1 //===- MemorySSA.cpp - Memory SSA Builder ---------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the MemorySSA class.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Analysis/MemorySSA.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/DenseMapInfo.h"
17 #include "llvm/ADT/DenseSet.h"
18 #include "llvm/ADT/DepthFirstIterator.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"
26 #include "llvm/ADT/iterator_range.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/IteratedDominanceFrontier.h"
29 #include "llvm/Analysis/MemoryLocation.h"
30 #include "llvm/IR/AssemblyAnnotationWriter.h"
31 #include "llvm/IR/BasicBlock.h"
32 #include "llvm/IR/CallSite.h"
33 #include "llvm/IR/Dominators.h"
34 #include "llvm/IR/Function.h"
35 #include "llvm/IR/Instruction.h"
36 #include "llvm/IR/Instructions.h"
37 #include "llvm/IR/IntrinsicInst.h"
38 #include "llvm/IR/Intrinsics.h"
39 #include "llvm/IR/LLVMContext.h"
40 #include "llvm/IR/PassManager.h"
41 #include "llvm/IR/Use.h"
42 #include "llvm/Pass.h"
43 #include "llvm/Support/AtomicOrdering.h"
44 #include "llvm/Support/Casting.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Compiler.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/FormattedStream.h"
50 #include "llvm/Support/raw_ostream.h"
59 #define DEBUG_TYPE "memoryssa"
61 INITIALIZE_PASS_BEGIN(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,
63 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
64 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
65 INITIALIZE_PASS_END(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,
68 INITIALIZE_PASS_BEGIN(MemorySSAPrinterLegacyPass, "print-memoryssa",
69 "Memory SSA Printer", false, false)
70 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
71 INITIALIZE_PASS_END(MemorySSAPrinterLegacyPass, "print-memoryssa",
72 "Memory SSA Printer", false, false)
74 static cl::opt<unsigned> MaxCheckLimit(
75 "memssa-check-limit", cl::Hidden, cl::init(100),
76 cl::desc("The maximum number of stores/phis MemorySSA"
77 "will consider trying to walk past (default = 100)"));
80 VerifyMemorySSA("verify-memoryssa", cl::init(false), cl::Hidden,
81 cl::desc("Verify MemorySSA in legacy printer pass."));
85 /// \brief An assembly annotator class to print Memory SSA information in
87 class MemorySSAAnnotatedWriter : public AssemblyAnnotationWriter {
88 friend class MemorySSA;
90 const MemorySSA *MSSA;
93 MemorySSAAnnotatedWriter(const MemorySSA *M) : MSSA(M) {}
95 void emitBasicBlockStartAnnot(const BasicBlock *BB,
96 formatted_raw_ostream &OS) override {
97 if (MemoryAccess *MA = MSSA->getMemoryAccess(BB))
98 OS << "; " << *MA << "\n";
101 void emitInstructionAnnot(const Instruction *I,
102 formatted_raw_ostream &OS) override {
103 if (MemoryAccess *MA = MSSA->getMemoryAccess(I))
104 OS << "; " << *MA << "\n";
108 } // end namespace llvm
112 /// Our current alias analysis API differentiates heavily between calls and
113 /// non-calls, and functions called on one usually assert on the other.
114 /// This class encapsulates the distinction to simplify other code that wants
115 /// "Memory affecting instructions and related data" to use as a key.
116 /// For example, this class is used as a densemap key in the use optimizer.
117 class MemoryLocOrCall {
121 MemoryLocOrCall() = default;
122 MemoryLocOrCall(MemoryUseOrDef *MUD)
123 : MemoryLocOrCall(MUD->getMemoryInst()) {}
124 MemoryLocOrCall(const MemoryUseOrDef *MUD)
125 : MemoryLocOrCall(MUD->getMemoryInst()) {}
127 MemoryLocOrCall(Instruction *Inst) {
128 if (ImmutableCallSite(Inst)) {
130 CS = ImmutableCallSite(Inst);
133 // There is no such thing as a memorylocation for a fence inst, and it is
134 // unique in that regard.
135 if (!isa<FenceInst>(Inst))
136 Loc = MemoryLocation::get(Inst);
140 explicit MemoryLocOrCall(const MemoryLocation &Loc) : Loc(Loc) {}
142 ImmutableCallSite getCS() const {
147 MemoryLocation getLoc() const {
152 bool operator==(const MemoryLocOrCall &Other) const {
153 if (IsCall != Other.IsCall)
157 return CS.getCalledValue() == Other.CS.getCalledValue();
158 return Loc == Other.Loc;
163 ImmutableCallSite CS;
168 } // end anonymous namespace
172 template <> struct DenseMapInfo<MemoryLocOrCall> {
173 static inline MemoryLocOrCall getEmptyKey() {
174 return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getEmptyKey());
177 static inline MemoryLocOrCall getTombstoneKey() {
178 return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getTombstoneKey());
181 static unsigned getHashValue(const MemoryLocOrCall &MLOC) {
183 return hash_combine(MLOC.IsCall,
184 DenseMapInfo<const Value *>::getHashValue(
185 MLOC.getCS().getCalledValue()));
187 MLOC.IsCall, DenseMapInfo<MemoryLocation>::getHashValue(MLOC.getLoc()));
190 static bool isEqual(const MemoryLocOrCall &LHS, const MemoryLocOrCall &RHS) {
195 enum class Reorderability { Always, IfNoAlias, Never };
197 } // end namespace llvm
199 /// This does one-way checks to see if Use could theoretically be hoisted above
200 /// MayClobber. This will not check the other way around.
202 /// This assumes that, for the purposes of MemorySSA, Use comes directly after
203 /// MayClobber, with no potentially clobbering operations in between them.
204 /// (Where potentially clobbering ops are memory barriers, aliased stores, etc.)
205 static Reorderability getLoadReorderability(const LoadInst *Use,
206 const LoadInst *MayClobber) {
207 bool VolatileUse = Use->isVolatile();
208 bool VolatileClobber = MayClobber->isVolatile();
209 // Volatile operations may never be reordered with other volatile operations.
210 if (VolatileUse && VolatileClobber)
211 return Reorderability::Never;
213 // The lang ref allows reordering of volatile and non-volatile operations.
214 // Whether an aliasing nonvolatile load and volatile load can be reordered,
215 // though, is ambiguous. Because it may not be best to exploit this ambiguity,
216 // we only allow volatile/non-volatile reordering if the volatile and
217 // non-volatile operations don't alias.
218 Reorderability Result = VolatileUse || VolatileClobber
219 ? Reorderability::IfNoAlias
220 : Reorderability::Always;
222 // If a load is seq_cst, it cannot be moved above other loads. If its ordering
223 // is weaker, it can be moved above other loads. We just need to be sure that
224 // MayClobber isn't an acquire load, because loads can't be moved above
227 // Note that this explicitly *does* allow the free reordering of monotonic (or
228 // weaker) loads of the same address.
229 bool SeqCstUse = Use->getOrdering() == AtomicOrdering::SequentiallyConsistent;
230 bool MayClobberIsAcquire = isAtLeastOrStrongerThan(MayClobber->getOrdering(),
231 AtomicOrdering::Acquire);
232 if (SeqCstUse || MayClobberIsAcquire)
233 return Reorderability::Never;
237 static bool instructionClobbersQuery(MemoryDef *MD,
238 const MemoryLocation &UseLoc,
239 const Instruction *UseInst,
241 Instruction *DefInst = MD->getMemoryInst();
242 assert(DefInst && "Defining instruction not actually an instruction");
243 ImmutableCallSite UseCS(UseInst);
245 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(DefInst)) {
246 // These intrinsics will show up as affecting memory, but they are just
248 switch (II->getIntrinsicID()) {
249 case Intrinsic::lifetime_start:
252 return AA.isMustAlias(MemoryLocation(II->getArgOperand(1)), UseLoc);
253 case Intrinsic::lifetime_end:
254 case Intrinsic::invariant_start:
255 case Intrinsic::invariant_end:
256 case Intrinsic::assume:
264 ModRefInfo I = AA.getModRefInfo(DefInst, UseCS);
265 return isModOrRefSet(I);
268 if (auto *DefLoad = dyn_cast<LoadInst>(DefInst)) {
269 if (auto *UseLoad = dyn_cast<LoadInst>(UseInst)) {
270 switch (getLoadReorderability(UseLoad, DefLoad)) {
271 case Reorderability::Always:
273 case Reorderability::Never:
275 case Reorderability::IfNoAlias:
276 return !AA.isNoAlias(UseLoc, MemoryLocation::get(DefLoad));
281 return isModSet(AA.getModRefInfo(DefInst, UseLoc));
284 static bool instructionClobbersQuery(MemoryDef *MD, const MemoryUseOrDef *MU,
285 const MemoryLocOrCall &UseMLOC,
287 // FIXME: This is a temporary hack to allow a single instructionClobbersQuery
288 // to exist while MemoryLocOrCall is pushed through places.
290 return instructionClobbersQuery(MD, MemoryLocation(), MU->getMemoryInst(),
292 return instructionClobbersQuery(MD, UseMLOC.getLoc(), MU->getMemoryInst(),
296 // Return true when MD may alias MU, return false otherwise.
297 bool MemorySSAUtil::defClobbersUseOrDef(MemoryDef *MD, const MemoryUseOrDef *MU,
299 return instructionClobbersQuery(MD, MU, MemoryLocOrCall(MU), AA);
304 struct UpwardsMemoryQuery {
305 // True if our original query started off as a call
307 // The pointer location we started the query with. This will be empty if
309 MemoryLocation StartingLoc;
310 // This is the instruction we were querying about.
311 const Instruction *Inst = nullptr;
312 // The MemoryAccess we actually got called with, used to test local domination
313 const MemoryAccess *OriginalAccess = nullptr;
315 UpwardsMemoryQuery() = default;
317 UpwardsMemoryQuery(const Instruction *Inst, const MemoryAccess *Access)
318 : IsCall(ImmutableCallSite(Inst)), Inst(Inst), OriginalAccess(Access) {
320 StartingLoc = MemoryLocation::get(Inst);
324 } // end anonymous namespace
326 static bool lifetimeEndsAt(MemoryDef *MD, const MemoryLocation &Loc,
328 Instruction *Inst = MD->getMemoryInst();
329 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
330 switch (II->getIntrinsicID()) {
331 case Intrinsic::lifetime_end:
332 return AA.isMustAlias(MemoryLocation(II->getArgOperand(1)), Loc);
340 static bool isUseTriviallyOptimizableToLiveOnEntry(AliasAnalysis &AA,
341 const Instruction *I) {
342 // If the memory can't be changed, then loads of the memory can't be
345 // FIXME: We should handle invariant groups, as well. It's a bit harder,
346 // because we need to pay close attention to invariant group barriers.
347 return isa<LoadInst>(I) && (I->getMetadata(LLVMContext::MD_invariant_load) ||
348 AA.pointsToConstantMemory(cast<LoadInst>(I)->
349 getPointerOperand()));
352 /// Verifies that `Start` is clobbered by `ClobberAt`, and that nothing
353 /// inbetween `Start` and `ClobberAt` can clobbers `Start`.
355 /// This is meant to be as simple and self-contained as possible. Because it
356 /// uses no cache, etc., it can be relatively expensive.
358 /// \param Start The MemoryAccess that we want to walk from.
359 /// \param ClobberAt A clobber for Start.
360 /// \param StartLoc The MemoryLocation for Start.
361 /// \param MSSA The MemorySSA isntance that Start and ClobberAt belong to.
362 /// \param Query The UpwardsMemoryQuery we used for our search.
363 /// \param AA The AliasAnalysis we used for our search.
364 static void LLVM_ATTRIBUTE_UNUSED
365 checkClobberSanity(MemoryAccess *Start, MemoryAccess *ClobberAt,
366 const MemoryLocation &StartLoc, const MemorySSA &MSSA,
367 const UpwardsMemoryQuery &Query, AliasAnalysis &AA) {
368 assert(MSSA.dominates(ClobberAt, Start) && "Clobber doesn't dominate start?");
370 if (MSSA.isLiveOnEntryDef(Start)) {
371 assert(MSSA.isLiveOnEntryDef(ClobberAt) &&
372 "liveOnEntry must clobber itself");
376 bool FoundClobber = false;
377 DenseSet<MemoryAccessPair> VisitedPhis;
378 SmallVector<MemoryAccessPair, 8> Worklist;
379 Worklist.emplace_back(Start, StartLoc);
380 // Walk all paths from Start to ClobberAt, while looking for clobbers. If one
381 // is found, complain.
382 while (!Worklist.empty()) {
383 MemoryAccessPair MAP = Worklist.pop_back_val();
384 // All we care about is that nothing from Start to ClobberAt clobbers Start.
385 // We learn nothing from revisiting nodes.
386 if (!VisitedPhis.insert(MAP).second)
389 for (MemoryAccess *MA : def_chain(MAP.first)) {
390 if (MA == ClobberAt) {
391 if (auto *MD = dyn_cast<MemoryDef>(MA)) {
392 // instructionClobbersQuery isn't essentially free, so don't use `|=`,
393 // since it won't let us short-circuit.
395 // Also, note that this can't be hoisted out of the `Worklist` loop,
396 // since MD may only act as a clobber for 1 of N MemoryLocations.
398 FoundClobber || MSSA.isLiveOnEntryDef(MD) ||
399 instructionClobbersQuery(MD, MAP.second, Query.Inst, AA);
404 // We should never hit liveOnEntry, unless it's the clobber.
405 assert(!MSSA.isLiveOnEntryDef(MA) && "Hit liveOnEntry before clobber?");
407 if (auto *MD = dyn_cast<MemoryDef>(MA)) {
409 assert(!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA) &&
410 "Found clobber before reaching ClobberAt!");
414 assert(isa<MemoryPhi>(MA));
415 Worklist.append(upward_defs_begin({MA, MAP.second}), upward_defs_end());
419 // If ClobberAt is a MemoryPhi, we can assume something above it acted as a
420 // clobber. Otherwise, `ClobberAt` should've acted as a clobber at some point.
421 assert((isa<MemoryPhi>(ClobberAt) || FoundClobber) &&
422 "ClobberAt never acted as a clobber");
427 /// Our algorithm for walking (and trying to optimize) clobbers, all wrapped up
429 class ClobberWalker {
430 /// Save a few bytes by using unsigned instead of size_t.
431 using ListIndex = unsigned;
433 /// Represents a span of contiguous MemoryDefs, potentially ending in a
437 // Note that, because we always walk in reverse, Last will always dominate
438 // First. Also note that First and Last are inclusive.
441 Optional<ListIndex> Previous;
443 DefPath(const MemoryLocation &Loc, MemoryAccess *First, MemoryAccess *Last,
444 Optional<ListIndex> Previous)
445 : Loc(Loc), First(First), Last(Last), Previous(Previous) {}
447 DefPath(const MemoryLocation &Loc, MemoryAccess *Init,
448 Optional<ListIndex> Previous)
449 : DefPath(Loc, Init, Init, Previous) {}
452 const MemorySSA &MSSA;
455 UpwardsMemoryQuery *Query;
457 // Phi optimization bookkeeping
458 SmallVector<DefPath, 32> Paths;
459 DenseSet<ConstMemoryAccessPair> VisitedPhis;
461 /// Find the nearest def or phi that `From` can legally be optimized to.
462 const MemoryAccess *getWalkTarget(const MemoryPhi *From) const {
463 assert(From->getNumOperands() && "Phi with no operands?");
465 BasicBlock *BB = From->getBlock();
466 MemoryAccess *Result = MSSA.getLiveOnEntryDef();
467 DomTreeNode *Node = DT.getNode(BB);
468 while ((Node = Node->getIDom())) {
469 auto *Defs = MSSA.getBlockDefs(Node->getBlock());
471 return &*Defs->rbegin();
476 /// Result of calling walkToPhiOrClobber.
477 struct UpwardsWalkResult {
478 /// The "Result" of the walk. Either a clobber, the last thing we walked, or
480 MemoryAccess *Result;
484 /// Walk to the next Phi or Clobber in the def chain starting at Desc.Last.
485 /// This will update Desc.Last as it walks. It will (optionally) also stop at
488 /// This does not test for whether StopAt is a clobber
490 walkToPhiOrClobber(DefPath &Desc,
491 const MemoryAccess *StopAt = nullptr) const {
492 assert(!isa<MemoryUse>(Desc.Last) && "Uses don't exist in my world");
494 for (MemoryAccess *Current : def_chain(Desc.Last)) {
496 if (Current == StopAt)
497 return {Current, false};
499 if (auto *MD = dyn_cast<MemoryDef>(Current))
500 if (MSSA.isLiveOnEntryDef(MD) ||
501 instructionClobbersQuery(MD, Desc.Loc, Query->Inst, AA))
505 assert(isa<MemoryPhi>(Desc.Last) &&
506 "Ended at a non-clobber that's not a phi?");
507 return {Desc.Last, false};
510 void addSearches(MemoryPhi *Phi, SmallVectorImpl<ListIndex> &PausedSearches,
511 ListIndex PriorNode) {
512 auto UpwardDefs = make_range(upward_defs_begin({Phi, Paths[PriorNode].Loc}),
514 for (const MemoryAccessPair &P : UpwardDefs) {
515 PausedSearches.push_back(Paths.size());
516 Paths.emplace_back(P.second, P.first, PriorNode);
520 /// Represents a search that terminated after finding a clobber. This clobber
521 /// may or may not be present in the path of defs from LastNode..SearchStart,
522 /// since it may have been retrieved from cache.
523 struct TerminatedPath {
524 MemoryAccess *Clobber;
528 /// Get an access that keeps us from optimizing to the given phi.
530 /// PausedSearches is an array of indices into the Paths array. Its incoming
531 /// value is the indices of searches that stopped at the last phi optimization
532 /// target. It's left in an unspecified state.
534 /// If this returns None, NewPaused is a vector of searches that terminated
535 /// at StopWhere. Otherwise, NewPaused is left in an unspecified state.
536 Optional<TerminatedPath>
537 getBlockingAccess(const MemoryAccess *StopWhere,
538 SmallVectorImpl<ListIndex> &PausedSearches,
539 SmallVectorImpl<ListIndex> &NewPaused,
540 SmallVectorImpl<TerminatedPath> &Terminated) {
541 assert(!PausedSearches.empty() && "No searches to continue?");
543 // BFS vs DFS really doesn't make a difference here, so just do a DFS with
544 // PausedSearches as our stack.
545 while (!PausedSearches.empty()) {
546 ListIndex PathIndex = PausedSearches.pop_back_val();
547 DefPath &Node = Paths[PathIndex];
549 // If we've already visited this path with this MemoryLocation, we don't
550 // need to do so again.
552 // NOTE: That we just drop these paths on the ground makes caching
553 // behavior sporadic. e.g. given a diamond:
558 // ...If we walk D, B, A, C, we'll only cache the result of phi
559 // optimization for A, B, and D; C will be skipped because it dies here.
560 // This arguably isn't the worst thing ever, since:
561 // - We generally query things in a top-down order, so if we got below D
562 // without needing cache entries for {C, MemLoc}, then chances are
563 // that those cache entries would end up ultimately unused.
564 // - We still cache things for A, so C only needs to walk up a bit.
565 // If this behavior becomes problematic, we can fix without a ton of extra
567 if (!VisitedPhis.insert({Node.Last, Node.Loc}).second)
570 UpwardsWalkResult Res = walkToPhiOrClobber(Node, /*StopAt=*/StopWhere);
571 if (Res.IsKnownClobber) {
572 assert(Res.Result != StopWhere);
573 // If this wasn't a cache hit, we hit a clobber when walking. That's a
575 TerminatedPath Term{Res.Result, PathIndex};
576 if (!MSSA.dominates(Res.Result, StopWhere))
579 // Otherwise, it's a valid thing to potentially optimize to.
580 Terminated.push_back(Term);
584 if (Res.Result == StopWhere) {
585 // We've hit our target. Save this path off for if we want to continue
587 NewPaused.push_back(PathIndex);
591 assert(!MSSA.isLiveOnEntryDef(Res.Result) && "liveOnEntry is a clobber");
592 addSearches(cast<MemoryPhi>(Res.Result), PausedSearches, PathIndex);
598 template <typename T, typename Walker>
599 struct generic_def_path_iterator
600 : public iterator_facade_base<generic_def_path_iterator<T, Walker>,
601 std::forward_iterator_tag, T *> {
602 generic_def_path_iterator() = default;
603 generic_def_path_iterator(Walker *W, ListIndex N) : W(W), N(N) {}
605 T &operator*() const { return curNode(); }
607 generic_def_path_iterator &operator++() {
608 N = curNode().Previous;
612 bool operator==(const generic_def_path_iterator &O) const {
613 if (N.hasValue() != O.N.hasValue())
615 return !N.hasValue() || *N == *O.N;
619 T &curNode() const { return W->Paths[*N]; }
622 Optional<ListIndex> N = None;
625 using def_path_iterator = generic_def_path_iterator<DefPath, ClobberWalker>;
626 using const_def_path_iterator =
627 generic_def_path_iterator<const DefPath, const ClobberWalker>;
629 iterator_range<def_path_iterator> def_path(ListIndex From) {
630 return make_range(def_path_iterator(this, From), def_path_iterator());
633 iterator_range<const_def_path_iterator> const_def_path(ListIndex From) const {
634 return make_range(const_def_path_iterator(this, From),
635 const_def_path_iterator());
639 /// The path that contains our result.
640 TerminatedPath PrimaryClobber;
641 /// The paths that we can legally cache back from, but that aren't
642 /// necessarily the result of the Phi optimization.
643 SmallVector<TerminatedPath, 4> OtherClobbers;
646 ListIndex defPathIndex(const DefPath &N) const {
647 // The assert looks nicer if we don't need to do &N
648 const DefPath *NP = &N;
649 assert(!Paths.empty() && NP >= &Paths.front() && NP <= &Paths.back() &&
650 "Out of bounds DefPath!");
651 return NP - &Paths.front();
654 /// Try to optimize a phi as best as we can. Returns a SmallVector of Paths
655 /// that act as legal clobbers. Note that this won't return *all* clobbers.
657 /// Phi optimization algorithm tl;dr:
658 /// - Find the earliest def/phi, A, we can optimize to
659 /// - Find if all paths from the starting memory access ultimately reach A
660 /// - If not, optimization isn't possible.
661 /// - Otherwise, walk from A to another clobber or phi, A'.
662 /// - If A' is a def, we're done.
663 /// - If A' is a phi, try to optimize it.
665 /// A path is a series of {MemoryAccess, MemoryLocation} pairs. A path
666 /// terminates when a MemoryAccess that clobbers said MemoryLocation is found.
667 OptznResult tryOptimizePhi(MemoryPhi *Phi, MemoryAccess *Start,
668 const MemoryLocation &Loc) {
669 assert(Paths.empty() && VisitedPhis.empty() &&
670 "Reset the optimization state.");
672 Paths.emplace_back(Loc, Start, Phi, None);
673 // Stores how many "valid" optimization nodes we had prior to calling
674 // addSearches/getBlockingAccess. Necessary for caching if we had a blocker.
675 auto PriorPathsSize = Paths.size();
677 SmallVector<ListIndex, 16> PausedSearches;
678 SmallVector<ListIndex, 8> NewPaused;
679 SmallVector<TerminatedPath, 4> TerminatedPaths;
681 addSearches(Phi, PausedSearches, 0);
683 // Moves the TerminatedPath with the "most dominated" Clobber to the end of
685 auto MoveDominatedPathToEnd = [&](SmallVectorImpl<TerminatedPath> &Paths) {
686 assert(!Paths.empty() && "Need a path to move");
687 auto Dom = Paths.begin();
688 for (auto I = std::next(Dom), E = Paths.end(); I != E; ++I)
689 if (!MSSA.dominates(I->Clobber, Dom->Clobber))
691 auto Last = Paths.end() - 1;
693 std::iter_swap(Last, Dom);
696 MemoryPhi *Current = Phi;
698 assert(!MSSA.isLiveOnEntryDef(Current) &&
699 "liveOnEntry wasn't treated as a clobber?");
701 const auto *Target = getWalkTarget(Current);
702 // If a TerminatedPath doesn't dominate Target, then it wasn't a legal
703 // optimization for the prior phi.
704 assert(all_of(TerminatedPaths, [&](const TerminatedPath &P) {
705 return MSSA.dominates(P.Clobber, Target);
708 // FIXME: This is broken, because the Blocker may be reported to be
709 // liveOnEntry, and we'll happily wait for that to disappear (read: never)
710 // For the moment, this is fine, since we do nothing with blocker info.
711 if (Optional<TerminatedPath> Blocker = getBlockingAccess(
712 Target, PausedSearches, NewPaused, TerminatedPaths)) {
714 // Find the node we started at. We can't search based on N->Last, since
715 // we may have gone around a loop with a different MemoryLocation.
716 auto Iter = find_if(def_path(Blocker->LastNode), [&](const DefPath &N) {
717 return defPathIndex(N) < PriorPathsSize;
719 assert(Iter != def_path_iterator());
721 DefPath &CurNode = *Iter;
722 assert(CurNode.Last == Current);
725 // A. We can't reliably cache all of NewPaused back. Consider a case
726 // where we have two paths in NewPaused; one of which can't optimize
727 // above this phi, whereas the other can. If we cache the second path
728 // back, we'll end up with suboptimal cache entries. We can handle
729 // cases like this a bit better when we either try to find all
730 // clobbers that block phi optimization, or when our cache starts
731 // supporting unfinished searches.
732 // B. We can't reliably cache TerminatedPaths back here without doing
733 // extra checks; consider a case like:
739 // Where T is our target, C is a node with a clobber on it, D is a
740 // diamond (with a clobber *only* on the left or right node, N), and
741 // S is our start. Say we walk to D, through the node opposite N
742 // (read: ignoring the clobber), and see a cache entry in the top
743 // node of D. That cache entry gets put into TerminatedPaths. We then
744 // walk up to C (N is later in our worklist), find the clobber, and
745 // quit. If we append TerminatedPaths to OtherClobbers, we'll cache
746 // the bottom part of D to the cached clobber, ignoring the clobber
747 // in N. Again, this problem goes away if we start tracking all
748 // blockers for a given phi optimization.
749 TerminatedPath Result{CurNode.Last, defPathIndex(CurNode)};
753 // If there's nothing left to search, then all paths led to valid clobbers
754 // that we got from our cache; pick the nearest to the start, and allow
755 // the rest to be cached back.
756 if (NewPaused.empty()) {
757 MoveDominatedPathToEnd(TerminatedPaths);
758 TerminatedPath Result = TerminatedPaths.pop_back_val();
759 return {Result, std::move(TerminatedPaths)};
762 MemoryAccess *DefChainEnd = nullptr;
763 SmallVector<TerminatedPath, 4> Clobbers;
764 for (ListIndex Paused : NewPaused) {
765 UpwardsWalkResult WR = walkToPhiOrClobber(Paths[Paused]);
766 if (WR.IsKnownClobber)
767 Clobbers.push_back({WR.Result, Paused});
769 // Micro-opt: If we hit the end of the chain, save it.
770 DefChainEnd = WR.Result;
773 if (!TerminatedPaths.empty()) {
774 // If we couldn't find the dominating phi/liveOnEntry in the above loop,
777 for (auto *MA : def_chain(const_cast<MemoryAccess *>(Target)))
780 // If any of the terminated paths don't dominate the phi we'll try to
781 // optimize, we need to figure out what they are and quit.
782 const BasicBlock *ChainBB = DefChainEnd->getBlock();
783 for (const TerminatedPath &TP : TerminatedPaths) {
784 // Because we know that DefChainEnd is as "high" as we can go, we
785 // don't need local dominance checks; BB dominance is sufficient.
786 if (DT.dominates(ChainBB, TP.Clobber->getBlock()))
787 Clobbers.push_back(TP);
791 // If we have clobbers in the def chain, find the one closest to Current
793 if (!Clobbers.empty()) {
794 MoveDominatedPathToEnd(Clobbers);
795 TerminatedPath Result = Clobbers.pop_back_val();
796 return {Result, std::move(Clobbers)};
799 assert(all_of(NewPaused,
800 [&](ListIndex I) { return Paths[I].Last == DefChainEnd; }));
802 // Because liveOnEntry is a clobber, this must be a phi.
803 auto *DefChainPhi = cast<MemoryPhi>(DefChainEnd);
805 PriorPathsSize = Paths.size();
806 PausedSearches.clear();
807 for (ListIndex I : NewPaused)
808 addSearches(DefChainPhi, PausedSearches, I);
811 Current = DefChainPhi;
815 void verifyOptResult(const OptznResult &R) const {
816 assert(all_of(R.OtherClobbers, [&](const TerminatedPath &P) {
817 return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber);
821 void resetPhiOptznState() {
827 ClobberWalker(const MemorySSA &MSSA, AliasAnalysis &AA, DominatorTree &DT)
828 : MSSA(MSSA), AA(AA), DT(DT) {}
832 /// Finds the nearest clobber for the given query, optimizing phis if
834 MemoryAccess *findClobber(MemoryAccess *Start, UpwardsMemoryQuery &Q) {
837 MemoryAccess *Current = Start;
838 // This walker pretends uses don't exist. If we're handed one, silently grab
839 // its def. (This has the nice side-effect of ensuring we never cache uses)
840 if (auto *MU = dyn_cast<MemoryUse>(Start))
841 Current = MU->getDefiningAccess();
843 DefPath FirstDesc(Q.StartingLoc, Current, Current, None);
844 // Fast path for the overly-common case (no crazy phi optimization
846 UpwardsWalkResult WalkResult = walkToPhiOrClobber(FirstDesc);
847 MemoryAccess *Result;
848 if (WalkResult.IsKnownClobber) {
849 Result = WalkResult.Result;
851 OptznResult OptRes = tryOptimizePhi(cast<MemoryPhi>(FirstDesc.Last),
852 Current, Q.StartingLoc);
853 verifyOptResult(OptRes);
854 resetPhiOptznState();
855 Result = OptRes.PrimaryClobber.Clobber;
858 #ifdef EXPENSIVE_CHECKS
859 checkClobberSanity(Current, Result, Q.StartingLoc, MSSA, Q, AA);
864 void verify(const MemorySSA *MSSA) { assert(MSSA == &this->MSSA); }
867 struct RenamePassData {
869 DomTreeNode::const_iterator ChildIt;
870 MemoryAccess *IncomingVal;
872 RenamePassData(DomTreeNode *D, DomTreeNode::const_iterator It,
874 : DTN(D), ChildIt(It), IncomingVal(M) {}
876 void swap(RenamePassData &RHS) {
877 std::swap(DTN, RHS.DTN);
878 std::swap(ChildIt, RHS.ChildIt);
879 std::swap(IncomingVal, RHS.IncomingVal);
883 } // end anonymous namespace
887 /// \brief A MemorySSAWalker that does AA walks to disambiguate accesses. It no
888 /// longer does caching on its own,
889 /// but the name has been retained for the moment.
890 class MemorySSA::CachingWalker final : public MemorySSAWalker {
891 ClobberWalker Walker;
892 bool AutoResetWalker = true;
894 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *, UpwardsMemoryQuery &);
897 CachingWalker(MemorySSA *, AliasAnalysis *, DominatorTree *);
898 ~CachingWalker() override = default;
900 using MemorySSAWalker::getClobberingMemoryAccess;
902 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *) override;
903 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *,
904 const MemoryLocation &) override;
905 void invalidateInfo(MemoryAccess *) override;
907 /// Whether we call resetClobberWalker() after each time we *actually* walk to
908 /// answer a clobber query.
909 void setAutoResetWalker(bool AutoReset) { AutoResetWalker = AutoReset; }
911 /// Drop the walker's persistent data structures.
912 void resetClobberWalker() { Walker.reset(); }
914 void verify(const MemorySSA *MSSA) override {
915 MemorySSAWalker::verify(MSSA);
920 } // end namespace llvm
922 void MemorySSA::renameSuccessorPhis(BasicBlock *BB, MemoryAccess *IncomingVal,
923 bool RenameAllUses) {
924 // Pass through values to our successors
925 for (const BasicBlock *S : successors(BB)) {
926 auto It = PerBlockAccesses.find(S);
927 // Rename the phi nodes in our successor block
928 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
930 AccessList *Accesses = It->second.get();
931 auto *Phi = cast<MemoryPhi>(&Accesses->front());
933 int PhiIndex = Phi->getBasicBlockIndex(BB);
934 assert(PhiIndex != -1 && "Incomplete phi during partial rename");
935 Phi->setIncomingValue(PhiIndex, IncomingVal);
937 Phi->addIncoming(IncomingVal, BB);
941 /// \brief Rename a single basic block into MemorySSA form.
942 /// Uses the standard SSA renaming algorithm.
943 /// \returns The new incoming value.
944 MemoryAccess *MemorySSA::renameBlock(BasicBlock *BB, MemoryAccess *IncomingVal,
945 bool RenameAllUses) {
946 auto It = PerBlockAccesses.find(BB);
947 // Skip most processing if the list is empty.
948 if (It != PerBlockAccesses.end()) {
949 AccessList *Accesses = It->second.get();
950 for (MemoryAccess &L : *Accesses) {
951 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(&L)) {
952 if (MUD->getDefiningAccess() == nullptr || RenameAllUses)
953 MUD->setDefiningAccess(IncomingVal);
954 if (isa<MemoryDef>(&L))
964 /// \brief This is the standard SSA renaming algorithm.
966 /// We walk the dominator tree in preorder, renaming accesses, and then filling
967 /// in phi nodes in our successors.
968 void MemorySSA::renamePass(DomTreeNode *Root, MemoryAccess *IncomingVal,
969 SmallPtrSetImpl<BasicBlock *> &Visited,
970 bool SkipVisited, bool RenameAllUses) {
971 SmallVector<RenamePassData, 32> WorkStack;
972 // Skip everything if we already renamed this block and we are skipping.
973 // Note: You can't sink this into the if, because we need it to occur
974 // regardless of whether we skip blocks or not.
975 bool AlreadyVisited = !Visited.insert(Root->getBlock()).second;
976 if (SkipVisited && AlreadyVisited)
979 IncomingVal = renameBlock(Root->getBlock(), IncomingVal, RenameAllUses);
980 renameSuccessorPhis(Root->getBlock(), IncomingVal, RenameAllUses);
981 WorkStack.push_back({Root, Root->begin(), IncomingVal});
983 while (!WorkStack.empty()) {
984 DomTreeNode *Node = WorkStack.back().DTN;
985 DomTreeNode::const_iterator ChildIt = WorkStack.back().ChildIt;
986 IncomingVal = WorkStack.back().IncomingVal;
988 if (ChildIt == Node->end()) {
989 WorkStack.pop_back();
991 DomTreeNode *Child = *ChildIt;
992 ++WorkStack.back().ChildIt;
993 BasicBlock *BB = Child->getBlock();
994 // Note: You can't sink this into the if, because we need it to occur
995 // regardless of whether we skip blocks or not.
996 AlreadyVisited = !Visited.insert(BB).second;
997 if (SkipVisited && AlreadyVisited) {
998 // We already visited this during our renaming, which can happen when
999 // being asked to rename multiple blocks. Figure out the incoming val,
1000 // which is the last def.
1001 // Incoming value can only change if there is a block def, and in that
1002 // case, it's the last block def in the list.
1003 if (auto *BlockDefs = getWritableBlockDefs(BB))
1004 IncomingVal = &*BlockDefs->rbegin();
1006 IncomingVal = renameBlock(BB, IncomingVal, RenameAllUses);
1007 renameSuccessorPhis(BB, IncomingVal, RenameAllUses);
1008 WorkStack.push_back({Child, Child->begin(), IncomingVal});
1013 /// \brief This handles unreachable block accesses by deleting phi nodes in
1014 /// unreachable blocks, and marking all other unreachable MemoryAccess's as
1015 /// being uses of the live on entry definition.
1016 void MemorySSA::markUnreachableAsLiveOnEntry(BasicBlock *BB) {
1017 assert(!DT->isReachableFromEntry(BB) &&
1018 "Reachable block found while handling unreachable blocks");
1020 // Make sure phi nodes in our reachable successors end up with a
1021 // LiveOnEntryDef for our incoming edge, even though our block is forward
1022 // unreachable. We could just disconnect these blocks from the CFG fully,
1023 // but we do not right now.
1024 for (const BasicBlock *S : successors(BB)) {
1025 if (!DT->isReachableFromEntry(S))
1027 auto It = PerBlockAccesses.find(S);
1028 // Rename the phi nodes in our successor block
1029 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
1031 AccessList *Accesses = It->second.get();
1032 auto *Phi = cast<MemoryPhi>(&Accesses->front());
1033 Phi->addIncoming(LiveOnEntryDef.get(), BB);
1036 auto It = PerBlockAccesses.find(BB);
1037 if (It == PerBlockAccesses.end())
1040 auto &Accesses = It->second;
1041 for (auto AI = Accesses->begin(), AE = Accesses->end(); AI != AE;) {
1042 auto Next = std::next(AI);
1043 // If we have a phi, just remove it. We are going to replace all
1044 // users with live on entry.
1045 if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(AI))
1046 UseOrDef->setDefiningAccess(LiveOnEntryDef.get());
1048 Accesses->erase(AI);
1053 MemorySSA::MemorySSA(Function &Func, AliasAnalysis *AA, DominatorTree *DT)
1054 : AA(AA), DT(DT), F(Func), LiveOnEntryDef(nullptr), Walker(nullptr),
1055 NextID(INVALID_MEMORYACCESS_ID) {
1059 MemorySSA::~MemorySSA() {
1060 // Drop all our references
1061 for (const auto &Pair : PerBlockAccesses)
1062 for (MemoryAccess &MA : *Pair.second)
1063 MA.dropAllReferences();
1066 MemorySSA::AccessList *MemorySSA::getOrCreateAccessList(const BasicBlock *BB) {
1067 auto Res = PerBlockAccesses.insert(std::make_pair(BB, nullptr));
1070 Res.first->second = llvm::make_unique<AccessList>();
1071 return Res.first->second.get();
1074 MemorySSA::DefsList *MemorySSA::getOrCreateDefsList(const BasicBlock *BB) {
1075 auto Res = PerBlockDefs.insert(std::make_pair(BB, nullptr));
1078 Res.first->second = llvm::make_unique<DefsList>();
1079 return Res.first->second.get();
1084 /// This class is a batch walker of all MemoryUse's in the program, and points
1085 /// their defining access at the thing that actually clobbers them. Because it
1086 /// is a batch walker that touches everything, it does not operate like the
1087 /// other walkers. This walker is basically performing a top-down SSA renaming
1088 /// pass, where the version stack is used as the cache. This enables it to be
1089 /// significantly more time and memory efficient than using the regular walker,
1090 /// which is walking bottom-up.
1091 class MemorySSA::OptimizeUses {
1093 OptimizeUses(MemorySSA *MSSA, MemorySSAWalker *Walker, AliasAnalysis *AA,
1095 : MSSA(MSSA), Walker(Walker), AA(AA), DT(DT) {
1096 Walker = MSSA->getWalker();
1099 void optimizeUses();
1102 /// This represents where a given memorylocation is in the stack.
1103 struct MemlocStackInfo {
1104 // This essentially is keeping track of versions of the stack. Whenever
1105 // the stack changes due to pushes or pops, these versions increase.
1106 unsigned long StackEpoch;
1107 unsigned long PopEpoch;
1108 // This is the lower bound of places on the stack to check. It is equal to
1109 // the place the last stack walk ended.
1110 // Note: Correctness depends on this being initialized to 0, which densemap
1112 unsigned long LowerBound;
1113 const BasicBlock *LowerBoundBlock;
1114 // This is where the last walk for this memory location ended.
1115 unsigned long LastKill;
1119 void optimizeUsesInBlock(const BasicBlock *, unsigned long &, unsigned long &,
1120 SmallVectorImpl<MemoryAccess *> &,
1121 DenseMap<MemoryLocOrCall, MemlocStackInfo> &);
1124 MemorySSAWalker *Walker;
1129 } // end namespace llvm
1131 /// Optimize the uses in a given block This is basically the SSA renaming
1132 /// algorithm, with one caveat: We are able to use a single stack for all
1133 /// MemoryUses. This is because the set of *possible* reaching MemoryDefs is
1134 /// the same for every MemoryUse. The *actual* clobbering MemoryDef is just
1135 /// going to be some position in that stack of possible ones.
1137 /// We track the stack positions that each MemoryLocation needs
1138 /// to check, and last ended at. This is because we only want to check the
1139 /// things that changed since last time. The same MemoryLocation should
1140 /// get clobbered by the same store (getModRefInfo does not use invariantness or
1141 /// things like this, and if they start, we can modify MemoryLocOrCall to
1142 /// include relevant data)
1143 void MemorySSA::OptimizeUses::optimizeUsesInBlock(
1144 const BasicBlock *BB, unsigned long &StackEpoch, unsigned long &PopEpoch,
1145 SmallVectorImpl<MemoryAccess *> &VersionStack,
1146 DenseMap<MemoryLocOrCall, MemlocStackInfo> &LocStackInfo) {
1148 /// If no accesses, nothing to do.
1149 MemorySSA::AccessList *Accesses = MSSA->getWritableBlockAccesses(BB);
1150 if (Accesses == nullptr)
1153 // Pop everything that doesn't dominate the current block off the stack,
1154 // increment the PopEpoch to account for this.
1157 !VersionStack.empty() &&
1158 "Version stack should have liveOnEntry sentinel dominating everything");
1159 BasicBlock *BackBlock = VersionStack.back()->getBlock();
1160 if (DT->dominates(BackBlock, BB))
1162 while (VersionStack.back()->getBlock() == BackBlock)
1163 VersionStack.pop_back();
1167 for (MemoryAccess &MA : *Accesses) {
1168 auto *MU = dyn_cast<MemoryUse>(&MA);
1170 VersionStack.push_back(&MA);
1175 if (isUseTriviallyOptimizableToLiveOnEntry(*AA, MU->getMemoryInst())) {
1176 MU->setDefiningAccess(MSSA->getLiveOnEntryDef(), true);
1180 MemoryLocOrCall UseMLOC(MU);
1181 auto &LocInfo = LocStackInfo[UseMLOC];
1182 // If the pop epoch changed, it means we've removed stuff from top of
1183 // stack due to changing blocks. We may have to reset the lower bound or
1185 if (LocInfo.PopEpoch != PopEpoch) {
1186 LocInfo.PopEpoch = PopEpoch;
1187 LocInfo.StackEpoch = StackEpoch;
1188 // If the lower bound was in something that no longer dominates us, we
1189 // have to reset it.
1190 // We can't simply track stack size, because the stack may have had
1191 // pushes/pops in the meantime.
1192 // XXX: This is non-optimal, but only is slower cases with heavily
1193 // branching dominator trees. To get the optimal number of queries would
1194 // be to make lowerbound and lastkill a per-loc stack, and pop it until
1195 // the top of that stack dominates us. This does not seem worth it ATM.
1196 // A much cheaper optimization would be to always explore the deepest
1197 // branch of the dominator tree first. This will guarantee this resets on
1198 // the smallest set of blocks.
1199 if (LocInfo.LowerBoundBlock && LocInfo.LowerBoundBlock != BB &&
1200 !DT->dominates(LocInfo.LowerBoundBlock, BB)) {
1201 // Reset the lower bound of things to check.
1202 // TODO: Some day we should be able to reset to last kill, rather than
1204 LocInfo.LowerBound = 0;
1205 LocInfo.LowerBoundBlock = VersionStack[0]->getBlock();
1206 LocInfo.LastKillValid = false;
1208 } else if (LocInfo.StackEpoch != StackEpoch) {
1209 // If all that has changed is the StackEpoch, we only have to check the
1210 // new things on the stack, because we've checked everything before. In
1211 // this case, the lower bound of things to check remains the same.
1212 LocInfo.PopEpoch = PopEpoch;
1213 LocInfo.StackEpoch = StackEpoch;
1215 if (!LocInfo.LastKillValid) {
1216 LocInfo.LastKill = VersionStack.size() - 1;
1217 LocInfo.LastKillValid = true;
1220 // At this point, we should have corrected last kill and LowerBound to be
1222 assert(LocInfo.LowerBound < VersionStack.size() &&
1223 "Lower bound out of range");
1224 assert(LocInfo.LastKill < VersionStack.size() &&
1225 "Last kill info out of range");
1226 // In any case, the new upper bound is the top of the stack.
1227 unsigned long UpperBound = VersionStack.size() - 1;
1229 if (UpperBound - LocInfo.LowerBound > MaxCheckLimit) {
1230 DEBUG(dbgs() << "MemorySSA skipping optimization of " << *MU << " ("
1231 << *(MU->getMemoryInst()) << ")"
1232 << " because there are " << UpperBound - LocInfo.LowerBound
1233 << " stores to disambiguate\n");
1234 // Because we did not walk, LastKill is no longer valid, as this may
1235 // have been a kill.
1236 LocInfo.LastKillValid = false;
1239 bool FoundClobberResult = false;
1240 while (UpperBound > LocInfo.LowerBound) {
1241 if (isa<MemoryPhi>(VersionStack[UpperBound])) {
1242 // For phis, use the walker, see where we ended up, go there
1243 Instruction *UseInst = MU->getMemoryInst();
1244 MemoryAccess *Result = Walker->getClobberingMemoryAccess(UseInst);
1245 // We are guaranteed to find it or something is wrong
1246 while (VersionStack[UpperBound] != Result) {
1247 assert(UpperBound != 0);
1250 FoundClobberResult = true;
1254 MemoryDef *MD = cast<MemoryDef>(VersionStack[UpperBound]);
1255 // If the lifetime of the pointer ends at this instruction, it's live on
1257 if (!UseMLOC.IsCall && lifetimeEndsAt(MD, UseMLOC.getLoc(), *AA)) {
1258 // Reset UpperBound to liveOnEntryDef's place in the stack
1260 FoundClobberResult = true;
1263 if (instructionClobbersQuery(MD, MU, UseMLOC, *AA)) {
1264 FoundClobberResult = true;
1269 // At the end of this loop, UpperBound is either a clobber, or lower bound
1270 // PHI walking may cause it to be < LowerBound, and in fact, < LastKill.
1271 if (FoundClobberResult || UpperBound < LocInfo.LastKill) {
1272 MU->setDefiningAccess(VersionStack[UpperBound], true);
1273 // We were last killed now by where we got to
1274 LocInfo.LastKill = UpperBound;
1276 // Otherwise, we checked all the new ones, and now we know we can get to
1278 MU->setDefiningAccess(VersionStack[LocInfo.LastKill], true);
1280 LocInfo.LowerBound = VersionStack.size() - 1;
1281 LocInfo.LowerBoundBlock = BB;
1285 /// Optimize uses to point to their actual clobbering definitions.
1286 void MemorySSA::OptimizeUses::optimizeUses() {
1287 SmallVector<MemoryAccess *, 16> VersionStack;
1288 DenseMap<MemoryLocOrCall, MemlocStackInfo> LocStackInfo;
1289 VersionStack.push_back(MSSA->getLiveOnEntryDef());
1291 unsigned long StackEpoch = 1;
1292 unsigned long PopEpoch = 1;
1293 // We perform a non-recursive top-down dominator tree walk.
1294 for (const auto *DomNode : depth_first(DT->getRootNode()))
1295 optimizeUsesInBlock(DomNode->getBlock(), StackEpoch, PopEpoch, VersionStack,
1299 void MemorySSA::placePHINodes(
1300 const SmallPtrSetImpl<BasicBlock *> &DefiningBlocks,
1301 const DenseMap<const BasicBlock *, unsigned int> &BBNumbers) {
1302 // Determine where our MemoryPhi's should go
1303 ForwardIDFCalculator IDFs(*DT);
1304 IDFs.setDefiningBlocks(DefiningBlocks);
1305 SmallVector<BasicBlock *, 32> IDFBlocks;
1306 IDFs.calculate(IDFBlocks);
1308 std::sort(IDFBlocks.begin(), IDFBlocks.end(),
1309 [&BBNumbers](const BasicBlock *A, const BasicBlock *B) {
1310 return BBNumbers.lookup(A) < BBNumbers.lookup(B);
1313 // Now place MemoryPhi nodes.
1314 for (auto &BB : IDFBlocks)
1315 createMemoryPhi(BB);
1318 void MemorySSA::buildMemorySSA() {
1319 // We create an access to represent "live on entry", for things like
1320 // arguments or users of globals, where the memory they use is defined before
1321 // the beginning of the function. We do not actually insert it into the IR.
1322 // We do not define a live on exit for the immediate uses, and thus our
1323 // semantics do *not* imply that something with no immediate uses can simply
1325 BasicBlock &StartingPoint = F.getEntryBlock();
1327 llvm::make_unique<MemoryDef>(F.getContext(), nullptr, nullptr,
1328 &StartingPoint, NextID++);
1329 DenseMap<const BasicBlock *, unsigned int> BBNumbers;
1330 unsigned NextBBNum = 0;
1332 // We maintain lists of memory accesses per-block, trading memory for time. We
1333 // could just look up the memory access for every possible instruction in the
1335 SmallPtrSet<BasicBlock *, 32> DefiningBlocks;
1336 // Go through each block, figure out where defs occur, and chain together all
1338 for (BasicBlock &B : F) {
1339 BBNumbers[&B] = NextBBNum++;
1340 bool InsertIntoDef = false;
1341 AccessList *Accesses = nullptr;
1342 DefsList *Defs = nullptr;
1343 for (Instruction &I : B) {
1344 MemoryUseOrDef *MUD = createNewAccess(&I);
1349 Accesses = getOrCreateAccessList(&B);
1350 Accesses->push_back(MUD);
1351 if (isa<MemoryDef>(MUD)) {
1352 InsertIntoDef = true;
1354 Defs = getOrCreateDefsList(&B);
1355 Defs->push_back(*MUD);
1359 DefiningBlocks.insert(&B);
1361 placePHINodes(DefiningBlocks, BBNumbers);
1363 // Now do regular SSA renaming on the MemoryDef/MemoryUse. Visited will get
1364 // filled in with all blocks.
1365 SmallPtrSet<BasicBlock *, 16> Visited;
1366 renamePass(DT->getRootNode(), LiveOnEntryDef.get(), Visited);
1368 CachingWalker *Walker = getWalkerImpl();
1370 // We're doing a batch of updates; don't drop useful caches between them.
1371 Walker->setAutoResetWalker(false);
1372 OptimizeUses(this, Walker, AA, DT).optimizeUses();
1373 Walker->setAutoResetWalker(true);
1374 Walker->resetClobberWalker();
1376 // Mark the uses in unreachable blocks as live on entry, so that they go
1379 if (!Visited.count(&BB))
1380 markUnreachableAsLiveOnEntry(&BB);
1383 MemorySSAWalker *MemorySSA::getWalker() { return getWalkerImpl(); }
1385 MemorySSA::CachingWalker *MemorySSA::getWalkerImpl() {
1387 return Walker.get();
1389 Walker = llvm::make_unique<CachingWalker>(this, AA, DT);
1390 return Walker.get();
1393 // This is a helper function used by the creation routines. It places NewAccess
1394 // into the access and defs lists for a given basic block, at the given
1396 void MemorySSA::insertIntoListsForBlock(MemoryAccess *NewAccess,
1397 const BasicBlock *BB,
1398 InsertionPlace Point) {
1399 auto *Accesses = getOrCreateAccessList(BB);
1400 if (Point == Beginning) {
1401 // If it's a phi node, it goes first, otherwise, it goes after any phi
1403 if (isa<MemoryPhi>(NewAccess)) {
1404 Accesses->push_front(NewAccess);
1405 auto *Defs = getOrCreateDefsList(BB);
1406 Defs->push_front(*NewAccess);
1408 auto AI = find_if_not(
1409 *Accesses, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
1410 Accesses->insert(AI, NewAccess);
1411 if (!isa<MemoryUse>(NewAccess)) {
1412 auto *Defs = getOrCreateDefsList(BB);
1413 auto DI = find_if_not(
1414 *Defs, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
1415 Defs->insert(DI, *NewAccess);
1419 Accesses->push_back(NewAccess);
1420 if (!isa<MemoryUse>(NewAccess)) {
1421 auto *Defs = getOrCreateDefsList(BB);
1422 Defs->push_back(*NewAccess);
1425 BlockNumberingValid.erase(BB);
1428 void MemorySSA::insertIntoListsBefore(MemoryAccess *What, const BasicBlock *BB,
1429 AccessList::iterator InsertPt) {
1430 auto *Accesses = getWritableBlockAccesses(BB);
1431 bool WasEnd = InsertPt == Accesses->end();
1432 Accesses->insert(AccessList::iterator(InsertPt), What);
1433 if (!isa<MemoryUse>(What)) {
1434 auto *Defs = getOrCreateDefsList(BB);
1435 // If we got asked to insert at the end, we have an easy job, just shove it
1436 // at the end. If we got asked to insert before an existing def, we also get
1437 // an terator. If we got asked to insert before a use, we have to hunt for
1440 Defs->push_back(*What);
1441 } else if (isa<MemoryDef>(InsertPt)) {
1442 Defs->insert(InsertPt->getDefsIterator(), *What);
1444 while (InsertPt != Accesses->end() && !isa<MemoryDef>(InsertPt))
1446 // Either we found a def, or we are inserting at the end
1447 if (InsertPt == Accesses->end())
1448 Defs->push_back(*What);
1450 Defs->insert(InsertPt->getDefsIterator(), *What);
1453 BlockNumberingValid.erase(BB);
1456 // Move What before Where in the IR. The end result is taht What will belong to
1457 // the right lists and have the right Block set, but will not otherwise be
1458 // correct. It will not have the right defining access, and if it is a def,
1459 // things below it will not properly be updated.
1460 void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB,
1461 AccessList::iterator Where) {
1462 // Keep it in the lookup tables, remove from the lists
1463 removeFromLists(What, false);
1465 insertIntoListsBefore(What, BB, Where);
1468 void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB,
1469 InsertionPlace Point) {
1470 removeFromLists(What, false);
1472 insertIntoListsForBlock(What, BB, Point);
1475 MemoryPhi *MemorySSA::createMemoryPhi(BasicBlock *BB) {
1476 assert(!getMemoryAccess(BB) && "MemoryPhi already exists for this BB");
1477 MemoryPhi *Phi = new MemoryPhi(BB->getContext(), BB, NextID++);
1478 // Phi's always are placed at the front of the block.
1479 insertIntoListsForBlock(Phi, BB, Beginning);
1480 ValueToMemoryAccess[BB] = Phi;
1484 MemoryUseOrDef *MemorySSA::createDefinedAccess(Instruction *I,
1485 MemoryAccess *Definition) {
1486 assert(!isa<PHINode>(I) && "Cannot create a defined access for a PHI");
1487 MemoryUseOrDef *NewAccess = createNewAccess(I);
1489 NewAccess != nullptr &&
1490 "Tried to create a memory access for a non-memory touching instruction");
1491 NewAccess->setDefiningAccess(Definition);
1495 // Return true if the instruction has ordering constraints.
1496 // Note specifically that this only considers stores and loads
1497 // because others are still considered ModRef by getModRefInfo.
1498 static inline bool isOrdered(const Instruction *I) {
1499 if (auto *SI = dyn_cast<StoreInst>(I)) {
1500 if (!SI->isUnordered())
1502 } else if (auto *LI = dyn_cast<LoadInst>(I)) {
1503 if (!LI->isUnordered())
1509 /// \brief Helper function to create new memory accesses
1510 MemoryUseOrDef *MemorySSA::createNewAccess(Instruction *I) {
1511 // The assume intrinsic has a control dependency which we model by claiming
1512 // that it writes arbitrarily. Ignore that fake memory dependency here.
1513 // FIXME: Replace this special casing with a more accurate modelling of
1514 // assume's control dependency.
1515 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1516 if (II->getIntrinsicID() == Intrinsic::assume)
1519 // Find out what affect this instruction has on memory.
1520 ModRefInfo ModRef = AA->getModRefInfo(I, None);
1521 // The isOrdered check is used to ensure that volatiles end up as defs
1522 // (atomics end up as ModRef right now anyway). Until we separate the
1523 // ordering chain from the memory chain, this enables people to see at least
1524 // some relative ordering to volatiles. Note that getClobberingMemoryAccess
1525 // will still give an answer that bypasses other volatile loads. TODO:
1526 // Separate memory aliasing and ordering into two different chains so that we
1527 // can precisely represent both "what memory will this read/write/is clobbered
1528 // by" and "what instructions can I move this past".
1529 bool Def = isModSet(ModRef) || isOrdered(I);
1530 bool Use = isRefSet(ModRef);
1532 // It's possible for an instruction to not modify memory at all. During
1533 // construction, we ignore them.
1537 assert((Def || Use) &&
1538 "Trying to create a memory access with a non-memory instruction");
1540 MemoryUseOrDef *MUD;
1542 MUD = new MemoryDef(I->getContext(), nullptr, I, I->getParent(), NextID++);
1544 MUD = new MemoryUse(I->getContext(), nullptr, I, I->getParent());
1545 ValueToMemoryAccess[I] = MUD;
1549 /// \brief Returns true if \p Replacer dominates \p Replacee .
1550 bool MemorySSA::dominatesUse(const MemoryAccess *Replacer,
1551 const MemoryAccess *Replacee) const {
1552 if (isa<MemoryUseOrDef>(Replacee))
1553 return DT->dominates(Replacer->getBlock(), Replacee->getBlock());
1554 const auto *MP = cast<MemoryPhi>(Replacee);
1555 // For a phi node, the use occurs in the predecessor block of the phi node.
1556 // Since we may occur multiple times in the phi node, we have to check each
1557 // operand to ensure Replacer dominates each operand where Replacee occurs.
1558 for (const Use &Arg : MP->operands()) {
1559 if (Arg.get() != Replacee &&
1560 !DT->dominates(Replacer->getBlock(), MP->getIncomingBlock(Arg)))
1566 /// \brief Properly remove \p MA from all of MemorySSA's lookup tables.
1567 void MemorySSA::removeFromLookups(MemoryAccess *MA) {
1568 assert(MA->use_empty() &&
1569 "Trying to remove memory access that still has uses");
1570 BlockNumbering.erase(MA);
1571 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA))
1572 MUD->setDefiningAccess(nullptr);
1573 // Invalidate our walker's cache if necessary
1574 if (!isa<MemoryUse>(MA))
1575 Walker->invalidateInfo(MA);
1576 // The call below to erase will destroy MA, so we can't change the order we
1577 // are doing things here
1579 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA)) {
1580 MemoryInst = MUD->getMemoryInst();
1582 MemoryInst = MA->getBlock();
1584 auto VMA = ValueToMemoryAccess.find(MemoryInst);
1585 if (VMA->second == MA)
1586 ValueToMemoryAccess.erase(VMA);
1589 /// \brief Properly remove \p MA from all of MemorySSA's lists.
1591 /// Because of the way the intrusive list and use lists work, it is important to
1592 /// do removal in the right order.
1593 /// ShouldDelete defaults to true, and will cause the memory access to also be
1594 /// deleted, not just removed.
1595 void MemorySSA::removeFromLists(MemoryAccess *MA, bool ShouldDelete) {
1596 // The access list owns the reference, so we erase it from the non-owning list
1598 if (!isa<MemoryUse>(MA)) {
1599 auto DefsIt = PerBlockDefs.find(MA->getBlock());
1600 std::unique_ptr<DefsList> &Defs = DefsIt->second;
1603 PerBlockDefs.erase(DefsIt);
1606 // The erase call here will delete it. If we don't want it deleted, we call
1608 auto AccessIt = PerBlockAccesses.find(MA->getBlock());
1609 std::unique_ptr<AccessList> &Accesses = AccessIt->second;
1611 Accesses->erase(MA);
1613 Accesses->remove(MA);
1615 if (Accesses->empty())
1616 PerBlockAccesses.erase(AccessIt);
1619 void MemorySSA::print(raw_ostream &OS) const {
1620 MemorySSAAnnotatedWriter Writer(this);
1621 F.print(OS, &Writer);
1624 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1625 LLVM_DUMP_METHOD void MemorySSA::dump() const { print(dbgs()); }
1628 void MemorySSA::verifyMemorySSA() const {
1630 verifyDomination(F);
1632 Walker->verify(this);
1635 /// \brief Verify that the order and existence of MemoryAccesses matches the
1636 /// order and existence of memory affecting instructions.
1637 void MemorySSA::verifyOrdering(Function &F) const {
1638 // Walk all the blocks, comparing what the lookups think and what the access
1639 // lists think, as well as the order in the blocks vs the order in the access
1641 SmallVector<MemoryAccess *, 32> ActualAccesses;
1642 SmallVector<MemoryAccess *, 32> ActualDefs;
1643 for (BasicBlock &B : F) {
1644 const AccessList *AL = getBlockAccesses(&B);
1645 const auto *DL = getBlockDefs(&B);
1646 MemoryAccess *Phi = getMemoryAccess(&B);
1648 ActualAccesses.push_back(Phi);
1649 ActualDefs.push_back(Phi);
1652 for (Instruction &I : B) {
1653 MemoryAccess *MA = getMemoryAccess(&I);
1654 assert((!MA || (AL && (isa<MemoryUse>(MA) || DL))) &&
1655 "We have memory affecting instructions "
1656 "in this block but they are not in the "
1657 "access list or defs list");
1659 ActualAccesses.push_back(MA);
1660 if (isa<MemoryDef>(MA))
1661 ActualDefs.push_back(MA);
1664 // Either we hit the assert, really have no accesses, or we have both
1665 // accesses and an access list.
1669 assert(AL->size() == ActualAccesses.size() &&
1670 "We don't have the same number of accesses in the block as on the "
1672 assert((DL || ActualDefs.size() == 0) &&
1673 "Either we should have a defs list, or we should have no defs");
1674 assert((!DL || DL->size() == ActualDefs.size()) &&
1675 "We don't have the same number of defs in the block as on the "
1677 auto ALI = AL->begin();
1678 auto AAI = ActualAccesses.begin();
1679 while (ALI != AL->end() && AAI != ActualAccesses.end()) {
1680 assert(&*ALI == *AAI && "Not the same accesses in the same order");
1684 ActualAccesses.clear();
1686 auto DLI = DL->begin();
1687 auto ADI = ActualDefs.begin();
1688 while (DLI != DL->end() && ADI != ActualDefs.end()) {
1689 assert(&*DLI == *ADI && "Not the same defs in the same order");
1698 /// \brief Verify the domination properties of MemorySSA by checking that each
1699 /// definition dominates all of its uses.
1700 void MemorySSA::verifyDomination(Function &F) const {
1702 for (BasicBlock &B : F) {
1703 // Phi nodes are attached to basic blocks
1704 if (MemoryPhi *MP = getMemoryAccess(&B))
1705 for (const Use &U : MP->uses())
1706 assert(dominates(MP, U) && "Memory PHI does not dominate it's uses");
1708 for (Instruction &I : B) {
1709 MemoryAccess *MD = dyn_cast_or_null<MemoryDef>(getMemoryAccess(&I));
1713 for (const Use &U : MD->uses())
1714 assert(dominates(MD, U) && "Memory Def does not dominate it's uses");
1720 /// \brief Verify the def-use lists in MemorySSA, by verifying that \p Use
1721 /// appears in the use list of \p Def.
1722 void MemorySSA::verifyUseInDefs(MemoryAccess *Def, MemoryAccess *Use) const {
1724 // The live on entry use may cause us to get a NULL def here
1726 assert(isLiveOnEntryDef(Use) &&
1727 "Null def but use not point to live on entry def");
1729 assert(is_contained(Def->users(), Use) &&
1730 "Did not find use in def's use list");
1734 /// \brief Verify the immediate use information, by walking all the memory
1735 /// accesses and verifying that, for each use, it appears in the
1736 /// appropriate def's use list
1737 void MemorySSA::verifyDefUses(Function &F) const {
1738 for (BasicBlock &B : F) {
1739 // Phi nodes are attached to basic blocks
1740 if (MemoryPhi *Phi = getMemoryAccess(&B)) {
1741 assert(Phi->getNumOperands() == static_cast<unsigned>(std::distance(
1742 pred_begin(&B), pred_end(&B))) &&
1743 "Incomplete MemoryPhi Node");
1744 for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I)
1745 verifyUseInDefs(Phi->getIncomingValue(I), Phi);
1748 for (Instruction &I : B) {
1749 if (MemoryUseOrDef *MA = getMemoryAccess(&I)) {
1750 verifyUseInDefs(MA->getDefiningAccess(), MA);
1756 MemoryUseOrDef *MemorySSA::getMemoryAccess(const Instruction *I) const {
1757 return cast_or_null<MemoryUseOrDef>(ValueToMemoryAccess.lookup(I));
1760 MemoryPhi *MemorySSA::getMemoryAccess(const BasicBlock *BB) const {
1761 return cast_or_null<MemoryPhi>(ValueToMemoryAccess.lookup(cast<Value>(BB)));
1764 /// Perform a local numbering on blocks so that instruction ordering can be
1765 /// determined in constant time.
1766 /// TODO: We currently just number in order. If we numbered by N, we could
1767 /// allow at least N-1 sequences of insertBefore or insertAfter (and at least
1768 /// log2(N) sequences of mixed before and after) without needing to invalidate
1770 void MemorySSA::renumberBlock(const BasicBlock *B) const {
1771 // The pre-increment ensures the numbers really start at 1.
1772 unsigned long CurrentNumber = 0;
1773 const AccessList *AL = getBlockAccesses(B);
1774 assert(AL != nullptr && "Asking to renumber an empty block");
1775 for (const auto &I : *AL)
1776 BlockNumbering[&I] = ++CurrentNumber;
1777 BlockNumberingValid.insert(B);
1780 /// \brief Determine, for two memory accesses in the same block,
1781 /// whether \p Dominator dominates \p Dominatee.
1782 /// \returns True if \p Dominator dominates \p Dominatee.
1783 bool MemorySSA::locallyDominates(const MemoryAccess *Dominator,
1784 const MemoryAccess *Dominatee) const {
1785 const BasicBlock *DominatorBlock = Dominator->getBlock();
1787 assert((DominatorBlock == Dominatee->getBlock()) &&
1788 "Asking for local domination when accesses are in different blocks!");
1789 // A node dominates itself.
1790 if (Dominatee == Dominator)
1793 // When Dominatee is defined on function entry, it is not dominated by another
1795 if (isLiveOnEntryDef(Dominatee))
1798 // When Dominator is defined on function entry, it dominates the other memory
1800 if (isLiveOnEntryDef(Dominator))
1803 if (!BlockNumberingValid.count(DominatorBlock))
1804 renumberBlock(DominatorBlock);
1806 unsigned long DominatorNum = BlockNumbering.lookup(Dominator);
1807 // All numbers start with 1
1808 assert(DominatorNum != 0 && "Block was not numbered properly");
1809 unsigned long DominateeNum = BlockNumbering.lookup(Dominatee);
1810 assert(DominateeNum != 0 && "Block was not numbered properly");
1811 return DominatorNum < DominateeNum;
1814 bool MemorySSA::dominates(const MemoryAccess *Dominator,
1815 const MemoryAccess *Dominatee) const {
1816 if (Dominator == Dominatee)
1819 if (isLiveOnEntryDef(Dominatee))
1822 if (Dominator->getBlock() != Dominatee->getBlock())
1823 return DT->dominates(Dominator->getBlock(), Dominatee->getBlock());
1824 return locallyDominates(Dominator, Dominatee);
1827 bool MemorySSA::dominates(const MemoryAccess *Dominator,
1828 const Use &Dominatee) const {
1829 if (MemoryPhi *MP = dyn_cast<MemoryPhi>(Dominatee.getUser())) {
1830 BasicBlock *UseBB = MP->getIncomingBlock(Dominatee);
1831 // The def must dominate the incoming block of the phi.
1832 if (UseBB != Dominator->getBlock())
1833 return DT->dominates(Dominator->getBlock(), UseBB);
1834 // If the UseBB and the DefBB are the same, compare locally.
1835 return locallyDominates(Dominator, cast<MemoryAccess>(Dominatee));
1837 // If it's not a PHI node use, the normal dominates can already handle it.
1838 return dominates(Dominator, cast<MemoryAccess>(Dominatee.getUser()));
1841 const static char LiveOnEntryStr[] = "liveOnEntry";
1843 void MemoryAccess::print(raw_ostream &OS) const {
1844 switch (getValueID()) {
1845 case MemoryPhiVal: return static_cast<const MemoryPhi *>(this)->print(OS);
1846 case MemoryDefVal: return static_cast<const MemoryDef *>(this)->print(OS);
1847 case MemoryUseVal: return static_cast<const MemoryUse *>(this)->print(OS);
1849 llvm_unreachable("invalid value id");
1852 void MemoryDef::print(raw_ostream &OS) const {
1853 MemoryAccess *UO = getDefiningAccess();
1855 OS << getID() << " = MemoryDef(";
1856 if (UO && UO->getID())
1859 OS << LiveOnEntryStr;
1863 void MemoryPhi::print(raw_ostream &OS) const {
1865 OS << getID() << " = MemoryPhi(";
1866 for (const auto &Op : operands()) {
1867 BasicBlock *BB = getIncomingBlock(Op);
1868 MemoryAccess *MA = cast<MemoryAccess>(Op);
1876 OS << BB->getName();
1878 BB->printAsOperand(OS, false);
1880 if (unsigned ID = MA->getID())
1883 OS << LiveOnEntryStr;
1889 void MemoryUse::print(raw_ostream &OS) const {
1890 MemoryAccess *UO = getDefiningAccess();
1892 if (UO && UO->getID())
1895 OS << LiveOnEntryStr;
1899 void MemoryAccess::dump() const {
1900 // Cannot completely remove virtual function even in release mode.
1901 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1907 char MemorySSAPrinterLegacyPass::ID = 0;
1909 MemorySSAPrinterLegacyPass::MemorySSAPrinterLegacyPass() : FunctionPass(ID) {
1910 initializeMemorySSAPrinterLegacyPassPass(*PassRegistry::getPassRegistry());
1913 void MemorySSAPrinterLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
1914 AU.setPreservesAll();
1915 AU.addRequired<MemorySSAWrapperPass>();
1918 bool MemorySSAPrinterLegacyPass::runOnFunction(Function &F) {
1919 auto &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA();
1921 if (VerifyMemorySSA)
1922 MSSA.verifyMemorySSA();
1926 AnalysisKey MemorySSAAnalysis::Key;
1928 MemorySSAAnalysis::Result MemorySSAAnalysis::run(Function &F,
1929 FunctionAnalysisManager &AM) {
1930 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1931 auto &AA = AM.getResult<AAManager>(F);
1932 return MemorySSAAnalysis::Result(llvm::make_unique<MemorySSA>(F, &AA, &DT));
1935 PreservedAnalyses MemorySSAPrinterPass::run(Function &F,
1936 FunctionAnalysisManager &AM) {
1937 OS << "MemorySSA for function: " << F.getName() << "\n";
1938 AM.getResult<MemorySSAAnalysis>(F).getMSSA().print(OS);
1940 return PreservedAnalyses::all();
1943 PreservedAnalyses MemorySSAVerifierPass::run(Function &F,
1944 FunctionAnalysisManager &AM) {
1945 AM.getResult<MemorySSAAnalysis>(F).getMSSA().verifyMemorySSA();
1947 return PreservedAnalyses::all();
1950 char MemorySSAWrapperPass::ID = 0;
1952 MemorySSAWrapperPass::MemorySSAWrapperPass() : FunctionPass(ID) {
1953 initializeMemorySSAWrapperPassPass(*PassRegistry::getPassRegistry());
1956 void MemorySSAWrapperPass::releaseMemory() { MSSA.reset(); }
1958 void MemorySSAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1959 AU.setPreservesAll();
1960 AU.addRequiredTransitive<DominatorTreeWrapperPass>();
1961 AU.addRequiredTransitive<AAResultsWrapperPass>();
1964 bool MemorySSAWrapperPass::runOnFunction(Function &F) {
1965 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1966 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
1967 MSSA.reset(new MemorySSA(F, &AA, &DT));
1971 void MemorySSAWrapperPass::verifyAnalysis() const { MSSA->verifyMemorySSA(); }
1973 void MemorySSAWrapperPass::print(raw_ostream &OS, const Module *M) const {
1977 MemorySSAWalker::MemorySSAWalker(MemorySSA *M) : MSSA(M) {}
1979 MemorySSA::CachingWalker::CachingWalker(MemorySSA *M, AliasAnalysis *A,
1981 : MemorySSAWalker(M), Walker(*M, *A, *D) {}
1983 void MemorySSA::CachingWalker::invalidateInfo(MemoryAccess *MA) {
1984 if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
1985 MUD->resetOptimized();
1988 /// \brief Walk the use-def chains starting at \p MA and find
1989 /// the MemoryAccess that actually clobbers Loc.
1991 /// \returns our clobbering memory access
1992 MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess(
1993 MemoryAccess *StartingAccess, UpwardsMemoryQuery &Q) {
1994 MemoryAccess *New = Walker.findClobber(StartingAccess, Q);
1995 #ifdef EXPENSIVE_CHECKS
1996 MemoryAccess *NewNoCache = Walker.findClobber(StartingAccess, Q);
1997 assert(NewNoCache == New && "Cache made us hand back a different result?");
2000 if (AutoResetWalker)
2001 resetClobberWalker();
2005 MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess(
2006 MemoryAccess *StartingAccess, const MemoryLocation &Loc) {
2007 if (isa<MemoryPhi>(StartingAccess))
2008 return StartingAccess;
2010 auto *StartingUseOrDef = cast<MemoryUseOrDef>(StartingAccess);
2011 if (MSSA->isLiveOnEntryDef(StartingUseOrDef))
2012 return StartingUseOrDef;
2014 Instruction *I = StartingUseOrDef->getMemoryInst();
2016 // Conservatively, fences are always clobbers, so don't perform the walk if we
2018 if (!ImmutableCallSite(I) && I->isFenceLike())
2019 return StartingUseOrDef;
2021 UpwardsMemoryQuery Q;
2022 Q.OriginalAccess = StartingUseOrDef;
2023 Q.StartingLoc = Loc;
2027 // Unlike the other function, do not walk to the def of a def, because we are
2028 // handed something we already believe is the clobbering access.
2029 MemoryAccess *DefiningAccess = isa<MemoryUse>(StartingUseOrDef)
2030 ? StartingUseOrDef->getDefiningAccess()
2033 MemoryAccess *Clobber = getClobberingMemoryAccess(DefiningAccess, Q);
2034 DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
2035 DEBUG(dbgs() << *StartingUseOrDef << "\n");
2036 DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ");
2037 DEBUG(dbgs() << *Clobber << "\n");
2042 MemorySSA::CachingWalker::getClobberingMemoryAccess(MemoryAccess *MA) {
2043 auto *StartingAccess = dyn_cast<MemoryUseOrDef>(MA);
2044 // If this is a MemoryPhi, we can't do anything.
2045 if (!StartingAccess)
2048 // If this is an already optimized use or def, return the optimized result.
2049 // Note: Currently, we do not store the optimized def result because we'd need
2050 // a separate field, since we can't use it as the defining access.
2051 if (auto *MUD = dyn_cast<MemoryUseOrDef>(StartingAccess))
2052 if (MUD->isOptimized())
2053 return MUD->getOptimized();
2055 const Instruction *I = StartingAccess->getMemoryInst();
2056 UpwardsMemoryQuery Q(I, StartingAccess);
2057 // We can't sanely do anything with a fences, they conservatively
2058 // clobber all memory, and have no locations to get pointers from to
2059 // try to disambiguate.
2060 if (!Q.IsCall && I->isFenceLike())
2061 return StartingAccess;
2063 if (isUseTriviallyOptimizableToLiveOnEntry(*MSSA->AA, I)) {
2064 MemoryAccess *LiveOnEntry = MSSA->getLiveOnEntryDef();
2065 if (auto *MUD = dyn_cast<MemoryUseOrDef>(StartingAccess))
2066 MUD->setOptimized(LiveOnEntry);
2070 // Start with the thing we already think clobbers this location
2071 MemoryAccess *DefiningAccess = StartingAccess->getDefiningAccess();
2073 // At this point, DefiningAccess may be the live on entry def.
2074 // If it is, we will not get a better result.
2075 if (MSSA->isLiveOnEntryDef(DefiningAccess))
2076 return DefiningAccess;
2078 MemoryAccess *Result = getClobberingMemoryAccess(DefiningAccess, Q);
2079 DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
2080 DEBUG(dbgs() << *DefiningAccess << "\n");
2081 DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ");
2082 DEBUG(dbgs() << *Result << "\n");
2083 if (auto *MUD = dyn_cast<MemoryUseOrDef>(StartingAccess))
2084 MUD->setOptimized(Result);
2090 DoNothingMemorySSAWalker::getClobberingMemoryAccess(MemoryAccess *MA) {
2091 if (auto *Use = dyn_cast<MemoryUseOrDef>(MA))
2092 return Use->getDefiningAccess();
2096 MemoryAccess *DoNothingMemorySSAWalker::getClobberingMemoryAccess(
2097 MemoryAccess *StartingAccess, const MemoryLocation &) {
2098 if (auto *Use = dyn_cast<MemoryUseOrDef>(StartingAccess))
2099 return Use->getDefiningAccess();
2100 return StartingAccess;
2103 void MemoryPhi::deleteMe(DerivedUser *Self) {
2104 delete static_cast<MemoryPhi *>(Self);
2107 void MemoryDef::deleteMe(DerivedUser *Self) {
2108 delete static_cast<MemoryDef *>(Self);
2111 void MemoryUse::deleteMe(DerivedUser *Self) {
2112 delete static_cast<MemoryUse *>(Self);