1 //===- JumpThreading.cpp - Thread control through conditional blocks ------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the Jump Threading pass.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "jump-threading"
15 #include "llvm/Transforms/Scalar.h"
16 #include "llvm/IntrinsicInst.h"
17 #include "llvm/LLVMContext.h"
18 #include "llvm/Pass.h"
19 #include "llvm/Analysis/InstructionSimplify.h"
20 #include "llvm/Analysis/LazyValueInfo.h"
21 #include "llvm/Analysis/Loads.h"
22 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
23 #include "llvm/Transforms/Utils/Local.h"
24 #include "llvm/Transforms/Utils/SSAUpdater.h"
25 #include "llvm/Target/TargetData.h"
26 #include "llvm/ADT/DenseMap.h"
27 #include "llvm/ADT/Statistic.h"
28 #include "llvm/ADT/STLExtras.h"
29 #include "llvm/ADT/SmallPtrSet.h"
30 #include "llvm/ADT/SmallSet.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ValueHandle.h"
34 #include "llvm/Support/raw_ostream.h"
37 STATISTIC(NumThreads, "Number of jumps threaded");
38 STATISTIC(NumFolds, "Number of terminators folded");
39 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
41 static cl::opt<unsigned>
42 Threshold("jump-threading-threshold",
43 cl::desc("Max block size to duplicate for jump threading"),
44 cl::init(6), cl::Hidden);
46 // Turn on use of LazyValueInfo.
48 EnableLVI("enable-jump-threading-lvi", cl::ReallyHidden);
53 /// This pass performs 'jump threading', which looks at blocks that have
54 /// multiple predecessors and multiple successors. If one or more of the
55 /// predecessors of the block can be proven to always jump to one of the
56 /// successors, we forward the edge from the predecessor to the successor by
57 /// duplicating the contents of this block.
59 /// An example of when this can occur is code like this:
66 /// In this case, the unconditional branch at the end of the first if can be
67 /// revectored to the false side of the second if.
69 class JumpThreading : public FunctionPass {
73 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
75 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
78 static char ID; // Pass identification
79 JumpThreading() : FunctionPass(&ID) {}
81 bool runOnFunction(Function &F);
83 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
85 AU.addRequired<LazyValueInfo>();
88 void FindLoopHeaders(Function &F);
89 bool ProcessBlock(BasicBlock *BB);
90 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
92 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
93 const SmallVectorImpl<BasicBlock *> &PredBBs);
95 typedef SmallVectorImpl<std::pair<ConstantInt*,
96 BasicBlock*> > PredValueInfo;
98 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
99 PredValueInfo &Result);
100 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB);
103 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
104 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB);
106 bool ProcessBranchOnPHI(PHINode *PN);
107 bool ProcessBranchOnXOR(BinaryOperator *BO);
109 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
113 char JumpThreading::ID = 0;
114 static RegisterPass<JumpThreading>
115 X("jump-threading", "Jump Threading");
117 // Public interface to the Jump Threading pass
118 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
120 /// runOnFunction - Top level algorithm.
122 bool JumpThreading::runOnFunction(Function &F) {
123 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
124 TD = getAnalysisIfAvailable<TargetData>();
125 LVI = EnableLVI ? &getAnalysis<LazyValueInfo>() : 0;
129 bool Changed, EverChanged = false;
132 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
134 // Thread all of the branches we can over this block.
135 while (ProcessBlock(BB))
140 // If the block is trivially dead, zap it. This eliminates the successor
141 // edges which simplifies the CFG.
142 if (pred_begin(BB) == pred_end(BB) &&
143 BB != &BB->getParent()->getEntryBlock()) {
144 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
145 << "' with terminator: " << *BB->getTerminator() << '\n');
146 LoopHeaders.erase(BB);
149 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
150 // Can't thread an unconditional jump, but if the block is "almost
151 // empty", we can replace uses of it with uses of the successor and make
153 if (BI->isUnconditional() &&
154 BB != &BB->getParent()->getEntryBlock()) {
155 BasicBlock::iterator BBI = BB->getFirstNonPHI();
156 // Ignore dbg intrinsics.
157 while (isa<DbgInfoIntrinsic>(BBI))
159 // If the terminator is the only non-phi instruction, try to nuke it.
160 if (BBI->isTerminator()) {
161 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
162 // block, we have to make sure it isn't in the LoopHeaders set. We
163 // reinsert afterward if needed.
164 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
165 BasicBlock *Succ = BI->getSuccessor(0);
167 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
169 // If we deleted BB and BB was the header of a loop, then the
170 // successor is now the header of the loop.
174 if (ErasedFromLoopHeaders)
175 LoopHeaders.insert(BB);
180 EverChanged |= Changed;
187 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
188 /// thread across it.
189 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) {
190 /// Ignore PHI nodes, these will be flattened when duplication happens.
191 BasicBlock::const_iterator I = BB->getFirstNonPHI();
193 // FIXME: THREADING will delete values that are just used to compute the
194 // branch, so they shouldn't count against the duplication cost.
197 // Sum up the cost of each instruction until we get to the terminator. Don't
198 // include the terminator because the copy won't include it.
200 for (; !isa<TerminatorInst>(I); ++I) {
201 // Debugger intrinsics don't incur code size.
202 if (isa<DbgInfoIntrinsic>(I)) continue;
204 // If this is a pointer->pointer bitcast, it is free.
205 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
208 // All other instructions count for at least one unit.
211 // Calls are more expensive. If they are non-intrinsic calls, we model them
212 // as having cost of 4. If they are a non-vector intrinsic, we model them
213 // as having cost of 2 total, and if they are a vector intrinsic, we model
214 // them as having cost 1.
215 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
216 if (!isa<IntrinsicInst>(CI))
218 else if (!CI->getType()->isVectorTy())
223 // Threading through a switch statement is particularly profitable. If this
224 // block ends in a switch, decrease its cost to make it more likely to happen.
225 if (isa<SwitchInst>(I))
226 Size = Size > 6 ? Size-6 : 0;
231 /// FindLoopHeaders - We do not want jump threading to turn proper loop
232 /// structures into irreducible loops. Doing this breaks up the loop nesting
233 /// hierarchy and pessimizes later transformations. To prevent this from
234 /// happening, we first have to find the loop headers. Here we approximate this
235 /// by finding targets of backedges in the CFG.
237 /// Note that there definitely are cases when we want to allow threading of
238 /// edges across a loop header. For example, threading a jump from outside the
239 /// loop (the preheader) to an exit block of the loop is definitely profitable.
240 /// It is also almost always profitable to thread backedges from within the loop
241 /// to exit blocks, and is often profitable to thread backedges to other blocks
242 /// within the loop (forming a nested loop). This simple analysis is not rich
243 /// enough to track all of these properties and keep it up-to-date as the CFG
244 /// mutates, so we don't allow any of these transformations.
246 void JumpThreading::FindLoopHeaders(Function &F) {
247 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
248 FindFunctionBackedges(F, Edges);
250 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
251 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
254 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
255 /// if we can infer that the value is a known ConstantInt in any of our
256 /// predecessors. If so, return the known list of value and pred BB in the
257 /// result vector. If a value is known to be undef, it is returned as null.
259 /// This returns true if there were any known values.
262 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,PredValueInfo &Result){
263 // If V is a constantint, then it is known in all predecessors.
264 if (isa<ConstantInt>(V) || isa<UndefValue>(V)) {
265 ConstantInt *CI = dyn_cast<ConstantInt>(V);
267 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
268 Result.push_back(std::make_pair(CI, *PI));
272 // If V is a non-instruction value, or an instruction in a different block,
273 // then it can't be derived from a PHI.
274 Instruction *I = dyn_cast<Instruction>(V);
275 if (I == 0 || I->getParent() != BB) {
277 // Okay, if this is a live-in value, see if it has a known value at the end
278 // of any of our predecessors.
280 // FIXME: This should be an edge property, not a block end property.
281 /// TODO: Per PR2563, we could infer value range information about a
282 /// predecessor based on its terminator.
285 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
286 // "I" is a non-local compare-with-a-constant instruction. This would be
287 // able to handle value inequalities better, for example if the compare is
288 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
289 // Perhaps getConstantOnEdge should be smart enough to do this?
291 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
292 // If the value is known by LazyValueInfo to be a constant in a
293 // predecessor, use that information to try to thread this block.
294 Constant *PredCst = LVI->getConstantOnEdge(V, *PI, BB);
296 (!isa<ConstantInt>(PredCst) && !isa<UndefValue>(PredCst)))
299 Result.push_back(std::make_pair(dyn_cast<ConstantInt>(PredCst), *PI));
302 return !Result.empty();
308 /// If I is a PHI node, then we know the incoming values for any constants.
309 if (PHINode *PN = dyn_cast<PHINode>(I)) {
310 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
311 Value *InVal = PN->getIncomingValue(i);
312 if (isa<ConstantInt>(InVal) || isa<UndefValue>(InVal)) {
313 ConstantInt *CI = dyn_cast<ConstantInt>(InVal);
314 Result.push_back(std::make_pair(CI, PN->getIncomingBlock(i)));
317 return !Result.empty();
320 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> LHSVals, RHSVals;
322 // Handle some boolean conditions.
323 if (I->getType()->getPrimitiveSizeInBits() == 1) {
325 // X & false -> false
326 if (I->getOpcode() == Instruction::Or ||
327 I->getOpcode() == Instruction::And) {
328 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals);
329 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals);
331 if (LHSVals.empty() && RHSVals.empty())
334 ConstantInt *InterestingVal;
335 if (I->getOpcode() == Instruction::Or)
336 InterestingVal = ConstantInt::getTrue(I->getContext());
338 InterestingVal = ConstantInt::getFalse(I->getContext());
340 // Scan for the sentinel. If we find an undef, force it to the
341 // interesting value: x|undef -> true and x&undef -> false.
342 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
343 if (LHSVals[i].first == InterestingVal || LHSVals[i].first == 0) {
344 Result.push_back(LHSVals[i]);
345 Result.back().first = InterestingVal;
347 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
348 if (RHSVals[i].first == InterestingVal || RHSVals[i].first == 0) {
349 // If we already inferred a value for this block on the LHS, don't
351 bool HasValue = false;
352 for (unsigned r = 0, e = Result.size(); r != e; ++r)
353 if (Result[r].second == RHSVals[i].second) {
359 Result.push_back(RHSVals[i]);
360 Result.back().first = InterestingVal;
363 return !Result.empty();
366 // Handle the NOT form of XOR.
367 if (I->getOpcode() == Instruction::Xor &&
368 isa<ConstantInt>(I->getOperand(1)) &&
369 cast<ConstantInt>(I->getOperand(1))->isOne()) {
370 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result);
374 // Invert the known values.
375 for (unsigned i = 0, e = Result.size(); i != e; ++i)
378 cast<ConstantInt>(ConstantExpr::getNot(Result[i].first));
383 // Handle compare with phi operand, where the PHI is defined in this block.
384 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
385 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
386 if (PN && PN->getParent() == BB) {
387 // We can do this simplification if any comparisons fold to true or false.
389 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
390 BasicBlock *PredBB = PN->getIncomingBlock(i);
391 Value *LHS = PN->getIncomingValue(i);
392 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
394 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, TD);
396 if (!LVI || !isa<Constant>(RHS))
399 LazyValueInfo::Tristate
400 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
401 cast<Constant>(RHS), PredBB, BB);
402 if (ResT == LazyValueInfo::Unknown)
404 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
407 if (isa<UndefValue>(Res))
408 Result.push_back(std::make_pair((ConstantInt*)0, PredBB));
409 else if (ConstantInt *CI = dyn_cast<ConstantInt>(Res))
410 Result.push_back(std::make_pair(CI, PredBB));
413 return !Result.empty();
417 // If comparing a live-in value against a constant, see if we know the
418 // live-in value on any predecessors.
419 if (LVI && isa<Constant>(Cmp->getOperand(1)) &&
420 Cmp->getType()->isIntegerTy() && // Not vector compare.
421 (!isa<Instruction>(Cmp->getOperand(0)) ||
422 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB)) {
423 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
425 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
426 // If the value is known by LazyValueInfo to be a constant in a
427 // predecessor, use that information to try to thread this block.
428 LazyValueInfo::Tristate
429 Res = LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
431 if (Res == LazyValueInfo::Unknown)
434 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
435 Result.push_back(std::make_pair(cast<ConstantInt>(ResC), *PI));
438 return !Result.empty();
446 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
447 /// in an undefined jump, decide which block is best to revector to.
449 /// Since we can pick an arbitrary destination, we pick the successor with the
450 /// fewest predecessors. This should reduce the in-degree of the others.
452 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
453 TerminatorInst *BBTerm = BB->getTerminator();
454 unsigned MinSucc = 0;
455 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
456 // Compute the successor with the minimum number of predecessors.
457 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
458 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
459 TestBB = BBTerm->getSuccessor(i);
460 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
461 if (NumPreds < MinNumPreds)
468 /// ProcessBlock - If there are any predecessors whose control can be threaded
469 /// through to a successor, transform them now.
470 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
471 // If the block is trivially dead, just return and let the caller nuke it.
472 // This simplifies other transformations.
473 if (pred_begin(BB) == pred_end(BB) &&
474 BB != &BB->getParent()->getEntryBlock())
477 // If this block has a single predecessor, and if that pred has a single
478 // successor, merge the blocks. This encourages recursive jump threading
479 // because now the condition in this block can be threaded through
480 // predecessors of our predecessor block.
481 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
482 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
484 // If SinglePred was a loop header, BB becomes one.
485 if (LoopHeaders.erase(SinglePred))
486 LoopHeaders.insert(BB);
488 // Remember if SinglePred was the entry block of the function. If so, we
489 // will need to move BB back to the entry position.
490 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
491 MergeBasicBlockIntoOnlyPred(BB);
493 if (isEntry && BB != &BB->getParent()->getEntryBlock())
494 BB->moveBefore(&BB->getParent()->getEntryBlock());
499 // Look to see if the terminator is a branch of switch, if not we can't thread
502 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
503 // Can't thread an unconditional jump.
504 if (BI->isUnconditional()) return false;
505 Condition = BI->getCondition();
506 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
507 Condition = SI->getCondition();
509 return false; // Must be an invoke.
511 // If the terminator of this block is branching on a constant, simplify the
512 // terminator to an unconditional branch. This can occur due to threading in
514 if (isa<ConstantInt>(Condition)) {
515 DEBUG(dbgs() << " In block '" << BB->getName()
516 << "' folding terminator: " << *BB->getTerminator() << '\n');
518 ConstantFoldTerminator(BB);
522 // If the terminator is branching on an undef, we can pick any of the
523 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
524 if (isa<UndefValue>(Condition)) {
525 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
527 // Fold the branch/switch.
528 TerminatorInst *BBTerm = BB->getTerminator();
529 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
530 if (i == BestSucc) continue;
531 RemovePredecessorAndSimplify(BBTerm->getSuccessor(i), BB, TD);
534 DEBUG(dbgs() << " In block '" << BB->getName()
535 << "' folding undef terminator: " << *BBTerm << '\n');
536 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
537 BBTerm->eraseFromParent();
541 Instruction *CondInst = dyn_cast<Instruction>(Condition);
543 // If the condition is an instruction defined in another block, see if a
544 // predecessor has the same condition:
549 !Condition->hasOneUse() && // Multiple uses.
550 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition.
551 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
552 if (isa<BranchInst>(BB->getTerminator())) {
553 for (; PI != E; ++PI)
554 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
555 if (PBI->isConditional() && PBI->getCondition() == Condition &&
556 ProcessBranchOnDuplicateCond(*PI, BB))
559 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator");
560 for (; PI != E; ++PI)
561 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator()))
562 if (PSI->getCondition() == Condition &&
563 ProcessSwitchOnDuplicateCond(*PI, BB))
568 // All the rest of our checks depend on the condition being an instruction.
570 // FIXME: Unify this with code below.
571 if (LVI && ProcessThreadableEdges(Condition, BB))
577 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
579 (!isa<PHINode>(CondCmp->getOperand(0)) ||
580 cast<PHINode>(CondCmp->getOperand(0))->getParent() != BB)) {
581 // If we have a comparison, loop over the predecessors to see if there is
582 // a condition with a lexically identical value.
583 pred_iterator PI = pred_begin(BB), E = pred_end(BB);
584 for (; PI != E; ++PI)
585 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator()))
586 if (PBI->isConditional() && *PI != BB) {
587 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) {
588 if (CI->getOperand(0) == CondCmp->getOperand(0) &&
589 CI->getOperand(1) == CondCmp->getOperand(1) &&
590 CI->getPredicate() == CondCmp->getPredicate()) {
591 // TODO: Could handle things like (x != 4) --> (x == 17)
592 if (ProcessBranchOnDuplicateCond(*PI, BB))
600 // Check for some cases that are worth simplifying. Right now we want to look
601 // for loads that are used by a switch or by the condition for the branch. If
602 // we see one, check to see if it's partially redundant. If so, insert a PHI
603 // which can then be used to thread the values.
605 Value *SimplifyValue = CondInst;
606 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
607 if (isa<Constant>(CondCmp->getOperand(1)))
608 SimplifyValue = CondCmp->getOperand(0);
610 // TODO: There are other places where load PRE would be profitable, such as
611 // more complex comparisons.
612 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
613 if (SimplifyPartiallyRedundantLoad(LI))
617 // Handle a variety of cases where we are branching on something derived from
618 // a PHI node in the current block. If we can prove that any predecessors
619 // compute a predictable value based on a PHI node, thread those predecessors.
621 if (ProcessThreadableEdges(CondInst, BB))
624 // If this is an otherwise-unfoldable branch on a phi node in the current
625 // block, see if we can simplify.
626 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
627 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
628 return ProcessBranchOnPHI(PN);
631 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
632 if (CondInst->getOpcode() == Instruction::Xor &&
633 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
634 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
637 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
638 // "(X == 4)", thread through this block.
643 /// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that
644 /// block that jump on exactly the same condition. This means that we almost
645 /// always know the direction of the edge in the DESTBB:
647 /// br COND, DESTBB, BBY
649 /// br COND, BBZ, BBW
651 /// If DESTBB has multiple predecessors, we can't just constant fold the branch
652 /// in DESTBB, we have to thread over it.
653 bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB,
655 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator());
657 // If both successors of PredBB go to DESTBB, we don't know anything. We can
658 // fold the branch to an unconditional one, which allows other recursive
661 if (PredBI->getSuccessor(1) != BB)
663 else if (PredBI->getSuccessor(0) != BB)
666 DEBUG(dbgs() << " In block '" << PredBB->getName()
667 << "' folding terminator: " << *PredBB->getTerminator() << '\n');
669 ConstantFoldTerminator(PredBB);
673 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator());
675 // If the dest block has one predecessor, just fix the branch condition to a
676 // constant and fold it.
677 if (BB->getSinglePredecessor()) {
678 DEBUG(dbgs() << " In block '" << BB->getName()
679 << "' folding condition to '" << BranchDir << "': "
680 << *BB->getTerminator() << '\n');
682 Value *OldCond = DestBI->getCondition();
683 DestBI->setCondition(ConstantInt::get(Type::getInt1Ty(BB->getContext()),
685 // Delete dead instructions before we fold the branch. Folding the branch
686 // can eliminate edges from the CFG which can end up deleting OldCond.
687 RecursivelyDeleteTriviallyDeadInstructions(OldCond);
688 ConstantFoldTerminator(BB);
693 // Next, figure out which successor we are threading to.
694 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir);
696 SmallVector<BasicBlock*, 2> Preds;
697 Preds.push_back(PredBB);
699 // Ok, try to thread it!
700 return ThreadEdge(BB, Preds, SuccBB);
703 /// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that
704 /// block that switch on exactly the same condition. This means that we almost
705 /// always know the direction of the edge in the DESTBB:
707 /// switch COND [... DESTBB, BBY ... ]
709 /// switch COND [... BBZ, BBW ]
711 /// Optimizing switches like this is very important, because simplifycfg builds
712 /// switches out of repeated 'if' conditions.
713 bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB,
714 BasicBlock *DestBB) {
715 // Can't thread edge to self.
716 if (PredBB == DestBB)
719 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator());
720 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator());
722 // There are a variety of optimizations that we can potentially do on these
723 // blocks: we order them from most to least preferable.
725 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB
726 // directly to their destination. This does not introduce *any* code size
727 // growth. Skip debug info first.
728 BasicBlock::iterator BBI = DestBB->begin();
729 while (isa<DbgInfoIntrinsic>(BBI))
732 // FIXME: Thread if it just contains a PHI.
733 if (isa<SwitchInst>(BBI)) {
734 bool MadeChange = false;
735 // Ignore the default edge for now.
736 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) {
737 ConstantInt *DestVal = DestSI->getCaseValue(i);
738 BasicBlock *DestSucc = DestSI->getSuccessor(i);
740 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if
741 // PredSI has an explicit case for it. If so, forward. If it is covered
742 // by the default case, we can't update PredSI.
743 unsigned PredCase = PredSI->findCaseValue(DestVal);
744 if (PredCase == 0) continue;
746 // If PredSI doesn't go to DestBB on this value, then it won't reach the
747 // case on this condition.
748 if (PredSI->getSuccessor(PredCase) != DestBB &&
749 DestSI->getSuccessor(i) != DestBB)
752 // Do not forward this if it already goes to this destination, this would
753 // be an infinite loop.
754 if (PredSI->getSuccessor(PredCase) == DestSucc)
757 // Otherwise, we're safe to make the change. Make sure that the edge from
758 // DestSI to DestSucc is not critical and has no PHI nodes.
759 DEBUG(dbgs() << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI);
760 DEBUG(dbgs() << "THROUGH: " << *DestSI);
762 // If the destination has PHI nodes, just split the edge for updating
764 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){
765 SplitCriticalEdge(DestSI, i, this);
766 DestSucc = DestSI->getSuccessor(i);
768 FoldSingleEntryPHINodes(DestSucc);
769 PredSI->setSuccessor(PredCase, DestSucc);
781 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
782 /// load instruction, eliminate it by replacing it with a PHI node. This is an
783 /// important optimization that encourages jump threading, and needs to be run
784 /// interlaced with other jump threading tasks.
785 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
786 // Don't hack volatile loads.
787 if (LI->isVolatile()) return false;
789 // If the load is defined in a block with exactly one predecessor, it can't be
790 // partially redundant.
791 BasicBlock *LoadBB = LI->getParent();
792 if (LoadBB->getSinglePredecessor())
795 Value *LoadedPtr = LI->getOperand(0);
797 // If the loaded operand is defined in the LoadBB, it can't be available.
798 // TODO: Could do simple PHI translation, that would be fun :)
799 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
800 if (PtrOp->getParent() == LoadBB)
803 // Scan a few instructions up from the load, to see if it is obviously live at
804 // the entry to its block.
805 BasicBlock::iterator BBIt = LI;
807 if (Value *AvailableVal =
808 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
809 // If the value if the load is locally available within the block, just use
810 // it. This frequently occurs for reg2mem'd allocas.
811 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
813 // If the returned value is the load itself, replace with an undef. This can
814 // only happen in dead loops.
815 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
816 LI->replaceAllUsesWith(AvailableVal);
817 LI->eraseFromParent();
821 // Otherwise, if we scanned the whole block and got to the top of the block,
822 // we know the block is locally transparent to the load. If not, something
823 // might clobber its value.
824 if (BBIt != LoadBB->begin())
828 SmallPtrSet<BasicBlock*, 8> PredsScanned;
829 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
830 AvailablePredsTy AvailablePreds;
831 BasicBlock *OneUnavailablePred = 0;
833 // If we got here, the loaded value is transparent through to the start of the
834 // block. Check to see if it is available in any of the predecessor blocks.
835 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
837 BasicBlock *PredBB = *PI;
839 // If we already scanned this predecessor, skip it.
840 if (!PredsScanned.insert(PredBB))
843 // Scan the predecessor to see if the value is available in the pred.
844 BBIt = PredBB->end();
845 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6);
846 if (!PredAvailable) {
847 OneUnavailablePred = PredBB;
851 // If so, this load is partially redundant. Remember this info so that we
852 // can create a PHI node.
853 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
856 // If the loaded value isn't available in any predecessor, it isn't partially
858 if (AvailablePreds.empty()) return false;
860 // Okay, the loaded value is available in at least one (and maybe all!)
861 // predecessors. If the value is unavailable in more than one unique
862 // predecessor, we want to insert a merge block for those common predecessors.
863 // This ensures that we only have to insert one reload, thus not increasing
865 BasicBlock *UnavailablePred = 0;
867 // If there is exactly one predecessor where the value is unavailable, the
868 // already computed 'OneUnavailablePred' block is it. If it ends in an
869 // unconditional branch, we know that it isn't a critical edge.
870 if (PredsScanned.size() == AvailablePreds.size()+1 &&
871 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
872 UnavailablePred = OneUnavailablePred;
873 } else if (PredsScanned.size() != AvailablePreds.size()) {
874 // Otherwise, we had multiple unavailable predecessors or we had a critical
875 // edge from the one.
876 SmallVector<BasicBlock*, 8> PredsToSplit;
877 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
879 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
880 AvailablePredSet.insert(AvailablePreds[i].first);
882 // Add all the unavailable predecessors to the PredsToSplit list.
883 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
885 // If the predecessor is an indirect goto, we can't split the edge.
886 if (isa<IndirectBrInst>((*PI)->getTerminator()))
889 if (!AvailablePredSet.count(*PI))
890 PredsToSplit.push_back(*PI);
893 // Split them out to their own block.
895 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(),
896 "thread-pre-split", this);
899 // If the value isn't available in all predecessors, then there will be
900 // exactly one where it isn't available. Insert a load on that edge and add
901 // it to the AvailablePreds list.
902 if (UnavailablePred) {
903 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
904 "Can't handle critical edge here!");
905 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
907 UnavailablePred->getTerminator());
908 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
911 // Now we know that each predecessor of this block has a value in
912 // AvailablePreds, sort them for efficient access as we're walking the preds.
913 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
915 // Create a PHI node at the start of the block for the PRE'd load value.
916 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin());
919 // Insert new entries into the PHI for each predecessor. A single block may
920 // have multiple entries here.
921 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E;
923 AvailablePredsTy::iterator I =
924 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
925 std::make_pair(*PI, (Value*)0));
927 assert(I != AvailablePreds.end() && I->first == *PI &&
928 "Didn't find entry for predecessor!");
930 PN->addIncoming(I->second, I->first);
933 //cerr << "PRE: " << *LI << *PN << "\n";
935 LI->replaceAllUsesWith(PN);
936 LI->eraseFromParent();
941 /// FindMostPopularDest - The specified list contains multiple possible
942 /// threadable destinations. Pick the one that occurs the most frequently in
945 FindMostPopularDest(BasicBlock *BB,
946 const SmallVectorImpl<std::pair<BasicBlock*,
947 BasicBlock*> > &PredToDestList) {
948 assert(!PredToDestList.empty());
950 // Determine popularity. If there are multiple possible destinations, we
951 // explicitly choose to ignore 'undef' destinations. We prefer to thread
952 // blocks with known and real destinations to threading undef. We'll handle
953 // them later if interesting.
954 DenseMap<BasicBlock*, unsigned> DestPopularity;
955 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
956 if (PredToDestList[i].second)
957 DestPopularity[PredToDestList[i].second]++;
959 // Find the most popular dest.
960 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
961 BasicBlock *MostPopularDest = DPI->first;
962 unsigned Popularity = DPI->second;
963 SmallVector<BasicBlock*, 4> SamePopularity;
965 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
966 // If the popularity of this entry isn't higher than the popularity we've
967 // seen so far, ignore it.
968 if (DPI->second < Popularity)
970 else if (DPI->second == Popularity) {
971 // If it is the same as what we've seen so far, keep track of it.
972 SamePopularity.push_back(DPI->first);
974 // If it is more popular, remember it.
975 SamePopularity.clear();
976 MostPopularDest = DPI->first;
977 Popularity = DPI->second;
981 // Okay, now we know the most popular destination. If there is more than
982 // destination, we need to determine one. This is arbitrary, but we need
983 // to make a deterministic decision. Pick the first one that appears in the
985 if (!SamePopularity.empty()) {
986 SamePopularity.push_back(MostPopularDest);
987 TerminatorInst *TI = BB->getTerminator();
988 for (unsigned i = 0; ; ++i) {
989 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
991 if (std::find(SamePopularity.begin(), SamePopularity.end(),
992 TI->getSuccessor(i)) == SamePopularity.end())
995 MostPopularDest = TI->getSuccessor(i);
1000 // Okay, we have finally picked the most popular destination.
1001 return MostPopularDest;
1004 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB) {
1005 // If threading this would thread across a loop header, don't even try to
1007 if (LoopHeaders.count(BB))
1010 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> PredValues;
1011 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues))
1013 assert(!PredValues.empty() &&
1014 "ComputeValueKnownInPredecessors returned true with no values");
1016 DEBUG(dbgs() << "IN BB: " << *BB;
1017 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1018 dbgs() << " BB '" << BB->getName() << "': FOUND condition = ";
1019 if (PredValues[i].first)
1020 dbgs() << *PredValues[i].first;
1023 dbgs() << " for pred '" << PredValues[i].second->getName()
1027 // Decide what we want to thread through. Convert our list of known values to
1028 // a list of known destinations for each pred. This also discards duplicate
1029 // predecessors and keeps track of the undefined inputs (which are represented
1030 // as a null dest in the PredToDestList).
1031 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1032 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1034 BasicBlock *OnlyDest = 0;
1035 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1037 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1038 BasicBlock *Pred = PredValues[i].second;
1039 if (!SeenPreds.insert(Pred))
1040 continue; // Duplicate predecessor entry.
1042 // If the predecessor ends with an indirect goto, we can't change its
1044 if (isa<IndirectBrInst>(Pred->getTerminator()))
1047 ConstantInt *Val = PredValues[i].first;
1050 if (Val == 0) // Undef.
1052 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1053 DestBB = BI->getSuccessor(Val->isZero());
1055 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator());
1056 DestBB = SI->getSuccessor(SI->findCaseValue(Val));
1059 // If we have exactly one destination, remember it for efficiency below.
1062 else if (OnlyDest != DestBB)
1063 OnlyDest = MultipleDestSentinel;
1065 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1068 // If all edges were unthreadable, we fail.
1069 if (PredToDestList.empty())
1072 // Determine which is the most common successor. If we have many inputs and
1073 // this block is a switch, we want to start by threading the batch that goes
1074 // to the most popular destination first. If we only know about one
1075 // threadable destination (the common case) we can avoid this.
1076 BasicBlock *MostPopularDest = OnlyDest;
1078 if (MostPopularDest == MultipleDestSentinel)
1079 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1081 // Now that we know what the most popular destination is, factor all
1082 // predecessors that will jump to it into a single predecessor.
1083 SmallVector<BasicBlock*, 16> PredsToFactor;
1084 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1085 if (PredToDestList[i].second == MostPopularDest) {
1086 BasicBlock *Pred = PredToDestList[i].first;
1088 // This predecessor may be a switch or something else that has multiple
1089 // edges to the block. Factor each of these edges by listing them
1090 // according to # occurrences in PredsToFactor.
1091 TerminatorInst *PredTI = Pred->getTerminator();
1092 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1093 if (PredTI->getSuccessor(i) == BB)
1094 PredsToFactor.push_back(Pred);
1097 // If the threadable edges are branching on an undefined value, we get to pick
1098 // the destination that these predecessors should get to.
1099 if (MostPopularDest == 0)
1100 MostPopularDest = BB->getTerminator()->
1101 getSuccessor(GetBestDestForJumpOnUndef(BB));
1103 // Ok, try to thread it!
1104 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1107 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1108 /// a PHI node in the current block. See if there are any simplifications we
1109 /// can do based on inputs to the phi node.
1111 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1112 BasicBlock *BB = PN->getParent();
1114 // TODO: We could make use of this to do it once for blocks with common PHI
1116 SmallVector<BasicBlock*, 1> PredBBs;
1119 // If any of the predecessor blocks end in an unconditional branch, we can
1120 // *duplicate* the conditional branch into that block in order to further
1121 // encourage jump threading and to eliminate cases where we have branch on a
1122 // phi of an icmp (branch on icmp is much better).
1123 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1124 BasicBlock *PredBB = PN->getIncomingBlock(i);
1125 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1126 if (PredBr->isUnconditional()) {
1127 PredBBs[0] = PredBB;
1128 // Try to duplicate BB into PredBB.
1129 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1137 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1138 /// a xor instruction in the current block. See if there are any
1139 /// simplifications we can do based on inputs to the xor.
1141 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1142 BasicBlock *BB = BO->getParent();
1144 // If either the LHS or RHS of the xor is a constant, don't do this
1146 if (isa<ConstantInt>(BO->getOperand(0)) ||
1147 isa<ConstantInt>(BO->getOperand(1)))
1150 // If the first instruction in BB isn't a phi, we won't be able to infer
1151 // anything special about any particular predecessor.
1152 if (!isa<PHINode>(BB->front()))
1155 // If we have a xor as the branch input to this block, and we know that the
1156 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1157 // the condition into the predecessor and fix that value to true, saving some
1158 // logical ops on that path and encouraging other paths to simplify.
1160 // This copies something like this:
1163 // %X = phi i1 [1], [%X']
1164 // %Y = icmp eq i32 %A, %B
1165 // %Z = xor i1 %X, %Y
1170 // %Y = icmp ne i32 %A, %B
1173 SmallVector<std::pair<ConstantInt*, BasicBlock*>, 8> XorOpValues;
1175 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues)) {
1176 assert(XorOpValues.empty());
1177 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues))
1182 assert(!XorOpValues.empty() &&
1183 "ComputeValueKnownInPredecessors returned true with no values");
1185 // Scan the information to see which is most popular: true or false. The
1186 // predecessors can be of the set true, false, or undef.
1187 unsigned NumTrue = 0, NumFalse = 0;
1188 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1189 if (!XorOpValues[i].first) continue; // Ignore undefs for the count.
1190 if (XorOpValues[i].first->isZero())
1196 // Determine which value to split on, true, false, or undef if neither.
1197 ConstantInt *SplitVal = 0;
1198 if (NumTrue > NumFalse)
1199 SplitVal = ConstantInt::getTrue(BB->getContext());
1200 else if (NumTrue != 0 || NumFalse != 0)
1201 SplitVal = ConstantInt::getFalse(BB->getContext());
1203 // Collect all of the blocks that this can be folded into so that we can
1204 // factor this once and clone it once.
1205 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1206 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1207 if (XorOpValues[i].first != SplitVal && XorOpValues[i].first != 0) continue;
1209 BlocksToFoldInto.push_back(XorOpValues[i].second);
1212 // If we inferred a value for all of the predecessors, then duplication won't
1213 // help us. However, we can just replace the LHS or RHS with the constant.
1214 if (BlocksToFoldInto.size() ==
1215 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1216 if (SplitVal == 0) {
1217 // If all preds provide undef, just nuke the xor, because it is undef too.
1218 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1219 BO->eraseFromParent();
1220 } else if (SplitVal->isZero()) {
1221 // If all preds provide 0, replace the xor with the other input.
1222 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1223 BO->eraseFromParent();
1225 // If all preds provide 1, set the computed value to 1.
1226 BO->setOperand(!isLHS, SplitVal);
1232 // Try to duplicate BB into PredBB.
1233 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1237 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1238 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1239 /// NewPred using the entries from OldPred (suitably mapped).
1240 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1241 BasicBlock *OldPred,
1242 BasicBlock *NewPred,
1243 DenseMap<Instruction*, Value*> &ValueMap) {
1244 for (BasicBlock::iterator PNI = PHIBB->begin();
1245 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1246 // Ok, we have a PHI node. Figure out what the incoming value was for the
1248 Value *IV = PN->getIncomingValueForBlock(OldPred);
1250 // Remap the value if necessary.
1251 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1252 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1253 if (I != ValueMap.end())
1257 PN->addIncoming(IV, NewPred);
1261 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1262 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1263 /// across BB. Transform the IR to reflect this change.
1264 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1265 const SmallVectorImpl<BasicBlock*> &PredBBs,
1266 BasicBlock *SuccBB) {
1267 // If threading to the same block as we come from, we would infinite loop.
1269 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1270 << "' - would thread to self!\n");
1274 // If threading this would thread across a loop header, don't thread the edge.
1275 // See the comments above FindLoopHeaders for justifications and caveats.
1276 if (LoopHeaders.count(BB)) {
1277 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1278 << "' to dest BB '" << SuccBB->getName()
1279 << "' - it might create an irreducible loop!\n");
1283 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB);
1284 if (JumpThreadCost > Threshold) {
1285 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1286 << "' - Cost is too high: " << JumpThreadCost << "\n");
1290 // And finally, do it! Start by factoring the predecessors is needed.
1292 if (PredBBs.size() == 1)
1293 PredBB = PredBBs[0];
1295 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1296 << " common predecessors.\n");
1297 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1301 // And finally, do it!
1302 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1303 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1304 << ", across block:\n "
1307 // We are going to have to map operands from the original BB block to the new
1308 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1309 // account for entry from PredBB.
1310 DenseMap<Instruction*, Value*> ValueMapping;
1312 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1313 BB->getName()+".thread",
1314 BB->getParent(), BB);
1315 NewBB->moveAfter(PredBB);
1317 BasicBlock::iterator BI = BB->begin();
1318 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1319 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1321 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1322 // mapping and using it to remap operands in the cloned instructions.
1323 for (; !isa<TerminatorInst>(BI); ++BI) {
1324 Instruction *New = BI->clone();
1325 New->setName(BI->getName());
1326 NewBB->getInstList().push_back(New);
1327 ValueMapping[BI] = New;
1329 // Remap operands to patch up intra-block references.
1330 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1331 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1332 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1333 if (I != ValueMapping.end())
1334 New->setOperand(i, I->second);
1338 // We didn't copy the terminator from BB over to NewBB, because there is now
1339 // an unconditional jump to SuccBB. Insert the unconditional jump.
1340 BranchInst::Create(SuccBB, NewBB);
1342 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1343 // PHI nodes for NewBB now.
1344 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1346 // If there were values defined in BB that are used outside the block, then we
1347 // now have to update all uses of the value to use either the original value,
1348 // the cloned value, or some PHI derived value. This can require arbitrary
1349 // PHI insertion, of which we are prepared to do, clean these up now.
1350 SSAUpdater SSAUpdate;
1351 SmallVector<Use*, 16> UsesToRename;
1352 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1353 // Scan all uses of this instruction to see if it is used outside of its
1354 // block, and if so, record them in UsesToRename.
1355 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1357 Instruction *User = cast<Instruction>(*UI);
1358 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1359 if (UserPN->getIncomingBlock(UI) == BB)
1361 } else if (User->getParent() == BB)
1364 UsesToRename.push_back(&UI.getUse());
1367 // If there are no uses outside the block, we're done with this instruction.
1368 if (UsesToRename.empty())
1371 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1373 // We found a use of I outside of BB. Rename all uses of I that are outside
1374 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1375 // with the two values we know.
1376 SSAUpdate.Initialize(I);
1377 SSAUpdate.AddAvailableValue(BB, I);
1378 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1380 while (!UsesToRename.empty())
1381 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1382 DEBUG(dbgs() << "\n");
1386 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1387 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1388 // us to simplify any PHI nodes in BB.
1389 TerminatorInst *PredTerm = PredBB->getTerminator();
1390 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1391 if (PredTerm->getSuccessor(i) == BB) {
1392 RemovePredecessorAndSimplify(BB, PredBB, TD);
1393 PredTerm->setSuccessor(i, NewBB);
1396 // At this point, the IR is fully up to date and consistent. Do a quick scan
1397 // over the new instructions and zap any that are constants or dead. This
1398 // frequently happens because of phi translation.
1399 SimplifyInstructionsInBlock(NewBB, TD);
1401 // Threaded an edge!
1406 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1407 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1408 /// If we can duplicate the contents of BB up into PredBB do so now, this
1409 /// improves the odds that the branch will be on an analyzable instruction like
1411 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1412 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1413 assert(!PredBBs.empty() && "Can't handle an empty set");
1415 // If BB is a loop header, then duplicating this block outside the loop would
1416 // cause us to transform this into an irreducible loop, don't do this.
1417 // See the comments above FindLoopHeaders for justifications and caveats.
1418 if (LoopHeaders.count(BB)) {
1419 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1420 << "' into predecessor block '" << PredBBs[0]->getName()
1421 << "' - it might create an irreducible loop!\n");
1425 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB);
1426 if (DuplicationCost > Threshold) {
1427 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1428 << "' - Cost is too high: " << DuplicationCost << "\n");
1432 // And finally, do it! Start by factoring the predecessors is needed.
1434 if (PredBBs.size() == 1)
1435 PredBB = PredBBs[0];
1437 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1438 << " common predecessors.\n");
1439 PredBB = SplitBlockPredecessors(BB, &PredBBs[0], PredBBs.size(),
1443 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1445 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1446 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1447 << DuplicationCost << " block is:" << *BB << "\n");
1449 // Unless PredBB ends with an unconditional branch, split the edge so that we
1450 // can just clone the bits from BB into the end of the new PredBB.
1451 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1453 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1454 PredBB = SplitEdge(PredBB, BB, this);
1455 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1458 // We are going to have to map operands from the original BB block into the
1459 // PredBB block. Evaluate PHI nodes in BB.
1460 DenseMap<Instruction*, Value*> ValueMapping;
1462 BasicBlock::iterator BI = BB->begin();
1463 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1464 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1466 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1467 // mapping and using it to remap operands in the cloned instructions.
1468 for (; BI != BB->end(); ++BI) {
1469 Instruction *New = BI->clone();
1471 // Remap operands to patch up intra-block references.
1472 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1473 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1474 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1475 if (I != ValueMapping.end())
1476 New->setOperand(i, I->second);
1479 // If this instruction can be simplified after the operands are updated,
1480 // just use the simplified value instead. This frequently happens due to
1482 if (Value *IV = SimplifyInstruction(New, TD)) {
1484 ValueMapping[BI] = IV;
1486 // Otherwise, insert the new instruction into the block.
1487 New->setName(BI->getName());
1488 PredBB->getInstList().insert(OldPredBranch, New);
1489 ValueMapping[BI] = New;
1493 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1494 // add entries to the PHI nodes for branch from PredBB now.
1495 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1496 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1498 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1501 // If there were values defined in BB that are used outside the block, then we
1502 // now have to update all uses of the value to use either the original value,
1503 // the cloned value, or some PHI derived value. This can require arbitrary
1504 // PHI insertion, of which we are prepared to do, clean these up now.
1505 SSAUpdater SSAUpdate;
1506 SmallVector<Use*, 16> UsesToRename;
1507 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1508 // Scan all uses of this instruction to see if it is used outside of its
1509 // block, and if so, record them in UsesToRename.
1510 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
1512 Instruction *User = cast<Instruction>(*UI);
1513 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1514 if (UserPN->getIncomingBlock(UI) == BB)
1516 } else if (User->getParent() == BB)
1519 UsesToRename.push_back(&UI.getUse());
1522 // If there are no uses outside the block, we're done with this instruction.
1523 if (UsesToRename.empty())
1526 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1528 // We found a use of I outside of BB. Rename all uses of I that are outside
1529 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1530 // with the two values we know.
1531 SSAUpdate.Initialize(I);
1532 SSAUpdate.AddAvailableValue(BB, I);
1533 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1535 while (!UsesToRename.empty())
1536 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1537 DEBUG(dbgs() << "\n");
1540 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1542 RemovePredecessorAndSimplify(BB, PredBB, TD);
1544 // Remove the unconditional branch at the end of the PredBB block.
1545 OldPredBranch->eraseFromParent();