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