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() : nullptr;
158 TLI = &getAnalysis<TargetLibraryInfo>();
159 LVI = &getAnalysis<LazyValueInfo>();
161 // Remove unreachable blocks from function as they may result in infinite loop.
162 removeUnreachableBlocks(F);
166 bool Changed, EverChanged = false;
169 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
171 // Thread all of the branches we can over this block.
172 while (ProcessBlock(BB))
177 // If the block is trivially dead, zap it. This eliminates the successor
178 // edges which simplifies the CFG.
179 if (pred_begin(BB) == pred_end(BB) &&
180 BB != &BB->getParent()->getEntryBlock()) {
181 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
182 << "' with terminator: " << *BB->getTerminator() << '\n');
183 LoopHeaders.erase(BB);
190 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
192 // Can't thread an unconditional jump, but if the block is "almost
193 // empty", we can replace uses of it with uses of the successor and make
195 if (BI && BI->isUnconditional() &&
196 BB != &BB->getParent()->getEntryBlock() &&
197 // If the terminator is the only non-phi instruction, try to nuke it.
198 BB->getFirstNonPHIOrDbg()->isTerminator()) {
199 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
200 // block, we have to make sure it isn't in the LoopHeaders set. We
201 // reinsert afterward if needed.
202 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
203 BasicBlock *Succ = BI->getSuccessor(0);
205 // FIXME: It is always conservatively correct to drop the info
206 // for a block even if it doesn't get erased. This isn't totally
207 // awesome, but it allows us to use AssertingVH to prevent nasty
208 // dangling pointer issues within LazyValueInfo.
210 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
212 // If we deleted BB and BB was the header of a loop, then the
213 // successor is now the header of the loop.
217 if (ErasedFromLoopHeaders)
218 LoopHeaders.insert(BB);
221 EverChanged |= Changed;
228 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
229 /// thread across it. Stop scanning the block when passing the threshold.
230 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
231 unsigned Threshold) {
232 /// Ignore PHI nodes, these will be flattened when duplication happens.
233 BasicBlock::const_iterator I = BB->getFirstNonPHI();
235 // FIXME: THREADING will delete values that are just used to compute the
236 // branch, so they shouldn't count against the duplication cost.
238 // Sum up the cost of each instruction until we get to the terminator. Don't
239 // include the terminator because the copy won't include it.
241 for (; !isa<TerminatorInst>(I); ++I) {
243 // Stop scanning the block if we've reached the threshold.
244 if (Size > Threshold)
247 // Debugger intrinsics don't incur code size.
248 if (isa<DbgInfoIntrinsic>(I)) continue;
250 // If this is a pointer->pointer bitcast, it is free.
251 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
254 // All other instructions count for at least one unit.
257 // Calls are more expensive. If they are non-intrinsic calls, we model them
258 // as having cost of 4. If they are a non-vector intrinsic, we model them
259 // as having cost of 2 total, and if they are a vector intrinsic, we model
260 // them as having cost 1.
261 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
262 if (CI->cannotDuplicate())
263 // Blocks with NoDuplicate are modelled as having infinite cost, so they
264 // are never duplicated.
266 else if (!isa<IntrinsicInst>(CI))
268 else if (!CI->getType()->isVectorTy())
273 // Threading through a switch statement is particularly profitable. If this
274 // block ends in a switch, decrease its cost to make it more likely to happen.
275 if (isa<SwitchInst>(I))
276 Size = Size > 6 ? Size-6 : 0;
278 // The same holds for indirect branches, but slightly more so.
279 if (isa<IndirectBrInst>(I))
280 Size = Size > 8 ? Size-8 : 0;
285 /// FindLoopHeaders - We do not want jump threading to turn proper loop
286 /// structures into irreducible loops. Doing this breaks up the loop nesting
287 /// hierarchy and pessimizes later transformations. To prevent this from
288 /// happening, we first have to find the loop headers. Here we approximate this
289 /// by finding targets of backedges in the CFG.
291 /// Note that there definitely are cases when we want to allow threading of
292 /// edges across a loop header. For example, threading a jump from outside the
293 /// loop (the preheader) to an exit block of the loop is definitely profitable.
294 /// It is also almost always profitable to thread backedges from within the loop
295 /// to exit blocks, and is often profitable to thread backedges to other blocks
296 /// within the loop (forming a nested loop). This simple analysis is not rich
297 /// enough to track all of these properties and keep it up-to-date as the CFG
298 /// mutates, so we don't allow any of these transformations.
300 void JumpThreading::FindLoopHeaders(Function &F) {
301 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
302 FindFunctionBackedges(F, Edges);
304 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
305 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
308 /// getKnownConstant - Helper method to determine if we can thread over a
309 /// terminator with the given value as its condition, and if so what value to
310 /// use for that. What kind of value this is depends on whether we want an
311 /// integer or a block address, but an undef is always accepted.
312 /// Returns null if Val is null or not an appropriate constant.
313 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
317 // Undef is "known" enough.
318 if (UndefValue *U = dyn_cast<UndefValue>(Val))
321 if (Preference == WantBlockAddress)
322 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
324 return dyn_cast<ConstantInt>(Val);
327 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
328 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
329 /// in any of our predecessors. If so, return the known list of value and pred
330 /// BB in the result vector.
332 /// This returns true if there were any known values.
335 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
336 ConstantPreference Preference) {
337 // This method walks up use-def chains recursively. Because of this, we could
338 // get into an infinite loop going around loops in the use-def chain. To
339 // prevent this, keep track of what (value, block) pairs we've already visited
340 // and terminate the search if we loop back to them
341 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
344 // An RAII help to remove this pair from the recursion set once the recursion
345 // stack pops back out again.
346 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
348 // If V is a constant, then it is known in all predecessors.
349 if (Constant *KC = getKnownConstant(V, Preference)) {
350 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
351 Result.push_back(std::make_pair(KC, *PI));
356 // If V is a non-instruction value, or an instruction in a different block,
357 // then it can't be derived from a PHI.
358 Instruction *I = dyn_cast<Instruction>(V);
359 if (!I || I->getParent() != BB) {
361 // Okay, if this is a live-in value, see if it has a known value at the end
362 // of any of our predecessors.
364 // FIXME: This should be an edge property, not a block end property.
365 /// TODO: Per PR2563, we could infer value range information about a
366 /// predecessor based on its terminator.
368 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
369 // "I" is a non-local compare-with-a-constant instruction. This would be
370 // able to handle value inequalities better, for example if the compare is
371 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
372 // Perhaps getConstantOnEdge should be smart enough to do this?
374 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
376 // If the value is known by LazyValueInfo to be a constant in a
377 // predecessor, use that information to try to thread this block.
378 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
379 if (Constant *KC = getKnownConstant(PredCst, Preference))
380 Result.push_back(std::make_pair(KC, P));
383 return !Result.empty();
386 /// If I is a PHI node, then we know the incoming values for any constants.
387 if (PHINode *PN = dyn_cast<PHINode>(I)) {
388 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
389 Value *InVal = PN->getIncomingValue(i);
390 if (Constant *KC = getKnownConstant(InVal, Preference)) {
391 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
393 Constant *CI = LVI->getConstantOnEdge(InVal,
394 PN->getIncomingBlock(i), BB);
395 if (Constant *KC = getKnownConstant(CI, Preference))
396 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
400 return !Result.empty();
403 PredValueInfoTy LHSVals, RHSVals;
405 // Handle some boolean conditions.
406 if (I->getType()->getPrimitiveSizeInBits() == 1) {
407 assert(Preference == WantInteger && "One-bit non-integer type?");
409 // X & false -> false
410 if (I->getOpcode() == Instruction::Or ||
411 I->getOpcode() == Instruction::And) {
412 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
414 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
417 if (LHSVals.empty() && RHSVals.empty())
420 ConstantInt *InterestingVal;
421 if (I->getOpcode() == Instruction::Or)
422 InterestingVal = ConstantInt::getTrue(I->getContext());
424 InterestingVal = ConstantInt::getFalse(I->getContext());
426 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
428 // Scan for the sentinel. If we find an undef, force it to the
429 // interesting value: x|undef -> true and x&undef -> false.
430 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
431 if (LHSVals[i].first == InterestingVal ||
432 isa<UndefValue>(LHSVals[i].first)) {
433 Result.push_back(LHSVals[i]);
434 Result.back().first = InterestingVal;
435 LHSKnownBBs.insert(LHSVals[i].second);
437 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
438 if (RHSVals[i].first == InterestingVal ||
439 isa<UndefValue>(RHSVals[i].first)) {
440 // If we already inferred a value for this block on the LHS, don't
442 if (!LHSKnownBBs.count(RHSVals[i].second)) {
443 Result.push_back(RHSVals[i]);
444 Result.back().first = InterestingVal;
448 return !Result.empty();
451 // Handle the NOT form of XOR.
452 if (I->getOpcode() == Instruction::Xor &&
453 isa<ConstantInt>(I->getOperand(1)) &&
454 cast<ConstantInt>(I->getOperand(1))->isOne()) {
455 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
460 // Invert the known values.
461 for (unsigned i = 0, e = Result.size(); i != e; ++i)
462 Result[i].first = ConstantExpr::getNot(Result[i].first);
467 // Try to simplify some other binary operator values.
468 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
469 assert(Preference != WantBlockAddress
470 && "A binary operator creating a block address?");
471 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
472 PredValueInfoTy LHSVals;
473 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
476 // Try to use constant folding to simplify the binary operator.
477 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
478 Constant *V = LHSVals[i].first;
479 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
481 if (Constant *KC = getKnownConstant(Folded, WantInteger))
482 Result.push_back(std::make_pair(KC, LHSVals[i].second));
486 return !Result.empty();
489 // Handle compare with phi operand, where the PHI is defined in this block.
490 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
491 assert(Preference == WantInteger && "Compares only produce integers");
492 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
493 if (PN && PN->getParent() == BB) {
494 // We can do this simplification if any comparisons fold to true or false.
496 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
497 BasicBlock *PredBB = PN->getIncomingBlock(i);
498 Value *LHS = PN->getIncomingValue(i);
499 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
501 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL);
503 if (!isa<Constant>(RHS))
506 LazyValueInfo::Tristate
507 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
508 cast<Constant>(RHS), PredBB, BB);
509 if (ResT == LazyValueInfo::Unknown)
511 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
514 if (Constant *KC = getKnownConstant(Res, WantInteger))
515 Result.push_back(std::make_pair(KC, PredBB));
518 return !Result.empty();
522 // If comparing a live-in value against a constant, see if we know the
523 // live-in value on any predecessors.
524 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
525 if (!isa<Instruction>(Cmp->getOperand(0)) ||
526 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
527 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
529 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
531 // If the value is known by LazyValueInfo to be a constant in a
532 // predecessor, use that information to try to thread this block.
533 LazyValueInfo::Tristate Res =
534 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
536 if (Res == LazyValueInfo::Unknown)
539 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
540 Result.push_back(std::make_pair(ResC, P));
543 return !Result.empty();
546 // Try to find a constant value for the LHS of a comparison,
547 // and evaluate it statically if we can.
548 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
549 PredValueInfoTy LHSVals;
550 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
553 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
554 Constant *V = LHSVals[i].first;
555 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
557 if (Constant *KC = getKnownConstant(Folded, WantInteger))
558 Result.push_back(std::make_pair(KC, LHSVals[i].second));
561 return !Result.empty();
566 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
567 // Handle select instructions where at least one operand is a known constant
568 // and we can figure out the condition value for any predecessor block.
569 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
570 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
571 PredValueInfoTy Conds;
572 if ((TrueVal || FalseVal) &&
573 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
575 for (unsigned i = 0, e = Conds.size(); i != e; ++i) {
576 Constant *Cond = Conds[i].first;
578 // Figure out what value to use for the condition.
580 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
582 KnownCond = CI->isOne();
584 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
585 // Either operand will do, so be sure to pick the one that's a known
587 // FIXME: Do this more cleverly if both values are known constants?
588 KnownCond = (TrueVal != nullptr);
591 // See if the select has a known constant value for this predecessor.
592 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
593 Result.push_back(std::make_pair(Val, Conds[i].second));
596 return !Result.empty();
600 // If all else fails, see if LVI can figure out a constant value for us.
601 Constant *CI = LVI->getConstant(V, BB);
602 if (Constant *KC = getKnownConstant(CI, Preference)) {
603 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
604 Result.push_back(std::make_pair(KC, *PI));
607 return !Result.empty();
612 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
613 /// in an undefined jump, decide which block is best to revector to.
615 /// Since we can pick an arbitrary destination, we pick the successor with the
616 /// fewest predecessors. This should reduce the in-degree of the others.
618 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
619 TerminatorInst *BBTerm = BB->getTerminator();
620 unsigned MinSucc = 0;
621 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
622 // Compute the successor with the minimum number of predecessors.
623 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
624 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
625 TestBB = BBTerm->getSuccessor(i);
626 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
627 if (NumPreds < MinNumPreds) {
629 MinNumPreds = NumPreds;
636 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
637 if (!BB->hasAddressTaken()) return false;
639 // If the block has its address taken, it may be a tree of dead constants
640 // hanging off of it. These shouldn't keep the block alive.
641 BlockAddress *BA = BlockAddress::get(BB);
642 BA->removeDeadConstantUsers();
643 return !BA->use_empty();
646 /// ProcessBlock - If there are any predecessors whose control can be threaded
647 /// through to a successor, transform them now.
648 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
649 // If the block is trivially dead, just return and let the caller nuke it.
650 // This simplifies other transformations.
651 if (pred_begin(BB) == pred_end(BB) &&
652 BB != &BB->getParent()->getEntryBlock())
655 // If this block has a single predecessor, and if that pred has a single
656 // successor, merge the blocks. This encourages recursive jump threading
657 // because now the condition in this block can be threaded through
658 // predecessors of our predecessor block.
659 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
660 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
661 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
662 // If SinglePred was a loop header, BB becomes one.
663 if (LoopHeaders.erase(SinglePred))
664 LoopHeaders.insert(BB);
666 // Remember if SinglePred was the entry block of the function. If so, we
667 // will need to move BB back to the entry position.
668 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
669 LVI->eraseBlock(SinglePred);
670 MergeBasicBlockIntoOnlyPred(BB);
672 if (isEntry && BB != &BB->getParent()->getEntryBlock())
673 BB->moveBefore(&BB->getParent()->getEntryBlock());
678 // What kind of constant we're looking for.
679 ConstantPreference Preference = WantInteger;
681 // Look to see if the terminator is a conditional branch, switch or indirect
682 // branch, if not we can't thread it.
684 Instruction *Terminator = BB->getTerminator();
685 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
686 // Can't thread an unconditional jump.
687 if (BI->isUnconditional()) return false;
688 Condition = BI->getCondition();
689 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
690 Condition = SI->getCondition();
691 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
692 // Can't thread indirect branch with no successors.
693 if (IB->getNumSuccessors() == 0) return false;
694 Condition = IB->getAddress()->stripPointerCasts();
695 Preference = WantBlockAddress;
697 return false; // Must be an invoke.
700 // Run constant folding to see if we can reduce the condition to a simple
702 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
703 Value *SimpleVal = ConstantFoldInstruction(I, DL, TLI);
705 I->replaceAllUsesWith(SimpleVal);
706 I->eraseFromParent();
707 Condition = SimpleVal;
711 // If the terminator is branching on an undef, we can pick any of the
712 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
713 if (isa<UndefValue>(Condition)) {
714 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
716 // Fold the branch/switch.
717 TerminatorInst *BBTerm = BB->getTerminator();
718 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
719 if (i == BestSucc) continue;
720 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
723 DEBUG(dbgs() << " In block '" << BB->getName()
724 << "' folding undef terminator: " << *BBTerm << '\n');
725 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
726 BBTerm->eraseFromParent();
730 // If the terminator of this block is branching on a constant, simplify the
731 // terminator to an unconditional branch. This can occur due to threading in
733 if (getKnownConstant(Condition, Preference)) {
734 DEBUG(dbgs() << " In block '" << BB->getName()
735 << "' folding terminator: " << *BB->getTerminator() << '\n');
737 ConstantFoldTerminator(BB, true);
741 Instruction *CondInst = dyn_cast<Instruction>(Condition);
743 // All the rest of our checks depend on the condition being an instruction.
745 // FIXME: Unify this with code below.
746 if (ProcessThreadableEdges(Condition, BB, Preference))
752 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
753 // For a comparison where the LHS is outside this block, it's possible
754 // that we've branched on it before. Used LVI to see if we can simplify
755 // the branch based on that.
756 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
757 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
758 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
759 if (CondBr && CondConst && CondBr->isConditional() && PI != PE &&
760 (!isa<Instruction>(CondCmp->getOperand(0)) ||
761 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
762 // For predecessor edge, determine if the comparison is true or false
763 // on that edge. If they're all true or all false, we can simplify the
765 // FIXME: We could handle mixed true/false by duplicating code.
766 LazyValueInfo::Tristate Baseline =
767 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
769 if (Baseline != LazyValueInfo::Unknown) {
770 // Check that all remaining incoming values match the first one.
772 LazyValueInfo::Tristate Ret =
773 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
774 CondCmp->getOperand(0), CondConst, *PI, BB);
775 if (Ret != Baseline) break;
778 // If we terminated early, then one of the values didn't match.
780 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
781 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
782 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
783 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
784 CondBr->eraseFromParent();
791 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB))
795 // Check for some cases that are worth simplifying. Right now we want to look
796 // for loads that are used by a switch or by the condition for the branch. If
797 // we see one, check to see if it's partially redundant. If so, insert a PHI
798 // which can then be used to thread the values.
800 Value *SimplifyValue = CondInst;
801 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
802 if (isa<Constant>(CondCmp->getOperand(1)))
803 SimplifyValue = CondCmp->getOperand(0);
805 // TODO: There are other places where load PRE would be profitable, such as
806 // more complex comparisons.
807 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
808 if (SimplifyPartiallyRedundantLoad(LI))
812 // Handle a variety of cases where we are branching on something derived from
813 // a PHI node in the current block. If we can prove that any predecessors
814 // compute a predictable value based on a PHI node, thread those predecessors.
816 if (ProcessThreadableEdges(CondInst, BB, Preference))
819 // If this is an otherwise-unfoldable branch on a phi node in the current
820 // block, see if we can simplify.
821 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
822 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
823 return ProcessBranchOnPHI(PN);
826 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
827 if (CondInst->getOpcode() == Instruction::Xor &&
828 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
829 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
832 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
833 // "(X == 4)", thread through this block.
838 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
839 /// load instruction, eliminate it by replacing it with a PHI node. This is an
840 /// important optimization that encourages jump threading, and needs to be run
841 /// interlaced with other jump threading tasks.
842 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
843 // Don't hack volatile/atomic loads.
844 if (!LI->isSimple()) return false;
846 // If the load is defined in a block with exactly one predecessor, it can't be
847 // partially redundant.
848 BasicBlock *LoadBB = LI->getParent();
849 if (LoadBB->getSinglePredecessor())
852 // If the load is defined in a landing pad, it can't be partially redundant,
853 // because the edges between the invoke and the landing pad cannot have other
854 // instructions between them.
855 if (LoadBB->isLandingPad())
858 Value *LoadedPtr = LI->getOperand(0);
860 // If the loaded operand is defined in the LoadBB, it can't be available.
861 // TODO: Could do simple PHI translation, that would be fun :)
862 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
863 if (PtrOp->getParent() == LoadBB)
866 // Scan a few instructions up from the load, to see if it is obviously live at
867 // the entry to its block.
868 BasicBlock::iterator BBIt = LI;
870 if (Value *AvailableVal =
871 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
872 // If the value if the load is locally available within the block, just use
873 // it. This frequently occurs for reg2mem'd allocas.
874 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
876 // If the returned value is the load itself, replace with an undef. This can
877 // only happen in dead loops.
878 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
879 LI->replaceAllUsesWith(AvailableVal);
880 LI->eraseFromParent();
884 // Otherwise, if we scanned the whole block and got to the top of the block,
885 // we know the block is locally transparent to the load. If not, something
886 // might clobber its value.
887 if (BBIt != LoadBB->begin())
890 // If all of the loads and stores that feed the value have the same TBAA tag,
891 // then we can propagate it onto any newly inserted loads.
892 MDNode *TBAATag = LI->getMetadata(LLVMContext::MD_tbaa);
894 SmallPtrSet<BasicBlock*, 8> PredsScanned;
895 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
896 AvailablePredsTy AvailablePreds;
897 BasicBlock *OneUnavailablePred = nullptr;
899 // If we got here, the loaded value is transparent through to the start of the
900 // block. Check to see if it is available in any of the predecessor blocks.
901 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
903 BasicBlock *PredBB = *PI;
905 // If we already scanned this predecessor, skip it.
906 if (!PredsScanned.insert(PredBB))
909 // Scan the predecessor to see if the value is available in the pred.
910 BBIt = PredBB->end();
911 MDNode *ThisTBAATag = nullptr;
912 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6,
913 nullptr, &ThisTBAATag);
914 if (!PredAvailable) {
915 OneUnavailablePred = PredBB;
919 // If tbaa tags disagree or are not present, forget about them.
920 if (TBAATag != ThisTBAATag) TBAATag = nullptr;
922 // If so, this load is partially redundant. Remember this info so that we
923 // can create a PHI node.
924 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
927 // If the loaded value isn't available in any predecessor, it isn't partially
929 if (AvailablePreds.empty()) return false;
931 // Okay, the loaded value is available in at least one (and maybe all!)
932 // predecessors. If the value is unavailable in more than one unique
933 // predecessor, we want to insert a merge block for those common predecessors.
934 // This ensures that we only have to insert one reload, thus not increasing
936 BasicBlock *UnavailablePred = nullptr;
938 // If there is exactly one predecessor where the value is unavailable, the
939 // already computed 'OneUnavailablePred' block is it. If it ends in an
940 // unconditional branch, we know that it isn't a critical edge.
941 if (PredsScanned.size() == AvailablePreds.size()+1 &&
942 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
943 UnavailablePred = OneUnavailablePred;
944 } else if (PredsScanned.size() != AvailablePreds.size()) {
945 // Otherwise, we had multiple unavailable predecessors or we had a critical
946 // edge from the one.
947 SmallVector<BasicBlock*, 8> PredsToSplit;
948 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
950 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
951 AvailablePredSet.insert(AvailablePreds[i].first);
953 // Add all the unavailable predecessors to the PredsToSplit list.
954 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
957 // If the predecessor is an indirect goto, we can't split the edge.
958 if (isa<IndirectBrInst>(P->getTerminator()))
961 if (!AvailablePredSet.count(P))
962 PredsToSplit.push_back(P);
965 // Split them out to their own block.
967 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split", this);
970 // If the value isn't available in all predecessors, then there will be
971 // exactly one where it isn't available. Insert a load on that edge and add
972 // it to the AvailablePreds list.
973 if (UnavailablePred) {
974 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
975 "Can't handle critical edge here!");
976 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
978 UnavailablePred->getTerminator());
979 NewVal->setDebugLoc(LI->getDebugLoc());
981 NewVal->setMetadata(LLVMContext::MD_tbaa, TBAATag);
983 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
986 // Now we know that each predecessor of this block has a value in
987 // AvailablePreds, sort them for efficient access as we're walking the preds.
988 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
990 // Create a PHI node at the start of the block for the PRE'd load value.
991 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
992 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
995 PN->setDebugLoc(LI->getDebugLoc());
997 // Insert new entries into the PHI for each predecessor. A single block may
998 // have multiple entries here.
999 for (pred_iterator PI = PB; PI != PE; ++PI) {
1000 BasicBlock *P = *PI;
1001 AvailablePredsTy::iterator I =
1002 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1003 std::make_pair(P, (Value*)nullptr));
1005 assert(I != AvailablePreds.end() && I->first == P &&
1006 "Didn't find entry for predecessor!");
1008 PN->addIncoming(I->second, I->first);
1011 //cerr << "PRE: " << *LI << *PN << "\n";
1013 LI->replaceAllUsesWith(PN);
1014 LI->eraseFromParent();
1019 /// FindMostPopularDest - The specified list contains multiple possible
1020 /// threadable destinations. Pick the one that occurs the most frequently in
1023 FindMostPopularDest(BasicBlock *BB,
1024 const SmallVectorImpl<std::pair<BasicBlock*,
1025 BasicBlock*> > &PredToDestList) {
1026 assert(!PredToDestList.empty());
1028 // Determine popularity. If there are multiple possible destinations, we
1029 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1030 // blocks with known and real destinations to threading undef. We'll handle
1031 // them later if interesting.
1032 DenseMap<BasicBlock*, unsigned> DestPopularity;
1033 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1034 if (PredToDestList[i].second)
1035 DestPopularity[PredToDestList[i].second]++;
1037 // Find the most popular dest.
1038 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1039 BasicBlock *MostPopularDest = DPI->first;
1040 unsigned Popularity = DPI->second;
1041 SmallVector<BasicBlock*, 4> SamePopularity;
1043 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1044 // If the popularity of this entry isn't higher than the popularity we've
1045 // seen so far, ignore it.
1046 if (DPI->second < Popularity)
1048 else if (DPI->second == Popularity) {
1049 // If it is the same as what we've seen so far, keep track of it.
1050 SamePopularity.push_back(DPI->first);
1052 // If it is more popular, remember it.
1053 SamePopularity.clear();
1054 MostPopularDest = DPI->first;
1055 Popularity = DPI->second;
1059 // Okay, now we know the most popular destination. If there is more than one
1060 // destination, we need to determine one. This is arbitrary, but we need
1061 // to make a deterministic decision. Pick the first one that appears in the
1063 if (!SamePopularity.empty()) {
1064 SamePopularity.push_back(MostPopularDest);
1065 TerminatorInst *TI = BB->getTerminator();
1066 for (unsigned i = 0; ; ++i) {
1067 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1069 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1070 TI->getSuccessor(i)) == SamePopularity.end())
1073 MostPopularDest = TI->getSuccessor(i);
1078 // Okay, we have finally picked the most popular destination.
1079 return MostPopularDest;
1082 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1083 ConstantPreference Preference) {
1084 // If threading this would thread across a loop header, don't even try to
1086 if (LoopHeaders.count(BB))
1089 PredValueInfoTy PredValues;
1090 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference))
1093 assert(!PredValues.empty() &&
1094 "ComputeValueKnownInPredecessors returned true with no values");
1096 DEBUG(dbgs() << "IN BB: " << *BB;
1097 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1098 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1099 << *PredValues[i].first
1100 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1103 // Decide what we want to thread through. Convert our list of known values to
1104 // a list of known destinations for each pred. This also discards duplicate
1105 // predecessors and keeps track of the undefined inputs (which are represented
1106 // as a null dest in the PredToDestList).
1107 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1108 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1110 BasicBlock *OnlyDest = nullptr;
1111 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1113 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1114 BasicBlock *Pred = PredValues[i].second;
1115 if (!SeenPreds.insert(Pred))
1116 continue; // Duplicate predecessor entry.
1118 // If the predecessor ends with an indirect goto, we can't change its
1120 if (isa<IndirectBrInst>(Pred->getTerminator()))
1123 Constant *Val = PredValues[i].first;
1126 if (isa<UndefValue>(Val))
1128 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1129 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1130 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1131 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1133 assert(isa<IndirectBrInst>(BB->getTerminator())
1134 && "Unexpected terminator");
1135 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1138 // If we have exactly one destination, remember it for efficiency below.
1139 if (PredToDestList.empty())
1141 else if (OnlyDest != DestBB)
1142 OnlyDest = MultipleDestSentinel;
1144 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1147 // If all edges were unthreadable, we fail.
1148 if (PredToDestList.empty())
1151 // Determine which is the most common successor. If we have many inputs and
1152 // this block is a switch, we want to start by threading the batch that goes
1153 // to the most popular destination first. If we only know about one
1154 // threadable destination (the common case) we can avoid this.
1155 BasicBlock *MostPopularDest = OnlyDest;
1157 if (MostPopularDest == MultipleDestSentinel)
1158 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1160 // Now that we know what the most popular destination is, factor all
1161 // predecessors that will jump to it into a single predecessor.
1162 SmallVector<BasicBlock*, 16> PredsToFactor;
1163 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1164 if (PredToDestList[i].second == MostPopularDest) {
1165 BasicBlock *Pred = PredToDestList[i].first;
1167 // This predecessor may be a switch or something else that has multiple
1168 // edges to the block. Factor each of these edges by listing them
1169 // according to # occurrences in PredsToFactor.
1170 TerminatorInst *PredTI = Pred->getTerminator();
1171 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1172 if (PredTI->getSuccessor(i) == BB)
1173 PredsToFactor.push_back(Pred);
1176 // If the threadable edges are branching on an undefined value, we get to pick
1177 // the destination that these predecessors should get to.
1178 if (!MostPopularDest)
1179 MostPopularDest = BB->getTerminator()->
1180 getSuccessor(GetBestDestForJumpOnUndef(BB));
1182 // Ok, try to thread it!
1183 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1186 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1187 /// a PHI node in the current block. See if there are any simplifications we
1188 /// can do based on inputs to the phi node.
1190 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1191 BasicBlock *BB = PN->getParent();
1193 // TODO: We could make use of this to do it once for blocks with common PHI
1195 SmallVector<BasicBlock*, 1> PredBBs;
1198 // If any of the predecessor blocks end in an unconditional branch, we can
1199 // *duplicate* the conditional branch into that block in order to further
1200 // encourage jump threading and to eliminate cases where we have branch on a
1201 // phi of an icmp (branch on icmp is much better).
1202 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1203 BasicBlock *PredBB = PN->getIncomingBlock(i);
1204 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1205 if (PredBr->isUnconditional()) {
1206 PredBBs[0] = PredBB;
1207 // Try to duplicate BB into PredBB.
1208 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1216 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1217 /// a xor instruction in the current block. See if there are any
1218 /// simplifications we can do based on inputs to the xor.
1220 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1221 BasicBlock *BB = BO->getParent();
1223 // If either the LHS or RHS of the xor is a constant, don't do this
1225 if (isa<ConstantInt>(BO->getOperand(0)) ||
1226 isa<ConstantInt>(BO->getOperand(1)))
1229 // If the first instruction in BB isn't a phi, we won't be able to infer
1230 // anything special about any particular predecessor.
1231 if (!isa<PHINode>(BB->front()))
1234 // If we have a xor as the branch input to this block, and we know that the
1235 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1236 // the condition into the predecessor and fix that value to true, saving some
1237 // logical ops on that path and encouraging other paths to simplify.
1239 // This copies something like this:
1242 // %X = phi i1 [1], [%X']
1243 // %Y = icmp eq i32 %A, %B
1244 // %Z = xor i1 %X, %Y
1249 // %Y = icmp ne i32 %A, %B
1252 PredValueInfoTy XorOpValues;
1254 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1256 assert(XorOpValues.empty());
1257 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1263 assert(!XorOpValues.empty() &&
1264 "ComputeValueKnownInPredecessors returned true with no values");
1266 // Scan the information to see which is most popular: true or false. The
1267 // predecessors can be of the set true, false, or undef.
1268 unsigned NumTrue = 0, NumFalse = 0;
1269 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1270 if (isa<UndefValue>(XorOpValues[i].first))
1271 // Ignore undefs for the count.
1273 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1279 // Determine which value to split on, true, false, or undef if neither.
1280 ConstantInt *SplitVal = nullptr;
1281 if (NumTrue > NumFalse)
1282 SplitVal = ConstantInt::getTrue(BB->getContext());
1283 else if (NumTrue != 0 || NumFalse != 0)
1284 SplitVal = ConstantInt::getFalse(BB->getContext());
1286 // Collect all of the blocks that this can be folded into so that we can
1287 // factor this once and clone it once.
1288 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1289 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1290 if (XorOpValues[i].first != SplitVal &&
1291 !isa<UndefValue>(XorOpValues[i].first))
1294 BlocksToFoldInto.push_back(XorOpValues[i].second);
1297 // If we inferred a value for all of the predecessors, then duplication won't
1298 // help us. However, we can just replace the LHS or RHS with the constant.
1299 if (BlocksToFoldInto.size() ==
1300 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1302 // If all preds provide undef, just nuke the xor, because it is undef too.
1303 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1304 BO->eraseFromParent();
1305 } else if (SplitVal->isZero()) {
1306 // If all preds provide 0, replace the xor with the other input.
1307 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1308 BO->eraseFromParent();
1310 // If all preds provide 1, set the computed value to 1.
1311 BO->setOperand(!isLHS, SplitVal);
1317 // Try to duplicate BB into PredBB.
1318 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1322 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1323 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1324 /// NewPred using the entries from OldPred (suitably mapped).
1325 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1326 BasicBlock *OldPred,
1327 BasicBlock *NewPred,
1328 DenseMap<Instruction*, Value*> &ValueMap) {
1329 for (BasicBlock::iterator PNI = PHIBB->begin();
1330 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1331 // Ok, we have a PHI node. Figure out what the incoming value was for the
1333 Value *IV = PN->getIncomingValueForBlock(OldPred);
1335 // Remap the value if necessary.
1336 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1337 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1338 if (I != ValueMap.end())
1342 PN->addIncoming(IV, NewPred);
1346 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1347 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1348 /// across BB. Transform the IR to reflect this change.
1349 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1350 const SmallVectorImpl<BasicBlock*> &PredBBs,
1351 BasicBlock *SuccBB) {
1352 // If threading to the same block as we come from, we would infinite loop.
1354 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1355 << "' - would thread to self!\n");
1359 // If threading this would thread across a loop header, don't thread the edge.
1360 // See the comments above FindLoopHeaders for justifications and caveats.
1361 if (LoopHeaders.count(BB)) {
1362 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1363 << "' to dest BB '" << SuccBB->getName()
1364 << "' - it might create an irreducible loop!\n");
1368 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, Threshold);
1369 if (JumpThreadCost > Threshold) {
1370 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1371 << "' - Cost is too high: " << JumpThreadCost << "\n");
1375 // And finally, do it! Start by factoring the predecessors is needed.
1377 if (PredBBs.size() == 1)
1378 PredBB = PredBBs[0];
1380 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1381 << " common predecessors.\n");
1382 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1385 // And finally, do it!
1386 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1387 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1388 << ", across block:\n "
1391 LVI->threadEdge(PredBB, BB, SuccBB);
1393 // We are going to have to map operands from the original BB block to the new
1394 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1395 // account for entry from PredBB.
1396 DenseMap<Instruction*, Value*> ValueMapping;
1398 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1399 BB->getName()+".thread",
1400 BB->getParent(), BB);
1401 NewBB->moveAfter(PredBB);
1403 BasicBlock::iterator BI = BB->begin();
1404 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1405 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1407 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1408 // mapping and using it to remap operands in the cloned instructions.
1409 for (; !isa<TerminatorInst>(BI); ++BI) {
1410 Instruction *New = BI->clone();
1411 New->setName(BI->getName());
1412 NewBB->getInstList().push_back(New);
1413 ValueMapping[BI] = New;
1415 // Remap operands to patch up intra-block references.
1416 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1417 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1418 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1419 if (I != ValueMapping.end())
1420 New->setOperand(i, I->second);
1424 // We didn't copy the terminator from BB over to NewBB, because there is now
1425 // an unconditional jump to SuccBB. Insert the unconditional jump.
1426 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB);
1427 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1429 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1430 // PHI nodes for NewBB now.
1431 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1433 // If there were values defined in BB that are used outside the block, then we
1434 // now have to update all uses of the value to use either the original value,
1435 // the cloned value, or some PHI derived value. This can require arbitrary
1436 // PHI insertion, of which we are prepared to do, clean these up now.
1437 SSAUpdater SSAUpdate;
1438 SmallVector<Use*, 16> UsesToRename;
1439 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1440 // Scan all uses of this instruction to see if it is used outside of its
1441 // block, and if so, record them in UsesToRename.
1442 for (Use &U : I->uses()) {
1443 Instruction *User = cast<Instruction>(U.getUser());
1444 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1445 if (UserPN->getIncomingBlock(U) == BB)
1447 } else if (User->getParent() == BB)
1450 UsesToRename.push_back(&U);
1453 // If there are no uses outside the block, we're done with this instruction.
1454 if (UsesToRename.empty())
1457 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1459 // We found a use of I outside of BB. Rename all uses of I that are outside
1460 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1461 // with the two values we know.
1462 SSAUpdate.Initialize(I->getType(), I->getName());
1463 SSAUpdate.AddAvailableValue(BB, I);
1464 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1466 while (!UsesToRename.empty())
1467 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1468 DEBUG(dbgs() << "\n");
1472 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1473 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1474 // us to simplify any PHI nodes in BB.
1475 TerminatorInst *PredTerm = PredBB->getTerminator();
1476 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1477 if (PredTerm->getSuccessor(i) == BB) {
1478 BB->removePredecessor(PredBB, true);
1479 PredTerm->setSuccessor(i, NewBB);
1482 // At this point, the IR is fully up to date and consistent. Do a quick scan
1483 // over the new instructions and zap any that are constants or dead. This
1484 // frequently happens because of phi translation.
1485 SimplifyInstructionsInBlock(NewBB, DL, TLI);
1487 // Threaded an edge!
1492 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1493 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1494 /// If we can duplicate the contents of BB up into PredBB do so now, this
1495 /// improves the odds that the branch will be on an analyzable instruction like
1497 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1498 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1499 assert(!PredBBs.empty() && "Can't handle an empty set");
1501 // If BB is a loop header, then duplicating this block outside the loop would
1502 // cause us to transform this into an irreducible loop, don't do this.
1503 // See the comments above FindLoopHeaders for justifications and caveats.
1504 if (LoopHeaders.count(BB)) {
1505 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1506 << "' into predecessor block '" << PredBBs[0]->getName()
1507 << "' - it might create an irreducible loop!\n");
1511 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, Threshold);
1512 if (DuplicationCost > Threshold) {
1513 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1514 << "' - Cost is too high: " << DuplicationCost << "\n");
1518 // And finally, do it! Start by factoring the predecessors is needed.
1520 if (PredBBs.size() == 1)
1521 PredBB = PredBBs[0];
1523 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1524 << " common predecessors.\n");
1525 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1528 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1530 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1531 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1532 << DuplicationCost << " block is:" << *BB << "\n");
1534 // Unless PredBB ends with an unconditional branch, split the edge so that we
1535 // can just clone the bits from BB into the end of the new PredBB.
1536 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1538 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
1539 PredBB = SplitEdge(PredBB, BB, this);
1540 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1543 // We are going to have to map operands from the original BB block into the
1544 // PredBB block. Evaluate PHI nodes in BB.
1545 DenseMap<Instruction*, Value*> ValueMapping;
1547 BasicBlock::iterator BI = BB->begin();
1548 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1549 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1551 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1552 // mapping and using it to remap operands in the cloned instructions.
1553 for (; BI != BB->end(); ++BI) {
1554 Instruction *New = BI->clone();
1556 // Remap operands to patch up intra-block references.
1557 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1558 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1559 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1560 if (I != ValueMapping.end())
1561 New->setOperand(i, I->second);
1564 // If this instruction can be simplified after the operands are updated,
1565 // just use the simplified value instead. This frequently happens due to
1567 if (Value *IV = SimplifyInstruction(New, DL)) {
1569 ValueMapping[BI] = IV;
1571 // Otherwise, insert the new instruction into the block.
1572 New->setName(BI->getName());
1573 PredBB->getInstList().insert(OldPredBranch, New);
1574 ValueMapping[BI] = New;
1578 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1579 // add entries to the PHI nodes for branch from PredBB now.
1580 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1581 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1583 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1586 // If there were values defined in BB that are used outside the block, then we
1587 // now have to update all uses of the value to use either the original value,
1588 // the cloned value, or some PHI derived value. This can require arbitrary
1589 // PHI insertion, of which we are prepared to do, clean these up now.
1590 SSAUpdater SSAUpdate;
1591 SmallVector<Use*, 16> UsesToRename;
1592 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1593 // Scan all uses of this instruction to see if it is used outside of its
1594 // block, and if so, record them in UsesToRename.
1595 for (Use &U : I->uses()) {
1596 Instruction *User = cast<Instruction>(U.getUser());
1597 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1598 if (UserPN->getIncomingBlock(U) == BB)
1600 } else if (User->getParent() == BB)
1603 UsesToRename.push_back(&U);
1606 // If there are no uses outside the block, we're done with this instruction.
1607 if (UsesToRename.empty())
1610 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1612 // We found a use of I outside of BB. Rename all uses of I that are outside
1613 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1614 // with the two values we know.
1615 SSAUpdate.Initialize(I->getType(), I->getName());
1616 SSAUpdate.AddAvailableValue(BB, I);
1617 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1619 while (!UsesToRename.empty())
1620 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1621 DEBUG(dbgs() << "\n");
1624 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1626 BB->removePredecessor(PredBB, true);
1628 // Remove the unconditional branch at the end of the PredBB block.
1629 OldPredBranch->eraseFromParent();
1635 /// TryToUnfoldSelect - Look for blocks of the form
1641 /// %p = phi [%a, %bb] ...
1645 /// And expand the select into a branch structure if one of its arms allows %c
1646 /// to be folded. This later enables threading from bb1 over bb2.
1647 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1648 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1649 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1650 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1652 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1653 CondLHS->getParent() != BB)
1656 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1657 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1658 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1660 // Look if one of the incoming values is a select in the corresponding
1662 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1665 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1666 if (!PredTerm || !PredTerm->isUnconditional())
1669 // Now check if one of the select values would allow us to constant fold the
1670 // terminator in BB. We don't do the transform if both sides fold, those
1671 // cases will be threaded in any case.
1672 LazyValueInfo::Tristate LHSFolds =
1673 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1675 LazyValueInfo::Tristate RHSFolds =
1676 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
1678 if ((LHSFolds != LazyValueInfo::Unknown ||
1679 RHSFolds != LazyValueInfo::Unknown) &&
1680 LHSFolds != RHSFolds) {
1681 // Expand the select.
1690 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
1691 BB->getParent(), BB);
1692 // Move the unconditional branch to NewBB.
1693 PredTerm->removeFromParent();
1694 NewBB->getInstList().insert(NewBB->end(), PredTerm);
1695 // Create a conditional branch and update PHI nodes.
1696 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
1697 CondLHS->setIncomingValue(I, SI->getFalseValue());
1698 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
1699 // The select is now dead.
1700 SI->eraseFromParent();
1702 // Update any other PHI nodes in BB.
1703 for (BasicBlock::iterator BI = BB->begin();
1704 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
1706 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);