1 //===-- Local.cpp - Functions to perform local transformations ------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This family of functions perform various local transformations to the
11 // program.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "llvm/Transforms/Utils/Local.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/LibCallSemantics.h"
22 #include "llvm/Analysis/MemoryBuiltins.h"
23 #include "llvm/Analysis/ValueTracking.h"
24 #include "llvm/IR/CFG.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DIBuilder.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/DebugInfo.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/Dominators.h"
31 #include "llvm/IR/GetElementPtrTypeIterator.h"
32 #include "llvm/IR/GlobalAlias.h"
33 #include "llvm/IR/GlobalVariable.h"
34 #include "llvm/IR/IRBuilder.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/IntrinsicInst.h"
37 #include "llvm/IR/Intrinsics.h"
38 #include "llvm/IR/MDBuilder.h"
39 #include "llvm/IR/Metadata.h"
40 #include "llvm/IR/Operator.h"
41 #include "llvm/IR/ValueHandle.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/MathExtras.h"
44 #include "llvm/Support/raw_ostream.h"
45 using namespace llvm;
46 
47 #define DEBUG_TYPE "local"
48 
49 STATISTIC(NumRemoved, "Number of unreachable basic blocks removed");
50 
51 //===----------------------------------------------------------------------===//
52 //  Local constant propagation.
53 //
54 
55 /// ConstantFoldTerminator - If a terminator instruction is predicated on a
56 /// constant value, convert it into an unconditional branch to the constant
57 /// destination.  This is a nontrivial operation because the successors of this
58 /// basic block must have their PHI nodes updated.
59 /// Also calls RecursivelyDeleteTriviallyDeadInstructions() on any branch/switch
60 /// conditions and indirectbr addresses this might make dead if
61 /// DeleteDeadConditions is true.
ConstantFoldTerminator(BasicBlock * BB,bool DeleteDeadConditions,const TargetLibraryInfo * TLI)62 bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions,
63                                   const TargetLibraryInfo *TLI) {
64   TerminatorInst *T = BB->getTerminator();
65   IRBuilder<> Builder(T);
66 
67   // Branch - See if we are conditional jumping on constant
68   if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
69     if (BI->isUnconditional()) return false;  // Can't optimize uncond branch
70     BasicBlock *Dest1 = BI->getSuccessor(0);
71     BasicBlock *Dest2 = BI->getSuccessor(1);
72 
73     if (ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition())) {
74       // Are we branching on constant?
75       // YES.  Change to unconditional branch...
76       BasicBlock *Destination = Cond->getZExtValue() ? Dest1 : Dest2;
77       BasicBlock *OldDest     = Cond->getZExtValue() ? Dest2 : Dest1;
78 
79       //cerr << "Function: " << T->getParent()->getParent()
80       //     << "\nRemoving branch from " << T->getParent()
81       //     << "\n\nTo: " << OldDest << endl;
82 
83       // Let the basic block know that we are letting go of it.  Based on this,
84       // it will adjust it's PHI nodes.
85       OldDest->removePredecessor(BB);
86 
87       // Replace the conditional branch with an unconditional one.
88       Builder.CreateBr(Destination);
89       BI->eraseFromParent();
90       return true;
91     }
92 
93     if (Dest2 == Dest1) {       // Conditional branch to same location?
94       // This branch matches something like this:
95       //     br bool %cond, label %Dest, label %Dest
96       // and changes it into:  br label %Dest
97 
98       // Let the basic block know that we are letting go of one copy of it.
99       assert(BI->getParent() && "Terminator not inserted in block!");
100       Dest1->removePredecessor(BI->getParent());
101 
102       // Replace the conditional branch with an unconditional one.
103       Builder.CreateBr(Dest1);
104       Value *Cond = BI->getCondition();
105       BI->eraseFromParent();
106       if (DeleteDeadConditions)
107         RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI);
108       return true;
109     }
110     return false;
111   }
112 
113   if (SwitchInst *SI = dyn_cast<SwitchInst>(T)) {
114     // If we are switching on a constant, we can convert the switch to an
115     // unconditional branch.
116     ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition());
117     BasicBlock *DefaultDest = SI->getDefaultDest();
118     BasicBlock *TheOnlyDest = DefaultDest;
119 
120     // If the default is unreachable, ignore it when searching for TheOnlyDest.
121     if (isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg()) &&
122         SI->getNumCases() > 0) {
123       TheOnlyDest = SI->case_begin().getCaseSuccessor();
124     }
125 
126     // Figure out which case it goes to.
127     for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
128          i != e; ++i) {
129       // Found case matching a constant operand?
130       if (i.getCaseValue() == CI) {
131         TheOnlyDest = i.getCaseSuccessor();
132         break;
133       }
134 
135       // Check to see if this branch is going to the same place as the default
136       // dest.  If so, eliminate it as an explicit compare.
137       if (i.getCaseSuccessor() == DefaultDest) {
138         MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
139         unsigned NCases = SI->getNumCases();
140         // Fold the case metadata into the default if there will be any branches
141         // left, unless the metadata doesn't match the switch.
142         if (NCases > 1 && MD && MD->getNumOperands() == 2 + NCases) {
143           // Collect branch weights into a vector.
144           SmallVector<uint32_t, 8> Weights;
145           for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
146                ++MD_i) {
147             ConstantInt *CI =
148                 mdconst::dyn_extract<ConstantInt>(MD->getOperand(MD_i));
149             assert(CI);
150             Weights.push_back(CI->getValue().getZExtValue());
151           }
152           // Merge weight of this case to the default weight.
153           unsigned idx = i.getCaseIndex();
154           Weights[0] += Weights[idx+1];
155           // Remove weight for this case.
156           std::swap(Weights[idx+1], Weights.back());
157           Weights.pop_back();
158           SI->setMetadata(LLVMContext::MD_prof,
159                           MDBuilder(BB->getContext()).
160                           createBranchWeights(Weights));
161         }
162         // Remove this entry.
163         DefaultDest->removePredecessor(SI->getParent());
164         SI->removeCase(i);
165         --i; --e;
166         continue;
167       }
168 
169       // Otherwise, check to see if the switch only branches to one destination.
170       // We do this by reseting "TheOnlyDest" to null when we find two non-equal
171       // destinations.
172       if (i.getCaseSuccessor() != TheOnlyDest) TheOnlyDest = nullptr;
173     }
174 
175     if (CI && !TheOnlyDest) {
176       // Branching on a constant, but not any of the cases, go to the default
177       // successor.
178       TheOnlyDest = SI->getDefaultDest();
179     }
180 
181     // If we found a single destination that we can fold the switch into, do so
182     // now.
183     if (TheOnlyDest) {
184       // Insert the new branch.
185       Builder.CreateBr(TheOnlyDest);
186       BasicBlock *BB = SI->getParent();
187 
188       // Remove entries from PHI nodes which we no longer branch to...
189       for (unsigned i = 0, e = SI->getNumSuccessors(); i != e; ++i) {
190         // Found case matching a constant operand?
191         BasicBlock *Succ = SI->getSuccessor(i);
192         if (Succ == TheOnlyDest)
193           TheOnlyDest = nullptr; // Don't modify the first branch to TheOnlyDest
194         else
195           Succ->removePredecessor(BB);
196       }
197 
198       // Delete the old switch.
199       Value *Cond = SI->getCondition();
200       SI->eraseFromParent();
201       if (DeleteDeadConditions)
202         RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI);
203       return true;
204     }
205 
206     if (SI->getNumCases() == 1) {
207       // Otherwise, we can fold this switch into a conditional branch
208       // instruction if it has only one non-default destination.
209       SwitchInst::CaseIt FirstCase = SI->case_begin();
210       Value *Cond = Builder.CreateICmpEQ(SI->getCondition(),
211           FirstCase.getCaseValue(), "cond");
212 
213       // Insert the new branch.
214       BranchInst *NewBr = Builder.CreateCondBr(Cond,
215                                                FirstCase.getCaseSuccessor(),
216                                                SI->getDefaultDest());
217       MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
218       if (MD && MD->getNumOperands() == 3) {
219         ConstantInt *SICase =
220             mdconst::dyn_extract<ConstantInt>(MD->getOperand(2));
221         ConstantInt *SIDef =
222             mdconst::dyn_extract<ConstantInt>(MD->getOperand(1));
223         assert(SICase && SIDef);
224         // The TrueWeight should be the weight for the single case of SI.
225         NewBr->setMetadata(LLVMContext::MD_prof,
226                         MDBuilder(BB->getContext()).
227                         createBranchWeights(SICase->getValue().getZExtValue(),
228                                             SIDef->getValue().getZExtValue()));
229       }
230 
231       // Delete the old switch.
232       SI->eraseFromParent();
233       return true;
234     }
235     return false;
236   }
237 
238   if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(T)) {
239     // indirectbr blockaddress(@F, @BB) -> br label @BB
240     if (BlockAddress *BA =
241           dyn_cast<BlockAddress>(IBI->getAddress()->stripPointerCasts())) {
242       BasicBlock *TheOnlyDest = BA->getBasicBlock();
243       // Insert the new branch.
244       Builder.CreateBr(TheOnlyDest);
245 
246       for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
247         if (IBI->getDestination(i) == TheOnlyDest)
248           TheOnlyDest = nullptr;
249         else
250           IBI->getDestination(i)->removePredecessor(IBI->getParent());
251       }
252       Value *Address = IBI->getAddress();
253       IBI->eraseFromParent();
254       if (DeleteDeadConditions)
255         RecursivelyDeleteTriviallyDeadInstructions(Address, TLI);
256 
257       // If we didn't find our destination in the IBI successor list, then we
258       // have undefined behavior.  Replace the unconditional branch with an
259       // 'unreachable' instruction.
260       if (TheOnlyDest) {
261         BB->getTerminator()->eraseFromParent();
262         new UnreachableInst(BB->getContext(), BB);
263       }
264 
265       return true;
266     }
267   }
268 
269   return false;
270 }
271 
272 
273 //===----------------------------------------------------------------------===//
274 //  Local dead code elimination.
275 //
276 
277 /// isInstructionTriviallyDead - Return true if the result produced by the
278 /// instruction is not used, and the instruction has no side effects.
279 ///
isInstructionTriviallyDead(Instruction * I,const TargetLibraryInfo * TLI)280 bool llvm::isInstructionTriviallyDead(Instruction *I,
281                                       const TargetLibraryInfo *TLI) {
282   if (!I->use_empty() || isa<TerminatorInst>(I)) return false;
283 
284   // We don't want the landingpad instruction removed by anything this general.
285   if (isa<LandingPadInst>(I))
286     return false;
287 
288   // We don't want debug info removed by anything this general, unless
289   // debug info is empty.
290   if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(I)) {
291     if (DDI->getAddress())
292       return false;
293     return true;
294   }
295   if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(I)) {
296     if (DVI->getValue())
297       return false;
298     return true;
299   }
300 
301   if (!I->mayHaveSideEffects()) return true;
302 
303   // Special case intrinsics that "may have side effects" but can be deleted
304   // when dead.
305   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
306     // Safe to delete llvm.stacksave if dead.
307     if (II->getIntrinsicID() == Intrinsic::stacksave)
308       return true;
309 
310     // Lifetime intrinsics are dead when their right-hand is undef.
311     if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
312         II->getIntrinsicID() == Intrinsic::lifetime_end)
313       return isa<UndefValue>(II->getArgOperand(1));
314 
315     // Assumptions are dead if their condition is trivially true.
316     if (II->getIntrinsicID() == Intrinsic::assume) {
317       if (ConstantInt *Cond = dyn_cast<ConstantInt>(II->getArgOperand(0)))
318         return !Cond->isZero();
319 
320       return false;
321     }
322   }
323 
324   if (isAllocLikeFn(I, TLI)) return true;
325 
326   if (CallInst *CI = isFreeCall(I, TLI))
327     if (Constant *C = dyn_cast<Constant>(CI->getArgOperand(0)))
328       return C->isNullValue() || isa<UndefValue>(C);
329 
330   return false;
331 }
332 
333 /// RecursivelyDeleteTriviallyDeadInstructions - If the specified value is a
334 /// trivially dead instruction, delete it.  If that makes any of its operands
335 /// trivially dead, delete them too, recursively.  Return true if any
336 /// instructions were deleted.
337 bool
RecursivelyDeleteTriviallyDeadInstructions(Value * V,const TargetLibraryInfo * TLI)338 llvm::RecursivelyDeleteTriviallyDeadInstructions(Value *V,
339                                                  const TargetLibraryInfo *TLI) {
340   Instruction *I = dyn_cast<Instruction>(V);
341   if (!I || !I->use_empty() || !isInstructionTriviallyDead(I, TLI))
342     return false;
343 
344   SmallVector<Instruction*, 16> DeadInsts;
345   DeadInsts.push_back(I);
346 
347   do {
348     I = DeadInsts.pop_back_val();
349 
350     // Null out all of the instruction's operands to see if any operand becomes
351     // dead as we go.
352     for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
353       Value *OpV = I->getOperand(i);
354       I->setOperand(i, nullptr);
355 
356       if (!OpV->use_empty()) continue;
357 
358       // If the operand is an instruction that became dead as we nulled out the
359       // operand, and if it is 'trivially' dead, delete it in a future loop
360       // iteration.
361       if (Instruction *OpI = dyn_cast<Instruction>(OpV))
362         if (isInstructionTriviallyDead(OpI, TLI))
363           DeadInsts.push_back(OpI);
364     }
365 
366     I->eraseFromParent();
367   } while (!DeadInsts.empty());
368 
369   return true;
370 }
371 
372 /// areAllUsesEqual - Check whether the uses of a value are all the same.
373 /// This is similar to Instruction::hasOneUse() except this will also return
374 /// true when there are no uses or multiple uses that all refer to the same
375 /// value.
areAllUsesEqual(Instruction * I)376 static bool areAllUsesEqual(Instruction *I) {
377   Value::user_iterator UI = I->user_begin();
378   Value::user_iterator UE = I->user_end();
379   if (UI == UE)
380     return true;
381 
382   User *TheUse = *UI;
383   for (++UI; UI != UE; ++UI) {
384     if (*UI != TheUse)
385       return false;
386   }
387   return true;
388 }
389 
390 /// RecursivelyDeleteDeadPHINode - If the specified value is an effectively
391 /// dead PHI node, due to being a def-use chain of single-use nodes that
392 /// either forms a cycle or is terminated by a trivially dead instruction,
393 /// delete it.  If that makes any of its operands trivially dead, delete them
394 /// too, recursively.  Return true if a change was made.
RecursivelyDeleteDeadPHINode(PHINode * PN,const TargetLibraryInfo * TLI)395 bool llvm::RecursivelyDeleteDeadPHINode(PHINode *PN,
396                                         const TargetLibraryInfo *TLI) {
397   SmallPtrSet<Instruction*, 4> Visited;
398   for (Instruction *I = PN; areAllUsesEqual(I) && !I->mayHaveSideEffects();
399        I = cast<Instruction>(*I->user_begin())) {
400     if (I->use_empty())
401       return RecursivelyDeleteTriviallyDeadInstructions(I, TLI);
402 
403     // If we find an instruction more than once, we're on a cycle that
404     // won't prove fruitful.
405     if (!Visited.insert(I).second) {
406       // Break the cycle and delete the instruction and its operands.
407       I->replaceAllUsesWith(UndefValue::get(I->getType()));
408       (void)RecursivelyDeleteTriviallyDeadInstructions(I, TLI);
409       return true;
410     }
411   }
412   return false;
413 }
414 
415 /// SimplifyInstructionsInBlock - Scan the specified basic block and try to
416 /// simplify any instructions in it and recursively delete dead instructions.
417 ///
418 /// This returns true if it changed the code, note that it can delete
419 /// instructions in other blocks as well in this block.
SimplifyInstructionsInBlock(BasicBlock * BB,const TargetLibraryInfo * TLI)420 bool llvm::SimplifyInstructionsInBlock(BasicBlock *BB,
421                                        const TargetLibraryInfo *TLI) {
422   bool MadeChange = false;
423 
424 #ifndef NDEBUG
425   // In debug builds, ensure that the terminator of the block is never replaced
426   // or deleted by these simplifications. The idea of simplification is that it
427   // cannot introduce new instructions, and there is no way to replace the
428   // terminator of a block without introducing a new instruction.
429   AssertingVH<Instruction> TerminatorVH(--BB->end());
430 #endif
431 
432   for (BasicBlock::iterator BI = BB->begin(), E = --BB->end(); BI != E; ) {
433     assert(!BI->isTerminator());
434     Instruction *Inst = BI++;
435 
436     WeakVH BIHandle(BI);
437     if (recursivelySimplifyInstruction(Inst, TLI)) {
438       MadeChange = true;
439       if (BIHandle != BI)
440         BI = BB->begin();
441       continue;
442     }
443 
444     MadeChange |= RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI);
445     if (BIHandle != BI)
446       BI = BB->begin();
447   }
448   return MadeChange;
449 }
450 
451 //===----------------------------------------------------------------------===//
452 //  Control Flow Graph Restructuring.
453 //
454 
455 
456 /// RemovePredecessorAndSimplify - Like BasicBlock::removePredecessor, this
457 /// method is called when we're about to delete Pred as a predecessor of BB.  If
458 /// BB contains any PHI nodes, this drops the entries in the PHI nodes for Pred.
459 ///
460 /// Unlike the removePredecessor method, this attempts to simplify uses of PHI
461 /// nodes that collapse into identity values.  For example, if we have:
462 ///   x = phi(1, 0, 0, 0)
463 ///   y = and x, z
464 ///
465 /// .. and delete the predecessor corresponding to the '1', this will attempt to
466 /// recursively fold the and to 0.
RemovePredecessorAndSimplify(BasicBlock * BB,BasicBlock * Pred)467 void llvm::RemovePredecessorAndSimplify(BasicBlock *BB, BasicBlock *Pred) {
468   // This only adjusts blocks with PHI nodes.
469   if (!isa<PHINode>(BB->begin()))
470     return;
471 
472   // Remove the entries for Pred from the PHI nodes in BB, but do not simplify
473   // them down.  This will leave us with single entry phi nodes and other phis
474   // that can be removed.
475   BB->removePredecessor(Pred, true);
476 
477   WeakVH PhiIt = &BB->front();
478   while (PHINode *PN = dyn_cast<PHINode>(PhiIt)) {
479     PhiIt = &*++BasicBlock::iterator(cast<Instruction>(PhiIt));
480     Value *OldPhiIt = PhiIt;
481 
482     if (!recursivelySimplifyInstruction(PN))
483       continue;
484 
485     // If recursive simplification ended up deleting the next PHI node we would
486     // iterate to, then our iterator is invalid, restart scanning from the top
487     // of the block.
488     if (PhiIt != OldPhiIt) PhiIt = &BB->front();
489   }
490 }
491 
492 
493 /// MergeBasicBlockIntoOnlyPred - DestBB is a block with one predecessor and its
494 /// predecessor is known to have one successor (DestBB!).  Eliminate the edge
495 /// between them, moving the instructions in the predecessor into DestBB and
496 /// deleting the predecessor block.
497 ///
MergeBasicBlockIntoOnlyPred(BasicBlock * DestBB,DominatorTree * DT)498 void llvm::MergeBasicBlockIntoOnlyPred(BasicBlock *DestBB, DominatorTree *DT) {
499   // If BB has single-entry PHI nodes, fold them.
500   while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) {
501     Value *NewVal = PN->getIncomingValue(0);
502     // Replace self referencing PHI with undef, it must be dead.
503     if (NewVal == PN) NewVal = UndefValue::get(PN->getType());
504     PN->replaceAllUsesWith(NewVal);
505     PN->eraseFromParent();
506   }
507 
508   BasicBlock *PredBB = DestBB->getSinglePredecessor();
509   assert(PredBB && "Block doesn't have a single predecessor!");
510 
511   // Zap anything that took the address of DestBB.  Not doing this will give the
512   // address an invalid value.
513   if (DestBB->hasAddressTaken()) {
514     BlockAddress *BA = BlockAddress::get(DestBB);
515     Constant *Replacement =
516       ConstantInt::get(llvm::Type::getInt32Ty(BA->getContext()), 1);
517     BA->replaceAllUsesWith(ConstantExpr::getIntToPtr(Replacement,
518                                                      BA->getType()));
519     BA->destroyConstant();
520   }
521 
522   // Anything that branched to PredBB now branches to DestBB.
523   PredBB->replaceAllUsesWith(DestBB);
524 
525   // Splice all the instructions from PredBB to DestBB.
526   PredBB->getTerminator()->eraseFromParent();
527   DestBB->getInstList().splice(DestBB->begin(), PredBB->getInstList());
528 
529   // If the PredBB is the entry block of the function, move DestBB up to
530   // become the entry block after we erase PredBB.
531   if (PredBB == &DestBB->getParent()->getEntryBlock())
532     DestBB->moveAfter(PredBB);
533 
534   if (DT) {
535     BasicBlock *PredBBIDom = DT->getNode(PredBB)->getIDom()->getBlock();
536     DT->changeImmediateDominator(DestBB, PredBBIDom);
537     DT->eraseNode(PredBB);
538   }
539   // Nuke BB.
540   PredBB->eraseFromParent();
541 }
542 
543 /// CanMergeValues - Return true if we can choose one of these values to use
544 /// in place of the other. Note that we will always choose the non-undef
545 /// value to keep.
CanMergeValues(Value * First,Value * Second)546 static bool CanMergeValues(Value *First, Value *Second) {
547   return First == Second || isa<UndefValue>(First) || isa<UndefValue>(Second);
548 }
549 
550 /// CanPropagatePredecessorsForPHIs - Return true if we can fold BB, an
551 /// almost-empty BB ending in an unconditional branch to Succ, into Succ.
552 ///
553 /// Assumption: Succ is the single successor for BB.
554 ///
CanPropagatePredecessorsForPHIs(BasicBlock * BB,BasicBlock * Succ)555 static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) {
556   assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!");
557 
558   DEBUG(dbgs() << "Looking to fold " << BB->getName() << " into "
559         << Succ->getName() << "\n");
560   // Shortcut, if there is only a single predecessor it must be BB and merging
561   // is always safe
562   if (Succ->getSinglePredecessor()) return true;
563 
564   // Make a list of the predecessors of BB
565   SmallPtrSet<BasicBlock*, 16> BBPreds(pred_begin(BB), pred_end(BB));
566 
567   // Look at all the phi nodes in Succ, to see if they present a conflict when
568   // merging these blocks
569   for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
570     PHINode *PN = cast<PHINode>(I);
571 
572     // If the incoming value from BB is again a PHINode in
573     // BB which has the same incoming value for *PI as PN does, we can
574     // merge the phi nodes and then the blocks can still be merged
575     PHINode *BBPN = dyn_cast<PHINode>(PN->getIncomingValueForBlock(BB));
576     if (BBPN && BBPN->getParent() == BB) {
577       for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) {
578         BasicBlock *IBB = PN->getIncomingBlock(PI);
579         if (BBPreds.count(IBB) &&
580             !CanMergeValues(BBPN->getIncomingValueForBlock(IBB),
581                             PN->getIncomingValue(PI))) {
582           DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in "
583                 << Succ->getName() << " is conflicting with "
584                 << BBPN->getName() << " with regard to common predecessor "
585                 << IBB->getName() << "\n");
586           return false;
587         }
588       }
589     } else {
590       Value* Val = PN->getIncomingValueForBlock(BB);
591       for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) {
592         // See if the incoming value for the common predecessor is equal to the
593         // one for BB, in which case this phi node will not prevent the merging
594         // of the block.
595         BasicBlock *IBB = PN->getIncomingBlock(PI);
596         if (BBPreds.count(IBB) &&
597             !CanMergeValues(Val, PN->getIncomingValue(PI))) {
598           DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in "
599                 << Succ->getName() << " is conflicting with regard to common "
600                 << "predecessor " << IBB->getName() << "\n");
601           return false;
602         }
603       }
604     }
605   }
606 
607   return true;
608 }
609 
610 typedef SmallVector<BasicBlock *, 16> PredBlockVector;
611 typedef DenseMap<BasicBlock *, Value *> IncomingValueMap;
612 
613 /// \brief Determines the value to use as the phi node input for a block.
614 ///
615 /// Select between \p OldVal any value that we know flows from \p BB
616 /// to a particular phi on the basis of which one (if either) is not
617 /// undef. Update IncomingValues based on the selected value.
618 ///
619 /// \param OldVal The value we are considering selecting.
620 /// \param BB The block that the value flows in from.
621 /// \param IncomingValues A map from block-to-value for other phi inputs
622 /// that we have examined.
623 ///
624 /// \returns the selected value.
selectIncomingValueForBlock(Value * OldVal,BasicBlock * BB,IncomingValueMap & IncomingValues)625 static Value *selectIncomingValueForBlock(Value *OldVal, BasicBlock *BB,
626                                           IncomingValueMap &IncomingValues) {
627   if (!isa<UndefValue>(OldVal)) {
628     assert((!IncomingValues.count(BB) ||
629             IncomingValues.find(BB)->second == OldVal) &&
630            "Expected OldVal to match incoming value from BB!");
631 
632     IncomingValues.insert(std::make_pair(BB, OldVal));
633     return OldVal;
634   }
635 
636   IncomingValueMap::const_iterator It = IncomingValues.find(BB);
637   if (It != IncomingValues.end()) return It->second;
638 
639   return OldVal;
640 }
641 
642 /// \brief Create a map from block to value for the operands of a
643 /// given phi.
644 ///
645 /// Create a map from block to value for each non-undef value flowing
646 /// into \p PN.
647 ///
648 /// \param PN The phi we are collecting the map for.
649 /// \param IncomingValues [out] The map from block to value for this phi.
gatherIncomingValuesToPhi(PHINode * PN,IncomingValueMap & IncomingValues)650 static void gatherIncomingValuesToPhi(PHINode *PN,
651                                       IncomingValueMap &IncomingValues) {
652   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
653     BasicBlock *BB = PN->getIncomingBlock(i);
654     Value *V = PN->getIncomingValue(i);
655 
656     if (!isa<UndefValue>(V))
657       IncomingValues.insert(std::make_pair(BB, V));
658   }
659 }
660 
661 /// \brief Replace the incoming undef values to a phi with the values
662 /// from a block-to-value map.
663 ///
664 /// \param PN The phi we are replacing the undefs in.
665 /// \param IncomingValues A map from block to value.
replaceUndefValuesInPhi(PHINode * PN,const IncomingValueMap & IncomingValues)666 static void replaceUndefValuesInPhi(PHINode *PN,
667                                     const IncomingValueMap &IncomingValues) {
668   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
669     Value *V = PN->getIncomingValue(i);
670 
671     if (!isa<UndefValue>(V)) continue;
672 
673     BasicBlock *BB = PN->getIncomingBlock(i);
674     IncomingValueMap::const_iterator It = IncomingValues.find(BB);
675     if (It == IncomingValues.end()) continue;
676 
677     PN->setIncomingValue(i, It->second);
678   }
679 }
680 
681 /// \brief Replace a value flowing from a block to a phi with
682 /// potentially multiple instances of that value flowing from the
683 /// block's predecessors to the phi.
684 ///
685 /// \param BB The block with the value flowing into the phi.
686 /// \param BBPreds The predecessors of BB.
687 /// \param PN The phi that we are updating.
redirectValuesFromPredecessorsToPhi(BasicBlock * BB,const PredBlockVector & BBPreds,PHINode * PN)688 static void redirectValuesFromPredecessorsToPhi(BasicBlock *BB,
689                                                 const PredBlockVector &BBPreds,
690                                                 PHINode *PN) {
691   Value *OldVal = PN->removeIncomingValue(BB, false);
692   assert(OldVal && "No entry in PHI for Pred BB!");
693 
694   IncomingValueMap IncomingValues;
695 
696   // We are merging two blocks - BB, and the block containing PN - and
697   // as a result we need to redirect edges from the predecessors of BB
698   // to go to the block containing PN, and update PN
699   // accordingly. Since we allow merging blocks in the case where the
700   // predecessor and successor blocks both share some predecessors,
701   // and where some of those common predecessors might have undef
702   // values flowing into PN, we want to rewrite those values to be
703   // consistent with the non-undef values.
704 
705   gatherIncomingValuesToPhi(PN, IncomingValues);
706 
707   // If this incoming value is one of the PHI nodes in BB, the new entries
708   // in the PHI node are the entries from the old PHI.
709   if (isa<PHINode>(OldVal) && cast<PHINode>(OldVal)->getParent() == BB) {
710     PHINode *OldValPN = cast<PHINode>(OldVal);
711     for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) {
712       // Note that, since we are merging phi nodes and BB and Succ might
713       // have common predecessors, we could end up with a phi node with
714       // identical incoming branches. This will be cleaned up later (and
715       // will trigger asserts if we try to clean it up now, without also
716       // simplifying the corresponding conditional branch).
717       BasicBlock *PredBB = OldValPN->getIncomingBlock(i);
718       Value *PredVal = OldValPN->getIncomingValue(i);
719       Value *Selected = selectIncomingValueForBlock(PredVal, PredBB,
720                                                     IncomingValues);
721 
722       // And add a new incoming value for this predecessor for the
723       // newly retargeted branch.
724       PN->addIncoming(Selected, PredBB);
725     }
726   } else {
727     for (unsigned i = 0, e = BBPreds.size(); i != e; ++i) {
728       // Update existing incoming values in PN for this
729       // predecessor of BB.
730       BasicBlock *PredBB = BBPreds[i];
731       Value *Selected = selectIncomingValueForBlock(OldVal, PredBB,
732                                                     IncomingValues);
733 
734       // And add a new incoming value for this predecessor for the
735       // newly retargeted branch.
736       PN->addIncoming(Selected, PredBB);
737     }
738   }
739 
740   replaceUndefValuesInPhi(PN, IncomingValues);
741 }
742 
743 /// TryToSimplifyUncondBranchFromEmptyBlock - BB is known to contain an
744 /// unconditional branch, and contains no instructions other than PHI nodes,
745 /// potential side-effect free intrinsics and the branch.  If possible,
746 /// eliminate BB by rewriting all the predecessors to branch to the successor
747 /// block and return true.  If we can't transform, return false.
TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock * BB)748 bool llvm::TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB) {
749   assert(BB != &BB->getParent()->getEntryBlock() &&
750          "TryToSimplifyUncondBranchFromEmptyBlock called on entry block!");
751 
752   // We can't eliminate infinite loops.
753   BasicBlock *Succ = cast<BranchInst>(BB->getTerminator())->getSuccessor(0);
754   if (BB == Succ) return false;
755 
756   // Check to see if merging these blocks would cause conflicts for any of the
757   // phi nodes in BB or Succ. If not, we can safely merge.
758   if (!CanPropagatePredecessorsForPHIs(BB, Succ)) return false;
759 
760   // Check for cases where Succ has multiple predecessors and a PHI node in BB
761   // has uses which will not disappear when the PHI nodes are merged.  It is
762   // possible to handle such cases, but difficult: it requires checking whether
763   // BB dominates Succ, which is non-trivial to calculate in the case where
764   // Succ has multiple predecessors.  Also, it requires checking whether
765   // constructing the necessary self-referential PHI node doesn't introduce any
766   // conflicts; this isn't too difficult, but the previous code for doing this
767   // was incorrect.
768   //
769   // Note that if this check finds a live use, BB dominates Succ, so BB is
770   // something like a loop pre-header (or rarely, a part of an irreducible CFG);
771   // folding the branch isn't profitable in that case anyway.
772   if (!Succ->getSinglePredecessor()) {
773     BasicBlock::iterator BBI = BB->begin();
774     while (isa<PHINode>(*BBI)) {
775       for (Use &U : BBI->uses()) {
776         if (PHINode* PN = dyn_cast<PHINode>(U.getUser())) {
777           if (PN->getIncomingBlock(U) != BB)
778             return false;
779         } else {
780           return false;
781         }
782       }
783       ++BBI;
784     }
785   }
786 
787   DEBUG(dbgs() << "Killing Trivial BB: \n" << *BB);
788 
789   if (isa<PHINode>(Succ->begin())) {
790     // If there is more than one pred of succ, and there are PHI nodes in
791     // the successor, then we need to add incoming edges for the PHI nodes
792     //
793     const PredBlockVector BBPreds(pred_begin(BB), pred_end(BB));
794 
795     // Loop over all of the PHI nodes in the successor of BB.
796     for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
797       PHINode *PN = cast<PHINode>(I);
798 
799       redirectValuesFromPredecessorsToPhi(BB, BBPreds, PN);
800     }
801   }
802 
803   if (Succ->getSinglePredecessor()) {
804     // BB is the only predecessor of Succ, so Succ will end up with exactly
805     // the same predecessors BB had.
806 
807     // Copy over any phi, debug or lifetime instruction.
808     BB->getTerminator()->eraseFromParent();
809     Succ->getInstList().splice(Succ->getFirstNonPHI(), BB->getInstList());
810   } else {
811     while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) {
812       // We explicitly check for such uses in CanPropagatePredecessorsForPHIs.
813       assert(PN->use_empty() && "There shouldn't be any uses here!");
814       PN->eraseFromParent();
815     }
816   }
817 
818   // Everything that jumped to BB now goes to Succ.
819   BB->replaceAllUsesWith(Succ);
820   if (!Succ->hasName()) Succ->takeName(BB);
821   BB->eraseFromParent();              // Delete the old basic block.
822   return true;
823 }
824 
825 /// EliminateDuplicatePHINodes - Check for and eliminate duplicate PHI
826 /// nodes in this block. This doesn't try to be clever about PHI nodes
827 /// which differ only in the order of the incoming values, but instcombine
828 /// orders them so it usually won't matter.
829 ///
EliminateDuplicatePHINodes(BasicBlock * BB)830 bool llvm::EliminateDuplicatePHINodes(BasicBlock *BB) {
831   bool Changed = false;
832 
833   // This implementation doesn't currently consider undef operands
834   // specially. Theoretically, two phis which are identical except for
835   // one having an undef where the other doesn't could be collapsed.
836 
837   // Map from PHI hash values to PHI nodes. If multiple PHIs have
838   // the same hash value, the element is the first PHI in the
839   // linked list in CollisionMap.
840   DenseMap<uintptr_t, PHINode *> HashMap;
841 
842   // Maintain linked lists of PHI nodes with common hash values.
843   DenseMap<PHINode *, PHINode *> CollisionMap;
844 
845   // Examine each PHI.
846   for (BasicBlock::iterator I = BB->begin();
847        PHINode *PN = dyn_cast<PHINode>(I++); ) {
848     // Compute a hash value on the operands. Instcombine will likely have sorted
849     // them, which helps expose duplicates, but we have to check all the
850     // operands to be safe in case instcombine hasn't run.
851     uintptr_t Hash = 0;
852     // This hash algorithm is quite weak as hash functions go, but it seems
853     // to do a good enough job for this particular purpose, and is very quick.
854     for (User::op_iterator I = PN->op_begin(), E = PN->op_end(); I != E; ++I) {
855       Hash ^= reinterpret_cast<uintptr_t>(static_cast<Value *>(*I));
856       Hash = (Hash << 7) | (Hash >> (sizeof(uintptr_t) * CHAR_BIT - 7));
857     }
858     for (PHINode::block_iterator I = PN->block_begin(), E = PN->block_end();
859          I != E; ++I) {
860       Hash ^= reinterpret_cast<uintptr_t>(static_cast<BasicBlock *>(*I));
861       Hash = (Hash << 7) | (Hash >> (sizeof(uintptr_t) * CHAR_BIT - 7));
862     }
863     // Avoid colliding with the DenseMap sentinels ~0 and ~0-1.
864     Hash >>= 1;
865     // If we've never seen this hash value before, it's a unique PHI.
866     std::pair<DenseMap<uintptr_t, PHINode *>::iterator, bool> Pair =
867       HashMap.insert(std::make_pair(Hash, PN));
868     if (Pair.second) continue;
869     // Otherwise it's either a duplicate or a hash collision.
870     for (PHINode *OtherPN = Pair.first->second; ; ) {
871       if (OtherPN->isIdenticalTo(PN)) {
872         // A duplicate. Replace this PHI with its duplicate.
873         PN->replaceAllUsesWith(OtherPN);
874         PN->eraseFromParent();
875         Changed = true;
876         break;
877       }
878       // A non-duplicate hash collision.
879       DenseMap<PHINode *, PHINode *>::iterator I = CollisionMap.find(OtherPN);
880       if (I == CollisionMap.end()) {
881         // Set this PHI to be the head of the linked list of colliding PHIs.
882         PHINode *Old = Pair.first->second;
883         Pair.first->second = PN;
884         CollisionMap[PN] = Old;
885         break;
886       }
887       // Proceed to the next PHI in the list.
888       OtherPN = I->second;
889     }
890   }
891 
892   return Changed;
893 }
894 
895 /// enforceKnownAlignment - If the specified pointer points to an object that
896 /// we control, modify the object's alignment to PrefAlign. This isn't
897 /// often possible though. If alignment is important, a more reliable approach
898 /// is to simply align all global variables and allocation instructions to
899 /// their preferred alignment from the beginning.
900 ///
enforceKnownAlignment(Value * V,unsigned Align,unsigned PrefAlign,const DataLayout & DL)901 static unsigned enforceKnownAlignment(Value *V, unsigned Align,
902                                       unsigned PrefAlign,
903                                       const DataLayout &DL) {
904   V = V->stripPointerCasts();
905 
906   if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
907     // If the preferred alignment is greater than the natural stack alignment
908     // then don't round up. This avoids dynamic stack realignment.
909     if (DL.exceedsNaturalStackAlignment(PrefAlign))
910       return Align;
911     // If there is a requested alignment and if this is an alloca, round up.
912     if (AI->getAlignment() >= PrefAlign)
913       return AI->getAlignment();
914     AI->setAlignment(PrefAlign);
915     return PrefAlign;
916   }
917 
918   if (auto *GO = dyn_cast<GlobalObject>(V)) {
919     // If there is a large requested alignment and we can, bump up the alignment
920     // of the global.
921     if (GO->isDeclaration())
922       return Align;
923     // If the memory we set aside for the global may not be the memory used by
924     // the final program then it is impossible for us to reliably enforce the
925     // preferred alignment.
926     if (GO->isWeakForLinker())
927       return Align;
928 
929     if (GO->getAlignment() >= PrefAlign)
930       return GO->getAlignment();
931     // We can only increase the alignment of the global if it has no alignment
932     // specified or if it is not assigned a section.  If it is assigned a
933     // section, the global could be densely packed with other objects in the
934     // section, increasing the alignment could cause padding issues.
935     if (!GO->hasSection() || GO->getAlignment() == 0)
936       GO->setAlignment(PrefAlign);
937     return GO->getAlignment();
938   }
939 
940   return Align;
941 }
942 
943 /// getOrEnforceKnownAlignment - If the specified pointer has an alignment that
944 /// we can determine, return it, otherwise return 0.  If PrefAlign is specified,
945 /// and it is more than the alignment of the ultimate object, see if we can
946 /// increase the alignment of the ultimate object, making this check succeed.
getOrEnforceKnownAlignment(Value * V,unsigned PrefAlign,const DataLayout & DL,const Instruction * CxtI,AssumptionCache * AC,const DominatorTree * DT)947 unsigned llvm::getOrEnforceKnownAlignment(Value *V, unsigned PrefAlign,
948                                           const DataLayout &DL,
949                                           const Instruction *CxtI,
950                                           AssumptionCache *AC,
951                                           const DominatorTree *DT) {
952   assert(V->getType()->isPointerTy() &&
953          "getOrEnforceKnownAlignment expects a pointer!");
954   unsigned BitWidth = DL.getPointerTypeSizeInBits(V->getType());
955 
956   APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
957   computeKnownBits(V, KnownZero, KnownOne, DL, 0, AC, CxtI, DT);
958   unsigned TrailZ = KnownZero.countTrailingOnes();
959 
960   // Avoid trouble with ridiculously large TrailZ values, such as
961   // those computed from a null pointer.
962   TrailZ = std::min(TrailZ, unsigned(sizeof(unsigned) * CHAR_BIT - 1));
963 
964   unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
965 
966   // LLVM doesn't support alignments larger than this currently.
967   Align = std::min(Align, +Value::MaximumAlignment);
968 
969   if (PrefAlign > Align)
970     Align = enforceKnownAlignment(V, Align, PrefAlign, DL);
971 
972   // We don't need to make any adjustment.
973   return Align;
974 }
975 
976 ///===---------------------------------------------------------------------===//
977 ///  Dbg Intrinsic utilities
978 ///
979 
980 /// See if there is a dbg.value intrinsic for DIVar before I.
LdStHasDebugValue(DIVariable & DIVar,Instruction * I)981 static bool LdStHasDebugValue(DIVariable &DIVar, Instruction *I) {
982   // Since we can't guarantee that the original dbg.declare instrinsic
983   // is removed by LowerDbgDeclare(), we need to make sure that we are
984   // not inserting the same dbg.value intrinsic over and over.
985   llvm::BasicBlock::InstListType::iterator PrevI(I);
986   if (PrevI != I->getParent()->getInstList().begin()) {
987     --PrevI;
988     if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(PrevI))
989       if (DVI->getValue() == I->getOperand(0) &&
990           DVI->getOffset() == 0 &&
991           DVI->getVariable() == DIVar)
992         return true;
993   }
994   return false;
995 }
996 
997 /// Inserts a llvm.dbg.value intrinsic before a store to an alloca'd value
998 /// that has an associated llvm.dbg.decl intrinsic.
ConvertDebugDeclareToDebugValue(DbgDeclareInst * DDI,StoreInst * SI,DIBuilder & Builder)999 bool llvm::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI,
1000                                            StoreInst *SI, DIBuilder &Builder) {
1001   DIVariable DIVar = DDI->getVariable();
1002   DIExpression DIExpr = DDI->getExpression();
1003   if (!DIVar)
1004     return false;
1005 
1006   if (LdStHasDebugValue(DIVar, SI))
1007     return true;
1008 
1009   // If an argument is zero extended then use argument directly. The ZExt
1010   // may be zapped by an optimization pass in future.
1011   Argument *ExtendedArg = nullptr;
1012   if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1013     ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0));
1014   if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1015     ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0));
1016   if (ExtendedArg)
1017     Builder.insertDbgValueIntrinsic(ExtendedArg, 0, DIVar, DIExpr,
1018                                     DDI->getDebugLoc(), SI);
1019   else
1020     Builder.insertDbgValueIntrinsic(SI->getOperand(0), 0, DIVar, DIExpr,
1021                                     DDI->getDebugLoc(), SI);
1022   return true;
1023 }
1024 
1025 /// Inserts a llvm.dbg.value intrinsic before a load of an alloca'd value
1026 /// that has an associated llvm.dbg.decl intrinsic.
ConvertDebugDeclareToDebugValue(DbgDeclareInst * DDI,LoadInst * LI,DIBuilder & Builder)1027 bool llvm::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI,
1028                                            LoadInst *LI, DIBuilder &Builder) {
1029   DIVariable DIVar = DDI->getVariable();
1030   DIExpression DIExpr = DDI->getExpression();
1031   if (!DIVar)
1032     return false;
1033 
1034   if (LdStHasDebugValue(DIVar, LI))
1035     return true;
1036 
1037   Builder.insertDbgValueIntrinsic(LI->getOperand(0), 0, DIVar, DIExpr,
1038                                   DDI->getDebugLoc(), LI);
1039   return true;
1040 }
1041 
1042 /// Determine whether this alloca is either a VLA or an array.
isArray(AllocaInst * AI)1043 static bool isArray(AllocaInst *AI) {
1044   return AI->isArrayAllocation() ||
1045     AI->getType()->getElementType()->isArrayTy();
1046 }
1047 
1048 /// LowerDbgDeclare - Lowers llvm.dbg.declare intrinsics into appropriate set
1049 /// of llvm.dbg.value intrinsics.
LowerDbgDeclare(Function & F)1050 bool llvm::LowerDbgDeclare(Function &F) {
1051   DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false);
1052   SmallVector<DbgDeclareInst *, 4> Dbgs;
1053   for (auto &FI : F)
1054     for (BasicBlock::iterator BI : FI)
1055       if (auto DDI = dyn_cast<DbgDeclareInst>(BI))
1056         Dbgs.push_back(DDI);
1057 
1058   if (Dbgs.empty())
1059     return false;
1060 
1061   for (auto &I : Dbgs) {
1062     DbgDeclareInst *DDI = I;
1063     AllocaInst *AI = dyn_cast_or_null<AllocaInst>(DDI->getAddress());
1064     // If this is an alloca for a scalar variable, insert a dbg.value
1065     // at each load and store to the alloca and erase the dbg.declare.
1066     // The dbg.values allow tracking a variable even if it is not
1067     // stored on the stack, while the dbg.declare can only describe
1068     // the stack slot (and at a lexical-scope granularity). Later
1069     // passes will attempt to elide the stack slot.
1070     if (AI && !isArray(AI)) {
1071       for (User *U : AI->users())
1072         if (StoreInst *SI = dyn_cast<StoreInst>(U))
1073           ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1074         else if (LoadInst *LI = dyn_cast<LoadInst>(U))
1075           ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1076         else if (CallInst *CI = dyn_cast<CallInst>(U)) {
1077           // This is a call by-value or some other instruction that
1078           // takes a pointer to the variable. Insert a *value*
1079           // intrinsic that describes the alloca.
1080           DIB.insertDbgValueIntrinsic(AI, 0, DIVariable(DDI->getVariable()),
1081                                       DIExpression(DDI->getExpression()),
1082                                       DDI->getDebugLoc(), CI);
1083         }
1084       DDI->eraseFromParent();
1085     }
1086   }
1087   return true;
1088 }
1089 
1090 /// FindAllocaDbgDeclare - Finds the llvm.dbg.declare intrinsic describing the
1091 /// alloca 'V', if any.
FindAllocaDbgDeclare(Value * V)1092 DbgDeclareInst *llvm::FindAllocaDbgDeclare(Value *V) {
1093   if (auto *L = LocalAsMetadata::getIfExists(V))
1094     if (auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L))
1095       for (User *U : MDV->users())
1096         if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
1097           return DDI;
1098 
1099   return nullptr;
1100 }
1101 
replaceDbgDeclareForAlloca(AllocaInst * AI,Value * NewAllocaAddress,DIBuilder & Builder,bool Deref)1102 bool llvm::replaceDbgDeclareForAlloca(AllocaInst *AI, Value *NewAllocaAddress,
1103                                       DIBuilder &Builder, bool Deref) {
1104   DbgDeclareInst *DDI = FindAllocaDbgDeclare(AI);
1105   if (!DDI)
1106     return false;
1107   DebugLoc Loc = DDI->getDebugLoc();
1108   DIVariable DIVar = DDI->getVariable();
1109   DIExpression DIExpr = DDI->getExpression();
1110   if (!DIVar)
1111     return false;
1112 
1113   if (Deref) {
1114     // Create a copy of the original DIDescriptor for user variable, prepending
1115     // "deref" operation to a list of address elements, as new llvm.dbg.declare
1116     // will take a value storing address of the memory for variable, not
1117     // alloca itself.
1118     SmallVector<uint64_t, 4> NewDIExpr;
1119     NewDIExpr.push_back(dwarf::DW_OP_deref);
1120     if (DIExpr)
1121       NewDIExpr.append(DIExpr->elements_begin(), DIExpr->elements_end());
1122     DIExpr = Builder.createExpression(NewDIExpr);
1123   }
1124 
1125   // Insert llvm.dbg.declare in the same basic block as the original alloca,
1126   // and remove old llvm.dbg.declare.
1127   BasicBlock *BB = AI->getParent();
1128   Builder.insertDeclare(NewAllocaAddress, DIVar, DIExpr, Loc, BB);
1129   DDI->eraseFromParent();
1130   return true;
1131 }
1132 
1133 /// changeToUnreachable - Insert an unreachable instruction before the specified
1134 /// instruction, making it and the rest of the code in the block dead.
changeToUnreachable(Instruction * I,bool UseLLVMTrap)1135 static void changeToUnreachable(Instruction *I, bool UseLLVMTrap) {
1136   BasicBlock *BB = I->getParent();
1137   // Loop over all of the successors, removing BB's entry from any PHI
1138   // nodes.
1139   for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI)
1140     (*SI)->removePredecessor(BB);
1141 
1142   // Insert a call to llvm.trap right before this.  This turns the undefined
1143   // behavior into a hard fail instead of falling through into random code.
1144   if (UseLLVMTrap) {
1145     Function *TrapFn =
1146       Intrinsic::getDeclaration(BB->getParent()->getParent(), Intrinsic::trap);
1147     CallInst *CallTrap = CallInst::Create(TrapFn, "", I);
1148     CallTrap->setDebugLoc(I->getDebugLoc());
1149   }
1150   new UnreachableInst(I->getContext(), I);
1151 
1152   // All instructions after this are dead.
1153   BasicBlock::iterator BBI = I, BBE = BB->end();
1154   while (BBI != BBE) {
1155     if (!BBI->use_empty())
1156       BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
1157     BB->getInstList().erase(BBI++);
1158   }
1159 }
1160 
1161 /// changeToCall - Convert the specified invoke into a normal call.
changeToCall(InvokeInst * II)1162 static void changeToCall(InvokeInst *II) {
1163   SmallVector<Value*, 8> Args(II->op_begin(), II->op_end() - 3);
1164   CallInst *NewCall = CallInst::Create(II->getCalledValue(), Args, "", II);
1165   NewCall->takeName(II);
1166   NewCall->setCallingConv(II->getCallingConv());
1167   NewCall->setAttributes(II->getAttributes());
1168   NewCall->setDebugLoc(II->getDebugLoc());
1169   II->replaceAllUsesWith(NewCall);
1170 
1171   // Follow the call by a branch to the normal destination.
1172   BranchInst::Create(II->getNormalDest(), II);
1173 
1174   // Update PHI nodes in the unwind destination
1175   II->getUnwindDest()->removePredecessor(II->getParent());
1176   II->eraseFromParent();
1177 }
1178 
markAliveBlocks(BasicBlock * BB,SmallPtrSetImpl<BasicBlock * > & Reachable)1179 static bool markAliveBlocks(BasicBlock *BB,
1180                             SmallPtrSetImpl<BasicBlock*> &Reachable) {
1181 
1182   SmallVector<BasicBlock*, 128> Worklist;
1183   Worklist.push_back(BB);
1184   Reachable.insert(BB);
1185   bool Changed = false;
1186   do {
1187     BB = Worklist.pop_back_val();
1188 
1189     // Do a quick scan of the basic block, turning any obviously unreachable
1190     // instructions into LLVM unreachable insts.  The instruction combining pass
1191     // canonicalizes unreachable insts into stores to null or undef.
1192     for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E;++BBI){
1193       // Assumptions that are known to be false are equivalent to unreachable.
1194       // Also, if the condition is undefined, then we make the choice most
1195       // beneficial to the optimizer, and choose that to also be unreachable.
1196       if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BBI))
1197         if (II->getIntrinsicID() == Intrinsic::assume) {
1198           bool MakeUnreachable = false;
1199           if (isa<UndefValue>(II->getArgOperand(0)))
1200             MakeUnreachable = true;
1201           else if (ConstantInt *Cond =
1202                    dyn_cast<ConstantInt>(II->getArgOperand(0)))
1203             MakeUnreachable = Cond->isZero();
1204 
1205           if (MakeUnreachable) {
1206             // Don't insert a call to llvm.trap right before the unreachable.
1207             changeToUnreachable(BBI, false);
1208             Changed = true;
1209             break;
1210           }
1211         }
1212 
1213       if (CallInst *CI = dyn_cast<CallInst>(BBI)) {
1214         if (CI->doesNotReturn()) {
1215           // If we found a call to a no-return function, insert an unreachable
1216           // instruction after it.  Make sure there isn't *already* one there
1217           // though.
1218           ++BBI;
1219           if (!isa<UnreachableInst>(BBI)) {
1220             // Don't insert a call to llvm.trap right before the unreachable.
1221             changeToUnreachable(BBI, false);
1222             Changed = true;
1223           }
1224           break;
1225         }
1226       }
1227 
1228       // Store to undef and store to null are undefined and used to signal that
1229       // they should be changed to unreachable by passes that can't modify the
1230       // CFG.
1231       if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
1232         // Don't touch volatile stores.
1233         if (SI->isVolatile()) continue;
1234 
1235         Value *Ptr = SI->getOperand(1);
1236 
1237         if (isa<UndefValue>(Ptr) ||
1238             (isa<ConstantPointerNull>(Ptr) &&
1239              SI->getPointerAddressSpace() == 0)) {
1240           changeToUnreachable(SI, true);
1241           Changed = true;
1242           break;
1243         }
1244       }
1245     }
1246 
1247     // Turn invokes that call 'nounwind' functions into ordinary calls.
1248     if (InvokeInst *II = dyn_cast<InvokeInst>(BB->getTerminator())) {
1249       Value *Callee = II->getCalledValue();
1250       if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1251         changeToUnreachable(II, true);
1252         Changed = true;
1253       } else if (II->doesNotThrow() && canSimplifyInvokeNoUnwind(II)) {
1254         if (II->use_empty() && II->onlyReadsMemory()) {
1255           // jump to the normal destination branch.
1256           BranchInst::Create(II->getNormalDest(), II);
1257           II->getUnwindDest()->removePredecessor(II->getParent());
1258           II->eraseFromParent();
1259         } else
1260           changeToCall(II);
1261         Changed = true;
1262       }
1263     }
1264 
1265     Changed |= ConstantFoldTerminator(BB, true);
1266     for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI)
1267       if (Reachable.insert(*SI).second)
1268         Worklist.push_back(*SI);
1269   } while (!Worklist.empty());
1270   return Changed;
1271 }
1272 
1273 /// removeUnreachableBlocksFromFn - Remove blocks that are not reachable, even
1274 /// if they are in a dead cycle.  Return true if a change was made, false
1275 /// otherwise.
removeUnreachableBlocks(Function & F)1276 bool llvm::removeUnreachableBlocks(Function &F) {
1277   SmallPtrSet<BasicBlock*, 128> Reachable;
1278   bool Changed = markAliveBlocks(F.begin(), Reachable);
1279 
1280   // If there are unreachable blocks in the CFG...
1281   if (Reachable.size() == F.size())
1282     return Changed;
1283 
1284   assert(Reachable.size() < F.size());
1285   NumRemoved += F.size()-Reachable.size();
1286 
1287   // Loop over all of the basic blocks that are not reachable, dropping all of
1288   // their internal references...
1289   for (Function::iterator BB = ++F.begin(), E = F.end(); BB != E; ++BB) {
1290     if (Reachable.count(BB))
1291       continue;
1292 
1293     for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI)
1294       if (Reachable.count(*SI))
1295         (*SI)->removePredecessor(BB);
1296     BB->dropAllReferences();
1297   }
1298 
1299   for (Function::iterator I = ++F.begin(); I != F.end();)
1300     if (!Reachable.count(I))
1301       I = F.getBasicBlockList().erase(I);
1302     else
1303       ++I;
1304 
1305   return true;
1306 }
1307 
combineMetadata(Instruction * K,const Instruction * J,ArrayRef<unsigned> KnownIDs)1308 void llvm::combineMetadata(Instruction *K, const Instruction *J, ArrayRef<unsigned> KnownIDs) {
1309   SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
1310   K->dropUnknownMetadata(KnownIDs);
1311   K->getAllMetadataOtherThanDebugLoc(Metadata);
1312   for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
1313     unsigned Kind = Metadata[i].first;
1314     MDNode *JMD = J->getMetadata(Kind);
1315     MDNode *KMD = Metadata[i].second;
1316 
1317     switch (Kind) {
1318       default:
1319         K->setMetadata(Kind, nullptr); // Remove unknown metadata
1320         break;
1321       case LLVMContext::MD_dbg:
1322         llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg");
1323       case LLVMContext::MD_tbaa:
1324         K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
1325         break;
1326       case LLVMContext::MD_alias_scope:
1327         K->setMetadata(Kind, MDNode::getMostGenericAliasScope(JMD, KMD));
1328         break;
1329       case LLVMContext::MD_noalias:
1330         K->setMetadata(Kind, MDNode::intersect(JMD, KMD));
1331         break;
1332       case LLVMContext::MD_range:
1333         K->setMetadata(Kind, MDNode::getMostGenericRange(JMD, KMD));
1334         break;
1335       case LLVMContext::MD_fpmath:
1336         K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
1337         break;
1338       case LLVMContext::MD_invariant_load:
1339         // Only set the !invariant.load if it is present in both instructions.
1340         K->setMetadata(Kind, JMD);
1341         break;
1342       case LLVMContext::MD_nonnull:
1343         // Only set the !nonnull if it is present in both instructions.
1344         K->setMetadata(Kind, JMD);
1345         break;
1346     }
1347   }
1348 }
1349