1 //===- MemorySSA.h - Build Memory SSA ---------------------------*- C++ -*-===//
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 /// \file
10 /// This file exposes an interface to building/using memory SSA to
11 /// walk memory instructions using a use/def graph.
12 ///
13 /// Memory SSA class builds an SSA form that links together memory access
14 /// instructions such as loads, stores, atomics, and calls. Additionally, it
15 /// does a trivial form of "heap versioning" Every time the memory state changes
16 /// in the program, we generate a new heap version. It generates
17 /// MemoryDef/Uses/Phis that are overlayed on top of the existing instructions.
18 ///
19 /// As a trivial example,
20 /// define i32 @main() #0 {
21 /// entry:
22 ///   %call = call noalias i8* @_Znwm(i64 4) #2
23 ///   %0 = bitcast i8* %call to i32*
24 ///   %call1 = call noalias i8* @_Znwm(i64 4) #2
25 ///   %1 = bitcast i8* %call1 to i32*
26 ///   store i32 5, i32* %0, align 4
27 ///   store i32 7, i32* %1, align 4
28 ///   %2 = load i32* %0, align 4
29 ///   %3 = load i32* %1, align 4
30 ///   %add = add nsw i32 %2, %3
31 ///   ret i32 %add
32 /// }
33 ///
34 /// Will become
35 /// define i32 @main() #0 {
36 /// entry:
37 ///   ; 1 = MemoryDef(0)
38 ///   %call = call noalias i8* @_Znwm(i64 4) #3
39 ///   %2 = bitcast i8* %call to i32*
40 ///   ; 2 = MemoryDef(1)
41 ///   %call1 = call noalias i8* @_Znwm(i64 4) #3
42 ///   %4 = bitcast i8* %call1 to i32*
43 ///   ; 3 = MemoryDef(2)
44 ///   store i32 5, i32* %2, align 4
45 ///   ; 4 = MemoryDef(3)
46 ///   store i32 7, i32* %4, align 4
47 ///   ; MemoryUse(3)
48 ///   %7 = load i32* %2, align 4
49 ///   ; MemoryUse(4)
50 ///   %8 = load i32* %4, align 4
51 ///   %add = add nsw i32 %7, %8
52 ///   ret i32 %add
53 /// }
54 ///
55 /// Given this form, all the stores that could ever effect the load at %8 can be
56 /// gotten by using the MemoryUse associated with it, and walking from use to
57 /// def until you hit the top of the function.
58 ///
59 /// Each def also has a list of users associated with it, so you can walk from
60 /// both def to users, and users to defs. Note that we disambiguate MemoryUses,
61 /// but not the RHS of MemoryDefs. You can see this above at %7, which would
62 /// otherwise be a MemoryUse(4). Being disambiguated means that for a given
63 /// store, all the MemoryUses on its use lists are may-aliases of that store
64 /// (but the MemoryDefs on its use list may not be).
65 ///
66 /// MemoryDefs are not disambiguated because it would require multiple reaching
67 /// definitions, which would require multiple phis, and multiple memoryaccesses
68 /// per instruction.
69 //
70 //===----------------------------------------------------------------------===//
71 
72 #ifndef LLVM_ANALYSIS_MEMORYSSA_H
73 #define LLVM_ANALYSIS_MEMORYSSA_H
74 
75 #include "llvm/ADT/DenseMap.h"
76 #include "llvm/ADT/GraphTraits.h"
77 #include "llvm/ADT/SmallPtrSet.h"
78 #include "llvm/ADT/SmallVector.h"
79 #include "llvm/ADT/ilist.h"
80 #include "llvm/ADT/ilist_node.h"
81 #include "llvm/ADT/iterator.h"
82 #include "llvm/ADT/iterator_range.h"
83 #include "llvm/ADT/simple_ilist.h"
84 #include "llvm/Analysis/AliasAnalysis.h"
85 #include "llvm/Analysis/MemoryLocation.h"
86 #include "llvm/Analysis/PHITransAddr.h"
87 #include "llvm/IR/BasicBlock.h"
88 #include "llvm/IR/DerivedUser.h"
89 #include "llvm/IR/Dominators.h"
90 #include "llvm/IR/Module.h"
91 #include "llvm/IR/Type.h"
92 #include "llvm/IR/Use.h"
93 #include "llvm/IR/User.h"
94 #include "llvm/IR/Value.h"
95 #include "llvm/IR/ValueHandle.h"
96 #include "llvm/Pass.h"
97 #include "llvm/Support/Casting.h"
98 #include "llvm/Support/CommandLine.h"
99 #include <algorithm>
100 #include <cassert>
101 #include <cstddef>
102 #include <iterator>
103 #include <memory>
104 #include <utility>
105 
106 namespace llvm {
107 
108 /// Enables memory ssa as a dependency for loop passes.
109 extern cl::opt<bool> EnableMSSALoopDependency;
110 
111 class Function;
112 class Instruction;
113 class MemoryAccess;
114 class MemorySSAWalker;
115 class LLVMContext;
116 class raw_ostream;
117 
118 namespace MSSAHelpers {
119 
120 struct AllAccessTag {};
121 struct DefsOnlyTag {};
122 
123 } // end namespace MSSAHelpers
124 
125 enum : unsigned {
126   // Used to signify what the default invalid ID is for MemoryAccess's
127   // getID()
128   INVALID_MEMORYACCESS_ID = -1U
129 };
130 
131 template <class T> class memoryaccess_def_iterator_base;
132 using memoryaccess_def_iterator = memoryaccess_def_iterator_base<MemoryAccess>;
133 using const_memoryaccess_def_iterator =
134     memoryaccess_def_iterator_base<const MemoryAccess>;
135 
136 // The base for all memory accesses. All memory accesses in a block are
137 // linked together using an intrusive list.
138 class MemoryAccess
139     : public DerivedUser,
140       public ilist_node<MemoryAccess, ilist_tag<MSSAHelpers::AllAccessTag>>,
141       public ilist_node<MemoryAccess, ilist_tag<MSSAHelpers::DefsOnlyTag>> {
142 public:
143   using AllAccessType =
144       ilist_node<MemoryAccess, ilist_tag<MSSAHelpers::AllAccessTag>>;
145   using DefsOnlyType =
146       ilist_node<MemoryAccess, ilist_tag<MSSAHelpers::DefsOnlyTag>>;
147 
148   MemoryAccess(const MemoryAccess &) = delete;
149   MemoryAccess &operator=(const MemoryAccess &) = delete;
150 
151   void *operator new(size_t) = delete;
152 
153   // Methods for support type inquiry through isa, cast, and
154   // dyn_cast
155   static bool classof(const Value *V) {
156     unsigned ID = V->getValueID();
157     return ID == MemoryUseVal || ID == MemoryPhiVal || ID == MemoryDefVal;
158   }
159 
160   BasicBlock *getBlock() const { return Block; }
161 
162   void print(raw_ostream &OS) const;
163   void dump() const;
164 
165   /// The user iterators for a memory access
166   using iterator = user_iterator;
167   using const_iterator = const_user_iterator;
168 
169   /// This iterator walks over all of the defs in a given
170   /// MemoryAccess. For MemoryPhi nodes, this walks arguments. For
171   /// MemoryUse/MemoryDef, this walks the defining access.
172   memoryaccess_def_iterator defs_begin();
173   const_memoryaccess_def_iterator defs_begin() const;
174   memoryaccess_def_iterator defs_end();
175   const_memoryaccess_def_iterator defs_end() const;
176 
177   /// Get the iterators for the all access list and the defs only list
178   /// We default to the all access list.
179   AllAccessType::self_iterator getIterator() {
180     return this->AllAccessType::getIterator();
181   }
182   AllAccessType::const_self_iterator getIterator() const {
183     return this->AllAccessType::getIterator();
184   }
185   AllAccessType::reverse_self_iterator getReverseIterator() {
186     return this->AllAccessType::getReverseIterator();
187   }
188   AllAccessType::const_reverse_self_iterator getReverseIterator() const {
189     return this->AllAccessType::getReverseIterator();
190   }
191   DefsOnlyType::self_iterator getDefsIterator() {
192     return this->DefsOnlyType::getIterator();
193   }
194   DefsOnlyType::const_self_iterator getDefsIterator() const {
195     return this->DefsOnlyType::getIterator();
196   }
197   DefsOnlyType::reverse_self_iterator getReverseDefsIterator() {
198     return this->DefsOnlyType::getReverseIterator();
199   }
200   DefsOnlyType::const_reverse_self_iterator getReverseDefsIterator() const {
201     return this->DefsOnlyType::getReverseIterator();
202   }
203 
204 protected:
205   friend class MemoryDef;
206   friend class MemoryPhi;
207   friend class MemorySSA;
208   friend class MemoryUse;
209   friend class MemoryUseOrDef;
210 
211   /// Used by MemorySSA to change the block of a MemoryAccess when it is
212   /// moved.
213   void setBlock(BasicBlock *BB) { Block = BB; }
214 
215   /// Used for debugging and tracking things about MemoryAccesses.
216   /// Guaranteed unique among MemoryAccesses, no guarantees otherwise.
217   inline unsigned getID() const;
218 
219   MemoryAccess(LLVMContext &C, unsigned Vty, DeleteValueTy DeleteValue,
220                BasicBlock *BB, unsigned NumOperands)
221       : DerivedUser(Type::getVoidTy(C), Vty, nullptr, NumOperands, DeleteValue),
222         Block(BB) {}
223 
224   // Use deleteValue() to delete a generic MemoryAccess.
225   ~MemoryAccess() = default;
226 
227 private:
228   BasicBlock *Block;
229 };
230 
231 template <>
232 struct ilist_alloc_traits<MemoryAccess> {
233   static void deleteNode(MemoryAccess *MA) { MA->deleteValue(); }
234 };
235 
236 inline raw_ostream &operator<<(raw_ostream &OS, const MemoryAccess &MA) {
237   MA.print(OS);
238   return OS;
239 }
240 
241 /// Class that has the common methods + fields of memory uses/defs. It's
242 /// a little awkward to have, but there are many cases where we want either a
243 /// use or def, and there are many cases where uses are needed (defs aren't
244 /// acceptable), and vice-versa.
245 ///
246 /// This class should never be instantiated directly; make a MemoryUse or
247 /// MemoryDef instead.
248 class MemoryUseOrDef : public MemoryAccess {
249 public:
250   void *operator new(size_t) = delete;
251 
252   DECLARE_TRANSPARENT_OPERAND_ACCESSORS(MemoryAccess);
253 
254   /// Get the instruction that this MemoryUse represents.
255   Instruction *getMemoryInst() const { return MemoryInstruction; }
256 
257   /// Get the access that produces the memory state used by this Use.
258   MemoryAccess *getDefiningAccess() const { return getOperand(0); }
259 
260   static bool classof(const Value *MA) {
261     return MA->getValueID() == MemoryUseVal || MA->getValueID() == MemoryDefVal;
262   }
263 
264   // Sadly, these have to be public because they are needed in some of the
265   // iterators.
266   inline bool isOptimized() const;
267   inline MemoryAccess *getOptimized() const;
268   inline void setOptimized(MemoryAccess *);
269 
270   // Retrieve AliasResult type of the optimized access. Ideally this would be
271   // returned by the caching walker and may go away in the future.
272   Optional<AliasResult> getOptimizedAccessType() const {
273     return OptimizedAccessAlias;
274   }
275 
276   /// Reset the ID of what this MemoryUse was optimized to, causing it to
277   /// be rewalked by the walker if necessary.
278   /// This really should only be called by tests.
279   inline void resetOptimized();
280 
281 protected:
282   friend class MemorySSA;
283   friend class MemorySSAUpdater;
284 
285   MemoryUseOrDef(LLVMContext &C, MemoryAccess *DMA, unsigned Vty,
286                  DeleteValueTy DeleteValue, Instruction *MI, BasicBlock *BB,
287                  unsigned NumOperands)
288       : MemoryAccess(C, Vty, DeleteValue, BB, NumOperands),
289         MemoryInstruction(MI), OptimizedAccessAlias(MayAlias) {
290     setDefiningAccess(DMA);
291   }
292 
293   // Use deleteValue() to delete a generic MemoryUseOrDef.
294   ~MemoryUseOrDef() = default;
295 
296   void setOptimizedAccessType(Optional<AliasResult> AR) {
297     OptimizedAccessAlias = AR;
298   }
299 
300   void setDefiningAccess(MemoryAccess *DMA, bool Optimized = false,
301                          Optional<AliasResult> AR = MayAlias) {
302     if (!Optimized) {
303       setOperand(0, DMA);
304       return;
305     }
306     setOptimized(DMA);
307     setOptimizedAccessType(AR);
308   }
309 
310 private:
311   Instruction *MemoryInstruction;
312   Optional<AliasResult> OptimizedAccessAlias;
313 };
314 
315 /// Represents read-only accesses to memory
316 ///
317 /// In particular, the set of Instructions that will be represented by
318 /// MemoryUse's is exactly the set of Instructions for which
319 /// AliasAnalysis::getModRefInfo returns "Ref".
320 class MemoryUse final : public MemoryUseOrDef {
321 public:
322   DECLARE_TRANSPARENT_OPERAND_ACCESSORS(MemoryAccess);
323 
324   MemoryUse(LLVMContext &C, MemoryAccess *DMA, Instruction *MI, BasicBlock *BB)
325       : MemoryUseOrDef(C, DMA, MemoryUseVal, deleteMe, MI, BB,
326                        /*NumOperands=*/1) {}
327 
328   // allocate space for exactly one operand
329   void *operator new(size_t s) { return User::operator new(s, 1); }
330 
331   static bool classof(const Value *MA) {
332     return MA->getValueID() == MemoryUseVal;
333   }
334 
335   void print(raw_ostream &OS) const;
336 
337   void setOptimized(MemoryAccess *DMA) {
338     OptimizedID = DMA->getID();
339     setOperand(0, DMA);
340   }
341 
342   bool isOptimized() const {
343     return getDefiningAccess() && OptimizedID == getDefiningAccess()->getID();
344   }
345 
346   MemoryAccess *getOptimized() const {
347     return getDefiningAccess();
348   }
349 
350   void resetOptimized() {
351     OptimizedID = INVALID_MEMORYACCESS_ID;
352   }
353 
354 protected:
355   friend class MemorySSA;
356 
357 private:
358   static void deleteMe(DerivedUser *Self);
359 
360   unsigned OptimizedID = INVALID_MEMORYACCESS_ID;
361 };
362 
363 template <>
364 struct OperandTraits<MemoryUse> : public FixedNumOperandTraits<MemoryUse, 1> {};
365 DEFINE_TRANSPARENT_OPERAND_ACCESSORS(MemoryUse, MemoryAccess)
366 
367 /// Represents a read-write access to memory, whether it is a must-alias,
368 /// or a may-alias.
369 ///
370 /// In particular, the set of Instructions that will be represented by
371 /// MemoryDef's is exactly the set of Instructions for which
372 /// AliasAnalysis::getModRefInfo returns "Mod" or "ModRef".
373 /// Note that, in order to provide def-def chains, all defs also have a use
374 /// associated with them. This use points to the nearest reaching
375 /// MemoryDef/MemoryPhi.
376 class MemoryDef final : public MemoryUseOrDef {
377 public:
378   friend class MemorySSA;
379 
380   DECLARE_TRANSPARENT_OPERAND_ACCESSORS(MemoryAccess);
381 
382   MemoryDef(LLVMContext &C, MemoryAccess *DMA, Instruction *MI, BasicBlock *BB,
383             unsigned Ver)
384       : MemoryUseOrDef(C, DMA, MemoryDefVal, deleteMe, MI, BB,
385                        /*NumOperands=*/2),
386         ID(Ver) {}
387 
388   // allocate space for exactly two operands
389   void *operator new(size_t s) { return User::operator new(s, 2); }
390 
391   static bool classof(const Value *MA) {
392     return MA->getValueID() == MemoryDefVal;
393   }
394 
395   void setOptimized(MemoryAccess *MA) {
396     setOperand(1, MA);
397     OptimizedID = MA->getID();
398   }
399 
400   MemoryAccess *getOptimized() const {
401     return cast_or_null<MemoryAccess>(getOperand(1));
402   }
403 
404   bool isOptimized() const {
405     return getOptimized() && OptimizedID == getOptimized()->getID();
406   }
407 
408   void resetOptimized() {
409     OptimizedID = INVALID_MEMORYACCESS_ID;
410     setOperand(1, nullptr);
411   }
412 
413   void print(raw_ostream &OS) const;
414 
415   unsigned getID() const { return ID; }
416 
417 private:
418   static void deleteMe(DerivedUser *Self);
419 
420   const unsigned ID;
421   unsigned OptimizedID = INVALID_MEMORYACCESS_ID;
422 };
423 
424 template <>
425 struct OperandTraits<MemoryDef> : public FixedNumOperandTraits<MemoryDef, 2> {};
426 DEFINE_TRANSPARENT_OPERAND_ACCESSORS(MemoryDef, MemoryAccess)
427 
428 template <>
429 struct OperandTraits<MemoryUseOrDef> {
430   static Use *op_begin(MemoryUseOrDef *MUD) {
431     if (auto *MU = dyn_cast<MemoryUse>(MUD))
432       return OperandTraits<MemoryUse>::op_begin(MU);
433     return OperandTraits<MemoryDef>::op_begin(cast<MemoryDef>(MUD));
434   }
435 
436   static Use *op_end(MemoryUseOrDef *MUD) {
437     if (auto *MU = dyn_cast<MemoryUse>(MUD))
438       return OperandTraits<MemoryUse>::op_end(MU);
439     return OperandTraits<MemoryDef>::op_end(cast<MemoryDef>(MUD));
440   }
441 
442   static unsigned operands(const MemoryUseOrDef *MUD) {
443     if (const auto *MU = dyn_cast<MemoryUse>(MUD))
444       return OperandTraits<MemoryUse>::operands(MU);
445     return OperandTraits<MemoryDef>::operands(cast<MemoryDef>(MUD));
446   }
447 };
448 DEFINE_TRANSPARENT_OPERAND_ACCESSORS(MemoryUseOrDef, MemoryAccess)
449 
450 /// Represents phi nodes for memory accesses.
451 ///
452 /// These have the same semantic as regular phi nodes, with the exception that
453 /// only one phi will ever exist in a given basic block.
454 /// Guaranteeing one phi per block means guaranteeing there is only ever one
455 /// valid reaching MemoryDef/MemoryPHI along each path to the phi node.
456 /// This is ensured by not allowing disambiguation of the RHS of a MemoryDef or
457 /// a MemoryPhi's operands.
458 /// That is, given
459 /// if (a) {
460 ///   store %a
461 ///   store %b
462 /// }
463 /// it *must* be transformed into
464 /// if (a) {
465 ///    1 = MemoryDef(liveOnEntry)
466 ///    store %a
467 ///    2 = MemoryDef(1)
468 ///    store %b
469 /// }
470 /// and *not*
471 /// if (a) {
472 ///    1 = MemoryDef(liveOnEntry)
473 ///    store %a
474 ///    2 = MemoryDef(liveOnEntry)
475 ///    store %b
476 /// }
477 /// even if the two stores do not conflict. Otherwise, both 1 and 2 reach the
478 /// end of the branch, and if there are not two phi nodes, one will be
479 /// disconnected completely from the SSA graph below that point.
480 /// Because MemoryUse's do not generate new definitions, they do not have this
481 /// issue.
482 class MemoryPhi final : public MemoryAccess {
483   // allocate space for exactly zero operands
484   void *operator new(size_t s) { return User::operator new(s); }
485 
486 public:
487   /// Provide fast operand accessors
488   DECLARE_TRANSPARENT_OPERAND_ACCESSORS(MemoryAccess);
489 
490   MemoryPhi(LLVMContext &C, BasicBlock *BB, unsigned Ver, unsigned NumPreds = 0)
491       : MemoryAccess(C, MemoryPhiVal, deleteMe, BB, 0), ID(Ver),
492         ReservedSpace(NumPreds) {
493     allocHungoffUses(ReservedSpace);
494   }
495 
496   // Block iterator interface. This provides access to the list of incoming
497   // basic blocks, which parallels the list of incoming values.
498   using block_iterator = BasicBlock **;
499   using const_block_iterator = BasicBlock *const *;
500 
501   block_iterator block_begin() {
502     auto *Ref = reinterpret_cast<Use::UserRef *>(op_begin() + ReservedSpace);
503     return reinterpret_cast<block_iterator>(Ref + 1);
504   }
505 
506   const_block_iterator block_begin() const {
507     const auto *Ref =
508         reinterpret_cast<const Use::UserRef *>(op_begin() + ReservedSpace);
509     return reinterpret_cast<const_block_iterator>(Ref + 1);
510   }
511 
512   block_iterator block_end() { return block_begin() + getNumOperands(); }
513 
514   const_block_iterator block_end() const {
515     return block_begin() + getNumOperands();
516   }
517 
518   iterator_range<block_iterator> blocks() {
519     return make_range(block_begin(), block_end());
520   }
521 
522   iterator_range<const_block_iterator> blocks() const {
523     return make_range(block_begin(), block_end());
524   }
525 
526   op_range incoming_values() { return operands(); }
527 
528   const_op_range incoming_values() const { return operands(); }
529 
530   /// Return the number of incoming edges
531   unsigned getNumIncomingValues() const { return getNumOperands(); }
532 
533   /// Return incoming value number x
534   MemoryAccess *getIncomingValue(unsigned I) const { return getOperand(I); }
535   void setIncomingValue(unsigned I, MemoryAccess *V) {
536     assert(V && "PHI node got a null value!");
537     setOperand(I, V);
538   }
539 
540   static unsigned getOperandNumForIncomingValue(unsigned I) { return I; }
541   static unsigned getIncomingValueNumForOperand(unsigned I) { return I; }
542 
543   /// Return incoming basic block number @p i.
544   BasicBlock *getIncomingBlock(unsigned I) const { return block_begin()[I]; }
545 
546   /// Return incoming basic block corresponding
547   /// to an operand of the PHI.
548   BasicBlock *getIncomingBlock(const Use &U) const {
549     assert(this == U.getUser() && "Iterator doesn't point to PHI's Uses?");
550     return getIncomingBlock(unsigned(&U - op_begin()));
551   }
552 
553   /// Return incoming basic block corresponding
554   /// to value use iterator.
555   BasicBlock *getIncomingBlock(MemoryAccess::const_user_iterator I) const {
556     return getIncomingBlock(I.getUse());
557   }
558 
559   void setIncomingBlock(unsigned I, BasicBlock *BB) {
560     assert(BB && "PHI node got a null basic block!");
561     block_begin()[I] = BB;
562   }
563 
564   /// Add an incoming value to the end of the PHI list
565   void addIncoming(MemoryAccess *V, BasicBlock *BB) {
566     if (getNumOperands() == ReservedSpace)
567       growOperands(); // Get more space!
568     // Initialize some new operands.
569     setNumHungOffUseOperands(getNumOperands() + 1);
570     setIncomingValue(getNumOperands() - 1, V);
571     setIncomingBlock(getNumOperands() - 1, BB);
572   }
573 
574   /// Return the first index of the specified basic
575   /// block in the value list for this PHI.  Returns -1 if no instance.
576   int getBasicBlockIndex(const BasicBlock *BB) const {
577     for (unsigned I = 0, E = getNumOperands(); I != E; ++I)
578       if (block_begin()[I] == BB)
579         return I;
580     return -1;
581   }
582 
583   MemoryAccess *getIncomingValueForBlock(const BasicBlock *BB) const {
584     int Idx = getBasicBlockIndex(BB);
585     assert(Idx >= 0 && "Invalid basic block argument!");
586     return getIncomingValue(Idx);
587   }
588 
589   // After deleting incoming position I, the order of incoming may be changed.
590   void unorderedDeleteIncoming(unsigned I) {
591     unsigned E = getNumOperands();
592     assert(I < E && "Cannot remove out of bounds Phi entry.");
593     // MemoryPhi must have at least two incoming values, otherwise the MemoryPhi
594     // itself should be deleted.
595     assert(E >= 2 && "Cannot only remove incoming values in MemoryPhis with "
596                      "at least 2 values.");
597     setIncomingValue(I, getIncomingValue(E - 1));
598     setIncomingBlock(I, block_begin()[E - 1]);
599     setOperand(E - 1, nullptr);
600     block_begin()[E - 1] = nullptr;
601     setNumHungOffUseOperands(getNumOperands() - 1);
602   }
603 
604   // After deleting entries that satisfy Pred, remaining entries may have
605   // changed order.
606   template <typename Fn> void unorderedDeleteIncomingIf(Fn &&Pred) {
607     for (unsigned I = 0, E = getNumOperands(); I != E; ++I)
608       if (Pred(getIncomingValue(I), getIncomingBlock(I))) {
609         unorderedDeleteIncoming(I);
610         E = getNumOperands();
611         --I;
612       }
613     assert(getNumOperands() >= 1 &&
614            "Cannot remove all incoming blocks in a MemoryPhi.");
615   }
616 
617   // After deleting incoming block BB, the incoming blocks order may be changed.
618   void unorderedDeleteIncomingBlock(const BasicBlock *BB) {
619     unorderedDeleteIncomingIf(
620         [&](const MemoryAccess *, const BasicBlock *B) { return BB == B; });
621   }
622 
623   // After deleting incoming memory access MA, the incoming accesses order may
624   // be changed.
625   void unorderedDeleteIncomingValue(const MemoryAccess *MA) {
626     unorderedDeleteIncomingIf(
627         [&](const MemoryAccess *M, const BasicBlock *) { return MA == M; });
628   }
629 
630   static bool classof(const Value *V) {
631     return V->getValueID() == MemoryPhiVal;
632   }
633 
634   void print(raw_ostream &OS) const;
635 
636   unsigned getID() const { return ID; }
637 
638 protected:
639   friend class MemorySSA;
640 
641   /// this is more complicated than the generic
642   /// User::allocHungoffUses, because we have to allocate Uses for the incoming
643   /// values and pointers to the incoming blocks, all in one allocation.
644   void allocHungoffUses(unsigned N) {
645     User::allocHungoffUses(N, /* IsPhi */ true);
646   }
647 
648 private:
649   // For debugging only
650   const unsigned ID;
651   unsigned ReservedSpace;
652 
653   /// This grows the operand list in response to a push_back style of
654   /// operation.  This grows the number of ops by 1.5 times.
655   void growOperands() {
656     unsigned E = getNumOperands();
657     // 2 op PHI nodes are VERY common, so reserve at least enough for that.
658     ReservedSpace = std::max(E + E / 2, 2u);
659     growHungoffUses(ReservedSpace, /* IsPhi */ true);
660   }
661 
662   static void deleteMe(DerivedUser *Self);
663 };
664 
665 inline unsigned MemoryAccess::getID() const {
666   assert((isa<MemoryDef>(this) || isa<MemoryPhi>(this)) &&
667          "only memory defs and phis have ids");
668   if (const auto *MD = dyn_cast<MemoryDef>(this))
669     return MD->getID();
670   return cast<MemoryPhi>(this)->getID();
671 }
672 
673 inline bool MemoryUseOrDef::isOptimized() const {
674   if (const auto *MD = dyn_cast<MemoryDef>(this))
675     return MD->isOptimized();
676   return cast<MemoryUse>(this)->isOptimized();
677 }
678 
679 inline MemoryAccess *MemoryUseOrDef::getOptimized() const {
680   if (const auto *MD = dyn_cast<MemoryDef>(this))
681     return MD->getOptimized();
682   return cast<MemoryUse>(this)->getOptimized();
683 }
684 
685 inline void MemoryUseOrDef::setOptimized(MemoryAccess *MA) {
686   if (auto *MD = dyn_cast<MemoryDef>(this))
687     MD->setOptimized(MA);
688   else
689     cast<MemoryUse>(this)->setOptimized(MA);
690 }
691 
692 inline void MemoryUseOrDef::resetOptimized() {
693   if (auto *MD = dyn_cast<MemoryDef>(this))
694     MD->resetOptimized();
695   else
696     cast<MemoryUse>(this)->resetOptimized();
697 }
698 
699 template <> struct OperandTraits<MemoryPhi> : public HungoffOperandTraits<2> {};
700 DEFINE_TRANSPARENT_OPERAND_ACCESSORS(MemoryPhi, MemoryAccess)
701 
702 /// Encapsulates MemorySSA, including all data associated with memory
703 /// accesses.
704 class MemorySSA {
705 public:
706   MemorySSA(Function &, AliasAnalysis *, DominatorTree *);
707 
708   // MemorySSA must remain where it's constructed; Walkers it creates store
709   // pointers to it.
710   MemorySSA(MemorySSA &&) = delete;
711 
712   ~MemorySSA();
713 
714   MemorySSAWalker *getWalker();
715   MemorySSAWalker *getSkipSelfWalker();
716 
717   /// Given a memory Mod/Ref'ing instruction, get the MemorySSA
718   /// access associated with it. If passed a basic block gets the memory phi
719   /// node that exists for that block, if there is one. Otherwise, this will get
720   /// a MemoryUseOrDef.
721   MemoryUseOrDef *getMemoryAccess(const Instruction *I) const {
722     return cast_or_null<MemoryUseOrDef>(ValueToMemoryAccess.lookup(I));
723   }
724 
725   MemoryPhi *getMemoryAccess(const BasicBlock *BB) const {
726     return cast_or_null<MemoryPhi>(ValueToMemoryAccess.lookup(cast<Value>(BB)));
727   }
728 
729   void dump() const;
730   void print(raw_ostream &) const;
731 
732   /// Return true if \p MA represents the live on entry value
733   ///
734   /// Loads and stores from pointer arguments and other global values may be
735   /// defined by memory operations that do not occur in the current function, so
736   /// they may be live on entry to the function. MemorySSA represents such
737   /// memory state by the live on entry definition, which is guaranteed to occur
738   /// before any other memory access in the function.
739   inline bool isLiveOnEntryDef(const MemoryAccess *MA) const {
740     return MA == LiveOnEntryDef.get();
741   }
742 
743   inline MemoryAccess *getLiveOnEntryDef() const {
744     return LiveOnEntryDef.get();
745   }
746 
747   // Sadly, iplists, by default, owns and deletes pointers added to the
748   // list. It's not currently possible to have two iplists for the same type,
749   // where one owns the pointers, and one does not. This is because the traits
750   // are per-type, not per-tag.  If this ever changes, we should make the
751   // DefList an iplist.
752   using AccessList = iplist<MemoryAccess, ilist_tag<MSSAHelpers::AllAccessTag>>;
753   using DefsList =
754       simple_ilist<MemoryAccess, ilist_tag<MSSAHelpers::DefsOnlyTag>>;
755 
756   /// Return the list of MemoryAccess's for a given basic block.
757   ///
758   /// This list is not modifiable by the user.
759   const AccessList *getBlockAccesses(const BasicBlock *BB) const {
760     return getWritableBlockAccesses(BB);
761   }
762 
763   /// Return the list of MemoryDef's and MemoryPhi's for a given basic
764   /// block.
765   ///
766   /// This list is not modifiable by the user.
767   const DefsList *getBlockDefs(const BasicBlock *BB) const {
768     return getWritableBlockDefs(BB);
769   }
770 
771   /// Given two memory accesses in the same basic block, determine
772   /// whether MemoryAccess \p A dominates MemoryAccess \p B.
773   bool locallyDominates(const MemoryAccess *A, const MemoryAccess *B) const;
774 
775   /// Given two memory accesses in potentially different blocks,
776   /// determine whether MemoryAccess \p A dominates MemoryAccess \p B.
777   bool dominates(const MemoryAccess *A, const MemoryAccess *B) const;
778 
779   /// Given a MemoryAccess and a Use, determine whether MemoryAccess \p A
780   /// dominates Use \p B.
781   bool dominates(const MemoryAccess *A, const Use &B) const;
782 
783   /// Verify that MemorySSA is self consistent (IE definitions dominate
784   /// all uses, uses appear in the right places).  This is used by unit tests.
785   void verifyMemorySSA() const;
786 
787   /// Used in various insertion functions to specify whether we are talking
788   /// about the beginning or end of a block.
789   enum InsertionPlace { Beginning, End, BeforeTerminator };
790 
791 protected:
792   // Used by Memory SSA annotater, dumpers, and wrapper pass
793   friend class MemorySSAAnnotatedWriter;
794   friend class MemorySSAPrinterLegacyPass;
795   friend class MemorySSAUpdater;
796 
797   void verifyOrderingDominationAndDefUses(Function &F) const;
798   void verifyDominationNumbers(const Function &F) const;
799   void verifyPrevDefInPhis(Function &F) const;
800 
801   // This is used by the use optimizer and updater.
802   AccessList *getWritableBlockAccesses(const BasicBlock *BB) const {
803     auto It = PerBlockAccesses.find(BB);
804     return It == PerBlockAccesses.end() ? nullptr : It->second.get();
805   }
806 
807   // This is used by the use optimizer and updater.
808   DefsList *getWritableBlockDefs(const BasicBlock *BB) const {
809     auto It = PerBlockDefs.find(BB);
810     return It == PerBlockDefs.end() ? nullptr : It->second.get();
811   }
812 
813   // These is used by the updater to perform various internal MemorySSA
814   // machinsations.  They do not always leave the IR in a correct state, and
815   // relies on the updater to fixup what it breaks, so it is not public.
816 
817   void moveTo(MemoryUseOrDef *What, BasicBlock *BB, AccessList::iterator Where);
818   void moveTo(MemoryAccess *What, BasicBlock *BB, InsertionPlace Point);
819 
820   // Rename the dominator tree branch rooted at BB.
821   void renamePass(BasicBlock *BB, MemoryAccess *IncomingVal,
822                   SmallPtrSetImpl<BasicBlock *> &Visited) {
823     renamePass(DT->getNode(BB), IncomingVal, Visited, true, true);
824   }
825 
826   void removeFromLookups(MemoryAccess *);
827   void removeFromLists(MemoryAccess *, bool ShouldDelete = true);
828   void insertIntoListsForBlock(MemoryAccess *, const BasicBlock *,
829                                InsertionPlace);
830   void insertIntoListsBefore(MemoryAccess *, const BasicBlock *,
831                              AccessList::iterator);
832   MemoryUseOrDef *createDefinedAccess(Instruction *, MemoryAccess *,
833                                       const MemoryUseOrDef *Template = nullptr,
834                                       bool CreationMustSucceed = true);
835 
836 private:
837   template <class AliasAnalysisType> class ClobberWalkerBase;
838   template <class AliasAnalysisType> class CachingWalker;
839   template <class AliasAnalysisType> class SkipSelfWalker;
840   class OptimizeUses;
841 
842   CachingWalker<AliasAnalysis> *getWalkerImpl();
843   void buildMemorySSA(BatchAAResults &BAA);
844   void optimizeUses();
845 
846   void prepareForMoveTo(MemoryAccess *, BasicBlock *);
847   void verifyUseInDefs(MemoryAccess *, MemoryAccess *) const;
848 
849   using AccessMap = DenseMap<const BasicBlock *, std::unique_ptr<AccessList>>;
850   using DefsMap = DenseMap<const BasicBlock *, std::unique_ptr<DefsList>>;
851 
852   void
853   determineInsertionPoint(const SmallPtrSetImpl<BasicBlock *> &DefiningBlocks);
854   void markUnreachableAsLiveOnEntry(BasicBlock *BB);
855   bool dominatesUse(const MemoryAccess *, const MemoryAccess *) const;
856   MemoryPhi *createMemoryPhi(BasicBlock *BB);
857   template <typename AliasAnalysisType>
858   MemoryUseOrDef *createNewAccess(Instruction *, AliasAnalysisType *,
859                                   const MemoryUseOrDef *Template = nullptr);
860   MemoryAccess *findDominatingDef(BasicBlock *, enum InsertionPlace);
861   void placePHINodes(const SmallPtrSetImpl<BasicBlock *> &);
862   MemoryAccess *renameBlock(BasicBlock *, MemoryAccess *, bool);
863   void renameSuccessorPhis(BasicBlock *, MemoryAccess *, bool);
864   void renamePass(DomTreeNode *, MemoryAccess *IncomingVal,
865                   SmallPtrSetImpl<BasicBlock *> &Visited,
866                   bool SkipVisited = false, bool RenameAllUses = false);
867   AccessList *getOrCreateAccessList(const BasicBlock *);
868   DefsList *getOrCreateDefsList(const BasicBlock *);
869   void renumberBlock(const BasicBlock *) const;
870   AliasAnalysis *AA;
871   DominatorTree *DT;
872   Function &F;
873 
874   // Memory SSA mappings
875   DenseMap<const Value *, MemoryAccess *> ValueToMemoryAccess;
876 
877   // These two mappings contain the main block to access/def mappings for
878   // MemorySSA. The list contained in PerBlockAccesses really owns all the
879   // MemoryAccesses.
880   // Both maps maintain the invariant that if a block is found in them, the
881   // corresponding list is not empty, and if a block is not found in them, the
882   // corresponding list is empty.
883   AccessMap PerBlockAccesses;
884   DefsMap PerBlockDefs;
885   std::unique_ptr<MemoryAccess, ValueDeleter> LiveOnEntryDef;
886 
887   // Domination mappings
888   // Note that the numbering is local to a block, even though the map is
889   // global.
890   mutable SmallPtrSet<const BasicBlock *, 16> BlockNumberingValid;
891   mutable DenseMap<const MemoryAccess *, unsigned long> BlockNumbering;
892 
893   // Memory SSA building info
894   std::unique_ptr<ClobberWalkerBase<AliasAnalysis>> WalkerBase;
895   std::unique_ptr<CachingWalker<AliasAnalysis>> Walker;
896   std::unique_ptr<SkipSelfWalker<AliasAnalysis>> SkipWalker;
897   unsigned NextID;
898 };
899 
900 // Internal MemorySSA utils, for use by MemorySSA classes and walkers
901 class MemorySSAUtil {
902 protected:
903   friend class GVNHoist;
904   friend class MemorySSAWalker;
905 
906   // This function should not be used by new passes.
907   static bool defClobbersUseOrDef(MemoryDef *MD, const MemoryUseOrDef *MU,
908                                   AliasAnalysis &AA);
909 };
910 
911 // This pass does eager building and then printing of MemorySSA. It is used by
912 // the tests to be able to build, dump, and verify Memory SSA.
913 class MemorySSAPrinterLegacyPass : public FunctionPass {
914 public:
915   MemorySSAPrinterLegacyPass();
916 
917   bool runOnFunction(Function &) override;
918   void getAnalysisUsage(AnalysisUsage &AU) const override;
919 
920   static char ID;
921 };
922 
923 /// An analysis that produces \c MemorySSA for a function.
924 ///
925 class MemorySSAAnalysis : public AnalysisInfoMixin<MemorySSAAnalysis> {
926   friend AnalysisInfoMixin<MemorySSAAnalysis>;
927 
928   static AnalysisKey Key;
929 
930 public:
931   // Wrap MemorySSA result to ensure address stability of internal MemorySSA
932   // pointers after construction.  Use a wrapper class instead of plain
933   // unique_ptr<MemorySSA> to avoid build breakage on MSVC.
934   struct Result {
935     Result(std::unique_ptr<MemorySSA> &&MSSA) : MSSA(std::move(MSSA)) {}
936 
937     MemorySSA &getMSSA() { return *MSSA.get(); }
938 
939     std::unique_ptr<MemorySSA> MSSA;
940 
941     bool invalidate(Function &F, const PreservedAnalyses &PA,
942                     FunctionAnalysisManager::Invalidator &Inv);
943   };
944 
945   Result run(Function &F, FunctionAnalysisManager &AM);
946 };
947 
948 /// Printer pass for \c MemorySSA.
949 class MemorySSAPrinterPass : public PassInfoMixin<MemorySSAPrinterPass> {
950   raw_ostream &OS;
951 
952 public:
953   explicit MemorySSAPrinterPass(raw_ostream &OS) : OS(OS) {}
954 
955   PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
956 };
957 
958 /// Verifier pass for \c MemorySSA.
959 struct MemorySSAVerifierPass : PassInfoMixin<MemorySSAVerifierPass> {
960   PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
961 };
962 
963 /// Legacy analysis pass which computes \c MemorySSA.
964 class MemorySSAWrapperPass : public FunctionPass {
965 public:
966   MemorySSAWrapperPass();
967 
968   static char ID;
969 
970   bool runOnFunction(Function &) override;
971   void releaseMemory() override;
972   MemorySSA &getMSSA() { return *MSSA; }
973   const MemorySSA &getMSSA() const { return *MSSA; }
974 
975   void getAnalysisUsage(AnalysisUsage &AU) const override;
976 
977   void verifyAnalysis() const override;
978   void print(raw_ostream &OS, const Module *M = nullptr) const override;
979 
980 private:
981   std::unique_ptr<MemorySSA> MSSA;
982 };
983 
984 /// This is the generic walker interface for walkers of MemorySSA.
985 /// Walkers are used to be able to further disambiguate the def-use chains
986 /// MemorySSA gives you, or otherwise produce better info than MemorySSA gives
987 /// you.
988 /// In particular, while the def-use chains provide basic information, and are
989 /// guaranteed to give, for example, the nearest may-aliasing MemoryDef for a
990 /// MemoryUse as AliasAnalysis considers it, a user mant want better or other
991 /// information. In particular, they may want to use SCEV info to further
992 /// disambiguate memory accesses, or they may want the nearest dominating
993 /// may-aliasing MemoryDef for a call or a store. This API enables a
994 /// standardized interface to getting and using that info.
995 class MemorySSAWalker {
996 public:
997   MemorySSAWalker(MemorySSA *);
998   virtual ~MemorySSAWalker() = default;
999 
1000   using MemoryAccessSet = SmallVector<MemoryAccess *, 8>;
1001 
1002   /// Given a memory Mod/Ref/ModRef'ing instruction, calling this
1003   /// will give you the nearest dominating MemoryAccess that Mod's the location
1004   /// the instruction accesses (by skipping any def which AA can prove does not
1005   /// alias the location(s) accessed by the instruction given).
1006   ///
1007   /// Note that this will return a single access, and it must dominate the
1008   /// Instruction, so if an operand of a MemoryPhi node Mod's the instruction,
1009   /// this will return the MemoryPhi, not the operand. This means that
1010   /// given:
1011   /// if (a) {
1012   ///   1 = MemoryDef(liveOnEntry)
1013   ///   store %a
1014   /// } else {
1015   ///   2 = MemoryDef(liveOnEntry)
1016   ///   store %b
1017   /// }
1018   /// 3 = MemoryPhi(2, 1)
1019   /// MemoryUse(3)
1020   /// load %a
1021   ///
1022   /// calling this API on load(%a) will return the MemoryPhi, not the MemoryDef
1023   /// in the if (a) branch.
1024   MemoryAccess *getClobberingMemoryAccess(const Instruction *I) {
1025     MemoryAccess *MA = MSSA->getMemoryAccess(I);
1026     assert(MA && "Handed an instruction that MemorySSA doesn't recognize?");
1027     return getClobberingMemoryAccess(MA);
1028   }
1029 
1030   /// Does the same thing as getClobberingMemoryAccess(const Instruction *I),
1031   /// but takes a MemoryAccess instead of an Instruction.
1032   virtual MemoryAccess *getClobberingMemoryAccess(MemoryAccess *) = 0;
1033 
1034   /// Given a potentially clobbering memory access and a new location,
1035   /// calling this will give you the nearest dominating clobbering MemoryAccess
1036   /// (by skipping non-aliasing def links).
1037   ///
1038   /// This version of the function is mainly used to disambiguate phi translated
1039   /// pointers, where the value of a pointer may have changed from the initial
1040   /// memory access. Note that this expects to be handed either a MemoryUse,
1041   /// or an already potentially clobbering access. Unlike the above API, if
1042   /// given a MemoryDef that clobbers the pointer as the starting access, it
1043   /// will return that MemoryDef, whereas the above would return the clobber
1044   /// starting from the use side of  the memory def.
1045   virtual MemoryAccess *getClobberingMemoryAccess(MemoryAccess *,
1046                                                   const MemoryLocation &) = 0;
1047 
1048   /// Given a memory access, invalidate anything this walker knows about
1049   /// that access.
1050   /// This API is used by walkers that store information to perform basic cache
1051   /// invalidation.  This will be called by MemorySSA at appropriate times for
1052   /// the walker it uses or returns.
1053   virtual void invalidateInfo(MemoryAccess *) {}
1054 
1055 protected:
1056   friend class MemorySSA; // For updating MSSA pointer in MemorySSA move
1057                           // constructor.
1058   MemorySSA *MSSA;
1059 };
1060 
1061 /// A MemorySSAWalker that does no alias queries, or anything else. It
1062 /// simply returns the links as they were constructed by the builder.
1063 class DoNothingMemorySSAWalker final : public MemorySSAWalker {
1064 public:
1065   // Keep the overrides below from hiding the Instruction overload of
1066   // getClobberingMemoryAccess.
1067   using MemorySSAWalker::getClobberingMemoryAccess;
1068 
1069   MemoryAccess *getClobberingMemoryAccess(MemoryAccess *) override;
1070   MemoryAccess *getClobberingMemoryAccess(MemoryAccess *,
1071                                           const MemoryLocation &) override;
1072 };
1073 
1074 using MemoryAccessPair = std::pair<MemoryAccess *, MemoryLocation>;
1075 using ConstMemoryAccessPair = std::pair<const MemoryAccess *, MemoryLocation>;
1076 
1077 /// Iterator base class used to implement const and non-const iterators
1078 /// over the defining accesses of a MemoryAccess.
1079 template <class T>
1080 class memoryaccess_def_iterator_base
1081     : public iterator_facade_base<memoryaccess_def_iterator_base<T>,
1082                                   std::forward_iterator_tag, T, ptrdiff_t, T *,
1083                                   T *> {
1084   using BaseT = typename memoryaccess_def_iterator_base::iterator_facade_base;
1085 
1086 public:
1087   memoryaccess_def_iterator_base(T *Start) : Access(Start) {}
1088   memoryaccess_def_iterator_base() = default;
1089 
1090   bool operator==(const memoryaccess_def_iterator_base &Other) const {
1091     return Access == Other.Access && (!Access || ArgNo == Other.ArgNo);
1092   }
1093 
1094   // This is a bit ugly, but for MemoryPHI's, unlike PHINodes, you can't get the
1095   // block from the operand in constant time (In a PHINode, the uselist has
1096   // both, so it's just subtraction). We provide it as part of the
1097   // iterator to avoid callers having to linear walk to get the block.
1098   // If the operation becomes constant time on MemoryPHI's, this bit of
1099   // abstraction breaking should be removed.
1100   BasicBlock *getPhiArgBlock() const {
1101     MemoryPhi *MP = dyn_cast<MemoryPhi>(Access);
1102     assert(MP && "Tried to get phi arg block when not iterating over a PHI");
1103     return MP->getIncomingBlock(ArgNo);
1104   }
1105 
1106   typename BaseT::iterator::pointer operator*() const {
1107     assert(Access && "Tried to access past the end of our iterator");
1108     // Go to the first argument for phis, and the defining access for everything
1109     // else.
1110     if (const MemoryPhi *MP = dyn_cast<MemoryPhi>(Access))
1111       return MP->getIncomingValue(ArgNo);
1112     return cast<MemoryUseOrDef>(Access)->getDefiningAccess();
1113   }
1114 
1115   using BaseT::operator++;
1116   memoryaccess_def_iterator_base &operator++() {
1117     assert(Access && "Hit end of iterator");
1118     if (const MemoryPhi *MP = dyn_cast<MemoryPhi>(Access)) {
1119       if (++ArgNo >= MP->getNumIncomingValues()) {
1120         ArgNo = 0;
1121         Access = nullptr;
1122       }
1123     } else {
1124       Access = nullptr;
1125     }
1126     return *this;
1127   }
1128 
1129 private:
1130   T *Access = nullptr;
1131   unsigned ArgNo = 0;
1132 };
1133 
1134 inline memoryaccess_def_iterator MemoryAccess::defs_begin() {
1135   return memoryaccess_def_iterator(this);
1136 }
1137 
1138 inline const_memoryaccess_def_iterator MemoryAccess::defs_begin() const {
1139   return const_memoryaccess_def_iterator(this);
1140 }
1141 
1142 inline memoryaccess_def_iterator MemoryAccess::defs_end() {
1143   return memoryaccess_def_iterator();
1144 }
1145 
1146 inline const_memoryaccess_def_iterator MemoryAccess::defs_end() const {
1147   return const_memoryaccess_def_iterator();
1148 }
1149 
1150 /// GraphTraits for a MemoryAccess, which walks defs in the normal case,
1151 /// and uses in the inverse case.
1152 template <> struct GraphTraits<MemoryAccess *> {
1153   using NodeRef = MemoryAccess *;
1154   using ChildIteratorType = memoryaccess_def_iterator;
1155 
1156   static NodeRef getEntryNode(NodeRef N) { return N; }
1157   static ChildIteratorType child_begin(NodeRef N) { return N->defs_begin(); }
1158   static ChildIteratorType child_end(NodeRef N) { return N->defs_end(); }
1159 };
1160 
1161 template <> struct GraphTraits<Inverse<MemoryAccess *>> {
1162   using NodeRef = MemoryAccess *;
1163   using ChildIteratorType = MemoryAccess::iterator;
1164 
1165   static NodeRef getEntryNode(NodeRef N) { return N; }
1166   static ChildIteratorType child_begin(NodeRef N) { return N->user_begin(); }
1167   static ChildIteratorType child_end(NodeRef N) { return N->user_end(); }
1168 };
1169 
1170 /// Provide an iterator that walks defs, giving both the memory access,
1171 /// and the current pointer location, updating the pointer location as it
1172 /// changes due to phi node translation.
1173 ///
1174 /// This iterator, while somewhat specialized, is what most clients actually
1175 /// want when walking upwards through MemorySSA def chains. It takes a pair of
1176 /// <MemoryAccess,MemoryLocation>, and walks defs, properly translating the
1177 /// memory location through phi nodes for the user.
1178 class upward_defs_iterator
1179     : public iterator_facade_base<upward_defs_iterator,
1180                                   std::forward_iterator_tag,
1181                                   const MemoryAccessPair> {
1182   using BaseT = upward_defs_iterator::iterator_facade_base;
1183 
1184 public:
1185   upward_defs_iterator(const MemoryAccessPair &Info)
1186       : DefIterator(Info.first), Location(Info.second),
1187         OriginalAccess(Info.first) {
1188     CurrentPair.first = nullptr;
1189 
1190     WalkingPhi = Info.first && isa<MemoryPhi>(Info.first);
1191     fillInCurrentPair();
1192   }
1193 
1194   upward_defs_iterator() { CurrentPair.first = nullptr; }
1195 
1196   bool operator==(const upward_defs_iterator &Other) const {
1197     return DefIterator == Other.DefIterator;
1198   }
1199 
1200   BaseT::iterator::reference operator*() const {
1201     assert(DefIterator != OriginalAccess->defs_end() &&
1202            "Tried to access past the end of our iterator");
1203     return CurrentPair;
1204   }
1205 
1206   using BaseT::operator++;
1207   upward_defs_iterator &operator++() {
1208     assert(DefIterator != OriginalAccess->defs_end() &&
1209            "Tried to access past the end of the iterator");
1210     ++DefIterator;
1211     if (DefIterator != OriginalAccess->defs_end())
1212       fillInCurrentPair();
1213     return *this;
1214   }
1215 
1216   BasicBlock *getPhiArgBlock() const { return DefIterator.getPhiArgBlock(); }
1217 
1218 private:
1219   void fillInCurrentPair() {
1220     CurrentPair.first = *DefIterator;
1221     if (WalkingPhi && Location.Ptr) {
1222       PHITransAddr Translator(
1223           const_cast<Value *>(Location.Ptr),
1224           OriginalAccess->getBlock()->getModule()->getDataLayout(), nullptr);
1225       if (!Translator.PHITranslateValue(OriginalAccess->getBlock(),
1226                                         DefIterator.getPhiArgBlock(), nullptr,
1227                                         false))
1228         if (Translator.getAddr() != Location.Ptr) {
1229           CurrentPair.second = Location.getWithNewPtr(Translator.getAddr());
1230           return;
1231         }
1232     }
1233     CurrentPair.second = Location;
1234   }
1235 
1236   MemoryAccessPair CurrentPair;
1237   memoryaccess_def_iterator DefIterator;
1238   MemoryLocation Location;
1239   MemoryAccess *OriginalAccess = nullptr;
1240   bool WalkingPhi = false;
1241 };
1242 
1243 inline upward_defs_iterator upward_defs_begin(const MemoryAccessPair &Pair) {
1244   return upward_defs_iterator(Pair);
1245 }
1246 
1247 inline upward_defs_iterator upward_defs_end() { return upward_defs_iterator(); }
1248 
1249 inline iterator_range<upward_defs_iterator>
1250 upward_defs(const MemoryAccessPair &Pair) {
1251   return make_range(upward_defs_begin(Pair), upward_defs_end());
1252 }
1253 
1254 /// Walks the defining accesses of MemoryDefs. Stops after we hit something that
1255 /// has no defining use (e.g. a MemoryPhi or liveOnEntry). Note that, when
1256 /// comparing against a null def_chain_iterator, this will compare equal only
1257 /// after walking said Phi/liveOnEntry.
1258 ///
1259 /// The UseOptimizedChain flag specifies whether to walk the clobbering
1260 /// access chain, or all the accesses.
1261 ///
1262 /// Normally, MemoryDef are all just def/use linked together, so a def_chain on
1263 /// a MemoryDef will walk all MemoryDefs above it in the program until it hits
1264 /// a phi node.  The optimized chain walks the clobbering access of a store.
1265 /// So if you are just trying to find, given a store, what the next
1266 /// thing that would clobber the same memory is, you want the optimized chain.
1267 template <class T, bool UseOptimizedChain = false>
1268 struct def_chain_iterator
1269     : public iterator_facade_base<def_chain_iterator<T, UseOptimizedChain>,
1270                                   std::forward_iterator_tag, MemoryAccess *> {
1271   def_chain_iterator() : MA(nullptr) {}
1272   def_chain_iterator(T MA) : MA(MA) {}
1273 
1274   T operator*() const { return MA; }
1275 
1276   def_chain_iterator &operator++() {
1277     // N.B. liveOnEntry has a null defining access.
1278     if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA)) {
1279       if (UseOptimizedChain && MUD->isOptimized())
1280         MA = MUD->getOptimized();
1281       else
1282         MA = MUD->getDefiningAccess();
1283     } else {
1284       MA = nullptr;
1285     }
1286 
1287     return *this;
1288   }
1289 
1290   bool operator==(const def_chain_iterator &O) const { return MA == O.MA; }
1291 
1292 private:
1293   T MA;
1294 };
1295 
1296 template <class T>
1297 inline iterator_range<def_chain_iterator<T>>
1298 def_chain(T MA, MemoryAccess *UpTo = nullptr) {
1299 #ifdef EXPENSIVE_CHECKS
1300   assert((!UpTo || find(def_chain(MA), UpTo) != def_chain_iterator<T>()) &&
1301          "UpTo isn't in the def chain!");
1302 #endif
1303   return make_range(def_chain_iterator<T>(MA), def_chain_iterator<T>(UpTo));
1304 }
1305 
1306 template <class T>
1307 inline iterator_range<def_chain_iterator<T, true>> optimized_def_chain(T MA) {
1308   return make_range(def_chain_iterator<T, true>(MA),
1309                     def_chain_iterator<T, true>(nullptr));
1310 }
1311 
1312 } // end namespace llvm
1313 
1314 #endif // LLVM_ANALYSIS_MEMORYSSA_H
1315