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>();
107 void FindLoopHeaders(Function &F);
108 bool ProcessBlock(BasicBlock *BB);
109 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
111 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
112 const SmallVectorImpl<BasicBlock *> &PredBBs);
114 typedef SmallVectorImpl<std::pair<ConstantInt*,
115 BasicBlock*> > PredValueInfo;
117 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
118 PredValueInfo &Result);
119 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB);
122 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
123 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
125 bool ProcessBranchOnPHI(PHINode *PN);
126 bool ProcessBranchOnXOR(BinaryOperator *BO);
128 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
132 char JumpThreading::ID = 0;
133 INITIALIZE_PASS(JumpThreading, "jump-threading",
134 "Jump Threading", false, false);
136 // Public interface to the Jump Threading pass
137 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
139 /// runOnFunction - Top level algorithm.
141 bool JumpThreading::runOnFunction(Function &F) {
142 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
143 TD = getAnalysisIfAvailable<TargetData>();
144 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0;
148 bool Changed, EverChanged = false;
151 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
153 // Thread all of the branches we can over this block.
154 while (ProcessBlock(BB))
159 // If the block is trivially dead, zap it. This eliminates the successor
160 // edges which simplifies the CFG.
161 if (pred_begin(BB) == pred_end(BB) &&
162 BB != &BB->getParent()->getEntryBlock()) {
163 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
164 << "' with terminator: " << *BB->getTerminator() << '\n');
165 LoopHeaders.erase(BB);
166 if (LVI) LVI->eraseBlock(BB);
169 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
170 // Can't thread an unconditional jump, but if the block is "almost
171 // empty", we can replace uses of it with uses of the successor and make
173 if (BI->isUnconditional() &&
174 BB != &BB->getParent()->getEntryBlock()) {
175 BasicBlock::iterator BBI = BB->getFirstNonPHI();
176 // Ignore dbg intrinsics.
177 while (isa<DbgInfoIntrinsic>(BBI))
179 // If the terminator is the only non-phi instruction, try to nuke it.
180 if (BBI->isTerminator()) {
181 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
182 // block, we have to make sure it isn't in the LoopHeaders set. We
183 // reinsert afterward if needed.
184 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
185 BasicBlock *Succ = BI->getSuccessor(0);
187 // FIXME: It is always conservatively correct to drop the info
188 // for a block even if it doesn't get erased. This isn't totally
189 // awesome, but it allows us to use AssertingVH to prevent nasty
190 // dangling pointer issues within LazyValueInfo.
191 if (LVI) LVI->eraseBlock(BB);
192 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
194 // If we deleted BB and BB was the header of a loop, then the
195 // successor is now the header of the loop.
199 if (ErasedFromLoopHeaders)
200 LoopHeaders.insert(BB);
205 EverChanged |= Changed;
212 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
213 /// thread across it.
214 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
215 /// Ignore PHI nodes, these will be flattened when duplication happens.
216 BasicBlock::const_iterator I = BB->getFirstNonPHI();
218 // FIXME: THREADING will delete values that are just used to compute the
219 // branch, so they shouldn't count against the duplication cost.
222 // Sum up the cost of each instruction until we get to the terminator. Don't
223 // include the terminator because the copy won't include it.
225 for (; !isa<TerminatorInst>(I); ++I) {
226 // Debugger intrinsics don't incur code size.
227 if (isa<DbgInfoIntrinsic>(I)) continue;
229 // If this is a pointer->pointer bitcast, it is free.
230 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
233 // All other instructions count for at least one unit.
236 // Calls are more expensive. If they are non-intrinsic calls, we model them
237 // as having cost of 4. If they are a non-vector intrinsic, we model them
238 // as having cost of 2 total, and if they are a vector intrinsic, we model
239 // them as having cost 1.
240 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
241 if (!isa<IntrinsicInst>(CI))
243 else if (!CI->getType()->isVectorTy())
248 // Threading through a switch statement is particularly profitable. If this
249 // block ends in a switch, decrease its cost to make it more likely to happen.
250 if (isa<SwitchInst>(I))
251 Size = Size > 6 ? Size-6 : 0;
256 /// FindLoopHeaders - We do not want jump threading to turn proper loop
257 /// structures into irreducible loops. Doing this breaks up the loop nesting
258 /// hierarchy and pessimizes later transformations. To prevent this from
259 /// happening, we first have to find the loop headers. Here we approximate this
260 /// by finding targets of backedges in the CFG.
262 /// Note that there definitely are cases when we want to allow threading of
263 /// edges across a loop header. For example, threading a jump from outside the
264 /// loop (the preheader) to an exit block of the loop is definitely profitable.
265 /// It is also almost always profitable to thread backedges from within the loop
266 /// to exit blocks, and is often profitable to thread backedges to other blocks
267 /// within the loop (forming a nested loop). This simple analysis is not rich
268 /// enough to track all of these properties and keep it up-to-date as the CFG
269 /// mutates, so we don't allow any of these transformations.
271 void JumpThreading::FindLoopHeaders(Function &F) {
272 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
273 FindFunctionBackedges(F, Edges);
275 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
276 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
279 // Helper method for ComputeValueKnownInPredecessors. If Value is a
280 // ConstantInt, push it. If it's an undef, push 0. Otherwise, do nothing.
281 static void PushConstantIntOrUndef(SmallVectorImpl<std::pair<ConstantInt*,
282 BasicBlock*> > &Result,
283 Constant *Value, BasicBlock* BB){
284 if (ConstantInt *FoldedCInt = dyn_cast<ConstantInt>(Value))
285 Result.push_back(std::make_pair(FoldedCInt, BB));
286 else if (isa<UndefValue>(Value))
287 Result.push_back(std::make_pair((ConstantInt*)0, BB));
290 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
291 /// if we can infer that the value is a known ConstantInt in any of our
292 /// predecessors. If so, return the known list of value and pred BB in the
293 /// result vector. If a value is known to be undef, it is returned as null.
295 /// This returns true if there were any known values.
298 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){
299 // This method walks up use-def chains recursively. Because of this, we could
300 // get into an infinite loop going around loops in the use-def chain. To
301 // prevent this, keep track of what (value, block) pairs we've already visited
302 // and terminate the search if we loop back to them
303 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
306 // An RAII help to remove this pair from the recursion set once the recursion
307 // stack pops back out again.
308 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
310 // If V is a constantint, then it is known in all predecessors.
311 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) {
312 ConstantInt *CI = dyn_cast<ConstantInt>(V);
314 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
315 Result.push_back(std::make_pair(CI, *PI));
320 // If V is a non-instruction value, or an instruction in a different block,
321 // then it can't be derived from a PHI.
322 Instruction *I = dyn_cast<Instruction>(V);
323 if (I == 0 || I->getParent() != BB) {
325 // Okay, if this is a live-in value, see if it has a known value at the end
326 // of any of our predecessors.
328 // FIXME: This should be an edge property, not a block end property.
329 /// TODO: Per PR2563, we could infer value range information about a
330 /// predecessor based on its terminator.
333 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
334 // "I" is a non-local compare-with-a-constant instruction. This would be
335 // able to handle value inequalities better, for example if the compare is
336 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
337 // Perhaps getConstantOnEdge should be smart enough to do this?
339 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
341 // If the value is known by LazyValueInfo to be a constant in a
342 // predecessor, use that information to try to thread this block.
343 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
345 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst)))
348 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), P));
351 return !Result.empty();
357 /// If I is a PHI node, then we know the incoming values for any constants.
358 if (PHINode *PN = dyn_cast<PHINode>(I)) {
359 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
360 Value *InVal = PN->getIncomingValue(i);
361 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) {
362 ConstantInt *CI = dyn_cast<ConstantInt>(InVal);
363 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i)));
365 Constant *CI = LVI->getConstantOnEdge(InVal,
366 PN->getIncomingBlock(i), BB);
367 // LVI returns null is no value could be determined.
369 PushConstantIntOrUndef(Result, CI, PN->getIncomingBlock(i));
373 return !Result.empty();
376 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals;
378 // Handle some boolean conditions.
379 if (I->getType()->getPrimitiveSizeInBits() == 1) {
381 // X & false -> false
382 if (I->getOpcode() == Instruction::Or ||
383 I->getOpcode() == Instruction::And) {
384 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
385 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals);
387 if (LHSVals.empty() && RHSVals.empty())
390 ConstantInt *InterestingVal;
391 if (I->getOpcode() == Instruction::Or)
392 InterestingVal = ConstantInt::getTrue(I->getContext());
394 InterestingVal = ConstantInt::getFalse(I->getContext());
396 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
398 // Scan for the sentinel. If we find an undef, force it to the
399 // interesting value: x|undef -> true and x&undef -> false.
400 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
401 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) {
402 Result.push_back(LHSVals[i]);
403 Result.back().first = InterestingVal;
404 LHSKnownBBs.insert(LHSVals[i].second);
406 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
407 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) {
408 // If we already inferred a value for this block on the LHS, don't
410 if (!LHSKnownBBs.count(RHSVals[i].second)) {
411 Result.push_back(RHSVals[i]);
412 Result.back().first = InterestingVal;
416 return !Result.empty();
419 // Handle the NOT form of XOR.
420 if (I->getOpcode() == Instruction::Xor &&
421 isa<ConstantInt>(I->getOperand(1)) &&
422 cast<ConstantInt>(I->getOperand(1))->isOne()) {
423 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result);
427 // Invert the known values.
428 for (unsigned i = 0, e = Result.size(); i != e; ++i)
431 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first));
436 // Try to simplify some other binary operator values.
437 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
438 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
439 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals;
440 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals);
442 // Try to use constant folding to simplify the binary operator.
443 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
444 Constant *V = LHSVals[i].first ? LHSVals[i].first :
445 cast<Constant>(UndefValue::get(BO->getType()));
446 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
448 PushConstantIntOrUndef(Result, Folded, LHSVals[i].second);
452 return !Result.empty();
455 // Handle compare with phi operand, where the PHI is defined in this block.
456 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
457 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
458 if (PN && PN->getParent() == BB) {
459 // We can do this simplification if any comparisons fold to true or false.
461 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
462 BasicBlock *PredBB = PN->getIncomingBlock(i);
463 Value *LHS = PN->getIncomingValue(i);
464 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
466 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
468 if (!LVI || !isa<Constant>(RHS))
471 LazyValueInfo::Tristate
472 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
473 cast<Constant>(RHS), PredBB, BB);
474 if (ResT == LazyValueInfo::Unknown)
476 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
479 if (Constant *ConstRes = dyn_cast<Constant>(Res))
480 PushConstantIntOrUndef(Result, ConstRes, PredBB);
483 return !Result.empty();
487 // If comparing a live-in value against a constant, see if we know the
488 // live-in value on any predecessors.
489 if (LVI && isa<Constant>(Cmp->getOperand(1)) &&
490 Cmp->getType()->isIntegerTy()) {
491 if (!isa<Instruction>(Cmp->getOperand(0)) ||
492 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
493 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
495 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
497 // If the value is known by LazyValueInfo to be a constant in a
498 // predecessor, use that information to try to thread this block.
499 LazyValueInfo::Tristate Res =
500 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
502 if (Res == LazyValueInfo::Unknown)
505 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
506 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), P));
509 return !Result.empty();
512 // Try to find a constant value for the LHS of a comparison,
513 // and evaluate it statically if we can.
514 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
515 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals;
516 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
518 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
519 Constant *V = LHSVals[i].first ? LHSVals[i].first :
520 cast<Constant>(UndefValue::get(CmpConst->getType()));
521 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
523 PushConstantIntOrUndef(Result, Folded, LHSVals[i].second);
526 return !Result.empty();
532 // If all else fails, see if LVI can figure out a constant value for us.
533 Constant *CI = LVI->getConstant(V, BB);
534 ConstantInt *CInt = dyn_cast_or_null<ConstantInt>(CI);
536 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
537 Result.push_back(std::make_pair(CInt, *PI));
540 return !Result.empty();
548 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
549 /// in an undefined jump, decide which block is best to revector to.
551 /// Since we can pick an arbitrary destination, we pick the successor with the
552 /// fewest predecessors. This should reduce the in-degree of the others.
554 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
555 TerminatorInst *BBTerm = BB->getTerminator();
556 unsigned MinSucc = 0;
557 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
558 // Compute the successor with the minimum number of predecessors.
559 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
560 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
561 TestBB = BBTerm->getSuccessor(i);
562 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
563 if (NumPreds < MinNumPreds)
570 /// ProcessBlock - If there are any predecessors whose control can be threaded
571 /// through to a successor, transform them now.
572 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
573 // If the block is trivially dead, just return and let the caller nuke it.
574 // This simplifies other transformations.
575 if (pred_begin(BB) == pred_end(BB) &&
576 BB != &BB->getParent()->getEntryBlock())
579 // If this block has a single predecessor, and if that pred has a single
580 // successor, merge the blocks. This encourages recursive jump threading
581 // because now the condition in this block can be threaded through
582 // predecessors of our predecessor block.
583 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
584 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
586 // If SinglePred was a loop header, BB becomes one.
587 if (LoopHeaders.erase(SinglePred))
588 LoopHeaders.insert(BB);
590 // Remember if SinglePred was the entry block of the function. If so, we
591 // will need to move BB back to the entry position.
592 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
593 if (LVI) LVI->eraseBlock(SinglePred);
594 MergeBasicBlockIntoOnlyPred(BB);
596 if (isEntry && BB != &BB->getParent()->getEntryBlock())
597 BB->moveBefore(&BB->getParent()->getEntryBlock());
602 // Look to see if the terminator is a branch of switch, if not we can't thread
605 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
606 // Can't thread an unconditional jump.
607 if (BI->isUnconditional()) return false;
608 Condition = BI->getCondition();
609 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
610 Condition = SI->getCondition();
612 return false; // Must be an invoke.
614 // If the terminator of this block is branching on a constant, simplify the
615 // terminator to an unconditional branch. This can occur due to threading in
617 if (isa<ConstantInt>(Condition)) {
618 DEBUG(dbgs() << " In block '" << BB->getName()
619 << "' folding terminator: " << *BB->getTerminator() << '\n');
621 ConstantFoldTerminator(BB);
625 // If the terminator is branching on an undef, we can pick any of the
626 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
627 if (isa<UndefValue>(Condition)) {
628 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
630 // Fold the branch/switch.
631 TerminatorInst *BBTerm = BB->getTerminator();
632 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
633 if (i == BestSucc) continue;
634 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD);
637 DEBUG(dbgs() << " In block '" << BB->getName()
638 << "' folding undef terminator: " << *BBTerm << '\n');
639 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
640 BBTerm->eraseFromParent();
644 Instruction *CondInst = dyn_cast<Instruction>(Condition);
646 // If the condition is an instruction defined in another block, see if a
647 // predecessor has the same condition:
652 !Condition->hasOneUse() && // Multiple uses.
653 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition.
654 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
655 if (isa<BranchInst>(BB->getTerminator())) {
656 for (; PI != E; ++PI) {
658 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator()))
659 if (PBI->isConditional() && PBI->getCondition() == Condition &&
660 ProcessBranchOnDuplicateCond(P, BB))
664 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator");
665 for (; PI != E; ++PI) {
667 if (SwitchInst *PSI = dyn_cast<SwitchInst>(P->getTerminator()))
668 if (PSI->getCondition() == Condition &&
669 ProcessSwitchOnDuplicateCond(P, BB))
675 // All the rest of our checks depend on the condition being an instruction.
677 // FIXME: Unify this with code below.
678 if (LVI && ProcessThreadableEdges(Condition, BB))
684 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
686 (!isa<PHINode>(CondCmp->getOperand(0)) ||
687 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) {
688 // If we have a comparison, loop over the predecessors to see if there is
689 // a condition with a lexically identical value.
690 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
691 for (; PI != E; ++PI) {
693 if (BranchInst *PBI = dyn_cast<BranchInst>(P->getTerminator()))
694 if (PBI->isConditional() && P != BB) {
695 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) {
696 if (CI->getOperand(0) == CondCmp->getOperand(0) &&
697 CI->getOperand(1) == CondCmp->getOperand(1) &&
698 CI->getPredicate() == CondCmp->getPredicate()) {
699 // TODO: Could handle things like (x != 4) --> (x == 17)
700 if (ProcessBranchOnDuplicateCond(P, BB))
708 // For a comparison where the LHS is outside this block, it's possible
709 // that we've branched on it before. Used LVI to see if we can simplify
710 // the branch based on that.
711 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
712 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
713 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
714 if (LVI && CondBr && CondConst && CondBr->isConditional() && PI != PE &&
715 (!isa<Instruction>(CondCmp->getOperand(0)) ||
716 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
717 // For predecessor edge, determine if the comparison is true or false
718 // on that edge. If they're all true or all false, we can simplify the
720 // FIXME: We could handle mixed true/false by duplicating code.
721 LazyValueInfo::Tristate Baseline =
722 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
724 if (Baseline != LazyValueInfo::Unknown) {
725 // Check that all remaining incoming values match the first one.
727 LazyValueInfo::Tristate Ret = LVI->getPredicateOnEdge(
728 CondCmp->getPredicate(),
729 CondCmp->getOperand(0),
731 if (Ret != Baseline) break;
734 // If we terminated early, then one of the values didn't match.
736 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
737 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
738 RemovePredecessorAndSimplify(CondBr->getSuccessor(ToRemove), BB, TD);
739 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
740 CondBr->eraseFromParent();
747 // Check for some cases that are worth simplifying. Right now we want to look
748 // for loads that are used by a switch or by the condition for the branch. If
749 // we see one, check to see if it's partially redundant. If so, insert a PHI
750 // which can then be used to thread the values.
752 Value *SimplifyValue = CondInst;
753 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
754 if (isa<Constant>(CondCmp->getOperand(1)))
755 SimplifyValue = CondCmp->getOperand(0);
757 // TODO: There are other places where load PRE would be profitable, such as
758 // more complex comparisons.
759 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
760 if (SimplifyPartiallyRedundantLoad(LI))
764 // Handle a variety of cases where we are branching on something derived from
765 // a PHI node in the current block. If we can prove that any predecessors
766 // compute a predictable value based on a PHI node, thread those predecessors.
768 if (ProcessThreadableEdges(CondInst, BB))
771 // If this is an otherwise-unfoldable branch on a phi node in the current
772 // block, see if we can simplify.
773 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
774 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
775 return ProcessBranchOnPHI(PN);
778 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
779 if (CondInst->getOpcode() == Instruction::Xor &&
780 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
781 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
784 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
785 // "(X == 4)", thread through this block.
790 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
791 /// block that jump on exactly the same condition. This means that we almost
792 /// always know the direction of the edge in the DESTBB:
794 /// br COND, DESTBB, BBY
796 /// br COND, BBZ, BBW
798 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
799 /// in DESTBB, we have to thread over it.
800 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
802 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
804 // If both successors of PredBB go to DESTBB, we don't know anything. We can
805 // fold the branch to an unconditional one, which allows other recursive
808 if (PredBI->getSuccessor(1) != BB)
810 else if (PredBI->getSuccessor(0) != BB)
813 DEBUG(dbgs() << " In block '" << PredBB->getName()
814 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
816 ConstantFoldTerminator(PredBB);
820 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
822 // If the dest block has one predecessor, just fix the branch condition to a
823 // constant and fold it.
824 if (BB->getSinglePredecessor()) {
825 DEBUG(dbgs() << " In block '" << BB->getName()
826 << "' folding condition to '" << BranchDir << "': "
827 << *BB->getTerminator() << '\n');
829 Value *OldCond = DestBI->getCondition();
830 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
832 // Delete dead instructions before we fold the branch. Folding the branch
833 // can eliminate edges from the CFG which can end up deleting OldCond.
834 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
835 ConstantFoldTerminator(BB);
840 // Next, figure out which successor we are threading to.
841 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
843 SmallVector<BasicBlock*, 2> Preds;
844 Preds.push_back(PredBB);
846 // Ok, try to thread it!
847 return ThreadEdge(BB, Preds, SuccBB);
850 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
851 /// block that switch on exactly the same condition. This means that we almost
852 /// always know the direction of the edge in the DESTBB:
854 /// switch COND [... DESTBB, BBY ... ]
856 /// switch COND [... BBZ, BBW ]
858 /// Optimizing switches like this is very important, because simplifycfg builds
859 /// switches out of repeated 'if' conditions.
860 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
861 BasicBlock *DestBB) {
862 // Can't thread edge to self.
863 if (PredBB == DestBB)
866 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
867 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
869 // There are a variety of optimizations that we can potentially do on these
870 // blocks: we order them from most to least preferable.
872 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
873 // directly to their destination. This does not introduce *any* code size
874 // growth. Skip debug info first.
875 BasicBlock::iterator BBI = DestBB->begin();
876 while (isa<DbgInfoIntrinsic>(BBI))
879 // FIXME: Thread if it just contains a PHI.
880 if (isa<SwitchInst>(BBI)) {
881 bool MadeChange = false;
882 // Ignore the default edge for now.
883 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
884 ConstantInt *DestVal = DestSI->getCaseValue(i);
885 BasicBlock *DestSucc = DestSI->getSuccessor(i);
887 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
888 // PredSI has an explicit case for it. If so, forward. If it is covered
889 // by the default case, we can't update PredSI.
890 unsigned PredCase = PredSI->findCaseValue(DestVal);
891 if (PredCase == 0) continue;
893 // If PredSI doesn't go to DestBB on this value, then it won't reach the
894 // case on this condition.
895 if (PredSI->getSuccessor(PredCase) != DestBB &&
896 DestSI->getSuccessor(i) != DestBB)
899 // Do not forward this if it already goes to this destination, this would
900 // be an infinite loop.
901 if (PredSI->getSuccessor(PredCase) == DestSucc)
904 // Otherwise, we're safe to make the change. Make sure that the edge from
905 // DestSI to DestSucc is not critical and has no PHI nodes.
906 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
907 DEBUG(dbgs() << "THROUGH: " << *DestSI);
909 // If the destination has PHI nodes, just split the edge for updating
911 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
912 SplitCriticalEdge(DestSI, i, this);
913 DestSucc = DestSI->getSuccessor(i);
915 FoldSingleEntryPHINodes(DestSucc);
916 PredSI->setSuccessor(PredCase, DestSucc);
928 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
929 /// load instruction, eliminate it by replacing it with a PHI node. This is an
930 /// important optimization that encourages jump threading, and needs to be run
931 /// interlaced with other jump threading tasks.
932 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
933 // Don't hack volatile loads.
934 if (LI->isVolatile()) return false;
936 // If the load is defined in a block with exactly one predecessor, it can't be
937 // partially redundant.
938 BasicBlock *LoadBB = LI->getParent();
939 if (LoadBB->getSinglePredecessor())
942 Value *LoadedPtr = LI->getOperand(0);
944 // If the loaded operand is defined in the LoadBB, it can't be available.
945 // TODO: Could do simple PHI translation, that would be fun :)
946 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
947 if (PtrOp->getParent() == LoadBB)
950 // Scan a few instructions up from the load, to see if it is obviously live at
951 // the entry to its block.
952 BasicBlock::iterator BBIt = LI;
954 if (Value *AvailableVal =
955 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
956 // If the value if the load is locally available within the block, just use
957 // it. This frequently occurs for reg2mem'd allocas.
958 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
960 // If the returned value is the load itself, replace with an undef. This can
961 // only happen in dead loops.
962 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
963 LI->replaceAllUsesWith(AvailableVal);
964 LI->eraseFromParent();
968 // Otherwise, if we scanned the whole block and got to the top of the block,
969 // we know the block is locally transparent to the load. If not, something
970 // might clobber its value.
971 if (BBIt != LoadBB->begin())
975 SmallPtrSet<BasicBlock*, 8> PredsScanned;
976 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
977 AvailablePredsTy AvailablePreds;
978 BasicBlock *OneUnavailablePred = 0;
980 // If we got here, the loaded value is transparent through to the start of the
981 // block. Check to see if it is available in any of the predecessor blocks.
982 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
984 BasicBlock *PredBB = *PI;
986 // If we already scanned this predecessor, skip it.
987 if (!PredsScanned.insert(PredBB))
990 // Scan the predecessor to see if the value is available in the pred.
991 BBIt = PredBB->end();
992 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
993 if (!PredAvailable) {
994 OneUnavailablePred = PredBB;
998 // If so, this load is partially redundant. Remember this info so that we
999 // can create a PHI node.
1000 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1003 // If the loaded value isn't available in any predecessor, it isn't partially
1005 if (AvailablePreds.empty()) return false;
1007 // Okay, the loaded value is available in at least one (and maybe all!)
1008 // predecessors. If the value is unavailable in more than one unique
1009 // predecessor, we want to insert a merge block for those common predecessors.
1010 // This ensures that we only have to insert one reload, thus not increasing
1012 BasicBlock *UnavailablePred = 0;
1014 // If there is exactly one predecessor where the value is unavailable, the
1015 // already computed 'OneUnavailablePred' block is it. If it ends in an
1016 // unconditional branch, we know that it isn't a critical edge.
1017 if (PredsScanned.size() == AvailablePreds.size()+1 &&
1018 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1019 UnavailablePred = OneUnavailablePred;
1020 } else if (PredsScanned.size() != AvailablePreds.size()) {
1021 // Otherwise, we had multiple unavailable predecessors or we had a critical
1022 // edge from the one.
1023 SmallVector<BasicBlock*, 8> PredsToSplit;
1024 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1026 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
1027 AvailablePredSet.insert(AvailablePreds[i].first);
1029 // Add all the unavailable predecessors to the PredsToSplit list.
1030 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
1032 BasicBlock *P = *PI;
1033 // If the predecessor is an indirect goto, we can't split the edge.
1034 if (isa<IndirectBrInst>(P->getTerminator()))
1037 if (!AvailablePredSet.count(P))
1038 PredsToSplit.push_back(P);
1041 // Split them out to their own block.
1043 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
1044 "thread-pre-split", this);
1047 // If the value isn't available in all predecessors, then there will be
1048 // exactly one where it isn't available. Insert a load on that edge and add
1049 // it to the AvailablePreds list.
1050 if (UnavailablePred) {
1051 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1052 "Can't handle critical edge here!");
1053 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
1055 UnavailablePred->getTerminator());
1056 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1059 // Now we know that each predecessor of this block has a value in
1060 // AvailablePreds, sort them for efficient access as we're walking the preds.
1061 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1063 // Create a PHI node at the start of the block for the PRE'd load value.
1064 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
1067 // Insert new entries into the PHI for each predecessor. A single block may
1068 // have multiple entries here.
1069 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
1071 BasicBlock *P = *PI;
1072 AvailablePredsTy::iterator I =
1073 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1074 std::make_pair(P, (Value*)0));
1076 assert(I != AvailablePreds.end() && I->first == P &&
1077 "Didn't find entry for predecessor!");
1079 PN->addIncoming(I->second, I->first);
1082 //cerr << "PRE: " << *LI << *PN << "\n";
1084 LI->replaceAllUsesWith(PN);
1085 LI->eraseFromParent();
1090 /// FindMostPopularDest - The specified list contains multiple possible
1091 /// threadable destinations. Pick the one that occurs the most frequently in
1094 FindMostPopularDest(BasicBlock *BB,
1095 const SmallVectorImpl<std::pair<BasicBlock*,
1096 BasicBlock*> > &PredToDestList) {
1097 assert(!PredToDestList.empty());
1099 // Determine popularity. If there are multiple possible destinations, we
1100 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1101 // blocks with known and real destinations to threading undef. We'll handle
1102 // them later if interesting.
1103 DenseMap<BasicBlock*, unsigned> DestPopularity;
1104 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1105 if (PredToDestList[i].second)
1106 DestPopularity[PredToDestList[i].second]++;
1108 // Find the most popular dest.
1109 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1110 BasicBlock *MostPopularDest = DPI->first;
1111 unsigned Popularity = DPI->second;
1112 SmallVector<BasicBlock*, 4> SamePopularity;
1114 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1115 // If the popularity of this entry isn't higher than the popularity we've
1116 // seen so far, ignore it.
1117 if (DPI->second < Popularity)
1119 else if (DPI->second == Popularity) {
1120 // If it is the same as what we've seen so far, keep track of it.
1121 SamePopularity.push_back(DPI->first);
1123 // If it is more popular, remember it.
1124 SamePopularity.clear();
1125 MostPopularDest = DPI->first;
1126 Popularity = DPI->second;
1130 // Okay, now we know the most popular destination. If there is more than
1131 // destination, we need to determine one. This is arbitrary, but we need
1132 // to make a deterministic decision. Pick the first one that appears in the
1134 if (!SamePopularity.empty()) {
1135 SamePopularity.push_back(MostPopularDest);
1136 TerminatorInst *TI = BB->getTerminator();
1137 for (unsigned i = 0; ; ++i) {
1138 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1140 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1141 TI->getSuccessor(i)) == SamePopularity.end())
1144 MostPopularDest = TI->getSuccessor(i);
1149 // Okay, we have finally picked the most popular destination.
1150 return MostPopularDest;
1153 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) {
1154 // If threading this would thread across a loop header, don't even try to
1156 if (LoopHeaders.count(BB))
1159 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues;
1160 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues))
1163 assert(!PredValues.empty() &&
1164 "ComputeValueKnownInPredecessors returned true with no values");
1166 DEBUG(dbgs() << "IN BB: " << *BB;
1167 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1168 dbgs() << " BB '" << BB->getName() << "': FOUND condition = ";
1169 if (PredValues[i].first)
1170 dbgs() << *PredValues[i].first;
1173 dbgs() << " for pred '" << PredValues[i].second->getName()
1177 // Decide what we want to thread through. Convert our list of known values to
1178 // a list of known destinations for each pred. This also discards duplicate
1179 // predecessors and keeps track of the undefined inputs (which are represented
1180 // as a null dest in the PredToDestList).
1181 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1182 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1184 BasicBlock *OnlyDest = 0;
1185 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1187 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1188 BasicBlock *Pred = PredValues[i].second;
1189 if (!SeenPreds.insert(Pred))
1190 continue; // Duplicate predecessor entry.
1192 // If the predecessor ends with an indirect goto, we can't change its
1194 if (isa<IndirectBrInst>(Pred->getTerminator()))
1197 ConstantInt *Val = PredValues[i].first;
1200 if (Val == 0) // Undef.
1202 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1203 DestBB = BI->getSuccessor(Val->isZero());
1205 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
1206 DestBB = SI->getSuccessor(SI->findCaseValue(Val));
1209 // If we have exactly one destination, remember it for efficiency below.
1212 else if (OnlyDest != DestBB)
1213 OnlyDest = MultipleDestSentinel;
1215 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1218 // If all edges were unthreadable, we fail.
1219 if (PredToDestList.empty())
1222 // Determine which is the most common successor. If we have many inputs and
1223 // this block is a switch, we want to start by threading the batch that goes
1224 // to the most popular destination first. If we only know about one
1225 // threadable destination (the common case) we can avoid this.
1226 BasicBlock *MostPopularDest = OnlyDest;
1228 if (MostPopularDest == MultipleDestSentinel)
1229 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1231 // Now that we know what the most popular destination is, factor all
1232 // predecessors that will jump to it into a single predecessor.
1233 SmallVector<BasicBlock*, 16> PredsToFactor;
1234 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1235 if (PredToDestList[i].second == MostPopularDest) {
1236 BasicBlock *Pred = PredToDestList[i].first;
1238 // This predecessor may be a switch or something else that has multiple
1239 // edges to the block. Factor each of these edges by listing them
1240 // according to # occurrences in PredsToFactor.
1241 TerminatorInst *PredTI = Pred->getTerminator();
1242 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1243 if (PredTI->getSuccessor(i) == BB)
1244 PredsToFactor.push_back(Pred);
1247 // If the threadable edges are branching on an undefined value, we get to pick
1248 // the destination that these predecessors should get to.
1249 if (MostPopularDest == 0)
1250 MostPopularDest = BB->getTerminator()->
1251 getSuccessor(GetBestDestForJumpOnUndef(BB));
1253 // Ok, try to thread it!
1254 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1257 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1258 /// a PHI node in the current block. See if there are any simplifications we
1259 /// can do based on inputs to the phi node.
1261 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1262 BasicBlock *BB = PN->getParent();
1264 // TODO: We could make use of this to do it once for blocks with common PHI
1266 SmallVector<BasicBlock*, 1> PredBBs;
1269 // If any of the predecessor blocks end in an unconditional branch, we can
1270 // *duplicate* the conditional branch into that block in order to further
1271 // encourage jump threading and to eliminate cases where we have branch on a
1272 // phi of an icmp (branch on icmp is much better).
1273 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1274 BasicBlock *PredBB = PN->getIncomingBlock(i);
1275 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1276 if (PredBr->isUnconditional()) {
1277 PredBBs[0] = PredBB;
1278 // Try to duplicate BB into PredBB.
1279 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1287 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1288 /// a xor instruction in the current block. See if there are any
1289 /// simplifications we can do based on inputs to the xor.
1291 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1292 BasicBlock *BB = BO->getParent();
1294 // If either the LHS or RHS of the xor is a constant, don't do this
1296 if (isa<ConstantInt>(BO->getOperand(0)) ||
1297 isa<ConstantInt>(BO->getOperand(1)))
1300 // If the first instruction in BB isn't a phi, we won't be able to infer
1301 // anything special about any particular predecessor.
1302 if (!isa<PHINode>(BB->front()))
1305 // If we have a xor as the branch input to this block, and we know that the
1306 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1307 // the condition into the predecessor and fix that value to true, saving some
1308 // logical ops on that path and encouraging other paths to simplify.
1310 // This copies something like this:
1313 // %X = phi i1 [1], [%X']
1314 // %Y = icmp eq i32 %A, %B
1315 // %Z = xor i1 %X, %Y
1320 // %Y = icmp ne i32 %A, %B
1323 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues;
1325 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) {
1326 assert(XorOpValues.empty());
1327 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues))
1332 assert(!XorOpValues.empty() &&
1333 "ComputeValueKnownInPredecessors returned true with no values");
1335 // Scan the information to see which is most popular: true or false. The
1336 // predecessors can be of the set true, false, or undef.
1337 unsigned NumTrue = 0, NumFalse = 0;
1338 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1339 if (!XorOpValues[i].first) continue; // Ignore undefs for the count.
1340 if (XorOpValues[i].first->isZero())
1346 // Determine which value to split on, true, false, or undef if neither.
1347 ConstantInt *SplitVal = 0;
1348 if (NumTrue > NumFalse)
1349 SplitVal = ConstantInt::getTrue(BB->getContext());
1350 else if (NumTrue != 0 || NumFalse != 0)
1351 SplitVal = ConstantInt::getFalse(BB->getContext());
1353 // Collect all of the blocks that this can be folded into so that we can
1354 // factor this once and clone it once.
1355 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1356 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1357 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue;
1359 BlocksToFoldInto.push_back(XorOpValues[i].second);
1362 // If we inferred a value for all of the predecessors, then duplication won't
1363 // help us. However, we can just replace the LHS or RHS with the constant.
1364 if (BlocksToFoldInto.size() ==
1365 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1366 if (SplitVal == 0) {
1367 // If all preds provide undef, just nuke the xor, because it is undef too.
1368 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1369 BO->eraseFromParent();
1370 } else if (SplitVal->isZero()) {
1371 // If all preds provide 0, replace the xor with the other input.
1372 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1373 BO->eraseFromParent();
1375 // If all preds provide 1, set the computed value to 1.
1376 BO->setOperand(!isLHS, SplitVal);
1382 // Try to duplicate BB into PredBB.
1383 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1387 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1388 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1389 /// NewPred using the entries from OldPred (suitably mapped).
1390 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1391 BasicBlock *OldPred,
1392 BasicBlock *NewPred,
1393 DenseMap<Instruction*, Value*> &ValueMap) {
1394 for (BasicBlock::iterator PNI = PHIBB->begin();
1395 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1396 // Ok, we have a PHI node. Figure out what the incoming value was for the
1398 Value *IV = PN->getIncomingValueForBlock(OldPred);
1400 // Remap the value if necessary.
1401 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1402 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1403 if (I != ValueMap.end())
1407 PN->addIncoming(IV, NewPred);
1411 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1412 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1413 /// across BB. Transform the IR to reflect this change.
1414 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1415 const SmallVectorImpl<BasicBlock*> &PredBBs,
1416 BasicBlock *SuccBB) {
1417 // If threading to the same block as we come from, we would infinite loop.
1419 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1420 << "' - would thread to self!\n");
1424 // If threading this would thread across a loop header, don't thread the edge.
1425 // See the comments above FindLoopHeaders for justifications and caveats.
1426 if (LoopHeaders.count(BB)) {
1427 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1428 << "' to dest BB '" << SuccBB->getName()
1429 << "' - it might create an irreducible loop!\n");
1433 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1434 if (JumpThreadCost > Threshold) {
1435 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1436 << "' - Cost is too high: " << JumpThreadCost << "\n");
1440 // And finally, do it! Start by factoring the predecessors is needed.
1442 if (PredBBs.size() == 1)
1443 PredBB = PredBBs[0];
1445 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1446 << " common predecessors.\n");
1447 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1451 // And finally, do it!
1452 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1453 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1454 << ", across block:\n "
1458 LVI->threadEdge(PredBB, BB, SuccBB);
1460 // We are going to have to map operands from the original BB block to the new
1461 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1462 // account for entry from PredBB.
1463 DenseMap<Instruction*, Value*> ValueMapping;
1465 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1466 BB->getName()+".thread",
1467 BB->getParent(), BB);
1468 NewBB->moveAfter(PredBB);
1470 BasicBlock::iterator BI = BB->begin();
1471 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1472 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1474 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1475 // mapping and using it to remap operands in the cloned instructions.
1476 for (; !isa<TerminatorInst>(BI); ++BI) {
1477 Instruction *New = BI->clone();
1478 New->setName(BI->getName());
1479 NewBB->getInstList().push_back(New);
1480 ValueMapping[BI] = New;
1482 // Remap operands to patch up intra-block references.
1483 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1484 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1485 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1486 if (I != ValueMapping.end())
1487 New->setOperand(i, I->second);
1491 // We didn't copy the terminator from BB over to NewBB, because there is now
1492 // an unconditional jump to SuccBB. Insert the unconditional jump.
1493 BranchInst::Create(SuccBB, NewBB);
1495 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1496 // PHI nodes for NewBB now.
1497 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1499 // If there were values defined in BB that are used outside the block, then we
1500 // now have to update all uses of the value to use either the original value,
1501 // the cloned value, or some PHI derived value. This can require arbitrary
1502 // PHI insertion, of which we are prepared to do, clean these up now.
1503 SSAUpdater SSAUpdate;
1504 SmallVector<Use*, 16> UsesToRename;
1505 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1506 // Scan all uses of this instruction to see if it is used outside of its
1507 // block, and if so, record them in UsesToRename.
1508 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1510 Instruction *User = cast<Instruction>(*UI);
1511 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1512 if (UserPN->getIncomingBlock(UI) == BB)
1514 } else if (User->getParent() == BB)
1517 UsesToRename.push_back(&UI.getUse());
1520 // If there are no uses outside the block, we're done with this instruction.
1521 if (UsesToRename.empty())
1524 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1526 // We found a use of I outside of BB. Rename all uses of I that are outside
1527 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1528 // with the two values we know.
1529 SSAUpdate.Initialize(I);
1530 SSAUpdate.AddAvailableValue(BB, I);
1531 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1533 while (!UsesToRename.empty())
1534 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1535 DEBUG(dbgs() << "\n");
1539 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1540 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1541 // us to simplify any PHI nodes in BB.
1542 TerminatorInst *PredTerm = PredBB->getTerminator();
1543 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1544 if (PredTerm->getSuccessor(i) == BB) {
1545 RemovePredecessorAndSimplify(BB, PredBB, TD);
1546 PredTerm->setSuccessor(i, NewBB);
1549 // At this point, the IR is fully up to date and consistent. Do a quick scan
1550 // over the new instructions and zap any that are constants or dead. This
1551 // frequently happens because of phi translation.
1552 SimplifyInstructionsInBlock(NewBB, TD);
1554 // Threaded an edge!
1559 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1560 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1561 /// If we can duplicate the contents of BB up into PredBB do so now, this
1562 /// improves the odds that the branch will be on an analyzable instruction like
1564 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1565 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1566 assert(!PredBBs.empty() && "Can't handle an empty set");
1568 // If BB is a loop header, then duplicating this block outside the loop would
1569 // cause us to transform this into an irreducible loop, don't do this.
1570 // See the comments above FindLoopHeaders for justifications and caveats.
1571 if (LoopHeaders.count(BB)) {
1572 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1573 << "' into predecessor block '" << PredBBs[0]->getName()
1574 << "' - it might create an irreducible loop!\n");
1578 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1579 if (DuplicationCost > Threshold) {
1580 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1581 << "' - Cost is too high: " << DuplicationCost << "\n");
1585 // And finally, do it! Start by factoring the predecessors is needed.
1587 if (PredBBs.size() == 1)
1588 PredBB = PredBBs[0];
1590 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1591 << " common predecessors.\n");
1592 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1596 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1598 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1599 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1600 << DuplicationCost << " block is:" << *BB << "\n");
1602 // Unless PredBB ends with an unconditional branch, split the edge so that we
1603 // can just clone the bits from BB into the end of the new PredBB.
1604 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1606 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1607 PredBB = SplitEdge(PredBB, BB, this);
1608 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1611 // We are going to have to map operands from the original BB block into the
1612 // PredBB block. Evaluate PHI nodes in BB.
1613 DenseMap<Instruction*, Value*> ValueMapping;
1615 BasicBlock::iterator BI = BB->begin();
1616 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1617 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1619 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1620 // mapping and using it to remap operands in the cloned instructions.
1621 for (; BI != BB->end(); ++BI) {
1622 Instruction *New = BI->clone();
1624 // Remap operands to patch up intra-block references.
1625 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1626 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1627 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1628 if (I != ValueMapping.end())
1629 New->setOperand(i, I->second);
1632 // If this instruction can be simplified after the operands are updated,
1633 // just use the simplified value instead. This frequently happens due to
1635 if (Value *IV = SimplifyInstruction(New, TD)) {
1637 ValueMapping[BI] = IV;
1639 // Otherwise, insert the new instruction into the block.
1640 New->setName(BI->getName());
1641 PredBB->getInstList().insert(OldPredBranch, New);
1642 ValueMapping[BI] = New;
1646 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1647 // add entries to the PHI nodes for branch from PredBB now.
1648 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1649 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1651 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1654 // If there were values defined in BB that are used outside the block, then we
1655 // now have to update all uses of the value to use either the original value,
1656 // the cloned value, or some PHI derived value. This can require arbitrary
1657 // PHI insertion, of which we are prepared to do, clean these up now.
1658 SSAUpdater SSAUpdate;
1659 SmallVector<Use*, 16> UsesToRename;
1660 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1661 // Scan all uses of this instruction to see if it is used outside of its
1662 // block, and if so, record them in UsesToRename.
1663 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1665 Instruction *User = cast<Instruction>(*UI);
1666 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1667 if (UserPN->getIncomingBlock(UI) == BB)
1669 } else if (User->getParent() == BB)
1672 UsesToRename.push_back(&UI.getUse());
1675 // If there are no uses outside the block, we're done with this instruction.
1676 if (UsesToRename.empty())
1679 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1681 // We found a use of I outside of BB. Rename all uses of I that are outside
1682 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1683 // with the two values we know.
1684 SSAUpdate.Initialize(I);
1685 SSAUpdate.AddAvailableValue(BB, I);
1686 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1688 while (!UsesToRename.empty())
1689 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1690 DEBUG(dbgs() << "\n");
1693 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1695 RemovePredecessorAndSimplify(BB, PredBB, TD);
1697 // Remove the unconditional branch at the end of the PredBB block.
1698 OldPredBranch->eraseFromParent();