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/IntrinsicInst.h"
17 #include "llvm/LLVMContext.h"
18 #include "llvm/Pass.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/Analysis/LazyValueInfo.h"
21 #include "llvm/Analysis/Loads.h"
22 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
23 #include "llvm/Transforms/Utils/Local.h"
24 #include "llvm/Transforms/Utils/SSAUpdater.h"
25 #include "llvm/Target/TargetData.h"
26 #include "llvm/ADT/DenseMap.h"
27 #include "llvm/ADT/DenseSet.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/ADT/STLExtras.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/SmallSet.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ValueHandle.h"
35 #include "llvm/Support/raw_ostream.h"
38 STATISTIC(NumThreads, "Number of jumps threaded");
39 STATISTIC(NumFolds, "Number of terminators folded");
40 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
42 static cl::opt<unsigned>
43 Threshold("jump-threading-threshold",
44 cl::desc("Max block size to duplicate for jump threading"),
45 cl::init(6), cl::Hidden);
47 // Turn on use of LazyValueInfo.
49 EnableLVI("enable-jump-threading-lvi",
50 cl::desc("Use LVI for jump threading"),
57 /// This pass performs 'jump threading', which looks at blocks that have
58 /// multiple predecessors and multiple successors. If one or more of the
59 /// predecessors of the block can be proven to always jump to one of the
60 /// successors, we forward the edge from the predecessor to the successor by
61 /// duplicating the contents of this block.
63 /// An example of when this can occur is code like this:
70 /// In this case, the unconditional branch at the end of the first if can be
71 /// revectored to the false side of the second if.
73 class JumpThreading : public FunctionPass {
77 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
79 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
81 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet;
83 static char ID; // Pass identification
84 JumpThreading() : FunctionPass(ID) {}
86 bool runOnFunction(Function &F);
88 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
90 AU.addRequired<LazyValueInfo>();
93 void FindLoopHeaders(Function &F);
94 bool ProcessBlock(BasicBlock *BB);
95 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
97 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
98 const SmallVectorImpl<BasicBlock *> &PredBBs);
100 typedef SmallVectorImpl<std::pair<ConstantInt*,
101 BasicBlock*> > PredValueInfo;
103 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
104 PredValueInfo &Result);
105 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB);
108 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
109 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
111 bool ProcessBranchOnPHI(PHINode *PN);
112 bool ProcessBranchOnXOR(BinaryOperator *BO);
114 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
118 char JumpThreading::ID = 0;
119 INITIALIZE_PASS(JumpThreading, "jump-threading",
120 "Jump Threading", false, false);
122 // Public interface to the Jump Threading pass
123 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
125 /// runOnFunction - Top level algorithm.
127 bool JumpThreading::runOnFunction(Function &F) {
128 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
129 TD = getAnalysisIfAvailable<TargetData>();
130 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0;
134 bool Changed, EverChanged = false;
137 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
139 // Thread all of the branches we can over this block.
140 while (ProcessBlock(BB))
145 // If the block is trivially dead, zap it. This eliminates the successor
146 // edges which simplifies the CFG.
147 if (pred_begin(BB) == pred_end(BB) &&
148 BB != &BB->getParent()->getEntryBlock()) {
149 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
150 << "' with terminator: " << *BB->getTerminator() << '\n');
151 LoopHeaders.erase(BB);
152 if (LVI) LVI->eraseBlock(BB);
155 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
156 // Can't thread an unconditional jump, but if the block is "almost
157 // empty", we can replace uses of it with uses of the successor and make
159 if (BI->isUnconditional() &&
160 BB != &BB->getParent()->getEntryBlock()) {
161 BasicBlock::iterator BBI = BB->getFirstNonPHI();
162 // Ignore dbg intrinsics.
163 while (isa<DbgInfoIntrinsic>(BBI))
165 // If the terminator is the only non-phi instruction, try to nuke it.
166 if (BBI->isTerminator()) {
167 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
168 // block, we have to make sure it isn't in the LoopHeaders set. We
169 // reinsert afterward if needed.
170 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
171 BasicBlock *Succ = BI->getSuccessor(0);
173 // FIXME: It is always conservatively correct to drop the info
174 // for a block even if it doesn't get erased. This isn't totally
175 // awesome, but it allows us to use AssertingVH to prevent nasty
176 // dangling pointer issues within LazyValueInfo.
177 if (LVI) LVI->eraseBlock(BB);
178 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
180 // If we deleted BB and BB was the header of a loop, then the
181 // successor is now the header of the loop.
185 if (ErasedFromLoopHeaders)
186 LoopHeaders.insert(BB);
191 EverChanged |= Changed;
198 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
199 /// thread across it.
200 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
201 /// Ignore PHI nodes, these will be flattened when duplication happens.
202 BasicBlock::const_iterator I = BB->getFirstNonPHI();
204 // FIXME: THREADING will delete values that are just used to compute the
205 // branch, so they shouldn't count against the duplication cost.
208 // Sum up the cost of each instruction until we get to the terminator. Don't
209 // include the terminator because the copy won't include it.
211 for (; !isa<TerminatorInst>(I); ++I) {
212 // Debugger intrinsics don't incur code size.
213 if (isa<DbgInfoIntrinsic>(I)) continue;
215 // If this is a pointer->pointer bitcast, it is free.
216 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
219 // All other instructions count for at least one unit.
222 // Calls are more expensive. If they are non-intrinsic calls, we model them
223 // as having cost of 4. If they are a non-vector intrinsic, we model them
224 // as having cost of 2 total, and if they are a vector intrinsic, we model
225 // them as having cost 1.
226 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
227 if (!isa<IntrinsicInst>(CI))
229 else if (!CI->getType()->isVectorTy())
234 // Threading through a switch statement is particularly profitable. If this
235 // block ends in a switch, decrease its cost to make it more likely to happen.
236 if (isa<SwitchInst>(I))
237 Size = Size > 6 ? Size-6 : 0;
242 /// FindLoopHeaders - We do not want jump threading to turn proper loop
243 /// structures into irreducible loops. Doing this breaks up the loop nesting
244 /// hierarchy and pessimizes later transformations. To prevent this from
245 /// happening, we first have to find the loop headers. Here we approximate this
246 /// by finding targets of backedges in the CFG.
248 /// Note that there definitely are cases when we want to allow threading of
249 /// edges across a loop header. For example, threading a jump from outside the
250 /// loop (the preheader) to an exit block of the loop is definitely profitable.
251 /// It is also almost always profitable to thread backedges from within the loop
252 /// to exit blocks, and is often profitable to thread backedges to other blocks
253 /// within the loop (forming a nested loop). This simple analysis is not rich
254 /// enough to track all of these properties and keep it up-to-date as the CFG
255 /// mutates, so we don't allow any of these transformations.
257 void JumpThreading::FindLoopHeaders(Function &F) {
258 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
259 FindFunctionBackedges(F, Edges);
261 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
262 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
265 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
266 /// if we can infer that the value is a known ConstantInt in any of our
267 /// predecessors. If so, return the known list of value and pred BB in the
268 /// result vector. If a value is known to be undef, it is returned as null.
270 /// This returns true if there were any known values.
273 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){
274 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
277 // If V is a constantint, then it is known in all predecessors.
278 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) {
279 ConstantInt *CI = dyn_cast<ConstantInt>(V);
281 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
282 Result.push_back(std::make_pair(CI, *PI));
284 RecursionSet.erase(std::make_pair(V, BB));
288 // If V is a non-instruction value, or an instruction in a different block,
289 // then it can't be derived from a PHI.
290 Instruction *I = dyn_cast<Instruction>(V);
291 if (I == 0 || I->getParent() != BB) {
293 // Okay, if this is a live-in value, see if it has a known value at the end
294 // of any of our predecessors.
296 // FIXME: This should be an edge property, not a block end property.
297 /// TODO: Per PR2563, we could infer value range information about a
298 /// predecessor based on its terminator.
301 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
302 // "I" is a non-local compare-with-a-constant instruction. This would be
303 // able to handle value inequalities better, for example if the compare is
304 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
305 // Perhaps getConstantOnEdge should be smart enough to do this?
307 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
309 // If the value is known by LazyValueInfo to be a constant in a
310 // predecessor, use that information to try to thread this block.
311 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
313 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst)))
316 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), P));
319 RecursionSet.erase(std::make_pair(V, BB));
320 return !Result.empty();
323 RecursionSet.erase(std::make_pair(V, BB));
327 /// If I is a PHI node, then we know the incoming values for any constants.
328 if (PHINode *PN = dyn_cast<PHINode>(I)) {
329 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
330 Value *InVal = PN->getIncomingValue(i);
331 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) {
332 ConstantInt *CI = dyn_cast<ConstantInt>(InVal);
333 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i)));
335 Constant *CI = LVI->getConstantOnEdge(InVal,
336 PN->getIncomingBlock(i), BB);
337 // LVI returns null is no value could be determined.
339 if (ConstantInt *CInt = dyn_cast<ConstantInt>(CI))
340 Result.push_back(std::make_pair(CInt, PN->getIncomingBlock(i)));
341 else if (isa<UndefValue>(CI))
342 Result.push_back(std::make_pair((ConstantInt*)0,
343 PN->getIncomingBlock(i)));
347 RecursionSet.erase(std::make_pair(V, BB));
348 return !Result.empty();
351 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals;
353 // Handle some boolean conditions.
354 if (I->getType()->getPrimitiveSizeInBits() == 1) {
356 // X & false -> false
357 if (I->getOpcode() == Instruction::Or ||
358 I->getOpcode() == Instruction::And) {
359 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
360 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals);
362 if (LHSVals.empty() && RHSVals.empty()) {
363 RecursionSet.erase(std::make_pair(V, BB));
367 ConstantInt *InterestingVal;
368 if (I->getOpcode() == Instruction::Or)
369 InterestingVal = ConstantInt::getTrue(I->getContext());
371 InterestingVal = ConstantInt::getFalse(I->getContext());
373 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
375 // Scan for the sentinel. If we find an undef, force it to the
376 // interesting value: x|undef -> true and x&undef -> false.
377 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
378 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) {
379 Result.push_back(LHSVals[i]);
380 Result.back().first = InterestingVal;
381 LHSKnownBBs.insert(LHSVals[i].second);
383 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
384 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) {
385 // If we already inferred a value for this block on the LHS, don't
387 if (!LHSKnownBBs.count(RHSVals[i].second)) {
388 Result.push_back(RHSVals[i]);
389 Result.back().first = InterestingVal;
393 RecursionSet.erase(std::make_pair(V, BB));
394 return !Result.empty();
397 // Handle the NOT form of XOR.
398 if (I->getOpcode() == Instruction::Xor &&
399 isa<ConstantInt>(I->getOperand(1)) &&
400 cast<ConstantInt>(I->getOperand(1))->isOne()) {
401 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result);
402 if (Result.empty()) {
403 RecursionSet.erase(std::make_pair(V, BB));
407 // Invert the known values.
408 for (unsigned i = 0, e = Result.size(); i != e; ++i)
411 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first));
413 RecursionSet.erase(std::make_pair(V, BB));
417 // Try to simplify some other binary operator values.
418 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
419 ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1));
421 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals;
422 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals);
424 // Try to use constant folding to simplify the binary operator.
425 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
426 Constant *Folded = 0;
427 if (LHSVals[i].first == 0) {
428 Folded = ConstantExpr::get(BO->getOpcode(),
429 UndefValue::get(BO->getType()),
432 Folded = ConstantExpr::get(BO->getOpcode(), LHSVals[i].first, CI);
435 if (ConstantInt *FoldedCInt = dyn_cast<ConstantInt>(Folded))
436 Result.push_back(std::make_pair(FoldedCInt, LHSVals[i].second));
437 else if (isa<UndefValue>(Folded))
438 Result.push_back(std::make_pair((ConstantInt*)0, LHSVals[i].second));
442 RecursionSet.erase(std::make_pair(V, BB));
443 return !Result.empty();
446 // Handle compare with phi operand, where the PHI is defined in this block.
447 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
448 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
449 if (PN && PN->getParent() == BB) {
450 // We can do this simplification if any comparisons fold to true or false.
452 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
453 BasicBlock *PredBB = PN->getIncomingBlock(i);
454 Value *LHS = PN->getIncomingValue(i);
455 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
457 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
459 if (!LVI || !isa<Constant>(RHS))
462 LazyValueInfo::Tristate
463 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
464 cast<Constant>(RHS), PredBB, BB);
465 if (ResT == LazyValueInfo::Unknown)
467 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
470 if (isa<UndefValue>(Res))
471 Result.push_back(std::make_pair((ConstantInt*)0, PredBB));
472 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res))
473 Result.push_back(std::make_pair(CI, PredBB));
476 RecursionSet.erase(std::make_pair(V, BB));
477 return !Result.empty();
481 // If comparing a live-in value against a constant, see if we know the
482 // live-in value on any predecessors.
483 if (LVI && isa<Constant>(Cmp->getOperand(1)) &&
484 Cmp->getType()->isIntegerTy()) {
485 if (!isa<Instruction>(Cmp->getOperand(0)) ||
486 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
487 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
489 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
491 // If the value is known by LazyValueInfo to be a constant in a
492 // predecessor, use that information to try to thread this block.
493 LazyValueInfo::Tristate Res =
494 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
496 if (Res == LazyValueInfo::Unknown)
499 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
500 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), P));
503 RecursionSet.erase(std::make_pair(V, BB));
504 return !Result.empty();
507 // Try to find a constant value for the LHS of a comparison,
508 // and evaluate it statically if we can.
509 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
510 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals;
511 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
513 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
514 Constant * Folded = 0;
515 if (LHSVals[i].first == 0)
516 Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
517 UndefValue::get(CmpConst->getType()), CmpConst);
519 Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
520 LHSVals[i].first, CmpConst);
522 if (ConstantInt *FoldedCInt = dyn_cast<ConstantInt>(Folded))
523 Result.push_back(std::make_pair(FoldedCInt, LHSVals[i].second));
524 else if (isa<UndefValue>(Folded))
525 Result.push_back(std::make_pair((ConstantInt*)0,LHSVals[i].second));
528 RecursionSet.erase(std::make_pair(V, BB));
529 return !Result.empty();
535 // If all else fails, see if LVI can figure out a constant value for us.
536 Constant *CI = LVI->getConstant(V, BB);
537 ConstantInt *CInt = dyn_cast_or_null<ConstantInt>(CI);
539 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
540 Result.push_back(std::make_pair(CInt, *PI));
543 RecursionSet.erase(std::make_pair(V, BB));
544 return !Result.empty();
547 RecursionSet.erase(std::make_pair(V, BB));
553 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
554 /// in an undefined jump, decide which block is best to revector to.
556 /// Since we can pick an arbitrary destination, we pick the successor with the
557 /// fewest predecessors. This should reduce the in-degree of the others.
559 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
560 TerminatorInst *BBTerm = BB->getTerminator();
561 unsigned MinSucc = 0;
562 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
563 // Compute the successor with the minimum number of predecessors.
564 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
565 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
566 TestBB = BBTerm->getSuccessor(i);
567 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
568 if (NumPreds < MinNumPreds)
575 /// ProcessBlock - If there are any predecessors whose control can be threaded
576 /// through to a successor, transform them now.
577 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
578 // If the block is trivially dead, just return and let the caller nuke it.
579 // This simplifies other transformations.
580 if (pred_begin(BB) == pred_end(BB) &&
581 BB != &BB->getParent()->getEntryBlock())
584 // If this block has a single predecessor, and if that pred has a single
585 // successor, merge the blocks. This encourages recursive jump threading
586 // because now the condition in this block can be threaded through
587 // predecessors of our predecessor block.
588 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
589 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
591 // If SinglePred was a loop header, BB becomes one.
592 if (LoopHeaders.erase(SinglePred))
593 LoopHeaders.insert(BB);
595 // Remember if SinglePred was the entry block of the function. If so, we
596 // will need to move BB back to the entry position.
597 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
598 if (LVI) LVI->eraseBlock(SinglePred);
599 MergeBasicBlockIntoOnlyPred(BB);
601 if (isEntry && BB != &BB->getParent()->getEntryBlock())
602 BB->moveBefore(&BB->getParent()->getEntryBlock());
607 // Look to see if the terminator is a branch of switch, if not we can't thread
610 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
611 // Can't thread an unconditional jump.
612 if (BI->isUnconditional()) return false;
613 Condition = BI->getCondition();
614 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
615 Condition = SI->getCondition();
617 return false; // Must be an invoke.
619 // If the terminator of this block is branching on a constant, simplify the
620 // terminator to an unconditional branch. This can occur due to threading in
622 if (isa<ConstantInt>(Condition)) {
623 DEBUG(dbgs() << " In block '" << BB->getName()
624 << "' folding terminator: " << *BB->getTerminator() << '\n');
626 ConstantFoldTerminator(BB);
630 // If the terminator is branching on an undef, we can pick any of the
631 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
632 if (isa<UndefValue>(Condition)) {
633 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
635 // Fold the branch/switch.
636 TerminatorInst *BBTerm = BB->getTerminator();
637 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
638 if (i == BestSucc) continue;
639 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD);
642 DEBUG(dbgs() << " In block '" << BB->getName()
643 << "' folding undef terminator: " << *BBTerm << '\n');
644 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
645 BBTerm->eraseFromParent();
649 Instruction *CondInst = dyn_cast<Instruction>(Condition);
651 // If the condition is an instruction defined in another block, see if a
652 // predecessor has the same condition:
657 !Condition->hasOneUse() && // Multiple uses.
658 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition.
659 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
660 if (isa<BranchInst>(BB->getTerminator())) {
661 for (; PI != E; ++PI) {
663 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator()))
664 if (PBI->isConditional() && PBI->getCondition() == Condition &&
665 ProcessBranchOnDuplicateCond(P, BB))
669 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator");
670 for (; PI != E; ++PI) {
672 if (SwitchInst *PSI = dyn_cast<SwitchInst>(P->getTerminator()))
673 if (PSI->getCondition() == Condition &&
674 ProcessSwitchOnDuplicateCond(P, BB))
680 // All the rest of our checks depend on the condition being an instruction.
682 // FIXME: Unify this with code below.
683 if (LVI && ProcessThreadableEdges(Condition, BB))
689 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
691 (!isa<PHINode>(CondCmp->getOperand(0)) ||
692 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) {
693 // If we have a comparison, loop over the predecessors to see if there is
694 // a condition with a lexically identical value.
695 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
696 for (; PI != E; ++PI) {
698 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator()))
699 if (PBI->isConditional() && P != BB) {
700 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) {
701 if (CI->getOperand(0) == CondCmp->getOperand(0) &&
702 CI->getOperand(1) == CondCmp->getOperand(1) &&
703 CI->getPredicate() == CondCmp->getPredicate()) {
704 // TODO: Could handle things like (x != 4) --> (x == 17)
705 if (ProcessBranchOnDuplicateCond(P, BB))
713 // For a comparison where the LHS is outside this block, it's possible
714 // that we've branched on it before. Used LVI to see if we can simplify
715 // the branch based on that.
716 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
717 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
718 if (LVI && CondBr && CondConst && CondBr->isConditional() &&
719 (!isa<Instruction>(CondCmp->getOperand(0)) ||
720 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
721 // For predecessor edge, determine if the comparison is true or false
722 // on that edge. If they're all true or all false, we can simplify the
724 // FIXME: We could handle mixed true/false by duplicating code.
725 unsigned Trues = 0, Falses = 0, predcount = 0;
726 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB);PI != PE; ++PI){
728 LazyValueInfo::Tristate Ret =
729 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
730 CondCmp->getOperand(0), CondConst, *PI, BB);
731 if (Ret == LazyValueInfo::True)
733 else if (Ret == LazyValueInfo::False)
737 // If we can determine the branch direction statically, convert
738 // the conditional branch to an unconditional one.
739 if (Trues && Trues == predcount) {
740 RemovePredecessorAndSimplify(CondBr->getSuccessor(1), BB, TD);
741 BranchInst::Create(CondBr->getSuccessor(0), CondBr);
742 CondBr->eraseFromParent();
744 } else if (Falses && Falses == predcount) {
745 RemovePredecessorAndSimplify(CondBr->getSuccessor(0), BB, TD);
746 BranchInst::Create(CondBr->getSuccessor(1), CondBr);
747 CondBr->eraseFromParent();
753 // Check for some cases that are worth simplifying. Right now we want to look
754 // for loads that are used by a switch or by the condition for the branch. If
755 // we see one, check to see if it's partially redundant. If so, insert a PHI
756 // which can then be used to thread the values.
758 Value *SimplifyValue = CondInst;
759 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
760 if (isa<Constant>(CondCmp->getOperand(1)))
761 SimplifyValue = CondCmp->getOperand(0);
763 // TODO: There are other places where load PRE would be profitable, such as
764 // more complex comparisons.
765 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
766 if (SimplifyPartiallyRedundantLoad(LI))
770 // Handle a variety of cases where we are branching on something derived from
771 // a PHI node in the current block. If we can prove that any predecessors
772 // compute a predictable value based on a PHI node, thread those predecessors.
774 if (ProcessThreadableEdges(CondInst, BB))
777 // If this is an otherwise-unfoldable branch on a phi node in the current
778 // block, see if we can simplify.
779 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
780 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
781 return ProcessBranchOnPHI(PN);
784 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
785 if (CondInst->getOpcode() == Instruction::Xor &&
786 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
787 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
790 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
791 // "(X == 4)", thread through this block.
796 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
797 /// block that jump on exactly the same condition. This means that we almost
798 /// always know the direction of the edge in the DESTBB:
800 /// br COND, DESTBB, BBY
802 /// br COND, BBZ, BBW
804 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
805 /// in DESTBB, we have to thread over it.
806 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
808 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
810 // If both successors of PredBB go to DESTBB, we don't know anything. We can
811 // fold the branch to an unconditional one, which allows other recursive
814 if (PredBI->getSuccessor(1) != BB)
816 else if (PredBI->getSuccessor(0) != BB)
819 DEBUG(dbgs() << " In block '" << PredBB->getName()
820 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
822 ConstantFoldTerminator(PredBB);
826 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
828 // If the dest block has one predecessor, just fix the branch condition to a
829 // constant and fold it.
830 if (BB->getSinglePredecessor()) {
831 DEBUG(dbgs() << " In block '" << BB->getName()
832 << "' folding condition to '" << BranchDir << "': "
833 << *BB->getTerminator() << '\n');
835 Value *OldCond = DestBI->getCondition();
836 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
838 // Delete dead instructions before we fold the branch. Folding the branch
839 // can eliminate edges from the CFG which can end up deleting OldCond.
840 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
841 ConstantFoldTerminator(BB);
846 // Next, figure out which successor we are threading to.
847 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
849 SmallVector<BasicBlock*, 2> Preds;
850 Preds.push_back(PredBB);
852 // Ok, try to thread it!
853 return ThreadEdge(BB, Preds, SuccBB);
856 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
857 /// block that switch on exactly the same condition. This means that we almost
858 /// always know the direction of the edge in the DESTBB:
860 /// switch COND [... DESTBB, BBY ... ]
862 /// switch COND [... BBZ, BBW ]
864 /// Optimizing switches like this is very important, because simplifycfg builds
865 /// switches out of repeated 'if' conditions.
866 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
867 BasicBlock *DestBB) {
868 // Can't thread edge to self.
869 if (PredBB == DestBB)
872 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
873 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
875 // There are a variety of optimizations that we can potentially do on these
876 // blocks: we order them from most to least preferable.
878 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
879 // directly to their destination. This does not introduce *any* code size
880 // growth. Skip debug info first.
881 BasicBlock::iterator BBI = DestBB->begin();
882 while (isa<DbgInfoIntrinsic>(BBI))
885 // FIXME: Thread if it just contains a PHI.
886 if (isa<SwitchInst>(BBI)) {
887 bool MadeChange = false;
888 // Ignore the default edge for now.
889 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
890 ConstantInt *DestVal = DestSI->getCaseValue(i);
891 BasicBlock *DestSucc = DestSI->getSuccessor(i);
893 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
894 // PredSI has an explicit case for it. If so, forward. If it is covered
895 // by the default case, we can't update PredSI.
896 unsigned PredCase = PredSI->findCaseValue(DestVal);
897 if (PredCase == 0) continue;
899 // If PredSI doesn't go to DestBB on this value, then it won't reach the
900 // case on this condition.
901 if (PredSI->getSuccessor(PredCase) != DestBB &&
902 DestSI->getSuccessor(i) != DestBB)
905 // Do not forward this if it already goes to this destination, this would
906 // be an infinite loop.
907 if (PredSI->getSuccessor(PredCase) == DestSucc)
910 // Otherwise, we're safe to make the change. Make sure that the edge from
911 // DestSI to DestSucc is not critical and has no PHI nodes.
912 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
913 DEBUG(dbgs() << "THROUGH: " << *DestSI);
915 // If the destination has PHI nodes, just split the edge for updating
917 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
918 SplitCriticalEdge(DestSI, i, this);
919 DestSucc = DestSI->getSuccessor(i);
921 FoldSingleEntryPHINodes(DestSucc);
922 PredSI->setSuccessor(PredCase, DestSucc);
934 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
935 /// load instruction, eliminate it by replacing it with a PHI node. This is an
936 /// important optimization that encourages jump threading, and needs to be run
937 /// interlaced with other jump threading tasks.
938 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
939 // Don't hack volatile loads.
940 if (LI->isVolatile()) return false;
942 // If the load is defined in a block with exactly one predecessor, it can't be
943 // partially redundant.
944 BasicBlock *LoadBB = LI->getParent();
945 if (LoadBB->getSinglePredecessor())
948 Value *LoadedPtr = LI->getOperand(0);
950 // If the loaded operand is defined in the LoadBB, it can't be available.
951 // TODO: Could do simple PHI translation, that would be fun :)
952 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
953 if (PtrOp->getParent() == LoadBB)
956 // Scan a few instructions up from the load, to see if it is obviously live at
957 // the entry to its block.
958 BasicBlock::iterator BBIt = LI;
960 if (Value *AvailableVal =
961 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
962 // If the value if the load is locally available within the block, just use
963 // it. This frequently occurs for reg2mem'd allocas.
964 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
966 // If the returned value is the load itself, replace with an undef. This can
967 // only happen in dead loops.
968 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
969 LI->replaceAllUsesWith(AvailableVal);
970 LI->eraseFromParent();
974 // Otherwise, if we scanned the whole block and got to the top of the block,
975 // we know the block is locally transparent to the load. If not, something
976 // might clobber its value.
977 if (BBIt != LoadBB->begin())
981 SmallPtrSet<BasicBlock*, 8> PredsScanned;
982 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
983 AvailablePredsTy AvailablePreds;
984 BasicBlock *OneUnavailablePred = 0;
986 // If we got here, the loaded value is transparent through to the start of the
987 // block. Check to see if it is available in any of the predecessor blocks.
988 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
990 BasicBlock *PredBB = *PI;
992 // If we already scanned this predecessor, skip it.
993 if (!PredsScanned.insert(PredBB))
996 // Scan the predecessor to see if the value is available in the pred.
997 BBIt = PredBB->end();
998 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
999 if (!PredAvailable) {
1000 OneUnavailablePred = PredBB;
1004 // If so, this load is partially redundant. Remember this info so that we
1005 // can create a PHI node.
1006 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1009 // If the loaded value isn't available in any predecessor, it isn't partially
1011 if (AvailablePreds.empty()) return false;
1013 // Okay, the loaded value is available in at least one (and maybe all!)
1014 // predecessors. If the value is unavailable in more than one unique
1015 // predecessor, we want to insert a merge block for those common predecessors.
1016 // This ensures that we only have to insert one reload, thus not increasing
1018 BasicBlock *UnavailablePred = 0;
1020 // If there is exactly one predecessor where the value is unavailable, the
1021 // already computed 'OneUnavailablePred' block is it. If it ends in an
1022 // unconditional branch, we know that it isn't a critical edge.
1023 if (PredsScanned.size() == AvailablePreds.size()+1 &&
1024 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1025 UnavailablePred = OneUnavailablePred;
1026 } else if (PredsScanned.size() != AvailablePreds.size()) {
1027 // Otherwise, we had multiple unavailable predecessors or we had a critical
1028 // edge from the one.
1029 SmallVector<BasicBlock*, 8> PredsToSplit;
1030 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1032 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
1033 AvailablePredSet.insert(AvailablePreds[i].first);
1035 // Add all the unavailable predecessors to the PredsToSplit list.
1036 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
1038 BasicBlock *P = *PI;
1039 // If the predecessor is an indirect goto, we can't split the edge.
1040 if (isa<IndirectBrInst>(P->getTerminator()))
1043 if (!AvailablePredSet.count(P))
1044 PredsToSplit.push_back(P);
1047 // Split them out to their own block.
1049 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
1050 "thread-pre-split", this);
1053 // If the value isn't available in all predecessors, then there will be
1054 // exactly one where it isn't available. Insert a load on that edge and add
1055 // it to the AvailablePreds list.
1056 if (UnavailablePred) {
1057 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1058 "Can't handle critical edge here!");
1059 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
1061 UnavailablePred->getTerminator());
1062 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1065 // Now we know that each predecessor of this block has a value in
1066 // AvailablePreds, sort them for efficient access as we're walking the preds.
1067 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1069 // Create a PHI node at the start of the block for the PRE'd load value.
1070 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
1073 // Insert new entries into the PHI for each predecessor. A single block may
1074 // have multiple entries here.
1075 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
1077 BasicBlock *P = *PI;
1078 AvailablePredsTy::iterator I =
1079 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1080 std::make_pair(P, (Value*)0));
1082 assert(I != AvailablePreds.end() && I->first == P &&
1083 "Didn't find entry for predecessor!");
1085 PN->addIncoming(I->second, I->first);
1088 //cerr << "PRE: " << *LI << *PN << "\n";
1090 LI->replaceAllUsesWith(PN);
1091 LI->eraseFromParent();
1096 /// FindMostPopularDest - The specified list contains multiple possible
1097 /// threadable destinations. Pick the one that occurs the most frequently in
1100 FindMostPopularDest(BasicBlock *BB,
1101 const SmallVectorImpl<std::pair<BasicBlock*,
1102 BasicBlock*> > &PredToDestList) {
1103 assert(!PredToDestList.empty());
1105 // Determine popularity. If there are multiple possible destinations, we
1106 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1107 // blocks with known and real destinations to threading undef. We'll handle
1108 // them later if interesting.
1109 DenseMap<BasicBlock*, unsigned> DestPopularity;
1110 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1111 if (PredToDestList[i].second)
1112 DestPopularity[PredToDestList[i].second]++;
1114 // Find the most popular dest.
1115 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1116 BasicBlock *MostPopularDest = DPI->first;
1117 unsigned Popularity = DPI->second;
1118 SmallVector<BasicBlock*, 4> SamePopularity;
1120 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1121 // If the popularity of this entry isn't higher than the popularity we've
1122 // seen so far, ignore it.
1123 if (DPI->second < Popularity)
1125 else if (DPI->second == Popularity) {
1126 // If it is the same as what we've seen so far, keep track of it.
1127 SamePopularity.push_back(DPI->first);
1129 // If it is more popular, remember it.
1130 SamePopularity.clear();
1131 MostPopularDest = DPI->first;
1132 Popularity = DPI->second;
1136 // Okay, now we know the most popular destination. If there is more than
1137 // destination, we need to determine one. This is arbitrary, but we need
1138 // to make a deterministic decision. Pick the first one that appears in the
1140 if (!SamePopularity.empty()) {
1141 SamePopularity.push_back(MostPopularDest);
1142 TerminatorInst *TI = BB->getTerminator();
1143 for (unsigned i = 0; ; ++i) {
1144 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1146 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1147 TI->getSuccessor(i)) == SamePopularity.end())
1150 MostPopularDest = TI->getSuccessor(i);
1155 // Okay, we have finally picked the most popular destination.
1156 return MostPopularDest;
1159 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) {
1160 // If threading this would thread across a loop header, don't even try to
1162 if (LoopHeaders.count(BB))
1165 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues;
1166 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues)) {
1169 assert(!PredValues.empty() &&
1170 "ComputeValueKnownInPredecessors returned true with no values");
1172 DEBUG(dbgs() << "IN BB: " << *BB;
1173 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1174 dbgs() << " BB '" << BB->getName() << "': FOUND condition = ";
1175 if (PredValues[i].first)
1176 dbgs() << *PredValues[i].first;
1179 dbgs() << " for pred '" << PredValues[i].second->getName()
1183 // Decide what we want to thread through. Convert our list of known values to
1184 // a list of known destinations for each pred. This also discards duplicate
1185 // predecessors and keeps track of the undefined inputs (which are represented
1186 // as a null dest in the PredToDestList).
1187 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1188 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1190 BasicBlock *OnlyDest = 0;
1191 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1193 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1194 BasicBlock *Pred = PredValues[i].second;
1195 if (!SeenPreds.insert(Pred))
1196 continue; // Duplicate predecessor entry.
1198 // If the predecessor ends with an indirect goto, we can't change its
1200 if (isa<IndirectBrInst>(Pred->getTerminator()))
1203 ConstantInt *Val = PredValues[i].first;
1206 if (Val == 0) // Undef.
1208 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1209 DestBB = BI->getSuccessor(Val->isZero());
1211 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
1212 DestBB = SI->getSuccessor(SI->findCaseValue(Val));
1215 // If we have exactly one destination, remember it for efficiency below.
1218 else if (OnlyDest != DestBB)
1219 OnlyDest = MultipleDestSentinel;
1221 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1224 // If all edges were unthreadable, we fail.
1225 if (PredToDestList.empty())
1228 // Determine which is the most common successor. If we have many inputs and
1229 // this block is a switch, we want to start by threading the batch that goes
1230 // to the most popular destination first. If we only know about one
1231 // threadable destination (the common case) we can avoid this.
1232 BasicBlock *MostPopularDest = OnlyDest;
1234 if (MostPopularDest == MultipleDestSentinel)
1235 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1237 // Now that we know what the most popular destination is, factor all
1238 // predecessors that will jump to it into a single predecessor.
1239 SmallVector<BasicBlock*, 16> PredsToFactor;
1240 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1241 if (PredToDestList[i].second == MostPopularDest) {
1242 BasicBlock *Pred = PredToDestList[i].first;
1244 // This predecessor may be a switch or something else that has multiple
1245 // edges to the block. Factor each of these edges by listing them
1246 // according to # occurrences in PredsToFactor.
1247 TerminatorInst *PredTI = Pred->getTerminator();
1248 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1249 if (PredTI->getSuccessor(i) == BB)
1250 PredsToFactor.push_back(Pred);
1253 // If the threadable edges are branching on an undefined value, we get to pick
1254 // the destination that these predecessors should get to.
1255 if (MostPopularDest == 0)
1256 MostPopularDest = BB->getTerminator()->
1257 getSuccessor(GetBestDestForJumpOnUndef(BB));
1259 // Ok, try to thread it!
1260 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1263 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1264 /// a PHI node in the current block. See if there are any simplifications we
1265 /// can do based on inputs to the phi node.
1267 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1268 BasicBlock *BB = PN->getParent();
1270 // TODO: We could make use of this to do it once for blocks with common PHI
1272 SmallVector<BasicBlock*, 1> PredBBs;
1275 // If any of the predecessor blocks end in an unconditional branch, we can
1276 // *duplicate* the conditional branch into that block in order to further
1277 // encourage jump threading and to eliminate cases where we have branch on a
1278 // phi of an icmp (branch on icmp is much better).
1279 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1280 BasicBlock *PredBB = PN->getIncomingBlock(i);
1281 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1282 if (PredBr->isUnconditional()) {
1283 PredBBs[0] = PredBB;
1284 // Try to duplicate BB into PredBB.
1285 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1293 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1294 /// a xor instruction in the current block. See if there are any
1295 /// simplifications we can do based on inputs to the xor.
1297 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1298 BasicBlock *BB = BO->getParent();
1300 // If either the LHS or RHS of the xor is a constant, don't do this
1302 if (isa<ConstantInt>(BO->getOperand(0)) ||
1303 isa<ConstantInt>(BO->getOperand(1)))
1306 // If the first instruction in BB isn't a phi, we won't be able to infer
1307 // anything special about any particular predecessor.
1308 if (!isa<PHINode>(BB->front()))
1311 // If we have a xor as the branch input to this block, and we know that the
1312 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1313 // the condition into the predecessor and fix that value to true, saving some
1314 // logical ops on that path and encouraging other paths to simplify.
1316 // This copies something like this:
1319 // %X = phi i1 [1], [%X']
1320 // %Y = icmp eq i32 %A, %B
1321 // %Z = xor i1 %X, %Y
1326 // %Y = icmp ne i32 %A, %B
1329 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues;
1331 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) {
1332 assert(XorOpValues.empty());
1333 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues))
1338 assert(!XorOpValues.empty() &&
1339 "ComputeValueKnownInPredecessors returned true with no values");
1341 // Scan the information to see which is most popular: true or false. The
1342 // predecessors can be of the set true, false, or undef.
1343 unsigned NumTrue = 0, NumFalse = 0;
1344 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1345 if (!XorOpValues[i].first) continue; // Ignore undefs for the count.
1346 if (XorOpValues[i].first->isZero())
1352 // Determine which value to split on, true, false, or undef if neither.
1353 ConstantInt *SplitVal = 0;
1354 if (NumTrue > NumFalse)
1355 SplitVal = ConstantInt::getTrue(BB->getContext());
1356 else if (NumTrue != 0 || NumFalse != 0)
1357 SplitVal = ConstantInt::getFalse(BB->getContext());
1359 // Collect all of the blocks that this can be folded into so that we can
1360 // factor this once and clone it once.
1361 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1362 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1363 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue;
1365 BlocksToFoldInto.push_back(XorOpValues[i].second);
1368 // If we inferred a value for all of the predecessors, then duplication won't
1369 // help us. However, we can just replace the LHS or RHS with the constant.
1370 if (BlocksToFoldInto.size() ==
1371 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1372 if (SplitVal == 0) {
1373 // If all preds provide undef, just nuke the xor, because it is undef too.
1374 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1375 BO->eraseFromParent();
1376 } else if (SplitVal->isZero()) {
1377 // If all preds provide 0, replace the xor with the other input.
1378 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1379 BO->eraseFromParent();
1381 // If all preds provide 1, set the computed value to 1.
1382 BO->setOperand(!isLHS, SplitVal);
1388 // Try to duplicate BB into PredBB.
1389 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1393 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1394 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1395 /// NewPred using the entries from OldPred (suitably mapped).
1396 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1397 BasicBlock *OldPred,
1398 BasicBlock *NewPred,
1399 DenseMap<Instruction*, Value*> &ValueMap) {
1400 for (BasicBlock::iterator PNI = PHIBB->begin();
1401 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1402 // Ok, we have a PHI node. Figure out what the incoming value was for the
1404 Value *IV = PN->getIncomingValueForBlock(OldPred);
1406 // Remap the value if necessary.
1407 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1408 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1409 if (I != ValueMap.end())
1413 PN->addIncoming(IV, NewPred);
1417 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1418 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1419 /// across BB. Transform the IR to reflect this change.
1420 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1421 const SmallVectorImpl<BasicBlock*> &PredBBs,
1422 BasicBlock *SuccBB) {
1423 // If threading to the same block as we come from, we would infinite loop.
1425 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1426 << "' - would thread to self!\n");
1430 // If threading this would thread across a loop header, don't thread the edge.
1431 // See the comments above FindLoopHeaders for justifications and caveats.
1432 if (LoopHeaders.count(BB)) {
1433 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1434 << "' to dest BB '" << SuccBB->getName()
1435 << "' - it might create an irreducible loop!\n");
1439 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1440 if (JumpThreadCost > Threshold) {
1441 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1442 << "' - Cost is too high: " << JumpThreadCost << "\n");
1446 // And finally, do it! Start by factoring the predecessors is needed.
1448 if (PredBBs.size() == 1)
1449 PredBB = PredBBs[0];
1451 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1452 << " common predecessors.\n");
1453 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1457 // And finally, do it!
1458 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1459 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1460 << ", across block:\n "
1464 LVI->threadEdge(PredBB, BB, SuccBB);
1466 // We are going to have to map operands from the original BB block to the new
1467 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1468 // account for entry from PredBB.
1469 DenseMap<Instruction*, Value*> ValueMapping;
1471 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1472 BB->getName()+".thread",
1473 BB->getParent(), BB);
1474 NewBB->moveAfter(PredBB);
1476 BasicBlock::iterator BI = BB->begin();
1477 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1478 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1480 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1481 // mapping and using it to remap operands in the cloned instructions.
1482 for (; !isa<TerminatorInst>(BI); ++BI) {
1483 Instruction *New = BI->clone();
1484 New->setName(BI->getName());
1485 NewBB->getInstList().push_back(New);
1486 ValueMapping[BI] = New;
1488 // Remap operands to patch up intra-block references.
1489 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1490 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1491 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1492 if (I != ValueMapping.end())
1493 New->setOperand(i, I->second);
1497 // We didn't copy the terminator from BB over to NewBB, because there is now
1498 // an unconditional jump to SuccBB. Insert the unconditional jump.
1499 BranchInst::Create(SuccBB, NewBB);
1501 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1502 // PHI nodes for NewBB now.
1503 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1505 // If there were values defined in BB that are used outside the block, then we
1506 // now have to update all uses of the value to use either the original value,
1507 // the cloned value, or some PHI derived value. This can require arbitrary
1508 // PHI insertion, of which we are prepared to do, clean these up now.
1509 SSAUpdater SSAUpdate;
1510 SmallVector<Use*, 16> UsesToRename;
1511 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1512 // Scan all uses of this instruction to see if it is used outside of its
1513 // block, and if so, record them in UsesToRename.
1514 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1516 Instruction *User = cast<Instruction>(*UI);
1517 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1518 if (UserPN->getIncomingBlock(UI) == BB)
1520 } else if (User->getParent() == BB)
1523 UsesToRename.push_back(&UI.getUse());
1526 // If there are no uses outside the block, we're done with this instruction.
1527 if (UsesToRename.empty())
1530 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1532 // We found a use of I outside of BB. Rename all uses of I that are outside
1533 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1534 // with the two values we know.
1535 SSAUpdate.Initialize(I);
1536 SSAUpdate.AddAvailableValue(BB, I);
1537 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1539 while (!UsesToRename.empty())
1540 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1541 DEBUG(dbgs() << "\n");
1545 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1546 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1547 // us to simplify any PHI nodes in BB.
1548 TerminatorInst *PredTerm = PredBB->getTerminator();
1549 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1550 if (PredTerm->getSuccessor(i) == BB) {
1551 RemovePredecessorAndSimplify(BB, PredBB, TD);
1552 PredTerm->setSuccessor(i, NewBB);
1555 // At this point, the IR is fully up to date and consistent. Do a quick scan
1556 // over the new instructions and zap any that are constants or dead. This
1557 // frequently happens because of phi translation.
1558 SimplifyInstructionsInBlock(NewBB, TD);
1560 // Threaded an edge!
1565 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1566 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1567 /// If we can duplicate the contents of BB up into PredBB do so now, this
1568 /// improves the odds that the branch will be on an analyzable instruction like
1570 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1571 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1572 assert(!PredBBs.empty() && "Can't handle an empty set");
1574 // If BB is a loop header, then duplicating this block outside the loop would
1575 // cause us to transform this into an irreducible loop, don't do this.
1576 // See the comments above FindLoopHeaders for justifications and caveats.
1577 if (LoopHeaders.count(BB)) {
1578 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1579 << "' into predecessor block '" << PredBBs[0]->getName()
1580 << "' - it might create an irreducible loop!\n");
1584 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1585 if (DuplicationCost > Threshold) {
1586 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1587 << "' - Cost is too high: " << DuplicationCost << "\n");
1591 // And finally, do it! Start by factoring the predecessors is needed.
1593 if (PredBBs.size() == 1)
1594 PredBB = PredBBs[0];
1596 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1597 << " common predecessors.\n");
1598 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1602 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1604 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1605 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1606 << DuplicationCost << " block is:" << *BB << "\n");
1608 // Unless PredBB ends with an unconditional branch, split the edge so that we
1609 // can just clone the bits from BB into the end of the new PredBB.
1610 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1612 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1613 PredBB = SplitEdge(PredBB, BB, this);
1614 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1617 // We are going to have to map operands from the original BB block into the
1618 // PredBB block. Evaluate PHI nodes in BB.
1619 DenseMap<Instruction*, Value*> ValueMapping;
1621 BasicBlock::iterator BI = BB->begin();
1622 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1623 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1625 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1626 // mapping and using it to remap operands in the cloned instructions.
1627 for (; BI != BB->end(); ++BI) {
1628 Instruction *New = BI->clone();
1630 // Remap operands to patch up intra-block references.
1631 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1632 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1633 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1634 if (I != ValueMapping.end())
1635 New->setOperand(i, I->second);
1638 // If this instruction can be simplified after the operands are updated,
1639 // just use the simplified value instead. This frequently happens due to
1641 if (Value *IV = SimplifyInstruction(New, TD)) {
1643 ValueMapping[BI] = IV;
1645 // Otherwise, insert the new instruction into the block.
1646 New->setName(BI->getName());
1647 PredBB->getInstList().insert(OldPredBranch, New);
1648 ValueMapping[BI] = New;
1652 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1653 // add entries to the PHI nodes for branch from PredBB now.
1654 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1655 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1657 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1660 // If there were values defined in BB that are used outside the block, then we
1661 // now have to update all uses of the value to use either the original value,
1662 // the cloned value, or some PHI derived value. This can require arbitrary
1663 // PHI insertion, of which we are prepared to do, clean these up now.
1664 SSAUpdater SSAUpdate;
1665 SmallVector<Use*, 16> UsesToRename;
1666 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1667 // Scan all uses of this instruction to see if it is used outside of its
1668 // block, and if so, record them in UsesToRename.
1669 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1671 Instruction *User = cast<Instruction>(*UI);
1672 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1673 if (UserPN->getIncomingBlock(UI) == BB)
1675 } else if (User->getParent() == BB)
1678 UsesToRename.push_back(&UI.getUse());
1681 // If there are no uses outside the block, we're done with this instruction.
1682 if (UsesToRename.empty())
1685 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1687 // We found a use of I outside of BB. Rename all uses of I that are outside
1688 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1689 // with the two values we know.
1690 SSAUpdate.Initialize(I);
1691 SSAUpdate.AddAvailableValue(BB, I);
1692 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1694 while (!UsesToRename.empty())
1695 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1696 DEBUG(dbgs() << "\n");
1699 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1701 RemovePredecessorAndSimplify(BB, PredBB, TD);
1703 // Remove the unconditional branch at the end of the PredBB block.
1704 OldPredBranch->eraseFromParent();