Bug Summary

File:lib/Transforms/Utils/MemorySSA.cpp
Warning:line 916, column 37
Called C++ object pointer is null

Annotated Source Code

1//===-- MemorySSA.cpp - Memory SSA Builder---------------------------===//
2//
3// The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------===//
9//
10// This file implements the MemorySSA class.
11//
12//===----------------------------------------------------------------===//
13#include "llvm/Transforms/Utils/MemorySSA.h"
14#include "llvm/ADT/DenseMap.h"
15#include "llvm/ADT/DenseSet.h"
16#include "llvm/ADT/DepthFirstIterator.h"
17#include "llvm/ADT/GraphTraits.h"
18#include "llvm/ADT/PostOrderIterator.h"
19#include "llvm/ADT/STLExtras.h"
20#include "llvm/ADT/SmallBitVector.h"
21#include "llvm/ADT/SmallPtrSet.h"
22#include "llvm/ADT/SmallSet.h"
23#include "llvm/ADT/Statistic.h"
24#include "llvm/Analysis/AliasAnalysis.h"
25#include "llvm/Analysis/CFG.h"
26#include "llvm/Analysis/GlobalsModRef.h"
27#include "llvm/Analysis/IteratedDominanceFrontier.h"
28#include "llvm/Analysis/MemoryLocation.h"
29#include "llvm/Analysis/PHITransAddr.h"
30#include "llvm/IR/AssemblyAnnotationWriter.h"
31#include "llvm/IR/DataLayout.h"
32#include "llvm/IR/Dominators.h"
33#include "llvm/IR/GlobalVariable.h"
34#include "llvm/IR/IRBuilder.h"
35#include "llvm/IR/IntrinsicInst.h"
36#include "llvm/IR/LLVMContext.h"
37#include "llvm/IR/Metadata.h"
38#include "llvm/IR/Module.h"
39#include "llvm/IR/PatternMatch.h"
40#include "llvm/Support/Debug.h"
41#include "llvm/Support/FormattedStream.h"
42#include "llvm/Transforms/Scalar.h"
43#include <algorithm>
44
45#define DEBUG_TYPE"memoryssa" "memoryssa"
46using namespace llvm;
47STATISTIC(NumClobberCacheLookups, "Number of Memory SSA version cache lookups")static llvm::Statistic NumClobberCacheLookups = {"memoryssa",
"NumClobberCacheLookups", "Number of Memory SSA version cache lookups"
, {0}, false}
;
48STATISTIC(NumClobberCacheHits, "Number of Memory SSA version cache hits")static llvm::Statistic NumClobberCacheHits = {"memoryssa", "NumClobberCacheHits"
, "Number of Memory SSA version cache hits", {0}, false}
;
49STATISTIC(NumClobberCacheInserts, "Number of MemorySSA version cache inserts")static llvm::Statistic NumClobberCacheInserts = {"memoryssa",
"NumClobberCacheInserts", "Number of MemorySSA version cache inserts"
, {0}, false}
;
50
51INITIALIZE_PASS_BEGIN(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,static void *initializeMemorySSAWrapperPassPassOnce(PassRegistry
&Registry) {
52 true)static void *initializeMemorySSAWrapperPassPassOnce(PassRegistry
&Registry) {
53INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
54INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry);
55INITIALIZE_PASS_END(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,PassInfo *PI = new PassInfo( "Memory SSA", "memoryssa", &
MemorySSAWrapperPass::ID, PassInfo::NormalCtor_t(callDefaultCtor
<MemorySSAWrapperPass>), false, true); Registry.registerPass
(*PI, true); return PI; } static once_flag InitializeMemorySSAWrapperPassPassFlag
; void llvm::initializeMemorySSAWrapperPassPass(PassRegistry &
Registry) { llvm::call_once(InitializeMemorySSAWrapperPassPassFlag
, initializeMemorySSAWrapperPassPassOnce, std::ref(Registry))
; }
56 true)PassInfo *PI = new PassInfo( "Memory SSA", "memoryssa", &
MemorySSAWrapperPass::ID, PassInfo::NormalCtor_t(callDefaultCtor
<MemorySSAWrapperPass>), false, true); Registry.registerPass
(*PI, true); return PI; } static once_flag InitializeMemorySSAWrapperPassPassFlag
; void llvm::initializeMemorySSAWrapperPassPass(PassRegistry &
Registry) { llvm::call_once(InitializeMemorySSAWrapperPassPassFlag
, initializeMemorySSAWrapperPassPassOnce, std::ref(Registry))
; }
57
58INITIALIZE_PASS_BEGIN(MemorySSAPrinterLegacyPass, "print-memoryssa",static void *initializeMemorySSAPrinterLegacyPassPassOnce(PassRegistry
&Registry) {
59 "Memory SSA Printer", false, false)static void *initializeMemorySSAPrinterLegacyPassPassOnce(PassRegistry
&Registry) {
60INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)initializeMemorySSAWrapperPassPass(Registry);
61INITIALIZE_PASS_END(MemorySSAPrinterLegacyPass, "print-memoryssa",PassInfo *PI = new PassInfo( "Memory SSA Printer", "print-memoryssa"
, &MemorySSAPrinterLegacyPass::ID, PassInfo::NormalCtor_t
(callDefaultCtor<MemorySSAPrinterLegacyPass>), false, false
); Registry.registerPass(*PI, true); return PI; } static once_flag
InitializeMemorySSAPrinterLegacyPassPassFlag; void llvm::initializeMemorySSAPrinterLegacyPassPass
(PassRegistry &Registry) { llvm::call_once(InitializeMemorySSAPrinterLegacyPassPassFlag
, initializeMemorySSAPrinterLegacyPassPassOnce, std::ref(Registry
)); }
62 "Memory SSA Printer", false, false)PassInfo *PI = new PassInfo( "Memory SSA Printer", "print-memoryssa"
, &MemorySSAPrinterLegacyPass::ID, PassInfo::NormalCtor_t
(callDefaultCtor<MemorySSAPrinterLegacyPass>), false, false
); Registry.registerPass(*PI, true); return PI; } static once_flag
InitializeMemorySSAPrinterLegacyPassPassFlag; void llvm::initializeMemorySSAPrinterLegacyPassPass
(PassRegistry &Registry) { llvm::call_once(InitializeMemorySSAPrinterLegacyPassPassFlag
, initializeMemorySSAPrinterLegacyPassPassOnce, std::ref(Registry
)); }
63
64static cl::opt<unsigned> MaxCheckLimit(
65 "memssa-check-limit", cl::Hidden, cl::init(100),
66 cl::desc("The maximum number of stores/phis MemorySSA"
67 "will consider trying to walk past (default = 100)"));
68
69static cl::opt<bool>
70 VerifyMemorySSA("verify-memoryssa", cl::init(false), cl::Hidden,
71 cl::desc("Verify MemorySSA in legacy printer pass."));
72
73namespace llvm {
74/// \brief An assembly annotator class to print Memory SSA information in
75/// comments.
76class MemorySSAAnnotatedWriter : public AssemblyAnnotationWriter {
77 friend class MemorySSA;
78 const MemorySSA *MSSA;
79
80public:
81 MemorySSAAnnotatedWriter(const MemorySSA *M) : MSSA(M) {}
82
83 virtual void emitBasicBlockStartAnnot(const BasicBlock *BB,
84 formatted_raw_ostream &OS) {
85 if (MemoryAccess *MA = MSSA->getMemoryAccess(BB))
86 OS << "; " << *MA << "\n";
87 }
88
89 virtual void emitInstructionAnnot(const Instruction *I,
90 formatted_raw_ostream &OS) {
91 if (MemoryAccess *MA = MSSA->getMemoryAccess(I))
92 OS << "; " << *MA << "\n";
93 }
94};
95}
96
97namespace {
98/// Our current alias analysis API differentiates heavily between calls and
99/// non-calls, and functions called on one usually assert on the other.
100/// This class encapsulates the distinction to simplify other code that wants
101/// "Memory affecting instructions and related data" to use as a key.
102/// For example, this class is used as a densemap key in the use optimizer.
103class MemoryLocOrCall {
104public:
105 MemoryLocOrCall() : IsCall(false) {}
106 MemoryLocOrCall(MemoryUseOrDef *MUD)
107 : MemoryLocOrCall(MUD->getMemoryInst()) {}
108 MemoryLocOrCall(const MemoryUseOrDef *MUD)
109 : MemoryLocOrCall(MUD->getMemoryInst()) {}
110
111 MemoryLocOrCall(Instruction *Inst) {
112 if (ImmutableCallSite(Inst)) {
113 IsCall = true;
114 CS = ImmutableCallSite(Inst);
115 } else {
116 IsCall = false;
117 // There is no such thing as a memorylocation for a fence inst, and it is
118 // unique in that regard.
119 if (!isa<FenceInst>(Inst))
120 Loc = MemoryLocation::get(Inst);
121 }
122 }
123
124 explicit MemoryLocOrCall(const MemoryLocation &Loc)
125 : IsCall(false), Loc(Loc) {}
126
127 bool IsCall;
128 ImmutableCallSite getCS() const {
129 assert(IsCall)((IsCall) ? static_cast<void> (0) : __assert_fail ("IsCall"
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 129, __PRETTY_FUNCTION__))
;
130 return CS;
131 }
132 MemoryLocation getLoc() const {
133 assert(!IsCall)((!IsCall) ? static_cast<void> (0) : __assert_fail ("!IsCall"
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 133, __PRETTY_FUNCTION__))
;
134 return Loc;
135 }
136
137 bool operator==(const MemoryLocOrCall &Other) const {
138 if (IsCall != Other.IsCall)
139 return false;
140
141 if (IsCall)
142 return CS.getCalledValue() == Other.CS.getCalledValue();
143 return Loc == Other.Loc;
144 }
145
146private:
147 union {
148 ImmutableCallSite CS;
149 MemoryLocation Loc;
150 };
151};
152}
153
154namespace llvm {
155template <> struct DenseMapInfo<MemoryLocOrCall> {
156 static inline MemoryLocOrCall getEmptyKey() {
157 return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getEmptyKey());
158 }
159 static inline MemoryLocOrCall getTombstoneKey() {
160 return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getTombstoneKey());
161 }
162 static unsigned getHashValue(const MemoryLocOrCall &MLOC) {
163 if (MLOC.IsCall)
164 return hash_combine(MLOC.IsCall,
165 DenseMapInfo<const Value *>::getHashValue(
166 MLOC.getCS().getCalledValue()));
167 return hash_combine(
168 MLOC.IsCall, DenseMapInfo<MemoryLocation>::getHashValue(MLOC.getLoc()));
169 }
170 static bool isEqual(const MemoryLocOrCall &LHS, const MemoryLocOrCall &RHS) {
171 return LHS == RHS;
172 }
173};
174
175enum class Reorderability { Always, IfNoAlias, Never };
176
177/// This does one-way checks to see if Use could theoretically be hoisted above
178/// MayClobber. This will not check the other way around.
179///
180/// This assumes that, for the purposes of MemorySSA, Use comes directly after
181/// MayClobber, with no potentially clobbering operations in between them.
182/// (Where potentially clobbering ops are memory barriers, aliased stores, etc.)
183static Reorderability getLoadReorderability(const LoadInst *Use,
184 const LoadInst *MayClobber) {
185 bool VolatileUse = Use->isVolatile();
186 bool VolatileClobber = MayClobber->isVolatile();
187 // Volatile operations may never be reordered with other volatile operations.
188 if (VolatileUse && VolatileClobber)
189 return Reorderability::Never;
190
191 // The lang ref allows reordering of volatile and non-volatile operations.
192 // Whether an aliasing nonvolatile load and volatile load can be reordered,
193 // though, is ambiguous. Because it may not be best to exploit this ambiguity,
194 // we only allow volatile/non-volatile reordering if the volatile and
195 // non-volatile operations don't alias.
196 Reorderability Result = VolatileUse || VolatileClobber
197 ? Reorderability::IfNoAlias
198 : Reorderability::Always;
199
200 // If a load is seq_cst, it cannot be moved above other loads. If its ordering
201 // is weaker, it can be moved above other loads. We just need to be sure that
202 // MayClobber isn't an acquire load, because loads can't be moved above
203 // acquire loads.
204 //
205 // Note that this explicitly *does* allow the free reordering of monotonic (or
206 // weaker) loads of the same address.
207 bool SeqCstUse = Use->getOrdering() == AtomicOrdering::SequentiallyConsistent;
208 bool MayClobberIsAcquire = isAtLeastOrStrongerThan(MayClobber->getOrdering(),
209 AtomicOrdering::Acquire);
210 if (SeqCstUse || MayClobberIsAcquire)
211 return Reorderability::Never;
212 return Result;
213}
214
215static bool instructionClobbersQuery(MemoryDef *MD,
216 const MemoryLocation &UseLoc,
217 const Instruction *UseInst,
218 AliasAnalysis &AA) {
219 Instruction *DefInst = MD->getMemoryInst();
220 assert(DefInst && "Defining instruction not actually an instruction")((DefInst && "Defining instruction not actually an instruction"
) ? static_cast<void> (0) : __assert_fail ("DefInst && \"Defining instruction not actually an instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 220, __PRETTY_FUNCTION__))
;
221
222 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(DefInst)) {
223 // These intrinsics will show up as affecting memory, but they are just
224 // markers.
225 switch (II->getIntrinsicID()) {
226 case Intrinsic::lifetime_start:
227 case Intrinsic::lifetime_end:
228 case Intrinsic::invariant_start:
229 case Intrinsic::invariant_end:
230 case Intrinsic::assume:
231 return false;
232 default:
233 break;
234 }
235 }
236
237 ImmutableCallSite UseCS(UseInst);
238 if (UseCS) {
239 ModRefInfo I = AA.getModRefInfo(DefInst, UseCS);
240 return I != MRI_NoModRef;
241 }
242
243 if (auto *DefLoad = dyn_cast<LoadInst>(DefInst)) {
244 if (auto *UseLoad = dyn_cast<LoadInst>(UseInst)) {
245 switch (getLoadReorderability(UseLoad, DefLoad)) {
246 case Reorderability::Always:
247 return false;
248 case Reorderability::Never:
249 return true;
250 case Reorderability::IfNoAlias:
251 return !AA.isNoAlias(UseLoc, MemoryLocation::get(DefLoad));
252 }
253 }
254 }
255
256 return AA.getModRefInfo(DefInst, UseLoc) & MRI_Mod;
257}
258
259static bool instructionClobbersQuery(MemoryDef *MD, const MemoryUseOrDef *MU,
260 const MemoryLocOrCall &UseMLOC,
261 AliasAnalysis &AA) {
262 // FIXME: This is a temporary hack to allow a single instructionClobbersQuery
263 // to exist while MemoryLocOrCall is pushed through places.
264 if (UseMLOC.IsCall)
265 return instructionClobbersQuery(MD, MemoryLocation(), MU->getMemoryInst(),
266 AA);
267 return instructionClobbersQuery(MD, UseMLOC.getLoc(), MU->getMemoryInst(),
268 AA);
269}
270
271// Return true when MD may alias MU, return false otherwise.
272bool defClobbersUseOrDef(MemoryDef *MD, const MemoryUseOrDef *MU,
273 AliasAnalysis &AA) {
274 return instructionClobbersQuery(MD, MU, MemoryLocOrCall(MU), AA);
275}
276}
277
278namespace {
279struct UpwardsMemoryQuery {
280 // True if our original query started off as a call
281 bool IsCall;
282 // The pointer location we started the query with. This will be empty if
283 // IsCall is true.
284 MemoryLocation StartingLoc;
285 // This is the instruction we were querying about.
286 const Instruction *Inst;
287 // The MemoryAccess we actually got called with, used to test local domination
288 const MemoryAccess *OriginalAccess;
289
290 UpwardsMemoryQuery()
291 : IsCall(false), Inst(nullptr), OriginalAccess(nullptr) {}
292
293 UpwardsMemoryQuery(const Instruction *Inst, const MemoryAccess *Access)
294 : IsCall(ImmutableCallSite(Inst)), Inst(Inst), OriginalAccess(Access) {
295 if (!IsCall)
296 StartingLoc = MemoryLocation::get(Inst);
297 }
298};
299
300static bool lifetimeEndsAt(MemoryDef *MD, const MemoryLocation &Loc,
301 AliasAnalysis &AA) {
302 Instruction *Inst = MD->getMemoryInst();
303 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
304 switch (II->getIntrinsicID()) {
305 case Intrinsic::lifetime_start:
306 case Intrinsic::lifetime_end:
307 return AA.isMustAlias(MemoryLocation(II->getArgOperand(1)), Loc);
308 default:
309 return false;
310 }
311 }
312 return false;
313}
314
315static bool isUseTriviallyOptimizableToLiveOnEntry(AliasAnalysis &AA,
316 const Instruction *I) {
317 // If the memory can't be changed, then loads of the memory can't be
318 // clobbered.
319 //
320 // FIXME: We should handle invariant groups, as well. It's a bit harder,
321 // because we need to pay close attention to invariant group barriers.
322 return isa<LoadInst>(I) && (I->getMetadata(LLVMContext::MD_invariant_load) ||
323 AA.pointsToConstantMemory(I));
324}
325
326/// Cache for our caching MemorySSA walker.
327class WalkerCache {
328 DenseMap<ConstMemoryAccessPair, MemoryAccess *> Accesses;
329 DenseMap<const MemoryAccess *, MemoryAccess *> Calls;
330
331public:
332 MemoryAccess *lookup(const MemoryAccess *MA, const MemoryLocation &Loc,
333 bool IsCall) const {
334 ++NumClobberCacheLookups;
335 MemoryAccess *R = IsCall ? Calls.lookup(MA) : Accesses.lookup({MA, Loc});
336 if (R)
337 ++NumClobberCacheHits;
338 return R;
339 }
340
341 bool insert(const MemoryAccess *MA, MemoryAccess *To,
342 const MemoryLocation &Loc, bool IsCall) {
343 // This is fine for Phis, since there are times where we can't optimize
344 // them. Making a def its own clobber is never correct, though.
345 assert((MA != To || isa<MemoryPhi>(MA)) &&(((MA != To || isa<MemoryPhi>(MA)) && "Something can't clobber itself!"
) ? static_cast<void> (0) : __assert_fail ("(MA != To || isa<MemoryPhi>(MA)) && \"Something can't clobber itself!\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 346, __PRETTY_FUNCTION__))
346 "Something can't clobber itself!")(((MA != To || isa<MemoryPhi>(MA)) && "Something can't clobber itself!"
) ? static_cast<void> (0) : __assert_fail ("(MA != To || isa<MemoryPhi>(MA)) && \"Something can't clobber itself!\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 346, __PRETTY_FUNCTION__))
;
347
348 ++NumClobberCacheInserts;
349 bool Inserted;
350 if (IsCall)
351 Inserted = Calls.insert({MA, To}).second;
352 else
353 Inserted = Accesses.insert({{MA, Loc}, To}).second;
354
355 return Inserted;
356 }
357
358 bool remove(const MemoryAccess *MA, const MemoryLocation &Loc, bool IsCall) {
359 return IsCall ? Calls.erase(MA) : Accesses.erase({MA, Loc});
360 }
361
362 void clear() {
363 Accesses.clear();
364 Calls.clear();
365 }
366
367 bool contains(const MemoryAccess *MA) const {
368 for (auto &P : Accesses)
369 if (P.first.first == MA || P.second == MA)
370 return true;
371 for (auto &P : Calls)
372 if (P.first == MA || P.second == MA)
373 return true;
374 return false;
375 }
376};
377
378/// Walks the defining uses of MemoryDefs. Stops after we hit something that has
379/// no defining use (e.g. a MemoryPhi or liveOnEntry). Note that, when comparing
380/// against a null def_chain_iterator, this will compare equal only after
381/// walking said Phi/liveOnEntry.
382struct def_chain_iterator
383 : public iterator_facade_base<def_chain_iterator, std::forward_iterator_tag,
384 MemoryAccess *> {
385 def_chain_iterator() : MA(nullptr) {}
386 def_chain_iterator(MemoryAccess *MA) : MA(MA) {}
387
388 MemoryAccess *operator*() const { return MA; }
389
390 def_chain_iterator &operator++() {
391 // N.B. liveOnEntry has a null defining access.
392 if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA))
393 MA = MUD->getDefiningAccess();
394 else
395 MA = nullptr;
396 return *this;
397 }
398
399 bool operator==(const def_chain_iterator &O) const { return MA == O.MA; }
400
401private:
402 MemoryAccess *MA;
403};
404
405static iterator_range<def_chain_iterator>
406def_chain(MemoryAccess *MA, MemoryAccess *UpTo = nullptr) {
407#ifdef EXPENSIVE_CHECKS
408 assert((!UpTo || find(def_chain(MA), UpTo) != def_chain_iterator()) &&(((!UpTo || find(def_chain(MA), UpTo) != def_chain_iterator()
) && "UpTo isn't in the def chain!") ? static_cast<
void> (0) : __assert_fail ("(!UpTo || find(def_chain(MA), UpTo) != def_chain_iterator()) && \"UpTo isn't in the def chain!\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 409, __PRETTY_FUNCTION__))
409 "UpTo isn't in the def chain!")(((!UpTo || find(def_chain(MA), UpTo) != def_chain_iterator()
) && "UpTo isn't in the def chain!") ? static_cast<
void> (0) : __assert_fail ("(!UpTo || find(def_chain(MA), UpTo) != def_chain_iterator()) && \"UpTo isn't in the def chain!\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 409, __PRETTY_FUNCTION__))
;
410#endif
411 return make_range(def_chain_iterator(MA), def_chain_iterator(UpTo));
412}
413
414/// Verifies that `Start` is clobbered by `ClobberAt`, and that nothing
415/// inbetween `Start` and `ClobberAt` can clobbers `Start`.
416///
417/// This is meant to be as simple and self-contained as possible. Because it
418/// uses no cache, etc., it can be relatively expensive.
419///
420/// \param Start The MemoryAccess that we want to walk from.
421/// \param ClobberAt A clobber for Start.
422/// \param StartLoc The MemoryLocation for Start.
423/// \param MSSA The MemorySSA isntance that Start and ClobberAt belong to.
424/// \param Query The UpwardsMemoryQuery we used for our search.
425/// \param AA The AliasAnalysis we used for our search.
426static void LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__))
427checkClobberSanity(MemoryAccess *Start, MemoryAccess *ClobberAt,
428 const MemoryLocation &StartLoc, const MemorySSA &MSSA,
429 const UpwardsMemoryQuery &Query, AliasAnalysis &AA) {
430 assert(MSSA.dominates(ClobberAt, Start) && "Clobber doesn't dominate start?")((MSSA.dominates(ClobberAt, Start) && "Clobber doesn't dominate start?"
) ? static_cast<void> (0) : __assert_fail ("MSSA.dominates(ClobberAt, Start) && \"Clobber doesn't dominate start?\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 430, __PRETTY_FUNCTION__))
;
431
432 if (MSSA.isLiveOnEntryDef(Start)) {
433 assert(MSSA.isLiveOnEntryDef(ClobberAt) &&((MSSA.isLiveOnEntryDef(ClobberAt) && "liveOnEntry must clobber itself"
) ? static_cast<void> (0) : __assert_fail ("MSSA.isLiveOnEntryDef(ClobberAt) && \"liveOnEntry must clobber itself\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 434, __PRETTY_FUNCTION__))
434 "liveOnEntry must clobber itself")((MSSA.isLiveOnEntryDef(ClobberAt) && "liveOnEntry must clobber itself"
) ? static_cast<void> (0) : __assert_fail ("MSSA.isLiveOnEntryDef(ClobberAt) && \"liveOnEntry must clobber itself\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 434, __PRETTY_FUNCTION__))
;
435 return;
436 }
437
438 bool FoundClobber = false;
439 DenseSet<MemoryAccessPair> VisitedPhis;
440 SmallVector<MemoryAccessPair, 8> Worklist;
441 Worklist.emplace_back(Start, StartLoc);
442 // Walk all paths from Start to ClobberAt, while looking for clobbers. If one
443 // is found, complain.
444 while (!Worklist.empty()) {
445 MemoryAccessPair MAP = Worklist.pop_back_val();
446 // All we care about is that nothing from Start to ClobberAt clobbers Start.
447 // We learn nothing from revisiting nodes.
448 if (!VisitedPhis.insert(MAP).second)
449 continue;
450
451 for (MemoryAccess *MA : def_chain(MAP.first)) {
452 if (MA == ClobberAt) {
453 if (auto *MD = dyn_cast<MemoryDef>(MA)) {
454 // instructionClobbersQuery isn't essentially free, so don't use `|=`,
455 // since it won't let us short-circuit.
456 //
457 // Also, note that this can't be hoisted out of the `Worklist` loop,
458 // since MD may only act as a clobber for 1 of N MemoryLocations.
459 FoundClobber =
460 FoundClobber || MSSA.isLiveOnEntryDef(MD) ||
461 instructionClobbersQuery(MD, MAP.second, Query.Inst, AA);
462 }
463 break;
464 }
465
466 // We should never hit liveOnEntry, unless it's the clobber.
467 assert(!MSSA.isLiveOnEntryDef(MA) && "Hit liveOnEntry before clobber?")((!MSSA.isLiveOnEntryDef(MA) && "Hit liveOnEntry before clobber?"
) ? static_cast<void> (0) : __assert_fail ("!MSSA.isLiveOnEntryDef(MA) && \"Hit liveOnEntry before clobber?\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 467, __PRETTY_FUNCTION__))
;
468
469 if (auto *MD = dyn_cast<MemoryDef>(MA)) {
470 (void)MD;
471 assert(!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA) &&((!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA) &&
"Found clobber before reaching ClobberAt!") ? static_cast<
void> (0) : __assert_fail ("!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA) && \"Found clobber before reaching ClobberAt!\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 472, __PRETTY_FUNCTION__))
472 "Found clobber before reaching ClobberAt!")((!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA) &&
"Found clobber before reaching ClobberAt!") ? static_cast<
void> (0) : __assert_fail ("!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA) && \"Found clobber before reaching ClobberAt!\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 472, __PRETTY_FUNCTION__))
;
473 continue;
474 }
475
476 assert(isa<MemoryPhi>(MA))((isa<MemoryPhi>(MA)) ? static_cast<void> (0) : __assert_fail
("isa<MemoryPhi>(MA)", "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 476, __PRETTY_FUNCTION__))
;
477 Worklist.append(upward_defs_begin({MA, MAP.second}), upward_defs_end());
478 }
479 }
480
481 // If ClobberAt is a MemoryPhi, we can assume something above it acted as a
482 // clobber. Otherwise, `ClobberAt` should've acted as a clobber at some point.
483 assert((isa<MemoryPhi>(ClobberAt) || FoundClobber) &&(((isa<MemoryPhi>(ClobberAt) || FoundClobber) &&
"ClobberAt never acted as a clobber") ? static_cast<void>
(0) : __assert_fail ("(isa<MemoryPhi>(ClobberAt) || FoundClobber) && \"ClobberAt never acted as a clobber\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 484, __PRETTY_FUNCTION__))
484 "ClobberAt never acted as a clobber")(((isa<MemoryPhi>(ClobberAt) || FoundClobber) &&
"ClobberAt never acted as a clobber") ? static_cast<void>
(0) : __assert_fail ("(isa<MemoryPhi>(ClobberAt) || FoundClobber) && \"ClobberAt never acted as a clobber\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 484, __PRETTY_FUNCTION__))
;
485}
486
487/// Our algorithm for walking (and trying to optimize) clobbers, all wrapped up
488/// in one class.
489class ClobberWalker {
490 /// Save a few bytes by using unsigned instead of size_t.
491 using ListIndex = unsigned;
492
493 /// Represents a span of contiguous MemoryDefs, potentially ending in a
494 /// MemoryPhi.
495 struct DefPath {
496 MemoryLocation Loc;
497 // Note that, because we always walk in reverse, Last will always dominate
498 // First. Also note that First and Last are inclusive.
499 MemoryAccess *First;
500 MemoryAccess *Last;
501 Optional<ListIndex> Previous;
502
503 DefPath(const MemoryLocation &Loc, MemoryAccess *First, MemoryAccess *Last,
504 Optional<ListIndex> Previous)
505 : Loc(Loc), First(First), Last(Last), Previous(Previous) {}
506
507 DefPath(const MemoryLocation &Loc, MemoryAccess *Init,
508 Optional<ListIndex> Previous)
509 : DefPath(Loc, Init, Init, Previous) {}
510 };
511
512 const MemorySSA &MSSA;
513 AliasAnalysis &AA;
514 DominatorTree &DT;
515 WalkerCache &WC;
516 UpwardsMemoryQuery *Query;
517 bool UseCache;
518
519 // Phi optimization bookkeeping
520 SmallVector<DefPath, 32> Paths;
521 DenseSet<ConstMemoryAccessPair> VisitedPhis;
522 DenseMap<const BasicBlock *, MemoryAccess *> WalkTargetCache;
523
524 void setUseCache(bool Use) { UseCache = Use; }
525 bool shouldIgnoreCache() const {
526 // UseCache will only be false when we're debugging, or when expensive
527 // checks are enabled. In either case, we don't care deeply about speed.
528 return LLVM_UNLIKELY(!UseCache)__builtin_expect((bool)(!UseCache), false);
529 }
530
531 void addCacheEntry(const MemoryAccess *What, MemoryAccess *To,
532 const MemoryLocation &Loc) const {
533// EXPENSIVE_CHECKS because most of these queries are redundant.
534#ifdef EXPENSIVE_CHECKS
535 assert(MSSA.dominates(To, What))((MSSA.dominates(To, What)) ? static_cast<void> (0) : __assert_fail
("MSSA.dominates(To, What)", "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 535, __PRETTY_FUNCTION__))
;
536#endif
537 if (shouldIgnoreCache())
538 return;
539 WC.insert(What, To, Loc, Query->IsCall);
540 }
541
542 MemoryAccess *lookupCache(const MemoryAccess *MA, const MemoryLocation &Loc) {
543 return shouldIgnoreCache() ? nullptr : WC.lookup(MA, Loc, Query->IsCall);
544 }
545
546 void cacheDefPath(const DefPath &DN, MemoryAccess *Target) const {
547 if (shouldIgnoreCache())
548 return;
549
550 for (MemoryAccess *MA : def_chain(DN.First, DN.Last))
551 addCacheEntry(MA, Target, DN.Loc);
552
553 // DefPaths only express the path we walked. So, DN.Last could either be a
554 // thing we want to cache, or not.
555 if (DN.Last != Target)
556 addCacheEntry(DN.Last, Target, DN.Loc);
557 }
558
559 /// Find the nearest def or phi that `From` can legally be optimized to.
560 ///
561 /// FIXME: Deduplicate this with MSSA::findDominatingDef. Ideally, MSSA should
562 /// keep track of this information for us, and allow us O(1) lookups of this
563 /// info.
564 MemoryAccess *getWalkTarget(const MemoryPhi *From) {
565 assert(From->getNumOperands() && "Phi with no operands?")((From->getNumOperands() && "Phi with no operands?"
) ? static_cast<void> (0) : __assert_fail ("From->getNumOperands() && \"Phi with no operands?\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 565, __PRETTY_FUNCTION__))
;
566
567 BasicBlock *BB = From->getBlock();
568 auto At = WalkTargetCache.find(BB);
569 if (At != WalkTargetCache.end())
570 return At->second;
571
572 SmallVector<const BasicBlock *, 8> ToCache;
573 ToCache.push_back(BB);
574
575 MemoryAccess *Result = MSSA.getLiveOnEntryDef();
576 DomTreeNode *Node = DT.getNode(BB);
577 while ((Node = Node->getIDom())) {
578 auto At = WalkTargetCache.find(BB);
579 if (At != WalkTargetCache.end()) {
580 Result = At->second;
581 break;
582 }
583
584 auto *Accesses = MSSA.getBlockAccesses(Node->getBlock());
585 if (Accesses) {
586 auto Iter = find_if(reverse(*Accesses), [](const MemoryAccess &MA) {
587 return !isa<MemoryUse>(MA);
588 });
589 if (Iter != Accesses->rend()) {
590 Result = const_cast<MemoryAccess *>(&*Iter);
591 break;
592 }
593 }
594
595 ToCache.push_back(Node->getBlock());
596 }
597
598 for (const BasicBlock *BB : ToCache)
599 WalkTargetCache.insert({BB, Result});
600 return Result;
601 }
602
603 /// Result of calling walkToPhiOrClobber.
604 struct UpwardsWalkResult {
605 /// The "Result" of the walk. Either a clobber, the last thing we walked, or
606 /// both.
607 MemoryAccess *Result;
608 bool IsKnownClobber;
609 bool FromCache;
610 };
611
612 /// Walk to the next Phi or Clobber in the def chain starting at Desc.Last.
613 /// This will update Desc.Last as it walks. It will (optionally) also stop at
614 /// StopAt.
615 ///
616 /// This does not test for whether StopAt is a clobber
617 UpwardsWalkResult walkToPhiOrClobber(DefPath &Desc,
618 MemoryAccess *StopAt = nullptr) {
619 assert(!isa<MemoryUse>(Desc.Last) && "Uses don't exist in my world")((!isa<MemoryUse>(Desc.Last) && "Uses don't exist in my world"
) ? static_cast<void> (0) : __assert_fail ("!isa<MemoryUse>(Desc.Last) && \"Uses don't exist in my world\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 619, __PRETTY_FUNCTION__))
;
620
621 for (MemoryAccess *Current : def_chain(Desc.Last)) {
622 Desc.Last = Current;
623 if (Current == StopAt)
624 return {Current, false, false};
625
626 if (auto *MD = dyn_cast<MemoryDef>(Current))
627 if (MSSA.isLiveOnEntryDef(MD) ||
628 instructionClobbersQuery(MD, Desc.Loc, Query->Inst, AA))
629 return {MD, true, false};
630
631 // Cache checks must be done last, because if Current is a clobber, the
632 // cache will contain the clobber for Current.
633 if (MemoryAccess *MA = lookupCache(Current, Desc.Loc))
634 return {MA, true, true};
635 }
636
637 assert(isa<MemoryPhi>(Desc.Last) &&((isa<MemoryPhi>(Desc.Last) && "Ended at a non-clobber that's not a phi?"
) ? static_cast<void> (0) : __assert_fail ("isa<MemoryPhi>(Desc.Last) && \"Ended at a non-clobber that's not a phi?\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 638, __PRETTY_FUNCTION__))
638 "Ended at a non-clobber that's not a phi?")((isa<MemoryPhi>(Desc.Last) && "Ended at a non-clobber that's not a phi?"
) ? static_cast<void> (0) : __assert_fail ("isa<MemoryPhi>(Desc.Last) && \"Ended at a non-clobber that's not a phi?\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 638, __PRETTY_FUNCTION__))
;
639 return {Desc.Last, false, false};
640 }
641
642 void addSearches(MemoryPhi *Phi, SmallVectorImpl<ListIndex> &PausedSearches,
643 ListIndex PriorNode) {
644 auto UpwardDefs = make_range(upward_defs_begin({Phi, Paths[PriorNode].Loc}),
645 upward_defs_end());
646 for (const MemoryAccessPair &P : UpwardDefs) {
647 PausedSearches.push_back(Paths.size());
648 Paths.emplace_back(P.second, P.first, PriorNode);
649 }
650 }
651
652 /// Represents a search that terminated after finding a clobber. This clobber
653 /// may or may not be present in the path of defs from LastNode..SearchStart,
654 /// since it may have been retrieved from cache.
655 struct TerminatedPath {
656 MemoryAccess *Clobber;
657 ListIndex LastNode;
658 };
659
660 /// Get an access that keeps us from optimizing to the given phi.
661 ///
662 /// PausedSearches is an array of indices into the Paths array. Its incoming
663 /// value is the indices of searches that stopped at the last phi optimization
664 /// target. It's left in an unspecified state.
665 ///
666 /// If this returns None, NewPaused is a vector of searches that terminated
667 /// at StopWhere. Otherwise, NewPaused is left in an unspecified state.
668 Optional<TerminatedPath>
669 getBlockingAccess(MemoryAccess *StopWhere,
670 SmallVectorImpl<ListIndex> &PausedSearches,
671 SmallVectorImpl<ListIndex> &NewPaused,
672 SmallVectorImpl<TerminatedPath> &Terminated) {
673 assert(!PausedSearches.empty() && "No searches to continue?")((!PausedSearches.empty() && "No searches to continue?"
) ? static_cast<void> (0) : __assert_fail ("!PausedSearches.empty() && \"No searches to continue?\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 673, __PRETTY_FUNCTION__))
;
674
675 // BFS vs DFS really doesn't make a difference here, so just do a DFS with
676 // PausedSearches as our stack.
677 while (!PausedSearches.empty()) {
678 ListIndex PathIndex = PausedSearches.pop_back_val();
679 DefPath &Node = Paths[PathIndex];
680
681 // If we've already visited this path with this MemoryLocation, we don't
682 // need to do so again.
683 //
684 // NOTE: That we just drop these paths on the ground makes caching
685 // behavior sporadic. e.g. given a diamond:
686 // A
687 // B C
688 // D
689 //
690 // ...If we walk D, B, A, C, we'll only cache the result of phi
691 // optimization for A, B, and D; C will be skipped because it dies here.
692 // This arguably isn't the worst thing ever, since:
693 // - We generally query things in a top-down order, so if we got below D
694 // without needing cache entries for {C, MemLoc}, then chances are
695 // that those cache entries would end up ultimately unused.
696 // - We still cache things for A, so C only needs to walk up a bit.
697 // If this behavior becomes problematic, we can fix without a ton of extra
698 // work.
699 if (!VisitedPhis.insert({Node.Last, Node.Loc}).second)
700 continue;
701
702 UpwardsWalkResult Res = walkToPhiOrClobber(Node, /*StopAt=*/StopWhere);
703 if (Res.IsKnownClobber) {
704 assert(Res.Result != StopWhere || Res.FromCache)((Res.Result != StopWhere || Res.FromCache) ? static_cast<
void> (0) : __assert_fail ("Res.Result != StopWhere || Res.FromCache"
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 704, __PRETTY_FUNCTION__))
;
705 // If this wasn't a cache hit, we hit a clobber when walking. That's a
706 // failure.
707 TerminatedPath Term{Res.Result, PathIndex};
708 if (!Res.FromCache || !MSSA.dominates(Res.Result, StopWhere))
709 return Term;
710
711 // Otherwise, it's a valid thing to potentially optimize to.
712 Terminated.push_back(Term);
713 continue;
714 }
715
716 if (Res.Result == StopWhere) {
717 // We've hit our target. Save this path off for if we want to continue
718 // walking.
719 NewPaused.push_back(PathIndex);
720 continue;
721 }
722
723 assert(!MSSA.isLiveOnEntryDef(Res.Result) && "liveOnEntry is a clobber")((!MSSA.isLiveOnEntryDef(Res.Result) && "liveOnEntry is a clobber"
) ? static_cast<void> (0) : __assert_fail ("!MSSA.isLiveOnEntryDef(Res.Result) && \"liveOnEntry is a clobber\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 723, __PRETTY_FUNCTION__))
;
724 addSearches(cast<MemoryPhi>(Res.Result), PausedSearches, PathIndex);
725 }
726
727 return None;
728 }
729
730 template <typename T, typename Walker>
731 struct generic_def_path_iterator
732 : public iterator_facade_base<generic_def_path_iterator<T, Walker>,
733 std::forward_iterator_tag, T *> {
734 generic_def_path_iterator() : W(nullptr), N(None) {}
735 generic_def_path_iterator(Walker *W, ListIndex N) : W(W), N(N) {}
736
737 T &operator*() const { return curNode(); }
738
739 generic_def_path_iterator &operator++() {
740 N = curNode().Previous;
741 return *this;
742 }
743
744 bool operator==(const generic_def_path_iterator &O) const {
745 if (N.hasValue() != O.N.hasValue())
746 return false;
747 return !N.hasValue() || *N == *O.N;
748 }
749
750 private:
751 T &curNode() const { return W->Paths[*N]; }
752
753 Walker *W;
754 Optional<ListIndex> N;
755 };
756
757 using def_path_iterator = generic_def_path_iterator<DefPath, ClobberWalker>;
758 using const_def_path_iterator =
759 generic_def_path_iterator<const DefPath, const ClobberWalker>;
760
761 iterator_range<def_path_iterator> def_path(ListIndex From) {
762 return make_range(def_path_iterator(this, From), def_path_iterator());
763 }
764
765 iterator_range<const_def_path_iterator> const_def_path(ListIndex From) const {
766 return make_range(const_def_path_iterator(this, From),
767 const_def_path_iterator());
768 }
769
770 struct OptznResult {
771 /// The path that contains our result.
772 TerminatedPath PrimaryClobber;
773 /// The paths that we can legally cache back from, but that aren't
774 /// necessarily the result of the Phi optimization.
775 SmallVector<TerminatedPath, 4> OtherClobbers;
776 };
777
778 ListIndex defPathIndex(const DefPath &N) const {
779 // The assert looks nicer if we don't need to do &N
780 const DefPath *NP = &N;
781 assert(!Paths.empty() && NP >= &Paths.front() && NP <= &Paths.back() &&((!Paths.empty() && NP >= &Paths.front() &&
NP <= &Paths.back() && "Out of bounds DefPath!"
) ? static_cast<void> (0) : __assert_fail ("!Paths.empty() && NP >= &Paths.front() && NP <= &Paths.back() && \"Out of bounds DefPath!\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 782, __PRETTY_FUNCTION__))
782 "Out of bounds DefPath!")((!Paths.empty() && NP >= &Paths.front() &&
NP <= &Paths.back() && "Out of bounds DefPath!"
) ? static_cast<void> (0) : __assert_fail ("!Paths.empty() && NP >= &Paths.front() && NP <= &Paths.back() && \"Out of bounds DefPath!\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 782, __PRETTY_FUNCTION__))
;
783 return NP - &Paths.front();
784 }
785
786 /// Try to optimize a phi as best as we can. Returns a SmallVector of Paths
787 /// that act as legal clobbers. Note that this won't return *all* clobbers.
788 ///
789 /// Phi optimization algorithm tl;dr:
790 /// - Find the earliest def/phi, A, we can optimize to
791 /// - Find if all paths from the starting memory access ultimately reach A
792 /// - If not, optimization isn't possible.
793 /// - Otherwise, walk from A to another clobber or phi, A'.
794 /// - If A' is a def, we're done.
795 /// - If A' is a phi, try to optimize it.
796 ///
797 /// A path is a series of {MemoryAccess, MemoryLocation} pairs. A path
798 /// terminates when a MemoryAccess that clobbers said MemoryLocation is found.
799 OptznResult tryOptimizePhi(MemoryPhi *Phi, MemoryAccess *Start,
800 const MemoryLocation &Loc) {
801 assert(Paths.empty() && VisitedPhis.empty() &&((Paths.empty() && VisitedPhis.empty() && "Reset the optimization state."
) ? static_cast<void> (0) : __assert_fail ("Paths.empty() && VisitedPhis.empty() && \"Reset the optimization state.\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 802, __PRETTY_FUNCTION__))
802 "Reset the optimization state.")((Paths.empty() && VisitedPhis.empty() && "Reset the optimization state."
) ? static_cast<void> (0) : __assert_fail ("Paths.empty() && VisitedPhis.empty() && \"Reset the optimization state.\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 802, __PRETTY_FUNCTION__))
;
803
804 Paths.emplace_back(Loc, Start, Phi, None);
805 // Stores how many "valid" optimization nodes we had prior to calling
806 // addSearches/getBlockingAccess. Necessary for caching if we had a blocker.
807 auto PriorPathsSize = Paths.size();
808
809 SmallVector<ListIndex, 16> PausedSearches;
810 SmallVector<ListIndex, 8> NewPaused;
811 SmallVector<TerminatedPath, 4> TerminatedPaths;
812
813 addSearches(Phi, PausedSearches, 0);
814
815 // Moves the TerminatedPath with the "most dominated" Clobber to the end of
816 // Paths.
817 auto MoveDominatedPathToEnd = [&](SmallVectorImpl<TerminatedPath> &Paths) {
818 assert(!Paths.empty() && "Need a path to move")((!Paths.empty() && "Need a path to move") ? static_cast
<void> (0) : __assert_fail ("!Paths.empty() && \"Need a path to move\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 818, __PRETTY_FUNCTION__))
;
819 auto Dom = Paths.begin();
820 for (auto I = std::next(Dom), E = Paths.end(); I != E; ++I)
821 if (!MSSA.dominates(I->Clobber, Dom->Clobber))
822 Dom = I;
823 auto Last = Paths.end() - 1;
824 if (Last != Dom)
825 std::iter_swap(Last, Dom);
826 };
827
828 MemoryPhi *Current = Phi;
829 while (1) {
1
Loop condition is true. Entering loop body
830 assert(!MSSA.isLiveOnEntryDef(Current) &&((!MSSA.isLiveOnEntryDef(Current) && "liveOnEntry wasn't treated as a clobber?"
) ? static_cast<void> (0) : __assert_fail ("!MSSA.isLiveOnEntryDef(Current) && \"liveOnEntry wasn't treated as a clobber?\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 831, __PRETTY_FUNCTION__))
831 "liveOnEntry wasn't treated as a clobber?")((!MSSA.isLiveOnEntryDef(Current) && "liveOnEntry wasn't treated as a clobber?"
) ? static_cast<void> (0) : __assert_fail ("!MSSA.isLiveOnEntryDef(Current) && \"liveOnEntry wasn't treated as a clobber?\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 831, __PRETTY_FUNCTION__))
;
832
833 MemoryAccess *Target = getWalkTarget(Current);
834 // If a TerminatedPath doesn't dominate Target, then it wasn't a legal
835 // optimization for the prior phi.
836 assert(all_of(TerminatedPaths, [&](const TerminatedPath &P) {((all_of(TerminatedPaths, [&](const TerminatedPath &P
) { return MSSA.dominates(P.Clobber, Target); })) ? static_cast
<void> (0) : __assert_fail ("all_of(TerminatedPaths, [&](const TerminatedPath &P) { return MSSA.dominates(P.Clobber, Target); })"
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 838, __PRETTY_FUNCTION__))
837 return MSSA.dominates(P.Clobber, Target);((all_of(TerminatedPaths, [&](const TerminatedPath &P
) { return MSSA.dominates(P.Clobber, Target); })) ? static_cast
<void> (0) : __assert_fail ("all_of(TerminatedPaths, [&](const TerminatedPath &P) { return MSSA.dominates(P.Clobber, Target); })"
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 838, __PRETTY_FUNCTION__))
838 }))((all_of(TerminatedPaths, [&](const TerminatedPath &P
) { return MSSA.dominates(P.Clobber, Target); })) ? static_cast
<void> (0) : __assert_fail ("all_of(TerminatedPaths, [&](const TerminatedPath &P) { return MSSA.dominates(P.Clobber, Target); })"
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 838, __PRETTY_FUNCTION__))
;
839
840 // FIXME: This is broken, because the Blocker may be reported to be
841 // liveOnEntry, and we'll happily wait for that to disappear (read: never)
842 // For the moment, this is fine, since we do nothing with blocker info.
843 if (Optional<TerminatedPath> Blocker = getBlockingAccess(
2
Taking false branch
844 Target, PausedSearches, NewPaused, TerminatedPaths)) {
845 // Cache our work on the blocking node, since we know that's correct.
846 cacheDefPath(Paths[Blocker->LastNode], Blocker->Clobber);
847
848 // Find the node we started at. We can't search based on N->Last, since
849 // we may have gone around a loop with a different MemoryLocation.
850 auto Iter = find_if(def_path(Blocker->LastNode), [&](const DefPath &N) {
851 return defPathIndex(N) < PriorPathsSize;
852 });
853 assert(Iter != def_path_iterator())((Iter != def_path_iterator()) ? static_cast<void> (0) :
__assert_fail ("Iter != def_path_iterator()", "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 853, __PRETTY_FUNCTION__))
;
854
855 DefPath &CurNode = *Iter;
856 assert(CurNode.Last == Current)((CurNode.Last == Current) ? static_cast<void> (0) : __assert_fail
("CurNode.Last == Current", "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 856, __PRETTY_FUNCTION__))
;
857
858 // Two things:
859 // A. We can't reliably cache all of NewPaused back. Consider a case
860 // where we have two paths in NewPaused; one of which can't optimize
861 // above this phi, whereas the other can. If we cache the second path
862 // back, we'll end up with suboptimal cache entries. We can handle
863 // cases like this a bit better when we either try to find all
864 // clobbers that block phi optimization, or when our cache starts
865 // supporting unfinished searches.
866 // B. We can't reliably cache TerminatedPaths back here without doing
867 // extra checks; consider a case like:
868 // T
869 // / \
870 // D C
871 // \ /
872 // S
873 // Where T is our target, C is a node with a clobber on it, D is a
874 // diamond (with a clobber *only* on the left or right node, N), and
875 // S is our start. Say we walk to D, through the node opposite N
876 // (read: ignoring the clobber), and see a cache entry in the top
877 // node of D. That cache entry gets put into TerminatedPaths. We then
878 // walk up to C (N is later in our worklist), find the clobber, and
879 // quit. If we append TerminatedPaths to OtherClobbers, we'll cache
880 // the bottom part of D to the cached clobber, ignoring the clobber
881 // in N. Again, this problem goes away if we start tracking all
882 // blockers for a given phi optimization.
883 TerminatedPath Result{CurNode.Last, defPathIndex(CurNode)};
884 return {Result, {}};
885 }
886
887 // If there's nothing left to search, then all paths led to valid clobbers
888 // that we got from our cache; pick the nearest to the start, and allow
889 // the rest to be cached back.
890 if (NewPaused.empty()) {
3
Taking false branch
891 MoveDominatedPathToEnd(TerminatedPaths);
892 TerminatedPath Result = TerminatedPaths.pop_back_val();
893 return {Result, std::move(TerminatedPaths)};
894 }
895
896 MemoryAccess *DefChainEnd = nullptr;
4
'DefChainEnd' initialized to a null pointer value
897 SmallVector<TerminatedPath, 4> Clobbers;
898 for (ListIndex Paused : NewPaused) {
5
Assuming '__begin' is equal to '__end'
899 UpwardsWalkResult WR = walkToPhiOrClobber(Paths[Paused]);
900 if (WR.IsKnownClobber)
901 Clobbers.push_back({WR.Result, Paused});
902 else
903 // Micro-opt: If we hit the end of the chain, save it.
904 DefChainEnd = WR.Result;
905 }
906
907 if (!TerminatedPaths.empty()) {
6
Taking true branch
908 // If we couldn't find the dominating phi/liveOnEntry in the above loop,
909 // do it now.
910 if (!DefChainEnd)
7
Taking true branch
911 for (MemoryAccess *MA : def_chain(Target))
912 DefChainEnd = MA;
913
914 // If any of the terminated paths don't dominate the phi we'll try to
915 // optimize, we need to figure out what they are and quit.
916 const BasicBlock *ChainBB = DefChainEnd->getBlock();
8
Called C++ object pointer is null
917 for (const TerminatedPath &TP : TerminatedPaths) {
918 // Because we know that DefChainEnd is as "high" as we can go, we
919 // don't need local dominance checks; BB dominance is sufficient.
920 if (DT.dominates(ChainBB, TP.Clobber->getBlock()))
921 Clobbers.push_back(TP);
922 }
923 }
924
925 // If we have clobbers in the def chain, find the one closest to Current
926 // and quit.
927 if (!Clobbers.empty()) {
928 MoveDominatedPathToEnd(Clobbers);
929 TerminatedPath Result = Clobbers.pop_back_val();
930 return {Result, std::move(Clobbers)};
931 }
932
933 assert(all_of(NewPaused,((all_of(NewPaused, [&](ListIndex I) { return Paths[I].Last
== DefChainEnd; })) ? static_cast<void> (0) : __assert_fail
("all_of(NewPaused, [&](ListIndex I) { return Paths[I].Last == DefChainEnd; })"
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 934, __PRETTY_FUNCTION__))
934 [&](ListIndex I) { return Paths[I].Last == DefChainEnd; }))((all_of(NewPaused, [&](ListIndex I) { return Paths[I].Last
== DefChainEnd; })) ? static_cast<void> (0) : __assert_fail
("all_of(NewPaused, [&](ListIndex I) { return Paths[I].Last == DefChainEnd; })"
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 934, __PRETTY_FUNCTION__))
;
935
936 // Because liveOnEntry is a clobber, this must be a phi.
937 auto *DefChainPhi = cast<MemoryPhi>(DefChainEnd);
938
939 PriorPathsSize = Paths.size();
940 PausedSearches.clear();
941 for (ListIndex I : NewPaused)
942 addSearches(DefChainPhi, PausedSearches, I);
943 NewPaused.clear();
944
945 Current = DefChainPhi;
946 }
947 }
948
949 /// Caches everything in an OptznResult.
950 void cacheOptResult(const OptznResult &R) {
951 if (R.OtherClobbers.empty()) {
952 // If we're not going to be caching OtherClobbers, don't bother with
953 // marking visited/etc.
954 for (const DefPath &N : const_def_path(R.PrimaryClobber.LastNode))
955 cacheDefPath(N, R.PrimaryClobber.Clobber);
956 return;
957 }
958
959 // PrimaryClobber is our answer. If we can cache anything back, we need to
960 // stop caching when we visit PrimaryClobber.
961 SmallBitVector Visited(Paths.size());
962 for (const DefPath &N : const_def_path(R.PrimaryClobber.LastNode)) {
963 Visited[defPathIndex(N)] = true;
964 cacheDefPath(N, R.PrimaryClobber.Clobber);
965 }
966
967 for (const TerminatedPath &P : R.OtherClobbers) {
968 for (const DefPath &N : const_def_path(P.LastNode)) {
969 ListIndex NIndex = defPathIndex(N);
970 if (Visited[NIndex])
971 break;
972 Visited[NIndex] = true;
973 cacheDefPath(N, P.Clobber);
974 }
975 }
976 }
977
978 void verifyOptResult(const OptznResult &R) const {
979 assert(all_of(R.OtherClobbers, [&](const TerminatedPath &P) {((all_of(R.OtherClobbers, [&](const TerminatedPath &P
) { return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber
); })) ? static_cast<void> (0) : __assert_fail ("all_of(R.OtherClobbers, [&](const TerminatedPath &P) { return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber); })"
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 981, __PRETTY_FUNCTION__))
980 return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber);((all_of(R.OtherClobbers, [&](const TerminatedPath &P
) { return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber
); })) ? static_cast<void> (0) : __assert_fail ("all_of(R.OtherClobbers, [&](const TerminatedPath &P) { return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber); })"
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 981, __PRETTY_FUNCTION__))
981 }))((all_of(R.OtherClobbers, [&](const TerminatedPath &P
) { return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber
); })) ? static_cast<void> (0) : __assert_fail ("all_of(R.OtherClobbers, [&](const TerminatedPath &P) { return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber); })"
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 981, __PRETTY_FUNCTION__))
;
982 }
983
984 void resetPhiOptznState() {
985 Paths.clear();
986 VisitedPhis.clear();
987 }
988
989public:
990 ClobberWalker(const MemorySSA &MSSA, AliasAnalysis &AA, DominatorTree &DT,
991 WalkerCache &WC)
992 : MSSA(MSSA), AA(AA), DT(DT), WC(WC), UseCache(true) {}
993
994 void reset() { WalkTargetCache.clear(); }
995
996 /// Finds the nearest clobber for the given query, optimizing phis if
997 /// possible.
998 MemoryAccess *findClobber(MemoryAccess *Start, UpwardsMemoryQuery &Q,
999 bool UseWalkerCache = true) {
1000 setUseCache(UseWalkerCache);
1001 Query = &Q;
1002
1003 MemoryAccess *Current = Start;
1004 // This walker pretends uses don't exist. If we're handed one, silently grab
1005 // its def. (This has the nice side-effect of ensuring we never cache uses)
1006 if (auto *MU = dyn_cast<MemoryUse>(Start))
1007 Current = MU->getDefiningAccess();
1008
1009 DefPath FirstDesc(Q.StartingLoc, Current, Current, None);
1010 // Fast path for the overly-common case (no crazy phi optimization
1011 // necessary)
1012 UpwardsWalkResult WalkResult = walkToPhiOrClobber(FirstDesc);
1013 MemoryAccess *Result;
1014 if (WalkResult.IsKnownClobber) {
1015 cacheDefPath(FirstDesc, WalkResult.Result);
1016 Result = WalkResult.Result;
1017 } else {
1018 OptznResult OptRes = tryOptimizePhi(cast<MemoryPhi>(FirstDesc.Last),
1019 Current, Q.StartingLoc);
1020 verifyOptResult(OptRes);
1021 cacheOptResult(OptRes);
1022 resetPhiOptznState();
1023 Result = OptRes.PrimaryClobber.Clobber;
1024 }
1025
1026#ifdef EXPENSIVE_CHECKS
1027 checkClobberSanity(Current, Result, Q.StartingLoc, MSSA, Q, AA);
1028#endif
1029 return Result;
1030 }
1031
1032 void verify(const MemorySSA *MSSA) { assert(MSSA == &this->MSSA)((MSSA == &this->MSSA) ? static_cast<void> (0) :
__assert_fail ("MSSA == &this->MSSA", "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1032, __PRETTY_FUNCTION__))
; }
1033};
1034
1035struct RenamePassData {
1036 DomTreeNode *DTN;
1037 DomTreeNode::const_iterator ChildIt;
1038 MemoryAccess *IncomingVal;
1039
1040 RenamePassData(DomTreeNode *D, DomTreeNode::const_iterator It,
1041 MemoryAccess *M)
1042 : DTN(D), ChildIt(It), IncomingVal(M) {}
1043 void swap(RenamePassData &RHS) {
1044 std::swap(DTN, RHS.DTN);
1045 std::swap(ChildIt, RHS.ChildIt);
1046 std::swap(IncomingVal, RHS.IncomingVal);
1047 }
1048};
1049} // anonymous namespace
1050
1051namespace llvm {
1052/// \brief A MemorySSAWalker that does AA walks and caching of lookups to
1053/// disambiguate accesses.
1054///
1055/// FIXME: The current implementation of this can take quadratic space in rare
1056/// cases. This can be fixed, but it is something to note until it is fixed.
1057///
1058/// In order to trigger this behavior, you need to store to N distinct locations
1059/// (that AA can prove don't alias), perform M stores to other memory
1060/// locations that AA can prove don't alias any of the initial N locations, and
1061/// then load from all of the N locations. In this case, we insert M cache
1062/// entries for each of the N loads.
1063///
1064/// For example:
1065/// define i32 @foo() {
1066/// %a = alloca i32, align 4
1067/// %b = alloca i32, align 4
1068/// store i32 0, i32* %a, align 4
1069/// store i32 0, i32* %b, align 4
1070///
1071/// ; Insert M stores to other memory that doesn't alias %a or %b here
1072///
1073/// %c = load i32, i32* %a, align 4 ; Caches M entries in
1074/// ; CachedUpwardsClobberingAccess for the
1075/// ; MemoryLocation %a
1076/// %d = load i32, i32* %b, align 4 ; Caches M entries in
1077/// ; CachedUpwardsClobberingAccess for the
1078/// ; MemoryLocation %b
1079///
1080/// ; For completeness' sake, loading %a or %b again would not cache *another*
1081/// ; M entries.
1082/// %r = add i32 %c, %d
1083/// ret i32 %r
1084/// }
1085class MemorySSA::CachingWalker final : public MemorySSAWalker {
1086 WalkerCache Cache;
1087 ClobberWalker Walker;
1088 bool AutoResetWalker;
1089
1090 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *, UpwardsMemoryQuery &);
1091 void verifyRemoved(MemoryAccess *);
1092
1093public:
1094 CachingWalker(MemorySSA *, AliasAnalysis *, DominatorTree *);
1095 ~CachingWalker() override;
1096
1097 using MemorySSAWalker::getClobberingMemoryAccess;
1098 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *) override;
1099 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *,
1100 const MemoryLocation &) override;
1101 void invalidateInfo(MemoryAccess *) override;
1102
1103 /// Whether we call resetClobberWalker() after each time we *actually* walk to
1104 /// answer a clobber query.
1105 void setAutoResetWalker(bool AutoReset) { AutoResetWalker = AutoReset; }
1106
1107 /// Drop the walker's persistent data structures. At the moment, this means
1108 /// "drop the walker's cache of BasicBlocks ->
1109 /// earliest-MemoryAccess-we-can-optimize-to". This is necessary if we're
1110 /// going to have DT updates, if we remove MemoryAccesses, etc.
1111 void resetClobberWalker() { Walker.reset(); }
1112
1113 void verify(const MemorySSA *MSSA) override {
1114 MemorySSAWalker::verify(MSSA);
1115 Walker.verify(MSSA);
1116 }
1117};
1118
1119/// \brief Rename a single basic block into MemorySSA form.
1120/// Uses the standard SSA renaming algorithm.
1121/// \returns The new incoming value.
1122MemoryAccess *MemorySSA::renameBlock(BasicBlock *BB,
1123 MemoryAccess *IncomingVal) {
1124 auto It = PerBlockAccesses.find(BB);
1125 // Skip most processing if the list is empty.
1126 if (It != PerBlockAccesses.end()) {
1127 AccessList *Accesses = It->second.get();
1128 for (MemoryAccess &L : *Accesses) {
1129 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(&L)) {
1130 if (MUD->getDefiningAccess() == nullptr)
1131 MUD->setDefiningAccess(IncomingVal);
1132 if (isa<MemoryDef>(&L))
1133 IncomingVal = &L;
1134 } else {
1135 IncomingVal = &L;
1136 }
1137 }
1138 }
1139
1140 // Pass through values to our successors
1141 for (const BasicBlock *S : successors(BB)) {
1142 auto It = PerBlockAccesses.find(S);
1143 // Rename the phi nodes in our successor block
1144 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
1145 continue;
1146 AccessList *Accesses = It->second.get();
1147 auto *Phi = cast<MemoryPhi>(&Accesses->front());
1148 Phi->addIncoming(IncomingVal, BB);
1149 }
1150
1151 return IncomingVal;
1152}
1153
1154/// \brief This is the standard SSA renaming algorithm.
1155///
1156/// We walk the dominator tree in preorder, renaming accesses, and then filling
1157/// in phi nodes in our successors.
1158void MemorySSA::renamePass(DomTreeNode *Root, MemoryAccess *IncomingVal,
1159 SmallPtrSet<BasicBlock *, 16> &Visited) {
1160 SmallVector<RenamePassData, 32> WorkStack;
1161 IncomingVal = renameBlock(Root->getBlock(), IncomingVal);
1162 WorkStack.push_back({Root, Root->begin(), IncomingVal});
1163 Visited.insert(Root->getBlock());
1164
1165 while (!WorkStack.empty()) {
1166 DomTreeNode *Node = WorkStack.back().DTN;
1167 DomTreeNode::const_iterator ChildIt = WorkStack.back().ChildIt;
1168 IncomingVal = WorkStack.back().IncomingVal;
1169
1170 if (ChildIt == Node->end()) {
1171 WorkStack.pop_back();
1172 } else {
1173 DomTreeNode *Child = *ChildIt;
1174 ++WorkStack.back().ChildIt;
1175 BasicBlock *BB = Child->getBlock();
1176 Visited.insert(BB);
1177 IncomingVal = renameBlock(BB, IncomingVal);
1178 WorkStack.push_back({Child, Child->begin(), IncomingVal});
1179 }
1180 }
1181}
1182
1183/// \brief Compute dominator levels, used by the phi insertion algorithm above.
1184void MemorySSA::computeDomLevels(DenseMap<DomTreeNode *, unsigned> &DomLevels) {
1185 for (auto DFI = df_begin(DT->getRootNode()), DFE = df_end(DT->getRootNode());
1186 DFI != DFE; ++DFI)
1187 DomLevels[*DFI] = DFI.getPathLength() - 1;
1188}
1189
1190/// \brief This handles unreachable block accesses by deleting phi nodes in
1191/// unreachable blocks, and marking all other unreachable MemoryAccess's as
1192/// being uses of the live on entry definition.
1193void MemorySSA::markUnreachableAsLiveOnEntry(BasicBlock *BB) {
1194 assert(!DT->isReachableFromEntry(BB) &&((!DT->isReachableFromEntry(BB) && "Reachable block found while handling unreachable blocks"
) ? static_cast<void> (0) : __assert_fail ("!DT->isReachableFromEntry(BB) && \"Reachable block found while handling unreachable blocks\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1195, __PRETTY_FUNCTION__))
1195 "Reachable block found while handling unreachable blocks")((!DT->isReachableFromEntry(BB) && "Reachable block found while handling unreachable blocks"
) ? static_cast<void> (0) : __assert_fail ("!DT->isReachableFromEntry(BB) && \"Reachable block found while handling unreachable blocks\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1195, __PRETTY_FUNCTION__))
;
1196
1197 // Make sure phi nodes in our reachable successors end up with a
1198 // LiveOnEntryDef for our incoming edge, even though our block is forward
1199 // unreachable. We could just disconnect these blocks from the CFG fully,
1200 // but we do not right now.
1201 for (const BasicBlock *S : successors(BB)) {
1202 if (!DT->isReachableFromEntry(S))
1203 continue;
1204 auto It = PerBlockAccesses.find(S);
1205 // Rename the phi nodes in our successor block
1206 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
1207 continue;
1208 AccessList *Accesses = It->second.get();
1209 auto *Phi = cast<MemoryPhi>(&Accesses->front());
1210 Phi->addIncoming(LiveOnEntryDef.get(), BB);
1211 }
1212
1213 auto It = PerBlockAccesses.find(BB);
1214 if (It == PerBlockAccesses.end())
1215 return;
1216
1217 auto &Accesses = It->second;
1218 for (auto AI = Accesses->begin(), AE = Accesses->end(); AI != AE;) {
1219 auto Next = std::next(AI);
1220 // If we have a phi, just remove it. We are going to replace all
1221 // users with live on entry.
1222 if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(AI))
1223 UseOrDef->setDefiningAccess(LiveOnEntryDef.get());
1224 else
1225 Accesses->erase(AI);
1226 AI = Next;
1227 }
1228}
1229
1230MemorySSA::MemorySSA(Function &Func, AliasAnalysis *AA, DominatorTree *DT)
1231 : AA(AA), DT(DT), F(Func), LiveOnEntryDef(nullptr), Walker(nullptr),
1232 NextID(INVALID_MEMORYACCESS_ID) {
1233 buildMemorySSA();
1234}
1235
1236MemorySSA::~MemorySSA() {
1237 // Drop all our references
1238 for (const auto &Pair : PerBlockAccesses)
1239 for (MemoryAccess &MA : *Pair.second)
1240 MA.dropAllReferences();
1241}
1242
1243MemorySSA::AccessList *MemorySSA::getOrCreateAccessList(const BasicBlock *BB) {
1244 auto Res = PerBlockAccesses.insert(std::make_pair(BB, nullptr));
1245
1246 if (Res.second)
1247 Res.first->second = make_unique<AccessList>();
1248 return Res.first->second.get();
1249}
1250
1251/// This class is a batch walker of all MemoryUse's in the program, and points
1252/// their defining access at the thing that actually clobbers them. Because it
1253/// is a batch walker that touches everything, it does not operate like the
1254/// other walkers. This walker is basically performing a top-down SSA renaming
1255/// pass, where the version stack is used as the cache. This enables it to be
1256/// significantly more time and memory efficient than using the regular walker,
1257/// which is walking bottom-up.
1258class MemorySSA::OptimizeUses {
1259public:
1260 OptimizeUses(MemorySSA *MSSA, MemorySSAWalker *Walker, AliasAnalysis *AA,
1261 DominatorTree *DT)
1262 : MSSA(MSSA), Walker(Walker), AA(AA), DT(DT) {
1263 Walker = MSSA->getWalker();
1264 }
1265
1266 void optimizeUses();
1267
1268private:
1269 /// This represents where a given memorylocation is in the stack.
1270 struct MemlocStackInfo {
1271 // This essentially is keeping track of versions of the stack. Whenever
1272 // the stack changes due to pushes or pops, these versions increase.
1273 unsigned long StackEpoch;
1274 unsigned long PopEpoch;
1275 // This is the lower bound of places on the stack to check. It is equal to
1276 // the place the last stack walk ended.
1277 // Note: Correctness depends on this being initialized to 0, which densemap
1278 // does
1279 unsigned long LowerBound;
1280 const BasicBlock *LowerBoundBlock;
1281 // This is where the last walk for this memory location ended.
1282 unsigned long LastKill;
1283 bool LastKillValid;
1284 };
1285 void optimizeUsesInBlock(const BasicBlock *, unsigned long &, unsigned long &,
1286 SmallVectorImpl<MemoryAccess *> &,
1287 DenseMap<MemoryLocOrCall, MemlocStackInfo> &);
1288 MemorySSA *MSSA;
1289 MemorySSAWalker *Walker;
1290 AliasAnalysis *AA;
1291 DominatorTree *DT;
1292};
1293
1294/// Optimize the uses in a given block This is basically the SSA renaming
1295/// algorithm, with one caveat: We are able to use a single stack for all
1296/// MemoryUses. This is because the set of *possible* reaching MemoryDefs is
1297/// the same for every MemoryUse. The *actual* clobbering MemoryDef is just
1298/// going to be some position in that stack of possible ones.
1299///
1300/// We track the stack positions that each MemoryLocation needs
1301/// to check, and last ended at. This is because we only want to check the
1302/// things that changed since last time. The same MemoryLocation should
1303/// get clobbered by the same store (getModRefInfo does not use invariantness or
1304/// things like this, and if they start, we can modify MemoryLocOrCall to
1305/// include relevant data)
1306void MemorySSA::OptimizeUses::optimizeUsesInBlock(
1307 const BasicBlock *BB, unsigned long &StackEpoch, unsigned long &PopEpoch,
1308 SmallVectorImpl<MemoryAccess *> &VersionStack,
1309 DenseMap<MemoryLocOrCall, MemlocStackInfo> &LocStackInfo) {
1310
1311 /// If no accesses, nothing to do.
1312 MemorySSA::AccessList *Accesses = MSSA->getWritableBlockAccesses(BB);
1313 if (Accesses == nullptr)
1314 return;
1315
1316 // Pop everything that doesn't dominate the current block off the stack,
1317 // increment the PopEpoch to account for this.
1318 while (!VersionStack.empty()) {
1319 BasicBlock *BackBlock = VersionStack.back()->getBlock();
1320 if (DT->dominates(BackBlock, BB))
1321 break;
1322 while (VersionStack.back()->getBlock() == BackBlock)
1323 VersionStack.pop_back();
1324 ++PopEpoch;
1325 }
1326 for (MemoryAccess &MA : *Accesses) {
1327 auto *MU = dyn_cast<MemoryUse>(&MA);
1328 if (!MU) {
1329 VersionStack.push_back(&MA);
1330 ++StackEpoch;
1331 continue;
1332 }
1333
1334 if (isUseTriviallyOptimizableToLiveOnEntry(*AA, MU->getMemoryInst())) {
1335 MU->setDefiningAccess(MSSA->getLiveOnEntryDef(), true);
1336 continue;
1337 }
1338
1339 MemoryLocOrCall UseMLOC(MU);
1340 auto &LocInfo = LocStackInfo[UseMLOC];
1341 // If the pop epoch changed, it means we've removed stuff from top of
1342 // stack due to changing blocks. We may have to reset the lower bound or
1343 // last kill info.
1344 if (LocInfo.PopEpoch != PopEpoch) {
1345 LocInfo.PopEpoch = PopEpoch;
1346 LocInfo.StackEpoch = StackEpoch;
1347 // If the lower bound was in something that no longer dominates us, we
1348 // have to reset it.
1349 // We can't simply track stack size, because the stack may have had
1350 // pushes/pops in the meantime.
1351 // XXX: This is non-optimal, but only is slower cases with heavily
1352 // branching dominator trees. To get the optimal number of queries would
1353 // be to make lowerbound and lastkill a per-loc stack, and pop it until
1354 // the top of that stack dominates us. This does not seem worth it ATM.
1355 // A much cheaper optimization would be to always explore the deepest
1356 // branch of the dominator tree first. This will guarantee this resets on
1357 // the smallest set of blocks.
1358 if (LocInfo.LowerBoundBlock && LocInfo.LowerBoundBlock != BB &&
1359 !DT->dominates(LocInfo.LowerBoundBlock, BB)) {
1360 // Reset the lower bound of things to check.
1361 // TODO: Some day we should be able to reset to last kill, rather than
1362 // 0.
1363 LocInfo.LowerBound = 0;
1364 LocInfo.LowerBoundBlock = VersionStack[0]->getBlock();
1365 LocInfo.LastKillValid = false;
1366 }
1367 } else if (LocInfo.StackEpoch != StackEpoch) {
1368 // If all that has changed is the StackEpoch, we only have to check the
1369 // new things on the stack, because we've checked everything before. In
1370 // this case, the lower bound of things to check remains the same.
1371 LocInfo.PopEpoch = PopEpoch;
1372 LocInfo.StackEpoch = StackEpoch;
1373 }
1374 if (!LocInfo.LastKillValid) {
1375 LocInfo.LastKill = VersionStack.size() - 1;
1376 LocInfo.LastKillValid = true;
1377 }
1378
1379 // At this point, we should have corrected last kill and LowerBound to be
1380 // in bounds.
1381 assert(LocInfo.LowerBound < VersionStack.size() &&((LocInfo.LowerBound < VersionStack.size() && "Lower bound out of range"
) ? static_cast<void> (0) : __assert_fail ("LocInfo.LowerBound < VersionStack.size() && \"Lower bound out of range\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1382, __PRETTY_FUNCTION__))
1382 "Lower bound out of range")((LocInfo.LowerBound < VersionStack.size() && "Lower bound out of range"
) ? static_cast<void> (0) : __assert_fail ("LocInfo.LowerBound < VersionStack.size() && \"Lower bound out of range\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1382, __PRETTY_FUNCTION__))
;
1383 assert(LocInfo.LastKill < VersionStack.size() &&((LocInfo.LastKill < VersionStack.size() && "Last kill info out of range"
) ? static_cast<void> (0) : __assert_fail ("LocInfo.LastKill < VersionStack.size() && \"Last kill info out of range\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1384, __PRETTY_FUNCTION__))
1384 "Last kill info out of range")((LocInfo.LastKill < VersionStack.size() && "Last kill info out of range"
) ? static_cast<void> (0) : __assert_fail ("LocInfo.LastKill < VersionStack.size() && \"Last kill info out of range\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1384, __PRETTY_FUNCTION__))
;
1385 // In any case, the new upper bound is the top of the stack.
1386 unsigned long UpperBound = VersionStack.size() - 1;
1387
1388 if (UpperBound - LocInfo.LowerBound > MaxCheckLimit) {
1389 DEBUG(dbgs() << "MemorySSA skipping optimization of " << *MU << " ("do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memoryssa")) { dbgs() << "MemorySSA skipping optimization of "
<< *MU << " (" << *(MU->getMemoryInst()
) << ")" << " because there are " << UpperBound
- LocInfo.LowerBound << " stores to disambiguate\n"; }
} while (false)
1390 << *(MU->getMemoryInst()) << ")"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memoryssa")) { dbgs() << "MemorySSA skipping optimization of "
<< *MU << " (" << *(MU->getMemoryInst()
) << ")" << " because there are " << UpperBound
- LocInfo.LowerBound << " stores to disambiguate\n"; }
} while (false)
1391 << " because there are " << UpperBound - LocInfo.LowerBounddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memoryssa")) { dbgs() << "MemorySSA skipping optimization of "
<< *MU << " (" << *(MU->getMemoryInst()
) << ")" << " because there are " << UpperBound
- LocInfo.LowerBound << " stores to disambiguate\n"; }
} while (false)
1392 << " stores to disambiguate\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memoryssa")) { dbgs() << "MemorySSA skipping optimization of "
<< *MU << " (" << *(MU->getMemoryInst()
) << ")" << " because there are " << UpperBound
- LocInfo.LowerBound << " stores to disambiguate\n"; }
} while (false)
;
1393 // Because we did not walk, LastKill is no longer valid, as this may
1394 // have been a kill.
1395 LocInfo.LastKillValid = false;
1396 continue;
1397 }
1398 bool FoundClobberResult = false;
1399 while (UpperBound > LocInfo.LowerBound) {
1400 if (isa<MemoryPhi>(VersionStack[UpperBound])) {
1401 // For phis, use the walker, see where we ended up, go there
1402 Instruction *UseInst = MU->getMemoryInst();
1403 MemoryAccess *Result = Walker->getClobberingMemoryAccess(UseInst);
1404 // We are guaranteed to find it or something is wrong
1405 while (VersionStack[UpperBound] != Result) {
1406 assert(UpperBound != 0)((UpperBound != 0) ? static_cast<void> (0) : __assert_fail
("UpperBound != 0", "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1406, __PRETTY_FUNCTION__))
;
1407 --UpperBound;
1408 }
1409 FoundClobberResult = true;
1410 break;
1411 }
1412
1413 MemoryDef *MD = cast<MemoryDef>(VersionStack[UpperBound]);
1414 // If the lifetime of the pointer ends at this instruction, it's live on
1415 // entry.
1416 if (!UseMLOC.IsCall && lifetimeEndsAt(MD, UseMLOC.getLoc(), *AA)) {
1417 // Reset UpperBound to liveOnEntryDef's place in the stack
1418 UpperBound = 0;
1419 FoundClobberResult = true;
1420 break;
1421 }
1422 if (instructionClobbersQuery(MD, MU, UseMLOC, *AA)) {
1423 FoundClobberResult = true;
1424 break;
1425 }
1426 --UpperBound;
1427 }
1428 // At the end of this loop, UpperBound is either a clobber, or lower bound
1429 // PHI walking may cause it to be < LowerBound, and in fact, < LastKill.
1430 if (FoundClobberResult || UpperBound < LocInfo.LastKill) {
1431 MU->setDefiningAccess(VersionStack[UpperBound], true);
1432 // We were last killed now by where we got to
1433 LocInfo.LastKill = UpperBound;
1434 } else {
1435 // Otherwise, we checked all the new ones, and now we know we can get to
1436 // LastKill.
1437 MU->setDefiningAccess(VersionStack[LocInfo.LastKill], true);
1438 }
1439 LocInfo.LowerBound = VersionStack.size() - 1;
1440 LocInfo.LowerBoundBlock = BB;
1441 }
1442}
1443
1444/// Optimize uses to point to their actual clobbering definitions.
1445void MemorySSA::OptimizeUses::optimizeUses() {
1446
1447 // We perform a non-recursive top-down dominator tree walk
1448 struct StackInfo {
1449 const DomTreeNode *Node;
1450 DomTreeNode::const_iterator Iter;
1451 };
1452
1453 SmallVector<MemoryAccess *, 16> VersionStack;
1454 SmallVector<StackInfo, 16> DomTreeWorklist;
1455 DenseMap<MemoryLocOrCall, MemlocStackInfo> LocStackInfo;
1456 VersionStack.push_back(MSSA->getLiveOnEntryDef());
1457
1458 unsigned long StackEpoch = 1;
1459 unsigned long PopEpoch = 1;
1460 for (const auto *DomNode : depth_first(DT->getRootNode()))
1461 optimizeUsesInBlock(DomNode->getBlock(), StackEpoch, PopEpoch, VersionStack,
1462 LocStackInfo);
1463}
1464
1465void MemorySSA::placePHINodes(
1466 const SmallPtrSetImpl<BasicBlock *> &DefiningBlocks,
1467 const DenseMap<const BasicBlock *, unsigned int> &BBNumbers) {
1468 // Determine where our MemoryPhi's should go
1469 ForwardIDFCalculator IDFs(*DT);
1470 IDFs.setDefiningBlocks(DefiningBlocks);
1471 SmallVector<BasicBlock *, 32> IDFBlocks;
1472 IDFs.calculate(IDFBlocks);
1473
1474 std::sort(IDFBlocks.begin(), IDFBlocks.end(),
1475 [&BBNumbers](const BasicBlock *A, const BasicBlock *B) {
1476 return BBNumbers.lookup(A) < BBNumbers.lookup(B);
1477 });
1478
1479 // Now place MemoryPhi nodes.
1480 for (auto &BB : IDFBlocks) {
1481 // Insert phi node
1482 AccessList *Accesses = getOrCreateAccessList(BB);
1483 MemoryPhi *Phi = new MemoryPhi(BB->getContext(), BB, NextID++);
1484 ValueToMemoryAccess[BB] = Phi;
1485 // Phi's always are placed at the front of the block.
1486 Accesses->push_front(Phi);
1487 }
1488}
1489
1490void MemorySSA::buildMemorySSA() {
1491 // We create an access to represent "live on entry", for things like
1492 // arguments or users of globals, where the memory they use is defined before
1493 // the beginning of the function. We do not actually insert it into the IR.
1494 // We do not define a live on exit for the immediate uses, and thus our
1495 // semantics do *not* imply that something with no immediate uses can simply
1496 // be removed.
1497 BasicBlock &StartingPoint = F.getEntryBlock();
1498 LiveOnEntryDef = make_unique<MemoryDef>(F.getContext(), nullptr, nullptr,
1499 &StartingPoint, NextID++);
1500 DenseMap<const BasicBlock *, unsigned int> BBNumbers;
1501 unsigned NextBBNum = 0;
1502
1503 // We maintain lists of memory accesses per-block, trading memory for time. We
1504 // could just look up the memory access for every possible instruction in the
1505 // stream.
1506 SmallPtrSet<BasicBlock *, 32> DefiningBlocks;
1507 SmallPtrSet<BasicBlock *, 32> DefUseBlocks;
1508 // Go through each block, figure out where defs occur, and chain together all
1509 // the accesses.
1510 for (BasicBlock &B : F) {
1511 BBNumbers[&B] = NextBBNum++;
1512 bool InsertIntoDef = false;
1513 AccessList *Accesses = nullptr;
1514 for (Instruction &I : B) {
1515 MemoryUseOrDef *MUD = createNewAccess(&I);
1516 if (!MUD)
1517 continue;
1518 InsertIntoDef |= isa<MemoryDef>(MUD);
1519
1520 if (!Accesses)
1521 Accesses = getOrCreateAccessList(&B);
1522 Accesses->push_back(MUD);
1523 }
1524 if (InsertIntoDef)
1525 DefiningBlocks.insert(&B);
1526 if (Accesses)
1527 DefUseBlocks.insert(&B);
1528 }
1529 placePHINodes(DefiningBlocks, BBNumbers);
1530
1531 // Now do regular SSA renaming on the MemoryDef/MemoryUse. Visited will get
1532 // filled in with all blocks.
1533 SmallPtrSet<BasicBlock *, 16> Visited;
1534 renamePass(DT->getRootNode(), LiveOnEntryDef.get(), Visited);
1535
1536 CachingWalker *Walker = getWalkerImpl();
1537
1538 // We're doing a batch of updates; don't drop useful caches between them.
1539 Walker->setAutoResetWalker(false);
1540 OptimizeUses(this, Walker, AA, DT).optimizeUses();
1541 Walker->setAutoResetWalker(true);
1542 Walker->resetClobberWalker();
1543
1544 // Mark the uses in unreachable blocks as live on entry, so that they go
1545 // somewhere.
1546 for (auto &BB : F)
1547 if (!Visited.count(&BB))
1548 markUnreachableAsLiveOnEntry(&BB);
1549}
1550
1551MemorySSAWalker *MemorySSA::getWalker() { return getWalkerImpl(); }
1552
1553MemorySSA::CachingWalker *MemorySSA::getWalkerImpl() {
1554 if (Walker)
1555 return Walker.get();
1556
1557 Walker = make_unique<CachingWalker>(this, AA, DT);
1558 return Walker.get();
1559}
1560
1561MemoryPhi *MemorySSA::createMemoryPhi(BasicBlock *BB) {
1562 assert(!getMemoryAccess(BB) && "MemoryPhi already exists for this BB")((!getMemoryAccess(BB) && "MemoryPhi already exists for this BB"
) ? static_cast<void> (0) : __assert_fail ("!getMemoryAccess(BB) && \"MemoryPhi already exists for this BB\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1562, __PRETTY_FUNCTION__))
;
1563 AccessList *Accesses = getOrCreateAccessList(BB);
1564 MemoryPhi *Phi = new MemoryPhi(BB->getContext(), BB, NextID++);
1565 ValueToMemoryAccess[BB] = Phi;
1566 // Phi's always are placed at the front of the block.
1567 Accesses->push_front(Phi);
1568 BlockNumberingValid.erase(BB);
1569 return Phi;
1570}
1571
1572MemoryUseOrDef *MemorySSA::createDefinedAccess(Instruction *I,
1573 MemoryAccess *Definition) {
1574 assert(!isa<PHINode>(I) && "Cannot create a defined access for a PHI")((!isa<PHINode>(I) && "Cannot create a defined access for a PHI"
) ? static_cast<void> (0) : __assert_fail ("!isa<PHINode>(I) && \"Cannot create a defined access for a PHI\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1574, __PRETTY_FUNCTION__))
;
1575 MemoryUseOrDef *NewAccess = createNewAccess(I);
1576 assert(((NewAccess != nullptr && "Tried to create a memory access for a non-memory touching instruction"
) ? static_cast<void> (0) : __assert_fail ("NewAccess != nullptr && \"Tried to create a memory access for a non-memory touching instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1578, __PRETTY_FUNCTION__))
1577 NewAccess != nullptr &&((NewAccess != nullptr && "Tried to create a memory access for a non-memory touching instruction"
) ? static_cast<void> (0) : __assert_fail ("NewAccess != nullptr && \"Tried to create a memory access for a non-memory touching instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1578, __PRETTY_FUNCTION__))
1578 "Tried to create a memory access for a non-memory touching instruction")((NewAccess != nullptr && "Tried to create a memory access for a non-memory touching instruction"
) ? static_cast<void> (0) : __assert_fail ("NewAccess != nullptr && \"Tried to create a memory access for a non-memory touching instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1578, __PRETTY_FUNCTION__))
;
1579 NewAccess->setDefiningAccess(Definition);
1580 return NewAccess;
1581}
1582
1583MemoryAccess *MemorySSA::createMemoryAccessInBB(Instruction *I,
1584 MemoryAccess *Definition,
1585 const BasicBlock *BB,
1586 InsertionPlace Point) {
1587 MemoryUseOrDef *NewAccess = createDefinedAccess(I, Definition);
1588 auto *Accesses = getOrCreateAccessList(BB);
1589 if (Point == Beginning) {
1590 // It goes after any phi nodes
1591 auto AI = find_if(
1592 *Accesses, [](const MemoryAccess &MA) { return !isa<MemoryPhi>(MA); });
1593
1594 Accesses->insert(AI, NewAccess);
1595 } else {
1596 Accesses->push_back(NewAccess);
1597 }
1598 BlockNumberingValid.erase(BB);
1599 return NewAccess;
1600}
1601
1602MemoryUseOrDef *MemorySSA::createMemoryAccessBefore(Instruction *I,
1603 MemoryAccess *Definition,
1604 MemoryUseOrDef *InsertPt) {
1605 assert(I->getParent() == InsertPt->getBlock() &&((I->getParent() == InsertPt->getBlock() && "New and old access must be in the same block"
) ? static_cast<void> (0) : __assert_fail ("I->getParent() == InsertPt->getBlock() && \"New and old access must be in the same block\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1606, __PRETTY_FUNCTION__))
1606 "New and old access must be in the same block")((I->getParent() == InsertPt->getBlock() && "New and old access must be in the same block"
) ? static_cast<void> (0) : __assert_fail ("I->getParent() == InsertPt->getBlock() && \"New and old access must be in the same block\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1606, __PRETTY_FUNCTION__))
;
1607 MemoryUseOrDef *NewAccess = createDefinedAccess(I, Definition);
1608 auto *Accesses = getOrCreateAccessList(InsertPt->getBlock());
1609 Accesses->insert(AccessList::iterator(InsertPt), NewAccess);
1610 BlockNumberingValid.erase(InsertPt->getBlock());
1611 return NewAccess;
1612}
1613
1614MemoryUseOrDef *MemorySSA::createMemoryAccessAfter(Instruction *I,
1615 MemoryAccess *Definition,
1616 MemoryAccess *InsertPt) {
1617 assert(I->getParent() == InsertPt->getBlock() &&((I->getParent() == InsertPt->getBlock() && "New and old access must be in the same block"
) ? static_cast<void> (0) : __assert_fail ("I->getParent() == InsertPt->getBlock() && \"New and old access must be in the same block\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1618, __PRETTY_FUNCTION__))
1618 "New and old access must be in the same block")((I->getParent() == InsertPt->getBlock() && "New and old access must be in the same block"
) ? static_cast<void> (0) : __assert_fail ("I->getParent() == InsertPt->getBlock() && \"New and old access must be in the same block\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1618, __PRETTY_FUNCTION__))
;
1619 MemoryUseOrDef *NewAccess = createDefinedAccess(I, Definition);
1620 auto *Accesses = getOrCreateAccessList(InsertPt->getBlock());
1621 Accesses->insertAfter(AccessList::iterator(InsertPt), NewAccess);
1622 BlockNumberingValid.erase(InsertPt->getBlock());
1623 return NewAccess;
1624}
1625
1626void MemorySSA::spliceMemoryAccessAbove(MemoryDef *Where,
1627 MemoryUseOrDef *What) {
1628 assert(What != getLiveOnEntryDef() &&((What != getLiveOnEntryDef() && Where != getLiveOnEntryDef
() && "Can't splice (above) LOE.") ? static_cast<void
> (0) : __assert_fail ("What != getLiveOnEntryDef() && Where != getLiveOnEntryDef() && \"Can't splice (above) LOE.\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1629, __PRETTY_FUNCTION__))
1629 Where != getLiveOnEntryDef() && "Can't splice (above) LOE.")((What != getLiveOnEntryDef() && Where != getLiveOnEntryDef
() && "Can't splice (above) LOE.") ? static_cast<void
> (0) : __assert_fail ("What != getLiveOnEntryDef() && Where != getLiveOnEntryDef() && \"Can't splice (above) LOE.\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1629, __PRETTY_FUNCTION__))
;
1630 assert(dominates(Where, What) && "Only upwards splices are permitted.")((dominates(Where, What) && "Only upwards splices are permitted."
) ? static_cast<void> (0) : __assert_fail ("dominates(Where, What) && \"Only upwards splices are permitted.\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1630, __PRETTY_FUNCTION__))
;
1631
1632 if (Where == What)
1633 return;
1634 if (isa<MemoryDef>(What)) {
1635 // TODO: possibly use removeMemoryAccess' more efficient RAUW
1636 What->replaceAllUsesWith(What->getDefiningAccess());
1637 What->setDefiningAccess(Where->getDefiningAccess());
1638 Where->setDefiningAccess(What);
1639 }
1640 AccessList *Src = getWritableBlockAccesses(What->getBlock());
1641 AccessList *Dest = getWritableBlockAccesses(Where->getBlock());
1642 Dest->splice(AccessList::iterator(Where), *Src, What);
1643
1644 BlockNumberingValid.erase(What->getBlock());
1645 if (What->getBlock() != Where->getBlock())
1646 BlockNumberingValid.erase(Where->getBlock());
1647}
1648
1649/// \brief Helper function to create new memory accesses
1650MemoryUseOrDef *MemorySSA::createNewAccess(Instruction *I) {
1651 // The assume intrinsic has a control dependency which we model by claiming
1652 // that it writes arbitrarily. Ignore that fake memory dependency here.
1653 // FIXME: Replace this special casing with a more accurate modelling of
1654 // assume's control dependency.
1655 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1656 if (II->getIntrinsicID() == Intrinsic::assume)
1657 return nullptr;
1658
1659 // Find out what affect this instruction has on memory.
1660 ModRefInfo ModRef = AA->getModRefInfo(I);
1661 bool Def = bool(ModRef & MRI_Mod);
1662 bool Use = bool(ModRef & MRI_Ref);
1663
1664 // It's possible for an instruction to not modify memory at all. During
1665 // construction, we ignore them.
1666 if (!Def && !Use)
1667 return nullptr;
1668
1669 assert((Def || Use) &&(((Def || Use) && "Trying to create a memory access with a non-memory instruction"
) ? static_cast<void> (0) : __assert_fail ("(Def || Use) && \"Trying to create a memory access with a non-memory instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1670, __PRETTY_FUNCTION__))
1670 "Trying to create a memory access with a non-memory instruction")(((Def || Use) && "Trying to create a memory access with a non-memory instruction"
) ? static_cast<void> (0) : __assert_fail ("(Def || Use) && \"Trying to create a memory access with a non-memory instruction\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1670, __PRETTY_FUNCTION__))
;
1671
1672 MemoryUseOrDef *MUD;
1673 if (Def)
1674 MUD = new MemoryDef(I->getContext(), nullptr, I, I->getParent(), NextID++);
1675 else
1676 MUD = new MemoryUse(I->getContext(), nullptr, I, I->getParent());
1677 ValueToMemoryAccess[I] = MUD;
1678 return MUD;
1679}
1680
1681MemoryAccess *MemorySSA::findDominatingDef(BasicBlock *UseBlock,
1682 enum InsertionPlace Where) {
1683 // Handle the initial case
1684 if (Where == Beginning)
1685 // The only thing that could define us at the beginning is a phi node
1686 if (MemoryPhi *Phi = getMemoryAccess(UseBlock))
1687 return Phi;
1688
1689 DomTreeNode *CurrNode = DT->getNode(UseBlock);
1690 // Need to be defined by our dominator
1691 if (Where == Beginning)
1692 CurrNode = CurrNode->getIDom();
1693 Where = End;
1694 while (CurrNode) {
1695 auto It = PerBlockAccesses.find(CurrNode->getBlock());
1696 if (It != PerBlockAccesses.end()) {
1697 auto &Accesses = It->second;
1698 for (MemoryAccess &RA : reverse(*Accesses)) {
1699 if (isa<MemoryDef>(RA) || isa<MemoryPhi>(RA))
1700 return &RA;
1701 }
1702 }
1703 CurrNode = CurrNode->getIDom();
1704 }
1705 return LiveOnEntryDef.get();
1706}
1707
1708/// \brief Returns true if \p Replacer dominates \p Replacee .
1709bool MemorySSA::dominatesUse(const MemoryAccess *Replacer,
1710 const MemoryAccess *Replacee) const {
1711 if (isa<MemoryUseOrDef>(Replacee))
1712 return DT->dominates(Replacer->getBlock(), Replacee->getBlock());
1713 const auto *MP = cast<MemoryPhi>(Replacee);
1714 // For a phi node, the use occurs in the predecessor block of the phi node.
1715 // Since we may occur multiple times in the phi node, we have to check each
1716 // operand to ensure Replacer dominates each operand where Replacee occurs.
1717 for (const Use &Arg : MP->operands()) {
1718 if (Arg.get() != Replacee &&
1719 !DT->dominates(Replacer->getBlock(), MP->getIncomingBlock(Arg)))
1720 return false;
1721 }
1722 return true;
1723}
1724
1725/// \brief If all arguments of a MemoryPHI are defined by the same incoming
1726/// argument, return that argument.
1727static MemoryAccess *onlySingleValue(MemoryPhi *MP) {
1728 MemoryAccess *MA = nullptr;
1729
1730 for (auto &Arg : MP->operands()) {
1731 if (!MA)
1732 MA = cast<MemoryAccess>(Arg);
1733 else if (MA != Arg)
1734 return nullptr;
1735 }
1736 return MA;
1737}
1738
1739/// \brief Properly remove \p MA from all of MemorySSA's lookup tables.
1740///
1741/// Because of the way the intrusive list and use lists work, it is important to
1742/// do removal in the right order.
1743void MemorySSA::removeFromLookups(MemoryAccess *MA) {
1744 assert(MA->use_empty() &&((MA->use_empty() && "Trying to remove memory access that still has uses"
) ? static_cast<void> (0) : __assert_fail ("MA->use_empty() && \"Trying to remove memory access that still has uses\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1745, __PRETTY_FUNCTION__))
1745 "Trying to remove memory access that still has uses")((MA->use_empty() && "Trying to remove memory access that still has uses"
) ? static_cast<void> (0) : __assert_fail ("MA->use_empty() && \"Trying to remove memory access that still has uses\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1745, __PRETTY_FUNCTION__))
;
1746 BlockNumbering.erase(MA);
1747 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA))
1748 MUD->setDefiningAccess(nullptr);
1749 // Invalidate our walker's cache if necessary
1750 if (!isa<MemoryUse>(MA))
1751 Walker->invalidateInfo(MA);
1752 // The call below to erase will destroy MA, so we can't change the order we
1753 // are doing things here
1754 Value *MemoryInst;
1755 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA)) {
1756 MemoryInst = MUD->getMemoryInst();
1757 } else {
1758 MemoryInst = MA->getBlock();
1759 }
1760 auto VMA = ValueToMemoryAccess.find(MemoryInst);
1761 if (VMA->second == MA)
1762 ValueToMemoryAccess.erase(VMA);
1763
1764 auto AccessIt = PerBlockAccesses.find(MA->getBlock());
1765 std::unique_ptr<AccessList> &Accesses = AccessIt->second;
1766 Accesses->erase(MA);
1767 if (Accesses->empty())
1768 PerBlockAccesses.erase(AccessIt);
1769}
1770
1771void MemorySSA::removeMemoryAccess(MemoryAccess *MA) {
1772 assert(!isLiveOnEntryDef(MA) && "Trying to remove the live on entry def")((!isLiveOnEntryDef(MA) && "Trying to remove the live on entry def"
) ? static_cast<void> (0) : __assert_fail ("!isLiveOnEntryDef(MA) && \"Trying to remove the live on entry def\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1772, __PRETTY_FUNCTION__))
;
1773 // We can only delete phi nodes if they have no uses, or we can replace all
1774 // uses with a single definition.
1775 MemoryAccess *NewDefTarget = nullptr;
1776 if (MemoryPhi *MP = dyn_cast<MemoryPhi>(MA)) {
1777 // Note that it is sufficient to know that all edges of the phi node have
1778 // the same argument. If they do, by the definition of dominance frontiers
1779 // (which we used to place this phi), that argument must dominate this phi,
1780 // and thus, must dominate the phi's uses, and so we will not hit the assert
1781 // below.
1782 NewDefTarget = onlySingleValue(MP);
1783 assert((NewDefTarget || MP->use_empty()) &&(((NewDefTarget || MP->use_empty()) && "We can't delete this memory phi"
) ? static_cast<void> (0) : __assert_fail ("(NewDefTarget || MP->use_empty()) && \"We can't delete this memory phi\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1784, __PRETTY_FUNCTION__))
1784 "We can't delete this memory phi")(((NewDefTarget || MP->use_empty()) && "We can't delete this memory phi"
) ? static_cast<void> (0) : __assert_fail ("(NewDefTarget || MP->use_empty()) && \"We can't delete this memory phi\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1784, __PRETTY_FUNCTION__))
;
1785 } else {
1786 NewDefTarget = cast<MemoryUseOrDef>(MA)->getDefiningAccess();
1787 }
1788
1789 // Re-point the uses at our defining access
1790 if (!MA->use_empty()) {
1791 // Reset optimized on users of this store, and reset the uses.
1792 // A few notes:
1793 // 1. This is a slightly modified version of RAUW to avoid walking the
1794 // uses twice here.
1795 // 2. If we wanted to be complete, we would have to reset the optimized
1796 // flags on users of phi nodes if doing the below makes a phi node have all
1797 // the same arguments. Instead, we prefer users to removeMemoryAccess those
1798 // phi nodes, because doing it here would be N^3.
1799 if (MA->hasValueHandle())
1800 ValueHandleBase::ValueIsRAUWd(MA, NewDefTarget);
1801 // Note: We assume MemorySSA is not used in metadata since it's not really
1802 // part of the IR.
1803
1804 while (!MA->use_empty()) {
1805 Use &U = *MA->use_begin();
1806 if (MemoryUse *MU = dyn_cast<MemoryUse>(U.getUser()))
1807 MU->resetOptimized();
1808 U.set(NewDefTarget);
1809 }
1810 }
1811
1812 // The call below to erase will destroy MA, so we can't change the order we
1813 // are doing things here
1814 removeFromLookups(MA);
1815}
1816
1817void MemorySSA::print(raw_ostream &OS) const {
1818 MemorySSAAnnotatedWriter Writer(this);
1819 F.print(OS, &Writer);
1820}
1821
1822void MemorySSA::dump() const {
1823 MemorySSAAnnotatedWriter Writer(this);
1824 F.print(dbgs(), &Writer);
1825}
1826
1827void MemorySSA::verifyMemorySSA() const {
1828 verifyDefUses(F);
1829 verifyDomination(F);
1830 verifyOrdering(F);
1831 Walker->verify(this);
1832}
1833
1834/// \brief Verify that the order and existence of MemoryAccesses matches the
1835/// order and existence of memory affecting instructions.
1836void MemorySSA::verifyOrdering(Function &F) const {
1837 // Walk all the blocks, comparing what the lookups think and what the access
1838 // lists think, as well as the order in the blocks vs the order in the access
1839 // lists.
1840 SmallVector<MemoryAccess *, 32> ActualAccesses;
1841 for (BasicBlock &B : F) {
1842 const AccessList *AL = getBlockAccesses(&B);
1843 MemoryAccess *Phi = getMemoryAccess(&B);
1844 if (Phi)
1845 ActualAccesses.push_back(Phi);
1846 for (Instruction &I : B) {
1847 MemoryAccess *MA = getMemoryAccess(&I);
1848 assert((!MA || AL) && "We have memory affecting instructions "(((!MA || AL) && "We have memory affecting instructions "
"in this block but they are not in the " "access list") ? static_cast
<void> (0) : __assert_fail ("(!MA || AL) && \"We have memory affecting instructions \" \"in this block but they are not in the \" \"access list\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1850, __PRETTY_FUNCTION__))
1849 "in this block but they are not in the "(((!MA || AL) && "We have memory affecting instructions "
"in this block but they are not in the " "access list") ? static_cast
<void> (0) : __assert_fail ("(!MA || AL) && \"We have memory affecting instructions \" \"in this block but they are not in the \" \"access list\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1850, __PRETTY_FUNCTION__))
1850 "access list")(((!MA || AL) && "We have memory affecting instructions "
"in this block but they are not in the " "access list") ? static_cast
<void> (0) : __assert_fail ("(!MA || AL) && \"We have memory affecting instructions \" \"in this block but they are not in the \" \"access list\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1850, __PRETTY_FUNCTION__))
;
1851 if (MA)
1852 ActualAccesses.push_back(MA);
1853 }
1854 // Either we hit the assert, really have no accesses, or we have both
1855 // accesses and an access list
1856 if (!AL)
1857 continue;
1858 assert(AL->size() == ActualAccesses.size() &&((AL->size() == ActualAccesses.size() && "We don't have the same number of accesses in the block as on the "
"access list") ? static_cast<void> (0) : __assert_fail
("AL->size() == ActualAccesses.size() && \"We don't have the same number of accesses in the block as on the \" \"access list\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1860, __PRETTY_FUNCTION__))
1859 "We don't have the same number of accesses in the block as on the "((AL->size() == ActualAccesses.size() && "We don't have the same number of accesses in the block as on the "
"access list") ? static_cast<void> (0) : __assert_fail
("AL->size() == ActualAccesses.size() && \"We don't have the same number of accesses in the block as on the \" \"access list\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1860, __PRETTY_FUNCTION__))
1860 "access list")((AL->size() == ActualAccesses.size() && "We don't have the same number of accesses in the block as on the "
"access list") ? static_cast<void> (0) : __assert_fail
("AL->size() == ActualAccesses.size() && \"We don't have the same number of accesses in the block as on the \" \"access list\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1860, __PRETTY_FUNCTION__))
;
1861 auto ALI = AL->begin();
1862 auto AAI = ActualAccesses.begin();
1863 while (ALI != AL->end() && AAI != ActualAccesses.end()) {
1864 assert(&*ALI == *AAI && "Not the same accesses in the same order")((&*ALI == *AAI && "Not the same accesses in the same order"
) ? static_cast<void> (0) : __assert_fail ("&*ALI == *AAI && \"Not the same accesses in the same order\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1864, __PRETTY_FUNCTION__))
;
1865 ++ALI;
1866 ++AAI;
1867 }
1868 ActualAccesses.clear();
1869 }
1870}
1871
1872/// \brief Verify the domination properties of MemorySSA by checking that each
1873/// definition dominates all of its uses.
1874void MemorySSA::verifyDomination(Function &F) const {
1875#ifndef NDEBUG
1876 for (BasicBlock &B : F) {
1877 // Phi nodes are attached to basic blocks
1878 if (MemoryPhi *MP = getMemoryAccess(&B))
1879 for (const Use &U : MP->uses())
1880 assert(dominates(MP, U) && "Memory PHI does not dominate it's uses")((dominates(MP, U) && "Memory PHI does not dominate it's uses"
) ? static_cast<void> (0) : __assert_fail ("dominates(MP, U) && \"Memory PHI does not dominate it's uses\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1880, __PRETTY_FUNCTION__))
;
1881
1882 for (Instruction &I : B) {
1883 MemoryAccess *MD = dyn_cast_or_null<MemoryDef>(getMemoryAccess(&I));
1884 if (!MD)
1885 continue;
1886
1887 for (const Use &U : MD->uses())
1888 assert(dominates(MD, U) && "Memory Def does not dominate it's uses")((dominates(MD, U) && "Memory Def does not dominate it's uses"
) ? static_cast<void> (0) : __assert_fail ("dominates(MD, U) && \"Memory Def does not dominate it's uses\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1888, __PRETTY_FUNCTION__))
;
1889 }
1890 }
1891#endif
1892}
1893
1894/// \brief Verify the def-use lists in MemorySSA, by verifying that \p Use
1895/// appears in the use list of \p Def.
1896
1897void MemorySSA::verifyUseInDefs(MemoryAccess *Def, MemoryAccess *Use) const {
1898#ifndef NDEBUG
1899 // The live on entry use may cause us to get a NULL def here
1900 if (!Def)
1901 assert(isLiveOnEntryDef(Use) &&((isLiveOnEntryDef(Use) && "Null def but use not point to live on entry def"
) ? static_cast<void> (0) : __assert_fail ("isLiveOnEntryDef(Use) && \"Null def but use not point to live on entry def\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1902, __PRETTY_FUNCTION__))
1902 "Null def but use not point to live on entry def")((isLiveOnEntryDef(Use) && "Null def but use not point to live on entry def"
) ? static_cast<void> (0) : __assert_fail ("isLiveOnEntryDef(Use) && \"Null def but use not point to live on entry def\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1902, __PRETTY_FUNCTION__))
;
1903 else
1904 assert(is_contained(Def->users(), Use) &&((is_contained(Def->users(), Use) && "Did not find use in def's use list"
) ? static_cast<void> (0) : __assert_fail ("is_contained(Def->users(), Use) && \"Did not find use in def's use list\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1905, __PRETTY_FUNCTION__))
1905 "Did not find use in def's use list")((is_contained(Def->users(), Use) && "Did not find use in def's use list"
) ? static_cast<void> (0) : __assert_fail ("is_contained(Def->users(), Use) && \"Did not find use in def's use list\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1905, __PRETTY_FUNCTION__))
;
1906#endif
1907}
1908
1909/// \brief Verify the immediate use information, by walking all the memory
1910/// accesses and verifying that, for each use, it appears in the
1911/// appropriate def's use list
1912void MemorySSA::verifyDefUses(Function &F) const {
1913 for (BasicBlock &B : F) {
1914 // Phi nodes are attached to basic blocks
1915 if (MemoryPhi *Phi = getMemoryAccess(&B)) {
1916 assert(Phi->getNumOperands() == static_cast<unsigned>(std::distance(((Phi->getNumOperands() == static_cast<unsigned>(std
::distance( pred_begin(&B), pred_end(&B))) &&
"Incomplete MemoryPhi Node") ? static_cast<void> (0) :
__assert_fail ("Phi->getNumOperands() == static_cast<unsigned>(std::distance( pred_begin(&B), pred_end(&B))) && \"Incomplete MemoryPhi Node\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1918, __PRETTY_FUNCTION__))
1917 pred_begin(&B), pred_end(&B))) &&((Phi->getNumOperands() == static_cast<unsigned>(std
::distance( pred_begin(&B), pred_end(&B))) &&
"Incomplete MemoryPhi Node") ? static_cast<void> (0) :
__assert_fail ("Phi->getNumOperands() == static_cast<unsigned>(std::distance( pred_begin(&B), pred_end(&B))) && \"Incomplete MemoryPhi Node\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1918, __PRETTY_FUNCTION__))
1918 "Incomplete MemoryPhi Node")((Phi->getNumOperands() == static_cast<unsigned>(std
::distance( pred_begin(&B), pred_end(&B))) &&
"Incomplete MemoryPhi Node") ? static_cast<void> (0) :
__assert_fail ("Phi->getNumOperands() == static_cast<unsigned>(std::distance( pred_begin(&B), pred_end(&B))) && \"Incomplete MemoryPhi Node\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1918, __PRETTY_FUNCTION__))
;
1919 for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I)
1920 verifyUseInDefs(Phi->getIncomingValue(I), Phi);
1921 }
1922
1923 for (Instruction &I : B) {
1924 if (MemoryUseOrDef *MA = getMemoryAccess(&I)) {
1925 verifyUseInDefs(MA->getDefiningAccess(), MA);
1926 }
1927 }
1928 }
1929}
1930
1931MemoryUseOrDef *MemorySSA::getMemoryAccess(const Instruction *I) const {
1932 return cast_or_null<MemoryUseOrDef>(ValueToMemoryAccess.lookup(I));
1933}
1934
1935MemoryPhi *MemorySSA::getMemoryAccess(const BasicBlock *BB) const {
1936 return cast_or_null<MemoryPhi>(ValueToMemoryAccess.lookup(cast<Value>(BB)));
1937}
1938
1939/// Perform a local numbering on blocks so that instruction ordering can be
1940/// determined in constant time.
1941/// TODO: We currently just number in order. If we numbered by N, we could
1942/// allow at least N-1 sequences of insertBefore or insertAfter (and at least
1943/// log2(N) sequences of mixed before and after) without needing to invalidate
1944/// the numbering.
1945void MemorySSA::renumberBlock(const BasicBlock *B) const {
1946 // The pre-increment ensures the numbers really start at 1.
1947 unsigned long CurrentNumber = 0;
1948 const AccessList *AL = getBlockAccesses(B);
1949 assert(AL != nullptr && "Asking to renumber an empty block")((AL != nullptr && "Asking to renumber an empty block"
) ? static_cast<void> (0) : __assert_fail ("AL != nullptr && \"Asking to renumber an empty block\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1949, __PRETTY_FUNCTION__))
;
1950 for (const auto &I : *AL)
1951 BlockNumbering[&I] = ++CurrentNumber;
1952 BlockNumberingValid.insert(B);
1953}
1954
1955/// \brief Determine, for two memory accesses in the same block,
1956/// whether \p Dominator dominates \p Dominatee.
1957/// \returns True if \p Dominator dominates \p Dominatee.
1958bool MemorySSA::locallyDominates(const MemoryAccess *Dominator,
1959 const MemoryAccess *Dominatee) const {
1960
1961 const BasicBlock *DominatorBlock = Dominator->getBlock();
1962
1963 assert((DominatorBlock == Dominatee->getBlock()) &&(((DominatorBlock == Dominatee->getBlock()) && "Asking for local domination when accesses are in different blocks!"
) ? static_cast<void> (0) : __assert_fail ("(DominatorBlock == Dominatee->getBlock()) && \"Asking for local domination when accesses are in different blocks!\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1964, __PRETTY_FUNCTION__))
1964 "Asking for local domination when accesses are in different blocks!")(((DominatorBlock == Dominatee->getBlock()) && "Asking for local domination when accesses are in different blocks!"
) ? static_cast<void> (0) : __assert_fail ("(DominatorBlock == Dominatee->getBlock()) && \"Asking for local domination when accesses are in different blocks!\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1964, __PRETTY_FUNCTION__))
;
1965 // A node dominates itself.
1966 if (Dominatee == Dominator)
1967 return true;
1968
1969 // When Dominatee is defined on function entry, it is not dominated by another
1970 // memory access.
1971 if (isLiveOnEntryDef(Dominatee))
1972 return false;
1973
1974 // When Dominator is defined on function entry, it dominates the other memory
1975 // access.
1976 if (isLiveOnEntryDef(Dominator))
1977 return true;
1978
1979 if (!BlockNumberingValid.count(DominatorBlock))
1980 renumberBlock(DominatorBlock);
1981
1982 unsigned long DominatorNum = BlockNumbering.lookup(Dominator);
1983 // All numbers start with 1
1984 assert(DominatorNum != 0 && "Block was not numbered properly")((DominatorNum != 0 && "Block was not numbered properly"
) ? static_cast<void> (0) : __assert_fail ("DominatorNum != 0 && \"Block was not numbered properly\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1984, __PRETTY_FUNCTION__))
;
1985 unsigned long DominateeNum = BlockNumbering.lookup(Dominatee);
1986 assert(DominateeNum != 0 && "Block was not numbered properly")((DominateeNum != 0 && "Block was not numbered properly"
) ? static_cast<void> (0) : __assert_fail ("DominateeNum != 0 && \"Block was not numbered properly\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 1986, __PRETTY_FUNCTION__))
;
1987 return DominatorNum < DominateeNum;
1988}
1989
1990bool MemorySSA::dominates(const MemoryAccess *Dominator,
1991 const MemoryAccess *Dominatee) const {
1992 if (Dominator == Dominatee)
1993 return true;
1994
1995 if (isLiveOnEntryDef(Dominatee))
1996 return false;
1997
1998 if (Dominator->getBlock() != Dominatee->getBlock())
1999 return DT->dominates(Dominator->getBlock(), Dominatee->getBlock());
2000 return locallyDominates(Dominator, Dominatee);
2001}
2002
2003bool MemorySSA::dominates(const MemoryAccess *Dominator,
2004 const Use &Dominatee) const {
2005 if (MemoryPhi *MP = dyn_cast<MemoryPhi>(Dominatee.getUser())) {
2006 BasicBlock *UseBB = MP->getIncomingBlock(Dominatee);
2007 // The def must dominate the incoming block of the phi.
2008 if (UseBB != Dominator->getBlock())
2009 return DT->dominates(Dominator->getBlock(), UseBB);
2010 // If the UseBB and the DefBB are the same, compare locally.
2011 return locallyDominates(Dominator, cast<MemoryAccess>(Dominatee));
2012 }
2013 // If it's not a PHI node use, the normal dominates can already handle it.
2014 return dominates(Dominator, cast<MemoryAccess>(Dominatee.getUser()));
2015}
2016
2017const static char LiveOnEntryStr[] = "liveOnEntry";
2018
2019void MemoryDef::print(raw_ostream &OS) const {
2020 MemoryAccess *UO = getDefiningAccess();
2021
2022 OS << getID() << " = MemoryDef(";
2023 if (UO && UO->getID())
2024 OS << UO->getID();
2025 else
2026 OS << LiveOnEntryStr;
2027 OS << ')';
2028}
2029
2030void MemoryPhi::print(raw_ostream &OS) const {
2031 bool First = true;
2032 OS << getID() << " = MemoryPhi(";
2033 for (const auto &Op : operands()) {
2034 BasicBlock *BB = getIncomingBlock(Op);
2035 MemoryAccess *MA = cast<MemoryAccess>(Op);
2036 if (!First)
2037 OS << ',';
2038 else
2039 First = false;
2040
2041 OS << '{';
2042 if (BB->hasName())
2043 OS << BB->getName();
2044 else
2045 BB->printAsOperand(OS, false);
2046 OS << ',';
2047 if (unsigned ID = MA->getID())
2048 OS << ID;
2049 else
2050 OS << LiveOnEntryStr;
2051 OS << '}';
2052 }
2053 OS << ')';
2054}
2055
2056MemoryAccess::~MemoryAccess() {}
2057
2058void MemoryUse::print(raw_ostream &OS) const {
2059 MemoryAccess *UO = getDefiningAccess();
2060 OS << "MemoryUse(";
2061 if (UO && UO->getID())
2062 OS << UO->getID();
2063 else
2064 OS << LiveOnEntryStr;
2065 OS << ')';
2066}
2067
2068void MemoryAccess::dump() const {
2069 print(dbgs());
2070 dbgs() << "\n";
2071}
2072
2073char MemorySSAPrinterLegacyPass::ID = 0;
2074
2075MemorySSAPrinterLegacyPass::MemorySSAPrinterLegacyPass() : FunctionPass(ID) {
2076 initializeMemorySSAPrinterLegacyPassPass(*PassRegistry::getPassRegistry());
2077}
2078
2079void MemorySSAPrinterLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const {
2080 AU.setPreservesAll();
2081 AU.addRequired<MemorySSAWrapperPass>();
2082 AU.addPreserved<MemorySSAWrapperPass>();
2083}
2084
2085bool MemorySSAPrinterLegacyPass::runOnFunction(Function &F) {
2086 auto &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA();
2087 MSSA.print(dbgs());
2088 if (VerifyMemorySSA)
2089 MSSA.verifyMemorySSA();
2090 return false;
2091}
2092
2093AnalysisKey MemorySSAAnalysis::Key;
2094
2095MemorySSAAnalysis::Result MemorySSAAnalysis::run(Function &F,
2096 FunctionAnalysisManager &AM) {
2097 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
2098 auto &AA = AM.getResult<AAManager>(F);
2099 return MemorySSAAnalysis::Result(make_unique<MemorySSA>(F, &AA, &DT));
2100}
2101
2102PreservedAnalyses MemorySSAPrinterPass::run(Function &F,
2103 FunctionAnalysisManager &AM) {
2104 OS << "MemorySSA for function: " << F.getName() << "\n";
2105 AM.getResult<MemorySSAAnalysis>(F).getMSSA().print(OS);
2106
2107 return PreservedAnalyses::all();
2108}
2109
2110PreservedAnalyses MemorySSAVerifierPass::run(Function &F,
2111 FunctionAnalysisManager &AM) {
2112 AM.getResult<MemorySSAAnalysis>(F).getMSSA().verifyMemorySSA();
2113
2114 return PreservedAnalyses::all();
2115}
2116
2117char MemorySSAWrapperPass::ID = 0;
2118
2119MemorySSAWrapperPass::MemorySSAWrapperPass() : FunctionPass(ID) {
2120 initializeMemorySSAWrapperPassPass(*PassRegistry::getPassRegistry());
2121}
2122
2123void MemorySSAWrapperPass::releaseMemory() { MSSA.reset(); }
2124
2125void MemorySSAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
2126 AU.setPreservesAll();
2127 AU.addRequiredTransitive<DominatorTreeWrapperPass>();
2128 AU.addRequiredTransitive<AAResultsWrapperPass>();
2129}
2130
2131bool MemorySSAWrapperPass::runOnFunction(Function &F) {
2132 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2133 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
2134 MSSA.reset(new MemorySSA(F, &AA, &DT));
2135 return false;
2136}
2137
2138void MemorySSAWrapperPass::verifyAnalysis() const { MSSA->verifyMemorySSA(); }
2139
2140void MemorySSAWrapperPass::print(raw_ostream &OS, const Module *M) const {
2141 MSSA->print(OS);
2142}
2143
2144MemorySSAWalker::MemorySSAWalker(MemorySSA *M) : MSSA(M) {}
2145
2146MemorySSA::CachingWalker::CachingWalker(MemorySSA *M, AliasAnalysis *A,
2147 DominatorTree *D)
2148 : MemorySSAWalker(M), Walker(*M, *A, *D, Cache), AutoResetWalker(true) {}
2149
2150MemorySSA::CachingWalker::~CachingWalker() {}
2151
2152void MemorySSA::CachingWalker::invalidateInfo(MemoryAccess *MA) {
2153 // TODO: We can do much better cache invalidation with differently stored
2154 // caches. For now, for MemoryUses, we simply remove them
2155 // from the cache, and kill the entire call/non-call cache for everything
2156 // else. The problem is for phis or defs, currently we'd need to follow use
2157 // chains down and invalidate anything below us in the chain that currently
2158 // terminates at this access.
2159
2160 // See if this is a MemoryUse, if so, just remove the cached info. MemoryUse
2161 // is by definition never a barrier, so nothing in the cache could point to
2162 // this use. In that case, we only need invalidate the info for the use
2163 // itself.
2164
2165 if (MemoryUse *MU = dyn_cast<MemoryUse>(MA)) {
2166 UpwardsMemoryQuery Q(MU->getMemoryInst(), MU);
2167 Cache.remove(MU, Q.StartingLoc, Q.IsCall);
2168 MU->resetOptimized();
2169 } else {
2170 // If it is not a use, the best we can do right now is destroy the cache.
2171 Cache.clear();
2172 }
2173
2174#ifdef EXPENSIVE_CHECKS
2175 verifyRemoved(MA);
2176#endif
2177}
2178
2179/// \brief Walk the use-def chains starting at \p MA and find
2180/// the MemoryAccess that actually clobbers Loc.
2181///
2182/// \returns our clobbering memory access
2183MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess(
2184 MemoryAccess *StartingAccess, UpwardsMemoryQuery &Q) {
2185 MemoryAccess *New = Walker.findClobber(StartingAccess, Q);
2186#ifdef EXPENSIVE_CHECKS
2187 MemoryAccess *NewNoCache =
2188 Walker.findClobber(StartingAccess, Q, /*UseWalkerCache=*/false);
2189 assert(NewNoCache == New && "Cache made us hand back a different result?")((NewNoCache == New && "Cache made us hand back a different result?"
) ? static_cast<void> (0) : __assert_fail ("NewNoCache == New && \"Cache made us hand back a different result?\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 2189, __PRETTY_FUNCTION__))
;
2190#endif
2191 if (AutoResetWalker)
2192 resetClobberWalker();
2193 return New;
2194}
2195
2196MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess(
2197 MemoryAccess *StartingAccess, const MemoryLocation &Loc) {
2198 if (isa<MemoryPhi>(StartingAccess))
2199 return StartingAccess;
2200
2201 auto *StartingUseOrDef = cast<MemoryUseOrDef>(StartingAccess);
2202 if (MSSA->isLiveOnEntryDef(StartingUseOrDef))
2203 return StartingUseOrDef;
2204
2205 Instruction *I = StartingUseOrDef->getMemoryInst();
2206
2207 // Conservatively, fences are always clobbers, so don't perform the walk if we
2208 // hit a fence.
2209 if (!ImmutableCallSite(I) && I->isFenceLike())
2210 return StartingUseOrDef;
2211
2212 UpwardsMemoryQuery Q;
2213 Q.OriginalAccess = StartingUseOrDef;
2214 Q.StartingLoc = Loc;
2215 Q.Inst = I;
2216 Q.IsCall = false;
2217
2218 if (auto *CacheResult = Cache.lookup(StartingUseOrDef, Loc, Q.IsCall))
2219 return CacheResult;
2220
2221 // Unlike the other function, do not walk to the def of a def, because we are
2222 // handed something we already believe is the clobbering access.
2223 MemoryAccess *DefiningAccess = isa<MemoryUse>(StartingUseOrDef)
2224 ? StartingUseOrDef->getDefiningAccess()
2225 : StartingUseOrDef;
2226
2227 MemoryAccess *Clobber = getClobberingMemoryAccess(DefiningAccess, Q);
2228 DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memoryssa")) { dbgs() << "Starting Memory SSA clobber for "
<< *I << " is "; } } while (false)
;
2229 DEBUG(dbgs() << *StartingUseOrDef << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memoryssa")) { dbgs() << *StartingUseOrDef << "\n"
; } } while (false)
;
2230 DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memoryssa")) { dbgs() << "Final Memory SSA clobber for "
<< *I << " is "; } } while (false)
;
2231 DEBUG(dbgs() << *Clobber << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memoryssa")) { dbgs() << *Clobber << "\n"; } } while
(false)
;
2232 return Clobber;
2233}
2234
2235MemoryAccess *
2236MemorySSA::CachingWalker::getClobberingMemoryAccess(MemoryAccess *MA) {
2237 auto *StartingAccess = dyn_cast<MemoryUseOrDef>(MA);
2238 // If this is a MemoryPhi, we can't do anything.
2239 if (!StartingAccess)
2240 return MA;
2241
2242 // If this is an already optimized use or def, return the optimized result.
2243 // Note: Currently, we do not store the optimized def result because we'd need
2244 // a separate field, since we can't use it as the defining access.
2245 if (MemoryUse *MU = dyn_cast<MemoryUse>(StartingAccess))
2246 if (MU->isOptimized())
2247 return MU->getDefiningAccess();
2248
2249 const Instruction *I = StartingAccess->getMemoryInst();
2250 UpwardsMemoryQuery Q(I, StartingAccess);
2251 // We can't sanely do anything with a fences, they conservatively
2252 // clobber all memory, and have no locations to get pointers from to
2253 // try to disambiguate.
2254 if (!Q.IsCall && I->isFenceLike())
2255 return StartingAccess;
2256
2257 if (auto *CacheResult = Cache.lookup(StartingAccess, Q.StartingLoc, Q.IsCall))
2258 return CacheResult;
2259
2260 if (isUseTriviallyOptimizableToLiveOnEntry(*MSSA->AA, I)) {
2261 MemoryAccess *LiveOnEntry = MSSA->getLiveOnEntryDef();
2262 Cache.insert(StartingAccess, LiveOnEntry, Q.StartingLoc, Q.IsCall);
2263 if (MemoryUse *MU = dyn_cast<MemoryUse>(StartingAccess))
2264 MU->setDefiningAccess(LiveOnEntry, true);
2265 return LiveOnEntry;
2266 }
2267
2268 // Start with the thing we already think clobbers this location
2269 MemoryAccess *DefiningAccess = StartingAccess->getDefiningAccess();
2270
2271 // At this point, DefiningAccess may be the live on entry def.
2272 // If it is, we will not get a better result.
2273 if (MSSA->isLiveOnEntryDef(DefiningAccess))
2274 return DefiningAccess;
2275
2276 MemoryAccess *Result = getClobberingMemoryAccess(DefiningAccess, Q);
2277 DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memoryssa")) { dbgs() << "Starting Memory SSA clobber for "
<< *I << " is "; } } while (false)
;
2278 DEBUG(dbgs() << *DefiningAccess << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memoryssa")) { dbgs() << *DefiningAccess << "\n"
; } } while (false)
;
2279 DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memoryssa")) { dbgs() << "Final Memory SSA clobber for "
<< *I << " is "; } } while (false)
;
2280 DEBUG(dbgs() << *Result << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memoryssa")) { dbgs() << *Result << "\n"; } } while
(false)
;
2281 if (MemoryUse *MU = dyn_cast<MemoryUse>(StartingAccess))
2282 MU->setDefiningAccess(Result, true);
2283
2284 return Result;
2285}
2286
2287// Verify that MA doesn't exist in any of the caches.
2288void MemorySSA::CachingWalker::verifyRemoved(MemoryAccess *MA) {
2289 assert(!Cache.contains(MA) && "Found removed MemoryAccess in cache.")((!Cache.contains(MA) && "Found removed MemoryAccess in cache."
) ? static_cast<void> (0) : __assert_fail ("!Cache.contains(MA) && \"Found removed MemoryAccess in cache.\""
, "/tmp/buildd/llvm-toolchain-snapshot-4.0~svn290870/lib/Transforms/Utils/MemorySSA.cpp"
, 2289, __PRETTY_FUNCTION__))
;
2290}
2291
2292MemoryAccess *
2293DoNothingMemorySSAWalker::getClobberingMemoryAccess(MemoryAccess *MA) {
2294 if (auto *Use = dyn_cast<MemoryUseOrDef>(MA))
2295 return Use->getDefiningAccess();
2296 return MA;
2297}
2298
2299MemoryAccess *DoNothingMemorySSAWalker::getClobberingMemoryAccess(
2300 MemoryAccess *StartingAccess, const MemoryLocation &) {
2301 if (auto *Use = dyn_cast<MemoryUseOrDef>(StartingAccess))
2302 return Use->getDefiningAccess();
2303 return StartingAccess;
2304}
2305} // namespace llvm