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
162 // loop. We do threading if we found something profitable. Jump threading a
163 // branch can create other opportunities. If these opportunities form a cycle
164 // i.e. if any jump treading is undoing previous threading in the path, then
165 // we will loop forever. We take care of this issue by not jump threading for
166 // back edges. This works for normal cases but not for unreachable blocks as
167 // they may have cycle with no back edge.
168 removeUnreachableBlocks(F);
172 bool Changed, EverChanged = false;
175 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
177 // Thread all of the branches we can over this block.
178 while (ProcessBlock(BB))
183 // If the block is trivially dead, zap it. This eliminates the successor
184 // edges which simplifies the CFG.
185 if (pred_begin(BB) == pred_end(BB) &&
186 BB != &BB->getParent()->getEntryBlock()) {
187 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
188 << "' with terminator: " << *BB->getTerminator() << '\n');
189 LoopHeaders.erase(BB);
196 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
198 // Can't thread an unconditional jump, but if the block is "almost
199 // empty", we can replace uses of it with uses of the successor and make
201 if (BI && BI->isUnconditional() &&
202 BB != &BB->getParent()->getEntryBlock() &&
203 // If the terminator is the only non-phi instruction, try to nuke it.
204 BB->getFirstNonPHIOrDbg()->isTerminator()) {
205 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
206 // block, we have to make sure it isn't in the LoopHeaders set. We
207 // reinsert afterward if needed.
208 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
209 BasicBlock *Succ = BI->getSuccessor(0);
211 // FIXME: It is always conservatively correct to drop the info
212 // for a block even if it doesn't get erased. This isn't totally
213 // awesome, but it allows us to use AssertingVH to prevent nasty
214 // dangling pointer issues within LazyValueInfo.
216 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
218 // If we deleted BB and BB was the header of a loop, then the
219 // successor is now the header of the loop.
223 if (ErasedFromLoopHeaders)
224 LoopHeaders.insert(BB);
227 EverChanged |= Changed;
234 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
235 /// thread across it. Stop scanning the block when passing the threshold.
236 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
237 unsigned Threshold) {
238 /// Ignore PHI nodes, these will be flattened when duplication happens.
239 BasicBlock::const_iterator I = BB->getFirstNonPHI();
241 // FIXME: THREADING will delete values that are just used to compute the
242 // branch, so they shouldn't count against the duplication cost.
244 // Sum up the cost of each instruction until we get to the terminator. Don't
245 // include the terminator because the copy won't include it.
247 for (; !isa<TerminatorInst>(I); ++I) {
249 // Stop scanning the block if we've reached the threshold.
250 if (Size > Threshold)
253 // Debugger intrinsics don't incur code size.
254 if (isa<DbgInfoIntrinsic>(I)) continue;
256 // If this is a pointer->pointer bitcast, it is free.
257 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
260 // All other instructions count for at least one unit.
263 // Calls are more expensive. If they are non-intrinsic calls, we model them
264 // as having cost of 4. If they are a non-vector intrinsic, we model them
265 // as having cost of 2 total, and if they are a vector intrinsic, we model
266 // them as having cost 1.
267 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
268 if (CI->cannotDuplicate())
269 // Blocks with NoDuplicate are modelled as having infinite cost, so they
270 // are never duplicated.
272 else if (!isa<IntrinsicInst>(CI))
274 else if (!CI->getType()->isVectorTy())
279 // Threading through a switch statement is particularly profitable. If this
280 // block ends in a switch, decrease its cost to make it more likely to happen.
281 if (isa<SwitchInst>(I))
282 Size = Size > 6 ? Size-6 : 0;
284 // The same holds for indirect branches, but slightly more so.
285 if (isa<IndirectBrInst>(I))
286 Size = Size > 8 ? Size-8 : 0;
291 /// FindLoopHeaders - We do not want jump threading to turn proper loop
292 /// structures into irreducible loops. Doing this breaks up the loop nesting
293 /// hierarchy and pessimizes later transformations. To prevent this from
294 /// happening, we first have to find the loop headers. Here we approximate this
295 /// by finding targets of backedges in the CFG.
297 /// Note that there definitely are cases when we want to allow threading of
298 /// edges across a loop header. For example, threading a jump from outside the
299 /// loop (the preheader) to an exit block of the loop is definitely profitable.
300 /// It is also almost always profitable to thread backedges from within the loop
301 /// to exit blocks, and is often profitable to thread backedges to other blocks
302 /// within the loop (forming a nested loop). This simple analysis is not rich
303 /// enough to track all of these properties and keep it up-to-date as the CFG
304 /// mutates, so we don't allow any of these transformations.
306 void JumpThreading::FindLoopHeaders(Function &F) {
307 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
308 FindFunctionBackedges(F, Edges);
310 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
311 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
314 /// getKnownConstant - Helper method to determine if we can thread over a
315 /// terminator with the given value as its condition, and if so what value to
316 /// use for that. What kind of value this is depends on whether we want an
317 /// integer or a block address, but an undef is always accepted.
318 /// Returns null if Val is null or not an appropriate constant.
319 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
323 // Undef is "known" enough.
324 if (UndefValue *U = dyn_cast<UndefValue>(Val))
327 if (Preference == WantBlockAddress)
328 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
330 return dyn_cast<ConstantInt>(Val);
333 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
334 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
335 /// in any of our predecessors. If so, return the known list of value and pred
336 /// BB in the result vector.
338 /// This returns true if there were any known values.
341 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
342 ConstantPreference Preference) {
343 // This method walks up use-def chains recursively. Because of this, we could
344 // get into an infinite loop going around loops in the use-def chain. To
345 // prevent this, keep track of what (value, block) pairs we've already visited
346 // and terminate the search if we loop back to them
347 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
350 // An RAII help to remove this pair from the recursion set once the recursion
351 // stack pops back out again.
352 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
354 // If V is a constant, then it is known in all predecessors.
355 if (Constant *KC = getKnownConstant(V, Preference)) {
356 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
357 Result.push_back(std::make_pair(KC, *PI));
362 // If V is a non-instruction value, or an instruction in a different block,
363 // then it can't be derived from a PHI.
364 Instruction *I = dyn_cast<Instruction>(V);
365 if (!I || I->getParent() != BB) {
367 // Okay, if this is a live-in value, see if it has a known value at the end
368 // of any of our predecessors.
370 // FIXME: This should be an edge property, not a block end property.
371 /// TODO: Per PR2563, we could infer value range information about a
372 /// predecessor based on its terminator.
374 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
375 // "I" is a non-local compare-with-a-constant instruction. This would be
376 // able to handle value inequalities better, for example if the compare is
377 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
378 // Perhaps getConstantOnEdge should be smart enough to do this?
380 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
382 // If the value is known by LazyValueInfo to be a constant in a
383 // predecessor, use that information to try to thread this block.
384 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
385 if (Constant *KC = getKnownConstant(PredCst, Preference))
386 Result.push_back(std::make_pair(KC, P));
389 return !Result.empty();
392 /// If I is a PHI node, then we know the incoming values for any constants.
393 if (PHINode *PN = dyn_cast<PHINode>(I)) {
394 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
395 Value *InVal = PN->getIncomingValue(i);
396 if (Constant *KC = getKnownConstant(InVal, Preference)) {
397 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
399 Constant *CI = LVI->getConstantOnEdge(InVal,
400 PN->getIncomingBlock(i), BB);
401 if (Constant *KC = getKnownConstant(CI, Preference))
402 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
406 return !Result.empty();
409 PredValueInfoTy LHSVals, RHSVals;
411 // Handle some boolean conditions.
412 if (I->getType()->getPrimitiveSizeInBits() == 1) {
413 assert(Preference == WantInteger && "One-bit non-integer type?");
415 // X & false -> false
416 if (I->getOpcode() == Instruction::Or ||
417 I->getOpcode() == Instruction::And) {
418 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
420 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
423 if (LHSVals.empty() && RHSVals.empty())
426 ConstantInt *InterestingVal;
427 if (I->getOpcode() == Instruction::Or)
428 InterestingVal = ConstantInt::getTrue(I->getContext());
430 InterestingVal = ConstantInt::getFalse(I->getContext());
432 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
434 // Scan for the sentinel. If we find an undef, force it to the
435 // interesting value: x|undef -> true and x&undef -> false.
436 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
437 if (LHSVals[i].first == InterestingVal ||
438 isa<UndefValue>(LHSVals[i].first)) {
439 Result.push_back(LHSVals[i]);
440 Result.back().first = InterestingVal;
441 LHSKnownBBs.insert(LHSVals[i].second);
443 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
444 if (RHSVals[i].first == InterestingVal ||
445 isa<UndefValue>(RHSVals[i].first)) {
446 // If we already inferred a value for this block on the LHS, don't
448 if (!LHSKnownBBs.count(RHSVals[i].second)) {
449 Result.push_back(RHSVals[i]);
450 Result.back().first = InterestingVal;
454 return !Result.empty();
457 // Handle the NOT form of XOR.
458 if (I->getOpcode() == Instruction::Xor &&
459 isa<ConstantInt>(I->getOperand(1)) &&
460 cast<ConstantInt>(I->getOperand(1))->isOne()) {
461 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
466 // Invert the known values.
467 for (unsigned i = 0, e = Result.size(); i != e; ++i)
468 Result[i].first = ConstantExpr::getNot(Result[i].first);
473 // Try to simplify some other binary operator values.
474 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
475 assert(Preference != WantBlockAddress
476 && "A binary operator creating a block address?");
477 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
478 PredValueInfoTy LHSVals;
479 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
482 // Try to use constant folding to simplify the binary operator.
483 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
484 Constant *V = LHSVals[i].first;
485 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
487 if (Constant *KC = getKnownConstant(Folded, WantInteger))
488 Result.push_back(std::make_pair(KC, LHSVals[i].second));
492 return !Result.empty();
495 // Handle compare with phi operand, where the PHI is defined in this block.
496 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
497 assert(Preference == WantInteger && "Compares only produce integers");
498 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
499 if (PN && PN->getParent() == BB) {
500 // We can do this simplification if any comparisons fold to true or false.
502 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
503 BasicBlock *PredBB = PN->getIncomingBlock(i);
504 Value *LHS = PN->getIncomingValue(i);
505 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
507 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL);
509 if (!isa<Constant>(RHS))
512 LazyValueInfo::Tristate
513 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
514 cast<Constant>(RHS), PredBB, BB);
515 if (ResT == LazyValueInfo::Unknown)
517 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
520 if (Constant *KC = getKnownConstant(Res, WantInteger))
521 Result.push_back(std::make_pair(KC, PredBB));
524 return !Result.empty();
528 // If comparing a live-in value against a constant, see if we know the
529 // live-in value on any predecessors.
530 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
531 if (!isa<Instruction>(Cmp->getOperand(0)) ||
532 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
533 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
535 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
537 // If the value is known by LazyValueInfo to be a constant in a
538 // predecessor, use that information to try to thread this block.
539 LazyValueInfo::Tristate Res =
540 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
542 if (Res == LazyValueInfo::Unknown)
545 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
546 Result.push_back(std::make_pair(ResC, P));
549 return !Result.empty();
552 // Try to find a constant value for the LHS of a comparison,
553 // and evaluate it statically if we can.
554 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
555 PredValueInfoTy LHSVals;
556 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
559 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
560 Constant *V = LHSVals[i].first;
561 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
563 if (Constant *KC = getKnownConstant(Folded, WantInteger))
564 Result.push_back(std::make_pair(KC, LHSVals[i].second));
567 return !Result.empty();
572 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
573 // Handle select instructions where at least one operand is a known constant
574 // and we can figure out the condition value for any predecessor block.
575 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
576 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
577 PredValueInfoTy Conds;
578 if ((TrueVal || FalseVal) &&
579 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
581 for (unsigned i = 0, e = Conds.size(); i != e; ++i) {
582 Constant *Cond = Conds[i].first;
584 // Figure out what value to use for the condition.
586 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
588 KnownCond = CI->isOne();
590 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
591 // Either operand will do, so be sure to pick the one that's a known
593 // FIXME: Do this more cleverly if both values are known constants?
594 KnownCond = (TrueVal != nullptr);
597 // See if the select has a known constant value for this predecessor.
598 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
599 Result.push_back(std::make_pair(Val, Conds[i].second));
602 return !Result.empty();
606 // If all else fails, see if LVI can figure out a constant value for us.
607 Constant *CI = LVI->getConstant(V, BB);
608 if (Constant *KC = getKnownConstant(CI, Preference)) {
609 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
610 Result.push_back(std::make_pair(KC, *PI));
613 return !Result.empty();
618 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
619 /// in an undefined jump, decide which block is best to revector to.
621 /// Since we can pick an arbitrary destination, we pick the successor with the
622 /// fewest predecessors. This should reduce the in-degree of the others.
624 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
625 TerminatorInst *BBTerm = BB->getTerminator();
626 unsigned MinSucc = 0;
627 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
628 // Compute the successor with the minimum number of predecessors.
629 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
630 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
631 TestBB = BBTerm->getSuccessor(i);
632 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
633 if (NumPreds < MinNumPreds) {
635 MinNumPreds = NumPreds;
642 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
643 if (!BB->hasAddressTaken()) return false;
645 // If the block has its address taken, it may be a tree of dead constants
646 // hanging off of it. These shouldn't keep the block alive.
647 BlockAddress *BA = BlockAddress::get(BB);
648 BA->removeDeadConstantUsers();
649 return !BA->use_empty();
652 /// ProcessBlock - If there are any predecessors whose control can be threaded
653 /// through to a successor, transform them now.
654 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
655 // If the block is trivially dead, just return and let the caller nuke it.
656 // This simplifies other transformations.
657 if (pred_begin(BB) == pred_end(BB) &&
658 BB != &BB->getParent()->getEntryBlock())
661 // If this block has a single predecessor, and if that pred has a single
662 // successor, merge the blocks. This encourages recursive jump threading
663 // because now the condition in this block can be threaded through
664 // predecessors of our predecessor block.
665 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
666 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
667 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
668 // If SinglePred was a loop header, BB becomes one.
669 if (LoopHeaders.erase(SinglePred))
670 LoopHeaders.insert(BB);
672 LVI->eraseBlock(SinglePred);
673 MergeBasicBlockIntoOnlyPred(BB);
679 // What kind of constant we're looking for.
680 ConstantPreference Preference = WantInteger;
682 // Look to see if the terminator is a conditional branch, switch or indirect
683 // branch, if not we can't thread it.
685 Instruction *Terminator = BB->getTerminator();
686 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
687 // Can't thread an unconditional jump.
688 if (BI->isUnconditional()) return false;
689 Condition = BI->getCondition();
690 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
691 Condition = SI->getCondition();
692 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
693 // Can't thread indirect branch with no successors.
694 if (IB->getNumSuccessors() == 0) return false;
695 Condition = IB->getAddress()->stripPointerCasts();
696 Preference = WantBlockAddress;
698 return false; // Must be an invoke.
701 // Run constant folding to see if we can reduce the condition to a simple
703 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
704 Value *SimpleVal = ConstantFoldInstruction(I, DL, TLI);
706 I->replaceAllUsesWith(SimpleVal);
707 I->eraseFromParent();
708 Condition = SimpleVal;
712 // If the terminator is branching on an undef, we can pick any of the
713 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
714 if (isa<UndefValue>(Condition)) {
715 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
717 // Fold the branch/switch.
718 TerminatorInst *BBTerm = BB->getTerminator();
719 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
720 if (i == BestSucc) continue;
721 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
724 DEBUG(dbgs() << " In block '" << BB->getName()
725 << "' folding undef terminator: " << *BBTerm << '\n');
726 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
727 BBTerm->eraseFromParent();
731 // If the terminator of this block is branching on a constant, simplify the
732 // terminator to an unconditional branch. This can occur due to threading in
734 if (getKnownConstant(Condition, Preference)) {
735 DEBUG(dbgs() << " In block '" << BB->getName()
736 << "' folding terminator: " << *BB->getTerminator() << '\n');
738 ConstantFoldTerminator(BB, true);
742 Instruction *CondInst = dyn_cast<Instruction>(Condition);
744 // All the rest of our checks depend on the condition being an instruction.
746 // FIXME: Unify this with code below.
747 if (ProcessThreadableEdges(Condition, BB, Preference))
753 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
754 // For a comparison where the LHS is outside this block, it's possible
755 // that we've branched on it before. Used LVI to see if we can simplify
756 // the branch based on that.
757 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
758 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
759 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
760 if (CondBr && CondConst && CondBr->isConditional() && PI != PE &&
761 (!isa<Instruction>(CondCmp->getOperand(0)) ||
762 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
763 // For predecessor edge, determine if the comparison is true or false
764 // on that edge. If they're all true or all false, we can simplify the
766 // FIXME: We could handle mixed true/false by duplicating code.
767 LazyValueInfo::Tristate Baseline =
768 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
770 if (Baseline != LazyValueInfo::Unknown) {
771 // Check that all remaining incoming values match the first one.
773 LazyValueInfo::Tristate Ret =
774 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
775 CondCmp->getOperand(0), CondConst, *PI, BB);
776 if (Ret != Baseline) break;
779 // If we terminated early, then one of the values didn't match.
781 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
782 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
783 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
784 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
785 CondBr->eraseFromParent();
792 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB))
796 // Check for some cases that are worth simplifying. Right now we want to look
797 // for loads that are used by a switch or by the condition for the branch. If
798 // we see one, check to see if it's partially redundant. If so, insert a PHI
799 // which can then be used to thread the values.
801 Value *SimplifyValue = CondInst;
802 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
803 if (isa<Constant>(CondCmp->getOperand(1)))
804 SimplifyValue = CondCmp->getOperand(0);
806 // TODO: There are other places where load PRE would be profitable, such as
807 // more complex comparisons.
808 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
809 if (SimplifyPartiallyRedundantLoad(LI))
813 // Handle a variety of cases where we are branching on something derived from
814 // a PHI node in the current block. If we can prove that any predecessors
815 // compute a predictable value based on a PHI node, thread those predecessors.
817 if (ProcessThreadableEdges(CondInst, BB, Preference))
820 // If this is an otherwise-unfoldable branch on a phi node in the current
821 // block, see if we can simplify.
822 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
823 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
824 return ProcessBranchOnPHI(PN);
827 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
828 if (CondInst->getOpcode() == Instruction::Xor &&
829 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
830 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
833 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
834 // "(X == 4)", thread through this block.
839 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
840 /// load instruction, eliminate it by replacing it with a PHI node. This is an
841 /// important optimization that encourages jump threading, and needs to be run
842 /// interlaced with other jump threading tasks.
843 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
844 // Don't hack volatile/atomic loads.
845 if (!LI->isSimple()) return false;
847 // If the load is defined in a block with exactly one predecessor, it can't be
848 // partially redundant.
849 BasicBlock *LoadBB = LI->getParent();
850 if (LoadBB->getSinglePredecessor())
853 // If the load is defined in a landing pad, it can't be partially redundant,
854 // because the edges between the invoke and the landing pad cannot have other
855 // instructions between them.
856 if (LoadBB->isLandingPad())
859 Value *LoadedPtr = LI->getOperand(0);
861 // If the loaded operand is defined in the LoadBB, it can't be available.
862 // TODO: Could do simple PHI translation, that would be fun :)
863 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
864 if (PtrOp->getParent() == LoadBB)
867 // Scan a few instructions up from the load, to see if it is obviously live at
868 // the entry to its block.
869 BasicBlock::iterator BBIt = LI;
871 if (Value *AvailableVal =
872 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
873 // If the value if the load is locally available within the block, just use
874 // it. This frequently occurs for reg2mem'd allocas.
875 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
877 // If the returned value is the load itself, replace with an undef. This can
878 // only happen in dead loops.
879 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
880 LI->replaceAllUsesWith(AvailableVal);
881 LI->eraseFromParent();
885 // Otherwise, if we scanned the whole block and got to the top of the block,
886 // we know the block is locally transparent to the load. If not, something
887 // might clobber its value.
888 if (BBIt != LoadBB->begin())
891 // If all of the loads and stores that feed the value have the same TBAA tag,
892 // then we can propagate it onto any newly inserted loads.
893 MDNode *TBAATag = LI->getMetadata(LLVMContext::MD_tbaa);
895 SmallPtrSet<BasicBlock*, 8> PredsScanned;
896 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
897 AvailablePredsTy AvailablePreds;
898 BasicBlock *OneUnavailablePred = nullptr;
900 // If we got here, the loaded value is transparent through to the start of the
901 // block. Check to see if it is available in any of the predecessor blocks.
902 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
904 BasicBlock *PredBB = *PI;
906 // If we already scanned this predecessor, skip it.
907 if (!PredsScanned.insert(PredBB))
910 // Scan the predecessor to see if the value is available in the pred.
911 BBIt = PredBB->end();
912 MDNode *ThisTBAATag = nullptr;
913 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6,
914 nullptr, &ThisTBAATag);
915 if (!PredAvailable) {
916 OneUnavailablePred = PredBB;
920 // If tbaa tags disagree or are not present, forget about them.
921 if (TBAATag != ThisTBAATag) TBAATag = nullptr;
923 // If so, this load is partially redundant. Remember this info so that we
924 // can create a PHI node.
925 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
928 // If the loaded value isn't available in any predecessor, it isn't partially
930 if (AvailablePreds.empty()) return false;
932 // Okay, the loaded value is available in at least one (and maybe all!)
933 // predecessors. If the value is unavailable in more than one unique
934 // predecessor, we want to insert a merge block for those common predecessors.
935 // This ensures that we only have to insert one reload, thus not increasing
937 BasicBlock *UnavailablePred = nullptr;
939 // If there is exactly one predecessor where the value is unavailable, the
940 // already computed 'OneUnavailablePred' block is it. If it ends in an
941 // unconditional branch, we know that it isn't a critical edge.
942 if (PredsScanned.size() == AvailablePreds.size()+1 &&
943 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
944 UnavailablePred = OneUnavailablePred;
945 } else if (PredsScanned.size() != AvailablePreds.size()) {
946 // Otherwise, we had multiple unavailable predecessors or we had a critical
947 // edge from the one.
948 SmallVector<BasicBlock*, 8> PredsToSplit;
949 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
951 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
952 AvailablePredSet.insert(AvailablePreds[i].first);
954 // Add all the unavailable predecessors to the PredsToSplit list.
955 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
958 // If the predecessor is an indirect goto, we can't split the edge.
959 if (isa<IndirectBrInst>(P->getTerminator()))
962 if (!AvailablePredSet.count(P))
963 PredsToSplit.push_back(P);
966 // Split them out to their own block.
968 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split", this);
971 // If the value isn't available in all predecessors, then there will be
972 // exactly one where it isn't available. Insert a load on that edge and add
973 // it to the AvailablePreds list.
974 if (UnavailablePred) {
975 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
976 "Can't handle critical edge here!");
977 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
979 UnavailablePred->getTerminator());
980 NewVal->setDebugLoc(LI->getDebugLoc());
982 NewVal->setMetadata(LLVMContext::MD_tbaa, TBAATag);
984 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
987 // Now we know that each predecessor of this block has a value in
988 // AvailablePreds, sort them for efficient access as we're walking the preds.
989 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
991 // Create a PHI node at the start of the block for the PRE'd load value.
992 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
993 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
996 PN->setDebugLoc(LI->getDebugLoc());
998 // Insert new entries into the PHI for each predecessor. A single block may
999 // have multiple entries here.
1000 for (pred_iterator PI = PB; PI != PE; ++PI) {
1001 BasicBlock *P = *PI;
1002 AvailablePredsTy::iterator I =
1003 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1004 std::make_pair(P, (Value*)nullptr));
1006 assert(I != AvailablePreds.end() && I->first == P &&
1007 "Didn't find entry for predecessor!");
1009 PN->addIncoming(I->second, I->first);
1012 //cerr << "PRE: " << *LI << *PN << "\n";
1014 LI->replaceAllUsesWith(PN);
1015 LI->eraseFromParent();
1020 /// FindMostPopularDest - The specified list contains multiple possible
1021 /// threadable destinations. Pick the one that occurs the most frequently in
1024 FindMostPopularDest(BasicBlock *BB,
1025 const SmallVectorImpl<std::pair<BasicBlock*,
1026 BasicBlock*> > &PredToDestList) {
1027 assert(!PredToDestList.empty());
1029 // Determine popularity. If there are multiple possible destinations, we
1030 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1031 // blocks with known and real destinations to threading undef. We'll handle
1032 // them later if interesting.
1033 DenseMap<BasicBlock*, unsigned> DestPopularity;
1034 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1035 if (PredToDestList[i].second)
1036 DestPopularity[PredToDestList[i].second]++;
1038 // Find the most popular dest.
1039 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1040 BasicBlock *MostPopularDest = DPI->first;
1041 unsigned Popularity = DPI->second;
1042 SmallVector<BasicBlock*, 4> SamePopularity;
1044 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1045 // If the popularity of this entry isn't higher than the popularity we've
1046 // seen so far, ignore it.
1047 if (DPI->second < Popularity)
1049 else if (DPI->second == Popularity) {
1050 // If it is the same as what we've seen so far, keep track of it.
1051 SamePopularity.push_back(DPI->first);
1053 // If it is more popular, remember it.
1054 SamePopularity.clear();
1055 MostPopularDest = DPI->first;
1056 Popularity = DPI->second;
1060 // Okay, now we know the most popular destination. If there is more than one
1061 // destination, we need to determine one. This is arbitrary, but we need
1062 // to make a deterministic decision. Pick the first one that appears in the
1064 if (!SamePopularity.empty()) {
1065 SamePopularity.push_back(MostPopularDest);
1066 TerminatorInst *TI = BB->getTerminator();
1067 for (unsigned i = 0; ; ++i) {
1068 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1070 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1071 TI->getSuccessor(i)) == SamePopularity.end())
1074 MostPopularDest = TI->getSuccessor(i);
1079 // Okay, we have finally picked the most popular destination.
1080 return MostPopularDest;
1083 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1084 ConstantPreference Preference) {
1085 // If threading this would thread across a loop header, don't even try to
1087 if (LoopHeaders.count(BB))
1090 PredValueInfoTy PredValues;
1091 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference))
1094 assert(!PredValues.empty() &&
1095 "ComputeValueKnownInPredecessors returned true with no values");
1097 DEBUG(dbgs() << "IN BB: " << *BB;
1098 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1099 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1100 << *PredValues[i].first
1101 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1104 // Decide what we want to thread through. Convert our list of known values to
1105 // a list of known destinations for each pred. This also discards duplicate
1106 // predecessors and keeps track of the undefined inputs (which are represented
1107 // as a null dest in the PredToDestList).
1108 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1109 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1111 BasicBlock *OnlyDest = nullptr;
1112 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1114 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1115 BasicBlock *Pred = PredValues[i].second;
1116 if (!SeenPreds.insert(Pred))
1117 continue; // Duplicate predecessor entry.
1119 // If the predecessor ends with an indirect goto, we can't change its
1121 if (isa<IndirectBrInst>(Pred->getTerminator()))
1124 Constant *Val = PredValues[i].first;
1127 if (isa<UndefValue>(Val))
1129 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1130 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1131 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1132 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1134 assert(isa<IndirectBrInst>(BB->getTerminator())
1135 && "Unexpected terminator");
1136 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1139 // If we have exactly one destination, remember it for efficiency below.
1140 if (PredToDestList.empty())
1142 else if (OnlyDest != DestBB)
1143 OnlyDest = MultipleDestSentinel;
1145 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1148 // If all edges were unthreadable, we fail.
1149 if (PredToDestList.empty())
1152 // Determine which is the most common successor. If we have many inputs and
1153 // this block is a switch, we want to start by threading the batch that goes
1154 // to the most popular destination first. If we only know about one
1155 // threadable destination (the common case) we can avoid this.
1156 BasicBlock *MostPopularDest = OnlyDest;
1158 if (MostPopularDest == MultipleDestSentinel)
1159 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1161 // Now that we know what the most popular destination is, factor all
1162 // predecessors that will jump to it into a single predecessor.
1163 SmallVector<BasicBlock*, 16> PredsToFactor;
1164 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1165 if (PredToDestList[i].second == MostPopularDest) {
1166 BasicBlock *Pred = PredToDestList[i].first;
1168 // This predecessor may be a switch or something else that has multiple
1169 // edges to the block. Factor each of these edges by listing them
1170 // according to # occurrences in PredsToFactor.
1171 TerminatorInst *PredTI = Pred->getTerminator();
1172 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1173 if (PredTI->getSuccessor(i) == BB)
1174 PredsToFactor.push_back(Pred);
1177 // If the threadable edges are branching on an undefined value, we get to pick
1178 // the destination that these predecessors should get to.
1179 if (!MostPopularDest)
1180 MostPopularDest = BB->getTerminator()->
1181 getSuccessor(GetBestDestForJumpOnUndef(BB));
1183 // Ok, try to thread it!
1184 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1187 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1188 /// a PHI node in the current block. See if there are any simplifications we
1189 /// can do based on inputs to the phi node.
1191 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1192 BasicBlock *BB = PN->getParent();
1194 // TODO: We could make use of this to do it once for blocks with common PHI
1196 SmallVector<BasicBlock*, 1> PredBBs;
1199 // If any of the predecessor blocks end in an unconditional branch, we can
1200 // *duplicate* the conditional branch into that block in order to further
1201 // encourage jump threading and to eliminate cases where we have branch on a
1202 // phi of an icmp (branch on icmp is much better).
1203 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1204 BasicBlock *PredBB = PN->getIncomingBlock(i);
1205 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1206 if (PredBr->isUnconditional()) {
1207 PredBBs[0] = PredBB;
1208 // Try to duplicate BB into PredBB.
1209 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1217 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1218 /// a xor instruction in the current block. See if there are any
1219 /// simplifications we can do based on inputs to the xor.
1221 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1222 BasicBlock *BB = BO->getParent();
1224 // If either the LHS or RHS of the xor is a constant, don't do this
1226 if (isa<ConstantInt>(BO->getOperand(0)) ||
1227 isa<ConstantInt>(BO->getOperand(1)))
1230 // If the first instruction in BB isn't a phi, we won't be able to infer
1231 // anything special about any particular predecessor.
1232 if (!isa<PHINode>(BB->front()))
1235 // If we have a xor as the branch input to this block, and we know that the
1236 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1237 // the condition into the predecessor and fix that value to true, saving some
1238 // logical ops on that path and encouraging other paths to simplify.
1240 // This copies something like this:
1243 // %X = phi i1 [1], [%X']
1244 // %Y = icmp eq i32 %A, %B
1245 // %Z = xor i1 %X, %Y
1250 // %Y = icmp ne i32 %A, %B
1253 PredValueInfoTy XorOpValues;
1255 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1257 assert(XorOpValues.empty());
1258 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1264 assert(!XorOpValues.empty() &&
1265 "ComputeValueKnownInPredecessors returned true with no values");
1267 // Scan the information to see which is most popular: true or false. The
1268 // predecessors can be of the set true, false, or undef.
1269 unsigned NumTrue = 0, NumFalse = 0;
1270 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1271 if (isa<UndefValue>(XorOpValues[i].first))
1272 // Ignore undefs for the count.
1274 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1280 // Determine which value to split on, true, false, or undef if neither.
1281 ConstantInt *SplitVal = nullptr;
1282 if (NumTrue > NumFalse)
1283 SplitVal = ConstantInt::getTrue(BB->getContext());
1284 else if (NumTrue != 0 || NumFalse != 0)
1285 SplitVal = ConstantInt::getFalse(BB->getContext());
1287 // Collect all of the blocks that this can be folded into so that we can
1288 // factor this once and clone it once.
1289 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1290 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1291 if (XorOpValues[i].first != SplitVal &&
1292 !isa<UndefValue>(XorOpValues[i].first))
1295 BlocksToFoldInto.push_back(XorOpValues[i].second);
1298 // If we inferred a value for all of the predecessors, then duplication won't
1299 // help us. However, we can just replace the LHS or RHS with the constant.
1300 if (BlocksToFoldInto.size() ==
1301 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1303 // If all preds provide undef, just nuke the xor, because it is undef too.
1304 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1305 BO->eraseFromParent();
1306 } else if (SplitVal->isZero()) {
1307 // If all preds provide 0, replace the xor with the other input.
1308 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1309 BO->eraseFromParent();
1311 // If all preds provide 1, set the computed value to 1.
1312 BO->setOperand(!isLHS, SplitVal);
1318 // Try to duplicate BB into PredBB.
1319 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1323 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1324 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1325 /// NewPred using the entries from OldPred (suitably mapped).
1326 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1327 BasicBlock *OldPred,
1328 BasicBlock *NewPred,
1329 DenseMap<Instruction*, Value*> &ValueMap) {
1330 for (BasicBlock::iterator PNI = PHIBB->begin();
1331 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1332 // Ok, we have a PHI node. Figure out what the incoming value was for the
1334 Value *IV = PN->getIncomingValueForBlock(OldPred);
1336 // Remap the value if necessary.
1337 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1338 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1339 if (I != ValueMap.end())
1343 PN->addIncoming(IV, NewPred);
1347 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1348 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1349 /// across BB. Transform the IR to reflect this change.
1350 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1351 const SmallVectorImpl<BasicBlock*> &PredBBs,
1352 BasicBlock *SuccBB) {
1353 // If threading to the same block as we come from, we would infinite loop.
1355 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1356 << "' - would thread to self!\n");
1360 // If threading this would thread across a loop header, don't thread the edge.
1361 // See the comments above FindLoopHeaders for justifications and caveats.
1362 if (LoopHeaders.count(BB)) {
1363 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1364 << "' to dest BB '" << SuccBB->getName()
1365 << "' - it might create an irreducible loop!\n");
1369 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, Threshold);
1370 if (JumpThreadCost > Threshold) {
1371 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1372 << "' - Cost is too high: " << JumpThreadCost << "\n");
1376 // And finally, do it! Start by factoring the predecessors is needed.
1378 if (PredBBs.size() == 1)
1379 PredBB = PredBBs[0];
1381 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1382 << " common predecessors.\n");
1383 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1386 // And finally, do it!
1387 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1388 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1389 << ", across block:\n "
1392 LVI->threadEdge(PredBB, BB, SuccBB);
1394 // We are going to have to map operands from the original BB block to the new
1395 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1396 // account for entry from PredBB.
1397 DenseMap<Instruction*, Value*> ValueMapping;
1399 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1400 BB->getName()+".thread",
1401 BB->getParent(), BB);
1402 NewBB->moveAfter(PredBB);
1404 BasicBlock::iterator BI = BB->begin();
1405 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1406 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1408 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1409 // mapping and using it to remap operands in the cloned instructions.
1410 for (; !isa<TerminatorInst>(BI); ++BI) {
1411 Instruction *New = BI->clone();
1412 New->setName(BI->getName());
1413 NewBB->getInstList().push_back(New);
1414 ValueMapping[BI] = New;
1416 // Remap operands to patch up intra-block references.
1417 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1418 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1419 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1420 if (I != ValueMapping.end())
1421 New->setOperand(i, I->second);
1425 // We didn't copy the terminator from BB over to NewBB, because there is now
1426 // an unconditional jump to SuccBB. Insert the unconditional jump.
1427 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB);
1428 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1430 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1431 // PHI nodes for NewBB now.
1432 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1434 // If there were values defined in BB that are used outside the block, then we
1435 // now have to update all uses of the value to use either the original value,
1436 // the cloned value, or some PHI derived value. This can require arbitrary
1437 // PHI insertion, of which we are prepared to do, clean these up now.
1438 SSAUpdater SSAUpdate;
1439 SmallVector<Use*, 16> UsesToRename;
1440 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1441 // Scan all uses of this instruction to see if it is used outside of its
1442 // block, and if so, record them in UsesToRename.
1443 for (Use &U : I->uses()) {
1444 Instruction *User = cast<Instruction>(U.getUser());
1445 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1446 if (UserPN->getIncomingBlock(U) == BB)
1448 } else if (User->getParent() == BB)
1451 UsesToRename.push_back(&U);
1454 // If there are no uses outside the block, we're done with this instruction.
1455 if (UsesToRename.empty())
1458 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1460 // We found a use of I outside of BB. Rename all uses of I that are outside
1461 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1462 // with the two values we know.
1463 SSAUpdate.Initialize(I->getType(), I->getName());
1464 SSAUpdate.AddAvailableValue(BB, I);
1465 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1467 while (!UsesToRename.empty())
1468 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1469 DEBUG(dbgs() << "\n");
1473 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1474 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1475 // us to simplify any PHI nodes in BB.
1476 TerminatorInst *PredTerm = PredBB->getTerminator();
1477 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1478 if (PredTerm->getSuccessor(i) == BB) {
1479 BB->removePredecessor(PredBB, true);
1480 PredTerm->setSuccessor(i, NewBB);
1483 // At this point, the IR is fully up to date and consistent. Do a quick scan
1484 // over the new instructions and zap any that are constants or dead. This
1485 // frequently happens because of phi translation.
1486 SimplifyInstructionsInBlock(NewBB, DL, TLI);
1488 // Threaded an edge!
1493 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1494 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1495 /// If we can duplicate the contents of BB up into PredBB do so now, this
1496 /// improves the odds that the branch will be on an analyzable instruction like
1498 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1499 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1500 assert(!PredBBs.empty() && "Can't handle an empty set");
1502 // If BB is a loop header, then duplicating this block outside the loop would
1503 // cause us to transform this into an irreducible loop, don't do this.
1504 // See the comments above FindLoopHeaders for justifications and caveats.
1505 if (LoopHeaders.count(BB)) {
1506 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1507 << "' into predecessor block '" << PredBBs[0]->getName()
1508 << "' - it might create an irreducible loop!\n");
1512 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, Threshold);
1513 if (DuplicationCost > Threshold) {
1514 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1515 << "' - Cost is too high: " << DuplicationCost << "\n");
1519 // And finally, do it! Start by factoring the predecessors is needed.
1521 if (PredBBs.size() == 1)
1522 PredBB = PredBBs[0];
1524 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1525 << " common predecessors.\n");
1526 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1529 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1531 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1532 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1533 << DuplicationCost << " block is:" << *BB << "\n");
1535 // Unless PredBB ends with an unconditional branch, split the edge so that we
1536 // can just clone the bits from BB into the end of the new PredBB.
1537 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1539 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
1540 PredBB = SplitEdge(PredBB, BB, this);
1541 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1544 // We are going to have to map operands from the original BB block into the
1545 // PredBB block. Evaluate PHI nodes in BB.
1546 DenseMap<Instruction*, Value*> ValueMapping;
1548 BasicBlock::iterator BI = BB->begin();
1549 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1550 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1552 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1553 // mapping and using it to remap operands in the cloned instructions.
1554 for (; BI != BB->end(); ++BI) {
1555 Instruction *New = BI->clone();
1557 // Remap operands to patch up intra-block references.
1558 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1559 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1560 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1561 if (I != ValueMapping.end())
1562 New->setOperand(i, I->second);
1565 // If this instruction can be simplified after the operands are updated,
1566 // just use the simplified value instead. This frequently happens due to
1568 if (Value *IV = SimplifyInstruction(New, DL)) {
1570 ValueMapping[BI] = IV;
1572 // Otherwise, insert the new instruction into the block.
1573 New->setName(BI->getName());
1574 PredBB->getInstList().insert(OldPredBranch, New);
1575 ValueMapping[BI] = New;
1579 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1580 // add entries to the PHI nodes for branch from PredBB now.
1581 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1582 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1584 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1587 // If there were values defined in BB that are used outside the block, then we
1588 // now have to update all uses of the value to use either the original value,
1589 // the cloned value, or some PHI derived value. This can require arbitrary
1590 // PHI insertion, of which we are prepared to do, clean these up now.
1591 SSAUpdater SSAUpdate;
1592 SmallVector<Use*, 16> UsesToRename;
1593 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1594 // Scan all uses of this instruction to see if it is used outside of its
1595 // block, and if so, record them in UsesToRename.
1596 for (Use &U : I->uses()) {
1597 Instruction *User = cast<Instruction>(U.getUser());
1598 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1599 if (UserPN->getIncomingBlock(U) == BB)
1601 } else if (User->getParent() == BB)
1604 UsesToRename.push_back(&U);
1607 // If there are no uses outside the block, we're done with this instruction.
1608 if (UsesToRename.empty())
1611 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1613 // We found a use of I outside of BB. Rename all uses of I that are outside
1614 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1615 // with the two values we know.
1616 SSAUpdate.Initialize(I->getType(), I->getName());
1617 SSAUpdate.AddAvailableValue(BB, I);
1618 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1620 while (!UsesToRename.empty())
1621 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1622 DEBUG(dbgs() << "\n");
1625 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1627 BB->removePredecessor(PredBB, true);
1629 // Remove the unconditional branch at the end of the PredBB block.
1630 OldPredBranch->eraseFromParent();
1636 /// TryToUnfoldSelect - Look for blocks of the form
1642 /// %p = phi [%a, %bb] ...
1646 /// And expand the select into a branch structure if one of its arms allows %c
1647 /// to be folded. This later enables threading from bb1 over bb2.
1648 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1649 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1650 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1651 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1653 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1654 CondLHS->getParent() != BB)
1657 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1658 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1659 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1661 // Look if one of the incoming values is a select in the corresponding
1663 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1666 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1667 if (!PredTerm || !PredTerm->isUnconditional())
1670 // Now check if one of the select values would allow us to constant fold the
1671 // terminator in BB. We don't do the transform if both sides fold, those
1672 // cases will be threaded in any case.
1673 LazyValueInfo::Tristate LHSFolds =
1674 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1676 LazyValueInfo::Tristate RHSFolds =
1677 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
1679 if ((LHSFolds != LazyValueInfo::Unknown ||
1680 RHSFolds != LazyValueInfo::Unknown) &&
1681 LHSFolds != RHSFolds) {
1682 // Expand the select.
1691 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
1692 BB->getParent(), BB);
1693 // Move the unconditional branch to NewBB.
1694 PredTerm->removeFromParent();
1695 NewBB->getInstList().insert(NewBB->end(), PredTerm);
1696 // Create a conditional branch and update PHI nodes.
1697 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
1698 CondLHS->setIncomingValue(I, SI->getFalseValue());
1699 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
1700 // The select is now dead.
1701 SI->eraseFromParent();
1703 // Update any other PHI nodes in BB.
1704 for (BasicBlock::iterator BI = BB->begin();
1705 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
1707 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);