1 | //===- MemorySSA.cpp - Memory SSA Builder ---------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file implements the MemorySSA class. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "llvm/Analysis/MemorySSA.h" |
14 | #include "llvm/ADT/DenseMap.h" |
15 | #include "llvm/ADT/DenseMapInfo.h" |
16 | #include "llvm/ADT/DenseSet.h" |
17 | #include "llvm/ADT/DepthFirstIterator.h" |
18 | #include "llvm/ADT/Hashing.h" |
19 | #include "llvm/ADT/STLExtras.h" |
20 | #include "llvm/ADT/SmallPtrSet.h" |
21 | #include "llvm/ADT/SmallVector.h" |
22 | #include "llvm/ADT/StringExtras.h" |
23 | #include "llvm/ADT/iterator.h" |
24 | #include "llvm/ADT/iterator_range.h" |
25 | #include "llvm/Analysis/AliasAnalysis.h" |
26 | #include "llvm/Analysis/CFGPrinter.h" |
27 | #include "llvm/Analysis/IteratedDominanceFrontier.h" |
28 | #include "llvm/Analysis/MemoryLocation.h" |
29 | #include "llvm/Config/llvm-config.h" |
30 | #include "llvm/IR/AssemblyAnnotationWriter.h" |
31 | #include "llvm/IR/BasicBlock.h" |
32 | #include "llvm/IR/Dominators.h" |
33 | #include "llvm/IR/Function.h" |
34 | #include "llvm/IR/Instruction.h" |
35 | #include "llvm/IR/Instructions.h" |
36 | #include "llvm/IR/IntrinsicInst.h" |
37 | #include "llvm/IR/LLVMContext.h" |
38 | #include "llvm/IR/Operator.h" |
39 | #include "llvm/IR/PassManager.h" |
40 | #include "llvm/IR/Use.h" |
41 | #include "llvm/InitializePasses.h" |
42 | #include "llvm/Pass.h" |
43 | #include "llvm/Support/AtomicOrdering.h" |
44 | #include "llvm/Support/Casting.h" |
45 | #include "llvm/Support/CommandLine.h" |
46 | #include "llvm/Support/Compiler.h" |
47 | #include "llvm/Support/Debug.h" |
48 | #include "llvm/Support/ErrorHandling.h" |
49 | #include "llvm/Support/FormattedStream.h" |
50 | #include "llvm/Support/GraphWriter.h" |
51 | #include "llvm/Support/raw_ostream.h" |
52 | #include <algorithm> |
53 | #include <cassert> |
54 | #include <iterator> |
55 | #include <memory> |
56 | #include <utility> |
57 | |
58 | using namespace llvm; |
59 | |
60 | #define DEBUG_TYPE "memoryssa" |
61 | |
62 | static cl::opt<std::string> |
63 | DotCFGMSSA("dot-cfg-mssa" , |
64 | cl::value_desc("file name for generated dot file" ), |
65 | cl::desc("file name for generated dot file" ), cl::init(Val: "" )); |
66 | |
67 | INITIALIZE_PASS_BEGIN(MemorySSAWrapperPass, "memoryssa" , "Memory SSA" , false, |
68 | true) |
69 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) |
70 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) |
71 | INITIALIZE_PASS_END(MemorySSAWrapperPass, "memoryssa" , "Memory SSA" , false, |
72 | true) |
73 | |
74 | static cl::opt<unsigned> MaxCheckLimit( |
75 | "memssa-check-limit" , cl::Hidden, cl::init(Val: 100), |
76 | cl::desc("The maximum number of stores/phis MemorySSA" |
77 | "will consider trying to walk past (default = 100)" )); |
78 | |
79 | // Always verify MemorySSA if expensive checking is enabled. |
80 | #ifdef EXPENSIVE_CHECKS |
81 | bool llvm::VerifyMemorySSA = true; |
82 | #else |
83 | bool llvm::VerifyMemorySSA = false; |
84 | #endif |
85 | |
86 | static cl::opt<bool, true> |
87 | VerifyMemorySSAX("verify-memoryssa" , cl::location(L&: VerifyMemorySSA), |
88 | cl::Hidden, cl::desc("Enable verification of MemorySSA." )); |
89 | |
90 | const static char LiveOnEntryStr[] = "liveOnEntry" ; |
91 | |
92 | namespace { |
93 | |
94 | /// An assembly annotator class to print Memory SSA information in |
95 | /// comments. |
96 | class MemorySSAAnnotatedWriter : public AssemblyAnnotationWriter { |
97 | const MemorySSA *MSSA; |
98 | |
99 | public: |
100 | MemorySSAAnnotatedWriter(const MemorySSA *M) : MSSA(M) {} |
101 | |
102 | void emitBasicBlockStartAnnot(const BasicBlock *BB, |
103 | formatted_raw_ostream &OS) override { |
104 | if (MemoryAccess *MA = MSSA->getMemoryAccess(BB)) |
105 | OS << "; " << *MA << "\n" ; |
106 | } |
107 | |
108 | void emitInstructionAnnot(const Instruction *I, |
109 | formatted_raw_ostream &OS) override { |
110 | if (MemoryAccess *MA = MSSA->getMemoryAccess(I)) |
111 | OS << "; " << *MA << "\n" ; |
112 | } |
113 | }; |
114 | |
115 | /// An assembly annotator class to print Memory SSA information in |
116 | /// comments. |
117 | class MemorySSAWalkerAnnotatedWriter : public AssemblyAnnotationWriter { |
118 | MemorySSA *MSSA; |
119 | MemorySSAWalker *Walker; |
120 | BatchAAResults BAA; |
121 | |
122 | public: |
123 | MemorySSAWalkerAnnotatedWriter(MemorySSA *M) |
124 | : MSSA(M), Walker(M->getWalker()), BAA(M->getAA()) {} |
125 | |
126 | void emitBasicBlockStartAnnot(const BasicBlock *BB, |
127 | formatted_raw_ostream &OS) override { |
128 | if (MemoryAccess *MA = MSSA->getMemoryAccess(BB)) |
129 | OS << "; " << *MA << "\n" ; |
130 | } |
131 | |
132 | void emitInstructionAnnot(const Instruction *I, |
133 | formatted_raw_ostream &OS) override { |
134 | if (MemoryAccess *MA = MSSA->getMemoryAccess(I)) { |
135 | MemoryAccess *Clobber = Walker->getClobberingMemoryAccess(MA, AA&: BAA); |
136 | OS << "; " << *MA; |
137 | if (Clobber) { |
138 | OS << " - clobbered by " ; |
139 | if (MSSA->isLiveOnEntryDef(MA: Clobber)) |
140 | OS << LiveOnEntryStr; |
141 | else |
142 | OS << *Clobber; |
143 | } |
144 | OS << "\n" ; |
145 | } |
146 | } |
147 | }; |
148 | |
149 | } // namespace |
150 | |
151 | namespace { |
152 | |
153 | /// Our current alias analysis API differentiates heavily between calls and |
154 | /// non-calls, and functions called on one usually assert on the other. |
155 | /// This class encapsulates the distinction to simplify other code that wants |
156 | /// "Memory affecting instructions and related data" to use as a key. |
157 | /// For example, this class is used as a densemap key in the use optimizer. |
158 | class MemoryLocOrCall { |
159 | public: |
160 | bool IsCall = false; |
161 | |
162 | MemoryLocOrCall(MemoryUseOrDef *MUD) |
163 | : MemoryLocOrCall(MUD->getMemoryInst()) {} |
164 | MemoryLocOrCall(const MemoryUseOrDef *MUD) |
165 | : MemoryLocOrCall(MUD->getMemoryInst()) {} |
166 | |
167 | MemoryLocOrCall(Instruction *Inst) { |
168 | if (auto *C = dyn_cast<CallBase>(Val: Inst)) { |
169 | IsCall = true; |
170 | Call = C; |
171 | } else { |
172 | IsCall = false; |
173 | // There is no such thing as a memorylocation for a fence inst, and it is |
174 | // unique in that regard. |
175 | if (!isa<FenceInst>(Val: Inst)) |
176 | Loc = MemoryLocation::get(Inst); |
177 | } |
178 | } |
179 | |
180 | explicit MemoryLocOrCall(const MemoryLocation &Loc) : Loc(Loc) {} |
181 | |
182 | const CallBase *getCall() const { |
183 | assert(IsCall); |
184 | return Call; |
185 | } |
186 | |
187 | MemoryLocation getLoc() const { |
188 | assert(!IsCall); |
189 | return Loc; |
190 | } |
191 | |
192 | bool operator==(const MemoryLocOrCall &Other) const { |
193 | if (IsCall != Other.IsCall) |
194 | return false; |
195 | |
196 | if (!IsCall) |
197 | return Loc == Other.Loc; |
198 | |
199 | if (Call->getCalledOperand() != Other.Call->getCalledOperand()) |
200 | return false; |
201 | |
202 | return Call->arg_size() == Other.Call->arg_size() && |
203 | std::equal(first1: Call->arg_begin(), last1: Call->arg_end(), |
204 | first2: Other.Call->arg_begin()); |
205 | } |
206 | |
207 | private: |
208 | union { |
209 | const CallBase *Call; |
210 | MemoryLocation Loc; |
211 | }; |
212 | }; |
213 | |
214 | } // end anonymous namespace |
215 | |
216 | namespace llvm { |
217 | |
218 | template <> struct DenseMapInfo<MemoryLocOrCall> { |
219 | static inline MemoryLocOrCall getEmptyKey() { |
220 | return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getEmptyKey()); |
221 | } |
222 | |
223 | static inline MemoryLocOrCall getTombstoneKey() { |
224 | return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getTombstoneKey()); |
225 | } |
226 | |
227 | static unsigned getHashValue(const MemoryLocOrCall &MLOC) { |
228 | if (!MLOC.IsCall) |
229 | return hash_combine( |
230 | args: MLOC.IsCall, |
231 | args: DenseMapInfo<MemoryLocation>::getHashValue(Val: MLOC.getLoc())); |
232 | |
233 | hash_code hash = |
234 | hash_combine(args: MLOC.IsCall, args: DenseMapInfo<const Value *>::getHashValue( |
235 | PtrVal: MLOC.getCall()->getCalledOperand())); |
236 | |
237 | for (const Value *Arg : MLOC.getCall()->args()) |
238 | hash = hash_combine(args: hash, args: DenseMapInfo<const Value *>::getHashValue(PtrVal: Arg)); |
239 | return hash; |
240 | } |
241 | |
242 | static bool isEqual(const MemoryLocOrCall &LHS, const MemoryLocOrCall &RHS) { |
243 | return LHS == RHS; |
244 | } |
245 | }; |
246 | |
247 | } // end namespace llvm |
248 | |
249 | /// This does one-way checks to see if Use could theoretically be hoisted above |
250 | /// MayClobber. This will not check the other way around. |
251 | /// |
252 | /// This assumes that, for the purposes of MemorySSA, Use comes directly after |
253 | /// MayClobber, with no potentially clobbering operations in between them. |
254 | /// (Where potentially clobbering ops are memory barriers, aliased stores, etc.) |
255 | static bool areLoadsReorderable(const LoadInst *Use, |
256 | const LoadInst *MayClobber) { |
257 | bool VolatileUse = Use->isVolatile(); |
258 | bool VolatileClobber = MayClobber->isVolatile(); |
259 | // Volatile operations may never be reordered with other volatile operations. |
260 | if (VolatileUse && VolatileClobber) |
261 | return false; |
262 | // Otherwise, volatile doesn't matter here. From the language reference: |
263 | // 'optimizers may change the order of volatile operations relative to |
264 | // non-volatile operations.'" |
265 | |
266 | // If a load is seq_cst, it cannot be moved above other loads. If its ordering |
267 | // is weaker, it can be moved above other loads. We just need to be sure that |
268 | // MayClobber isn't an acquire load, because loads can't be moved above |
269 | // acquire loads. |
270 | // |
271 | // Note that this explicitly *does* allow the free reordering of monotonic (or |
272 | // weaker) loads of the same address. |
273 | bool SeqCstUse = Use->getOrdering() == AtomicOrdering::SequentiallyConsistent; |
274 | bool MayClobberIsAcquire = isAtLeastOrStrongerThan(AO: MayClobber->getOrdering(), |
275 | Other: AtomicOrdering::Acquire); |
276 | return !(SeqCstUse || MayClobberIsAcquire); |
277 | } |
278 | |
279 | template <typename AliasAnalysisType> |
280 | static bool |
281 | instructionClobbersQuery(const MemoryDef *MD, const MemoryLocation &UseLoc, |
282 | const Instruction *UseInst, AliasAnalysisType &AA) { |
283 | Instruction *DefInst = MD->getMemoryInst(); |
284 | assert(DefInst && "Defining instruction not actually an instruction" ); |
285 | |
286 | if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: DefInst)) { |
287 | // These intrinsics will show up as affecting memory, but they are just |
288 | // markers, mostly. |
289 | // |
290 | // FIXME: We probably don't actually want MemorySSA to model these at all |
291 | // (including creating MemoryAccesses for them): we just end up inventing |
292 | // clobbers where they don't really exist at all. Please see D43269 for |
293 | // context. |
294 | switch (II->getIntrinsicID()) { |
295 | case Intrinsic::allow_runtime_check: |
296 | case Intrinsic::allow_ubsan_check: |
297 | case Intrinsic::invariant_start: |
298 | case Intrinsic::invariant_end: |
299 | case Intrinsic::assume: |
300 | case Intrinsic::experimental_noalias_scope_decl: |
301 | case Intrinsic::pseudoprobe: |
302 | return false; |
303 | case Intrinsic::dbg_declare: |
304 | case Intrinsic::dbg_label: |
305 | case Intrinsic::dbg_value: |
306 | llvm_unreachable("debuginfo shouldn't have associated defs!" ); |
307 | default: |
308 | break; |
309 | } |
310 | } |
311 | |
312 | if (auto *CB = dyn_cast_or_null<CallBase>(Val: UseInst)) { |
313 | ModRefInfo I = AA.getModRefInfo(DefInst, CB); |
314 | return isModOrRefSet(MRI: I); |
315 | } |
316 | |
317 | if (auto *DefLoad = dyn_cast<LoadInst>(Val: DefInst)) |
318 | if (auto *UseLoad = dyn_cast_or_null<LoadInst>(Val: UseInst)) |
319 | return !areLoadsReorderable(Use: UseLoad, MayClobber: DefLoad); |
320 | |
321 | ModRefInfo I = AA.getModRefInfo(DefInst, UseLoc); |
322 | return isModSet(MRI: I); |
323 | } |
324 | |
325 | template <typename AliasAnalysisType> |
326 | static bool instructionClobbersQuery(MemoryDef *MD, const MemoryUseOrDef *MU, |
327 | const MemoryLocOrCall &UseMLOC, |
328 | AliasAnalysisType &AA) { |
329 | // FIXME: This is a temporary hack to allow a single instructionClobbersQuery |
330 | // to exist while MemoryLocOrCall is pushed through places. |
331 | if (UseMLOC.IsCall) |
332 | return instructionClobbersQuery(MD, MemoryLocation(), MU->getMemoryInst(), |
333 | AA); |
334 | return instructionClobbersQuery(MD, UseMLOC.getLoc(), MU->getMemoryInst(), |
335 | AA); |
336 | } |
337 | |
338 | // Return true when MD may alias MU, return false otherwise. |
339 | bool MemorySSAUtil::defClobbersUseOrDef(MemoryDef *MD, const MemoryUseOrDef *MU, |
340 | AliasAnalysis &AA) { |
341 | return instructionClobbersQuery(MD, MU, UseMLOC: MemoryLocOrCall(MU), AA); |
342 | } |
343 | |
344 | namespace { |
345 | |
346 | struct UpwardsMemoryQuery { |
347 | // True if our original query started off as a call |
348 | bool IsCall = false; |
349 | // The pointer location we started the query with. This will be empty if |
350 | // IsCall is true. |
351 | MemoryLocation StartingLoc; |
352 | // This is the instruction we were querying about. |
353 | const Instruction *Inst = nullptr; |
354 | // The MemoryAccess we actually got called with, used to test local domination |
355 | const MemoryAccess *OriginalAccess = nullptr; |
356 | bool SkipSelfAccess = false; |
357 | |
358 | UpwardsMemoryQuery() = default; |
359 | |
360 | UpwardsMemoryQuery(const Instruction *Inst, const MemoryAccess *Access) |
361 | : IsCall(isa<CallBase>(Val: Inst)), Inst(Inst), OriginalAccess(Access) { |
362 | if (!IsCall) |
363 | StartingLoc = MemoryLocation::get(Inst); |
364 | } |
365 | }; |
366 | |
367 | } // end anonymous namespace |
368 | |
369 | template <typename AliasAnalysisType> |
370 | static bool isUseTriviallyOptimizableToLiveOnEntry(AliasAnalysisType &AA, |
371 | const Instruction *I) { |
372 | // If the memory can't be changed, then loads of the memory can't be |
373 | // clobbered. |
374 | if (auto *LI = dyn_cast<LoadInst>(Val: I)) { |
375 | return I->hasMetadata(KindID: LLVMContext::MD_invariant_load) || |
376 | !isModSet(AA.getModRefInfoMask(MemoryLocation::get(LI))); |
377 | } |
378 | return false; |
379 | } |
380 | |
381 | /// Verifies that `Start` is clobbered by `ClobberAt`, and that nothing |
382 | /// inbetween `Start` and `ClobberAt` can clobbers `Start`. |
383 | /// |
384 | /// This is meant to be as simple and self-contained as possible. Because it |
385 | /// uses no cache, etc., it can be relatively expensive. |
386 | /// |
387 | /// \param Start The MemoryAccess that we want to walk from. |
388 | /// \param ClobberAt A clobber for Start. |
389 | /// \param StartLoc The MemoryLocation for Start. |
390 | /// \param MSSA The MemorySSA instance that Start and ClobberAt belong to. |
391 | /// \param Query The UpwardsMemoryQuery we used for our search. |
392 | /// \param AA The AliasAnalysis we used for our search. |
393 | /// \param AllowImpreciseClobber Always false, unless we do relaxed verify. |
394 | |
395 | LLVM_ATTRIBUTE_UNUSED static void |
396 | checkClobberSanity(const MemoryAccess *Start, MemoryAccess *ClobberAt, |
397 | const MemoryLocation &StartLoc, const MemorySSA &MSSA, |
398 | const UpwardsMemoryQuery &Query, BatchAAResults &AA, |
399 | bool AllowImpreciseClobber = false) { |
400 | assert(MSSA.dominates(ClobberAt, Start) && "Clobber doesn't dominate start?" ); |
401 | |
402 | if (MSSA.isLiveOnEntryDef(MA: Start)) { |
403 | assert(MSSA.isLiveOnEntryDef(ClobberAt) && |
404 | "liveOnEntry must clobber itself" ); |
405 | return; |
406 | } |
407 | |
408 | bool FoundClobber = false; |
409 | DenseSet<ConstMemoryAccessPair> VisitedPhis; |
410 | SmallVector<ConstMemoryAccessPair, 8> Worklist; |
411 | Worklist.emplace_back(Args&: Start, Args: StartLoc); |
412 | // Walk all paths from Start to ClobberAt, while looking for clobbers. If one |
413 | // is found, complain. |
414 | while (!Worklist.empty()) { |
415 | auto MAP = Worklist.pop_back_val(); |
416 | // All we care about is that nothing from Start to ClobberAt clobbers Start. |
417 | // We learn nothing from revisiting nodes. |
418 | if (!VisitedPhis.insert(V: MAP).second) |
419 | continue; |
420 | |
421 | for (const auto *MA : def_chain(MA: MAP.first)) { |
422 | if (MA == ClobberAt) { |
423 | if (const auto *MD = dyn_cast<MemoryDef>(Val: MA)) { |
424 | // instructionClobbersQuery isn't essentially free, so don't use `|=`, |
425 | // since it won't let us short-circuit. |
426 | // |
427 | // Also, note that this can't be hoisted out of the `Worklist` loop, |
428 | // since MD may only act as a clobber for 1 of N MemoryLocations. |
429 | FoundClobber = FoundClobber || MSSA.isLiveOnEntryDef(MA: MD); |
430 | if (!FoundClobber) { |
431 | if (instructionClobbersQuery(MD, UseLoc: MAP.second, UseInst: Query.Inst, AA)) |
432 | FoundClobber = true; |
433 | } |
434 | } |
435 | break; |
436 | } |
437 | |
438 | // We should never hit liveOnEntry, unless it's the clobber. |
439 | assert(!MSSA.isLiveOnEntryDef(MA) && "Hit liveOnEntry before clobber?" ); |
440 | |
441 | if (const auto *MD = dyn_cast<MemoryDef>(Val: MA)) { |
442 | // If Start is a Def, skip self. |
443 | if (MD == Start) |
444 | continue; |
445 | |
446 | assert(!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA) && |
447 | "Found clobber before reaching ClobberAt!" ); |
448 | continue; |
449 | } |
450 | |
451 | if (const auto *MU = dyn_cast<MemoryUse>(Val: MA)) { |
452 | (void)MU; |
453 | assert (MU == Start && |
454 | "Can only find use in def chain if Start is a use" ); |
455 | continue; |
456 | } |
457 | |
458 | assert(isa<MemoryPhi>(MA)); |
459 | |
460 | // Add reachable phi predecessors |
461 | for (auto ItB = upward_defs_begin( |
462 | Pair: {const_cast<MemoryAccess *>(MA), MAP.second}, |
463 | DT&: MSSA.getDomTree()), |
464 | ItE = upward_defs_end(); |
465 | ItB != ItE; ++ItB) |
466 | if (MSSA.getDomTree().isReachableFromEntry(A: ItB.getPhiArgBlock())) |
467 | Worklist.emplace_back(Args: *ItB); |
468 | } |
469 | } |
470 | |
471 | // If the verify is done following an optimization, it's possible that |
472 | // ClobberAt was a conservative clobbering, that we can now infer is not a |
473 | // true clobbering access. Don't fail the verify if that's the case. |
474 | // We do have accesses that claim they're optimized, but could be optimized |
475 | // further. Updating all these can be expensive, so allow it for now (FIXME). |
476 | if (AllowImpreciseClobber) |
477 | return; |
478 | |
479 | // If ClobberAt is a MemoryPhi, we can assume something above it acted as a |
480 | // clobber. Otherwise, `ClobberAt` should've acted as a clobber at some point. |
481 | assert((isa<MemoryPhi>(ClobberAt) || FoundClobber) && |
482 | "ClobberAt never acted as a clobber" ); |
483 | } |
484 | |
485 | namespace { |
486 | |
487 | /// Our algorithm for walking (and trying to optimize) clobbers, all wrapped up |
488 | /// in one class. |
489 | class ClobberWalker { |
490 | /// Save a few bytes by using unsigned instead of size_t. |
491 | using ListIndex = unsigned; |
492 | |
493 | /// Represents a span of contiguous MemoryDefs, potentially ending in a |
494 | /// MemoryPhi. |
495 | struct DefPath { |
496 | MemoryLocation Loc; |
497 | // Note that, because we always walk in reverse, Last will always dominate |
498 | // First. Also note that First and Last are inclusive. |
499 | MemoryAccess *First; |
500 | MemoryAccess *Last; |
501 | std::optional<ListIndex> Previous; |
502 | |
503 | DefPath(const MemoryLocation &Loc, MemoryAccess *First, MemoryAccess *Last, |
504 | std::optional<ListIndex> Previous) |
505 | : Loc(Loc), First(First), Last(Last), Previous(Previous) {} |
506 | |
507 | DefPath(const MemoryLocation &Loc, MemoryAccess *Init, |
508 | std::optional<ListIndex> Previous) |
509 | : DefPath(Loc, Init, Init, Previous) {} |
510 | }; |
511 | |
512 | const MemorySSA &MSSA; |
513 | DominatorTree &DT; |
514 | BatchAAResults *AA; |
515 | UpwardsMemoryQuery *Query; |
516 | unsigned *UpwardWalkLimit; |
517 | |
518 | // Phi optimization bookkeeping: |
519 | // List of DefPath to process during the current phi optimization walk. |
520 | SmallVector<DefPath, 32> Paths; |
521 | // List of visited <Access, Location> pairs; we can skip paths already |
522 | // visited with the same memory location. |
523 | DenseSet<ConstMemoryAccessPair> VisitedPhis; |
524 | |
525 | /// Find the nearest def or phi that `From` can legally be optimized to. |
526 | const MemoryAccess *getWalkTarget(const MemoryPhi *From) const { |
527 | assert(From->getNumOperands() && "Phi with no operands?" ); |
528 | |
529 | BasicBlock *BB = From->getBlock(); |
530 | MemoryAccess *Result = MSSA.getLiveOnEntryDef(); |
531 | DomTreeNode *Node = DT.getNode(BB); |
532 | while ((Node = Node->getIDom())) { |
533 | auto *Defs = MSSA.getBlockDefs(BB: Node->getBlock()); |
534 | if (Defs) |
535 | return &*Defs->rbegin(); |
536 | } |
537 | return Result; |
538 | } |
539 | |
540 | /// Result of calling walkToPhiOrClobber. |
541 | struct UpwardsWalkResult { |
542 | /// The "Result" of the walk. Either a clobber, the last thing we walked, or |
543 | /// both. Include alias info when clobber found. |
544 | MemoryAccess *Result; |
545 | bool IsKnownClobber; |
546 | }; |
547 | |
548 | /// Walk to the next Phi or Clobber in the def chain starting at Desc.Last. |
549 | /// This will update Desc.Last as it walks. It will (optionally) also stop at |
550 | /// StopAt. |
551 | /// |
552 | /// This does not test for whether StopAt is a clobber |
553 | UpwardsWalkResult |
554 | walkToPhiOrClobber(DefPath &Desc, const MemoryAccess *StopAt = nullptr, |
555 | const MemoryAccess *SkipStopAt = nullptr) const { |
556 | assert(!isa<MemoryUse>(Desc.Last) && "Uses don't exist in my world" ); |
557 | assert(UpwardWalkLimit && "Need a valid walk limit" ); |
558 | bool LimitAlreadyReached = false; |
559 | // (*UpwardWalkLimit) may be 0 here, due to the loop in tryOptimizePhi. Set |
560 | // it to 1. This will not do any alias() calls. It either returns in the |
561 | // first iteration in the loop below, or is set back to 0 if all def chains |
562 | // are free of MemoryDefs. |
563 | if (!*UpwardWalkLimit) { |
564 | *UpwardWalkLimit = 1; |
565 | LimitAlreadyReached = true; |
566 | } |
567 | |
568 | for (MemoryAccess *Current : def_chain(MA: Desc.Last)) { |
569 | Desc.Last = Current; |
570 | if (Current == StopAt || Current == SkipStopAt) |
571 | return {.Result: Current, .IsKnownClobber: false}; |
572 | |
573 | if (auto *MD = dyn_cast<MemoryDef>(Val: Current)) { |
574 | if (MSSA.isLiveOnEntryDef(MA: MD)) |
575 | return {.Result: MD, .IsKnownClobber: true}; |
576 | |
577 | if (!--*UpwardWalkLimit) |
578 | return {.Result: Current, .IsKnownClobber: true}; |
579 | |
580 | if (instructionClobbersQuery(MD, UseLoc: Desc.Loc, UseInst: Query->Inst, AA&: *AA)) |
581 | return {.Result: MD, .IsKnownClobber: true}; |
582 | } |
583 | } |
584 | |
585 | if (LimitAlreadyReached) |
586 | *UpwardWalkLimit = 0; |
587 | |
588 | assert(isa<MemoryPhi>(Desc.Last) && |
589 | "Ended at a non-clobber that's not a phi?" ); |
590 | return {.Result: Desc.Last, .IsKnownClobber: false}; |
591 | } |
592 | |
593 | void addSearches(MemoryPhi *Phi, SmallVectorImpl<ListIndex> &PausedSearches, |
594 | ListIndex PriorNode) { |
595 | auto UpwardDefsBegin = upward_defs_begin(Pair: {Phi, Paths[PriorNode].Loc}, DT); |
596 | auto UpwardDefs = make_range(x: UpwardDefsBegin, y: upward_defs_end()); |
597 | for (const MemoryAccessPair &P : UpwardDefs) { |
598 | PausedSearches.push_back(Elt: Paths.size()); |
599 | Paths.emplace_back(Args: P.second, Args: P.first, Args&: PriorNode); |
600 | } |
601 | } |
602 | |
603 | /// Represents a search that terminated after finding a clobber. This clobber |
604 | /// may or may not be present in the path of defs from LastNode..SearchStart, |
605 | /// since it may have been retrieved from cache. |
606 | struct TerminatedPath { |
607 | MemoryAccess *Clobber; |
608 | ListIndex LastNode; |
609 | }; |
610 | |
611 | /// Get an access that keeps us from optimizing to the given phi. |
612 | /// |
613 | /// PausedSearches is an array of indices into the Paths array. Its incoming |
614 | /// value is the indices of searches that stopped at the last phi optimization |
615 | /// target. It's left in an unspecified state. |
616 | /// |
617 | /// If this returns std::nullopt, NewPaused is a vector of searches that |
618 | /// terminated at StopWhere. Otherwise, NewPaused is left in an unspecified |
619 | /// state. |
620 | std::optional<TerminatedPath> |
621 | getBlockingAccess(const MemoryAccess *StopWhere, |
622 | SmallVectorImpl<ListIndex> &PausedSearches, |
623 | SmallVectorImpl<ListIndex> &NewPaused, |
624 | SmallVectorImpl<TerminatedPath> &Terminated) { |
625 | assert(!PausedSearches.empty() && "No searches to continue?" ); |
626 | |
627 | // BFS vs DFS really doesn't make a difference here, so just do a DFS with |
628 | // PausedSearches as our stack. |
629 | while (!PausedSearches.empty()) { |
630 | ListIndex PathIndex = PausedSearches.pop_back_val(); |
631 | DefPath &Node = Paths[PathIndex]; |
632 | |
633 | // If we've already visited this path with this MemoryLocation, we don't |
634 | // need to do so again. |
635 | // |
636 | // NOTE: That we just drop these paths on the ground makes caching |
637 | // behavior sporadic. e.g. given a diamond: |
638 | // A |
639 | // B C |
640 | // D |
641 | // |
642 | // ...If we walk D, B, A, C, we'll only cache the result of phi |
643 | // optimization for A, B, and D; C will be skipped because it dies here. |
644 | // This arguably isn't the worst thing ever, since: |
645 | // - We generally query things in a top-down order, so if we got below D |
646 | // without needing cache entries for {C, MemLoc}, then chances are |
647 | // that those cache entries would end up ultimately unused. |
648 | // - We still cache things for A, so C only needs to walk up a bit. |
649 | // If this behavior becomes problematic, we can fix without a ton of extra |
650 | // work. |
651 | if (!VisitedPhis.insert(V: {Node.Last, Node.Loc}).second) |
652 | continue; |
653 | |
654 | const MemoryAccess *SkipStopWhere = nullptr; |
655 | if (Query->SkipSelfAccess && Node.Loc == Query->StartingLoc) { |
656 | assert(isa<MemoryDef>(Query->OriginalAccess)); |
657 | SkipStopWhere = Query->OriginalAccess; |
658 | } |
659 | |
660 | UpwardsWalkResult Res = walkToPhiOrClobber(Desc&: Node, |
661 | /*StopAt=*/StopWhere, |
662 | /*SkipStopAt=*/SkipStopWhere); |
663 | if (Res.IsKnownClobber) { |
664 | assert(Res.Result != StopWhere && Res.Result != SkipStopWhere); |
665 | |
666 | // If this wasn't a cache hit, we hit a clobber when walking. That's a |
667 | // failure. |
668 | TerminatedPath Term{.Clobber: Res.Result, .LastNode: PathIndex}; |
669 | if (!MSSA.dominates(A: Res.Result, B: StopWhere)) |
670 | return Term; |
671 | |
672 | // Otherwise, it's a valid thing to potentially optimize to. |
673 | Terminated.push_back(Elt: Term); |
674 | continue; |
675 | } |
676 | |
677 | if (Res.Result == StopWhere || Res.Result == SkipStopWhere) { |
678 | // We've hit our target. Save this path off for if we want to continue |
679 | // walking. If we are in the mode of skipping the OriginalAccess, and |
680 | // we've reached back to the OriginalAccess, do not save path, we've |
681 | // just looped back to self. |
682 | if (Res.Result != SkipStopWhere) |
683 | NewPaused.push_back(Elt: PathIndex); |
684 | continue; |
685 | } |
686 | |
687 | assert(!MSSA.isLiveOnEntryDef(Res.Result) && "liveOnEntry is a clobber" ); |
688 | addSearches(Phi: cast<MemoryPhi>(Val: Res.Result), PausedSearches, PriorNode: PathIndex); |
689 | } |
690 | |
691 | return std::nullopt; |
692 | } |
693 | |
694 | template <typename T, typename Walker> |
695 | struct generic_def_path_iterator |
696 | : public iterator_facade_base<generic_def_path_iterator<T, Walker>, |
697 | std::forward_iterator_tag, T *> { |
698 | generic_def_path_iterator() = default; |
699 | generic_def_path_iterator(Walker *W, ListIndex N) : W(W), N(N) {} |
700 | |
701 | T &operator*() const { return curNode(); } |
702 | |
703 | generic_def_path_iterator &operator++() { |
704 | N = curNode().Previous; |
705 | return *this; |
706 | } |
707 | |
708 | bool operator==(const generic_def_path_iterator &O) const { |
709 | if (N.has_value() != O.N.has_value()) |
710 | return false; |
711 | return !N || *N == *O.N; |
712 | } |
713 | |
714 | private: |
715 | T &curNode() const { return W->Paths[*N]; } |
716 | |
717 | Walker *W = nullptr; |
718 | std::optional<ListIndex> N; |
719 | }; |
720 | |
721 | using def_path_iterator = generic_def_path_iterator<DefPath, ClobberWalker>; |
722 | using const_def_path_iterator = |
723 | generic_def_path_iterator<const DefPath, const ClobberWalker>; |
724 | |
725 | iterator_range<def_path_iterator> def_path(ListIndex From) { |
726 | return make_range(x: def_path_iterator(this, From), y: def_path_iterator()); |
727 | } |
728 | |
729 | iterator_range<const_def_path_iterator> const_def_path(ListIndex From) const { |
730 | return make_range(x: const_def_path_iterator(this, From), |
731 | y: const_def_path_iterator()); |
732 | } |
733 | |
734 | struct OptznResult { |
735 | /// The path that contains our result. |
736 | TerminatedPath PrimaryClobber; |
737 | /// The paths that we can legally cache back from, but that aren't |
738 | /// necessarily the result of the Phi optimization. |
739 | SmallVector<TerminatedPath, 4> OtherClobbers; |
740 | }; |
741 | |
742 | ListIndex defPathIndex(const DefPath &N) const { |
743 | // The assert looks nicer if we don't need to do &N |
744 | const DefPath *NP = &N; |
745 | assert(!Paths.empty() && NP >= &Paths.front() && NP <= &Paths.back() && |
746 | "Out of bounds DefPath!" ); |
747 | return NP - &Paths.front(); |
748 | } |
749 | |
750 | /// Try to optimize a phi as best as we can. Returns a SmallVector of Paths |
751 | /// that act as legal clobbers. Note that this won't return *all* clobbers. |
752 | /// |
753 | /// Phi optimization algorithm tl;dr: |
754 | /// - Find the earliest def/phi, A, we can optimize to |
755 | /// - Find if all paths from the starting memory access ultimately reach A |
756 | /// - If not, optimization isn't possible. |
757 | /// - Otherwise, walk from A to another clobber or phi, A'. |
758 | /// - If A' is a def, we're done. |
759 | /// - If A' is a phi, try to optimize it. |
760 | /// |
761 | /// A path is a series of {MemoryAccess, MemoryLocation} pairs. A path |
762 | /// terminates when a MemoryAccess that clobbers said MemoryLocation is found. |
763 | OptznResult tryOptimizePhi(MemoryPhi *Phi, MemoryAccess *Start, |
764 | const MemoryLocation &Loc) { |
765 | assert(Paths.empty() && VisitedPhis.empty() && |
766 | "Reset the optimization state." ); |
767 | |
768 | Paths.emplace_back(Args: Loc, Args&: Start, Args&: Phi, Args: std::nullopt); |
769 | // Stores how many "valid" optimization nodes we had prior to calling |
770 | // addSearches/getBlockingAccess. Necessary for caching if we had a blocker. |
771 | auto PriorPathsSize = Paths.size(); |
772 | |
773 | SmallVector<ListIndex, 16> PausedSearches; |
774 | SmallVector<ListIndex, 8> NewPaused; |
775 | SmallVector<TerminatedPath, 4> TerminatedPaths; |
776 | |
777 | addSearches(Phi, PausedSearches, PriorNode: 0); |
778 | |
779 | // Moves the TerminatedPath with the "most dominated" Clobber to the end of |
780 | // Paths. |
781 | auto MoveDominatedPathToEnd = [&](SmallVectorImpl<TerminatedPath> &Paths) { |
782 | assert(!Paths.empty() && "Need a path to move" ); |
783 | auto Dom = Paths.begin(); |
784 | for (auto I = std::next(x: Dom), E = Paths.end(); I != E; ++I) |
785 | if (!MSSA.dominates(A: I->Clobber, B: Dom->Clobber)) |
786 | Dom = I; |
787 | auto Last = Paths.end() - 1; |
788 | if (Last != Dom) |
789 | std::iter_swap(a: Last, b: Dom); |
790 | }; |
791 | |
792 | MemoryPhi *Current = Phi; |
793 | while (true) { |
794 | assert(!MSSA.isLiveOnEntryDef(Current) && |
795 | "liveOnEntry wasn't treated as a clobber?" ); |
796 | |
797 | const auto *Target = getWalkTarget(From: Current); |
798 | // If a TerminatedPath doesn't dominate Target, then it wasn't a legal |
799 | // optimization for the prior phi. |
800 | assert(all_of(TerminatedPaths, [&](const TerminatedPath &P) { |
801 | return MSSA.dominates(P.Clobber, Target); |
802 | })); |
803 | |
804 | // FIXME: This is broken, because the Blocker may be reported to be |
805 | // liveOnEntry, and we'll happily wait for that to disappear (read: never) |
806 | // For the moment, this is fine, since we do nothing with blocker info. |
807 | if (std::optional<TerminatedPath> Blocker = getBlockingAccess( |
808 | StopWhere: Target, PausedSearches, NewPaused, Terminated&: TerminatedPaths)) { |
809 | |
810 | // Find the node we started at. We can't search based on N->Last, since |
811 | // we may have gone around a loop with a different MemoryLocation. |
812 | auto Iter = find_if(Range: def_path(From: Blocker->LastNode), P: [&](const DefPath &N) { |
813 | return defPathIndex(N) < PriorPathsSize; |
814 | }); |
815 | assert(Iter != def_path_iterator()); |
816 | |
817 | DefPath &CurNode = *Iter; |
818 | assert(CurNode.Last == Current); |
819 | |
820 | // Two things: |
821 | // A. We can't reliably cache all of NewPaused back. Consider a case |
822 | // where we have two paths in NewPaused; one of which can't optimize |
823 | // above this phi, whereas the other can. If we cache the second path |
824 | // back, we'll end up with suboptimal cache entries. We can handle |
825 | // cases like this a bit better when we either try to find all |
826 | // clobbers that block phi optimization, or when our cache starts |
827 | // supporting unfinished searches. |
828 | // B. We can't reliably cache TerminatedPaths back here without doing |
829 | // extra checks; consider a case like: |
830 | // T |
831 | // / \ |
832 | // D C |
833 | // \ / |
834 | // S |
835 | // Where T is our target, C is a node with a clobber on it, D is a |
836 | // diamond (with a clobber *only* on the left or right node, N), and |
837 | // S is our start. Say we walk to D, through the node opposite N |
838 | // (read: ignoring the clobber), and see a cache entry in the top |
839 | // node of D. That cache entry gets put into TerminatedPaths. We then |
840 | // walk up to C (N is later in our worklist), find the clobber, and |
841 | // quit. If we append TerminatedPaths to OtherClobbers, we'll cache |
842 | // the bottom part of D to the cached clobber, ignoring the clobber |
843 | // in N. Again, this problem goes away if we start tracking all |
844 | // blockers for a given phi optimization. |
845 | TerminatedPath Result{.Clobber: CurNode.Last, .LastNode: defPathIndex(N: CurNode)}; |
846 | return {.PrimaryClobber: Result, .OtherClobbers: {}}; |
847 | } |
848 | |
849 | // If there's nothing left to search, then all paths led to valid clobbers |
850 | // that we got from our cache; pick the nearest to the start, and allow |
851 | // the rest to be cached back. |
852 | if (NewPaused.empty()) { |
853 | MoveDominatedPathToEnd(TerminatedPaths); |
854 | TerminatedPath Result = TerminatedPaths.pop_back_val(); |
855 | return {.PrimaryClobber: Result, .OtherClobbers: std::move(TerminatedPaths)}; |
856 | } |
857 | |
858 | MemoryAccess *DefChainEnd = nullptr; |
859 | SmallVector<TerminatedPath, 4> Clobbers; |
860 | for (ListIndex Paused : NewPaused) { |
861 | UpwardsWalkResult WR = walkToPhiOrClobber(Desc&: Paths[Paused]); |
862 | if (WR.IsKnownClobber) |
863 | Clobbers.push_back(Elt: {.Clobber: WR.Result, .LastNode: Paused}); |
864 | else |
865 | // Micro-opt: If we hit the end of the chain, save it. |
866 | DefChainEnd = WR.Result; |
867 | } |
868 | |
869 | if (!TerminatedPaths.empty()) { |
870 | // If we couldn't find the dominating phi/liveOnEntry in the above loop, |
871 | // do it now. |
872 | if (!DefChainEnd) |
873 | for (auto *MA : def_chain(MA: const_cast<MemoryAccess *>(Target))) |
874 | DefChainEnd = MA; |
875 | assert(DefChainEnd && "Failed to find dominating phi/liveOnEntry" ); |
876 | |
877 | // If any of the terminated paths don't dominate the phi we'll try to |
878 | // optimize, we need to figure out what they are and quit. |
879 | const BasicBlock *ChainBB = DefChainEnd->getBlock(); |
880 | for (const TerminatedPath &TP : TerminatedPaths) { |
881 | // Because we know that DefChainEnd is as "high" as we can go, we |
882 | // don't need local dominance checks; BB dominance is sufficient. |
883 | if (DT.dominates(A: ChainBB, B: TP.Clobber->getBlock())) |
884 | Clobbers.push_back(Elt: TP); |
885 | } |
886 | } |
887 | |
888 | // If we have clobbers in the def chain, find the one closest to Current |
889 | // and quit. |
890 | if (!Clobbers.empty()) { |
891 | MoveDominatedPathToEnd(Clobbers); |
892 | TerminatedPath Result = Clobbers.pop_back_val(); |
893 | return {.PrimaryClobber: Result, .OtherClobbers: std::move(Clobbers)}; |
894 | } |
895 | |
896 | assert(all_of(NewPaused, |
897 | [&](ListIndex I) { return Paths[I].Last == DefChainEnd; })); |
898 | |
899 | // Because liveOnEntry is a clobber, this must be a phi. |
900 | auto *DefChainPhi = cast<MemoryPhi>(Val: DefChainEnd); |
901 | |
902 | PriorPathsSize = Paths.size(); |
903 | PausedSearches.clear(); |
904 | for (ListIndex I : NewPaused) |
905 | addSearches(Phi: DefChainPhi, PausedSearches, PriorNode: I); |
906 | NewPaused.clear(); |
907 | |
908 | Current = DefChainPhi; |
909 | } |
910 | } |
911 | |
912 | void verifyOptResult(const OptznResult &R) const { |
913 | assert(all_of(R.OtherClobbers, [&](const TerminatedPath &P) { |
914 | return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber); |
915 | })); |
916 | } |
917 | |
918 | void resetPhiOptznState() { |
919 | Paths.clear(); |
920 | VisitedPhis.clear(); |
921 | } |
922 | |
923 | public: |
924 | ClobberWalker(const MemorySSA &MSSA, DominatorTree &DT) |
925 | : MSSA(MSSA), DT(DT) {} |
926 | |
927 | /// Finds the nearest clobber for the given query, optimizing phis if |
928 | /// possible. |
929 | MemoryAccess *findClobber(BatchAAResults &BAA, MemoryAccess *Start, |
930 | UpwardsMemoryQuery &Q, unsigned &UpWalkLimit) { |
931 | AA = &BAA; |
932 | Query = &Q; |
933 | UpwardWalkLimit = &UpWalkLimit; |
934 | // Starting limit must be > 0. |
935 | if (!UpWalkLimit) |
936 | UpWalkLimit++; |
937 | |
938 | MemoryAccess *Current = Start; |
939 | // This walker pretends uses don't exist. If we're handed one, silently grab |
940 | // its def. (This has the nice side-effect of ensuring we never cache uses) |
941 | if (auto *MU = dyn_cast<MemoryUse>(Val: Start)) |
942 | Current = MU->getDefiningAccess(); |
943 | |
944 | DefPath FirstDesc(Q.StartingLoc, Current, Current, std::nullopt); |
945 | // Fast path for the overly-common case (no crazy phi optimization |
946 | // necessary) |
947 | UpwardsWalkResult WalkResult = walkToPhiOrClobber(Desc&: FirstDesc); |
948 | MemoryAccess *Result; |
949 | if (WalkResult.IsKnownClobber) { |
950 | Result = WalkResult.Result; |
951 | } else { |
952 | OptznResult OptRes = tryOptimizePhi(Phi: cast<MemoryPhi>(Val: FirstDesc.Last), |
953 | Start: Current, Loc: Q.StartingLoc); |
954 | verifyOptResult(R: OptRes); |
955 | resetPhiOptznState(); |
956 | Result = OptRes.PrimaryClobber.Clobber; |
957 | } |
958 | |
959 | #ifdef EXPENSIVE_CHECKS |
960 | if (!Q.SkipSelfAccess && *UpwardWalkLimit > 0) |
961 | checkClobberSanity(Current, Result, Q.StartingLoc, MSSA, Q, BAA); |
962 | #endif |
963 | return Result; |
964 | } |
965 | }; |
966 | |
967 | struct RenamePassData { |
968 | DomTreeNode *DTN; |
969 | DomTreeNode::const_iterator ChildIt; |
970 | MemoryAccess *IncomingVal; |
971 | |
972 | RenamePassData(DomTreeNode *D, DomTreeNode::const_iterator It, |
973 | MemoryAccess *M) |
974 | : DTN(D), ChildIt(It), IncomingVal(M) {} |
975 | |
976 | void swap(RenamePassData &RHS) { |
977 | std::swap(a&: DTN, b&: RHS.DTN); |
978 | std::swap(a&: ChildIt, b&: RHS.ChildIt); |
979 | std::swap(a&: IncomingVal, b&: RHS.IncomingVal); |
980 | } |
981 | }; |
982 | |
983 | } // end anonymous namespace |
984 | |
985 | namespace llvm { |
986 | |
987 | class MemorySSA::ClobberWalkerBase { |
988 | ClobberWalker Walker; |
989 | MemorySSA *MSSA; |
990 | |
991 | public: |
992 | ClobberWalkerBase(MemorySSA *M, DominatorTree *D) : Walker(*M, *D), MSSA(M) {} |
993 | |
994 | MemoryAccess *getClobberingMemoryAccessBase(MemoryAccess *, |
995 | const MemoryLocation &, |
996 | BatchAAResults &, unsigned &); |
997 | // Third argument (bool), defines whether the clobber search should skip the |
998 | // original queried access. If true, there will be a follow-up query searching |
999 | // for a clobber access past "self". Note that the Optimized access is not |
1000 | // updated if a new clobber is found by this SkipSelf search. If this |
1001 | // additional query becomes heavily used we may decide to cache the result. |
1002 | // Walker instantiations will decide how to set the SkipSelf bool. |
1003 | MemoryAccess *getClobberingMemoryAccessBase(MemoryAccess *, BatchAAResults &, |
1004 | unsigned &, bool, |
1005 | bool UseInvariantGroup = true); |
1006 | }; |
1007 | |
1008 | /// A MemorySSAWalker that does AA walks to disambiguate accesses. It no |
1009 | /// longer does caching on its own, but the name has been retained for the |
1010 | /// moment. |
1011 | class MemorySSA::CachingWalker final : public MemorySSAWalker { |
1012 | ClobberWalkerBase *Walker; |
1013 | |
1014 | public: |
1015 | CachingWalker(MemorySSA *M, ClobberWalkerBase *W) |
1016 | : MemorySSAWalker(M), Walker(W) {} |
1017 | ~CachingWalker() override = default; |
1018 | |
1019 | using MemorySSAWalker::getClobberingMemoryAccess; |
1020 | |
1021 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, BatchAAResults &BAA, |
1022 | unsigned &UWL) { |
1023 | return Walker->getClobberingMemoryAccessBase(MA, BAA, UWL, false); |
1024 | } |
1025 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, |
1026 | const MemoryLocation &Loc, |
1027 | BatchAAResults &BAA, unsigned &UWL) { |
1028 | return Walker->getClobberingMemoryAccessBase(MA, Loc, BAA, UWL); |
1029 | } |
1030 | // This method is not accessible outside of this file. |
1031 | MemoryAccess *getClobberingMemoryAccessWithoutInvariantGroup( |
1032 | MemoryAccess *MA, BatchAAResults &BAA, unsigned &UWL) { |
1033 | return Walker->getClobberingMemoryAccessBase(MA, BAA, UWL, false, UseInvariantGroup: false); |
1034 | } |
1035 | |
1036 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, |
1037 | BatchAAResults &BAA) override { |
1038 | unsigned UpwardWalkLimit = MaxCheckLimit; |
1039 | return getClobberingMemoryAccess(MA, BAA, UWL&: UpwardWalkLimit); |
1040 | } |
1041 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, |
1042 | const MemoryLocation &Loc, |
1043 | BatchAAResults &BAA) override { |
1044 | unsigned UpwardWalkLimit = MaxCheckLimit; |
1045 | return getClobberingMemoryAccess(MA, Loc, BAA, UWL&: UpwardWalkLimit); |
1046 | } |
1047 | |
1048 | void invalidateInfo(MemoryAccess *MA) override { |
1049 | if (auto *MUD = dyn_cast<MemoryUseOrDef>(Val: MA)) |
1050 | MUD->resetOptimized(); |
1051 | } |
1052 | }; |
1053 | |
1054 | class MemorySSA::SkipSelfWalker final : public MemorySSAWalker { |
1055 | ClobberWalkerBase *Walker; |
1056 | |
1057 | public: |
1058 | SkipSelfWalker(MemorySSA *M, ClobberWalkerBase *W) |
1059 | : MemorySSAWalker(M), Walker(W) {} |
1060 | ~SkipSelfWalker() override = default; |
1061 | |
1062 | using MemorySSAWalker::getClobberingMemoryAccess; |
1063 | |
1064 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, BatchAAResults &BAA, |
1065 | unsigned &UWL) { |
1066 | return Walker->getClobberingMemoryAccessBase(MA, BAA, UWL, true); |
1067 | } |
1068 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, |
1069 | const MemoryLocation &Loc, |
1070 | BatchAAResults &BAA, unsigned &UWL) { |
1071 | return Walker->getClobberingMemoryAccessBase(MA, Loc, BAA, UWL); |
1072 | } |
1073 | |
1074 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, |
1075 | BatchAAResults &BAA) override { |
1076 | unsigned UpwardWalkLimit = MaxCheckLimit; |
1077 | return getClobberingMemoryAccess(MA, BAA, UWL&: UpwardWalkLimit); |
1078 | } |
1079 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, |
1080 | const MemoryLocation &Loc, |
1081 | BatchAAResults &BAA) override { |
1082 | unsigned UpwardWalkLimit = MaxCheckLimit; |
1083 | return getClobberingMemoryAccess(MA, Loc, BAA, UWL&: UpwardWalkLimit); |
1084 | } |
1085 | |
1086 | void invalidateInfo(MemoryAccess *MA) override { |
1087 | if (auto *MUD = dyn_cast<MemoryUseOrDef>(Val: MA)) |
1088 | MUD->resetOptimized(); |
1089 | } |
1090 | }; |
1091 | |
1092 | } // end namespace llvm |
1093 | |
1094 | void MemorySSA::renameSuccessorPhis(BasicBlock *BB, MemoryAccess *IncomingVal, |
1095 | bool RenameAllUses) { |
1096 | // Pass through values to our successors |
1097 | for (const BasicBlock *S : successors(BB)) { |
1098 | auto It = PerBlockAccesses.find(Val: S); |
1099 | // Rename the phi nodes in our successor block |
1100 | if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(Val: It->second->front())) |
1101 | continue; |
1102 | AccessList *Accesses = It->second.get(); |
1103 | auto *Phi = cast<MemoryPhi>(Val: &Accesses->front()); |
1104 | if (RenameAllUses) { |
1105 | bool ReplacementDone = false; |
1106 | for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) |
1107 | if (Phi->getIncomingBlock(I) == BB) { |
1108 | Phi->setIncomingValue(I, V: IncomingVal); |
1109 | ReplacementDone = true; |
1110 | } |
1111 | (void) ReplacementDone; |
1112 | assert(ReplacementDone && "Incomplete phi during partial rename" ); |
1113 | } else |
1114 | Phi->addIncoming(V: IncomingVal, BB); |
1115 | } |
1116 | } |
1117 | |
1118 | /// Rename a single basic block into MemorySSA form. |
1119 | /// Uses the standard SSA renaming algorithm. |
1120 | /// \returns The new incoming value. |
1121 | MemoryAccess *MemorySSA::renameBlock(BasicBlock *BB, MemoryAccess *IncomingVal, |
1122 | bool RenameAllUses) { |
1123 | auto It = PerBlockAccesses.find(Val: BB); |
1124 | // Skip most processing if the list is empty. |
1125 | if (It != PerBlockAccesses.end()) { |
1126 | AccessList *Accesses = It->second.get(); |
1127 | for (MemoryAccess &L : *Accesses) { |
1128 | if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(Val: &L)) { |
1129 | if (MUD->getDefiningAccess() == nullptr || RenameAllUses) |
1130 | MUD->setDefiningAccess(DMA: IncomingVal); |
1131 | if (isa<MemoryDef>(Val: &L)) |
1132 | IncomingVal = &L; |
1133 | } else { |
1134 | IncomingVal = &L; |
1135 | } |
1136 | } |
1137 | } |
1138 | return IncomingVal; |
1139 | } |
1140 | |
1141 | /// This is the standard SSA renaming algorithm. |
1142 | /// |
1143 | /// We walk the dominator tree in preorder, renaming accesses, and then filling |
1144 | /// in phi nodes in our successors. |
1145 | void MemorySSA::renamePass(DomTreeNode *Root, MemoryAccess *IncomingVal, |
1146 | SmallPtrSetImpl<BasicBlock *> &Visited, |
1147 | bool SkipVisited, bool RenameAllUses) { |
1148 | assert(Root && "Trying to rename accesses in an unreachable block" ); |
1149 | |
1150 | SmallVector<RenamePassData, 32> WorkStack; |
1151 | // Skip everything if we already renamed this block and we are skipping. |
1152 | // Note: You can't sink this into the if, because we need it to occur |
1153 | // regardless of whether we skip blocks or not. |
1154 | bool AlreadyVisited = !Visited.insert(Ptr: Root->getBlock()).second; |
1155 | if (SkipVisited && AlreadyVisited) |
1156 | return; |
1157 | |
1158 | IncomingVal = renameBlock(BB: Root->getBlock(), IncomingVal, RenameAllUses); |
1159 | renameSuccessorPhis(BB: Root->getBlock(), IncomingVal, RenameAllUses); |
1160 | WorkStack.push_back(Elt: {Root, Root->begin(), IncomingVal}); |
1161 | |
1162 | while (!WorkStack.empty()) { |
1163 | DomTreeNode *Node = WorkStack.back().DTN; |
1164 | DomTreeNode::const_iterator ChildIt = WorkStack.back().ChildIt; |
1165 | IncomingVal = WorkStack.back().IncomingVal; |
1166 | |
1167 | if (ChildIt == Node->end()) { |
1168 | WorkStack.pop_back(); |
1169 | } else { |
1170 | DomTreeNode *Child = *ChildIt; |
1171 | ++WorkStack.back().ChildIt; |
1172 | BasicBlock *BB = Child->getBlock(); |
1173 | // Note: You can't sink this into the if, because we need it to occur |
1174 | // regardless of whether we skip blocks or not. |
1175 | AlreadyVisited = !Visited.insert(Ptr: BB).second; |
1176 | if (SkipVisited && AlreadyVisited) { |
1177 | // We already visited this during our renaming, which can happen when |
1178 | // being asked to rename multiple blocks. Figure out the incoming val, |
1179 | // which is the last def. |
1180 | // Incoming value can only change if there is a block def, and in that |
1181 | // case, it's the last block def in the list. |
1182 | if (auto *BlockDefs = getWritableBlockDefs(BB)) |
1183 | IncomingVal = &*BlockDefs->rbegin(); |
1184 | } else |
1185 | IncomingVal = renameBlock(BB, IncomingVal, RenameAllUses); |
1186 | renameSuccessorPhis(BB, IncomingVal, RenameAllUses); |
1187 | WorkStack.push_back(Elt: {Child, Child->begin(), IncomingVal}); |
1188 | } |
1189 | } |
1190 | } |
1191 | |
1192 | /// This handles unreachable block accesses by deleting phi nodes in |
1193 | /// unreachable blocks, and marking all other unreachable MemoryAccess's as |
1194 | /// being uses of the live on entry definition. |
1195 | void MemorySSA::markUnreachableAsLiveOnEntry(BasicBlock *BB) { |
1196 | assert(!DT->isReachableFromEntry(BB) && |
1197 | "Reachable block found while handling unreachable blocks" ); |
1198 | |
1199 | // Make sure phi nodes in our reachable successors end up with a |
1200 | // LiveOnEntryDef for our incoming edge, even though our block is forward |
1201 | // unreachable. We could just disconnect these blocks from the CFG fully, |
1202 | // but we do not right now. |
1203 | for (const BasicBlock *S : successors(BB)) { |
1204 | if (!DT->isReachableFromEntry(A: S)) |
1205 | continue; |
1206 | auto It = PerBlockAccesses.find(Val: S); |
1207 | // Rename the phi nodes in our successor block |
1208 | if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(Val: It->second->front())) |
1209 | continue; |
1210 | AccessList *Accesses = It->second.get(); |
1211 | auto *Phi = cast<MemoryPhi>(Val: &Accesses->front()); |
1212 | Phi->addIncoming(V: LiveOnEntryDef.get(), BB); |
1213 | } |
1214 | |
1215 | auto It = PerBlockAccesses.find(Val: BB); |
1216 | if (It == PerBlockAccesses.end()) |
1217 | return; |
1218 | |
1219 | auto &Accesses = It->second; |
1220 | for (auto AI = Accesses->begin(), AE = Accesses->end(); AI != AE;) { |
1221 | auto Next = std::next(x: AI); |
1222 | // If we have a phi, just remove it. We are going to replace all |
1223 | // users with live on entry. |
1224 | if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(Val&: AI)) |
1225 | UseOrDef->setDefiningAccess(DMA: LiveOnEntryDef.get()); |
1226 | else |
1227 | Accesses->erase(where: AI); |
1228 | AI = Next; |
1229 | } |
1230 | } |
1231 | |
1232 | MemorySSA::MemorySSA(Function &Func, AliasAnalysis *AA, DominatorTree *DT) |
1233 | : DT(DT), F(Func), LiveOnEntryDef(nullptr), Walker(nullptr), |
1234 | SkipWalker(nullptr) { |
1235 | // Build MemorySSA using a batch alias analysis. This reuses the internal |
1236 | // state that AA collects during an alias()/getModRefInfo() call. This is |
1237 | // safe because there are no CFG changes while building MemorySSA and can |
1238 | // significantly reduce the time spent by the compiler in AA, because we will |
1239 | // make queries about all the instructions in the Function. |
1240 | assert(AA && "No alias analysis?" ); |
1241 | BatchAAResults BatchAA(*AA); |
1242 | buildMemorySSA(BAA&: BatchAA); |
1243 | // Intentionally leave AA to nullptr while building so we don't accidently |
1244 | // use non-batch AliasAnalysis. |
1245 | this->AA = AA; |
1246 | // Also create the walker here. |
1247 | getWalker(); |
1248 | } |
1249 | |
1250 | MemorySSA::~MemorySSA() { |
1251 | // Drop all our references |
1252 | for (const auto &Pair : PerBlockAccesses) |
1253 | for (MemoryAccess &MA : *Pair.second) |
1254 | MA.dropAllReferences(); |
1255 | } |
1256 | |
1257 | MemorySSA::AccessList *MemorySSA::getOrCreateAccessList(const BasicBlock *BB) { |
1258 | auto Res = PerBlockAccesses.insert(KV: std::make_pair(x&: BB, y: nullptr)); |
1259 | |
1260 | if (Res.second) |
1261 | Res.first->second = std::make_unique<AccessList>(); |
1262 | return Res.first->second.get(); |
1263 | } |
1264 | |
1265 | MemorySSA::DefsList *MemorySSA::getOrCreateDefsList(const BasicBlock *BB) { |
1266 | auto Res = PerBlockDefs.insert(KV: std::make_pair(x&: BB, y: nullptr)); |
1267 | |
1268 | if (Res.second) |
1269 | Res.first->second = std::make_unique<DefsList>(); |
1270 | return Res.first->second.get(); |
1271 | } |
1272 | |
1273 | namespace llvm { |
1274 | |
1275 | /// This class is a batch walker of all MemoryUse's in the program, and points |
1276 | /// their defining access at the thing that actually clobbers them. Because it |
1277 | /// is a batch walker that touches everything, it does not operate like the |
1278 | /// other walkers. This walker is basically performing a top-down SSA renaming |
1279 | /// pass, where the version stack is used as the cache. This enables it to be |
1280 | /// significantly more time and memory efficient than using the regular walker, |
1281 | /// which is walking bottom-up. |
1282 | class MemorySSA::OptimizeUses { |
1283 | public: |
1284 | OptimizeUses(MemorySSA *MSSA, CachingWalker *Walker, BatchAAResults *BAA, |
1285 | DominatorTree *DT) |
1286 | : MSSA(MSSA), Walker(Walker), AA(BAA), DT(DT) {} |
1287 | |
1288 | void optimizeUses(); |
1289 | |
1290 | private: |
1291 | /// This represents where a given memorylocation is in the stack. |
1292 | struct MemlocStackInfo { |
1293 | // This essentially is keeping track of versions of the stack. Whenever |
1294 | // the stack changes due to pushes or pops, these versions increase. |
1295 | unsigned long StackEpoch; |
1296 | unsigned long PopEpoch; |
1297 | // This is the lower bound of places on the stack to check. It is equal to |
1298 | // the place the last stack walk ended. |
1299 | // Note: Correctness depends on this being initialized to 0, which densemap |
1300 | // does |
1301 | unsigned long LowerBound; |
1302 | const BasicBlock *LowerBoundBlock; |
1303 | // This is where the last walk for this memory location ended. |
1304 | unsigned long LastKill; |
1305 | bool LastKillValid; |
1306 | }; |
1307 | |
1308 | void optimizeUsesInBlock(const BasicBlock *, unsigned long &, unsigned long &, |
1309 | SmallVectorImpl<MemoryAccess *> &, |
1310 | DenseMap<MemoryLocOrCall, MemlocStackInfo> &); |
1311 | |
1312 | MemorySSA *MSSA; |
1313 | CachingWalker *Walker; |
1314 | BatchAAResults *AA; |
1315 | DominatorTree *DT; |
1316 | }; |
1317 | |
1318 | } // end namespace llvm |
1319 | |
1320 | /// Optimize the uses in a given block This is basically the SSA renaming |
1321 | /// algorithm, with one caveat: We are able to use a single stack for all |
1322 | /// MemoryUses. This is because the set of *possible* reaching MemoryDefs is |
1323 | /// the same for every MemoryUse. The *actual* clobbering MemoryDef is just |
1324 | /// going to be some position in that stack of possible ones. |
1325 | /// |
1326 | /// We track the stack positions that each MemoryLocation needs |
1327 | /// to check, and last ended at. This is because we only want to check the |
1328 | /// things that changed since last time. The same MemoryLocation should |
1329 | /// get clobbered by the same store (getModRefInfo does not use invariantness or |
1330 | /// things like this, and if they start, we can modify MemoryLocOrCall to |
1331 | /// include relevant data) |
1332 | void MemorySSA::OptimizeUses::optimizeUsesInBlock( |
1333 | const BasicBlock *BB, unsigned long &StackEpoch, unsigned long &PopEpoch, |
1334 | SmallVectorImpl<MemoryAccess *> &VersionStack, |
1335 | DenseMap<MemoryLocOrCall, MemlocStackInfo> &LocStackInfo) { |
1336 | |
1337 | /// If no accesses, nothing to do. |
1338 | MemorySSA::AccessList *Accesses = MSSA->getWritableBlockAccesses(BB); |
1339 | if (Accesses == nullptr) |
1340 | return; |
1341 | |
1342 | // Pop everything that doesn't dominate the current block off the stack, |
1343 | // increment the PopEpoch to account for this. |
1344 | while (true) { |
1345 | assert( |
1346 | !VersionStack.empty() && |
1347 | "Version stack should have liveOnEntry sentinel dominating everything" ); |
1348 | BasicBlock *BackBlock = VersionStack.back()->getBlock(); |
1349 | if (DT->dominates(A: BackBlock, B: BB)) |
1350 | break; |
1351 | while (VersionStack.back()->getBlock() == BackBlock) |
1352 | VersionStack.pop_back(); |
1353 | ++PopEpoch; |
1354 | } |
1355 | |
1356 | for (MemoryAccess &MA : *Accesses) { |
1357 | auto *MU = dyn_cast<MemoryUse>(Val: &MA); |
1358 | if (!MU) { |
1359 | VersionStack.push_back(Elt: &MA); |
1360 | ++StackEpoch; |
1361 | continue; |
1362 | } |
1363 | |
1364 | if (MU->isOptimized()) |
1365 | continue; |
1366 | |
1367 | MemoryLocOrCall UseMLOC(MU); |
1368 | auto &LocInfo = LocStackInfo[UseMLOC]; |
1369 | // If the pop epoch changed, it means we've removed stuff from top of |
1370 | // stack due to changing blocks. We may have to reset the lower bound or |
1371 | // last kill info. |
1372 | if (LocInfo.PopEpoch != PopEpoch) { |
1373 | LocInfo.PopEpoch = PopEpoch; |
1374 | LocInfo.StackEpoch = StackEpoch; |
1375 | // If the lower bound was in something that no longer dominates us, we |
1376 | // have to reset it. |
1377 | // We can't simply track stack size, because the stack may have had |
1378 | // pushes/pops in the meantime. |
1379 | // XXX: This is non-optimal, but only is slower cases with heavily |
1380 | // branching dominator trees. To get the optimal number of queries would |
1381 | // be to make lowerbound and lastkill a per-loc stack, and pop it until |
1382 | // the top of that stack dominates us. This does not seem worth it ATM. |
1383 | // A much cheaper optimization would be to always explore the deepest |
1384 | // branch of the dominator tree first. This will guarantee this resets on |
1385 | // the smallest set of blocks. |
1386 | if (LocInfo.LowerBoundBlock && LocInfo.LowerBoundBlock != BB && |
1387 | !DT->dominates(A: LocInfo.LowerBoundBlock, B: BB)) { |
1388 | // Reset the lower bound of things to check. |
1389 | // TODO: Some day we should be able to reset to last kill, rather than |
1390 | // 0. |
1391 | LocInfo.LowerBound = 0; |
1392 | LocInfo.LowerBoundBlock = VersionStack[0]->getBlock(); |
1393 | LocInfo.LastKillValid = false; |
1394 | } |
1395 | } else if (LocInfo.StackEpoch != StackEpoch) { |
1396 | // If all that has changed is the StackEpoch, we only have to check the |
1397 | // new things on the stack, because we've checked everything before. In |
1398 | // this case, the lower bound of things to check remains the same. |
1399 | LocInfo.PopEpoch = PopEpoch; |
1400 | LocInfo.StackEpoch = StackEpoch; |
1401 | } |
1402 | if (!LocInfo.LastKillValid) { |
1403 | LocInfo.LastKill = VersionStack.size() - 1; |
1404 | LocInfo.LastKillValid = true; |
1405 | } |
1406 | |
1407 | // At this point, we should have corrected last kill and LowerBound to be |
1408 | // in bounds. |
1409 | assert(LocInfo.LowerBound < VersionStack.size() && |
1410 | "Lower bound out of range" ); |
1411 | assert(LocInfo.LastKill < VersionStack.size() && |
1412 | "Last kill info out of range" ); |
1413 | // In any case, the new upper bound is the top of the stack. |
1414 | unsigned long UpperBound = VersionStack.size() - 1; |
1415 | |
1416 | if (UpperBound - LocInfo.LowerBound > MaxCheckLimit) { |
1417 | LLVM_DEBUG(dbgs() << "MemorySSA skipping optimization of " << *MU << " (" |
1418 | << *(MU->getMemoryInst()) << ")" |
1419 | << " because there are " |
1420 | << UpperBound - LocInfo.LowerBound |
1421 | << " stores to disambiguate\n" ); |
1422 | // Because we did not walk, LastKill is no longer valid, as this may |
1423 | // have been a kill. |
1424 | LocInfo.LastKillValid = false; |
1425 | continue; |
1426 | } |
1427 | bool FoundClobberResult = false; |
1428 | unsigned UpwardWalkLimit = MaxCheckLimit; |
1429 | while (UpperBound > LocInfo.LowerBound) { |
1430 | if (isa<MemoryPhi>(Val: VersionStack[UpperBound])) { |
1431 | // For phis, use the walker, see where we ended up, go there. |
1432 | // The invariant.group handling in MemorySSA is ad-hoc and doesn't |
1433 | // support updates, so don't use it to optimize uses. |
1434 | MemoryAccess *Result = |
1435 | Walker->getClobberingMemoryAccessWithoutInvariantGroup( |
1436 | MA: MU, BAA&: *AA, UWL&: UpwardWalkLimit); |
1437 | // We are guaranteed to find it or something is wrong. |
1438 | while (VersionStack[UpperBound] != Result) { |
1439 | assert(UpperBound != 0); |
1440 | --UpperBound; |
1441 | } |
1442 | FoundClobberResult = true; |
1443 | break; |
1444 | } |
1445 | |
1446 | MemoryDef *MD = cast<MemoryDef>(Val: VersionStack[UpperBound]); |
1447 | if (instructionClobbersQuery(MD, MU, UseMLOC, AA&: *AA)) { |
1448 | FoundClobberResult = true; |
1449 | break; |
1450 | } |
1451 | --UpperBound; |
1452 | } |
1453 | |
1454 | // At the end of this loop, UpperBound is either a clobber, or lower bound |
1455 | // PHI walking may cause it to be < LowerBound, and in fact, < LastKill. |
1456 | if (FoundClobberResult || UpperBound < LocInfo.LastKill) { |
1457 | MU->setDefiningAccess(DMA: VersionStack[UpperBound], Optimized: true); |
1458 | LocInfo.LastKill = UpperBound; |
1459 | } else { |
1460 | // Otherwise, we checked all the new ones, and now we know we can get to |
1461 | // LastKill. |
1462 | MU->setDefiningAccess(DMA: VersionStack[LocInfo.LastKill], Optimized: true); |
1463 | } |
1464 | LocInfo.LowerBound = VersionStack.size() - 1; |
1465 | LocInfo.LowerBoundBlock = BB; |
1466 | } |
1467 | } |
1468 | |
1469 | /// Optimize uses to point to their actual clobbering definitions. |
1470 | void MemorySSA::OptimizeUses::optimizeUses() { |
1471 | SmallVector<MemoryAccess *, 16> VersionStack; |
1472 | DenseMap<MemoryLocOrCall, MemlocStackInfo> LocStackInfo; |
1473 | VersionStack.push_back(Elt: MSSA->getLiveOnEntryDef()); |
1474 | |
1475 | unsigned long StackEpoch = 1; |
1476 | unsigned long PopEpoch = 1; |
1477 | // We perform a non-recursive top-down dominator tree walk. |
1478 | for (const auto *DomNode : depth_first(G: DT->getRootNode())) |
1479 | optimizeUsesInBlock(BB: DomNode->getBlock(), StackEpoch, PopEpoch, VersionStack, |
1480 | LocStackInfo); |
1481 | } |
1482 | |
1483 | void MemorySSA::placePHINodes( |
1484 | const SmallPtrSetImpl<BasicBlock *> &DefiningBlocks) { |
1485 | // Determine where our MemoryPhi's should go |
1486 | ForwardIDFCalculator IDFs(*DT); |
1487 | IDFs.setDefiningBlocks(DefiningBlocks); |
1488 | SmallVector<BasicBlock *, 32> IDFBlocks; |
1489 | IDFs.calculate(IDFBlocks); |
1490 | |
1491 | // Now place MemoryPhi nodes. |
1492 | for (auto &BB : IDFBlocks) |
1493 | createMemoryPhi(BB); |
1494 | } |
1495 | |
1496 | void MemorySSA::buildMemorySSA(BatchAAResults &BAA) { |
1497 | // We create an access to represent "live on entry", for things like |
1498 | // arguments or users of globals, where the memory they use is defined before |
1499 | // the beginning of the function. We do not actually insert it into the IR. |
1500 | // We do not define a live on exit for the immediate uses, and thus our |
1501 | // semantics do *not* imply that something with no immediate uses can simply |
1502 | // be removed. |
1503 | BasicBlock &StartingPoint = F.getEntryBlock(); |
1504 | LiveOnEntryDef.reset(p: new MemoryDef(F.getContext(), nullptr, nullptr, |
1505 | &StartingPoint, NextID++)); |
1506 | |
1507 | // We maintain lists of memory accesses per-block, trading memory for time. We |
1508 | // could just look up the memory access for every possible instruction in the |
1509 | // stream. |
1510 | SmallPtrSet<BasicBlock *, 32> DefiningBlocks; |
1511 | // Go through each block, figure out where defs occur, and chain together all |
1512 | // the accesses. |
1513 | for (BasicBlock &B : F) { |
1514 | bool InsertIntoDef = false; |
1515 | AccessList *Accesses = nullptr; |
1516 | DefsList *Defs = nullptr; |
1517 | for (Instruction &I : B) { |
1518 | MemoryUseOrDef *MUD = createNewAccess(I: &I, AAP: &BAA); |
1519 | if (!MUD) |
1520 | continue; |
1521 | |
1522 | if (!Accesses) |
1523 | Accesses = getOrCreateAccessList(BB: &B); |
1524 | Accesses->push_back(val: MUD); |
1525 | if (isa<MemoryDef>(Val: MUD)) { |
1526 | InsertIntoDef = true; |
1527 | if (!Defs) |
1528 | Defs = getOrCreateDefsList(BB: &B); |
1529 | Defs->push_back(Node&: *MUD); |
1530 | } |
1531 | } |
1532 | if (InsertIntoDef) |
1533 | DefiningBlocks.insert(Ptr: &B); |
1534 | } |
1535 | placePHINodes(DefiningBlocks); |
1536 | |
1537 | // Now do regular SSA renaming on the MemoryDef/MemoryUse. Visited will get |
1538 | // filled in with all blocks. |
1539 | SmallPtrSet<BasicBlock *, 16> Visited; |
1540 | renamePass(Root: DT->getRootNode(), IncomingVal: LiveOnEntryDef.get(), Visited); |
1541 | |
1542 | // Mark the uses in unreachable blocks as live on entry, so that they go |
1543 | // somewhere. |
1544 | for (auto &BB : F) |
1545 | if (!Visited.count(Ptr: &BB)) |
1546 | markUnreachableAsLiveOnEntry(BB: &BB); |
1547 | } |
1548 | |
1549 | MemorySSAWalker *MemorySSA::getWalker() { return getWalkerImpl(); } |
1550 | |
1551 | MemorySSA::CachingWalker *MemorySSA::getWalkerImpl() { |
1552 | if (Walker) |
1553 | return Walker.get(); |
1554 | |
1555 | if (!WalkerBase) |
1556 | WalkerBase = std::make_unique<ClobberWalkerBase>(args: this, args&: DT); |
1557 | |
1558 | Walker = std::make_unique<CachingWalker>(args: this, args: WalkerBase.get()); |
1559 | return Walker.get(); |
1560 | } |
1561 | |
1562 | MemorySSAWalker *MemorySSA::getSkipSelfWalker() { |
1563 | if (SkipWalker) |
1564 | return SkipWalker.get(); |
1565 | |
1566 | if (!WalkerBase) |
1567 | WalkerBase = std::make_unique<ClobberWalkerBase>(args: this, args&: DT); |
1568 | |
1569 | SkipWalker = std::make_unique<SkipSelfWalker>(args: this, args: WalkerBase.get()); |
1570 | return SkipWalker.get(); |
1571 | } |
1572 | |
1573 | |
1574 | // This is a helper function used by the creation routines. It places NewAccess |
1575 | // into the access and defs lists for a given basic block, at the given |
1576 | // insertion point. |
1577 | void MemorySSA::insertIntoListsForBlock(MemoryAccess *NewAccess, |
1578 | const BasicBlock *BB, |
1579 | InsertionPlace Point) { |
1580 | auto *Accesses = getOrCreateAccessList(BB); |
1581 | if (Point == Beginning) { |
1582 | // If it's a phi node, it goes first, otherwise, it goes after any phi |
1583 | // nodes. |
1584 | if (isa<MemoryPhi>(Val: NewAccess)) { |
1585 | Accesses->push_front(val: NewAccess); |
1586 | auto *Defs = getOrCreateDefsList(BB); |
1587 | Defs->push_front(Node&: *NewAccess); |
1588 | } else { |
1589 | auto AI = find_if_not( |
1590 | Range&: *Accesses, P: [](const MemoryAccess &MA) { return isa<MemoryPhi>(Val: MA); }); |
1591 | Accesses->insert(where: AI, New: NewAccess); |
1592 | if (!isa<MemoryUse>(Val: NewAccess)) { |
1593 | auto *Defs = getOrCreateDefsList(BB); |
1594 | auto DI = find_if_not( |
1595 | Range&: *Defs, P: [](const MemoryAccess &MA) { return isa<MemoryPhi>(Val: MA); }); |
1596 | Defs->insert(I: DI, Node&: *NewAccess); |
1597 | } |
1598 | } |
1599 | } else { |
1600 | Accesses->push_back(val: NewAccess); |
1601 | if (!isa<MemoryUse>(Val: NewAccess)) { |
1602 | auto *Defs = getOrCreateDefsList(BB); |
1603 | Defs->push_back(Node&: *NewAccess); |
1604 | } |
1605 | } |
1606 | BlockNumberingValid.erase(Ptr: BB); |
1607 | } |
1608 | |
1609 | void MemorySSA::insertIntoListsBefore(MemoryAccess *What, const BasicBlock *BB, |
1610 | AccessList::iterator InsertPt) { |
1611 | auto *Accesses = getWritableBlockAccesses(BB); |
1612 | bool WasEnd = InsertPt == Accesses->end(); |
1613 | Accesses->insert(where: AccessList::iterator(InsertPt), New: What); |
1614 | if (!isa<MemoryUse>(Val: What)) { |
1615 | auto *Defs = getOrCreateDefsList(BB); |
1616 | // If we got asked to insert at the end, we have an easy job, just shove it |
1617 | // at the end. If we got asked to insert before an existing def, we also get |
1618 | // an iterator. If we got asked to insert before a use, we have to hunt for |
1619 | // the next def. |
1620 | if (WasEnd) { |
1621 | Defs->push_back(Node&: *What); |
1622 | } else if (isa<MemoryDef>(Val: InsertPt)) { |
1623 | Defs->insert(I: InsertPt->getDefsIterator(), Node&: *What); |
1624 | } else { |
1625 | while (InsertPt != Accesses->end() && !isa<MemoryDef>(Val: InsertPt)) |
1626 | ++InsertPt; |
1627 | // Either we found a def, or we are inserting at the end |
1628 | if (InsertPt == Accesses->end()) |
1629 | Defs->push_back(Node&: *What); |
1630 | else |
1631 | Defs->insert(I: InsertPt->getDefsIterator(), Node&: *What); |
1632 | } |
1633 | } |
1634 | BlockNumberingValid.erase(Ptr: BB); |
1635 | } |
1636 | |
1637 | void MemorySSA::prepareForMoveTo(MemoryAccess *What, BasicBlock *BB) { |
1638 | // Keep it in the lookup tables, remove from the lists |
1639 | removeFromLists(What, ShouldDelete: false); |
1640 | |
1641 | // Note that moving should implicitly invalidate the optimized state of a |
1642 | // MemoryUse (and Phis can't be optimized). However, it doesn't do so for a |
1643 | // MemoryDef. |
1644 | if (auto *MD = dyn_cast<MemoryDef>(Val: What)) |
1645 | MD->resetOptimized(); |
1646 | What->setBlock(BB); |
1647 | } |
1648 | |
1649 | // Move What before Where in the IR. The end result is that What will belong to |
1650 | // the right lists and have the right Block set, but will not otherwise be |
1651 | // correct. It will not have the right defining access, and if it is a def, |
1652 | // things below it will not properly be updated. |
1653 | void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB, |
1654 | AccessList::iterator Where) { |
1655 | prepareForMoveTo(What, BB); |
1656 | insertIntoListsBefore(What, BB, InsertPt: Where); |
1657 | } |
1658 | |
1659 | void MemorySSA::moveTo(MemoryAccess *What, BasicBlock *BB, |
1660 | InsertionPlace Point) { |
1661 | if (isa<MemoryPhi>(Val: What)) { |
1662 | assert(Point == Beginning && |
1663 | "Can only move a Phi at the beginning of the block" ); |
1664 | // Update lookup table entry |
1665 | ValueToMemoryAccess.erase(Val: What->getBlock()); |
1666 | bool Inserted = ValueToMemoryAccess.insert(KV: {BB, What}).second; |
1667 | (void)Inserted; |
1668 | assert(Inserted && "Cannot move a Phi to a block that already has one" ); |
1669 | } |
1670 | |
1671 | prepareForMoveTo(What, BB); |
1672 | insertIntoListsForBlock(NewAccess: What, BB, Point); |
1673 | } |
1674 | |
1675 | MemoryPhi *MemorySSA::createMemoryPhi(BasicBlock *BB) { |
1676 | assert(!getMemoryAccess(BB) && "MemoryPhi already exists for this BB" ); |
1677 | MemoryPhi *Phi = new MemoryPhi(BB->getContext(), BB, NextID++); |
1678 | // Phi's always are placed at the front of the block. |
1679 | insertIntoListsForBlock(NewAccess: Phi, BB, Point: Beginning); |
1680 | ValueToMemoryAccess[BB] = Phi; |
1681 | return Phi; |
1682 | } |
1683 | |
1684 | MemoryUseOrDef *MemorySSA::createDefinedAccess(Instruction *I, |
1685 | MemoryAccess *Definition, |
1686 | const MemoryUseOrDef *Template, |
1687 | bool CreationMustSucceed) { |
1688 | assert(!isa<PHINode>(I) && "Cannot create a defined access for a PHI" ); |
1689 | MemoryUseOrDef *NewAccess = createNewAccess(I, AAP: AA, Template); |
1690 | if (CreationMustSucceed) |
1691 | assert(NewAccess != nullptr && "Tried to create a memory access for a " |
1692 | "non-memory touching instruction" ); |
1693 | if (NewAccess) { |
1694 | assert((!Definition || !isa<MemoryUse>(Definition)) && |
1695 | "A use cannot be a defining access" ); |
1696 | NewAccess->setDefiningAccess(DMA: Definition); |
1697 | } |
1698 | return NewAccess; |
1699 | } |
1700 | |
1701 | // Return true if the instruction has ordering constraints. |
1702 | // Note specifically that this only considers stores and loads |
1703 | // because others are still considered ModRef by getModRefInfo. |
1704 | static inline bool isOrdered(const Instruction *I) { |
1705 | if (auto *SI = dyn_cast<StoreInst>(Val: I)) { |
1706 | if (!SI->isUnordered()) |
1707 | return true; |
1708 | } else if (auto *LI = dyn_cast<LoadInst>(Val: I)) { |
1709 | if (!LI->isUnordered()) |
1710 | return true; |
1711 | } |
1712 | return false; |
1713 | } |
1714 | |
1715 | /// Helper function to create new memory accesses |
1716 | template <typename AliasAnalysisType> |
1717 | MemoryUseOrDef *MemorySSA::createNewAccess(Instruction *I, |
1718 | AliasAnalysisType *AAP, |
1719 | const MemoryUseOrDef *Template) { |
1720 | // The assume intrinsic has a control dependency which we model by claiming |
1721 | // that it writes arbitrarily. Debuginfo intrinsics may be considered |
1722 | // clobbers when we have a nonstandard AA pipeline. Ignore these fake memory |
1723 | // dependencies here. |
1724 | // FIXME: Replace this special casing with a more accurate modelling of |
1725 | // assume's control dependency. |
1726 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: I)) { |
1727 | switch (II->getIntrinsicID()) { |
1728 | default: |
1729 | break; |
1730 | case Intrinsic::allow_runtime_check: |
1731 | case Intrinsic::allow_ubsan_check: |
1732 | case Intrinsic::assume: |
1733 | case Intrinsic::experimental_noalias_scope_decl: |
1734 | case Intrinsic::pseudoprobe: |
1735 | return nullptr; |
1736 | } |
1737 | } |
1738 | |
1739 | // Using a nonstandard AA pipelines might leave us with unexpected modref |
1740 | // results for I, so add a check to not model instructions that may not read |
1741 | // from or write to memory. This is necessary for correctness. |
1742 | if (!I->mayReadFromMemory() && !I->mayWriteToMemory()) |
1743 | return nullptr; |
1744 | |
1745 | bool Def, Use; |
1746 | if (Template) { |
1747 | Def = isa<MemoryDef>(Val: Template); |
1748 | Use = isa<MemoryUse>(Val: Template); |
1749 | #if !defined(NDEBUG) |
1750 | ModRefInfo ModRef = AAP->getModRefInfo(I, std::nullopt); |
1751 | bool DefCheck, UseCheck; |
1752 | DefCheck = isModSet(MRI: ModRef) || isOrdered(I); |
1753 | UseCheck = isRefSet(MRI: ModRef); |
1754 | // Memory accesses should only be reduced and can not be increased since AA |
1755 | // just might return better results as a result of some transformations. |
1756 | assert((Def == DefCheck || !DefCheck) && |
1757 | "Memory accesses should only be reduced" ); |
1758 | if (!Def && Use != UseCheck) { |
1759 | // New Access should not have more power than template access |
1760 | assert(!UseCheck && "Invalid template" ); |
1761 | } |
1762 | #endif |
1763 | } else { |
1764 | // Find out what affect this instruction has on memory. |
1765 | ModRefInfo ModRef = AAP->getModRefInfo(I, std::nullopt); |
1766 | // The isOrdered check is used to ensure that volatiles end up as defs |
1767 | // (atomics end up as ModRef right now anyway). Until we separate the |
1768 | // ordering chain from the memory chain, this enables people to see at least |
1769 | // some relative ordering to volatiles. Note that getClobberingMemoryAccess |
1770 | // will still give an answer that bypasses other volatile loads. TODO: |
1771 | // Separate memory aliasing and ordering into two different chains so that |
1772 | // we can precisely represent both "what memory will this read/write/is |
1773 | // clobbered by" and "what instructions can I move this past". |
1774 | Def = isModSet(MRI: ModRef) || isOrdered(I); |
1775 | Use = isRefSet(MRI: ModRef); |
1776 | } |
1777 | |
1778 | // It's possible for an instruction to not modify memory at all. During |
1779 | // construction, we ignore them. |
1780 | if (!Def && !Use) |
1781 | return nullptr; |
1782 | |
1783 | MemoryUseOrDef *MUD; |
1784 | if (Def) { |
1785 | MUD = new MemoryDef(I->getContext(), nullptr, I, I->getParent(), NextID++); |
1786 | } else { |
1787 | MUD = new MemoryUse(I->getContext(), nullptr, I, I->getParent()); |
1788 | if (isUseTriviallyOptimizableToLiveOnEntry(*AAP, I)) { |
1789 | MemoryAccess *LiveOnEntry = getLiveOnEntryDef(); |
1790 | MUD->setOptimized(LiveOnEntry); |
1791 | } |
1792 | } |
1793 | ValueToMemoryAccess[I] = MUD; |
1794 | return MUD; |
1795 | } |
1796 | |
1797 | /// Properly remove \p MA from all of MemorySSA's lookup tables. |
1798 | void MemorySSA::removeFromLookups(MemoryAccess *MA) { |
1799 | assert(MA->use_empty() && |
1800 | "Trying to remove memory access that still has uses" ); |
1801 | BlockNumbering.erase(Val: MA); |
1802 | if (auto *MUD = dyn_cast<MemoryUseOrDef>(Val: MA)) |
1803 | MUD->setDefiningAccess(DMA: nullptr); |
1804 | // Invalidate our walker's cache if necessary |
1805 | if (!isa<MemoryUse>(Val: MA)) |
1806 | getWalker()->invalidateInfo(MA); |
1807 | |
1808 | Value *MemoryInst; |
1809 | if (const auto *MUD = dyn_cast<MemoryUseOrDef>(Val: MA)) |
1810 | MemoryInst = MUD->getMemoryInst(); |
1811 | else |
1812 | MemoryInst = MA->getBlock(); |
1813 | |
1814 | auto VMA = ValueToMemoryAccess.find(Val: MemoryInst); |
1815 | if (VMA->second == MA) |
1816 | ValueToMemoryAccess.erase(I: VMA); |
1817 | } |
1818 | |
1819 | /// Properly remove \p MA from all of MemorySSA's lists. |
1820 | /// |
1821 | /// Because of the way the intrusive list and use lists work, it is important to |
1822 | /// do removal in the right order. |
1823 | /// ShouldDelete defaults to true, and will cause the memory access to also be |
1824 | /// deleted, not just removed. |
1825 | void MemorySSA::removeFromLists(MemoryAccess *MA, bool ShouldDelete) { |
1826 | BasicBlock *BB = MA->getBlock(); |
1827 | // The access list owns the reference, so we erase it from the non-owning list |
1828 | // first. |
1829 | if (!isa<MemoryUse>(Val: MA)) { |
1830 | auto DefsIt = PerBlockDefs.find(Val: BB); |
1831 | std::unique_ptr<DefsList> &Defs = DefsIt->second; |
1832 | Defs->remove(N&: *MA); |
1833 | if (Defs->empty()) |
1834 | PerBlockDefs.erase(I: DefsIt); |
1835 | } |
1836 | |
1837 | // The erase call here will delete it. If we don't want it deleted, we call |
1838 | // remove instead. |
1839 | auto AccessIt = PerBlockAccesses.find(Val: BB); |
1840 | std::unique_ptr<AccessList> &Accesses = AccessIt->second; |
1841 | if (ShouldDelete) |
1842 | Accesses->erase(IT: MA); |
1843 | else |
1844 | Accesses->remove(IT: MA); |
1845 | |
1846 | if (Accesses->empty()) { |
1847 | PerBlockAccesses.erase(I: AccessIt); |
1848 | BlockNumberingValid.erase(Ptr: BB); |
1849 | } |
1850 | } |
1851 | |
1852 | void MemorySSA::print(raw_ostream &OS) const { |
1853 | MemorySSAAnnotatedWriter Writer(this); |
1854 | F.print(OS, AAW: &Writer); |
1855 | } |
1856 | |
1857 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
1858 | LLVM_DUMP_METHOD void MemorySSA::dump() const { print(OS&: dbgs()); } |
1859 | #endif |
1860 | |
1861 | void MemorySSA::verifyMemorySSA(VerificationLevel VL) const { |
1862 | #if !defined(NDEBUG) && defined(EXPENSIVE_CHECKS) |
1863 | VL = VerificationLevel::Full; |
1864 | #endif |
1865 | |
1866 | #ifndef NDEBUG |
1867 | verifyOrderingDominationAndDefUses(F, VL); |
1868 | verifyDominationNumbers(F); |
1869 | if (VL == VerificationLevel::Full) |
1870 | verifyPrevDefInPhis(F); |
1871 | #endif |
1872 | // Previously, the verification used to also verify that the clobberingAccess |
1873 | // cached by MemorySSA is the same as the clobberingAccess found at a later |
1874 | // query to AA. This does not hold true in general due to the current fragility |
1875 | // of BasicAA which has arbitrary caps on the things it analyzes before giving |
1876 | // up. As a result, transformations that are correct, will lead to BasicAA |
1877 | // returning different Alias answers before and after that transformation. |
1878 | // Invalidating MemorySSA is not an option, as the results in BasicAA can be so |
1879 | // random, in the worst case we'd need to rebuild MemorySSA from scratch after |
1880 | // every transformation, which defeats the purpose of using it. For such an |
1881 | // example, see test4 added in D51960. |
1882 | } |
1883 | |
1884 | void MemorySSA::verifyPrevDefInPhis(Function &F) const { |
1885 | for (const BasicBlock &BB : F) { |
1886 | if (MemoryPhi *Phi = getMemoryAccess(BB: &BB)) { |
1887 | for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) { |
1888 | auto *Pred = Phi->getIncomingBlock(I); |
1889 | auto *IncAcc = Phi->getIncomingValue(I); |
1890 | // If Pred has no unreachable predecessors, get last def looking at |
1891 | // IDoms. If, while walkings IDoms, any of these has an unreachable |
1892 | // predecessor, then the incoming def can be any access. |
1893 | if (auto *DTNode = DT->getNode(BB: Pred)) { |
1894 | while (DTNode) { |
1895 | if (auto *DefList = getBlockDefs(BB: DTNode->getBlock())) { |
1896 | auto *LastAcc = &*(--DefList->end()); |
1897 | assert(LastAcc == IncAcc && |
1898 | "Incorrect incoming access into phi." ); |
1899 | (void)IncAcc; |
1900 | (void)LastAcc; |
1901 | break; |
1902 | } |
1903 | DTNode = DTNode->getIDom(); |
1904 | } |
1905 | } else { |
1906 | // If Pred has unreachable predecessors, but has at least a Def, the |
1907 | // incoming access can be the last Def in Pred, or it could have been |
1908 | // optimized to LoE. After an update, though, the LoE may have been |
1909 | // replaced by another access, so IncAcc may be any access. |
1910 | // If Pred has unreachable predecessors and no Defs, incoming access |
1911 | // should be LoE; However, after an update, it may be any access. |
1912 | } |
1913 | } |
1914 | } |
1915 | } |
1916 | } |
1917 | |
1918 | /// Verify that all of the blocks we believe to have valid domination numbers |
1919 | /// actually have valid domination numbers. |
1920 | void MemorySSA::verifyDominationNumbers(const Function &F) const { |
1921 | if (BlockNumberingValid.empty()) |
1922 | return; |
1923 | |
1924 | SmallPtrSet<const BasicBlock *, 16> ValidBlocks = BlockNumberingValid; |
1925 | for (const BasicBlock &BB : F) { |
1926 | if (!ValidBlocks.count(Ptr: &BB)) |
1927 | continue; |
1928 | |
1929 | ValidBlocks.erase(Ptr: &BB); |
1930 | |
1931 | const AccessList *Accesses = getBlockAccesses(BB: &BB); |
1932 | // It's correct to say an empty block has valid numbering. |
1933 | if (!Accesses) |
1934 | continue; |
1935 | |
1936 | // Block numbering starts at 1. |
1937 | unsigned long LastNumber = 0; |
1938 | for (const MemoryAccess &MA : *Accesses) { |
1939 | auto ThisNumberIter = BlockNumbering.find(Val: &MA); |
1940 | assert(ThisNumberIter != BlockNumbering.end() && |
1941 | "MemoryAccess has no domination number in a valid block!" ); |
1942 | |
1943 | unsigned long ThisNumber = ThisNumberIter->second; |
1944 | assert(ThisNumber > LastNumber && |
1945 | "Domination numbers should be strictly increasing!" ); |
1946 | (void)LastNumber; |
1947 | LastNumber = ThisNumber; |
1948 | } |
1949 | } |
1950 | |
1951 | assert(ValidBlocks.empty() && |
1952 | "All valid BasicBlocks should exist in F -- dangling pointers?" ); |
1953 | } |
1954 | |
1955 | /// Verify ordering: the order and existence of MemoryAccesses matches the |
1956 | /// order and existence of memory affecting instructions. |
1957 | /// Verify domination: each definition dominates all of its uses. |
1958 | /// Verify def-uses: the immediate use information - walk all the memory |
1959 | /// accesses and verifying that, for each use, it appears in the appropriate |
1960 | /// def's use list |
1961 | void MemorySSA::verifyOrderingDominationAndDefUses(Function &F, |
1962 | VerificationLevel VL) const { |
1963 | // Walk all the blocks, comparing what the lookups think and what the access |
1964 | // lists think, as well as the order in the blocks vs the order in the access |
1965 | // lists. |
1966 | SmallVector<MemoryAccess *, 32> ActualAccesses; |
1967 | SmallVector<MemoryAccess *, 32> ActualDefs; |
1968 | for (BasicBlock &B : F) { |
1969 | const AccessList *AL = getBlockAccesses(BB: &B); |
1970 | const auto *DL = getBlockDefs(BB: &B); |
1971 | MemoryPhi *Phi = getMemoryAccess(BB: &B); |
1972 | if (Phi) { |
1973 | // Verify ordering. |
1974 | ActualAccesses.push_back(Elt: Phi); |
1975 | ActualDefs.push_back(Elt: Phi); |
1976 | // Verify domination |
1977 | for (const Use &U : Phi->uses()) { |
1978 | assert(dominates(Phi, U) && "Memory PHI does not dominate it's uses" ); |
1979 | (void)U; |
1980 | } |
1981 | // Verify def-uses for full verify. |
1982 | if (VL == VerificationLevel::Full) { |
1983 | assert(Phi->getNumOperands() == pred_size(&B) && |
1984 | "Incomplete MemoryPhi Node" ); |
1985 | for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) { |
1986 | verifyUseInDefs(Phi->getIncomingValue(I), Phi); |
1987 | assert(is_contained(predecessors(&B), Phi->getIncomingBlock(I)) && |
1988 | "Incoming phi block not a block predecessor" ); |
1989 | } |
1990 | } |
1991 | } |
1992 | |
1993 | for (Instruction &I : B) { |
1994 | MemoryUseOrDef *MA = getMemoryAccess(I: &I); |
1995 | assert((!MA || (AL && (isa<MemoryUse>(MA) || DL))) && |
1996 | "We have memory affecting instructions " |
1997 | "in this block but they are not in the " |
1998 | "access list or defs list" ); |
1999 | if (MA) { |
2000 | // Verify ordering. |
2001 | ActualAccesses.push_back(Elt: MA); |
2002 | if (MemoryAccess *MD = dyn_cast<MemoryDef>(Val: MA)) { |
2003 | // Verify ordering. |
2004 | ActualDefs.push_back(Elt: MA); |
2005 | // Verify domination. |
2006 | for (const Use &U : MD->uses()) { |
2007 | assert(dominates(MD, U) && |
2008 | "Memory Def does not dominate it's uses" ); |
2009 | (void)U; |
2010 | } |
2011 | } |
2012 | // Verify def-uses for full verify. |
2013 | if (VL == VerificationLevel::Full) |
2014 | verifyUseInDefs(MA->getDefiningAccess(), MA); |
2015 | } |
2016 | } |
2017 | // Either we hit the assert, really have no accesses, or we have both |
2018 | // accesses and an access list. Same with defs. |
2019 | if (!AL && !DL) |
2020 | continue; |
2021 | // Verify ordering. |
2022 | assert(AL->size() == ActualAccesses.size() && |
2023 | "We don't have the same number of accesses in the block as on the " |
2024 | "access list" ); |
2025 | assert((DL || ActualDefs.size() == 0) && |
2026 | "Either we should have a defs list, or we should have no defs" ); |
2027 | assert((!DL || DL->size() == ActualDefs.size()) && |
2028 | "We don't have the same number of defs in the block as on the " |
2029 | "def list" ); |
2030 | auto ALI = AL->begin(); |
2031 | auto AAI = ActualAccesses.begin(); |
2032 | while (ALI != AL->end() && AAI != ActualAccesses.end()) { |
2033 | assert(&*ALI == *AAI && "Not the same accesses in the same order" ); |
2034 | ++ALI; |
2035 | ++AAI; |
2036 | } |
2037 | ActualAccesses.clear(); |
2038 | if (DL) { |
2039 | auto DLI = DL->begin(); |
2040 | auto ADI = ActualDefs.begin(); |
2041 | while (DLI != DL->end() && ADI != ActualDefs.end()) { |
2042 | assert(&*DLI == *ADI && "Not the same defs in the same order" ); |
2043 | ++DLI; |
2044 | ++ADI; |
2045 | } |
2046 | } |
2047 | ActualDefs.clear(); |
2048 | } |
2049 | } |
2050 | |
2051 | /// Verify the def-use lists in MemorySSA, by verifying that \p Use |
2052 | /// appears in the use list of \p Def. |
2053 | void MemorySSA::verifyUseInDefs(MemoryAccess *Def, MemoryAccess *Use) const { |
2054 | // The live on entry use may cause us to get a NULL def here |
2055 | if (!Def) |
2056 | assert(isLiveOnEntryDef(Use) && |
2057 | "Null def but use not point to live on entry def" ); |
2058 | else |
2059 | assert(is_contained(Def->users(), Use) && |
2060 | "Did not find use in def's use list" ); |
2061 | } |
2062 | |
2063 | /// Perform a local numbering on blocks so that instruction ordering can be |
2064 | /// determined in constant time. |
2065 | /// TODO: We currently just number in order. If we numbered by N, we could |
2066 | /// allow at least N-1 sequences of insertBefore or insertAfter (and at least |
2067 | /// log2(N) sequences of mixed before and after) without needing to invalidate |
2068 | /// the numbering. |
2069 | void MemorySSA::renumberBlock(const BasicBlock *B) const { |
2070 | // The pre-increment ensures the numbers really start at 1. |
2071 | unsigned long CurrentNumber = 0; |
2072 | const AccessList *AL = getBlockAccesses(BB: B); |
2073 | assert(AL != nullptr && "Asking to renumber an empty block" ); |
2074 | for (const auto &I : *AL) |
2075 | BlockNumbering[&I] = ++CurrentNumber; |
2076 | BlockNumberingValid.insert(Ptr: B); |
2077 | } |
2078 | |
2079 | /// Determine, for two memory accesses in the same block, |
2080 | /// whether \p Dominator dominates \p Dominatee. |
2081 | /// \returns True if \p Dominator dominates \p Dominatee. |
2082 | bool MemorySSA::locallyDominates(const MemoryAccess *Dominator, |
2083 | const MemoryAccess *Dominatee) const { |
2084 | const BasicBlock *DominatorBlock = Dominator->getBlock(); |
2085 | |
2086 | assert((DominatorBlock == Dominatee->getBlock()) && |
2087 | "Asking for local domination when accesses are in different blocks!" ); |
2088 | // A node dominates itself. |
2089 | if (Dominatee == Dominator) |
2090 | return true; |
2091 | |
2092 | // When Dominatee is defined on function entry, it is not dominated by another |
2093 | // memory access. |
2094 | if (isLiveOnEntryDef(MA: Dominatee)) |
2095 | return false; |
2096 | |
2097 | // When Dominator is defined on function entry, it dominates the other memory |
2098 | // access. |
2099 | if (isLiveOnEntryDef(MA: Dominator)) |
2100 | return true; |
2101 | |
2102 | if (!BlockNumberingValid.count(Ptr: DominatorBlock)) |
2103 | renumberBlock(B: DominatorBlock); |
2104 | |
2105 | unsigned long DominatorNum = BlockNumbering.lookup(Val: Dominator); |
2106 | // All numbers start with 1 |
2107 | assert(DominatorNum != 0 && "Block was not numbered properly" ); |
2108 | unsigned long DominateeNum = BlockNumbering.lookup(Val: Dominatee); |
2109 | assert(DominateeNum != 0 && "Block was not numbered properly" ); |
2110 | return DominatorNum < DominateeNum; |
2111 | } |
2112 | |
2113 | bool MemorySSA::dominates(const MemoryAccess *Dominator, |
2114 | const MemoryAccess *Dominatee) const { |
2115 | if (Dominator == Dominatee) |
2116 | return true; |
2117 | |
2118 | if (isLiveOnEntryDef(MA: Dominatee)) |
2119 | return false; |
2120 | |
2121 | if (Dominator->getBlock() != Dominatee->getBlock()) |
2122 | return DT->dominates(A: Dominator->getBlock(), B: Dominatee->getBlock()); |
2123 | return locallyDominates(Dominator, Dominatee); |
2124 | } |
2125 | |
2126 | bool MemorySSA::dominates(const MemoryAccess *Dominator, |
2127 | const Use &Dominatee) const { |
2128 | if (MemoryPhi *MP = dyn_cast<MemoryPhi>(Val: Dominatee.getUser())) { |
2129 | BasicBlock *UseBB = MP->getIncomingBlock(U: Dominatee); |
2130 | // The def must dominate the incoming block of the phi. |
2131 | if (UseBB != Dominator->getBlock()) |
2132 | return DT->dominates(A: Dominator->getBlock(), B: UseBB); |
2133 | // If the UseBB and the DefBB are the same, compare locally. |
2134 | return locallyDominates(Dominator, Dominatee: cast<MemoryAccess>(Val: Dominatee)); |
2135 | } |
2136 | // If it's not a PHI node use, the normal dominates can already handle it. |
2137 | return dominates(Dominator, Dominatee: cast<MemoryAccess>(Val: Dominatee.getUser())); |
2138 | } |
2139 | |
2140 | void MemorySSA::ensureOptimizedUses() { |
2141 | if (IsOptimized) |
2142 | return; |
2143 | |
2144 | BatchAAResults BatchAA(*AA); |
2145 | ClobberWalkerBase WalkerBase(this, DT); |
2146 | CachingWalker WalkerLocal(this, &WalkerBase); |
2147 | OptimizeUses(this, &WalkerLocal, &BatchAA, DT).optimizeUses(); |
2148 | IsOptimized = true; |
2149 | } |
2150 | |
2151 | void MemoryAccess::print(raw_ostream &OS) const { |
2152 | switch (getValueID()) { |
2153 | case MemoryPhiVal: return static_cast<const MemoryPhi *>(this)->print(OS); |
2154 | case MemoryDefVal: return static_cast<const MemoryDef *>(this)->print(OS); |
2155 | case MemoryUseVal: return static_cast<const MemoryUse *>(this)->print(OS); |
2156 | } |
2157 | llvm_unreachable("invalid value id" ); |
2158 | } |
2159 | |
2160 | void MemoryDef::print(raw_ostream &OS) const { |
2161 | MemoryAccess *UO = getDefiningAccess(); |
2162 | |
2163 | auto printID = [&OS](MemoryAccess *A) { |
2164 | if (A && A->getID()) |
2165 | OS << A->getID(); |
2166 | else |
2167 | OS << LiveOnEntryStr; |
2168 | }; |
2169 | |
2170 | OS << getID() << " = MemoryDef(" ; |
2171 | printID(UO); |
2172 | OS << ")" ; |
2173 | |
2174 | if (isOptimized()) { |
2175 | OS << "->" ; |
2176 | printID(getOptimized()); |
2177 | } |
2178 | } |
2179 | |
2180 | void MemoryPhi::print(raw_ostream &OS) const { |
2181 | ListSeparator LS("," ); |
2182 | OS << getID() << " = MemoryPhi(" ; |
2183 | for (const auto &Op : operands()) { |
2184 | BasicBlock *BB = getIncomingBlock(U: Op); |
2185 | MemoryAccess *MA = cast<MemoryAccess>(Val: Op); |
2186 | |
2187 | OS << LS << '{'; |
2188 | if (BB->hasName()) |
2189 | OS << BB->getName(); |
2190 | else |
2191 | BB->printAsOperand(O&: OS, PrintType: false); |
2192 | OS << ','; |
2193 | if (unsigned ID = MA->getID()) |
2194 | OS << ID; |
2195 | else |
2196 | OS << LiveOnEntryStr; |
2197 | OS << '}'; |
2198 | } |
2199 | OS << ')'; |
2200 | } |
2201 | |
2202 | void MemoryUse::print(raw_ostream &OS) const { |
2203 | MemoryAccess *UO = getDefiningAccess(); |
2204 | OS << "MemoryUse(" ; |
2205 | if (UO && UO->getID()) |
2206 | OS << UO->getID(); |
2207 | else |
2208 | OS << LiveOnEntryStr; |
2209 | OS << ')'; |
2210 | } |
2211 | |
2212 | void MemoryAccess::dump() const { |
2213 | // Cannot completely remove virtual function even in release mode. |
2214 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
2215 | print(OS&: dbgs()); |
2216 | dbgs() << "\n" ; |
2217 | #endif |
2218 | } |
2219 | |
2220 | class DOTFuncMSSAInfo { |
2221 | private: |
2222 | const Function &F; |
2223 | MemorySSAAnnotatedWriter MSSAWriter; |
2224 | |
2225 | public: |
2226 | DOTFuncMSSAInfo(const Function &F, MemorySSA &MSSA) |
2227 | : F(F), MSSAWriter(&MSSA) {} |
2228 | |
2229 | const Function *getFunction() { return &F; } |
2230 | MemorySSAAnnotatedWriter &getWriter() { return MSSAWriter; } |
2231 | }; |
2232 | |
2233 | namespace llvm { |
2234 | |
2235 | template <> |
2236 | struct GraphTraits<DOTFuncMSSAInfo *> : public GraphTraits<const BasicBlock *> { |
2237 | static NodeRef getEntryNode(DOTFuncMSSAInfo *CFGInfo) { |
2238 | return &(CFGInfo->getFunction()->getEntryBlock()); |
2239 | } |
2240 | |
2241 | // nodes_iterator/begin/end - Allow iteration over all nodes in the graph |
2242 | using nodes_iterator = pointer_iterator<Function::const_iterator>; |
2243 | |
2244 | static nodes_iterator nodes_begin(DOTFuncMSSAInfo *CFGInfo) { |
2245 | return nodes_iterator(CFGInfo->getFunction()->begin()); |
2246 | } |
2247 | |
2248 | static nodes_iterator nodes_end(DOTFuncMSSAInfo *CFGInfo) { |
2249 | return nodes_iterator(CFGInfo->getFunction()->end()); |
2250 | } |
2251 | |
2252 | static size_t size(DOTFuncMSSAInfo *CFGInfo) { |
2253 | return CFGInfo->getFunction()->size(); |
2254 | } |
2255 | }; |
2256 | |
2257 | template <> |
2258 | struct DOTGraphTraits<DOTFuncMSSAInfo *> : public DefaultDOTGraphTraits { |
2259 | |
2260 | DOTGraphTraits(bool IsSimple = false) : DefaultDOTGraphTraits(IsSimple) {} |
2261 | |
2262 | static std::string getGraphName(DOTFuncMSSAInfo *CFGInfo) { |
2263 | return "MSSA CFG for '" + CFGInfo->getFunction()->getName().str() + |
2264 | "' function" ; |
2265 | } |
2266 | |
2267 | std::string getNodeLabel(const BasicBlock *Node, DOTFuncMSSAInfo *CFGInfo) { |
2268 | return DOTGraphTraits<DOTFuncInfo *>::getCompleteNodeLabel( |
2269 | Node, nullptr, |
2270 | HandleBasicBlock: [CFGInfo](raw_string_ostream &OS, const BasicBlock &BB) -> void { |
2271 | BB.print(OS, AAW: &CFGInfo->getWriter(), ShouldPreserveUseListOrder: true, IsForDebug: true); |
2272 | }, |
2273 | HandleComment: [](std::string &S, unsigned &I, unsigned Idx) -> void { |
2274 | std::string Str = S.substr(pos: I, n: Idx - I); |
2275 | StringRef SR = Str; |
2276 | if (SR.count(Str: " = MemoryDef(" ) || SR.count(Str: " = MemoryPhi(" ) || |
2277 | SR.count(Str: "MemoryUse(" )) |
2278 | return; |
2279 | DOTGraphTraits<DOTFuncInfo *>::eraseComment(OutStr&: S, I, Idx); |
2280 | }); |
2281 | } |
2282 | |
2283 | static std::string getEdgeSourceLabel(const BasicBlock *Node, |
2284 | const_succ_iterator I) { |
2285 | return DOTGraphTraits<DOTFuncInfo *>::getEdgeSourceLabel(Node, I); |
2286 | } |
2287 | |
2288 | /// Display the raw branch weights from PGO. |
2289 | std::string getEdgeAttributes(const BasicBlock *Node, const_succ_iterator I, |
2290 | DOTFuncMSSAInfo *CFGInfo) { |
2291 | return "" ; |
2292 | } |
2293 | |
2294 | std::string getNodeAttributes(const BasicBlock *Node, |
2295 | DOTFuncMSSAInfo *CFGInfo) { |
2296 | return getNodeLabel(Node, CFGInfo).find(c: ';') != std::string::npos |
2297 | ? "style=filled, fillcolor=lightpink" |
2298 | : "" ; |
2299 | } |
2300 | }; |
2301 | |
2302 | } // namespace llvm |
2303 | |
2304 | AnalysisKey MemorySSAAnalysis::Key; |
2305 | |
2306 | MemorySSAAnalysis::Result MemorySSAAnalysis::run(Function &F, |
2307 | FunctionAnalysisManager &AM) { |
2308 | auto &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F); |
2309 | auto &AA = AM.getResult<AAManager>(IR&: F); |
2310 | return MemorySSAAnalysis::Result(std::make_unique<MemorySSA>(args&: F, args: &AA, args: &DT)); |
2311 | } |
2312 | |
2313 | bool MemorySSAAnalysis::Result::invalidate( |
2314 | Function &F, const PreservedAnalyses &PA, |
2315 | FunctionAnalysisManager::Invalidator &Inv) { |
2316 | auto PAC = PA.getChecker<MemorySSAAnalysis>(); |
2317 | return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>()) || |
2318 | Inv.invalidate<AAManager>(IR&: F, PA) || |
2319 | Inv.invalidate<DominatorTreeAnalysis>(IR&: F, PA); |
2320 | } |
2321 | |
2322 | PreservedAnalyses MemorySSAPrinterPass::run(Function &F, |
2323 | FunctionAnalysisManager &AM) { |
2324 | auto &MSSA = AM.getResult<MemorySSAAnalysis>(IR&: F).getMSSA(); |
2325 | if (EnsureOptimizedUses) |
2326 | MSSA.ensureOptimizedUses(); |
2327 | if (DotCFGMSSA != "" ) { |
2328 | DOTFuncMSSAInfo CFGInfo(F, MSSA); |
2329 | WriteGraph(G: &CFGInfo, Name: "" , ShortNames: false, Title: "MSSA" , Filename: DotCFGMSSA); |
2330 | } else { |
2331 | OS << "MemorySSA for function: " << F.getName() << "\n" ; |
2332 | MSSA.print(OS); |
2333 | } |
2334 | |
2335 | return PreservedAnalyses::all(); |
2336 | } |
2337 | |
2338 | PreservedAnalyses MemorySSAWalkerPrinterPass::run(Function &F, |
2339 | FunctionAnalysisManager &AM) { |
2340 | auto &MSSA = AM.getResult<MemorySSAAnalysis>(IR&: F).getMSSA(); |
2341 | OS << "MemorySSA (walker) for function: " << F.getName() << "\n" ; |
2342 | MemorySSAWalkerAnnotatedWriter Writer(&MSSA); |
2343 | F.print(OS, AAW: &Writer); |
2344 | |
2345 | return PreservedAnalyses::all(); |
2346 | } |
2347 | |
2348 | PreservedAnalyses MemorySSAVerifierPass::run(Function &F, |
2349 | FunctionAnalysisManager &AM) { |
2350 | AM.getResult<MemorySSAAnalysis>(IR&: F).getMSSA().verifyMemorySSA(); |
2351 | |
2352 | return PreservedAnalyses::all(); |
2353 | } |
2354 | |
2355 | char MemorySSAWrapperPass::ID = 0; |
2356 | |
2357 | MemorySSAWrapperPass::MemorySSAWrapperPass() : FunctionPass(ID) { |
2358 | initializeMemorySSAWrapperPassPass(Registry&: *PassRegistry::getPassRegistry()); |
2359 | } |
2360 | |
2361 | void MemorySSAWrapperPass::releaseMemory() { MSSA.reset(); } |
2362 | |
2363 | void MemorySSAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { |
2364 | AU.setPreservesAll(); |
2365 | AU.addRequiredTransitive<DominatorTreeWrapperPass>(); |
2366 | AU.addRequiredTransitive<AAResultsWrapperPass>(); |
2367 | } |
2368 | |
2369 | bool MemorySSAWrapperPass::runOnFunction(Function &F) { |
2370 | auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
2371 | auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults(); |
2372 | MSSA.reset(p: new MemorySSA(F, &AA, &DT)); |
2373 | return false; |
2374 | } |
2375 | |
2376 | void MemorySSAWrapperPass::verifyAnalysis() const { |
2377 | if (VerifyMemorySSA) |
2378 | MSSA->verifyMemorySSA(); |
2379 | } |
2380 | |
2381 | void MemorySSAWrapperPass::print(raw_ostream &OS, const Module *M) const { |
2382 | MSSA->print(OS); |
2383 | } |
2384 | |
2385 | MemorySSAWalker::MemorySSAWalker(MemorySSA *M) : MSSA(M) {} |
2386 | |
2387 | /// Walk the use-def chains starting at \p StartingAccess and find |
2388 | /// the MemoryAccess that actually clobbers Loc. |
2389 | /// |
2390 | /// \returns our clobbering memory access |
2391 | MemoryAccess *MemorySSA::ClobberWalkerBase::getClobberingMemoryAccessBase( |
2392 | MemoryAccess *StartingAccess, const MemoryLocation &Loc, |
2393 | BatchAAResults &BAA, unsigned &UpwardWalkLimit) { |
2394 | assert(!isa<MemoryUse>(StartingAccess) && "Use cannot be defining access" ); |
2395 | |
2396 | // If location is undefined, conservatively return starting access. |
2397 | if (Loc.Ptr == nullptr) |
2398 | return StartingAccess; |
2399 | |
2400 | Instruction *I = nullptr; |
2401 | if (auto *StartingUseOrDef = dyn_cast<MemoryUseOrDef>(Val: StartingAccess)) { |
2402 | if (MSSA->isLiveOnEntryDef(MA: StartingUseOrDef)) |
2403 | return StartingUseOrDef; |
2404 | |
2405 | I = StartingUseOrDef->getMemoryInst(); |
2406 | |
2407 | // Conservatively, fences are always clobbers, so don't perform the walk if |
2408 | // we hit a fence. |
2409 | if (!isa<CallBase>(Val: I) && I->isFenceLike()) |
2410 | return StartingUseOrDef; |
2411 | } |
2412 | |
2413 | UpwardsMemoryQuery Q; |
2414 | Q.OriginalAccess = StartingAccess; |
2415 | Q.StartingLoc = Loc; |
2416 | Q.Inst = nullptr; |
2417 | Q.IsCall = false; |
2418 | |
2419 | // Unlike the other function, do not walk to the def of a def, because we are |
2420 | // handed something we already believe is the clobbering access. |
2421 | // We never set SkipSelf to true in Q in this method. |
2422 | MemoryAccess *Clobber = |
2423 | Walker.findClobber(BAA, Start: StartingAccess, Q, UpWalkLimit&: UpwardWalkLimit); |
2424 | LLVM_DEBUG({ |
2425 | dbgs() << "Clobber starting at access " << *StartingAccess << "\n" ; |
2426 | if (I) |
2427 | dbgs() << " for instruction " << *I << "\n" ; |
2428 | dbgs() << " is " << *Clobber << "\n" ; |
2429 | }); |
2430 | return Clobber; |
2431 | } |
2432 | |
2433 | static const Instruction * |
2434 | getInvariantGroupClobberingInstruction(Instruction &I, DominatorTree &DT) { |
2435 | if (!I.hasMetadata(KindID: LLVMContext::MD_invariant_group) || I.isVolatile()) |
2436 | return nullptr; |
2437 | |
2438 | // We consider bitcasts and zero GEPs to be the same pointer value. Start by |
2439 | // stripping bitcasts and zero GEPs, then we will recursively look at loads |
2440 | // and stores through bitcasts and zero GEPs. |
2441 | Value *PointerOperand = getLoadStorePointerOperand(V: &I)->stripPointerCasts(); |
2442 | |
2443 | // It's not safe to walk the use list of a global value because function |
2444 | // passes aren't allowed to look outside their functions. |
2445 | // FIXME: this could be fixed by filtering instructions from outside of |
2446 | // current function. |
2447 | if (isa<Constant>(Val: PointerOperand)) |
2448 | return nullptr; |
2449 | |
2450 | // Queue to process all pointers that are equivalent to load operand. |
2451 | SmallVector<const Value *, 8> PointerUsesQueue; |
2452 | PointerUsesQueue.push_back(Elt: PointerOperand); |
2453 | |
2454 | const Instruction *MostDominatingInstruction = &I; |
2455 | |
2456 | // FIXME: This loop is O(n^2) because dominates can be O(n) and in worst case |
2457 | // we will see all the instructions. It may not matter in practice. If it |
2458 | // does, we will have to support MemorySSA construction and updates. |
2459 | while (!PointerUsesQueue.empty()) { |
2460 | const Value *Ptr = PointerUsesQueue.pop_back_val(); |
2461 | assert(Ptr && !isa<GlobalValue>(Ptr) && |
2462 | "Null or GlobalValue should not be inserted" ); |
2463 | |
2464 | for (const User *Us : Ptr->users()) { |
2465 | auto *U = dyn_cast<Instruction>(Val: Us); |
2466 | if (!U || U == &I || !DT.dominates(Def: U, User: MostDominatingInstruction)) |
2467 | continue; |
2468 | |
2469 | // Add bitcasts and zero GEPs to queue. |
2470 | if (isa<BitCastInst>(Val: U)) { |
2471 | PointerUsesQueue.push_back(Elt: U); |
2472 | continue; |
2473 | } |
2474 | if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: U)) { |
2475 | if (GEP->hasAllZeroIndices()) |
2476 | PointerUsesQueue.push_back(Elt: U); |
2477 | continue; |
2478 | } |
2479 | |
2480 | // If we hit a load/store with an invariant.group metadata and the same |
2481 | // pointer operand, we can assume that value pointed to by the pointer |
2482 | // operand didn't change. |
2483 | if (U->hasMetadata(KindID: LLVMContext::MD_invariant_group) && |
2484 | getLoadStorePointerOperand(V: U) == Ptr && !U->isVolatile()) { |
2485 | MostDominatingInstruction = U; |
2486 | } |
2487 | } |
2488 | } |
2489 | return MostDominatingInstruction == &I ? nullptr : MostDominatingInstruction; |
2490 | } |
2491 | |
2492 | MemoryAccess *MemorySSA::ClobberWalkerBase::getClobberingMemoryAccessBase( |
2493 | MemoryAccess *MA, BatchAAResults &BAA, unsigned &UpwardWalkLimit, |
2494 | bool SkipSelf, bool UseInvariantGroup) { |
2495 | auto *StartingAccess = dyn_cast<MemoryUseOrDef>(Val: MA); |
2496 | // If this is a MemoryPhi, we can't do anything. |
2497 | if (!StartingAccess) |
2498 | return MA; |
2499 | |
2500 | if (UseInvariantGroup) { |
2501 | if (auto *I = getInvariantGroupClobberingInstruction( |
2502 | I&: *StartingAccess->getMemoryInst(), DT&: MSSA->getDomTree())) { |
2503 | assert(isa<LoadInst>(I) || isa<StoreInst>(I)); |
2504 | |
2505 | auto *ClobberMA = MSSA->getMemoryAccess(I); |
2506 | assert(ClobberMA); |
2507 | if (isa<MemoryUse>(Val: ClobberMA)) |
2508 | return ClobberMA->getDefiningAccess(); |
2509 | return ClobberMA; |
2510 | } |
2511 | } |
2512 | |
2513 | bool IsOptimized = false; |
2514 | |
2515 | // If this is an already optimized use or def, return the optimized result. |
2516 | // Note: Currently, we store the optimized def result in a separate field, |
2517 | // since we can't use the defining access. |
2518 | if (StartingAccess->isOptimized()) { |
2519 | if (!SkipSelf || !isa<MemoryDef>(Val: StartingAccess)) |
2520 | return StartingAccess->getOptimized(); |
2521 | IsOptimized = true; |
2522 | } |
2523 | |
2524 | const Instruction *I = StartingAccess->getMemoryInst(); |
2525 | // We can't sanely do anything with a fence, since they conservatively clobber |
2526 | // all memory, and have no locations to get pointers from to try to |
2527 | // disambiguate. |
2528 | if (!isa<CallBase>(Val: I) && I->isFenceLike()) |
2529 | return StartingAccess; |
2530 | |
2531 | UpwardsMemoryQuery Q(I, StartingAccess); |
2532 | |
2533 | if (isUseTriviallyOptimizableToLiveOnEntry(AA&: BAA, I)) { |
2534 | MemoryAccess *LiveOnEntry = MSSA->getLiveOnEntryDef(); |
2535 | StartingAccess->setOptimized(LiveOnEntry); |
2536 | return LiveOnEntry; |
2537 | } |
2538 | |
2539 | MemoryAccess *OptimizedAccess; |
2540 | if (!IsOptimized) { |
2541 | // Start with the thing we already think clobbers this location |
2542 | MemoryAccess *DefiningAccess = StartingAccess->getDefiningAccess(); |
2543 | |
2544 | // At this point, DefiningAccess may be the live on entry def. |
2545 | // If it is, we will not get a better result. |
2546 | if (MSSA->isLiveOnEntryDef(MA: DefiningAccess)) { |
2547 | StartingAccess->setOptimized(DefiningAccess); |
2548 | return DefiningAccess; |
2549 | } |
2550 | |
2551 | OptimizedAccess = |
2552 | Walker.findClobber(BAA, Start: DefiningAccess, Q, UpWalkLimit&: UpwardWalkLimit); |
2553 | StartingAccess->setOptimized(OptimizedAccess); |
2554 | } else |
2555 | OptimizedAccess = StartingAccess->getOptimized(); |
2556 | |
2557 | LLVM_DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is " ); |
2558 | LLVM_DEBUG(dbgs() << *StartingAccess << "\n" ); |
2559 | LLVM_DEBUG(dbgs() << "Optimized Memory SSA clobber for " << *I << " is " ); |
2560 | LLVM_DEBUG(dbgs() << *OptimizedAccess << "\n" ); |
2561 | |
2562 | MemoryAccess *Result; |
2563 | if (SkipSelf && isa<MemoryPhi>(Val: OptimizedAccess) && |
2564 | isa<MemoryDef>(Val: StartingAccess) && UpwardWalkLimit) { |
2565 | assert(isa<MemoryDef>(Q.OriginalAccess)); |
2566 | Q.SkipSelfAccess = true; |
2567 | Result = Walker.findClobber(BAA, Start: OptimizedAccess, Q, UpWalkLimit&: UpwardWalkLimit); |
2568 | } else |
2569 | Result = OptimizedAccess; |
2570 | |
2571 | LLVM_DEBUG(dbgs() << "Result Memory SSA clobber [SkipSelf = " << SkipSelf); |
2572 | LLVM_DEBUG(dbgs() << "] for " << *I << " is " << *Result << "\n" ); |
2573 | |
2574 | return Result; |
2575 | } |
2576 | |
2577 | MemoryAccess * |
2578 | DoNothingMemorySSAWalker::getClobberingMemoryAccess(MemoryAccess *MA, |
2579 | BatchAAResults &) { |
2580 | if (auto *Use = dyn_cast<MemoryUseOrDef>(Val: MA)) |
2581 | return Use->getDefiningAccess(); |
2582 | return MA; |
2583 | } |
2584 | |
2585 | MemoryAccess *DoNothingMemorySSAWalker::getClobberingMemoryAccess( |
2586 | MemoryAccess *StartingAccess, const MemoryLocation &, BatchAAResults &) { |
2587 | if (auto *Use = dyn_cast<MemoryUseOrDef>(Val: StartingAccess)) |
2588 | return Use->getDefiningAccess(); |
2589 | return StartingAccess; |
2590 | } |
2591 | |
2592 | void MemoryPhi::deleteMe(DerivedUser *Self) { |
2593 | delete static_cast<MemoryPhi *>(Self); |
2594 | } |
2595 | |
2596 | void MemoryDef::deleteMe(DerivedUser *Self) { |
2597 | delete static_cast<MemoryDef *>(Self); |
2598 | } |
2599 | |
2600 | void MemoryUse::deleteMe(DerivedUser *Self) { |
2601 | delete static_cast<MemoryUse *>(Self); |
2602 | } |
2603 | |
2604 | bool upward_defs_iterator::IsGuaranteedLoopInvariant(const Value *Ptr) const { |
2605 | auto IsGuaranteedLoopInvariantBase = [](const Value *Ptr) { |
2606 | Ptr = Ptr->stripPointerCasts(); |
2607 | if (!isa<Instruction>(Val: Ptr)) |
2608 | return true; |
2609 | return isa<AllocaInst>(Val: Ptr); |
2610 | }; |
2611 | |
2612 | Ptr = Ptr->stripPointerCasts(); |
2613 | if (auto *I = dyn_cast<Instruction>(Val: Ptr)) { |
2614 | if (I->getParent()->isEntryBlock()) |
2615 | return true; |
2616 | } |
2617 | if (auto *GEP = dyn_cast<GEPOperator>(Val: Ptr)) { |
2618 | return IsGuaranteedLoopInvariantBase(GEP->getPointerOperand()) && |
2619 | GEP->hasAllConstantIndices(); |
2620 | } |
2621 | return IsGuaranteedLoopInvariantBase(Ptr); |
2622 | } |
2623 | |