LLVM  8.0.0svn
MemorySSA.cpp
Go to the documentation of this file.
1 //===- MemorySSA.cpp - Memory SSA Builder ---------------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the MemorySSA class.
11 //
12 //===----------------------------------------------------------------------===//
13 
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/DenseMapInfo.h"
17 #include "llvm/ADT/DenseSet.h"
19 #include "llvm/ADT/Hashing.h"
20 #include "llvm/ADT/None.h"
21 #include "llvm/ADT/Optional.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/ADT/SmallPtrSet.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/iterator.h"
30 #include "llvm/Config/llvm-config.h"
32 #include "llvm/IR/BasicBlock.h"
33 #include "llvm/IR/CallSite.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/Instruction.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/IntrinsicInst.h"
39 #include "llvm/IR/Intrinsics.h"
40 #include "llvm/IR/LLVMContext.h"
41 #include "llvm/IR/PassManager.h"
42 #include "llvm/IR/Use.h"
43 #include "llvm/Pass.h"
45 #include "llvm/Support/Casting.h"
47 #include "llvm/Support/Compiler.h"
48 #include "llvm/Support/Debug.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 INITIALIZE_PASS_BEGIN(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,
63  true)
67  true)
68 
70  "Memory SSA Printer", false, false)
71 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
72 INITIALIZE_PASS_END(MemorySSAPrinterLegacyPass, "print-memoryssa",
73  "Memory SSA Printer", false, false)
74 
75 static cl::opt<unsigned> MaxCheckLimit(
76  "memssa-check-limit", cl::Hidden, cl::init(100),
77  cl::desc("The maximum number of stores/phis MemorySSA"
78  "will consider trying to walk past (default = 100)"));
79 
80 // Always verify MemorySSA if expensive checking is enabled.
81 #ifdef EXPENSIVE_CHECKS
82 bool llvm::VerifyMemorySSA = true;
83 #else
84 bool llvm::VerifyMemorySSA = false;
85 #endif
87  VerifyMemorySSAX("verify-memoryssa", cl::location(VerifyMemorySSA),
88  cl::Hidden, cl::desc("Enable verification of MemorySSA."));
89 
90 namespace llvm {
91 
92 /// An assembly annotator class to print Memory SSA information in
93 /// comments.
95  friend class MemorySSA;
96 
97  const MemorySSA *MSSA;
98 
99 public:
100  MemorySSAAnnotatedWriter(const MemorySSA *M) : MSSA(M) {}
101 
103  formatted_raw_ostream &OS) override {
104  if (MemoryAccess *MA = MSSA->getMemoryAccess(BB))
105  OS << "; " << *MA << "\n";
106  }
107 
109  formatted_raw_ostream &OS) override {
110  if (MemoryAccess *MA = MSSA->getMemoryAccess(I))
111  OS << "; " << *MA << "\n";
112  }
113 };
114 
115 } // end namespace llvm
116 
117 namespace {
118 
119 /// Our current alias analysis API differentiates heavily between calls and
120 /// non-calls, and functions called on one usually assert on the other.
121 /// This class encapsulates the distinction to simplify other code that wants
122 /// "Memory affecting instructions and related data" to use as a key.
123 /// For example, this class is used as a densemap key in the use optimizer.
124 class MemoryLocOrCall {
125 public:
126  bool IsCall = false;
127 
128  MemoryLocOrCall(MemoryUseOrDef *MUD)
129  : MemoryLocOrCall(MUD->getMemoryInst()) {}
130  MemoryLocOrCall(const MemoryUseOrDef *MUD)
131  : MemoryLocOrCall(MUD->getMemoryInst()) {}
132 
133  MemoryLocOrCall(Instruction *Inst) {
134  if (ImmutableCallSite(Inst)) {
135  IsCall = true;
136  CS = ImmutableCallSite(Inst);
137  } else {
138  IsCall = false;
139  // There is no such thing as a memorylocation for a fence inst, and it is
140  // unique in that regard.
141  if (!isa<FenceInst>(Inst))
142  Loc = MemoryLocation::get(Inst);
143  }
144  }
145 
146  explicit MemoryLocOrCall(const MemoryLocation &Loc) : Loc(Loc) {}
147 
148  ImmutableCallSite getCS() const {
149  assert(IsCall);
150  return CS;
151  }
152 
153  MemoryLocation getLoc() const {
154  assert(!IsCall);
155  return Loc;
156  }
157 
158  bool operator==(const MemoryLocOrCall &Other) const {
159  if (IsCall != Other.IsCall)
160  return false;
161 
162  if (!IsCall)
163  return Loc == Other.Loc;
164 
165  if (CS.getCalledValue() != Other.CS.getCalledValue())
166  return false;
167 
168  return CS.arg_size() == Other.CS.arg_size() &&
169  std::equal(CS.arg_begin(), CS.arg_end(), Other.CS.arg_begin());
170  }
171 
172 private:
173  union {
175  MemoryLocation Loc;
176  };
177 };
178 
179 } // end anonymous namespace
180 
181 namespace llvm {
182 
183 template <> struct DenseMapInfo<MemoryLocOrCall> {
184  static inline MemoryLocOrCall getEmptyKey() {
185  return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getEmptyKey());
186  }
187 
188  static inline MemoryLocOrCall getTombstoneKey() {
189  return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getTombstoneKey());
190  }
191 
192  static unsigned getHashValue(const MemoryLocOrCall &MLOC) {
193  if (!MLOC.IsCall)
194  return hash_combine(
195  MLOC.IsCall,
197 
198  hash_code hash =
200  MLOC.getCS().getCalledValue()));
201 
202  for (const Value *Arg : MLOC.getCS().args())
204  return hash;
205  }
206 
207  static bool isEqual(const MemoryLocOrCall &LHS, const MemoryLocOrCall &RHS) {
208  return LHS == RHS;
209  }
210 };
211 
212 } // end namespace llvm
213 
214 /// This does one-way checks to see if Use could theoretically be hoisted above
215 /// MayClobber. This will not check the other way around.
216 ///
217 /// This assumes that, for the purposes of MemorySSA, Use comes directly after
218 /// MayClobber, with no potentially clobbering operations in between them.
219 /// (Where potentially clobbering ops are memory barriers, aliased stores, etc.)
220 static bool areLoadsReorderable(const LoadInst *Use,
221  const LoadInst *MayClobber) {
222  bool VolatileUse = Use->isVolatile();
223  bool VolatileClobber = MayClobber->isVolatile();
224  // Volatile operations may never be reordered with other volatile operations.
225  if (VolatileUse && VolatileClobber)
226  return false;
227  // Otherwise, volatile doesn't matter here. From the language reference:
228  // 'optimizers may change the order of volatile operations relative to
229  // non-volatile operations.'"
230 
231  // If a load is seq_cst, it cannot be moved above other loads. If its ordering
232  // is weaker, it can be moved above other loads. We just need to be sure that
233  // MayClobber isn't an acquire load, because loads can't be moved above
234  // acquire loads.
235  //
236  // Note that this explicitly *does* allow the free reordering of monotonic (or
237  // weaker) loads of the same address.
238  bool SeqCstUse = Use->getOrdering() == AtomicOrdering::SequentiallyConsistent;
239  bool MayClobberIsAcquire = isAtLeastOrStrongerThan(MayClobber->getOrdering(),
241  return !(SeqCstUse || MayClobberIsAcquire);
242 }
243 
244 namespace {
245 
246 struct ClobberAlias {
247  bool IsClobber;
249 };
250 
251 } // end anonymous namespace
252 
253 // Return a pair of {IsClobber (bool), AR (AliasResult)}. It relies on AR being
254 // ignored if IsClobber = false.
255 static ClobberAlias instructionClobbersQuery(const MemoryDef *MD,
256  const MemoryLocation &UseLoc,
257  const Instruction *UseInst,
258  AliasAnalysis &AA) {
259  Instruction *DefInst = MD->getMemoryInst();
260  assert(DefInst && "Defining instruction not actually an instruction");
261  ImmutableCallSite UseCS(UseInst);
263 
264  if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(DefInst)) {
265  // These intrinsics will show up as affecting memory, but they are just
266  // markers, mostly.
267  //
268  // FIXME: We probably don't actually want MemorySSA to model these at all
269  // (including creating MemoryAccesses for them): we just end up inventing
270  // clobbers where they don't really exist at all. Please see D43269 for
271  // context.
272  switch (II->getIntrinsicID()) {
273  case Intrinsic::lifetime_start:
274  if (UseCS)
275  return {false, NoAlias};
276  AR = AA.alias(MemoryLocation(II->getArgOperand(1)), UseLoc);
277  return {AR != NoAlias, AR};
278  case Intrinsic::lifetime_end:
279  case Intrinsic::invariant_start:
280  case Intrinsic::invariant_end:
281  case Intrinsic::assume:
282  return {false, NoAlias};
283  default:
284  break;
285  }
286  }
287 
288  if (UseCS) {
289  ModRefInfo I = AA.getModRefInfo(DefInst, UseCS);
290  AR = isMustSet(I) ? MustAlias : MayAlias;
291  return {isModOrRefSet(I), AR};
292  }
293 
294  if (auto *DefLoad = dyn_cast<LoadInst>(DefInst))
295  if (auto *UseLoad = dyn_cast<LoadInst>(UseInst))
296  return {!areLoadsReorderable(UseLoad, DefLoad), MayAlias};
297 
298  ModRefInfo I = AA.getModRefInfo(DefInst, UseLoc);
299  AR = isMustSet(I) ? MustAlias : MayAlias;
300  return {isModSet(I), AR};
301 }
302 
303 static ClobberAlias instructionClobbersQuery(MemoryDef *MD,
304  const MemoryUseOrDef *MU,
305  const MemoryLocOrCall &UseMLOC,
306  AliasAnalysis &AA) {
307  // FIXME: This is a temporary hack to allow a single instructionClobbersQuery
308  // to exist while MemoryLocOrCall is pushed through places.
309  if (UseMLOC.IsCall)
311  AA);
312  return instructionClobbersQuery(MD, UseMLOC.getLoc(), MU->getMemoryInst(),
313  AA);
314 }
315 
316 // Return true when MD may alias MU, return false otherwise.
318  AliasAnalysis &AA) {
319  return instructionClobbersQuery(MD, MU, MemoryLocOrCall(MU), AA).IsClobber;
320 }
321 
322 namespace {
323 
324 struct UpwardsMemoryQuery {
325  // True if our original query started off as a call
326  bool IsCall = false;
327  // The pointer location we started the query with. This will be empty if
328  // IsCall is true.
329  MemoryLocation StartingLoc;
330  // This is the instruction we were querying about.
331  const Instruction *Inst = nullptr;
332  // The MemoryAccess we actually got called with, used to test local domination
333  const MemoryAccess *OriginalAccess = nullptr;
335 
336  UpwardsMemoryQuery() = default;
337 
338  UpwardsMemoryQuery(const Instruction *Inst, const MemoryAccess *Access)
339  : IsCall(ImmutableCallSite(Inst)), Inst(Inst), OriginalAccess(Access) {
340  if (!IsCall)
341  StartingLoc = MemoryLocation::get(Inst);
342  }
343 };
344 
345 } // end anonymous namespace
346 
347 static bool lifetimeEndsAt(MemoryDef *MD, const MemoryLocation &Loc,
348  AliasAnalysis &AA) {
349  Instruction *Inst = MD->getMemoryInst();
350  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
351  switch (II->getIntrinsicID()) {
352  case Intrinsic::lifetime_end:
353  return AA.isMustAlias(MemoryLocation(II->getArgOperand(1)), Loc);
354  default:
355  return false;
356  }
357  }
358  return false;
359 }
360 
362  const Instruction *I) {
363  // If the memory can't be changed, then loads of the memory can't be
364  // clobbered.
365  return isa<LoadInst>(I) && (I->getMetadata(LLVMContext::MD_invariant_load) ||
366  AA.pointsToConstantMemory(cast<LoadInst>(I)->
367  getPointerOperand()));
368 }
369 
370 /// Verifies that `Start` is clobbered by `ClobberAt`, and that nothing
371 /// inbetween `Start` and `ClobberAt` can clobbers `Start`.
372 ///
373 /// This is meant to be as simple and self-contained as possible. Because it
374 /// uses no cache, etc., it can be relatively expensive.
375 ///
376 /// \param Start The MemoryAccess that we want to walk from.
377 /// \param ClobberAt A clobber for Start.
378 /// \param StartLoc The MemoryLocation for Start.
379 /// \param MSSA The MemorySSA instance that Start and ClobberAt belong to.
380 /// \param Query The UpwardsMemoryQuery we used for our search.
381 /// \param AA The AliasAnalysis we used for our search.
382 /// \param AllowImpreciseClobber Always false, unless we do relaxed verify.
383 static void
385  const MemoryLocation &StartLoc, const MemorySSA &MSSA,
386  const UpwardsMemoryQuery &Query, AliasAnalysis &AA,
387  bool AllowImpreciseClobber = false) {
388  assert(MSSA.dominates(ClobberAt, Start) && "Clobber doesn't dominate start?");
389 
390  if (MSSA.isLiveOnEntryDef(Start)) {
391  assert(MSSA.isLiveOnEntryDef(ClobberAt) &&
392  "liveOnEntry must clobber itself");
393  return;
394  }
395 
396  bool FoundClobber = false;
399  Worklist.emplace_back(Start, StartLoc);
400  // Walk all paths from Start to ClobberAt, while looking for clobbers. If one
401  // is found, complain.
402  while (!Worklist.empty()) {
403  auto MAP = Worklist.pop_back_val();
404  // All we care about is that nothing from Start to ClobberAt clobbers Start.
405  // We learn nothing from revisiting nodes.
406  if (!VisitedPhis.insert(MAP).second)
407  continue;
408 
409  for (const auto *MA : def_chain(MAP.first)) {
410  if (MA == ClobberAt) {
411  if (const auto *MD = dyn_cast<MemoryDef>(MA)) {
412  // instructionClobbersQuery isn't essentially free, so don't use `|=`,
413  // since it won't let us short-circuit.
414  //
415  // Also, note that this can't be hoisted out of the `Worklist` loop,
416  // since MD may only act as a clobber for 1 of N MemoryLocations.
417  FoundClobber = FoundClobber || MSSA.isLiveOnEntryDef(MD);
418  if (!FoundClobber) {
419  ClobberAlias CA =
420  instructionClobbersQuery(MD, MAP.second, Query.Inst, AA);
421  if (CA.IsClobber) {
422  FoundClobber = true;
423  // Not used: CA.AR;
424  }
425  }
426  }
427  break;
428  }
429 
430  // We should never hit liveOnEntry, unless it's the clobber.
431  assert(!MSSA.isLiveOnEntryDef(MA) && "Hit liveOnEntry before clobber?");
432 
433  if (const auto *MD = dyn_cast<MemoryDef>(MA)) {
434  // If Start is a Def, skip self.
435  if (MD == Start)
436  continue;
437 
438  assert(!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA)
439  .IsClobber &&
440  "Found clobber before reaching ClobberAt!");
441  continue;
442  }
443 
444  if (const auto *MU = dyn_cast<MemoryUse>(MA)) {
445  (void)MU;
446  assert (MU == Start &&
447  "Can only find use in def chain if Start is a use");
448  continue;
449  }
450 
451  assert(isa<MemoryPhi>(MA));
452  Worklist.append(
453  upward_defs_begin({const_cast<MemoryAccess *>(MA), MAP.second}),
454  upward_defs_end());
455  }
456  }
457 
458  // If the verify is done following an optimization, it's possible that
459  // ClobberAt was a conservative clobbering, that we can now infer is not a
460  // true clobbering access. Don't fail the verify if that's the case.
461  // We do have accesses that claim they're optimized, but could be optimized
462  // further. Updating all these can be expensive, so allow it for now (FIXME).
463  if (AllowImpreciseClobber)
464  return;
465 
466  // If ClobberAt is a MemoryPhi, we can assume something above it acted as a
467  // clobber. Otherwise, `ClobberAt` should've acted as a clobber at some point.
468  assert((isa<MemoryPhi>(ClobberAt) || FoundClobber) &&
469  "ClobberAt never acted as a clobber");
470 }
471 
472 namespace {
473 
474 /// Our algorithm for walking (and trying to optimize) clobbers, all wrapped up
475 /// in one class.
476 class ClobberWalker {
477  /// Save a few bytes by using unsigned instead of size_t.
478  using ListIndex = unsigned;
479 
480  /// Represents a span of contiguous MemoryDefs, potentially ending in a
481  /// MemoryPhi.
482  struct DefPath {
483  MemoryLocation Loc;
484  // Note that, because we always walk in reverse, Last will always dominate
485  // First. Also note that First and Last are inclusive.
486  MemoryAccess *First;
487  MemoryAccess *Last;
488  Optional<ListIndex> Previous;
489 
490  DefPath(const MemoryLocation &Loc, MemoryAccess *First, MemoryAccess *Last,
491  Optional<ListIndex> Previous)
492  : Loc(Loc), First(First), Last(Last), Previous(Previous) {}
493 
494  DefPath(const MemoryLocation &Loc, MemoryAccess *Init,
495  Optional<ListIndex> Previous)
496  : DefPath(Loc, Init, Init, Previous) {}
497  };
498 
499  const MemorySSA &MSSA;
500  AliasAnalysis &AA;
501  DominatorTree &DT;
502  UpwardsMemoryQuery *Query;
503 
504  // Phi optimization bookkeeping
507 
508  /// Find the nearest def or phi that `From` can legally be optimized to.
509  const MemoryAccess *getWalkTarget(const MemoryPhi *From) const {
510  assert(From->getNumOperands() && "Phi with no operands?");
511 
512  BasicBlock *BB = From->getBlock();
513  MemoryAccess *Result = MSSA.getLiveOnEntryDef();
514  DomTreeNode *Node = DT.getNode(BB);
515  while ((Node = Node->getIDom())) {
516  auto *Defs = MSSA.getBlockDefs(Node->getBlock());
517  if (Defs)
518  return &*Defs->rbegin();
519  }
520  return Result;
521  }
522 
523  /// Result of calling walkToPhiOrClobber.
524  struct UpwardsWalkResult {
525  /// The "Result" of the walk. Either a clobber, the last thing we walked, or
526  /// both. Include alias info when clobber found.
527  MemoryAccess *Result;
528  bool IsKnownClobber;
530  };
531 
532  /// Walk to the next Phi or Clobber in the def chain starting at Desc.Last.
533  /// This will update Desc.Last as it walks. It will (optionally) also stop at
534  /// StopAt.
535  ///
536  /// This does not test for whether StopAt is a clobber
537  UpwardsWalkResult
538  walkToPhiOrClobber(DefPath &Desc,
539  const MemoryAccess *StopAt = nullptr) const {
540  assert(!isa<MemoryUse>(Desc.Last) && "Uses don't exist in my world");
541 
542  for (MemoryAccess *Current : def_chain(Desc.Last)) {
543  Desc.Last = Current;
544  if (Current == StopAt)
545  return {Current, false, MayAlias};
546 
547  if (auto *MD = dyn_cast<MemoryDef>(Current)) {
548  if (MSSA.isLiveOnEntryDef(MD))
549  return {MD, true, MustAlias};
550  ClobberAlias CA =
551  instructionClobbersQuery(MD, Desc.Loc, Query->Inst, AA);
552  if (CA.IsClobber)
553  return {MD, true, CA.AR};
554  }
555  }
556 
557  assert(isa<MemoryPhi>(Desc.Last) &&
558  "Ended at a non-clobber that's not a phi?");
559  return {Desc.Last, false, MayAlias};
560  }
561 
562  void addSearches(MemoryPhi *Phi, SmallVectorImpl<ListIndex> &PausedSearches,
563  ListIndex PriorNode) {
564  auto UpwardDefs = make_range(upward_defs_begin({Phi, Paths[PriorNode].Loc}),
565  upward_defs_end());
566  for (const MemoryAccessPair &P : UpwardDefs) {
567  PausedSearches.push_back(Paths.size());
568  Paths.emplace_back(P.second, P.first, PriorNode);
569  }
570  }
571 
572  /// Represents a search that terminated after finding a clobber. This clobber
573  /// may or may not be present in the path of defs from LastNode..SearchStart,
574  /// since it may have been retrieved from cache.
575  struct TerminatedPath {
576  MemoryAccess *Clobber;
577  ListIndex LastNode;
578  };
579 
580  /// Get an access that keeps us from optimizing to the given phi.
581  ///
582  /// PausedSearches is an array of indices into the Paths array. Its incoming
583  /// value is the indices of searches that stopped at the last phi optimization
584  /// target. It's left in an unspecified state.
585  ///
586  /// If this returns None, NewPaused is a vector of searches that terminated
587  /// at StopWhere. Otherwise, NewPaused is left in an unspecified state.
589  getBlockingAccess(const MemoryAccess *StopWhere,
590  SmallVectorImpl<ListIndex> &PausedSearches,
591  SmallVectorImpl<ListIndex> &NewPaused,
592  SmallVectorImpl<TerminatedPath> &Terminated) {
593  assert(!PausedSearches.empty() && "No searches to continue?");
594 
595  // BFS vs DFS really doesn't make a difference here, so just do a DFS with
596  // PausedSearches as our stack.
597  while (!PausedSearches.empty()) {
598  ListIndex PathIndex = PausedSearches.pop_back_val();
599  DefPath &Node = Paths[PathIndex];
600 
601  // If we've already visited this path with this MemoryLocation, we don't
602  // need to do so again.
603  //
604  // NOTE: That we just drop these paths on the ground makes caching
605  // behavior sporadic. e.g. given a diamond:
606  // A
607  // B C
608  // D
609  //
610  // ...If we walk D, B, A, C, we'll only cache the result of phi
611  // optimization for A, B, and D; C will be skipped because it dies here.
612  // This arguably isn't the worst thing ever, since:
613  // - We generally query things in a top-down order, so if we got below D
614  // without needing cache entries for {C, MemLoc}, then chances are
615  // that those cache entries would end up ultimately unused.
616  // - We still cache things for A, so C only needs to walk up a bit.
617  // If this behavior becomes problematic, we can fix without a ton of extra
618  // work.
619  if (!VisitedPhis.insert({Node.Last, Node.Loc}).second)
620  continue;
621 
622  UpwardsWalkResult Res = walkToPhiOrClobber(Node, /*StopAt=*/StopWhere);
623  if (Res.IsKnownClobber) {
624  assert(Res.Result != StopWhere);
625  // If this wasn't a cache hit, we hit a clobber when walking. That's a
626  // failure.
627  TerminatedPath Term{Res.Result, PathIndex};
628  if (!MSSA.dominates(Res.Result, StopWhere))
629  return Term;
630 
631  // Otherwise, it's a valid thing to potentially optimize to.
632  Terminated.push_back(Term);
633  continue;
634  }
635 
636  if (Res.Result == StopWhere) {
637  // We've hit our target. Save this path off for if we want to continue
638  // walking.
639  NewPaused.push_back(PathIndex);
640  continue;
641  }
642 
643  assert(!MSSA.isLiveOnEntryDef(Res.Result) && "liveOnEntry is a clobber");
644  addSearches(cast<MemoryPhi>(Res.Result), PausedSearches, PathIndex);
645  }
646 
647  return None;
648  }
649 
650  template <typename T, typename Walker>
651  struct generic_def_path_iterator
652  : public iterator_facade_base<generic_def_path_iterator<T, Walker>,
653  std::forward_iterator_tag, T *> {
654  generic_def_path_iterator() = default;
655  generic_def_path_iterator(Walker *W, ListIndex N) : W(W), N(N) {}
656 
657  T &operator*() const { return curNode(); }
658 
659  generic_def_path_iterator &operator++() {
660  N = curNode().Previous;
661  return *this;
662  }
663 
664  bool operator==(const generic_def_path_iterator &O) const {
665  if (N.hasValue() != O.N.hasValue())
666  return false;
667  return !N.hasValue() || *N == *O.N;
668  }
669 
670  private:
671  T &curNode() const { return W->Paths[*N]; }
672 
673  Walker *W = nullptr;
675  };
676 
677  using def_path_iterator = generic_def_path_iterator<DefPath, ClobberWalker>;
678  using const_def_path_iterator =
679  generic_def_path_iterator<const DefPath, const ClobberWalker>;
680 
681  iterator_range<def_path_iterator> def_path(ListIndex From) {
682  return make_range(def_path_iterator(this, From), def_path_iterator());
683  }
684 
685  iterator_range<const_def_path_iterator> const_def_path(ListIndex From) const {
686  return make_range(const_def_path_iterator(this, From),
687  const_def_path_iterator());
688  }
689 
690  struct OptznResult {
691  /// The path that contains our result.
692  TerminatedPath PrimaryClobber;
693  /// The paths that we can legally cache back from, but that aren't
694  /// necessarily the result of the Phi optimization.
695  SmallVector<TerminatedPath, 4> OtherClobbers;
696  };
697 
698  ListIndex defPathIndex(const DefPath &N) const {
699  // The assert looks nicer if we don't need to do &N
700  const DefPath *NP = &N;
701  assert(!Paths.empty() && NP >= &Paths.front() && NP <= &Paths.back() &&
702  "Out of bounds DefPath!");
703  return NP - &Paths.front();
704  }
705 
706  /// Try to optimize a phi as best as we can. Returns a SmallVector of Paths
707  /// that act as legal clobbers. Note that this won't return *all* clobbers.
708  ///
709  /// Phi optimization algorithm tl;dr:
710  /// - Find the earliest def/phi, A, we can optimize to
711  /// - Find if all paths from the starting memory access ultimately reach A
712  /// - If not, optimization isn't possible.
713  /// - Otherwise, walk from A to another clobber or phi, A'.
714  /// - If A' is a def, we're done.
715  /// - If A' is a phi, try to optimize it.
716  ///
717  /// A path is a series of {MemoryAccess, MemoryLocation} pairs. A path
718  /// terminates when a MemoryAccess that clobbers said MemoryLocation is found.
719  OptznResult tryOptimizePhi(MemoryPhi *Phi, MemoryAccess *Start,
720  const MemoryLocation &Loc) {
721  assert(Paths.empty() && VisitedPhis.empty() &&
722  "Reset the optimization state.");
723 
724  Paths.emplace_back(Loc, Start, Phi, None);
725  // Stores how many "valid" optimization nodes we had prior to calling
726  // addSearches/getBlockingAccess. Necessary for caching if we had a blocker.
727  auto PriorPathsSize = Paths.size();
728 
729  SmallVector<ListIndex, 16> PausedSearches;
730  SmallVector<ListIndex, 8> NewPaused;
731  SmallVector<TerminatedPath, 4> TerminatedPaths;
732 
733  addSearches(Phi, PausedSearches, 0);
734 
735  // Moves the TerminatedPath with the "most dominated" Clobber to the end of
736  // Paths.
737  auto MoveDominatedPathToEnd = [&](SmallVectorImpl<TerminatedPath> &Paths) {
738  assert(!Paths.empty() && "Need a path to move");
739  auto Dom = Paths.begin();
740  for (auto I = std::next(Dom), E = Paths.end(); I != E; ++I)
741  if (!MSSA.dominates(I->Clobber, Dom->Clobber))
742  Dom = I;
743  auto Last = Paths.end() - 1;
744  if (Last != Dom)
745  std::iter_swap(Last, Dom);
746  };
747 
748  MemoryPhi *Current = Phi;
749  while (true) {
750  assert(!MSSA.isLiveOnEntryDef(Current) &&
751  "liveOnEntry wasn't treated as a clobber?");
752 
753  const auto *Target = getWalkTarget(Current);
754  // If a TerminatedPath doesn't dominate Target, then it wasn't a legal
755  // optimization for the prior phi.
756  assert(all_of(TerminatedPaths, [&](const TerminatedPath &P) {
757  return MSSA.dominates(P.Clobber, Target);
758  }));
759 
760  // FIXME: This is broken, because the Blocker may be reported to be
761  // liveOnEntry, and we'll happily wait for that to disappear (read: never)
762  // For the moment, this is fine, since we do nothing with blocker info.
763  if (Optional<TerminatedPath> Blocker = getBlockingAccess(
764  Target, PausedSearches, NewPaused, TerminatedPaths)) {
765 
766  // Find the node we started at. We can't search based on N->Last, since
767  // we may have gone around a loop with a different MemoryLocation.
768  auto Iter = find_if(def_path(Blocker->LastNode), [&](const DefPath &N) {
769  return defPathIndex(N) < PriorPathsSize;
770  });
771  assert(Iter != def_path_iterator());
772 
773  DefPath &CurNode = *Iter;
774  assert(CurNode.Last == Current);
775 
776  // Two things:
777  // A. We can't reliably cache all of NewPaused back. Consider a case
778  // where we have two paths in NewPaused; one of which can't optimize
779  // above this phi, whereas the other can. If we cache the second path
780  // back, we'll end up with suboptimal cache entries. We can handle
781  // cases like this a bit better when we either try to find all
782  // clobbers that block phi optimization, or when our cache starts
783  // supporting unfinished searches.
784  // B. We can't reliably cache TerminatedPaths back here without doing
785  // extra checks; consider a case like:
786  // T
787  // / \
788  // D C
789  // \ /
790  // S
791  // Where T is our target, C is a node with a clobber on it, D is a
792  // diamond (with a clobber *only* on the left or right node, N), and
793  // S is our start. Say we walk to D, through the node opposite N
794  // (read: ignoring the clobber), and see a cache entry in the top
795  // node of D. That cache entry gets put into TerminatedPaths. We then
796  // walk up to C (N is later in our worklist), find the clobber, and
797  // quit. If we append TerminatedPaths to OtherClobbers, we'll cache
798  // the bottom part of D to the cached clobber, ignoring the clobber
799  // in N. Again, this problem goes away if we start tracking all
800  // blockers for a given phi optimization.
801  TerminatedPath Result{CurNode.Last, defPathIndex(CurNode)};
802  return {Result, {}};
803  }
804 
805  // If there's nothing left to search, then all paths led to valid clobbers
806  // that we got from our cache; pick the nearest to the start, and allow
807  // the rest to be cached back.
808  if (NewPaused.empty()) {
809  MoveDominatedPathToEnd(TerminatedPaths);
810  TerminatedPath Result = TerminatedPaths.pop_back_val();
811  return {Result, std::move(TerminatedPaths)};
812  }
813 
814  MemoryAccess *DefChainEnd = nullptr;
816  for (ListIndex Paused : NewPaused) {
817  UpwardsWalkResult WR = walkToPhiOrClobber(Paths[Paused]);
818  if (WR.IsKnownClobber)
819  Clobbers.push_back({WR.Result, Paused});
820  else
821  // Micro-opt: If we hit the end of the chain, save it.
822  DefChainEnd = WR.Result;
823  }
824 
825  if (!TerminatedPaths.empty()) {
826  // If we couldn't find the dominating phi/liveOnEntry in the above loop,
827  // do it now.
828  if (!DefChainEnd)
829  for (auto *MA : def_chain(const_cast<MemoryAccess *>(Target)))
830  DefChainEnd = MA;
831 
832  // If any of the terminated paths don't dominate the phi we'll try to
833  // optimize, we need to figure out what they are and quit.
834  const BasicBlock *ChainBB = DefChainEnd->getBlock();
835  for (const TerminatedPath &TP : TerminatedPaths) {
836  // Because we know that DefChainEnd is as "high" as we can go, we
837  // don't need local dominance checks; BB dominance is sufficient.
838  if (DT.dominates(ChainBB, TP.Clobber->getBlock()))
839  Clobbers.push_back(TP);
840  }
841  }
842 
843  // If we have clobbers in the def chain, find the one closest to Current
844  // and quit.
845  if (!Clobbers.empty()) {
846  MoveDominatedPathToEnd(Clobbers);
847  TerminatedPath Result = Clobbers.pop_back_val();
848  return {Result, std::move(Clobbers)};
849  }
850 
851  assert(all_of(NewPaused,
852  [&](ListIndex I) { return Paths[I].Last == DefChainEnd; }));
853 
854  // Because liveOnEntry is a clobber, this must be a phi.
855  auto *DefChainPhi = cast<MemoryPhi>(DefChainEnd);
856 
857  PriorPathsSize = Paths.size();
858  PausedSearches.clear();
859  for (ListIndex I : NewPaused)
860  addSearches(DefChainPhi, PausedSearches, I);
861  NewPaused.clear();
862 
863  Current = DefChainPhi;
864  }
865  }
866 
867  void verifyOptResult(const OptznResult &R) const {
868  assert(all_of(R.OtherClobbers, [&](const TerminatedPath &P) {
869  return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber);
870  }));
871  }
872 
873  void resetPhiOptznState() {
874  Paths.clear();
875  VisitedPhis.clear();
876  }
877 
878 public:
879  ClobberWalker(const MemorySSA &MSSA, AliasAnalysis &AA, DominatorTree &DT)
880  : MSSA(MSSA), AA(AA), DT(DT) {}
881 
882  /// Finds the nearest clobber for the given query, optimizing phis if
883  /// possible.
884  MemoryAccess *findClobber(MemoryAccess *Start, UpwardsMemoryQuery &Q) {
885  Query = &Q;
886 
887  MemoryAccess *Current = Start;
888  // This walker pretends uses don't exist. If we're handed one, silently grab
889  // its def. (This has the nice side-effect of ensuring we never cache uses)
890  if (auto *MU = dyn_cast<MemoryUse>(Start))
891  Current = MU->getDefiningAccess();
892 
893  DefPath FirstDesc(Q.StartingLoc, Current, Current, None);
894  // Fast path for the overly-common case (no crazy phi optimization
895  // necessary)
896  UpwardsWalkResult WalkResult = walkToPhiOrClobber(FirstDesc);
897  MemoryAccess *Result;
898  if (WalkResult.IsKnownClobber) {
899  Result = WalkResult.Result;
900  Q.AR = WalkResult.AR;
901  } else {
902  OptznResult OptRes = tryOptimizePhi(cast<MemoryPhi>(FirstDesc.Last),
903  Current, Q.StartingLoc);
904  verifyOptResult(OptRes);
905  resetPhiOptznState();
906  Result = OptRes.PrimaryClobber.Clobber;
907  }
908 
909 #ifdef EXPENSIVE_CHECKS
910  checkClobberSanity(Current, Result, Q.StartingLoc, MSSA, Q, AA);
911 #endif
912  return Result;
913  }
914 
915  void verify(const MemorySSA *MSSA) { assert(MSSA == &this->MSSA); }
916 };
917 
918 struct RenamePassData {
919  DomTreeNode *DTN;
921  MemoryAccess *IncomingVal;
922 
923  RenamePassData(DomTreeNode *D, DomTreeNode::const_iterator It,
924  MemoryAccess *M)
925  : DTN(D), ChildIt(It), IncomingVal(M) {}
926 
927  void swap(RenamePassData &RHS) {
928  std::swap(DTN, RHS.DTN);
929  std::swap(ChildIt, RHS.ChildIt);
930  std::swap(IncomingVal, RHS.IncomingVal);
931  }
932 };
933 
934 } // end anonymous namespace
935 
936 namespace llvm {
937 
938 /// A MemorySSAWalker that does AA walks to disambiguate accesses. It no
939 /// longer does caching on its own, but the name has been retained for the
940 /// moment.
942  ClobberWalker Walker;
943 
944  MemoryAccess *getClobberingMemoryAccess(MemoryAccess *, UpwardsMemoryQuery &);
945 
946 public:
948  ~CachingWalker() override = default;
949 
951 
952  MemoryAccess *getClobberingMemoryAccess(MemoryAccess *) override;
953  MemoryAccess *getClobberingMemoryAccess(MemoryAccess *,
954  const MemoryLocation &) override;
955  void invalidateInfo(MemoryAccess *) override;
956 
957  void verify(const MemorySSA *MSSA) override {
959  Walker.verify(MSSA);
960  }
961 };
962 
963 } // end namespace llvm
964 
965 void MemorySSA::renameSuccessorPhis(BasicBlock *BB, MemoryAccess *IncomingVal,
966  bool RenameAllUses) {
967  // Pass through values to our successors
968  for (const BasicBlock *S : successors(BB)) {
969  auto It = PerBlockAccesses.find(S);
970  // Rename the phi nodes in our successor block
971  if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
972  continue;
973  AccessList *Accesses = It->second.get();
974  auto *Phi = cast<MemoryPhi>(&Accesses->front());
975  if (RenameAllUses) {
976  int PhiIndex = Phi->getBasicBlockIndex(BB);
977  assert(PhiIndex != -1 && "Incomplete phi during partial rename");
978  Phi->setIncomingValue(PhiIndex, IncomingVal);
979  } else
980  Phi->addIncoming(IncomingVal, BB);
981  }
982 }
983 
984 /// Rename a single basic block into MemorySSA form.
985 /// Uses the standard SSA renaming algorithm.
986 /// \returns The new incoming value.
987 MemoryAccess *MemorySSA::renameBlock(BasicBlock *BB, MemoryAccess *IncomingVal,
988  bool RenameAllUses) {
989  auto It = PerBlockAccesses.find(BB);
990  // Skip most processing if the list is empty.
991  if (It != PerBlockAccesses.end()) {
992  AccessList *Accesses = It->second.get();
993  for (MemoryAccess &L : *Accesses) {
994  if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(&L)) {
995  if (MUD->getDefiningAccess() == nullptr || RenameAllUses)
996  MUD->setDefiningAccess(IncomingVal);
997  if (isa<MemoryDef>(&L))
998  IncomingVal = &L;
999  } else {
1000  IncomingVal = &L;
1001  }
1002  }
1003  }
1004  return IncomingVal;
1005 }
1006 
1007 /// This is the standard SSA renaming algorithm.
1008 ///
1009 /// We walk the dominator tree in preorder, renaming accesses, and then filling
1010 /// in phi nodes in our successors.
1011 void MemorySSA::renamePass(DomTreeNode *Root, MemoryAccess *IncomingVal,
1013  bool SkipVisited, bool RenameAllUses) {
1015  // Skip everything if we already renamed this block and we are skipping.
1016  // Note: You can't sink this into the if, because we need it to occur
1017  // regardless of whether we skip blocks or not.
1018  bool AlreadyVisited = !Visited.insert(Root->getBlock()).second;
1019  if (SkipVisited && AlreadyVisited)
1020  return;
1021 
1022  IncomingVal = renameBlock(Root->getBlock(), IncomingVal, RenameAllUses);
1023  renameSuccessorPhis(Root->getBlock(), IncomingVal, RenameAllUses);
1024  WorkStack.push_back({Root, Root->begin(), IncomingVal});
1025 
1026  while (!WorkStack.empty()) {
1027  DomTreeNode *Node = WorkStack.back().DTN;
1028  DomTreeNode::const_iterator ChildIt = WorkStack.back().ChildIt;
1029  IncomingVal = WorkStack.back().IncomingVal;
1030 
1031  if (ChildIt == Node->end()) {
1032  WorkStack.pop_back();
1033  } else {
1034  DomTreeNode *Child = *ChildIt;
1035  ++WorkStack.back().ChildIt;
1036  BasicBlock *BB = Child->getBlock();
1037  // Note: You can't sink this into the if, because we need it to occur
1038  // regardless of whether we skip blocks or not.
1039  AlreadyVisited = !Visited.insert(BB).second;
1040  if (SkipVisited && AlreadyVisited) {
1041  // We already visited this during our renaming, which can happen when
1042  // being asked to rename multiple blocks. Figure out the incoming val,
1043  // which is the last def.
1044  // Incoming value can only change if there is a block def, and in that
1045  // case, it's the last block def in the list.
1046  if (auto *BlockDefs = getWritableBlockDefs(BB))
1047  IncomingVal = &*BlockDefs->rbegin();
1048  } else
1049  IncomingVal = renameBlock(BB, IncomingVal, RenameAllUses);
1050  renameSuccessorPhis(BB, IncomingVal, RenameAllUses);
1051  WorkStack.push_back({Child, Child->begin(), IncomingVal});
1052  }
1053  }
1054 }
1055 
1056 /// This handles unreachable block accesses by deleting phi nodes in
1057 /// unreachable blocks, and marking all other unreachable MemoryAccess's as
1058 /// being uses of the live on entry definition.
1059 void MemorySSA::markUnreachableAsLiveOnEntry(BasicBlock *BB) {
1060  assert(!DT->isReachableFromEntry(BB) &&
1061  "Reachable block found while handling unreachable blocks");
1062 
1063  // Make sure phi nodes in our reachable successors end up with a
1064  // LiveOnEntryDef for our incoming edge, even though our block is forward
1065  // unreachable. We could just disconnect these blocks from the CFG fully,
1066  // but we do not right now.
1067  for (const BasicBlock *S : successors(BB)) {
1068  if (!DT->isReachableFromEntry(S))
1069  continue;
1070  auto It = PerBlockAccesses.find(S);
1071  // Rename the phi nodes in our successor block
1072  if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
1073  continue;
1074  AccessList *Accesses = It->second.get();
1075  auto *Phi = cast<MemoryPhi>(&Accesses->front());
1076  Phi->addIncoming(LiveOnEntryDef.get(), BB);
1077  }
1078 
1079  auto It = PerBlockAccesses.find(BB);
1080  if (It == PerBlockAccesses.end())
1081  return;
1082 
1083  auto &Accesses = It->second;
1084  for (auto AI = Accesses->begin(), AE = Accesses->end(); AI != AE;) {
1085  auto Next = std::next(AI);
1086  // If we have a phi, just remove it. We are going to replace all
1087  // users with live on entry.
1088  if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(AI))
1089  UseOrDef->setDefiningAccess(LiveOnEntryDef.get());
1090  else
1091  Accesses->erase(AI);
1092  AI = Next;
1093  }
1094 }
1095 
1097  : AA(AA), DT(DT), F(Func), LiveOnEntryDef(nullptr), Walker(nullptr),
1098  NextID(0) {
1099  buildMemorySSA();
1100 }
1101 
1103  // Drop all our references
1104  for (const auto &Pair : PerBlockAccesses)
1105  for (MemoryAccess &MA : *Pair.second)
1106  MA.dropAllReferences();
1107 }
1108 
1109 MemorySSA::AccessList *MemorySSA::getOrCreateAccessList(const BasicBlock *BB) {
1110  auto Res = PerBlockAccesses.insert(std::make_pair(BB, nullptr));
1111 
1112  if (Res.second)
1113  Res.first->second = llvm::make_unique<AccessList>();
1114  return Res.first->second.get();
1115 }
1116 
1117 MemorySSA::DefsList *MemorySSA::getOrCreateDefsList(const BasicBlock *BB) {
1118  auto Res = PerBlockDefs.insert(std::make_pair(BB, nullptr));
1119 
1120  if (Res.second)
1121  Res.first->second = llvm::make_unique<DefsList>();
1122  return Res.first->second.get();
1123 }
1124 
1125 namespace llvm {
1126 
1127 /// This class is a batch walker of all MemoryUse's in the program, and points
1128 /// their defining access at the thing that actually clobbers them. Because it
1129 /// is a batch walker that touches everything, it does not operate like the
1130 /// other walkers. This walker is basically performing a top-down SSA renaming
1131 /// pass, where the version stack is used as the cache. This enables it to be
1132 /// significantly more time and memory efficient than using the regular walker,
1133 /// which is walking bottom-up.
1135 public:
1137  DominatorTree *DT)
1138  : MSSA(MSSA), Walker(Walker), AA(AA), DT(DT) {
1139  Walker = MSSA->getWalker();
1140  }
1141 
1142  void optimizeUses();
1143 
1144 private:
1145  /// This represents where a given memorylocation is in the stack.
1146  struct MemlocStackInfo {
1147  // This essentially is keeping track of versions of the stack. Whenever
1148  // the stack changes due to pushes or pops, these versions increase.
1149  unsigned long StackEpoch;
1150  unsigned long PopEpoch;
1151  // This is the lower bound of places on the stack to check. It is equal to
1152  // the place the last stack walk ended.
1153  // Note: Correctness depends on this being initialized to 0, which densemap
1154  // does
1155  unsigned long LowerBound;
1156  const BasicBlock *LowerBoundBlock;
1157  // This is where the last walk for this memory location ended.
1158  unsigned long LastKill;
1159  bool LastKillValid;
1161  };
1162 
1163  void optimizeUsesInBlock(const BasicBlock *, unsigned long &, unsigned long &,
1166 
1167  MemorySSA *MSSA;
1168  MemorySSAWalker *Walker;
1169  AliasAnalysis *AA;
1170  DominatorTree *DT;
1171 };
1172 
1173 } // end namespace llvm
1174 
1175 /// Optimize the uses in a given block This is basically the SSA renaming
1176 /// algorithm, with one caveat: We are able to use a single stack for all
1177 /// MemoryUses. This is because the set of *possible* reaching MemoryDefs is
1178 /// the same for every MemoryUse. The *actual* clobbering MemoryDef is just
1179 /// going to be some position in that stack of possible ones.
1180 ///
1181 /// We track the stack positions that each MemoryLocation needs
1182 /// to check, and last ended at. This is because we only want to check the
1183 /// things that changed since last time. The same MemoryLocation should
1184 /// get clobbered by the same store (getModRefInfo does not use invariantness or
1185 /// things like this, and if they start, we can modify MemoryLocOrCall to
1186 /// include relevant data)
1187 void MemorySSA::OptimizeUses::optimizeUsesInBlock(
1188  const BasicBlock *BB, unsigned long &StackEpoch, unsigned long &PopEpoch,
1189  SmallVectorImpl<MemoryAccess *> &VersionStack,
1191 
1192  /// If no accesses, nothing to do.
1193  MemorySSA::AccessList *Accesses = MSSA->getWritableBlockAccesses(BB);
1194  if (Accesses == nullptr)
1195  return;
1196 
1197  // Pop everything that doesn't dominate the current block off the stack,
1198  // increment the PopEpoch to account for this.
1199  while (true) {
1200  assert(
1201  !VersionStack.empty() &&
1202  "Version stack should have liveOnEntry sentinel dominating everything");
1203  BasicBlock *BackBlock = VersionStack.back()->getBlock();
1204  if (DT->dominates(BackBlock, BB))
1205  break;
1206  while (VersionStack.back()->getBlock() == BackBlock)
1207  VersionStack.pop_back();
1208  ++PopEpoch;
1209  }
1210 
1211  for (MemoryAccess &MA : *Accesses) {
1212  auto *MU = dyn_cast<MemoryUse>(&MA);
1213  if (!MU) {
1214  VersionStack.push_back(&MA);
1215  ++StackEpoch;
1216  continue;
1217  }
1218 
1219  if (isUseTriviallyOptimizableToLiveOnEntry(*AA, MU->getMemoryInst())) {
1220  MU->setDefiningAccess(MSSA->getLiveOnEntryDef(), true, None);
1221  continue;
1222  }
1223 
1224  MemoryLocOrCall UseMLOC(MU);
1225  auto &LocInfo = LocStackInfo[UseMLOC];
1226  // If the pop epoch changed, it means we've removed stuff from top of
1227  // stack due to changing blocks. We may have to reset the lower bound or
1228  // last kill info.
1229  if (LocInfo.PopEpoch != PopEpoch) {
1230  LocInfo.PopEpoch = PopEpoch;
1231  LocInfo.StackEpoch = StackEpoch;
1232  // If the lower bound was in something that no longer dominates us, we
1233  // have to reset it.
1234  // We can't simply track stack size, because the stack may have had
1235  // pushes/pops in the meantime.
1236  // XXX: This is non-optimal, but only is slower cases with heavily
1237  // branching dominator trees. To get the optimal number of queries would
1238  // be to make lowerbound and lastkill a per-loc stack, and pop it until
1239  // the top of that stack dominates us. This does not seem worth it ATM.
1240  // A much cheaper optimization would be to always explore the deepest
1241  // branch of the dominator tree first. This will guarantee this resets on
1242  // the smallest set of blocks.
1243  if (LocInfo.LowerBoundBlock && LocInfo.LowerBoundBlock != BB &&
1244  !DT->dominates(LocInfo.LowerBoundBlock, BB)) {
1245  // Reset the lower bound of things to check.
1246  // TODO: Some day we should be able to reset to last kill, rather than
1247  // 0.
1248  LocInfo.LowerBound = 0;
1249  LocInfo.LowerBoundBlock = VersionStack[0]->getBlock();
1250  LocInfo.LastKillValid = false;
1251  }
1252  } else if (LocInfo.StackEpoch != StackEpoch) {
1253  // If all that has changed is the StackEpoch, we only have to check the
1254  // new things on the stack, because we've checked everything before. In
1255  // this case, the lower bound of things to check remains the same.
1256  LocInfo.PopEpoch = PopEpoch;
1257  LocInfo.StackEpoch = StackEpoch;
1258  }
1259  if (!LocInfo.LastKillValid) {
1260  LocInfo.LastKill = VersionStack.size() - 1;
1261  LocInfo.LastKillValid = true;
1262  LocInfo.AR = MayAlias;
1263  }
1264 
1265  // At this point, we should have corrected last kill and LowerBound to be
1266  // in bounds.
1267  assert(LocInfo.LowerBound < VersionStack.size() &&
1268  "Lower bound out of range");
1269  assert(LocInfo.LastKill < VersionStack.size() &&
1270  "Last kill info out of range");
1271  // In any case, the new upper bound is the top of the stack.
1272  unsigned long UpperBound = VersionStack.size() - 1;
1273 
1274  if (UpperBound - LocInfo.LowerBound > MaxCheckLimit) {
1275  LLVM_DEBUG(dbgs() << "MemorySSA skipping optimization of " << *MU << " ("
1276  << *(MU->getMemoryInst()) << ")"
1277  << " because there are "
1278  << UpperBound - LocInfo.LowerBound
1279  << " stores to disambiguate\n");
1280  // Because we did not walk, LastKill is no longer valid, as this may
1281  // have been a kill.
1282  LocInfo.LastKillValid = false;
1283  continue;
1284  }
1285  bool FoundClobberResult = false;
1286  while (UpperBound > LocInfo.LowerBound) {
1287  if (isa<MemoryPhi>(VersionStack[UpperBound])) {
1288  // For phis, use the walker, see where we ended up, go there
1289  Instruction *UseInst = MU->getMemoryInst();
1290  MemoryAccess *Result = Walker->getClobberingMemoryAccess(UseInst);
1291  // We are guaranteed to find it or something is wrong
1292  while (VersionStack[UpperBound] != Result) {
1293  assert(UpperBound != 0);
1294  --UpperBound;
1295  }
1296  FoundClobberResult = true;
1297  break;
1298  }
1299 
1300  MemoryDef *MD = cast<MemoryDef>(VersionStack[UpperBound]);
1301  // If the lifetime of the pointer ends at this instruction, it's live on
1302  // entry.
1303  if (!UseMLOC.IsCall && lifetimeEndsAt(MD, UseMLOC.getLoc(), *AA)) {
1304  // Reset UpperBound to liveOnEntryDef's place in the stack
1305  UpperBound = 0;
1306  FoundClobberResult = true;
1307  LocInfo.AR = MustAlias;
1308  break;
1309  }
1310  ClobberAlias CA = instructionClobbersQuery(MD, MU, UseMLOC, *AA);
1311  if (CA.IsClobber) {
1312  FoundClobberResult = true;
1313  LocInfo.AR = CA.AR;
1314  break;
1315  }
1316  --UpperBound;
1317  }
1318 
1319  // Note: Phis always have AliasResult AR set to MayAlias ATM.
1320 
1321  // At the end of this loop, UpperBound is either a clobber, or lower bound
1322  // PHI walking may cause it to be < LowerBound, and in fact, < LastKill.
1323  if (FoundClobberResult || UpperBound < LocInfo.LastKill) {
1324  // We were last killed now by where we got to
1325  if (MSSA->isLiveOnEntryDef(VersionStack[UpperBound]))
1326  LocInfo.AR = None;
1327  MU->setDefiningAccess(VersionStack[UpperBound], true, LocInfo.AR);
1328  LocInfo.LastKill = UpperBound;
1329  } else {
1330  // Otherwise, we checked all the new ones, and now we know we can get to
1331  // LastKill.
1332  MU->setDefiningAccess(VersionStack[LocInfo.LastKill], true, LocInfo.AR);
1333  }
1334  LocInfo.LowerBound = VersionStack.size() - 1;
1335  LocInfo.LowerBoundBlock = BB;
1336  }
1337 }
1338 
1339 /// Optimize uses to point to their actual clobbering definitions.
1341  SmallVector<MemoryAccess *, 16> VersionStack;
1343  VersionStack.push_back(MSSA->getLiveOnEntryDef());
1344 
1345  unsigned long StackEpoch = 1;
1346  unsigned long PopEpoch = 1;
1347  // We perform a non-recursive top-down dominator tree walk.
1348  for (const auto *DomNode : depth_first(DT->getRootNode()))
1349  optimizeUsesInBlock(DomNode->getBlock(), StackEpoch, PopEpoch, VersionStack,
1350  LocStackInfo);
1351 }
1352 
1353 void MemorySSA::placePHINodes(
1354  const SmallPtrSetImpl<BasicBlock *> &DefiningBlocks) {
1355  // Determine where our MemoryPhi's should go
1356  ForwardIDFCalculator IDFs(*DT);
1357  IDFs.setDefiningBlocks(DefiningBlocks);
1359  IDFs.calculate(IDFBlocks);
1360 
1361  // Now place MemoryPhi nodes.
1362  for (auto &BB : IDFBlocks)
1363  createMemoryPhi(BB);
1364 }
1365 
1366 void MemorySSA::buildMemorySSA() {
1367  // We create an access to represent "live on entry", for things like
1368  // arguments or users of globals, where the memory they use is defined before
1369  // the beginning of the function. We do not actually insert it into the IR.
1370  // We do not define a live on exit for the immediate uses, and thus our
1371  // semantics do *not* imply that something with no immediate uses can simply
1372  // be removed.
1373  BasicBlock &StartingPoint = F.getEntryBlock();
1374  LiveOnEntryDef.reset(new MemoryDef(F.getContext(), nullptr, nullptr,
1375  &StartingPoint, NextID++));
1376 
1377  // We maintain lists of memory accesses per-block, trading memory for time. We
1378  // could just look up the memory access for every possible instruction in the
1379  // stream.
1380  SmallPtrSet<BasicBlock *, 32> DefiningBlocks;
1381  // Go through each block, figure out where defs occur, and chain together all
1382  // the accesses.
1383  for (BasicBlock &B : F) {
1384  bool InsertIntoDef = false;
1385  AccessList *Accesses = nullptr;
1386  DefsList *Defs = nullptr;
1387  for (Instruction &I : B) {
1388  MemoryUseOrDef *MUD = createNewAccess(&I);
1389  if (!MUD)
1390  continue;
1391 
1392  if (!Accesses)
1393  Accesses = getOrCreateAccessList(&B);
1394  Accesses->push_back(MUD);
1395  if (isa<MemoryDef>(MUD)) {
1396  InsertIntoDef = true;
1397  if (!Defs)
1398  Defs = getOrCreateDefsList(&B);
1399  Defs->push_back(*MUD);
1400  }
1401  }
1402  if (InsertIntoDef)
1403  DefiningBlocks.insert(&B);
1404  }
1405  placePHINodes(DefiningBlocks);
1406 
1407  // Now do regular SSA renaming on the MemoryDef/MemoryUse. Visited will get
1408  // filled in with all blocks.
1410  renamePass(DT->getRootNode(), LiveOnEntryDef.get(), Visited);
1411 
1412  CachingWalker *Walker = getWalkerImpl();
1413 
1414  OptimizeUses(this, Walker, AA, DT).optimizeUses();
1415 
1416  // Mark the uses in unreachable blocks as live on entry, so that they go
1417  // somewhere.
1418  for (auto &BB : F)
1419  if (!Visited.count(&BB))
1420  markUnreachableAsLiveOnEntry(&BB);
1421 }
1422 
1423 MemorySSAWalker *MemorySSA::getWalker() { return getWalkerImpl(); }
1424 
1425 MemorySSA::CachingWalker *MemorySSA::getWalkerImpl() {
1426  if (Walker)
1427  return Walker.get();
1428 
1429  Walker = llvm::make_unique<CachingWalker>(this, AA, DT);
1430  return Walker.get();
1431 }
1432 
1433 // This is a helper function used by the creation routines. It places NewAccess
1434 // into the access and defs lists for a given basic block, at the given
1435 // insertion point.
1437  const BasicBlock *BB,
1438  InsertionPlace Point) {
1439  auto *Accesses = getOrCreateAccessList(BB);
1440  if (Point == Beginning) {
1441  // If it's a phi node, it goes first, otherwise, it goes after any phi
1442  // nodes.
1443  if (isa<MemoryPhi>(NewAccess)) {
1444  Accesses->push_front(NewAccess);
1445  auto *Defs = getOrCreateDefsList(BB);
1446  Defs->push_front(*NewAccess);
1447  } else {
1448  auto AI = find_if_not(
1449  *Accesses, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
1450  Accesses->insert(AI, NewAccess);
1451  if (!isa<MemoryUse>(NewAccess)) {
1452  auto *Defs = getOrCreateDefsList(BB);
1453  auto DI = find_if_not(
1454  *Defs, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
1455  Defs->insert(DI, *NewAccess);
1456  }
1457  }
1458  } else {
1459  Accesses->push_back(NewAccess);
1460  if (!isa<MemoryUse>(NewAccess)) {
1461  auto *Defs = getOrCreateDefsList(BB);
1462  Defs->push_back(*NewAccess);
1463  }
1464  }
1465  BlockNumberingValid.erase(BB);
1466 }
1467 
1469  AccessList::iterator InsertPt) {
1470  auto *Accesses = getWritableBlockAccesses(BB);
1471  bool WasEnd = InsertPt == Accesses->end();
1472  Accesses->insert(AccessList::iterator(InsertPt), What);
1473  if (!isa<MemoryUse>(What)) {
1474  auto *Defs = getOrCreateDefsList(BB);
1475  // If we got asked to insert at the end, we have an easy job, just shove it
1476  // at the end. If we got asked to insert before an existing def, we also get
1477  // an iterator. If we got asked to insert before a use, we have to hunt for
1478  // the next def.
1479  if (WasEnd) {
1480  Defs->push_back(*What);
1481  } else if (isa<MemoryDef>(InsertPt)) {
1482  Defs->insert(InsertPt->getDefsIterator(), *What);
1483  } else {
1484  while (InsertPt != Accesses->end() && !isa<MemoryDef>(InsertPt))
1485  ++InsertPt;
1486  // Either we found a def, or we are inserting at the end
1487  if (InsertPt == Accesses->end())
1488  Defs->push_back(*What);
1489  else
1490  Defs->insert(InsertPt->getDefsIterator(), *What);
1491  }
1492  }
1493  BlockNumberingValid.erase(BB);
1494 }
1495 
1496 void MemorySSA::prepareForMoveTo(MemoryAccess *What, BasicBlock *BB) {
1497  // Keep it in the lookup tables, remove from the lists
1498  removeFromLists(What, false);
1499 
1500  // Note that moving should implicitly invalidate the optimized state of a
1501  // MemoryUse (and Phis can't be optimized). However, it doesn't do so for a
1502  // MemoryDef.
1503  if (auto *MD = dyn_cast<MemoryDef>(What))
1504  MD->resetOptimized();
1505  What->setBlock(BB);
1506 }
1507 
1508 // Move What before Where in the IR. The end result is that What will belong to
1509 // the right lists and have the right Block set, but will not otherwise be
1510 // correct. It will not have the right defining access, and if it is a def,
1511 // things below it will not properly be updated.
1513  AccessList::iterator Where) {
1514  prepareForMoveTo(What, BB);
1515  insertIntoListsBefore(What, BB, Where);
1516 }
1517 
1519  InsertionPlace Point) {
1520  if (isa<MemoryPhi>(What)) {
1521  assert(Point == Beginning &&
1522  "Can only move a Phi at the beginning of the block");
1523  // Update lookup table entry
1524  ValueToMemoryAccess.erase(What->getBlock());
1525  bool Inserted = ValueToMemoryAccess.insert({BB, What}).second;
1526  (void)Inserted;
1527  assert(Inserted && "Cannot move a Phi to a block that already has one");
1528  }
1529 
1530  prepareForMoveTo(What, BB);
1531  insertIntoListsForBlock(What, BB, Point);
1532 }
1533 
1534 MemoryPhi *MemorySSA::createMemoryPhi(BasicBlock *BB) {
1535  assert(!getMemoryAccess(BB) && "MemoryPhi already exists for this BB");
1536  MemoryPhi *Phi = new MemoryPhi(BB->getContext(), BB, NextID++);
1537  // Phi's always are placed at the front of the block.
1539  ValueToMemoryAccess[BB] = Phi;
1540  return Phi;
1541 }
1542 
1544  MemoryAccess *Definition,
1545  const MemoryUseOrDef *Template) {
1546  assert(!isa<PHINode>(I) && "Cannot create a defined access for a PHI");
1547  MemoryUseOrDef *NewAccess = createNewAccess(I, Template);
1548  assert(
1549  NewAccess != nullptr &&
1550  "Tried to create a memory access for a non-memory touching instruction");
1551  NewAccess->setDefiningAccess(Definition);
1552  return NewAccess;
1553 }
1554 
1555 // Return true if the instruction has ordering constraints.
1556 // Note specifically that this only considers stores and loads
1557 // because others are still considered ModRef by getModRefInfo.
1558 static inline bool isOrdered(const Instruction *I) {
1559  if (auto *SI = dyn_cast<StoreInst>(I)) {
1560  if (!SI->isUnordered())
1561  return true;
1562  } else if (auto *LI = dyn_cast<LoadInst>(I)) {
1563  if (!LI->isUnordered())
1564  return true;
1565  }
1566  return false;
1567 }
1568 
1569 /// Helper function to create new memory accesses
1570 MemoryUseOrDef *MemorySSA::createNewAccess(Instruction *I,
1571  const MemoryUseOrDef *Template) {
1572  // The assume intrinsic has a control dependency which we model by claiming
1573  // that it writes arbitrarily. Ignore that fake memory dependency here.
1574  // FIXME: Replace this special casing with a more accurate modelling of
1575  // assume's control dependency.
1576  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1577  if (II->getIntrinsicID() == Intrinsic::assume)
1578  return nullptr;
1579 
1580  bool Def, Use;
1581  if (Template) {
1582  Def = dyn_cast_or_null<MemoryDef>(Template) != nullptr;
1583  Use = dyn_cast_or_null<MemoryUse>(Template) != nullptr;
1584 #if !defined(NDEBUG)
1585  ModRefInfo ModRef = AA->getModRefInfo(I, None);
1586  bool DefCheck, UseCheck;
1587  DefCheck = isModSet(ModRef) || isOrdered(I);
1588  UseCheck = isRefSet(ModRef);
1589  assert(Def == DefCheck && (Def || Use == UseCheck) && "Invalid template");
1590 #endif
1591  } else {
1592  // Find out what affect this instruction has on memory.
1593  ModRefInfo ModRef = AA->getModRefInfo(I, None);
1594  // The isOrdered check is used to ensure that volatiles end up as defs
1595  // (atomics end up as ModRef right now anyway). Until we separate the
1596  // ordering chain from the memory chain, this enables people to see at least
1597  // some relative ordering to volatiles. Note that getClobberingMemoryAccess
1598  // will still give an answer that bypasses other volatile loads. TODO:
1599  // Separate memory aliasing and ordering into two different chains so that
1600  // we can precisely represent both "what memory will this read/write/is
1601  // clobbered by" and "what instructions can I move this past".
1602  Def = isModSet(ModRef) || isOrdered(I);
1603  Use = isRefSet(ModRef);
1604  }
1605 
1606  // It's possible for an instruction to not modify memory at all. During
1607  // construction, we ignore them.
1608  if (!Def && !Use)
1609  return nullptr;
1610 
1611  MemoryUseOrDef *MUD;
1612  if (Def)
1613  MUD = new MemoryDef(I->getContext(), nullptr, I, I->getParent(), NextID++);
1614  else
1615  MUD = new MemoryUse(I->getContext(), nullptr, I, I->getParent());
1616  ValueToMemoryAccess[I] = MUD;
1617  return MUD;
1618 }
1619 
1620 /// Returns true if \p Replacer dominates \p Replacee .
1621 bool MemorySSA::dominatesUse(const MemoryAccess *Replacer,
1622  const MemoryAccess *Replacee) const {
1623  if (isa<MemoryUseOrDef>(Replacee))
1624  return DT->dominates(Replacer->getBlock(), Replacee->getBlock());
1625  const auto *MP = cast<MemoryPhi>(Replacee);
1626  // For a phi node, the use occurs in the predecessor block of the phi node.
1627  // Since we may occur multiple times in the phi node, we have to check each
1628  // operand to ensure Replacer dominates each operand where Replacee occurs.
1629  for (const Use &Arg : MP->operands()) {
1630  if (Arg.get() != Replacee &&
1631  !DT->dominates(Replacer->getBlock(), MP->getIncomingBlock(Arg)))
1632  return false;
1633  }
1634  return true;
1635 }
1636 
1637 /// Properly remove \p MA from all of MemorySSA's lookup tables.
1639  assert(MA->use_empty() &&
1640  "Trying to remove memory access that still has uses");
1641  BlockNumbering.erase(MA);
1642  if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
1643  MUD->setDefiningAccess(nullptr);
1644  // Invalidate our walker's cache if necessary
1645  if (!isa<MemoryUse>(MA))
1646  Walker->invalidateInfo(MA);
1647 
1648  Value *MemoryInst;
1649  if (const auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
1650  MemoryInst = MUD->getMemoryInst();
1651  else
1652  MemoryInst = MA->getBlock();
1653 
1654  auto VMA = ValueToMemoryAccess.find(MemoryInst);
1655  if (VMA->second == MA)
1656  ValueToMemoryAccess.erase(VMA);
1657 }
1658 
1659 /// Properly remove \p MA from all of MemorySSA's lists.
1660 ///
1661 /// Because of the way the intrusive list and use lists work, it is important to
1662 /// do removal in the right order.
1663 /// ShouldDelete defaults to true, and will cause the memory access to also be
1664 /// deleted, not just removed.
1665 void MemorySSA::removeFromLists(MemoryAccess *MA, bool ShouldDelete) {
1666  BasicBlock *BB = MA->getBlock();
1667  // The access list owns the reference, so we erase it from the non-owning list
1668  // first.
1669  if (!isa<MemoryUse>(MA)) {
1670  auto DefsIt = PerBlockDefs.find(BB);
1671  std::unique_ptr<DefsList> &Defs = DefsIt->second;
1672  Defs->remove(*MA);
1673  if (Defs->empty())
1674  PerBlockDefs.erase(DefsIt);
1675  }
1676 
1677  // The erase call here will delete it. If we don't want it deleted, we call
1678  // remove instead.
1679  auto AccessIt = PerBlockAccesses.find(BB);
1680  std::unique_ptr<AccessList> &Accesses = AccessIt->second;
1681  if (ShouldDelete)
1682  Accesses->erase(MA);
1683  else
1684  Accesses->remove(MA);
1685 
1686  if (Accesses->empty()) {
1687  PerBlockAccesses.erase(AccessIt);
1688  BlockNumberingValid.erase(BB);
1689  }
1690 }
1691 
1693  MemorySSAAnnotatedWriter Writer(this);
1694  F.print(OS, &Writer);
1695 }
1696 
1697 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1699 #endif
1700 
1702  verifyDefUses(F);
1704  verifyOrdering(F);
1706  Walker->verify(this);
1708 }
1709 
1710 /// Check sanity of the clobbering instruction for access MA.
1712  if (const auto *MUD = dyn_cast<MemoryUseOrDef>(MA)) {
1713  if (!MUD->isOptimized())
1714  return;
1715  auto *I = MUD->getMemoryInst();
1716  auto Loc = MemoryLocation::getOrNone(I);
1717  if (Loc == None)
1718  return;
1719  auto *Clobber = MUD->getOptimized();
1720  UpwardsMemoryQuery Q(I, MUD);
1721  checkClobberSanity(MUD, Clobber, *Loc, *this, Q, *AA, true);
1722  }
1723 }
1724 
1726 #if !defined(NDEBUG) && defined(EXPENSIVE_CHECKS)
1727  for (const BasicBlock &BB : F) {
1728  const AccessList *Accesses = getBlockAccesses(&BB);
1729  if (!Accesses)
1730  continue;
1731  for (const MemoryAccess &MA : *Accesses)
1733  }
1734 #endif
1735 }
1736 
1737 /// Verify that all of the blocks we believe to have valid domination numbers
1738 /// actually have valid domination numbers.
1740 #ifndef NDEBUG
1741  if (BlockNumberingValid.empty())
1742  return;
1743 
1744  SmallPtrSet<const BasicBlock *, 16> ValidBlocks = BlockNumberingValid;
1745  for (const BasicBlock &BB : F) {
1746  if (!ValidBlocks.count(&BB))
1747  continue;
1748 
1749  ValidBlocks.erase(&BB);
1750 
1751  const AccessList *Accesses = getBlockAccesses(&BB);
1752  // It's correct to say an empty block has valid numbering.
1753  if (!Accesses)
1754  continue;
1755 
1756  // Block numbering starts at 1.
1757  unsigned long LastNumber = 0;
1758  for (const MemoryAccess &MA : *Accesses) {
1759  auto ThisNumberIter = BlockNumbering.find(&MA);
1760  assert(ThisNumberIter != BlockNumbering.end() &&
1761  "MemoryAccess has no domination number in a valid block!");
1762 
1763  unsigned long ThisNumber = ThisNumberIter->second;
1764  assert(ThisNumber > LastNumber &&
1765  "Domination numbers should be strictly increasing!");
1766  LastNumber = ThisNumber;
1767  }
1768  }
1769 
1770  assert(ValidBlocks.empty() &&
1771  "All valid BasicBlocks should exist in F -- dangling pointers?");
1772 #endif
1773 }
1774 
1775 /// Verify that the order and existence of MemoryAccesses matches the
1776 /// order and existence of memory affecting instructions.
1778 #ifndef NDEBUG
1779  // Walk all the blocks, comparing what the lookups think and what the access
1780  // lists think, as well as the order in the blocks vs the order in the access
1781  // lists.
1782  SmallVector<MemoryAccess *, 32> ActualAccesses;
1784  for (BasicBlock &B : F) {
1785  const AccessList *AL = getBlockAccesses(&B);
1786  const auto *DL = getBlockDefs(&B);
1787  MemoryAccess *Phi = getMemoryAccess(&B);
1788  if (Phi) {
1789  ActualAccesses.push_back(Phi);
1790  ActualDefs.push_back(Phi);
1791  }
1792 
1793  for (Instruction &I : B) {
1794  MemoryAccess *MA = getMemoryAccess(&I);
1795  assert((!MA || (AL && (isa<MemoryUse>(MA) || DL))) &&
1796  "We have memory affecting instructions "
1797  "in this block but they are not in the "
1798  "access list or defs list");
1799  if (MA) {
1800  ActualAccesses.push_back(MA);
1801  if (isa<MemoryDef>(MA))
1802  ActualDefs.push_back(MA);
1803  }
1804  }
1805  // Either we hit the assert, really have no accesses, or we have both
1806  // accesses and an access list.
1807  // Same with defs.
1808  if (!AL && !DL)
1809  continue;
1810  assert(AL->size() == ActualAccesses.size() &&
1811  "We don't have the same number of accesses in the block as on the "
1812  "access list");
1813  assert((DL || ActualDefs.size() == 0) &&
1814  "Either we should have a defs list, or we should have no defs");
1815  assert((!DL || DL->size() == ActualDefs.size()) &&
1816  "We don't have the same number of defs in the block as on the "
1817  "def list");
1818  auto ALI = AL->begin();
1819  auto AAI = ActualAccesses.begin();
1820  while (ALI != AL->end() && AAI != ActualAccesses.end()) {
1821  assert(&*ALI == *AAI && "Not the same accesses in the same order");
1822  ++ALI;
1823  ++AAI;
1824  }
1825  ActualAccesses.clear();
1826  if (DL) {
1827  auto DLI = DL->begin();
1828  auto ADI = ActualDefs.begin();
1829  while (DLI != DL->end() && ADI != ActualDefs.end()) {
1830  assert(&*DLI == *ADI && "Not the same defs in the same order");
1831  ++DLI;
1832  ++ADI;
1833  }
1834  }
1835  ActualDefs.clear();
1836  }
1837 #endif
1838 }
1839 
1840 /// Verify the domination properties of MemorySSA by checking that each
1841 /// definition dominates all of its uses.
1843 #ifndef NDEBUG
1844  for (BasicBlock &B : F) {
1845  // Phi nodes are attached to basic blocks
1846  if (MemoryPhi *MP = getMemoryAccess(&B))
1847  for (const Use &U : MP->uses())
1848  assert(dominates(MP, U) && "Memory PHI does not dominate it's uses");
1849 
1850  for (Instruction &I : B) {
1851  MemoryAccess *MD = dyn_cast_or_null<MemoryDef>(getMemoryAccess(&I));
1852  if (!MD)
1853  continue;
1854 
1855  for (const Use &U : MD->uses())
1856  assert(dominates(MD, U) && "Memory Def does not dominate it's uses");
1857  }
1858  }
1859 #endif
1860 }
1861 
1862 /// Verify the def-use lists in MemorySSA, by verifying that \p Use
1863 /// appears in the use list of \p Def.
1864 void MemorySSA::verifyUseInDefs(MemoryAccess *Def, MemoryAccess *Use) const {
1865 #ifndef NDEBUG
1866  // The live on entry use may cause us to get a NULL def here
1867  if (!Def)
1868  assert(isLiveOnEntryDef(Use) &&
1869  "Null def but use not point to live on entry def");
1870  else
1871  assert(is_contained(Def->users(), Use) &&
1872  "Did not find use in def's use list");
1873 #endif
1874 }
1875 
1876 /// Verify the immediate use information, by walking all the memory
1877 /// accesses and verifying that, for each use, it appears in the
1878 /// appropriate def's use list
1880 #ifndef NDEBUG
1881  for (BasicBlock &B : F) {
1882  // Phi nodes are attached to basic blocks
1883  if (MemoryPhi *Phi = getMemoryAccess(&B)) {
1884  assert(Phi->getNumOperands() == static_cast<unsigned>(std::distance(
1885  pred_begin(&B), pred_end(&B))) &&
1886  "Incomplete MemoryPhi Node");
1887  for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) {
1888  verifyUseInDefs(Phi->getIncomingValue(I), Phi);
1889  assert(find(predecessors(&B), Phi->getIncomingBlock(I)) !=
1890  pred_end(&B) &&
1891  "Incoming phi block not a block predecessor");
1892  }
1893  }
1894 
1895  for (Instruction &I : B) {
1896  if (MemoryUseOrDef *MA = getMemoryAccess(&I)) {
1897  verifyUseInDefs(MA->getDefiningAccess(), MA);
1898  }
1899  }
1900  }
1901 #endif
1902 }
1903 
1904 /// Perform a local numbering on blocks so that instruction ordering can be
1905 /// determined in constant time.
1906 /// TODO: We currently just number in order. If we numbered by N, we could
1907 /// allow at least N-1 sequences of insertBefore or insertAfter (and at least
1908 /// log2(N) sequences of mixed before and after) without needing to invalidate
1909 /// the numbering.
1910 void MemorySSA::renumberBlock(const BasicBlock *B) const {
1911  // The pre-increment ensures the numbers really start at 1.
1912  unsigned long CurrentNumber = 0;
1913  const AccessList *AL = getBlockAccesses(B);
1914  assert(AL != nullptr && "Asking to renumber an empty block");
1915  for (const auto &I : *AL)
1916  BlockNumbering[&I] = ++CurrentNumber;
1917  BlockNumberingValid.insert(B);
1918 }
1919 
1920 /// Determine, for two memory accesses in the same block,
1921 /// whether \p Dominator dominates \p Dominatee.
1922 /// \returns True if \p Dominator dominates \p Dominatee.
1924  const MemoryAccess *Dominatee) const {
1925  const BasicBlock *DominatorBlock = Dominator->getBlock();
1926 
1927  assert((DominatorBlock == Dominatee->getBlock()) &&
1928  "Asking for local domination when accesses are in different blocks!");
1929  // A node dominates itself.
1930  if (Dominatee == Dominator)
1931  return true;
1932 
1933  // When Dominatee is defined on function entry, it is not dominated by another
1934  // memory access.
1935  if (isLiveOnEntryDef(Dominatee))
1936  return false;
1937 
1938  // When Dominator is defined on function entry, it dominates the other memory
1939  // access.
1940  if (isLiveOnEntryDef(Dominator))
1941  return true;
1942 
1943  if (!BlockNumberingValid.count(DominatorBlock))
1944  renumberBlock(DominatorBlock);
1945 
1946  unsigned long DominatorNum = BlockNumbering.lookup(Dominator);
1947  // All numbers start with 1
1948  assert(DominatorNum != 0 && "Block was not numbered properly");
1949  unsigned long DominateeNum = BlockNumbering.lookup(Dominatee);
1950  assert(DominateeNum != 0 && "Block was not numbered properly");
1951  return DominatorNum < DominateeNum;
1952 }
1953 
1954 bool MemorySSA::dominates(const MemoryAccess *Dominator,
1955  const MemoryAccess *Dominatee) const {
1956  if (Dominator == Dominatee)
1957  return true;
1958 
1959  if (isLiveOnEntryDef(Dominatee))
1960  return false;
1961 
1962  if (Dominator->getBlock() != Dominatee->getBlock())
1963  return DT->dominates(Dominator->getBlock(), Dominatee->getBlock());
1964  return locallyDominates(Dominator, Dominatee);
1965 }
1966 
1967 bool MemorySSA::dominates(const MemoryAccess *Dominator,
1968  const Use &Dominatee) const {
1969  if (MemoryPhi *MP = dyn_cast<MemoryPhi>(Dominatee.getUser())) {
1970  BasicBlock *UseBB = MP->getIncomingBlock(Dominatee);
1971  // The def must dominate the incoming block of the phi.
1972  if (UseBB != Dominator->getBlock())
1973  return DT->dominates(Dominator->getBlock(), UseBB);
1974  // If the UseBB and the DefBB are the same, compare locally.
1975  return locallyDominates(Dominator, cast<MemoryAccess>(Dominatee));
1976  }
1977  // If it's not a PHI node use, the normal dominates can already handle it.
1978  return dominates(Dominator, cast<MemoryAccess>(Dominatee.getUser()));
1979 }
1980 
1981 const static char LiveOnEntryStr[] = "liveOnEntry";
1982 
1984  switch (getValueID()) {
1985  case MemoryPhiVal: return static_cast<const MemoryPhi *>(this)->print(OS);
1986  case MemoryDefVal: return static_cast<const MemoryDef *>(this)->print(OS);
1987  case MemoryUseVal: return static_cast<const MemoryUse *>(this)->print(OS);
1988  }
1989  llvm_unreachable("invalid value id");
1990 }
1991 
1993  MemoryAccess *UO = getDefiningAccess();
1994 
1995  auto printID = [&OS](MemoryAccess *A) {
1996  if (A && A->getID())
1997  OS << A->getID();
1998  else
1999  OS << LiveOnEntryStr;
2000  };
2001 
2002  OS << getID() << " = MemoryDef(";
2003  printID(UO);
2004  OS << ")";
2005 
2006  if (isOptimized()) {
2007  OS << "->";
2008  printID(getOptimized());
2009 
2010  if (Optional<AliasResult> AR = getOptimizedAccessType())
2011  OS << " " << *AR;
2012  }
2013 }
2014 
2016  bool First = true;
2017  OS << getID() << " = MemoryPhi(";
2018  for (const auto &Op : operands()) {
2019  BasicBlock *BB = getIncomingBlock(Op);
2020  MemoryAccess *MA = cast<MemoryAccess>(Op);
2021  if (!First)
2022  OS << ',';
2023  else
2024  First = false;
2025 
2026  OS << '{';
2027  if (BB->hasName())
2028  OS << BB->getName();
2029  else
2030  BB->printAsOperand(OS, false);
2031  OS << ',';
2032  if (unsigned ID = MA->getID())
2033  OS << ID;
2034  else
2035  OS << LiveOnEntryStr;
2036  OS << '}';
2037  }
2038  OS << ')';
2039 }
2040 
2042  MemoryAccess *UO = getDefiningAccess();
2043  OS << "MemoryUse(";
2044  if (UO && UO->getID())
2045  OS << UO->getID();
2046  else
2047  OS << LiveOnEntryStr;
2048  OS << ')';
2049 
2050  if (Optional<AliasResult> AR = getOptimizedAccessType())
2051  OS << " " << *AR;
2052 }
2053 
2054 void MemoryAccess::dump() const {
2055 // Cannot completely remove virtual function even in release mode.
2056 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2057  print(dbgs());
2058  dbgs() << "\n";
2059 #endif
2060 }
2061 
2063 
2066 }
2067 
2069  AU.setPreservesAll();
2070  AU.addRequired<MemorySSAWrapperPass>();
2071 }
2072 
2074  auto &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA();
2075  MSSA.print(dbgs());
2076  if (VerifyMemorySSA)
2077  MSSA.verifyMemorySSA();
2078  return false;
2079 }
2080 
2081 AnalysisKey MemorySSAAnalysis::Key;
2082 
2085  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
2086  auto &AA = AM.getResult<AAManager>(F);
2087  return MemorySSAAnalysis::Result(llvm::make_unique<MemorySSA>(F, &AA, &DT));
2088 }
2089 
2092  OS << "MemorySSA for function: " << F.getName() << "\n";
2093  AM.getResult<MemorySSAAnalysis>(F).getMSSA().print(OS);
2094 
2095  return PreservedAnalyses::all();
2096 }
2097 
2100  AM.getResult<MemorySSAAnalysis>(F).getMSSA().verifyMemorySSA();
2101 
2102  return PreservedAnalyses::all();
2103 }
2104 
2105 char MemorySSAWrapperPass::ID = 0;
2106 
2109 }
2110 
2111 void MemorySSAWrapperPass::releaseMemory() { MSSA.reset(); }
2112 
2114  AU.setPreservesAll();
2117 }
2118 
2120  auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2121  auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
2122  MSSA.reset(new MemorySSA(F, &AA, &DT));
2123  return false;
2124 }
2125 
2126 void MemorySSAWrapperPass::verifyAnalysis() const { MSSA->verifyMemorySSA(); }
2127 
2129  MSSA->print(OS);
2130 }
2131 
2133 
2135  DominatorTree *D)
2136  : MemorySSAWalker(M), Walker(*M, *A, *D) {}
2137 
2139  if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
2140  MUD->resetOptimized();
2141 }
2142 
2143 /// Walk the use-def chains starting at \p MA and find
2144 /// the MemoryAccess that actually clobbers Loc.
2145 ///
2146 /// \returns our clobbering memory access
2147 MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess(
2148  MemoryAccess *StartingAccess, UpwardsMemoryQuery &Q) {
2149  return Walker.findClobber(StartingAccess, Q);
2150 }
2151 
2152 MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess(
2153  MemoryAccess *StartingAccess, const MemoryLocation &Loc) {
2154  if (isa<MemoryPhi>(StartingAccess))
2155  return StartingAccess;
2156 
2157  auto *StartingUseOrDef = cast<MemoryUseOrDef>(StartingAccess);
2158  if (MSSA->isLiveOnEntryDef(StartingUseOrDef))
2159  return StartingUseOrDef;
2160 
2161  Instruction *I = StartingUseOrDef->getMemoryInst();
2162 
2163  // Conservatively, fences are always clobbers, so don't perform the walk if we
2164  // hit a fence.
2165  if (!ImmutableCallSite(I) && I->isFenceLike())
2166  return StartingUseOrDef;
2167 
2168  UpwardsMemoryQuery Q;
2169  Q.OriginalAccess = StartingUseOrDef;
2170  Q.StartingLoc = Loc;
2171  Q.Inst = I;
2172  Q.IsCall = false;
2173 
2174  // Unlike the other function, do not walk to the def of a def, because we are
2175  // handed something we already believe is the clobbering access.
2176  MemoryAccess *DefiningAccess = isa<MemoryUse>(StartingUseOrDef)
2177  ? StartingUseOrDef->getDefiningAccess()
2178  : StartingUseOrDef;
2179 
2180  MemoryAccess *Clobber = getClobberingMemoryAccess(DefiningAccess, Q);
2181  LLVM_DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
2182  LLVM_DEBUG(dbgs() << *StartingUseOrDef << "\n");
2183  LLVM_DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ");
2184  LLVM_DEBUG(dbgs() << *Clobber << "\n");
2185  return Clobber;
2186 }
2187 
2188 MemoryAccess *
2189 MemorySSA::CachingWalker::getClobberingMemoryAccess(MemoryAccess *MA) {
2190  auto *StartingAccess = dyn_cast<MemoryUseOrDef>(MA);
2191  // If this is a MemoryPhi, we can't do anything.
2192  if (!StartingAccess)
2193  return MA;
2194 
2195  // If this is an already optimized use or def, return the optimized result.
2196  // Note: Currently, we store the optimized def result in a separate field,
2197  // since we can't use the defining access.
2198  if (StartingAccess->isOptimized())
2199  return StartingAccess->getOptimized();
2200 
2201  const Instruction *I = StartingAccess->getMemoryInst();
2202  // We can't sanely do anything with a fence, since they conservatively clobber
2203  // all memory, and have no locations to get pointers from to try to
2204  // disambiguate.
2205  if (!ImmutableCallSite(I) && I->isFenceLike())
2206  return StartingAccess;
2207 
2208  UpwardsMemoryQuery Q(I, StartingAccess);
2209 
2211  MemoryAccess *LiveOnEntry = MSSA->getLiveOnEntryDef();
2212  StartingAccess->setOptimized(LiveOnEntry);
2213  StartingAccess->setOptimizedAccessType(None);
2214  return LiveOnEntry;
2215  }
2216 
2217  // Start with the thing we already think clobbers this location
2218  MemoryAccess *DefiningAccess = StartingAccess->getDefiningAccess();
2219 
2220  // At this point, DefiningAccess may be the live on entry def.
2221  // If it is, we will not get a better result.
2222  if (MSSA->isLiveOnEntryDef(DefiningAccess)) {
2223  StartingAccess->setOptimized(DefiningAccess);
2224  StartingAccess->setOptimizedAccessType(None);
2225  return DefiningAccess;
2226  }
2227 
2228  MemoryAccess *Result = getClobberingMemoryAccess(DefiningAccess, Q);
2229  LLVM_DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
2230  LLVM_DEBUG(dbgs() << *DefiningAccess << "\n");
2231  LLVM_DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ");
2232  LLVM_DEBUG(dbgs() << *Result << "\n");
2233 
2234  StartingAccess->setOptimized(Result);
2235  if (MSSA->isLiveOnEntryDef(Result))
2236  StartingAccess->setOptimizedAccessType(None);
2237  else if (Q.AR == MustAlias)
2238  StartingAccess->setOptimizedAccessType(MustAlias);
2239 
2240  return Result;
2241 }
2242 
2243 MemoryAccess *
2245  if (auto *Use = dyn_cast<MemoryUseOrDef>(MA))
2246  return Use->getDefiningAccess();
2247  return MA;
2248 }
2249 
2251  MemoryAccess *StartingAccess, const MemoryLocation &) {
2252  if (auto *Use = dyn_cast<MemoryUseOrDef>(StartingAccess))
2253  return Use->getDefiningAccess();
2254  return StartingAccess;
2255 }
2256 
2257 void MemoryPhi::deleteMe(DerivedUser *Self) {
2258  delete static_cast<MemoryPhi *>(Self);
2259 }
2260 
2261 void MemoryDef::deleteMe(DerivedUser *Self) {
2262  delete static_cast<MemoryDef *>(Self);
2263 }
2264 
2265 void MemoryUse::deleteMe(DerivedUser *Self) {
2266  delete static_cast<MemoryUse *>(Self);
2267 }
MemorySSAWalker * getWalker()
Definition: MemorySSA.cpp:1423
bool runOnFunction(Function &) override
runOnFunction - Virtual method overriden by subclasses to do the per-function processing of the pass...
Definition: MemorySSA.cpp:2119
The access may reference and may modify the value stored in memory.
void initializeMemorySSAWrapperPassPass(PassRegistry &)
AccessList * getWritableBlockAccesses(const BasicBlock *BB) const
Definition: MemorySSA.h:797
typename std::vector< DomTreeNodeBase *>::const_iterator const_iterator
iterator_range< use_iterator > uses()
Definition: Value.h:355
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
Value * getPointerOperand(Value *V)
A helper function that returns the pointer operand of a load, store or GEP instruction.
virtual void verify(const MemorySSA *MSSA)
Definition: MemorySSA.h:1041
void dropAllReferences()
Drop all references to operands.
Definition: User.h:295
Atomic ordering constants.
bool VerifyMemorySSA
Enables verification of MemorySSA.
Definition: MemorySSA.cpp:84
bool isFenceLike() const
Return true if this instruction behaves like a memory fence: it can load or store to memory location ...
Definition: Instruction.h:541
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:770
void print(raw_ostream &OS) const
Definition: MemorySSA.cpp:2015
This class represents lattice values for constants.
Definition: AllocatorList.h:24
bool dominates(const MemoryAccess *A, const MemoryAccess *B) const
Given two memory accesses in potentially different blocks, determine whether MemoryAccess A dominates...
Definition: MemorySSA.cpp:1954
A Module instance is used to store all the information related to an LLVM module. ...
Definition: Module.h:64
void push_back(reference Node)
Insert a node at the back; never copies.
Definition: simple_ilist.h:148
formatted_raw_ostream - A raw_ostream that wraps another one and keeps track of line and column posit...
This class provides various memory handling functions that manipulate MemoryBlock instances...
Definition: Memory.h:46
Implements a dense probed hash-table based set.
Definition: DenseSet.h:250
const AccessList * getBlockAccesses(const BasicBlock *BB) const
Return the list of MemoryAccess&#39;s for a given basic block.
Definition: MemorySSA.h:749
This provides a very simple, boring adaptor for a begin and end iterator into a range type...
void getAnalysisUsage(AnalysisUsage &AU) const override
getAnalysisUsage - This function should be overriden by passes that need analysis information to do t...
Definition: MemorySSA.cpp:2068
The two locations do not alias at all.
Definition: AliasAnalysis.h:85
Extension point for the Value hierarchy.
Definition: DerivedUser.h:28
Represents a read-write access to memory, whether it is a must-alias, or a may-alias.
Definition: MemorySSA.h:373
AtomicOrdering getOrdering() const
Returns the ordering constraint of this load instruction.
Definition: Instructions.h:237
static const char LiveOnEntryStr[]
Definition: MemorySSA.cpp:1981
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:705
unsigned second
void getAnalysisUsage(AnalysisUsage &AU) const override
getAnalysisUsage - This function should be overriden by passes that need analysis information to do t...
Definition: MemorySSA.cpp:2113
bool all_of(R &&range, UnaryPredicate P)
Provide wrappers to std::all_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1180
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:231
F(f)
OptimizeUses(MemorySSA *MSSA, MemorySSAWalker *Walker, AliasAnalysis *AA, DominatorTree *DT)
Definition: MemorySSA.cpp:1136
block Block Frequency true
An instruction for reading from memory.
Definition: Instructions.h:168
memoryssa
Definition: MemorySSA.cpp:66
LLVM_READONLY APFloat maximum(const APFloat &A, const APFloat &B)
Implements IEEE 754-2018 maximum semantics.
Definition: APFloat.h:1262
This defines the Use class.
void print(raw_ostream &OS) const
Definition: MemorySSA.cpp:1992
static Optional< MemoryLocation > getOrNone(const Instruction *Inst)
MemorySSA(Function &, AliasAnalysis *, DominatorTree *)
Definition: MemorySSA.cpp:1096
LLVMContext & getContext() const
Get the context in which this basic block lives.
Definition: BasicBlock.cpp:33
INITIALIZE_PASS_BEGIN(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false, true) INITIALIZE_PASS_END(MemorySSAWrapperPass
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Definition: DenseMap.h:221
Represents read-only accesses to memory.
Definition: MemorySSA.h:317
AliasResult alias(const MemoryLocation &LocA, const MemoryLocation &LocB)
The main low level interface to the alias analysis implementation.
This class is a batch walker of all MemoryUse&#39;s in the program, and points their defining access at t...
Definition: MemorySSA.cpp:1134
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:51
Legacy analysis pass which computes MemorySSA.
Definition: MemorySSA.h:950
bool isVolatile() const
Return true if this is a load from a volatile memory location.
Definition: Instructions.h:221
void renamePass(BasicBlock *BB, MemoryAccess *IncomingVal, SmallPtrSetImpl< BasicBlock *> &Visited)
Definition: MemorySSA.h:816
A MemorySSAWalker that does AA walks to disambiguate accesses.
Definition: MemorySSA.cpp:941
A Use represents the edge between a Value definition and its users.
Definition: Use.h:56
void setDefiningBlocks(const SmallPtrSetImpl< BasicBlock *> &Blocks)
Give the IDF calculator the set of blocks in which the value is defined.
static bool defClobbersUseOrDef(MemoryDef *MD, const MemoryUseOrDef *MU, AliasAnalysis &AA)
Definition: MemorySSA.cpp:317
MemorySSAAnnotatedWriter(const MemorySSA *M)
Definition: MemorySSA.cpp:100
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:42
Encapsulates MemorySSA, including all data associated with memory accesses.
Definition: MemorySSA.h:700
const DefsList * getBlockDefs(const BasicBlock *BB) const
Return the list of MemoryDef&#39;s and MemoryPhi&#39;s for a given basic block.
Definition: MemorySSA.h:757
void insertIntoListsForBlock(MemoryAccess *, const BasicBlock *, InsertionPlace)
Definition: MemorySSA.cpp:1436
#define LLVM_DUMP_METHOD
Definition: Compiler.h:74
static bool isOrdered(const Instruction *I)
Definition: MemorySSA.cpp:1558
ELFYAML::ELF_STO Other
Definition: ELFYAML.cpp:783
APInt operator*(APInt a, uint64_t RHS)
Definition: APInt.h:2091
void verifyDefUses(Function &F) const
Verify the immediate use information, by walking all the memory accesses and verifying that...
Definition: MemorySSA.cpp:1879
void verifyAnalysis() const override
verifyAnalysis() - This member can be implemented by a analysis pass to check state of analysis infor...
Definition: MemorySSA.cpp:2126
A simple intrusive list implementation.
Definition: simple_ilist.h:79
LLVM_NODISCARD bool isMustSet(const ModRefInfo MRI)
User * getUser() const LLVM_READONLY
Returns the User that contains this Use.
Definition: Use.cpp:41
#define F(x, y, z)
Definition: MD5.cpp:55
static int getID(struct InternalInstruction *insn, const void *miiArg)
MemoryUseOrDef * getMemoryAccess(const Instruction *I) const
Given a memory Mod/Ref&#39;ing instruction, get the MemorySSA access associated with it.
Definition: MemorySSA.h:711
upward_defs_iterator upward_defs_end()
Definition: MemorySSA.h:1235
early cse memssa
Definition: EarlyCSE.cpp:1333
Memory SSA
Definition: MemorySSA.cpp:66
MDNode * getMetadata(unsigned KindID) const
Get the metadata of given kind attached to this Instruction.
Definition: Instruction.h:221
void emitBasicBlockStartAnnot(const BasicBlock *BB, formatted_raw_ostream &OS) override
emitBasicBlockStartAnnot - This may be implemented to emit a string right after the basic block label...
Definition: MemorySSA.cpp:102
void dump() const
Definition: MemorySSA.cpp:1698
static cl::opt< bool, true > VerifyMemorySSAX("verify-memoryssa", cl::location(VerifyMemorySSA), cl::Hidden, cl::desc("Enable verification of MemorySSA."))
This is the generic walker interface for walkers of MemorySSA.
Definition: MemorySSA.h:981
static void checkClobberSanity(const MemoryAccess *Start, MemoryAccess *ClobberAt, const MemoryLocation &StartLoc, const MemorySSA &MSSA, const UpwardsMemoryQuery &Query, AliasAnalysis &AA, bool AllowImpreciseClobber=false)
Verifies that Start is clobbered by ClobberAt, and that nothing inbetween Start and ClobberAt can clo...
Definition: MemorySSA.cpp:384
CRTP base class which implements the entire standard iterator facade in terms of a minimal subset of ...
Definition: iterator.h:68
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:145
static bool lifetimeEndsAt(MemoryDef *MD, const MemoryLocation &Loc, AliasAnalysis &AA)
Definition: MemorySSA.cpp:347
bool isAtLeastOrStrongerThan(AtomicOrdering ao, AtomicOrdering other)
bool pointsToConstantMemory(const MemoryLocation &Loc, bool OrLocal=false)
Checks whether the given location points to constant memory, or if OrLocal is true whether it points ...
An assembly annotator class to print Memory SSA information in comments.
Definition: MemorySSA.cpp:94
void removeFromLookups(MemoryAccess *)
Properly remove MA from all of MemorySSA&#39;s lookup tables.
Definition: MemorySSA.cpp:1638
NodeT * getBlock() const
#define P(N)
static MemoryLocation get(const LoadInst *LI)
Return a location with information about the memory reference by the given instruction.
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:419
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:154
bool hasName() const
Definition: Value.h:251
LLVM Basic Block Representation.
Definition: BasicBlock.h:58
DomTreeNodeBase * getIDom() const
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
LLVM_NODISCARD bool empty() const
Definition: SmallPtrSet.h:92
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator begin()
Definition: SmallVector.h:129
A manager for alias analyses.
static bool isEqual(const MemoryLocOrCall &LHS, const MemoryLocOrCall &RHS)
Definition: MemorySSA.cpp:207
std::pair< iterator, bool > insert(PtrType Ptr)
Inserts Ptr if and only if there is no element in the container equal to Ptr.
Definition: SmallPtrSet.h:371
early cse Early CSE w MemorySSA
Definition: EarlyCSE.cpp:1333
void dump() const
Definition: MemorySSA.cpp:2054
Interval::pred_iterator pred_begin(Interval *I)
pred_begin/pred_end - define methods so that Intervals may be used just like BasicBlocks can with the...
Definition: Interval.h:113
InsertionPlace
Used in various insertion functions to specify whether we are talking about the beginning or end of a...
Definition: MemorySSA.h:782
Represent the analysis usage information of a pass.
void print(raw_ostream &OS) const
Definition: MemorySSA.cpp:1983
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:285
Interval::pred_iterator pred_end(Interval *I)
Definition: Interval.h:116
void verifyDomination(Function &F) const
Verify the domination properties of MemorySSA by checking that each definition dominates all of its u...
Definition: MemorySSA.cpp:1842
static void print(raw_ostream &Out, object::Archive::Kind Kind, T Val)
size_type count(ConstPtrType Ptr) const
count - Return 1 if the specified pointer is in the set, 0 otherwise.
Definition: SmallPtrSet.h:382
Memory true print Memory SSA Printer
Definition: MemorySSA.cpp:72
void invalidateInfo(MemoryAccess *) override
Given a memory access, invalidate anything this walker knows about that access.
Definition: MemorySSA.cpp:2138
auto find_if(R &&Range, UnaryPredicate P) -> decltype(adl_begin(Range))
Provide wrappers to std::find_if which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1208
void checkClobberSanityAccess(const MemoryAccess *MA) const
Check clobber sanity for an access.
Definition: MemorySSA.cpp:1711
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
Definition: MemorySSA.cpp:2098
auto find_if_not(R &&Range, UnaryPredicate P) -> decltype(adl_begin(Range))
Definition: STLExtras.h:1213
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:160
size_t size() const
Definition: SmallVector.h:53
Compute iterated dominance frontiers using a linear time algorithm.
DomTreeNodeBase< NodeT > * getNode(const NodeT *BB) const
getNode - return the (Post)DominatorTree node for the specified basic block.
auto find(R &&Range, const T &Val) -> decltype(adl_begin(Range))
Provide wrappers to std::find which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1201
An intrusive list with ownership and callbacks specified/controlled by ilist_traits, only with API safe for polymorphic types.
Definition: ilist.h:390
void printAsOperand(raw_ostream &O, bool PrintType=true, const Module *M=nullptr) const
Print the name of this Value out to the specified raw_ostream.
Definition: AsmWriter.cpp:4222
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
bool verify(const TargetRegisterInfo &TRI) const
Check that information hold by this instance make sense for the given TRI.
Memory true print Memory SSA static false cl::opt< unsigned > MaxCheckLimit("memssa-check-limit", cl::Hidden, cl::init(100), cl::desc("The maximum number of stores/phis MemorySSA" "will consider trying to walk past (default = 100)"))
MemoryUseOrDef * createDefinedAccess(Instruction *, MemoryAccess *, const MemoryUseOrDef *Template=nullptr)
Definition: MemorySSA.cpp:1543
void optimizeUses()
Optimize uses to point to their actual clobbering definitions.
Definition: MemorySSA.cpp:1340
The two locations may or may not alias. This is the least precise result.
Definition: AliasAnalysis.h:87
Representation for a specific memory location.
The two locations precisely alias each other.
Definition: AliasAnalysis.h:91
iterator_range< T > make_range(T x, T y)
Convenience function for iterating over sub-ranges.
void verifyOrdering(Function &F) const
Verify that the order and existence of MemoryAccesses matches the order and existence of memory affec...
Definition: MemorySSA.cpp:1777
void setDefiningAccess(MemoryAccess *DMA, bool Optimized=false, Optional< AliasResult > AR=MayAlias)
Definition: MemorySSA.h:297
unsigned getNumOperands() const
Definition: User.h:192
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements...
Definition: SmallPtrSet.h:418
BlockVerifier::State From
void verifyMemorySSA() const
Verify that MemorySSA is self consistent (IE definitions dominate all uses, uses appear in the right ...
Definition: MemorySSA.cpp:1701
bool erase(PtrType Ptr)
erase - If the set contains the specified pointer, remove it and return true, otherwise return false...
Definition: SmallPtrSet.h:378
void calculate(SmallVectorImpl< BasicBlock *> &IDFBlocks)
Calculate iterated dominance frontiers.
bool locallyDominates(const MemoryAccess *A, const MemoryAccess *B) const
Given two memory accesses in the same basic block, determine whether MemoryAccess A dominates MemoryA...
Definition: MemorySSA.cpp:1923
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:847
bool dominates(const Instruction *Def, const Use &U) const
Return true if Def dominates a use in User.
Definition: Dominators.cpp:249
bool runOnFunction(Function &) override
runOnFunction - Virtual method overriden by subclasses to do the per-function processing of the pass...
Definition: MemorySSA.cpp:2073
An analysis that produces MemorySSA for a function.
Definition: MemorySSA.h:914
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:381
static ClobberAlias instructionClobbersQuery(const MemoryDef *MD, const MemoryLocation &UseLoc, const Instruction *UseInst, AliasAnalysis &AA)
Definition: MemorySSA.cpp:255
BasicBlock * getBlock() const
Definition: MemorySSA.h:157
static GCRegistry::Add< StatepointGC > D("statepoint-example", "an example strategy for statepoint")
pred_range predecessors(BasicBlock *BB)
Definition: CFG.h:125
void verify(const MemorySSA *MSSA) override
Definition: MemorySSA.cpp:957
#define MAP(n)
void emitInstructionAnnot(const Instruction *I, formatted_raw_ostream &OS) override
emitInstructionAnnot - This may be implemented to emit a string right before an instruction is emitte...
Definition: MemorySSA.cpp:108
MemoryAccess * getLiveOnEntryDef() const
Definition: MemorySSA.h:733
void verifyClobberSanity(const Function &F) const
Definition: MemorySSA.cpp:1725
void print(raw_ostream &OS) const
Definition: MemorySSA.cpp:2041
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:133
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:941
A range adaptor for a pair of iterators.
void removeFromLists(MemoryAccess *, bool ShouldDelete=true)
Properly remove MA from all of MemorySSA&#39;s lists.
Definition: MemorySSA.cpp:1665
Target - Wrapper for Target specific information.
void push_back(pointer val)
Definition: ilist.h:313
hash_code hash_combine(const Ts &...args)
Combine values into a single hash_code.
Definition: Hashing.h:601
Class that has the common methods + fields of memory uses/defs.
Definition: MemorySSA.h:245
ModRefInfo getModRefInfo(ImmutableCallSite CS, const MemoryLocation &Loc)
getModRefInfo (for call sites) - Return information about whether a particular call site modifies or ...
void setPreservesAll()
Set by analyses that do not transform their input at all.
iterator_range< user_iterator > users()
Definition: Value.h:400
bool isMustAlias(const MemoryLocation &LocA, const MemoryLocation &LocB)
A trivial helper function to check to see if the specified pointers are must-alias.
An opaque object representing a hash code.
Definition: Hashing.h:72
Instruction * getMemoryInst() const
Get the instruction that this MemoryUse represents.
Definition: MemorySSA.h:252
void append(in_iter in_start, in_iter in_end)
Add the specified range to the end of the SmallVector.
Definition: SmallVector.h:394
amdgpu Simplify well known AMD library false Value Value * Arg
void initializeMemorySSAPrinterLegacyPassPass(PassRegistry &)
LLVM_NODISCARD bool isModSet(const ModRefInfo MRI)
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator end()
Definition: SmallVector.h:133
MemorySSAWalker(MemorySSA *)
Definition: MemorySSA.cpp:2132
iterator_range< def_chain_iterator< T > > def_chain(T MA, MemoryAccess *UpTo=nullptr)
Definition: MemorySSA.h:1286
void emplace_back(ArgTypes &&... Args)
Definition: SmallVector.h:652
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:56
void releaseMemory() override
releaseMemory() - This member can be implemented by a pass if it wants to be able to release its memo...
Definition: MemorySSA.cpp:2111
This file provides utility analysis objects describing memory locations.
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:215
Establish a view to a call site for examination.
Definition: CallSite.h:711
CachingWalker(MemorySSA *, AliasAnalysis *, DominatorTree *)
Definition: MemorySSA.cpp:2134
#define I(x, y, z)
Definition: MD5.cpp:58
#define N
void moveTo(MemoryUseOrDef *What, BasicBlock *BB, AccessList::iterator Where)
Definition: MemorySSA.cpp:1512
MemoryAccess * getClobberingMemoryAccess(MemoryAccess *) override
Does the same thing as getClobberingMemoryAccess(const Instruction *I), but takes a MemoryAccess inst...
Definition: MemorySSA.cpp:2244
static void Query(const MachineInstr &MI, AliasAnalysis &AA, bool &Read, bool &Write, bool &Effects, bool &StackPointer)
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:323
bool isLiveOnEntryDef(const MemoryAccess *MA) const
Return true if MA represents the live on entry value.
Definition: MemorySSA.h:729
AnalysisUsage & addRequiredTransitive()
iterator_range< df_iterator< T > > depth_first(const T &G)
MemoryAccess * getClobberingMemoryAccess(const Instruction *I)
Given a memory Mod/Ref/ModRef&#39;ing instruction, calling this will give you the nearest dominating Memo...
Definition: MemorySSA.h:1010
Determine the iterated dominance frontier, given a set of defining blocks, and optionally, a set of live-in blocks.
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
Definition: MemorySSA.cpp:2090
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
Result run(Function &F, FunctionAnalysisManager &AM)
Definition: MemorySSA.cpp:2083
static bool areLoadsReorderable(const LoadInst *Use, const LoadInst *MayClobber)
This does one-way checks to see if Use could theoretically be hoisted above MayClobber.
Definition: MemorySSA.cpp:220
LLVM Value Representation.
Definition: Value.h:73
static MemoryLocOrCall getTombstoneKey()
Definition: MemorySSA.cpp:188
succ_range successors(Instruction *I)
Definition: CFG.h:264
upward_defs_iterator upward_defs_begin(const MemoryAccessPair &Pair)
Definition: MemorySSA.h:1231
unsigned getID() const
Used for debugging and tracking things about MemoryAccesses.
Definition: MemorySSA.h:661
static bool isUseTriviallyOptimizableToLiveOnEntry(AliasAnalysis &AA, const Instruction *I)
Definition: MemorySSA.cpp:361
This class implements an extremely fast bulk output stream that can only output to a stream...
Definition: raw_ostream.h:46
ModRefInfo
Flags indicating whether a memory access modifies or references memory.
static unsigned getHashValue(const MemoryLocOrCall &MLOC)
Definition: MemorySSA.cpp:192
This file exposes an interface to building/using memory SSA to walk memory instructions using a use/d...
hexagon cext opt
A container for analyses that lazily runs them and caches their results.
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:260
void setBlock(BasicBlock *BB)
Used by MemorySSA to change the block of a MemoryAccess when it is moved.
Definition: MemorySSA.h:210
LLVM_NODISCARD bool isModOrRefSet(const ModRefInfo MRI)
bool operator==(uint64_t V1, const APInt &V2)
Definition: APInt.h:1967
void verifyDominationNumbers(const Function &F) const
Verify that all of the blocks we believe to have valid domination numbers actually have valid dominat...
Definition: MemorySSA.cpp:1739
A wrapper pass to provide the legacy pass manager access to a suitably prepared AAResults object...
Represents phi nodes for memory accesses.
Definition: MemorySSA.h:478
void print(raw_ostream &) const
Definition: MemorySSA.cpp:1692
This header defines various interfaces for pass management in LLVM.
#define LLVM_DEBUG(X)
Definition: Debug.h:123
static MemoryLocOrCall getEmptyKey()
Definition: MemorySSA.cpp:184
void insertIntoListsBefore(MemoryAccess *, const BasicBlock *, AccessList::iterator)
Definition: MemorySSA.cpp:1468
A special type used by analysis passes to provide an address that identifies that particular analysis...
Definition: PassManager.h:71
bool use_empty() const
Definition: Value.h:323
LocationClass< Ty > location(Ty &L)
Definition: CommandLine.h:435
hexagon widen stores
reverse_iterator rbegin()
Definition: simple_ilist.h:122
std::pair< MemoryAccess *, MemoryLocation > MemoryAccessPair
Definition: MemorySSA.h:1062
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:44
const BasicBlock * getParent() const
Definition: Instruction.h:67
void print(raw_ostream &OS, const Module *M=nullptr) const override
print - Print out the internal state of the pass.
Definition: MemorySSA.cpp:2128
bool is_contained(R &&Range, const E &Element)
Wrapper function around std::find to detect if an element exists in a container.
Definition: STLExtras.h:1239
LLVM_NODISCARD bool isRefSet(const ModRefInfo MRI)