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/Transforms/Utils/BasicBlockUtils.h"
22 #include "llvm/Transforms/Utils/Local.h"
23 #include "llvm/Transforms/Utils/SSAUpdater.h"
24 #include "llvm/Target/TargetData.h"
25 #include "llvm/ADT/DenseMap.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/ADT/STLExtras.h"
28 #include "llvm/ADT/SmallPtrSet.h"
29 #include "llvm/ADT/SmallSet.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ValueHandle.h"
33 #include "llvm/Support/raw_ostream.h"
36 STATISTIC(NumThreads, "Number of jumps threaded");
37 STATISTIC(NumFolds, "Number of terminators folded");
38 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
40 static cl::opt<unsigned>
41 Threshold("jump-threading-threshold",
42 cl::desc("Max block size to duplicate for jump threading"),
43 cl::init(6), cl::Hidden);
45 // Turn on use of LazyValueInfo.
47 EnableLVI("enable-jump-threading-lvi", cl::ReallyHidden);
52 /// This pass performs 'jump threading', which looks at blocks that have
53 /// multiple predecessors and multiple successors. If one or more of the
54 /// predecessors of the block can be proven to always jump to one of the
55 /// successors, we forward the edge from the predecessor to the successor by
56 /// duplicating the contents of this block.
58 /// An example of when this can occur is code like this:
65 /// In this case, the unconditional branch at the end of the first if can be
66 /// revectored to the false side of the second if.
68 class JumpThreading : public FunctionPass {
72 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
74 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
77 static char ID; // Pass identification
78 JumpThreading() : FunctionPass(&ID) {}
80 bool runOnFunction(Function &F);
82 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
84 AU.addRequired<LazyValueInfo>();
87 void FindLoopHeaders(Function &F);
88 bool ProcessBlock(BasicBlock *BB);
89 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
91 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
94 typedef SmallVectorImpl<std::pair<ConstantInt*,
95 BasicBlock*> > PredValueInfo;
97 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
98 PredValueInfo &Result);
99 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB);
102 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
103 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
105 bool ProcessJumpOnPHI(PHINode *PN);
107 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
111 char JumpThreading::ID = 0;
112 static RegisterPass<JumpThreading>
113 X("jump-threading", "Jump Threading");
115 // Public interface to the Jump Threading pass
116 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
118 /// runOnFunction - Top level algorithm.
120 bool JumpThreading::runOnFunction(Function &F) {
121 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
122 TD = getAnalysisIfAvailable<TargetData>();
123 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0;
127 bool Changed, EverChanged = false;
130 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
132 // Thread all of the branches we can over this block.
133 while (ProcessBlock(BB))
138 // If the block is trivially dead, zap it. This eliminates the successor
139 // edges which simplifies the CFG.
140 if (pred_begin(BB) == pred_end(BB) &&
141 BB != &BB->getParent()->getEntryBlock()) {
142 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
143 << "' with terminator: " << *BB->getTerminator() << '\n');
144 LoopHeaders.erase(BB);
147 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
148 // Can't thread an unconditional jump, but if the block is "almost
149 // empty", we can replace uses of it with uses of the successor and make
151 if (BI->isUnconditional() &&
152 BB != &BB->getParent()->getEntryBlock()) {
153 BasicBlock::iterator BBI = BB->getFirstNonPHI();
154 // Ignore dbg intrinsics.
155 while (isa<DbgInfoIntrinsic>(BBI))
157 // If the terminator is the only non-phi instruction, try to nuke it.
158 if (BBI->isTerminator()) {
159 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
160 // block, we have to make sure it isn't in the LoopHeaders set. We
161 // reinsert afterward if needed.
162 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
163 BasicBlock *Succ = BI->getSuccessor(0);
165 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
167 // If we deleted BB and BB was the header of a loop, then the
168 // successor is now the header of the loop.
172 if (ErasedFromLoopHeaders)
173 LoopHeaders.insert(BB);
178 EverChanged |= Changed;
185 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
186 /// thread across it.
187 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
188 /// Ignore PHI nodes, these will be flattened when duplication happens.
189 BasicBlock::const_iterator I = BB->getFirstNonPHI();
191 // FIXME: THREADING will delete values that are just used to compute the
192 // branch, so they shouldn't count against the duplication cost.
195 // Sum up the cost of each instruction until we get to the terminator. Don't
196 // include the terminator because the copy won't include it.
198 for (; !isa<TerminatorInst>(I); ++I) {
199 // Debugger intrinsics don't incur code size.
200 if (isa<DbgInfoIntrinsic>(I)) continue;
202 // If this is a pointer->pointer bitcast, it is free.
203 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType()))
206 // All other instructions count for at least one unit.
209 // Calls are more expensive. If they are non-intrinsic calls, we model them
210 // as having cost of 4. If they are a non-vector intrinsic, we model them
211 // as having cost of 2 total, and if they are a vector intrinsic, we model
212 // them as having cost 1.
213 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
214 if (!isa<IntrinsicInst>(CI))
216 else if (!isa<VectorType>(CI->getType()))
221 // Threading through a switch statement is particularly profitable. If this
222 // block ends in a switch, decrease its cost to make it more likely to happen.
223 if (isa<SwitchInst>(I))
224 Size = Size > 6 ? Size-6 : 0;
229 /// FindLoopHeaders - We do not want jump threading to turn proper loop
230 /// structures into irreducible loops. Doing this breaks up the loop nesting
231 /// hierarchy and pessimizes later transformations. To prevent this from
232 /// happening, we first have to find the loop headers. Here we approximate this
233 /// by finding targets of backedges in the CFG.
235 /// Note that there definitely are cases when we want to allow threading of
236 /// edges across a loop header. For example, threading a jump from outside the
237 /// loop (the preheader) to an exit block of the loop is definitely profitable.
238 /// It is also almost always profitable to thread backedges from within the loop
239 /// to exit blocks, and is often profitable to thread backedges to other blocks
240 /// within the loop (forming a nested loop). This simple analysis is not rich
241 /// enough to track all of these properties and keep it up-to-date as the CFG
242 /// mutates, so we don't allow any of these transformations.
244 void JumpThreading::FindLoopHeaders(Function &F) {
245 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
246 FindFunctionBackedges(F, Edges);
248 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
249 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
252 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
253 /// if we can infer that the value is a known ConstantInt in any of our
254 /// predecessors. If so, return the known list of value and pred BB in the
255 /// result vector. If a value is known to be undef, it is returned as null.
257 /// This returns true if there were any known values.
260 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){
261 // If V is a constantint, then it is known in all predecessors.
262 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) {
263 ConstantInt *CI = dyn_cast<ConstantInt>(V);
265 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
266 Result.push_back(std::make_pair(CI, *PI));
270 // If V is a non-instruction value, or an instruction in a different block,
271 // then it can't be derived from a PHI.
272 Instruction *I = dyn_cast<Instruction>(V);
273 if (I == 0 || I->getParent() != BB) {
275 // Okay, if this is a live-in value, see if it has a known value at the end
276 // of any of our predecessors.
278 // FIXME: This should be an edge property, not a block end property.
279 /// TODO: Per PR2563, we could infer value range information about a
280 /// predecessor based on its terminator.
283 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
284 // "I" is a non-local compare-with-a-constant instruction. This would be
285 // able to handle value inequalities better, for example if the compare is
286 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
287 // Perhaps getConstantOnEdge should be smart enough to do this?
289 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
290 // If the value is known by LazyValueInfo to be a constant in a
291 // predecessor, use that information to try to thread this block.
292 Constant *PredCst = LVI->getConstantOnEdge(V, *PI, BB);
294 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst)))
297 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), *PI));
300 return !Result.empty();
306 /// If I is a PHI node, then we know the incoming values for any constants.
307 if (PHINode *PN = dyn_cast<PHINode>(I)) {
308 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
309 Value *InVal = PN->getIncomingValue(i);
310 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) {
311 ConstantInt *CI = dyn_cast<ConstantInt>(InVal);
312 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i)));
315 return !Result.empty();
318 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals;
320 // Handle some boolean conditions.
321 if (I->getType()->getPrimitiveSizeInBits() == 1) {
323 // X & false -> false
324 if (I->getOpcode() == Instruction::Or ||
325 I->getOpcode() == Instruction::And) {
326 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
327 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals);
329 if (LHSVals.empty() && RHSVals.empty())
332 ConstantInt *InterestingVal;
333 if (I->getOpcode() == Instruction::Or)
334 InterestingVal = ConstantInt::getTrue(I->getContext());
336 InterestingVal = ConstantInt::getFalse(I->getContext());
338 // Scan for the sentinel.
339 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
340 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0)
341 Result.push_back(LHSVals[i]);
342 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
343 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0)
344 Result.push_back(RHSVals[i]);
345 return !Result.empty();
348 // Handle the NOT form of XOR.
349 if (I->getOpcode() == Instruction::Xor &&
350 isa<ConstantInt>(I->getOperand(1)) &&
351 cast<ConstantInt>(I->getOperand(1))->isOne()) {
352 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result);
356 // Invert the known values.
357 for (unsigned i = 0, e = Result.size(); i != e; ++i)
360 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first));
365 // Handle compare with phi operand, where the PHI is defined in this block.
366 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
367 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
368 if (PN && PN->getParent() == BB) {
369 // We can do this simplification if any comparisons fold to true or false.
371 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
372 BasicBlock *PredBB = PN->getIncomingBlock(i);
373 Value *LHS = PN->getIncomingValue(i);
374 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
376 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
378 if (!LVI || !isa<Constant>(RHS))
381 LazyValueInfo::Tristate
382 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
383 cast<Constant>(RHS), PredBB, BB);
384 if (ResT == LazyValueInfo::Unknown)
386 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
389 if (isa<UndefValue>(Res))
390 Result.push_back(std::make_pair((ConstantInt*)0, PredBB));
391 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res))
392 Result.push_back(std::make_pair(CI, PredBB));
395 return !Result.empty();
399 // If comparing a live-in value against a constant, see if we know the
400 // live-in value on any predecessors.
401 if (LVI && isa<Constant>(Cmp->getOperand(1)) &&
402 Cmp->getType()->isInteger() && // Not vector compare.
403 (!isa<Instruction>(Cmp->getOperand(0)) ||
404 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB)) {
405 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
407 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
408 // If the value is known by LazyValueInfo to be a constant in a
409 // predecessor, use that information to try to thread this block.
410 LazyValueInfo::Tristate
411 Res = LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
413 if (Res == LazyValueInfo::Unknown)
416 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
417 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), *PI));
420 return !Result.empty();
428 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
429 /// in an undefined jump, decide which block is best to revector to.
431 /// Since we can pick an arbitrary destination, we pick the successor with the
432 /// fewest predecessors. This should reduce the in-degree of the others.
434 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
435 TerminatorInst *BBTerm = BB->getTerminator();
436 unsigned MinSucc = 0;
437 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
438 // Compute the successor with the minimum number of predecessors.
439 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
440 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
441 TestBB = BBTerm->getSuccessor(i);
442 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
443 if (NumPreds < MinNumPreds)
450 /// ProcessBlock - If there are any predecessors whose control can be threaded
451 /// through to a successor, transform them now.
452 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
453 // If this block has a single predecessor, and if that pred has a single
454 // successor, merge the blocks. This encourages recursive jump threading
455 // because now the condition in this block can be threaded through
456 // predecessors of our predecessor block.
457 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
458 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
460 // If SinglePred was a loop header, BB becomes one.
461 if (LoopHeaders.erase(SinglePred))
462 LoopHeaders.insert(BB);
464 // Remember if SinglePred was the entry block of the function. If so, we
465 // will need to move BB back to the entry position.
466 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
467 MergeBasicBlockIntoOnlyPred(BB);
469 if (isEntry && BB != &BB->getParent()->getEntryBlock())
470 BB->moveBefore(&BB->getParent()->getEntryBlock());
475 // Look to see if the terminator is a branch of switch, if not we can't thread
478 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
479 // Can't thread an unconditional jump.
480 if (BI->isUnconditional()) return false;
481 Condition = BI->getCondition();
482 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
483 Condition = SI->getCondition();
485 return false; // Must be an invoke.
487 // If the terminator of this block is branching on a constant, simplify the
488 // terminator to an unconditional branch. This can occur due to threading in
490 if (isa<ConstantInt>(Condition)) {
491 DEBUG(dbgs() << " In block '" << BB->getName()
492 << "' folding terminator: " << *BB->getTerminator() << '\n');
494 ConstantFoldTerminator(BB);
498 // If the terminator is branching on an undef, we can pick any of the
499 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
500 if (isa<UndefValue>(Condition)) {
501 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
503 // Fold the branch/switch.
504 TerminatorInst *BBTerm = BB->getTerminator();
505 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
506 if (i == BestSucc) continue;
507 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD);
510 DEBUG(dbgs() << " In block '" << BB->getName()
511 << "' folding undef terminator: " << *BBTerm << '\n');
512 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
513 BBTerm->eraseFromParent();
517 Instruction *CondInst = dyn_cast<Instruction>(Condition);
519 // If the condition is an instruction defined in another block, see if a
520 // predecessor has the same condition:
525 !Condition->hasOneUse() && // Multiple uses.
526 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition.
527 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
528 if (isa<BranchInst>(BB->getTerminator())) {
529 for (; PI != E; ++PI)
530 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
531 if (PBI->isConditional() && PBI->getCondition() == Condition &&
532 ProcessBranchOnDuplicateCond(*PI, BB))
535 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator");
536 for (; PI != E; ++PI)
537 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator()))
538 if (PSI->getCondition() == Condition &&
539 ProcessSwitchOnDuplicateCond(*PI, BB))
544 // All the rest of our checks depend on the condition being an instruction.
546 // FIXME: Unify this with code below.
547 if (LVI && ProcessThreadableEdges(Condition, BB))
553 // See if this is a phi node in the current block.
554 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
555 if (PN->getParent() == BB)
556 return ProcessJumpOnPHI(PN);
558 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
560 (!isa<PHINode>(CondCmp->getOperand(0)) ||
561 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) {
562 // If we have a comparison, loop over the predecessors to see if there is
563 // a condition with a lexically identical value.
564 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
565 for (; PI != E; ++PI)
566 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
567 if (PBI->isConditional() && *PI != BB) {
568 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) {
569 if (CI->getOperand(0) == CondCmp->getOperand(0) &&
570 CI->getOperand(1) == CondCmp->getOperand(1) &&
571 CI->getPredicate() == CondCmp->getPredicate()) {
572 // TODO: Could handle things like (x != 4) --> (x == 17)
573 if (ProcessBranchOnDuplicateCond(*PI, BB))
581 // Check for some cases that are worth simplifying. Right now we want to look
582 // for loads that are used by a switch or by the condition for the branch. If
583 // we see one, check to see if it's partially redundant. If so, insert a PHI
584 // which can then be used to thread the values.
586 // This is particularly important because reg2mem inserts loads and stores all
587 // over the place, and this blocks jump threading if we don't zap them.
588 Value *SimplifyValue = CondInst;
589 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
590 if (isa<Constant>(CondCmp->getOperand(1)))
591 SimplifyValue = CondCmp->getOperand(0);
593 // TODO: There are other places where load PRE would be profitable, such as
594 // more complex comparisons.
595 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
596 if (SimplifyPartiallyRedundantLoad(LI))
600 // Handle a variety of cases where we are branching on something derived from
601 // a PHI node in the current block. If we can prove that any predecessors
602 // compute a predictable value based on a PHI node, thread those predecessors.
604 if (ProcessThreadableEdges(CondInst, BB))
608 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
609 // "(X == 4)" thread through this block.
614 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
615 /// block that jump on exactly the same condition. This means that we almost
616 /// always know the direction of the edge in the DESTBB:
618 /// br COND, DESTBB, BBY
620 /// br COND, BBZ, BBW
622 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
623 /// in DESTBB, we have to thread over it.
624 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
626 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
628 // If both successors of PredBB go to DESTBB, we don't know anything. We can
629 // fold the branch to an unconditional one, which allows other recursive
632 if (PredBI->getSuccessor(1) != BB)
634 else if (PredBI->getSuccessor(0) != BB)
637 DEBUG(dbgs() << " In block '" << PredBB->getName()
638 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
640 ConstantFoldTerminator(PredBB);
644 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
646 // If the dest block has one predecessor, just fix the branch condition to a
647 // constant and fold it.
648 if (BB->getSinglePredecessor()) {
649 DEBUG(dbgs() << " In block '" << BB->getName()
650 << "' folding condition to '" << BranchDir << "': "
651 << *BB->getTerminator() << '\n');
653 Value *OldCond = DestBI->getCondition();
654 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
656 ConstantFoldTerminator(BB);
657 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
662 // Next, figure out which successor we are threading to.
663 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
665 SmallVector<BasicBlock*, 2> Preds;
666 Preds.push_back(PredBB);
668 // Ok, try to thread it!
669 return ThreadEdge(BB, Preds, SuccBB);
672 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
673 /// block that switch on exactly the same condition. This means that we almost
674 /// always know the direction of the edge in the DESTBB:
676 /// switch COND [... DESTBB, BBY ... ]
678 /// switch COND [... BBZ, BBW ]
680 /// Optimizing switches like this is very important, because simplifycfg builds
681 /// switches out of repeated 'if' conditions.
682 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
683 BasicBlock *DestBB) {
684 // Can't thread edge to self.
685 if (PredBB == DestBB)
688 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
689 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
691 // There are a variety of optimizations that we can potentially do on these
692 // blocks: we order them from most to least preferable.
694 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
695 // directly to their destination. This does not introduce *any* code size
696 // growth. Skip debug info first.
697 BasicBlock::iterator BBI = DestBB->begin();
698 while (isa<DbgInfoIntrinsic>(BBI))
701 // FIXME: Thread if it just contains a PHI.
702 if (isa<SwitchInst>(BBI)) {
703 bool MadeChange = false;
704 // Ignore the default edge for now.
705 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
706 ConstantInt *DestVal = DestSI->getCaseValue(i);
707 BasicBlock *DestSucc = DestSI->getSuccessor(i);
709 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
710 // PredSI has an explicit case for it. If so, forward. If it is covered
711 // by the default case, we can't update PredSI.
712 unsigned PredCase = PredSI->findCaseValue(DestVal);
713 if (PredCase == 0) continue;
715 // If PredSI doesn't go to DestBB on this value, then it won't reach the
716 // case on this condition.
717 if (PredSI->getSuccessor(PredCase) != DestBB &&
718 DestSI->getSuccessor(i) != DestBB)
721 // Do not forward this if it already goes to this destination, this would
722 // be an infinite loop.
723 if (PredSI->getSuccessor(PredCase) == DestSucc)
726 // Otherwise, we're safe to make the change. Make sure that the edge from
727 // DestSI to DestSucc is not critical and has no PHI nodes.
728 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
729 DEBUG(dbgs() << "THROUGH: " << *DestSI);
731 // If the destination has PHI nodes, just split the edge for updating
733 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
734 SplitCriticalEdge(DestSI, i, this);
735 DestSucc = DestSI->getSuccessor(i);
737 FoldSingleEntryPHINodes(DestSucc);
738 PredSI->setSuccessor(PredCase, DestSucc);
750 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
751 /// load instruction, eliminate it by replacing it with a PHI node. This is an
752 /// important optimization that encourages jump threading, and needs to be run
753 /// interlaced with other jump threading tasks.
754 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
755 // Don't hack volatile loads.
756 if (LI->isVolatile()) return false;
758 // If the load is defined in a block with exactly one predecessor, it can't be
759 // partially redundant.
760 BasicBlock *LoadBB = LI->getParent();
761 if (LoadBB->getSinglePredecessor())
764 Value *LoadedPtr = LI->getOperand(0);
766 // If the loaded operand is defined in the LoadBB, it can't be available.
767 // TODO: Could do simple PHI translation, that would be fun :)
768 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
769 if (PtrOp->getParent() == LoadBB)
772 // Scan a few instructions up from the load, to see if it is obviously live at
773 // the entry to its block.
774 BasicBlock::iterator BBIt = LI;
776 if (Value *AvailableVal =
777 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
778 // If the value if the load is locally available within the block, just use
779 // it. This frequently occurs for reg2mem'd allocas.
780 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
782 // If the returned value is the load itself, replace with an undef. This can
783 // only happen in dead loops.
784 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
785 LI->replaceAllUsesWith(AvailableVal);
786 LI->eraseFromParent();
790 // Otherwise, if we scanned the whole block and got to the top of the block,
791 // we know the block is locally transparent to the load. If not, something
792 // might clobber its value.
793 if (BBIt != LoadBB->begin())
797 SmallPtrSet<BasicBlock*, 8> PredsScanned;
798 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
799 AvailablePredsTy AvailablePreds;
800 BasicBlock *OneUnavailablePred = 0;
802 // If we got here, the loaded value is transparent through to the start of the
803 // block. Check to see if it is available in any of the predecessor blocks.
804 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
806 BasicBlock *PredBB = *PI;
808 // If we already scanned this predecessor, skip it.
809 if (!PredsScanned.insert(PredBB))
812 // Scan the predecessor to see if the value is available in the pred.
813 BBIt = PredBB->end();
814 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
815 if (!PredAvailable) {
816 OneUnavailablePred = PredBB;
820 // If so, this load is partially redundant. Remember this info so that we
821 // can create a PHI node.
822 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
825 // If the loaded value isn't available in any predecessor, it isn't partially
827 if (AvailablePreds.empty()) return false;
829 // Okay, the loaded value is available in at least one (and maybe all!)
830 // predecessors. If the value is unavailable in more than one unique
831 // predecessor, we want to insert a merge block for those common predecessors.
832 // This ensures that we only have to insert one reload, thus not increasing
834 BasicBlock *UnavailablePred = 0;
836 // If there is exactly one predecessor where the value is unavailable, the
837 // already computed 'OneUnavailablePred' block is it. If it ends in an
838 // unconditional branch, we know that it isn't a critical edge.
839 if (PredsScanned.size() == AvailablePreds.size()+1 &&
840 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
841 UnavailablePred = OneUnavailablePred;
842 } else if (PredsScanned.size() != AvailablePreds.size()) {
843 // Otherwise, we had multiple unavailable predecessors or we had a critical
844 // edge from the one.
845 SmallVector<BasicBlock*, 8> PredsToSplit;
846 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
848 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
849 AvailablePredSet.insert(AvailablePreds[i].first);
851 // Add all the unavailable predecessors to the PredsToSplit list.
852 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
854 if (!AvailablePredSet.count(*PI))
855 PredsToSplit.push_back(*PI);
857 // Split them out to their own block.
859 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
860 "thread-pre-split", this);
863 // If the value isn't available in all predecessors, then there will be
864 // exactly one where it isn't available. Insert a load on that edge and add
865 // it to the AvailablePreds list.
866 if (UnavailablePred) {
867 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
868 "Can't handle critical edge here!");
869 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
871 UnavailablePred->getTerminator());
872 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
875 // Now we know that each predecessor of this block has a value in
876 // AvailablePreds, sort them for efficient access as we're walking the preds.
877 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
879 // Create a PHI node at the start of the block for the PRE'd load value.
880 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
883 // Insert new entries into the PHI for each predecessor. A single block may
884 // have multiple entries here.
885 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
887 AvailablePredsTy::iterator I =
888 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
889 std::make_pair(*PI, (Value*)0));
891 assert(I != AvailablePreds.end() && I->first == *PI &&
892 "Didn't find entry for predecessor!");
894 PN->addIncoming(I->second, I->first);
897 //cerr << "PRE: " << *LI << *PN << "\n";
899 LI->replaceAllUsesWith(PN);
900 LI->eraseFromParent();
905 /// FindMostPopularDest - The specified list contains multiple possible
906 /// threadable destinations. Pick the one that occurs the most frequently in
909 FindMostPopularDest(BasicBlock *BB,
910 const SmallVectorImpl<std::pair<BasicBlock*,
911 BasicBlock*> > &PredToDestList) {
912 assert(!PredToDestList.empty());
914 // Determine popularity. If there are multiple possible destinations, we
915 // explicitly choose to ignore 'undef' destinations. We prefer to thread
916 // blocks with known and real destinations to threading undef. We'll handle
917 // them later if interesting.
918 DenseMap<BasicBlock*, unsigned> DestPopularity;
919 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
920 if (PredToDestList[i].second)
921 DestPopularity[PredToDestList[i].second]++;
923 // Find the most popular dest.
924 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
925 BasicBlock *MostPopularDest = DPI->first;
926 unsigned Popularity = DPI->second;
927 SmallVector<BasicBlock*, 4> SamePopularity;
929 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
930 // If the popularity of this entry isn't higher than the popularity we've
931 // seen so far, ignore it.
932 if (DPI->second < Popularity)
934 else if (DPI->second == Popularity) {
935 // If it is the same as what we've seen so far, keep track of it.
936 SamePopularity.push_back(DPI->first);
938 // If it is more popular, remember it.
939 SamePopularity.clear();
940 MostPopularDest = DPI->first;
941 Popularity = DPI->second;
945 // Okay, now we know the most popular destination. If there is more than
946 // destination, we need to determine one. This is arbitrary, but we need
947 // to make a deterministic decision. Pick the first one that appears in the
949 if (!SamePopularity.empty()) {
950 SamePopularity.push_back(MostPopularDest);
951 TerminatorInst *TI = BB->getTerminator();
952 for (unsigned i = 0; ; ++i) {
953 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
955 if (std::find(SamePopularity.begin(), SamePopularity.end(),
956 TI->getSuccessor(i)) == SamePopularity.end())
959 MostPopularDest = TI->getSuccessor(i);
964 // Okay, we have finally picked the most popular destination.
965 return MostPopularDest;
968 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) {
969 // If threading this would thread across a loop header, don't even try to
971 if (LoopHeaders.count(BB))
974 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues;
975 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues))
977 assert(!PredValues.empty() &&
978 "ComputeValueKnownInPredecessors returned true with no values");
980 DEBUG(dbgs() << "IN BB: " << *BB;
981 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
982 dbgs() << " BB '" << BB->getName() << "': FOUND condition = ";
983 if (PredValues[i].first)
984 dbgs() << *PredValues[i].first;
987 dbgs() << " for pred '" << PredValues[i].second->getName()
991 // Decide what we want to thread through. Convert our list of known values to
992 // a list of known destinations for each pred. This also discards duplicate
993 // predecessors and keeps track of the undefined inputs (which are represented
994 // as a null dest in the PredToDestList).
995 SmallPtrSet<BasicBlock*, 16> SeenPreds;
996 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
998 BasicBlock *OnlyDest = 0;
999 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1001 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1002 BasicBlock *Pred = PredValues[i].second;
1003 if (!SeenPreds.insert(Pred))
1004 continue; // Duplicate predecessor entry.
1006 // If the predecessor ends with an indirect goto, we can't change its
1008 if (isa<IndirectBrInst>(Pred->getTerminator()))
1011 ConstantInt *Val = PredValues[i].first;
1014 if (Val == 0) // Undef.
1016 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1017 DestBB = BI->getSuccessor(Val->isZero());
1019 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
1020 DestBB = SI->getSuccessor(SI->findCaseValue(Val));
1023 // If we have exactly one destination, remember it for efficiency below.
1026 else if (OnlyDest != DestBB)
1027 OnlyDest = MultipleDestSentinel;
1029 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1032 // If all edges were unthreadable, we fail.
1033 if (PredToDestList.empty())
1036 // Determine which is the most common successor. If we have many inputs and
1037 // this block is a switch, we want to start by threading the batch that goes
1038 // to the most popular destination first. If we only know about one
1039 // threadable destination (the common case) we can avoid this.
1040 BasicBlock *MostPopularDest = OnlyDest;
1042 if (MostPopularDest == MultipleDestSentinel)
1043 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1045 // Now that we know what the most popular destination is, factor all
1046 // predecessors that will jump to it into a single predecessor.
1047 SmallVector<BasicBlock*, 16> PredsToFactor;
1048 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1049 if (PredToDestList[i].second == MostPopularDest) {
1050 BasicBlock *Pred = PredToDestList[i].first;
1052 // This predecessor may be a switch or something else that has multiple
1053 // edges to the block. Factor each of these edges by listing them
1054 // according to # occurrences in PredsToFactor.
1055 TerminatorInst *PredTI = Pred->getTerminator();
1056 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1057 if (PredTI->getSuccessor(i) == BB)
1058 PredsToFactor.push_back(Pred);
1061 // If the threadable edges are branching on an undefined value, we get to pick
1062 // the destination that these predecessors should get to.
1063 if (MostPopularDest == 0)
1064 MostPopularDest = BB->getTerminator()->
1065 getSuccessor(GetBestDestForJumpOnUndef(BB));
1067 // Ok, try to thread it!
1068 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1071 /// ProcessJumpOnPHI - We have a conditional branch or switch on a PHI node in
1072 /// the current block. See if there are any simplifications we can do based on
1073 /// inputs to the phi node.
1075 bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) {
1076 BasicBlock *BB = PN->getParent();
1078 // If any of the predecessor blocks end in an unconditional branch, we can
1079 // *duplicate* the jump into that block in order to further encourage jump
1080 // threading and to eliminate cases where we have branch on a phi of an icmp
1081 // (branch on icmp is much better).
1083 // We don't want to do this tranformation for switches, because we don't
1084 // really want to duplicate a switch.
1085 if (isa<SwitchInst>(BB->getTerminator()))
1088 // Look for unconditional branch predecessors.
1089 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1090 BasicBlock *PredBB = PN->getIncomingBlock(i);
1091 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1092 if (PredBr->isUnconditional() &&
1093 // Try to duplicate BB into PredBB.
1094 DuplicateCondBranchOnPHIIntoPred(BB, PredBB))
1102 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1103 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1104 /// NewPred using the entries from OldPred (suitably mapped).
1105 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1106 BasicBlock *OldPred,
1107 BasicBlock *NewPred,
1108 DenseMap<Instruction*, Value*> &ValueMap) {
1109 for (BasicBlock::iterator PNI = PHIBB->begin();
1110 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1111 // Ok, we have a PHI node. Figure out what the incoming value was for the
1113 Value *IV = PN->getIncomingValueForBlock(OldPred);
1115 // Remap the value if necessary.
1116 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1117 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1118 if (I != ValueMap.end())
1122 PN->addIncoming(IV, NewPred);
1126 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1127 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1128 /// across BB. Transform the IR to reflect this change.
1129 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1130 const SmallVectorImpl<BasicBlock*> &PredBBs,
1131 BasicBlock *SuccBB) {
1132 // If threading to the same block as we come from, we would infinite loop.
1134 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1135 << "' - would thread to self!\n");
1139 // If threading this would thread across a loop header, don't thread the edge.
1140 // See the comments above FindLoopHeaders for justifications and caveats.
1141 if (LoopHeaders.count(BB)) {
1142 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1143 << "' to dest BB '" << SuccBB->getName()
1144 << "' - it might create an irreducible loop!\n");
1148 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1149 if (JumpThreadCost > Threshold) {
1150 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1151 << "' - Cost is too high: " << JumpThreadCost << "\n");
1155 // And finally, do it! Start by factoring the predecessors is needed.
1157 if (PredBBs.size() == 1)
1158 PredBB = PredBBs[0];
1160 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1161 << " common predecessors.\n");
1162 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1166 // And finally, do it!
1167 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1168 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1169 << ", across block:\n "
1172 // We are going to have to map operands from the original BB block to the new
1173 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1174 // account for entry from PredBB.
1175 DenseMap<Instruction*, Value*> ValueMapping;
1177 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1178 BB->getName()+".thread",
1179 BB->getParent(), BB);
1180 NewBB->moveAfter(PredBB);
1182 BasicBlock::iterator BI = BB->begin();
1183 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1184 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1186 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1187 // mapping and using it to remap operands in the cloned instructions.
1188 for (; !isa<TerminatorInst>(BI); ++BI) {
1189 Instruction *New = BI->clone();
1190 New->setName(BI->getName());
1191 NewBB->getInstList().push_back(New);
1192 ValueMapping[BI] = New;
1194 // Remap operands to patch up intra-block references.
1195 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1196 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1197 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1198 if (I != ValueMapping.end())
1199 New->setOperand(i, I->second);
1203 // We didn't copy the terminator from BB over to NewBB, because there is now
1204 // an unconditional jump to SuccBB. Insert the unconditional jump.
1205 BranchInst::Create(SuccBB, NewBB);
1207 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1208 // PHI nodes for NewBB now.
1209 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1211 // If there were values defined in BB that are used outside the block, then we
1212 // now have to update all uses of the value to use either the original value,
1213 // the cloned value, or some PHI derived value. This can require arbitrary
1214 // PHI insertion, of which we are prepared to do, clean these up now.
1215 SSAUpdater SSAUpdate;
1216 SmallVector<Use*, 16> UsesToRename;
1217 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1218 // Scan all uses of this instruction to see if it is used outside of its
1219 // block, and if so, record them in UsesToRename.
1220 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1222 Instruction *User = cast<Instruction>(*UI);
1223 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1224 if (UserPN->getIncomingBlock(UI) == BB)
1226 } else if (User->getParent() == BB)
1229 UsesToRename.push_back(&UI.getUse());
1232 // If there are no uses outside the block, we're done with this instruction.
1233 if (UsesToRename.empty())
1236 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1238 // We found a use of I outside of BB. Rename all uses of I that are outside
1239 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1240 // with the two values we know.
1241 SSAUpdate.Initialize(I);
1242 SSAUpdate.AddAvailableValue(BB, I);
1243 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1245 while (!UsesToRename.empty())
1246 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1247 DEBUG(dbgs() << "\n");
1251 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1252 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1253 // us to simplify any PHI nodes in BB.
1254 TerminatorInst *PredTerm = PredBB->getTerminator();
1255 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1256 if (PredTerm->getSuccessor(i) == BB) {
1257 RemovePredecessorAndSimplify(BB, PredBB, TD);
1258 PredTerm->setSuccessor(i, NewBB);
1261 // At this point, the IR is fully up to date and consistent. Do a quick scan
1262 // over the new instructions and zap any that are constants or dead. This
1263 // frequently happens because of phi translation.
1264 BI = NewBB->begin();
1265 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) {
1266 Instruction *Inst = BI++;
1268 if (Value *V = SimplifyInstruction(Inst, TD)) {
1269 WeakVH BIHandle(BI);
1270 ReplaceAndSimplifyAllUses(Inst, V, TD);
1272 BI = NewBB->begin();
1276 RecursivelyDeleteTriviallyDeadInstructions(Inst);
1279 // Threaded an edge!
1284 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1285 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1286 /// If we can duplicate the contents of BB up into PredBB do so now, this
1287 /// improves the odds that the branch will be on an analyzable instruction like
1289 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1290 BasicBlock *PredBB) {
1291 // If BB is a loop header, then duplicating this block outside the loop would
1292 // cause us to transform this into an irreducible loop, don't do this.
1293 // See the comments above FindLoopHeaders for justifications and caveats.
1294 if (LoopHeaders.count(BB)) {
1295 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1296 << "' into predecessor block '" << PredBB->getName()
1297 << "' - it might create an irreducible loop!\n");
1301 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1302 if (DuplicationCost > Threshold) {
1303 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1304 << "' - Cost is too high: " << DuplicationCost << "\n");
1308 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1310 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1311 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1312 << DuplicationCost << " block is:" << *BB << "\n");
1314 // We are going to have to map operands from the original BB block into the
1315 // PredBB block. Evaluate PHI nodes in BB.
1316 DenseMap<Instruction*, Value*> ValueMapping;
1318 BasicBlock::iterator BI = BB->begin();
1319 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1320 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1322 BranchInst *OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1324 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1325 // mapping and using it to remap operands in the cloned instructions.
1326 for (; BI != BB->end(); ++BI) {
1327 Instruction *New = BI->clone();
1328 New->setName(BI->getName());
1329 PredBB->getInstList().insert(OldPredBranch, New);
1330 ValueMapping[BI] = New;
1332 // Remap operands to patch up intra-block references.
1333 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1334 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1335 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1336 if (I != ValueMapping.end())
1337 New->setOperand(i, I->second);
1341 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1342 // add entries to the PHI nodes for branch from PredBB now.
1343 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1344 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1346 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1349 // If there were values defined in BB that are used outside the block, then we
1350 // now have to update all uses of the value to use either the original value,
1351 // the cloned value, or some PHI derived value. This can require arbitrary
1352 // PHI insertion, of which we are prepared to do, clean these up now.
1353 SSAUpdater SSAUpdate;
1354 SmallVector<Use*, 16> UsesToRename;
1355 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1356 // Scan all uses of this instruction to see if it is used outside of its
1357 // block, and if so, record them in UsesToRename.
1358 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1360 Instruction *User = cast<Instruction>(*UI);
1361 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1362 if (UserPN->getIncomingBlock(UI) == BB)
1364 } else if (User->getParent() == BB)
1367 UsesToRename.push_back(&UI.getUse());
1370 // If there are no uses outside the block, we're done with this instruction.
1371 if (UsesToRename.empty())
1374 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1376 // We found a use of I outside of BB. Rename all uses of I that are outside
1377 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1378 // with the two values we know.
1379 SSAUpdate.Initialize(I);
1380 SSAUpdate.AddAvailableValue(BB, I);
1381 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1383 while (!UsesToRename.empty())
1384 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1385 DEBUG(dbgs() << "\n");
1388 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1390 RemovePredecessorAndSimplify(BB, PredBB, TD);
1392 // Remove the unconditional branch at the end of the PredBB block.
1393 OldPredBranch->eraseFromParent();