->getTerminator());
974 NewVal->setDebugLoc(LI->getDebugLoc());
976 NewVal->setMetadata(LLVMContext::MD_tbaa, TBAATag);
978 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
981 // Now we know that each predecessor of this block has a value in
982 // AvailablePreds, sort them for efficient access as we're walking the preds.
983 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
985 // Create a PHI node at the start of the block for the PRE'd load value.
986 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
987 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
990 PN->setDebugLoc(LI->getDebugLoc());
992 // Insert new entries into the PHI for each predecessor. A single block may
993 // have multiple entries here.
994 for (pred_iterator PI = PB; PI != PE; ++PI) {
996 AvailablePredsTy::iterator I =
997 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
998 std::make_pair(P, (Value*)0));
1000 assert(I != AvailablePreds.end() && I->first == P &&
1001 "Didn't find entry for predecessor!");
1003 PN->addIncoming(I->second, I->first);
1006 //cerr << "PRE: " << *LI << *PN << "\n";
1008 LI->replaceAllUsesWith(PN);
1009 LI->eraseFromParent();
1014 /// FindMostPopularDest - The specified list contains multiple possible
1015 /// threadable destinations. Pick the one that occurs the most frequently in
1018 FindMostPopularDest(BasicBlock *BB,
1019 const SmallVectorImpl<std::pair<BasicBlock*,
1020 BasicBlock*> > &PredToDestList) {
1021 assert(!PredToDestList.empty());
1023 // Determine popularity. If there are multiple possible destinations, we
1024 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1025 // blocks with known and real destinations to threading undef. We'll handle
1026 // them later if interesting.
1027 DenseMap<BasicBlock*, unsigned> DestPopularity;
1028 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1029 if (PredToDestList[i].second)
1030 DestPopularity[PredToDestList[i].second]++;
1032 // Find the most popular dest.
1033 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1034 BasicBlock *MostPopularDest = DPI->first;
1035 unsigned Popularity = DPI->second;
1036 SmallVector<BasicBlock*, 4> SamePopularity;
1038 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1039 // If the popularity of this entry isn't higher than the popularity we've
1040 // seen so far, ignore it.
1041 if (DPI->second < Popularity)
1043 else if (DPI->second == Popularity) {
1044 // If it is the same as what we've seen so far, keep track of it.
1045 SamePopularity.push_back(DPI->first);
1047 // If it is more popular, remember it.
1048 SamePopularity.clear();
1049 MostPopularDest = DPI->first;
1050 Popularity = DPI->second;
1054 // Okay, now we know the most popular destination. If there is more than one
1055 // destination, we need to determine one. This is arbitrary, but we need
1056 // to make a deterministic decision. Pick the first one that appears in the
1058 if (!SamePopularity.empty()) {
1059 SamePopularity.push_back(MostPopularDest);
1060 TerminatorInst *TI = BB->getTerminator();
1061 for (unsigned i = 0; ; ++i) {
1062 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1064 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1065 TI->getSuccessor(i)) == SamePopularity.end())
1068 MostPopularDest = TI->getSuccessor(i);
1073 // Okay, we have finally picked the most popular destination.
1074 return MostPopularDest;
1077 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1078 ConstantPreference Preference) {
1079 // If threading this would thread across a loop header, don't even try to
1081 if (LoopHeaders.count(BB))
1084 PredValueInfoTy PredValues;
1085 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference))
1088 assert(!PredValues.empty() &&
1089 "ComputeValueKnownInPredecessors returned true with no values");
1091 DEBUG(dbgs() << "IN BB: " << *BB;
1092 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1093 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1094 << *PredValues[i].first
1095 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1098 // Decide what we want to thread through. Convert our list of known values to
1099 // a list of known destinations for each pred. This also discards duplicate
1100 // predecessors and keeps track of the undefined inputs (which are represented
1101 // as a null dest in the PredToDestList).
1102 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1103 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1105 BasicBlock *OnlyDest = 0;
1106 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1108 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1109 BasicBlock *Pred = PredValues[i].second;
1110 if (!SeenPreds.insert(Pred))
1111 continue; // Duplicate predecessor entry.
1113 // If the predecessor ends with an indirect goto, we can't change its
1115 if (isa<IndirectBrInst>(Pred->getTerminator()))
1118 Constant *Val = PredValues[i].first;
1121 if (isa<UndefValue>(Val))
1123 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1124 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1125 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1126 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1128 assert(isa<IndirectBrInst>(BB->getTerminator())
1129 && "Unexpected terminator");
1130 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1133 // If we have exactly one destination, remember it for efficiency below.
1134 if (PredToDestList.empty())
1136 else if (OnlyDest != DestBB)
1137 OnlyDest = MultipleDestSentinel;
1139 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1142 // If all edges were unthreadable, we fail.
1143 if (PredToDestList.empty())
1146 // Determine which is the most common successor. If we have many inputs and
1147 // this block is a switch, we want to start by threading the batch that goes
1148 // to the most popular destination first. If we only know about one
1149 // threadable destination (the common case) we can avoid this.
1150 BasicBlock *MostPopularDest = OnlyDest;
1152 if (MostPopularDest == MultipleDestSentinel)
1153 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1155 // Now that we know what the most popular destination is, factor all
1156 // predecessors that will jump to it into a single predecessor.
1157 SmallVector<BasicBlock*, 16> PredsToFactor;
1158 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1159 if (PredToDestList[i].second == MostPopularDest) {
1160 BasicBlock *Pred = PredToDestList[i].first;
1162 // This predecessor may be a switch or something else that has multiple
1163 // edges to the block. Factor each of these edges by listing them
1164 // according to # occurrences in PredsToFactor.
1165 TerminatorInst *PredTI = Pred->getTerminator();
1166 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1167 if (PredTI->getSuccessor(i) == BB)
1168 PredsToFactor.push_back(Pred);
1171 // If the threadable edges are branching on an undefined value, we get to pick
1172 // the destination that these predecessors should get to.
1173 if (MostPopularDest == 0)
1174 MostPopularDest = BB->getTerminator()->
1175 getSuccessor(GetBestDestForJumpOnUndef(BB));
1177 // Ok, try to thread it!
1178 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1181 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1182 /// a PHI node in the current block. See if there are any simplifications we
1183 /// can do based on inputs to the phi node.
1185 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1186 BasicBlock *BB = PN->getParent();
1188 // TODO: We could make use of this to do it once for blocks with common PHI
1190 SmallVector<BasicBlock*, 1> PredBBs;
1193 // If any of the predecessor blocks end in an unconditional branch, we can
1194 // *duplicate* the conditional branch into that block in order to further
1195 // encourage jump threading and to eliminate cases where we have branch on a
1196 // phi of an icmp (branch on icmp is much better).
1197 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1198 BasicBlock *PredBB = PN->getIncomingBlock(i);
1199 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1200 if (PredBr->isUnconditional()) {
1201 PredBBs[0] = PredBB;
1202 // Try to duplicate BB into PredBB.
1203 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1211 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1212 /// a xor instruction in the current block. See if there are any
1213 /// simplifications we can do based on inputs to the xor.
1215 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1216 BasicBlock *BB = BO->getParent();
1218 // If either the LHS or RHS of the xor is a constant, don't do this
1220 if (isa<ConstantInt>(BO->getOperand(0)) ||
1221 isa<ConstantInt>(BO->getOperand(1)))
1224 // If the first instruction in BB isn't a phi, we won't be able to infer
1225 // anything special about any particular predecessor.
1226 if (!isa<PHINode>(BB->front()))
1229 // If we have a xor as the branch input to this block, and we know that the
1230 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1231 // the condition into the predecessor and fix that value to true, saving some
1232 // logical ops on that path and encouraging other paths to simplify.
1234 // This copies something like this:
1237 // %X = phi i1 [1], [%X']
1238 // %Y = icmp eq i32 %A, %B
1239 // %Z = xor i1 %X, %Y
1244 // %Y = icmp ne i32 %A, %B
1247 PredValueInfoTy XorOpValues;
1249 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1251 assert(XorOpValues.empty());
1252 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1258 assert(!XorOpValues.empty() &&
1259 "ComputeValueKnownInPredecessors returned true with no values");
1261 // Scan the information to see which is most popular: true or false. The
1262 // predecessors can be of the set true, false, or undef.
1263 unsigned NumTrue = 0, NumFalse = 0;
1264 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1265 if (isa<UndefValue>(XorOpValues[i].first))
1266 // Ignore undefs for the count.
1268 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1274 // Determine which value to split on, true, false, or undef if neither.
1275 ConstantInt *SplitVal = 0;
1276 if (NumTrue > NumFalse)
1277 SplitVal = ConstantInt::getTrue(BB->getContext());
1278 else if (NumTrue != 0 || NumFalse != 0)
1279 SplitVal = ConstantInt::getFalse(BB->getContext());
1281 // Collect all of the blocks that this can be folded into so that we can
1282 // factor this once and clone it once.
1283 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1284 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1285 if (XorOpValues[i].first != SplitVal &&
1286 !isa<UndefValue>(XorOpValues[i].first))
1289 BlocksToFoldInto.push_back(XorOpValues[i].second);
1292 // If we inferred a value for all of the predecessors, then duplication won't
1293 // help us. However, we can just replace the LHS or RHS with the constant.
1294 if (BlocksToFoldInto.size() ==
1295 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1296 if (SplitVal == 0) {
1297 // If all preds provide undef, just nuke the xor, because it is undef too.
1298 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1299 BO->eraseFromParent();
1300 } else if (SplitVal->isZero()) {
1301 // If all preds provide 0, replace the xor with the other input.
1302 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1303 BO->eraseFromParent();
1305 // If all preds provide 1, set the computed value to 1.
1306 BO->setOperand(!isLHS, SplitVal);
1312 // Try to duplicate BB into PredBB.
1313 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1317 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1318 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1319 /// NewPred using the entries from OldPred (suitably mapped).
1320 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1321 BasicBlock *OldPred,
1322 BasicBlock *NewPred,
1323 DenseMap<Instruction*, Value*> &ValueMap) {
1324 for (BasicBlock::iterator PNI = PHIBB->begin();
1325 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1326 // Ok, we have a PHI node. Figure out what the incoming value was for the
1328 Value *IV = PN->getIncomingValueForBlock(OldPred);
1330 // Remap the value if necessary.
1331 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1332 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1333 if (I != ValueMap.end())
1337 PN->addIncoming(IV, NewPred);
1341 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1342 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1343 /// across BB. Transform the IR to reflect this change.
1344 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1345 const SmallVectorImpl<BasicBlock*> &PredBBs,
1346 BasicBlock *SuccBB) {
1347 // If threading to the same block as we come from, we would infinite loop.
1349 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1350 << "' - would thread to self!\n");
1354 // If threading this would thread across a loop header, don't thread the edge.
1355 // See the comments above FindLoopHeaders for justifications and caveats.
1356 if (LoopHeaders.count(BB)) {
1357 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1358 << "' to dest BB '" << SuccBB->getName()
1359 << "' - it might create an irreducible loop!\n");
1363 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, Threshold);
1364 if (JumpThreadCost > Threshold) {
1365 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1366 << "' - Cost is too high: " << JumpThreadCost << "\n");
1370 // And finally, do it! Start by factoring the predecessors is needed.
1372 if (PredBBs.size() == 1)
1373 PredBB = PredBBs[0];
1375 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1376 << " common predecessors.\n");
1377 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1380 // And finally, do it!
1381 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1382 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1383 << ", across block:\n "
1386 LVI->threadEdge(PredBB, BB, SuccBB);
1388 // We are going to have to map operands from the original BB block to the new
1389 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1390 // account for entry from PredBB.
1391 DenseMap<Instruction*, Value*> ValueMapping;
1393 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1394 BB->getName()+".thread",
1395 BB->getParent(), BB);
1396 NewBB->moveAfter(PredBB);
1398 BasicBlock::iterator BI = BB->begin();
1399 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1400 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1402 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1403 // mapping and using it to remap operands in the cloned instructions.
1404 for (; !isa<TerminatorInst>(BI); ++BI) {
1405 Instruction *New = BI->clone();
1406 New->setName(BI->getName());
1407 NewBB->getInstList().push_back(New);
1408 ValueMapping[BI] = New;
1410 // Remap operands to patch up intra-block references.
1411 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1412 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1413 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1414 if (I != ValueMapping.end())
1415 New->setOperand(i, I->second);
1419 // We didn't copy the terminator from BB over to NewBB, because there is now
1420 // an unconditional jump to SuccBB. Insert the unconditional jump.
1421 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB);
1422 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1424 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1425 // PHI nodes for NewBB now.
1426 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1428 // If there were values defined in BB that are used outside the block, then we
1429 // now have to update all uses of the value to use either the original value,
1430 // the cloned value, or some PHI derived value. This can require arbitrary
1431 // PHI insertion, of which we are prepared to do, clean these up now.
1432 SSAUpdater SSAUpdate;
1433 SmallVector<Use*, 16> UsesToRename;
1434 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1435 // Scan all uses of this instruction to see if it is used outside of its
1436 // block, and if so, record them in UsesToRename.
1437 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1439 Instruction *User = cast<Instruction>(*UI);
1440 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1441 if (UserPN->getIncomingBlock(UI) == BB)
1443 } else if (User->getParent() == BB)
1446 UsesToRename.push_back(&UI.getUse());
1449 // If there are no uses outside the block, we're done with this instruction.
1450 if (UsesToRename.empty())
1453 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1455 // We found a use of I outside of BB. Rename all uses of I that are outside
1456 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1457 // with the two values we know.
1458 SSAUpdate.Initialize(I->getType(), I->getName());
1459 SSAUpdate.AddAvailableValue(BB, I);
1460 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1462 while (!UsesToRename.empty())
1463 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1464 DEBUG(dbgs() << "\n");
1468 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1469 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1470 // us to simplify any PHI nodes in BB.
1471 TerminatorInst *PredTerm = PredBB->getTerminator();
1472 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1473 if (PredTerm->getSuccessor(i) == BB) {
1474 BB->removePredecessor(PredBB, true);
1475 PredTerm->setSuccessor(i, NewBB);
1478 // At this point, the IR is fully up to date and consistent. Do a quick scan
1479 // over the new instructions and zap any that are constants or dead. This
1480 // frequently happens because of phi translation.
1481 SimplifyInstructionsInBlock(NewBB, TD, TLI);
1483 // Threaded an edge!
1488 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1489 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1490 /// If we can duplicate the contents of BB up into PredBB do so now, this
1491 /// improves the odds that the branch will be on an analyzable instruction like
1493 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1494 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1495 assert(!PredBBs.empty() && "Can't handle an empty set");
1497 // If BB is a loop header, then duplicating this block outside the loop would
1498 // cause us to transform this into an irreducible loop, don't do this.
1499 // See the comments above FindLoopHeaders for justifications and caveats.
1500 if (LoopHeaders.count(BB)) {
1501 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1502 << "' into predecessor block '" << PredBBs[0]->getName()
1503 << "' - it might create an irreducible loop!\n");
1507 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, Threshold);
1508 if (DuplicationCost > Threshold) {
1509 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1510 << "' - Cost is too high: " << DuplicationCost << "\n");
1514 // And finally, do it! Start by factoring the predecessors is needed.
1516 if (PredBBs.size() == 1)
1517 PredBB = PredBBs[0];
1519 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1520 << " common predecessors.\n");
1521 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1524 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1526 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1527 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1528 << DuplicationCost << " block is:" << *BB << "\n");
1530 // Unless PredBB ends with an unconditional branch, split the edge so that we
1531 // can just clone the bits from BB into the end of the new PredBB.
1532 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1534 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1535 PredBB = SplitEdge(PredBB, BB, this);
1536 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1539 // We are going to have to map operands from the original BB block into the
1540 // PredBB block. Evaluate PHI nodes in BB.
1541 DenseMap<Instruction*, Value*> ValueMapping;
1543 BasicBlock::iterator BI = BB->begin();
1544 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1545 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1547 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1548 // mapping and using it to remap operands in the cloned instructions.
1549 for (; BI != BB->end(); ++BI) {
1550 Instruction *New = BI->clone();
1552 // Remap operands to patch up intra-block references.
1553 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1554 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1555 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1556 if (I != ValueMapping.end())
1557 New->setOperand(i, I->second);
1560 // If this instruction can be simplified after the operands are updated,
1561 // just use the simplified value instead. This frequently happens due to
1563 if (Value *IV = SimplifyInstruction(New, TD)) {
1565 ValueMapping[BI] = IV;
1567 // Otherwise, insert the new instruction into the block.
1568 New->setName(BI->getName());
1569 PredBB->getInstList().insert(OldPredBranch, New);
1570 ValueMapping[BI] = New;
1574 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1575 // add entries to the PHI nodes for branch from PredBB now.
1576 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1577 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1579 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1582 // If there were values defined in BB that are used outside the block, then we
1583 // now have to update all uses of the value to use either the original value,
1584 // the cloned value, or some PHI derived value. This can require arbitrary
1585 // PHI insertion, of which we are prepared to do, clean these up now.
1586 SSAUpdater SSAUpdate;
1587 SmallVector<Use*, 16> UsesToRename;
1588 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1589 // Scan all uses of this instruction to see if it is used outside of its
1590 // block, and if so, record them in UsesToRename.
1591 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1593 Instruction *User = cast<Instruction>(*UI);
1594 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1595 if (UserPN->getIncomingBlock(UI) == BB)
1597 } else if (User->getParent() == BB)
1600 UsesToRename.push_back(&UI.getUse());
1603 // If there are no uses outside the block, we're done with this instruction.
1604 if (UsesToRename.empty())
1607 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1609 // We found a use of I outside of BB. Rename all uses of I that are outside
1610 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1611 // with the two values we know.
1612 SSAUpdate.Initialize(I->getType(), I->getName());
1613 SSAUpdate.AddAvailableValue(BB, I);
1614 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1616 while (!UsesToRename.empty())
1617 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1618 DEBUG(dbgs() << "\n");
1621 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1623 BB->removePredecessor(PredBB, true);
1625 // Remove the unconditional branch at the end of the PredBB block.
1626 OldPredBranch->eraseFromParent();
1632 /// TryToUnfoldSelect - Look for blocks of the form
1638 /// %p = phi [%a, %bb] ...
1642 /// And expand the select into a branch structure if one of its arms allows %c
1643 /// to be folded. This later enables threading from bb1 over bb2.
1644 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1645 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1646 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1647 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1649 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1650 CondLHS->getParent() != BB)
1653 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1654 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1655 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1657 // Look if one of the incoming values is a select in the corresponding
1659 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1662 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1663 if (!PredTerm || !PredTerm->isUnconditional())
1666 // Now check if one of the select values would allow us to constant fold the
1667 // terminator in BB. We don't do the transform if both sides fold, those
1668 // cases will be threaded in any case.
1669 LazyValueInfo::Tristate LHSFolds =
1670 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1672 LazyValueInfo::Tristate RHSFolds =
1673 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
1675 if ((LHSFolds != LazyValueInfo::Unknown ||
1676 RHSFolds != LazyValueInfo::Unknown) &&
1677 LHSFolds != RHSFolds) {
1678 // Expand the select.
1687 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
1688 BB->getParent(), BB);
1689 // Move the unconditional branch to NewBB.
1690 PredTerm->removeFromParent();
1691 NewBB->getInstList().insert(NewBB->end(), PredTerm);
1692 // Create a conditional branch and update PHI nodes.
1693 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
1694 CondLHS->setIncomingValue(I, SI->getFalseValue());
1695 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
1696 // The select is now dead.
1697 SI->eraseFromParent();
1699 // Update any other PHI nodes in BB.
1700 for (BasicBlock::iterator BI = BB->begin();
1701 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
1703 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);