• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/DenseSet.h"
18 #include "llvm/ADT/Hashing.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/EHPersonalities.h"
24 #include "llvm/Analysis/InstructionSimplify.h"
25 #include "llvm/Analysis/MemoryBuiltins.h"
26 #include "llvm/Analysis/ValueTracking.h"
27 #include "llvm/IR/CFG.h"
28 #include "llvm/IR/Constants.h"
29 #include "llvm/IR/DIBuilder.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/DebugInfo.h"
32 #include "llvm/IR/DerivedTypes.h"
33 #include "llvm/IR/Dominators.h"
34 #include "llvm/IR/GetElementPtrTypeIterator.h"
35 #include "llvm/IR/GlobalAlias.h"
36 #include "llvm/IR/GlobalVariable.h"
37 #include "llvm/IR/IRBuilder.h"
38 #include "llvm/IR/Instructions.h"
39 #include "llvm/IR/IntrinsicInst.h"
40 #include "llvm/IR/Intrinsics.h"
41 #include "llvm/IR/MDBuilder.h"
42 #include "llvm/IR/Metadata.h"
43 #include "llvm/IR/Operator.h"
44 #include "llvm/IR/ValueHandle.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/MathExtras.h"
47 #include "llvm/Support/raw_ostream.h"
48 using namespace llvm;
49 
50 #define DEBUG_TYPE "local"
51 
52 STATISTIC(NumRemoved, "Number of unreachable basic blocks removed");
53 
54 //===----------------------------------------------------------------------===//
55 //  Local constant propagation.
56 //
57 
58 /// ConstantFoldTerminator - If a terminator instruction is predicated on a
59 /// constant value, convert it into an unconditional branch to the constant
60 /// destination.  This is a nontrivial operation because the successors of this
61 /// basic block must have their PHI nodes updated.
62 /// Also calls RecursivelyDeleteTriviallyDeadInstructions() on any branch/switch
63 /// conditions and indirectbr addresses this might make dead if
64 /// DeleteDeadConditions is true.
ConstantFoldTerminator(BasicBlock * BB,bool DeleteDeadConditions,const TargetLibraryInfo * TLI)65 bool llvm::ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions,
66                                   const TargetLibraryInfo *TLI) {
67   TerminatorInst *T = BB->getTerminator();
68   IRBuilder<> Builder(T);
69 
70   // Branch - See if we are conditional jumping on constant
71   if (BranchInst *BI = dyn_cast<BranchInst>(T)) {
72     if (BI->isUnconditional()) return false;  // Can't optimize uncond branch
73     BasicBlock *Dest1 = BI->getSuccessor(0);
74     BasicBlock *Dest2 = BI->getSuccessor(1);
75 
76     if (ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition())) {
77       // Are we branching on constant?
78       // YES.  Change to unconditional branch...
79       BasicBlock *Destination = Cond->getZExtValue() ? Dest1 : Dest2;
80       BasicBlock *OldDest     = Cond->getZExtValue() ? Dest2 : Dest1;
81 
82       //cerr << "Function: " << T->getParent()->getParent()
83       //     << "\nRemoving branch from " << T->getParent()
84       //     << "\n\nTo: " << OldDest << endl;
85 
86       // Let the basic block know that we are letting go of it.  Based on this,
87       // it will adjust it's PHI nodes.
88       OldDest->removePredecessor(BB);
89 
90       // Replace the conditional branch with an unconditional one.
91       Builder.CreateBr(Destination);
92       BI->eraseFromParent();
93       return true;
94     }
95 
96     if (Dest2 == Dest1) {       // Conditional branch to same location?
97       // This branch matches something like this:
98       //     br bool %cond, label %Dest, label %Dest
99       // and changes it into:  br label %Dest
100 
101       // Let the basic block know that we are letting go of one copy of it.
102       assert(BI->getParent() && "Terminator not inserted in block!");
103       Dest1->removePredecessor(BI->getParent());
104 
105       // Replace the conditional branch with an unconditional one.
106       Builder.CreateBr(Dest1);
107       Value *Cond = BI->getCondition();
108       BI->eraseFromParent();
109       if (DeleteDeadConditions)
110         RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI);
111       return true;
112     }
113     return false;
114   }
115 
116   if (SwitchInst *SI = dyn_cast<SwitchInst>(T)) {
117     // If we are switching on a constant, we can convert the switch to an
118     // unconditional branch.
119     ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition());
120     BasicBlock *DefaultDest = SI->getDefaultDest();
121     BasicBlock *TheOnlyDest = DefaultDest;
122 
123     // If the default is unreachable, ignore it when searching for TheOnlyDest.
124     if (isa<UnreachableInst>(DefaultDest->getFirstNonPHIOrDbg()) &&
125         SI->getNumCases() > 0) {
126       TheOnlyDest = SI->case_begin().getCaseSuccessor();
127     }
128 
129     // Figure out which case it goes to.
130     for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
131          i != e; ++i) {
132       // Found case matching a constant operand?
133       if (i.getCaseValue() == CI) {
134         TheOnlyDest = i.getCaseSuccessor();
135         break;
136       }
137 
138       // Check to see if this branch is going to the same place as the default
139       // dest.  If so, eliminate it as an explicit compare.
140       if (i.getCaseSuccessor() == DefaultDest) {
141         MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
142         unsigned NCases = SI->getNumCases();
143         // Fold the case metadata into the default if there will be any branches
144         // left, unless the metadata doesn't match the switch.
145         if (NCases > 1 && MD && MD->getNumOperands() == 2 + NCases) {
146           // Collect branch weights into a vector.
147           SmallVector<uint32_t, 8> Weights;
148           for (unsigned MD_i = 1, MD_e = MD->getNumOperands(); MD_i < MD_e;
149                ++MD_i) {
150             ConstantInt *CI =
151                 mdconst::dyn_extract<ConstantInt>(MD->getOperand(MD_i));
152             assert(CI);
153             Weights.push_back(CI->getValue().getZExtValue());
154           }
155           // Merge weight of this case to the default weight.
156           unsigned idx = i.getCaseIndex();
157           Weights[0] += Weights[idx+1];
158           // Remove weight for this case.
159           std::swap(Weights[idx+1], Weights.back());
160           Weights.pop_back();
161           SI->setMetadata(LLVMContext::MD_prof,
162                           MDBuilder(BB->getContext()).
163                           createBranchWeights(Weights));
164         }
165         // Remove this entry.
166         DefaultDest->removePredecessor(SI->getParent());
167         SI->removeCase(i);
168         --i; --e;
169         continue;
170       }
171 
172       // Otherwise, check to see if the switch only branches to one destination.
173       // We do this by reseting "TheOnlyDest" to null when we find two non-equal
174       // destinations.
175       if (i.getCaseSuccessor() != TheOnlyDest) TheOnlyDest = nullptr;
176     }
177 
178     if (CI && !TheOnlyDest) {
179       // Branching on a constant, but not any of the cases, go to the default
180       // successor.
181       TheOnlyDest = SI->getDefaultDest();
182     }
183 
184     // If we found a single destination that we can fold the switch into, do so
185     // now.
186     if (TheOnlyDest) {
187       // Insert the new branch.
188       Builder.CreateBr(TheOnlyDest);
189       BasicBlock *BB = SI->getParent();
190 
191       // Remove entries from PHI nodes which we no longer branch to...
192       for (BasicBlock *Succ : SI->successors()) {
193         // Found case matching a constant operand?
194         if (Succ == TheOnlyDest)
195           TheOnlyDest = nullptr; // Don't modify the first branch to TheOnlyDest
196         else
197           Succ->removePredecessor(BB);
198       }
199 
200       // Delete the old switch.
201       Value *Cond = SI->getCondition();
202       SI->eraseFromParent();
203       if (DeleteDeadConditions)
204         RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI);
205       return true;
206     }
207 
208     if (SI->getNumCases() == 1) {
209       // Otherwise, we can fold this switch into a conditional branch
210       // instruction if it has only one non-default destination.
211       SwitchInst::CaseIt FirstCase = SI->case_begin();
212       Value *Cond = Builder.CreateICmpEQ(SI->getCondition(),
213           FirstCase.getCaseValue(), "cond");
214 
215       // Insert the new branch.
216       BranchInst *NewBr = Builder.CreateCondBr(Cond,
217                                                FirstCase.getCaseSuccessor(),
218                                                SI->getDefaultDest());
219       MDNode *MD = SI->getMetadata(LLVMContext::MD_prof);
220       if (MD && MD->getNumOperands() == 3) {
221         ConstantInt *SICase =
222             mdconst::dyn_extract<ConstantInt>(MD->getOperand(2));
223         ConstantInt *SIDef =
224             mdconst::dyn_extract<ConstantInt>(MD->getOperand(1));
225         assert(SICase && SIDef);
226         // The TrueWeight should be the weight for the single case of SI.
227         NewBr->setMetadata(LLVMContext::MD_prof,
228                         MDBuilder(BB->getContext()).
229                         createBranchWeights(SICase->getValue().getZExtValue(),
230                                             SIDef->getValue().getZExtValue()));
231       }
232 
233       // Update make.implicit metadata to the newly-created conditional branch.
234       MDNode *MakeImplicitMD = SI->getMetadata(LLVMContext::MD_make_implicit);
235       if (MakeImplicitMD)
236         NewBr->setMetadata(LLVMContext::MD_make_implicit, MakeImplicitMD);
237 
238       // Delete the old switch.
239       SI->eraseFromParent();
240       return true;
241     }
242     return false;
243   }
244 
245   if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(T)) {
246     // indirectbr blockaddress(@F, @BB) -> br label @BB
247     if (BlockAddress *BA =
248           dyn_cast<BlockAddress>(IBI->getAddress()->stripPointerCasts())) {
249       BasicBlock *TheOnlyDest = BA->getBasicBlock();
250       // Insert the new branch.
251       Builder.CreateBr(TheOnlyDest);
252 
253       for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) {
254         if (IBI->getDestination(i) == TheOnlyDest)
255           TheOnlyDest = nullptr;
256         else
257           IBI->getDestination(i)->removePredecessor(IBI->getParent());
258       }
259       Value *Address = IBI->getAddress();
260       IBI->eraseFromParent();
261       if (DeleteDeadConditions)
262         RecursivelyDeleteTriviallyDeadInstructions(Address, TLI);
263 
264       // If we didn't find our destination in the IBI successor list, then we
265       // have undefined behavior.  Replace the unconditional branch with an
266       // 'unreachable' instruction.
267       if (TheOnlyDest) {
268         BB->getTerminator()->eraseFromParent();
269         new UnreachableInst(BB->getContext(), BB);
270       }
271 
272       return true;
273     }
274   }
275 
276   return false;
277 }
278 
279 
280 //===----------------------------------------------------------------------===//
281 //  Local dead code elimination.
282 //
283 
284 /// isInstructionTriviallyDead - Return true if the result produced by the
285 /// instruction is not used, and the instruction has no side effects.
286 ///
isInstructionTriviallyDead(Instruction * I,const TargetLibraryInfo * TLI)287 bool llvm::isInstructionTriviallyDead(Instruction *I,
288                                       const TargetLibraryInfo *TLI) {
289   if (!I->use_empty() || isa<TerminatorInst>(I)) return false;
290 
291   // We don't want the landingpad-like instructions removed by anything this
292   // general.
293   if (I->isEHPad())
294     return false;
295 
296   // We don't want debug info removed by anything this general, unless
297   // debug info is empty.
298   if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(I)) {
299     if (DDI->getAddress())
300       return false;
301     return true;
302   }
303   if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(I)) {
304     if (DVI->getValue())
305       return false;
306     return true;
307   }
308 
309   if (!I->mayHaveSideEffects()) return true;
310 
311   // Special case intrinsics that "may have side effects" but can be deleted
312   // when dead.
313   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
314     // Safe to delete llvm.stacksave if dead.
315     if (II->getIntrinsicID() == Intrinsic::stacksave)
316       return true;
317 
318     // Lifetime intrinsics are dead when their right-hand is undef.
319     if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
320         II->getIntrinsicID() == Intrinsic::lifetime_end)
321       return isa<UndefValue>(II->getArgOperand(1));
322 
323     // Assumptions are dead if their condition is trivially true.
324     if (II->getIntrinsicID() == Intrinsic::assume) {
325       if (ConstantInt *Cond = dyn_cast<ConstantInt>(II->getArgOperand(0)))
326         return !Cond->isZero();
327 
328       return false;
329     }
330   }
331 
332   if (isAllocLikeFn(I, TLI)) return true;
333 
334   if (CallInst *CI = isFreeCall(I, TLI))
335     if (Constant *C = dyn_cast<Constant>(CI->getArgOperand(0)))
336       return C->isNullValue() || isa<UndefValue>(C);
337 
338   return false;
339 }
340 
341 /// RecursivelyDeleteTriviallyDeadInstructions - If the specified value is a
342 /// trivially dead instruction, delete it.  If that makes any of its operands
343 /// trivially dead, delete them too, recursively.  Return true if any
344 /// instructions were deleted.
345 bool
RecursivelyDeleteTriviallyDeadInstructions(Value * V,const TargetLibraryInfo * TLI)346 llvm::RecursivelyDeleteTriviallyDeadInstructions(Value *V,
347                                                  const TargetLibraryInfo *TLI) {
348   Instruction *I = dyn_cast<Instruction>(V);
349   if (!I || !I->use_empty() || !isInstructionTriviallyDead(I, TLI))
350     return false;
351 
352   SmallVector<Instruction*, 16> DeadInsts;
353   DeadInsts.push_back(I);
354 
355   do {
356     I = DeadInsts.pop_back_val();
357 
358     // Null out all of the instruction's operands to see if any operand becomes
359     // dead as we go.
360     for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
361       Value *OpV = I->getOperand(i);
362       I->setOperand(i, nullptr);
363 
364       if (!OpV->use_empty()) continue;
365 
366       // If the operand is an instruction that became dead as we nulled out the
367       // operand, and if it is 'trivially' dead, delete it in a future loop
368       // iteration.
369       if (Instruction *OpI = dyn_cast<Instruction>(OpV))
370         if (isInstructionTriviallyDead(OpI, TLI))
371           DeadInsts.push_back(OpI);
372     }
373 
374     I->eraseFromParent();
375   } while (!DeadInsts.empty());
376 
377   return true;
378 }
379 
380 /// areAllUsesEqual - Check whether the uses of a value are all the same.
381 /// This is similar to Instruction::hasOneUse() except this will also return
382 /// true when there are no uses or multiple uses that all refer to the same
383 /// value.
areAllUsesEqual(Instruction * I)384 static bool areAllUsesEqual(Instruction *I) {
385   Value::user_iterator UI = I->user_begin();
386   Value::user_iterator UE = I->user_end();
387   if (UI == UE)
388     return true;
389 
390   User *TheUse = *UI;
391   for (++UI; UI != UE; ++UI) {
392     if (*UI != TheUse)
393       return false;
394   }
395   return true;
396 }
397 
398 /// RecursivelyDeleteDeadPHINode - If the specified value is an effectively
399 /// dead PHI node, due to being a def-use chain of single-use nodes that
400 /// either forms a cycle or is terminated by a trivially dead instruction,
401 /// delete it.  If that makes any of its operands trivially dead, delete them
402 /// too, recursively.  Return true if a change was made.
RecursivelyDeleteDeadPHINode(PHINode * PN,const TargetLibraryInfo * TLI)403 bool llvm::RecursivelyDeleteDeadPHINode(PHINode *PN,
404                                         const TargetLibraryInfo *TLI) {
405   SmallPtrSet<Instruction*, 4> Visited;
406   for (Instruction *I = PN; areAllUsesEqual(I) && !I->mayHaveSideEffects();
407        I = cast<Instruction>(*I->user_begin())) {
408     if (I->use_empty())
409       return RecursivelyDeleteTriviallyDeadInstructions(I, TLI);
410 
411     // If we find an instruction more than once, we're on a cycle that
412     // won't prove fruitful.
413     if (!Visited.insert(I).second) {
414       // Break the cycle and delete the instruction and its operands.
415       I->replaceAllUsesWith(UndefValue::get(I->getType()));
416       (void)RecursivelyDeleteTriviallyDeadInstructions(I, TLI);
417       return true;
418     }
419   }
420   return false;
421 }
422 
423 static bool
simplifyAndDCEInstruction(Instruction * I,SmallSetVector<Instruction *,16> & WorkList,const DataLayout & DL,const TargetLibraryInfo * TLI)424 simplifyAndDCEInstruction(Instruction *I,
425                           SmallSetVector<Instruction *, 16> &WorkList,
426                           const DataLayout &DL,
427                           const TargetLibraryInfo *TLI) {
428   if (isInstructionTriviallyDead(I, TLI)) {
429     // Null out all of the instruction's operands to see if any operand becomes
430     // dead as we go.
431     for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
432       Value *OpV = I->getOperand(i);
433       I->setOperand(i, nullptr);
434 
435       if (!OpV->use_empty() || I == OpV)
436         continue;
437 
438       // If the operand is an instruction that became dead as we nulled out the
439       // operand, and if it is 'trivially' dead, delete it in a future loop
440       // iteration.
441       if (Instruction *OpI = dyn_cast<Instruction>(OpV))
442         if (isInstructionTriviallyDead(OpI, TLI))
443           WorkList.insert(OpI);
444     }
445 
446     I->eraseFromParent();
447 
448     return true;
449   }
450 
451   if (Value *SimpleV = SimplifyInstruction(I, DL)) {
452     // Add the users to the worklist. CAREFUL: an instruction can use itself,
453     // in the case of a phi node.
454     for (User *U : I->users())
455       if (U != I)
456         WorkList.insert(cast<Instruction>(U));
457 
458     // Replace the instruction with its simplified value.
459     I->replaceAllUsesWith(SimpleV);
460     I->eraseFromParent();
461     return true;
462   }
463   return false;
464 }
465 
466 /// SimplifyInstructionsInBlock - Scan the specified basic block and try to
467 /// simplify any instructions in it and recursively delete dead instructions.
468 ///
469 /// This returns true if it changed the code, note that it can delete
470 /// instructions in other blocks as well in this block.
SimplifyInstructionsInBlock(BasicBlock * BB,const TargetLibraryInfo * TLI)471 bool llvm::SimplifyInstructionsInBlock(BasicBlock *BB,
472                                        const TargetLibraryInfo *TLI) {
473   bool MadeChange = false;
474   const DataLayout &DL = BB->getModule()->getDataLayout();
475 
476 #ifndef NDEBUG
477   // In debug builds, ensure that the terminator of the block is never replaced
478   // or deleted by these simplifications. The idea of simplification is that it
479   // cannot introduce new instructions, and there is no way to replace the
480   // terminator of a block without introducing a new instruction.
481   AssertingVH<Instruction> TerminatorVH(&BB->back());
482 #endif
483 
484   SmallSetVector<Instruction *, 16> WorkList;
485   // Iterate over the original function, only adding insts to the worklist
486   // if they actually need to be revisited. This avoids having to pre-init
487   // the worklist with the entire function's worth of instructions.
488   for (BasicBlock::iterator BI = BB->begin(), E = std::prev(BB->end()); BI != E;) {
489     assert(!BI->isTerminator());
490     Instruction *I = &*BI;
491     ++BI;
492 
493     // We're visiting this instruction now, so make sure it's not in the
494     // worklist from an earlier visit.
495     if (!WorkList.count(I))
496       MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI);
497   }
498 
499   while (!WorkList.empty()) {
500     Instruction *I = WorkList.pop_back_val();
501     MadeChange |= simplifyAndDCEInstruction(I, WorkList, DL, TLI);
502   }
503   return MadeChange;
504 }
505 
506 //===----------------------------------------------------------------------===//
507 //  Control Flow Graph Restructuring.
508 //
509 
510 
511 /// RemovePredecessorAndSimplify - Like BasicBlock::removePredecessor, this
512 /// method is called when we're about to delete Pred as a predecessor of BB.  If
513 /// BB contains any PHI nodes, this drops the entries in the PHI nodes for Pred.
514 ///
515 /// Unlike the removePredecessor method, this attempts to simplify uses of PHI
516 /// nodes that collapse into identity values.  For example, if we have:
517 ///   x = phi(1, 0, 0, 0)
518 ///   y = and x, z
519 ///
520 /// .. and delete the predecessor corresponding to the '1', this will attempt to
521 /// recursively fold the and to 0.
RemovePredecessorAndSimplify(BasicBlock * BB,BasicBlock * Pred)522 void llvm::RemovePredecessorAndSimplify(BasicBlock *BB, BasicBlock *Pred) {
523   // This only adjusts blocks with PHI nodes.
524   if (!isa<PHINode>(BB->begin()))
525     return;
526 
527   // Remove the entries for Pred from the PHI nodes in BB, but do not simplify
528   // them down.  This will leave us with single entry phi nodes and other phis
529   // that can be removed.
530   BB->removePredecessor(Pred, true);
531 
532   WeakVH PhiIt = &BB->front();
533   while (PHINode *PN = dyn_cast<PHINode>(PhiIt)) {
534     PhiIt = &*++BasicBlock::iterator(cast<Instruction>(PhiIt));
535     Value *OldPhiIt = PhiIt;
536 
537     if (!recursivelySimplifyInstruction(PN))
538       continue;
539 
540     // If recursive simplification ended up deleting the next PHI node we would
541     // iterate to, then our iterator is invalid, restart scanning from the top
542     // of the block.
543     if (PhiIt != OldPhiIt) PhiIt = &BB->front();
544   }
545 }
546 
547 
548 /// MergeBasicBlockIntoOnlyPred - DestBB is a block with one predecessor and its
549 /// predecessor is known to have one successor (DestBB!).  Eliminate the edge
550 /// between them, moving the instructions in the predecessor into DestBB and
551 /// deleting the predecessor block.
552 ///
MergeBasicBlockIntoOnlyPred(BasicBlock * DestBB,DominatorTree * DT)553 void llvm::MergeBasicBlockIntoOnlyPred(BasicBlock *DestBB, DominatorTree *DT) {
554   // If BB has single-entry PHI nodes, fold them.
555   while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) {
556     Value *NewVal = PN->getIncomingValue(0);
557     // Replace self referencing PHI with undef, it must be dead.
558     if (NewVal == PN) NewVal = UndefValue::get(PN->getType());
559     PN->replaceAllUsesWith(NewVal);
560     PN->eraseFromParent();
561   }
562 
563   BasicBlock *PredBB = DestBB->getSinglePredecessor();
564   assert(PredBB && "Block doesn't have a single predecessor!");
565 
566   // Zap anything that took the address of DestBB.  Not doing this will give the
567   // address an invalid value.
568   if (DestBB->hasAddressTaken()) {
569     BlockAddress *BA = BlockAddress::get(DestBB);
570     Constant *Replacement =
571       ConstantInt::get(llvm::Type::getInt32Ty(BA->getContext()), 1);
572     BA->replaceAllUsesWith(ConstantExpr::getIntToPtr(Replacement,
573                                                      BA->getType()));
574     BA->destroyConstant();
575   }
576 
577   // Anything that branched to PredBB now branches to DestBB.
578   PredBB->replaceAllUsesWith(DestBB);
579 
580   // Splice all the instructions from PredBB to DestBB.
581   PredBB->getTerminator()->eraseFromParent();
582   DestBB->getInstList().splice(DestBB->begin(), PredBB->getInstList());
583 
584   // If the PredBB is the entry block of the function, move DestBB up to
585   // become the entry block after we erase PredBB.
586   if (PredBB == &DestBB->getParent()->getEntryBlock())
587     DestBB->moveAfter(PredBB);
588 
589   if (DT) {
590     BasicBlock *PredBBIDom = DT->getNode(PredBB)->getIDom()->getBlock();
591     DT->changeImmediateDominator(DestBB, PredBBIDom);
592     DT->eraseNode(PredBB);
593   }
594   // Nuke BB.
595   PredBB->eraseFromParent();
596 }
597 
598 /// CanMergeValues - Return true if we can choose one of these values to use
599 /// in place of the other. Note that we will always choose the non-undef
600 /// value to keep.
CanMergeValues(Value * First,Value * Second)601 static bool CanMergeValues(Value *First, Value *Second) {
602   return First == Second || isa<UndefValue>(First) || isa<UndefValue>(Second);
603 }
604 
605 /// CanPropagatePredecessorsForPHIs - Return true if we can fold BB, an
606 /// almost-empty BB ending in an unconditional branch to Succ, into Succ.
607 ///
608 /// Assumption: Succ is the single successor for BB.
609 ///
CanPropagatePredecessorsForPHIs(BasicBlock * BB,BasicBlock * Succ)610 static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) {
611   assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!");
612 
613   DEBUG(dbgs() << "Looking to fold " << BB->getName() << " into "
614         << Succ->getName() << "\n");
615   // Shortcut, if there is only a single predecessor it must be BB and merging
616   // is always safe
617   if (Succ->getSinglePredecessor()) return true;
618 
619   // Make a list of the predecessors of BB
620   SmallPtrSet<BasicBlock*, 16> BBPreds(pred_begin(BB), pred_end(BB));
621 
622   // Look at all the phi nodes in Succ, to see if they present a conflict when
623   // merging these blocks
624   for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
625     PHINode *PN = cast<PHINode>(I);
626 
627     // If the incoming value from BB is again a PHINode in
628     // BB which has the same incoming value for *PI as PN does, we can
629     // merge the phi nodes and then the blocks can still be merged
630     PHINode *BBPN = dyn_cast<PHINode>(PN->getIncomingValueForBlock(BB));
631     if (BBPN && BBPN->getParent() == BB) {
632       for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) {
633         BasicBlock *IBB = PN->getIncomingBlock(PI);
634         if (BBPreds.count(IBB) &&
635             !CanMergeValues(BBPN->getIncomingValueForBlock(IBB),
636                             PN->getIncomingValue(PI))) {
637           DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in "
638                 << Succ->getName() << " is conflicting with "
639                 << BBPN->getName() << " with regard to common predecessor "
640                 << IBB->getName() << "\n");
641           return false;
642         }
643       }
644     } else {
645       Value* Val = PN->getIncomingValueForBlock(BB);
646       for (unsigned PI = 0, PE = PN->getNumIncomingValues(); PI != PE; ++PI) {
647         // See if the incoming value for the common predecessor is equal to the
648         // one for BB, in which case this phi node will not prevent the merging
649         // of the block.
650         BasicBlock *IBB = PN->getIncomingBlock(PI);
651         if (BBPreds.count(IBB) &&
652             !CanMergeValues(Val, PN->getIncomingValue(PI))) {
653           DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in "
654                 << Succ->getName() << " is conflicting with regard to common "
655                 << "predecessor " << IBB->getName() << "\n");
656           return false;
657         }
658       }
659     }
660   }
661 
662   return true;
663 }
664 
665 typedef SmallVector<BasicBlock *, 16> PredBlockVector;
666 typedef DenseMap<BasicBlock *, Value *> IncomingValueMap;
667 
668 /// \brief Determines the value to use as the phi node input for a block.
669 ///
670 /// Select between \p OldVal any value that we know flows from \p BB
671 /// to a particular phi on the basis of which one (if either) is not
672 /// undef. Update IncomingValues based on the selected value.
673 ///
674 /// \param OldVal The value we are considering selecting.
675 /// \param BB The block that the value flows in from.
676 /// \param IncomingValues A map from block-to-value for other phi inputs
677 /// that we have examined.
678 ///
679 /// \returns the selected value.
selectIncomingValueForBlock(Value * OldVal,BasicBlock * BB,IncomingValueMap & IncomingValues)680 static Value *selectIncomingValueForBlock(Value *OldVal, BasicBlock *BB,
681                                           IncomingValueMap &IncomingValues) {
682   if (!isa<UndefValue>(OldVal)) {
683     assert((!IncomingValues.count(BB) ||
684             IncomingValues.find(BB)->second == OldVal) &&
685            "Expected OldVal to match incoming value from BB!");
686 
687     IncomingValues.insert(std::make_pair(BB, OldVal));
688     return OldVal;
689   }
690 
691   IncomingValueMap::const_iterator It = IncomingValues.find(BB);
692   if (It != IncomingValues.end()) return It->second;
693 
694   return OldVal;
695 }
696 
697 /// \brief Create a map from block to value for the operands of a
698 /// given phi.
699 ///
700 /// Create a map from block to value for each non-undef value flowing
701 /// into \p PN.
702 ///
703 /// \param PN The phi we are collecting the map for.
704 /// \param IncomingValues [out] The map from block to value for this phi.
gatherIncomingValuesToPhi(PHINode * PN,IncomingValueMap & IncomingValues)705 static void gatherIncomingValuesToPhi(PHINode *PN,
706                                       IncomingValueMap &IncomingValues) {
707   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
708     BasicBlock *BB = PN->getIncomingBlock(i);
709     Value *V = PN->getIncomingValue(i);
710 
711     if (!isa<UndefValue>(V))
712       IncomingValues.insert(std::make_pair(BB, V));
713   }
714 }
715 
716 /// \brief Replace the incoming undef values to a phi with the values
717 /// from a block-to-value map.
718 ///
719 /// \param PN The phi we are replacing the undefs in.
720 /// \param IncomingValues A map from block to value.
replaceUndefValuesInPhi(PHINode * PN,const IncomingValueMap & IncomingValues)721 static void replaceUndefValuesInPhi(PHINode *PN,
722                                     const IncomingValueMap &IncomingValues) {
723   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
724     Value *V = PN->getIncomingValue(i);
725 
726     if (!isa<UndefValue>(V)) continue;
727 
728     BasicBlock *BB = PN->getIncomingBlock(i);
729     IncomingValueMap::const_iterator It = IncomingValues.find(BB);
730     if (It == IncomingValues.end()) continue;
731 
732     PN->setIncomingValue(i, It->second);
733   }
734 }
735 
736 /// \brief Replace a value flowing from a block to a phi with
737 /// potentially multiple instances of that value flowing from the
738 /// block's predecessors to the phi.
739 ///
740 /// \param BB The block with the value flowing into the phi.
741 /// \param BBPreds The predecessors of BB.
742 /// \param PN The phi that we are updating.
redirectValuesFromPredecessorsToPhi(BasicBlock * BB,const PredBlockVector & BBPreds,PHINode * PN)743 static void redirectValuesFromPredecessorsToPhi(BasicBlock *BB,
744                                                 const PredBlockVector &BBPreds,
745                                                 PHINode *PN) {
746   Value *OldVal = PN->removeIncomingValue(BB, false);
747   assert(OldVal && "No entry in PHI for Pred BB!");
748 
749   IncomingValueMap IncomingValues;
750 
751   // We are merging two blocks - BB, and the block containing PN - and
752   // as a result we need to redirect edges from the predecessors of BB
753   // to go to the block containing PN, and update PN
754   // accordingly. Since we allow merging blocks in the case where the
755   // predecessor and successor blocks both share some predecessors,
756   // and where some of those common predecessors might have undef
757   // values flowing into PN, we want to rewrite those values to be
758   // consistent with the non-undef values.
759 
760   gatherIncomingValuesToPhi(PN, IncomingValues);
761 
762   // If this incoming value is one of the PHI nodes in BB, the new entries
763   // in the PHI node are the entries from the old PHI.
764   if (isa<PHINode>(OldVal) && cast<PHINode>(OldVal)->getParent() == BB) {
765     PHINode *OldValPN = cast<PHINode>(OldVal);
766     for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) {
767       // Note that, since we are merging phi nodes and BB and Succ might
768       // have common predecessors, we could end up with a phi node with
769       // identical incoming branches. This will be cleaned up later (and
770       // will trigger asserts if we try to clean it up now, without also
771       // simplifying the corresponding conditional branch).
772       BasicBlock *PredBB = OldValPN->getIncomingBlock(i);
773       Value *PredVal = OldValPN->getIncomingValue(i);
774       Value *Selected = selectIncomingValueForBlock(PredVal, PredBB,
775                                                     IncomingValues);
776 
777       // And add a new incoming value for this predecessor for the
778       // newly retargeted branch.
779       PN->addIncoming(Selected, PredBB);
780     }
781   } else {
782     for (unsigned i = 0, e = BBPreds.size(); i != e; ++i) {
783       // Update existing incoming values in PN for this
784       // predecessor of BB.
785       BasicBlock *PredBB = BBPreds[i];
786       Value *Selected = selectIncomingValueForBlock(OldVal, PredBB,
787                                                     IncomingValues);
788 
789       // And add a new incoming value for this predecessor for the
790       // newly retargeted branch.
791       PN->addIncoming(Selected, PredBB);
792     }
793   }
794 
795   replaceUndefValuesInPhi(PN, IncomingValues);
796 }
797 
798 /// TryToSimplifyUncondBranchFromEmptyBlock - BB is known to contain an
799 /// unconditional branch, and contains no instructions other than PHI nodes,
800 /// potential side-effect free intrinsics and the branch.  If possible,
801 /// eliminate BB by rewriting all the predecessors to branch to the successor
802 /// block and return true.  If we can't transform, return false.
TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock * BB)803 bool llvm::TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB) {
804   assert(BB != &BB->getParent()->getEntryBlock() &&
805          "TryToSimplifyUncondBranchFromEmptyBlock called on entry block!");
806 
807   // We can't eliminate infinite loops.
808   BasicBlock *Succ = cast<BranchInst>(BB->getTerminator())->getSuccessor(0);
809   if (BB == Succ) return false;
810 
811   // Check to see if merging these blocks would cause conflicts for any of the
812   // phi nodes in BB or Succ. If not, we can safely merge.
813   if (!CanPropagatePredecessorsForPHIs(BB, Succ)) return false;
814 
815   // Check for cases where Succ has multiple predecessors and a PHI node in BB
816   // has uses which will not disappear when the PHI nodes are merged.  It is
817   // possible to handle such cases, but difficult: it requires checking whether
818   // BB dominates Succ, which is non-trivial to calculate in the case where
819   // Succ has multiple predecessors.  Also, it requires checking whether
820   // constructing the necessary self-referential PHI node doesn't introduce any
821   // conflicts; this isn't too difficult, but the previous code for doing this
822   // was incorrect.
823   //
824   // Note that if this check finds a live use, BB dominates Succ, so BB is
825   // something like a loop pre-header (or rarely, a part of an irreducible CFG);
826   // folding the branch isn't profitable in that case anyway.
827   if (!Succ->getSinglePredecessor()) {
828     BasicBlock::iterator BBI = BB->begin();
829     while (isa<PHINode>(*BBI)) {
830       for (Use &U : BBI->uses()) {
831         if (PHINode* PN = dyn_cast<PHINode>(U.getUser())) {
832           if (PN->getIncomingBlock(U) != BB)
833             return false;
834         } else {
835           return false;
836         }
837       }
838       ++BBI;
839     }
840   }
841 
842   DEBUG(dbgs() << "Killing Trivial BB: \n" << *BB);
843 
844   if (isa<PHINode>(Succ->begin())) {
845     // If there is more than one pred of succ, and there are PHI nodes in
846     // the successor, then we need to add incoming edges for the PHI nodes
847     //
848     const PredBlockVector BBPreds(pred_begin(BB), pred_end(BB));
849 
850     // Loop over all of the PHI nodes in the successor of BB.
851     for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
852       PHINode *PN = cast<PHINode>(I);
853 
854       redirectValuesFromPredecessorsToPhi(BB, BBPreds, PN);
855     }
856   }
857 
858   if (Succ->getSinglePredecessor()) {
859     // BB is the only predecessor of Succ, so Succ will end up with exactly
860     // the same predecessors BB had.
861 
862     // Copy over any phi, debug or lifetime instruction.
863     BB->getTerminator()->eraseFromParent();
864     Succ->getInstList().splice(Succ->getFirstNonPHI()->getIterator(),
865                                BB->getInstList());
866   } else {
867     while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) {
868       // We explicitly check for such uses in CanPropagatePredecessorsForPHIs.
869       assert(PN->use_empty() && "There shouldn't be any uses here!");
870       PN->eraseFromParent();
871     }
872   }
873 
874   // Everything that jumped to BB now goes to Succ.
875   BB->replaceAllUsesWith(Succ);
876   if (!Succ->hasName()) Succ->takeName(BB);
877   BB->eraseFromParent();              // Delete the old basic block.
878   return true;
879 }
880 
881 /// EliminateDuplicatePHINodes - Check for and eliminate duplicate PHI
882 /// nodes in this block. This doesn't try to be clever about PHI nodes
883 /// which differ only in the order of the incoming values, but instcombine
884 /// orders them so it usually won't matter.
885 ///
EliminateDuplicatePHINodes(BasicBlock * BB)886 bool llvm::EliminateDuplicatePHINodes(BasicBlock *BB) {
887   // This implementation doesn't currently consider undef operands
888   // specially. Theoretically, two phis which are identical except for
889   // one having an undef where the other doesn't could be collapsed.
890 
891   struct PHIDenseMapInfo {
892     static PHINode *getEmptyKey() {
893       return DenseMapInfo<PHINode *>::getEmptyKey();
894     }
895     static PHINode *getTombstoneKey() {
896       return DenseMapInfo<PHINode *>::getTombstoneKey();
897     }
898     static unsigned getHashValue(PHINode *PN) {
899       // Compute a hash value on the operands. Instcombine will likely have
900       // sorted them, which helps expose duplicates, but we have to check all
901       // the operands to be safe in case instcombine hasn't run.
902       return static_cast<unsigned>(hash_combine(
903           hash_combine_range(PN->value_op_begin(), PN->value_op_end()),
904           hash_combine_range(PN->block_begin(), PN->block_end())));
905     }
906     static bool isEqual(PHINode *LHS, PHINode *RHS) {
907       if (LHS == getEmptyKey() || LHS == getTombstoneKey() ||
908           RHS == getEmptyKey() || RHS == getTombstoneKey())
909         return LHS == RHS;
910       return LHS->isIdenticalTo(RHS);
911     }
912   };
913 
914   // Set of unique PHINodes.
915   DenseSet<PHINode *, PHIDenseMapInfo> PHISet;
916 
917   // Examine each PHI.
918   bool Changed = false;
919   for (auto I = BB->begin(); PHINode *PN = dyn_cast<PHINode>(I++);) {
920     auto Inserted = PHISet.insert(PN);
921     if (!Inserted.second) {
922       // A duplicate. Replace this PHI with its duplicate.
923       PN->replaceAllUsesWith(*Inserted.first);
924       PN->eraseFromParent();
925       Changed = true;
926 
927       // The RAUW can change PHIs that we already visited. Start over from the
928       // beginning.
929       PHISet.clear();
930       I = BB->begin();
931     }
932   }
933 
934   return Changed;
935 }
936 
937 /// enforceKnownAlignment - If the specified pointer points to an object that
938 /// we control, modify the object's alignment to PrefAlign. This isn't
939 /// often possible though. If alignment is important, a more reliable approach
940 /// is to simply align all global variables and allocation instructions to
941 /// their preferred alignment from the beginning.
942 ///
enforceKnownAlignment(Value * V,unsigned Align,unsigned PrefAlign,const DataLayout & DL)943 static unsigned enforceKnownAlignment(Value *V, unsigned Align,
944                                       unsigned PrefAlign,
945                                       const DataLayout &DL) {
946   V = V->stripPointerCasts();
947 
948   if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
949     // If the preferred alignment is greater than the natural stack alignment
950     // then don't round up. This avoids dynamic stack realignment.
951     if (DL.exceedsNaturalStackAlignment(PrefAlign))
952       return Align;
953     // If there is a requested alignment and if this is an alloca, round up.
954     if (AI->getAlignment() >= PrefAlign)
955       return AI->getAlignment();
956     AI->setAlignment(PrefAlign);
957     return PrefAlign;
958   }
959 
960   if (auto *GO = dyn_cast<GlobalObject>(V)) {
961     // If there is a large requested alignment and we can, bump up the alignment
962     // of the global.  If the memory we set aside for the global may not be the
963     // memory used by the final program then it is impossible for us to reliably
964     // enforce the preferred alignment.
965     if (!GO->isStrongDefinitionForLinker())
966       return Align;
967 
968     if (GO->getAlignment() >= PrefAlign)
969       return GO->getAlignment();
970     // We can only increase the alignment of the global if it has no alignment
971     // specified or if it is not assigned a section.  If it is assigned a
972     // section, the global could be densely packed with other objects in the
973     // section, increasing the alignment could cause padding issues.
974     if (!GO->hasSection() || GO->getAlignment() == 0)
975       GO->setAlignment(PrefAlign);
976     return GO->getAlignment();
977   }
978 
979   return Align;
980 }
981 
982 /// getOrEnforceKnownAlignment - If the specified pointer has an alignment that
983 /// we can determine, return it, otherwise return 0.  If PrefAlign is specified,
984 /// and it is more than the alignment of the ultimate object, see if we can
985 /// 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)986 unsigned llvm::getOrEnforceKnownAlignment(Value *V, unsigned PrefAlign,
987                                           const DataLayout &DL,
988                                           const Instruction *CxtI,
989                                           AssumptionCache *AC,
990                                           const DominatorTree *DT) {
991   assert(V->getType()->isPointerTy() &&
992          "getOrEnforceKnownAlignment expects a pointer!");
993   unsigned BitWidth = DL.getPointerTypeSizeInBits(V->getType());
994 
995   APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
996   computeKnownBits(V, KnownZero, KnownOne, DL, 0, AC, CxtI, DT);
997   unsigned TrailZ = KnownZero.countTrailingOnes();
998 
999   // Avoid trouble with ridiculously large TrailZ values, such as
1000   // those computed from a null pointer.
1001   TrailZ = std::min(TrailZ, unsigned(sizeof(unsigned) * CHAR_BIT - 1));
1002 
1003   unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
1004 
1005   // LLVM doesn't support alignments larger than this currently.
1006   Align = std::min(Align, +Value::MaximumAlignment);
1007 
1008   if (PrefAlign > Align)
1009     Align = enforceKnownAlignment(V, Align, PrefAlign, DL);
1010 
1011   // We don't need to make any adjustment.
1012   return Align;
1013 }
1014 
1015 ///===---------------------------------------------------------------------===//
1016 ///  Dbg Intrinsic utilities
1017 ///
1018 
1019 /// See if there is a dbg.value intrinsic for DIVar before I.
LdStHasDebugValue(const DILocalVariable * DIVar,Instruction * I)1020 static bool LdStHasDebugValue(const DILocalVariable *DIVar, Instruction *I) {
1021   // Since we can't guarantee that the original dbg.declare instrinsic
1022   // is removed by LowerDbgDeclare(), we need to make sure that we are
1023   // not inserting the same dbg.value intrinsic over and over.
1024   llvm::BasicBlock::InstListType::iterator PrevI(I);
1025   if (PrevI != I->getParent()->getInstList().begin()) {
1026     --PrevI;
1027     if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(PrevI))
1028       if (DVI->getValue() == I->getOperand(0) &&
1029           DVI->getOffset() == 0 &&
1030           DVI->getVariable() == DIVar)
1031         return true;
1032   }
1033   return false;
1034 }
1035 
1036 /// Inserts a llvm.dbg.value intrinsic before a store to an alloca'd value
1037 /// that has an associated llvm.dbg.decl intrinsic.
ConvertDebugDeclareToDebugValue(DbgDeclareInst * DDI,StoreInst * SI,DIBuilder & Builder)1038 bool llvm::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI,
1039                                            StoreInst *SI, DIBuilder &Builder) {
1040   auto *DIVar = DDI->getVariable();
1041   auto *DIExpr = DDI->getExpression();
1042   assert(DIVar && "Missing variable");
1043 
1044   if (LdStHasDebugValue(DIVar, SI))
1045     return true;
1046 
1047   // If an argument is zero extended then use argument directly. The ZExt
1048   // may be zapped by an optimization pass in future.
1049   Argument *ExtendedArg = nullptr;
1050   if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1051     ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0));
1052   if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1053     ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0));
1054   if (ExtendedArg)
1055     Builder.insertDbgValueIntrinsic(ExtendedArg, 0, DIVar, DIExpr,
1056                                     DDI->getDebugLoc(), SI);
1057   else
1058     Builder.insertDbgValueIntrinsic(SI->getOperand(0), 0, DIVar, DIExpr,
1059                                     DDI->getDebugLoc(), SI);
1060   return true;
1061 }
1062 
1063 /// Inserts a llvm.dbg.value intrinsic before a load of an alloca'd value
1064 /// that has an associated llvm.dbg.decl intrinsic.
ConvertDebugDeclareToDebugValue(DbgDeclareInst * DDI,LoadInst * LI,DIBuilder & Builder)1065 bool llvm::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI,
1066                                            LoadInst *LI, DIBuilder &Builder) {
1067   auto *DIVar = DDI->getVariable();
1068   auto *DIExpr = DDI->getExpression();
1069   assert(DIVar && "Missing variable");
1070 
1071   if (LdStHasDebugValue(DIVar, LI))
1072     return true;
1073 
1074   // We are now tracking the loaded value instead of the address. In the
1075   // future if multi-location support is added to the IR, it might be
1076   // preferable to keep tracking both the loaded value and the original
1077   // address in case the alloca can not be elided.
1078   Instruction *DbgValue = Builder.insertDbgValueIntrinsic(
1079       LI, 0, DIVar, DIExpr, DDI->getDebugLoc(), (Instruction *)nullptr);
1080   DbgValue->insertAfter(LI);
1081   return true;
1082 }
1083 
1084 /// Determine whether this alloca is either a VLA or an array.
isArray(AllocaInst * AI)1085 static bool isArray(AllocaInst *AI) {
1086   return AI->isArrayAllocation() ||
1087     AI->getType()->getElementType()->isArrayTy();
1088 }
1089 
1090 /// LowerDbgDeclare - Lowers llvm.dbg.declare intrinsics into appropriate set
1091 /// of llvm.dbg.value intrinsics.
LowerDbgDeclare(Function & F)1092 bool llvm::LowerDbgDeclare(Function &F) {
1093   DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false);
1094   SmallVector<DbgDeclareInst *, 4> Dbgs;
1095   for (auto &FI : F)
1096     for (Instruction &BI : FI)
1097       if (auto DDI = dyn_cast<DbgDeclareInst>(&BI))
1098         Dbgs.push_back(DDI);
1099 
1100   if (Dbgs.empty())
1101     return false;
1102 
1103   for (auto &I : Dbgs) {
1104     DbgDeclareInst *DDI = I;
1105     AllocaInst *AI = dyn_cast_or_null<AllocaInst>(DDI->getAddress());
1106     // If this is an alloca for a scalar variable, insert a dbg.value
1107     // at each load and store to the alloca and erase the dbg.declare.
1108     // The dbg.values allow tracking a variable even if it is not
1109     // stored on the stack, while the dbg.declare can only describe
1110     // the stack slot (and at a lexical-scope granularity). Later
1111     // passes will attempt to elide the stack slot.
1112     if (AI && !isArray(AI)) {
1113       for (User *U : AI->users())
1114         if (StoreInst *SI = dyn_cast<StoreInst>(U))
1115           ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1116         else if (LoadInst *LI = dyn_cast<LoadInst>(U))
1117           ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1118         else if (CallInst *CI = dyn_cast<CallInst>(U)) {
1119           // This is a call by-value or some other instruction that
1120           // takes a pointer to the variable. Insert a *value*
1121           // intrinsic that describes the alloca.
1122           SmallVector<uint64_t, 1> NewDIExpr;
1123           auto *DIExpr = DDI->getExpression();
1124           NewDIExpr.push_back(dwarf::DW_OP_deref);
1125           NewDIExpr.append(DIExpr->elements_begin(), DIExpr->elements_end());
1126           DIB.insertDbgValueIntrinsic(AI, 0, DDI->getVariable(),
1127                                       DIB.createExpression(NewDIExpr),
1128                                       DDI->getDebugLoc(), CI);
1129         }
1130       DDI->eraseFromParent();
1131     }
1132   }
1133   return true;
1134 }
1135 
1136 /// FindAllocaDbgDeclare - Finds the llvm.dbg.declare intrinsic describing the
1137 /// alloca 'V', if any.
FindAllocaDbgDeclare(Value * V)1138 DbgDeclareInst *llvm::FindAllocaDbgDeclare(Value *V) {
1139   if (auto *L = LocalAsMetadata::getIfExists(V))
1140     if (auto *MDV = MetadataAsValue::getIfExists(V->getContext(), L))
1141       for (User *U : MDV->users())
1142         if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
1143           return DDI;
1144 
1145   return nullptr;
1146 }
1147 
replaceDbgDeclare(Value * Address,Value * NewAddress,Instruction * InsertBefore,DIBuilder & Builder,bool Deref,int Offset)1148 bool llvm::replaceDbgDeclare(Value *Address, Value *NewAddress,
1149                              Instruction *InsertBefore, DIBuilder &Builder,
1150                              bool Deref, int Offset) {
1151   DbgDeclareInst *DDI = FindAllocaDbgDeclare(Address);
1152   if (!DDI)
1153     return false;
1154   DebugLoc Loc = DDI->getDebugLoc();
1155   auto *DIVar = DDI->getVariable();
1156   auto *DIExpr = DDI->getExpression();
1157   assert(DIVar && "Missing variable");
1158 
1159   if (Deref || Offset) {
1160     // Create a copy of the original DIDescriptor for user variable, prepending
1161     // "deref" operation to a list of address elements, as new llvm.dbg.declare
1162     // will take a value storing address of the memory for variable, not
1163     // alloca itself.
1164     SmallVector<uint64_t, 4> NewDIExpr;
1165     if (Deref)
1166       NewDIExpr.push_back(dwarf::DW_OP_deref);
1167     if (Offset > 0) {
1168       NewDIExpr.push_back(dwarf::DW_OP_plus);
1169       NewDIExpr.push_back(Offset);
1170     } else if (Offset < 0) {
1171       NewDIExpr.push_back(dwarf::DW_OP_minus);
1172       NewDIExpr.push_back(-Offset);
1173     }
1174     if (DIExpr)
1175       NewDIExpr.append(DIExpr->elements_begin(), DIExpr->elements_end());
1176     DIExpr = Builder.createExpression(NewDIExpr);
1177   }
1178 
1179   // Insert llvm.dbg.declare immediately after the original alloca, and remove
1180   // old llvm.dbg.declare.
1181   Builder.insertDeclare(NewAddress, DIVar, DIExpr, Loc, InsertBefore);
1182   DDI->eraseFromParent();
1183   return true;
1184 }
1185 
replaceDbgDeclareForAlloca(AllocaInst * AI,Value * NewAllocaAddress,DIBuilder & Builder,bool Deref,int Offset)1186 bool llvm::replaceDbgDeclareForAlloca(AllocaInst *AI, Value *NewAllocaAddress,
1187                                       DIBuilder &Builder, bool Deref, int Offset) {
1188   return replaceDbgDeclare(AI, NewAllocaAddress, AI->getNextNode(), Builder,
1189                            Deref, Offset);
1190 }
1191 
changeToUnreachable(Instruction * I,bool UseLLVMTrap)1192 void llvm::changeToUnreachable(Instruction *I, bool UseLLVMTrap) {
1193   BasicBlock *BB = I->getParent();
1194   // Loop over all of the successors, removing BB's entry from any PHI
1195   // nodes.
1196   for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI)
1197     (*SI)->removePredecessor(BB);
1198 
1199   // Insert a call to llvm.trap right before this.  This turns the undefined
1200   // behavior into a hard fail instead of falling through into random code.
1201   if (UseLLVMTrap) {
1202     Function *TrapFn =
1203       Intrinsic::getDeclaration(BB->getParent()->getParent(), Intrinsic::trap);
1204     CallInst *CallTrap = CallInst::Create(TrapFn, "", I);
1205     CallTrap->setDebugLoc(I->getDebugLoc());
1206   }
1207   new UnreachableInst(I->getContext(), I);
1208 
1209   // All instructions after this are dead.
1210   BasicBlock::iterator BBI = I->getIterator(), BBE = BB->end();
1211   while (BBI != BBE) {
1212     if (!BBI->use_empty())
1213       BBI->replaceAllUsesWith(UndefValue::get(BBI->getType()));
1214     BB->getInstList().erase(BBI++);
1215   }
1216 }
1217 
1218 /// changeToCall - Convert the specified invoke into a normal call.
changeToCall(InvokeInst * II)1219 static void changeToCall(InvokeInst *II) {
1220   SmallVector<Value*, 8> Args(II->arg_begin(), II->arg_end());
1221   SmallVector<OperandBundleDef, 1> OpBundles;
1222   II->getOperandBundlesAsDefs(OpBundles);
1223   CallInst *NewCall = CallInst::Create(II->getCalledValue(), Args, OpBundles,
1224                                        "", II);
1225   NewCall->takeName(II);
1226   NewCall->setCallingConv(II->getCallingConv());
1227   NewCall->setAttributes(II->getAttributes());
1228   NewCall->setDebugLoc(II->getDebugLoc());
1229   II->replaceAllUsesWith(NewCall);
1230 
1231   // Follow the call by a branch to the normal destination.
1232   BranchInst::Create(II->getNormalDest(), II);
1233 
1234   // Update PHI nodes in the unwind destination
1235   II->getUnwindDest()->removePredecessor(II->getParent());
1236   II->eraseFromParent();
1237 }
1238 
markAliveBlocks(Function & F,SmallPtrSetImpl<BasicBlock * > & Reachable)1239 static bool markAliveBlocks(Function &F,
1240                             SmallPtrSetImpl<BasicBlock*> &Reachable) {
1241 
1242   SmallVector<BasicBlock*, 128> Worklist;
1243   BasicBlock *BB = &F.front();
1244   Worklist.push_back(BB);
1245   Reachable.insert(BB);
1246   bool Changed = false;
1247   do {
1248     BB = Worklist.pop_back_val();
1249 
1250     // Do a quick scan of the basic block, turning any obviously unreachable
1251     // instructions into LLVM unreachable insts.  The instruction combining pass
1252     // canonicalizes unreachable insts into stores to null or undef.
1253     for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E;++BBI){
1254       // Assumptions that are known to be false are equivalent to unreachable.
1255       // Also, if the condition is undefined, then we make the choice most
1256       // beneficial to the optimizer, and choose that to also be unreachable.
1257       if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BBI))
1258         if (II->getIntrinsicID() == Intrinsic::assume) {
1259           bool MakeUnreachable = false;
1260           if (isa<UndefValue>(II->getArgOperand(0)))
1261             MakeUnreachable = true;
1262           else if (ConstantInt *Cond =
1263                    dyn_cast<ConstantInt>(II->getArgOperand(0)))
1264             MakeUnreachable = Cond->isZero();
1265 
1266           if (MakeUnreachable) {
1267             // Don't insert a call to llvm.trap right before the unreachable.
1268             changeToUnreachable(&*BBI, false);
1269             Changed = true;
1270             break;
1271           }
1272         }
1273 
1274       if (CallInst *CI = dyn_cast<CallInst>(BBI)) {
1275         if (CI->doesNotReturn()) {
1276           // If we found a call to a no-return function, insert an unreachable
1277           // instruction after it.  Make sure there isn't *already* one there
1278           // though.
1279           ++BBI;
1280           if (!isa<UnreachableInst>(BBI)) {
1281             // Don't insert a call to llvm.trap right before the unreachable.
1282             changeToUnreachable(&*BBI, false);
1283             Changed = true;
1284           }
1285           break;
1286         }
1287       }
1288 
1289       // Store to undef and store to null are undefined and used to signal that
1290       // they should be changed to unreachable by passes that can't modify the
1291       // CFG.
1292       if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
1293         // Don't touch volatile stores.
1294         if (SI->isVolatile()) continue;
1295 
1296         Value *Ptr = SI->getOperand(1);
1297 
1298         if (isa<UndefValue>(Ptr) ||
1299             (isa<ConstantPointerNull>(Ptr) &&
1300              SI->getPointerAddressSpace() == 0)) {
1301           changeToUnreachable(SI, true);
1302           Changed = true;
1303           break;
1304         }
1305       }
1306     }
1307 
1308     // Turn invokes that call 'nounwind' functions into ordinary calls.
1309     if (InvokeInst *II = dyn_cast<InvokeInst>(BB->getTerminator())) {
1310       Value *Callee = II->getCalledValue();
1311       if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
1312         changeToUnreachable(II, true);
1313         Changed = true;
1314       } else if (II->doesNotThrow() && canSimplifyInvokeNoUnwind(&F)) {
1315         if (II->use_empty() && II->onlyReadsMemory()) {
1316           // jump to the normal destination branch.
1317           BranchInst::Create(II->getNormalDest(), II);
1318           II->getUnwindDest()->removePredecessor(II->getParent());
1319           II->eraseFromParent();
1320         } else
1321           changeToCall(II);
1322         Changed = true;
1323       }
1324     }
1325 
1326     Changed |= ConstantFoldTerminator(BB, true);
1327     for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI)
1328       if (Reachable.insert(*SI).second)
1329         Worklist.push_back(*SI);
1330   } while (!Worklist.empty());
1331   return Changed;
1332 }
1333 
removeUnwindEdge(BasicBlock * BB)1334 void llvm::removeUnwindEdge(BasicBlock *BB) {
1335   TerminatorInst *TI = BB->getTerminator();
1336 
1337   if (auto *II = dyn_cast<InvokeInst>(TI)) {
1338     changeToCall(II);
1339     return;
1340   }
1341 
1342   TerminatorInst *NewTI;
1343   BasicBlock *UnwindDest;
1344 
1345   if (auto *CRI = dyn_cast<CleanupReturnInst>(TI)) {
1346     NewTI = CleanupReturnInst::Create(CRI->getCleanupPad(), nullptr, CRI);
1347     UnwindDest = CRI->getUnwindDest();
1348   } else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(TI)) {
1349     auto *NewCatchSwitch = CatchSwitchInst::Create(
1350         CatchSwitch->getParentPad(), nullptr, CatchSwitch->getNumHandlers(),
1351         CatchSwitch->getName(), CatchSwitch);
1352     for (BasicBlock *PadBB : CatchSwitch->handlers())
1353       NewCatchSwitch->addHandler(PadBB);
1354 
1355     NewTI = NewCatchSwitch;
1356     UnwindDest = CatchSwitch->getUnwindDest();
1357   } else {
1358     llvm_unreachable("Could not find unwind successor");
1359   }
1360 
1361   NewTI->takeName(TI);
1362   NewTI->setDebugLoc(TI->getDebugLoc());
1363   UnwindDest->removePredecessor(BB);
1364   TI->replaceAllUsesWith(NewTI);
1365   TI->eraseFromParent();
1366 }
1367 
1368 /// removeUnreachableBlocksFromFn - Remove blocks that are not reachable, even
1369 /// if they are in a dead cycle.  Return true if a change was made, false
1370 /// otherwise.
removeUnreachableBlocks(Function & F)1371 bool llvm::removeUnreachableBlocks(Function &F) {
1372   SmallPtrSet<BasicBlock*, 128> Reachable;
1373   bool Changed = markAliveBlocks(F, Reachable);
1374 
1375   // If there are unreachable blocks in the CFG...
1376   if (Reachable.size() == F.size())
1377     return Changed;
1378 
1379   assert(Reachable.size() < F.size());
1380   NumRemoved += F.size()-Reachable.size();
1381 
1382   // Loop over all of the basic blocks that are not reachable, dropping all of
1383   // their internal references...
1384   for (Function::iterator BB = ++F.begin(), E = F.end(); BB != E; ++BB) {
1385     if (Reachable.count(&*BB))
1386       continue;
1387 
1388     for (succ_iterator SI = succ_begin(&*BB), SE = succ_end(&*BB); SI != SE;
1389          ++SI)
1390       if (Reachable.count(*SI))
1391         (*SI)->removePredecessor(&*BB);
1392     BB->dropAllReferences();
1393   }
1394 
1395   for (Function::iterator I = ++F.begin(); I != F.end();)
1396     if (!Reachable.count(&*I))
1397       I = F.getBasicBlockList().erase(I);
1398     else
1399       ++I;
1400 
1401   return true;
1402 }
1403 
combineMetadata(Instruction * K,const Instruction * J,ArrayRef<unsigned> KnownIDs)1404 void llvm::combineMetadata(Instruction *K, const Instruction *J,
1405                            ArrayRef<unsigned> KnownIDs) {
1406   SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
1407   K->dropUnknownNonDebugMetadata(KnownIDs);
1408   K->getAllMetadataOtherThanDebugLoc(Metadata);
1409   for (unsigned i = 0, n = Metadata.size(); i < n; ++i) {
1410     unsigned Kind = Metadata[i].first;
1411     MDNode *JMD = J->getMetadata(Kind);
1412     MDNode *KMD = Metadata[i].second;
1413 
1414     switch (Kind) {
1415       default:
1416         K->setMetadata(Kind, nullptr); // Remove unknown metadata
1417         break;
1418       case LLVMContext::MD_dbg:
1419         llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg");
1420       case LLVMContext::MD_tbaa:
1421         K->setMetadata(Kind, MDNode::getMostGenericTBAA(JMD, KMD));
1422         break;
1423       case LLVMContext::MD_alias_scope:
1424         K->setMetadata(Kind, MDNode::getMostGenericAliasScope(JMD, KMD));
1425         break;
1426       case LLVMContext::MD_noalias:
1427         K->setMetadata(Kind, MDNode::intersect(JMD, KMD));
1428         break;
1429       case LLVMContext::MD_range:
1430         K->setMetadata(Kind, MDNode::getMostGenericRange(JMD, KMD));
1431         break;
1432       case LLVMContext::MD_fpmath:
1433         K->setMetadata(Kind, MDNode::getMostGenericFPMath(JMD, KMD));
1434         break;
1435       case LLVMContext::MD_invariant_load:
1436         // Only set the !invariant.load if it is present in both instructions.
1437         K->setMetadata(Kind, JMD);
1438         break;
1439       case LLVMContext::MD_nonnull:
1440         // Only set the !nonnull if it is present in both instructions.
1441         K->setMetadata(Kind, JMD);
1442         break;
1443       case LLVMContext::MD_invariant_group:
1444         // Preserve !invariant.group in K.
1445         break;
1446       case LLVMContext::MD_align:
1447         K->setMetadata(Kind,
1448           MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD));
1449         break;
1450       case LLVMContext::MD_dereferenceable:
1451       case LLVMContext::MD_dereferenceable_or_null:
1452         K->setMetadata(Kind,
1453           MDNode::getMostGenericAlignmentOrDereferenceable(JMD, KMD));
1454         break;
1455     }
1456   }
1457   // Set !invariant.group from J if J has it. If both instructions have it
1458   // then we will just pick it from J - even when they are different.
1459   // Also make sure that K is load or store - f.e. combining bitcast with load
1460   // could produce bitcast with invariant.group metadata, which is invalid.
1461   // FIXME: we should try to preserve both invariant.group md if they are
1462   // different, but right now instruction can only have one invariant.group.
1463   if (auto *JMD = J->getMetadata(LLVMContext::MD_invariant_group))
1464     if (isa<LoadInst>(K) || isa<StoreInst>(K))
1465       K->setMetadata(LLVMContext::MD_invariant_group, JMD);
1466 }
1467 
replaceDominatedUsesWith(Value * From,Value * To,DominatorTree & DT,const BasicBlockEdge & Root)1468 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To,
1469                                         DominatorTree &DT,
1470                                         const BasicBlockEdge &Root) {
1471   assert(From->getType() == To->getType());
1472 
1473   unsigned Count = 0;
1474   for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
1475        UI != UE; ) {
1476     Use &U = *UI++;
1477     if (DT.dominates(Root, U)) {
1478       U.set(To);
1479       DEBUG(dbgs() << "Replace dominated use of '"
1480             << From->getName() << "' as "
1481             << *To << " in " << *U << "\n");
1482       ++Count;
1483     }
1484   }
1485   return Count;
1486 }
1487 
replaceDominatedUsesWith(Value * From,Value * To,DominatorTree & DT,const BasicBlock * BB)1488 unsigned llvm::replaceDominatedUsesWith(Value *From, Value *To,
1489                                         DominatorTree &DT,
1490                                         const BasicBlock *BB) {
1491   assert(From->getType() == To->getType());
1492 
1493   unsigned Count = 0;
1494   for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
1495        UI != UE;) {
1496     Use &U = *UI++;
1497     auto *I = cast<Instruction>(U.getUser());
1498     if (DT.dominates(BB, I->getParent())) {
1499       U.set(To);
1500       DEBUG(dbgs() << "Replace dominated use of '" << From->getName() << "' as "
1501                    << *To << " in " << *U << "\n");
1502       ++Count;
1503     }
1504   }
1505   return Count;
1506 }
1507 
callsGCLeafFunction(ImmutableCallSite CS)1508 bool llvm::callsGCLeafFunction(ImmutableCallSite CS) {
1509   if (isa<IntrinsicInst>(CS.getInstruction()))
1510     // Most LLVM intrinsics are things which can never take a safepoint.
1511     // As a result, we don't need to have the stack parsable at the
1512     // callsite.  This is a highly useful optimization since intrinsic
1513     // calls are fairly prevalent, particularly in debug builds.
1514     return true;
1515 
1516   // Check if the function is specifically marked as a gc leaf function.
1517   //
1518   // TODO: we should be checking the attributes on the call site as well.
1519   if (const Function *F = CS.getCalledFunction())
1520     return F->hasFnAttribute("gc-leaf-function");
1521 
1522   return false;
1523 }
1524