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/IR/ValueHandle.h"
31 #include "llvm/Pass.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.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) override;
110 void getAnalysisUsage(AnalysisUsage &AU) const override {
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 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
156 DL = DLP ? &DLP->getDataLayout() : 0;
157 TLI = &getAnalysis<TargetLibraryInfo>();
158 LVI = &getAnalysis<LazyValueInfo>();
162 bool Changed, EverChanged = false;
165 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
167 // Thread all of the branches we can over this block.
168 while (ProcessBlock(BB))
173 // If the block is trivially dead, zap it. This eliminates the successor
174 // edges which simplifies the CFG.
175 if (pred_begin(BB) == pred_end(BB) &&
176 BB != &BB->getParent()->getEntryBlock()) {
177 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
178 << "' with terminator: " << *BB->getTerminator() << '\n');
179 LoopHeaders.erase(BB);
186 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
188 // Can't thread an unconditional jump, but if the block is "almost
189 // empty", we can replace uses of it with uses of the successor and make
191 if (BI && BI->isUnconditional() &&
192 BB != &BB->getParent()->getEntryBlock() &&
193 // If the terminator is the only non-phi instruction, try to nuke it.
194 BB->getFirstNonPHIOrDbg()->isTerminator()) {
195 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
196 // block, we have to make sure it isn't in the LoopHeaders set. We
197 // reinsert afterward if needed.
198 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
199 BasicBlock *Succ = BI->getSuccessor(0);
201 // FIXME: It is always conservatively correct to drop the info
202 // for a block even if it doesn't get erased. This isn't totally
203 // awesome, but it allows us to use AssertingVH to prevent nasty
204 // dangling pointer issues within LazyValueInfo.
206 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
208 // If we deleted BB and BB was the header of a loop, then the
209 // successor is now the header of the loop.
213 if (ErasedFromLoopHeaders)
214 LoopHeaders.insert(BB);
217 EverChanged |= Changed;
224 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
225 /// thread across it. Stop scanning the block when passing the threshold.
226 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
227 unsigned Threshold) {
228 /// Ignore PHI nodes, these will be flattened when duplication happens.
229 BasicBlock::const_iterator I = BB->getFirstNonPHI();
231 // FIXME: THREADING will delete values that are just used to compute the
232 // branch, so they shouldn't count against the duplication cost.
234 // Sum up the cost of each instruction until we get to the terminator. Don't
235 // include the terminator because the copy won't include it.
237 for (; !isa<TerminatorInst>(I); ++I) {
239 // Stop scanning the block if we've reached the threshold.
240 if (Size > Threshold)
243 // Debugger intrinsics don't incur code size.
244 if (isa<DbgInfoIntrinsic>(I)) continue;
246 // If this is a pointer->pointer bitcast, it is free.
247 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
250 // All other instructions count for at least one unit.
253 // Calls are more expensive. If they are non-intrinsic calls, we model them
254 // as having cost of 4. If they are a non-vector intrinsic, we model them
255 // as having cost of 2 total, and if they are a vector intrinsic, we model
256 // them as having cost 1.
257 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
258 if (CI->hasFnAttr(Attribute::NoDuplicate))
259 // Blocks with NoDuplicate are modelled as having infinite cost, so they
260 // are never duplicated.
262 else if (!isa<IntrinsicInst>(CI))
264 else if (!CI->getType()->isVectorTy())
269 // Threading through a switch statement is particularly profitable. If this
270 // block ends in a switch, decrease its cost to make it more likely to happen.
271 if (isa<SwitchInst>(I))
272 Size = Size > 6 ? Size-6 : 0;
274 // The same holds for indirect branches, but slightly more so.
275 if (isa<IndirectBrInst>(I))
276 Size = Size > 8 ? Size-8 : 0;
281 /// FindLoopHeaders - We do not want jump threading to turn proper loop
282 /// structures into irreducible loops. Doing this breaks up the loop nesting
283 /// hierarchy and pessimizes later transformations. To prevent this from
284 /// happening, we first have to find the loop headers. Here we approximate this
285 /// by finding targets of backedges in the CFG.
287 /// Note that there definitely are cases when we want to allow threading of
288 /// edges across a loop header. For example, threading a jump from outside the
289 /// loop (the preheader) to an exit block of the loop is definitely profitable.
290 /// It is also almost always profitable to thread backedges from within the loop
291 /// to exit blocks, and is often profitable to thread backedges to other blocks
292 /// within the loop (forming a nested loop). This simple analysis is not rich
293 /// enough to track all of these properties and keep it up-to-date as the CFG
294 /// mutates, so we don't allow any of these transformations.
296 void JumpThreading::FindLoopHeaders(Function &F) {
297 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
298 FindFunctionBackedges(F, Edges);
300 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
301 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
304 /// getKnownConstant - Helper method to determine if we can thread over a
305 /// terminator with the given value as its condition, and if so what value to
306 /// use for that. What kind of value this is depends on whether we want an
307 /// integer or a block address, but an undef is always accepted.
308 /// Returns null if Val is null or not an appropriate constant.
309 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
313 // Undef is "known" enough.
314 if (UndefValue *U = dyn_cast<UndefValue>(Val))
317 if (Preference == WantBlockAddress)
318 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
320 return dyn_cast<ConstantInt>(Val);
323 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
324 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
325 /// in any of our predecessors. If so, return the known list of value and pred
326 /// BB in the result vector.
328 /// This returns true if there were any known values.
331 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
332 ConstantPreference Preference) {
333 // This method walks up use-def chains recursively. Because of this, we could
334 // get into an infinite loop going around loops in the use-def chain. To
335 // prevent this, keep track of what (value, block) pairs we've already visited
336 // and terminate the search if we loop back to them
337 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
340 // An RAII help to remove this pair from the recursion set once the recursion
341 // stack pops back out again.
342 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
344 // If V is a constant, then it is known in all predecessors.
345 if (Constant *KC = getKnownConstant(V, Preference)) {
346 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
347 Result.push_back(std::make_pair(KC, *PI));
352 // If V is a non-instruction value, or an instruction in a different block,
353 // then it can't be derived from a PHI.
354 Instruction *I = dyn_cast<Instruction>(V);
355 if (I == 0 || I->getParent() != BB) {
357 // Okay, if this is a live-in value, see if it has a known value at the end
358 // of any of our predecessors.
360 // FIXME: This should be an edge property, not a block end property.
361 /// TODO: Per PR2563, we could infer value range information about a
362 /// predecessor based on its terminator.
364 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
365 // "I" is a non-local compare-with-a-constant instruction. This would be
366 // able to handle value inequalities better, for example if the compare is
367 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
368 // Perhaps getConstantOnEdge should be smart enough to do this?
370 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
372 // If the value is known by LazyValueInfo to be a constant in a
373 // predecessor, use that information to try to thread this block.
374 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
375 if (Constant *KC = getKnownConstant(PredCst, Preference))
376 Result.push_back(std::make_pair(KC, P));
379 return !Result.empty();
382 /// If I is a PHI node, then we know the incoming values for any constants.
383 if (PHINode *PN = dyn_cast<PHINode>(I)) {
384 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
385 Value *InVal = PN->getIncomingValue(i);
386 if (Constant *KC = getKnownConstant(InVal, Preference)) {
387 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
389 Constant *CI = LVI->getConstantOnEdge(InVal,
390 PN->getIncomingBlock(i), BB);
391 if (Constant *KC = getKnownConstant(CI, Preference))
392 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
396 return !Result.empty();
399 PredValueInfoTy LHSVals, RHSVals;
401 // Handle some boolean conditions.
402 if (I->getType()->getPrimitiveSizeInBits() == 1) {
403 assert(Preference == WantInteger && "One-bit non-integer type?");
405 // X & false -> false
406 if (I->getOpcode() == Instruction::Or ||
407 I->getOpcode() == Instruction::And) {
408 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
410 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
413 if (LHSVals.empty() && RHSVals.empty())
416 ConstantInt *InterestingVal;
417 if (I->getOpcode() == Instruction::Or)
418 InterestingVal = ConstantInt::getTrue(I->getContext());
420 InterestingVal = ConstantInt::getFalse(I->getContext());
422 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
424 // Scan for the sentinel. If we find an undef, force it to the
425 // interesting value: x|undef -> true and x&undef -> false.
426 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
427 if (LHSVals[i].first == InterestingVal ||
428 isa<UndefValue>(LHSVals[i].first)) {
429 Result.push_back(LHSVals[i]);
430 Result.back().first = InterestingVal;
431 LHSKnownBBs.insert(LHSVals[i].second);
433 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
434 if (RHSVals[i].first == InterestingVal ||
435 isa<UndefValue>(RHSVals[i].first)) {
436 // If we already inferred a value for this block on the LHS, don't
438 if (!LHSKnownBBs.count(RHSVals[i].second)) {
439 Result.push_back(RHSVals[i]);
440 Result.back().first = InterestingVal;
444 return !Result.empty();
447 // Handle the NOT form of XOR.
448 if (I->getOpcode() == Instruction::Xor &&
449 isa<ConstantInt>(I->getOperand(1)) &&
450 cast<ConstantInt>(I->getOperand(1))->isOne()) {
451 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
456 // Invert the known values.
457 for (unsigned i = 0, e = Result.size(); i != e; ++i)
458 Result[i].first = ConstantExpr::getNot(Result[i].first);
463 // Try to simplify some other binary operator values.
464 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
465 assert(Preference != WantBlockAddress
466 && "A binary operator creating a block address?");
467 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
468 PredValueInfoTy LHSVals;
469 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
472 // Try to use constant folding to simplify the binary operator.
473 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
474 Constant *V = LHSVals[i].first;
475 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
477 if (Constant *KC = getKnownConstant(Folded, WantInteger))
478 Result.push_back(std::make_pair(KC, LHSVals[i].second));
482 return !Result.empty();
485 // Handle compare with phi operand, where the PHI is defined in this block.
486 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
487 assert(Preference == WantInteger && "Compares only produce integers");
488 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
489 if (PN && PN->getParent() == BB) {
490 // We can do this simplification if any comparisons fold to true or false.
492 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
493 BasicBlock *PredBB = PN->getIncomingBlock(i);
494 Value *LHS = PN->getIncomingValue(i);
495 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
497 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL);
499 if (!isa<Constant>(RHS))
502 LazyValueInfo::Tristate
503 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
504 cast<Constant>(RHS), PredBB, BB);
505 if (ResT == LazyValueInfo::Unknown)
507 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
510 if (Constant *KC = getKnownConstant(Res, WantInteger))
511 Result.push_back(std::make_pair(KC, PredBB));
514 return !Result.empty();
518 // If comparing a live-in value against a constant, see if we know the
519 // live-in value on any predecessors.
520 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
521 if (!isa<Instruction>(Cmp->getOperand(0)) ||
522 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
523 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
525 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
527 // If the value is known by LazyValueInfo to be a constant in a
528 // predecessor, use that information to try to thread this block.
529 LazyValueInfo::Tristate Res =
530 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
532 if (Res == LazyValueInfo::Unknown)
535 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
536 Result.push_back(std::make_pair(ResC, P));
539 return !Result.empty();
542 // Try to find a constant value for the LHS of a comparison,
543 // and evaluate it statically if we can.
544 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
545 PredValueInfoTy LHSVals;
546 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
549 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
550 Constant *V = LHSVals[i].first;
551 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
553 if (Constant *KC = getKnownConstant(Folded, WantInteger))
554 Result.push_back(std::make_pair(KC, LHSVals[i].second));
557 return !Result.empty();
562 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
563 // Handle select instructions where at least one operand is a known constant
564 // and we can figure out the condition value for any predecessor block.
565 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
566 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
567 PredValueInfoTy Conds;
568 if ((TrueVal || FalseVal) &&
569 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
571 for (unsigned i = 0, e = Conds.size(); i != e; ++i) {
572 Constant *Cond = Conds[i].first;
574 // Figure out what value to use for the condition.
576 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
578 KnownCond = CI->isOne();
580 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
581 // Either operand will do, so be sure to pick the one that's a known
583 // FIXME: Do this more cleverly if both values are known constants?
584 KnownCond = (TrueVal != 0);
587 // See if the select has a known constant value for this predecessor.
588 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
589 Result.push_back(std::make_pair(Val, Conds[i].second));
592 return !Result.empty();
596 // If all else fails, see if LVI can figure out a constant value for us.
597 Constant *CI = LVI->getConstant(V, BB);
598 if (Constant *KC = getKnownConstant(CI, Preference)) {
599 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
600 Result.push_back(std::make_pair(KC, *PI));
603 return !Result.empty();
608 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
609 /// in an undefined jump, decide which block is best to revector to.
611 /// Since we can pick an arbitrary destination, we pick the successor with the
612 /// fewest predecessors. This should reduce the in-degree of the others.
614 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
615 TerminatorInst *BBTerm = BB->getTerminator();
616 unsigned MinSucc = 0;
617 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
618 // Compute the successor with the minimum number of predecessors.
619 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
620 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
621 TestBB = BBTerm->getSuccessor(i);
622 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
623 if (NumPreds < MinNumPreds) {
625 MinNumPreds = NumPreds;
632 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
633 if (!BB->hasAddressTaken()) return false;
635 // If the block has its address taken, it may be a tree of dead constants
636 // hanging off of it. These shouldn't keep the block alive.
637 BlockAddress *BA = BlockAddress::get(BB);
638 BA->removeDeadConstantUsers();
639 return !BA->use_empty();
642 /// ProcessBlock - If there are any predecessors whose control can be threaded
643 /// through to a successor, transform them now.
644 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
645 // If the block is trivially dead, just return and let the caller nuke it.
646 // This simplifies other transformations.
647 if (pred_begin(BB) == pred_end(BB) &&
648 BB != &BB->getParent()->getEntryBlock())
651 // If this block has a single predecessor, and if that pred has a single
652 // successor, merge the blocks. This encourages recursive jump threading
653 // because now the condition in this block can be threaded through
654 // predecessors of our predecessor block.
655 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
656 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
657 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
658 // If SinglePred was a loop header, BB becomes one.
659 if (LoopHeaders.erase(SinglePred))
660 LoopHeaders.insert(BB);
662 // Remember if SinglePred was the entry block of the function. If so, we
663 // will need to move BB back to the entry position.
664 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
665 LVI->eraseBlock(SinglePred);
666 MergeBasicBlockIntoOnlyPred(BB);
668 if (isEntry && BB != &BB->getParent()->getEntryBlock())
669 BB->moveBefore(&BB->getParent()->getEntryBlock());
674 // What kind of constant we're looking for.
675 ConstantPreference Preference = WantInteger;
677 // Look to see if the terminator is a conditional branch, switch or indirect
678 // branch, if not we can't thread it.
680 Instruction *Terminator = BB->getTerminator();
681 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
682 // Can't thread an unconditional jump.
683 if (BI->isUnconditional()) return false;
684 Condition = BI->getCondition();
685 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
686 Condition = SI->getCondition();
687 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
688 // Can't thread indirect branch with no successors.
689 if (IB->getNumSuccessors() == 0) return false;
690 Condition = IB->getAddress()->stripPointerCasts();
691 Preference = WantBlockAddress;
693 return false; // Must be an invoke.
696 // Run constant folding to see if we can reduce the condition to a simple
698 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
699 Value *SimpleVal = ConstantFoldInstruction(I, DL, TLI);
701 I->replaceAllUsesWith(SimpleVal);
702 I->eraseFromParent();
703 Condition = SimpleVal;
707 // If the terminator is branching on an undef, we can pick any of the
708 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
709 if (isa<UndefValue>(Condition)) {
710 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
712 // Fold the branch/switch.
713 TerminatorInst *BBTerm = BB->getTerminator();
714 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
715 if (i == BestSucc) continue;
716 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
719 DEBUG(dbgs() << " In block '" << BB->getName()
720 << "' folding undef terminator: " << *BBTerm << '\n');
721 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
722 BBTerm->eraseFromParent();
726 // If the terminator of this block is branching on a constant, simplify the
727 // terminator to an unconditional branch. This can occur due to threading in
729 if (getKnownConstant(Condition, Preference)) {
730 DEBUG(dbgs() << " In block '" << BB->getName()
731 << "' folding terminator: " << *BB->getTerminator() << '\n');
733 ConstantFoldTerminator(BB, true);
737 Instruction *CondInst = dyn_cast<Instruction>(Condition);
739 // All the rest of our checks depend on the condition being an instruction.
741 // FIXME: Unify this with code below.
742 if (ProcessThreadableEdges(Condition, BB, Preference))
748 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
749 // For a comparison where the LHS is outside this block, it's possible
750 // that we've branched on it before. Used LVI to see if we can simplify
751 // the branch based on that.
752 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
753 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
754 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
755 if (CondBr && CondConst && CondBr->isConditional() && PI != PE &&
756 (!isa<Instruction>(CondCmp->getOperand(0)) ||
757 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
758 // For predecessor edge, determine if the comparison is true or false
759 // on that edge. If they're all true or all false, we can simplify the
761 // FIXME: We could handle mixed true/false by duplicating code.
762 LazyValueInfo::Tristate Baseline =
763 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
765 if (Baseline != LazyValueInfo::Unknown) {
766 // Check that all remaining incoming values match the first one.
768 LazyValueInfo::Tristate Ret =
769 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
770 CondCmp->getOperand(0), CondConst, *PI, BB);
771 if (Ret != Baseline) break;
774 // If we terminated early, then one of the values didn't match.
776 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
777 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
778 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
779 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
780 CondBr->eraseFromParent();
787 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB))
791 // Check for some cases that are worth simplifying. Right now we want to look
792 // for loads that are used by a switch or by the condition for the branch. If
793 // we see one, check to see if it's partially redundant. If so, insert a PHI
794 // which can then be used to thread the values.
796 Value *SimplifyValue = CondInst;
797 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
798 if (isa<Constant>(CondCmp->getOperand(1)))
799 SimplifyValue = CondCmp->getOperand(0);
801 // TODO: There are other places where load PRE would be profitable, such as
802 // more complex comparisons.
803 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
804 if (SimplifyPartiallyRedundantLoad(LI))
808 // Handle a variety of cases where we are branching on something derived from
809 // a PHI node in the current block. If we can prove that any predecessors
810 // compute a predictable value based on a PHI node, thread those predecessors.
812 if (ProcessThreadableEdges(CondInst, BB, Preference))
815 // If this is an otherwise-unfoldable branch on a phi node in the current
816 // block, see if we can simplify.
817 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
818 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
819 return ProcessBranchOnPHI(PN);
822 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
823 if (CondInst->getOpcode() == Instruction::Xor &&
824 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
825 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
828 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
829 // "(X == 4)", thread through this block.
834 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
835 /// load instruction, eliminate it by replacing it with a PHI node. This is an
836 /// important optimization that encourages jump threading, and needs to be run
837 /// interlaced with other jump threading tasks.
838 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
839 // Don't hack volatile/atomic loads.
840 if (!LI->isSimple()) return false;
842 // If the load is defined in a block with exactly one predecessor, it can't be
843 // partially redundant.
844 BasicBlock *LoadBB = LI->getParent();
845 if (LoadBB->getSinglePredecessor())
848 // If the load is defined in a landing pad, it can't be partially redundant,
849 // because the edges between the invoke and the landing pad cannot have other
850 // instructions between them.
851 if (LoadBB->isLandingPad())
854 Value *LoadedPtr = LI->getOperand(0);
856 // If the loaded operand is defined in the LoadBB, it can't be available.
857 // TODO: Could do simple PHI translation, that would be fun :)
858 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
859 if (PtrOp->getParent() == LoadBB)
862 // Scan a few instructions up from the load, to see if it is obviously live at
863 // the entry to its block.
864 BasicBlock::iterator BBIt = LI;
866 if (Value *AvailableVal =
867 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
868 // If the value if the load is locally available within the block, just use
869 // it. This frequently occurs for reg2mem'd allocas.
870 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
872 // If the returned value is the load itself, replace with an undef. This can
873 // only happen in dead loops.
874 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
875 LI->replaceAllUsesWith(AvailableVal);
876 LI->eraseFromParent();
880 // Otherwise, if we scanned the whole block and got to the top of the block,
881 // we know the block is locally transparent to the load. If not, something
882 // might clobber its value.
883 if (BBIt != LoadBB->begin())
886 // If all of the loads and stores that feed the value have the same TBAA tag,
887 // then we can propagate it onto any newly inserted loads.
888 MDNode *TBAATag = LI->getMetadata(LLVMContext::MD_tbaa);
890 SmallPtrSet<BasicBlock*, 8> PredsScanned;
891 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
892 AvailablePredsTy AvailablePreds;
893 BasicBlock *OneUnavailablePred = 0;
895 // If we got here, the loaded value is transparent through to the start of the
896 // block. Check to see if it is available in any of the predecessor blocks.
897 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
899 BasicBlock *PredBB = *PI;
901 // If we already scanned this predecessor, skip it.
902 if (!PredsScanned.insert(PredBB))
905 // Scan the predecessor to see if the value is available in the pred.
906 BBIt = PredBB->end();
907 MDNode *ThisTBAATag = 0;
908 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6,
910 if (!PredAvailable) {
911 OneUnavailablePred = PredBB;
915 // If tbaa tags disagree or are not present, forget about them.
916 if (TBAATag != ThisTBAATag) TBAATag = 0;
918 // If so, this load is partially redundant. Remember this info so that we
919 // can create a PHI node.
920 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
923 // If the loaded value isn't available in any predecessor, it isn't partially
925 if (AvailablePreds.empty()) return false;
927 // Okay, the loaded value is available in at least one (and maybe all!)
928 // predecessors. If the value is unavailable in more than one unique
929 // predecessor, we want to insert a merge block for those common predecessors.
930 // This ensures that we only have to insert one reload, thus not increasing
932 BasicBlock *UnavailablePred = 0;
934 // If there is exactly one predecessor where the value is unavailable, the
935 // already computed 'OneUnavailablePred' block is it. If it ends in an
936 // unconditional branch, we know that it isn't a critical edge.
937 if (PredsScanned.size() == AvailablePreds.size()+1 &&
938 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
939 UnavailablePred = OneUnavailablePred;
940 } else if (PredsScanned.size() != AvailablePreds.size()) {
941 // Otherwise, we had multiple unavailable predecessors or we had a critical
942 // edge from the one.
943 SmallVector<BasicBlock*, 8> PredsToSplit;
944 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
946 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
947 AvailablePredSet.insert(AvailablePreds[i].first);
949 // Add all the unavailable predecessors to the PredsToSplit list.
950 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
953 // If the predecessor is an indirect goto, we can't split the edge.
954 if (isa<IndirectBrInst>(P->getTerminator()))
957 if (!AvailablePredSet.count(P))
958 PredsToSplit.push_back(P);
961 // Split them out to their own block.
963 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split", this);
966 // If the value isn't available in all predecessors, then there will be
967 // exactly one where it isn't available. Insert a load on that edge and add
968 // it to the AvailablePreds list.
969 if (UnavailablePred) {
970 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
971 "Can't handle critical edge here!");
972 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
974 UnavailablePred->getTerminator());
975 NewVal->setDebugLoc(LI->getDebugLoc());
977 NewVal->setMetadata(LLVMContext::MD_tbaa, TBAATag);
979 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
982 // Now we know that each predecessor of this block has a value in
983 // AvailablePreds, sort them for efficient access as we're walking the preds.
984 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
986 // Create a PHI node at the start of the block for the PRE'd load value.
987 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
988 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
991 PN->setDebugLoc(LI->getDebugLoc());
993 // Insert new entries into the PHI for each predecessor. A single block may
994 // have multiple entries here.
995 for (pred_iterator PI = PB; PI != PE; ++PI) {
997 AvailablePredsTy::iterator I =
998 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
999 std::make_pair(P, (Value*)0));
1001 assert(I != AvailablePreds.end() && I->first == P &&
1002 "Didn't find entry for predecessor!");
1004 PN->addIncoming(I->second, I->first);
1007 //cerr << "PRE: " << *LI << *PN << "\n";
1009 LI->replaceAllUsesWith(PN);
1010 LI->eraseFromParent();
1015 /// FindMostPopularDest - The specified list contains multiple possible
1016 /// threadable destinations. Pick the one that occurs the most frequently in
1019 FindMostPopularDest(BasicBlock *BB,
1020 const SmallVectorImpl<std::pair<BasicBlock*,
1021 BasicBlock*> > &PredToDestList) {
1022 assert(!PredToDestList.empty());
1024 // Determine popularity. If there are multiple possible destinations, we
1025 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1026 // blocks with known and real destinations to threading undef. We'll handle
1027 // them later if interesting.
1028 DenseMap<BasicBlock*, unsigned> DestPopularity;
1029 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1030 if (PredToDestList[i].second)
1031 DestPopularity[PredToDestList[i].second]++;
1033 // Find the most popular dest.
1034 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1035 BasicBlock *MostPopularDest = DPI->first;
1036 unsigned Popularity = DPI->second;
1037 SmallVector<BasicBlock*, 4> SamePopularity;
1039 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1040 // If the popularity of this entry isn't higher than the popularity we've
1041 // seen so far, ignore it.
1042 if (DPI->second < Popularity)
1044 else if (DPI->second == Popularity) {
1045 // If it is the same as what we've seen so far, keep track of it.
1046 SamePopularity.push_back(DPI->first);
1048 // If it is more popular, remember it.
1049 SamePopularity.clear();
1050 MostPopularDest = DPI->first;
1051 Popularity = DPI->second;
1055 // Okay, now we know the most popular destination. If there is more than one
1056 // destination, we need to determine one. This is arbitrary, but we need
1057 // to make a deterministic decision. Pick the first one that appears in the
1059 if (!SamePopularity.empty()) {
1060 SamePopularity.push_back(MostPopularDest);
1061 TerminatorInst *TI = BB->getTerminator();
1062 for (unsigned i = 0; ; ++i) {
1063 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1065 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1066 TI->getSuccessor(i)) == SamePopularity.end())
1069 MostPopularDest = TI->getSuccessor(i);
1074 // Okay, we have finally picked the most popular destination.
1075 return MostPopularDest;
1078 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1079 ConstantPreference Preference) {
1080 // If threading this would thread across a loop header, don't even try to
1082 if (LoopHeaders.count(BB))
1085 PredValueInfoTy PredValues;
1086 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference))
1089 assert(!PredValues.empty() &&
1090 "ComputeValueKnownInPredecessors returned true with no values");
1092 DEBUG(dbgs() << "IN BB: " << *BB;
1093 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1094 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1095 << *PredValues[i].first
1096 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1099 // Decide what we want to thread through. Convert our list of known values to
1100 // a list of known destinations for each pred. This also discards duplicate
1101 // predecessors and keeps track of the undefined inputs (which are represented
1102 // as a null dest in the PredToDestList).
1103 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1104 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1106 BasicBlock *OnlyDest = 0;
1107 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1109 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1110 BasicBlock *Pred = PredValues[i].second;
1111 if (!SeenPreds.insert(Pred))
1112 continue; // Duplicate predecessor entry.
1114 // If the predecessor ends with an indirect goto, we can't change its
1116 if (isa<IndirectBrInst>(Pred->getTerminator()))
1119 Constant *Val = PredValues[i].first;
1122 if (isa<UndefValue>(Val))
1124 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1125 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1126 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1127 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1129 assert(isa<IndirectBrInst>(BB->getTerminator())
1130 && "Unexpected terminator");
1131 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1134 // If we have exactly one destination, remember it for efficiency below.
1135 if (PredToDestList.empty())
1137 else if (OnlyDest != DestBB)
1138 OnlyDest = MultipleDestSentinel;
1140 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1143 // If all edges were unthreadable, we fail.
1144 if (PredToDestList.empty())
1147 // Determine which is the most common successor. If we have many inputs and
1148 // this block is a switch, we want to start by threading the batch that goes
1149 // to the most popular destination first. If we only know about one
1150 // threadable destination (the common case) we can avoid this.
1151 BasicBlock *MostPopularDest = OnlyDest;
1153 if (MostPopularDest == MultipleDestSentinel)
1154 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1156 // Now that we know what the most popular destination is, factor all
1157 // predecessors that will jump to it into a single predecessor.
1158 SmallVector<BasicBlock*, 16> PredsToFactor;
1159 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1160 if (PredToDestList[i].second == MostPopularDest) {
1161 BasicBlock *Pred = PredToDestList[i].first;
1163 // This predecessor may be a switch or something else that has multiple
1164 // edges to the block. Factor each of these edges by listing them
1165 // according to # occurrences in PredsToFactor.
1166 TerminatorInst *PredTI = Pred->getTerminator();
1167 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1168 if (PredTI->getSuccessor(i) == BB)
1169 PredsToFactor.push_back(Pred);
1172 // If the threadable edges are branching on an undefined value, we get to pick
1173 // the destination that these predecessors should get to.
1174 if (MostPopularDest == 0)
1175 MostPopularDest = BB->getTerminator()->
1176 getSuccessor(GetBestDestForJumpOnUndef(BB));
1178 // Ok, try to thread it!
1179 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1182 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1183 /// a PHI node in the current block. See if there are any simplifications we
1184 /// can do based on inputs to the phi node.
1186 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1187 BasicBlock *BB = PN->getParent();
1189 // TODO: We could make use of this to do it once for blocks with common PHI
1191 SmallVector<BasicBlock*, 1> PredBBs;
1194 // If any of the predecessor blocks end in an unconditional branch, we can
1195 // *duplicate* the conditional branch into that block in order to further
1196 // encourage jump threading and to eliminate cases where we have branch on a
1197 // phi of an icmp (branch on icmp is much better).
1198 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1199 BasicBlock *PredBB = PN->getIncomingBlock(i);
1200 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1201 if (PredBr->isUnconditional()) {
1202 PredBBs[0] = PredBB;
1203 // Try to duplicate BB into PredBB.
1204 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1212 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1213 /// a xor instruction in the current block. See if there are any
1214 /// simplifications we can do based on inputs to the xor.
1216 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1217 BasicBlock *BB = BO->getParent();
1219 // If either the LHS or RHS of the xor is a constant, don't do this
1221 if (isa<ConstantInt>(BO->getOperand(0)) ||
1222 isa<ConstantInt>(BO->getOperand(1)))
1225 // If the first instruction in BB isn't a phi, we won't be able to infer
1226 // anything special about any particular predecessor.
1227 if (!isa<PHINode>(BB->front()))
1230 // If we have a xor as the branch input to this block, and we know that the
1231 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1232 // the condition into the predecessor and fix that value to true, saving some
1233 // logical ops on that path and encouraging other paths to simplify.
1235 // This copies something like this:
1238 // %X = phi i1 [1], [%X']
1239 // %Y = icmp eq i32 %A, %B
1240 // %Z = xor i1 %X, %Y
1245 // %Y = icmp ne i32 %A, %B
1248 PredValueInfoTy XorOpValues;
1250 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1252 assert(XorOpValues.empty());
1253 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1259 assert(!XorOpValues.empty() &&
1260 "ComputeValueKnownInPredecessors returned true with no values");
1262 // Scan the information to see which is most popular: true or false. The
1263 // predecessors can be of the set true, false, or undef.
1264 unsigned NumTrue = 0, NumFalse = 0;
1265 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1266 if (isa<UndefValue>(XorOpValues[i].first))
1267 // Ignore undefs for the count.
1269 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1275 // Determine which value to split on, true, false, or undef if neither.
1276 ConstantInt *SplitVal = 0;
1277 if (NumTrue > NumFalse)
1278 SplitVal = ConstantInt::getTrue(BB->getContext());
1279 else if (NumTrue != 0 || NumFalse != 0)
1280 SplitVal = ConstantInt::getFalse(BB->getContext());
1282 // Collect all of the blocks that this can be folded into so that we can
1283 // factor this once and clone it once.
1284 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1285 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1286 if (XorOpValues[i].first != SplitVal &&
1287 !isa<UndefValue>(XorOpValues[i].first))
1290 BlocksToFoldInto.push_back(XorOpValues[i].second);
1293 // If we inferred a value for all of the predecessors, then duplication won't
1294 // help us. However, we can just replace the LHS or RHS with the constant.
1295 if (BlocksToFoldInto.size() ==
1296 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1297 if (SplitVal == 0) {
1298 // If all preds provide undef, just nuke the xor, because it is undef too.
1299 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1300 BO->eraseFromParent();
1301 } else if (SplitVal->isZero()) {
1302 // If all preds provide 0, replace the xor with the other input.
1303 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1304 BO->eraseFromParent();
1306 // If all preds provide 1, set the computed value to 1.
1307 BO->setOperand(!isLHS, SplitVal);
1313 // Try to duplicate BB into PredBB.
1314 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1318 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1319 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1320 /// NewPred using the entries from OldPred (suitably mapped).
1321 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1322 BasicBlock *OldPred,
1323 BasicBlock *NewPred,
1324 DenseMap<Instruction*, Value*> &ValueMap) {
1325 for (BasicBlock::iterator PNI = PHIBB->begin();
1326 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1327 // Ok, we have a PHI node. Figure out what the incoming value was for the
1329 Value *IV = PN->getIncomingValueForBlock(OldPred);
1331 // Remap the value if necessary.
1332 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1333 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1334 if (I != ValueMap.end())
1338 PN->addIncoming(IV, NewPred);
1342 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1343 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1344 /// across BB. Transform the IR to reflect this change.
1345 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1346 const SmallVectorImpl<BasicBlock*> &PredBBs,
1347 BasicBlock *SuccBB) {
1348 // If threading to the same block as we come from, we would infinite loop.
1350 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1351 << "' - would thread to self!\n");
1355 // If threading this would thread across a loop header, don't thread the edge.
1356 // See the comments above FindLoopHeaders for justifications and caveats.
1357 if (LoopHeaders.count(BB)) {
1358 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1359 << "' to dest BB '" << SuccBB->getName()
1360 << "' - it might create an irreducible loop!\n");
1364 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, Threshold);
1365 if (JumpThreadCost > Threshold) {
1366 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1367 << "' - Cost is too high: " << JumpThreadCost << "\n");
1371 // And finally, do it! Start by factoring the predecessors is needed.
1373 if (PredBBs.size() == 1)
1374 PredBB = PredBBs[0];
1376 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1377 << " common predecessors.\n");
1378 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1381 // And finally, do it!
1382 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1383 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1384 << ", across block:\n "
1387 LVI->threadEdge(PredBB, BB, SuccBB);
1389 // We are going to have to map operands from the original BB block to the new
1390 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1391 // account for entry from PredBB.
1392 DenseMap<Instruction*, Value*> ValueMapping;
1394 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1395 BB->getName()+".thread",
1396 BB->getParent(), BB);
1397 NewBB->moveAfter(PredBB);
1399 BasicBlock::iterator BI = BB->begin();
1400 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1401 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1403 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1404 // mapping and using it to remap operands in the cloned instructions.
1405 for (; !isa<TerminatorInst>(BI); ++BI) {
1406 Instruction *New = BI->clone();
1407 New->setName(BI->getName());
1408 NewBB->getInstList().push_back(New);
1409 ValueMapping[BI] = New;
1411 // Remap operands to patch up intra-block references.
1412 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1413 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1414 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1415 if (I != ValueMapping.end())
1416 New->setOperand(i, I->second);
1420 // We didn't copy the terminator from BB over to NewBB, because there is now
1421 // an unconditional jump to SuccBB. Insert the unconditional jump.
1422 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB);
1423 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1425 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1426 // PHI nodes for NewBB now.
1427 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1429 // If there were values defined in BB that are used outside the block, then we
1430 // now have to update all uses of the value to use either the original value,
1431 // the cloned value, or some PHI derived value. This can require arbitrary
1432 // PHI insertion, of which we are prepared to do, clean these up now.
1433 SSAUpdater SSAUpdate;
1434 SmallVector<Use*, 16> UsesToRename;
1435 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1436 // Scan all uses of this instruction to see if it is used outside of its
1437 // block, and if so, record them in UsesToRename.
1438 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1440 Instruction *User = cast<Instruction>(*UI);
1441 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1442 if (UserPN->getIncomingBlock(UI) == BB)
1444 } else if (User->getParent() == BB)
1447 UsesToRename.push_back(&UI.getUse());
1450 // If there are no uses outside the block, we're done with this instruction.
1451 if (UsesToRename.empty())
1454 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1456 // We found a use of I outside of BB. Rename all uses of I that are outside
1457 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1458 // with the two values we know.
1459 SSAUpdate.Initialize(I->getType(), I->getName());
1460 SSAUpdate.AddAvailableValue(BB, I);
1461 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1463 while (!UsesToRename.empty())
1464 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1465 DEBUG(dbgs() << "\n");
1469 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1470 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1471 // us to simplify any PHI nodes in BB.
1472 TerminatorInst *PredTerm = PredBB->getTerminator();
1473 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1474 if (PredTerm->getSuccessor(i) == BB) {
1475 BB->removePredecessor(PredBB, true);
1476 PredTerm->setSuccessor(i, NewBB);
1479 // At this point, the IR is fully up to date and consistent. Do a quick scan
1480 // over the new instructions and zap any that are constants or dead. This
1481 // frequently happens because of phi translation.
1482 SimplifyInstructionsInBlock(NewBB, DL, TLI);
1484 // Threaded an edge!
1489 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1490 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1491 /// If we can duplicate the contents of BB up into PredBB do so now, this
1492 /// improves the odds that the branch will be on an analyzable instruction like
1494 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1495 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1496 assert(!PredBBs.empty() && "Can't handle an empty set");
1498 // If BB is a loop header, then duplicating this block outside the loop would
1499 // cause us to transform this into an irreducible loop, don't do this.
1500 // See the comments above FindLoopHeaders for justifications and caveats.
1501 if (LoopHeaders.count(BB)) {
1502 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1503 << "' into predecessor block '" << PredBBs[0]->getName()
1504 << "' - it might create an irreducible loop!\n");
1508 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, Threshold);
1509 if (DuplicationCost > Threshold) {
1510 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1511 << "' - Cost is too high: " << DuplicationCost << "\n");
1515 // And finally, do it! Start by factoring the predecessors is needed.
1517 if (PredBBs.size() == 1)
1518 PredBB = PredBBs[0];
1520 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1521 << " common predecessors.\n");
1522 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1525 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1527 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1528 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1529 << DuplicationCost << " block is:" << *BB << "\n");
1531 // Unless PredBB ends with an unconditional branch, split the edge so that we
1532 // can just clone the bits from BB into the end of the new PredBB.
1533 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1535 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1536 PredBB = SplitEdge(PredBB, BB, this);
1537 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1540 // We are going to have to map operands from the original BB block into the
1541 // PredBB block. Evaluate PHI nodes in BB.
1542 DenseMap<Instruction*, Value*> ValueMapping;
1544 BasicBlock::iterator BI = BB->begin();
1545 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1546 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1548 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1549 // mapping and using it to remap operands in the cloned instructions.
1550 for (; BI != BB->end(); ++BI) {
1551 Instruction *New = BI->clone();
1553 // Remap operands to patch up intra-block references.
1554 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1555 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1556 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1557 if (I != ValueMapping.end())
1558 New->setOperand(i, I->second);
1561 // If this instruction can be simplified after the operands are updated,
1562 // just use the simplified value instead. This frequently happens due to
1564 if (Value *IV = SimplifyInstruction(New, DL)) {
1566 ValueMapping[BI] = IV;
1568 // Otherwise, insert the new instruction into the block.
1569 New->setName(BI->getName());
1570 PredBB->getInstList().insert(OldPredBranch, New);
1571 ValueMapping[BI] = New;
1575 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1576 // add entries to the PHI nodes for branch from PredBB now.
1577 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1578 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1580 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1583 // If there were values defined in BB that are used outside the block, then we
1584 // now have to update all uses of the value to use either the original value,
1585 // the cloned value, or some PHI derived value. This can require arbitrary
1586 // PHI insertion, of which we are prepared to do, clean these up now.
1587 SSAUpdater SSAUpdate;
1588 SmallVector<Use*, 16> UsesToRename;
1589 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1590 // Scan all uses of this instruction to see if it is used outside of its
1591 // block, and if so, record them in UsesToRename.
1592 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1594 Instruction *User = cast<Instruction>(*UI);
1595 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1596 if (UserPN->getIncomingBlock(UI) == BB)
1598 } else if (User->getParent() == BB)
1601 UsesToRename.push_back(&UI.getUse());
1604 // If there are no uses outside the block, we're done with this instruction.
1605 if (UsesToRename.empty())
1608 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1610 // We found a use of I outside of BB. Rename all uses of I that are outside
1611 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1612 // with the two values we know.
1613 SSAUpdate.Initialize(I->getType(), I->getName());
1614 SSAUpdate.AddAvailableValue(BB, I);
1615 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1617 while (!UsesToRename.empty())
1618 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1619 DEBUG(dbgs() << "\n");
1622 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1624 BB->removePredecessor(PredBB, true);
1626 // Remove the unconditional branch at the end of the PredBB block.
1627 OldPredBranch->eraseFromParent();
1633 /// TryToUnfoldSelect - Look for blocks of the form
1639 /// %p = phi [%a, %bb] ...
1643 /// And expand the select into a branch structure if one of its arms allows %c
1644 /// to be folded. This later enables threading from bb1 over bb2.
1645 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1646 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1647 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1648 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1650 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1651 CondLHS->getParent() != BB)
1654 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1655 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1656 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1658 // Look if one of the incoming values is a select in the corresponding
1660 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1663 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1664 if (!PredTerm || !PredTerm->isUnconditional())
1667 // Now check if one of the select values would allow us to constant fold the
1668 // terminator in BB. We don't do the transform if both sides fold, those
1669 // cases will be threaded in any case.
1670 LazyValueInfo::Tristate LHSFolds =
1671 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1673 LazyValueInfo::Tristate RHSFolds =
1674 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
1676 if ((LHSFolds != LazyValueInfo::Unknown ||
1677 RHSFolds != LazyValueInfo::Unknown) &&
1678 LHSFolds != RHSFolds) {
1679 // Expand the select.
1688 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
1689 BB->getParent(), BB);
1690 // Move the unconditional branch to NewBB.
1691 PredTerm->removeFromParent();
1692 NewBB->getInstList().insert(NewBB->end(), PredTerm);
1693 // Create a conditional branch and update PHI nodes.
1694 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
1695 CondLHS->setIncomingValue(I, SI->getFalseValue());
1696 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
1697 // The select is now dead.
1698 SI->eraseFromParent();
1700 // Update any other PHI nodes in BB.
1701 for (BasicBlock::iterator BI = BB->begin();
1702 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
1704 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);