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/Config/llvm-config.h"
31 #include "llvm/IR/AssemblyAnnotationWriter.h"
32 #include "llvm/IR/BasicBlock.h"
33 #include "llvm/IR/CallSite.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/Instruction.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/IntrinsicInst.h"
39 #include "llvm/IR/Intrinsics.h"
40 #include "llvm/IR/LLVMContext.h"
41 #include "llvm/IR/PassManager.h"
42 #include "llvm/IR/Use.h"
43 #include "llvm/Pass.h"
44 #include "llvm/Support/AtomicOrdering.h"
45 #include "llvm/Support/Casting.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Compiler.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/FormattedStream.h"
51 #include "llvm/Support/raw_ostream.h"
60 #define DEBUG_TYPE "memoryssa"
62 INITIALIZE_PASS_BEGIN(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,
64 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
65 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
66 INITIALIZE_PASS_END(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,
69 INITIALIZE_PASS_BEGIN(MemorySSAPrinterLegacyPass, "print-memoryssa",
70 "Memory SSA Printer", false, false)
71 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
72 INITIALIZE_PASS_END(MemorySSAPrinterLegacyPass, "print-memoryssa",
73 "Memory SSA Printer", false, false)
75 static cl::opt<unsigned> MaxCheckLimit(
76 "memssa-check-limit", cl::Hidden, cl::init(100),
77 cl::desc("The maximum number of stores/phis MemorySSA"
78 "will consider trying to walk past (default = 100)"));
81 VerifyMemorySSA("verify-memoryssa", cl::init(false), cl::Hidden,
82 cl::desc("Verify MemorySSA in legacy printer pass."));
86 /// An assembly annotator class to print Memory SSA information in
88 class MemorySSAAnnotatedWriter : public AssemblyAnnotationWriter {
89 friend class MemorySSA;
91 const MemorySSA *MSSA;
94 MemorySSAAnnotatedWriter(const MemorySSA *M) : MSSA(M) {}
96 void emitBasicBlockStartAnnot(const BasicBlock *BB,
97 formatted_raw_ostream &OS) override {
98 if (MemoryAccess *MA = MSSA->getMemoryAccess(BB))
99 OS << "; " << *MA << "\n";
102 void emitInstructionAnnot(const Instruction *I,
103 formatted_raw_ostream &OS) override {
104 if (MemoryAccess *MA = MSSA->getMemoryAccess(I))
105 OS << "; " << *MA << "\n";
109 } // end namespace llvm
113 /// Our current alias analysis API differentiates heavily between calls and
114 /// non-calls, and functions called on one usually assert on the other.
115 /// This class encapsulates the distinction to simplify other code that wants
116 /// "Memory affecting instructions and related data" to use as a key.
117 /// For example, this class is used as a densemap key in the use optimizer.
118 class MemoryLocOrCall {
122 MemoryLocOrCall() = default;
123 MemoryLocOrCall(MemoryUseOrDef *MUD)
124 : MemoryLocOrCall(MUD->getMemoryInst()) {}
125 MemoryLocOrCall(const MemoryUseOrDef *MUD)
126 : MemoryLocOrCall(MUD->getMemoryInst()) {}
128 MemoryLocOrCall(Instruction *Inst) {
129 if (ImmutableCallSite(Inst)) {
131 CS = ImmutableCallSite(Inst);
134 // There is no such thing as a memorylocation for a fence inst, and it is
135 // unique in that regard.
136 if (!isa<FenceInst>(Inst))
137 Loc = MemoryLocation::get(Inst);
141 explicit MemoryLocOrCall(const MemoryLocation &Loc) : Loc(Loc) {}
143 ImmutableCallSite getCS() const {
148 MemoryLocation getLoc() const {
153 bool operator==(const MemoryLocOrCall &Other) const {
154 if (IsCall != Other.IsCall)
158 return Loc == Other.Loc;
160 if (CS.getCalledValue() != Other.CS.getCalledValue())
163 return CS.arg_size() == Other.CS.arg_size() &&
164 std::equal(CS.arg_begin(), CS.arg_end(), Other.CS.arg_begin());
169 ImmutableCallSite CS;
174 } // end anonymous namespace
178 template <> struct DenseMapInfo<MemoryLocOrCall> {
179 static inline MemoryLocOrCall getEmptyKey() {
180 return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getEmptyKey());
183 static inline MemoryLocOrCall getTombstoneKey() {
184 return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getTombstoneKey());
187 static unsigned getHashValue(const MemoryLocOrCall &MLOC) {
191 DenseMapInfo<MemoryLocation>::getHashValue(MLOC.getLoc()));
194 hash_combine(MLOC.IsCall, DenseMapInfo<const Value *>::getHashValue(
195 MLOC.getCS().getCalledValue()));
197 for (const Value *Arg : MLOC.getCS().args())
198 hash = hash_combine(hash, DenseMapInfo<const Value *>::getHashValue(Arg));
202 static bool isEqual(const MemoryLocOrCall &LHS, const MemoryLocOrCall &RHS) {
207 } // end namespace llvm
209 /// This does one-way checks to see if Use could theoretically be hoisted above
210 /// MayClobber. This will not check the other way around.
212 /// This assumes that, for the purposes of MemorySSA, Use comes directly after
213 /// MayClobber, with no potentially clobbering operations in between them.
214 /// (Where potentially clobbering ops are memory barriers, aliased stores, etc.)
215 static bool areLoadsReorderable(const LoadInst *Use,
216 const LoadInst *MayClobber) {
217 bool VolatileUse = Use->isVolatile();
218 bool VolatileClobber = MayClobber->isVolatile();
219 // Volatile operations may never be reordered with other volatile operations.
220 if (VolatileUse && VolatileClobber)
222 // Otherwise, volatile doesn't matter here. From the language reference:
223 // 'optimizers may change the order of volatile operations relative to
224 // non-volatile operations.'"
226 // If a load is seq_cst, it cannot be moved above other loads. If its ordering
227 // is weaker, it can be moved above other loads. We just need to be sure that
228 // MayClobber isn't an acquire load, because loads can't be moved above
231 // Note that this explicitly *does* allow the free reordering of monotonic (or
232 // weaker) loads of the same address.
233 bool SeqCstUse = Use->getOrdering() == AtomicOrdering::SequentiallyConsistent;
234 bool MayClobberIsAcquire = isAtLeastOrStrongerThan(MayClobber->getOrdering(),
235 AtomicOrdering::Acquire);
236 return !(SeqCstUse || MayClobberIsAcquire);
241 struct ClobberAlias {
243 Optional<AliasResult> AR;
246 } // end anonymous namespace
248 // Return a pair of {IsClobber (bool), AR (AliasResult)}. It relies on AR being
249 // ignored if IsClobber = false.
250 static ClobberAlias instructionClobbersQuery(MemoryDef *MD,
251 const MemoryLocation &UseLoc,
252 const Instruction *UseInst,
254 Instruction *DefInst = MD->getMemoryInst();
255 assert(DefInst && "Defining instruction not actually an instruction");
256 ImmutableCallSite UseCS(UseInst);
257 Optional<AliasResult> AR;
259 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(DefInst)) {
260 // These intrinsics will show up as affecting memory, but they are just
262 switch (II->getIntrinsicID()) {
263 case Intrinsic::lifetime_start:
265 return {false, NoAlias};
266 AR = AA.alias(MemoryLocation(II->getArgOperand(1)), UseLoc);
267 return {AR == MustAlias, AR};
268 case Intrinsic::lifetime_end:
269 case Intrinsic::invariant_start:
270 case Intrinsic::invariant_end:
271 case Intrinsic::assume:
272 return {false, NoAlias};
279 ModRefInfo I = AA.getModRefInfo(DefInst, UseCS);
280 AR = isMustSet(I) ? MustAlias : MayAlias;
281 return {isModOrRefSet(I), AR};
284 if (auto *DefLoad = dyn_cast<LoadInst>(DefInst))
285 if (auto *UseLoad = dyn_cast<LoadInst>(UseInst))
286 return {!areLoadsReorderable(UseLoad, DefLoad), MayAlias};
288 ModRefInfo I = AA.getModRefInfo(DefInst, UseLoc);
289 AR = isMustSet(I) ? MustAlias : MayAlias;
290 return {isModSet(I), AR};
293 static ClobberAlias instructionClobbersQuery(MemoryDef *MD,
294 const MemoryUseOrDef *MU,
295 const MemoryLocOrCall &UseMLOC,
297 // FIXME: This is a temporary hack to allow a single instructionClobbersQuery
298 // to exist while MemoryLocOrCall is pushed through places.
300 return instructionClobbersQuery(MD, MemoryLocation(), MU->getMemoryInst(),
302 return instructionClobbersQuery(MD, UseMLOC.getLoc(), MU->getMemoryInst(),
306 // Return true when MD may alias MU, return false otherwise.
307 bool MemorySSAUtil::defClobbersUseOrDef(MemoryDef *MD, const MemoryUseOrDef *MU,
309 return instructionClobbersQuery(MD, MU, MemoryLocOrCall(MU), AA).IsClobber;
314 struct UpwardsMemoryQuery {
315 // True if our original query started off as a call
317 // The pointer location we started the query with. This will be empty if
319 MemoryLocation StartingLoc;
320 // This is the instruction we were querying about.
321 const Instruction *Inst = nullptr;
322 // The MemoryAccess we actually got called with, used to test local domination
323 const MemoryAccess *OriginalAccess = nullptr;
324 Optional<AliasResult> AR = MayAlias;
326 UpwardsMemoryQuery() = default;
328 UpwardsMemoryQuery(const Instruction *Inst, const MemoryAccess *Access)
329 : IsCall(ImmutableCallSite(Inst)), Inst(Inst), OriginalAccess(Access) {
331 StartingLoc = MemoryLocation::get(Inst);
335 } // end anonymous namespace
337 static bool lifetimeEndsAt(MemoryDef *MD, const MemoryLocation &Loc,
339 Instruction *Inst = MD->getMemoryInst();
340 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
341 switch (II->getIntrinsicID()) {
342 case Intrinsic::lifetime_end:
343 return AA.isMustAlias(MemoryLocation(II->getArgOperand(1)), Loc);
351 static bool isUseTriviallyOptimizableToLiveOnEntry(AliasAnalysis &AA,
352 const Instruction *I) {
353 // If the memory can't be changed, then loads of the memory can't be
355 return isa<LoadInst>(I) && (I->getMetadata(LLVMContext::MD_invariant_load) ||
356 AA.pointsToConstantMemory(cast<LoadInst>(I)->
357 getPointerOperand()));
360 /// Verifies that `Start` is clobbered by `ClobberAt`, and that nothing
361 /// inbetween `Start` and `ClobberAt` can clobbers `Start`.
363 /// This is meant to be as simple and self-contained as possible. Because it
364 /// uses no cache, etc., it can be relatively expensive.
366 /// \param Start The MemoryAccess that we want to walk from.
367 /// \param ClobberAt A clobber for Start.
368 /// \param StartLoc The MemoryLocation for Start.
369 /// \param MSSA The MemorySSA isntance that Start and ClobberAt belong to.
370 /// \param Query The UpwardsMemoryQuery we used for our search.
371 /// \param AA The AliasAnalysis we used for our search.
372 static void LLVM_ATTRIBUTE_UNUSED
373 checkClobberSanity(MemoryAccess *Start, MemoryAccess *ClobberAt,
374 const MemoryLocation &StartLoc, const MemorySSA &MSSA,
375 const UpwardsMemoryQuery &Query, AliasAnalysis &AA) {
376 assert(MSSA.dominates(ClobberAt, Start) && "Clobber doesn't dominate start?");
378 if (MSSA.isLiveOnEntryDef(Start)) {
379 assert(MSSA.isLiveOnEntryDef(ClobberAt) &&
380 "liveOnEntry must clobber itself");
384 bool FoundClobber = false;
385 DenseSet<MemoryAccessPair> VisitedPhis;
386 SmallVector<MemoryAccessPair, 8> Worklist;
387 Worklist.emplace_back(Start, StartLoc);
388 // Walk all paths from Start to ClobberAt, while looking for clobbers. If one
389 // is found, complain.
390 while (!Worklist.empty()) {
391 MemoryAccessPair MAP = Worklist.pop_back_val();
392 // All we care about is that nothing from Start to ClobberAt clobbers Start.
393 // We learn nothing from revisiting nodes.
394 if (!VisitedPhis.insert(MAP).second)
397 for (MemoryAccess *MA : def_chain(MAP.first)) {
398 if (MA == ClobberAt) {
399 if (auto *MD = dyn_cast<MemoryDef>(MA)) {
400 // instructionClobbersQuery isn't essentially free, so don't use `|=`,
401 // since it won't let us short-circuit.
403 // Also, note that this can't be hoisted out of the `Worklist` loop,
404 // since MD may only act as a clobber for 1 of N MemoryLocations.
405 FoundClobber = FoundClobber || MSSA.isLiveOnEntryDef(MD);
408 instructionClobbersQuery(MD, MAP.second, Query.Inst, AA);
418 // We should never hit liveOnEntry, unless it's the clobber.
419 assert(!MSSA.isLiveOnEntryDef(MA) && "Hit liveOnEntry before clobber?");
421 if (auto *MD = dyn_cast<MemoryDef>(MA)) {
423 assert(!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA)
425 "Found clobber before reaching ClobberAt!");
429 assert(isa<MemoryPhi>(MA));
430 Worklist.append(upward_defs_begin({MA, MAP.second}), upward_defs_end());
434 // If ClobberAt is a MemoryPhi, we can assume something above it acted as a
435 // clobber. Otherwise, `ClobberAt` should've acted as a clobber at some point.
436 assert((isa<MemoryPhi>(ClobberAt) || FoundClobber) &&
437 "ClobberAt never acted as a clobber");
442 /// Our algorithm for walking (and trying to optimize) clobbers, all wrapped up
444 class ClobberWalker {
445 /// Save a few bytes by using unsigned instead of size_t.
446 using ListIndex = unsigned;
448 /// Represents a span of contiguous MemoryDefs, potentially ending in a
452 // Note that, because we always walk in reverse, Last will always dominate
453 // First. Also note that First and Last are inclusive.
456 Optional<ListIndex> Previous;
458 DefPath(const MemoryLocation &Loc, MemoryAccess *First, MemoryAccess *Last,
459 Optional<ListIndex> Previous)
460 : Loc(Loc), First(First), Last(Last), Previous(Previous) {}
462 DefPath(const MemoryLocation &Loc, MemoryAccess *Init,
463 Optional<ListIndex> Previous)
464 : DefPath(Loc, Init, Init, Previous) {}
467 const MemorySSA &MSSA;
470 UpwardsMemoryQuery *Query;
472 // Phi optimization bookkeeping
473 SmallVector<DefPath, 32> Paths;
474 DenseSet<ConstMemoryAccessPair> VisitedPhis;
476 /// Find the nearest def or phi that `From` can legally be optimized to.
477 const MemoryAccess *getWalkTarget(const MemoryPhi *From) const {
478 assert(From->getNumOperands() && "Phi with no operands?");
480 BasicBlock *BB = From->getBlock();
481 MemoryAccess *Result = MSSA.getLiveOnEntryDef();
482 DomTreeNode *Node = DT.getNode(BB);
483 while ((Node = Node->getIDom())) {
484 auto *Defs = MSSA.getBlockDefs(Node->getBlock());
486 return &*Defs->rbegin();
491 /// Result of calling walkToPhiOrClobber.
492 struct UpwardsWalkResult {
493 /// The "Result" of the walk. Either a clobber, the last thing we walked, or
494 /// both. Include alias info when clobber found.
495 MemoryAccess *Result;
497 Optional<AliasResult> AR;
500 /// Walk to the next Phi or Clobber in the def chain starting at Desc.Last.
501 /// This will update Desc.Last as it walks. It will (optionally) also stop at
504 /// This does not test for whether StopAt is a clobber
506 walkToPhiOrClobber(DefPath &Desc,
507 const MemoryAccess *StopAt = nullptr) const {
508 assert(!isa<MemoryUse>(Desc.Last) && "Uses don't exist in my world");
510 for (MemoryAccess *Current : def_chain(Desc.Last)) {
512 if (Current == StopAt)
513 return {Current, false, MayAlias};
515 if (auto *MD = dyn_cast<MemoryDef>(Current)) {
516 if (MSSA.isLiveOnEntryDef(MD))
517 return {MD, true, MustAlias};
519 instructionClobbersQuery(MD, Desc.Loc, Query->Inst, AA);
521 return {MD, true, CA.AR};
525 assert(isa<MemoryPhi>(Desc.Last) &&
526 "Ended at a non-clobber that's not a phi?");
527 return {Desc.Last, false, MayAlias};
530 void addSearches(MemoryPhi *Phi, SmallVectorImpl<ListIndex> &PausedSearches,
531 ListIndex PriorNode) {
532 auto UpwardDefs = make_range(upward_defs_begin({Phi, Paths[PriorNode].Loc}),
534 for (const MemoryAccessPair &P : UpwardDefs) {
535 PausedSearches.push_back(Paths.size());
536 Paths.emplace_back(P.second, P.first, PriorNode);
540 /// Represents a search that terminated after finding a clobber. This clobber
541 /// may or may not be present in the path of defs from LastNode..SearchStart,
542 /// since it may have been retrieved from cache.
543 struct TerminatedPath {
544 MemoryAccess *Clobber;
548 /// Get an access that keeps us from optimizing to the given phi.
550 /// PausedSearches is an array of indices into the Paths array. Its incoming
551 /// value is the indices of searches that stopped at the last phi optimization
552 /// target. It's left in an unspecified state.
554 /// If this returns None, NewPaused is a vector of searches that terminated
555 /// at StopWhere. Otherwise, NewPaused is left in an unspecified state.
556 Optional<TerminatedPath>
557 getBlockingAccess(const MemoryAccess *StopWhere,
558 SmallVectorImpl<ListIndex> &PausedSearches,
559 SmallVectorImpl<ListIndex> &NewPaused,
560 SmallVectorImpl<TerminatedPath> &Terminated) {
561 assert(!PausedSearches.empty() && "No searches to continue?");
563 // BFS vs DFS really doesn't make a difference here, so just do a DFS with
564 // PausedSearches as our stack.
565 while (!PausedSearches.empty()) {
566 ListIndex PathIndex = PausedSearches.pop_back_val();
567 DefPath &Node = Paths[PathIndex];
569 // If we've already visited this path with this MemoryLocation, we don't
570 // need to do so again.
572 // NOTE: That we just drop these paths on the ground makes caching
573 // behavior sporadic. e.g. given a diamond:
578 // ...If we walk D, B, A, C, we'll only cache the result of phi
579 // optimization for A, B, and D; C will be skipped because it dies here.
580 // This arguably isn't the worst thing ever, since:
581 // - We generally query things in a top-down order, so if we got below D
582 // without needing cache entries for {C, MemLoc}, then chances are
583 // that those cache entries would end up ultimately unused.
584 // - We still cache things for A, so C only needs to walk up a bit.
585 // If this behavior becomes problematic, we can fix without a ton of extra
587 if (!VisitedPhis.insert({Node.Last, Node.Loc}).second)
590 UpwardsWalkResult Res = walkToPhiOrClobber(Node, /*StopAt=*/StopWhere);
591 if (Res.IsKnownClobber) {
592 assert(Res.Result != StopWhere);
593 // If this wasn't a cache hit, we hit a clobber when walking. That's a
595 TerminatedPath Term{Res.Result, PathIndex};
596 if (!MSSA.dominates(Res.Result, StopWhere))
599 // Otherwise, it's a valid thing to potentially optimize to.
600 Terminated.push_back(Term);
604 if (Res.Result == StopWhere) {
605 // We've hit our target. Save this path off for if we want to continue
607 NewPaused.push_back(PathIndex);
611 assert(!MSSA.isLiveOnEntryDef(Res.Result) && "liveOnEntry is a clobber");
612 addSearches(cast<MemoryPhi>(Res.Result), PausedSearches, PathIndex);
618 template <typename T, typename Walker>
619 struct generic_def_path_iterator
620 : public iterator_facade_base<generic_def_path_iterator<T, Walker>,
621 std::forward_iterator_tag, T *> {
622 generic_def_path_iterator() = default;
623 generic_def_path_iterator(Walker *W, ListIndex N) : W(W), N(N) {}
625 T &operator*() const { return curNode(); }
627 generic_def_path_iterator &operator++() {
628 N = curNode().Previous;
632 bool operator==(const generic_def_path_iterator &O) const {
633 if (N.hasValue() != O.N.hasValue())
635 return !N.hasValue() || *N == *O.N;
639 T &curNode() const { return W->Paths[*N]; }
642 Optional<ListIndex> N = None;
645 using def_path_iterator = generic_def_path_iterator<DefPath, ClobberWalker>;
646 using const_def_path_iterator =
647 generic_def_path_iterator<const DefPath, const ClobberWalker>;
649 iterator_range<def_path_iterator> def_path(ListIndex From) {
650 return make_range(def_path_iterator(this, From), def_path_iterator());
653 iterator_range<const_def_path_iterator> const_def_path(ListIndex From) const {
654 return make_range(const_def_path_iterator(this, From),
655 const_def_path_iterator());
659 /// The path that contains our result.
660 TerminatedPath PrimaryClobber;
661 /// The paths that we can legally cache back from, but that aren't
662 /// necessarily the result of the Phi optimization.
663 SmallVector<TerminatedPath, 4> OtherClobbers;
666 ListIndex defPathIndex(const DefPath &N) const {
667 // The assert looks nicer if we don't need to do &N
668 const DefPath *NP = &N;
669 assert(!Paths.empty() && NP >= &Paths.front() && NP <= &Paths.back() &&
670 "Out of bounds DefPath!");
671 return NP - &Paths.front();
674 /// Try to optimize a phi as best as we can. Returns a SmallVector of Paths
675 /// that act as legal clobbers. Note that this won't return *all* clobbers.
677 /// Phi optimization algorithm tl;dr:
678 /// - Find the earliest def/phi, A, we can optimize to
679 /// - Find if all paths from the starting memory access ultimately reach A
680 /// - If not, optimization isn't possible.
681 /// - Otherwise, walk from A to another clobber or phi, A'.
682 /// - If A' is a def, we're done.
683 /// - If A' is a phi, try to optimize it.
685 /// A path is a series of {MemoryAccess, MemoryLocation} pairs. A path
686 /// terminates when a MemoryAccess that clobbers said MemoryLocation is found.
687 OptznResult tryOptimizePhi(MemoryPhi *Phi, MemoryAccess *Start,
688 const MemoryLocation &Loc) {
689 assert(Paths.empty() && VisitedPhis.empty() &&
690 "Reset the optimization state.");
692 Paths.emplace_back(Loc, Start, Phi, None);
693 // Stores how many "valid" optimization nodes we had prior to calling
694 // addSearches/getBlockingAccess. Necessary for caching if we had a blocker.
695 auto PriorPathsSize = Paths.size();
697 SmallVector<ListIndex, 16> PausedSearches;
698 SmallVector<ListIndex, 8> NewPaused;
699 SmallVector<TerminatedPath, 4> TerminatedPaths;
701 addSearches(Phi, PausedSearches, 0);
703 // Moves the TerminatedPath with the "most dominated" Clobber to the end of
705 auto MoveDominatedPathToEnd = [&](SmallVectorImpl<TerminatedPath> &Paths) {
706 assert(!Paths.empty() && "Need a path to move");
707 auto Dom = Paths.begin();
708 for (auto I = std::next(Dom), E = Paths.end(); I != E; ++I)
709 if (!MSSA.dominates(I->Clobber, Dom->Clobber))
711 auto Last = Paths.end() - 1;
713 std::iter_swap(Last, Dom);
716 MemoryPhi *Current = Phi;
718 assert(!MSSA.isLiveOnEntryDef(Current) &&
719 "liveOnEntry wasn't treated as a clobber?");
721 const auto *Target = getWalkTarget(Current);
722 // If a TerminatedPath doesn't dominate Target, then it wasn't a legal
723 // optimization for the prior phi.
724 assert(all_of(TerminatedPaths, [&](const TerminatedPath &P) {
725 return MSSA.dominates(P.Clobber, Target);
728 // FIXME: This is broken, because the Blocker may be reported to be
729 // liveOnEntry, and we'll happily wait for that to disappear (read: never)
730 // For the moment, this is fine, since we do nothing with blocker info.
731 if (Optional<TerminatedPath> Blocker = getBlockingAccess(
732 Target, PausedSearches, NewPaused, TerminatedPaths)) {
734 // Find the node we started at. We can't search based on N->Last, since
735 // we may have gone around a loop with a different MemoryLocation.
736 auto Iter = find_if(def_path(Blocker->LastNode), [&](const DefPath &N) {
737 return defPathIndex(N) < PriorPathsSize;
739 assert(Iter != def_path_iterator());
741 DefPath &CurNode = *Iter;
742 assert(CurNode.Last == Current);
745 // A. We can't reliably cache all of NewPaused back. Consider a case
746 // where we have two paths in NewPaused; one of which can't optimize
747 // above this phi, whereas the other can. If we cache the second path
748 // back, we'll end up with suboptimal cache entries. We can handle
749 // cases like this a bit better when we either try to find all
750 // clobbers that block phi optimization, or when our cache starts
751 // supporting unfinished searches.
752 // B. We can't reliably cache TerminatedPaths back here without doing
753 // extra checks; consider a case like:
759 // Where T is our target, C is a node with a clobber on it, D is a
760 // diamond (with a clobber *only* on the left or right node, N), and
761 // S is our start. Say we walk to D, through the node opposite N
762 // (read: ignoring the clobber), and see a cache entry in the top
763 // node of D. That cache entry gets put into TerminatedPaths. We then
764 // walk up to C (N is later in our worklist), find the clobber, and
765 // quit. If we append TerminatedPaths to OtherClobbers, we'll cache
766 // the bottom part of D to the cached clobber, ignoring the clobber
767 // in N. Again, this problem goes away if we start tracking all
768 // blockers for a given phi optimization.
769 TerminatedPath Result{CurNode.Last, defPathIndex(CurNode)};
773 // If there's nothing left to search, then all paths led to valid clobbers
774 // that we got from our cache; pick the nearest to the start, and allow
775 // the rest to be cached back.
776 if (NewPaused.empty()) {
777 MoveDominatedPathToEnd(TerminatedPaths);
778 TerminatedPath Result = TerminatedPaths.pop_back_val();
779 return {Result, std::move(TerminatedPaths)};
782 MemoryAccess *DefChainEnd = nullptr;
783 SmallVector<TerminatedPath, 4> Clobbers;
784 for (ListIndex Paused : NewPaused) {
785 UpwardsWalkResult WR = walkToPhiOrClobber(Paths[Paused]);
786 if (WR.IsKnownClobber)
787 Clobbers.push_back({WR.Result, Paused});
789 // Micro-opt: If we hit the end of the chain, save it.
790 DefChainEnd = WR.Result;
793 if (!TerminatedPaths.empty()) {
794 // If we couldn't find the dominating phi/liveOnEntry in the above loop,
797 for (auto *MA : def_chain(const_cast<MemoryAccess *>(Target)))
800 // If any of the terminated paths don't dominate the phi we'll try to
801 // optimize, we need to figure out what they are and quit.
802 const BasicBlock *ChainBB = DefChainEnd->getBlock();
803 for (const TerminatedPath &TP : TerminatedPaths) {
804 // Because we know that DefChainEnd is as "high" as we can go, we
805 // don't need local dominance checks; BB dominance is sufficient.
806 if (DT.dominates(ChainBB, TP.Clobber->getBlock()))
807 Clobbers.push_back(TP);
811 // If we have clobbers in the def chain, find the one closest to Current
813 if (!Clobbers.empty()) {
814 MoveDominatedPathToEnd(Clobbers);
815 TerminatedPath Result = Clobbers.pop_back_val();
816 return {Result, std::move(Clobbers)};
819 assert(all_of(NewPaused,
820 [&](ListIndex I) { return Paths[I].Last == DefChainEnd; }));
822 // Because liveOnEntry is a clobber, this must be a phi.
823 auto *DefChainPhi = cast<MemoryPhi>(DefChainEnd);
825 PriorPathsSize = Paths.size();
826 PausedSearches.clear();
827 for (ListIndex I : NewPaused)
828 addSearches(DefChainPhi, PausedSearches, I);
831 Current = DefChainPhi;
835 void verifyOptResult(const OptznResult &R) const {
836 assert(all_of(R.OtherClobbers, [&](const TerminatedPath &P) {
837 return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber);
841 void resetPhiOptznState() {
847 ClobberWalker(const MemorySSA &MSSA, AliasAnalysis &AA, DominatorTree &DT)
848 : MSSA(MSSA), AA(AA), DT(DT) {}
850 /// Finds the nearest clobber for the given query, optimizing phis if
852 MemoryAccess *findClobber(MemoryAccess *Start, UpwardsMemoryQuery &Q) {
855 MemoryAccess *Current = Start;
856 // This walker pretends uses don't exist. If we're handed one, silently grab
857 // its def. (This has the nice side-effect of ensuring we never cache uses)
858 if (auto *MU = dyn_cast<MemoryUse>(Start))
859 Current = MU->getDefiningAccess();
861 DefPath FirstDesc(Q.StartingLoc, Current, Current, None);
862 // Fast path for the overly-common case (no crazy phi optimization
864 UpwardsWalkResult WalkResult = walkToPhiOrClobber(FirstDesc);
865 MemoryAccess *Result;
866 if (WalkResult.IsKnownClobber) {
867 Result = WalkResult.Result;
868 Q.AR = WalkResult.AR;
870 OptznResult OptRes = tryOptimizePhi(cast<MemoryPhi>(FirstDesc.Last),
871 Current, Q.StartingLoc);
872 verifyOptResult(OptRes);
873 resetPhiOptznState();
874 Result = OptRes.PrimaryClobber.Clobber;
877 #ifdef EXPENSIVE_CHECKS
878 checkClobberSanity(Current, Result, Q.StartingLoc, MSSA, Q, AA);
883 void verify(const MemorySSA *MSSA) { assert(MSSA == &this->MSSA); }
886 struct RenamePassData {
888 DomTreeNode::const_iterator ChildIt;
889 MemoryAccess *IncomingVal;
891 RenamePassData(DomTreeNode *D, DomTreeNode::const_iterator It,
893 : DTN(D), ChildIt(It), IncomingVal(M) {}
895 void swap(RenamePassData &RHS) {
896 std::swap(DTN, RHS.DTN);
897 std::swap(ChildIt, RHS.ChildIt);
898 std::swap(IncomingVal, RHS.IncomingVal);
902 } // end anonymous namespace
906 /// A MemorySSAWalker that does AA walks to disambiguate accesses. It no
907 /// longer does caching on its own, but the name has been retained for the
909 class MemorySSA::CachingWalker final : public MemorySSAWalker {
910 ClobberWalker Walker;
912 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *, UpwardsMemoryQuery &);
915 CachingWalker(MemorySSA *, AliasAnalysis *, DominatorTree *);
916 ~CachingWalker() override = default;
918 using MemorySSAWalker::getClobberingMemoryAccess;
920 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *) override;
921 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *,
922 const MemoryLocation &) override;
923 void invalidateInfo(MemoryAccess *) override;
925 void verify(const MemorySSA *MSSA) override {
926 MemorySSAWalker::verify(MSSA);
931 } // end namespace llvm
933 void MemorySSA::renameSuccessorPhis(BasicBlock *BB, MemoryAccess *IncomingVal,
934 bool RenameAllUses) {
935 // Pass through values to our successors
936 for (const BasicBlock *S : successors(BB)) {
937 auto It = PerBlockAccesses.find(S);
938 // Rename the phi nodes in our successor block
939 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
941 AccessList *Accesses = It->second.get();
942 auto *Phi = cast<MemoryPhi>(&Accesses->front());
944 int PhiIndex = Phi->getBasicBlockIndex(BB);
945 assert(PhiIndex != -1 && "Incomplete phi during partial rename");
946 Phi->setIncomingValue(PhiIndex, IncomingVal);
948 Phi->addIncoming(IncomingVal, BB);
952 /// Rename a single basic block into MemorySSA form.
953 /// Uses the standard SSA renaming algorithm.
954 /// \returns The new incoming value.
955 MemoryAccess *MemorySSA::renameBlock(BasicBlock *BB, MemoryAccess *IncomingVal,
956 bool RenameAllUses) {
957 auto It = PerBlockAccesses.find(BB);
958 // Skip most processing if the list is empty.
959 if (It != PerBlockAccesses.end()) {
960 AccessList *Accesses = It->second.get();
961 for (MemoryAccess &L : *Accesses) {
962 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(&L)) {
963 if (MUD->getDefiningAccess() == nullptr || RenameAllUses)
964 MUD->setDefiningAccess(IncomingVal);
965 if (isa<MemoryDef>(&L))
975 /// This is the standard SSA renaming algorithm.
977 /// We walk the dominator tree in preorder, renaming accesses, and then filling
978 /// in phi nodes in our successors.
979 void MemorySSA::renamePass(DomTreeNode *Root, MemoryAccess *IncomingVal,
980 SmallPtrSetImpl<BasicBlock *> &Visited,
981 bool SkipVisited, bool RenameAllUses) {
982 SmallVector<RenamePassData, 32> WorkStack;
983 // Skip everything if we already renamed this block and we are skipping.
984 // Note: You can't sink this into the if, because we need it to occur
985 // regardless of whether we skip blocks or not.
986 bool AlreadyVisited = !Visited.insert(Root->getBlock()).second;
987 if (SkipVisited && AlreadyVisited)
990 IncomingVal = renameBlock(Root->getBlock(), IncomingVal, RenameAllUses);
991 renameSuccessorPhis(Root->getBlock(), IncomingVal, RenameAllUses);
992 WorkStack.push_back({Root, Root->begin(), IncomingVal});
994 while (!WorkStack.empty()) {
995 DomTreeNode *Node = WorkStack.back().DTN;
996 DomTreeNode::const_iterator ChildIt = WorkStack.back().ChildIt;
997 IncomingVal = WorkStack.back().IncomingVal;
999 if (ChildIt == Node->end()) {
1000 WorkStack.pop_back();
1002 DomTreeNode *Child = *ChildIt;
1003 ++WorkStack.back().ChildIt;
1004 BasicBlock *BB = Child->getBlock();
1005 // Note: You can't sink this into the if, because we need it to occur
1006 // regardless of whether we skip blocks or not.
1007 AlreadyVisited = !Visited.insert(BB).second;
1008 if (SkipVisited && AlreadyVisited) {
1009 // We already visited this during our renaming, which can happen when
1010 // being asked to rename multiple blocks. Figure out the incoming val,
1011 // which is the last def.
1012 // Incoming value can only change if there is a block def, and in that
1013 // case, it's the last block def in the list.
1014 if (auto *BlockDefs = getWritableBlockDefs(BB))
1015 IncomingVal = &*BlockDefs->rbegin();
1017 IncomingVal = renameBlock(BB, IncomingVal, RenameAllUses);
1018 renameSuccessorPhis(BB, IncomingVal, RenameAllUses);
1019 WorkStack.push_back({Child, Child->begin(), IncomingVal});
1024 /// This handles unreachable block accesses by deleting phi nodes in
1025 /// unreachable blocks, and marking all other unreachable MemoryAccess's as
1026 /// being uses of the live on entry definition.
1027 void MemorySSA::markUnreachableAsLiveOnEntry(BasicBlock *BB) {
1028 assert(!DT->isReachableFromEntry(BB) &&
1029 "Reachable block found while handling unreachable blocks");
1031 // Make sure phi nodes in our reachable successors end up with a
1032 // LiveOnEntryDef for our incoming edge, even though our block is forward
1033 // unreachable. We could just disconnect these blocks from the CFG fully,
1034 // but we do not right now.
1035 for (const BasicBlock *S : successors(BB)) {
1036 if (!DT->isReachableFromEntry(S))
1038 auto It = PerBlockAccesses.find(S);
1039 // Rename the phi nodes in our successor block
1040 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
1042 AccessList *Accesses = It->second.get();
1043 auto *Phi = cast<MemoryPhi>(&Accesses->front());
1044 Phi->addIncoming(LiveOnEntryDef.get(), BB);
1047 auto It = PerBlockAccesses.find(BB);
1048 if (It == PerBlockAccesses.end())
1051 auto &Accesses = It->second;
1052 for (auto AI = Accesses->begin(), AE = Accesses->end(); AI != AE;) {
1053 auto Next = std::next(AI);
1054 // If we have a phi, just remove it. We are going to replace all
1055 // users with live on entry.
1056 if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(AI))
1057 UseOrDef->setDefiningAccess(LiveOnEntryDef.get());
1059 Accesses->erase(AI);
1064 MemorySSA::MemorySSA(Function &Func, AliasAnalysis *AA, DominatorTree *DT)
1065 : AA(AA), DT(DT), F(Func), LiveOnEntryDef(nullptr), Walker(nullptr),
1070 MemorySSA::~MemorySSA() {
1071 // Drop all our references
1072 for (const auto &Pair : PerBlockAccesses)
1073 for (MemoryAccess &MA : *Pair.second)
1074 MA.dropAllReferences();
1077 MemorySSA::AccessList *MemorySSA::getOrCreateAccessList(const BasicBlock *BB) {
1078 auto Res = PerBlockAccesses.insert(std::make_pair(BB, nullptr));
1081 Res.first->second = llvm::make_unique<AccessList>();
1082 return Res.first->second.get();
1085 MemorySSA::DefsList *MemorySSA::getOrCreateDefsList(const BasicBlock *BB) {
1086 auto Res = PerBlockDefs.insert(std::make_pair(BB, nullptr));
1089 Res.first->second = llvm::make_unique<DefsList>();
1090 return Res.first->second.get();
1095 /// This class is a batch walker of all MemoryUse's in the program, and points
1096 /// their defining access at the thing that actually clobbers them. Because it
1097 /// is a batch walker that touches everything, it does not operate like the
1098 /// other walkers. This walker is basically performing a top-down SSA renaming
1099 /// pass, where the version stack is used as the cache. This enables it to be
1100 /// significantly more time and memory efficient than using the regular walker,
1101 /// which is walking bottom-up.
1102 class MemorySSA::OptimizeUses {
1104 OptimizeUses(MemorySSA *MSSA, MemorySSAWalker *Walker, AliasAnalysis *AA,
1106 : MSSA(MSSA), Walker(Walker), AA(AA), DT(DT) {
1107 Walker = MSSA->getWalker();
1110 void optimizeUses();
1113 /// This represents where a given memorylocation is in the stack.
1114 struct MemlocStackInfo {
1115 // This essentially is keeping track of versions of the stack. Whenever
1116 // the stack changes due to pushes or pops, these versions increase.
1117 unsigned long StackEpoch;
1118 unsigned long PopEpoch;
1119 // This is the lower bound of places on the stack to check. It is equal to
1120 // the place the last stack walk ended.
1121 // Note: Correctness depends on this being initialized to 0, which densemap
1123 unsigned long LowerBound;
1124 const BasicBlock *LowerBoundBlock;
1125 // This is where the last walk for this memory location ended.
1126 unsigned long LastKill;
1128 Optional<AliasResult> AR;
1131 void optimizeUsesInBlock(const BasicBlock *, unsigned long &, unsigned long &,
1132 SmallVectorImpl<MemoryAccess *> &,
1133 DenseMap<MemoryLocOrCall, MemlocStackInfo> &);
1136 MemorySSAWalker *Walker;
1141 } // end namespace llvm
1143 /// Optimize the uses in a given block This is basically the SSA renaming
1144 /// algorithm, with one caveat: We are able to use a single stack for all
1145 /// MemoryUses. This is because the set of *possible* reaching MemoryDefs is
1146 /// the same for every MemoryUse. The *actual* clobbering MemoryDef is just
1147 /// going to be some position in that stack of possible ones.
1149 /// We track the stack positions that each MemoryLocation needs
1150 /// to check, and last ended at. This is because we only want to check the
1151 /// things that changed since last time. The same MemoryLocation should
1152 /// get clobbered by the same store (getModRefInfo does not use invariantness or
1153 /// things like this, and if they start, we can modify MemoryLocOrCall to
1154 /// include relevant data)
1155 void MemorySSA::OptimizeUses::optimizeUsesInBlock(
1156 const BasicBlock *BB, unsigned long &StackEpoch, unsigned long &PopEpoch,
1157 SmallVectorImpl<MemoryAccess *> &VersionStack,
1158 DenseMap<MemoryLocOrCall, MemlocStackInfo> &LocStackInfo) {
1160 /// If no accesses, nothing to do.
1161 MemorySSA::AccessList *Accesses = MSSA->getWritableBlockAccesses(BB);
1162 if (Accesses == nullptr)
1165 // Pop everything that doesn't dominate the current block off the stack,
1166 // increment the PopEpoch to account for this.
1169 !VersionStack.empty() &&
1170 "Version stack should have liveOnEntry sentinel dominating everything");
1171 BasicBlock *BackBlock = VersionStack.back()->getBlock();
1172 if (DT->dominates(BackBlock, BB))
1174 while (VersionStack.back()->getBlock() == BackBlock)
1175 VersionStack.pop_back();
1179 for (MemoryAccess &MA : *Accesses) {
1180 auto *MU = dyn_cast<MemoryUse>(&MA);
1182 VersionStack.push_back(&MA);
1187 if (isUseTriviallyOptimizableToLiveOnEntry(*AA, MU->getMemoryInst())) {
1188 MU->setDefiningAccess(MSSA->getLiveOnEntryDef(), true, None);
1192 MemoryLocOrCall UseMLOC(MU);
1193 auto &LocInfo = LocStackInfo[UseMLOC];
1194 // If the pop epoch changed, it means we've removed stuff from top of
1195 // stack due to changing blocks. We may have to reset the lower bound or
1197 if (LocInfo.PopEpoch != PopEpoch) {
1198 LocInfo.PopEpoch = PopEpoch;
1199 LocInfo.StackEpoch = StackEpoch;
1200 // If the lower bound was in something that no longer dominates us, we
1201 // have to reset it.
1202 // We can't simply track stack size, because the stack may have had
1203 // pushes/pops in the meantime.
1204 // XXX: This is non-optimal, but only is slower cases with heavily
1205 // branching dominator trees. To get the optimal number of queries would
1206 // be to make lowerbound and lastkill a per-loc stack, and pop it until
1207 // the top of that stack dominates us. This does not seem worth it ATM.
1208 // A much cheaper optimization would be to always explore the deepest
1209 // branch of the dominator tree first. This will guarantee this resets on
1210 // the smallest set of blocks.
1211 if (LocInfo.LowerBoundBlock && LocInfo.LowerBoundBlock != BB &&
1212 !DT->dominates(LocInfo.LowerBoundBlock, BB)) {
1213 // Reset the lower bound of things to check.
1214 // TODO: Some day we should be able to reset to last kill, rather than
1216 LocInfo.LowerBound = 0;
1217 LocInfo.LowerBoundBlock = VersionStack[0]->getBlock();
1218 LocInfo.LastKillValid = false;
1220 } else if (LocInfo.StackEpoch != StackEpoch) {
1221 // If all that has changed is the StackEpoch, we only have to check the
1222 // new things on the stack, because we've checked everything before. In
1223 // this case, the lower bound of things to check remains the same.
1224 LocInfo.PopEpoch = PopEpoch;
1225 LocInfo.StackEpoch = StackEpoch;
1227 if (!LocInfo.LastKillValid) {
1228 LocInfo.LastKill = VersionStack.size() - 1;
1229 LocInfo.LastKillValid = true;
1230 LocInfo.AR = MayAlias;
1233 // At this point, we should have corrected last kill and LowerBound to be
1235 assert(LocInfo.LowerBound < VersionStack.size() &&
1236 "Lower bound out of range");
1237 assert(LocInfo.LastKill < VersionStack.size() &&
1238 "Last kill info out of range");
1239 // In any case, the new upper bound is the top of the stack.
1240 unsigned long UpperBound = VersionStack.size() - 1;
1242 if (UpperBound - LocInfo.LowerBound > MaxCheckLimit) {
1243 LLVM_DEBUG(dbgs() << "MemorySSA skipping optimization of " << *MU << " ("
1244 << *(MU->getMemoryInst()) << ")"
1245 << " because there are "
1246 << UpperBound - LocInfo.LowerBound
1247 << " stores to disambiguate\n");
1248 // Because we did not walk, LastKill is no longer valid, as this may
1249 // have been a kill.
1250 LocInfo.LastKillValid = false;
1253 bool FoundClobberResult = false;
1254 while (UpperBound > LocInfo.LowerBound) {
1255 if (isa<MemoryPhi>(VersionStack[UpperBound])) {
1256 // For phis, use the walker, see where we ended up, go there
1257 Instruction *UseInst = MU->getMemoryInst();
1258 MemoryAccess *Result = Walker->getClobberingMemoryAccess(UseInst);
1259 // We are guaranteed to find it or something is wrong
1260 while (VersionStack[UpperBound] != Result) {
1261 assert(UpperBound != 0);
1264 FoundClobberResult = true;
1268 MemoryDef *MD = cast<MemoryDef>(VersionStack[UpperBound]);
1269 // If the lifetime of the pointer ends at this instruction, it's live on
1271 if (!UseMLOC.IsCall && lifetimeEndsAt(MD, UseMLOC.getLoc(), *AA)) {
1272 // Reset UpperBound to liveOnEntryDef's place in the stack
1274 FoundClobberResult = true;
1275 LocInfo.AR = MustAlias;
1278 ClobberAlias CA = instructionClobbersQuery(MD, MU, UseMLOC, *AA);
1280 FoundClobberResult = true;
1287 // Note: Phis always have AliasResult AR set to MayAlias ATM.
1289 // At the end of this loop, UpperBound is either a clobber, or lower bound
1290 // PHI walking may cause it to be < LowerBound, and in fact, < LastKill.
1291 if (FoundClobberResult || UpperBound < LocInfo.LastKill) {
1292 // We were last killed now by where we got to
1293 if (MSSA->isLiveOnEntryDef(VersionStack[UpperBound]))
1295 MU->setDefiningAccess(VersionStack[UpperBound], true, LocInfo.AR);
1296 LocInfo.LastKill = UpperBound;
1298 // Otherwise, we checked all the new ones, and now we know we can get to
1300 MU->setDefiningAccess(VersionStack[LocInfo.LastKill], true, LocInfo.AR);
1302 LocInfo.LowerBound = VersionStack.size() - 1;
1303 LocInfo.LowerBoundBlock = BB;
1307 /// Optimize uses to point to their actual clobbering definitions.
1308 void MemorySSA::OptimizeUses::optimizeUses() {
1309 SmallVector<MemoryAccess *, 16> VersionStack;
1310 DenseMap<MemoryLocOrCall, MemlocStackInfo> LocStackInfo;
1311 VersionStack.push_back(MSSA->getLiveOnEntryDef());
1313 unsigned long StackEpoch = 1;
1314 unsigned long PopEpoch = 1;
1315 // We perform a non-recursive top-down dominator tree walk.
1316 for (const auto *DomNode : depth_first(DT->getRootNode()))
1317 optimizeUsesInBlock(DomNode->getBlock(), StackEpoch, PopEpoch, VersionStack,
1321 void MemorySSA::placePHINodes(
1322 const SmallPtrSetImpl<BasicBlock *> &DefiningBlocks) {
1323 // Determine where our MemoryPhi's should go
1324 ForwardIDFCalculator IDFs(*DT);
1325 IDFs.setDefiningBlocks(DefiningBlocks);
1326 SmallVector<BasicBlock *, 32> IDFBlocks;
1327 IDFs.calculate(IDFBlocks);
1329 // Now place MemoryPhi nodes.
1330 for (auto &BB : IDFBlocks)
1331 createMemoryPhi(BB);
1334 void MemorySSA::buildMemorySSA() {
1335 // We create an access to represent "live on entry", for things like
1336 // arguments or users of globals, where the memory they use is defined before
1337 // the beginning of the function. We do not actually insert it into the IR.
1338 // We do not define a live on exit for the immediate uses, and thus our
1339 // semantics do *not* imply that something with no immediate uses can simply
1341 BasicBlock &StartingPoint = F.getEntryBlock();
1342 LiveOnEntryDef.reset(new MemoryDef(F.getContext(), nullptr, nullptr,
1343 &StartingPoint, NextID++));
1345 // We maintain lists of memory accesses per-block, trading memory for time. We
1346 // could just look up the memory access for every possible instruction in the
1348 SmallPtrSet<BasicBlock *, 32> DefiningBlocks;
1349 // Go through each block, figure out where defs occur, and chain together all
1351 for (BasicBlock &B : F) {
1352 bool InsertIntoDef = false;
1353 AccessList *Accesses = nullptr;
1354 DefsList *Defs = nullptr;
1355 for (Instruction &I : B) {
1356 MemoryUseOrDef *MUD = createNewAccess(&I);
1361 Accesses = getOrCreateAccessList(&B);
1362 Accesses->push_back(MUD);
1363 if (isa<MemoryDef>(MUD)) {
1364 InsertIntoDef = true;
1366 Defs = getOrCreateDefsList(&B);
1367 Defs->push_back(*MUD);
1371 DefiningBlocks.insert(&B);
1373 placePHINodes(DefiningBlocks);
1375 // Now do regular SSA renaming on the MemoryDef/MemoryUse. Visited will get
1376 // filled in with all blocks.
1377 SmallPtrSet<BasicBlock *, 16> Visited;
1378 renamePass(DT->getRootNode(), LiveOnEntryDef.get(), Visited);
1380 CachingWalker *Walker = getWalkerImpl();
1382 OptimizeUses(this, Walker, AA, DT).optimizeUses();
1384 // Mark the uses in unreachable blocks as live on entry, so that they go
1387 if (!Visited.count(&BB))
1388 markUnreachableAsLiveOnEntry(&BB);
1391 MemorySSAWalker *MemorySSA::getWalker() { return getWalkerImpl(); }
1393 MemorySSA::CachingWalker *MemorySSA::getWalkerImpl() {
1395 return Walker.get();
1397 Walker = llvm::make_unique<CachingWalker>(this, AA, DT);
1398 return Walker.get();
1401 // This is a helper function used by the creation routines. It places NewAccess
1402 // into the access and defs lists for a given basic block, at the given
1404 void MemorySSA::insertIntoListsForBlock(MemoryAccess *NewAccess,
1405 const BasicBlock *BB,
1406 InsertionPlace Point) {
1407 auto *Accesses = getOrCreateAccessList(BB);
1408 if (Point == Beginning) {
1409 // If it's a phi node, it goes first, otherwise, it goes after any phi
1411 if (isa<MemoryPhi>(NewAccess)) {
1412 Accesses->push_front(NewAccess);
1413 auto *Defs = getOrCreateDefsList(BB);
1414 Defs->push_front(*NewAccess);
1416 auto AI = find_if_not(
1417 *Accesses, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
1418 Accesses->insert(AI, NewAccess);
1419 if (!isa<MemoryUse>(NewAccess)) {
1420 auto *Defs = getOrCreateDefsList(BB);
1421 auto DI = find_if_not(
1422 *Defs, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); });
1423 Defs->insert(DI, *NewAccess);
1427 Accesses->push_back(NewAccess);
1428 if (!isa<MemoryUse>(NewAccess)) {
1429 auto *Defs = getOrCreateDefsList(BB);
1430 Defs->push_back(*NewAccess);
1433 BlockNumberingValid.erase(BB);
1436 void MemorySSA::insertIntoListsBefore(MemoryAccess *What, const BasicBlock *BB,
1437 AccessList::iterator InsertPt) {
1438 auto *Accesses = getWritableBlockAccesses(BB);
1439 bool WasEnd = InsertPt == Accesses->end();
1440 Accesses->insert(AccessList::iterator(InsertPt), What);
1441 if (!isa<MemoryUse>(What)) {
1442 auto *Defs = getOrCreateDefsList(BB);
1443 // If we got asked to insert at the end, we have an easy job, just shove it
1444 // at the end. If we got asked to insert before an existing def, we also get
1445 // an iterator. If we got asked to insert before a use, we have to hunt for
1448 Defs->push_back(*What);
1449 } else if (isa<MemoryDef>(InsertPt)) {
1450 Defs->insert(InsertPt->getDefsIterator(), *What);
1452 while (InsertPt != Accesses->end() && !isa<MemoryDef>(InsertPt))
1454 // Either we found a def, or we are inserting at the end
1455 if (InsertPt == Accesses->end())
1456 Defs->push_back(*What);
1458 Defs->insert(InsertPt->getDefsIterator(), *What);
1461 BlockNumberingValid.erase(BB);
1464 // Move What before Where in the IR. The end result is that What will belong to
1465 // the right lists and have the right Block set, but will not otherwise be
1466 // correct. It will not have the right defining access, and if it is a def,
1467 // things below it will not properly be updated.
1468 void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB,
1469 AccessList::iterator Where) {
1470 // Keep it in the lookup tables, remove from the lists
1471 removeFromLists(What, false);
1473 insertIntoListsBefore(What, BB, Where);
1476 void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB,
1477 InsertionPlace Point) {
1478 removeFromLists(What, false);
1480 insertIntoListsForBlock(What, BB, Point);
1483 MemoryPhi *MemorySSA::createMemoryPhi(BasicBlock *BB) {
1484 assert(!getMemoryAccess(BB) && "MemoryPhi already exists for this BB");
1485 MemoryPhi *Phi = new MemoryPhi(BB->getContext(), BB, NextID++);
1486 // Phi's always are placed at the front of the block.
1487 insertIntoListsForBlock(Phi, BB, Beginning);
1488 ValueToMemoryAccess[BB] = Phi;
1492 MemoryUseOrDef *MemorySSA::createDefinedAccess(Instruction *I,
1493 MemoryAccess *Definition) {
1494 assert(!isa<PHINode>(I) && "Cannot create a defined access for a PHI");
1495 MemoryUseOrDef *NewAccess = createNewAccess(I);
1497 NewAccess != nullptr &&
1498 "Tried to create a memory access for a non-memory touching instruction");
1499 NewAccess->setDefiningAccess(Definition);
1503 // Return true if the instruction has ordering constraints.
1504 // Note specifically that this only considers stores and loads
1505 // because others are still considered ModRef by getModRefInfo.
1506 static inline bool isOrdered(const Instruction *I) {
1507 if (auto *SI = dyn_cast<StoreInst>(I)) {
1508 if (!SI->isUnordered())
1510 } else if (auto *LI = dyn_cast<LoadInst>(I)) {
1511 if (!LI->isUnordered())
1517 /// Helper function to create new memory accesses
1518 MemoryUseOrDef *MemorySSA::createNewAccess(Instruction *I) {
1519 // The assume intrinsic has a control dependency which we model by claiming
1520 // that it writes arbitrarily. Ignore that fake memory dependency here.
1521 // FIXME: Replace this special casing with a more accurate modelling of
1522 // assume's control dependency.
1523 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1524 if (II->getIntrinsicID() == Intrinsic::assume)
1527 // Find out what affect this instruction has on memory.
1528 ModRefInfo ModRef = AA->getModRefInfo(I, None);
1529 // The isOrdered check is used to ensure that volatiles end up as defs
1530 // (atomics end up as ModRef right now anyway). Until we separate the
1531 // ordering chain from the memory chain, this enables people to see at least
1532 // some relative ordering to volatiles. Note that getClobberingMemoryAccess
1533 // will still give an answer that bypasses other volatile loads. TODO:
1534 // Separate memory aliasing and ordering into two different chains so that we
1535 // can precisely represent both "what memory will this read/write/is clobbered
1536 // by" and "what instructions can I move this past".
1537 bool Def = isModSet(ModRef) || isOrdered(I);
1538 bool Use = isRefSet(ModRef);
1540 // It's possible for an instruction to not modify memory at all. During
1541 // construction, we ignore them.
1545 MemoryUseOrDef *MUD;
1547 MUD = new MemoryDef(I->getContext(), nullptr, I, I->getParent(), NextID++);
1549 MUD = new MemoryUse(I->getContext(), nullptr, I, I->getParent());
1550 ValueToMemoryAccess[I] = MUD;
1554 /// Returns true if \p Replacer dominates \p Replacee .
1555 bool MemorySSA::dominatesUse(const MemoryAccess *Replacer,
1556 const MemoryAccess *Replacee) const {
1557 if (isa<MemoryUseOrDef>(Replacee))
1558 return DT->dominates(Replacer->getBlock(), Replacee->getBlock());
1559 const auto *MP = cast<MemoryPhi>(Replacee);
1560 // For a phi node, the use occurs in the predecessor block of the phi node.
1561 // Since we may occur multiple times in the phi node, we have to check each
1562 // operand to ensure Replacer dominates each operand where Replacee occurs.
1563 for (const Use &Arg : MP->operands()) {
1564 if (Arg.get() != Replacee &&
1565 !DT->dominates(Replacer->getBlock(), MP->getIncomingBlock(Arg)))
1571 /// Properly remove \p MA from all of MemorySSA's lookup tables.
1572 void MemorySSA::removeFromLookups(MemoryAccess *MA) {
1573 assert(MA->use_empty() &&
1574 "Trying to remove memory access that still has uses");
1575 BlockNumbering.erase(MA);
1576 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA))
1577 MUD->setDefiningAccess(nullptr);
1578 // Invalidate our walker's cache if necessary
1579 if (!isa<MemoryUse>(MA))
1580 Walker->invalidateInfo(MA);
1581 // The call below to erase will destroy MA, so we can't change the order we
1582 // are doing things here
1584 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA)) {
1585 MemoryInst = MUD->getMemoryInst();
1587 MemoryInst = MA->getBlock();
1589 auto VMA = ValueToMemoryAccess.find(MemoryInst);
1590 if (VMA->second == MA)
1591 ValueToMemoryAccess.erase(VMA);
1594 /// Properly remove \p MA from all of MemorySSA's lists.
1596 /// Because of the way the intrusive list and use lists work, it is important to
1597 /// do removal in the right order.
1598 /// ShouldDelete defaults to true, and will cause the memory access to also be
1599 /// deleted, not just removed.
1600 void MemorySSA::removeFromLists(MemoryAccess *MA, bool ShouldDelete) {
1601 // The access list owns the reference, so we erase it from the non-owning list
1603 if (!isa<MemoryUse>(MA)) {
1604 auto DefsIt = PerBlockDefs.find(MA->getBlock());
1605 std::unique_ptr<DefsList> &Defs = DefsIt->second;
1608 PerBlockDefs.erase(DefsIt);
1611 // The erase call here will delete it. If we don't want it deleted, we call
1613 auto AccessIt = PerBlockAccesses.find(MA->getBlock());
1614 std::unique_ptr<AccessList> &Accesses = AccessIt->second;
1616 Accesses->erase(MA);
1618 Accesses->remove(MA);
1620 if (Accesses->empty())
1621 PerBlockAccesses.erase(AccessIt);
1624 void MemorySSA::print(raw_ostream &OS) const {
1625 MemorySSAAnnotatedWriter Writer(this);
1626 F.print(OS, &Writer);
1629 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1630 LLVM_DUMP_METHOD void MemorySSA::dump() const { print(dbgs()); }
1633 void MemorySSA::verifyMemorySSA() const {
1635 verifyDomination(F);
1637 Walker->verify(this);
1640 /// Verify that the order and existence of MemoryAccesses matches the
1641 /// order and existence of memory affecting instructions.
1642 void MemorySSA::verifyOrdering(Function &F) const {
1643 // Walk all the blocks, comparing what the lookups think and what the access
1644 // lists think, as well as the order in the blocks vs the order in the access
1646 SmallVector<MemoryAccess *, 32> ActualAccesses;
1647 SmallVector<MemoryAccess *, 32> ActualDefs;
1648 for (BasicBlock &B : F) {
1649 const AccessList *AL = getBlockAccesses(&B);
1650 const auto *DL = getBlockDefs(&B);
1651 MemoryAccess *Phi = getMemoryAccess(&B);
1653 ActualAccesses.push_back(Phi);
1654 ActualDefs.push_back(Phi);
1657 for (Instruction &I : B) {
1658 MemoryAccess *MA = getMemoryAccess(&I);
1659 assert((!MA || (AL && (isa<MemoryUse>(MA) || DL))) &&
1660 "We have memory affecting instructions "
1661 "in this block but they are not in the "
1662 "access list or defs list");
1664 ActualAccesses.push_back(MA);
1665 if (isa<MemoryDef>(MA))
1666 ActualDefs.push_back(MA);
1669 // Either we hit the assert, really have no accesses, or we have both
1670 // accesses and an access list.
1674 assert(AL->size() == ActualAccesses.size() &&
1675 "We don't have the same number of accesses in the block as on the "
1677 assert((DL || ActualDefs.size() == 0) &&
1678 "Either we should have a defs list, or we should have no defs");
1679 assert((!DL || DL->size() == ActualDefs.size()) &&
1680 "We don't have the same number of defs in the block as on the "
1682 auto ALI = AL->begin();
1683 auto AAI = ActualAccesses.begin();
1684 while (ALI != AL->end() && AAI != ActualAccesses.end()) {
1685 assert(&*ALI == *AAI && "Not the same accesses in the same order");
1689 ActualAccesses.clear();
1691 auto DLI = DL->begin();
1692 auto ADI = ActualDefs.begin();
1693 while (DLI != DL->end() && ADI != ActualDefs.end()) {
1694 assert(&*DLI == *ADI && "Not the same defs in the same order");
1703 /// Verify the domination properties of MemorySSA by checking that each
1704 /// definition dominates all of its uses.
1705 void MemorySSA::verifyDomination(Function &F) const {
1707 for (BasicBlock &B : F) {
1708 // Phi nodes are attached to basic blocks
1709 if (MemoryPhi *MP = getMemoryAccess(&B))
1710 for (const Use &U : MP->uses())
1711 assert(dominates(MP, U) && "Memory PHI does not dominate it's uses");
1713 for (Instruction &I : B) {
1714 MemoryAccess *MD = dyn_cast_or_null<MemoryDef>(getMemoryAccess(&I));
1718 for (const Use &U : MD->uses())
1719 assert(dominates(MD, U) && "Memory Def does not dominate it's uses");
1725 /// Verify the def-use lists in MemorySSA, by verifying that \p Use
1726 /// appears in the use list of \p Def.
1727 void MemorySSA::verifyUseInDefs(MemoryAccess *Def, MemoryAccess *Use) const {
1729 // The live on entry use may cause us to get a NULL def here
1731 assert(isLiveOnEntryDef(Use) &&
1732 "Null def but use not point to live on entry def");
1734 assert(is_contained(Def->users(), Use) &&
1735 "Did not find use in def's use list");
1739 /// Verify the immediate use information, by walking all the memory
1740 /// accesses and verifying that, for each use, it appears in the
1741 /// appropriate def's use list
1742 void MemorySSA::verifyDefUses(Function &F) const {
1743 for (BasicBlock &B : F) {
1744 // Phi nodes are attached to basic blocks
1745 if (MemoryPhi *Phi = getMemoryAccess(&B)) {
1746 assert(Phi->getNumOperands() == static_cast<unsigned>(std::distance(
1747 pred_begin(&B), pred_end(&B))) &&
1748 "Incomplete MemoryPhi Node");
1749 for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I)
1750 verifyUseInDefs(Phi->getIncomingValue(I), Phi);
1753 for (Instruction &I : B) {
1754 if (MemoryUseOrDef *MA = getMemoryAccess(&I)) {
1755 verifyUseInDefs(MA->getDefiningAccess(), MA);
1761 MemoryUseOrDef *MemorySSA::getMemoryAccess(const Instruction *I) const {
1762 return cast_or_null<MemoryUseOrDef>(ValueToMemoryAccess.lookup(I));
1765 MemoryPhi *MemorySSA::getMemoryAccess(const BasicBlock *BB) const {
1766 return cast_or_null<MemoryPhi>(ValueToMemoryAccess.lookup(cast<Value>(BB)));
1769 /// Perform a local numbering on blocks so that instruction ordering can be
1770 /// determined in constant time.
1771 /// TODO: We currently just number in order. If we numbered by N, we could
1772 /// allow at least N-1 sequences of insertBefore or insertAfter (and at least
1773 /// log2(N) sequences of mixed before and after) without needing to invalidate
1775 void MemorySSA::renumberBlock(const BasicBlock *B) const {
1776 // The pre-increment ensures the numbers really start at 1.
1777 unsigned long CurrentNumber = 0;
1778 const AccessList *AL = getBlockAccesses(B);
1779 assert(AL != nullptr && "Asking to renumber an empty block");
1780 for (const auto &I : *AL)
1781 BlockNumbering[&I] = ++CurrentNumber;
1782 BlockNumberingValid.insert(B);
1785 /// Determine, for two memory accesses in the same block,
1786 /// whether \p Dominator dominates \p Dominatee.
1787 /// \returns True if \p Dominator dominates \p Dominatee.
1788 bool MemorySSA::locallyDominates(const MemoryAccess *Dominator,
1789 const MemoryAccess *Dominatee) const {
1790 const BasicBlock *DominatorBlock = Dominator->getBlock();
1792 assert((DominatorBlock == Dominatee->getBlock()) &&
1793 "Asking for local domination when accesses are in different blocks!");
1794 // A node dominates itself.
1795 if (Dominatee == Dominator)
1798 // When Dominatee is defined on function entry, it is not dominated by another
1800 if (isLiveOnEntryDef(Dominatee))
1803 // When Dominator is defined on function entry, it dominates the other memory
1805 if (isLiveOnEntryDef(Dominator))
1808 if (!BlockNumberingValid.count(DominatorBlock))
1809 renumberBlock(DominatorBlock);
1811 unsigned long DominatorNum = BlockNumbering.lookup(Dominator);
1812 // All numbers start with 1
1813 assert(DominatorNum != 0 && "Block was not numbered properly");
1814 unsigned long DominateeNum = BlockNumbering.lookup(Dominatee);
1815 assert(DominateeNum != 0 && "Block was not numbered properly");
1816 return DominatorNum < DominateeNum;
1819 bool MemorySSA::dominates(const MemoryAccess *Dominator,
1820 const MemoryAccess *Dominatee) const {
1821 if (Dominator == Dominatee)
1824 if (isLiveOnEntryDef(Dominatee))
1827 if (Dominator->getBlock() != Dominatee->getBlock())
1828 return DT->dominates(Dominator->getBlock(), Dominatee->getBlock());
1829 return locallyDominates(Dominator, Dominatee);
1832 bool MemorySSA::dominates(const MemoryAccess *Dominator,
1833 const Use &Dominatee) const {
1834 if (MemoryPhi *MP = dyn_cast<MemoryPhi>(Dominatee.getUser())) {
1835 BasicBlock *UseBB = MP->getIncomingBlock(Dominatee);
1836 // The def must dominate the incoming block of the phi.
1837 if (UseBB != Dominator->getBlock())
1838 return DT->dominates(Dominator->getBlock(), UseBB);
1839 // If the UseBB and the DefBB are the same, compare locally.
1840 return locallyDominates(Dominator, cast<MemoryAccess>(Dominatee));
1842 // If it's not a PHI node use, the normal dominates can already handle it.
1843 return dominates(Dominator, cast<MemoryAccess>(Dominatee.getUser()));
1846 const static char LiveOnEntryStr[] = "liveOnEntry";
1848 void MemoryAccess::print(raw_ostream &OS) const {
1849 switch (getValueID()) {
1850 case MemoryPhiVal: return static_cast<const MemoryPhi *>(this)->print(OS);
1851 case MemoryDefVal: return static_cast<const MemoryDef *>(this)->print(OS);
1852 case MemoryUseVal: return static_cast<const MemoryUse *>(this)->print(OS);
1854 llvm_unreachable("invalid value id");
1857 void MemoryDef::print(raw_ostream &OS) const {
1858 MemoryAccess *UO = getDefiningAccess();
1860 OS << getID() << " = MemoryDef(";
1861 if (UO && UO->getID())
1864 OS << LiveOnEntryStr;
1868 void MemoryPhi::print(raw_ostream &OS) const {
1870 OS << getID() << " = MemoryPhi(";
1871 for (const auto &Op : operands()) {
1872 BasicBlock *BB = getIncomingBlock(Op);
1873 MemoryAccess *MA = cast<MemoryAccess>(Op);
1881 OS << BB->getName();
1883 BB->printAsOperand(OS, false);
1885 if (unsigned ID = MA->getID())
1888 OS << LiveOnEntryStr;
1894 void MemoryUse::print(raw_ostream &OS) const {
1895 MemoryAccess *UO = getDefiningAccess();
1897 if (UO && UO->getID())
1900 OS << LiveOnEntryStr;
1904 void MemoryAccess::dump() const {
1905 // Cannot completely remove virtual function even in release mode.
1906 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1912 char MemorySSAPrinterLegacyPass::ID = 0;
1914 MemorySSAPrinterLegacyPass::MemorySSAPrinterLegacyPass() : FunctionPass(ID) {
1915 initializeMemorySSAPrinterLegacyPassPass(*PassRegistry::getPassRegistry());
1918 void MemorySSAPrinterLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
1919 AU.setPreservesAll();
1920 AU.addRequired<MemorySSAWrapperPass>();
1923 bool MemorySSAPrinterLegacyPass::runOnFunction(Function &F) {
1924 auto &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA();
1926 if (VerifyMemorySSA)
1927 MSSA.verifyMemorySSA();
1931 AnalysisKey MemorySSAAnalysis::Key;
1933 MemorySSAAnalysis::Result MemorySSAAnalysis::run(Function &F,
1934 FunctionAnalysisManager &AM) {
1935 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1936 auto &AA = AM.getResult<AAManager>(F);
1937 return MemorySSAAnalysis::Result(llvm::make_unique<MemorySSA>(F, &AA, &DT));
1940 PreservedAnalyses MemorySSAPrinterPass::run(Function &F,
1941 FunctionAnalysisManager &AM) {
1942 OS << "MemorySSA for function: " << F.getName() << "\n";
1943 AM.getResult<MemorySSAAnalysis>(F).getMSSA().print(OS);
1945 return PreservedAnalyses::all();
1948 PreservedAnalyses MemorySSAVerifierPass::run(Function &F,
1949 FunctionAnalysisManager &AM) {
1950 AM.getResult<MemorySSAAnalysis>(F).getMSSA().verifyMemorySSA();
1952 return PreservedAnalyses::all();
1955 char MemorySSAWrapperPass::ID = 0;
1957 MemorySSAWrapperPass::MemorySSAWrapperPass() : FunctionPass(ID) {
1958 initializeMemorySSAWrapperPassPass(*PassRegistry::getPassRegistry());
1961 void MemorySSAWrapperPass::releaseMemory() { MSSA.reset(); }
1963 void MemorySSAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1964 AU.setPreservesAll();
1965 AU.addRequiredTransitive<DominatorTreeWrapperPass>();
1966 AU.addRequiredTransitive<AAResultsWrapperPass>();
1969 bool MemorySSAWrapperPass::runOnFunction(Function &F) {
1970 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1971 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
1972 MSSA.reset(new MemorySSA(F, &AA, &DT));
1976 void MemorySSAWrapperPass::verifyAnalysis() const { MSSA->verifyMemorySSA(); }
1978 void MemorySSAWrapperPass::print(raw_ostream &OS, const Module *M) const {
1982 MemorySSAWalker::MemorySSAWalker(MemorySSA *M) : MSSA(M) {}
1984 MemorySSA::CachingWalker::CachingWalker(MemorySSA *M, AliasAnalysis *A,
1986 : MemorySSAWalker(M), Walker(*M, *A, *D) {}
1988 void MemorySSA::CachingWalker::invalidateInfo(MemoryAccess *MA) {
1989 if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
1990 MUD->resetOptimized();
1993 /// Walk the use-def chains starting at \p MA and find
1994 /// the MemoryAccess that actually clobbers Loc.
1996 /// \returns our clobbering memory access
1997 MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess(
1998 MemoryAccess *StartingAccess, UpwardsMemoryQuery &Q) {
1999 return Walker.findClobber(StartingAccess, Q);
2002 MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess(
2003 MemoryAccess *StartingAccess, const MemoryLocation &Loc) {
2004 if (isa<MemoryPhi>(StartingAccess))
2005 return StartingAccess;
2007 auto *StartingUseOrDef = cast<MemoryUseOrDef>(StartingAccess);
2008 if (MSSA->isLiveOnEntryDef(StartingUseOrDef))
2009 return StartingUseOrDef;
2011 Instruction *I = StartingUseOrDef->getMemoryInst();
2013 // Conservatively, fences are always clobbers, so don't perform the walk if we
2015 if (!ImmutableCallSite(I) && I->isFenceLike())
2016 return StartingUseOrDef;
2018 UpwardsMemoryQuery Q;
2019 Q.OriginalAccess = StartingUseOrDef;
2020 Q.StartingLoc = Loc;
2024 // Unlike the other function, do not walk to the def of a def, because we are
2025 // handed something we already believe is the clobbering access.
2026 MemoryAccess *DefiningAccess = isa<MemoryUse>(StartingUseOrDef)
2027 ? StartingUseOrDef->getDefiningAccess()
2030 MemoryAccess *Clobber = getClobberingMemoryAccess(DefiningAccess, Q);
2031 LLVM_DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
2032 LLVM_DEBUG(dbgs() << *StartingUseOrDef << "\n");
2033 LLVM_DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ");
2034 LLVM_DEBUG(dbgs() << *Clobber << "\n");
2039 MemorySSA::CachingWalker::getClobberingMemoryAccess(MemoryAccess *MA) {
2040 auto *StartingAccess = dyn_cast<MemoryUseOrDef>(MA);
2041 // If this is a MemoryPhi, we can't do anything.
2042 if (!StartingAccess)
2045 // If this is an already optimized use or def, return the optimized result.
2046 // Note: Currently, we store the optimized def result in a separate field,
2047 // since we can't use the defining access.
2048 if (StartingAccess->isOptimized())
2049 return StartingAccess->getOptimized();
2051 const Instruction *I = StartingAccess->getMemoryInst();
2052 UpwardsMemoryQuery Q(I, StartingAccess);
2053 // We can't sanely do anything with a fence, since they conservatively clobber
2054 // all memory, and have no locations to get pointers from to try to
2056 if (!Q.IsCall && I->isFenceLike())
2057 return StartingAccess;
2059 if (isUseTriviallyOptimizableToLiveOnEntry(*MSSA->AA, I)) {
2060 MemoryAccess *LiveOnEntry = MSSA->getLiveOnEntryDef();
2061 StartingAccess->setOptimized(LiveOnEntry);
2062 StartingAccess->setOptimizedAccessType(None);
2066 // Start with the thing we already think clobbers this location
2067 MemoryAccess *DefiningAccess = StartingAccess->getDefiningAccess();
2069 // At this point, DefiningAccess may be the live on entry def.
2070 // If it is, we will not get a better result.
2071 if (MSSA->isLiveOnEntryDef(DefiningAccess)) {
2072 StartingAccess->setOptimized(DefiningAccess);
2073 StartingAccess->setOptimizedAccessType(None);
2074 return DefiningAccess;
2077 MemoryAccess *Result = getClobberingMemoryAccess(DefiningAccess, Q);
2078 LLVM_DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
2079 LLVM_DEBUG(dbgs() << *DefiningAccess << "\n");
2080 LLVM_DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ");
2081 LLVM_DEBUG(dbgs() << *Result << "\n");
2083 StartingAccess->setOptimized(Result);
2084 if (MSSA->isLiveOnEntryDef(Result))
2085 StartingAccess->setOptimizedAccessType(None);
2086 else if (Q.AR == MustAlias)
2087 StartingAccess->setOptimizedAccessType(MustAlias);
2093 DoNothingMemorySSAWalker::getClobberingMemoryAccess(MemoryAccess *MA) {
2094 if (auto *Use = dyn_cast<MemoryUseOrDef>(MA))
2095 return Use->getDefiningAccess();
2099 MemoryAccess *DoNothingMemorySSAWalker::getClobberingMemoryAccess(
2100 MemoryAccess *StartingAccess, const MemoryLocation &) {
2101 if (auto *Use = dyn_cast<MemoryUseOrDef>(StartingAccess))
2102 return Use->getDefiningAccess();
2103 return StartingAccess;
2106 void MemoryPhi::deleteMe(DerivedUser *Self) {
2107 delete static_cast<MemoryPhi *>(Self);
2110 void MemoryDef::deleteMe(DerivedUser *Self) {
2111 delete static_cast<MemoryDef *>(Self);
2114 void MemoryUse::deleteMe(DerivedUser *Self) {
2115 delete static_cast<MemoryUse *>(Self);