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/Statistic.h"
28 #include "llvm/ADT/STLExtras.h"
29 #include "llvm/ADT/SmallPtrSet.h"
30 #include "llvm/ADT/SmallSet.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ValueHandle.h"
34 #include "llvm/Support/raw_ostream.h"
37 STATISTIC(NumThreads, "Number of jumps threaded");
38 STATISTIC(NumFolds, "Number of terminators folded");
39 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
41 static cl::opt<unsigned>
42 Threshold("jump-threading-threshold",
43 cl::desc("Max block size to duplicate for jump threading"),
44 cl::init(6), cl::Hidden);
46 // Turn on use of LazyValueInfo.
48 EnableLVI("enable-jump-threading-lvi",
49 cl::desc("Use LVI for jump threading"),
56 /// This pass performs 'jump threading', which looks at blocks that have
57 /// multiple predecessors and multiple successors. If one or more of the
58 /// predecessors of the block can be proven to always jump to one of the
59 /// successors, we forward the edge from the predecessor to the successor by
60 /// duplicating the contents of this block.
62 /// An example of when this can occur is code like this:
69 /// In this case, the unconditional branch at the end of the first if can be
70 /// revectored to the false side of the second if.
72 class JumpThreading : public FunctionPass {
76 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
78 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
81 static char ID; // Pass identification
82 JumpThreading() : FunctionPass(ID) {}
84 bool runOnFunction(Function &F);
86 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
88 AU.addRequired<LazyValueInfo>();
91 void FindLoopHeaders(Function &F);
92 bool ProcessBlock(BasicBlock *BB);
93 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
95 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
96 const SmallVectorImpl<BasicBlock *> &PredBBs);
98 typedef SmallVectorImpl<std::pair<ConstantInt*,
99 BasicBlock*> > PredValueInfo;
101 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
102 PredValueInfo &Result);
103 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB);
106 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
107 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
109 bool ProcessBranchOnPHI(PHINode *PN);
110 bool ProcessBranchOnXOR(BinaryOperator *BO);
112 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
116 char JumpThreading::ID = 0;
117 INITIALIZE_PASS(JumpThreading, "jump-threading",
118 "Jump Threading", false, false);
120 // Public interface to the Jump Threading pass
121 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
123 /// runOnFunction - Top level algorithm.
125 bool JumpThreading::runOnFunction(Function &F) {
126 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
127 TD = getAnalysisIfAvailable<TargetData>();
128 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0;
132 bool Changed, EverChanged = false;
135 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
137 // Thread all of the branches we can over this block.
138 while (ProcessBlock(BB))
143 // If the block is trivially dead, zap it. This eliminates the successor
144 // edges which simplifies the CFG.
145 if (pred_begin(BB) == pred_end(BB) &&
146 BB != &BB->getParent()->getEntryBlock()) {
147 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
148 << "' with terminator: " << *BB->getTerminator() << '\n');
149 LoopHeaders.erase(BB);
150 if (LVI) LVI->eraseBlock(BB);
153 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
154 // Can't thread an unconditional jump, but if the block is "almost
155 // empty", we can replace uses of it with uses of the successor and make
157 if (BI->isUnconditional() &&
158 BB != &BB->getParent()->getEntryBlock()) {
159 BasicBlock::iterator BBI = BB->getFirstNonPHI();
160 // Ignore dbg intrinsics.
161 while (isa<DbgInfoIntrinsic>(BBI))
163 // If the terminator is the only non-phi instruction, try to nuke it.
164 if (BBI->isTerminator()) {
165 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
166 // block, we have to make sure it isn't in the LoopHeaders set. We
167 // reinsert afterward if needed.
168 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
169 BasicBlock *Succ = BI->getSuccessor(0);
171 // FIXME: It is always conservatively correct to drop the info
172 // for a block even if it doesn't get erased. This isn't totally
173 // awesome, but it allows us to use AssertingVH to prevent nasty
174 // dangling pointer issues within LazyValueInfo.
175 if (LVI) LVI->eraseBlock(BB);
176 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
178 // If we deleted BB and BB was the header of a loop, then the
179 // successor is now the header of the loop.
183 if (ErasedFromLoopHeaders)
184 LoopHeaders.insert(BB);
189 EverChanged |= Changed;
196 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
197 /// thread across it.
198 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
199 /// Ignore PHI nodes, these will be flattened when duplication happens.
200 BasicBlock::const_iterator I = BB->getFirstNonPHI();
202 // FIXME: THREADING will delete values that are just used to compute the
203 // branch, so they shouldn't count against the duplication cost.
206 // Sum up the cost of each instruction until we get to the terminator. Don't
207 // include the terminator because the copy won't include it.
209 for (; !isa<TerminatorInst>(I); ++I) {
210 // Debugger intrinsics don't incur code size.
211 if (isa<DbgInfoIntrinsic>(I)) continue;
213 // If this is a pointer->pointer bitcast, it is free.
214 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
217 // All other instructions count for at least one unit.
220 // Calls are more expensive. If they are non-intrinsic calls, we model them
221 // as having cost of 4. If they are a non-vector intrinsic, we model them
222 // as having cost of 2 total, and if they are a vector intrinsic, we model
223 // them as having cost 1.
224 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
225 if (!isa<IntrinsicInst>(CI))
227 else if (!CI->getType()->isVectorTy())
232 // Threading through a switch statement is particularly profitable. If this
233 // block ends in a switch, decrease its cost to make it more likely to happen.
234 if (isa<SwitchInst>(I))
235 Size = Size > 6 ? Size-6 : 0;
240 /// FindLoopHeaders - We do not want jump threading to turn proper loop
241 /// structures into irreducible loops. Doing this breaks up the loop nesting
242 /// hierarchy and pessimizes later transformations. To prevent this from
243 /// happening, we first have to find the loop headers. Here we approximate this
244 /// by finding targets of backedges in the CFG.
246 /// Note that there definitely are cases when we want to allow threading of
247 /// edges across a loop header. For example, threading a jump from outside the
248 /// loop (the preheader) to an exit block of the loop is definitely profitable.
249 /// It is also almost always profitable to thread backedges from within the loop
250 /// to exit blocks, and is often profitable to thread backedges to other blocks
251 /// within the loop (forming a nested loop). This simple analysis is not rich
252 /// enough to track all of these properties and keep it up-to-date as the CFG
253 /// mutates, so we don't allow any of these transformations.
255 void JumpThreading::FindLoopHeaders(Function &F) {
256 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
257 FindFunctionBackedges(F, Edges);
259 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
260 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
263 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
264 /// if we can infer that the value is a known ConstantInt in any of our
265 /// predecessors. If so, return the known list of value and pred BB in the
266 /// result vector. If a value is known to be undef, it is returned as null.
268 /// This returns true if there were any known values.
271 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){
272 // If V is a constantint, then it is known in all predecessors.
273 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) {
274 ConstantInt *CI = dyn_cast<ConstantInt>(V);
276 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
277 Result.push_back(std::make_pair(CI, *PI));
281 // If V is a non-instruction value, or an instruction in a different block,
282 // then it can't be derived from a PHI.
283 Instruction *I = dyn_cast<Instruction>(V);
284 if (I == 0 || I->getParent() != BB) {
286 // Okay, if this is a live-in value, see if it has a known value at the end
287 // of any of our predecessors.
289 // FIXME: This should be an edge property, not a block end property.
290 /// TODO: Per PR2563, we could infer value range information about a
291 /// predecessor based on its terminator.
294 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
295 // "I" is a non-local compare-with-a-constant instruction. This would be
296 // able to handle value inequalities better, for example if the compare is
297 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
298 // Perhaps getConstantOnEdge should be smart enough to do this?
300 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
302 // If the value is known by LazyValueInfo to be a constant in a
303 // predecessor, use that information to try to thread this block.
304 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
306 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst)))
309 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), P));
312 return !Result.empty();
318 /// If I is a PHI node, then we know the incoming values for any constants.
319 if (PHINode *PN = dyn_cast<PHINode>(I)) {
320 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
321 Value *InVal = PN->getIncomingValue(i);
322 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) {
323 ConstantInt *CI = dyn_cast<ConstantInt>(InVal);
324 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i)));
326 Constant *CI = LVI->getConstantOnEdge(InVal,
327 PN->getIncomingBlock(i), BB);
328 ConstantInt *CInt = dyn_cast_or_null<ConstantInt>(CI);
330 Result.push_back(std::make_pair(CInt, PN->getIncomingBlock(i)));
333 return !Result.empty();
336 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals;
338 // Handle some boolean conditions.
339 if (I->getType()->getPrimitiveSizeInBits() == 1) {
341 // X & false -> false
342 if (I->getOpcode() == Instruction::Or ||
343 I->getOpcode() == Instruction::And) {
344 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
345 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals);
347 if (LHSVals.empty() && RHSVals.empty())
350 ConstantInt *InterestingVal;
351 if (I->getOpcode() == Instruction::Or)
352 InterestingVal = ConstantInt::getTrue(I->getContext());
354 InterestingVal = ConstantInt::getFalse(I->getContext());
356 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
358 // Scan for the sentinel. If we find an undef, force it to the
359 // interesting value: x|undef -> true and x&undef -> false.
360 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
361 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) {
362 Result.push_back(LHSVals[i]);
363 Result.back().first = InterestingVal;
364 LHSKnownBBs.insert(LHSVals[i].second);
366 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
367 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) {
368 // If we already inferred a value for this block on the LHS, don't
370 if (!LHSKnownBBs.count(RHSVals[i].second)) {
371 Result.push_back(RHSVals[i]);
372 Result.back().first = InterestingVal;
375 return !Result.empty();
377 // Try to process a few other binary operator patterns.
378 } else if (isa<BinaryOperator>(I)) {
382 // Handle the NOT form of XOR.
383 if (I->getOpcode() == Instruction::Xor &&
384 isa<ConstantInt>(I->getOperand(1)) &&
385 cast<ConstantInt>(I->getOperand(1))->isOne()) {
386 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result);
390 // Invert the known values.
391 for (unsigned i = 0, e = Result.size(); i != e; ++i)
394 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first));
398 // Try to simplify some other binary operator values.
399 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
400 // AND or OR of a value with itself is that value.
401 ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1));
402 if (CI && (BO->getOpcode() == Instruction::And ||
403 BO->getOpcode() == Instruction::Or)) {
404 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals;
405 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals);
406 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
407 if (LHSVals[i].first == CI)
408 Result.push_back(std::make_pair(CI, LHSVals[i].second));
410 return !Result.empty();
414 // Handle compare with phi operand, where the PHI is defined in this block.
415 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
416 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
417 if (PN && PN->getParent() == BB) {
418 // We can do this simplification if any comparisons fold to true or false.
420 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
421 BasicBlock *PredBB = PN->getIncomingBlock(i);
422 Value *LHS = PN->getIncomingValue(i);
423 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
425 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
427 if (!LVI || !isa<Constant>(RHS))
430 LazyValueInfo::Tristate
431 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
432 cast<Constant>(RHS), PredBB, BB);
433 if (ResT == LazyValueInfo::Unknown)
435 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
438 if (isa<UndefValue>(Res))
439 Result.push_back(std::make_pair((ConstantInt*)0, PredBB));
440 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res))
441 Result.push_back(std::make_pair(CI, PredBB));
444 return !Result.empty();
448 // If comparing a live-in value against a constant, see if we know the
449 // live-in value on any predecessors.
450 if (LVI && isa<Constant>(Cmp->getOperand(1)) &&
451 Cmp->getType()->isIntegerTy()) {
452 if (!isa<Instruction>(Cmp->getOperand(0)) ||
453 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
454 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
456 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
458 // If the value is known by LazyValueInfo to be a constant in a
459 // predecessor, use that information to try to thread this block.
460 LazyValueInfo::Tristate Res =
461 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
463 if (Res == LazyValueInfo::Unknown)
466 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
467 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), P));
470 return !Result.empty();
473 // Try to find a constant value for the LHS of an equality comparison,
474 // and evaluate it statically if we can.
475 if (Cmp->getPredicate() == CmpInst::ICMP_EQ ||
476 Cmp->getPredicate() == CmpInst::ICMP_NE) {
477 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals;
478 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
480 ConstantInt *True = ConstantInt::getTrue(I->getContext());
481 ConstantInt *False = ConstantInt::getFalse(I->getContext());
482 if (Cmp->getPredicate() == CmpInst::ICMP_NE) std::swap(True, False);
484 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
485 if (LHSVals[i].first == Cmp->getOperand(1))
486 Result.push_back(std::make_pair(True, LHSVals[i].second));
488 Result.push_back(std::make_pair(False, LHSVals[i].second));
491 return !Result.empty();
497 // If all else fails, see if LVI can figure out a constant value for us.
498 Constant *CI = LVI->getConstant(V, BB);
499 ConstantInt *CInt = dyn_cast_or_null<ConstantInt>(CI);
501 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
502 Result.push_back(std::make_pair(CInt, *PI));
505 return !Result.empty();
513 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
514 /// in an undefined jump, decide which block is best to revector to.
516 /// Since we can pick an arbitrary destination, we pick the successor with the
517 /// fewest predecessors. This should reduce the in-degree of the others.
519 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
520 TerminatorInst *BBTerm = BB->getTerminator();
521 unsigned MinSucc = 0;
522 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
523 // Compute the successor with the minimum number of predecessors.
524 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
525 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
526 TestBB = BBTerm->getSuccessor(i);
527 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
528 if (NumPreds < MinNumPreds)
535 /// ProcessBlock - If there are any predecessors whose control can be threaded
536 /// through to a successor, transform them now.
537 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
538 // If the block is trivially dead, just return and let the caller nuke it.
539 // This simplifies other transformations.
540 if (pred_begin(BB) == pred_end(BB) &&
541 BB != &BB->getParent()->getEntryBlock())
544 // If this block has a single predecessor, and if that pred has a single
545 // successor, merge the blocks. This encourages recursive jump threading
546 // because now the condition in this block can be threaded through
547 // predecessors of our predecessor block.
548 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
549 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
551 // If SinglePred was a loop header, BB becomes one.
552 if (LoopHeaders.erase(SinglePred))
553 LoopHeaders.insert(BB);
555 // Remember if SinglePred was the entry block of the function. If so, we
556 // will need to move BB back to the entry position.
557 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
558 if (LVI) LVI->eraseBlock(SinglePred);
559 MergeBasicBlockIntoOnlyPred(BB);
561 if (isEntry && BB != &BB->getParent()->getEntryBlock())
562 BB->moveBefore(&BB->getParent()->getEntryBlock());
567 // Look to see if the terminator is a branch of switch, if not we can't thread
570 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
571 // Can't thread an unconditional jump.
572 if (BI->isUnconditional()) return false;
573 Condition = BI->getCondition();
574 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
575 Condition = SI->getCondition();
577 return false; // Must be an invoke.
579 // If the terminator of this block is branching on a constant, simplify the
580 // terminator to an unconditional branch. This can occur due to threading in
582 if (isa<ConstantInt>(Condition)) {
583 DEBUG(dbgs() << " In block '" << BB->getName()
584 << "' folding terminator: " << *BB->getTerminator() << '\n');
586 ConstantFoldTerminator(BB);
590 // If the terminator is branching on an undef, we can pick any of the
591 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
592 if (isa<UndefValue>(Condition)) {
593 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
595 // Fold the branch/switch.
596 TerminatorInst *BBTerm = BB->getTerminator();
597 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
598 if (i == BestSucc) continue;
599 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD);
602 DEBUG(dbgs() << " In block '" << BB->getName()
603 << "' folding undef terminator: " << *BBTerm << '\n');
604 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
605 BBTerm->eraseFromParent();
609 Instruction *CondInst = dyn_cast<Instruction>(Condition);
611 // If the condition is an instruction defined in another block, see if a
612 // predecessor has the same condition:
617 !Condition->hasOneUse() && // Multiple uses.
618 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition.
619 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
620 if (isa<BranchInst>(BB->getTerminator())) {
621 for (; PI != E; ++PI) {
623 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator()))
624 if (PBI->isConditional() && PBI->getCondition() == Condition &&
625 ProcessBranchOnDuplicateCond(P, BB))
629 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator");
630 for (; PI != E; ++PI) {
632 if (SwitchInst *PSI = dyn_cast<SwitchInst>(P->getTerminator()))
633 if (PSI->getCondition() == Condition &&
634 ProcessSwitchOnDuplicateCond(P, BB))
640 // All the rest of our checks depend on the condition being an instruction.
642 // FIXME: Unify this with code below.
643 if (LVI && ProcessThreadableEdges(Condition, BB))
649 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
651 (!isa<PHINode>(CondCmp->getOperand(0)) ||
652 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) {
653 // If we have a comparison, loop over the predecessors to see if there is
654 // a condition with a lexically identical value.
655 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
656 for (; PI != E; ++PI) {
658 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator()))
659 if (PBI->isConditional() && P != BB) {
660 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) {
661 if (CI->getOperand(0) == CondCmp->getOperand(0) &&
662 CI->getOperand(1) == CondCmp->getOperand(1) &&
663 CI->getPredicate() == CondCmp->getPredicate()) {
664 // TODO: Could handle things like (x != 4) --> (x == 17)
665 if (ProcessBranchOnDuplicateCond(P, BB))
673 // For a comparison where the LHS is outside this block, it's possible
674 // that we've branched on it before. Used LVI to see if we can simplify
675 // the branch based on that.
676 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
677 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
678 if (LVI && CondBr && CondConst && CondBr->isConditional() &&
679 (!isa<Instruction>(CondCmp->getOperand(0)) ||
680 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
681 // For predecessor edge, determine if the comparison is true or false
682 // on that edge. If they're all true or all false, we can simplify the
684 // FIXME: We could handle mixed true/false by duplicating code.
685 unsigned Trues = 0, Falses = 0, predcount = 0;
686 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB);PI != PE; ++PI){
688 LazyValueInfo::Tristate Ret =
689 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
690 CondCmp->getOperand(0), CondConst, *PI, BB);
691 if (Ret == LazyValueInfo::True)
693 else if (Ret == LazyValueInfo::False)
697 // If we can determine the branch direction statically, convert
698 // the conditional branch to an unconditional one.
699 if (Trues && Trues == predcount) {
700 RemovePredecessorAndSimplify(CondBr->getSuccessor(1), BB, TD);
701 BranchInst::Create(CondBr->getSuccessor(0), CondBr);
702 CondBr->eraseFromParent();
704 } else if (Falses && Falses == predcount) {
705 RemovePredecessorAndSimplify(CondBr->getSuccessor(0), BB, TD);
706 BranchInst::Create(CondBr->getSuccessor(1), CondBr);
707 CondBr->eraseFromParent();
713 // Check for some cases that are worth simplifying. Right now we want to look
714 // for loads that are used by a switch or by the condition for the branch. If
715 // we see one, check to see if it's partially redundant. If so, insert a PHI
716 // which can then be used to thread the values.
718 Value *SimplifyValue = CondInst;
719 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
720 if (isa<Constant>(CondCmp->getOperand(1)))
721 SimplifyValue = CondCmp->getOperand(0);
723 // TODO: There are other places where load PRE would be profitable, such as
724 // more complex comparisons.
725 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
726 if (SimplifyPartiallyRedundantLoad(LI))
730 // Handle a variety of cases where we are branching on something derived from
731 // a PHI node in the current block. If we can prove that any predecessors
732 // compute a predictable value based on a PHI node, thread those predecessors.
734 if (ProcessThreadableEdges(CondInst, BB))
737 // If this is an otherwise-unfoldable branch on a phi node in the current
738 // block, see if we can simplify.
739 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
740 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
741 return ProcessBranchOnPHI(PN);
744 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
745 if (CondInst->getOpcode() == Instruction::Xor &&
746 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
747 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
750 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
751 // "(X == 4)", thread through this block.
756 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
757 /// block that jump on exactly the same condition. This means that we almost
758 /// always know the direction of the edge in the DESTBB:
760 /// br COND, DESTBB, BBY
762 /// br COND, BBZ, BBW
764 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
765 /// in DESTBB, we have to thread over it.
766 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
768 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
770 // If both successors of PredBB go to DESTBB, we don't know anything. We can
771 // fold the branch to an unconditional one, which allows other recursive
774 if (PredBI->getSuccessor(1) != BB)
776 else if (PredBI->getSuccessor(0) != BB)
779 DEBUG(dbgs() << " In block '" << PredBB->getName()
780 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
782 ConstantFoldTerminator(PredBB);
786 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
788 // If the dest block has one predecessor, just fix the branch condition to a
789 // constant and fold it.
790 if (BB->getSinglePredecessor()) {
791 DEBUG(dbgs() << " In block '" << BB->getName()
792 << "' folding condition to '" << BranchDir << "': "
793 << *BB->getTerminator() << '\n');
795 Value *OldCond = DestBI->getCondition();
796 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
798 // Delete dead instructions before we fold the branch. Folding the branch
799 // can eliminate edges from the CFG which can end up deleting OldCond.
800 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
801 ConstantFoldTerminator(BB);
806 // Next, figure out which successor we are threading to.
807 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
809 SmallVector<BasicBlock*, 2> Preds;
810 Preds.push_back(PredBB);
812 // Ok, try to thread it!
813 return ThreadEdge(BB, Preds, SuccBB);
816 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
817 /// block that switch on exactly the same condition. This means that we almost
818 /// always know the direction of the edge in the DESTBB:
820 /// switch COND [... DESTBB, BBY ... ]
822 /// switch COND [... BBZ, BBW ]
824 /// Optimizing switches like this is very important, because simplifycfg builds
825 /// switches out of repeated 'if' conditions.
826 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
827 BasicBlock *DestBB) {
828 // Can't thread edge to self.
829 if (PredBB == DestBB)
832 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
833 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
835 // There are a variety of optimizations that we can potentially do on these
836 // blocks: we order them from most to least preferable.
838 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
839 // directly to their destination. This does not introduce *any* code size
840 // growth. Skip debug info first.
841 BasicBlock::iterator BBI = DestBB->begin();
842 while (isa<DbgInfoIntrinsic>(BBI))
845 // FIXME: Thread if it just contains a PHI.
846 if (isa<SwitchInst>(BBI)) {
847 bool MadeChange = false;
848 // Ignore the default edge for now.
849 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
850 ConstantInt *DestVal = DestSI->getCaseValue(i);
851 BasicBlock *DestSucc = DestSI->getSuccessor(i);
853 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
854 // PredSI has an explicit case for it. If so, forward. If it is covered
855 // by the default case, we can't update PredSI.
856 unsigned PredCase = PredSI->findCaseValue(DestVal);
857 if (PredCase == 0) continue;
859 // If PredSI doesn't go to DestBB on this value, then it won't reach the
860 // case on this condition.
861 if (PredSI->getSuccessor(PredCase) != DestBB &&
862 DestSI->getSuccessor(i) != DestBB)
865 // Do not forward this if it already goes to this destination, this would
866 // be an infinite loop.
867 if (PredSI->getSuccessor(PredCase) == DestSucc)
870 // Otherwise, we're safe to make the change. Make sure that the edge from
871 // DestSI to DestSucc is not critical and has no PHI nodes.
872 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
873 DEBUG(dbgs() << "THROUGH: " << *DestSI);
875 // If the destination has PHI nodes, just split the edge for updating
877 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
878 SplitCriticalEdge(DestSI, i, this);
879 DestSucc = DestSI->getSuccessor(i);
881 FoldSingleEntryPHINodes(DestSucc);
882 PredSI->setSuccessor(PredCase, DestSucc);
894 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
895 /// load instruction, eliminate it by replacing it with a PHI node. This is an
896 /// important optimization that encourages jump threading, and needs to be run
897 /// interlaced with other jump threading tasks.
898 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
899 // Don't hack volatile loads.
900 if (LI->isVolatile()) return false;
902 // If the load is defined in a block with exactly one predecessor, it can't be
903 // partially redundant.
904 BasicBlock *LoadBB = LI->getParent();
905 if (LoadBB->getSinglePredecessor())
908 Value *LoadedPtr = LI->getOperand(0);
910 // If the loaded operand is defined in the LoadBB, it can't be available.
911 // TODO: Could do simple PHI translation, that would be fun :)
912 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
913 if (PtrOp->getParent() == LoadBB)
916 // Scan a few instructions up from the load, to see if it is obviously live at
917 // the entry to its block.
918 BasicBlock::iterator BBIt = LI;
920 if (Value *AvailableVal =
921 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
922 // If the value if the load is locally available within the block, just use
923 // it. This frequently occurs for reg2mem'd allocas.
924 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
926 // If the returned value is the load itself, replace with an undef. This can
927 // only happen in dead loops.
928 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
929 LI->replaceAllUsesWith(AvailableVal);
930 LI->eraseFromParent();
934 // Otherwise, if we scanned the whole block and got to the top of the block,
935 // we know the block is locally transparent to the load. If not, something
936 // might clobber its value.
937 if (BBIt != LoadBB->begin())
941 SmallPtrSet<BasicBlock*, 8> PredsScanned;
942 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
943 AvailablePredsTy AvailablePreds;
944 BasicBlock *OneUnavailablePred = 0;
946 // If we got here, the loaded value is transparent through to the start of the
947 // block. Check to see if it is available in any of the predecessor blocks.
948 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
950 BasicBlock *PredBB = *PI;
952 // If we already scanned this predecessor, skip it.
953 if (!PredsScanned.insert(PredBB))
956 // Scan the predecessor to see if the value is available in the pred.
957 BBIt = PredBB->end();
958 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
959 if (!PredAvailable) {
960 OneUnavailablePred = PredBB;
964 // If so, this load is partially redundant. Remember this info so that we
965 // can create a PHI node.
966 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
969 // If the loaded value isn't available in any predecessor, it isn't partially
971 if (AvailablePreds.empty()) return false;
973 // Okay, the loaded value is available in at least one (and maybe all!)
974 // predecessors. If the value is unavailable in more than one unique
975 // predecessor, we want to insert a merge block for those common predecessors.
976 // This ensures that we only have to insert one reload, thus not increasing
978 BasicBlock *UnavailablePred = 0;
980 // If there is exactly one predecessor where the value is unavailable, the
981 // already computed 'OneUnavailablePred' block is it. If it ends in an
982 // unconditional branch, we know that it isn't a critical edge.
983 if (PredsScanned.size() == AvailablePreds.size()+1 &&
984 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
985 UnavailablePred = OneUnavailablePred;
986 } else if (PredsScanned.size() != AvailablePreds.size()) {
987 // Otherwise, we had multiple unavailable predecessors or we had a critical
988 // edge from the one.
989 SmallVector<BasicBlock*, 8> PredsToSplit;
990 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
992 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
993 AvailablePredSet.insert(AvailablePreds[i].first);
995 // Add all the unavailable predecessors to the PredsToSplit list.
996 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
999 // If the predecessor is an indirect goto, we can't split the edge.
1000 if (isa<IndirectBrInst>(P->getTerminator()))
1003 if (!AvailablePredSet.count(P))
1004 PredsToSplit.push_back(P);
1007 // Split them out to their own block.
1009 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
1010 "thread-pre-split", this);
1013 // If the value isn't available in all predecessors, then there will be
1014 // exactly one where it isn't available. Insert a load on that edge and add
1015 // it to the AvailablePreds list.
1016 if (UnavailablePred) {
1017 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1018 "Can't handle critical edge here!");
1019 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
1021 UnavailablePred->getTerminator());
1022 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1025 // Now we know that each predecessor of this block has a value in
1026 // AvailablePreds, sort them for efficient access as we're walking the preds.
1027 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1029 // Create a PHI node at the start of the block for the PRE'd load value.
1030 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
1033 // Insert new entries into the PHI for each predecessor. A single block may
1034 // have multiple entries here.
1035 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
1037 BasicBlock *P = *PI;
1038 AvailablePredsTy::iterator I =
1039 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1040 std::make_pair(P, (Value*)0));
1042 assert(I != AvailablePreds.end() && I->first == P &&
1043 "Didn't find entry for predecessor!");
1045 PN->addIncoming(I->second, I->first);
1048 //cerr << "PRE: " << *LI << *PN << "\n";
1050 LI->replaceAllUsesWith(PN);
1051 LI->eraseFromParent();
1056 /// FindMostPopularDest - The specified list contains multiple possible
1057 /// threadable destinations. Pick the one that occurs the most frequently in
1060 FindMostPopularDest(BasicBlock *BB,
1061 const SmallVectorImpl<std::pair<BasicBlock*,
1062 BasicBlock*> > &PredToDestList) {
1063 assert(!PredToDestList.empty());
1065 // Determine popularity. If there are multiple possible destinations, we
1066 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1067 // blocks with known and real destinations to threading undef. We'll handle
1068 // them later if interesting.
1069 DenseMap<BasicBlock*, unsigned> DestPopularity;
1070 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1071 if (PredToDestList[i].second)
1072 DestPopularity[PredToDestList[i].second]++;
1074 // Find the most popular dest.
1075 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1076 BasicBlock *MostPopularDest = DPI->first;
1077 unsigned Popularity = DPI->second;
1078 SmallVector<BasicBlock*, 4> SamePopularity;
1080 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1081 // If the popularity of this entry isn't higher than the popularity we've
1082 // seen so far, ignore it.
1083 if (DPI->second < Popularity)
1085 else if (DPI->second == Popularity) {
1086 // If it is the same as what we've seen so far, keep track of it.
1087 SamePopularity.push_back(DPI->first);
1089 // If it is more popular, remember it.
1090 SamePopularity.clear();
1091 MostPopularDest = DPI->first;
1092 Popularity = DPI->second;
1096 // Okay, now we know the most popular destination. If there is more than
1097 // destination, we need to determine one. This is arbitrary, but we need
1098 // to make a deterministic decision. Pick the first one that appears in the
1100 if (!SamePopularity.empty()) {
1101 SamePopularity.push_back(MostPopularDest);
1102 TerminatorInst *TI = BB->getTerminator();
1103 for (unsigned i = 0; ; ++i) {
1104 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1106 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1107 TI->getSuccessor(i)) == SamePopularity.end())
1110 MostPopularDest = TI->getSuccessor(i);
1115 // Okay, we have finally picked the most popular destination.
1116 return MostPopularDest;
1119 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) {
1120 // If threading this would thread across a loop header, don't even try to
1122 if (LoopHeaders.count(BB))
1125 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues;
1126 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues))
1128 assert(!PredValues.empty() &&
1129 "ComputeValueKnownInPredecessors returned true with no values");
1131 DEBUG(dbgs() << "IN BB: " << *BB;
1132 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1133 dbgs() << " BB '" << BB->getName() << "': FOUND condition = ";
1134 if (PredValues[i].first)
1135 dbgs() << *PredValues[i].first;
1138 dbgs() << " for pred '" << PredValues[i].second->getName()
1142 // Decide what we want to thread through. Convert our list of known values to
1143 // a list of known destinations for each pred. This also discards duplicate
1144 // predecessors and keeps track of the undefined inputs (which are represented
1145 // as a null dest in the PredToDestList).
1146 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1147 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1149 BasicBlock *OnlyDest = 0;
1150 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1152 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1153 BasicBlock *Pred = PredValues[i].second;
1154 if (!SeenPreds.insert(Pred))
1155 continue; // Duplicate predecessor entry.
1157 // If the predecessor ends with an indirect goto, we can't change its
1159 if (isa<IndirectBrInst>(Pred->getTerminator()))
1162 ConstantInt *Val = PredValues[i].first;
1165 if (Val == 0) // Undef.
1167 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1168 DestBB = BI->getSuccessor(Val->isZero());
1170 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
1171 DestBB = SI->getSuccessor(SI->findCaseValue(Val));
1174 // If we have exactly one destination, remember it for efficiency below.
1177 else if (OnlyDest != DestBB)
1178 OnlyDest = MultipleDestSentinel;
1180 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1183 // If all edges were unthreadable, we fail.
1184 if (PredToDestList.empty())
1187 // Determine which is the most common successor. If we have many inputs and
1188 // this block is a switch, we want to start by threading the batch that goes
1189 // to the most popular destination first. If we only know about one
1190 // threadable destination (the common case) we can avoid this.
1191 BasicBlock *MostPopularDest = OnlyDest;
1193 if (MostPopularDest == MultipleDestSentinel)
1194 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1196 // Now that we know what the most popular destination is, factor all
1197 // predecessors that will jump to it into a single predecessor.
1198 SmallVector<BasicBlock*, 16> PredsToFactor;
1199 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1200 if (PredToDestList[i].second == MostPopularDest) {
1201 BasicBlock *Pred = PredToDestList[i].first;
1203 // This predecessor may be a switch or something else that has multiple
1204 // edges to the block. Factor each of these edges by listing them
1205 // according to # occurrences in PredsToFactor.
1206 TerminatorInst *PredTI = Pred->getTerminator();
1207 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1208 if (PredTI->getSuccessor(i) == BB)
1209 PredsToFactor.push_back(Pred);
1212 // If the threadable edges are branching on an undefined value, we get to pick
1213 // the destination that these predecessors should get to.
1214 if (MostPopularDest == 0)
1215 MostPopularDest = BB->getTerminator()->
1216 getSuccessor(GetBestDestForJumpOnUndef(BB));
1218 // Ok, try to thread it!
1219 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1222 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1223 /// a PHI node in the current block. See if there are any simplifications we
1224 /// can do based on inputs to the phi node.
1226 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1227 BasicBlock *BB = PN->getParent();
1229 // TODO: We could make use of this to do it once for blocks with common PHI
1231 SmallVector<BasicBlock*, 1> PredBBs;
1234 // If any of the predecessor blocks end in an unconditional branch, we can
1235 // *duplicate* the conditional branch into that block in order to further
1236 // encourage jump threading and to eliminate cases where we have branch on a
1237 // phi of an icmp (branch on icmp is much better).
1238 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1239 BasicBlock *PredBB = PN->getIncomingBlock(i);
1240 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1241 if (PredBr->isUnconditional()) {
1242 PredBBs[0] = PredBB;
1243 // Try to duplicate BB into PredBB.
1244 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1252 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1253 /// a xor instruction in the current block. See if there are any
1254 /// simplifications we can do based on inputs to the xor.
1256 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1257 BasicBlock *BB = BO->getParent();
1259 // If either the LHS or RHS of the xor is a constant, don't do this
1261 if (isa<ConstantInt>(BO->getOperand(0)) ||
1262 isa<ConstantInt>(BO->getOperand(1)))
1265 // If the first instruction in BB isn't a phi, we won't be able to infer
1266 // anything special about any particular predecessor.
1267 if (!isa<PHINode>(BB->front()))
1270 // If we have a xor as the branch input to this block, and we know that the
1271 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1272 // the condition into the predecessor and fix that value to true, saving some
1273 // logical ops on that path and encouraging other paths to simplify.
1275 // This copies something like this:
1278 // %X = phi i1 [1], [%X']
1279 // %Y = icmp eq i32 %A, %B
1280 // %Z = xor i1 %X, %Y
1285 // %Y = icmp ne i32 %A, %B
1288 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues;
1290 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) {
1291 assert(XorOpValues.empty());
1292 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues))
1297 assert(!XorOpValues.empty() &&
1298 "ComputeValueKnownInPredecessors returned true with no values");
1300 // Scan the information to see which is most popular: true or false. The
1301 // predecessors can be of the set true, false, or undef.
1302 unsigned NumTrue = 0, NumFalse = 0;
1303 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1304 if (!XorOpValues[i].first) continue; // Ignore undefs for the count.
1305 if (XorOpValues[i].first->isZero())
1311 // Determine which value to split on, true, false, or undef if neither.
1312 ConstantInt *SplitVal = 0;
1313 if (NumTrue > NumFalse)
1314 SplitVal = ConstantInt::getTrue(BB->getContext());
1315 else if (NumTrue != 0 || NumFalse != 0)
1316 SplitVal = ConstantInt::getFalse(BB->getContext());
1318 // Collect all of the blocks that this can be folded into so that we can
1319 // factor this once and clone it once.
1320 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1321 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1322 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue;
1324 BlocksToFoldInto.push_back(XorOpValues[i].second);
1327 // If we inferred a value for all of the predecessors, then duplication won't
1328 // help us. However, we can just replace the LHS or RHS with the constant.
1329 if (BlocksToFoldInto.size() ==
1330 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1331 if (SplitVal == 0) {
1332 // If all preds provide undef, just nuke the xor, because it is undef too.
1333 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1334 BO->eraseFromParent();
1335 } else if (SplitVal->isZero()) {
1336 // If all preds provide 0, replace the xor with the other input.
1337 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1338 BO->eraseFromParent();
1340 // If all preds provide 1, set the computed value to 1.
1341 BO->setOperand(!isLHS, SplitVal);
1347 // Try to duplicate BB into PredBB.
1348 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1352 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1353 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1354 /// NewPred using the entries from OldPred (suitably mapped).
1355 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1356 BasicBlock *OldPred,
1357 BasicBlock *NewPred,
1358 DenseMap<Instruction*, Value*> &ValueMap) {
1359 for (BasicBlock::iterator PNI = PHIBB->begin();
1360 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1361 // Ok, we have a PHI node. Figure out what the incoming value was for the
1363 Value *IV = PN->getIncomingValueForBlock(OldPred);
1365 // Remap the value if necessary.
1366 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1367 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1368 if (I != ValueMap.end())
1372 PN->addIncoming(IV, NewPred);
1376 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1377 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1378 /// across BB. Transform the IR to reflect this change.
1379 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1380 const SmallVectorImpl<BasicBlock*> &PredBBs,
1381 BasicBlock *SuccBB) {
1382 // If threading to the same block as we come from, we would infinite loop.
1384 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1385 << "' - would thread to self!\n");
1389 // If threading this would thread across a loop header, don't thread the edge.
1390 // See the comments above FindLoopHeaders for justifications and caveats.
1391 if (LoopHeaders.count(BB)) {
1392 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1393 << "' to dest BB '" << SuccBB->getName()
1394 << "' - it might create an irreducible loop!\n");
1398 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1399 if (JumpThreadCost > Threshold) {
1400 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1401 << "' - Cost is too high: " << JumpThreadCost << "\n");
1405 // And finally, do it! Start by factoring the predecessors is needed.
1407 if (PredBBs.size() == 1)
1408 PredBB = PredBBs[0];
1410 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1411 << " common predecessors.\n");
1412 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1416 // And finally, do it!
1417 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1418 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1419 << ", across block:\n "
1423 LVI->threadEdge(PredBB, BB, SuccBB);
1425 // We are going to have to map operands from the original BB block to the new
1426 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1427 // account for entry from PredBB.
1428 DenseMap<Instruction*, Value*> ValueMapping;
1430 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1431 BB->getName()+".thread",
1432 BB->getParent(), BB);
1433 NewBB->moveAfter(PredBB);
1435 BasicBlock::iterator BI = BB->begin();
1436 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1437 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1439 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1440 // mapping and using it to remap operands in the cloned instructions.
1441 for (; !isa<TerminatorInst>(BI); ++BI) {
1442 Instruction *New = BI->clone();
1443 New->setName(BI->getName());
1444 NewBB->getInstList().push_back(New);
1445 ValueMapping[BI] = New;
1447 // Remap operands to patch up intra-block references.
1448 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1449 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1450 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1451 if (I != ValueMapping.end())
1452 New->setOperand(i, I->second);
1456 // We didn't copy the terminator from BB over to NewBB, because there is now
1457 // an unconditional jump to SuccBB. Insert the unconditional jump.
1458 BranchInst::Create(SuccBB, NewBB);
1460 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1461 // PHI nodes for NewBB now.
1462 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1464 // If there were values defined in BB that are used outside the block, then we
1465 // now have to update all uses of the value to use either the original value,
1466 // the cloned value, or some PHI derived value. This can require arbitrary
1467 // PHI insertion, of which we are prepared to do, clean these up now.
1468 SSAUpdater SSAUpdate;
1469 SmallVector<Use*, 16> UsesToRename;
1470 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1471 // Scan all uses of this instruction to see if it is used outside of its
1472 // block, and if so, record them in UsesToRename.
1473 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1475 Instruction *User = cast<Instruction>(*UI);
1476 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1477 if (UserPN->getIncomingBlock(UI) == BB)
1479 } else if (User->getParent() == BB)
1482 UsesToRename.push_back(&UI.getUse());
1485 // If there are no uses outside the block, we're done with this instruction.
1486 if (UsesToRename.empty())
1489 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1491 // We found a use of I outside of BB. Rename all uses of I that are outside
1492 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1493 // with the two values we know.
1494 SSAUpdate.Initialize(I);
1495 SSAUpdate.AddAvailableValue(BB, I);
1496 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1498 while (!UsesToRename.empty())
1499 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1500 DEBUG(dbgs() << "\n");
1504 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1505 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1506 // us to simplify any PHI nodes in BB.
1507 TerminatorInst *PredTerm = PredBB->getTerminator();
1508 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1509 if (PredTerm->getSuccessor(i) == BB) {
1510 RemovePredecessorAndSimplify(BB, PredBB, TD);
1511 PredTerm->setSuccessor(i, NewBB);
1514 // At this point, the IR is fully up to date and consistent. Do a quick scan
1515 // over the new instructions and zap any that are constants or dead. This
1516 // frequently happens because of phi translation.
1517 SimplifyInstructionsInBlock(NewBB, TD);
1519 // Threaded an edge!
1524 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1525 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1526 /// If we can duplicate the contents of BB up into PredBB do so now, this
1527 /// improves the odds that the branch will be on an analyzable instruction like
1529 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1530 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1531 assert(!PredBBs.empty() && "Can't handle an empty set");
1533 // If BB is a loop header, then duplicating this block outside the loop would
1534 // cause us to transform this into an irreducible loop, don't do this.
1535 // See the comments above FindLoopHeaders for justifications and caveats.
1536 if (LoopHeaders.count(BB)) {
1537 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1538 << "' into predecessor block '" << PredBBs[0]->getName()
1539 << "' - it might create an irreducible loop!\n");
1543 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1544 if (DuplicationCost > Threshold) {
1545 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1546 << "' - Cost is too high: " << DuplicationCost << "\n");
1550 // And finally, do it! Start by factoring the predecessors is needed.
1552 if (PredBBs.size() == 1)
1553 PredBB = PredBBs[0];
1555 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1556 << " common predecessors.\n");
1557 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1561 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1563 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1564 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1565 << DuplicationCost << " block is:" << *BB << "\n");
1567 // Unless PredBB ends with an unconditional branch, split the edge so that we
1568 // can just clone the bits from BB into the end of the new PredBB.
1569 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1571 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1572 PredBB = SplitEdge(PredBB, BB, this);
1573 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1576 // We are going to have to map operands from the original BB block into the
1577 // PredBB block. Evaluate PHI nodes in BB.
1578 DenseMap<Instruction*, Value*> ValueMapping;
1580 BasicBlock::iterator BI = BB->begin();
1581 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1582 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1584 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1585 // mapping and using it to remap operands in the cloned instructions.
1586 for (; BI != BB->end(); ++BI) {
1587 Instruction *New = BI->clone();
1589 // Remap operands to patch up intra-block references.
1590 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1591 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1592 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1593 if (I != ValueMapping.end())
1594 New->setOperand(i, I->second);
1597 // If this instruction can be simplified after the operands are updated,
1598 // just use the simplified value instead. This frequently happens due to
1600 if (Value *IV = SimplifyInstruction(New, TD)) {
1602 ValueMapping[BI] = IV;
1604 // Otherwise, insert the new instruction into the block.
1605 New->setName(BI->getName());
1606 PredBB->getInstList().insert(OldPredBranch, New);
1607 ValueMapping[BI] = New;
1611 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1612 // add entries to the PHI nodes for branch from PredBB now.
1613 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1614 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1616 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1619 // If there were values defined in BB that are used outside the block, then we
1620 // now have to update all uses of the value to use either the original value,
1621 // the cloned value, or some PHI derived value. This can require arbitrary
1622 // PHI insertion, of which we are prepared to do, clean these up now.
1623 SSAUpdater SSAUpdate;
1624 SmallVector<Use*, 16> UsesToRename;
1625 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1626 // Scan all uses of this instruction to see if it is used outside of its
1627 // block, and if so, record them in UsesToRename.
1628 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1630 Instruction *User = cast<Instruction>(*UI);
1631 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1632 if (UserPN->getIncomingBlock(UI) == BB)
1634 } else if (User->getParent() == BB)
1637 UsesToRename.push_back(&UI.getUse());
1640 // If there are no uses outside the block, we're done with this instruction.
1641 if (UsesToRename.empty())
1644 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1646 // We found a use of I outside of BB. Rename all uses of I that are outside
1647 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1648 // with the two values we know.
1649 SSAUpdate.Initialize(I);
1650 SSAUpdate.AddAvailableValue(BB, I);
1651 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1653 while (!UsesToRename.empty())
1654 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1655 DEBUG(dbgs() << "\n");
1658 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1660 RemovePredecessorAndSimplify(BB, PredBB, TD);
1662 // Remove the unconditional branch at the end of the PredBB block.
1663 OldPredBranch->eraseFromParent();