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 #include "llvm/Transforms/Scalar.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/DenseSet.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/CFG.h"
22 #include "llvm/Analysis/ConstantFolding.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/LazyValueInfo.h"
25 #include "llvm/Analysis/Loads.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/IR/ValueHandle.h"
30 #include "llvm/Pass.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetLibraryInfo.h"
35 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
36 #include "llvm/Transforms/Utils/Local.h"
37 #include "llvm/Transforms/Utils/SSAUpdater.h"
40 #define DEBUG_TYPE "jump-threading"
42 STATISTIC(NumThreads, "Number of jumps threaded");
43 STATISTIC(NumFolds, "Number of terminators folded");
44 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
46 static cl::opt<unsigned>
47 Threshold("jump-threading-threshold",
48 cl::desc("Max block size to duplicate for jump threading"),
49 cl::init(6), cl::Hidden);
52 // These are at global scope so static functions can use them too.
53 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo;
54 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy;
56 // This is used to keep track of what kind of constant we're currently hoping
58 enum ConstantPreference {
63 /// This pass performs 'jump threading', which looks at blocks that have
64 /// multiple predecessors and multiple successors. If one or more of the
65 /// predecessors of the block can be proven to always jump to one of the
66 /// successors, we forward the edge from the predecessor to the successor by
67 /// duplicating the contents of this block.
69 /// An example of when this can occur is code like this:
76 /// In this case, the unconditional branch at the end of the first if can be
77 /// revectored to the false side of the second if.
79 class JumpThreading : public FunctionPass {
81 TargetLibraryInfo *TLI;
84 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
86 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
88 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet;
90 // RAII helper for updating the recursion stack.
91 struct RecursionSetRemover {
92 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet;
93 std::pair<Value*, BasicBlock*> ThePair;
95 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S,
96 std::pair<Value*, BasicBlock*> P)
97 : TheSet(S), ThePair(P) { }
99 ~RecursionSetRemover() {
100 TheSet.erase(ThePair);
104 static char ID; // Pass identification
105 JumpThreading() : FunctionPass(ID) {
106 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
109 bool runOnFunction(Function &F) override;
111 void getAnalysisUsage(AnalysisUsage &AU) const override {
112 AU.addRequired<LazyValueInfo>();
113 AU.addPreserved<LazyValueInfo>();
114 AU.addRequired<TargetLibraryInfo>();
117 void FindLoopHeaders(Function &F);
118 bool ProcessBlock(BasicBlock *BB);
119 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
121 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
122 const SmallVectorImpl<BasicBlock *> &PredBBs);
124 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
125 PredValueInfo &Result,
126 ConstantPreference Preference);
127 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
128 ConstantPreference Preference);
130 bool ProcessBranchOnPHI(PHINode *PN);
131 bool ProcessBranchOnXOR(BinaryOperator *BO);
133 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
134 bool TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB);
138 char JumpThreading::ID = 0;
139 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
140 "Jump Threading", false, false)
141 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)
142 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
143 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
144 "Jump Threading", false, false)
146 // Public interface to the Jump Threading pass
147 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
149 /// runOnFunction - Top level algorithm.
151 bool JumpThreading::runOnFunction(Function &F) {
152 if (skipOptnoneFunction(F))
155 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
156 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
157 DL = DLP ? &DLP->getDataLayout() : 0;
158 TLI = &getAnalysis<TargetLibraryInfo>();
159 LVI = &getAnalysis<LazyValueInfo>();
163 bool Changed, EverChanged = false;
166 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
168 // Thread all of the branches we can over this block.
169 while (ProcessBlock(BB))
174 // If the block is trivially dead, zap it. This eliminates the successor
175 // edges which simplifies the CFG.
176 if (pred_begin(BB) == pred_end(BB) &&
177 BB != &BB->getParent()->getEntryBlock()) {
178 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
179 << "' with terminator: " << *BB->getTerminator() << '\n');
180 LoopHeaders.erase(BB);
187 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
189 // Can't thread an unconditional jump, but if the block is "almost
190 // empty", we can replace uses of it with uses of the successor and make
192 if (BI && BI->isUnconditional() &&
193 BB != &BB->getParent()->getEntryBlock() &&
194 // If the terminator is the only non-phi instruction, try to nuke it.
195 BB->getFirstNonPHIOrDbg()->isTerminator()) {
196 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
197 // block, we have to make sure it isn't in the LoopHeaders set. We
198 // reinsert afterward if needed.
199 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
200 BasicBlock *Succ = BI->getSuccessor(0);
202 // FIXME: It is always conservatively correct to drop the info
203 // for a block even if it doesn't get erased. This isn't totally
204 // awesome, but it allows us to use AssertingVH to prevent nasty
205 // dangling pointer issues within LazyValueInfo.
207 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
209 // If we deleted BB and BB was the header of a loop, then the
210 // successor is now the header of the loop.
214 if (ErasedFromLoopHeaders)
215 LoopHeaders.insert(BB);
218 EverChanged |= Changed;
225 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
226 /// thread across it. Stop scanning the block when passing the threshold.
227 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
228 unsigned Threshold) {
229 /// Ignore PHI nodes, these will be flattened when duplication happens.
230 BasicBlock::const_iterator I = BB->getFirstNonPHI();
232 // FIXME: THREADING will delete values that are just used to compute the
233 // branch, so they shouldn't count against the duplication cost.
235 // Sum up the cost of each instruction until we get to the terminator. Don't
236 // include the terminator because the copy won't include it.
238 for (; !isa<TerminatorInst>(I); ++I) {
240 // Stop scanning the block if we've reached the threshold.
241 if (Size > Threshold)
244 // Debugger intrinsics don't incur code size.
245 if (isa<DbgInfoIntrinsic>(I)) continue;
247 // If this is a pointer->pointer bitcast, it is free.
248 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
251 // All other instructions count for at least one unit.
254 // Calls are more expensive. If they are non-intrinsic calls, we model them
255 // as having cost of 4. If they are a non-vector intrinsic, we model them
256 // as having cost of 2 total, and if they are a vector intrinsic, we model
257 // them as having cost 1.
258 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
259 if (CI->cannotDuplicate())
260 // Blocks with NoDuplicate are modelled as having infinite cost, so they
261 // are never duplicated.
263 else if (!isa<IntrinsicInst>(CI))
265 else if (!CI->getType()->isVectorTy())
270 // Threading through a switch statement is particularly profitable. If this
271 // block ends in a switch, decrease its cost to make it more likely to happen.
272 if (isa<SwitchInst>(I))
273 Size = Size > 6 ? Size-6 : 0;
275 // The same holds for indirect branches, but slightly more so.
276 if (isa<IndirectBrInst>(I))
277 Size = Size > 8 ? Size-8 : 0;
282 /// FindLoopHeaders - We do not want jump threading to turn proper loop
283 /// structures into irreducible loops. Doing this breaks up the loop nesting
284 /// hierarchy and pessimizes later transformations. To prevent this from
285 /// happening, we first have to find the loop headers. Here we approximate this
286 /// by finding targets of backedges in the CFG.
288 /// Note that there definitely are cases when we want to allow threading of
289 /// edges across a loop header. For example, threading a jump from outside the
290 /// loop (the preheader) to an exit block of the loop is definitely profitable.
291 /// It is also almost always profitable to thread backedges from within the loop
292 /// to exit blocks, and is often profitable to thread backedges to other blocks
293 /// within the loop (forming a nested loop). This simple analysis is not rich
294 /// enough to track all of these properties and keep it up-to-date as the CFG
295 /// mutates, so we don't allow any of these transformations.
297 void JumpThreading::FindLoopHeaders(Function &F) {
298 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
299 FindFunctionBackedges(F, Edges);
301 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
302 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
305 /// getKnownConstant - Helper method to determine if we can thread over a
306 /// terminator with the given value as its condition, and if so what value to
307 /// use for that. What kind of value this is depends on whether we want an
308 /// integer or a block address, but an undef is always accepted.
309 /// Returns null if Val is null or not an appropriate constant.
310 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
314 // Undef is "known" enough.
315 if (UndefValue *U = dyn_cast<UndefValue>(Val))
318 if (Preference == WantBlockAddress)
319 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
321 return dyn_cast<ConstantInt>(Val);
324 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
325 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
326 /// in any of our predecessors. If so, return the known list of value and pred
327 /// BB in the result vector.
329 /// This returns true if there were any known values.
332 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
333 ConstantPreference Preference) {
334 // This method walks up use-def chains recursively. Because of this, we could
335 // get into an infinite loop going around loops in the use-def chain. To
336 // prevent this, keep track of what (value, block) pairs we've already visited
337 // and terminate the search if we loop back to them
338 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
341 // An RAII help to remove this pair from the recursion set once the recursion
342 // stack pops back out again.
343 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
345 // If V is a constant, then it is known in all predecessors.
346 if (Constant *KC = getKnownConstant(V, Preference)) {
347 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
348 Result.push_back(std::make_pair(KC, *PI));
353 // If V is a non-instruction value, or an instruction in a different block,
354 // then it can't be derived from a PHI.
355 Instruction *I = dyn_cast<Instruction>(V);
356 if (I == 0 || I->getParent() != BB) {
358 // Okay, if this is a live-in value, see if it has a known value at the end
359 // of any of our predecessors.
361 // FIXME: This should be an edge property, not a block end property.
362 /// TODO: Per PR2563, we could infer value range information about a
363 /// predecessor based on its terminator.
365 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
366 // "I" is a non-local compare-with-a-constant instruction. This would be
367 // able to handle value inequalities better, for example if the compare is
368 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
369 // Perhaps getConstantOnEdge should be smart enough to do this?
371 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
373 // If the value is known by LazyValueInfo to be a constant in a
374 // predecessor, use that information to try to thread this block.
375 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
376 if (Constant *KC = getKnownConstant(PredCst, Preference))
377 Result.push_back(std::make_pair(KC, P));
380 return !Result.empty();
383 /// If I is a PHI node, then we know the incoming values for any constants.
384 if (PHINode *PN = dyn_cast<PHINode>(I)) {
385 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
386 Value *InVal = PN->getIncomingValue(i);
387 if (Constant *KC = getKnownConstant(InVal, Preference)) {
388 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
390 Constant *CI = LVI->getConstantOnEdge(InVal,
391 PN->getIncomingBlock(i), BB);
392 if (Constant *KC = getKnownConstant(CI, Preference))
393 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
397 return !Result.empty();
400 PredValueInfoTy LHSVals, RHSVals;
402 // Handle some boolean conditions.
403 if (I->getType()->getPrimitiveSizeInBits() == 1) {
404 assert(Preference == WantInteger && "One-bit non-integer type?");
406 // X & false -> false
407 if (I->getOpcode() == Instruction::Or ||
408 I->getOpcode() == Instruction::And) {
409 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
411 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
414 if (LHSVals.empty() && RHSVals.empty())
417 ConstantInt *InterestingVal;
418 if (I->getOpcode() == Instruction::Or)
419 InterestingVal = ConstantInt::getTrue(I->getContext());
421 InterestingVal = ConstantInt::getFalse(I->getContext());
423 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
425 // Scan for the sentinel. If we find an undef, force it to the
426 // interesting value: x|undef -> true and x&undef -> false.
427 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
428 if (LHSVals[i].first == InterestingVal ||
429 isa<UndefValue>(LHSVals[i].first)) {
430 Result.push_back(LHSVals[i]);
431 Result.back().first = InterestingVal;
432 LHSKnownBBs.insert(LHSVals[i].second);
434 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
435 if (RHSVals[i].first == InterestingVal ||
436 isa<UndefValue>(RHSVals[i].first)) {
437 // If we already inferred a value for this block on the LHS, don't
439 if (!LHSKnownBBs.count(RHSVals[i].second)) {
440 Result.push_back(RHSVals[i]);
441 Result.back().first = InterestingVal;
445 return !Result.empty();
448 // Handle the NOT form of XOR.
449 if (I->getOpcode() == Instruction::Xor &&
450 isa<ConstantInt>(I->getOperand(1)) &&
451 cast<ConstantInt>(I->getOperand(1))->isOne()) {
452 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
457 // Invert the known values.
458 for (unsigned i = 0, e = Result.size(); i != e; ++i)
459 Result[i].first = ConstantExpr::getNot(Result[i].first);
464 // Try to simplify some other binary operator values.
465 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
466 assert(Preference != WantBlockAddress
467 && "A binary operator creating a block address?");
468 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
469 PredValueInfoTy LHSVals;
470 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
473 // Try to use constant folding to simplify the binary operator.
474 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
475 Constant *V = LHSVals[i].first;
476 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
478 if (Constant *KC = getKnownConstant(Folded, WantInteger))
479 Result.push_back(std::make_pair(KC, LHSVals[i].second));
483 return !Result.empty();
486 // Handle compare with phi operand, where the PHI is defined in this block.
487 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
488 assert(Preference == WantInteger && "Compares only produce integers");
489 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
490 if (PN && PN->getParent() == BB) {
491 // We can do this simplification if any comparisons fold to true or false.
493 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
494 BasicBlock *PredBB = PN->getIncomingBlock(i);
495 Value *LHS = PN->getIncomingValue(i);
496 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
498 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL);
500 if (!isa<Constant>(RHS))
503 LazyValueInfo::Tristate
504 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
505 cast<Constant>(RHS), PredBB, BB);
506 if (ResT == LazyValueInfo::Unknown)
508 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
511 if (Constant *KC = getKnownConstant(Res, WantInteger))
512 Result.push_back(std::make_pair(KC, PredBB));
515 return !Result.empty();
519 // If comparing a live-in value against a constant, see if we know the
520 // live-in value on any predecessors.
521 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
522 if (!isa<Instruction>(Cmp->getOperand(0)) ||
523 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
524 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
526 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
528 // If the value is known by LazyValueInfo to be a constant in a
529 // predecessor, use that information to try to thread this block.
530 LazyValueInfo::Tristate Res =
531 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
533 if (Res == LazyValueInfo::Unknown)
536 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
537 Result.push_back(std::make_pair(ResC, P));
540 return !Result.empty();
543 // Try to find a constant value for the LHS of a comparison,
544 // and evaluate it statically if we can.
545 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
546 PredValueInfoTy LHSVals;
547 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
550 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
551 Constant *V = LHSVals[i].first;
552 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
554 if (Constant *KC = getKnownConstant(Folded, WantInteger))
555 Result.push_back(std::make_pair(KC, LHSVals[i].second));
558 return !Result.empty();
563 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
564 // Handle select instructions where at least one operand is a known constant
565 // and we can figure out the condition value for any predecessor block.
566 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
567 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
568 PredValueInfoTy Conds;
569 if ((TrueVal || FalseVal) &&
570 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
572 for (unsigned i = 0, e = Conds.size(); i != e; ++i) {
573 Constant *Cond = Conds[i].first;
575 // Figure out what value to use for the condition.
577 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
579 KnownCond = CI->isOne();
581 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
582 // Either operand will do, so be sure to pick the one that's a known
584 // FIXME: Do this more cleverly if both values are known constants?
585 KnownCond = (TrueVal != 0);
588 // See if the select has a known constant value for this predecessor.
589 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
590 Result.push_back(std::make_pair(Val, Conds[i].second));
593 return !Result.empty();
597 // If all else fails, see if LVI can figure out a constant value for us.
598 Constant *CI = LVI->getConstant(V, BB);
599 if (Constant *KC = getKnownConstant(CI, Preference)) {
600 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
601 Result.push_back(std::make_pair(KC, *PI));
604 return !Result.empty();
609 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
610 /// in an undefined jump, decide which block is best to revector to.
612 /// Since we can pick an arbitrary destination, we pick the successor with the
613 /// fewest predecessors. This should reduce the in-degree of the others.
615 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
616 TerminatorInst *BBTerm = BB->getTerminator();
617 unsigned MinSucc = 0;
618 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
619 // Compute the successor with the minimum number of predecessors.
620 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
621 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
622 TestBB = BBTerm->getSuccessor(i);
623 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
624 if (NumPreds < MinNumPreds) {
626 MinNumPreds = NumPreds;
633 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
634 if (!BB->hasAddressTaken()) return false;
636 // If the block has its address taken, it may be a tree of dead constants
637 // hanging off of it. These shouldn't keep the block alive.
638 BlockAddress *BA = BlockAddress::get(BB);
639 BA->removeDeadConstantUsers();
640 return !BA->use_empty();
643 /// ProcessBlock - If there are any predecessors whose control can be threaded
644 /// through to a successor, transform them now.
645 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
646 // If the block is trivially dead, just return and let the caller nuke it.
647 // This simplifies other transformations.
648 if (pred_begin(BB) == pred_end(BB) &&
649 BB != &BB->getParent()->getEntryBlock())
652 // If this block has a single predecessor, and if that pred has a single
653 // successor, merge the blocks. This encourages recursive jump threading
654 // because now the condition in this block can be threaded through
655 // predecessors of our predecessor block.
656 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
657 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
658 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
659 // If SinglePred was a loop header, BB becomes one.
660 if (LoopHeaders.erase(SinglePred))
661 LoopHeaders.insert(BB);
663 // Remember if SinglePred was the entry block of the function. If so, we
664 // will need to move BB back to the entry position.
665 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
666 LVI->eraseBlock(SinglePred);
667 MergeBasicBlockIntoOnlyPred(BB);
669 if (isEntry && BB != &BB->getParent()->getEntryBlock())
670 BB->moveBefore(&BB->getParent()->getEntryBlock());
675 // What kind of constant we're looking for.
676 ConstantPreference Preference = WantInteger;
678 // Look to see if the terminator is a conditional branch, switch or indirect
679 // branch, if not we can't thread it.
681 Instruction *Terminator = BB->getTerminator();
682 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
683 // Can't thread an unconditional jump.
684 if (BI->isUnconditional()) return false;
685 Condition = BI->getCondition();
686 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
687 Condition = SI->getCondition();
688 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
689 // Can't thread indirect branch with no successors.
690 if (IB->getNumSuccessors() == 0) return false;
691 Condition = IB->getAddress()->stripPointerCasts();
692 Preference = WantBlockAddress;
694 return false; // Must be an invoke.
697 // Run constant folding to see if we can reduce the condition to a simple
699 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
700 Value *SimpleVal = ConstantFoldInstruction(I, DL, TLI);
702 I->replaceAllUsesWith(SimpleVal);
703 I->eraseFromParent();
704 Condition = SimpleVal;
708 // If the terminator is branching on an undef, we can pick any of the
709 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
710 if (isa<UndefValue>(Condition)) {
711 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
713 // Fold the branch/switch.
714 TerminatorInst *BBTerm = BB->getTerminator();
715 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
716 if (i == BestSucc) continue;
717 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
720 DEBUG(dbgs() << " In block '" << BB->getName()
721 << "' folding undef terminator: " << *BBTerm << '\n');
722 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
723 BBTerm->eraseFromParent();
727 // If the terminator of this block is branching on a constant, simplify the
728 // terminator to an unconditional branch. This can occur due to threading in
730 if (getKnownConstant(Condition, Preference)) {
731 DEBUG(dbgs() << " In block '" << BB->getName()
732 << "' folding terminator: " << *BB->getTerminator() << '\n');
734 ConstantFoldTerminator(BB, true);
738 Instruction *CondInst = dyn_cast<Instruction>(Condition);
740 // All the rest of our checks depend on the condition being an instruction.
742 // FIXME: Unify this with code below.
743 if (ProcessThreadableEdges(Condition, BB, Preference))
749 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
750 // For a comparison where the LHS is outside this block, it's possible
751 // that we've branched on it before. Used LVI to see if we can simplify
752 // the branch based on that.
753 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
754 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
755 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
756 if (CondBr && CondConst && CondBr->isConditional() && PI != PE &&
757 (!isa<Instruction>(CondCmp->getOperand(0)) ||
758 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
759 // For predecessor edge, determine if the comparison is true or false
760 // on that edge. If they're all true or all false, we can simplify the
762 // FIXME: We could handle mixed true/false by duplicating code.
763 LazyValueInfo::Tristate Baseline =
764 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
766 if (Baseline != LazyValueInfo::Unknown) {
767 // Check that all remaining incoming values match the first one.
769 LazyValueInfo::Tristate Ret =
770 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
771 CondCmp->getOperand(0), CondConst, *PI, BB);
772 if (Ret != Baseline) break;
775 // If we terminated early, then one of the values didn't match.
777 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
778 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
779 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
780 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
781 CondBr->eraseFromParent();
788 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB))
792 // Check for some cases that are worth simplifying. Right now we want to look
793 // for loads that are used by a switch or by the condition for the branch. If
794 // we see one, check to see if it's partially redundant. If so, insert a PHI
795 // which can then be used to thread the values.
797 Value *SimplifyValue = CondInst;
798 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
799 if (isa<Constant>(CondCmp->getOperand(1)))
800 SimplifyValue = CondCmp->getOperand(0);
802 // TODO: There are other places where load PRE would be profitable, such as
803 // more complex comparisons.
804 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
805 if (SimplifyPartiallyRedundantLoad(LI))
809 // Handle a variety of cases where we are branching on something derived from
810 // a PHI node in the current block. If we can prove that any predecessors
811 // compute a predictable value based on a PHI node, thread those predecessors.
813 if (ProcessThreadableEdges(CondInst, BB, Preference))
816 // If this is an otherwise-unfoldable branch on a phi node in the current
817 // block, see if we can simplify.
818 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
819 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
820 return ProcessBranchOnPHI(PN);
823 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
824 if (CondInst->getOpcode() == Instruction::Xor &&
825 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
826 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
829 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
830 // "(X == 4)", thread through this block.
835 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
836 /// load instruction, eliminate it by replacing it with a PHI node. This is an
837 /// important optimization that encourages jump threading, and needs to be run
838 /// interlaced with other jump threading tasks.
839 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
840 // Don't hack volatile/atomic loads.
841 if (!LI->isSimple()) return false;
843 // If the load is defined in a block with exactly one predecessor, it can't be
844 // partially redundant.
845 BasicBlock *LoadBB = LI->getParent();
846 if (LoadBB->getSinglePredecessor())
849 // If the load is defined in a landing pad, it can't be partially redundant,
850 // because the edges between the invoke and the landing pad cannot have other
851 // instructions between them.
852 if (LoadBB->isLandingPad())
855 Value *LoadedPtr = LI->getOperand(0);
857 // If the loaded operand is defined in the LoadBB, it can't be available.
858 // TODO: Could do simple PHI translation, that would be fun :)
859 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
860 if (PtrOp->getParent() == LoadBB)
863 // Scan a few instructions up from the load, to see if it is obviously live at
864 // the entry to its block.
865 BasicBlock::iterator BBIt = LI;
867 if (Value *AvailableVal =
868 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
869 // If the value if the load is locally available within the block, just use
870 // it. This frequently occurs for reg2mem'd allocas.
871 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
873 // If the returned value is the load itself, replace with an undef. This can
874 // only happen in dead loops.
875 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
876 LI->replaceAllUsesWith(AvailableVal);
877 LI->eraseFromParent();
881 // Otherwise, if we scanned the whole block and got to the top of the block,
882 // we know the block is locally transparent to the load. If not, something
883 // might clobber its value.
884 if (BBIt != LoadBB->begin())
887 // If all of the loads and stores that feed the value have the same TBAA tag,
888 // then we can propagate it onto any newly inserted loads.
889 MDNode *TBAATag = LI->getMetadata(LLVMContext::MD_tbaa);
891 SmallPtrSet<BasicBlock*, 8> PredsScanned;
892 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
893 AvailablePredsTy AvailablePreds;
894 BasicBlock *OneUnavailablePred = 0;
896 // If we got here, the loaded value is transparent through to the start of the
897 // block. Check to see if it is available in any of the predecessor blocks.
898 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
900 BasicBlock *PredBB = *PI;
902 // If we already scanned this predecessor, skip it.
903 if (!PredsScanned.insert(PredBB))
906 // Scan the predecessor to see if the value is available in the pred.
907 BBIt = PredBB->end();
908 MDNode *ThisTBAATag = 0;
909 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6,
911 if (!PredAvailable) {
912 OneUnavailablePred = PredBB;
916 // If tbaa tags disagree or are not present, forget about them.
917 if (TBAATag != ThisTBAATag) TBAATag = 0;
919 // If so, this load is partially redundant. Remember this info so that we
920 // can create a PHI node.
921 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
924 // If the loaded value isn't available in any predecessor, it isn't partially
926 if (AvailablePreds.empty()) return false;
928 // Okay, the loaded value is available in at least one (and maybe all!)
929 // predecessors. If the value is unavailable in more than one unique
930 // predecessor, we want to insert a merge block for those common predecessors.
931 // This ensures that we only have to insert one reload, thus not increasing
933 BasicBlock *UnavailablePred = 0;
935 // If there is exactly one predecessor where the value is unavailable, the
936 // already computed 'OneUnavailablePred' block is it. If it ends in an
937 // unconditional branch, we know that it isn't a critical edge.
938 if (PredsScanned.size() == AvailablePreds.size()+1 &&
939 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
940 UnavailablePred = OneUnavailablePred;
941 } else if (PredsScanned.size() != AvailablePreds.size()) {
942 // Otherwise, we had multiple unavailable predecessors or we had a critical
943 // edge from the one.
944 SmallVector<BasicBlock*, 8> PredsToSplit;
945 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
947 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
948 AvailablePredSet.insert(AvailablePreds[i].first);
950 // Add all the unavailable predecessors to the PredsToSplit list.
951 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
954 // If the predecessor is an indirect goto, we can't split the edge.
955 if (isa<IndirectBrInst>(P->getTerminator()))
958 if (!AvailablePredSet.count(P))
959 PredsToSplit.push_back(P);
962 // Split them out to their own block.
964 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split", this);
967 // If the value isn't available in all predecessors, then there will be
968 // exactly one where it isn't available. Insert a load on that edge and add
969 // it to the AvailablePreds list.
970 if (UnavailablePred) {
971 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
972 "Can't handle critical edge here!");
973 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
975 UnavailablePred->getTerminator());
976 NewVal->setDebugLoc(LI->getDebugLoc());
978 NewVal->setMetadata(LLVMContext::MD_tbaa, TBAATag);
980 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
983 // Now we know that each predecessor of this block has a value in
984 // AvailablePreds, sort them for efficient access as we're walking the preds.
985 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
987 // Create a PHI node at the start of the block for the PRE'd load value.
988 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
989 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
992 PN->setDebugLoc(LI->getDebugLoc());
994 // Insert new entries into the PHI for each predecessor. A single block may
995 // have multiple entries here.
996 for (pred_iterator PI = PB; PI != PE; ++PI) {
998 AvailablePredsTy::iterator I =
999 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1000 std::make_pair(P, (Value*)0));
1002 assert(I != AvailablePreds.end() && I->first == P &&
1003 "Didn't find entry for predecessor!");
1005 PN->addIncoming(I->second, I->first);
1008 //cerr << "PRE: " << *LI << *PN << "\n";
1010 LI->replaceAllUsesWith(PN);
1011 LI->eraseFromParent();
1016 /// FindMostPopularDest - The specified list contains multiple possible
1017 /// threadable destinations. Pick the one that occurs the most frequently in
1020 FindMostPopularDest(BasicBlock *BB,
1021 const SmallVectorImpl<std::pair<BasicBlock*,
1022 BasicBlock*> > &PredToDestList) {
1023 assert(!PredToDestList.empty());
1025 // Determine popularity. If there are multiple possible destinations, we
1026 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1027 // blocks with known and real destinations to threading undef. We'll handle
1028 // them later if interesting.
1029 DenseMap<BasicBlock*, unsigned> DestPopularity;
1030 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1031 if (PredToDestList[i].second)
1032 DestPopularity[PredToDestList[i].second]++;
1034 // Find the most popular dest.
1035 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1036 BasicBlock *MostPopularDest = DPI->first;
1037 unsigned Popularity = DPI->second;
1038 SmallVector<BasicBlock*, 4> SamePopularity;
1040 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1041 // If the popularity of this entry isn't higher than the popularity we've
1042 // seen so far, ignore it.
1043 if (DPI->second < Popularity)
1045 else if (DPI->second == Popularity) {
1046 // If it is the same as what we've seen so far, keep track of it.
1047 SamePopularity.push_back(DPI->first);
1049 // If it is more popular, remember it.
1050 SamePopularity.clear();
1051 MostPopularDest = DPI->first;
1052 Popularity = DPI->second;
1056 // Okay, now we know the most popular destination. If there is more than one
1057 // destination, we need to determine one. This is arbitrary, but we need
1058 // to make a deterministic decision. Pick the first one that appears in the
1060 if (!SamePopularity.empty()) {
1061 SamePopularity.push_back(MostPopularDest);
1062 TerminatorInst *TI = BB->getTerminator();
1063 for (unsigned i = 0; ; ++i) {
1064 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1066 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1067 TI->getSuccessor(i)) == SamePopularity.end())
1070 MostPopularDest = TI->getSuccessor(i);
1075 // Okay, we have finally picked the most popular destination.
1076 return MostPopularDest;
1079 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1080 ConstantPreference Preference) {
1081 // If threading this would thread across a loop header, don't even try to
1083 if (LoopHeaders.count(BB))
1086 PredValueInfoTy PredValues;
1087 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference))
1090 assert(!PredValues.empty() &&
1091 "ComputeValueKnownInPredecessors returned true with no values");
1093 DEBUG(dbgs() << "IN BB: " << *BB;
1094 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1095 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1096 << *PredValues[i].first
1097 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1100 // Decide what we want to thread through. Convert our list of known values to
1101 // a list of known destinations for each pred. This also discards duplicate
1102 // predecessors and keeps track of the undefined inputs (which are represented
1103 // as a null dest in the PredToDestList).
1104 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1105 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1107 BasicBlock *OnlyDest = 0;
1108 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1110 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1111 BasicBlock *Pred = PredValues[i].second;
1112 if (!SeenPreds.insert(Pred))
1113 continue; // Duplicate predecessor entry.
1115 // If the predecessor ends with an indirect goto, we can't change its
1117 if (isa<IndirectBrInst>(Pred->getTerminator()))
1120 Constant *Val = PredValues[i].first;
1123 if (isa<UndefValue>(Val))
1125 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1126 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1127 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1128 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1130 assert(isa<IndirectBrInst>(BB->getTerminator())
1131 && "Unexpected terminator");
1132 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1135 // If we have exactly one destination, remember it for efficiency below.
1136 if (PredToDestList.empty())
1138 else if (OnlyDest != DestBB)
1139 OnlyDest = MultipleDestSentinel;
1141 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1144 // If all edges were unthreadable, we fail.
1145 if (PredToDestList.empty())
1148 // Determine which is the most common successor. If we have many inputs and
1149 // this block is a switch, we want to start by threading the batch that goes
1150 // to the most popular destination first. If we only know about one
1151 // threadable destination (the common case) we can avoid this.
1152 BasicBlock *MostPopularDest = OnlyDest;
1154 if (MostPopularDest == MultipleDestSentinel)
1155 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1157 // Now that we know what the most popular destination is, factor all
1158 // predecessors that will jump to it into a single predecessor.
1159 SmallVector<BasicBlock*, 16> PredsToFactor;
1160 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1161 if (PredToDestList[i].second == MostPopularDest) {
1162 BasicBlock *Pred = PredToDestList[i].first;
1164 // This predecessor may be a switch or something else that has multiple
1165 // edges to the block. Factor each of these edges by listing them
1166 // according to # occurrences in PredsToFactor.
1167 TerminatorInst *PredTI = Pred->getTerminator();
1168 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1169 if (PredTI->getSuccessor(i) == BB)
1170 PredsToFactor.push_back(Pred);
1173 // If the threadable edges are branching on an undefined value, we get to pick
1174 // the destination that these predecessors should get to.
1175 if (MostPopularDest == 0)
1176 MostPopularDest = BB->getTerminator()->
1177 getSuccessor(GetBestDestForJumpOnUndef(BB));
1179 // Ok, try to thread it!
1180 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1183 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1184 /// a PHI node in the current block. See if there are any simplifications we
1185 /// can do based on inputs to the phi node.
1187 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1188 BasicBlock *BB = PN->getParent();
1190 // TODO: We could make use of this to do it once for blocks with common PHI
1192 SmallVector<BasicBlock*, 1> PredBBs;
1195 // If any of the predecessor blocks end in an unconditional branch, we can
1196 // *duplicate* the conditional branch into that block in order to further
1197 // encourage jump threading and to eliminate cases where we have branch on a
1198 // phi of an icmp (branch on icmp is much better).
1199 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1200 BasicBlock *PredBB = PN->getIncomingBlock(i);
1201 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1202 if (PredBr->isUnconditional()) {
1203 PredBBs[0] = PredBB;
1204 // Try to duplicate BB into PredBB.
1205 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1213 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1214 /// a xor instruction in the current block. See if there are any
1215 /// simplifications we can do based on inputs to the xor.
1217 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1218 BasicBlock *BB = BO->getParent();
1220 // If either the LHS or RHS of the xor is a constant, don't do this
1222 if (isa<ConstantInt>(BO->getOperand(0)) ||
1223 isa<ConstantInt>(BO->getOperand(1)))
1226 // If the first instruction in BB isn't a phi, we won't be able to infer
1227 // anything special about any particular predecessor.
1228 if (!isa<PHINode>(BB->front()))
1231 // If we have a xor as the branch input to this block, and we know that the
1232 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1233 // the condition into the predecessor and fix that value to true, saving some
1234 // logical ops on that path and encouraging other paths to simplify.
1236 // This copies something like this:
1239 // %X = phi i1 [1], [%X']
1240 // %Y = icmp eq i32 %A, %B
1241 // %Z = xor i1 %X, %Y
1246 // %Y = icmp ne i32 %A, %B
1249 PredValueInfoTy XorOpValues;
1251 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1253 assert(XorOpValues.empty());
1254 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1260 assert(!XorOpValues.empty() &&
1261 "ComputeValueKnownInPredecessors returned true with no values");
1263 // Scan the information to see which is most popular: true or false. The
1264 // predecessors can be of the set true, false, or undef.
1265 unsigned NumTrue = 0, NumFalse = 0;
1266 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1267 if (isa<UndefValue>(XorOpValues[i].first))
1268 // Ignore undefs for the count.
1270 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1276 // Determine which value to split on, true, false, or undef if neither.
1277 ConstantInt *SplitVal = 0;
1278 if (NumTrue > NumFalse)
1279 SplitVal = ConstantInt::getTrue(BB->getContext());
1280 else if (NumTrue != 0 || NumFalse != 0)
1281 SplitVal = ConstantInt::getFalse(BB->getContext());
1283 // Collect all of the blocks that this can be folded into so that we can
1284 // factor this once and clone it once.
1285 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1286 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1287 if (XorOpValues[i].first != SplitVal &&
1288 !isa<UndefValue>(XorOpValues[i].first))
1291 BlocksToFoldInto.push_back(XorOpValues[i].second);
1294 // If we inferred a value for all of the predecessors, then duplication won't
1295 // help us. However, we can just replace the LHS or RHS with the constant.
1296 if (BlocksToFoldInto.size() ==
1297 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1298 if (SplitVal == 0) {
1299 // If all preds provide undef, just nuke the xor, because it is undef too.
1300 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1301 BO->eraseFromParent();
1302 } else if (SplitVal->isZero()) {
1303 // If all preds provide 0, replace the xor with the other input.
1304 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1305 BO->eraseFromParent();
1307 // If all preds provide 1, set the computed value to 1.
1308 BO->setOperand(!isLHS, SplitVal);
1314 // Try to duplicate BB into PredBB.
1315 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1319 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1320 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1321 /// NewPred using the entries from OldPred (suitably mapped).
1322 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1323 BasicBlock *OldPred,
1324 BasicBlock *NewPred,
1325 DenseMap<Instruction*, Value*> &ValueMap) {
1326 for (BasicBlock::iterator PNI = PHIBB->begin();
1327 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1328 // Ok, we have a PHI node. Figure out what the incoming value was for the
1330 Value *IV = PN->getIncomingValueForBlock(OldPred);
1332 // Remap the value if necessary.
1333 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1334 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1335 if (I != ValueMap.end())
1339 PN->addIncoming(IV, NewPred);
1343 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1344 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1345 /// across BB. Transform the IR to reflect this change.
1346 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1347 const SmallVectorImpl<BasicBlock*> &PredBBs,
1348 BasicBlock *SuccBB) {
1349 // If threading to the same block as we come from, we would infinite loop.
1351 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1352 << "' - would thread to self!\n");
1356 // If threading this would thread across a loop header, don't thread the edge.
1357 // See the comments above FindLoopHeaders for justifications and caveats.
1358 if (LoopHeaders.count(BB)) {
1359 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1360 << "' to dest BB '" << SuccBB->getName()
1361 << "' - it might create an irreducible loop!\n");
1365 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, Threshold);
1366 if (JumpThreadCost > Threshold) {
1367 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1368 << "' - Cost is too high: " << JumpThreadCost << "\n");
1372 // And finally, do it! Start by factoring the predecessors is needed.
1374 if (PredBBs.size() == 1)
1375 PredBB = PredBBs[0];
1377 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1378 << " common predecessors.\n");
1379 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1382 // And finally, do it!
1383 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1384 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1385 << ", across block:\n "
1388 LVI->threadEdge(PredBB, BB, SuccBB);
1390 // We are going to have to map operands from the original BB block to the new
1391 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1392 // account for entry from PredBB.
1393 DenseMap<Instruction*, Value*> ValueMapping;
1395 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1396 BB->getName()+".thread",
1397 BB->getParent(), BB);
1398 NewBB->moveAfter(PredBB);
1400 BasicBlock::iterator BI = BB->begin();
1401 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1402 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1404 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1405 // mapping and using it to remap operands in the cloned instructions.
1406 for (; !isa<TerminatorInst>(BI); ++BI) {
1407 Instruction *New = BI->clone();
1408 New->setName(BI->getName());
1409 NewBB->getInstList().push_back(New);
1410 ValueMapping[BI] = New;
1412 // Remap operands to patch up intra-block references.
1413 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1414 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1415 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1416 if (I != ValueMapping.end())
1417 New->setOperand(i, I->second);
1421 // We didn't copy the terminator from BB over to NewBB, because there is now
1422 // an unconditional jump to SuccBB. Insert the unconditional jump.
1423 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB);
1424 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1426 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1427 // PHI nodes for NewBB now.
1428 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1430 // If there were values defined in BB that are used outside the block, then we
1431 // now have to update all uses of the value to use either the original value,
1432 // the cloned value, or some PHI derived value. This can require arbitrary
1433 // PHI insertion, of which we are prepared to do, clean these up now.
1434 SSAUpdater SSAUpdate;
1435 SmallVector<Use*, 16> UsesToRename;
1436 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1437 // Scan all uses of this instruction to see if it is used outside of its
1438 // block, and if so, record them in UsesToRename.
1439 for (Use &U : I->uses()) {
1440 Instruction *User = cast<Instruction>(U.getUser());
1441 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1442 if (UserPN->getIncomingBlock(U) == BB)
1444 } else if (User->getParent() == BB)
1447 UsesToRename.push_back(&U);
1450 // If there are no uses outside the block, we're done with this instruction.
1451 if (UsesToRename.empty())
1454 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1456 // We found a use of I outside of BB. Rename all uses of I that are outside
1457 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1458 // with the two values we know.
1459 SSAUpdate.Initialize(I->getType(), I->getName());
1460 SSAUpdate.AddAvailableValue(BB, I);
1461 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1463 while (!UsesToRename.empty())
1464 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1465 DEBUG(dbgs() << "\n");
1469 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1470 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1471 // us to simplify any PHI nodes in BB.
1472 TerminatorInst *PredTerm = PredBB->getTerminator();
1473 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1474 if (PredTerm->getSuccessor(i) == BB) {
1475 BB->removePredecessor(PredBB, true);
1476 PredTerm->setSuccessor(i, NewBB);
1479 // At this point, the IR is fully up to date and consistent. Do a quick scan
1480 // over the new instructions and zap any that are constants or dead. This
1481 // frequently happens because of phi translation.
1482 SimplifyInstructionsInBlock(NewBB, DL, TLI);
1484 // Threaded an edge!
1489 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1490 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1491 /// If we can duplicate the contents of BB up into PredBB do so now, this
1492 /// improves the odds that the branch will be on an analyzable instruction like
1494 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1495 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1496 assert(!PredBBs.empty() && "Can't handle an empty set");
1498 // If BB is a loop header, then duplicating this block outside the loop would
1499 // cause us to transform this into an irreducible loop, don't do this.
1500 // See the comments above FindLoopHeaders for justifications and caveats.
1501 if (LoopHeaders.count(BB)) {
1502 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1503 << "' into predecessor block '" << PredBBs[0]->getName()
1504 << "' - it might create an irreducible loop!\n");
1508 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, Threshold);
1509 if (DuplicationCost > Threshold) {
1510 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1511 << "' - Cost is too high: " << DuplicationCost << "\n");
1515 // And finally, do it! Start by factoring the predecessors is needed.
1517 if (PredBBs.size() == 1)
1518 PredBB = PredBBs[0];
1520 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1521 << " common predecessors.\n");
1522 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1525 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1527 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1528 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1529 << DuplicationCost << " block is:" << *BB << "\n");
1531 // Unless PredBB ends with an unconditional branch, split the edge so that we
1532 // can just clone the bits from BB into the end of the new PredBB.
1533 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1535 if (OldPredBranch == 0 || !OldPredBranch->isUnconditional()) {
1536 PredBB = SplitEdge(PredBB, BB, this);
1537 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1540 // We are going to have to map operands from the original BB block into the
1541 // PredBB block. Evaluate PHI nodes in BB.
1542 DenseMap<Instruction*, Value*> ValueMapping;
1544 BasicBlock::iterator BI = BB->begin();
1545 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1546 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1548 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1549 // mapping and using it to remap operands in the cloned instructions.
1550 for (; BI != BB->end(); ++BI) {
1551 Instruction *New = BI->clone();
1553 // Remap operands to patch up intra-block references.
1554 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1555 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1556 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1557 if (I != ValueMapping.end())
1558 New->setOperand(i, I->second);
1561 // If this instruction can be simplified after the operands are updated,
1562 // just use the simplified value instead. This frequently happens due to
1564 if (Value *IV = SimplifyInstruction(New, DL)) {
1566 ValueMapping[BI] = IV;
1568 // Otherwise, insert the new instruction into the block.
1569 New->setName(BI->getName());
1570 PredBB->getInstList().insert(OldPredBranch, New);
1571 ValueMapping[BI] = New;
1575 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1576 // add entries to the PHI nodes for branch from PredBB now.
1577 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1578 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1580 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1583 // If there were values defined in BB that are used outside the block, then we
1584 // now have to update all uses of the value to use either the original value,
1585 // the cloned value, or some PHI derived value. This can require arbitrary
1586 // PHI insertion, of which we are prepared to do, clean these up now.
1587 SSAUpdater SSAUpdate;
1588 SmallVector<Use*, 16> UsesToRename;
1589 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1590 // Scan all uses of this instruction to see if it is used outside of its
1591 // block, and if so, record them in UsesToRename.
1592 for (Use &U : I->uses()) {
1593 Instruction *User = cast<Instruction>(U.getUser());
1594 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1595 if (UserPN->getIncomingBlock(U) == BB)
1597 } else if (User->getParent() == BB)
1600 UsesToRename.push_back(&U);
1603 // If there are no uses outside the block, we're done with this instruction.
1604 if (UsesToRename.empty())
1607 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1609 // We found a use of I outside of BB. Rename all uses of I that are outside
1610 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1611 // with the two values we know.
1612 SSAUpdate.Initialize(I->getType(), I->getName());
1613 SSAUpdate.AddAvailableValue(BB, I);
1614 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1616 while (!UsesToRename.empty())
1617 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1618 DEBUG(dbgs() << "\n");
1621 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1623 BB->removePredecessor(PredBB, true);
1625 // Remove the unconditional branch at the end of the PredBB block.
1626 OldPredBranch->eraseFromParent();
1632 /// TryToUnfoldSelect - Look for blocks of the form
1638 /// %p = phi [%a, %bb] ...
1642 /// And expand the select into a branch structure if one of its arms allows %c
1643 /// to be folded. This later enables threading from bb1 over bb2.
1644 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1645 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1646 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1647 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1649 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1650 CondLHS->getParent() != BB)
1653 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1654 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1655 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1657 // Look if one of the incoming values is a select in the corresponding
1659 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1662 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1663 if (!PredTerm || !PredTerm->isUnconditional())
1666 // Now check if one of the select values would allow us to constant fold the
1667 // terminator in BB. We don't do the transform if both sides fold, those
1668 // cases will be threaded in any case.
1669 LazyValueInfo::Tristate LHSFolds =
1670 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1672 LazyValueInfo::Tristate RHSFolds =
1673 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
1675 if ((LHSFolds != LazyValueInfo::Unknown ||
1676 RHSFolds != LazyValueInfo::Unknown) &&
1677 LHSFolds != RHSFolds) {
1678 // Expand the select.
1687 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
1688 BB->getParent(), BB);
1689 // Move the unconditional branch to NewBB.
1690 PredTerm->removeFromParent();
1691 NewBB->getInstList().insert(NewBB->end(), PredTerm);
1692 // Create a conditional branch and update PHI nodes.
1693 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
1694 CondLHS->setIncomingValue(I, SI->getFalseValue());
1695 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
1696 // The select is now dead.
1697 SI->eraseFromParent();
1699 // Update any other PHI nodes in BB.
1700 for (BasicBlock::iterator BI = BB->begin();
1701 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
1703 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);