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 | /// In addition to the def/use graph described above, MemoryDefs also contain |
71 | /// an "optimized" definition use. The "optimized" use points to some def |
72 | /// reachable through the memory def chain. The optimized def *may* (but is |
73 | /// not required to) alias the original MemoryDef, but no def *closer* to the |
74 | /// source def may alias it. As the name implies, the purpose of the optimized |
75 | /// use is to allow caching of clobber searches for memory defs. The optimized |
76 | /// def may be nullptr, in which case clients must walk the defining access |
77 | /// chain. |
78 | /// |
79 | /// When iterating the uses of a MemoryDef, both defining uses and optimized |
80 | /// uses will be encountered. If only one type is needed, the client must |
81 | /// filter the use walk. |
82 | // |
83 | //===----------------------------------------------------------------------===// |
84 | |
85 | #ifndef LLVM_ANALYSIS_MEMORYSSA_H |
86 | #define LLVM_ANALYSIS_MEMORYSSA_H |
87 | |
88 | #include "llvm/ADT/DenseMap.h" |
89 | #include "llvm/ADT/SmallPtrSet.h" |
90 | #include "llvm/ADT/SmallVector.h" |
91 | #include "llvm/ADT/ilist_node.h" |
92 | #include "llvm/ADT/iterator_range.h" |
93 | #include "llvm/Analysis/AliasAnalysis.h" |
94 | #include "llvm/Analysis/MemoryLocation.h" |
95 | #include "llvm/Analysis/PHITransAddr.h" |
96 | #include "llvm/IR/DerivedUser.h" |
97 | #include "llvm/IR/Dominators.h" |
98 | #include "llvm/IR/Type.h" |
99 | #include "llvm/IR/User.h" |
100 | #include "llvm/Pass.h" |
101 | #include <algorithm> |
102 | #include <cassert> |
103 | #include <cstddef> |
104 | #include <iterator> |
105 | #include <memory> |
106 | #include <utility> |
107 | |
108 | namespace llvm { |
109 | |
110 | template <class GraphType> struct GraphTraits; |
111 | class BasicBlock; |
112 | class Function; |
113 | class Loop; |
114 | class Instruction; |
115 | class LLVMContext; |
116 | class MemoryAccess; |
117 | class MemorySSAWalker; |
118 | class Module; |
119 | class Use; |
120 | class Value; |
121 | class raw_ostream; |
122 | |
123 | namespace MSSAHelpers { |
124 | |
125 | struct AllAccessTag {}; |
126 | struct DefsOnlyTag {}; |
127 | |
128 | } // end namespace MSSAHelpers |
129 | |
130 | enum : unsigned { |
131 | // Used to signify what the default invalid ID is for MemoryAccess's |
132 | // getID() |
133 | INVALID_MEMORYACCESS_ID = -1U |
134 | }; |
135 | |
136 | template <class T> class memoryaccess_def_iterator_base; |
137 | using memoryaccess_def_iterator = memoryaccess_def_iterator_base<MemoryAccess>; |
138 | using const_memoryaccess_def_iterator = |
139 | memoryaccess_def_iterator_base<const MemoryAccess>; |
140 | |
141 | // The base for all memory accesses. All memory accesses in a block are |
142 | // linked together using an intrusive list. |
143 | class MemoryAccess |
144 | : public DerivedUser, |
145 | public ilist_node<MemoryAccess, ilist_tag<MSSAHelpers::AllAccessTag>>, |
146 | public ilist_node<MemoryAccess, ilist_tag<MSSAHelpers::DefsOnlyTag>> { |
147 | public: |
148 | using AllAccessType = |
149 | ilist_node<MemoryAccess, ilist_tag<MSSAHelpers::AllAccessTag>>; |
150 | using DefsOnlyType = |
151 | ilist_node<MemoryAccess, ilist_tag<MSSAHelpers::DefsOnlyTag>>; |
152 | |
153 | MemoryAccess(const MemoryAccess &) = delete; |
154 | MemoryAccess &operator=(const MemoryAccess &) = delete; |
155 | |
156 | void *operator new(size_t) = delete; |
157 | |
158 | // Methods for support type inquiry through isa, cast, and |
159 | // dyn_cast |
160 | static bool classof(const Value *V) { |
161 | unsigned ID = V->getValueID(); |
162 | return ID == MemoryUseVal || ID == MemoryPhiVal || ID == MemoryDefVal; |
163 | } |
164 | |
165 | BasicBlock *getBlock() const { return Block; } |
166 | |
167 | void print(raw_ostream &OS) const; |
168 | void dump() const; |
169 | |
170 | /// The user iterators for a memory access |
171 | using iterator = user_iterator; |
172 | using const_iterator = const_user_iterator; |
173 | |
174 | /// This iterator walks over all of the defs in a given |
175 | /// MemoryAccess. For MemoryPhi nodes, this walks arguments. For |
176 | /// MemoryUse/MemoryDef, this walks the defining access. |
177 | memoryaccess_def_iterator defs_begin(); |
178 | const_memoryaccess_def_iterator defs_begin() const; |
179 | memoryaccess_def_iterator defs_end(); |
180 | const_memoryaccess_def_iterator defs_end() const; |
181 | |
182 | /// Get the iterators for the all access list and the defs only list |
183 | /// We default to the all access list. |
184 | AllAccessType::self_iterator getIterator() { |
185 | return this->AllAccessType::getIterator(); |
186 | } |
187 | AllAccessType::const_self_iterator getIterator() const { |
188 | return this->AllAccessType::getIterator(); |
189 | } |
190 | AllAccessType::reverse_self_iterator getReverseIterator() { |
191 | return this->AllAccessType::getReverseIterator(); |
192 | } |
193 | AllAccessType::const_reverse_self_iterator getReverseIterator() const { |
194 | return this->AllAccessType::getReverseIterator(); |
195 | } |
196 | DefsOnlyType::self_iterator getDefsIterator() { |
197 | return this->DefsOnlyType::getIterator(); |
198 | } |
199 | DefsOnlyType::const_self_iterator getDefsIterator() const { |
200 | return this->DefsOnlyType::getIterator(); |
201 | } |
202 | DefsOnlyType::reverse_self_iterator getReverseDefsIterator() { |
203 | return this->DefsOnlyType::getReverseIterator(); |
204 | } |
205 | DefsOnlyType::const_reverse_self_iterator getReverseDefsIterator() const { |
206 | return this->DefsOnlyType::getReverseIterator(); |
207 | } |
208 | |
209 | protected: |
210 | friend class MemoryDef; |
211 | friend class MemoryPhi; |
212 | friend class MemorySSA; |
213 | friend class MemoryUse; |
214 | friend class MemoryUseOrDef; |
215 | |
216 | /// Used by MemorySSA to change the block of a MemoryAccess when it is |
217 | /// moved. |
218 | void setBlock(BasicBlock *BB) { Block = BB; } |
219 | |
220 | /// Used for debugging and tracking things about MemoryAccesses. |
221 | /// Guaranteed unique among MemoryAccesses, no guarantees otherwise. |
222 | inline unsigned getID() const; |
223 | |
224 | MemoryAccess(LLVMContext &C, unsigned Vty, DeleteValueTy DeleteValue, |
225 | BasicBlock *BB, unsigned NumOperands) |
226 | : DerivedUser(Type::getVoidTy(C), Vty, nullptr, NumOperands, DeleteValue), |
227 | Block(BB) {} |
228 | |
229 | // Use deleteValue() to delete a generic MemoryAccess. |
230 | ~MemoryAccess() = default; |
231 | |
232 | private: |
233 | BasicBlock *Block; |
234 | }; |
235 | |
236 | template <> |
237 | struct ilist_alloc_traits<MemoryAccess> { |
238 | static void deleteNode(MemoryAccess *MA) { MA->deleteValue(); } |
239 | }; |
240 | |
241 | inline raw_ostream &operator<<(raw_ostream &OS, const MemoryAccess &MA) { |
242 | MA.print(OS); |
243 | return OS; |
244 | } |
245 | |
246 | /// Class that has the common methods + fields of memory uses/defs. It's |
247 | /// a little awkward to have, but there are many cases where we want either a |
248 | /// use or def, and there are many cases where uses are needed (defs aren't |
249 | /// acceptable), and vice-versa. |
250 | /// |
251 | /// This class should never be instantiated directly; make a MemoryUse or |
252 | /// MemoryDef instead. |
253 | class MemoryUseOrDef : public MemoryAccess { |
254 | public: |
255 | void *operator new(size_t) = delete; |
256 | |
257 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(MemoryAccess); |
258 | |
259 | /// Get the instruction that this MemoryUse represents. |
260 | Instruction *getMemoryInst() const { return MemoryInstruction; } |
261 | |
262 | /// Get the access that produces the memory state used by this Use. |
263 | MemoryAccess *getDefiningAccess() const { return getOperand(0); } |
264 | |
265 | static bool classof(const Value *MA) { |
266 | return MA->getValueID() == MemoryUseVal || MA->getValueID() == MemoryDefVal; |
267 | } |
268 | |
269 | /// Do we have an optimized use? |
270 | inline bool isOptimized() const; |
271 | /// Return the MemoryAccess associated with the optimized use, or nullptr. |
272 | inline MemoryAccess *getOptimized() const; |
273 | /// Sets the optimized use for a MemoryDef. |
274 | inline void setOptimized(MemoryAccess *); |
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) { |
290 | setDefiningAccess(DMA); |
291 | } |
292 | |
293 | // Use deleteValue() to delete a generic MemoryUseOrDef. |
294 | ~MemoryUseOrDef() = default; |
295 | |
296 | void setDefiningAccess(MemoryAccess *DMA, bool Optimized = false) { |
297 | if (!Optimized) { |
298 | setOperand(0, DMA); |
299 | return; |
300 | } |
301 | setOptimized(DMA); |
302 | } |
303 | |
304 | private: |
305 | Instruction *MemoryInstruction; |
306 | }; |
307 | |
308 | /// Represents read-only accesses to memory |
309 | /// |
310 | /// In particular, the set of Instructions that will be represented by |
311 | /// MemoryUse's is exactly the set of Instructions for which |
312 | /// AliasAnalysis::getModRefInfo returns "Ref". |
313 | class MemoryUse final : public MemoryUseOrDef { |
314 | public: |
315 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(MemoryAccess); |
316 | |
317 | MemoryUse(LLVMContext &C, MemoryAccess *DMA, Instruction *MI, BasicBlock *BB) |
318 | : MemoryUseOrDef(C, DMA, MemoryUseVal, deleteMe, MI, BB, |
319 | /*NumOperands=*/1) {} |
320 | |
321 | // allocate space for exactly one operand |
322 | void *operator new(size_t S) { return User::operator new(Size: S, Us: 1); } |
323 | void operator delete(void *Ptr) { User::operator delete(Usr: Ptr); } |
324 | |
325 | static bool classof(const Value *MA) { |
326 | return MA->getValueID() == MemoryUseVal; |
327 | } |
328 | |
329 | void print(raw_ostream &OS) const; |
330 | |
331 | void setOptimized(MemoryAccess *DMA) { |
332 | OptimizedID = DMA->getID(); |
333 | setOperand(0, DMA); |
334 | } |
335 | |
336 | /// Whether the MemoryUse is optimized. If ensureOptimizedUses() was called, |
337 | /// uses will usually be optimized, but this is not guaranteed (e.g. due to |
338 | /// invalidation and optimization limits.) |
339 | bool isOptimized() const { |
340 | return getDefiningAccess() && OptimizedID == getDefiningAccess()->getID(); |
341 | } |
342 | |
343 | MemoryAccess *getOptimized() const { |
344 | return getDefiningAccess(); |
345 | } |
346 | |
347 | void resetOptimized() { |
348 | OptimizedID = INVALID_MEMORYACCESS_ID; |
349 | } |
350 | |
351 | protected: |
352 | friend class MemorySSA; |
353 | |
354 | private: |
355 | static void deleteMe(DerivedUser *Self); |
356 | |
357 | unsigned OptimizedID = INVALID_MEMORYACCESS_ID; |
358 | }; |
359 | |
360 | template <> |
361 | struct OperandTraits<MemoryUse> : public FixedNumOperandTraits<MemoryUse, 1> {}; |
362 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(MemoryUse, MemoryAccess) |
363 | |
364 | /// Represents a read-write access to memory, whether it is a must-alias, |
365 | /// or a may-alias. |
366 | /// |
367 | /// In particular, the set of Instructions that will be represented by |
368 | /// MemoryDef's is exactly the set of Instructions for which |
369 | /// AliasAnalysis::getModRefInfo returns "Mod" or "ModRef". |
370 | /// Note that, in order to provide def-def chains, all defs also have a use |
371 | /// associated with them. This use points to the nearest reaching |
372 | /// MemoryDef/MemoryPhi. |
373 | class MemoryDef final : public MemoryUseOrDef { |
374 | public: |
375 | friend class MemorySSA; |
376 | |
377 | DECLARE_TRANSPARENT_OPERAND_ACCESSORS(MemoryAccess); |
378 | |
379 | MemoryDef(LLVMContext &C, MemoryAccess *DMA, Instruction *MI, BasicBlock *BB, |
380 | unsigned Ver) |
381 | : MemoryUseOrDef(C, DMA, MemoryDefVal, deleteMe, MI, BB, |
382 | /*NumOperands=*/2), |
383 | ID(Ver) {} |
384 | |
385 | // allocate space for exactly two operands |
386 | void *operator new(size_t S) { return User::operator new(Size: S, Us: 2); } |
387 | void operator delete(void *Ptr) { User::operator delete(Usr: Ptr); } |
388 | |
389 | static bool classof(const Value *MA) { |
390 | return MA->getValueID() == MemoryDefVal; |
391 | } |
392 | |
393 | void setOptimized(MemoryAccess *MA) { |
394 | setOperand(1, MA); |
395 | OptimizedID = MA->getID(); |
396 | } |
397 | |
398 | MemoryAccess *getOptimized() const { |
399 | return cast_or_null<MemoryAccess>(Val: getOperand(1)); |
400 | } |
401 | |
402 | bool isOptimized() const { |
403 | return getOptimized() && OptimizedID == getOptimized()->getID(); |
404 | } |
405 | |
406 | void resetOptimized() { |
407 | OptimizedID = INVALID_MEMORYACCESS_ID; |
408 | setOperand(1, nullptr); |
409 | } |
410 | |
411 | void print(raw_ostream &OS) const; |
412 | |
413 | unsigned getID() const { return ID; } |
414 | |
415 | private: |
416 | static void deleteMe(DerivedUser *Self); |
417 | |
418 | const unsigned ID; |
419 | unsigned OptimizedID = INVALID_MEMORYACCESS_ID; |
420 | }; |
421 | |
422 | template <> |
423 | struct OperandTraits<MemoryDef> : public FixedNumOperandTraits<MemoryDef, 2> {}; |
424 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(MemoryDef, MemoryAccess) |
425 | |
426 | template <> |
427 | struct OperandTraits<MemoryUseOrDef> { |
428 | static Use *op_begin(MemoryUseOrDef *MUD) { |
429 | if (auto *MU = dyn_cast<MemoryUse>(Val: MUD)) |
430 | return OperandTraits<MemoryUse>::op_begin(U: MU); |
431 | return OperandTraits<MemoryDef>::op_begin(U: cast<MemoryDef>(Val: MUD)); |
432 | } |
433 | |
434 | static Use *op_end(MemoryUseOrDef *MUD) { |
435 | if (auto *MU = dyn_cast<MemoryUse>(Val: MUD)) |
436 | return OperandTraits<MemoryUse>::op_end(U: MU); |
437 | return OperandTraits<MemoryDef>::op_end(U: cast<MemoryDef>(Val: MUD)); |
438 | } |
439 | |
440 | static unsigned operands(const MemoryUseOrDef *MUD) { |
441 | if (const auto *MU = dyn_cast<MemoryUse>(Val: MUD)) |
442 | return OperandTraits<MemoryUse>::operands(MU); |
443 | return OperandTraits<MemoryDef>::operands(cast<MemoryDef>(Val: MUD)); |
444 | } |
445 | }; |
446 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(MemoryUseOrDef, MemoryAccess) |
447 | |
448 | /// Represents phi nodes for memory accesses. |
449 | /// |
450 | /// These have the same semantic as regular phi nodes, with the exception that |
451 | /// only one phi will ever exist in a given basic block. |
452 | /// Guaranteeing one phi per block means guaranteeing there is only ever one |
453 | /// valid reaching MemoryDef/MemoryPHI along each path to the phi node. |
454 | /// This is ensured by not allowing disambiguation of the RHS of a MemoryDef or |
455 | /// a MemoryPhi's operands. |
456 | /// That is, given |
457 | /// if (a) { |
458 | /// store %a |
459 | /// store %b |
460 | /// } |
461 | /// it *must* be transformed into |
462 | /// if (a) { |
463 | /// 1 = MemoryDef(liveOnEntry) |
464 | /// store %a |
465 | /// 2 = MemoryDef(1) |
466 | /// store %b |
467 | /// } |
468 | /// and *not* |
469 | /// if (a) { |
470 | /// 1 = MemoryDef(liveOnEntry) |
471 | /// store %a |
472 | /// 2 = MemoryDef(liveOnEntry) |
473 | /// store %b |
474 | /// } |
475 | /// even if the two stores do not conflict. Otherwise, both 1 and 2 reach the |
476 | /// end of the branch, and if there are not two phi nodes, one will be |
477 | /// disconnected completely from the SSA graph below that point. |
478 | /// Because MemoryUse's do not generate new definitions, they do not have this |
479 | /// issue. |
480 | class MemoryPhi final : public MemoryAccess { |
481 | // allocate space for exactly zero operands |
482 | void *operator new(size_t S) { return User::operator new(Size: S); } |
483 | |
484 | public: |
485 | void operator delete(void *Ptr) { User::operator delete(Usr: Ptr); } |
486 | |
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(N: 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 | return reinterpret_cast<block_iterator>(op_begin() + ReservedSpace); |
503 | } |
504 | |
505 | const_block_iterator block_begin() const { |
506 | return reinterpret_cast<const_block_iterator>(op_begin() + ReservedSpace); |
507 | } |
508 | |
509 | block_iterator block_end() { return block_begin() + getNumOperands(); } |
510 | |
511 | const_block_iterator block_end() const { |
512 | return block_begin() + getNumOperands(); |
513 | } |
514 | |
515 | iterator_range<block_iterator> blocks() { |
516 | return make_range(x: block_begin(), y: block_end()); |
517 | } |
518 | |
519 | iterator_range<const_block_iterator> blocks() const { |
520 | return make_range(x: block_begin(), y: block_end()); |
521 | } |
522 | |
523 | op_range incoming_values() { return operands(); } |
524 | |
525 | const_op_range incoming_values() const { return operands(); } |
526 | |
527 | /// Return the number of incoming edges |
528 | unsigned getNumIncomingValues() const { return getNumOperands(); } |
529 | |
530 | /// Return incoming value number x |
531 | MemoryAccess *getIncomingValue(unsigned I) const { return getOperand(I); } |
532 | void setIncomingValue(unsigned I, MemoryAccess *V) { |
533 | assert(V && "PHI node got a null value!" ); |
534 | setOperand(I, V); |
535 | } |
536 | |
537 | static unsigned getOperandNumForIncomingValue(unsigned I) { return I; } |
538 | static unsigned getIncomingValueNumForOperand(unsigned I) { return I; } |
539 | |
540 | /// Return incoming basic block number @p i. |
541 | BasicBlock *getIncomingBlock(unsigned I) const { return block_begin()[I]; } |
542 | |
543 | /// Return incoming basic block corresponding |
544 | /// to an operand of the PHI. |
545 | BasicBlock *getIncomingBlock(const Use &U) const { |
546 | assert(this == U.getUser() && "Iterator doesn't point to PHI's Uses?" ); |
547 | return getIncomingBlock(I: unsigned(&U - op_begin())); |
548 | } |
549 | |
550 | /// Return incoming basic block corresponding |
551 | /// to value use iterator. |
552 | BasicBlock *getIncomingBlock(MemoryAccess::const_user_iterator I) const { |
553 | return getIncomingBlock(U: I.getUse()); |
554 | } |
555 | |
556 | void setIncomingBlock(unsigned I, BasicBlock *BB) { |
557 | assert(BB && "PHI node got a null basic block!" ); |
558 | block_begin()[I] = BB; |
559 | } |
560 | |
561 | /// Add an incoming value to the end of the PHI list |
562 | void addIncoming(MemoryAccess *V, BasicBlock *BB) { |
563 | if (getNumOperands() == ReservedSpace) |
564 | growOperands(); // Get more space! |
565 | // Initialize some new operands. |
566 | setNumHungOffUseOperands(getNumOperands() + 1); |
567 | setIncomingValue(I: getNumOperands() - 1, V); |
568 | setIncomingBlock(I: getNumOperands() - 1, BB); |
569 | } |
570 | |
571 | /// Return the first index of the specified basic |
572 | /// block in the value list for this PHI. Returns -1 if no instance. |
573 | int getBasicBlockIndex(const BasicBlock *BB) const { |
574 | for (unsigned I = 0, E = getNumOperands(); I != E; ++I) |
575 | if (block_begin()[I] == BB) |
576 | return I; |
577 | return -1; |
578 | } |
579 | |
580 | MemoryAccess *getIncomingValueForBlock(const BasicBlock *BB) const { |
581 | int Idx = getBasicBlockIndex(BB); |
582 | assert(Idx >= 0 && "Invalid basic block argument!" ); |
583 | return getIncomingValue(I: Idx); |
584 | } |
585 | |
586 | // After deleting incoming position I, the order of incoming may be changed. |
587 | void unorderedDeleteIncoming(unsigned I) { |
588 | unsigned E = getNumOperands(); |
589 | assert(I < E && "Cannot remove out of bounds Phi entry." ); |
590 | // MemoryPhi must have at least two incoming values, otherwise the MemoryPhi |
591 | // itself should be deleted. |
592 | assert(E >= 2 && "Cannot only remove incoming values in MemoryPhis with " |
593 | "at least 2 values." ); |
594 | setIncomingValue(I, V: getIncomingValue(I: E - 1)); |
595 | setIncomingBlock(I, BB: block_begin()[E - 1]); |
596 | setOperand(E - 1, nullptr); |
597 | block_begin()[E - 1] = nullptr; |
598 | setNumHungOffUseOperands(getNumOperands() - 1); |
599 | } |
600 | |
601 | // After deleting entries that satisfy Pred, remaining entries may have |
602 | // changed order. |
603 | template <typename Fn> void unorderedDeleteIncomingIf(Fn &&Pred) { |
604 | for (unsigned I = 0, E = getNumOperands(); I != E; ++I) |
605 | if (Pred(getIncomingValue(I), getIncomingBlock(I))) { |
606 | unorderedDeleteIncoming(I); |
607 | E = getNumOperands(); |
608 | --I; |
609 | } |
610 | assert(getNumOperands() >= 1 && |
611 | "Cannot remove all incoming blocks in a MemoryPhi." ); |
612 | } |
613 | |
614 | // After deleting incoming block BB, the incoming blocks order may be changed. |
615 | void unorderedDeleteIncomingBlock(const BasicBlock *BB) { |
616 | unorderedDeleteIncomingIf( |
617 | Pred: [&](const MemoryAccess *, const BasicBlock *B) { return BB == B; }); |
618 | } |
619 | |
620 | // After deleting incoming memory access MA, the incoming accesses order may |
621 | // be changed. |
622 | void unorderedDeleteIncomingValue(const MemoryAccess *MA) { |
623 | unorderedDeleteIncomingIf( |
624 | Pred: [&](const MemoryAccess *M, const BasicBlock *) { return MA == M; }); |
625 | } |
626 | |
627 | static bool classof(const Value *V) { |
628 | return V->getValueID() == MemoryPhiVal; |
629 | } |
630 | |
631 | void print(raw_ostream &OS) const; |
632 | |
633 | unsigned getID() const { return ID; } |
634 | |
635 | protected: |
636 | friend class MemorySSA; |
637 | |
638 | /// this is more complicated than the generic |
639 | /// User::allocHungoffUses, because we have to allocate Uses for the incoming |
640 | /// values and pointers to the incoming blocks, all in one allocation. |
641 | void allocHungoffUses(unsigned N) { |
642 | User::allocHungoffUses(N, /* IsPhi */ IsPhi: true); |
643 | } |
644 | |
645 | private: |
646 | // For debugging only |
647 | const unsigned ID; |
648 | unsigned ReservedSpace; |
649 | |
650 | /// This grows the operand list in response to a push_back style of |
651 | /// operation. This grows the number of ops by 1.5 times. |
652 | void growOperands() { |
653 | unsigned E = getNumOperands(); |
654 | // 2 op PHI nodes are VERY common, so reserve at least enough for that. |
655 | ReservedSpace = std::max(a: E + E / 2, b: 2u); |
656 | growHungoffUses(N: ReservedSpace, /* IsPhi */ IsPhi: true); |
657 | } |
658 | |
659 | static void deleteMe(DerivedUser *Self); |
660 | }; |
661 | |
662 | inline unsigned MemoryAccess::getID() const { |
663 | assert((isa<MemoryDef>(this) || isa<MemoryPhi>(this)) && |
664 | "only memory defs and phis have ids" ); |
665 | if (const auto *MD = dyn_cast<MemoryDef>(Val: this)) |
666 | return MD->getID(); |
667 | return cast<MemoryPhi>(Val: this)->getID(); |
668 | } |
669 | |
670 | inline bool MemoryUseOrDef::isOptimized() const { |
671 | if (const auto *MD = dyn_cast<MemoryDef>(Val: this)) |
672 | return MD->isOptimized(); |
673 | return cast<MemoryUse>(Val: this)->isOptimized(); |
674 | } |
675 | |
676 | inline MemoryAccess *MemoryUseOrDef::getOptimized() const { |
677 | if (const auto *MD = dyn_cast<MemoryDef>(Val: this)) |
678 | return MD->getOptimized(); |
679 | return cast<MemoryUse>(Val: this)->getOptimized(); |
680 | } |
681 | |
682 | inline void MemoryUseOrDef::setOptimized(MemoryAccess *MA) { |
683 | if (auto *MD = dyn_cast<MemoryDef>(Val: this)) |
684 | MD->setOptimized(MA); |
685 | else |
686 | cast<MemoryUse>(Val: this)->setOptimized(MA); |
687 | } |
688 | |
689 | inline void MemoryUseOrDef::resetOptimized() { |
690 | if (auto *MD = dyn_cast<MemoryDef>(Val: this)) |
691 | MD->resetOptimized(); |
692 | else |
693 | cast<MemoryUse>(Val: this)->resetOptimized(); |
694 | } |
695 | |
696 | template <> struct OperandTraits<MemoryPhi> : public HungoffOperandTraits<2> {}; |
697 | DEFINE_TRANSPARENT_OPERAND_ACCESSORS(MemoryPhi, MemoryAccess) |
698 | |
699 | /// Encapsulates MemorySSA, including all data associated with memory |
700 | /// accesses. |
701 | class MemorySSA { |
702 | public: |
703 | MemorySSA(Function &, AliasAnalysis *, DominatorTree *); |
704 | MemorySSA(Loop &, AliasAnalysis *, DominatorTree *); |
705 | |
706 | // MemorySSA must remain where it's constructed; Walkers it creates store |
707 | // pointers to it. |
708 | MemorySSA(MemorySSA &&) = delete; |
709 | |
710 | ~MemorySSA(); |
711 | |
712 | MemorySSAWalker *getWalker(); |
713 | MemorySSAWalker *getSkipSelfWalker(); |
714 | |
715 | /// Given a memory Mod/Ref'ing instruction, get the MemorySSA |
716 | /// access associated with it. If passed a basic block gets the memory phi |
717 | /// node that exists for that block, if there is one. Otherwise, this will get |
718 | /// a MemoryUseOrDef. |
719 | MemoryUseOrDef *getMemoryAccess(const Instruction *I) const { |
720 | return cast_or_null<MemoryUseOrDef>(Val: ValueToMemoryAccess.lookup(Val: I)); |
721 | } |
722 | |
723 | MemoryPhi *getMemoryAccess(const BasicBlock *BB) const { |
724 | return cast_or_null<MemoryPhi>(Val: ValueToMemoryAccess.lookup(Val: cast<Value>(Val: BB))); |
725 | } |
726 | |
727 | DominatorTree &getDomTree() const { return *DT; } |
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 | enum class VerificationLevel { Fast, Full }; |
784 | /// Verify that MemorySSA is self consistent (IE definitions dominate |
785 | /// all uses, uses appear in the right places). This is used by unit tests. |
786 | void verifyMemorySSA(VerificationLevel = VerificationLevel::Fast) const; |
787 | |
788 | /// Used in various insertion functions to specify whether we are talking |
789 | /// about the beginning or end of a block. |
790 | enum InsertionPlace { Beginning, End, BeforeTerminator }; |
791 | |
792 | /// By default, uses are *not* optimized during MemorySSA construction. |
793 | /// Calling this method will attempt to optimize all MemoryUses, if this has |
794 | /// not happened yet for this MemorySSA instance. This should be done if you |
795 | /// plan to query the clobbering access for most uses, or if you walk the |
796 | /// def-use chain of uses. |
797 | void ensureOptimizedUses(); |
798 | |
799 | AliasAnalysis &getAA() { return *AA; } |
800 | |
801 | protected: |
802 | // Used by Memory SSA dumpers and wrapper pass |
803 | friend class MemorySSAUpdater; |
804 | |
805 | template <typename IterT> |
806 | void verifyOrderingDominationAndDefUses( |
807 | IterT Blocks, VerificationLevel = VerificationLevel::Fast) const; |
808 | template <typename IterT> void verifyDominationNumbers(IterT Blocks) const; |
809 | template <typename IterT> void verifyPrevDefInPhis(IterT Blocks) const; |
810 | |
811 | // This is used by the use optimizer and updater. |
812 | AccessList *getWritableBlockAccesses(const BasicBlock *BB) const { |
813 | auto It = PerBlockAccesses.find(Val: BB); |
814 | return It == PerBlockAccesses.end() ? nullptr : It->second.get(); |
815 | } |
816 | |
817 | // This is used by the use optimizer and updater. |
818 | DefsList *getWritableBlockDefs(const BasicBlock *BB) const { |
819 | auto It = PerBlockDefs.find(Val: BB); |
820 | return It == PerBlockDefs.end() ? nullptr : It->second.get(); |
821 | } |
822 | |
823 | // These is used by the updater to perform various internal MemorySSA |
824 | // machinsations. They do not always leave the IR in a correct state, and |
825 | // relies on the updater to fixup what it breaks, so it is not public. |
826 | |
827 | void moveTo(MemoryUseOrDef *What, BasicBlock *BB, AccessList::iterator Where); |
828 | void moveTo(MemoryAccess *What, BasicBlock *BB, InsertionPlace Point); |
829 | |
830 | // Rename the dominator tree branch rooted at BB. |
831 | void renamePass(BasicBlock *BB, MemoryAccess *IncomingVal, |
832 | SmallPtrSetImpl<BasicBlock *> &Visited) { |
833 | renamePass(DT->getNode(BB), IncomingVal, Visited, SkipVisited: true, RenameAllUses: true); |
834 | } |
835 | |
836 | void removeFromLookups(MemoryAccess *); |
837 | void removeFromLists(MemoryAccess *, bool ShouldDelete = true); |
838 | void insertIntoListsForBlock(MemoryAccess *, const BasicBlock *, |
839 | InsertionPlace); |
840 | void insertIntoListsBefore(MemoryAccess *, const BasicBlock *, |
841 | AccessList::iterator); |
842 | MemoryUseOrDef *createDefinedAccess(Instruction *, MemoryAccess *, |
843 | const MemoryUseOrDef *Template = nullptr, |
844 | bool CreationMustSucceed = true); |
845 | |
846 | private: |
847 | class ClobberWalkerBase; |
848 | class CachingWalker; |
849 | class SkipSelfWalker; |
850 | class OptimizeUses; |
851 | |
852 | CachingWalker *getWalkerImpl(); |
853 | template <typename IterT> |
854 | void buildMemorySSA(BatchAAResults &BAA, IterT Blocks); |
855 | |
856 | void prepareForMoveTo(MemoryAccess *, BasicBlock *); |
857 | void verifyUseInDefs(MemoryAccess *, MemoryAccess *) const; |
858 | |
859 | using AccessMap = DenseMap<const BasicBlock *, std::unique_ptr<AccessList>>; |
860 | using DefsMap = DenseMap<const BasicBlock *, std::unique_ptr<DefsList>>; |
861 | |
862 | void markUnreachableAsLiveOnEntry(BasicBlock *BB); |
863 | MemoryPhi *createMemoryPhi(BasicBlock *BB); |
864 | template <typename AliasAnalysisType> |
865 | MemoryUseOrDef *createNewAccess(Instruction *, AliasAnalysisType *, |
866 | const MemoryUseOrDef *Template = nullptr); |
867 | void placePHINodes(const SmallPtrSetImpl<BasicBlock *> &); |
868 | MemoryAccess *renameBlock(BasicBlock *, MemoryAccess *, bool); |
869 | void renameSuccessorPhis(BasicBlock *, MemoryAccess *, bool); |
870 | void renamePass(DomTreeNode *, MemoryAccess *IncomingVal, |
871 | SmallPtrSetImpl<BasicBlock *> &Visited, |
872 | bool SkipVisited = false, bool RenameAllUses = false); |
873 | AccessList *getOrCreateAccessList(const BasicBlock *); |
874 | DefsList *getOrCreateDefsList(const BasicBlock *); |
875 | void renumberBlock(const BasicBlock *) const; |
876 | AliasAnalysis *AA = nullptr; |
877 | DominatorTree *DT; |
878 | Function *F = nullptr; |
879 | Loop *L = nullptr; |
880 | |
881 | // Memory SSA mappings |
882 | DenseMap<const Value *, MemoryAccess *> ValueToMemoryAccess; |
883 | |
884 | // These two mappings contain the main block to access/def mappings for |
885 | // MemorySSA. The list contained in PerBlockAccesses really owns all the |
886 | // MemoryAccesses. |
887 | // Both maps maintain the invariant that if a block is found in them, the |
888 | // corresponding list is not empty, and if a block is not found in them, the |
889 | // corresponding list is empty. |
890 | AccessMap PerBlockAccesses; |
891 | DefsMap PerBlockDefs; |
892 | std::unique_ptr<MemoryAccess, ValueDeleter> LiveOnEntryDef; |
893 | |
894 | // Domination mappings |
895 | // Note that the numbering is local to a block, even though the map is |
896 | // global. |
897 | mutable SmallPtrSet<const BasicBlock *, 16> BlockNumberingValid; |
898 | mutable DenseMap<const MemoryAccess *, unsigned long> BlockNumbering; |
899 | |
900 | // Memory SSA building info |
901 | std::unique_ptr<ClobberWalkerBase> WalkerBase; |
902 | std::unique_ptr<CachingWalker> Walker; |
903 | std::unique_ptr<SkipSelfWalker> SkipWalker; |
904 | unsigned NextID = 0; |
905 | bool IsOptimized = false; |
906 | }; |
907 | |
908 | /// Enables verification of MemorySSA. |
909 | /// |
910 | /// The checks which this flag enables is exensive and disabled by default |
911 | /// unless `EXPENSIVE_CHECKS` is defined. The flag `-verify-memoryssa` can be |
912 | /// used to selectively enable the verification without re-compilation. |
913 | extern bool VerifyMemorySSA; |
914 | |
915 | // Internal MemorySSA utils, for use by MemorySSA classes and walkers |
916 | class MemorySSAUtil { |
917 | protected: |
918 | friend class GVNHoist; |
919 | friend class MemorySSAWalker; |
920 | |
921 | // This function should not be used by new passes. |
922 | static bool defClobbersUseOrDef(MemoryDef *MD, const MemoryUseOrDef *MU, |
923 | AliasAnalysis &AA); |
924 | }; |
925 | |
926 | /// An analysis that produces \c MemorySSA for a function. |
927 | /// |
928 | class MemorySSAAnalysis : public AnalysisInfoMixin<MemorySSAAnalysis> { |
929 | friend AnalysisInfoMixin<MemorySSAAnalysis>; |
930 | |
931 | static AnalysisKey Key; |
932 | |
933 | public: |
934 | // Wrap MemorySSA result to ensure address stability of internal MemorySSA |
935 | // pointers after construction. Use a wrapper class instead of plain |
936 | // unique_ptr<MemorySSA> to avoid build breakage on MSVC. |
937 | struct Result { |
938 | Result(std::unique_ptr<MemorySSA> &&MSSA) : MSSA(std::move(MSSA)) {} |
939 | |
940 | MemorySSA &getMSSA() { return *MSSA; } |
941 | |
942 | std::unique_ptr<MemorySSA> MSSA; |
943 | |
944 | bool invalidate(Function &F, const PreservedAnalyses &PA, |
945 | FunctionAnalysisManager::Invalidator &Inv); |
946 | }; |
947 | |
948 | Result run(Function &F, FunctionAnalysisManager &AM); |
949 | }; |
950 | |
951 | /// Printer pass for \c MemorySSA. |
952 | class MemorySSAPrinterPass : public PassInfoMixin<MemorySSAPrinterPass> { |
953 | raw_ostream &OS; |
954 | bool EnsureOptimizedUses; |
955 | |
956 | public: |
957 | explicit MemorySSAPrinterPass(raw_ostream &OS, bool EnsureOptimizedUses) |
958 | : OS(OS), EnsureOptimizedUses(EnsureOptimizedUses) {} |
959 | |
960 | PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM); |
961 | |
962 | static bool isRequired() { return true; } |
963 | }; |
964 | |
965 | /// Printer pass for \c MemorySSA via the walker. |
966 | class MemorySSAWalkerPrinterPass |
967 | : public PassInfoMixin<MemorySSAWalkerPrinterPass> { |
968 | raw_ostream &OS; |
969 | |
970 | public: |
971 | explicit MemorySSAWalkerPrinterPass(raw_ostream &OS) : OS(OS) {} |
972 | |
973 | PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM); |
974 | |
975 | static bool isRequired() { return true; } |
976 | }; |
977 | |
978 | /// Verifier pass for \c MemorySSA. |
979 | struct MemorySSAVerifierPass : PassInfoMixin<MemorySSAVerifierPass> { |
980 | PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM); |
981 | static bool isRequired() { return true; } |
982 | }; |
983 | |
984 | /// Legacy analysis pass which computes \c MemorySSA. |
985 | class MemorySSAWrapperPass : public FunctionPass { |
986 | public: |
987 | MemorySSAWrapperPass(); |
988 | |
989 | static char ID; |
990 | |
991 | bool runOnFunction(Function &) override; |
992 | void releaseMemory() override; |
993 | MemorySSA &getMSSA() { return *MSSA; } |
994 | const MemorySSA &getMSSA() const { return *MSSA; } |
995 | |
996 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
997 | |
998 | void verifyAnalysis() const override; |
999 | void print(raw_ostream &OS, const Module *M = nullptr) const override; |
1000 | |
1001 | private: |
1002 | std::unique_ptr<MemorySSA> MSSA; |
1003 | }; |
1004 | |
1005 | /// This is the generic walker interface for walkers of MemorySSA. |
1006 | /// Walkers are used to be able to further disambiguate the def-use chains |
1007 | /// MemorySSA gives you, or otherwise produce better info than MemorySSA gives |
1008 | /// you. |
1009 | /// In particular, while the def-use chains provide basic information, and are |
1010 | /// guaranteed to give, for example, the nearest may-aliasing MemoryDef for a |
1011 | /// MemoryUse as AliasAnalysis considers it, a user mant want better or other |
1012 | /// information. In particular, they may want to use SCEV info to further |
1013 | /// disambiguate memory accesses, or they may want the nearest dominating |
1014 | /// may-aliasing MemoryDef for a call or a store. This API enables a |
1015 | /// standardized interface to getting and using that info. |
1016 | class MemorySSAWalker { |
1017 | public: |
1018 | MemorySSAWalker(MemorySSA *); |
1019 | virtual ~MemorySSAWalker() = default; |
1020 | |
1021 | using MemoryAccessSet = SmallVector<MemoryAccess *, 8>; |
1022 | |
1023 | /// Given a memory Mod/Ref/ModRef'ing instruction, calling this |
1024 | /// will give you the nearest dominating MemoryAccess that Mod's the location |
1025 | /// the instruction accesses (by skipping any def which AA can prove does not |
1026 | /// alias the location(s) accessed by the instruction given). |
1027 | /// |
1028 | /// Note that this will return a single access, and it must dominate the |
1029 | /// Instruction, so if an operand of a MemoryPhi node Mod's the instruction, |
1030 | /// this will return the MemoryPhi, not the operand. This means that |
1031 | /// given: |
1032 | /// if (a) { |
1033 | /// 1 = MemoryDef(liveOnEntry) |
1034 | /// store %a |
1035 | /// } else { |
1036 | /// 2 = MemoryDef(liveOnEntry) |
1037 | /// store %b |
1038 | /// } |
1039 | /// 3 = MemoryPhi(2, 1) |
1040 | /// MemoryUse(3) |
1041 | /// load %a |
1042 | /// |
1043 | /// calling this API on load(%a) will return the MemoryPhi, not the MemoryDef |
1044 | /// in the if (a) branch. |
1045 | MemoryAccess *getClobberingMemoryAccess(const Instruction *I, |
1046 | BatchAAResults &AA) { |
1047 | MemoryAccess *MA = MSSA->getMemoryAccess(I); |
1048 | assert(MA && "Handed an instruction that MemorySSA doesn't recognize?" ); |
1049 | return getClobberingMemoryAccess(MA, AA); |
1050 | } |
1051 | |
1052 | /// Does the same thing as getClobberingMemoryAccess(const Instruction *I), |
1053 | /// but takes a MemoryAccess instead of an Instruction. |
1054 | virtual MemoryAccess *getClobberingMemoryAccess(MemoryAccess *, |
1055 | BatchAAResults &AA) = 0; |
1056 | |
1057 | /// Given a potentially clobbering memory access and a new location, |
1058 | /// calling this will give you the nearest dominating clobbering MemoryAccess |
1059 | /// (by skipping non-aliasing def links). |
1060 | /// |
1061 | /// This version of the function is mainly used to disambiguate phi translated |
1062 | /// pointers, where the value of a pointer may have changed from the initial |
1063 | /// memory access. Note that this expects to be handed either a MemoryUse, |
1064 | /// or an already potentially clobbering access. Unlike the above API, if |
1065 | /// given a MemoryDef that clobbers the pointer as the starting access, it |
1066 | /// will return that MemoryDef, whereas the above would return the clobber |
1067 | /// starting from the use side of the memory def. |
1068 | virtual MemoryAccess *getClobberingMemoryAccess(MemoryAccess *, |
1069 | const MemoryLocation &, |
1070 | BatchAAResults &AA) = 0; |
1071 | |
1072 | MemoryAccess *getClobberingMemoryAccess(const Instruction *I) { |
1073 | BatchAAResults BAA(MSSA->getAA()); |
1074 | return getClobberingMemoryAccess(I, AA&: BAA); |
1075 | } |
1076 | |
1077 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA) { |
1078 | BatchAAResults BAA(MSSA->getAA()); |
1079 | return getClobberingMemoryAccess(MA, AA&: BAA); |
1080 | } |
1081 | |
1082 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, |
1083 | const MemoryLocation &Loc) { |
1084 | BatchAAResults BAA(MSSA->getAA()); |
1085 | return getClobberingMemoryAccess(MA, Loc, AA&: BAA); |
1086 | } |
1087 | |
1088 | /// Given a memory access, invalidate anything this walker knows about |
1089 | /// that access. |
1090 | /// This API is used by walkers that store information to perform basic cache |
1091 | /// invalidation. This will be called by MemorySSA at appropriate times for |
1092 | /// the walker it uses or returns. |
1093 | virtual void invalidateInfo(MemoryAccess *) {} |
1094 | |
1095 | protected: |
1096 | friend class MemorySSA; // For updating MSSA pointer in MemorySSA move |
1097 | // constructor. |
1098 | MemorySSA *MSSA; |
1099 | }; |
1100 | |
1101 | /// A MemorySSAWalker that does no alias queries, or anything else. It |
1102 | /// simply returns the links as they were constructed by the builder. |
1103 | class DoNothingMemorySSAWalker final : public MemorySSAWalker { |
1104 | public: |
1105 | // Keep the overrides below from hiding the Instruction overload of |
1106 | // getClobberingMemoryAccess. |
1107 | using MemorySSAWalker::getClobberingMemoryAccess; |
1108 | |
1109 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *, |
1110 | BatchAAResults &) override; |
1111 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *, |
1112 | const MemoryLocation &, |
1113 | BatchAAResults &) override; |
1114 | }; |
1115 | |
1116 | using MemoryAccessPair = std::pair<MemoryAccess *, MemoryLocation>; |
1117 | using ConstMemoryAccessPair = std::pair<const MemoryAccess *, MemoryLocation>; |
1118 | |
1119 | /// Iterator base class used to implement const and non-const iterators |
1120 | /// over the defining accesses of a MemoryAccess. |
1121 | template <class T> |
1122 | class memoryaccess_def_iterator_base |
1123 | : public iterator_facade_base<memoryaccess_def_iterator_base<T>, |
1124 | std::forward_iterator_tag, T, ptrdiff_t, T *, |
1125 | T *> { |
1126 | using BaseT = typename memoryaccess_def_iterator_base::iterator_facade_base; |
1127 | |
1128 | public: |
1129 | memoryaccess_def_iterator_base(T *Start) : Access(Start) {} |
1130 | memoryaccess_def_iterator_base() = default; |
1131 | |
1132 | bool operator==(const memoryaccess_def_iterator_base &Other) const { |
1133 | return Access == Other.Access && (!Access || ArgNo == Other.ArgNo); |
1134 | } |
1135 | |
1136 | // This is a bit ugly, but for MemoryPHI's, unlike PHINodes, you can't get the |
1137 | // block from the operand in constant time (In a PHINode, the uselist has |
1138 | // both, so it's just subtraction). We provide it as part of the |
1139 | // iterator to avoid callers having to linear walk to get the block. |
1140 | // If the operation becomes constant time on MemoryPHI's, this bit of |
1141 | // abstraction breaking should be removed. |
1142 | BasicBlock *getPhiArgBlock() const { |
1143 | MemoryPhi *MP = dyn_cast<MemoryPhi>(Access); |
1144 | assert(MP && "Tried to get phi arg block when not iterating over a PHI" ); |
1145 | return MP->getIncomingBlock(I: ArgNo); |
1146 | } |
1147 | |
1148 | typename std::iterator_traits<BaseT>::pointer operator*() const { |
1149 | assert(Access && "Tried to access past the end of our iterator" ); |
1150 | // Go to the first argument for phis, and the defining access for everything |
1151 | // else. |
1152 | if (const MemoryPhi *MP = dyn_cast<MemoryPhi>(Access)) |
1153 | return MP->getIncomingValue(I: ArgNo); |
1154 | return cast<MemoryUseOrDef>(Access)->getDefiningAccess(); |
1155 | } |
1156 | |
1157 | using BaseT::operator++; |
1158 | memoryaccess_def_iterator_base &operator++() { |
1159 | assert(Access && "Hit end of iterator" ); |
1160 | if (const MemoryPhi *MP = dyn_cast<MemoryPhi>(Access)) { |
1161 | if (++ArgNo >= MP->getNumIncomingValues()) { |
1162 | ArgNo = 0; |
1163 | Access = nullptr; |
1164 | } |
1165 | } else { |
1166 | Access = nullptr; |
1167 | } |
1168 | return *this; |
1169 | } |
1170 | |
1171 | private: |
1172 | T *Access = nullptr; |
1173 | unsigned ArgNo = 0; |
1174 | }; |
1175 | |
1176 | inline memoryaccess_def_iterator MemoryAccess::defs_begin() { |
1177 | return memoryaccess_def_iterator(this); |
1178 | } |
1179 | |
1180 | inline const_memoryaccess_def_iterator MemoryAccess::defs_begin() const { |
1181 | return const_memoryaccess_def_iterator(this); |
1182 | } |
1183 | |
1184 | inline memoryaccess_def_iterator MemoryAccess::defs_end() { |
1185 | return memoryaccess_def_iterator(); |
1186 | } |
1187 | |
1188 | inline const_memoryaccess_def_iterator MemoryAccess::defs_end() const { |
1189 | return const_memoryaccess_def_iterator(); |
1190 | } |
1191 | |
1192 | /// GraphTraits for a MemoryAccess, which walks defs in the normal case, |
1193 | /// and uses in the inverse case. |
1194 | template <> struct GraphTraits<MemoryAccess *> { |
1195 | using NodeRef = MemoryAccess *; |
1196 | using ChildIteratorType = memoryaccess_def_iterator; |
1197 | |
1198 | static NodeRef getEntryNode(NodeRef N) { return N; } |
1199 | static ChildIteratorType child_begin(NodeRef N) { return N->defs_begin(); } |
1200 | static ChildIteratorType child_end(NodeRef N) { return N->defs_end(); } |
1201 | }; |
1202 | |
1203 | template <> struct GraphTraits<Inverse<MemoryAccess *>> { |
1204 | using NodeRef = MemoryAccess *; |
1205 | using ChildIteratorType = MemoryAccess::iterator; |
1206 | |
1207 | static NodeRef getEntryNode(NodeRef N) { return N; } |
1208 | static ChildIteratorType child_begin(NodeRef N) { return N->user_begin(); } |
1209 | static ChildIteratorType child_end(NodeRef N) { return N->user_end(); } |
1210 | }; |
1211 | |
1212 | /// Provide an iterator that walks defs, giving both the memory access, |
1213 | /// and the current pointer location, updating the pointer location as it |
1214 | /// changes due to phi node translation. |
1215 | /// |
1216 | /// This iterator, while somewhat specialized, is what most clients actually |
1217 | /// want when walking upwards through MemorySSA def chains. It takes a pair of |
1218 | /// <MemoryAccess,MemoryLocation>, and walks defs, properly translating the |
1219 | /// memory location through phi nodes for the user. |
1220 | class upward_defs_iterator |
1221 | : public iterator_facade_base<upward_defs_iterator, |
1222 | std::forward_iterator_tag, |
1223 | const MemoryAccessPair> { |
1224 | using BaseT = upward_defs_iterator::iterator_facade_base; |
1225 | |
1226 | public: |
1227 | upward_defs_iterator(const MemoryAccessPair &Info, DominatorTree *DT) |
1228 | : DefIterator(Info.first), Location(Info.second), |
1229 | OriginalAccess(Info.first), DT(DT) { |
1230 | CurrentPair.first = nullptr; |
1231 | |
1232 | WalkingPhi = Info.first && isa<MemoryPhi>(Val: Info.first); |
1233 | fillInCurrentPair(); |
1234 | } |
1235 | |
1236 | upward_defs_iterator() { CurrentPair.first = nullptr; } |
1237 | |
1238 | bool operator==(const upward_defs_iterator &Other) const { |
1239 | return DefIterator == Other.DefIterator; |
1240 | } |
1241 | |
1242 | typename std::iterator_traits<BaseT>::reference operator*() const { |
1243 | assert(DefIterator != OriginalAccess->defs_end() && |
1244 | "Tried to access past the end of our iterator" ); |
1245 | return CurrentPair; |
1246 | } |
1247 | |
1248 | using BaseT::operator++; |
1249 | upward_defs_iterator &operator++() { |
1250 | assert(DefIterator != OriginalAccess->defs_end() && |
1251 | "Tried to access past the end of the iterator" ); |
1252 | ++DefIterator; |
1253 | if (DefIterator != OriginalAccess->defs_end()) |
1254 | fillInCurrentPair(); |
1255 | return *this; |
1256 | } |
1257 | |
1258 | BasicBlock *getPhiArgBlock() const { return DefIterator.getPhiArgBlock(); } |
1259 | |
1260 | private: |
1261 | /// Returns true if \p Ptr is guaranteed to be loop invariant for any possible |
1262 | /// loop. In particular, this guarantees that it only references a single |
1263 | /// MemoryLocation during execution of the containing function. |
1264 | bool IsGuaranteedLoopInvariant(const Value *Ptr) const; |
1265 | |
1266 | void fillInCurrentPair() { |
1267 | CurrentPair.first = *DefIterator; |
1268 | CurrentPair.second = Location; |
1269 | if (WalkingPhi && Location.Ptr) { |
1270 | PHITransAddr Translator( |
1271 | const_cast<Value *>(Location.Ptr), |
1272 | OriginalAccess->getBlock()->getDataLayout(), nullptr); |
1273 | |
1274 | if (Value *Addr = |
1275 | Translator.translateValue(CurBB: OriginalAccess->getBlock(), |
1276 | PredBB: DefIterator.getPhiArgBlock(), DT, MustDominate: true)) |
1277 | if (Addr != CurrentPair.second.Ptr) |
1278 | CurrentPair.second = CurrentPair.second.getWithNewPtr(NewPtr: Addr); |
1279 | |
1280 | // Mark size as unknown, if the location is not guaranteed to be |
1281 | // loop-invariant for any possible loop in the function. Setting the size |
1282 | // to unknown guarantees that any memory accesses that access locations |
1283 | // after the pointer are considered as clobbers, which is important to |
1284 | // catch loop carried dependences. |
1285 | if (!IsGuaranteedLoopInvariant(Ptr: CurrentPair.second.Ptr)) |
1286 | CurrentPair.second = CurrentPair.second.getWithNewSize( |
1287 | NewSize: LocationSize::beforeOrAfterPointer()); |
1288 | } |
1289 | } |
1290 | |
1291 | MemoryAccessPair CurrentPair; |
1292 | memoryaccess_def_iterator DefIterator; |
1293 | MemoryLocation Location; |
1294 | MemoryAccess *OriginalAccess = nullptr; |
1295 | DominatorTree *DT = nullptr; |
1296 | bool WalkingPhi = false; |
1297 | }; |
1298 | |
1299 | inline upward_defs_iterator |
1300 | upward_defs_begin(const MemoryAccessPair &Pair, DominatorTree &DT) { |
1301 | return upward_defs_iterator(Pair, &DT); |
1302 | } |
1303 | |
1304 | inline upward_defs_iterator upward_defs_end() { return upward_defs_iterator(); } |
1305 | |
1306 | inline iterator_range<upward_defs_iterator> |
1307 | upward_defs(const MemoryAccessPair &Pair, DominatorTree &DT) { |
1308 | return make_range(x: upward_defs_begin(Pair, DT), y: upward_defs_end()); |
1309 | } |
1310 | |
1311 | /// Walks the defining accesses of MemoryDefs. Stops after we hit something that |
1312 | /// has no defining use (e.g. a MemoryPhi or liveOnEntry). Note that, when |
1313 | /// comparing against a null def_chain_iterator, this will compare equal only |
1314 | /// after walking said Phi/liveOnEntry. |
1315 | /// |
1316 | /// The UseOptimizedChain flag specifies whether to walk the clobbering |
1317 | /// access chain, or all the accesses. |
1318 | /// |
1319 | /// Normally, MemoryDef are all just def/use linked together, so a def_chain on |
1320 | /// a MemoryDef will walk all MemoryDefs above it in the program until it hits |
1321 | /// a phi node. The optimized chain walks the clobbering access of a store. |
1322 | /// So if you are just trying to find, given a store, what the next |
1323 | /// thing that would clobber the same memory is, you want the optimized chain. |
1324 | template <class T, bool UseOptimizedChain = false> |
1325 | struct def_chain_iterator |
1326 | : public iterator_facade_base<def_chain_iterator<T, UseOptimizedChain>, |
1327 | std::forward_iterator_tag, MemoryAccess *> { |
1328 | def_chain_iterator() : MA(nullptr) {} |
1329 | def_chain_iterator(T MA) : MA(MA) {} |
1330 | |
1331 | T operator*() const { return MA; } |
1332 | |
1333 | def_chain_iterator &operator++() { |
1334 | // N.B. liveOnEntry has a null defining access. |
1335 | if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA)) { |
1336 | if (UseOptimizedChain && MUD->isOptimized()) |
1337 | MA = MUD->getOptimized(); |
1338 | else |
1339 | MA = MUD->getDefiningAccess(); |
1340 | } else { |
1341 | MA = nullptr; |
1342 | } |
1343 | |
1344 | return *this; |
1345 | } |
1346 | |
1347 | bool operator==(const def_chain_iterator &O) const { return MA == O.MA; } |
1348 | |
1349 | private: |
1350 | T MA; |
1351 | }; |
1352 | |
1353 | template <class T> |
1354 | inline iterator_range<def_chain_iterator<T>> |
1355 | def_chain(T MA, MemoryAccess *UpTo = nullptr) { |
1356 | #ifdef EXPENSIVE_CHECKS |
1357 | assert((!UpTo || find(def_chain(MA), UpTo) != def_chain_iterator<T>()) && |
1358 | "UpTo isn't in the def chain!" ); |
1359 | #endif |
1360 | return make_range(def_chain_iterator<T>(MA), def_chain_iterator<T>(UpTo)); |
1361 | } |
1362 | |
1363 | template <class T> |
1364 | inline iterator_range<def_chain_iterator<T, true>> optimized_def_chain(T MA) { |
1365 | return make_range(def_chain_iterator<T, true>(MA), |
1366 | def_chain_iterator<T, true>(nullptr)); |
1367 | } |
1368 | |
1369 | } // end namespace llvm |
1370 | |
1371 | #endif // LLVM_ANALYSIS_MEMORYSSA_H |
1372 | |