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 // RAII helper for updating the recursion stack.
84 struct RecursionSetRemover {
85 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet;
86 std::pair<Value*, BasicBlock*> ThePair;
88 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S,
89 std::pair<Value*, BasicBlock*> P)
90 : TheSet(S), ThePair(P) { }
92 ~RecursionSetRemover() {
93 TheSet.erase(ThePair);
97 static char ID; // Pass identification
98 JumpThreading() : FunctionPass(ID) {}
100 bool runOnFunction(Function &F);
102 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
104 AU.addRequired<LazyValueInfo>();
105 AU.addPreserved<LazyValueInfo>();
109 void FindLoopHeaders(Function &F);
110 bool ProcessBlock(BasicBlock *BB);
111 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
113 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
114 const SmallVectorImpl<BasicBlock *> &PredBBs);
116 typedef SmallVectorImpl<std::pair<ConstantInt*,
117 BasicBlock*> > PredValueInfo;
119 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
120 PredValueInfo &Result);
121 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB);
124 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
125 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
127 bool ProcessBranchOnPHI(PHINode *PN);
128 bool ProcessBranchOnXOR(BinaryOperator *BO);
130 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
134 char JumpThreading::ID = 0;
135 INITIALIZE_PASS(JumpThreading, "jump-threading",
136 "Jump Threading", false, false);
138 // Public interface to the Jump Threading pass
139 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
141 /// runOnFunction - Top level algorithm.
143 bool JumpThreading::runOnFunction(Function &F) {
144 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
145 TD = getAnalysisIfAvailable<TargetData>();
146 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0;
150 bool Changed, EverChanged = false;
153 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
155 // Thread all of the branches we can over this block.
156 while (ProcessBlock(BB))
161 // If the block is trivially dead, zap it. This eliminates the successor
162 // edges which simplifies the CFG.
163 if (pred_begin(BB) == pred_end(BB) &&
164 BB != &BB->getParent()->getEntryBlock()) {
165 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
166 << "' with terminator: " << *BB->getTerminator() << '\n');
167 LoopHeaders.erase(BB);
168 if (LVI) LVI->eraseBlock(BB);
171 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
172 // Can't thread an unconditional jump, but if the block is "almost
173 // empty", we can replace uses of it with uses of the successor and make
175 if (BI->isUnconditional() &&
176 BB != &BB->getParent()->getEntryBlock()) {
177 BasicBlock::iterator BBI = BB->getFirstNonPHI();
178 // Ignore dbg intrinsics.
179 while (isa<DbgInfoIntrinsic>(BBI))
181 // If the terminator is the only non-phi instruction, try to nuke it.
182 if (BBI->isTerminator()) {
183 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
184 // block, we have to make sure it isn't in the LoopHeaders set. We
185 // reinsert afterward if needed.
186 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
187 BasicBlock *Succ = BI->getSuccessor(0);
189 // FIXME: It is always conservatively correct to drop the info
190 // for a block even if it doesn't get erased. This isn't totally
191 // awesome, but it allows us to use AssertingVH to prevent nasty
192 // dangling pointer issues within LazyValueInfo.
193 if (LVI) LVI->eraseBlock(BB);
194 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
196 // If we deleted BB and BB was the header of a loop, then the
197 // successor is now the header of the loop.
201 if (ErasedFromLoopHeaders)
202 LoopHeaders.insert(BB);
207 EverChanged |= Changed;
214 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
215 /// thread across it.
216 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
217 /// Ignore PHI nodes, these will be flattened when duplication happens.
218 BasicBlock::const_iterator I = BB->getFirstNonPHI();
220 // FIXME: THREADING will delete values that are just used to compute the
221 // branch, so they shouldn't count against the duplication cost.
224 // Sum up the cost of each instruction until we get to the terminator. Don't
225 // include the terminator because the copy won't include it.
227 for (; !isa<TerminatorInst>(I); ++I) {
228 // Debugger intrinsics don't incur code size.
229 if (isa<DbgInfoIntrinsic>(I)) continue;
231 // If this is a pointer->pointer bitcast, it is free.
232 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
235 // All other instructions count for at least one unit.
238 // Calls are more expensive. If they are non-intrinsic calls, we model them
239 // as having cost of 4. If they are a non-vector intrinsic, we model them
240 // as having cost of 2 total, and if they are a vector intrinsic, we model
241 // them as having cost 1.
242 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
243 if (!isa<IntrinsicInst>(CI))
245 else if (!CI->getType()->isVectorTy())
250 // Threading through a switch statement is particularly profitable. If this
251 // block ends in a switch, decrease its cost to make it more likely to happen.
252 if (isa<SwitchInst>(I))
253 Size = Size > 6 ? Size-6 : 0;
258 /// FindLoopHeaders - We do not want jump threading to turn proper loop
259 /// structures into irreducible loops. Doing this breaks up the loop nesting
260 /// hierarchy and pessimizes later transformations. To prevent this from
261 /// happening, we first have to find the loop headers. Here we approximate this
262 /// by finding targets of backedges in the CFG.
264 /// Note that there definitely are cases when we want to allow threading of
265 /// edges across a loop header. For example, threading a jump from outside the
266 /// loop (the preheader) to an exit block of the loop is definitely profitable.
267 /// It is also almost always profitable to thread backedges from within the loop
268 /// to exit blocks, and is often profitable to thread backedges to other blocks
269 /// within the loop (forming a nested loop). This simple analysis is not rich
270 /// enough to track all of these properties and keep it up-to-date as the CFG
271 /// mutates, so we don't allow any of these transformations.
273 void JumpThreading::FindLoopHeaders(Function &F) {
274 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
275 FindFunctionBackedges(F, Edges);
277 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
278 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
281 // Helper method for ComputeValueKnownInPredecessors. If Value is a
282 // ConstantInt, push it. If it's an undef, push 0. Otherwise, do nothing.
283 static void PushConstantIntOrUndef(SmallVectorImpl<std::pair<ConstantInt*,
284 BasicBlock*> > &Result,
285 Constant *Value, BasicBlock* BB){
286 if (ConstantInt *FoldedCInt = dyn_cast<ConstantInt>(Value))
287 Result.push_back(std::make_pair(FoldedCInt, BB));
288 else if (isa<UndefValue>(Value))
289 Result.push_back(std::make_pair((ConstantInt*)0, BB));
292 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
293 /// if we can infer that the value is a known ConstantInt in any of our
294 /// predecessors. If so, return the known list of value and pred BB in the
295 /// result vector. If a value is known to be undef, it is returned as null.
297 /// This returns true if there were any known values.
300 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){
301 // This method walks up use-def chains recursively. Because of this, we could
302 // get into an infinite loop going around loops in the use-def chain. To
303 // prevent this, keep track of what (value, block) pairs we've already visited
304 // and terminate the search if we loop back to them
305 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
308 // An RAII help to remove this pair from the recursion set once the recursion
309 // stack pops back out again.
310 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
312 // If V is a constantint, then it is known in all predecessors.
313 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) {
314 ConstantInt *CI = dyn_cast<ConstantInt>(V);
316 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
317 Result.push_back(std::make_pair(CI, *PI));
322 // If V is a non-instruction value, or an instruction in a different block,
323 // then it can't be derived from a PHI.
324 Instruction *I = dyn_cast<Instruction>(V);
325 if (I == 0 || I->getParent() != BB) {
327 // Okay, if this is a live-in value, see if it has a known value at the end
328 // of any of our predecessors.
330 // FIXME: This should be an edge property, not a block end property.
331 /// TODO: Per PR2563, we could infer value range information about a
332 /// predecessor based on its terminator.
335 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
336 // "I" is a non-local compare-with-a-constant instruction. This would be
337 // able to handle value inequalities better, for example if the compare is
338 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
339 // Perhaps getConstantOnEdge should be smart enough to do this?
341 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
343 // If the value is known by LazyValueInfo to be a constant in a
344 // predecessor, use that information to try to thread this block.
345 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
347 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst)))
350 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), P));
353 return !Result.empty();
359 /// If I is a PHI node, then we know the incoming values for any constants.
360 if (PHINode *PN = dyn_cast<PHINode>(I)) {
361 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
362 Value *InVal = PN->getIncomingValue(i);
363 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) {
364 ConstantInt *CI = dyn_cast<ConstantInt>(InVal);
365 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i)));
367 Constant *CI = LVI->getConstantOnEdge(InVal,
368 PN->getIncomingBlock(i), BB);
369 // LVI returns null is no value could be determined.
371 PushConstantIntOrUndef(Result, CI, PN->getIncomingBlock(i));
375 return !Result.empty();
378 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals;
380 // Handle some boolean conditions.
381 if (I->getType()->getPrimitiveSizeInBits() == 1) {
383 // X & false -> false
384 if (I->getOpcode() == Instruction::Or ||
385 I->getOpcode() == Instruction::And) {
386 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
387 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals);
389 if (LHSVals.empty() && RHSVals.empty())
392 ConstantInt *InterestingVal;
393 if (I->getOpcode() == Instruction::Or)
394 InterestingVal = ConstantInt::getTrue(I->getContext());
396 InterestingVal = ConstantInt::getFalse(I->getContext());
398 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
400 // Scan for the sentinel. If we find an undef, force it to the
401 // interesting value: x|undef -> true and x&undef -> false.
402 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
403 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) {
404 Result.push_back(LHSVals[i]);
405 Result.back().first = InterestingVal;
406 LHSKnownBBs.insert(LHSVals[i].second);
408 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
409 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) {
410 // If we already inferred a value for this block on the LHS, don't
412 if (!LHSKnownBBs.count(RHSVals[i].second)) {
413 Result.push_back(RHSVals[i]);
414 Result.back().first = InterestingVal;
418 return !Result.empty();
421 // Handle the NOT form of XOR.
422 if (I->getOpcode() == Instruction::Xor &&
423 isa<ConstantInt>(I->getOperand(1)) &&
424 cast<ConstantInt>(I->getOperand(1))->isOne()) {
425 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result);
429 // Invert the known values.
430 for (unsigned i = 0, e = Result.size(); i != e; ++i)
433 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first));
438 // Try to simplify some other binary operator values.
439 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
440 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
441 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals;
442 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals);
444 // Try to use constant folding to simplify the binary operator.
445 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
446 Constant *V = LHSVals[i].first;
447 if (V == 0) V = UndefValue::get(BO->getType());
448 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
450 PushConstantIntOrUndef(Result, Folded, LHSVals[i].second);
454 return !Result.empty();
457 // Handle compare with phi operand, where the PHI is defined in this block.
458 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
459 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
460 if (PN && PN->getParent() == BB) {
461 // We can do this simplification if any comparisons fold to true or false.
463 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
464 BasicBlock *PredBB = PN->getIncomingBlock(i);
465 Value *LHS = PN->getIncomingValue(i);
466 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
468 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
470 if (!LVI || !isa<Constant>(RHS))
473 LazyValueInfo::Tristate
474 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
475 cast<Constant>(RHS), PredBB, BB);
476 if (ResT == LazyValueInfo::Unknown)
478 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
481 if (Constant *ConstRes = dyn_cast<Constant>(Res))
482 PushConstantIntOrUndef(Result, ConstRes, PredBB);
485 return !Result.empty();
489 // If comparing a live-in value against a constant, see if we know the
490 // live-in value on any predecessors.
491 if (LVI && isa<Constant>(Cmp->getOperand(1)) &&
492 Cmp->getType()->isIntegerTy()) {
493 if (!isa<Instruction>(Cmp->getOperand(0)) ||
494 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
495 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
497 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
499 // If the value is known by LazyValueInfo to be a constant in a
500 // predecessor, use that information to try to thread this block.
501 LazyValueInfo::Tristate Res =
502 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
504 if (Res == LazyValueInfo::Unknown)
507 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
508 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), P));
511 return !Result.empty();
514 // Try to find a constant value for the LHS of a comparison,
515 // and evaluate it statically if we can.
516 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
517 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals;
518 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
520 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
521 Constant *V = LHSVals[i].first;
522 if (V == 0) V = UndefValue::get(CmpConst->getType());
523 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
525 PushConstantIntOrUndef(Result, Folded, LHSVals[i].second);
528 return !Result.empty();
534 // If all else fails, see if LVI can figure out a constant value for us.
535 Constant *CI = LVI->getConstant(V, BB);
536 ConstantInt *CInt = dyn_cast_or_null<ConstantInt>(CI);
538 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
539 Result.push_back(std::make_pair(CInt, *PI));
542 return !Result.empty();
550 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
551 /// in an undefined jump, decide which block is best to revector to.
553 /// Since we can pick an arbitrary destination, we pick the successor with the
554 /// fewest predecessors. This should reduce the in-degree of the others.
556 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
557 TerminatorInst *BBTerm = BB->getTerminator();
558 unsigned MinSucc = 0;
559 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
560 // Compute the successor with the minimum number of predecessors.
561 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
562 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
563 TestBB = BBTerm->getSuccessor(i);
564 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
565 if (NumPreds < MinNumPreds)
572 /// ProcessBlock - If there are any predecessors whose control can be threaded
573 /// through to a successor, transform them now.
574 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
575 // If the block is trivially dead, just return and let the caller nuke it.
576 // This simplifies other transformations.
577 if (pred_begin(BB) == pred_end(BB) &&
578 BB != &BB->getParent()->getEntryBlock())
581 // If this block has a single predecessor, and if that pred has a single
582 // successor, merge the blocks. This encourages recursive jump threading
583 // because now the condition in this block can be threaded through
584 // predecessors of our predecessor block.
585 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
586 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
588 // If SinglePred was a loop header, BB becomes one.
589 if (LoopHeaders.erase(SinglePred))
590 LoopHeaders.insert(BB);
592 // Remember if SinglePred was the entry block of the function. If so, we
593 // will need to move BB back to the entry position.
594 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
595 if (LVI) LVI->eraseBlock(SinglePred);
596 MergeBasicBlockIntoOnlyPred(BB);
598 if (isEntry && BB != &BB->getParent()->getEntryBlock())
599 BB->moveBefore(&BB->getParent()->getEntryBlock());
604 // Look to see if the terminator is a branch of switch, if not we can't thread
607 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
608 // Can't thread an unconditional jump.
609 if (BI->isUnconditional()) return false;
610 Condition = BI->getCondition();
611 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
612 Condition = SI->getCondition();
614 return false; // Must be an invoke.
616 // If the terminator of this block is branching on a constant, simplify the
617 // terminator to an unconditional branch. This can occur due to threading in
619 if (isa<ConstantInt>(Condition)) {
620 DEBUG(dbgs() << " In block '" << BB->getName()
621 << "' folding terminator: " << *BB->getTerminator() << '\n');
623 ConstantFoldTerminator(BB);
627 // If the terminator is branching on an undef, we can pick any of the
628 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
629 if (isa<UndefValue>(Condition)) {
630 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
632 // Fold the branch/switch.
633 TerminatorInst *BBTerm = BB->getTerminator();
634 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
635 if (i == BestSucc) continue;
636 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD);
639 DEBUG(dbgs() << " In block '" << BB->getName()
640 << "' folding undef terminator: " << *BBTerm << '\n');
641 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
642 BBTerm->eraseFromParent();
646 Instruction *CondInst = dyn_cast<Instruction>(Condition);
648 // If the condition is an instruction defined in another block, see if a
649 // predecessor has the same condition:
654 !Condition->hasOneUse() && // Multiple uses.
655 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition.
656 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
657 if (isa<BranchInst>(BB->getTerminator())) {
658 for (; PI != E; ++PI) {
660 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator()))
661 if (PBI->isConditional() && PBI->getCondition() == Condition &&
662 ProcessBranchOnDuplicateCond(P, BB))
666 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator");
667 for (; PI != E; ++PI) {
669 if (SwitchInst *PSI = dyn_cast<SwitchInst>(P->getTerminator()))
670 if (PSI->getCondition() == Condition &&
671 ProcessSwitchOnDuplicateCond(P, BB))
677 // All the rest of our checks depend on the condition being an instruction.
679 // FIXME: Unify this with code below.
680 if (LVI && ProcessThreadableEdges(Condition, BB))
686 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
688 (!isa<PHINode>(CondCmp->getOperand(0)) ||
689 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) {
690 // If we have a comparison, loop over the predecessors to see if there is
691 // a condition with a lexically identical value.
692 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
693 for (; PI != E; ++PI) {
695 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator()))
696 if (PBI->isConditional() && P != BB) {
697 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) {
698 if (CI->getOperand(0) == CondCmp->getOperand(0) &&
699 CI->getOperand(1) == CondCmp->getOperand(1) &&
700 CI->getPredicate() == CondCmp->getPredicate()) {
701 // TODO: Could handle things like (x != 4) --> (x == 17)
702 if (ProcessBranchOnDuplicateCond(P, BB))
710 // For a comparison where the LHS is outside this block, it's possible
711 // that we've branched on it before. Used LVI to see if we can simplify
712 // the branch based on that.
713 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
714 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
715 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
716 if (LVI && CondBr && CondConst && CondBr->isConditional() && PI != PE &&
717 (!isa<Instruction>(CondCmp->getOperand(0)) ||
718 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
719 // For predecessor edge, determine if the comparison is true or false
720 // on that edge. If they're all true or all false, we can simplify the
722 // FIXME: We could handle mixed true/false by duplicating code.
723 LazyValueInfo::Tristate Baseline =
724 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
726 if (Baseline != LazyValueInfo::Unknown) {
727 // Check that all remaining incoming values match the first one.
729 LazyValueInfo::Tristate Ret =
730 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
731 CondCmp->getOperand(0), CondConst, *PI, BB);
732 if (Ret != Baseline) break;
735 // If we terminated early, then one of the values didn't match.
737 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
738 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
739 RemovePredecessorAndSimplify(CondBr->getSuccessor(ToRemove), BB, TD);
740 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
741 CondBr->eraseFromParent();
748 // Check for some cases that are worth simplifying. Right now we want to look
749 // for loads that are used by a switch or by the condition for the branch. If
750 // we see one, check to see if it's partially redundant. If so, insert a PHI
751 // which can then be used to thread the values.
753 Value *SimplifyValue = CondInst;
754 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
755 if (isa<Constant>(CondCmp->getOperand(1)))
756 SimplifyValue = CondCmp->getOperand(0);
758 // TODO: There are other places where load PRE would be profitable, such as
759 // more complex comparisons.
760 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
761 if (SimplifyPartiallyRedundantLoad(LI))
765 // Handle a variety of cases where we are branching on something derived from
766 // a PHI node in the current block. If we can prove that any predecessors
767 // compute a predictable value based on a PHI node, thread those predecessors.
769 if (ProcessThreadableEdges(CondInst, BB))
772 // If this is an otherwise-unfoldable branch on a phi node in the current
773 // block, see if we can simplify.
774 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
775 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
776 return ProcessBranchOnPHI(PN);
779 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
780 if (CondInst->getOpcode() == Instruction::Xor &&
781 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
782 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
785 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
786 // "(X == 4)", thread through this block.
791 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
792 /// block that jump on exactly the same condition. This means that we almost
793 /// always know the direction of the edge in the DESTBB:
795 /// br COND, DESTBB, BBY
797 /// br COND, BBZ, BBW
799 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
800 /// in DESTBB, we have to thread over it.
801 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
803 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
805 // If both successors of PredBB go to DESTBB, we don't know anything. We can
806 // fold the branch to an unconditional one, which allows other recursive
809 if (PredBI->getSuccessor(1) != BB)
811 else if (PredBI->getSuccessor(0) != BB)
814 DEBUG(dbgs() << " In block '" << PredBB->getName()
815 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
817 ConstantFoldTerminator(PredBB);
821 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
823 // If the dest block has one predecessor, just fix the branch condition to a
824 // constant and fold it.
825 if (BB->getSinglePredecessor()) {
826 DEBUG(dbgs() << " In block '" << BB->getName()
827 << "' folding condition to '" << BranchDir << "': "
828 << *BB->getTerminator() << '\n');
830 Value *OldCond = DestBI->getCondition();
831 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
833 // Delete dead instructions before we fold the branch. Folding the branch
834 // can eliminate edges from the CFG which can end up deleting OldCond.
835 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
836 ConstantFoldTerminator(BB);
841 // Next, figure out which successor we are threading to.
842 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
844 SmallVector<BasicBlock*, 2> Preds;
845 Preds.push_back(PredBB);
847 // Ok, try to thread it!
848 return ThreadEdge(BB, Preds, SuccBB);
851 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
852 /// block that switch on exactly the same condition. This means that we almost
853 /// always know the direction of the edge in the DESTBB:
855 /// switch COND [... DESTBB, BBY ... ]
857 /// switch COND [... BBZ, BBW ]
859 /// Optimizing switches like this is very important, because simplifycfg builds
860 /// switches out of repeated 'if' conditions.
861 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
862 BasicBlock *DestBB) {
863 // Can't thread edge to self.
864 if (PredBB == DestBB)
867 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
868 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
870 // There are a variety of optimizations that we can potentially do on these
871 // blocks: we order them from most to least preferable.
873 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
874 // directly to their destination. This does not introduce *any* code size
875 // growth. Skip debug info first.
876 BasicBlock::iterator BBI = DestBB->begin();
877 while (isa<DbgInfoIntrinsic>(BBI))
880 // FIXME: Thread if it just contains a PHI.
881 if (isa<SwitchInst>(BBI)) {
882 bool MadeChange = false;
883 // Ignore the default edge for now.
884 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
885 ConstantInt *DestVal = DestSI->getCaseValue(i);
886 BasicBlock *DestSucc = DestSI->getSuccessor(i);
888 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
889 // PredSI has an explicit case for it. If so, forward. If it is covered
890 // by the default case, we can't update PredSI.
891 unsigned PredCase = PredSI->findCaseValue(DestVal);
892 if (PredCase == 0) continue;
894 // If PredSI doesn't go to DestBB on this value, then it won't reach the
895 // case on this condition.
896 if (PredSI->getSuccessor(PredCase) != DestBB &&
897 DestSI->getSuccessor(i) != DestBB)
900 // Do not forward this if it already goes to this destination, this would
901 // be an infinite loop.
902 if (PredSI->getSuccessor(PredCase) == DestSucc)
905 // Otherwise, we're safe to make the change. Make sure that the edge from
906 // DestSI to DestSucc is not critical and has no PHI nodes.
907 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
908 DEBUG(dbgs() << "THROUGH: " << *DestSI);
910 // If the destination has PHI nodes, just split the edge for updating
912 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
913 SplitCriticalEdge(DestSI, i, this);
914 DestSucc = DestSI->getSuccessor(i);
916 FoldSingleEntryPHINodes(DestSucc);
917 PredSI->setSuccessor(PredCase, DestSucc);
929 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
930 /// load instruction, eliminate it by replacing it with a PHI node. This is an
931 /// important optimization that encourages jump threading, and needs to be run
932 /// interlaced with other jump threading tasks.
933 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
934 // Don't hack volatile loads.
935 if (LI->isVolatile()) return false;
937 // If the load is defined in a block with exactly one predecessor, it can't be
938 // partially redundant.
939 BasicBlock *LoadBB = LI->getParent();
940 if (LoadBB->getSinglePredecessor())
943 Value *LoadedPtr = LI->getOperand(0);
945 // If the loaded operand is defined in the LoadBB, it can't be available.
946 // TODO: Could do simple PHI translation, that would be fun :)
947 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
948 if (PtrOp->getParent() == LoadBB)
951 // Scan a few instructions up from the load, to see if it is obviously live at
952 // the entry to its block.
953 BasicBlock::iterator BBIt = LI;
955 if (Value *AvailableVal =
956 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
957 // If the value if the load is locally available within the block, just use
958 // it. This frequently occurs for reg2mem'd allocas.
959 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
961 // If the returned value is the load itself, replace with an undef. This can
962 // only happen in dead loops.
963 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
964 LI->replaceAllUsesWith(AvailableVal);
965 LI->eraseFromParent();
969 // Otherwise, if we scanned the whole block and got to the top of the block,
970 // we know the block is locally transparent to the load. If not, something
971 // might clobber its value.
972 if (BBIt != LoadBB->begin())
976 SmallPtrSet<BasicBlock*, 8> PredsScanned;
977 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
978 AvailablePredsTy AvailablePreds;
979 BasicBlock *OneUnavailablePred = 0;
981 // If we got here, the loaded value is transparent through to the start of the
982 // block. Check to see if it is available in any of the predecessor blocks.
983 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
985 BasicBlock *PredBB = *PI;
987 // If we already scanned this predecessor, skip it.
988 if (!PredsScanned.insert(PredBB))
991 // Scan the predecessor to see if the value is available in the pred.
992 BBIt = PredBB->end();
993 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
994 if (!PredAvailable) {
995 OneUnavailablePred = PredBB;
999 // If so, this load is partially redundant. Remember this info so that we
1000 // can create a PHI node.
1001 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1004 // If the loaded value isn't available in any predecessor, it isn't partially
1006 if (AvailablePreds.empty()) return false;
1008 // Okay, the loaded value is available in at least one (and maybe all!)
1009 // predecessors. If the value is unavailable in more than one unique
1010 // predecessor, we want to insert a merge block for those common predecessors.
1011 // This ensures that we only have to insert one reload, thus not increasing
1013 BasicBlock *UnavailablePred = 0;
1015 // If there is exactly one predecessor where the value is unavailable, the
1016 // already computed 'OneUnavailablePred' block is it. If it ends in an
1017 // unconditional branch, we know that it isn't a critical edge.
1018 if (PredsScanned.size() == AvailablePreds.size()+1 &&
1019 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1020 UnavailablePred = OneUnavailablePred;
1021 } else if (PredsScanned.size() != AvailablePreds.size()) {
1022 // Otherwise, we had multiple unavailable predecessors or we had a critical
1023 // edge from the one.
1024 SmallVector<BasicBlock*, 8> PredsToSplit;
1025 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1027 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
1028 AvailablePredSet.insert(AvailablePreds[i].first);
1030 // Add all the unavailable predecessors to the PredsToSplit list.
1031 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
1033 BasicBlock *P = *PI;
1034 // If the predecessor is an indirect goto, we can't split the edge.
1035 if (isa<IndirectBrInst>(P->getTerminator()))
1038 if (!AvailablePredSet.count(P))
1039 PredsToSplit.push_back(P);
1042 // Split them out to their own block.
1044 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
1045 "thread-pre-split", this);
1048 // If the value isn't available in all predecessors, then there will be
1049 // exactly one where it isn't available. Insert a load on that edge and add
1050 // it to the AvailablePreds list.
1051 if (UnavailablePred) {
1052 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1053 "Can't handle critical edge here!");
1054 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
1056 UnavailablePred->getTerminator());
1057 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1060 // Now we know that each predecessor of this block has a value in
1061 // AvailablePreds, sort them for efficient access as we're walking the preds.
1062 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1064 // Create a PHI node at the start of the block for the PRE'd load value.
1065 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
1068 // Insert new entries into the PHI for each predecessor. A single block may
1069 // have multiple entries here.
1070 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
1072 BasicBlock *P = *PI;
1073 AvailablePredsTy::iterator I =
1074 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1075 std::make_pair(P, (Value*)0));
1077 assert(I != AvailablePreds.end() && I->first == P &&
1078 "Didn't find entry for predecessor!");
1080 PN->addIncoming(I->second, I->first);
1083 //cerr << "PRE: " << *LI << *PN << "\n";
1085 LI->replaceAllUsesWith(PN);
1086 LI->eraseFromParent();
1091 /// FindMostPopularDest - The specified list contains multiple possible
1092 /// threadable destinations. Pick the one that occurs the most frequently in
1095 FindMostPopularDest(BasicBlock *BB,
1096 const SmallVectorImpl<std::pair<BasicBlock*,
1097 BasicBlock*> > &PredToDestList) {
1098 assert(!PredToDestList.empty());
1100 // Determine popularity. If there are multiple possible destinations, we
1101 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1102 // blocks with known and real destinations to threading undef. We'll handle
1103 // them later if interesting.
1104 DenseMap<BasicBlock*, unsigned> DestPopularity;
1105 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1106 if (PredToDestList[i].second)
1107 DestPopularity[PredToDestList[i].second]++;
1109 // Find the most popular dest.
1110 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1111 BasicBlock *MostPopularDest = DPI->first;
1112 unsigned Popularity = DPI->second;
1113 SmallVector<BasicBlock*, 4> SamePopularity;
1115 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1116 // If the popularity of this entry isn't higher than the popularity we've
1117 // seen so far, ignore it.
1118 if (DPI->second < Popularity)
1120 else if (DPI->second == Popularity) {
1121 // If it is the same as what we've seen so far, keep track of it.
1122 SamePopularity.push_back(DPI->first);
1124 // If it is more popular, remember it.
1125 SamePopularity.clear();
1126 MostPopularDest = DPI->first;
1127 Popularity = DPI->second;
1131 // Okay, now we know the most popular destination. If there is more than
1132 // destination, we need to determine one. This is arbitrary, but we need
1133 // to make a deterministic decision. Pick the first one that appears in the
1135 if (!SamePopularity.empty()) {
1136 SamePopularity.push_back(MostPopularDest);
1137 TerminatorInst *TI = BB->getTerminator();
1138 for (unsigned i = 0; ; ++i) {
1139 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1141 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1142 TI->getSuccessor(i)) == SamePopularity.end())
1145 MostPopularDest = TI->getSuccessor(i);
1150 // Okay, we have finally picked the most popular destination.
1151 return MostPopularDest;
1154 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) {
1155 // If threading this would thread across a loop header, don't even try to
1157 if (LoopHeaders.count(BB))
1160 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues;
1161 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues))
1164 assert(!PredValues.empty() &&
1165 "ComputeValueKnownInPredecessors returned true with no values");
1167 DEBUG(dbgs() << "IN BB: " << *BB;
1168 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1169 dbgs() << " BB '" << BB->getName() << "': FOUND condition = ";
1170 if (PredValues[i].first)
1171 dbgs() << *PredValues[i].first;
1174 dbgs() << " for pred '" << PredValues[i].second->getName()
1178 // Decide what we want to thread through. Convert our list of known values to
1179 // a list of known destinations for each pred. This also discards duplicate
1180 // predecessors and keeps track of the undefined inputs (which are represented
1181 // as a null dest in the PredToDestList).
1182 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1183 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1185 BasicBlock *OnlyDest = 0;
1186 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1188 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1189 BasicBlock *Pred = PredValues[i].second;
1190 if (!SeenPreds.insert(Pred))
1191 continue; // Duplicate predecessor entry.
1193 // If the predecessor ends with an indirect goto, we can't change its
1195 if (isa<IndirectBrInst>(Pred->getTerminator()))
1198 ConstantInt *Val = PredValues[i].first;
1201 if (Val == 0) // Undef.
1203 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1204 DestBB = BI->getSuccessor(Val->isZero());
1206 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
1207 DestBB = SI->getSuccessor(SI->findCaseValue(Val));
1210 // If we have exactly one destination, remember it for efficiency below.
1213 else if (OnlyDest != DestBB)
1214 OnlyDest = MultipleDestSentinel;
1216 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1219 // If all edges were unthreadable, we fail.
1220 if (PredToDestList.empty())
1223 // Determine which is the most common successor. If we have many inputs and
1224 // this block is a switch, we want to start by threading the batch that goes
1225 // to the most popular destination first. If we only know about one
1226 // threadable destination (the common case) we can avoid this.
1227 BasicBlock *MostPopularDest = OnlyDest;
1229 if (MostPopularDest == MultipleDestSentinel)
1230 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1232 // Now that we know what the most popular destination is, factor all
1233 // predecessors that will jump to it into a single predecessor.
1234 SmallVector<BasicBlock*, 16> PredsToFactor;
1235 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1236 if (PredToDestList[i].second == MostPopularDest) {
1237 BasicBlock *Pred = PredToDestList[i].first;
1239 // This predecessor may be a switch or something else that has multiple
1240 // edges to the block. Factor each of these edges by listing them
1241 // according to # occurrences in PredsToFactor.
1242 TerminatorInst *PredTI = Pred->getTerminator();
1243 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1244 if (PredTI->getSuccessor(i) == BB)
1245 PredsToFactor.push_back(Pred);
1248 // If the threadable edges are branching on an undefined value, we get to pick
1249 // the destination that these predecessors should get to.
1250 if (MostPopularDest == 0)
1251 MostPopularDest = BB->getTerminator()->
1252 getSuccessor(GetBestDestForJumpOnUndef(BB));
1254 // Ok, try to thread it!
1255 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1258 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1259 /// a PHI node in the current block. See if there are any simplifications we
1260 /// can do based on inputs to the phi node.
1262 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1263 BasicBlock *BB = PN->getParent();
1265 // TODO: We could make use of this to do it once for blocks with common PHI
1267 SmallVector<BasicBlock*, 1> PredBBs;
1270 // If any of the predecessor blocks end in an unconditional branch, we can
1271 // *duplicate* the conditional branch into that block in order to further
1272 // encourage jump threading and to eliminate cases where we have branch on a
1273 // phi of an icmp (branch on icmp is much better).
1274 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1275 BasicBlock *PredBB = PN->getIncomingBlock(i);
1276 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1277 if (PredBr->isUnconditional()) {
1278 PredBBs[0] = PredBB;
1279 // Try to duplicate BB into PredBB.
1280 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1288 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1289 /// a xor instruction in the current block. See if there are any
1290 /// simplifications we can do based on inputs to the xor.
1292 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1293 BasicBlock *BB = BO->getParent();
1295 // If either the LHS or RHS of the xor is a constant, don't do this
1297 if (isa<ConstantInt>(BO->getOperand(0)) ||
1298 isa<ConstantInt>(BO->getOperand(1)))
1301 // If the first instruction in BB isn't a phi, we won't be able to infer
1302 // anything special about any particular predecessor.
1303 if (!isa<PHINode>(BB->front()))
1306 // If we have a xor as the branch input to this block, and we know that the
1307 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1308 // the condition into the predecessor and fix that value to true, saving some
1309 // logical ops on that path and encouraging other paths to simplify.
1311 // This copies something like this:
1314 // %X = phi i1 [1], [%X']
1315 // %Y = icmp eq i32 %A, %B
1316 // %Z = xor i1 %X, %Y
1321 // %Y = icmp ne i32 %A, %B
1324 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues;
1326 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) {
1327 assert(XorOpValues.empty());
1328 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues))
1333 assert(!XorOpValues.empty() &&
1334 "ComputeValueKnownInPredecessors returned true with no values");
1336 // Scan the information to see which is most popular: true or false. The
1337 // predecessors can be of the set true, false, or undef.
1338 unsigned NumTrue = 0, NumFalse = 0;
1339 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1340 if (!XorOpValues[i].first) continue; // Ignore undefs for the count.
1341 if (XorOpValues[i].first->isZero())
1347 // Determine which value to split on, true, false, or undef if neither.
1348 ConstantInt *SplitVal = 0;
1349 if (NumTrue > NumFalse)
1350 SplitVal = ConstantInt::getTrue(BB->getContext());
1351 else if (NumTrue != 0 || NumFalse != 0)
1352 SplitVal = ConstantInt::getFalse(BB->getContext());
1354 // Collect all of the blocks that this can be folded into so that we can
1355 // factor this once and clone it once.
1356 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1357 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1358 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue;
1360 BlocksToFoldInto.push_back(XorOpValues[i].second);
1363 // If we inferred a value for all of the predecessors, then duplication won't
1364 // help us. However, we can just replace the LHS or RHS with the constant.
1365 if (BlocksToFoldInto.size() ==
1366 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1367 if (SplitVal == 0) {
1368 // If all preds provide undef, just nuke the xor, because it is undef too.
1369 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1370 BO->eraseFromParent();
1371 } else if (SplitVal->isZero()) {
1372 // If all preds provide 0, replace the xor with the other input.
1373 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1374 BO->eraseFromParent();
1376 // If all preds provide 1, set the computed value to 1.
1377 BO->setOperand(!isLHS, SplitVal);
1383 // Try to duplicate BB into PredBB.
1384 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1388 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1389 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1390 /// NewPred using the entries from OldPred (suitably mapped).
1391 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1392 BasicBlock *OldPred,
1393 BasicBlock *NewPred,
1394 DenseMap<Instruction*, Value*> &ValueMap) {
1395 for (BasicBlock::iterator PNI = PHIBB->begin();
1396 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1397 // Ok, we have a PHI node. Figure out what the incoming value was for the
1399 Value *IV = PN->getIncomingValueForBlock(OldPred);
1401 // Remap the value if necessary.
1402 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1403 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1404 if (I != ValueMap.end())
1408 PN->addIncoming(IV, NewPred);
1412 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1413 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1414 /// across BB. Transform the IR to reflect this change.
1415 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1416 const SmallVectorImpl<BasicBlock*> &PredBBs,
1417 BasicBlock *SuccBB) {
1418 // If threading to the same block as we come from, we would infinite loop.
1420 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1421 << "' - would thread to self!\n");
1425 // If threading this would thread across a loop header, don't thread the edge.
1426 // See the comments above FindLoopHeaders for justifications and caveats.
1427 if (LoopHeaders.count(BB)) {
1428 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1429 << "' to dest BB '" << SuccBB->getName()
1430 << "' - it might create an irreducible loop!\n");
1434 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1435 if (JumpThreadCost > Threshold) {
1436 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1437 << "' - Cost is too high: " << JumpThreadCost << "\n");
1441 // And finally, do it! Start by factoring the predecessors is needed.
1443 if (PredBBs.size() == 1)
1444 PredBB = PredBBs[0];
1446 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1447 << " common predecessors.\n");
1448 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1452 // And finally, do it!
1453 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1454 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1455 << ", across block:\n "
1459 LVI->threadEdge(PredBB, BB, SuccBB);
1461 // We are going to have to map operands from the original BB block to the new
1462 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1463 // account for entry from PredBB.
1464 DenseMap<Instruction*, Value*> ValueMapping;
1466 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1467 BB->getName()+".thread",
1468 BB->getParent(), BB);
1469 NewBB->moveAfter(PredBB);
1471 BasicBlock::iterator BI = BB->begin();
1472 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1473 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1475 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1476 // mapping and using it to remap operands in the cloned instructions.
1477 for (; !isa<TerminatorInst>(BI); ++BI) {
1478 Instruction *New = BI->clone();
1479 New->setName(BI->getName());
1480 NewBB->getInstList().push_back(New);
1481 ValueMapping[BI] = New;
1483 // Remap operands to patch up intra-block references.
1484 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1485 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1486 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1487 if (I != ValueMapping.end())
1488 New->setOperand(i, I->second);
1492 // We didn't copy the terminator from BB over to NewBB, because there is now
1493 // an unconditional jump to SuccBB. Insert the unconditional jump.
1494 BranchInst::Create(SuccBB, NewBB);
1496 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1497 // PHI nodes for NewBB now.
1498 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1500 // If there were values defined in BB that are used outside the block, then we
1501 // now have to update all uses of the value to use either the original value,
1502 // the cloned value, or some PHI derived value. This can require arbitrary
1503 // PHI insertion, of which we are prepared to do, clean these up now.
1504 SSAUpdater SSAUpdate;
1505 SmallVector<Use*, 16> UsesToRename;
1506 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1507 // Scan all uses of this instruction to see if it is used outside of its
1508 // block, and if so, record them in UsesToRename.
1509 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1511 Instruction *User = cast<Instruction>(*UI);
1512 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1513 if (UserPN->getIncomingBlock(UI) == BB)
1515 } else if (User->getParent() == BB)
1518 UsesToRename.push_back(&UI.getUse());
1521 // If there are no uses outside the block, we're done with this instruction.
1522 if (UsesToRename.empty())
1525 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1527 // We found a use of I outside of BB. Rename all uses of I that are outside
1528 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1529 // with the two values we know.
1530 SSAUpdate.Initialize(I->getType(), I->getName());
1531 SSAUpdate.AddAvailableValue(BB, I);
1532 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1534 while (!UsesToRename.empty())
1535 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1536 DEBUG(dbgs() << "\n");
1540 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1541 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1542 // us to simplify any PHI nodes in BB.
1543 TerminatorInst *PredTerm = PredBB->getTerminator();
1544 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1545 if (PredTerm->getSuccessor(i) == BB) {
1546 RemovePredecessorAndSimplify(BB, PredBB, TD);
1547 PredTerm->setSuccessor(i, NewBB);
1550 // At this point, the IR is fully up to date and consistent. Do a quick scan
1551 // over the new instructions and zap any that are constants or dead. This
1552 // frequently happens because of phi translation.
1553 SimplifyInstructionsInBlock(NewBB, TD);
1555 // Threaded an edge!
1560 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1561 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1562 /// If we can duplicate the contents of BB up into PredBB do so now, this
1563 /// improves the odds that the branch will be on an analyzable instruction like
1565 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1566 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1567 assert(!PredBBs.empty() && "Can't handle an empty set");
1569 // If BB is a loop header, then duplicating this block outside the loop would
1570 // cause us to transform this into an irreducible loop, don't do this.
1571 // See the comments above FindLoopHeaders for justifications and caveats.
1572 if (LoopHeaders.count(BB)) {
1573 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1574 << "' into predecessor block '" << PredBBs[0]->getName()
1575 << "' - it might create an irreducible loop!\n");
1579 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1580 if (DuplicationCost > Threshold) {
1581 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1582 << "' - Cost is too high: " << DuplicationCost << "\n");
1586 // And finally, do it! Start by factoring the predecessors is needed.
1588 if (PredBBs.size() == 1)
1589 PredBB = PredBBs[0];
1591 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1592 << " common predecessors.\n");
1593 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1597 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1599 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1600 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1601 << DuplicationCost << " block is:" << *BB << "\n");
1603 // Unless PredBB ends with an unconditional branch, split the edge so that we
1604 // can just clone the bits from BB into the end of the new PredBB.
1605 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1607 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1608 PredBB = SplitEdge(PredBB, BB, this);
1609 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1612 // We are going to have to map operands from the original BB block into the
1613 // PredBB block. Evaluate PHI nodes in BB.
1614 DenseMap<Instruction*, Value*> ValueMapping;
1616 BasicBlock::iterator BI = BB->begin();
1617 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1618 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1620 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1621 // mapping and using it to remap operands in the cloned instructions.
1622 for (; BI != BB->end(); ++BI) {
1623 Instruction *New = BI->clone();
1625 // Remap operands to patch up intra-block references.
1626 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1627 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1628 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1629 if (I != ValueMapping.end())
1630 New->setOperand(i, I->second);
1633 // If this instruction can be simplified after the operands are updated,
1634 // just use the simplified value instead. This frequently happens due to
1636 if (Value *IV = SimplifyInstruction(New, TD)) {
1638 ValueMapping[BI] = IV;
1640 // Otherwise, insert the new instruction into the block.
1641 New->setName(BI->getName());
1642 PredBB->getInstList().insert(OldPredBranch, New);
1643 ValueMapping[BI] = New;
1647 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1648 // add entries to the PHI nodes for branch from PredBB now.
1649 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1650 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1652 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1655 // If there were values defined in BB that are used outside the block, then we
1656 // now have to update all uses of the value to use either the original value,
1657 // the cloned value, or some PHI derived value. This can require arbitrary
1658 // PHI insertion, of which we are prepared to do, clean these up now.
1659 SSAUpdater SSAUpdate;
1660 SmallVector<Use*, 16> UsesToRename;
1661 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1662 // Scan all uses of this instruction to see if it is used outside of its
1663 // block, and if so, record them in UsesToRename.
1664 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1666 Instruction *User = cast<Instruction>(*UI);
1667 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1668 if (UserPN->getIncomingBlock(UI) == BB)
1670 } else if (User->getParent() == BB)
1673 UsesToRename.push_back(&UI.getUse());
1676 // If there are no uses outside the block, we're done with this instruction.
1677 if (UsesToRename.empty())
1680 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1682 // We found a use of I outside of BB. Rename all uses of I that are outside
1683 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1684 // with the two values we know.
1685 SSAUpdate.Initialize(I->getType(), I->getName());
1686 SSAUpdate.AddAvailableValue(BB, I);
1687 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1689 while (!UsesToRename.empty())
1690 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1691 DEBUG(dbgs() << "\n");
1694 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1696 RemovePredecessorAndSimplify(BB, PredBB, TD);
1698 // Remove the unconditional branch at the end of the PredBB block.
1699 OldPredBranch->eraseFromParent();