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);
136 char JumpThreading::ID = 0;
137 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
138 "Jump Threading", false, false)
139 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)
140 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
141 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
142 "Jump Threading", false, false)
144 // Public interface to the Jump Threading pass
145 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
147 /// runOnFunction - Top level algorithm.
149 bool JumpThreading::runOnFunction(Function &F) {
150 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
151 TD = getAnalysisIfAvailable<DataLayout>();
152 TLI = &getAnalysis<TargetLibraryInfo>();
153 LVI = &getAnalysis<LazyValueInfo>();
157 bool Changed, EverChanged = false;
160 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
162 // Thread all of the branches we can over this block.
163 while (ProcessBlock(BB))
168 // If the block is trivially dead, zap it. This eliminates the successor
169 // edges which simplifies the CFG.
170 if (pred_begin(BB) == pred_end(BB) &&
171 BB != &BB->getParent()->getEntryBlock()) {
172 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
173 << "' with terminator: " << *BB->getTerminator() << '\n');
174 LoopHeaders.erase(BB);
181 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
183 // Can't thread an unconditional jump, but if the block is "almost
184 // empty", we can replace uses of it with uses of the successor and make
186 if (BI && BI->isUnconditional() &&
187 BB != &BB->getParent()->getEntryBlock() &&
188 // If the terminator is the only non-phi instruction, try to nuke it.
189 BB->getFirstNonPHIOrDbg()->isTerminator()) {
190 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
191 // block, we have to make sure it isn't in the LoopHeaders set. We
192 // reinsert afterward if needed.
193 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
194 BasicBlock *Succ = BI->getSuccessor(0);
196 // FIXME: It is always conservatively correct to drop the info
197 // for a block even if it doesn't get erased. This isn't totally
198 // awesome, but it allows us to use AssertingVH to prevent nasty
199 // dangling pointer issues within LazyValueInfo.
201 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
203 // If we deleted BB and BB was the header of a loop, then the
204 // successor is now the header of the loop.
208 if (ErasedFromLoopHeaders)
209 LoopHeaders.insert(BB);
212 EverChanged |= Changed;
219 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
220 /// thread across it. Stop scanning the block when passing the threshold.
221 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
222 unsigned Threshold) {
223 /// Ignore PHI nodes, these will be flattened when duplication happens.
224 BasicBlock::const_iterator I = BB->getFirstNonPHI();
226 // FIXME: THREADING will delete values that are just used to compute the
227 // branch, so they shouldn't count against the duplication cost.
229 // Sum up the cost of each instruction until we get to the terminator. Don't
230 // include the terminator because the copy won't include it.
232 for (; !isa<TerminatorInst>(I); ++I) {
234 // Stop scanning the block if we've reached the threshold.
235 if (Size > Threshold)
238 // Debugger intrinsics don't incur code size.
239 if (isa<DbgInfoIntrinsic>(I)) continue;
241 // If this is a pointer->pointer bitcast, it is free.
242 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
245 // All other instructions count for at least one unit.
248 // Calls are more expensive. If they are non-intrinsic calls, we model them
249 // as having cost of 4. If they are a non-vector intrinsic, we model them
250 // as having cost of 2 total, and if they are a vector intrinsic, we model
251 // them as having cost 1.
252 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
253 if (CI->hasFnAttr(Attribute::NoDuplicate))
254 // Blocks with NoDuplicate are modelled as having infinite cost, so they
255 // are never duplicated.
257 else if (!isa<IntrinsicInst>(CI))
259 else if (!CI->getType()->isVectorTy())
264 // Threading through a switch statement is particularly profitable. If this
265 // block ends in a switch, decrease its cost to make it more likely to happen.
266 if (isa<SwitchInst>(I))
267 Size = Size > 6 ? Size-6 : 0;
269 // The same holds for indirect branches, but slightly more so.
270 if (isa<IndirectBrInst>(I))
271 Size = Size > 8 ? Size-8 : 0;
276 /// FindLoopHeaders - We do not want jump threading to turn proper loop
277 /// structures into irreducible loops. Doing this breaks up the loop nesting
278 /// hierarchy and pessimizes later transformations. To prevent this from
279 /// happening, we first have to find the loop headers. Here we approximate this
280 /// by finding targets of backedges in the CFG.
282 /// Note that there definitely are cases when we want to allow threading of
283 /// edges across a loop header. For example, threading a jump from outside the
284 /// loop (the preheader) to an exit block of the loop is definitely profitable.
285 /// It is also almost always profitable to thread backedges from within the loop
286 /// to exit blocks, and is often profitable to thread backedges to other blocks
287 /// within the loop (forming a nested loop). This simple analysis is not rich
288 /// enough to track all of these properties and keep it up-to-date as the CFG
289 /// mutates, so we don't allow any of these transformations.
291 void JumpThreading::FindLoopHeaders(Function &F) {
292 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
293 FindFunctionBackedges(F, Edges);
295 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
296 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
299 /// getKnownConstant - Helper method to determine if we can thread over a
300 /// terminator with the given value as its condition, and if so what value to
301 /// use for that. What kind of value this is depends on whether we want an
302 /// integer or a block address, but an undef is always accepted.
303 /// Returns null if Val is null or not an appropriate constant.
304 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
308 // Undef is "known" enough.
309 if (UndefValue *U = dyn_cast<UndefValue>(Val))
312 if (Preference == WantBlockAddress)
313 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
315 return dyn_cast<ConstantInt>(Val);
318 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
319 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
320 /// in any of our predecessors. If so, return the known list of value and pred
321 /// BB in the result vector.
323 /// This returns true if there were any known values.
326 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
327 ConstantPreference Preference) {
328 // This method walks up use-def chains recursively. Because of this, we could
329 // get into an infinite loop going around loops in the use-def chain. To
330 // prevent this, keep track of what (value, block) pairs we've already visited
331 // and terminate the search if we loop back to them
332 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
335 // An RAII help to remove this pair from the recursion set once the recursion
336 // stack pops back out again.
337 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
339 // If V is a constant, then it is known in all predecessors.
340 if (Constant *KC = getKnownConstant(V, Preference)) {
341 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
342 Result.push_back(std::make_pair(KC, *PI));
347 // If V is a non-instruction value, or an instruction in a different block,
348 // then it can't be derived from a PHI.
349 Instruction *I = dyn_cast<Instruction>(V);
350 if (I == 0 || I->getParent() != BB) {
352 // Okay, if this is a live-in value, see if it has a known value at the end
353 // of any of our predecessors.
355 // FIXME: This should be an edge property, not a block end property.
356 /// TODO: Per PR2563, we could infer value range information about a
357 /// predecessor based on its terminator.
359 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
360 // "I" is a non-local compare-with-a-constant instruction. This would be
361 // able to handle value inequalities better, for example if the compare is
362 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
363 // Perhaps getConstantOnEdge should be smart enough to do this?
365 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
367 // If the value is known by LazyValueInfo to be a constant in a
368 // predecessor, use that information to try to thread this block.
369 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
370 if (Constant *KC = getKnownConstant(PredCst, Preference))
371 Result.push_back(std::make_pair(KC, P));
374 return !Result.empty();
377 /// If I is a PHI node, then we know the incoming values for any constants.
378 if (PHINode *PN = dyn_cast<PHINode>(I)) {
379 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
380 Value *InVal = PN->getIncomingValue(i);
381 if (Constant *KC = getKnownConstant(InVal, Preference)) {
382 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
384 Constant *CI = LVI->getConstantOnEdge(InVal,
385 PN->getIncomingBlock(i), BB);
386 if (Constant *KC = getKnownConstant(CI, Preference))
387 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
391 return !Result.empty();
394 PredValueInfoTy LHSVals, RHSVals;
396 // Handle some boolean conditions.
397 if (I->getType()->getPrimitiveSizeInBits() == 1) {
398 assert(Preference == WantInteger && "One-bit non-integer type?");
400 // X & false -> false
401 if (I->getOpcode() == Instruction::Or ||
402 I->getOpcode() == Instruction::And) {
403 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
405 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
408 if (LHSVals.empty() && RHSVals.empty())
411 ConstantInt *InterestingVal;
412 if (I->getOpcode() == Instruction::Or)
413 InterestingVal = ConstantInt::getTrue(I->getContext());
415 InterestingVal = ConstantInt::getFalse(I->getContext());
417 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
419 // Scan for the sentinel. If we find an undef, force it to the
420 // interesting value: x|undef -> true and x&undef -> false.
421 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
422 if (LHSVals[i].first == InterestingVal ||
423 isa<UndefValue>(LHSVals[i].first)) {
424 Result.push_back(LHSVals[i]);
425 Result.back().first = InterestingVal;
426 LHSKnownBBs.insert(LHSVals[i].second);
428 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
429 if (RHSVals[i].first == InterestingVal ||
430 isa<UndefValue>(RHSVals[i].first)) {
431 // If we already inferred a value for this block on the LHS, don't
433 if (!LHSKnownBBs.count(RHSVals[i].second)) {
434 Result.push_back(RHSVals[i]);
435 Result.back().first = InterestingVal;
439 return !Result.empty();
442 // Handle the NOT form of XOR.
443 if (I->getOpcode() == Instruction::Xor &&
444 isa<ConstantInt>(I->getOperand(1)) &&
445 cast<ConstantInt>(I->getOperand(1))->isOne()) {
446 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
451 // Invert the known values.
452 for (unsigned i = 0, e = Result.size(); i != e; ++i)
453 Result[i].first = ConstantExpr::getNot(Result[i].first);
458 // Try to simplify some other binary operator values.
459 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
460 assert(Preference != WantBlockAddress
461 && "A binary operator creating a block address?");
462 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
463 PredValueInfoTy LHSVals;
464 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
467 // Try to use constant folding to simplify the binary operator.
468 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
469 Constant *V = LHSVals[i].first;
470 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
472 if (Constant *KC = getKnownConstant(Folded, WantInteger))
473 Result.push_back(std::make_pair(KC, LHSVals[i].second));
477 return !Result.empty();
480 // Handle compare with phi operand, where the PHI is defined in this block.
481 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
482 assert(Preference == WantInteger && "Compares only produce integers");
483 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
484 if (PN && PN->getParent() == BB) {
485 // We can do this simplification if any comparisons fold to true or false.
487 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
488 BasicBlock *PredBB = PN->getIncomingBlock(i);
489 Value *LHS = PN->getIncomingValue(i);
490 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
492 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
494 if (!isa<Constant>(RHS))
497 LazyValueInfo::Tristate
498 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
499 cast<Constant>(RHS), PredBB, BB);
500 if (ResT == LazyValueInfo::Unknown)
502 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
505 if (Constant *KC = getKnownConstant(Res, WantInteger))
506 Result.push_back(std::make_pair(KC, PredBB));
509 return !Result.empty();
513 // If comparing a live-in value against a constant, see if we know the
514 // live-in value on any predecessors.
515 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
516 if (!isa<Instruction>(Cmp->getOperand(0)) ||
517 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
518 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
520 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
522 // If the value is known by LazyValueInfo to be a constant in a
523 // predecessor, use that information to try to thread this block.
524 LazyValueInfo::Tristate Res =
525 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
527 if (Res == LazyValueInfo::Unknown)
530 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
531 Result.push_back(std::make_pair(ResC, P));
534 return !Result.empty();
537 // Try to find a constant value for the LHS of a comparison,
538 // and evaluate it statically if we can.
539 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
540 PredValueInfoTy LHSVals;
541 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
544 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
545 Constant *V = LHSVals[i].first;
546 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
548 if (Constant *KC = getKnownConstant(Folded, WantInteger))
549 Result.push_back(std::make_pair(KC, LHSVals[i].second));
552 return !Result.empty();
557 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
558 // Handle select instructions where at least one operand is a known constant
559 // and we can figure out the condition value for any predecessor block.
560 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
561 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
562 PredValueInfoTy Conds;
563 if ((TrueVal || FalseVal) &&
564 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
566 for (unsigned i = 0, e = Conds.size(); i != e; ++i) {
567 Constant *Cond = Conds[i].first;
569 // Figure out what value to use for the condition.
571 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
573 KnownCond = CI->isOne();
575 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
576 // Either operand will do, so be sure to pick the one that's a known
578 // FIXME: Do this more cleverly if both values are known constants?
579 KnownCond = (TrueVal != 0);
582 // See if the select has a known constant value for this predecessor.
583 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
584 Result.push_back(std::make_pair(Val, Conds[i].second));
587 return !Result.empty();
591 // If all else fails, see if LVI can figure out a constant value for us.
592 Constant *CI = LVI->getConstant(V, BB);
593 if (Constant *KC = getKnownConstant(CI, Preference)) {
594 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
595 Result.push_back(std::make_pair(KC, *PI));
598 return !Result.empty();
603 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
604 /// in an undefined jump, decide which block is best to revector to.
606 /// Since we can pick an arbitrary destination, we pick the successor with the
607 /// fewest predecessors. This should reduce the in-degree of the others.
609 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
610 TerminatorInst *BBTerm = BB->getTerminator();
611 unsigned MinSucc = 0;
612 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
613 // Compute the successor with the minimum number of predecessors.
614 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
615 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
616 TestBB = BBTerm->getSuccessor(i);
617 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
618 if (NumPreds < MinNumPreds) {
620 MinNumPreds = NumPreds;
627 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
628 if (!BB->hasAddressTaken()) return false;
630 // If the block has its address taken, it may be a tree of dead constants
631 // hanging off of it. These shouldn't keep the block alive.
632 BlockAddress *BA = BlockAddress::get(BB);
633 BA->removeDeadConstantUsers();
634 return !BA->use_empty();
637 /// ProcessBlock - If there are any predecessors whose control can be threaded
638 /// through to a successor, transform them now.
639 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
640 // If the block is trivially dead, just return and let the caller nuke it.
641 // This simplifies other transformations.
642 if (pred_begin(BB) == pred_end(BB) &&
643 BB != &BB->getParent()->getEntryBlock())
646 // If this block has a single predecessor, and if that pred has a single
647 // successor, merge the blocks. This encourages recursive jump threading
648 // because now the condition in this block can be threaded through
649 // predecessors of our predecessor block.
650 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
651 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
652 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
653 // If SinglePred was a loop header, BB becomes one.
654 if (LoopHeaders.erase(SinglePred))
655 LoopHeaders.insert(BB);
657 // Remember if SinglePred was the entry block of the function. If so, we
658 // will need to move BB back to the entry position.
659 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
660 LVI->eraseBlock(SinglePred);
661 MergeBasicBlockIntoOnlyPred(BB);
663 if (isEntry && BB != &BB->getParent()->getEntryBlock())
664 BB->moveBefore(&BB->getParent()->getEntryBlock());
669 // What kind of constant we're looking for.
670 ConstantPreference Preference = WantInteger;
672 // Look to see if the terminator is a conditional branch, switch or indirect
673 // branch, if not we can't thread it.
675 Instruction *Terminator = BB->getTerminator();
676 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
677 // Can't thread an unconditional jump.
678 if (BI->isUnconditional()) return false;
679 Condition = BI->getCondition();
680 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
681 Condition = SI->getCondition();
682 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
683 // Can't thread indirect branch with no successors.
684 if (IB->getNumSuccessors() == 0) return false;
685 Condition = IB->getAddress()->stripPointerCasts();
686 Preference = WantBlockAddress;
688 return false; // Must be an invoke.
691 // Run constant folding to see if we can reduce the condition to a simple
693 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
694 Value *SimpleVal = ConstantFoldInstruction(I, TD, TLI);
696 I->replaceAllUsesWith(SimpleVal);
697 I->eraseFromParent();
698 Condition = SimpleVal;
702 // If the terminator is branching on an undef, we can pick any of the
703 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
704 if (isa<UndefValue>(Condition)) {
705 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
707 // Fold the branch/switch.
708 TerminatorInst *BBTerm = BB->getTerminator();
709 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
710 if (i == BestSucc) continue;
711 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
714 DEBUG(dbgs() << " In block '" << BB->getName()
715 << "' folding undef terminator: " << *BBTerm << '\n');
716 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
717 BBTerm->eraseFromParent();
721 // If the terminator of this block is branching on a constant, simplify the
722 // terminator to an unconditional branch. This can occur due to threading in
724 if (getKnownConstant(Condition, Preference)) {
725 DEBUG(dbgs() << " In block '" << BB->getName()
726 << "' folding terminator: " << *BB->getTerminator() << '\n');
728 ConstantFoldTerminator(BB, true);
732 Instruction *CondInst = dyn_cast<Instruction>(Condition);
734 // All the rest of our checks depend on the condition being an instruction.
736 // FIXME: Unify this with code below.
737 if (ProcessThreadableEdges(Condition, BB, Preference))
743 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
744 // For a comparison where the LHS is outside this block, it's possible
745 // that we've branched on it before. Used LVI to see if we can simplify
746 // the branch based on that.
747 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
748 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
749 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
750 if (CondBr && CondConst && CondBr->isConditional() && PI != PE &&
751 (!isa<Instruction>(CondCmp->getOperand(0)) ||
752 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
753 // For predecessor edge, determine if the comparison is true or false
754 // on that edge. If they're all true or all false, we can simplify the
756 // FIXME: We could handle mixed true/false by duplicating code.
757 LazyValueInfo::Tristate Baseline =
758 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
760 if (Baseline != LazyValueInfo::Unknown) {
761 // Check that all remaining incoming values match the first one.
763 LazyValueInfo::Tristate Ret =
764 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
765 CondCmp->getOperand(0), CondConst, *PI, BB);
766 if (Ret != Baseline) break;
769 // If we terminated early, then one of the values didn't match.
771 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
772 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
773 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
774 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
775 CondBr->eraseFromParent();
782 // Check for some cases that are worth simplifying. Right now we want to look
783 // for loads that are used by a switch or by the condition for the branch. If
784 // we see one, check to see if it's partially redundant. If so, insert a PHI
785 // which can then be used to thread the values.
787 Value *SimplifyValue = CondInst;
788 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
789 if (isa<Constant>(CondCmp->getOperand(1)))
790 SimplifyValue = CondCmp->getOperand(0);
792 // TODO: There are other places where load PRE would be profitable, such as
793 // more complex comparisons.
794 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
795 if (SimplifyPartiallyRedundantLoad(LI))
799 // Handle a variety of cases where we are branching on something derived from
800 // a PHI node in the current block. If we can prove that any predecessors
801 // compute a predictable value based on a PHI node, thread those predecessors.
803 if (ProcessThreadableEdges(CondInst, BB, Preference))
806 // If this is an otherwise-unfoldable branch on a phi node in the current
807 // block, see if we can simplify.
808 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
809 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
810 return ProcessBranchOnPHI(PN);
813 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
814 if (CondInst->getOpcode() == Instruction::Xor &&
815 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
816 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
819 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
820 // "(X == 4)", thread through this block.
826 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
827 /// load instruction, eliminate it by replacing it with a PHI node. This is an
828 /// important optimization that encourages jump threading, and needs to be run
829 /// interlaced with other jump threading tasks.
830 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
831 // Don't hack volatile/atomic loads.
832 if (!LI->isSimple()) return false;
834 // If the load is defined in a block with exactly one predecessor, it can't be
835 // partially redundant.
836 BasicBlock *LoadBB = LI->getParent();
837 if (LoadBB->getSinglePredecessor())
840 Value *LoadedPtr = LI->getOperand(0);
842 // If the loaded operand is defined in the LoadBB, it can't be available.
843 // TODO: Could do simple PHI translation, that would be fun :)
844 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
845 if (PtrOp->getParent() == LoadBB)
848 // Scan a few instructions up from the load, to see if it is obviously live at
849 // the entry to its block.
850 BasicBlock::iterator BBIt = LI;
852 if (Value *AvailableVal =
853 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
854 // If the value if the load is locally available within the block, just use
855 // it. This frequently occurs for reg2mem'd allocas.
856 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
858 // If the returned value is the load itself, replace with an undef. This can
859 // only happen in dead loops.
860 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
861 LI->replaceAllUsesWith(AvailableVal);
862 LI->eraseFromParent();
866 // Otherwise, if we scanned the whole block and got to the top of the block,
867 // we know the block is locally transparent to the load. If not, something
868 // might clobber its value.
869 if (BBIt != LoadBB->begin())
872 // If all of the loads and stores that feed the value have the same TBAA tag,
873 // then we can propagate it onto any newly inserted loads.
874 MDNode *TBAATag = LI->getMetadata(LLVMContext::MD_tbaa);
876 SmallPtrSet<BasicBlock*, 8> PredsScanned;
877 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
878 AvailablePredsTy AvailablePreds;
879 BasicBlock *OneUnavailablePred = 0;
881 // If we got here, the loaded value is transparent through to the start of the
882 // block. Check to see if it is available in any of the predecessor blocks.
883 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
885 BasicBlock *PredBB = *PI;
887 // If we already scanned this predecessor, skip it.
888 if (!PredsScanned.insert(PredBB))
891 // Scan the predecessor to see if the value is available in the pred.
892 BBIt = PredBB->end();
893 MDNode *ThisTBAATag = 0;
894 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6,
896 if (!PredAvailable) {
897 OneUnavailablePred = PredBB;
901 // If tbaa tags disagree or are not present, forget about them.
902 if (TBAATag != ThisTBAATag) TBAATag = 0;
904 // If so, this load is partially redundant. Remember this info so that we
905 // can create a PHI node.
906 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
909 // If the loaded value isn't available in any predecessor, it isn't partially
911 if (AvailablePreds.empty()) return false;
913 // Okay, the loaded value is available in at least one (and maybe all!)
914 // predecessors. If the value is unavailable in more than one unique
915 // predecessor, we want to insert a merge block for those common predecessors.
916 // This ensures that we only have to insert one reload, thus not increasing
918 BasicBlock *UnavailablePred = 0;
920 // If there is exactly one predecessor where the value is unavailable, the
921 // already computed 'OneUnavailablePred' block is it. If it ends in an
922 // unconditional branch, we know that it isn't a critical edge.
923 if (PredsScanned.size() == AvailablePreds.size()+1 &&
924 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
925 UnavailablePred = OneUnavailablePred;
926 } else if (PredsScanned.size() != AvailablePreds.size()) {
927 // Otherwise, we had multiple unavailable predecessors or we had a critical
928 // edge from the one.
929 SmallVector<BasicBlock*, 8> PredsToSplit;
930 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
932 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
933 AvailablePredSet.insert(AvailablePreds[i].first);
935 // Add all the unavailable predecessors to the PredsToSplit list.
936 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
939 // If the predecessor is an indirect goto, we can't split the edge.
940 if (isa<IndirectBrInst>(P->getTerminator()))
943 if (!AvailablePredSet.count(P))
944 PredsToSplit.push_back(P);
947 // Split them out to their own block.
949 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split", this);
952 // If the value isn't available in all predecessors, then there will be
953 // exactly one where it isn't available. Insert a load on that edge and add
954 // it to the AvailablePreds list.
955 if (UnavailablePred) {
956 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
957 "Can't handle critical edge here!");
958 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
960 UnavailablePred->getTerminator());
961 NewVal->setDebugLoc(LI->getDebugLoc());
963 NewVal->setMetadata(LLVMContext::MD_tbaa, TBAATag);
965 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
968 // Now we know that each predecessor of this block has a value in
969 // AvailablePreds, sort them for efficient access as we're walking the preds.
970 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
972 // Create a PHI node at the start of the block for the PRE'd load value.
973 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
974 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
977 PN->setDebugLoc(LI->getDebugLoc());
979 // Insert new entries into the PHI for each predecessor. A single block may
980 // have multiple entries here.
981 for (pred_iterator PI = PB; PI != PE; ++PI) {
983 AvailablePredsTy::iterator I =
984 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
985 std::make_pair(P, (Value*)0));
987 assert(I != AvailablePreds.end() && I->first == P &&
988 "Didn't find entry for predecessor!");
990 PN->addIncoming(I->second, I->first);
993 //cerr << "PRE: " << *LI << *PN << "\n";
995 LI->replaceAllUsesWith(PN);
996 LI->eraseFromParent();
1001 /// FindMostPopularDest - The specified list contains multiple possible
1002 /// threadable destinations. Pick the one that occurs the most frequently in
1005 FindMostPopularDest(BasicBlock *BB,
1006 const SmallVectorImpl<std::pair<BasicBlock*,
1007 BasicBlock*> > &PredToDestList) {
1008 assert(!PredToDestList.empty());
1010 // Determine popularity. If there are multiple possible destinations, we
1011 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1012 // blocks with known and real destinations to threading undef. We'll handle
1013 // them later if interesting.
1014 DenseMap<BasicBlock*, unsigned> DestPopularity;
1015 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1016 if (PredToDestList[i].second)
1017 DestPopularity[PredToDestList[i].second]++;
1019 // Find the most popular dest.
1020 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1021 BasicBlock *MostPopularDest = DPI->first;
1022 unsigned Popularity = DPI->second;
1023 SmallVector<BasicBlock*, 4> SamePopularity;
1025 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1026 // If the popularity of this entry isn't higher than the popularity we've
1027 // seen so far, ignore it.
1028 if (DPI->second < Popularity)
1030 else if (DPI->second == Popularity) {
1031 // If it is the same as what we've seen so far, keep track of it.
1032 SamePopularity.push_back(DPI->first);
1034 // If it is more popular, remember it.
1035 SamePopularity.clear();
1036 MostPopularDest = DPI->first;
1037 Popularity = DPI->second;
1041 // Okay, now we know the most popular destination. If there is more than one
1042 // destination, we need to determine one. This is arbitrary, but we need
1043 // to make a deterministic decision. Pick the first one that appears in the
1045 if (!SamePopularity.empty()) {
1046 SamePopularity.push_back(MostPopularDest);
1047 TerminatorInst *TI = BB->getTerminator();
1048 for (unsigned i = 0; ; ++i) {
1049 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1051 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1052 TI->getSuccessor(i)) == SamePopularity.end())
1055 MostPopularDest = TI->getSuccessor(i);
1060 // Okay, we have finally picked the most popular destination.
1061 return MostPopularDest;
1064 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1065 ConstantPreference Preference) {
1066 // If threading this would thread across a loop header, don't even try to
1068 if (LoopHeaders.count(BB))
1071 PredValueInfoTy PredValues;
1072 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference))
1075 assert(!PredValues.empty() &&
1076 "ComputeValueKnownInPredecessors returned true with no values");
1078 DEBUG(dbgs() << "IN BB: " << *BB;
1079 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1080 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1081 << *PredValues[i].first
1082 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1085 // Decide what we want to thread through. Convert our list of known values to
1086 // a list of known destinations for each pred. This also discards duplicate
1087 // predecessors and keeps track of the undefined inputs (which are represented
1088 // as a null dest in the PredToDestList).
1089 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1090 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1092 BasicBlock *OnlyDest = 0;
1093 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1095 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1096 BasicBlock *Pred = PredValues[i].second;
1097 if (!SeenPreds.insert(Pred))
1098 continue; // Duplicate predecessor entry.
1100 // If the predecessor ends with an indirect goto, we can't change its
1102 if (isa<IndirectBrInst>(Pred->getTerminator()))
1105 Constant *Val = PredValues[i].first;
1108 if (isa<UndefValue>(Val))
1110 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1111 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1112 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1113 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1115 assert(isa<IndirectBrInst>(BB->getTerminator())
1116 && "Unexpected terminator");
1117 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1120 // If we have exactly one destination, remember it for efficiency below.
1121 if (PredToDestList.empty())
1123 else if (OnlyDest != DestBB)
1124 OnlyDest = MultipleDestSentinel;
1126 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1129 // If all edges were unthreadable, we fail.
1130 if (PredToDestList.empty())
1133 // Determine which is the most common successor. If we have many inputs and
1134 // this block is a switch, we want to start by threading the batch that goes
1135 // to the most popular destination first. If we only know about one
1136 // threadable destination (the common case) we can avoid this.
1137 BasicBlock *MostPopularDest = OnlyDest;
1139 if (MostPopularDest == MultipleDestSentinel)
1140 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1142 // Now that we know what the most popular destination is, factor all
1143 // predecessors that will jump to it into a single predecessor.
1144 SmallVector<BasicBlock*, 16> PredsToFactor;
1145 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1146 if (PredToDestList[i].second == MostPopularDest) {
1147 BasicBlock *Pred = PredToDestList[i].first;
1149 // This predecessor may be a switch or something else that has multiple
1150 // edges to the block. Factor each of these edges by listing them
1151 // according to # occurrences in PredsToFactor.
1152 TerminatorInst *PredTI = Pred->getTerminator();
1153 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1154 if (PredTI->getSuccessor(i) == BB)
1155 PredsToFactor.push_back(Pred);
1158 // If the threadable edges are branching on an undefined value, we get to pick
1159 // the destination that these predecessors should get to.
1160 if (MostPopularDest == 0)
1161 MostPopularDest = BB->getTerminator()->
1162 getSuccessor(GetBestDestForJumpOnUndef(BB));
1164 // Ok, try to thread it!
1165 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1168 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1169 /// a PHI node in the current block. See if there are any simplifications we
1170 /// can do based on inputs to the phi node.
1172 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1173 BasicBlock *BB = PN->getParent();
1175 // TODO: We could make use of this to do it once for blocks with common PHI
1177 SmallVector<BasicBlock*, 1> PredBBs;
1180 // If any of the predecessor blocks end in an unconditional branch, we can
1181 // *duplicate* the conditional branch into that block in order to further
1182 // encourage jump threading and to eliminate cases where we have branch on a
1183 // phi of an icmp (branch on icmp is much better).
1184 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1185 BasicBlock *PredBB = PN->getIncomingBlock(i);
1186 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1187 if (PredBr->isUnconditional()) {
1188 PredBBs[0] = PredBB;
1189 // Try to duplicate BB into PredBB.
1190 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1198 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1199 /// a xor instruction in the current block. See if there are any
1200 /// simplifications we can do based on inputs to the xor.
1202 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1203 BasicBlock *BB = BO->getParent();
1205 // If either the LHS or RHS of the xor is a constant, don't do this
1207 if (isa<ConstantInt>(BO->getOperand(0)) ||
1208 isa<ConstantInt>(BO->getOperand(1)))
1211 // If the first instruction in BB isn't a phi, we won't be able to infer
1212 // anything special about any particular predecessor.
1213 if (!isa<PHINode>(BB->front()))
1216 // If we have a xor as the branch input to this block, and we know that the
1217 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1218 // the condition into the predecessor and fix that value to true, saving some
1219 // logical ops on that path and encouraging other paths to simplify.
1221 // This copies something like this:
1224 // %X = phi i1 [1], [%X']
1225 // %Y = icmp eq i32 %A, %B
1226 // %Z = xor i1 %X, %Y
1231 // %Y = icmp ne i32 %A, %B
1234 PredValueInfoTy XorOpValues;
1236 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1238 assert(XorOpValues.empty());
1239 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1245 assert(!XorOpValues.empty() &&
1246 "ComputeValueKnownInPredecessors returned true with no values");
1248 // Scan the information to see which is most popular: true or false. The
1249 // predecessors can be of the set true, false, or undef.
1250 unsigned NumTrue = 0, NumFalse = 0;
1251 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1252 if (isa<UndefValue>(XorOpValues[i].first))
1253 // Ignore undefs for the count.
1255 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1261 // Determine which value to split on, true, false, or undef if neither.
1262 ConstantInt *SplitVal = 0;
1263 if (NumTrue > NumFalse)
1264 SplitVal = ConstantInt::getTrue(BB->getContext());
1265 else if (NumTrue != 0 || NumFalse != 0)
1266 SplitVal = ConstantInt::getFalse(BB->getContext());
1268 // Collect all of the blocks that this can be folded into so that we can
1269 // factor this once and clone it once.
1270 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1271 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1272 if (XorOpValues[i].first != SplitVal &&
1273 !isa<UndefValue>(XorOpValues[i].first))
1276 BlocksToFoldInto.push_back(XorOpValues[i].second);
1279 // If we inferred a value for all of the predecessors, then duplication won't
1280 // help us. However, we can just replace the LHS or RHS with the constant.
1281 if (BlocksToFoldInto.size() ==
1282 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1283 if (SplitVal == 0) {
1284 // If all preds provide undef, just nuke the xor, because it is undef too.
1285 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1286 BO->eraseFromParent();
1287 } else if (SplitVal->isZero()) {
1288 // If all preds provide 0, replace the xor with the other input.
1289 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1290 BO->eraseFromParent();
1292 // If all preds provide 1, set the computed value to 1.
1293 BO->setOperand(!isLHS, SplitVal);
1299 // Try to duplicate BB into PredBB.
1300 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1304 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1305 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1306 /// NewPred using the entries from OldPred (suitably mapped).
1307 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1308 BasicBlock *OldPred,
1309 BasicBlock *NewPred,
1310 DenseMap<Instruction*, Value*> &ValueMap) {
1311 for (BasicBlock::iterator PNI = PHIBB->begin();
1312 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1313 // Ok, we have a PHI node. Figure out what the incoming value was for the
1315 Value *IV = PN->getIncomingValueForBlock(OldPred);
1317 // Remap the value if necessary.
1318 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1319 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1320 if (I != ValueMap.end())
1324 PN->addIncoming(IV, NewPred);
1328 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1329 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1330 /// across BB. Transform the IR to reflect this change.
1331 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1332 const SmallVectorImpl<BasicBlock*> &PredBBs,
1333 BasicBlock *SuccBB) {
1334 // If threading to the same block as we come from, we would infinite loop.
1336 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1337 << "' - would thread to self!\n");
1341 // If threading this would thread across a loop header, don't thread the edge.
1342 // See the comments above FindLoopHeaders for justifications and caveats.
1343 if (LoopHeaders.count(BB)) {
1344 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1345 << "' to dest BB '" << SuccBB->getName()
1346 << "' - it might create an irreducible loop!\n");
1350 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, Threshold);
1351 if (JumpThreadCost > Threshold) {
1352 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1353 << "' - Cost is too high: " << JumpThreadCost << "\n");
1357 // And finally, do it! Start by factoring the predecessors is needed.
1359 if (PredBBs.size() == 1)
1360 PredBB = PredBBs[0];
1362 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1363 << " common predecessors.\n");
1364 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1367 // And finally, do it!
1368 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1369 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1370 << ", across block:\n "
1373 LVI->threadEdge(PredBB, BB, SuccBB);
1375 // We are going to have to map operands from the original BB block to the new
1376 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1377 // account for entry from PredBB.
1378 DenseMap<Instruction*, Value*> ValueMapping;
1380 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1381 BB->getName()+".thread",
1382 BB->getParent(), BB);
1383 NewBB->moveAfter(PredBB);
1385 BasicBlock::iterator BI = BB->begin();
1386 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1387 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1389 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1390 // mapping and using it to remap operands in the cloned instructions.
1391 for (; !isa<TerminatorInst>(BI); ++BI) {
1392 Instruction *New = BI->clone();
1393 New->setName(BI->getName());
1394 NewBB->getInstList().push_back(New);
1395 ValueMapping[BI] = New;
1397 // Remap operands to patch up intra-block references.
1398 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1399 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1400 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1401 if (I != ValueMapping.end())
1402 New->setOperand(i, I->second);
1406 // We didn't copy the terminator from BB over to NewBB, because there is now
1407 // an unconditional jump to SuccBB. Insert the unconditional jump.
1408 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB);
1409 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1411 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1412 // PHI nodes for NewBB now.
1413 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1415 // If there were values defined in BB that are used outside the block, then we
1416 // now have to update all uses of the value to use either the original value,
1417 // the cloned value, or some PHI derived value. This can require arbitrary
1418 // PHI insertion, of which we are prepared to do, clean these up now.
1419 SSAUpdater SSAUpdate;
1420 SmallVector<Use*, 16> UsesToRename;
1421 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1422 // Scan all uses of this instruction to see if it is used outside of its
1423 // block, and if so, record them in UsesToRename.
1424 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1426 Instruction *User = cast<Instruction>(*UI);
1427 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1428 if (UserPN->getIncomingBlock(UI) == BB)
1430 } else if (User->getParent() == BB)
1433 UsesToRename.push_back(&UI.getUse());
1436 // If there are no uses outside the block, we're done with this instruction.
1437 if (UsesToRename.empty())
1440 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1442 // We found a use of I outside of BB. Rename all uses of I that are outside
1443 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1444 // with the two values we know.
1445 SSAUpdate.Initialize(I->getType(), I->getName());
1446 SSAUpdate.AddAvailableValue(BB, I);
1447 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1449 while (!UsesToRename.empty())
1450 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1451 DEBUG(dbgs() << "\n");
1455 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1456 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1457 // us to simplify any PHI nodes in BB.
1458 TerminatorInst *PredTerm = PredBB->getTerminator();
1459 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1460 if (PredTerm->getSuccessor(i) == BB) {
1461 BB->removePredecessor(PredBB, true);
1462 PredTerm->setSuccessor(i, NewBB);
1465 // At this point, the IR is fully up to date and consistent. Do a quick scan
1466 // over the new instructions and zap any that are constants or dead. This
1467 // frequently happens because of phi translation.
1468 SimplifyInstructionsInBlock(NewBB, TD, TLI);
1470 // Threaded an edge!
1475 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1476 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1477 /// If we can duplicate the contents of BB up into PredBB do so now, this
1478 /// improves the odds that the branch will be on an analyzable instruction like
1480 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1481 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1482 assert(!PredBBs.empty() && "Can't handle an empty set");
1484 // If BB is a loop header, then duplicating this block outside the loop would
1485 // cause us to transform this into an irreducible loop, don't do this.
1486 // See the comments above FindLoopHeaders for justifications and caveats.
1487 if (LoopHeaders.count(BB)) {
1488 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1489 << "' into predecessor block '" << PredBBs[0]->getName()
1490 << "' - it might create an irreducible loop!\n");
1494 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, Threshold);
1495 if (DuplicationCost > Threshold) {
1496 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1497 << "' - Cost is too high: " << DuplicationCost << "\n");
1501 // And finally, do it! Start by factoring the predecessors is needed.
1503 if (PredBBs.size() == 1)
1504 PredBB = PredBBs[0];
1506 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1507 << " common predecessors.\n");
1508 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1511 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1513 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1514 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1515 << DuplicationCost << " block is:" << *BB << "\n");
1517 // Unless PredBB ends with an unconditional branch, split the edge so that we
1518 // can just clone the bits from BB into the end of the new PredBB.
1519 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1521 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1522 PredBB = SplitEdge(PredBB, BB, this);
1523 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1526 // We are going to have to map operands from the original BB block into the
1527 // PredBB block. Evaluate PHI nodes in BB.
1528 DenseMap<Instruction*, Value*> ValueMapping;
1530 BasicBlock::iterator BI = BB->begin();
1531 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1532 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1534 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1535 // mapping and using it to remap operands in the cloned instructions.
1536 for (; BI != BB->end(); ++BI) {
1537 Instruction *New = BI->clone();
1539 // Remap operands to patch up intra-block references.
1540 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1541 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1542 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1543 if (I != ValueMapping.end())
1544 New->setOperand(i, I->second);
1547 // If this instruction can be simplified after the operands are updated,
1548 // just use the simplified value instead. This frequently happens due to
1550 if (Value *IV = SimplifyInstruction(New, TD)) {
1552 ValueMapping[BI] = IV;
1554 // Otherwise, insert the new instruction into the block.
1555 New->setName(BI->getName());
1556 PredBB->getInstList().insert(OldPredBranch, New);
1557 ValueMapping[BI] = New;
1561 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1562 // add entries to the PHI nodes for branch from PredBB now.
1563 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1564 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1566 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1569 // If there were values defined in BB that are used outside the block, then we
1570 // now have to update all uses of the value to use either the original value,
1571 // the cloned value, or some PHI derived value. This can require arbitrary
1572 // PHI insertion, of which we are prepared to do, clean these up now.
1573 SSAUpdater SSAUpdate;
1574 SmallVector<Use*, 16> UsesToRename;
1575 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1576 // Scan all uses of this instruction to see if it is used outside of its
1577 // block, and if so, record them in UsesToRename.
1578 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1580 Instruction *User = cast<Instruction>(*UI);
1581 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1582 if (UserPN->getIncomingBlock(UI) == BB)
1584 } else if (User->getParent() == BB)
1587 UsesToRename.push_back(&UI.getUse());
1590 // If there are no uses outside the block, we're done with this instruction.
1591 if (UsesToRename.empty())
1594 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1596 // We found a use of I outside of BB. Rename all uses of I that are outside
1597 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1598 // with the two values we know.
1599 SSAUpdate.Initialize(I->getType(), I->getName());
1600 SSAUpdate.AddAvailableValue(BB, I);
1601 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1603 while (!UsesToRename.empty())
1604 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1605 DEBUG(dbgs() << "\n");
1608 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1610 BB->removePredecessor(PredBB, true);
1612 // Remove the unconditional branch at the end of the PredBB block.
1613 OldPredBranch->eraseFromParent();