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/Metadata.h"
30 #include "llvm/IR/ValueHandle.h"
31 #include "llvm/Pass.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetLibraryInfo.h"
36 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
37 #include "llvm/Transforms/Utils/Local.h"
38 #include "llvm/Transforms/Utils/SSAUpdater.h"
41 #define DEBUG_TYPE "jump-threading"
43 STATISTIC(NumThreads, "Number of jumps threaded");
44 STATISTIC(NumFolds, "Number of terminators folded");
45 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
47 static cl::opt<unsigned>
48 Threshold("jump-threading-threshold",
49 cl::desc("Max block size to duplicate for jump threading"),
50 cl::init(6), cl::Hidden);
53 // These are at global scope so static functions can use them too.
54 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo;
55 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy;
57 // This is used to keep track of what kind of constant we're currently hoping
59 enum ConstantPreference {
64 /// This pass performs 'jump threading', which looks at blocks that have
65 /// multiple predecessors and multiple successors. If one or more of the
66 /// predecessors of the block can be proven to always jump to one of the
67 /// successors, we forward the edge from the predecessor to the successor by
68 /// duplicating the contents of this block.
70 /// An example of when this can occur is code like this:
77 /// In this case, the unconditional branch at the end of the first if can be
78 /// revectored to the false side of the second if.
80 class JumpThreading : public FunctionPass {
82 TargetLibraryInfo *TLI;
85 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
87 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
89 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet;
91 // RAII helper for updating the recursion stack.
92 struct RecursionSetRemover {
93 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet;
94 std::pair<Value*, BasicBlock*> ThePair;
96 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S,
97 std::pair<Value*, BasicBlock*> P)
98 : TheSet(S), ThePair(P) { }
100 ~RecursionSetRemover() {
101 TheSet.erase(ThePair);
105 static char ID; // Pass identification
106 JumpThreading() : FunctionPass(ID) {
107 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
110 bool runOnFunction(Function &F) override;
112 void getAnalysisUsage(AnalysisUsage &AU) const override {
113 AU.addRequired<LazyValueInfo>();
114 AU.addPreserved<LazyValueInfo>();
115 AU.addRequired<TargetLibraryInfo>();
118 void FindLoopHeaders(Function &F);
119 bool ProcessBlock(BasicBlock *BB);
120 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
122 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
123 const SmallVectorImpl<BasicBlock *> &PredBBs);
125 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
126 PredValueInfo &Result,
127 ConstantPreference Preference,
128 Instruction *CxtI = nullptr);
129 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
130 ConstantPreference Preference,
131 Instruction *CxtI = nullptr);
133 bool ProcessBranchOnPHI(PHINode *PN);
134 bool ProcessBranchOnXOR(BinaryOperator *BO);
136 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
137 bool TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB);
141 char JumpThreading::ID = 0;
142 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
143 "Jump Threading", false, false)
144 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)
145 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
146 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
147 "Jump Threading", false, false)
149 // Public interface to the Jump Threading pass
150 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
152 /// runOnFunction - Top level algorithm.
154 bool JumpThreading::runOnFunction(Function &F) {
155 if (skipOptnoneFunction(F))
158 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
159 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
160 DL = DLP ? &DLP->getDataLayout() : nullptr;
161 TLI = &getAnalysis<TargetLibraryInfo>();
162 LVI = &getAnalysis<LazyValueInfo>();
164 // Remove unreachable blocks from function as they may result in infinite
165 // loop. We do threading if we found something profitable. Jump threading a
166 // branch can create other opportunities. If these opportunities form a cycle
167 // i.e. if any jump treading is undoing previous threading in the path, then
168 // we will loop forever. We take care of this issue by not jump threading for
169 // back edges. This works for normal cases but not for unreachable blocks as
170 // they may have cycle with no back edge.
171 removeUnreachableBlocks(F);
175 bool Changed, EverChanged = false;
178 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
180 // Thread all of the branches we can over this block.
181 while (ProcessBlock(BB))
186 // If the block is trivially dead, zap it. This eliminates the successor
187 // edges which simplifies the CFG.
188 if (pred_begin(BB) == pred_end(BB) &&
189 BB != &BB->getParent()->getEntryBlock()) {
190 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
191 << "' with terminator: " << *BB->getTerminator() << '\n');
192 LoopHeaders.erase(BB);
199 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
201 // Can't thread an unconditional jump, but if the block is "almost
202 // empty", we can replace uses of it with uses of the successor and make
204 if (BI && BI->isUnconditional() &&
205 BB != &BB->getParent()->getEntryBlock() &&
206 // If the terminator is the only non-phi instruction, try to nuke it.
207 BB->getFirstNonPHIOrDbg()->isTerminator()) {
208 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
209 // block, we have to make sure it isn't in the LoopHeaders set. We
210 // reinsert afterward if needed.
211 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
212 BasicBlock *Succ = BI->getSuccessor(0);
214 // FIXME: It is always conservatively correct to drop the info
215 // for a block even if it doesn't get erased. This isn't totally
216 // awesome, but it allows us to use AssertingVH to prevent nasty
217 // dangling pointer issues within LazyValueInfo.
219 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
221 // If we deleted BB and BB was the header of a loop, then the
222 // successor is now the header of the loop.
226 if (ErasedFromLoopHeaders)
227 LoopHeaders.insert(BB);
230 EverChanged |= Changed;
237 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
238 /// thread across it. Stop scanning the block when passing the threshold.
239 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
240 unsigned Threshold) {
241 /// Ignore PHI nodes, these will be flattened when duplication happens.
242 BasicBlock::const_iterator I = BB->getFirstNonPHI();
244 // FIXME: THREADING will delete values that are just used to compute the
245 // branch, so they shouldn't count against the duplication cost.
247 // Sum up the cost of each instruction until we get to the terminator. Don't
248 // include the terminator because the copy won't include it.
250 for (; !isa<TerminatorInst>(I); ++I) {
252 // Stop scanning the block if we've reached the threshold.
253 if (Size > Threshold)
256 // Debugger intrinsics don't incur code size.
257 if (isa<DbgInfoIntrinsic>(I)) continue;
259 // If this is a pointer->pointer bitcast, it is free.
260 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
263 // All other instructions count for at least one unit.
266 // Calls are more expensive. If they are non-intrinsic calls, we model them
267 // as having cost of 4. If they are a non-vector intrinsic, we model them
268 // as having cost of 2 total, and if they are a vector intrinsic, we model
269 // them as having cost 1.
270 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
271 if (CI->cannotDuplicate())
272 // Blocks with NoDuplicate are modelled as having infinite cost, so they
273 // are never duplicated.
275 else if (!isa<IntrinsicInst>(CI))
277 else if (!CI->getType()->isVectorTy())
282 // Threading through a switch statement is particularly profitable. If this
283 // block ends in a switch, decrease its cost to make it more likely to happen.
284 if (isa<SwitchInst>(I))
285 Size = Size > 6 ? Size-6 : 0;
287 // The same holds for indirect branches, but slightly more so.
288 if (isa<IndirectBrInst>(I))
289 Size = Size > 8 ? Size-8 : 0;
294 /// FindLoopHeaders - We do not want jump threading to turn proper loop
295 /// structures into irreducible loops. Doing this breaks up the loop nesting
296 /// hierarchy and pessimizes later transformations. To prevent this from
297 /// happening, we first have to find the loop headers. Here we approximate this
298 /// by finding targets of backedges in the CFG.
300 /// Note that there definitely are cases when we want to allow threading of
301 /// edges across a loop header. For example, threading a jump from outside the
302 /// loop (the preheader) to an exit block of the loop is definitely profitable.
303 /// It is also almost always profitable to thread backedges from within the loop
304 /// to exit blocks, and is often profitable to thread backedges to other blocks
305 /// within the loop (forming a nested loop). This simple analysis is not rich
306 /// enough to track all of these properties and keep it up-to-date as the CFG
307 /// mutates, so we don't allow any of these transformations.
309 void JumpThreading::FindLoopHeaders(Function &F) {
310 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
311 FindFunctionBackedges(F, Edges);
313 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
314 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
317 /// getKnownConstant - Helper method to determine if we can thread over a
318 /// terminator with the given value as its condition, and if so what value to
319 /// use for that. What kind of value this is depends on whether we want an
320 /// integer or a block address, but an undef is always accepted.
321 /// Returns null if Val is null or not an appropriate constant.
322 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
326 // Undef is "known" enough.
327 if (UndefValue *U = dyn_cast<UndefValue>(Val))
330 if (Preference == WantBlockAddress)
331 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
333 return dyn_cast<ConstantInt>(Val);
336 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
337 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
338 /// in any of our predecessors. If so, return the known list of value and pred
339 /// BB in the result vector.
341 /// This returns true if there were any known values.
344 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
345 ConstantPreference Preference,
347 // This method walks up use-def chains recursively. Because of this, we could
348 // get into an infinite loop going around loops in the use-def chain. To
349 // prevent this, keep track of what (value, block) pairs we've already visited
350 // and terminate the search if we loop back to them
351 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
354 // An RAII help to remove this pair from the recursion set once the recursion
355 // stack pops back out again.
356 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
358 // If V is a constant, then it is known in all predecessors.
359 if (Constant *KC = getKnownConstant(V, Preference)) {
360 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
361 Result.push_back(std::make_pair(KC, *PI));
366 // If V is a non-instruction value, or an instruction in a different block,
367 // then it can't be derived from a PHI.
368 Instruction *I = dyn_cast<Instruction>(V);
369 if (!I || I->getParent() != BB) {
371 // Okay, if this is a live-in value, see if it has a known value at the end
372 // of any of our predecessors.
374 // FIXME: This should be an edge property, not a block end property.
375 /// TODO: Per PR2563, we could infer value range information about a
376 /// predecessor based on its terminator.
378 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
379 // "I" is a non-local compare-with-a-constant instruction. This would be
380 // able to handle value inequalities better, for example if the compare is
381 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
382 // Perhaps getConstantOnEdge should be smart enough to do this?
384 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
386 // If the value is known by LazyValueInfo to be a constant in a
387 // predecessor, use that information to try to thread this block.
388 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
389 if (Constant *KC = getKnownConstant(PredCst, Preference))
390 Result.push_back(std::make_pair(KC, P));
393 return !Result.empty();
396 /// If I is a PHI node, then we know the incoming values for any constants.
397 if (PHINode *PN = dyn_cast<PHINode>(I)) {
398 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
399 Value *InVal = PN->getIncomingValue(i);
400 if (Constant *KC = getKnownConstant(InVal, Preference)) {
401 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
403 Constant *CI = LVI->getConstantOnEdge(InVal,
404 PN->getIncomingBlock(i),
406 if (Constant *KC = getKnownConstant(CI, Preference))
407 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
411 return !Result.empty();
414 PredValueInfoTy LHSVals, RHSVals;
416 // Handle some boolean conditions.
417 if (I->getType()->getPrimitiveSizeInBits() == 1) {
418 assert(Preference == WantInteger && "One-bit non-integer type?");
420 // X & false -> false
421 if (I->getOpcode() == Instruction::Or ||
422 I->getOpcode() == Instruction::And) {
423 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
425 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
428 if (LHSVals.empty() && RHSVals.empty())
431 ConstantInt *InterestingVal;
432 if (I->getOpcode() == Instruction::Or)
433 InterestingVal = ConstantInt::getTrue(I->getContext());
435 InterestingVal = ConstantInt::getFalse(I->getContext());
437 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
439 // Scan for the sentinel. If we find an undef, force it to the
440 // interesting value: x|undef -> true and x&undef -> false.
441 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
442 if (LHSVals[i].first == InterestingVal ||
443 isa<UndefValue>(LHSVals[i].first)) {
444 Result.push_back(LHSVals[i]);
445 Result.back().first = InterestingVal;
446 LHSKnownBBs.insert(LHSVals[i].second);
448 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
449 if (RHSVals[i].first == InterestingVal ||
450 isa<UndefValue>(RHSVals[i].first)) {
451 // If we already inferred a value for this block on the LHS, don't
453 if (!LHSKnownBBs.count(RHSVals[i].second)) {
454 Result.push_back(RHSVals[i]);
455 Result.back().first = InterestingVal;
459 return !Result.empty();
462 // Handle the NOT form of XOR.
463 if (I->getOpcode() == Instruction::Xor &&
464 isa<ConstantInt>(I->getOperand(1)) &&
465 cast<ConstantInt>(I->getOperand(1))->isOne()) {
466 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
471 // Invert the known values.
472 for (unsigned i = 0, e = Result.size(); i != e; ++i)
473 Result[i].first = ConstantExpr::getNot(Result[i].first);
478 // Try to simplify some other binary operator values.
479 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
480 assert(Preference != WantBlockAddress
481 && "A binary operator creating a block address?");
482 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
483 PredValueInfoTy LHSVals;
484 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
487 // Try to use constant folding to simplify the binary operator.
488 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
489 Constant *V = LHSVals[i].first;
490 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
492 if (Constant *KC = getKnownConstant(Folded, WantInteger))
493 Result.push_back(std::make_pair(KC, LHSVals[i].second));
497 return !Result.empty();
500 // Handle compare with phi operand, where the PHI is defined in this block.
501 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
502 assert(Preference == WantInteger && "Compares only produce integers");
503 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
504 if (PN && PN->getParent() == BB) {
505 // We can do this simplification if any comparisons fold to true or false.
507 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
508 BasicBlock *PredBB = PN->getIncomingBlock(i);
509 Value *LHS = PN->getIncomingValue(i);
510 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
512 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL);
514 if (!isa<Constant>(RHS))
517 LazyValueInfo::Tristate
518 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
519 cast<Constant>(RHS), PredBB, BB,
521 if (ResT == LazyValueInfo::Unknown)
523 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
526 if (Constant *KC = getKnownConstant(Res, WantInteger))
527 Result.push_back(std::make_pair(KC, PredBB));
530 return !Result.empty();
533 // If comparing a live-in value against a constant, see if we know the
534 // live-in value on any predecessors.
535 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
536 if (!isa<Instruction>(Cmp->getOperand(0)) ||
537 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
538 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
540 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
542 // If the value is known by LazyValueInfo to be a constant in a
543 // predecessor, use that information to try to thread this block.
544 LazyValueInfo::Tristate Res =
545 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
546 RHSCst, P, BB, CxtI ? CxtI : Cmp);
547 if (Res == LazyValueInfo::Unknown)
550 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
551 Result.push_back(std::make_pair(ResC, P));
554 return !Result.empty();
557 // Try to find a constant value for the LHS of a comparison,
558 // and evaluate it statically if we can.
559 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
560 PredValueInfoTy LHSVals;
561 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
564 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
565 Constant *V = LHSVals[i].first;
566 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
568 if (Constant *KC = getKnownConstant(Folded, WantInteger))
569 Result.push_back(std::make_pair(KC, LHSVals[i].second));
572 return !Result.empty();
577 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
578 // Handle select instructions where at least one operand is a known constant
579 // and we can figure out the condition value for any predecessor block.
580 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
581 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
582 PredValueInfoTy Conds;
583 if ((TrueVal || FalseVal) &&
584 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
585 WantInteger, CxtI)) {
586 for (unsigned i = 0, e = Conds.size(); i != e; ++i) {
587 Constant *Cond = Conds[i].first;
589 // Figure out what value to use for the condition.
591 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
593 KnownCond = CI->isOne();
595 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
596 // Either operand will do, so be sure to pick the one that's a known
598 // FIXME: Do this more cleverly if both values are known constants?
599 KnownCond = (TrueVal != nullptr);
602 // See if the select has a known constant value for this predecessor.
603 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
604 Result.push_back(std::make_pair(Val, Conds[i].second));
607 return !Result.empty();
611 // If all else fails, see if LVI can figure out a constant value for us.
612 Constant *CI = LVI->getConstant(V, BB, CxtI);
613 if (Constant *KC = getKnownConstant(CI, Preference)) {
614 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
615 Result.push_back(std::make_pair(KC, *PI));
618 return !Result.empty();
623 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
624 /// in an undefined jump, decide which block is best to revector to.
626 /// Since we can pick an arbitrary destination, we pick the successor with the
627 /// fewest predecessors. This should reduce the in-degree of the others.
629 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
630 TerminatorInst *BBTerm = BB->getTerminator();
631 unsigned MinSucc = 0;
632 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
633 // Compute the successor with the minimum number of predecessors.
634 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
635 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
636 TestBB = BBTerm->getSuccessor(i);
637 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
638 if (NumPreds < MinNumPreds) {
640 MinNumPreds = NumPreds;
647 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
648 if (!BB->hasAddressTaken()) return false;
650 // If the block has its address taken, it may be a tree of dead constants
651 // hanging off of it. These shouldn't keep the block alive.
652 BlockAddress *BA = BlockAddress::get(BB);
653 BA->removeDeadConstantUsers();
654 return !BA->use_empty();
657 /// ProcessBlock - If there are any predecessors whose control can be threaded
658 /// through to a successor, transform them now.
659 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
660 // If the block is trivially dead, just return and let the caller nuke it.
661 // This simplifies other transformations.
662 if (pred_begin(BB) == pred_end(BB) &&
663 BB != &BB->getParent()->getEntryBlock())
666 // If this block has a single predecessor, and if that pred has a single
667 // successor, merge the blocks. This encourages recursive jump threading
668 // because now the condition in this block can be threaded through
669 // predecessors of our predecessor block.
670 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
671 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
672 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
673 // If SinglePred was a loop header, BB becomes one.
674 if (LoopHeaders.erase(SinglePred))
675 LoopHeaders.insert(BB);
677 LVI->eraseBlock(SinglePred);
678 MergeBasicBlockIntoOnlyPred(BB);
684 // What kind of constant we're looking for.
685 ConstantPreference Preference = WantInteger;
687 // Look to see if the terminator is a conditional branch, switch or indirect
688 // branch, if not we can't thread it.
690 Instruction *Terminator = BB->getTerminator();
691 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
692 // Can't thread an unconditional jump.
693 if (BI->isUnconditional()) return false;
694 Condition = BI->getCondition();
695 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
696 Condition = SI->getCondition();
697 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
698 // Can't thread indirect branch with no successors.
699 if (IB->getNumSuccessors() == 0) return false;
700 Condition = IB->getAddress()->stripPointerCasts();
701 Preference = WantBlockAddress;
703 return false; // Must be an invoke.
706 // Run constant folding to see if we can reduce the condition to a simple
708 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
709 Value *SimpleVal = ConstantFoldInstruction(I, DL, TLI);
711 I->replaceAllUsesWith(SimpleVal);
712 I->eraseFromParent();
713 Condition = SimpleVal;
717 // If the terminator is branching on an undef, we can pick any of the
718 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
719 if (isa<UndefValue>(Condition)) {
720 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
722 // Fold the branch/switch.
723 TerminatorInst *BBTerm = BB->getTerminator();
724 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
725 if (i == BestSucc) continue;
726 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
729 DEBUG(dbgs() << " In block '" << BB->getName()
730 << "' folding undef terminator: " << *BBTerm << '\n');
731 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
732 BBTerm->eraseFromParent();
736 // If the terminator of this block is branching on a constant, simplify the
737 // terminator to an unconditional branch. This can occur due to threading in
739 if (getKnownConstant(Condition, Preference)) {
740 DEBUG(dbgs() << " In block '" << BB->getName()
741 << "' folding terminator: " << *BB->getTerminator() << '\n');
743 ConstantFoldTerminator(BB, true);
747 Instruction *CondInst = dyn_cast<Instruction>(Condition);
749 // All the rest of our checks depend on the condition being an instruction.
751 // FIXME: Unify this with code below.
752 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
758 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
759 // For a comparison where the LHS is outside this block, it's possible
760 // that we've branched on it before. Used LVI to see if we can simplify
761 // the branch based on that.
762 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
763 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
764 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
765 if (CondBr && CondConst && CondBr->isConditional() && PI != PE &&
766 (!isa<Instruction>(CondCmp->getOperand(0)) ||
767 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
768 // For predecessor edge, determine if the comparison is true or false
769 // on that edge. If they're all true or all false, we can simplify the
771 // FIXME: We could handle mixed true/false by duplicating code.
772 LazyValueInfo::Tristate Baseline =
773 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
774 CondConst, *PI, BB, CondCmp);
775 if (Baseline != LazyValueInfo::Unknown) {
776 // Check that all remaining incoming values match the first one.
778 LazyValueInfo::Tristate Ret =
779 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
780 CondCmp->getOperand(0), CondConst, *PI, BB,
782 if (Ret != Baseline) break;
785 // If we terminated early, then one of the values didn't match.
787 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
788 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
789 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
790 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
791 CondBr->eraseFromParent();
796 } else if (CondBr && CondConst && CondBr->isConditional()) {
797 // There might be an invairant in the same block with the conditional
798 // that can determine the predicate.
800 LazyValueInfo::Tristate Ret =
801 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
803 if (Ret != LazyValueInfo::Unknown) {
804 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
805 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
806 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
807 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
808 CondBr->eraseFromParent();
813 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB))
817 // Check for some cases that are worth simplifying. Right now we want to look
818 // for loads that are used by a switch or by the condition for the branch. If
819 // we see one, check to see if it's partially redundant. If so, insert a PHI
820 // which can then be used to thread the values.
822 Value *SimplifyValue = CondInst;
823 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
824 if (isa<Constant>(CondCmp->getOperand(1)))
825 SimplifyValue = CondCmp->getOperand(0);
827 // TODO: There are other places where load PRE would be profitable, such as
828 // more complex comparisons.
829 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
830 if (SimplifyPartiallyRedundantLoad(LI))
834 // Handle a variety of cases where we are branching on something derived from
835 // a PHI node in the current block. If we can prove that any predecessors
836 // compute a predictable value based on a PHI node, thread those predecessors.
838 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
841 // If this is an otherwise-unfoldable branch on a phi node in the current
842 // block, see if we can simplify.
843 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
844 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
845 return ProcessBranchOnPHI(PN);
848 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
849 if (CondInst->getOpcode() == Instruction::Xor &&
850 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
851 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
854 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
855 // "(X == 4)", thread through this block.
860 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
861 /// load instruction, eliminate it by replacing it with a PHI node. This is an
862 /// important optimization that encourages jump threading, and needs to be run
863 /// interlaced with other jump threading tasks.
864 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
865 // Don't hack volatile/atomic loads.
866 if (!LI->isSimple()) return false;
868 // If the load is defined in a block with exactly one predecessor, it can't be
869 // partially redundant.
870 BasicBlock *LoadBB = LI->getParent();
871 if (LoadBB->getSinglePredecessor())
874 // If the load is defined in a landing pad, it can't be partially redundant,
875 // because the edges between the invoke and the landing pad cannot have other
876 // instructions between them.
877 if (LoadBB->isLandingPad())
880 Value *LoadedPtr = LI->getOperand(0);
882 // If the loaded operand is defined in the LoadBB, it can't be available.
883 // TODO: Could do simple PHI translation, that would be fun :)
884 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
885 if (PtrOp->getParent() == LoadBB)
888 // Scan a few instructions up from the load, to see if it is obviously live at
889 // the entry to its block.
890 BasicBlock::iterator BBIt = LI;
892 if (Value *AvailableVal =
893 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
894 // If the value if the load is locally available within the block, just use
895 // it. This frequently occurs for reg2mem'd allocas.
896 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
898 // If the returned value is the load itself, replace with an undef. This can
899 // only happen in dead loops.
900 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
901 LI->replaceAllUsesWith(AvailableVal);
902 LI->eraseFromParent();
906 // Otherwise, if we scanned the whole block and got to the top of the block,
907 // we know the block is locally transparent to the load. If not, something
908 // might clobber its value.
909 if (BBIt != LoadBB->begin())
912 // If all of the loads and stores that feed the value have the same AA tags,
913 // then we can propagate them onto any newly inserted loads.
915 LI->getAAMetadata(AATags);
917 SmallPtrSet<BasicBlock*, 8> PredsScanned;
918 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
919 AvailablePredsTy AvailablePreds;
920 BasicBlock *OneUnavailablePred = nullptr;
922 // If we got here, the loaded value is transparent through to the start of the
923 // block. Check to see if it is available in any of the predecessor blocks.
924 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
926 BasicBlock *PredBB = *PI;
928 // If we already scanned this predecessor, skip it.
929 if (!PredsScanned.insert(PredBB))
932 // Scan the predecessor to see if the value is available in the pred.
933 BBIt = PredBB->end();
934 AAMDNodes ThisAATags;
935 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6,
936 nullptr, &ThisAATags);
937 if (!PredAvailable) {
938 OneUnavailablePred = PredBB;
942 // If AA tags disagree or are not present, forget about them.
943 if (AATags != ThisAATags) AATags = AAMDNodes();
945 // If so, this load is partially redundant. Remember this info so that we
946 // can create a PHI node.
947 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
950 // If the loaded value isn't available in any predecessor, it isn't partially
952 if (AvailablePreds.empty()) return false;
954 // Okay, the loaded value is available in at least one (and maybe all!)
955 // predecessors. If the value is unavailable in more than one unique
956 // predecessor, we want to insert a merge block for those common predecessors.
957 // This ensures that we only have to insert one reload, thus not increasing
959 BasicBlock *UnavailablePred = nullptr;
961 // If there is exactly one predecessor where the value is unavailable, the
962 // already computed 'OneUnavailablePred' block is it. If it ends in an
963 // unconditional branch, we know that it isn't a critical edge.
964 if (PredsScanned.size() == AvailablePreds.size()+1 &&
965 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
966 UnavailablePred = OneUnavailablePred;
967 } else if (PredsScanned.size() != AvailablePreds.size()) {
968 // Otherwise, we had multiple unavailable predecessors or we had a critical
969 // edge from the one.
970 SmallVector<BasicBlock*, 8> PredsToSplit;
971 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
973 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
974 AvailablePredSet.insert(AvailablePreds[i].first);
976 // Add all the unavailable predecessors to the PredsToSplit list.
977 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
980 // If the predecessor is an indirect goto, we can't split the edge.
981 if (isa<IndirectBrInst>(P->getTerminator()))
984 if (!AvailablePredSet.count(P))
985 PredsToSplit.push_back(P);
988 // Split them out to their own block.
990 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split", this);
993 // If the value isn't available in all predecessors, then there will be
994 // exactly one where it isn't available. Insert a load on that edge and add
995 // it to the AvailablePreds list.
996 if (UnavailablePred) {
997 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
998 "Can't handle critical edge here!");
999 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
1001 UnavailablePred->getTerminator());
1002 NewVal->setDebugLoc(LI->getDebugLoc());
1004 NewVal->setAAMetadata(AATags);
1006 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1009 // Now we know that each predecessor of this block has a value in
1010 // AvailablePreds, sort them for efficient access as we're walking the preds.
1011 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1013 // Create a PHI node at the start of the block for the PRE'd load value.
1014 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1015 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
1018 PN->setDebugLoc(LI->getDebugLoc());
1020 // Insert new entries into the PHI for each predecessor. A single block may
1021 // have multiple entries here.
1022 for (pred_iterator PI = PB; PI != PE; ++PI) {
1023 BasicBlock *P = *PI;
1024 AvailablePredsTy::iterator I =
1025 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1026 std::make_pair(P, (Value*)nullptr));
1028 assert(I != AvailablePreds.end() && I->first == P &&
1029 "Didn't find entry for predecessor!");
1031 PN->addIncoming(I->second, I->first);
1034 //cerr << "PRE: " << *LI << *PN << "\n";
1036 LI->replaceAllUsesWith(PN);
1037 LI->eraseFromParent();
1042 /// FindMostPopularDest - The specified list contains multiple possible
1043 /// threadable destinations. Pick the one that occurs the most frequently in
1046 FindMostPopularDest(BasicBlock *BB,
1047 const SmallVectorImpl<std::pair<BasicBlock*,
1048 BasicBlock*> > &PredToDestList) {
1049 assert(!PredToDestList.empty());
1051 // Determine popularity. If there are multiple possible destinations, we
1052 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1053 // blocks with known and real destinations to threading undef. We'll handle
1054 // them later if interesting.
1055 DenseMap<BasicBlock*, unsigned> DestPopularity;
1056 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1057 if (PredToDestList[i].second)
1058 DestPopularity[PredToDestList[i].second]++;
1060 // Find the most popular dest.
1061 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1062 BasicBlock *MostPopularDest = DPI->first;
1063 unsigned Popularity = DPI->second;
1064 SmallVector<BasicBlock*, 4> SamePopularity;
1066 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1067 // If the popularity of this entry isn't higher than the popularity we've
1068 // seen so far, ignore it.
1069 if (DPI->second < Popularity)
1071 else if (DPI->second == Popularity) {
1072 // If it is the same as what we've seen so far, keep track of it.
1073 SamePopularity.push_back(DPI->first);
1075 // If it is more popular, remember it.
1076 SamePopularity.clear();
1077 MostPopularDest = DPI->first;
1078 Popularity = DPI->second;
1082 // Okay, now we know the most popular destination. If there is more than one
1083 // destination, we need to determine one. This is arbitrary, but we need
1084 // to make a deterministic decision. Pick the first one that appears in the
1086 if (!SamePopularity.empty()) {
1087 SamePopularity.push_back(MostPopularDest);
1088 TerminatorInst *TI = BB->getTerminator();
1089 for (unsigned i = 0; ; ++i) {
1090 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1092 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1093 TI->getSuccessor(i)) == SamePopularity.end())
1096 MostPopularDest = TI->getSuccessor(i);
1101 // Okay, we have finally picked the most popular destination.
1102 return MostPopularDest;
1105 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1106 ConstantPreference Preference,
1107 Instruction *CxtI) {
1108 // If threading this would thread across a loop header, don't even try to
1110 if (LoopHeaders.count(BB))
1113 PredValueInfoTy PredValues;
1114 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1117 assert(!PredValues.empty() &&
1118 "ComputeValueKnownInPredecessors returned true with no values");
1120 DEBUG(dbgs() << "IN BB: " << *BB;
1121 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1122 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1123 << *PredValues[i].first
1124 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1127 // Decide what we want to thread through. Convert our list of known values to
1128 // a list of known destinations for each pred. This also discards duplicate
1129 // predecessors and keeps track of the undefined inputs (which are represented
1130 // as a null dest in the PredToDestList).
1131 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1132 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1134 BasicBlock *OnlyDest = nullptr;
1135 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1137 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1138 BasicBlock *Pred = PredValues[i].second;
1139 if (!SeenPreds.insert(Pred))
1140 continue; // Duplicate predecessor entry.
1142 // If the predecessor ends with an indirect goto, we can't change its
1144 if (isa<IndirectBrInst>(Pred->getTerminator()))
1147 Constant *Val = PredValues[i].first;
1150 if (isa<UndefValue>(Val))
1152 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1153 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1154 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1155 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1157 assert(isa<IndirectBrInst>(BB->getTerminator())
1158 && "Unexpected terminator");
1159 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1162 // If we have exactly one destination, remember it for efficiency below.
1163 if (PredToDestList.empty())
1165 else if (OnlyDest != DestBB)
1166 OnlyDest = MultipleDestSentinel;
1168 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1171 // If all edges were unthreadable, we fail.
1172 if (PredToDestList.empty())
1175 // Determine which is the most common successor. If we have many inputs and
1176 // this block is a switch, we want to start by threading the batch that goes
1177 // to the most popular destination first. If we only know about one
1178 // threadable destination (the common case) we can avoid this.
1179 BasicBlock *MostPopularDest = OnlyDest;
1181 if (MostPopularDest == MultipleDestSentinel)
1182 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1184 // Now that we know what the most popular destination is, factor all
1185 // predecessors that will jump to it into a single predecessor.
1186 SmallVector<BasicBlock*, 16> PredsToFactor;
1187 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1188 if (PredToDestList[i].second == MostPopularDest) {
1189 BasicBlock *Pred = PredToDestList[i].first;
1191 // This predecessor may be a switch or something else that has multiple
1192 // edges to the block. Factor each of these edges by listing them
1193 // according to # occurrences in PredsToFactor.
1194 TerminatorInst *PredTI = Pred->getTerminator();
1195 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1196 if (PredTI->getSuccessor(i) == BB)
1197 PredsToFactor.push_back(Pred);
1200 // If the threadable edges are branching on an undefined value, we get to pick
1201 // the destination that these predecessors should get to.
1202 if (!MostPopularDest)
1203 MostPopularDest = BB->getTerminator()->
1204 getSuccessor(GetBestDestForJumpOnUndef(BB));
1206 // Ok, try to thread it!
1207 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1210 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1211 /// a PHI node in the current block. See if there are any simplifications we
1212 /// can do based on inputs to the phi node.
1214 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1215 BasicBlock *BB = PN->getParent();
1217 // TODO: We could make use of this to do it once for blocks with common PHI
1219 SmallVector<BasicBlock*, 1> PredBBs;
1222 // If any of the predecessor blocks end in an unconditional branch, we can
1223 // *duplicate* the conditional branch into that block in order to further
1224 // encourage jump threading and to eliminate cases where we have branch on a
1225 // phi of an icmp (branch on icmp is much better).
1226 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1227 BasicBlock *PredBB = PN->getIncomingBlock(i);
1228 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1229 if (PredBr->isUnconditional()) {
1230 PredBBs[0] = PredBB;
1231 // Try to duplicate BB into PredBB.
1232 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1240 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1241 /// a xor instruction in the current block. See if there are any
1242 /// simplifications we can do based on inputs to the xor.
1244 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1245 BasicBlock *BB = BO->getParent();
1247 // If either the LHS or RHS of the xor is a constant, don't do this
1249 if (isa<ConstantInt>(BO->getOperand(0)) ||
1250 isa<ConstantInt>(BO->getOperand(1)))
1253 // If the first instruction in BB isn't a phi, we won't be able to infer
1254 // anything special about any particular predecessor.
1255 if (!isa<PHINode>(BB->front()))
1258 // If we have a xor as the branch input to this block, and we know that the
1259 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1260 // the condition into the predecessor and fix that value to true, saving some
1261 // logical ops on that path and encouraging other paths to simplify.
1263 // This copies something like this:
1266 // %X = phi i1 [1], [%X']
1267 // %Y = icmp eq i32 %A, %B
1268 // %Z = xor i1 %X, %Y
1273 // %Y = icmp ne i32 %A, %B
1276 PredValueInfoTy XorOpValues;
1278 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1280 assert(XorOpValues.empty());
1281 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1287 assert(!XorOpValues.empty() &&
1288 "ComputeValueKnownInPredecessors returned true with no values");
1290 // Scan the information to see which is most popular: true or false. The
1291 // predecessors can be of the set true, false, or undef.
1292 unsigned NumTrue = 0, NumFalse = 0;
1293 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1294 if (isa<UndefValue>(XorOpValues[i].first))
1295 // Ignore undefs for the count.
1297 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1303 // Determine which value to split on, true, false, or undef if neither.
1304 ConstantInt *SplitVal = nullptr;
1305 if (NumTrue > NumFalse)
1306 SplitVal = ConstantInt::getTrue(BB->getContext());
1307 else if (NumTrue != 0 || NumFalse != 0)
1308 SplitVal = ConstantInt::getFalse(BB->getContext());
1310 // Collect all of the blocks that this can be folded into so that we can
1311 // factor this once and clone it once.
1312 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1313 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1314 if (XorOpValues[i].first != SplitVal &&
1315 !isa<UndefValue>(XorOpValues[i].first))
1318 BlocksToFoldInto.push_back(XorOpValues[i].second);
1321 // If we inferred a value for all of the predecessors, then duplication won't
1322 // help us. However, we can just replace the LHS or RHS with the constant.
1323 if (BlocksToFoldInto.size() ==
1324 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1326 // If all preds provide undef, just nuke the xor, because it is undef too.
1327 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1328 BO->eraseFromParent();
1329 } else if (SplitVal->isZero()) {
1330 // If all preds provide 0, replace the xor with the other input.
1331 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1332 BO->eraseFromParent();
1334 // If all preds provide 1, set the computed value to 1.
1335 BO->setOperand(!isLHS, SplitVal);
1341 // Try to duplicate BB into PredBB.
1342 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1346 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1347 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1348 /// NewPred using the entries from OldPred (suitably mapped).
1349 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1350 BasicBlock *OldPred,
1351 BasicBlock *NewPred,
1352 DenseMap<Instruction*, Value*> &ValueMap) {
1353 for (BasicBlock::iterator PNI = PHIBB->begin();
1354 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1355 // Ok, we have a PHI node. Figure out what the incoming value was for the
1357 Value *IV = PN->getIncomingValueForBlock(OldPred);
1359 // Remap the value if necessary.
1360 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1361 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1362 if (I != ValueMap.end())
1366 PN->addIncoming(IV, NewPred);
1370 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1371 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1372 /// across BB. Transform the IR to reflect this change.
1373 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1374 const SmallVectorImpl<BasicBlock*> &PredBBs,
1375 BasicBlock *SuccBB) {
1376 // If threading to the same block as we come from, we would infinite loop.
1378 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1379 << "' - would thread to self!\n");
1383 // If threading this would thread across a loop header, don't thread the edge.
1384 // See the comments above FindLoopHeaders for justifications and caveats.
1385 if (LoopHeaders.count(BB)) {
1386 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1387 << "' to dest BB '" << SuccBB->getName()
1388 << "' - it might create an irreducible loop!\n");
1392 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, Threshold);
1393 if (JumpThreadCost > Threshold) {
1394 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1395 << "' - Cost is too high: " << JumpThreadCost << "\n");
1399 // And finally, do it! Start by factoring the predecessors is needed.
1401 if (PredBBs.size() == 1)
1402 PredBB = PredBBs[0];
1404 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1405 << " common predecessors.\n");
1406 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1409 // And finally, do it!
1410 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1411 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1412 << ", across block:\n "
1415 LVI->threadEdge(PredBB, BB, SuccBB);
1417 // We are going to have to map operands from the original BB block to the new
1418 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1419 // account for entry from PredBB.
1420 DenseMap<Instruction*, Value*> ValueMapping;
1422 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1423 BB->getName()+".thread",
1424 BB->getParent(), BB);
1425 NewBB->moveAfter(PredBB);
1427 BasicBlock::iterator BI = BB->begin();
1428 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1429 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1431 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1432 // mapping and using it to remap operands in the cloned instructions.
1433 for (; !isa<TerminatorInst>(BI); ++BI) {
1434 Instruction *New = BI->clone();
1435 New->setName(BI->getName());
1436 NewBB->getInstList().push_back(New);
1437 ValueMapping[BI] = New;
1439 // Remap operands to patch up intra-block references.
1440 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1441 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1442 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1443 if (I != ValueMapping.end())
1444 New->setOperand(i, I->second);
1448 // We didn't copy the terminator from BB over to NewBB, because there is now
1449 // an unconditional jump to SuccBB. Insert the unconditional jump.
1450 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB);
1451 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1453 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1454 // PHI nodes for NewBB now.
1455 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1457 // If there were values defined in BB that are used outside the block, then we
1458 // now have to update all uses of the value to use either the original value,
1459 // the cloned value, or some PHI derived value. This can require arbitrary
1460 // PHI insertion, of which we are prepared to do, clean these up now.
1461 SSAUpdater SSAUpdate;
1462 SmallVector<Use*, 16> UsesToRename;
1463 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1464 // Scan all uses of this instruction to see if it is used outside of its
1465 // block, and if so, record them in UsesToRename.
1466 for (Use &U : I->uses()) {
1467 Instruction *User = cast<Instruction>(U.getUser());
1468 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1469 if (UserPN->getIncomingBlock(U) == BB)
1471 } else if (User->getParent() == BB)
1474 UsesToRename.push_back(&U);
1477 // If there are no uses outside the block, we're done with this instruction.
1478 if (UsesToRename.empty())
1481 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1483 // We found a use of I outside of BB. Rename all uses of I that are outside
1484 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1485 // with the two values we know.
1486 SSAUpdate.Initialize(I->getType(), I->getName());
1487 SSAUpdate.AddAvailableValue(BB, I);
1488 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1490 while (!UsesToRename.empty())
1491 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1492 DEBUG(dbgs() << "\n");
1496 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1497 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1498 // us to simplify any PHI nodes in BB.
1499 TerminatorInst *PredTerm = PredBB->getTerminator();
1500 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1501 if (PredTerm->getSuccessor(i) == BB) {
1502 BB->removePredecessor(PredBB, true);
1503 PredTerm->setSuccessor(i, NewBB);
1506 // At this point, the IR is fully up to date and consistent. Do a quick scan
1507 // over the new instructions and zap any that are constants or dead. This
1508 // frequently happens because of phi translation.
1509 SimplifyInstructionsInBlock(NewBB, DL, TLI);
1511 // Threaded an edge!
1516 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1517 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1518 /// If we can duplicate the contents of BB up into PredBB do so now, this
1519 /// improves the odds that the branch will be on an analyzable instruction like
1521 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1522 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1523 assert(!PredBBs.empty() && "Can't handle an empty set");
1525 // If BB is a loop header, then duplicating this block outside the loop would
1526 // cause us to transform this into an irreducible loop, don't do this.
1527 // See the comments above FindLoopHeaders for justifications and caveats.
1528 if (LoopHeaders.count(BB)) {
1529 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1530 << "' into predecessor block '" << PredBBs[0]->getName()
1531 << "' - it might create an irreducible loop!\n");
1535 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, Threshold);
1536 if (DuplicationCost > Threshold) {
1537 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1538 << "' - Cost is too high: " << DuplicationCost << "\n");
1542 // And finally, do it! Start by factoring the predecessors is needed.
1544 if (PredBBs.size() == 1)
1545 PredBB = PredBBs[0];
1547 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1548 << " common predecessors.\n");
1549 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1552 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1554 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1555 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1556 << DuplicationCost << " block is:" << *BB << "\n");
1558 // Unless PredBB ends with an unconditional branch, split the edge so that we
1559 // can just clone the bits from BB into the end of the new PredBB.
1560 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1562 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
1563 PredBB = SplitEdge(PredBB, BB, this);
1564 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1567 // We are going to have to map operands from the original BB block into the
1568 // PredBB block. Evaluate PHI nodes in BB.
1569 DenseMap<Instruction*, Value*> ValueMapping;
1571 BasicBlock::iterator BI = BB->begin();
1572 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1573 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1575 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1576 // mapping and using it to remap operands in the cloned instructions.
1577 for (; BI != BB->end(); ++BI) {
1578 Instruction *New = BI->clone();
1580 // Remap operands to patch up intra-block references.
1581 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1582 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1583 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1584 if (I != ValueMapping.end())
1585 New->setOperand(i, I->second);
1588 // If this instruction can be simplified after the operands are updated,
1589 // just use the simplified value instead. This frequently happens due to
1591 if (Value *IV = SimplifyInstruction(New, DL)) {
1593 ValueMapping[BI] = IV;
1595 // Otherwise, insert the new instruction into the block.
1596 New->setName(BI->getName());
1597 PredBB->getInstList().insert(OldPredBranch, New);
1598 ValueMapping[BI] = New;
1602 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1603 // add entries to the PHI nodes for branch from PredBB now.
1604 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1605 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1607 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1610 // If there were values defined in BB that are used outside the block, then we
1611 // now have to update all uses of the value to use either the original value,
1612 // the cloned value, or some PHI derived value. This can require arbitrary
1613 // PHI insertion, of which we are prepared to do, clean these up now.
1614 SSAUpdater SSAUpdate;
1615 SmallVector<Use*, 16> UsesToRename;
1616 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1617 // Scan all uses of this instruction to see if it is used outside of its
1618 // block, and if so, record them in UsesToRename.
1619 for (Use &U : I->uses()) {
1620 Instruction *User = cast<Instruction>(U.getUser());
1621 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1622 if (UserPN->getIncomingBlock(U) == BB)
1624 } else if (User->getParent() == BB)
1627 UsesToRename.push_back(&U);
1630 // If there are no uses outside the block, we're done with this instruction.
1631 if (UsesToRename.empty())
1634 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1636 // We found a use of I outside of BB. Rename all uses of I that are outside
1637 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1638 // with the two values we know.
1639 SSAUpdate.Initialize(I->getType(), I->getName());
1640 SSAUpdate.AddAvailableValue(BB, I);
1641 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1643 while (!UsesToRename.empty())
1644 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1645 DEBUG(dbgs() << "\n");
1648 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1650 BB->removePredecessor(PredBB, true);
1652 // Remove the unconditional branch at the end of the PredBB block.
1653 OldPredBranch->eraseFromParent();
1659 /// TryToUnfoldSelect - Look for blocks of the form
1665 /// %p = phi [%a, %bb] ...
1669 /// And expand the select into a branch structure if one of its arms allows %c
1670 /// to be folded. This later enables threading from bb1 over bb2.
1671 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1672 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1673 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1674 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1676 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1677 CondLHS->getParent() != BB)
1680 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1681 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1682 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1684 // Look if one of the incoming values is a select in the corresponding
1686 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1689 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1690 if (!PredTerm || !PredTerm->isUnconditional())
1693 // Now check if one of the select values would allow us to constant fold the
1694 // terminator in BB. We don't do the transform if both sides fold, those
1695 // cases will be threaded in any case.
1696 LazyValueInfo::Tristate LHSFolds =
1697 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1698 CondRHS, Pred, BB, CondCmp);
1699 LazyValueInfo::Tristate RHSFolds =
1700 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
1701 CondRHS, Pred, BB, CondCmp);
1702 if ((LHSFolds != LazyValueInfo::Unknown ||
1703 RHSFolds != LazyValueInfo::Unknown) &&
1704 LHSFolds != RHSFolds) {
1705 // Expand the select.
1714 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
1715 BB->getParent(), BB);
1716 // Move the unconditional branch to NewBB.
1717 PredTerm->removeFromParent();
1718 NewBB->getInstList().insert(NewBB->end(), PredTerm);
1719 // Create a conditional branch and update PHI nodes.
1720 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
1721 CondLHS->setIncomingValue(I, SI->getFalseValue());
1722 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
1723 // The select is now dead.
1724 SI->eraseFromParent();
1726 // Update any other PHI nodes in BB.
1727 for (BasicBlock::iterator BI = BB->begin();
1728 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
1730 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);