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 Loc == Other.Loc;
159 if (CS.getCalledValue() != Other.CS.getCalledValue())
162 return CS.arg_size() == Other.CS.arg_size() &&
163 std::equal(CS.arg_begin(), CS.arg_end(), Other.CS.arg_begin());
168 ImmutableCallSite CS;
173 } // end anonymous namespace
177 template <> struct DenseMapInfo<MemoryLocOrCall> {
178 static inline MemoryLocOrCall getEmptyKey() {
179 return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getEmptyKey());
182 static inline MemoryLocOrCall getTombstoneKey() {
183 return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getTombstoneKey());
186 static unsigned getHashValue(const MemoryLocOrCall &MLOC) {
190 DenseMapInfo<MemoryLocation>::getHashValue(MLOC.getLoc()));
193 hash_combine(MLOC.IsCall, DenseMapInfo<const Value *>::getHashValue(
194 MLOC.getCS().getCalledValue()));
196 for (const Value *Arg : MLOC.getCS().args())
197 hash = hash_combine(hash, DenseMapInfo<const Value *>::getHashValue(Arg));
201 static bool isEqual(const MemoryLocOrCall &LHS, const MemoryLocOrCall &RHS) {
206 } // end namespace llvm
208 /// This does one-way checks to see if Use could theoretically be hoisted above
209 /// MayClobber. This will not check the other way around.
211 /// This assumes that, for the purposes of MemorySSA, Use comes directly after
212 /// MayClobber, with no potentially clobbering operations in between them.
213 /// (Where potentially clobbering ops are memory barriers, aliased stores, etc.)
214 static bool areLoadsReorderable(const LoadInst *Use,
215 const LoadInst *MayClobber) {
216 bool VolatileUse = Use->isVolatile();
217 bool VolatileClobber = MayClobber->isVolatile();
218 // Volatile operations may never be reordered with other volatile operations.
219 if (VolatileUse && VolatileClobber)
221 // Otherwise, volatile doesn't matter here. From the language reference:
222 // 'optimizers may change the order of volatile operations relative to
223 // non-volatile operations.'"
225 // If a load is seq_cst, it cannot be moved above other loads. If its ordering
226 // is weaker, it can be moved above other loads. We just need to be sure that
227 // MayClobber isn't an acquire load, because loads can't be moved above
230 // Note that this explicitly *does* allow the free reordering of monotonic (or
231 // weaker) loads of the same address.
232 bool SeqCstUse = Use->getOrdering() == AtomicOrdering::SequentiallyConsistent;
233 bool MayClobberIsAcquire = isAtLeastOrStrongerThan(MayClobber->getOrdering(),
234 AtomicOrdering::Acquire);
235 return !(SeqCstUse || MayClobberIsAcquire);
238 static bool instructionClobbersQuery(MemoryDef *MD,
239 const MemoryLocation &UseLoc,
240 const Instruction *UseInst,
242 Instruction *DefInst = MD->getMemoryInst();
243 assert(DefInst && "Defining instruction not actually an instruction");
244 ImmutableCallSite UseCS(UseInst);
246 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(DefInst)) {
247 // These intrinsics will show up as affecting memory, but they are just
249 switch (II->getIntrinsicID()) {
250 case Intrinsic::lifetime_start:
253 return AA.isMustAlias(MemoryLocation(II->getArgOperand(1)), UseLoc);
254 case Intrinsic::lifetime_end:
255 case Intrinsic::invariant_start:
256 case Intrinsic::invariant_end:
257 case Intrinsic::assume:
265 ModRefInfo I = AA.getModRefInfo(DefInst, UseCS);
266 return isModOrRefSet(I);
269 if (auto *DefLoad = dyn_cast<LoadInst>(DefInst))
270 if (auto *UseLoad = dyn_cast<LoadInst>(UseInst))
271 return !areLoadsReorderable(UseLoad, DefLoad);
273 return isModSet(AA.getModRefInfo(DefInst, UseLoc));
276 static bool instructionClobbersQuery(MemoryDef *MD, const MemoryUseOrDef *MU,
277 const MemoryLocOrCall &UseMLOC,
279 // FIXME: This is a temporary hack to allow a single instructionClobbersQuery
280 // to exist while MemoryLocOrCall is pushed through places.
282 return instructionClobbersQuery(MD, MemoryLocation(), MU->getMemoryInst(),
284 return instructionClobbersQuery(MD, UseMLOC.getLoc(), MU->getMemoryInst(),
288 // Return true when MD may alias MU, return false otherwise.
289 bool MemorySSAUtil::defClobbersUseOrDef(MemoryDef *MD, const MemoryUseOrDef *MU,
291 return instructionClobbersQuery(MD, MU, MemoryLocOrCall(MU), AA);
296 struct UpwardsMemoryQuery {
297 // True if our original query started off as a call
299 // The pointer location we started the query with. This will be empty if
301 MemoryLocation StartingLoc;
302 // This is the instruction we were querying about.
303 const Instruction *Inst = nullptr;
304 // The MemoryAccess we actually got called with, used to test local domination
305 const MemoryAccess *OriginalAccess = nullptr;
307 UpwardsMemoryQuery() = default;
309 UpwardsMemoryQuery(const Instruction *Inst, const MemoryAccess *Access)
310 : IsCall(ImmutableCallSite(Inst)), Inst(Inst), OriginalAccess(Access) {
312 StartingLoc = MemoryLocation::get(Inst);
316 } // end anonymous namespace
318 static bool lifetimeEndsAt(MemoryDef *MD, const MemoryLocation &Loc,
320 Instruction *Inst = MD->getMemoryInst();
321 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
322 switch (II->getIntrinsicID()) {
323 case Intrinsic::lifetime_end:
324 return AA.isMustAlias(MemoryLocation(II->getArgOperand(1)), Loc);
332 static bool isUseTriviallyOptimizableToLiveOnEntry(AliasAnalysis &AA,
333 const Instruction *I) {
334 // If the memory can't be changed, then loads of the memory can't be
337 // FIXME: We should handle invariant groups, as well. It's a bit harder,
338 // because we need to pay close attention to invariant group barriers.
339 return isa<LoadInst>(I) && (I->getMetadata(LLVMContext::MD_invariant_load) ||
340 AA.pointsToConstantMemory(cast<LoadInst>(I)->
341 getPointerOperand()));
344 /// Verifies that `Start` is clobbered by `ClobberAt`, and that nothing
345 /// inbetween `Start` and `ClobberAt` can clobbers `Start`.
347 /// This is meant to be as simple and self-contained as possible. Because it
348 /// uses no cache, etc., it can be relatively expensive.
350 /// \param Start The MemoryAccess that we want to walk from.
351 /// \param ClobberAt A clobber for Start.
352 /// \param StartLoc The MemoryLocation for Start.
353 /// \param MSSA The MemorySSA isntance that Start and ClobberAt belong to.
354 /// \param Query The UpwardsMemoryQuery we used for our search.
355 /// \param AA The AliasAnalysis we used for our search.
356 static void LLVM_ATTRIBUTE_UNUSED
357 checkClobberSanity(MemoryAccess *Start, MemoryAccess *ClobberAt,
358 const MemoryLocation &StartLoc, const MemorySSA &MSSA,
359 const UpwardsMemoryQuery &Query, AliasAnalysis &AA) {
360 assert(MSSA.dominates(ClobberAt, Start) && "Clobber doesn't dominate start?");
362 if (MSSA.isLiveOnEntryDef(Start)) {
363 assert(MSSA.isLiveOnEntryDef(ClobberAt) &&
364 "liveOnEntry must clobber itself");
368 bool FoundClobber = false;
369 DenseSet<MemoryAccessPair> VisitedPhis;
370 SmallVector<MemoryAccessPair, 8> Worklist;
371 Worklist.emplace_back(Start, StartLoc);
372 // Walk all paths from Start to ClobberAt, while looking for clobbers. If one
373 // is found, complain.
374 while (!Worklist.empty()) {
375 MemoryAccessPair MAP = Worklist.pop_back_val();
376 // All we care about is that nothing from Start to ClobberAt clobbers Start.
377 // We learn nothing from revisiting nodes.
378 if (!VisitedPhis.insert(MAP).second)
381 for (MemoryAccess *MA : def_chain(MAP.first)) {
382 if (MA == ClobberAt) {
383 if (auto *MD = dyn_cast<MemoryDef>(MA)) {
384 // instructionClobbersQuery isn't essentially free, so don't use `|=`,
385 // since it won't let us short-circuit.
387 // Also, note that this can't be hoisted out of the `Worklist` loop,
388 // since MD may only act as a clobber for 1 of N MemoryLocations.
390 FoundClobber || MSSA.isLiveOnEntryDef(MD) ||
391 instructionClobbersQuery(MD, MAP.second, Query.Inst, AA);
396 // We should never hit liveOnEntry, unless it's the clobber.
397 assert(!MSSA.isLiveOnEntryDef(MA) && "Hit liveOnEntry before clobber?");
399 if (auto *MD = dyn_cast<MemoryDef>(MA)) {
401 assert(!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA) &&
402 "Found clobber before reaching ClobberAt!");
406 assert(isa<MemoryPhi>(MA));
407 Worklist.append(upward_defs_begin({MA, MAP.second}), upward_defs_end());
411 // If ClobberAt is a MemoryPhi, we can assume something above it acted as a
412 // clobber. Otherwise, `ClobberAt` should've acted as a clobber at some point.
413 assert((isa<MemoryPhi>(ClobberAt) || FoundClobber) &&
414 "ClobberAt never acted as a clobber");
419 /// Our algorithm for walking (and trying to optimize) clobbers, all wrapped up
421 class ClobberWalker {
422 /// Save a few bytes by using unsigned instead of size_t.
423 using ListIndex = unsigned;
425 /// Represents a span of contiguous MemoryDefs, potentially ending in a
429 // Note that, because we always walk in reverse, Last will always dominate
430 // First. Also note that First and Last are inclusive.
433 Optional<ListIndex> Previous;
435 DefPath(const MemoryLocation &Loc, MemoryAccess *First, MemoryAccess *Last,
436 Optional<ListIndex> Previous)
437 : Loc(Loc), First(First), Last(Last), Previous(Previous) {}
439 DefPath(const MemoryLocation &Loc, MemoryAccess *Init,
440 Optional<ListIndex> Previous)
441 : DefPath(Loc, Init, Init, Previous) {}
444 const MemorySSA &MSSA;
447 UpwardsMemoryQuery *Query;
449 // Phi optimization bookkeeping
450 SmallVector<DefPath, 32> Paths;
451 DenseSet<ConstMemoryAccessPair> VisitedPhis;
453 /// Find the nearest def or phi that `From` can legally be optimized to.
454 const MemoryAccess *getWalkTarget(const MemoryPhi *From) const {
455 assert(From->getNumOperands() && "Phi with no operands?");
457 BasicBlock *BB = From->getBlock();
458 MemoryAccess *Result = MSSA.getLiveOnEntryDef();
459 DomTreeNode *Node = DT.getNode(BB);
460 while ((Node = Node->getIDom())) {
461 auto *Defs = MSSA.getBlockDefs(Node->getBlock());
463 return &*Defs->rbegin();
468 /// Result of calling walkToPhiOrClobber.
469 struct UpwardsWalkResult {
470 /// The "Result" of the walk. Either a clobber, the last thing we walked, or
472 MemoryAccess *Result;
476 /// Walk to the next Phi or Clobber in the def chain starting at Desc.Last.
477 /// This will update Desc.Last as it walks. It will (optionally) also stop at
480 /// This does not test for whether StopAt is a clobber
482 walkToPhiOrClobber(DefPath &Desc,
483 const MemoryAccess *StopAt = nullptr) const {
484 assert(!isa<MemoryUse>(Desc.Last) && "Uses don't exist in my world");
486 for (MemoryAccess *Current : def_chain(Desc.Last)) {
488 if (Current == StopAt)
489 return {Current, false};
491 if (auto *MD = dyn_cast<MemoryDef>(Current))
492 if (MSSA.isLiveOnEntryDef(MD) ||
493 instructionClobbersQuery(MD, Desc.Loc, Query->Inst, AA))
497 assert(isa<MemoryPhi>(Desc.Last) &&
498 "Ended at a non-clobber that's not a phi?");
499 return {Desc.Last, false};
502 void addSearches(MemoryPhi *Phi, SmallVectorImpl<ListIndex> &PausedSearches,
503 ListIndex PriorNode) {
504 auto UpwardDefs = make_range(upward_defs_begin({Phi, Paths[PriorNode].Loc}),
506 for (const MemoryAccessPair &P : UpwardDefs) {
507 PausedSearches.push_back(Paths.size());
508 Paths.emplace_back(P.second, P.first, PriorNode);
512 /// Represents a search that terminated after finding a clobber. This clobber
513 /// may or may not be present in the path of defs from LastNode..SearchStart,
514 /// since it may have been retrieved from cache.
515 struct TerminatedPath {
516 MemoryAccess *Clobber;
520 /// Get an access that keeps us from optimizing to the given phi.
522 /// PausedSearches is an array of indices into the Paths array. Its incoming
523 /// value is the indices of searches that stopped at the last phi optimization
524 /// target. It's left in an unspecified state.
526 /// If this returns None, NewPaused is a vector of searches that terminated
527 /// at StopWhere. Otherwise, NewPaused is left in an unspecified state.
528 Optional<TerminatedPath>
529 getBlockingAccess(const MemoryAccess *StopWhere,
530 SmallVectorImpl<ListIndex> &PausedSearches,
531 SmallVectorImpl<ListIndex> &NewPaused,
532 SmallVectorImpl<TerminatedPath> &Terminated) {
533 assert(!PausedSearches.empty() && "No searches to continue?");
535 // BFS vs DFS really doesn't make a difference here, so just do a DFS with
536 // PausedSearches as our stack.
537 while (!PausedSearches.empty()) {
538 ListIndex PathIndex = PausedSearches.pop_back_val();
539 DefPath &Node = Paths[PathIndex];
541 // If we've already visited this path with this MemoryLocation, we don't
542 // need to do so again.
544 // NOTE: That we just drop these paths on the ground makes caching
545 // behavior sporadic. e.g. given a diamond:
550 // ...If we walk D, B, A, C, we'll only cache the result of phi
551 // optimization for A, B, and D; C will be skipped because it dies here.
552 // This arguably isn't the worst thing ever, since:
553 // - We generally query things in a top-down order, so if we got below D
554 // without needing cache entries for {C, MemLoc}, then chances are
555 // that those cache entries would end up ultimately unused.
556 // - We still cache things for A, so C only needs to walk up a bit.
557 // If this behavior becomes problematic, we can fix without a ton of extra
559 if (!VisitedPhis.insert({Node.Last, Node.Loc}).second)
562 UpwardsWalkResult Res = walkToPhiOrClobber(Node, /*StopAt=*/StopWhere);
563 if (Res.IsKnownClobber) {
564 assert(Res.Result != StopWhere);
565 // If this wasn't a cache hit, we hit a clobber when walking. That's a
567 TerminatedPath Term{Res.Result, PathIndex};
568 if (!MSSA.dominates(Res.Result, StopWhere))
571 // Otherwise, it's a valid thing to potentially optimize to.
572 Terminated.push_back(Term);
576 if (Res.Result == StopWhere) {
577 // We've hit our target. Save this path off for if we want to continue
579 NewPaused.push_back(PathIndex);
583 assert(!MSSA.isLiveOnEntryDef(Res.Result) && "liveOnEntry is a clobber");
584 addSearches(cast<MemoryPhi>(Res.Result), PausedSearches, PathIndex);
590 template <typename T, typename Walker>
591 struct generic_def_path_iterator
592 : public iterator_facade_base<generic_def_path_iterator<T, Walker>,
593 std::forward_iterator_tag, T *> {
594 generic_def_path_iterator() = default;
595 generic_def_path_iterator(Walker *W, ListIndex N) : W(W), N(N) {}
597 T &operator*() const { return curNode(); }
599 generic_def_path_iterator &operator++() {
600 N = curNode().Previous;
604 bool operator==(const generic_def_path_iterator &O) const {
605 if (N.hasValue() != O.N.hasValue())
607 return !N.hasValue() || *N == *O.N;
611 T &curNode() const { return W->Paths[*N]; }
614 Optional<ListIndex> N = None;
617 using def_path_iterator = generic_def_path_iterator<DefPath, ClobberWalker>;
618 using const_def_path_iterator =
619 generic_def_path_iterator<const DefPath, const ClobberWalker>;
621 iterator_range<def_path_iterator> def_path(ListIndex From) {
622 return make_range(def_path_iterator(this, From), def_path_iterator());
625 iterator_range<const_def_path_iterator> const_def_path(ListIndex From) const {
626 return make_range(const_def_path_iterator(this, From),
627 const_def_path_iterator());
631 /// The path that contains our result.
632 TerminatedPath PrimaryClobber;
633 /// The paths that we can legally cache back from, but that aren't
634 /// necessarily the result of the Phi optimization.
635 SmallVector<TerminatedPath, 4> OtherClobbers;
638 ListIndex defPathIndex(const DefPath &N) const {
639 // The assert looks nicer if we don't need to do &N
640 const DefPath *NP = &N;
641 assert(!Paths.empty() && NP >= &Paths.front() && NP <= &Paths.back() &&
642 "Out of bounds DefPath!");
643 return NP - &Paths.front();
646 /// Try to optimize a phi as best as we can. Returns a SmallVector of Paths
647 /// that act as legal clobbers. Note that this won't return *all* clobbers.
649 /// Phi optimization algorithm tl;dr:
650 /// - Find the earliest def/phi, A, we can optimize to
651 /// - Find if all paths from the starting memory access ultimately reach A
652 /// - If not, optimization isn't possible.
653 /// - Otherwise, walk from A to another clobber or phi, A'.
654 /// - If A' is a def, we're done.
655 /// - If A' is a phi, try to optimize it.
657 /// A path is a series of {MemoryAccess, MemoryLocation} pairs. A path
658 /// terminates when a MemoryAccess that clobbers said MemoryLocation is found.
659 OptznResult tryOptimizePhi(MemoryPhi *Phi, MemoryAccess *Start,
660 const MemoryLocation &Loc) {
661 assert(Paths.empty() && VisitedPhis.empty() &&
662 "Reset the optimization state.");
664 Paths.emplace_back(Loc, Start, Phi, None);
665 // Stores how many "valid" optimization nodes we had prior to calling
666 // addSearches/getBlockingAccess. Necessary for caching if we had a blocker.
667 auto PriorPathsSize = Paths.size();
669 SmallVector<ListIndex, 16> PausedSearches;
670 SmallVector<ListIndex, 8> NewPaused;
671 SmallVector<TerminatedPath, 4> TerminatedPaths;
673 addSearches(Phi, PausedSearches, 0);
675 // Moves the TerminatedPath with the "most dominated" Clobber to the end of
677 auto MoveDominatedPathToEnd = [&](SmallVectorImpl<TerminatedPath> &Paths) {
678 assert(!Paths.empty() && "Need a path to move");
679 auto Dom = Paths.begin();
680 for (auto I = std::next(Dom), E = Paths.end(); I != E; ++I)
681 if (!MSSA.dominates(I->Clobber, Dom->Clobber))
683 auto Last = Paths.end() - 1;
685 std::iter_swap(Last, Dom);
688 MemoryPhi *Current = Phi;
690 assert(!MSSA.isLiveOnEntryDef(Current) &&
691 "liveOnEntry wasn't treated as a clobber?");
693 const auto *Target = getWalkTarget(Current);
694 // If a TerminatedPath doesn't dominate Target, then it wasn't a legal
695 // optimization for the prior phi.
696 assert(all_of(TerminatedPaths, [&](const TerminatedPath &P) {
697 return MSSA.dominates(P.Clobber, Target);
700 // FIXME: This is broken, because the Blocker may be reported to be
701 // liveOnEntry, and we'll happily wait for that to disappear (read: never)
702 // For the moment, this is fine, since we do nothing with blocker info.
703 if (Optional<TerminatedPath> Blocker = getBlockingAccess(
704 Target, PausedSearches, NewPaused, TerminatedPaths)) {
706 // Find the node we started at. We can't search based on N->Last, since
707 // we may have gone around a loop with a different MemoryLocation.
708 auto Iter = find_if(def_path(Blocker->LastNode), [&](const DefPath &N) {
709 return defPathIndex(N) < PriorPathsSize;
711 assert(Iter != def_path_iterator());
713 DefPath &CurNode = *Iter;
714 assert(CurNode.Last == Current);
717 // A. We can't reliably cache all of NewPaused back. Consider a case
718 // where we have two paths in NewPaused; one of which can't optimize
719 // above this phi, whereas the other can. If we cache the second path
720 // back, we'll end up with suboptimal cache entries. We can handle
721 // cases like this a bit better when we either try to find all
722 // clobbers that block phi optimization, or when our cache starts
723 // supporting unfinished searches.
724 // B. We can't reliably cache TerminatedPaths back here without doing
725 // extra checks; consider a case like:
731 // Where T is our target, C is a node with a clobber on it, D is a
732 // diamond (with a clobber *only* on the left or right node, N), and
733 // S is our start. Say we walk to D, through the node opposite N
734 // (read: ignoring the clobber), and see a cache entry in the top
735 // node of D. That cache entry gets put into TerminatedPaths. We then
736 // walk up to C (N is later in our worklist), find the clobber, and
737 // quit. If we append TerminatedPaths to OtherClobbers, we'll cache
738 // the bottom part of D to the cached clobber, ignoring the clobber
739 // in N. Again, this problem goes away if we start tracking all
740 // blockers for a given phi optimization.
741 TerminatedPath Result{CurNode.Last, defPathIndex(CurNode)};
745 // If there's nothing left to search, then all paths led to valid clobbers
746 // that we got from our cache; pick the nearest to the start, and allow
747 // the rest to be cached back.
748 if (NewPaused.empty()) {
749 MoveDominatedPathToEnd(TerminatedPaths);
750 TerminatedPath Result = TerminatedPaths.pop_back_val();
751 return {Result, std::move(TerminatedPaths)};
754 MemoryAccess *DefChainEnd = nullptr;
755 SmallVector<TerminatedPath, 4> Clobbers;
756 for (ListIndex Paused : NewPaused) {
757 UpwardsWalkResult WR = walkToPhiOrClobber(Paths[Paused]);
758 if (WR.IsKnownClobber)
759 Clobbers.push_back({WR.Result, Paused});
761 // Micro-opt: If we hit the end of the chain, save it.
762 DefChainEnd = WR.Result;
765 if (!TerminatedPaths.empty()) {
766 // If we couldn't find the dominating phi/liveOnEntry in the above loop,
769 for (auto *MA : def_chain(const_cast<MemoryAccess *>(Target)))
772 // If any of the terminated paths don't dominate the phi we'll try to
773 // optimize, we need to figure out what they are and quit.
774 const BasicBlock *ChainBB = DefChainEnd->getBlock();
775 for (const TerminatedPath &TP : TerminatedPaths) {
776 // Because we know that DefChainEnd is as "high" as we can go, we
777 // don't need local dominance checks; BB dominance is sufficient.
778 if (DT.dominates(ChainBB, TP.Clobber->getBlock()))
779 Clobbers.push_back(TP);
783 // If we have clobbers in the def chain, find the one closest to Current
785 if (!Clobbers.empty()) {
786 MoveDominatedPathToEnd(Clobbers);
787 TerminatedPath Result = Clobbers.pop_back_val();
788 return {Result, std::move(Clobbers)};
791 assert(all_of(NewPaused,
792 [&](ListIndex I) { return Paths[I].Last == DefChainEnd; }));
794 // Because liveOnEntry is a clobber, this must be a phi.
795 auto *DefChainPhi = cast<MemoryPhi>(DefChainEnd);
797 PriorPathsSize = Paths.size();
798 PausedSearches.clear();
799 for (ListIndex I : NewPaused)
800 addSearches(DefChainPhi, PausedSearches, I);
803 Current = DefChainPhi;
807 void verifyOptResult(const OptznResult &R) const {
808 assert(all_of(R.OtherClobbers, [&](const TerminatedPath &P) {
809 return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber);
813 void resetPhiOptznState() {
819 ClobberWalker(const MemorySSA &MSSA, AliasAnalysis &AA, DominatorTree &DT)
820 : MSSA(MSSA), AA(AA), DT(DT) {}
824 /// Finds the nearest clobber for the given query, optimizing phis if
826 MemoryAccess *findClobber(MemoryAccess *Start, UpwardsMemoryQuery &Q) {
829 MemoryAccess *Current = Start;
830 // This walker pretends uses don't exist. If we're handed one, silently grab
831 // its def. (This has the nice side-effect of ensuring we never cache uses)
832 if (auto *MU = dyn_cast<MemoryUse>(Start))
833 Current = MU->getDefiningAccess();
835 DefPath FirstDesc(Q.StartingLoc, Current, Current, None);
836 // Fast path for the overly-common case (no crazy phi optimization
838 UpwardsWalkResult WalkResult = walkToPhiOrClobber(FirstDesc);
839 MemoryAccess *Result;
840 if (WalkResult.IsKnownClobber) {
841 Result = WalkResult.Result;
843 OptznResult OptRes = tryOptimizePhi(cast<MemoryPhi>(FirstDesc.Last),
844 Current, Q.StartingLoc);
845 verifyOptResult(OptRes);
846 resetPhiOptznState();
847 Result = OptRes.PrimaryClobber.Clobber;
850 #ifdef EXPENSIVE_CHECKS
851 checkClobberSanity(Current, Result, Q.StartingLoc, MSSA, Q, AA);
856 void verify(const MemorySSA *MSSA) { assert(MSSA == &this->MSSA); }
859 struct RenamePassData {
861 DomTreeNode::const_iterator ChildIt;
862 MemoryAccess *IncomingVal;
864 RenamePassData(DomTreeNode *D, DomTreeNode::const_iterator It,
866 : DTN(D), ChildIt(It), IncomingVal(M) {}
868 void swap(RenamePassData &RHS) {
869 std::swap(DTN, RHS.DTN);
870 std::swap(ChildIt, RHS.ChildIt);
871 std::swap(IncomingVal, RHS.IncomingVal);
875 } // end anonymous namespace
879 /// \brief A MemorySSAWalker that does AA walks to disambiguate accesses. It no
880 /// longer does caching on its own,
881 /// but the name has been retained for the moment.
882 class MemorySSA::CachingWalker final : public MemorySSAWalker {
883 ClobberWalker Walker;
884 bool AutoResetWalker = true;
886 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *, UpwardsMemoryQuery &);
889 CachingWalker(MemorySSA *, AliasAnalysis *, DominatorTree *);
890 ~CachingWalker() override = default;
892 using MemorySSAWalker::getClobberingMemoryAccess;
894 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *) override;
895 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *,
896 const MemoryLocation &) override;
897 void invalidateInfo(MemoryAccess *) override;
899 /// Whether we call resetClobberWalker() after each time we *actually* walk to
900 /// answer a clobber query.
901 void setAutoResetWalker(bool AutoReset) { AutoResetWalker = AutoReset; }
903 /// Drop the walker's persistent data structures.
904 void resetClobberWalker() { Walker.reset(); }
906 void verify(const MemorySSA *MSSA) override {
907 MemorySSAWalker::verify(MSSA);
912 } // end namespace llvm
914 void MemorySSA::renameSuccessorPhis(BasicBlock *BB, MemoryAccess *IncomingVal,
915 bool RenameAllUses) {
916 // Pass through values to our successors
917 for (const BasicBlock *S : successors(BB)) {
918 auto It = PerBlockAccesses.find(S);
919 // Rename the phi nodes in our successor block
920 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
922 AccessList *Accesses = It->second.get();
923 auto *Phi = cast<MemoryPhi>(&Accesses->front());
925 int PhiIndex = Phi->getBasicBlockIndex(BB);
926 assert(PhiIndex != -1 && "Incomplete phi during partial rename");
927 Phi->setIncomingValue(PhiIndex, IncomingVal);
929 Phi->addIncoming(IncomingVal, BB);
933 /// \brief Rename a single basic block into MemorySSA form.
934 /// Uses the standard SSA renaming algorithm.
935 /// \returns The new incoming value.
936 MemoryAccess *MemorySSA::renameBlock(BasicBlock *BB, MemoryAccess *IncomingVal,
937 bool RenameAllUses) {
938 auto It = PerBlockAccesses.find(BB);
939 // Skip most processing if the list is empty.
940 if (It != PerBlockAccesses.end()) {
941 AccessList *Accesses = It->second.get();
942 for (MemoryAccess &L : *Accesses) {
943 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(&L)) {
944 if (MUD->getDefiningAccess() == nullptr || RenameAllUses)
945 MUD->setDefiningAccess(IncomingVal);
946 if (isa<MemoryDef>(&L))
956 /// \brief This is the standard SSA renaming algorithm.
958 /// We walk the dominator tree in preorder, renaming accesses, and then filling
959 /// in phi nodes in our successors.
960 void MemorySSA::renamePass(DomTreeNode *Root, MemoryAccess *IncomingVal,
961 SmallPtrSetImpl<BasicBlock *> &Visited,
962 bool SkipVisited, bool RenameAllUses) {
963 SmallVector<RenamePassData, 32> WorkStack;
964 // Skip everything if we already renamed this block and we are skipping.
965 // Note: You can't sink this into the if, because we need it to occur
966 // regardless of whether we skip blocks or not.
967 bool AlreadyVisited = !Visited.insert(Root->getBlock()).second;
968 if (SkipVisited && AlreadyVisited)
971 IncomingVal = renameBlock(Root->getBlock(), IncomingVal, RenameAllUses);
972 renameSuccessorPhis(Root->getBlock(), IncomingVal, RenameAllUses);
973 WorkStack.push_back({Root, Root->begin(), IncomingVal});
975 while (!WorkStack.empty()) {
976 DomTreeNode *Node = WorkStack.back().DTN;
977 DomTreeNode::const_iterator ChildIt = WorkStack.back().ChildIt;
978 IncomingVal = WorkStack.back().IncomingVal;
980 if (ChildIt == Node->end()) {
981 WorkStack.pop_back();
983 DomTreeNode *Child = *ChildIt;
984 ++WorkStack.back().ChildIt;
985 BasicBlock *BB = Child->getBlock();
986 // Note: You can't sink this into the if, because we need it to occur
987 // regardless of whether we skip blocks or not.
988 AlreadyVisited = !Visited.insert(BB).second;
989 if (SkipVisited && AlreadyVisited) {
990 // We already visited this during our renaming, which can happen when
991 // being asked to rename multiple blocks. Figure out the incoming val,
992 // which is the last def.
993 // Incoming value can only change if there is a block def, and in that
994 // case, it's the last block def in the list.
995 if (auto *BlockDefs = getWritableBlockDefs(BB))
996 IncomingVal = &*BlockDefs->rbegin();
998 IncomingVal = renameBlock(BB, IncomingVal, RenameAllUses);
999 renameSuccessorPhis(BB, IncomingVal, RenameAllUses);
1000 WorkStack.push_back({Child, Child->begin(), IncomingVal});
1005 /// \brief This handles unreachable block accesses by deleting phi nodes in
1006 /// unreachable blocks, and marking all other unreachable MemoryAccess's as
1007 /// being uses of the live on entry definition.
1008 void MemorySSA::markUnreachableAsLiveOnEntry(BasicBlock *BB) {
1009 assert(!DT->isReachableFromEntry(BB) &&
1010 "Reachable block found while handling unreachable blocks");
1012 // Make sure phi nodes in our reachable successors end up with a
1013 // LiveOnEntryDef for our incoming edge, even though our block is forward
1014 // unreachable. We could just disconnect these blocks from the CFG fully,
1015 // but we do not right now.
1016 for (const BasicBlock *S : successors(BB)) {
1017 if (!DT->isReachableFromEntry(S))
1019 auto It = PerBlockAccesses.find(S);
1020 // Rename the phi nodes in our successor block
1021 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
1023 AccessList *Accesses = It->second.get();
1024 auto *Phi = cast<MemoryPhi>(&Accesses->front());
1025 Phi->addIncoming(LiveOnEntryDef.get(), BB);
1028 auto It = PerBlockAccesses.find(BB);
1029 if (It == PerBlockAccesses.end())
1032 auto &Accesses = It->second;
1033 for (auto AI = Accesses->begin(), AE = Accesses->end(); AI != AE;) {
1034 auto Next = std::next(AI);
1035 // If we have a phi, just remove it. We are going to replace all
1036 // users with live on entry.
1037 if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(AI))
1038 UseOrDef->setDefiningAccess(LiveOnEntryDef.get());
1040 Accesses->erase(AI);
1045 MemorySSA::MemorySSA(Function &Func, AliasAnalysis *AA, DominatorTree *DT)
1046 : AA(AA), DT(DT), F(Func), LiveOnEntryDef(nullptr), Walker(nullptr),
1047 NextID(INVALID_MEMORYACCESS_ID) {
1051 MemorySSA::~MemorySSA() {
1052 // Drop all our references
1053 for (const auto &Pair : PerBlockAccesses)
1054 for (MemoryAccess &MA : *Pair.second)
1055 MA.dropAllReferences();
1058 MemorySSA::AccessList *MemorySSA::getOrCreateAccessList(const BasicBlock *BB) {
1059 auto Res = PerBlockAccesses.insert(std::make_pair(BB, nullptr));
1062 Res.first->second = llvm::make_unique<AccessList>();
1063 return Res.first->second.get();
1066 MemorySSA::DefsList *MemorySSA::getOrCreateDefsList(const BasicBlock *BB) {
1067 auto Res = PerBlockDefs.insert(std::make_pair(BB, nullptr));
1070 Res.first->second = llvm::make_unique<DefsList>();
1071 return Res.first->second.get();
1076 /// This class is a batch walker of all MemoryUse's in the program, and points
1077 /// their defining access at the thing that actually clobbers them. Because it
1078 /// is a batch walker that touches everything, it does not operate like the
1079 /// other walkers. This walker is basically performing a top-down SSA renaming
1080 /// pass, where the version stack is used as the cache. This enables it to be
1081 /// significantly more time and memory efficient than using the regular walker,
1082 /// which is walking bottom-up.
1083 class MemorySSA::OptimizeUses {
1085 OptimizeUses(MemorySSA *MSSA, MemorySSAWalker *Walker, AliasAnalysis *AA,
1087 : MSSA(MSSA), Walker(Walker), AA(AA), DT(DT) {
1088 Walker = MSSA->getWalker();
1091 void optimizeUses();
1094 /// This represents where a given memorylocation is in the stack.
1095 struct MemlocStackInfo {
1096 // This essentially is keeping track of versions of the stack. Whenever
1097 // the stack changes due to pushes or pops, these versions increase.
1098 unsigned long StackEpoch;
1099 unsigned long PopEpoch;
1100 // This is the lower bound of places on the stack to check. It is equal to
1101 // the place the last stack walk ended.
1102 // Note: Correctness depends on this being initialized to 0, which densemap
1104 unsigned long LowerBound;
1105 const BasicBlock *LowerBoundBlock;
1106 // This is where the last walk for this memory location ended.
1107 unsigned long LastKill;
1111 void optimizeUsesInBlock(const BasicBlock *, unsigned long &, unsigned long &,
1112 SmallVectorImpl<MemoryAccess *> &,
1113 DenseMap<MemoryLocOrCall, MemlocStackInfo> &);
1116 MemorySSAWalker *Walker;
1121 } // end namespace llvm
1123 /// Optimize the uses in a given block This is basically the SSA renaming
1124 /// algorithm, with one caveat: We are able to use a single stack for all
1125 /// MemoryUses. This is because the set of *possible* reaching MemoryDefs is
1126 /// the same for every MemoryUse. The *actual* clobbering MemoryDef is just
1127 /// going to be some position in that stack of possible ones.
1129 /// We track the stack positions that each MemoryLocation needs
1130 /// to check, and last ended at. This is because we only want to check the
1131 /// things that changed since last time. The same MemoryLocation should
1132 /// get clobbered by the same store (getModRefInfo does not use invariantness or
1133 /// things like this, and if they start, we can modify MemoryLocOrCall to
1134 /// include relevant data)
1135 void MemorySSA::OptimizeUses::optimizeUsesInBlock(
1136 const BasicBlock *BB, unsigned long &StackEpoch, unsigned long &PopEpoch,
1137 SmallVectorImpl<MemoryAccess *> &VersionStack,
1138 DenseMap<MemoryLocOrCall, MemlocStackInfo> &LocStackInfo) {
1140 /// If no accesses, nothing to do.
1141 MemorySSA::AccessList *Accesses = MSSA->getWritableBlockAccesses(BB);
1142 if (Accesses == nullptr)
1145 // Pop everything that doesn't dominate the current block off the stack,
1146 // increment the PopEpoch to account for this.
1149 !VersionStack.empty() &&
1150 "Version stack should have liveOnEntry sentinel dominating everything");
1151 BasicBlock *BackBlock = VersionStack.back()->getBlock();
1152 if (DT->dominates(BackBlock, BB))
1154 while (VersionStack.back()->getBlock() == BackBlock)
1155 VersionStack.pop_back();
1159 for (MemoryAccess &MA : *Accesses) {
1160 auto *MU = dyn_cast<MemoryUse>(&MA);
1162 VersionStack.push_back(&MA);
1167 if (isUseTriviallyOptimizableToLiveOnEntry(*AA, MU->getMemoryInst())) {
1168 MU->setDefiningAccess(MSSA->getLiveOnEntryDef(), true);
1172 MemoryLocOrCall UseMLOC(MU);
1173 auto &LocInfo = LocStackInfo[UseMLOC];
1174 // If the pop epoch changed, it means we've removed stuff from top of
1175 // stack due to changing blocks. We may have to reset the lower bound or
1177 if (LocInfo.PopEpoch != PopEpoch) {
1178 LocInfo.PopEpoch = PopEpoch;
1179 LocInfo.StackEpoch = StackEpoch;
1180 // If the lower bound was in something that no longer dominates us, we
1181 // have to reset it.
1182 // We can't simply track stack size, because the stack may have had
1183 // pushes/pops in the meantime.
1184 // XXX: This is non-optimal, but only is slower cases with heavily
1185 // branching dominator trees. To get the optimal number of queries would
1186 // be to make lowerbound and lastkill a per-loc stack, and pop it until
1187 // the top of that stack dominates us. This does not seem worth it ATM.
1188 // A much cheaper optimization would be to always explore the deepest
1189 // branch of the dominator tree first. This will guarantee this resets on
1190 // the smallest set of blocks.
1191 if (LocInfo.LowerBoundBlock && LocInfo.LowerBoundBlock != BB &&
1192 !DT->dominates(LocInfo.LowerBoundBlock, BB)) {
1193 // Reset the lower bound of things to check.
1194 // TODO: Some day we should be able to reset to last kill, rather than
1196 LocInfo.LowerBound = 0;
1197 LocInfo.LowerBoundBlock = VersionStack[0]->getBlock();
1198 LocInfo.LastKillValid = false;
1200 } else if (LocInfo.StackEpoch != StackEpoch) {
1201 // If all that has changed is the StackEpoch, we only have to check the
1202 // new things on the stack, because we've checked everything before. In
1203 // this case, the lower bound of things to check remains the same.
1204 LocInfo.PopEpoch = PopEpoch;
1205 LocInfo.StackEpoch = StackEpoch;
1207 if (!LocInfo.LastKillValid) {
1208 LocInfo.LastKill = VersionStack.size() - 1;
1209 LocInfo.LastKillValid = true;
1212 // At this point, we should have corrected last kill and LowerBound to be
1214 assert(LocInfo.LowerBound < VersionStack.size() &&
1215 "Lower bound out of range");
1216 assert(LocInfo.LastKill < VersionStack.size() &&
1217 "Last kill info out of range");
1218 // In any case, the new upper bound is the top of the stack.
1219 unsigned long UpperBound = VersionStack.size() - 1;
1221 if (UpperBound - LocInfo.LowerBound > MaxCheckLimit) {
1222 DEBUG(dbgs() << "MemorySSA skipping optimization of " << *MU << " ("
1223 << *(MU->getMemoryInst()) << ")"
1224 << " because there are " << UpperBound - LocInfo.LowerBound
1225 << " stores to disambiguate\n");
1226 // Because we did not walk, LastKill is no longer valid, as this may
1227 // have been a kill.
1228 LocInfo.LastKillValid = false;
1231 bool FoundClobberResult = false;
1232 while (UpperBound > LocInfo.LowerBound) {
1233 if (isa<MemoryPhi>(VersionStack[UpperBound])) {
1234 // For phis, use the walker, see where we ended up, go there
1235 Instruction *UseInst = MU->getMemoryInst();
1236 MemoryAccess *Result = Walker->getClobberingMemoryAccess(UseInst);
1237 // We are guaranteed to find it or something is wrong
1238 while (VersionStack[UpperBound] != Result) {
1239 assert(UpperBound != 0);
1242 FoundClobberResult = true;
1246 MemoryDef *MD = cast<MemoryDef>(VersionStack[UpperBound]);
1247 // If the lifetime of the pointer ends at this instruction, it's live on
1249 if (!UseMLOC.IsCall && lifetimeEndsAt(MD, UseMLOC.getLoc(), *AA)) {
1250 // Reset UpperBound to liveOnEntryDef's place in the stack
1252 FoundClobberResult = true;
1255 if (instructionClobbersQuery(MD, MU, UseMLOC, *AA)) {
1256 FoundClobberResult = true;
1261 // At the end of this loop, UpperBound is either a clobber, or lower bound
1262 // PHI walking may cause it to be < LowerBound, and in fact, < LastKill.
1263 if (FoundClobberResult || UpperBound < LocInfo.LastKill) {
1264 MU->setDefiningAccess(VersionStack[UpperBound], true);
1265 // We were last killed now by where we got to
1266 LocInfo.LastKill = UpperBound;
1268 // Otherwise, we checked all the new ones, and now we know we can get to
1270 MU->setDefiningAccess(VersionStack[LocInfo.LastKill], true);
1272 LocInfo.LowerBound = VersionStack.size() - 1;
1273 LocInfo.LowerBoundBlock = BB;
1277 /// Optimize uses to point to their actual clobbering definitions.
1278 void MemorySSA::OptimizeUses::optimizeUses() {
1279 SmallVector<MemoryAccess *, 16> VersionStack;
1280 DenseMap<MemoryLocOrCall, MemlocStackInfo> LocStackInfo;
1281 VersionStack.push_back(MSSA->getLiveOnEntryDef());
1283 unsigned long StackEpoch = 1;
1284 unsigned long PopEpoch = 1;
1285 // We perform a non-recursive top-down dominator tree walk.
1286 for (const auto *DomNode : depth_first(DT->getRootNode()))
1287 optimizeUsesInBlock(DomNode->getBlock(), StackEpoch, PopEpoch, VersionStack,
1291 void MemorySSA::placePHINodes(
1292 const SmallPtrSetImpl<BasicBlock *> &DefiningBlocks,
1293 const DenseMap<const BasicBlock *, unsigned int> &BBNumbers) {
1294 // Determine where our MemoryPhi's should go
1295 ForwardIDFCalculator IDFs(*DT);
1296 IDFs.setDefiningBlocks(DefiningBlocks);
1297 SmallVector<BasicBlock *, 32> IDFBlocks;
1298 IDFs.calculate(IDFBlocks);
1300 std::sort(IDFBlocks.begin(), IDFBlocks.end(),
1301 [&BBNumbers](const BasicBlock *A, const BasicBlock *B) {
1302 return BBNumbers.lookup(A) < BBNumbers.lookup(B);
1305 // Now place MemoryPhi nodes.
1306 for (auto &BB : IDFBlocks)
1307 createMemoryPhi(BB);
1310 void MemorySSA::buildMemorySSA() {
1311 // We create an access to represent "live on entry", for things like
1312 // arguments or users of globals, where the memory they use is defined before
1313 // the beginning of the function. We do not actually insert it into the IR.
1314 // We do not define a live on exit for the immediate uses, and thus our
1315 // semantics do *not* imply that something with no immediate uses can simply
1317 BasicBlock &StartingPoint = F.getEntryBlock();
1319 llvm::make_unique<MemoryDef>(F.getContext(), nullptr, nullptr,
1320 &StartingPoint, NextID++);
1321 DenseMap<const BasicBlock *, unsigned int> BBNumbers;
1322 unsigned NextBBNum = 0;
1324 // We maintain lists of memory accesses per-block, trading memory for time. We
1325 // could just look up the memory access for every possible instruction in the
1327 SmallPtrSet<BasicBlock *, 32> DefiningBlocks;
1328 // Go through each block, figure out where defs occur, and chain together all
1330 for (BasicBlock &B : F) {
1331 BBNumbers[&B] = NextBBNum++;
1332 bool InsertIntoDef = false;
1333 AccessList *Accesses = nullptr;
1334 DefsList *Defs = nullptr;
1335 for (Instruction &I : B) {
1336 MemoryUseOrDef *MUD = createNewAccess(&I);
1341 Accesses = getOrCreateAccessList(&B);
1342 Accesses->push_back(MUD);
1343 if (isa<MemoryDef>(MUD)) {
1344 InsertIntoDef = true;
1346 Defs = getOrCreateDefsList(&B);
1347 Defs->push_back(*MUD);
1351 DefiningBlocks.insert(&B);
1353 placePHINodes(DefiningBlocks, BBNumbers);
1355 // Now do regular SSA renaming on the MemoryDef/MemoryUse. Visited will get
1356 // filled in with all blocks.
1357 SmallPtrSet<BasicBlock *, 16> Visited;
1358 renamePass(DT->getRootNode(), LiveOnEntryDef.get(), Visited);
1360 CachingWalker *Walker = getWalkerImpl();
1362 // We're doing a batch of updates; don't drop useful caches between them.
1363 Walker->setAutoResetWalker(false);
1364 OptimizeUses(this, Walker, AA, DT).optimizeUses();
1365 Walker->setAutoResetWalker(true);
1366 Walker->resetClobberWalker();
1368 // Mark the uses in unreachable blocks as live on entry, so that they go
1371 if (!Visited.count(&BB))
1372 markUnreachableAsLiveOnEntry(&BB);
1375 MemorySSAWalker *MemorySSA::getWalker() { return getWalkerImpl(); }
1377 MemorySSA::CachingWalker *MemorySSA::getWalkerImpl() {
1379 return Walker.get();
1381 Walker = llvm::make_unique<CachingWalker>(this, AA, DT);
1382 return Walker.get();
1385 // This is a helper function used by the creation routines. It places NewAccess
1386 // into the access and defs lists for a given basic block, at the given
1388 void MemorySSA::insertIntoListsForBlock(MemoryAccess *NewAccess,
1389 const BasicBlock *BB,
1390 InsertionPlace Point) {
1391 auto *Accesses = getOrCreateAccessList(BB);
1392 if (Point == Beginning) {
1393 // If it's a phi node, it goes first, otherwise, it goes after any phi
1395 if (isa<MemoryPhi>(NewAccess)) {
1396 Accesses->push_front(NewAccess);
1397 auto *Defs = getOrCreateDefsList(BB);
1398 Defs->push_front(*NewAccess);
1400 auto AI = find_if_not(
1401 *Accesses, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
1402 Accesses->insert(AI, NewAccess);
1403 if (!isa<MemoryUse>(NewAccess)) {
1404 auto *Defs = getOrCreateDefsList(BB);
1405 auto DI = find_if_not(
1406 *Defs, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
1407 Defs->insert(DI, *NewAccess);
1411 Accesses->push_back(NewAccess);
1412 if (!isa<MemoryUse>(NewAccess)) {
1413 auto *Defs = getOrCreateDefsList(BB);
1414 Defs->push_back(*NewAccess);
1417 BlockNumberingValid.erase(BB);
1420 void MemorySSA::insertIntoListsBefore(MemoryAccess *What, const BasicBlock *BB,
1421 AccessList::iterator InsertPt) {
1422 auto *Accesses = getWritableBlockAccesses(BB);
1423 bool WasEnd = InsertPt == Accesses->end();
1424 Accesses->insert(AccessList::iterator(InsertPt), What);
1425 if (!isa<MemoryUse>(What)) {
1426 auto *Defs = getOrCreateDefsList(BB);
1427 // If we got asked to insert at the end, we have an easy job, just shove it
1428 // at the end. If we got asked to insert before an existing def, we also get
1429 // an terator. If we got asked to insert before a use, we have to hunt for
1432 Defs->push_back(*What);
1433 } else if (isa<MemoryDef>(InsertPt)) {
1434 Defs->insert(InsertPt->getDefsIterator(), *What);
1436 while (InsertPt != Accesses->end() && !isa<MemoryDef>(InsertPt))
1438 // Either we found a def, or we are inserting at the end
1439 if (InsertPt == Accesses->end())
1440 Defs->push_back(*What);
1442 Defs->insert(InsertPt->getDefsIterator(), *What);
1445 BlockNumberingValid.erase(BB);
1448 // Move What before Where in the IR. The end result is taht What will belong to
1449 // the right lists and have the right Block set, but will not otherwise be
1450 // correct. It will not have the right defining access, and if it is a def,
1451 // things below it will not properly be updated.
1452 void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB,
1453 AccessList::iterator Where) {
1454 // Keep it in the lookup tables, remove from the lists
1455 removeFromLists(What, false);
1457 insertIntoListsBefore(What, BB, Where);
1460 void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB,
1461 InsertionPlace Point) {
1462 removeFromLists(What, false);
1464 insertIntoListsForBlock(What, BB, Point);
1467 MemoryPhi *MemorySSA::createMemoryPhi(BasicBlock *BB) {
1468 assert(!getMemoryAccess(BB) && "MemoryPhi already exists for this BB");
1469 MemoryPhi *Phi = new MemoryPhi(BB->getContext(), BB, NextID++);
1470 // Phi's always are placed at the front of the block.
1471 insertIntoListsForBlock(Phi, BB, Beginning);
1472 ValueToMemoryAccess[BB] = Phi;
1476 MemoryUseOrDef *MemorySSA::createDefinedAccess(Instruction *I,
1477 MemoryAccess *Definition) {
1478 assert(!isa<PHINode>(I) && "Cannot create a defined access for a PHI");
1479 MemoryUseOrDef *NewAccess = createNewAccess(I);
1481 NewAccess != nullptr &&
1482 "Tried to create a memory access for a non-memory touching instruction");
1483 NewAccess->setDefiningAccess(Definition);
1487 // Return true if the instruction has ordering constraints.
1488 // Note specifically that this only considers stores and loads
1489 // because others are still considered ModRef by getModRefInfo.
1490 static inline bool isOrdered(const Instruction *I) {
1491 if (auto *SI = dyn_cast<StoreInst>(I)) {
1492 if (!SI->isUnordered())
1494 } else if (auto *LI = dyn_cast<LoadInst>(I)) {
1495 if (!LI->isUnordered())
1501 /// \brief Helper function to create new memory accesses
1502 MemoryUseOrDef *MemorySSA::createNewAccess(Instruction *I) {
1503 // The assume intrinsic has a control dependency which we model by claiming
1504 // that it writes arbitrarily. Ignore that fake memory dependency here.
1505 // FIXME: Replace this special casing with a more accurate modelling of
1506 // assume's control dependency.
1507 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1508 if (II->getIntrinsicID() == Intrinsic::assume)
1511 // Find out what affect this instruction has on memory.
1512 ModRefInfo ModRef = AA->getModRefInfo(I, None);
1513 // The isOrdered check is used to ensure that volatiles end up as defs
1514 // (atomics end up as ModRef right now anyway). Until we separate the
1515 // ordering chain from the memory chain, this enables people to see at least
1516 // some relative ordering to volatiles. Note that getClobberingMemoryAccess
1517 // will still give an answer that bypasses other volatile loads. TODO:
1518 // Separate memory aliasing and ordering into two different chains so that we
1519 // can precisely represent both "what memory will this read/write/is clobbered
1520 // by" and "what instructions can I move this past".
1521 bool Def = isModSet(ModRef) || isOrdered(I);
1522 bool Use = isRefSet(ModRef);
1524 // It's possible for an instruction to not modify memory at all. During
1525 // construction, we ignore them.
1529 assert((Def || Use) &&
1530 "Trying to create a memory access with a non-memory instruction");
1532 MemoryUseOrDef *MUD;
1534 MUD = new MemoryDef(I->getContext(), nullptr, I, I->getParent(), NextID++);
1536 MUD = new MemoryUse(I->getContext(), nullptr, I, I->getParent());
1537 ValueToMemoryAccess[I] = MUD;
1541 /// \brief Returns true if \p Replacer dominates \p Replacee .
1542 bool MemorySSA::dominatesUse(const MemoryAccess *Replacer,
1543 const MemoryAccess *Replacee) const {
1544 if (isa<MemoryUseOrDef>(Replacee))
1545 return DT->dominates(Replacer->getBlock(), Replacee->getBlock());
1546 const auto *MP = cast<MemoryPhi>(Replacee);
1547 // For a phi node, the use occurs in the predecessor block of the phi node.
1548 // Since we may occur multiple times in the phi node, we have to check each
1549 // operand to ensure Replacer dominates each operand where Replacee occurs.
1550 for (const Use &Arg : MP->operands()) {
1551 if (Arg.get() != Replacee &&
1552 !DT->dominates(Replacer->getBlock(), MP->getIncomingBlock(Arg)))
1558 /// \brief Properly remove \p MA from all of MemorySSA's lookup tables.
1559 void MemorySSA::removeFromLookups(MemoryAccess *MA) {
1560 assert(MA->use_empty() &&
1561 "Trying to remove memory access that still has uses");
1562 BlockNumbering.erase(MA);
1563 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA))
1564 MUD->setDefiningAccess(nullptr);
1565 // Invalidate our walker's cache if necessary
1566 if (!isa<MemoryUse>(MA))
1567 Walker->invalidateInfo(MA);
1568 // The call below to erase will destroy MA, so we can't change the order we
1569 // are doing things here
1571 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA)) {
1572 MemoryInst = MUD->getMemoryInst();
1574 MemoryInst = MA->getBlock();
1576 auto VMA = ValueToMemoryAccess.find(MemoryInst);
1577 if (VMA->second == MA)
1578 ValueToMemoryAccess.erase(VMA);
1581 /// \brief Properly remove \p MA from all of MemorySSA's lists.
1583 /// Because of the way the intrusive list and use lists work, it is important to
1584 /// do removal in the right order.
1585 /// ShouldDelete defaults to true, and will cause the memory access to also be
1586 /// deleted, not just removed.
1587 void MemorySSA::removeFromLists(MemoryAccess *MA, bool ShouldDelete) {
1588 // The access list owns the reference, so we erase it from the non-owning list
1590 if (!isa<MemoryUse>(MA)) {
1591 auto DefsIt = PerBlockDefs.find(MA->getBlock());
1592 std::unique_ptr<DefsList> &Defs = DefsIt->second;
1595 PerBlockDefs.erase(DefsIt);
1598 // The erase call here will delete it. If we don't want it deleted, we call
1600 auto AccessIt = PerBlockAccesses.find(MA->getBlock());
1601 std::unique_ptr<AccessList> &Accesses = AccessIt->second;
1603 Accesses->erase(MA);
1605 Accesses->remove(MA);
1607 if (Accesses->empty())
1608 PerBlockAccesses.erase(AccessIt);
1611 void MemorySSA::print(raw_ostream &OS) const {
1612 MemorySSAAnnotatedWriter Writer(this);
1613 F.print(OS, &Writer);
1616 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1617 LLVM_DUMP_METHOD void MemorySSA::dump() const { print(dbgs()); }
1620 void MemorySSA::verifyMemorySSA() const {
1622 verifyDomination(F);
1624 Walker->verify(this);
1627 /// \brief Verify that the order and existence of MemoryAccesses matches the
1628 /// order and existence of memory affecting instructions.
1629 void MemorySSA::verifyOrdering(Function &F) const {
1630 // Walk all the blocks, comparing what the lookups think and what the access
1631 // lists think, as well as the order in the blocks vs the order in the access
1633 SmallVector<MemoryAccess *, 32> ActualAccesses;
1634 SmallVector<MemoryAccess *, 32> ActualDefs;
1635 for (BasicBlock &B : F) {
1636 const AccessList *AL = getBlockAccesses(&B);
1637 const auto *DL = getBlockDefs(&B);
1638 MemoryAccess *Phi = getMemoryAccess(&B);
1640 ActualAccesses.push_back(Phi);
1641 ActualDefs.push_back(Phi);
1644 for (Instruction &I : B) {
1645 MemoryAccess *MA = getMemoryAccess(&I);
1646 assert((!MA || (AL && (isa<MemoryUse>(MA) || DL))) &&
1647 "We have memory affecting instructions "
1648 "in this block but they are not in the "
1649 "access list or defs list");
1651 ActualAccesses.push_back(MA);
1652 if (isa<MemoryDef>(MA))
1653 ActualDefs.push_back(MA);
1656 // Either we hit the assert, really have no accesses, or we have both
1657 // accesses and an access list.
1661 assert(AL->size() == ActualAccesses.size() &&
1662 "We don't have the same number of accesses in the block as on the "
1664 assert((DL || ActualDefs.size() == 0) &&
1665 "Either we should have a defs list, or we should have no defs");
1666 assert((!DL || DL->size() == ActualDefs.size()) &&
1667 "We don't have the same number of defs in the block as on the "
1669 auto ALI = AL->begin();
1670 auto AAI = ActualAccesses.begin();
1671 while (ALI != AL->end() && AAI != ActualAccesses.end()) {
1672 assert(&*ALI == *AAI && "Not the same accesses in the same order");
1676 ActualAccesses.clear();
1678 auto DLI = DL->begin();
1679 auto ADI = ActualDefs.begin();
1680 while (DLI != DL->end() && ADI != ActualDefs.end()) {
1681 assert(&*DLI == *ADI && "Not the same defs in the same order");
1690 /// \brief Verify the domination properties of MemorySSA by checking that each
1691 /// definition dominates all of its uses.
1692 void MemorySSA::verifyDomination(Function &F) const {
1694 for (BasicBlock &B : F) {
1695 // Phi nodes are attached to basic blocks
1696 if (MemoryPhi *MP = getMemoryAccess(&B))
1697 for (const Use &U : MP->uses())
1698 assert(dominates(MP, U) && "Memory PHI does not dominate it's uses");
1700 for (Instruction &I : B) {
1701 MemoryAccess *MD = dyn_cast_or_null<MemoryDef>(getMemoryAccess(&I));
1705 for (const Use &U : MD->uses())
1706 assert(dominates(MD, U) && "Memory Def does not dominate it's uses");
1712 /// \brief Verify the def-use lists in MemorySSA, by verifying that \p Use
1713 /// appears in the use list of \p Def.
1714 void MemorySSA::verifyUseInDefs(MemoryAccess *Def, MemoryAccess *Use) const {
1716 // The live on entry use may cause us to get a NULL def here
1718 assert(isLiveOnEntryDef(Use) &&
1719 "Null def but use not point to live on entry def");
1721 assert(is_contained(Def->users(), Use) &&
1722 "Did not find use in def's use list");
1726 /// \brief Verify the immediate use information, by walking all the memory
1727 /// accesses and verifying that, for each use, it appears in the
1728 /// appropriate def's use list
1729 void MemorySSA::verifyDefUses(Function &F) const {
1730 for (BasicBlock &B : F) {
1731 // Phi nodes are attached to basic blocks
1732 if (MemoryPhi *Phi = getMemoryAccess(&B)) {
1733 assert(Phi->getNumOperands() == static_cast<unsigned>(std::distance(
1734 pred_begin(&B), pred_end(&B))) &&
1735 "Incomplete MemoryPhi Node");
1736 for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I)
1737 verifyUseInDefs(Phi->getIncomingValue(I), Phi);
1740 for (Instruction &I : B) {
1741 if (MemoryUseOrDef *MA = getMemoryAccess(&I)) {
1742 verifyUseInDefs(MA->getDefiningAccess(), MA);
1748 MemoryUseOrDef *MemorySSA::getMemoryAccess(const Instruction *I) const {
1749 return cast_or_null<MemoryUseOrDef>(ValueToMemoryAccess.lookup(I));
1752 MemoryPhi *MemorySSA::getMemoryAccess(const BasicBlock *BB) const {
1753 return cast_or_null<MemoryPhi>(ValueToMemoryAccess.lookup(cast<Value>(BB)));
1756 /// Perform a local numbering on blocks so that instruction ordering can be
1757 /// determined in constant time.
1758 /// TODO: We currently just number in order. If we numbered by N, we could
1759 /// allow at least N-1 sequences of insertBefore or insertAfter (and at least
1760 /// log2(N) sequences of mixed before and after) without needing to invalidate
1762 void MemorySSA::renumberBlock(const BasicBlock *B) const {
1763 // The pre-increment ensures the numbers really start at 1.
1764 unsigned long CurrentNumber = 0;
1765 const AccessList *AL = getBlockAccesses(B);
1766 assert(AL != nullptr && "Asking to renumber an empty block");
1767 for (const auto &I : *AL)
1768 BlockNumbering[&I] = ++CurrentNumber;
1769 BlockNumberingValid.insert(B);
1772 /// \brief Determine, for two memory accesses in the same block,
1773 /// whether \p Dominator dominates \p Dominatee.
1774 /// \returns True if \p Dominator dominates \p Dominatee.
1775 bool MemorySSA::locallyDominates(const MemoryAccess *Dominator,
1776 const MemoryAccess *Dominatee) const {
1777 const BasicBlock *DominatorBlock = Dominator->getBlock();
1779 assert((DominatorBlock == Dominatee->getBlock()) &&
1780 "Asking for local domination when accesses are in different blocks!");
1781 // A node dominates itself.
1782 if (Dominatee == Dominator)
1785 // When Dominatee is defined on function entry, it is not dominated by another
1787 if (isLiveOnEntryDef(Dominatee))
1790 // When Dominator is defined on function entry, it dominates the other memory
1792 if (isLiveOnEntryDef(Dominator))
1795 if (!BlockNumberingValid.count(DominatorBlock))
1796 renumberBlock(DominatorBlock);
1798 unsigned long DominatorNum = BlockNumbering.lookup(Dominator);
1799 // All numbers start with 1
1800 assert(DominatorNum != 0 && "Block was not numbered properly");
1801 unsigned long DominateeNum = BlockNumbering.lookup(Dominatee);
1802 assert(DominateeNum != 0 && "Block was not numbered properly");
1803 return DominatorNum < DominateeNum;
1806 bool MemorySSA::dominates(const MemoryAccess *Dominator,
1807 const MemoryAccess *Dominatee) const {
1808 if (Dominator == Dominatee)
1811 if (isLiveOnEntryDef(Dominatee))
1814 if (Dominator->getBlock() != Dominatee->getBlock())
1815 return DT->dominates(Dominator->getBlock(), Dominatee->getBlock());
1816 return locallyDominates(Dominator, Dominatee);
1819 bool MemorySSA::dominates(const MemoryAccess *Dominator,
1820 const Use &Dominatee) const {
1821 if (MemoryPhi *MP = dyn_cast<MemoryPhi>(Dominatee.getUser())) {
1822 BasicBlock *UseBB = MP->getIncomingBlock(Dominatee);
1823 // The def must dominate the incoming block of the phi.
1824 if (UseBB != Dominator->getBlock())
1825 return DT->dominates(Dominator->getBlock(), UseBB);
1826 // If the UseBB and the DefBB are the same, compare locally.
1827 return locallyDominates(Dominator, cast<MemoryAccess>(Dominatee));
1829 // If it's not a PHI node use, the normal dominates can already handle it.
1830 return dominates(Dominator, cast<MemoryAccess>(Dominatee.getUser()));
1833 const static char LiveOnEntryStr[] = "liveOnEntry";
1835 void MemoryAccess::print(raw_ostream &OS) const {
1836 switch (getValueID()) {
1837 case MemoryPhiVal: return static_cast<const MemoryPhi *>(this)->print(OS);
1838 case MemoryDefVal: return static_cast<const MemoryDef *>(this)->print(OS);
1839 case MemoryUseVal: return static_cast<const MemoryUse *>(this)->print(OS);
1841 llvm_unreachable("invalid value id");
1844 void MemoryDef::print(raw_ostream &OS) const {
1845 MemoryAccess *UO = getDefiningAccess();
1847 OS << getID() << " = MemoryDef(";
1848 if (UO && UO->getID())
1851 OS << LiveOnEntryStr;
1855 void MemoryPhi::print(raw_ostream &OS) const {
1857 OS << getID() << " = MemoryPhi(";
1858 for (const auto &Op : operands()) {
1859 BasicBlock *BB = getIncomingBlock(Op);
1860 MemoryAccess *MA = cast<MemoryAccess>(Op);
1868 OS << BB->getName();
1870 BB->printAsOperand(OS, false);
1872 if (unsigned ID = MA->getID())
1875 OS << LiveOnEntryStr;
1881 void MemoryUse::print(raw_ostream &OS) const {
1882 MemoryAccess *UO = getDefiningAccess();
1884 if (UO && UO->getID())
1887 OS << LiveOnEntryStr;
1891 void MemoryAccess::dump() const {
1892 // Cannot completely remove virtual function even in release mode.
1893 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1899 char MemorySSAPrinterLegacyPass::ID = 0;
1901 MemorySSAPrinterLegacyPass::MemorySSAPrinterLegacyPass() : FunctionPass(ID) {
1902 initializeMemorySSAPrinterLegacyPassPass(*PassRegistry::getPassRegistry());
1905 void MemorySSAPrinterLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
1906 AU.setPreservesAll();
1907 AU.addRequired<MemorySSAWrapperPass>();
1910 bool MemorySSAPrinterLegacyPass::runOnFunction(Function &F) {
1911 auto &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA();
1913 if (VerifyMemorySSA)
1914 MSSA.verifyMemorySSA();
1918 AnalysisKey MemorySSAAnalysis::Key;
1920 MemorySSAAnalysis::Result MemorySSAAnalysis::run(Function &F,
1921 FunctionAnalysisManager &AM) {
1922 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1923 auto &AA = AM.getResult<AAManager>(F);
1924 return MemorySSAAnalysis::Result(llvm::make_unique<MemorySSA>(F, &AA, &DT));
1927 PreservedAnalyses MemorySSAPrinterPass::run(Function &F,
1928 FunctionAnalysisManager &AM) {
1929 OS << "MemorySSA for function: " << F.getName() << "\n";
1930 AM.getResult<MemorySSAAnalysis>(F).getMSSA().print(OS);
1932 return PreservedAnalyses::all();
1935 PreservedAnalyses MemorySSAVerifierPass::run(Function &F,
1936 FunctionAnalysisManager &AM) {
1937 AM.getResult<MemorySSAAnalysis>(F).getMSSA().verifyMemorySSA();
1939 return PreservedAnalyses::all();
1942 char MemorySSAWrapperPass::ID = 0;
1944 MemorySSAWrapperPass::MemorySSAWrapperPass() : FunctionPass(ID) {
1945 initializeMemorySSAWrapperPassPass(*PassRegistry::getPassRegistry());
1948 void MemorySSAWrapperPass::releaseMemory() { MSSA.reset(); }
1950 void MemorySSAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1951 AU.setPreservesAll();
1952 AU.addRequiredTransitive<DominatorTreeWrapperPass>();
1953 AU.addRequiredTransitive<AAResultsWrapperPass>();
1956 bool MemorySSAWrapperPass::runOnFunction(Function &F) {
1957 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1958 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
1959 MSSA.reset(new MemorySSA(F, &AA, &DT));
1963 void MemorySSAWrapperPass::verifyAnalysis() const { MSSA->verifyMemorySSA(); }
1965 void MemorySSAWrapperPass::print(raw_ostream &OS, const Module *M) const {
1969 MemorySSAWalker::MemorySSAWalker(MemorySSA *M) : MSSA(M) {}
1971 MemorySSA::CachingWalker::CachingWalker(MemorySSA *M, AliasAnalysis *A,
1973 : MemorySSAWalker(M), Walker(*M, *A, *D) {}
1975 void MemorySSA::CachingWalker::invalidateInfo(MemoryAccess *MA) {
1976 if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
1977 MUD->resetOptimized();
1980 /// \brief Walk the use-def chains starting at \p MA and find
1981 /// the MemoryAccess that actually clobbers Loc.
1983 /// \returns our clobbering memory access
1984 MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess(
1985 MemoryAccess *StartingAccess, UpwardsMemoryQuery &Q) {
1986 MemoryAccess *New = Walker.findClobber(StartingAccess, Q);
1987 #ifdef EXPENSIVE_CHECKS
1988 MemoryAccess *NewNoCache = Walker.findClobber(StartingAccess, Q);
1989 assert(NewNoCache == New && "Cache made us hand back a different result?");
1992 if (AutoResetWalker)
1993 resetClobberWalker();
1997 MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess(
1998 MemoryAccess *StartingAccess, const MemoryLocation &Loc) {
1999 if (isa<MemoryPhi>(StartingAccess))
2000 return StartingAccess;
2002 auto *StartingUseOrDef = cast<MemoryUseOrDef>(StartingAccess);
2003 if (MSSA->isLiveOnEntryDef(StartingUseOrDef))
2004 return StartingUseOrDef;
2006 Instruction *I = StartingUseOrDef->getMemoryInst();
2008 // Conservatively, fences are always clobbers, so don't perform the walk if we
2010 if (!ImmutableCallSite(I) && I->isFenceLike())
2011 return StartingUseOrDef;
2013 UpwardsMemoryQuery Q;
2014 Q.OriginalAccess = StartingUseOrDef;
2015 Q.StartingLoc = Loc;
2019 // Unlike the other function, do not walk to the def of a def, because we are
2020 // handed something we already believe is the clobbering access.
2021 MemoryAccess *DefiningAccess = isa<MemoryUse>(StartingUseOrDef)
2022 ? StartingUseOrDef->getDefiningAccess()
2025 MemoryAccess *Clobber = getClobberingMemoryAccess(DefiningAccess, Q);
2026 DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
2027 DEBUG(dbgs() << *StartingUseOrDef << "\n");
2028 DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ");
2029 DEBUG(dbgs() << *Clobber << "\n");
2034 MemorySSA::CachingWalker::getClobberingMemoryAccess(MemoryAccess *MA) {
2035 auto *StartingAccess = dyn_cast<MemoryUseOrDef>(MA);
2036 // If this is a MemoryPhi, we can't do anything.
2037 if (!StartingAccess)
2040 // If this is an already optimized use or def, return the optimized result.
2041 // Note: Currently, we do not store the optimized def result because we'd need
2042 // a separate field, since we can't use it as the defining access.
2043 if (auto *MUD = dyn_cast<MemoryUseOrDef>(StartingAccess))
2044 if (MUD->isOptimized())
2045 return MUD->getOptimized();
2047 const Instruction *I = StartingAccess->getMemoryInst();
2048 UpwardsMemoryQuery Q(I, StartingAccess);
2049 // We can't sanely do anything with a fences, they conservatively
2050 // clobber all memory, and have no locations to get pointers from to
2051 // try to disambiguate.
2052 if (!Q.IsCall && I->isFenceLike())
2053 return StartingAccess;
2055 if (isUseTriviallyOptimizableToLiveOnEntry(*MSSA->AA, I)) {
2056 MemoryAccess *LiveOnEntry = MSSA->getLiveOnEntryDef();
2057 if (auto *MUD = dyn_cast<MemoryUseOrDef>(StartingAccess))
2058 MUD->setOptimized(LiveOnEntry);
2062 // Start with the thing we already think clobbers this location
2063 MemoryAccess *DefiningAccess = StartingAccess->getDefiningAccess();
2065 // At this point, DefiningAccess may be the live on entry def.
2066 // If it is, we will not get a better result.
2067 if (MSSA->isLiveOnEntryDef(DefiningAccess))
2068 return DefiningAccess;
2070 MemoryAccess *Result = getClobberingMemoryAccess(DefiningAccess, Q);
2071 DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
2072 DEBUG(dbgs() << *DefiningAccess << "\n");
2073 DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ");
2074 DEBUG(dbgs() << *Result << "\n");
2075 if (auto *MUD = dyn_cast<MemoryUseOrDef>(StartingAccess))
2076 MUD->setOptimized(Result);
2082 DoNothingMemorySSAWalker::getClobberingMemoryAccess(MemoryAccess *MA) {
2083 if (auto *Use = dyn_cast<MemoryUseOrDef>(MA))
2084 return Use->getDefiningAccess();
2088 MemoryAccess *DoNothingMemorySSAWalker::getClobberingMemoryAccess(
2089 MemoryAccess *StartingAccess, const MemoryLocation &) {
2090 if (auto *Use = dyn_cast<MemoryUseOrDef>(StartingAccess))
2091 return Use->getDefiningAccess();
2092 return StartingAccess;
2095 void MemoryPhi::deleteMe(DerivedUser *Self) {
2096 delete static_cast<MemoryPhi *>(Self);
2099 void MemoryDef::deleteMe(DerivedUser *Self) {
2100 delete static_cast<MemoryDef *>(Self);
2103 void MemoryUse::deleteMe(DerivedUser *Self) {
2104 delete static_cast<MemoryUse *>(Self);