xref: /freebsd/contrib/llvm-project/llvm/lib/Analysis/MemorySSAUpdater.cpp (revision ebacd8013fe5f7fdf9f6a5b286f6680dd2891036)
1 //===-- MemorySSAUpdater.cpp - Memory SSA Updater--------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------===//
8 //
9 // This file implements the MemorySSAUpdater class.
10 //
11 //===----------------------------------------------------------------===//
12 #include "llvm/Analysis/MemorySSAUpdater.h"
13 #include "llvm/ADT/STLExtras.h"
14 #include "llvm/ADT/SetVector.h"
15 #include "llvm/ADT/SmallPtrSet.h"
16 #include "llvm/Analysis/IteratedDominanceFrontier.h"
17 #include "llvm/Analysis/LoopIterator.h"
18 #include "llvm/Analysis/MemorySSA.h"
19 #include "llvm/IR/BasicBlock.h"
20 #include "llvm/IR/Dominators.h"
21 #include "llvm/Support/Debug.h"
22 #include <algorithm>
23 
24 #define DEBUG_TYPE "memoryssa"
25 using namespace llvm;
26 
27 // This is the marker algorithm from "Simple and Efficient Construction of
28 // Static Single Assignment Form"
29 // The simple, non-marker algorithm places phi nodes at any join
30 // Here, we place markers, and only place phi nodes if they end up necessary.
31 // They are only necessary if they break a cycle (IE we recursively visit
32 // ourselves again), or we discover, while getting the value of the operands,
33 // that there are two or more definitions needing to be merged.
34 // This still will leave non-minimal form in the case of irreducible control
35 // flow, where phi nodes may be in cycles with themselves, but unnecessary.
36 MemoryAccess *MemorySSAUpdater::getPreviousDefRecursive(
37     BasicBlock *BB,
38     DenseMap<BasicBlock *, TrackingVH<MemoryAccess>> &CachedPreviousDef) {
39   // First, do a cache lookup. Without this cache, certain CFG structures
40   // (like a series of if statements) take exponential time to visit.
41   auto Cached = CachedPreviousDef.find(BB);
42   if (Cached != CachedPreviousDef.end())
43     return Cached->second;
44 
45   // If this method is called from an unreachable block, return LoE.
46   if (!MSSA->DT->isReachableFromEntry(BB))
47     return MSSA->getLiveOnEntryDef();
48 
49   if (BasicBlock *Pred = BB->getUniquePredecessor()) {
50     VisitedBlocks.insert(BB);
51     // Single predecessor case, just recurse, we can only have one definition.
52     MemoryAccess *Result = getPreviousDefFromEnd(Pred, CachedPreviousDef);
53     CachedPreviousDef.insert({BB, Result});
54     return Result;
55   }
56 
57   if (VisitedBlocks.count(BB)) {
58     // We hit our node again, meaning we had a cycle, we must insert a phi
59     // node to break it so we have an operand. The only case this will
60     // insert useless phis is if we have irreducible control flow.
61     MemoryAccess *Result = MSSA->createMemoryPhi(BB);
62     CachedPreviousDef.insert({BB, Result});
63     return Result;
64   }
65 
66   if (VisitedBlocks.insert(BB).second) {
67     // Mark us visited so we can detect a cycle
68     SmallVector<TrackingVH<MemoryAccess>, 8> PhiOps;
69 
70     // Recurse to get the values in our predecessors for placement of a
71     // potential phi node. This will insert phi nodes if we cycle in order to
72     // break the cycle and have an operand.
73     bool UniqueIncomingAccess = true;
74     MemoryAccess *SingleAccess = nullptr;
75     for (auto *Pred : predecessors(BB)) {
76       if (MSSA->DT->isReachableFromEntry(Pred)) {
77         auto *IncomingAccess = getPreviousDefFromEnd(Pred, CachedPreviousDef);
78         if (!SingleAccess)
79           SingleAccess = IncomingAccess;
80         else if (IncomingAccess != SingleAccess)
81           UniqueIncomingAccess = false;
82         PhiOps.push_back(IncomingAccess);
83       } else
84         PhiOps.push_back(MSSA->getLiveOnEntryDef());
85     }
86 
87     // Now try to simplify the ops to avoid placing a phi.
88     // This may return null if we never created a phi yet, that's okay
89     MemoryPhi *Phi = dyn_cast_or_null<MemoryPhi>(MSSA->getMemoryAccess(BB));
90 
91     // See if we can avoid the phi by simplifying it.
92     auto *Result = tryRemoveTrivialPhi(Phi, PhiOps);
93     // If we couldn't simplify, we may have to create a phi
94     if (Result == Phi && UniqueIncomingAccess && SingleAccess) {
95       // A concrete Phi only exists if we created an empty one to break a cycle.
96       if (Phi) {
97         assert(Phi->operands().empty() && "Expected empty Phi");
98         Phi->replaceAllUsesWith(SingleAccess);
99         removeMemoryAccess(Phi);
100       }
101       Result = SingleAccess;
102     } else if (Result == Phi && !(UniqueIncomingAccess && SingleAccess)) {
103       if (!Phi)
104         Phi = MSSA->createMemoryPhi(BB);
105 
106       // See if the existing phi operands match what we need.
107       // Unlike normal SSA, we only allow one phi node per block, so we can't just
108       // create a new one.
109       if (Phi->getNumOperands() != 0) {
110         // FIXME: Figure out whether this is dead code and if so remove it.
111         if (!std::equal(Phi->op_begin(), Phi->op_end(), PhiOps.begin())) {
112           // These will have been filled in by the recursive read we did above.
113           llvm::copy(PhiOps, Phi->op_begin());
114           std::copy(pred_begin(BB), pred_end(BB), Phi->block_begin());
115         }
116       } else {
117         unsigned i = 0;
118         for (auto *Pred : predecessors(BB))
119           Phi->addIncoming(&*PhiOps[i++], Pred);
120         InsertedPHIs.push_back(Phi);
121       }
122       Result = Phi;
123     }
124 
125     // Set ourselves up for the next variable by resetting visited state.
126     VisitedBlocks.erase(BB);
127     CachedPreviousDef.insert({BB, Result});
128     return Result;
129   }
130   llvm_unreachable("Should have hit one of the three cases above");
131 }
132 
133 // This starts at the memory access, and goes backwards in the block to find the
134 // previous definition. If a definition is not found the block of the access,
135 // it continues globally, creating phi nodes to ensure we have a single
136 // definition.
137 MemoryAccess *MemorySSAUpdater::getPreviousDef(MemoryAccess *MA) {
138   if (auto *LocalResult = getPreviousDefInBlock(MA))
139     return LocalResult;
140   DenseMap<BasicBlock *, TrackingVH<MemoryAccess>> CachedPreviousDef;
141   return getPreviousDefRecursive(MA->getBlock(), CachedPreviousDef);
142 }
143 
144 // This starts at the memory access, and goes backwards in the block to the find
145 // the previous definition. If the definition is not found in the block of the
146 // access, it returns nullptr.
147 MemoryAccess *MemorySSAUpdater::getPreviousDefInBlock(MemoryAccess *MA) {
148   auto *Defs = MSSA->getWritableBlockDefs(MA->getBlock());
149 
150   // It's possible there are no defs, or we got handed the first def to start.
151   if (Defs) {
152     // If this is a def, we can just use the def iterators.
153     if (!isa<MemoryUse>(MA)) {
154       auto Iter = MA->getReverseDefsIterator();
155       ++Iter;
156       if (Iter != Defs->rend())
157         return &*Iter;
158     } else {
159       // Otherwise, have to walk the all access iterator.
160       auto End = MSSA->getWritableBlockAccesses(MA->getBlock())->rend();
161       for (auto &U : make_range(++MA->getReverseIterator(), End))
162         if (!isa<MemoryUse>(U))
163           return cast<MemoryAccess>(&U);
164       // Note that if MA comes before Defs->begin(), we won't hit a def.
165       return nullptr;
166     }
167   }
168   return nullptr;
169 }
170 
171 // This starts at the end of block
172 MemoryAccess *MemorySSAUpdater::getPreviousDefFromEnd(
173     BasicBlock *BB,
174     DenseMap<BasicBlock *, TrackingVH<MemoryAccess>> &CachedPreviousDef) {
175   auto *Defs = MSSA->getWritableBlockDefs(BB);
176 
177   if (Defs) {
178     CachedPreviousDef.insert({BB, &*Defs->rbegin()});
179     return &*Defs->rbegin();
180   }
181 
182   return getPreviousDefRecursive(BB, CachedPreviousDef);
183 }
184 // Recurse over a set of phi uses to eliminate the trivial ones
185 MemoryAccess *MemorySSAUpdater::recursePhi(MemoryAccess *Phi) {
186   if (!Phi)
187     return nullptr;
188   TrackingVH<MemoryAccess> Res(Phi);
189   SmallVector<TrackingVH<Value>, 8> Uses;
190   std::copy(Phi->user_begin(), Phi->user_end(), std::back_inserter(Uses));
191   for (auto &U : Uses)
192     if (MemoryPhi *UsePhi = dyn_cast<MemoryPhi>(&*U))
193       tryRemoveTrivialPhi(UsePhi);
194   return Res;
195 }
196 
197 // Eliminate trivial phis
198 // Phis are trivial if they are defined either by themselves, or all the same
199 // argument.
200 // IE phi(a, a) or b = phi(a, b) or c = phi(a, a, c)
201 // We recursively try to remove them.
202 MemoryAccess *MemorySSAUpdater::tryRemoveTrivialPhi(MemoryPhi *Phi) {
203   assert(Phi && "Can only remove concrete Phi.");
204   auto OperRange = Phi->operands();
205   return tryRemoveTrivialPhi(Phi, OperRange);
206 }
207 template <class RangeType>
208 MemoryAccess *MemorySSAUpdater::tryRemoveTrivialPhi(MemoryPhi *Phi,
209                                                     RangeType &Operands) {
210   // Bail out on non-opt Phis.
211   if (NonOptPhis.count(Phi))
212     return Phi;
213 
214   // Detect equal or self arguments
215   MemoryAccess *Same = nullptr;
216   for (auto &Op : Operands) {
217     // If the same or self, good so far
218     if (Op == Phi || Op == Same)
219       continue;
220     // not the same, return the phi since it's not eliminatable by us
221     if (Same)
222       return Phi;
223     Same = cast<MemoryAccess>(&*Op);
224   }
225   // Never found a non-self reference, the phi is undef
226   if (Same == nullptr)
227     return MSSA->getLiveOnEntryDef();
228   if (Phi) {
229     Phi->replaceAllUsesWith(Same);
230     removeMemoryAccess(Phi);
231   }
232 
233   // We should only end up recursing in case we replaced something, in which
234   // case, we may have made other Phis trivial.
235   return recursePhi(Same);
236 }
237 
238 void MemorySSAUpdater::insertUse(MemoryUse *MU, bool RenameUses) {
239   VisitedBlocks.clear();
240   InsertedPHIs.clear();
241   MU->setDefiningAccess(getPreviousDef(MU));
242 
243   // In cases without unreachable blocks, because uses do not create new
244   // may-defs, there are only two cases:
245   // 1. There was a def already below us, and therefore, we should not have
246   // created a phi node because it was already needed for the def.
247   //
248   // 2. There is no def below us, and therefore, there is no extra renaming work
249   // to do.
250 
251   // In cases with unreachable blocks, where the unnecessary Phis were
252   // optimized out, adding the Use may re-insert those Phis. Hence, when
253   // inserting Uses outside of the MSSA creation process, and new Phis were
254   // added, rename all uses if we are asked.
255 
256   if (!RenameUses && !InsertedPHIs.empty()) {
257     auto *Defs = MSSA->getBlockDefs(MU->getBlock());
258     (void)Defs;
259     assert((!Defs || (++Defs->begin() == Defs->end())) &&
260            "Block may have only a Phi or no defs");
261   }
262 
263   if (RenameUses && InsertedPHIs.size()) {
264     SmallPtrSet<BasicBlock *, 16> Visited;
265     BasicBlock *StartBlock = MU->getBlock();
266 
267     if (auto *Defs = MSSA->getWritableBlockDefs(StartBlock)) {
268       MemoryAccess *FirstDef = &*Defs->begin();
269       // Convert to incoming value if it's a memorydef. A phi *is* already an
270       // incoming value.
271       if (auto *MD = dyn_cast<MemoryDef>(FirstDef))
272         FirstDef = MD->getDefiningAccess();
273 
274       MSSA->renamePass(MU->getBlock(), FirstDef, Visited);
275     }
276     // We just inserted a phi into this block, so the incoming value will
277     // become the phi anyway, so it does not matter what we pass.
278     for (auto &MP : InsertedPHIs)
279       if (MemoryPhi *Phi = cast_or_null<MemoryPhi>(MP))
280         MSSA->renamePass(Phi->getBlock(), nullptr, Visited);
281   }
282 }
283 
284 // Set every incoming edge {BB, MP->getBlock()} of MemoryPhi MP to NewDef.
285 static void setMemoryPhiValueForBlock(MemoryPhi *MP, const BasicBlock *BB,
286                                       MemoryAccess *NewDef) {
287   // Replace any operand with us an incoming block with the new defining
288   // access.
289   int i = MP->getBasicBlockIndex(BB);
290   assert(i != -1 && "Should have found the basic block in the phi");
291   // We can't just compare i against getNumOperands since one is signed and the
292   // other not. So use it to index into the block iterator.
293   for (const BasicBlock *BlockBB : llvm::drop_begin(MP->blocks(), i)) {
294     if (BlockBB != BB)
295       break;
296     MP->setIncomingValue(i, NewDef);
297     ++i;
298   }
299 }
300 
301 // A brief description of the algorithm:
302 // First, we compute what should define the new def, using the SSA
303 // construction algorithm.
304 // Then, we update the defs below us (and any new phi nodes) in the graph to
305 // point to the correct new defs, to ensure we only have one variable, and no
306 // disconnected stores.
307 void MemorySSAUpdater::insertDef(MemoryDef *MD, bool RenameUses) {
308   // Don't bother updating dead code.
309   if (!MSSA->DT->isReachableFromEntry(MD->getBlock())) {
310     MD->setDefiningAccess(MSSA->getLiveOnEntryDef());
311     return;
312   }
313 
314   VisitedBlocks.clear();
315   InsertedPHIs.clear();
316 
317   // See if we had a local def, and if not, go hunting.
318   MemoryAccess *DefBefore = getPreviousDef(MD);
319   bool DefBeforeSameBlock = false;
320   if (DefBefore->getBlock() == MD->getBlock() &&
321       !(isa<MemoryPhi>(DefBefore) &&
322         llvm::is_contained(InsertedPHIs, DefBefore)))
323     DefBeforeSameBlock = true;
324 
325   // There is a def before us, which means we can replace any store/phi uses
326   // of that thing with us, since we are in the way of whatever was there
327   // before.
328   // We now define that def's memorydefs and memoryphis
329   if (DefBeforeSameBlock) {
330     DefBefore->replaceUsesWithIf(MD, [MD](Use &U) {
331       // Leave the MemoryUses alone.
332       // Also make sure we skip ourselves to avoid self references.
333       User *Usr = U.getUser();
334       return !isa<MemoryUse>(Usr) && Usr != MD;
335       // Defs are automatically unoptimized when the user is set to MD below,
336       // because the isOptimized() call will fail to find the same ID.
337     });
338   }
339 
340   // and that def is now our defining access.
341   MD->setDefiningAccess(DefBefore);
342 
343   SmallVector<WeakVH, 8> FixupList(InsertedPHIs.begin(), InsertedPHIs.end());
344 
345   SmallSet<WeakVH, 8> ExistingPhis;
346 
347   // Remember the index where we may insert new phis.
348   unsigned NewPhiIndex = InsertedPHIs.size();
349   if (!DefBeforeSameBlock) {
350     // If there was a local def before us, we must have the same effect it
351     // did. Because every may-def is the same, any phis/etc we would create, it
352     // would also have created.  If there was no local def before us, we
353     // performed a global update, and have to search all successors and make
354     // sure we update the first def in each of them (following all paths until
355     // we hit the first def along each path). This may also insert phi nodes.
356     // TODO: There are other cases we can skip this work, such as when we have a
357     // single successor, and only used a straight line of single pred blocks
358     // backwards to find the def.  To make that work, we'd have to track whether
359     // getDefRecursive only ever used the single predecessor case.  These types
360     // of paths also only exist in between CFG simplifications.
361 
362     // If this is the first def in the block and this insert is in an arbitrary
363     // place, compute IDF and place phis.
364     SmallPtrSet<BasicBlock *, 2> DefiningBlocks;
365 
366     // If this is the last Def in the block, we may need additional Phis.
367     // Compute IDF in all cases, as renaming needs to be done even when MD is
368     // not the last access, because it can introduce a new access past which a
369     // previous access was optimized; that access needs to be reoptimized.
370     DefiningBlocks.insert(MD->getBlock());
371     for (const auto &VH : InsertedPHIs)
372       if (const auto *RealPHI = cast_or_null<MemoryPhi>(VH))
373         DefiningBlocks.insert(RealPHI->getBlock());
374     ForwardIDFCalculator IDFs(*MSSA->DT);
375     SmallVector<BasicBlock *, 32> IDFBlocks;
376     IDFs.setDefiningBlocks(DefiningBlocks);
377     IDFs.calculate(IDFBlocks);
378     SmallVector<AssertingVH<MemoryPhi>, 4> NewInsertedPHIs;
379     for (auto *BBIDF : IDFBlocks) {
380       auto *MPhi = MSSA->getMemoryAccess(BBIDF);
381       if (!MPhi) {
382         MPhi = MSSA->createMemoryPhi(BBIDF);
383         NewInsertedPHIs.push_back(MPhi);
384       } else {
385         ExistingPhis.insert(MPhi);
386       }
387       // Add the phis created into the IDF blocks to NonOptPhis, so they are not
388       // optimized out as trivial by the call to getPreviousDefFromEnd below.
389       // Once they are complete, all these Phis are added to the FixupList, and
390       // removed from NonOptPhis inside fixupDefs(). Existing Phis in IDF may
391       // need fixing as well, and potentially be trivial before this insertion,
392       // hence add all IDF Phis. See PR43044.
393       NonOptPhis.insert(MPhi);
394     }
395     for (auto &MPhi : NewInsertedPHIs) {
396       auto *BBIDF = MPhi->getBlock();
397       for (auto *Pred : predecessors(BBIDF)) {
398         DenseMap<BasicBlock *, TrackingVH<MemoryAccess>> CachedPreviousDef;
399         MPhi->addIncoming(getPreviousDefFromEnd(Pred, CachedPreviousDef), Pred);
400       }
401     }
402 
403     // Re-take the index where we're adding the new phis, because the above call
404     // to getPreviousDefFromEnd, may have inserted into InsertedPHIs.
405     NewPhiIndex = InsertedPHIs.size();
406     for (auto &MPhi : NewInsertedPHIs) {
407       InsertedPHIs.push_back(&*MPhi);
408       FixupList.push_back(&*MPhi);
409     }
410 
411     FixupList.push_back(MD);
412   }
413 
414   // Remember the index where we stopped inserting new phis above, since the
415   // fixupDefs call in the loop below may insert more, that are already minimal.
416   unsigned NewPhiIndexEnd = InsertedPHIs.size();
417 
418   while (!FixupList.empty()) {
419     unsigned StartingPHISize = InsertedPHIs.size();
420     fixupDefs(FixupList);
421     FixupList.clear();
422     // Put any new phis on the fixup list, and process them
423     FixupList.append(InsertedPHIs.begin() + StartingPHISize, InsertedPHIs.end());
424   }
425 
426   // Optimize potentially non-minimal phis added in this method.
427   unsigned NewPhiSize = NewPhiIndexEnd - NewPhiIndex;
428   if (NewPhiSize)
429     tryRemoveTrivialPhis(ArrayRef<WeakVH>(&InsertedPHIs[NewPhiIndex], NewPhiSize));
430 
431   // Now that all fixups are done, rename all uses if we are asked. The defs are
432   // guaranteed to be in reachable code due to the check at the method entry.
433   BasicBlock *StartBlock = MD->getBlock();
434   if (RenameUses) {
435     SmallPtrSet<BasicBlock *, 16> Visited;
436     // We are guaranteed there is a def in the block, because we just got it
437     // handed to us in this function.
438     MemoryAccess *FirstDef = &*MSSA->getWritableBlockDefs(StartBlock)->begin();
439     // Convert to incoming value if it's a memorydef. A phi *is* already an
440     // incoming value.
441     if (auto *MD = dyn_cast<MemoryDef>(FirstDef))
442       FirstDef = MD->getDefiningAccess();
443 
444     MSSA->renamePass(MD->getBlock(), FirstDef, Visited);
445     // We just inserted a phi into this block, so the incoming value will become
446     // the phi anyway, so it does not matter what we pass.
447     for (auto &MP : InsertedPHIs) {
448       MemoryPhi *Phi = dyn_cast_or_null<MemoryPhi>(MP);
449       if (Phi)
450         MSSA->renamePass(Phi->getBlock(), nullptr, Visited);
451     }
452     // Existing Phi blocks may need renaming too, if an access was previously
453     // optimized and the inserted Defs "covers" the Optimized value.
454     for (const auto &MP : ExistingPhis) {
455       MemoryPhi *Phi = dyn_cast_or_null<MemoryPhi>(MP);
456       if (Phi)
457         MSSA->renamePass(Phi->getBlock(), nullptr, Visited);
458     }
459   }
460 }
461 
462 void MemorySSAUpdater::fixupDefs(const SmallVectorImpl<WeakVH> &Vars) {
463   SmallPtrSet<const BasicBlock *, 8> Seen;
464   SmallVector<const BasicBlock *, 16> Worklist;
465   for (const auto &Var : Vars) {
466     MemoryAccess *NewDef = dyn_cast_or_null<MemoryAccess>(Var);
467     if (!NewDef)
468       continue;
469     // First, see if there is a local def after the operand.
470     auto *Defs = MSSA->getWritableBlockDefs(NewDef->getBlock());
471     auto DefIter = NewDef->getDefsIterator();
472 
473     // The temporary Phi is being fixed, unmark it for not to optimize.
474     if (MemoryPhi *Phi = dyn_cast<MemoryPhi>(NewDef))
475       NonOptPhis.erase(Phi);
476 
477     // If there is a local def after us, we only have to rename that.
478     if (++DefIter != Defs->end()) {
479       cast<MemoryDef>(DefIter)->setDefiningAccess(NewDef);
480       continue;
481     }
482 
483     // Otherwise, we need to search down through the CFG.
484     // For each of our successors, handle it directly if their is a phi, or
485     // place on the fixup worklist.
486     for (const auto *S : successors(NewDef->getBlock())) {
487       if (auto *MP = MSSA->getMemoryAccess(S))
488         setMemoryPhiValueForBlock(MP, NewDef->getBlock(), NewDef);
489       else
490         Worklist.push_back(S);
491     }
492 
493     while (!Worklist.empty()) {
494       const BasicBlock *FixupBlock = Worklist.pop_back_val();
495 
496       // Get the first def in the block that isn't a phi node.
497       if (auto *Defs = MSSA->getWritableBlockDefs(FixupBlock)) {
498         auto *FirstDef = &*Defs->begin();
499         // The loop above and below should have taken care of phi nodes
500         assert(!isa<MemoryPhi>(FirstDef) &&
501                "Should have already handled phi nodes!");
502         // We are now this def's defining access, make sure we actually dominate
503         // it
504         assert(MSSA->dominates(NewDef, FirstDef) &&
505                "Should have dominated the new access");
506 
507         // This may insert new phi nodes, because we are not guaranteed the
508         // block we are processing has a single pred, and depending where the
509         // store was inserted, it may require phi nodes below it.
510         cast<MemoryDef>(FirstDef)->setDefiningAccess(getPreviousDef(FirstDef));
511         return;
512       }
513       // We didn't find a def, so we must continue.
514       for (const auto *S : successors(FixupBlock)) {
515         // If there is a phi node, handle it.
516         // Otherwise, put the block on the worklist
517         if (auto *MP = MSSA->getMemoryAccess(S))
518           setMemoryPhiValueForBlock(MP, FixupBlock, NewDef);
519         else {
520           // If we cycle, we should have ended up at a phi node that we already
521           // processed.  FIXME: Double check this
522           if (!Seen.insert(S).second)
523             continue;
524           Worklist.push_back(S);
525         }
526       }
527     }
528   }
529 }
530 
531 void MemorySSAUpdater::removeEdge(BasicBlock *From, BasicBlock *To) {
532   if (MemoryPhi *MPhi = MSSA->getMemoryAccess(To)) {
533     MPhi->unorderedDeleteIncomingBlock(From);
534     tryRemoveTrivialPhi(MPhi);
535   }
536 }
537 
538 void MemorySSAUpdater::removeDuplicatePhiEdgesBetween(const BasicBlock *From,
539                                                       const BasicBlock *To) {
540   if (MemoryPhi *MPhi = MSSA->getMemoryAccess(To)) {
541     bool Found = false;
542     MPhi->unorderedDeleteIncomingIf([&](const MemoryAccess *, BasicBlock *B) {
543       if (From != B)
544         return false;
545       if (Found)
546         return true;
547       Found = true;
548       return false;
549     });
550     tryRemoveTrivialPhi(MPhi);
551   }
552 }
553 
554 /// If all arguments of a MemoryPHI are defined by the same incoming
555 /// argument, return that argument.
556 static MemoryAccess *onlySingleValue(MemoryPhi *MP) {
557   MemoryAccess *MA = nullptr;
558 
559   for (auto &Arg : MP->operands()) {
560     if (!MA)
561       MA = cast<MemoryAccess>(Arg);
562     else if (MA != Arg)
563       return nullptr;
564   }
565   return MA;
566 }
567 
568 static MemoryAccess *getNewDefiningAccessForClone(MemoryAccess *MA,
569                                                   const ValueToValueMapTy &VMap,
570                                                   PhiToDefMap &MPhiMap,
571                                                   bool CloneWasSimplified,
572                                                   MemorySSA *MSSA) {
573   MemoryAccess *InsnDefining = MA;
574   if (MemoryDef *DefMUD = dyn_cast<MemoryDef>(InsnDefining)) {
575     if (!MSSA->isLiveOnEntryDef(DefMUD)) {
576       Instruction *DefMUDI = DefMUD->getMemoryInst();
577       assert(DefMUDI && "Found MemoryUseOrDef with no Instruction.");
578       if (Instruction *NewDefMUDI =
579               cast_or_null<Instruction>(VMap.lookup(DefMUDI))) {
580         InsnDefining = MSSA->getMemoryAccess(NewDefMUDI);
581         if (!CloneWasSimplified)
582           assert(InsnDefining && "Defining instruction cannot be nullptr.");
583         else if (!InsnDefining || isa<MemoryUse>(InsnDefining)) {
584           // The clone was simplified, it's no longer a MemoryDef, look up.
585           auto DefIt = DefMUD->getDefsIterator();
586           // Since simplified clones only occur in single block cloning, a
587           // previous definition must exist, otherwise NewDefMUDI would not
588           // have been found in VMap.
589           assert(DefIt != MSSA->getBlockDefs(DefMUD->getBlock())->begin() &&
590                  "Previous def must exist");
591           InsnDefining = getNewDefiningAccessForClone(
592               &*(--DefIt), VMap, MPhiMap, CloneWasSimplified, MSSA);
593         }
594       }
595     }
596   } else {
597     MemoryPhi *DefPhi = cast<MemoryPhi>(InsnDefining);
598     if (MemoryAccess *NewDefPhi = MPhiMap.lookup(DefPhi))
599       InsnDefining = NewDefPhi;
600   }
601   assert(InsnDefining && "Defining instruction cannot be nullptr.");
602   return InsnDefining;
603 }
604 
605 void MemorySSAUpdater::cloneUsesAndDefs(BasicBlock *BB, BasicBlock *NewBB,
606                                         const ValueToValueMapTy &VMap,
607                                         PhiToDefMap &MPhiMap,
608                                         bool CloneWasSimplified) {
609   const MemorySSA::AccessList *Acc = MSSA->getBlockAccesses(BB);
610   if (!Acc)
611     return;
612   for (const MemoryAccess &MA : *Acc) {
613     if (const MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(&MA)) {
614       Instruction *Insn = MUD->getMemoryInst();
615       // Entry does not exist if the clone of the block did not clone all
616       // instructions. This occurs in LoopRotate when cloning instructions
617       // from the old header to the old preheader. The cloned instruction may
618       // also be a simplified Value, not an Instruction (see LoopRotate).
619       // Also in LoopRotate, even when it's an instruction, due to it being
620       // simplified, it may be a Use rather than a Def, so we cannot use MUD as
621       // template. Calls coming from updateForClonedBlockIntoPred, ensure this.
622       if (Instruction *NewInsn =
623               dyn_cast_or_null<Instruction>(VMap.lookup(Insn))) {
624         MemoryAccess *NewUseOrDef = MSSA->createDefinedAccess(
625             NewInsn,
626             getNewDefiningAccessForClone(MUD->getDefiningAccess(), VMap,
627                                          MPhiMap, CloneWasSimplified, MSSA),
628             /*Template=*/CloneWasSimplified ? nullptr : MUD,
629             /*CreationMustSucceed=*/CloneWasSimplified ? false : true);
630         if (NewUseOrDef)
631           MSSA->insertIntoListsForBlock(NewUseOrDef, NewBB, MemorySSA::End);
632       }
633     }
634   }
635 }
636 
637 void MemorySSAUpdater::updatePhisWhenInsertingUniqueBackedgeBlock(
638     BasicBlock *Header, BasicBlock *Preheader, BasicBlock *BEBlock) {
639   auto *MPhi = MSSA->getMemoryAccess(Header);
640   if (!MPhi)
641     return;
642 
643   // Create phi node in the backedge block and populate it with the same
644   // incoming values as MPhi. Skip incoming values coming from Preheader.
645   auto *NewMPhi = MSSA->createMemoryPhi(BEBlock);
646   bool HasUniqueIncomingValue = true;
647   MemoryAccess *UniqueValue = nullptr;
648   for (unsigned I = 0, E = MPhi->getNumIncomingValues(); I != E; ++I) {
649     BasicBlock *IBB = MPhi->getIncomingBlock(I);
650     MemoryAccess *IV = MPhi->getIncomingValue(I);
651     if (IBB != Preheader) {
652       NewMPhi->addIncoming(IV, IBB);
653       if (HasUniqueIncomingValue) {
654         if (!UniqueValue)
655           UniqueValue = IV;
656         else if (UniqueValue != IV)
657           HasUniqueIncomingValue = false;
658       }
659     }
660   }
661 
662   // Update incoming edges into MPhi. Remove all but the incoming edge from
663   // Preheader. Add an edge from NewMPhi
664   auto *AccFromPreheader = MPhi->getIncomingValueForBlock(Preheader);
665   MPhi->setIncomingValue(0, AccFromPreheader);
666   MPhi->setIncomingBlock(0, Preheader);
667   for (unsigned I = MPhi->getNumIncomingValues() - 1; I >= 1; --I)
668     MPhi->unorderedDeleteIncoming(I);
669   MPhi->addIncoming(NewMPhi, BEBlock);
670 
671   // If NewMPhi is a trivial phi, remove it. Its use in the header MPhi will be
672   // replaced with the unique value.
673   tryRemoveTrivialPhi(NewMPhi);
674 }
675 
676 void MemorySSAUpdater::updateForClonedLoop(const LoopBlocksRPO &LoopBlocks,
677                                            ArrayRef<BasicBlock *> ExitBlocks,
678                                            const ValueToValueMapTy &VMap,
679                                            bool IgnoreIncomingWithNoClones) {
680   PhiToDefMap MPhiMap;
681 
682   auto FixPhiIncomingValues = [&](MemoryPhi *Phi, MemoryPhi *NewPhi) {
683     assert(Phi && NewPhi && "Invalid Phi nodes.");
684     BasicBlock *NewPhiBB = NewPhi->getBlock();
685     SmallPtrSet<BasicBlock *, 4> NewPhiBBPreds(pred_begin(NewPhiBB),
686                                                pred_end(NewPhiBB));
687     for (unsigned It = 0, E = Phi->getNumIncomingValues(); It < E; ++It) {
688       MemoryAccess *IncomingAccess = Phi->getIncomingValue(It);
689       BasicBlock *IncBB = Phi->getIncomingBlock(It);
690 
691       if (BasicBlock *NewIncBB = cast_or_null<BasicBlock>(VMap.lookup(IncBB)))
692         IncBB = NewIncBB;
693       else if (IgnoreIncomingWithNoClones)
694         continue;
695 
696       // Now we have IncBB, and will need to add incoming from it to NewPhi.
697 
698       // If IncBB is not a predecessor of NewPhiBB, then do not add it.
699       // NewPhiBB was cloned without that edge.
700       if (!NewPhiBBPreds.count(IncBB))
701         continue;
702 
703       // Determine incoming value and add it as incoming from IncBB.
704       if (MemoryUseOrDef *IncMUD = dyn_cast<MemoryUseOrDef>(IncomingAccess)) {
705         if (!MSSA->isLiveOnEntryDef(IncMUD)) {
706           Instruction *IncI = IncMUD->getMemoryInst();
707           assert(IncI && "Found MemoryUseOrDef with no Instruction.");
708           if (Instruction *NewIncI =
709                   cast_or_null<Instruction>(VMap.lookup(IncI))) {
710             IncMUD = MSSA->getMemoryAccess(NewIncI);
711             assert(IncMUD &&
712                    "MemoryUseOrDef cannot be null, all preds processed.");
713           }
714         }
715         NewPhi->addIncoming(IncMUD, IncBB);
716       } else {
717         MemoryPhi *IncPhi = cast<MemoryPhi>(IncomingAccess);
718         if (MemoryAccess *NewDefPhi = MPhiMap.lookup(IncPhi))
719           NewPhi->addIncoming(NewDefPhi, IncBB);
720         else
721           NewPhi->addIncoming(IncPhi, IncBB);
722       }
723     }
724     if (auto *SingleAccess = onlySingleValue(NewPhi)) {
725       MPhiMap[Phi] = SingleAccess;
726       removeMemoryAccess(NewPhi);
727     }
728   };
729 
730   auto ProcessBlock = [&](BasicBlock *BB) {
731     BasicBlock *NewBlock = cast_or_null<BasicBlock>(VMap.lookup(BB));
732     if (!NewBlock)
733       return;
734 
735     assert(!MSSA->getWritableBlockAccesses(NewBlock) &&
736            "Cloned block should have no accesses");
737 
738     // Add MemoryPhi.
739     if (MemoryPhi *MPhi = MSSA->getMemoryAccess(BB)) {
740       MemoryPhi *NewPhi = MSSA->createMemoryPhi(NewBlock);
741       MPhiMap[MPhi] = NewPhi;
742     }
743     // Update Uses and Defs.
744     cloneUsesAndDefs(BB, NewBlock, VMap, MPhiMap);
745   };
746 
747   for (auto *BB : llvm::concat<BasicBlock *const>(LoopBlocks, ExitBlocks))
748     ProcessBlock(BB);
749 
750   for (auto *BB : llvm::concat<BasicBlock *const>(LoopBlocks, ExitBlocks))
751     if (MemoryPhi *MPhi = MSSA->getMemoryAccess(BB))
752       if (MemoryAccess *NewPhi = MPhiMap.lookup(MPhi))
753         FixPhiIncomingValues(MPhi, cast<MemoryPhi>(NewPhi));
754 }
755 
756 void MemorySSAUpdater::updateForClonedBlockIntoPred(
757     BasicBlock *BB, BasicBlock *P1, const ValueToValueMapTy &VM) {
758   // All defs/phis from outside BB that are used in BB, are valid uses in P1.
759   // Since those defs/phis must have dominated BB, and also dominate P1.
760   // Defs from BB being used in BB will be replaced with the cloned defs from
761   // VM. The uses of BB's Phi (if it exists) in BB will be replaced by the
762   // incoming def into the Phi from P1.
763   // Instructions cloned into the predecessor are in practice sometimes
764   // simplified, so disable the use of the template, and create an access from
765   // scratch.
766   PhiToDefMap MPhiMap;
767   if (MemoryPhi *MPhi = MSSA->getMemoryAccess(BB))
768     MPhiMap[MPhi] = MPhi->getIncomingValueForBlock(P1);
769   cloneUsesAndDefs(BB, P1, VM, MPhiMap, /*CloneWasSimplified=*/true);
770 }
771 
772 template <typename Iter>
773 void MemorySSAUpdater::privateUpdateExitBlocksForClonedLoop(
774     ArrayRef<BasicBlock *> ExitBlocks, Iter ValuesBegin, Iter ValuesEnd,
775     DominatorTree &DT) {
776   SmallVector<CFGUpdate, 4> Updates;
777   // Update/insert phis in all successors of exit blocks.
778   for (auto *Exit : ExitBlocks)
779     for (const ValueToValueMapTy *VMap : make_range(ValuesBegin, ValuesEnd))
780       if (BasicBlock *NewExit = cast_or_null<BasicBlock>(VMap->lookup(Exit))) {
781         BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
782         Updates.push_back({DT.Insert, NewExit, ExitSucc});
783       }
784   applyInsertUpdates(Updates, DT);
785 }
786 
787 void MemorySSAUpdater::updateExitBlocksForClonedLoop(
788     ArrayRef<BasicBlock *> ExitBlocks, const ValueToValueMapTy &VMap,
789     DominatorTree &DT) {
790   const ValueToValueMapTy *const Arr[] = {&VMap};
791   privateUpdateExitBlocksForClonedLoop(ExitBlocks, std::begin(Arr),
792                                        std::end(Arr), DT);
793 }
794 
795 void MemorySSAUpdater::updateExitBlocksForClonedLoop(
796     ArrayRef<BasicBlock *> ExitBlocks,
797     ArrayRef<std::unique_ptr<ValueToValueMapTy>> VMaps, DominatorTree &DT) {
798   auto GetPtr = [&](const std::unique_ptr<ValueToValueMapTy> &I) {
799     return I.get();
800   };
801   using MappedIteratorType =
802       mapped_iterator<const std::unique_ptr<ValueToValueMapTy> *,
803                       decltype(GetPtr)>;
804   auto MapBegin = MappedIteratorType(VMaps.begin(), GetPtr);
805   auto MapEnd = MappedIteratorType(VMaps.end(), GetPtr);
806   privateUpdateExitBlocksForClonedLoop(ExitBlocks, MapBegin, MapEnd, DT);
807 }
808 
809 void MemorySSAUpdater::applyUpdates(ArrayRef<CFGUpdate> Updates,
810                                     DominatorTree &DT, bool UpdateDT) {
811   SmallVector<CFGUpdate, 4> DeleteUpdates;
812   SmallVector<CFGUpdate, 4> RevDeleteUpdates;
813   SmallVector<CFGUpdate, 4> InsertUpdates;
814   for (const auto &Update : Updates) {
815     if (Update.getKind() == DT.Insert)
816       InsertUpdates.push_back({DT.Insert, Update.getFrom(), Update.getTo()});
817     else {
818       DeleteUpdates.push_back({DT.Delete, Update.getFrom(), Update.getTo()});
819       RevDeleteUpdates.push_back({DT.Insert, Update.getFrom(), Update.getTo()});
820     }
821   }
822 
823   if (!DeleteUpdates.empty()) {
824     if (!InsertUpdates.empty()) {
825       if (!UpdateDT) {
826         SmallVector<CFGUpdate, 0> Empty;
827         // Deletes are reversed applied, because this CFGView is pretending the
828         // deletes did not happen yet, hence the edges still exist.
829         DT.applyUpdates(Empty, RevDeleteUpdates);
830       } else {
831         // Apply all updates, with the RevDeleteUpdates as PostCFGView.
832         DT.applyUpdates(Updates, RevDeleteUpdates);
833       }
834 
835       // Note: the MSSA update below doesn't distinguish between a GD with
836       // (RevDelete,false) and (Delete, true), but this matters for the DT
837       // updates above; for "children" purposes they are equivalent; but the
838       // updates themselves convey the desired update, used inside DT only.
839       GraphDiff<BasicBlock *> GD(RevDeleteUpdates);
840       applyInsertUpdates(InsertUpdates, DT, &GD);
841       // Update DT to redelete edges; this matches the real CFG so we can
842       // perform the standard update without a postview of the CFG.
843       DT.applyUpdates(DeleteUpdates);
844     } else {
845       if (UpdateDT)
846         DT.applyUpdates(DeleteUpdates);
847     }
848   } else {
849     if (UpdateDT)
850       DT.applyUpdates(Updates);
851     GraphDiff<BasicBlock *> GD;
852     applyInsertUpdates(InsertUpdates, DT, &GD);
853   }
854 
855   // Update for deleted edges
856   for (auto &Update : DeleteUpdates)
857     removeEdge(Update.getFrom(), Update.getTo());
858 }
859 
860 void MemorySSAUpdater::applyInsertUpdates(ArrayRef<CFGUpdate> Updates,
861                                           DominatorTree &DT) {
862   GraphDiff<BasicBlock *> GD;
863   applyInsertUpdates(Updates, DT, &GD);
864 }
865 
866 void MemorySSAUpdater::applyInsertUpdates(ArrayRef<CFGUpdate> Updates,
867                                           DominatorTree &DT,
868                                           const GraphDiff<BasicBlock *> *GD) {
869   // Get recursive last Def, assuming well formed MSSA and updated DT.
870   auto GetLastDef = [&](BasicBlock *BB) -> MemoryAccess * {
871     while (true) {
872       MemorySSA::DefsList *Defs = MSSA->getWritableBlockDefs(BB);
873       // Return last Def or Phi in BB, if it exists.
874       if (Defs)
875         return &*(--Defs->end());
876 
877       // Check number of predecessors, we only care if there's more than one.
878       unsigned Count = 0;
879       BasicBlock *Pred = nullptr;
880       for (auto *Pi : GD->template getChildren</*InverseEdge=*/true>(BB)) {
881         Pred = Pi;
882         Count++;
883         if (Count == 2)
884           break;
885       }
886 
887       // If BB has multiple predecessors, get last definition from IDom.
888       if (Count != 1) {
889         // [SimpleLoopUnswitch] If BB is a dead block, about to be deleted, its
890         // DT is invalidated. Return LoE as its last def. This will be added to
891         // MemoryPhi node, and later deleted when the block is deleted.
892         if (!DT.getNode(BB))
893           return MSSA->getLiveOnEntryDef();
894         if (auto *IDom = DT.getNode(BB)->getIDom())
895           if (IDom->getBlock() != BB) {
896             BB = IDom->getBlock();
897             continue;
898           }
899         return MSSA->getLiveOnEntryDef();
900       } else {
901         // Single predecessor, BB cannot be dead. GetLastDef of Pred.
902         assert(Count == 1 && Pred && "Single predecessor expected.");
903         // BB can be unreachable though, return LoE if that is the case.
904         if (!DT.getNode(BB))
905           return MSSA->getLiveOnEntryDef();
906         BB = Pred;
907       }
908     };
909     llvm_unreachable("Unable to get last definition.");
910   };
911 
912   // Get nearest IDom given a set of blocks.
913   // TODO: this can be optimized by starting the search at the node with the
914   // lowest level (highest in the tree).
915   auto FindNearestCommonDominator =
916       [&](const SmallSetVector<BasicBlock *, 2> &BBSet) -> BasicBlock * {
917     BasicBlock *PrevIDom = *BBSet.begin();
918     for (auto *BB : BBSet)
919       PrevIDom = DT.findNearestCommonDominator(PrevIDom, BB);
920     return PrevIDom;
921   };
922 
923   // Get all blocks that dominate PrevIDom, stop when reaching CurrIDom. Do not
924   // include CurrIDom.
925   auto GetNoLongerDomBlocks =
926       [&](BasicBlock *PrevIDom, BasicBlock *CurrIDom,
927           SmallVectorImpl<BasicBlock *> &BlocksPrevDom) {
928         if (PrevIDom == CurrIDom)
929           return;
930         BlocksPrevDom.push_back(PrevIDom);
931         BasicBlock *NextIDom = PrevIDom;
932         while (BasicBlock *UpIDom =
933                    DT.getNode(NextIDom)->getIDom()->getBlock()) {
934           if (UpIDom == CurrIDom)
935             break;
936           BlocksPrevDom.push_back(UpIDom);
937           NextIDom = UpIDom;
938         }
939       };
940 
941   // Map a BB to its predecessors: added + previously existing. To get a
942   // deterministic order, store predecessors as SetVectors. The order in each
943   // will be defined by the order in Updates (fixed) and the order given by
944   // children<> (also fixed). Since we further iterate over these ordered sets,
945   // we lose the information of multiple edges possibly existing between two
946   // blocks, so we'll keep and EdgeCount map for that.
947   // An alternate implementation could keep unordered set for the predecessors,
948   // traverse either Updates or children<> each time to get  the deterministic
949   // order, and drop the usage of EdgeCount. This alternate approach would still
950   // require querying the maps for each predecessor, and children<> call has
951   // additional computation inside for creating the snapshot-graph predecessors.
952   // As such, we favor using a little additional storage and less compute time.
953   // This decision can be revisited if we find the alternative more favorable.
954 
955   struct PredInfo {
956     SmallSetVector<BasicBlock *, 2> Added;
957     SmallSetVector<BasicBlock *, 2> Prev;
958   };
959   SmallDenseMap<BasicBlock *, PredInfo> PredMap;
960 
961   for (const auto &Edge : Updates) {
962     BasicBlock *BB = Edge.getTo();
963     auto &AddedBlockSet = PredMap[BB].Added;
964     AddedBlockSet.insert(Edge.getFrom());
965   }
966 
967   // Store all existing predecessor for each BB, at least one must exist.
968   SmallDenseMap<std::pair<BasicBlock *, BasicBlock *>, int> EdgeCountMap;
969   SmallPtrSet<BasicBlock *, 2> NewBlocks;
970   for (auto &BBPredPair : PredMap) {
971     auto *BB = BBPredPair.first;
972     const auto &AddedBlockSet = BBPredPair.second.Added;
973     auto &PrevBlockSet = BBPredPair.second.Prev;
974     for (auto *Pi : GD->template getChildren</*InverseEdge=*/true>(BB)) {
975       if (!AddedBlockSet.count(Pi))
976         PrevBlockSet.insert(Pi);
977       EdgeCountMap[{Pi, BB}]++;
978     }
979 
980     if (PrevBlockSet.empty()) {
981       assert(pred_size(BB) == AddedBlockSet.size() && "Duplicate edges added.");
982       LLVM_DEBUG(
983           dbgs()
984           << "Adding a predecessor to a block with no predecessors. "
985              "This must be an edge added to a new, likely cloned, block. "
986              "Its memory accesses must be already correct, assuming completed "
987              "via the updateExitBlocksForClonedLoop API. "
988              "Assert a single such edge is added so no phi addition or "
989              "additional processing is required.\n");
990       assert(AddedBlockSet.size() == 1 &&
991              "Can only handle adding one predecessor to a new block.");
992       // Need to remove new blocks from PredMap. Remove below to not invalidate
993       // iterator here.
994       NewBlocks.insert(BB);
995     }
996   }
997   // Nothing to process for new/cloned blocks.
998   for (auto *BB : NewBlocks)
999     PredMap.erase(BB);
1000 
1001   SmallVector<BasicBlock *, 16> BlocksWithDefsToReplace;
1002   SmallVector<WeakVH, 8> InsertedPhis;
1003 
1004   // First create MemoryPhis in all blocks that don't have one. Create in the
1005   // order found in Updates, not in PredMap, to get deterministic numbering.
1006   for (const auto &Edge : Updates) {
1007     BasicBlock *BB = Edge.getTo();
1008     if (PredMap.count(BB) && !MSSA->getMemoryAccess(BB))
1009       InsertedPhis.push_back(MSSA->createMemoryPhi(BB));
1010   }
1011 
1012   // Now we'll fill in the MemoryPhis with the right incoming values.
1013   for (auto &BBPredPair : PredMap) {
1014     auto *BB = BBPredPair.first;
1015     const auto &PrevBlockSet = BBPredPair.second.Prev;
1016     const auto &AddedBlockSet = BBPredPair.second.Added;
1017     assert(!PrevBlockSet.empty() &&
1018            "At least one previous predecessor must exist.");
1019 
1020     // TODO: if this becomes a bottleneck, we can save on GetLastDef calls by
1021     // keeping this map before the loop. We can reuse already populated entries
1022     // if an edge is added from the same predecessor to two different blocks,
1023     // and this does happen in rotate. Note that the map needs to be updated
1024     // when deleting non-necessary phis below, if the phi is in the map by
1025     // replacing the value with DefP1.
1026     SmallDenseMap<BasicBlock *, MemoryAccess *> LastDefAddedPred;
1027     for (auto *AddedPred : AddedBlockSet) {
1028       auto *DefPn = GetLastDef(AddedPred);
1029       assert(DefPn != nullptr && "Unable to find last definition.");
1030       LastDefAddedPred[AddedPred] = DefPn;
1031     }
1032 
1033     MemoryPhi *NewPhi = MSSA->getMemoryAccess(BB);
1034     // If Phi is not empty, add an incoming edge from each added pred. Must
1035     // still compute blocks with defs to replace for this block below.
1036     if (NewPhi->getNumOperands()) {
1037       for (auto *Pred : AddedBlockSet) {
1038         auto *LastDefForPred = LastDefAddedPred[Pred];
1039         for (int I = 0, E = EdgeCountMap[{Pred, BB}]; I < E; ++I)
1040           NewPhi->addIncoming(LastDefForPred, Pred);
1041       }
1042     } else {
1043       // Pick any existing predecessor and get its definition. All other
1044       // existing predecessors should have the same one, since no phi existed.
1045       auto *P1 = *PrevBlockSet.begin();
1046       MemoryAccess *DefP1 = GetLastDef(P1);
1047 
1048       // Check DefP1 against all Defs in LastDefPredPair. If all the same,
1049       // nothing to add.
1050       bool InsertPhi = false;
1051       for (auto LastDefPredPair : LastDefAddedPred)
1052         if (DefP1 != LastDefPredPair.second) {
1053           InsertPhi = true;
1054           break;
1055         }
1056       if (!InsertPhi) {
1057         // Since NewPhi may be used in other newly added Phis, replace all uses
1058         // of NewPhi with the definition coming from all predecessors (DefP1),
1059         // before deleting it.
1060         NewPhi->replaceAllUsesWith(DefP1);
1061         removeMemoryAccess(NewPhi);
1062         continue;
1063       }
1064 
1065       // Update Phi with new values for new predecessors and old value for all
1066       // other predecessors. Since AddedBlockSet and PrevBlockSet are ordered
1067       // sets, the order of entries in NewPhi is deterministic.
1068       for (auto *Pred : AddedBlockSet) {
1069         auto *LastDefForPred = LastDefAddedPred[Pred];
1070         for (int I = 0, E = EdgeCountMap[{Pred, BB}]; I < E; ++I)
1071           NewPhi->addIncoming(LastDefForPred, Pred);
1072       }
1073       for (auto *Pred : PrevBlockSet)
1074         for (int I = 0, E = EdgeCountMap[{Pred, BB}]; I < E; ++I)
1075           NewPhi->addIncoming(DefP1, Pred);
1076     }
1077 
1078     // Get all blocks that used to dominate BB and no longer do after adding
1079     // AddedBlockSet, where PrevBlockSet are the previously known predecessors.
1080     assert(DT.getNode(BB)->getIDom() && "BB does not have valid idom");
1081     BasicBlock *PrevIDom = FindNearestCommonDominator(PrevBlockSet);
1082     assert(PrevIDom && "Previous IDom should exists");
1083     BasicBlock *NewIDom = DT.getNode(BB)->getIDom()->getBlock();
1084     assert(NewIDom && "BB should have a new valid idom");
1085     assert(DT.dominates(NewIDom, PrevIDom) &&
1086            "New idom should dominate old idom");
1087     GetNoLongerDomBlocks(PrevIDom, NewIDom, BlocksWithDefsToReplace);
1088   }
1089 
1090   tryRemoveTrivialPhis(InsertedPhis);
1091   // Create the set of blocks that now have a definition. We'll use this to
1092   // compute IDF and add Phis there next.
1093   SmallVector<BasicBlock *, 8> BlocksToProcess;
1094   for (auto &VH : InsertedPhis)
1095     if (auto *MPhi = cast_or_null<MemoryPhi>(VH))
1096       BlocksToProcess.push_back(MPhi->getBlock());
1097 
1098   // Compute IDF and add Phis in all IDF blocks that do not have one.
1099   SmallVector<BasicBlock *, 32> IDFBlocks;
1100   if (!BlocksToProcess.empty()) {
1101     ForwardIDFCalculator IDFs(DT, GD);
1102     SmallPtrSet<BasicBlock *, 16> DefiningBlocks(BlocksToProcess.begin(),
1103                                                  BlocksToProcess.end());
1104     IDFs.setDefiningBlocks(DefiningBlocks);
1105     IDFs.calculate(IDFBlocks);
1106 
1107     SmallSetVector<MemoryPhi *, 4> PhisToFill;
1108     // First create all needed Phis.
1109     for (auto *BBIDF : IDFBlocks)
1110       if (!MSSA->getMemoryAccess(BBIDF)) {
1111         auto *IDFPhi = MSSA->createMemoryPhi(BBIDF);
1112         InsertedPhis.push_back(IDFPhi);
1113         PhisToFill.insert(IDFPhi);
1114       }
1115     // Then update or insert their correct incoming values.
1116     for (auto *BBIDF : IDFBlocks) {
1117       auto *IDFPhi = MSSA->getMemoryAccess(BBIDF);
1118       assert(IDFPhi && "Phi must exist");
1119       if (!PhisToFill.count(IDFPhi)) {
1120         // Update existing Phi.
1121         // FIXME: some updates may be redundant, try to optimize and skip some.
1122         for (unsigned I = 0, E = IDFPhi->getNumIncomingValues(); I < E; ++I)
1123           IDFPhi->setIncomingValue(I, GetLastDef(IDFPhi->getIncomingBlock(I)));
1124       } else {
1125         for (auto *Pi : GD->template getChildren</*InverseEdge=*/true>(BBIDF))
1126           IDFPhi->addIncoming(GetLastDef(Pi), Pi);
1127       }
1128     }
1129   }
1130 
1131   // Now for all defs in BlocksWithDefsToReplace, if there are uses they no
1132   // longer dominate, replace those with the closest dominating def.
1133   // This will also update optimized accesses, as they're also uses.
1134   for (auto *BlockWithDefsToReplace : BlocksWithDefsToReplace) {
1135     if (auto DefsList = MSSA->getWritableBlockDefs(BlockWithDefsToReplace)) {
1136       for (auto &DefToReplaceUses : *DefsList) {
1137         BasicBlock *DominatingBlock = DefToReplaceUses.getBlock();
1138         for (Use &U : llvm::make_early_inc_range(DefToReplaceUses.uses())) {
1139           MemoryAccess *Usr = cast<MemoryAccess>(U.getUser());
1140           if (MemoryPhi *UsrPhi = dyn_cast<MemoryPhi>(Usr)) {
1141             BasicBlock *DominatedBlock = UsrPhi->getIncomingBlock(U);
1142             if (!DT.dominates(DominatingBlock, DominatedBlock))
1143               U.set(GetLastDef(DominatedBlock));
1144           } else {
1145             BasicBlock *DominatedBlock = Usr->getBlock();
1146             if (!DT.dominates(DominatingBlock, DominatedBlock)) {
1147               if (auto *DomBlPhi = MSSA->getMemoryAccess(DominatedBlock))
1148                 U.set(DomBlPhi);
1149               else {
1150                 auto *IDom = DT.getNode(DominatedBlock)->getIDom();
1151                 assert(IDom && "Block must have a valid IDom.");
1152                 U.set(GetLastDef(IDom->getBlock()));
1153               }
1154               cast<MemoryUseOrDef>(Usr)->resetOptimized();
1155             }
1156           }
1157         }
1158       }
1159     }
1160   }
1161   tryRemoveTrivialPhis(InsertedPhis);
1162 }
1163 
1164 // Move What before Where in the MemorySSA IR.
1165 template <class WhereType>
1166 void MemorySSAUpdater::moveTo(MemoryUseOrDef *What, BasicBlock *BB,
1167                               WhereType Where) {
1168   // Mark MemoryPhi users of What not to be optimized.
1169   for (auto *U : What->users())
1170     if (MemoryPhi *PhiUser = dyn_cast<MemoryPhi>(U))
1171       NonOptPhis.insert(PhiUser);
1172 
1173   // Replace all our users with our defining access.
1174   What->replaceAllUsesWith(What->getDefiningAccess());
1175 
1176   // Let MemorySSA take care of moving it around in the lists.
1177   MSSA->moveTo(What, BB, Where);
1178 
1179   // Now reinsert it into the IR and do whatever fixups needed.
1180   if (auto *MD = dyn_cast<MemoryDef>(What))
1181     insertDef(MD, /*RenameUses=*/true);
1182   else
1183     insertUse(cast<MemoryUse>(What), /*RenameUses=*/true);
1184 
1185   // Clear dangling pointers. We added all MemoryPhi users, but not all
1186   // of them are removed by fixupDefs().
1187   NonOptPhis.clear();
1188 }
1189 
1190 // Move What before Where in the MemorySSA IR.
1191 void MemorySSAUpdater::moveBefore(MemoryUseOrDef *What, MemoryUseOrDef *Where) {
1192   moveTo(What, Where->getBlock(), Where->getIterator());
1193 }
1194 
1195 // Move What after Where in the MemorySSA IR.
1196 void MemorySSAUpdater::moveAfter(MemoryUseOrDef *What, MemoryUseOrDef *Where) {
1197   moveTo(What, Where->getBlock(), ++Where->getIterator());
1198 }
1199 
1200 void MemorySSAUpdater::moveToPlace(MemoryUseOrDef *What, BasicBlock *BB,
1201                                    MemorySSA::InsertionPlace Where) {
1202   if (Where != MemorySSA::InsertionPlace::BeforeTerminator)
1203     return moveTo(What, BB, Where);
1204 
1205   if (auto *Where = MSSA->getMemoryAccess(BB->getTerminator()))
1206     return moveBefore(What, Where);
1207   else
1208     return moveTo(What, BB, MemorySSA::InsertionPlace::End);
1209 }
1210 
1211 // All accesses in To used to be in From. Move to end and update access lists.
1212 void MemorySSAUpdater::moveAllAccesses(BasicBlock *From, BasicBlock *To,
1213                                        Instruction *Start) {
1214 
1215   MemorySSA::AccessList *Accs = MSSA->getWritableBlockAccesses(From);
1216   if (!Accs)
1217     return;
1218 
1219   assert(Start->getParent() == To && "Incorrect Start instruction");
1220   MemoryAccess *FirstInNew = nullptr;
1221   for (Instruction &I : make_range(Start->getIterator(), To->end()))
1222     if ((FirstInNew = MSSA->getMemoryAccess(&I)))
1223       break;
1224   if (FirstInNew) {
1225     auto *MUD = cast<MemoryUseOrDef>(FirstInNew);
1226     do {
1227       auto NextIt = ++MUD->getIterator();
1228       MemoryUseOrDef *NextMUD = (!Accs || NextIt == Accs->end())
1229                                     ? nullptr
1230                                     : cast<MemoryUseOrDef>(&*NextIt);
1231       MSSA->moveTo(MUD, To, MemorySSA::End);
1232       // Moving MUD from Accs in the moveTo above, may delete Accs, so we need
1233       // to retrieve it again.
1234       Accs = MSSA->getWritableBlockAccesses(From);
1235       MUD = NextMUD;
1236     } while (MUD);
1237   }
1238 
1239   // If all accesses were moved and only a trivial Phi remains, we try to remove
1240   // that Phi. This is needed when From is going to be deleted.
1241   auto *Defs = MSSA->getWritableBlockDefs(From);
1242   if (Defs && !Defs->empty())
1243     if (auto *Phi = dyn_cast<MemoryPhi>(&*Defs->begin()))
1244       tryRemoveTrivialPhi(Phi);
1245 }
1246 
1247 void MemorySSAUpdater::moveAllAfterSpliceBlocks(BasicBlock *From,
1248                                                 BasicBlock *To,
1249                                                 Instruction *Start) {
1250   assert(MSSA->getBlockAccesses(To) == nullptr &&
1251          "To block is expected to be free of MemoryAccesses.");
1252   moveAllAccesses(From, To, Start);
1253   for (BasicBlock *Succ : successors(To))
1254     if (MemoryPhi *MPhi = MSSA->getMemoryAccess(Succ))
1255       MPhi->setIncomingBlock(MPhi->getBasicBlockIndex(From), To);
1256 }
1257 
1258 void MemorySSAUpdater::moveAllAfterMergeBlocks(BasicBlock *From, BasicBlock *To,
1259                                                Instruction *Start) {
1260   assert(From->getUniquePredecessor() == To &&
1261          "From block is expected to have a single predecessor (To).");
1262   moveAllAccesses(From, To, Start);
1263   for (BasicBlock *Succ : successors(From))
1264     if (MemoryPhi *MPhi = MSSA->getMemoryAccess(Succ))
1265       MPhi->setIncomingBlock(MPhi->getBasicBlockIndex(From), To);
1266 }
1267 
1268 void MemorySSAUpdater::wireOldPredecessorsToNewImmediatePredecessor(
1269     BasicBlock *Old, BasicBlock *New, ArrayRef<BasicBlock *> Preds,
1270     bool IdenticalEdgesWereMerged) {
1271   assert(!MSSA->getWritableBlockAccesses(New) &&
1272          "Access list should be null for a new block.");
1273   MemoryPhi *Phi = MSSA->getMemoryAccess(Old);
1274   if (!Phi)
1275     return;
1276   if (Old->hasNPredecessors(1)) {
1277     assert(pred_size(New) == Preds.size() &&
1278            "Should have moved all predecessors.");
1279     MSSA->moveTo(Phi, New, MemorySSA::Beginning);
1280   } else {
1281     assert(!Preds.empty() && "Must be moving at least one predecessor to the "
1282                              "new immediate predecessor.");
1283     MemoryPhi *NewPhi = MSSA->createMemoryPhi(New);
1284     SmallPtrSet<BasicBlock *, 16> PredsSet(Preds.begin(), Preds.end());
1285     // Currently only support the case of removing a single incoming edge when
1286     // identical edges were not merged.
1287     if (!IdenticalEdgesWereMerged)
1288       assert(PredsSet.size() == Preds.size() &&
1289              "If identical edges were not merged, we cannot have duplicate "
1290              "blocks in the predecessors");
1291     Phi->unorderedDeleteIncomingIf([&](MemoryAccess *MA, BasicBlock *B) {
1292       if (PredsSet.count(B)) {
1293         NewPhi->addIncoming(MA, B);
1294         if (!IdenticalEdgesWereMerged)
1295           PredsSet.erase(B);
1296         return true;
1297       }
1298       return false;
1299     });
1300     Phi->addIncoming(NewPhi, New);
1301     tryRemoveTrivialPhi(NewPhi);
1302   }
1303 }
1304 
1305 void MemorySSAUpdater::removeMemoryAccess(MemoryAccess *MA, bool OptimizePhis) {
1306   assert(!MSSA->isLiveOnEntryDef(MA) &&
1307          "Trying to remove the live on entry def");
1308   // We can only delete phi nodes if they have no uses, or we can replace all
1309   // uses with a single definition.
1310   MemoryAccess *NewDefTarget = nullptr;
1311   if (MemoryPhi *MP = dyn_cast<MemoryPhi>(MA)) {
1312     // Note that it is sufficient to know that all edges of the phi node have
1313     // the same argument.  If they do, by the definition of dominance frontiers
1314     // (which we used to place this phi), that argument must dominate this phi,
1315     // and thus, must dominate the phi's uses, and so we will not hit the assert
1316     // below.
1317     NewDefTarget = onlySingleValue(MP);
1318     assert((NewDefTarget || MP->use_empty()) &&
1319            "We can't delete this memory phi");
1320   } else {
1321     NewDefTarget = cast<MemoryUseOrDef>(MA)->getDefiningAccess();
1322   }
1323 
1324   SmallSetVector<MemoryPhi *, 4> PhisToCheck;
1325 
1326   // Re-point the uses at our defining access
1327   if (!isa<MemoryUse>(MA) && !MA->use_empty()) {
1328     // Reset optimized on users of this store, and reset the uses.
1329     // A few notes:
1330     // 1. This is a slightly modified version of RAUW to avoid walking the
1331     // uses twice here.
1332     // 2. If we wanted to be complete, we would have to reset the optimized
1333     // flags on users of phi nodes if doing the below makes a phi node have all
1334     // the same arguments. Instead, we prefer users to removeMemoryAccess those
1335     // phi nodes, because doing it here would be N^3.
1336     if (MA->hasValueHandle())
1337       ValueHandleBase::ValueIsRAUWd(MA, NewDefTarget);
1338     // Note: We assume MemorySSA is not used in metadata since it's not really
1339     // part of the IR.
1340 
1341     assert(NewDefTarget != MA && "Going into an infinite loop");
1342     while (!MA->use_empty()) {
1343       Use &U = *MA->use_begin();
1344       if (auto *MUD = dyn_cast<MemoryUseOrDef>(U.getUser()))
1345         MUD->resetOptimized();
1346       if (OptimizePhis)
1347         if (MemoryPhi *MP = dyn_cast<MemoryPhi>(U.getUser()))
1348           PhisToCheck.insert(MP);
1349       U.set(NewDefTarget);
1350     }
1351   }
1352 
1353   // The call below to erase will destroy MA, so we can't change the order we
1354   // are doing things here
1355   MSSA->removeFromLookups(MA);
1356   MSSA->removeFromLists(MA);
1357 
1358   // Optionally optimize Phi uses. This will recursively remove trivial phis.
1359   if (!PhisToCheck.empty()) {
1360     SmallVector<WeakVH, 16> PhisToOptimize{PhisToCheck.begin(),
1361                                            PhisToCheck.end()};
1362     PhisToCheck.clear();
1363 
1364     unsigned PhisSize = PhisToOptimize.size();
1365     while (PhisSize-- > 0)
1366       if (MemoryPhi *MP =
1367               cast_or_null<MemoryPhi>(PhisToOptimize.pop_back_val()))
1368         tryRemoveTrivialPhi(MP);
1369   }
1370 }
1371 
1372 void MemorySSAUpdater::removeBlocks(
1373     const SmallSetVector<BasicBlock *, 8> &DeadBlocks) {
1374   // First delete all uses of BB in MemoryPhis.
1375   for (BasicBlock *BB : DeadBlocks) {
1376     Instruction *TI = BB->getTerminator();
1377     assert(TI && "Basic block expected to have a terminator instruction");
1378     for (BasicBlock *Succ : successors(TI))
1379       if (!DeadBlocks.count(Succ))
1380         if (MemoryPhi *MP = MSSA->getMemoryAccess(Succ)) {
1381           MP->unorderedDeleteIncomingBlock(BB);
1382           tryRemoveTrivialPhi(MP);
1383         }
1384     // Drop all references of all accesses in BB
1385     if (MemorySSA::AccessList *Acc = MSSA->getWritableBlockAccesses(BB))
1386       for (MemoryAccess &MA : *Acc)
1387         MA.dropAllReferences();
1388   }
1389 
1390   // Next, delete all memory accesses in each block
1391   for (BasicBlock *BB : DeadBlocks) {
1392     MemorySSA::AccessList *Acc = MSSA->getWritableBlockAccesses(BB);
1393     if (!Acc)
1394       continue;
1395     for (MemoryAccess &MA : llvm::make_early_inc_range(*Acc)) {
1396       MSSA->removeFromLookups(&MA);
1397       MSSA->removeFromLists(&MA);
1398     }
1399   }
1400 }
1401 
1402 void MemorySSAUpdater::tryRemoveTrivialPhis(ArrayRef<WeakVH> UpdatedPHIs) {
1403   for (const auto &VH : UpdatedPHIs)
1404     if (auto *MPhi = cast_or_null<MemoryPhi>(VH))
1405       tryRemoveTrivialPhi(MPhi);
1406 }
1407 
1408 void MemorySSAUpdater::changeToUnreachable(const Instruction *I) {
1409   const BasicBlock *BB = I->getParent();
1410   // Remove memory accesses in BB for I and all following instructions.
1411   auto BBI = I->getIterator(), BBE = BB->end();
1412   // FIXME: If this becomes too expensive, iterate until the first instruction
1413   // with a memory access, then iterate over MemoryAccesses.
1414   while (BBI != BBE)
1415     removeMemoryAccess(&*(BBI++));
1416   // Update phis in BB's successors to remove BB.
1417   SmallVector<WeakVH, 16> UpdatedPHIs;
1418   for (const BasicBlock *Successor : successors(BB)) {
1419     removeDuplicatePhiEdgesBetween(BB, Successor);
1420     if (MemoryPhi *MPhi = MSSA->getMemoryAccess(Successor)) {
1421       MPhi->unorderedDeleteIncomingBlock(BB);
1422       UpdatedPHIs.push_back(MPhi);
1423     }
1424   }
1425   // Optimize trivial phis.
1426   tryRemoveTrivialPhis(UpdatedPHIs);
1427 }
1428 
1429 MemoryAccess *MemorySSAUpdater::createMemoryAccessInBB(
1430     Instruction *I, MemoryAccess *Definition, const BasicBlock *BB,
1431     MemorySSA::InsertionPlace Point) {
1432   MemoryUseOrDef *NewAccess = MSSA->createDefinedAccess(I, Definition);
1433   MSSA->insertIntoListsForBlock(NewAccess, BB, Point);
1434   return NewAccess;
1435 }
1436 
1437 MemoryUseOrDef *MemorySSAUpdater::createMemoryAccessBefore(
1438     Instruction *I, MemoryAccess *Definition, MemoryUseOrDef *InsertPt) {
1439   assert(I->getParent() == InsertPt->getBlock() &&
1440          "New and old access must be in the same block");
1441   MemoryUseOrDef *NewAccess = MSSA->createDefinedAccess(I, Definition);
1442   MSSA->insertIntoListsBefore(NewAccess, InsertPt->getBlock(),
1443                               InsertPt->getIterator());
1444   return NewAccess;
1445 }
1446 
1447 MemoryUseOrDef *MemorySSAUpdater::createMemoryAccessAfter(
1448     Instruction *I, MemoryAccess *Definition, MemoryAccess *InsertPt) {
1449   assert(I->getParent() == InsertPt->getBlock() &&
1450          "New and old access must be in the same block");
1451   MemoryUseOrDef *NewAccess = MSSA->createDefinedAccess(I, Definition);
1452   MSSA->insertIntoListsBefore(NewAccess, InsertPt->getBlock(),
1453                               ++InsertPt->getIterator());
1454   return NewAccess;
1455 }
1456