[ARM] MVE sext costs
[llvm-complete.git] / lib / Analysis / MemorySSA.cpp
blob62bd19133b7c94793ad6c9ad9bee136b74cd96ee
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.
11 //===----------------------------------------------------------------------===//
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/None.h"
20 #include "llvm/ADT/Optional.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/iterator.h"
25 #include "llvm/ADT/iterator_range.h"
26 #include "llvm/Analysis/AliasAnalysis.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/Intrinsics.h"
38 #include "llvm/IR/LLVMContext.h"
39 #include "llvm/IR/PassManager.h"
40 #include "llvm/IR/Use.h"
41 #include "llvm/Pass.h"
42 #include "llvm/Support/AtomicOrdering.h"
43 #include "llvm/Support/Casting.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Compiler.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/ErrorHandling.h"
48 #include "llvm/Support/FormattedStream.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include <algorithm>
51 #include <cassert>
52 #include <cstdlib>
53 #include <iterator>
54 #include <memory>
55 #include <utility>
57 using namespace llvm;
59 #define DEBUG_TYPE "memoryssa"
61 INITIALIZE_PASS_BEGIN(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,
62 true)
63 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
64 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
65 INITIALIZE_PASS_END(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,
66 true)
68 INITIALIZE_PASS_BEGIN(MemorySSAPrinterLegacyPass, "print-memoryssa",
69 "Memory SSA Printer", false, false)
70 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
71 INITIALIZE_PASS_END(MemorySSAPrinterLegacyPass, "print-memoryssa",
72 "Memory SSA Printer", false, false)
74 static cl::opt<unsigned> MaxCheckLimit(
75 "memssa-check-limit", cl::Hidden, cl::init(100),
76 cl::desc("The maximum number of stores/phis MemorySSA"
77 "will consider trying to walk past (default = 100)"));
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 /// Enables memory ssa as a dependency for loop passes in legacy pass manager.
86 cl::opt<bool> llvm::EnableMSSALoopDependency(
87 "enable-mssa-loop-dependency", cl::Hidden, cl::init(false),
88 cl::desc("Enable MemorySSA dependency for loop pass manager"));
90 static cl::opt<bool, true>
91 VerifyMemorySSAX("verify-memoryssa", cl::location(VerifyMemorySSA),
92 cl::Hidden, cl::desc("Enable verification of MemorySSA."));
94 namespace llvm {
96 /// An assembly annotator class to print Memory SSA information in
97 /// comments.
98 class MemorySSAAnnotatedWriter : public AssemblyAnnotationWriter {
99 friend class MemorySSA;
101 const MemorySSA *MSSA;
103 public:
104 MemorySSAAnnotatedWriter(const MemorySSA *M) : MSSA(M) {}
106 void emitBasicBlockStartAnnot(const BasicBlock *BB,
107 formatted_raw_ostream &OS) override {
108 if (MemoryAccess *MA = MSSA->getMemoryAccess(BB))
109 OS << "; " << *MA << "\n";
112 void emitInstructionAnnot(const Instruction *I,
113 formatted_raw_ostream &OS) override {
114 if (MemoryAccess *MA = MSSA->getMemoryAccess(I))
115 OS << "; " << *MA << "\n";
119 } // end namespace llvm
121 namespace {
123 /// Our current alias analysis API differentiates heavily between calls and
124 /// non-calls, and functions called on one usually assert on the other.
125 /// This class encapsulates the distinction to simplify other code that wants
126 /// "Memory affecting instructions and related data" to use as a key.
127 /// For example, this class is used as a densemap key in the use optimizer.
128 class MemoryLocOrCall {
129 public:
130 bool IsCall = false;
132 MemoryLocOrCall(MemoryUseOrDef *MUD)
133 : MemoryLocOrCall(MUD->getMemoryInst()) {}
134 MemoryLocOrCall(const MemoryUseOrDef *MUD)
135 : MemoryLocOrCall(MUD->getMemoryInst()) {}
137 MemoryLocOrCall(Instruction *Inst) {
138 if (auto *C = dyn_cast<CallBase>(Inst)) {
139 IsCall = true;
140 Call = C;
141 } else {
142 IsCall = false;
143 // There is no such thing as a memorylocation for a fence inst, and it is
144 // unique in that regard.
145 if (!isa<FenceInst>(Inst))
146 Loc = MemoryLocation::get(Inst);
150 explicit MemoryLocOrCall(const MemoryLocation &Loc) : Loc(Loc) {}
152 const CallBase *getCall() const {
153 assert(IsCall);
154 return Call;
157 MemoryLocation getLoc() const {
158 assert(!IsCall);
159 return Loc;
162 bool operator==(const MemoryLocOrCall &Other) const {
163 if (IsCall != Other.IsCall)
164 return false;
166 if (!IsCall)
167 return Loc == Other.Loc;
169 if (Call->getCalledValue() != Other.Call->getCalledValue())
170 return false;
172 return Call->arg_size() == Other.Call->arg_size() &&
173 std::equal(Call->arg_begin(), Call->arg_end(),
174 Other.Call->arg_begin());
177 private:
178 union {
179 const CallBase *Call;
180 MemoryLocation Loc;
184 } // end anonymous namespace
186 namespace llvm {
188 template <> struct DenseMapInfo<MemoryLocOrCall> {
189 static inline MemoryLocOrCall getEmptyKey() {
190 return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getEmptyKey());
193 static inline MemoryLocOrCall getTombstoneKey() {
194 return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getTombstoneKey());
197 static unsigned getHashValue(const MemoryLocOrCall &MLOC) {
198 if (!MLOC.IsCall)
199 return hash_combine(
200 MLOC.IsCall,
201 DenseMapInfo<MemoryLocation>::getHashValue(MLOC.getLoc()));
203 hash_code hash =
204 hash_combine(MLOC.IsCall, DenseMapInfo<const Value *>::getHashValue(
205 MLOC.getCall()->getCalledValue()));
207 for (const Value *Arg : MLOC.getCall()->args())
208 hash = hash_combine(hash, DenseMapInfo<const Value *>::getHashValue(Arg));
209 return hash;
212 static bool isEqual(const MemoryLocOrCall &LHS, const MemoryLocOrCall &RHS) {
213 return LHS == RHS;
217 } // end namespace llvm
219 /// This does one-way checks to see if Use could theoretically be hoisted above
220 /// MayClobber. This will not check the other way around.
222 /// This assumes that, for the purposes of MemorySSA, Use comes directly after
223 /// MayClobber, with no potentially clobbering operations in between them.
224 /// (Where potentially clobbering ops are memory barriers, aliased stores, etc.)
225 static bool areLoadsReorderable(const LoadInst *Use,
226 const LoadInst *MayClobber) {
227 bool VolatileUse = Use->isVolatile();
228 bool VolatileClobber = MayClobber->isVolatile();
229 // Volatile operations may never be reordered with other volatile operations.
230 if (VolatileUse && VolatileClobber)
231 return false;
232 // Otherwise, volatile doesn't matter here. From the language reference:
233 // 'optimizers may change the order of volatile operations relative to
234 // non-volatile operations.'"
236 // If a load is seq_cst, it cannot be moved above other loads. If its ordering
237 // is weaker, it can be moved above other loads. We just need to be sure that
238 // MayClobber isn't an acquire load, because loads can't be moved above
239 // acquire loads.
241 // Note that this explicitly *does* allow the free reordering of monotonic (or
242 // weaker) loads of the same address.
243 bool SeqCstUse = Use->getOrdering() == AtomicOrdering::SequentiallyConsistent;
244 bool MayClobberIsAcquire = isAtLeastOrStrongerThan(MayClobber->getOrdering(),
245 AtomicOrdering::Acquire);
246 return !(SeqCstUse || MayClobberIsAcquire);
249 namespace {
251 struct ClobberAlias {
252 bool IsClobber;
253 Optional<AliasResult> AR;
256 } // end anonymous namespace
258 // Return a pair of {IsClobber (bool), AR (AliasResult)}. It relies on AR being
259 // ignored if IsClobber = false.
260 template <typename AliasAnalysisType>
261 static ClobberAlias
262 instructionClobbersQuery(const MemoryDef *MD, const MemoryLocation &UseLoc,
263 const Instruction *UseInst, AliasAnalysisType &AA) {
264 Instruction *DefInst = MD->getMemoryInst();
265 assert(DefInst && "Defining instruction not actually an instruction");
266 const auto *UseCall = dyn_cast<CallBase>(UseInst);
267 Optional<AliasResult> AR;
269 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(DefInst)) {
270 // These intrinsics will show up as affecting memory, but they are just
271 // markers, mostly.
273 // FIXME: We probably don't actually want MemorySSA to model these at all
274 // (including creating MemoryAccesses for them): we just end up inventing
275 // clobbers where they don't really exist at all. Please see D43269 for
276 // context.
277 switch (II->getIntrinsicID()) {
278 case Intrinsic::lifetime_start:
279 if (UseCall)
280 return {false, NoAlias};
281 AR = AA.alias(MemoryLocation(II->getArgOperand(1)), UseLoc);
282 return {AR != NoAlias, AR};
283 case Intrinsic::lifetime_end:
284 case Intrinsic::invariant_start:
285 case Intrinsic::invariant_end:
286 case Intrinsic::assume:
287 return {false, NoAlias};
288 default:
289 break;
293 if (UseCall) {
294 ModRefInfo I = AA.getModRefInfo(DefInst, UseCall);
295 AR = isMustSet(I) ? MustAlias : MayAlias;
296 return {isModOrRefSet(I), AR};
299 if (auto *DefLoad = dyn_cast<LoadInst>(DefInst))
300 if (auto *UseLoad = dyn_cast<LoadInst>(UseInst))
301 return {!areLoadsReorderable(UseLoad, DefLoad), MayAlias};
303 ModRefInfo I = AA.getModRefInfo(DefInst, UseLoc);
304 AR = isMustSet(I) ? MustAlias : MayAlias;
305 return {isModSet(I), AR};
308 template <typename AliasAnalysisType>
309 static ClobberAlias instructionClobbersQuery(MemoryDef *MD,
310 const MemoryUseOrDef *MU,
311 const MemoryLocOrCall &UseMLOC,
312 AliasAnalysisType &AA) {
313 // FIXME: This is a temporary hack to allow a single instructionClobbersQuery
314 // to exist while MemoryLocOrCall is pushed through places.
315 if (UseMLOC.IsCall)
316 return instructionClobbersQuery(MD, MemoryLocation(), MU->getMemoryInst(),
317 AA);
318 return instructionClobbersQuery(MD, UseMLOC.getLoc(), MU->getMemoryInst(),
319 AA);
322 // Return true when MD may alias MU, return false otherwise.
323 bool MemorySSAUtil::defClobbersUseOrDef(MemoryDef *MD, const MemoryUseOrDef *MU,
324 AliasAnalysis &AA) {
325 return instructionClobbersQuery(MD, MU, MemoryLocOrCall(MU), AA).IsClobber;
328 namespace {
330 struct UpwardsMemoryQuery {
331 // True if our original query started off as a call
332 bool IsCall = false;
333 // The pointer location we started the query with. This will be empty if
334 // IsCall is true.
335 MemoryLocation StartingLoc;
336 // This is the instruction we were querying about.
337 const Instruction *Inst = nullptr;
338 // The MemoryAccess we actually got called with, used to test local domination
339 const MemoryAccess *OriginalAccess = nullptr;
340 Optional<AliasResult> AR = MayAlias;
341 bool SkipSelfAccess = false;
343 UpwardsMemoryQuery() = default;
345 UpwardsMemoryQuery(const Instruction *Inst, const MemoryAccess *Access)
346 : IsCall(isa<CallBase>(Inst)), Inst(Inst), OriginalAccess(Access) {
347 if (!IsCall)
348 StartingLoc = MemoryLocation::get(Inst);
352 } // end anonymous namespace
354 static bool lifetimeEndsAt(MemoryDef *MD, const MemoryLocation &Loc,
355 BatchAAResults &AA) {
356 Instruction *Inst = MD->getMemoryInst();
357 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
358 switch (II->getIntrinsicID()) {
359 case Intrinsic::lifetime_end:
360 return AA.alias(MemoryLocation(II->getArgOperand(1)), Loc) == MustAlias;
361 default:
362 return false;
365 return false;
368 template <typename AliasAnalysisType>
369 static bool isUseTriviallyOptimizableToLiveOnEntry(AliasAnalysisType &AA,
370 const Instruction *I) {
371 // If the memory can't be changed, then loads of the memory can't be
372 // clobbered.
373 return isa<LoadInst>(I) && (I->getMetadata(LLVMContext::MD_invariant_load) ||
374 AA.pointsToConstantMemory(MemoryLocation(
375 cast<LoadInst>(I)->getPointerOperand())));
378 /// Verifies that `Start` is clobbered by `ClobberAt`, and that nothing
379 /// inbetween `Start` and `ClobberAt` can clobbers `Start`.
381 /// This is meant to be as simple and self-contained as possible. Because it
382 /// uses no cache, etc., it can be relatively expensive.
384 /// \param Start The MemoryAccess that we want to walk from.
385 /// \param ClobberAt A clobber for Start.
386 /// \param StartLoc The MemoryLocation for Start.
387 /// \param MSSA The MemorySSA instance that Start and ClobberAt belong to.
388 /// \param Query The UpwardsMemoryQuery we used for our search.
389 /// \param AA The AliasAnalysis we used for our search.
390 /// \param AllowImpreciseClobber Always false, unless we do relaxed verify.
392 template <typename AliasAnalysisType>
393 LLVM_ATTRIBUTE_UNUSED static void
394 checkClobberSanity(const MemoryAccess *Start, MemoryAccess *ClobberAt,
395 const MemoryLocation &StartLoc, const MemorySSA &MSSA,
396 const UpwardsMemoryQuery &Query, AliasAnalysisType &AA,
397 bool AllowImpreciseClobber = false) {
398 assert(MSSA.dominates(ClobberAt, Start) && "Clobber doesn't dominate start?");
400 if (MSSA.isLiveOnEntryDef(Start)) {
401 assert(MSSA.isLiveOnEntryDef(ClobberAt) &&
402 "liveOnEntry must clobber itself");
403 return;
406 bool FoundClobber = false;
407 DenseSet<ConstMemoryAccessPair> VisitedPhis;
408 SmallVector<ConstMemoryAccessPair, 8> Worklist;
409 Worklist.emplace_back(Start, StartLoc);
410 // Walk all paths from Start to ClobberAt, while looking for clobbers. If one
411 // is found, complain.
412 while (!Worklist.empty()) {
413 auto MAP = Worklist.pop_back_val();
414 // All we care about is that nothing from Start to ClobberAt clobbers Start.
415 // We learn nothing from revisiting nodes.
416 if (!VisitedPhis.insert(MAP).second)
417 continue;
419 for (const auto *MA : def_chain(MAP.first)) {
420 if (MA == ClobberAt) {
421 if (const auto *MD = dyn_cast<MemoryDef>(MA)) {
422 // instructionClobbersQuery isn't essentially free, so don't use `|=`,
423 // since it won't let us short-circuit.
425 // Also, note that this can't be hoisted out of the `Worklist` loop,
426 // since MD may only act as a clobber for 1 of N MemoryLocations.
427 FoundClobber = FoundClobber || MSSA.isLiveOnEntryDef(MD);
428 if (!FoundClobber) {
429 ClobberAlias CA =
430 instructionClobbersQuery(MD, MAP.second, Query.Inst, AA);
431 if (CA.IsClobber) {
432 FoundClobber = true;
433 // Not used: CA.AR;
437 break;
440 // We should never hit liveOnEntry, unless it's the clobber.
441 assert(!MSSA.isLiveOnEntryDef(MA) && "Hit liveOnEntry before clobber?");
443 if (const auto *MD = dyn_cast<MemoryDef>(MA)) {
444 // If Start is a Def, skip self.
445 if (MD == Start)
446 continue;
448 assert(!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA)
449 .IsClobber &&
450 "Found clobber before reaching ClobberAt!");
451 continue;
454 if (const auto *MU = dyn_cast<MemoryUse>(MA)) {
455 (void)MU;
456 assert (MU == Start &&
457 "Can only find use in def chain if Start is a use");
458 continue;
461 assert(isa<MemoryPhi>(MA));
462 Worklist.append(
463 upward_defs_begin({const_cast<MemoryAccess *>(MA), MAP.second}),
464 upward_defs_end());
468 // If the verify is done following an optimization, it's possible that
469 // ClobberAt was a conservative clobbering, that we can now infer is not a
470 // true clobbering access. Don't fail the verify if that's the case.
471 // We do have accesses that claim they're optimized, but could be optimized
472 // further. Updating all these can be expensive, so allow it for now (FIXME).
473 if (AllowImpreciseClobber)
474 return;
476 // If ClobberAt is a MemoryPhi, we can assume something above it acted as a
477 // clobber. Otherwise, `ClobberAt` should've acted as a clobber at some point.
478 assert((isa<MemoryPhi>(ClobberAt) || FoundClobber) &&
479 "ClobberAt never acted as a clobber");
482 namespace {
484 /// Our algorithm for walking (and trying to optimize) clobbers, all wrapped up
485 /// in one class.
486 template <class AliasAnalysisType> class ClobberWalker {
487 /// Save a few bytes by using unsigned instead of size_t.
488 using ListIndex = unsigned;
490 /// Represents a span of contiguous MemoryDefs, potentially ending in a
491 /// MemoryPhi.
492 struct DefPath {
493 MemoryLocation Loc;
494 // Note that, because we always walk in reverse, Last will always dominate
495 // First. Also note that First and Last are inclusive.
496 MemoryAccess *First;
497 MemoryAccess *Last;
498 Optional<ListIndex> Previous;
500 DefPath(const MemoryLocation &Loc, MemoryAccess *First, MemoryAccess *Last,
501 Optional<ListIndex> Previous)
502 : Loc(Loc), First(First), Last(Last), Previous(Previous) {}
504 DefPath(const MemoryLocation &Loc, MemoryAccess *Init,
505 Optional<ListIndex> Previous)
506 : DefPath(Loc, Init, Init, Previous) {}
509 const MemorySSA &MSSA;
510 AliasAnalysisType &AA;
511 DominatorTree &DT;
512 UpwardsMemoryQuery *Query;
513 unsigned *UpwardWalkLimit;
515 // Phi optimization bookkeeping
516 SmallVector<DefPath, 32> Paths;
517 DenseSet<ConstMemoryAccessPair> VisitedPhis;
519 /// Find the nearest def or phi that `From` can legally be optimized to.
520 const MemoryAccess *getWalkTarget(const MemoryPhi *From) const {
521 assert(From->getNumOperands() && "Phi with no operands?");
523 BasicBlock *BB = From->getBlock();
524 MemoryAccess *Result = MSSA.getLiveOnEntryDef();
525 DomTreeNode *Node = DT.getNode(BB);
526 while ((Node = Node->getIDom())) {
527 auto *Defs = MSSA.getBlockDefs(Node->getBlock());
528 if (Defs)
529 return &*Defs->rbegin();
531 return Result;
534 /// Result of calling walkToPhiOrClobber.
535 struct UpwardsWalkResult {
536 /// The "Result" of the walk. Either a clobber, the last thing we walked, or
537 /// both. Include alias info when clobber found.
538 MemoryAccess *Result;
539 bool IsKnownClobber;
540 Optional<AliasResult> AR;
543 /// Walk to the next Phi or Clobber in the def chain starting at Desc.Last.
544 /// This will update Desc.Last as it walks. It will (optionally) also stop at
545 /// StopAt.
547 /// This does not test for whether StopAt is a clobber
548 UpwardsWalkResult
549 walkToPhiOrClobber(DefPath &Desc, const MemoryAccess *StopAt = nullptr,
550 const MemoryAccess *SkipStopAt = nullptr) const {
551 assert(!isa<MemoryUse>(Desc.Last) && "Uses don't exist in my world");
552 assert(UpwardWalkLimit && "Need a valid walk limit");
553 bool LimitAlreadyReached = false;
554 // (*UpwardWalkLimit) may be 0 here, due to the loop in tryOptimizePhi. Set
555 // it to 1. This will not do any alias() calls. It either returns in the
556 // first iteration in the loop below, or is set back to 0 if all def chains
557 // are free of MemoryDefs.
558 if (!*UpwardWalkLimit) {
559 *UpwardWalkLimit = 1;
560 LimitAlreadyReached = true;
563 for (MemoryAccess *Current : def_chain(Desc.Last)) {
564 Desc.Last = Current;
565 if (Current == StopAt || Current == SkipStopAt)
566 return {Current, false, MayAlias};
568 if (auto *MD = dyn_cast<MemoryDef>(Current)) {
569 if (MSSA.isLiveOnEntryDef(MD))
570 return {MD, true, MustAlias};
572 if (!--*UpwardWalkLimit)
573 return {Current, true, MayAlias};
575 ClobberAlias CA =
576 instructionClobbersQuery(MD, Desc.Loc, Query->Inst, AA);
577 if (CA.IsClobber)
578 return {MD, true, CA.AR};
582 if (LimitAlreadyReached)
583 *UpwardWalkLimit = 0;
585 assert(isa<MemoryPhi>(Desc.Last) &&
586 "Ended at a non-clobber that's not a phi?");
587 return {Desc.Last, false, MayAlias};
590 void addSearches(MemoryPhi *Phi, SmallVectorImpl<ListIndex> &PausedSearches,
591 ListIndex PriorNode) {
592 auto UpwardDefs = make_range(upward_defs_begin({Phi, Paths[PriorNode].Loc}),
593 upward_defs_end());
594 for (const MemoryAccessPair &P : UpwardDefs) {
595 PausedSearches.push_back(Paths.size());
596 Paths.emplace_back(P.second, P.first, PriorNode);
600 /// Represents a search that terminated after finding a clobber. This clobber
601 /// may or may not be present in the path of defs from LastNode..SearchStart,
602 /// since it may have been retrieved from cache.
603 struct TerminatedPath {
604 MemoryAccess *Clobber;
605 ListIndex LastNode;
608 /// Get an access that keeps us from optimizing to the given phi.
610 /// PausedSearches is an array of indices into the Paths array. Its incoming
611 /// value is the indices of searches that stopped at the last phi optimization
612 /// target. It's left in an unspecified state.
614 /// If this returns None, NewPaused is a vector of searches that terminated
615 /// at StopWhere. Otherwise, NewPaused is left in an unspecified state.
616 Optional<TerminatedPath>
617 getBlockingAccess(const MemoryAccess *StopWhere,
618 SmallVectorImpl<ListIndex> &PausedSearches,
619 SmallVectorImpl<ListIndex> &NewPaused,
620 SmallVectorImpl<TerminatedPath> &Terminated) {
621 assert(!PausedSearches.empty() && "No searches to continue?");
623 // BFS vs DFS really doesn't make a difference here, so just do a DFS with
624 // PausedSearches as our stack.
625 while (!PausedSearches.empty()) {
626 ListIndex PathIndex = PausedSearches.pop_back_val();
627 DefPath &Node = Paths[PathIndex];
629 // If we've already visited this path with this MemoryLocation, we don't
630 // need to do so again.
632 // NOTE: That we just drop these paths on the ground makes caching
633 // behavior sporadic. e.g. given a diamond:
634 // A
635 // B C
636 // D
638 // ...If we walk D, B, A, C, we'll only cache the result of phi
639 // optimization for A, B, and D; C will be skipped because it dies here.
640 // This arguably isn't the worst thing ever, since:
641 // - We generally query things in a top-down order, so if we got below D
642 // without needing cache entries for {C, MemLoc}, then chances are
643 // that those cache entries would end up ultimately unused.
644 // - We still cache things for A, so C only needs to walk up a bit.
645 // If this behavior becomes problematic, we can fix without a ton of extra
646 // work.
647 if (!VisitedPhis.insert({Node.Last, Node.Loc}).second)
648 continue;
650 const MemoryAccess *SkipStopWhere = nullptr;
651 if (Query->SkipSelfAccess && Node.Loc == Query->StartingLoc) {
652 assert(isa<MemoryDef>(Query->OriginalAccess));
653 SkipStopWhere = Query->OriginalAccess;
656 UpwardsWalkResult Res = walkToPhiOrClobber(Node,
657 /*StopAt=*/StopWhere,
658 /*SkipStopAt=*/SkipStopWhere);
659 if (Res.IsKnownClobber) {
660 assert(Res.Result != StopWhere && Res.Result != SkipStopWhere);
662 // If this wasn't a cache hit, we hit a clobber when walking. That's a
663 // failure.
664 TerminatedPath Term{Res.Result, PathIndex};
665 if (!MSSA.dominates(Res.Result, StopWhere))
666 return Term;
668 // Otherwise, it's a valid thing to potentially optimize to.
669 Terminated.push_back(Term);
670 continue;
673 if (Res.Result == StopWhere || Res.Result == SkipStopWhere) {
674 // We've hit our target. Save this path off for if we want to continue
675 // walking. If we are in the mode of skipping the OriginalAccess, and
676 // we've reached back to the OriginalAccess, do not save path, we've
677 // just looped back to self.
678 if (Res.Result != SkipStopWhere)
679 NewPaused.push_back(PathIndex);
680 continue;
683 assert(!MSSA.isLiveOnEntryDef(Res.Result) && "liveOnEntry is a clobber");
684 addSearches(cast<MemoryPhi>(Res.Result), PausedSearches, PathIndex);
687 return None;
690 template <typename T, typename Walker>
691 struct generic_def_path_iterator
692 : public iterator_facade_base<generic_def_path_iterator<T, Walker>,
693 std::forward_iterator_tag, T *> {
694 generic_def_path_iterator() {}
695 generic_def_path_iterator(Walker *W, ListIndex N) : W(W), N(N) {}
697 T &operator*() const { return curNode(); }
699 generic_def_path_iterator &operator++() {
700 N = curNode().Previous;
701 return *this;
704 bool operator==(const generic_def_path_iterator &O) const {
705 if (N.hasValue() != O.N.hasValue())
706 return false;
707 return !N.hasValue() || *N == *O.N;
710 private:
711 T &curNode() const { return W->Paths[*N]; }
713 Walker *W = nullptr;
714 Optional<ListIndex> N = None;
717 using def_path_iterator = generic_def_path_iterator<DefPath, ClobberWalker>;
718 using const_def_path_iterator =
719 generic_def_path_iterator<const DefPath, const ClobberWalker>;
721 iterator_range<def_path_iterator> def_path(ListIndex From) {
722 return make_range(def_path_iterator(this, From), def_path_iterator());
725 iterator_range<const_def_path_iterator> const_def_path(ListIndex From) const {
726 return make_range(const_def_path_iterator(this, From),
727 const_def_path_iterator());
730 struct OptznResult {
731 /// The path that contains our result.
732 TerminatedPath PrimaryClobber;
733 /// The paths that we can legally cache back from, but that aren't
734 /// necessarily the result of the Phi optimization.
735 SmallVector<TerminatedPath, 4> OtherClobbers;
738 ListIndex defPathIndex(const DefPath &N) const {
739 // The assert looks nicer if we don't need to do &N
740 const DefPath *NP = &N;
741 assert(!Paths.empty() && NP >= &Paths.front() && NP <= &Paths.back() &&
742 "Out of bounds DefPath!");
743 return NP - &Paths.front();
746 /// Try to optimize a phi as best as we can. Returns a SmallVector of Paths
747 /// that act as legal clobbers. Note that this won't return *all* clobbers.
749 /// Phi optimization algorithm tl;dr:
750 /// - Find the earliest def/phi, A, we can optimize to
751 /// - Find if all paths from the starting memory access ultimately reach A
752 /// - If not, optimization isn't possible.
753 /// - Otherwise, walk from A to another clobber or phi, A'.
754 /// - If A' is a def, we're done.
755 /// - If A' is a phi, try to optimize it.
757 /// A path is a series of {MemoryAccess, MemoryLocation} pairs. A path
758 /// terminates when a MemoryAccess that clobbers said MemoryLocation is found.
759 OptznResult tryOptimizePhi(MemoryPhi *Phi, MemoryAccess *Start,
760 const MemoryLocation &Loc) {
761 assert(Paths.empty() && VisitedPhis.empty() &&
762 "Reset the optimization state.");
764 Paths.emplace_back(Loc, Start, Phi, None);
765 // Stores how many "valid" optimization nodes we had prior to calling
766 // addSearches/getBlockingAccess. Necessary for caching if we had a blocker.
767 auto PriorPathsSize = Paths.size();
769 SmallVector<ListIndex, 16> PausedSearches;
770 SmallVector<ListIndex, 8> NewPaused;
771 SmallVector<TerminatedPath, 4> TerminatedPaths;
773 addSearches(Phi, PausedSearches, 0);
775 // Moves the TerminatedPath with the "most dominated" Clobber to the end of
776 // Paths.
777 auto MoveDominatedPathToEnd = [&](SmallVectorImpl<TerminatedPath> &Paths) {
778 assert(!Paths.empty() && "Need a path to move");
779 auto Dom = Paths.begin();
780 for (auto I = std::next(Dom), E = Paths.end(); I != E; ++I)
781 if (!MSSA.dominates(I->Clobber, Dom->Clobber))
782 Dom = I;
783 auto Last = Paths.end() - 1;
784 if (Last != Dom)
785 std::iter_swap(Last, Dom);
788 MemoryPhi *Current = Phi;
789 while (true) {
790 assert(!MSSA.isLiveOnEntryDef(Current) &&
791 "liveOnEntry wasn't treated as a clobber?");
793 const auto *Target = getWalkTarget(Current);
794 // If a TerminatedPath doesn't dominate Target, then it wasn't a legal
795 // optimization for the prior phi.
796 assert(all_of(TerminatedPaths, [&](const TerminatedPath &P) {
797 return MSSA.dominates(P.Clobber, Target);
798 }));
800 // FIXME: This is broken, because the Blocker may be reported to be
801 // liveOnEntry, and we'll happily wait for that to disappear (read: never)
802 // For the moment, this is fine, since we do nothing with blocker info.
803 if (Optional<TerminatedPath> Blocker = getBlockingAccess(
804 Target, PausedSearches, NewPaused, TerminatedPaths)) {
806 // Find the node we started at. We can't search based on N->Last, since
807 // we may have gone around a loop with a different MemoryLocation.
808 auto Iter = find_if(def_path(Blocker->LastNode), [&](const DefPath &N) {
809 return defPathIndex(N) < PriorPathsSize;
811 assert(Iter != def_path_iterator());
813 DefPath &CurNode = *Iter;
814 assert(CurNode.Last == Current);
816 // Two things:
817 // A. We can't reliably cache all of NewPaused back. Consider a case
818 // where we have two paths in NewPaused; one of which can't optimize
819 // above this phi, whereas the other can. If we cache the second path
820 // back, we'll end up with suboptimal cache entries. We can handle
821 // cases like this a bit better when we either try to find all
822 // clobbers that block phi optimization, or when our cache starts
823 // supporting unfinished searches.
824 // B. We can't reliably cache TerminatedPaths back here without doing
825 // extra checks; consider a case like:
826 // T
827 // / \
828 // D C
829 // \ /
830 // S
831 // Where T is our target, C is a node with a clobber on it, D is a
832 // diamond (with a clobber *only* on the left or right node, N), and
833 // S is our start. Say we walk to D, through the node opposite N
834 // (read: ignoring the clobber), and see a cache entry in the top
835 // node of D. That cache entry gets put into TerminatedPaths. We then
836 // walk up to C (N is later in our worklist), find the clobber, and
837 // quit. If we append TerminatedPaths to OtherClobbers, we'll cache
838 // the bottom part of D to the cached clobber, ignoring the clobber
839 // in N. Again, this problem goes away if we start tracking all
840 // blockers for a given phi optimization.
841 TerminatedPath Result{CurNode.Last, defPathIndex(CurNode)};
842 return {Result, {}};
845 // If there's nothing left to search, then all paths led to valid clobbers
846 // that we got from our cache; pick the nearest to the start, and allow
847 // the rest to be cached back.
848 if (NewPaused.empty()) {
849 MoveDominatedPathToEnd(TerminatedPaths);
850 TerminatedPath Result = TerminatedPaths.pop_back_val();
851 return {Result, std::move(TerminatedPaths)};
854 MemoryAccess *DefChainEnd = nullptr;
855 SmallVector<TerminatedPath, 4> Clobbers;
856 for (ListIndex Paused : NewPaused) {
857 UpwardsWalkResult WR = walkToPhiOrClobber(Paths[Paused]);
858 if (WR.IsKnownClobber)
859 Clobbers.push_back({WR.Result, Paused});
860 else
861 // Micro-opt: If we hit the end of the chain, save it.
862 DefChainEnd = WR.Result;
865 if (!TerminatedPaths.empty()) {
866 // If we couldn't find the dominating phi/liveOnEntry in the above loop,
867 // do it now.
868 if (!DefChainEnd)
869 for (auto *MA : def_chain(const_cast<MemoryAccess *>(Target)))
870 DefChainEnd = MA;
872 // If any of the terminated paths don't dominate the phi we'll try to
873 // optimize, we need to figure out what they are and quit.
874 const BasicBlock *ChainBB = DefChainEnd->getBlock();
875 for (const TerminatedPath &TP : TerminatedPaths) {
876 // Because we know that DefChainEnd is as "high" as we can go, we
877 // don't need local dominance checks; BB dominance is sufficient.
878 if (DT.dominates(ChainBB, TP.Clobber->getBlock()))
879 Clobbers.push_back(TP);
883 // If we have clobbers in the def chain, find the one closest to Current
884 // and quit.
885 if (!Clobbers.empty()) {
886 MoveDominatedPathToEnd(Clobbers);
887 TerminatedPath Result = Clobbers.pop_back_val();
888 return {Result, std::move(Clobbers)};
891 assert(all_of(NewPaused,
892 [&](ListIndex I) { return Paths[I].Last == DefChainEnd; }));
894 // Because liveOnEntry is a clobber, this must be a phi.
895 auto *DefChainPhi = cast<MemoryPhi>(DefChainEnd);
897 PriorPathsSize = Paths.size();
898 PausedSearches.clear();
899 for (ListIndex I : NewPaused)
900 addSearches(DefChainPhi, PausedSearches, I);
901 NewPaused.clear();
903 Current = DefChainPhi;
907 void verifyOptResult(const OptznResult &R) const {
908 assert(all_of(R.OtherClobbers, [&](const TerminatedPath &P) {
909 return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber);
910 }));
913 void resetPhiOptznState() {
914 Paths.clear();
915 VisitedPhis.clear();
918 public:
919 ClobberWalker(const MemorySSA &MSSA, AliasAnalysisType &AA, DominatorTree &DT)
920 : MSSA(MSSA), AA(AA), DT(DT) {}
922 AliasAnalysisType *getAA() { return &AA; }
923 /// Finds the nearest clobber for the given query, optimizing phis if
924 /// possible.
925 MemoryAccess *findClobber(MemoryAccess *Start, UpwardsMemoryQuery &Q,
926 unsigned &UpWalkLimit) {
927 Query = &Q;
928 UpwardWalkLimit = &UpWalkLimit;
929 // Starting limit must be > 0.
930 if (!UpWalkLimit)
931 UpWalkLimit++;
933 MemoryAccess *Current = Start;
934 // This walker pretends uses don't exist. If we're handed one, silently grab
935 // its def. (This has the nice side-effect of ensuring we never cache uses)
936 if (auto *MU = dyn_cast<MemoryUse>(Start))
937 Current = MU->getDefiningAccess();
939 DefPath FirstDesc(Q.StartingLoc, Current, Current, None);
940 // Fast path for the overly-common case (no crazy phi optimization
941 // necessary)
942 UpwardsWalkResult WalkResult = walkToPhiOrClobber(FirstDesc);
943 MemoryAccess *Result;
944 if (WalkResult.IsKnownClobber) {
945 Result = WalkResult.Result;
946 Q.AR = WalkResult.AR;
947 } else {
948 OptznResult OptRes = tryOptimizePhi(cast<MemoryPhi>(FirstDesc.Last),
949 Current, Q.StartingLoc);
950 verifyOptResult(OptRes);
951 resetPhiOptznState();
952 Result = OptRes.PrimaryClobber.Clobber;
955 #ifdef EXPENSIVE_CHECKS
956 if (!Q.SkipSelfAccess && *UpwardWalkLimit > 0)
957 checkClobberSanity(Current, Result, Q.StartingLoc, MSSA, Q, AA);
958 #endif
959 return Result;
963 struct RenamePassData {
964 DomTreeNode *DTN;
965 DomTreeNode::const_iterator ChildIt;
966 MemoryAccess *IncomingVal;
968 RenamePassData(DomTreeNode *D, DomTreeNode::const_iterator It,
969 MemoryAccess *M)
970 : DTN(D), ChildIt(It), IncomingVal(M) {}
972 void swap(RenamePassData &RHS) {
973 std::swap(DTN, RHS.DTN);
974 std::swap(ChildIt, RHS.ChildIt);
975 std::swap(IncomingVal, RHS.IncomingVal);
979 } // end anonymous namespace
981 namespace llvm {
983 template <class AliasAnalysisType> class MemorySSA::ClobberWalkerBase {
984 ClobberWalker<AliasAnalysisType> Walker;
985 MemorySSA *MSSA;
987 public:
988 ClobberWalkerBase(MemorySSA *M, AliasAnalysisType *A, DominatorTree *D)
989 : Walker(*M, *A, *D), MSSA(M) {}
991 MemoryAccess *getClobberingMemoryAccessBase(MemoryAccess *,
992 const MemoryLocation &,
993 unsigned &);
994 // Third argument (bool), defines whether the clobber search should skip the
995 // original queried access. If true, there will be a follow-up query searching
996 // for a clobber access past "self". Note that the Optimized access is not
997 // updated if a new clobber is found by this SkipSelf search. If this
998 // additional query becomes heavily used we may decide to cache the result.
999 // Walker instantiations will decide how to set the SkipSelf bool.
1000 MemoryAccess *getClobberingMemoryAccessBase(MemoryAccess *, unsigned &, bool);
1003 /// A MemorySSAWalker that does AA walks to disambiguate accesses. It no
1004 /// longer does caching on its own, but the name has been retained for the
1005 /// moment.
1006 template <class AliasAnalysisType>
1007 class MemorySSA::CachingWalker final : public MemorySSAWalker {
1008 ClobberWalkerBase<AliasAnalysisType> *Walker;
1010 public:
1011 CachingWalker(MemorySSA *M, ClobberWalkerBase<AliasAnalysisType> *W)
1012 : MemorySSAWalker(M), Walker(W) {}
1013 ~CachingWalker() override = default;
1015 using MemorySSAWalker::getClobberingMemoryAccess;
1017 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, unsigned &UWL) {
1018 return Walker->getClobberingMemoryAccessBase(MA, UWL, false);
1020 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA,
1021 const MemoryLocation &Loc,
1022 unsigned &UWL) {
1023 return Walker->getClobberingMemoryAccessBase(MA, Loc, UWL);
1026 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA) override {
1027 unsigned UpwardWalkLimit = MaxCheckLimit;
1028 return getClobberingMemoryAccess(MA, UpwardWalkLimit);
1030 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA,
1031 const MemoryLocation &Loc) override {
1032 unsigned UpwardWalkLimit = MaxCheckLimit;
1033 return getClobberingMemoryAccess(MA, Loc, UpwardWalkLimit);
1036 void invalidateInfo(MemoryAccess *MA) override {
1037 if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
1038 MUD->resetOptimized();
1042 template <class AliasAnalysisType>
1043 class MemorySSA::SkipSelfWalker final : public MemorySSAWalker {
1044 ClobberWalkerBase<AliasAnalysisType> *Walker;
1046 public:
1047 SkipSelfWalker(MemorySSA *M, ClobberWalkerBase<AliasAnalysisType> *W)
1048 : MemorySSAWalker(M), Walker(W) {}
1049 ~SkipSelfWalker() override = default;
1051 using MemorySSAWalker::getClobberingMemoryAccess;
1053 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, unsigned &UWL) {
1054 return Walker->getClobberingMemoryAccessBase(MA, UWL, true);
1056 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA,
1057 const MemoryLocation &Loc,
1058 unsigned &UWL) {
1059 return Walker->getClobberingMemoryAccessBase(MA, Loc, UWL);
1062 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA) override {
1063 unsigned UpwardWalkLimit = MaxCheckLimit;
1064 return getClobberingMemoryAccess(MA, UpwardWalkLimit);
1066 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA,
1067 const MemoryLocation &Loc) override {
1068 unsigned UpwardWalkLimit = MaxCheckLimit;
1069 return getClobberingMemoryAccess(MA, Loc, UpwardWalkLimit);
1072 void invalidateInfo(MemoryAccess *MA) override {
1073 if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
1074 MUD->resetOptimized();
1078 } // end namespace llvm
1080 void MemorySSA::renameSuccessorPhis(BasicBlock *BB, MemoryAccess *IncomingVal,
1081 bool RenameAllUses) {
1082 // Pass through values to our successors
1083 for (const BasicBlock *S : successors(BB)) {
1084 auto It = PerBlockAccesses.find(S);
1085 // Rename the phi nodes in our successor block
1086 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
1087 continue;
1088 AccessList *Accesses = It->second.get();
1089 auto *Phi = cast<MemoryPhi>(&Accesses->front());
1090 if (RenameAllUses) {
1091 int PhiIndex = Phi->getBasicBlockIndex(BB);
1092 assert(PhiIndex != -1 && "Incomplete phi during partial rename");
1093 Phi->setIncomingValue(PhiIndex, IncomingVal);
1094 } else
1095 Phi->addIncoming(IncomingVal, BB);
1099 /// Rename a single basic block into MemorySSA form.
1100 /// Uses the standard SSA renaming algorithm.
1101 /// \returns The new incoming value.
1102 MemoryAccess *MemorySSA::renameBlock(BasicBlock *BB, MemoryAccess *IncomingVal,
1103 bool RenameAllUses) {
1104 auto It = PerBlockAccesses.find(BB);
1105 // Skip most processing if the list is empty.
1106 if (It != PerBlockAccesses.end()) {
1107 AccessList *Accesses = It->second.get();
1108 for (MemoryAccess &L : *Accesses) {
1109 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(&L)) {
1110 if (MUD->getDefiningAccess() == nullptr || RenameAllUses)
1111 MUD->setDefiningAccess(IncomingVal);
1112 if (isa<MemoryDef>(&L))
1113 IncomingVal = &L;
1114 } else {
1115 IncomingVal = &L;
1119 return IncomingVal;
1122 /// This is the standard SSA renaming algorithm.
1124 /// We walk the dominator tree in preorder, renaming accesses, and then filling
1125 /// in phi nodes in our successors.
1126 void MemorySSA::renamePass(DomTreeNode *Root, MemoryAccess *IncomingVal,
1127 SmallPtrSetImpl<BasicBlock *> &Visited,
1128 bool SkipVisited, bool RenameAllUses) {
1129 assert(Root && "Trying to rename accesses in an unreachable block");
1131 SmallVector<RenamePassData, 32> WorkStack;
1132 // Skip everything if we already renamed this block and we are skipping.
1133 // Note: You can't sink this into the if, because we need it to occur
1134 // regardless of whether we skip blocks or not.
1135 bool AlreadyVisited = !Visited.insert(Root->getBlock()).second;
1136 if (SkipVisited && AlreadyVisited)
1137 return;
1139 IncomingVal = renameBlock(Root->getBlock(), IncomingVal, RenameAllUses);
1140 renameSuccessorPhis(Root->getBlock(), IncomingVal, RenameAllUses);
1141 WorkStack.push_back({Root, Root->begin(), IncomingVal});
1143 while (!WorkStack.empty()) {
1144 DomTreeNode *Node = WorkStack.back().DTN;
1145 DomTreeNode::const_iterator ChildIt = WorkStack.back().ChildIt;
1146 IncomingVal = WorkStack.back().IncomingVal;
1148 if (ChildIt == Node->end()) {
1149 WorkStack.pop_back();
1150 } else {
1151 DomTreeNode *Child = *ChildIt;
1152 ++WorkStack.back().ChildIt;
1153 BasicBlock *BB = Child->getBlock();
1154 // Note: You can't sink this into the if, because we need it to occur
1155 // regardless of whether we skip blocks or not.
1156 AlreadyVisited = !Visited.insert(BB).second;
1157 if (SkipVisited && AlreadyVisited) {
1158 // We already visited this during our renaming, which can happen when
1159 // being asked to rename multiple blocks. Figure out the incoming val,
1160 // which is the last def.
1161 // Incoming value can only change if there is a block def, and in that
1162 // case, it's the last block def in the list.
1163 if (auto *BlockDefs = getWritableBlockDefs(BB))
1164 IncomingVal = &*BlockDefs->rbegin();
1165 } else
1166 IncomingVal = renameBlock(BB, IncomingVal, RenameAllUses);
1167 renameSuccessorPhis(BB, IncomingVal, RenameAllUses);
1168 WorkStack.push_back({Child, Child->begin(), IncomingVal});
1173 /// This handles unreachable block accesses by deleting phi nodes in
1174 /// unreachable blocks, and marking all other unreachable MemoryAccess's as
1175 /// being uses of the live on entry definition.
1176 void MemorySSA::markUnreachableAsLiveOnEntry(BasicBlock *BB) {
1177 assert(!DT->isReachableFromEntry(BB) &&
1178 "Reachable block found while handling unreachable blocks");
1180 // Make sure phi nodes in our reachable successors end up with a
1181 // LiveOnEntryDef for our incoming edge, even though our block is forward
1182 // unreachable. We could just disconnect these blocks from the CFG fully,
1183 // but we do not right now.
1184 for (const BasicBlock *S : successors(BB)) {
1185 if (!DT->isReachableFromEntry(S))
1186 continue;
1187 auto It = PerBlockAccesses.find(S);
1188 // Rename the phi nodes in our successor block
1189 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
1190 continue;
1191 AccessList *Accesses = It->second.get();
1192 auto *Phi = cast<MemoryPhi>(&Accesses->front());
1193 Phi->addIncoming(LiveOnEntryDef.get(), BB);
1196 auto It = PerBlockAccesses.find(BB);
1197 if (It == PerBlockAccesses.end())
1198 return;
1200 auto &Accesses = It->second;
1201 for (auto AI = Accesses->begin(), AE = Accesses->end(); AI != AE;) {
1202 auto Next = std::next(AI);
1203 // If we have a phi, just remove it. We are going to replace all
1204 // users with live on entry.
1205 if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(AI))
1206 UseOrDef->setDefiningAccess(LiveOnEntryDef.get());
1207 else
1208 Accesses->erase(AI);
1209 AI = Next;
1213 MemorySSA::MemorySSA(Function &Func, AliasAnalysis *AA, DominatorTree *DT)
1214 : AA(nullptr), DT(DT), F(Func), LiveOnEntryDef(nullptr), Walker(nullptr),
1215 SkipWalker(nullptr), NextID(0) {
1216 // Build MemorySSA using a batch alias analysis. This reuses the internal
1217 // state that AA collects during an alias()/getModRefInfo() call. This is
1218 // safe because there are no CFG changes while building MemorySSA and can
1219 // significantly reduce the time spent by the compiler in AA, because we will
1220 // make queries about all the instructions in the Function.
1221 BatchAAResults BatchAA(*AA);
1222 buildMemorySSA(BatchAA);
1223 // Intentionally leave AA to nullptr while building so we don't accidently
1224 // use non-batch AliasAnalysis.
1225 this->AA = AA;
1226 // Also create the walker here.
1227 getWalker();
1230 MemorySSA::~MemorySSA() {
1231 // Drop all our references
1232 for (const auto &Pair : PerBlockAccesses)
1233 for (MemoryAccess &MA : *Pair.second)
1234 MA.dropAllReferences();
1237 MemorySSA::AccessList *MemorySSA::getOrCreateAccessList(const BasicBlock *BB) {
1238 auto Res = PerBlockAccesses.insert(std::make_pair(BB, nullptr));
1240 if (Res.second)
1241 Res.first->second = std::make_unique<AccessList>();
1242 return Res.first->second.get();
1245 MemorySSA::DefsList *MemorySSA::getOrCreateDefsList(const BasicBlock *BB) {
1246 auto Res = PerBlockDefs.insert(std::make_pair(BB, nullptr));
1248 if (Res.second)
1249 Res.first->second = std::make_unique<DefsList>();
1250 return Res.first->second.get();
1253 namespace llvm {
1255 /// This class is a batch walker of all MemoryUse's in the program, and points
1256 /// their defining access at the thing that actually clobbers them. Because it
1257 /// is a batch walker that touches everything, it does not operate like the
1258 /// other walkers. This walker is basically performing a top-down SSA renaming
1259 /// pass, where the version stack is used as the cache. This enables it to be
1260 /// significantly more time and memory efficient than using the regular walker,
1261 /// which is walking bottom-up.
1262 class MemorySSA::OptimizeUses {
1263 public:
1264 OptimizeUses(MemorySSA *MSSA, CachingWalker<BatchAAResults> *Walker,
1265 BatchAAResults *BAA, DominatorTree *DT)
1266 : MSSA(MSSA), Walker(Walker), AA(BAA), DT(DT) {}
1268 void optimizeUses();
1270 private:
1271 /// This represents where a given memorylocation is in the stack.
1272 struct MemlocStackInfo {
1273 // This essentially is keeping track of versions of the stack. Whenever
1274 // the stack changes due to pushes or pops, these versions increase.
1275 unsigned long StackEpoch;
1276 unsigned long PopEpoch;
1277 // This is the lower bound of places on the stack to check. It is equal to
1278 // the place the last stack walk ended.
1279 // Note: Correctness depends on this being initialized to 0, which densemap
1280 // does
1281 unsigned long LowerBound;
1282 const BasicBlock *LowerBoundBlock;
1283 // This is where the last walk for this memory location ended.
1284 unsigned long LastKill;
1285 bool LastKillValid;
1286 Optional<AliasResult> AR;
1289 void optimizeUsesInBlock(const BasicBlock *, unsigned long &, unsigned long &,
1290 SmallVectorImpl<MemoryAccess *> &,
1291 DenseMap<MemoryLocOrCall, MemlocStackInfo> &);
1293 MemorySSA *MSSA;
1294 CachingWalker<BatchAAResults> *Walker;
1295 BatchAAResults *AA;
1296 DominatorTree *DT;
1299 } // end namespace llvm
1301 /// Optimize the uses in a given block This is basically the SSA renaming
1302 /// algorithm, with one caveat: We are able to use a single stack for all
1303 /// MemoryUses. This is because the set of *possible* reaching MemoryDefs is
1304 /// the same for every MemoryUse. The *actual* clobbering MemoryDef is just
1305 /// going to be some position in that stack of possible ones.
1307 /// We track the stack positions that each MemoryLocation needs
1308 /// to check, and last ended at. This is because we only want to check the
1309 /// things that changed since last time. The same MemoryLocation should
1310 /// get clobbered by the same store (getModRefInfo does not use invariantness or
1311 /// things like this, and if they start, we can modify MemoryLocOrCall to
1312 /// include relevant data)
1313 void MemorySSA::OptimizeUses::optimizeUsesInBlock(
1314 const BasicBlock *BB, unsigned long &StackEpoch, unsigned long &PopEpoch,
1315 SmallVectorImpl<MemoryAccess *> &VersionStack,
1316 DenseMap<MemoryLocOrCall, MemlocStackInfo> &LocStackInfo) {
1318 /// If no accesses, nothing to do.
1319 MemorySSA::AccessList *Accesses = MSSA->getWritableBlockAccesses(BB);
1320 if (Accesses == nullptr)
1321 return;
1323 // Pop everything that doesn't dominate the current block off the stack,
1324 // increment the PopEpoch to account for this.
1325 while (true) {
1326 assert(
1327 !VersionStack.empty() &&
1328 "Version stack should have liveOnEntry sentinel dominating everything");
1329 BasicBlock *BackBlock = VersionStack.back()->getBlock();
1330 if (DT->dominates(BackBlock, BB))
1331 break;
1332 while (VersionStack.back()->getBlock() == BackBlock)
1333 VersionStack.pop_back();
1334 ++PopEpoch;
1337 for (MemoryAccess &MA : *Accesses) {
1338 auto *MU = dyn_cast<MemoryUse>(&MA);
1339 if (!MU) {
1340 VersionStack.push_back(&MA);
1341 ++StackEpoch;
1342 continue;
1345 if (isUseTriviallyOptimizableToLiveOnEntry(*AA, MU->getMemoryInst())) {
1346 MU->setDefiningAccess(MSSA->getLiveOnEntryDef(), true, None);
1347 continue;
1350 MemoryLocOrCall UseMLOC(MU);
1351 auto &LocInfo = LocStackInfo[UseMLOC];
1352 // If the pop epoch changed, it means we've removed stuff from top of
1353 // stack due to changing blocks. We may have to reset the lower bound or
1354 // last kill info.
1355 if (LocInfo.PopEpoch != PopEpoch) {
1356 LocInfo.PopEpoch = PopEpoch;
1357 LocInfo.StackEpoch = StackEpoch;
1358 // If the lower bound was in something that no longer dominates us, we
1359 // have to reset it.
1360 // We can't simply track stack size, because the stack may have had
1361 // pushes/pops in the meantime.
1362 // XXX: This is non-optimal, but only is slower cases with heavily
1363 // branching dominator trees. To get the optimal number of queries would
1364 // be to make lowerbound and lastkill a per-loc stack, and pop it until
1365 // the top of that stack dominates us. This does not seem worth it ATM.
1366 // A much cheaper optimization would be to always explore the deepest
1367 // branch of the dominator tree first. This will guarantee this resets on
1368 // the smallest set of blocks.
1369 if (LocInfo.LowerBoundBlock && LocInfo.LowerBoundBlock != BB &&
1370 !DT->dominates(LocInfo.LowerBoundBlock, BB)) {
1371 // Reset the lower bound of things to check.
1372 // TODO: Some day we should be able to reset to last kill, rather than
1373 // 0.
1374 LocInfo.LowerBound = 0;
1375 LocInfo.LowerBoundBlock = VersionStack[0]->getBlock();
1376 LocInfo.LastKillValid = false;
1378 } else if (LocInfo.StackEpoch != StackEpoch) {
1379 // If all that has changed is the StackEpoch, we only have to check the
1380 // new things on the stack, because we've checked everything before. In
1381 // this case, the lower bound of things to check remains the same.
1382 LocInfo.PopEpoch = PopEpoch;
1383 LocInfo.StackEpoch = StackEpoch;
1385 if (!LocInfo.LastKillValid) {
1386 LocInfo.LastKill = VersionStack.size() - 1;
1387 LocInfo.LastKillValid = true;
1388 LocInfo.AR = MayAlias;
1391 // At this point, we should have corrected last kill and LowerBound to be
1392 // in bounds.
1393 assert(LocInfo.LowerBound < VersionStack.size() &&
1394 "Lower bound out of range");
1395 assert(LocInfo.LastKill < VersionStack.size() &&
1396 "Last kill info out of range");
1397 // In any case, the new upper bound is the top of the stack.
1398 unsigned long UpperBound = VersionStack.size() - 1;
1400 if (UpperBound - LocInfo.LowerBound > MaxCheckLimit) {
1401 LLVM_DEBUG(dbgs() << "MemorySSA skipping optimization of " << *MU << " ("
1402 << *(MU->getMemoryInst()) << ")"
1403 << " because there are "
1404 << UpperBound - LocInfo.LowerBound
1405 << " stores to disambiguate\n");
1406 // Because we did not walk, LastKill is no longer valid, as this may
1407 // have been a kill.
1408 LocInfo.LastKillValid = false;
1409 continue;
1411 bool FoundClobberResult = false;
1412 unsigned UpwardWalkLimit = MaxCheckLimit;
1413 while (UpperBound > LocInfo.LowerBound) {
1414 if (isa<MemoryPhi>(VersionStack[UpperBound])) {
1415 // For phis, use the walker, see where we ended up, go there
1416 MemoryAccess *Result =
1417 Walker->getClobberingMemoryAccess(MU, UpwardWalkLimit);
1418 // We are guaranteed to find it or something is wrong
1419 while (VersionStack[UpperBound] != Result) {
1420 assert(UpperBound != 0);
1421 --UpperBound;
1423 FoundClobberResult = true;
1424 break;
1427 MemoryDef *MD = cast<MemoryDef>(VersionStack[UpperBound]);
1428 // If the lifetime of the pointer ends at this instruction, it's live on
1429 // entry.
1430 if (!UseMLOC.IsCall && lifetimeEndsAt(MD, UseMLOC.getLoc(), *AA)) {
1431 // Reset UpperBound to liveOnEntryDef's place in the stack
1432 UpperBound = 0;
1433 FoundClobberResult = true;
1434 LocInfo.AR = MustAlias;
1435 break;
1437 ClobberAlias CA = instructionClobbersQuery(MD, MU, UseMLOC, *AA);
1438 if (CA.IsClobber) {
1439 FoundClobberResult = true;
1440 LocInfo.AR = CA.AR;
1441 break;
1443 --UpperBound;
1446 // Note: Phis always have AliasResult AR set to MayAlias ATM.
1448 // At the end of this loop, UpperBound is either a clobber, or lower bound
1449 // PHI walking may cause it to be < LowerBound, and in fact, < LastKill.
1450 if (FoundClobberResult || UpperBound < LocInfo.LastKill) {
1451 // We were last killed now by where we got to
1452 if (MSSA->isLiveOnEntryDef(VersionStack[UpperBound]))
1453 LocInfo.AR = None;
1454 MU->setDefiningAccess(VersionStack[UpperBound], true, LocInfo.AR);
1455 LocInfo.LastKill = UpperBound;
1456 } else {
1457 // Otherwise, we checked all the new ones, and now we know we can get to
1458 // LastKill.
1459 MU->setDefiningAccess(VersionStack[LocInfo.LastKill], true, LocInfo.AR);
1461 LocInfo.LowerBound = VersionStack.size() - 1;
1462 LocInfo.LowerBoundBlock = BB;
1466 /// Optimize uses to point to their actual clobbering definitions.
1467 void MemorySSA::OptimizeUses::optimizeUses() {
1468 SmallVector<MemoryAccess *, 16> VersionStack;
1469 DenseMap<MemoryLocOrCall, MemlocStackInfo> LocStackInfo;
1470 VersionStack.push_back(MSSA->getLiveOnEntryDef());
1472 unsigned long StackEpoch = 1;
1473 unsigned long PopEpoch = 1;
1474 // We perform a non-recursive top-down dominator tree walk.
1475 for (const auto *DomNode : depth_first(DT->getRootNode()))
1476 optimizeUsesInBlock(DomNode->getBlock(), StackEpoch, PopEpoch, VersionStack,
1477 LocStackInfo);
1480 void MemorySSA::placePHINodes(
1481 const SmallPtrSetImpl<BasicBlock *> &DefiningBlocks) {
1482 // Determine where our MemoryPhi's should go
1483 ForwardIDFCalculator IDFs(*DT);
1484 IDFs.setDefiningBlocks(DefiningBlocks);
1485 SmallVector<BasicBlock *, 32> IDFBlocks;
1486 IDFs.calculate(IDFBlocks);
1488 // Now place MemoryPhi nodes.
1489 for (auto &BB : IDFBlocks)
1490 createMemoryPhi(BB);
1493 void MemorySSA::buildMemorySSA(BatchAAResults &BAA) {
1494 // We create an access to represent "live on entry", for things like
1495 // arguments or users of globals, where the memory they use is defined before
1496 // the beginning of the function. We do not actually insert it into the IR.
1497 // We do not define a live on exit for the immediate uses, and thus our
1498 // semantics do *not* imply that something with no immediate uses can simply
1499 // be removed.
1500 BasicBlock &StartingPoint = F.getEntryBlock();
1501 LiveOnEntryDef.reset(new MemoryDef(F.getContext(), nullptr, nullptr,
1502 &StartingPoint, NextID++));
1504 // We maintain lists of memory accesses per-block, trading memory for time. We
1505 // could just look up the memory access for every possible instruction in the
1506 // stream.
1507 SmallPtrSet<BasicBlock *, 32> DefiningBlocks;
1508 // Go through each block, figure out where defs occur, and chain together all
1509 // the accesses.
1510 for (BasicBlock &B : F) {
1511 bool InsertIntoDef = false;
1512 AccessList *Accesses = nullptr;
1513 DefsList *Defs = nullptr;
1514 for (Instruction &I : B) {
1515 MemoryUseOrDef *MUD = createNewAccess(&I, &BAA);
1516 if (!MUD)
1517 continue;
1519 if (!Accesses)
1520 Accesses = getOrCreateAccessList(&B);
1521 Accesses->push_back(MUD);
1522 if (isa<MemoryDef>(MUD)) {
1523 InsertIntoDef = true;
1524 if (!Defs)
1525 Defs = getOrCreateDefsList(&B);
1526 Defs->push_back(*MUD);
1529 if (InsertIntoDef)
1530 DefiningBlocks.insert(&B);
1532 placePHINodes(DefiningBlocks);
1534 // Now do regular SSA renaming on the MemoryDef/MemoryUse. Visited will get
1535 // filled in with all blocks.
1536 SmallPtrSet<BasicBlock *, 16> Visited;
1537 renamePass(DT->getRootNode(), LiveOnEntryDef.get(), Visited);
1539 ClobberWalkerBase<BatchAAResults> WalkerBase(this, &BAA, DT);
1540 CachingWalker<BatchAAResults> WalkerLocal(this, &WalkerBase);
1541 OptimizeUses(this, &WalkerLocal, &BAA, DT).optimizeUses();
1543 // Mark the uses in unreachable blocks as live on entry, so that they go
1544 // somewhere.
1545 for (auto &BB : F)
1546 if (!Visited.count(&BB))
1547 markUnreachableAsLiveOnEntry(&BB);
1550 MemorySSAWalker *MemorySSA::getWalker() { return getWalkerImpl(); }
1552 MemorySSA::CachingWalker<AliasAnalysis> *MemorySSA::getWalkerImpl() {
1553 if (Walker)
1554 return Walker.get();
1556 if (!WalkerBase)
1557 WalkerBase =
1558 std::make_unique<ClobberWalkerBase<AliasAnalysis>>(this, AA, DT);
1560 Walker =
1561 std::make_unique<CachingWalker<AliasAnalysis>>(this, WalkerBase.get());
1562 return Walker.get();
1565 MemorySSAWalker *MemorySSA::getSkipSelfWalker() {
1566 if (SkipWalker)
1567 return SkipWalker.get();
1569 if (!WalkerBase)
1570 WalkerBase =
1571 std::make_unique<ClobberWalkerBase<AliasAnalysis>>(this, AA, DT);
1573 SkipWalker =
1574 std::make_unique<SkipSelfWalker<AliasAnalysis>>(this, WalkerBase.get());
1575 return SkipWalker.get();
1579 // This is a helper function used by the creation routines. It places NewAccess
1580 // into the access and defs lists for a given basic block, at the given
1581 // insertion point.
1582 void MemorySSA::insertIntoListsForBlock(MemoryAccess *NewAccess,
1583 const BasicBlock *BB,
1584 InsertionPlace Point) {
1585 auto *Accesses = getOrCreateAccessList(BB);
1586 if (Point == Beginning) {
1587 // If it's a phi node, it goes first, otherwise, it goes after any phi
1588 // nodes.
1589 if (isa<MemoryPhi>(NewAccess)) {
1590 Accesses->push_front(NewAccess);
1591 auto *Defs = getOrCreateDefsList(BB);
1592 Defs->push_front(*NewAccess);
1593 } else {
1594 auto AI = find_if_not(
1595 *Accesses, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
1596 Accesses->insert(AI, NewAccess);
1597 if (!isa<MemoryUse>(NewAccess)) {
1598 auto *Defs = getOrCreateDefsList(BB);
1599 auto DI = find_if_not(
1600 *Defs, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
1601 Defs->insert(DI, *NewAccess);
1604 } else {
1605 Accesses->push_back(NewAccess);
1606 if (!isa<MemoryUse>(NewAccess)) {
1607 auto *Defs = getOrCreateDefsList(BB);
1608 Defs->push_back(*NewAccess);
1611 BlockNumberingValid.erase(BB);
1614 void MemorySSA::insertIntoListsBefore(MemoryAccess *What, const BasicBlock *BB,
1615 AccessList::iterator InsertPt) {
1616 auto *Accesses = getWritableBlockAccesses(BB);
1617 bool WasEnd = InsertPt == Accesses->end();
1618 Accesses->insert(AccessList::iterator(InsertPt), What);
1619 if (!isa<MemoryUse>(What)) {
1620 auto *Defs = getOrCreateDefsList(BB);
1621 // If we got asked to insert at the end, we have an easy job, just shove it
1622 // at the end. If we got asked to insert before an existing def, we also get
1623 // an iterator. If we got asked to insert before a use, we have to hunt for
1624 // the next def.
1625 if (WasEnd) {
1626 Defs->push_back(*What);
1627 } else if (isa<MemoryDef>(InsertPt)) {
1628 Defs->insert(InsertPt->getDefsIterator(), *What);
1629 } else {
1630 while (InsertPt != Accesses->end() && !isa<MemoryDef>(InsertPt))
1631 ++InsertPt;
1632 // Either we found a def, or we are inserting at the end
1633 if (InsertPt == Accesses->end())
1634 Defs->push_back(*What);
1635 else
1636 Defs->insert(InsertPt->getDefsIterator(), *What);
1639 BlockNumberingValid.erase(BB);
1642 void MemorySSA::prepareForMoveTo(MemoryAccess *What, BasicBlock *BB) {
1643 // Keep it in the lookup tables, remove from the lists
1644 removeFromLists(What, false);
1646 // Note that moving should implicitly invalidate the optimized state of a
1647 // MemoryUse (and Phis can't be optimized). However, it doesn't do so for a
1648 // MemoryDef.
1649 if (auto *MD = dyn_cast<MemoryDef>(What))
1650 MD->resetOptimized();
1651 What->setBlock(BB);
1654 // Move What before Where in the IR. The end result is that What will belong to
1655 // the right lists and have the right Block set, but will not otherwise be
1656 // correct. It will not have the right defining access, and if it is a def,
1657 // things below it will not properly be updated.
1658 void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB,
1659 AccessList::iterator Where) {
1660 prepareForMoveTo(What, BB);
1661 insertIntoListsBefore(What, BB, Where);
1664 void MemorySSA::moveTo(MemoryAccess *What, BasicBlock *BB,
1665 InsertionPlace Point) {
1666 if (isa<MemoryPhi>(What)) {
1667 assert(Point == Beginning &&
1668 "Can only move a Phi at the beginning of the block");
1669 // Update lookup table entry
1670 ValueToMemoryAccess.erase(What->getBlock());
1671 bool Inserted = ValueToMemoryAccess.insert({BB, What}).second;
1672 (void)Inserted;
1673 assert(Inserted && "Cannot move a Phi to a block that already has one");
1676 prepareForMoveTo(What, BB);
1677 insertIntoListsForBlock(What, BB, Point);
1680 MemoryPhi *MemorySSA::createMemoryPhi(BasicBlock *BB) {
1681 assert(!getMemoryAccess(BB) && "MemoryPhi already exists for this BB");
1682 MemoryPhi *Phi = new MemoryPhi(BB->getContext(), BB, NextID++);
1683 // Phi's always are placed at the front of the block.
1684 insertIntoListsForBlock(Phi, BB, Beginning);
1685 ValueToMemoryAccess[BB] = Phi;
1686 return Phi;
1689 MemoryUseOrDef *MemorySSA::createDefinedAccess(Instruction *I,
1690 MemoryAccess *Definition,
1691 const MemoryUseOrDef *Template,
1692 bool CreationMustSucceed) {
1693 assert(!isa<PHINode>(I) && "Cannot create a defined access for a PHI");
1694 MemoryUseOrDef *NewAccess = createNewAccess(I, AA, Template);
1695 if (CreationMustSucceed)
1696 assert(NewAccess != nullptr && "Tried to create a memory access for a "
1697 "non-memory touching instruction");
1698 if (NewAccess)
1699 NewAccess->setDefiningAccess(Definition);
1700 return NewAccess;
1703 // Return true if the instruction has ordering constraints.
1704 // Note specifically that this only considers stores and loads
1705 // because others are still considered ModRef by getModRefInfo.
1706 static inline bool isOrdered(const Instruction *I) {
1707 if (auto *SI = dyn_cast<StoreInst>(I)) {
1708 if (!SI->isUnordered())
1709 return true;
1710 } else if (auto *LI = dyn_cast<LoadInst>(I)) {
1711 if (!LI->isUnordered())
1712 return true;
1714 return false;
1717 /// Helper function to create new memory accesses
1718 template <typename AliasAnalysisType>
1719 MemoryUseOrDef *MemorySSA::createNewAccess(Instruction *I,
1720 AliasAnalysisType *AAP,
1721 const MemoryUseOrDef *Template) {
1722 // The assume intrinsic has a control dependency which we model by claiming
1723 // that it writes arbitrarily. Ignore that fake memory dependency 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>(I))
1727 if (II->getIntrinsicID() == Intrinsic::assume)
1728 return nullptr;
1730 bool Def, Use;
1731 if (Template) {
1732 Def = dyn_cast_or_null<MemoryDef>(Template) != nullptr;
1733 Use = dyn_cast_or_null<MemoryUse>(Template) != nullptr;
1734 #if !defined(NDEBUG)
1735 ModRefInfo ModRef = AAP->getModRefInfo(I, None);
1736 bool DefCheck, UseCheck;
1737 DefCheck = isModSet(ModRef) || isOrdered(I);
1738 UseCheck = isRefSet(ModRef);
1739 assert(Def == DefCheck && (Def || Use == UseCheck) && "Invalid template");
1740 #endif
1741 } else {
1742 // Find out what affect this instruction has on memory.
1743 ModRefInfo ModRef = AAP->getModRefInfo(I, None);
1744 // The isOrdered check is used to ensure that volatiles end up as defs
1745 // (atomics end up as ModRef right now anyway). Until we separate the
1746 // ordering chain from the memory chain, this enables people to see at least
1747 // some relative ordering to volatiles. Note that getClobberingMemoryAccess
1748 // will still give an answer that bypasses other volatile loads. TODO:
1749 // Separate memory aliasing and ordering into two different chains so that
1750 // we can precisely represent both "what memory will this read/write/is
1751 // clobbered by" and "what instructions can I move this past".
1752 Def = isModSet(ModRef) || isOrdered(I);
1753 Use = isRefSet(ModRef);
1756 // It's possible for an instruction to not modify memory at all. During
1757 // construction, we ignore them.
1758 if (!Def && !Use)
1759 return nullptr;
1761 MemoryUseOrDef *MUD;
1762 if (Def)
1763 MUD = new MemoryDef(I->getContext(), nullptr, I, I->getParent(), NextID++);
1764 else
1765 MUD = new MemoryUse(I->getContext(), nullptr, I, I->getParent());
1766 ValueToMemoryAccess[I] = MUD;
1767 return MUD;
1770 /// Returns true if \p Replacer dominates \p Replacee .
1771 bool MemorySSA::dominatesUse(const MemoryAccess *Replacer,
1772 const MemoryAccess *Replacee) const {
1773 if (isa<MemoryUseOrDef>(Replacee))
1774 return DT->dominates(Replacer->getBlock(), Replacee->getBlock());
1775 const auto *MP = cast<MemoryPhi>(Replacee);
1776 // For a phi node, the use occurs in the predecessor block of the phi node.
1777 // Since we may occur multiple times in the phi node, we have to check each
1778 // operand to ensure Replacer dominates each operand where Replacee occurs.
1779 for (const Use &Arg : MP->operands()) {
1780 if (Arg.get() != Replacee &&
1781 !DT->dominates(Replacer->getBlock(), MP->getIncomingBlock(Arg)))
1782 return false;
1784 return true;
1787 /// Properly remove \p MA from all of MemorySSA's lookup tables.
1788 void MemorySSA::removeFromLookups(MemoryAccess *MA) {
1789 assert(MA->use_empty() &&
1790 "Trying to remove memory access that still has uses");
1791 BlockNumbering.erase(MA);
1792 if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
1793 MUD->setDefiningAccess(nullptr);
1794 // Invalidate our walker's cache if necessary
1795 if (!isa<MemoryUse>(MA))
1796 getWalker()->invalidateInfo(MA);
1798 Value *MemoryInst;
1799 if (const auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
1800 MemoryInst = MUD->getMemoryInst();
1801 else
1802 MemoryInst = MA->getBlock();
1804 auto VMA = ValueToMemoryAccess.find(MemoryInst);
1805 if (VMA->second == MA)
1806 ValueToMemoryAccess.erase(VMA);
1809 /// Properly remove \p MA from all of MemorySSA's lists.
1811 /// Because of the way the intrusive list and use lists work, it is important to
1812 /// do removal in the right order.
1813 /// ShouldDelete defaults to true, and will cause the memory access to also be
1814 /// deleted, not just removed.
1815 void MemorySSA::removeFromLists(MemoryAccess *MA, bool ShouldDelete) {
1816 BasicBlock *BB = MA->getBlock();
1817 // The access list owns the reference, so we erase it from the non-owning list
1818 // first.
1819 if (!isa<MemoryUse>(MA)) {
1820 auto DefsIt = PerBlockDefs.find(BB);
1821 std::unique_ptr<DefsList> &Defs = DefsIt->second;
1822 Defs->remove(*MA);
1823 if (Defs->empty())
1824 PerBlockDefs.erase(DefsIt);
1827 // The erase call here will delete it. If we don't want it deleted, we call
1828 // remove instead.
1829 auto AccessIt = PerBlockAccesses.find(BB);
1830 std::unique_ptr<AccessList> &Accesses = AccessIt->second;
1831 if (ShouldDelete)
1832 Accesses->erase(MA);
1833 else
1834 Accesses->remove(MA);
1836 if (Accesses->empty()) {
1837 PerBlockAccesses.erase(AccessIt);
1838 BlockNumberingValid.erase(BB);
1842 void MemorySSA::print(raw_ostream &OS) const {
1843 MemorySSAAnnotatedWriter Writer(this);
1844 F.print(OS, &Writer);
1847 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1848 LLVM_DUMP_METHOD void MemorySSA::dump() const { print(dbgs()); }
1849 #endif
1851 void MemorySSA::verifyMemorySSA() const {
1852 verifyDefUses(F);
1853 verifyDomination(F);
1854 verifyOrdering(F);
1855 verifyDominationNumbers(F);
1856 verifyPrevDefInPhis(F);
1857 // Previously, the verification used to also verify that the clobberingAccess
1858 // cached by MemorySSA is the same as the clobberingAccess found at a later
1859 // query to AA. This does not hold true in general due to the current fragility
1860 // of BasicAA which has arbitrary caps on the things it analyzes before giving
1861 // up. As a result, transformations that are correct, will lead to BasicAA
1862 // returning different Alias answers before and after that transformation.
1863 // Invalidating MemorySSA is not an option, as the results in BasicAA can be so
1864 // random, in the worst case we'd need to rebuild MemorySSA from scratch after
1865 // every transformation, which defeats the purpose of using it. For such an
1866 // example, see test4 added in D51960.
1869 void MemorySSA::verifyPrevDefInPhis(Function &F) const {
1870 #ifndef NDEBUG
1871 for (const BasicBlock &BB : F) {
1872 if (MemoryPhi *Phi = getMemoryAccess(&BB)) {
1873 for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) {
1874 auto *Pred = Phi->getIncomingBlock(I);
1875 auto *IncAcc = Phi->getIncomingValue(I);
1876 // If Pred has no unreachable predecessors, get last def looking at
1877 // IDoms. If, while walkings IDoms, any of these has an unreachable
1878 // predecessor, then the expected incoming def is LoE.
1879 if (auto *DTNode = DT->getNode(Pred)) {
1880 while (DTNode) {
1881 if (auto *DefList = getBlockDefs(DTNode->getBlock())) {
1882 auto *LastAcc = &*(--DefList->end());
1883 assert(LastAcc == IncAcc &&
1884 "Incorrect incoming access into phi.");
1885 break;
1887 DTNode = DTNode->getIDom();
1889 } else if (auto *DefList = getBlockDefs(Pred)) {
1890 // If Pred has unreachable predecessors, but has at least a Def, the
1891 // incoming access can be the last Def in Pred, or it could have been
1892 // optimized to LoE.
1893 auto *LastAcc = &*(--DefList->end());
1894 assert((LastAcc == IncAcc || IncAcc == getLiveOnEntryDef()) &&
1895 "Incorrect incoming access into phi.");
1896 } else {
1897 // If Pred has unreachable predecessors and no Defs, incoming access
1898 // should be LoE; In practice, after an update, it may be any access.
1903 #endif
1906 /// Verify that all of the blocks we believe to have valid domination numbers
1907 /// actually have valid domination numbers.
1908 void MemorySSA::verifyDominationNumbers(const Function &F) const {
1909 #ifndef NDEBUG
1910 if (BlockNumberingValid.empty())
1911 return;
1913 SmallPtrSet<const BasicBlock *, 16> ValidBlocks = BlockNumberingValid;
1914 for (const BasicBlock &BB : F) {
1915 if (!ValidBlocks.count(&BB))
1916 continue;
1918 ValidBlocks.erase(&BB);
1920 const AccessList *Accesses = getBlockAccesses(&BB);
1921 // It's correct to say an empty block has valid numbering.
1922 if (!Accesses)
1923 continue;
1925 // Block numbering starts at 1.
1926 unsigned long LastNumber = 0;
1927 for (const MemoryAccess &MA : *Accesses) {
1928 auto ThisNumberIter = BlockNumbering.find(&MA);
1929 assert(ThisNumberIter != BlockNumbering.end() &&
1930 "MemoryAccess has no domination number in a valid block!");
1932 unsigned long ThisNumber = ThisNumberIter->second;
1933 assert(ThisNumber > LastNumber &&
1934 "Domination numbers should be strictly increasing!");
1935 LastNumber = ThisNumber;
1939 assert(ValidBlocks.empty() &&
1940 "All valid BasicBlocks should exist in F -- dangling pointers?");
1941 #endif
1944 /// Verify that the order and existence of MemoryAccesses matches the
1945 /// order and existence of memory affecting instructions.
1946 void MemorySSA::verifyOrdering(Function &F) const {
1947 #ifndef NDEBUG
1948 // Walk all the blocks, comparing what the lookups think and what the access
1949 // lists think, as well as the order in the blocks vs the order in the access
1950 // lists.
1951 SmallVector<MemoryAccess *, 32> ActualAccesses;
1952 SmallVector<MemoryAccess *, 32> ActualDefs;
1953 for (BasicBlock &B : F) {
1954 const AccessList *AL = getBlockAccesses(&B);
1955 const auto *DL = getBlockDefs(&B);
1956 MemoryAccess *Phi = getMemoryAccess(&B);
1957 if (Phi) {
1958 ActualAccesses.push_back(Phi);
1959 ActualDefs.push_back(Phi);
1962 for (Instruction &I : B) {
1963 MemoryAccess *MA = getMemoryAccess(&I);
1964 assert((!MA || (AL && (isa<MemoryUse>(MA) || DL))) &&
1965 "We have memory affecting instructions "
1966 "in this block but they are not in the "
1967 "access list or defs list");
1968 if (MA) {
1969 ActualAccesses.push_back(MA);
1970 if (isa<MemoryDef>(MA))
1971 ActualDefs.push_back(MA);
1974 // Either we hit the assert, really have no accesses, or we have both
1975 // accesses and an access list.
1976 // Same with defs.
1977 if (!AL && !DL)
1978 continue;
1979 assert(AL->size() == ActualAccesses.size() &&
1980 "We don't have the same number of accesses in the block as on the "
1981 "access list");
1982 assert((DL || ActualDefs.size() == 0) &&
1983 "Either we should have a defs list, or we should have no defs");
1984 assert((!DL || DL->size() == ActualDefs.size()) &&
1985 "We don't have the same number of defs in the block as on the "
1986 "def list");
1987 auto ALI = AL->begin();
1988 auto AAI = ActualAccesses.begin();
1989 while (ALI != AL->end() && AAI != ActualAccesses.end()) {
1990 assert(&*ALI == *AAI && "Not the same accesses in the same order");
1991 ++ALI;
1992 ++AAI;
1994 ActualAccesses.clear();
1995 if (DL) {
1996 auto DLI = DL->begin();
1997 auto ADI = ActualDefs.begin();
1998 while (DLI != DL->end() && ADI != ActualDefs.end()) {
1999 assert(&*DLI == *ADI && "Not the same defs in the same order");
2000 ++DLI;
2001 ++ADI;
2004 ActualDefs.clear();
2006 #endif
2009 /// Verify the domination properties of MemorySSA by checking that each
2010 /// definition dominates all of its uses.
2011 void MemorySSA::verifyDomination(Function &F) const {
2012 #ifndef NDEBUG
2013 for (BasicBlock &B : F) {
2014 // Phi nodes are attached to basic blocks
2015 if (MemoryPhi *MP = getMemoryAccess(&B))
2016 for (const Use &U : MP->uses())
2017 assert(dominates(MP, U) && "Memory PHI does not dominate it's uses");
2019 for (Instruction &I : B) {
2020 MemoryAccess *MD = dyn_cast_or_null<MemoryDef>(getMemoryAccess(&I));
2021 if (!MD)
2022 continue;
2024 for (const Use &U : MD->uses())
2025 assert(dominates(MD, U) && "Memory Def does not dominate it's uses");
2028 #endif
2031 /// Verify the def-use lists in MemorySSA, by verifying that \p Use
2032 /// appears in the use list of \p Def.
2033 void MemorySSA::verifyUseInDefs(MemoryAccess *Def, MemoryAccess *Use) const {
2034 #ifndef NDEBUG
2035 // The live on entry use may cause us to get a NULL def here
2036 if (!Def)
2037 assert(isLiveOnEntryDef(Use) &&
2038 "Null def but use not point to live on entry def");
2039 else
2040 assert(is_contained(Def->users(), Use) &&
2041 "Did not find use in def's use list");
2042 #endif
2045 /// Verify the immediate use information, by walking all the memory
2046 /// accesses and verifying that, for each use, it appears in the
2047 /// appropriate def's use list
2048 void MemorySSA::verifyDefUses(Function &F) const {
2049 #ifndef NDEBUG
2050 for (BasicBlock &B : F) {
2051 // Phi nodes are attached to basic blocks
2052 if (MemoryPhi *Phi = getMemoryAccess(&B)) {
2053 assert(Phi->getNumOperands() == static_cast<unsigned>(std::distance(
2054 pred_begin(&B), pred_end(&B))) &&
2055 "Incomplete MemoryPhi Node");
2056 for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) {
2057 verifyUseInDefs(Phi->getIncomingValue(I), Phi);
2058 assert(find(predecessors(&B), Phi->getIncomingBlock(I)) !=
2059 pred_end(&B) &&
2060 "Incoming phi block not a block predecessor");
2064 for (Instruction &I : B) {
2065 if (MemoryUseOrDef *MA = getMemoryAccess(&I)) {
2066 verifyUseInDefs(MA->getDefiningAccess(), MA);
2070 #endif
2073 /// Perform a local numbering on blocks so that instruction ordering can be
2074 /// determined in constant time.
2075 /// TODO: We currently just number in order. If we numbered by N, we could
2076 /// allow at least N-1 sequences of insertBefore or insertAfter (and at least
2077 /// log2(N) sequences of mixed before and after) without needing to invalidate
2078 /// the numbering.
2079 void MemorySSA::renumberBlock(const BasicBlock *B) const {
2080 // The pre-increment ensures the numbers really start at 1.
2081 unsigned long CurrentNumber = 0;
2082 const AccessList *AL = getBlockAccesses(B);
2083 assert(AL != nullptr && "Asking to renumber an empty block");
2084 for (const auto &I : *AL)
2085 BlockNumbering[&I] = ++CurrentNumber;
2086 BlockNumberingValid.insert(B);
2089 /// Determine, for two memory accesses in the same block,
2090 /// whether \p Dominator dominates \p Dominatee.
2091 /// \returns True if \p Dominator dominates \p Dominatee.
2092 bool MemorySSA::locallyDominates(const MemoryAccess *Dominator,
2093 const MemoryAccess *Dominatee) const {
2094 const BasicBlock *DominatorBlock = Dominator->getBlock();
2096 assert((DominatorBlock == Dominatee->getBlock()) &&
2097 "Asking for local domination when accesses are in different blocks!");
2098 // A node dominates itself.
2099 if (Dominatee == Dominator)
2100 return true;
2102 // When Dominatee is defined on function entry, it is not dominated by another
2103 // memory access.
2104 if (isLiveOnEntryDef(Dominatee))
2105 return false;
2107 // When Dominator is defined on function entry, it dominates the other memory
2108 // access.
2109 if (isLiveOnEntryDef(Dominator))
2110 return true;
2112 if (!BlockNumberingValid.count(DominatorBlock))
2113 renumberBlock(DominatorBlock);
2115 unsigned long DominatorNum = BlockNumbering.lookup(Dominator);
2116 // All numbers start with 1
2117 assert(DominatorNum != 0 && "Block was not numbered properly");
2118 unsigned long DominateeNum = BlockNumbering.lookup(Dominatee);
2119 assert(DominateeNum != 0 && "Block was not numbered properly");
2120 return DominatorNum < DominateeNum;
2123 bool MemorySSA::dominates(const MemoryAccess *Dominator,
2124 const MemoryAccess *Dominatee) const {
2125 if (Dominator == Dominatee)
2126 return true;
2128 if (isLiveOnEntryDef(Dominatee))
2129 return false;
2131 if (Dominator->getBlock() != Dominatee->getBlock())
2132 return DT->dominates(Dominator->getBlock(), Dominatee->getBlock());
2133 return locallyDominates(Dominator, Dominatee);
2136 bool MemorySSA::dominates(const MemoryAccess *Dominator,
2137 const Use &Dominatee) const {
2138 if (MemoryPhi *MP = dyn_cast<MemoryPhi>(Dominatee.getUser())) {
2139 BasicBlock *UseBB = MP->getIncomingBlock(Dominatee);
2140 // The def must dominate the incoming block of the phi.
2141 if (UseBB != Dominator->getBlock())
2142 return DT->dominates(Dominator->getBlock(), UseBB);
2143 // If the UseBB and the DefBB are the same, compare locally.
2144 return locallyDominates(Dominator, cast<MemoryAccess>(Dominatee));
2146 // If it's not a PHI node use, the normal dominates can already handle it.
2147 return dominates(Dominator, cast<MemoryAccess>(Dominatee.getUser()));
2150 const static char LiveOnEntryStr[] = "liveOnEntry";
2152 void MemoryAccess::print(raw_ostream &OS) const {
2153 switch (getValueID()) {
2154 case MemoryPhiVal: return static_cast<const MemoryPhi *>(this)->print(OS);
2155 case MemoryDefVal: return static_cast<const MemoryDef *>(this)->print(OS);
2156 case MemoryUseVal: return static_cast<const MemoryUse *>(this)->print(OS);
2158 llvm_unreachable("invalid value id");
2161 void MemoryDef::print(raw_ostream &OS) const {
2162 MemoryAccess *UO = getDefiningAccess();
2164 auto printID = [&OS](MemoryAccess *A) {
2165 if (A && A->getID())
2166 OS << A->getID();
2167 else
2168 OS << LiveOnEntryStr;
2171 OS << getID() << " = MemoryDef(";
2172 printID(UO);
2173 OS << ")";
2175 if (isOptimized()) {
2176 OS << "->";
2177 printID(getOptimized());
2179 if (Optional<AliasResult> AR = getOptimizedAccessType())
2180 OS << " " << *AR;
2184 void MemoryPhi::print(raw_ostream &OS) const {
2185 bool First = true;
2186 OS << getID() << " = MemoryPhi(";
2187 for (const auto &Op : operands()) {
2188 BasicBlock *BB = getIncomingBlock(Op);
2189 MemoryAccess *MA = cast<MemoryAccess>(Op);
2190 if (!First)
2191 OS << ',';
2192 else
2193 First = false;
2195 OS << '{';
2196 if (BB->hasName())
2197 OS << BB->getName();
2198 else
2199 BB->printAsOperand(OS, false);
2200 OS << ',';
2201 if (unsigned ID = MA->getID())
2202 OS << ID;
2203 else
2204 OS << LiveOnEntryStr;
2205 OS << '}';
2207 OS << ')';
2210 void MemoryUse::print(raw_ostream &OS) const {
2211 MemoryAccess *UO = getDefiningAccess();
2212 OS << "MemoryUse(";
2213 if (UO && UO->getID())
2214 OS << UO->getID();
2215 else
2216 OS << LiveOnEntryStr;
2217 OS << ')';
2219 if (Optional<AliasResult> AR = getOptimizedAccessType())
2220 OS << " " << *AR;
2223 void MemoryAccess::dump() const {
2224 // Cannot completely remove virtual function even in release mode.
2225 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2226 print(dbgs());
2227 dbgs() << "\n";
2228 #endif
2231 char MemorySSAPrinterLegacyPass::ID = 0;
2233 MemorySSAPrinterLegacyPass::MemorySSAPrinterLegacyPass() : FunctionPass(ID) {
2234 initializeMemorySSAPrinterLegacyPassPass(*PassRegistry::getPassRegistry());
2237 void MemorySSAPrinterLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
2238 AU.setPreservesAll();
2239 AU.addRequired<MemorySSAWrapperPass>();
2242 bool MemorySSAPrinterLegacyPass::runOnFunction(Function &F) {
2243 auto &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA();
2244 MSSA.print(dbgs());
2245 if (VerifyMemorySSA)
2246 MSSA.verifyMemorySSA();
2247 return false;
2250 AnalysisKey MemorySSAAnalysis::Key;
2252 MemorySSAAnalysis::Result MemorySSAAnalysis::run(Function &F,
2253 FunctionAnalysisManager &AM) {
2254 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
2255 auto &AA = AM.getResult<AAManager>(F);
2256 return MemorySSAAnalysis::Result(std::make_unique<MemorySSA>(F, &AA, &DT));
2259 bool MemorySSAAnalysis::Result::invalidate(
2260 Function &F, const PreservedAnalyses &PA,
2261 FunctionAnalysisManager::Invalidator &Inv) {
2262 auto PAC = PA.getChecker<MemorySSAAnalysis>();
2263 return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>()) ||
2264 Inv.invalidate<AAManager>(F, PA) ||
2265 Inv.invalidate<DominatorTreeAnalysis>(F, PA);
2268 PreservedAnalyses MemorySSAPrinterPass::run(Function &F,
2269 FunctionAnalysisManager &AM) {
2270 OS << "MemorySSA for function: " << F.getName() << "\n";
2271 AM.getResult<MemorySSAAnalysis>(F).getMSSA().print(OS);
2273 return PreservedAnalyses::all();
2276 PreservedAnalyses MemorySSAVerifierPass::run(Function &F,
2277 FunctionAnalysisManager &AM) {
2278 AM.getResult<MemorySSAAnalysis>(F).getMSSA().verifyMemorySSA();
2280 return PreservedAnalyses::all();
2283 char MemorySSAWrapperPass::ID = 0;
2285 MemorySSAWrapperPass::MemorySSAWrapperPass() : FunctionPass(ID) {
2286 initializeMemorySSAWrapperPassPass(*PassRegistry::getPassRegistry());
2289 void MemorySSAWrapperPass::releaseMemory() { MSSA.reset(); }
2291 void MemorySSAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
2292 AU.setPreservesAll();
2293 AU.addRequiredTransitive<DominatorTreeWrapperPass>();
2294 AU.addRequiredTransitive<AAResultsWrapperPass>();
2297 bool MemorySSAWrapperPass::runOnFunction(Function &F) {
2298 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2299 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
2300 MSSA.reset(new MemorySSA(F, &AA, &DT));
2301 return false;
2304 void MemorySSAWrapperPass::verifyAnalysis() const { MSSA->verifyMemorySSA(); }
2306 void MemorySSAWrapperPass::print(raw_ostream &OS, const Module *M) const {
2307 MSSA->print(OS);
2310 MemorySSAWalker::MemorySSAWalker(MemorySSA *M) : MSSA(M) {}
2312 /// Walk the use-def chains starting at \p StartingAccess and find
2313 /// the MemoryAccess that actually clobbers Loc.
2315 /// \returns our clobbering memory access
2316 template <typename AliasAnalysisType>
2317 MemoryAccess *
2318 MemorySSA::ClobberWalkerBase<AliasAnalysisType>::getClobberingMemoryAccessBase(
2319 MemoryAccess *StartingAccess, const MemoryLocation &Loc,
2320 unsigned &UpwardWalkLimit) {
2321 if (isa<MemoryPhi>(StartingAccess))
2322 return StartingAccess;
2324 auto *StartingUseOrDef = cast<MemoryUseOrDef>(StartingAccess);
2325 if (MSSA->isLiveOnEntryDef(StartingUseOrDef))
2326 return StartingUseOrDef;
2328 Instruction *I = StartingUseOrDef->getMemoryInst();
2330 // Conservatively, fences are always clobbers, so don't perform the walk if we
2331 // hit a fence.
2332 if (!isa<CallBase>(I) && I->isFenceLike())
2333 return StartingUseOrDef;
2335 UpwardsMemoryQuery Q;
2336 Q.OriginalAccess = StartingUseOrDef;
2337 Q.StartingLoc = Loc;
2338 Q.Inst = I;
2339 Q.IsCall = false;
2341 // Unlike the other function, do not walk to the def of a def, because we are
2342 // handed something we already believe is the clobbering access.
2343 // We never set SkipSelf to true in Q in this method.
2344 MemoryAccess *DefiningAccess = isa<MemoryUse>(StartingUseOrDef)
2345 ? StartingUseOrDef->getDefiningAccess()
2346 : StartingUseOrDef;
2348 MemoryAccess *Clobber =
2349 Walker.findClobber(DefiningAccess, Q, UpwardWalkLimit);
2350 LLVM_DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
2351 LLVM_DEBUG(dbgs() << *StartingUseOrDef << "\n");
2352 LLVM_DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ");
2353 LLVM_DEBUG(dbgs() << *Clobber << "\n");
2354 return Clobber;
2357 template <typename AliasAnalysisType>
2358 MemoryAccess *
2359 MemorySSA::ClobberWalkerBase<AliasAnalysisType>::getClobberingMemoryAccessBase(
2360 MemoryAccess *MA, unsigned &UpwardWalkLimit, bool SkipSelf) {
2361 auto *StartingAccess = dyn_cast<MemoryUseOrDef>(MA);
2362 // If this is a MemoryPhi, we can't do anything.
2363 if (!StartingAccess)
2364 return MA;
2366 bool IsOptimized = false;
2368 // If this is an already optimized use or def, return the optimized result.
2369 // Note: Currently, we store the optimized def result in a separate field,
2370 // since we can't use the defining access.
2371 if (StartingAccess->isOptimized()) {
2372 if (!SkipSelf || !isa<MemoryDef>(StartingAccess))
2373 return StartingAccess->getOptimized();
2374 IsOptimized = true;
2377 const Instruction *I = StartingAccess->getMemoryInst();
2378 // We can't sanely do anything with a fence, since they conservatively clobber
2379 // all memory, and have no locations to get pointers from to try to
2380 // disambiguate.
2381 if (!isa<CallBase>(I) && I->isFenceLike())
2382 return StartingAccess;
2384 UpwardsMemoryQuery Q(I, StartingAccess);
2386 if (isUseTriviallyOptimizableToLiveOnEntry(*Walker.getAA(), I)) {
2387 MemoryAccess *LiveOnEntry = MSSA->getLiveOnEntryDef();
2388 StartingAccess->setOptimized(LiveOnEntry);
2389 StartingAccess->setOptimizedAccessType(None);
2390 return LiveOnEntry;
2393 MemoryAccess *OptimizedAccess;
2394 if (!IsOptimized) {
2395 // Start with the thing we already think clobbers this location
2396 MemoryAccess *DefiningAccess = StartingAccess->getDefiningAccess();
2398 // At this point, DefiningAccess may be the live on entry def.
2399 // If it is, we will not get a better result.
2400 if (MSSA->isLiveOnEntryDef(DefiningAccess)) {
2401 StartingAccess->setOptimized(DefiningAccess);
2402 StartingAccess->setOptimizedAccessType(None);
2403 return DefiningAccess;
2406 OptimizedAccess = Walker.findClobber(DefiningAccess, Q, UpwardWalkLimit);
2407 StartingAccess->setOptimized(OptimizedAccess);
2408 if (MSSA->isLiveOnEntryDef(OptimizedAccess))
2409 StartingAccess->setOptimizedAccessType(None);
2410 else if (Q.AR == MustAlias)
2411 StartingAccess->setOptimizedAccessType(MustAlias);
2412 } else
2413 OptimizedAccess = StartingAccess->getOptimized();
2415 LLVM_DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
2416 LLVM_DEBUG(dbgs() << *StartingAccess << "\n");
2417 LLVM_DEBUG(dbgs() << "Optimized Memory SSA clobber for " << *I << " is ");
2418 LLVM_DEBUG(dbgs() << *OptimizedAccess << "\n");
2420 MemoryAccess *Result;
2421 if (SkipSelf && isa<MemoryPhi>(OptimizedAccess) &&
2422 isa<MemoryDef>(StartingAccess) && UpwardWalkLimit) {
2423 assert(isa<MemoryDef>(Q.OriginalAccess));
2424 Q.SkipSelfAccess = true;
2425 Result = Walker.findClobber(OptimizedAccess, Q, UpwardWalkLimit);
2426 } else
2427 Result = OptimizedAccess;
2429 LLVM_DEBUG(dbgs() << "Result Memory SSA clobber [SkipSelf = " << SkipSelf);
2430 LLVM_DEBUG(dbgs() << "] for " << *I << " is " << *Result << "\n");
2432 return Result;
2435 MemoryAccess *
2436 DoNothingMemorySSAWalker::getClobberingMemoryAccess(MemoryAccess *MA) {
2437 if (auto *Use = dyn_cast<MemoryUseOrDef>(MA))
2438 return Use->getDefiningAccess();
2439 return MA;
2442 MemoryAccess *DoNothingMemorySSAWalker::getClobberingMemoryAccess(
2443 MemoryAccess *StartingAccess, const MemoryLocation &) {
2444 if (auto *Use = dyn_cast<MemoryUseOrDef>(StartingAccess))
2445 return Use->getDefiningAccess();
2446 return StartingAccess;
2449 void MemoryPhi::deleteMe(DerivedUser *Self) {
2450 delete static_cast<MemoryPhi *>(Self);
2453 void MemoryDef::deleteMe(DerivedUser *Self) {
2454 delete static_cast<MemoryDef *>(Self);
2457 void MemoryUse::deleteMe(DerivedUser *Self) {
2458 delete static_cast<MemoryUse *>(Self);