blob: 85d21b1007e33a17d6346400efbde9e46ffdab4f [file] [log] [blame]
George Burgess IVe1100f52016-02-02 22:46:49 +00001//===-- 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//===----------------------------------------------------------------===//
Mehdi Aminib550cb12016-04-18 09:17:29 +000013#include "llvm/Transforms/Utils/MemorySSA.h"
George Burgess IVe1100f52016-02-02 22:46:49 +000014#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/SmallPtrSet.h"
21#include "llvm/ADT/SmallSet.h"
22#include "llvm/ADT/Statistic.h"
23#include "llvm/Analysis/AliasAnalysis.h"
24#include "llvm/Analysis/CFG.h"
25#include "llvm/Analysis/GlobalsModRef.h"
26#include "llvm/Analysis/IteratedDominanceFrontier.h"
27#include "llvm/Analysis/MemoryLocation.h"
28#include "llvm/Analysis/PHITransAddr.h"
29#include "llvm/IR/AssemblyAnnotationWriter.h"
30#include "llvm/IR/DataLayout.h"
31#include "llvm/IR/Dominators.h"
32#include "llvm/IR/GlobalVariable.h"
33#include "llvm/IR/IRBuilder.h"
34#include "llvm/IR/IntrinsicInst.h"
35#include "llvm/IR/LLVMContext.h"
36#include "llvm/IR/Metadata.h"
37#include "llvm/IR/Module.h"
38#include "llvm/IR/PatternMatch.h"
George Burgess IVe1100f52016-02-02 22:46:49 +000039#include "llvm/Support/Debug.h"
40#include "llvm/Support/FormattedStream.h"
41#include "llvm/Transforms/Scalar.h"
George Burgess IVe1100f52016-02-02 22:46:49 +000042#include <algorithm>
43
44#define DEBUG_TYPE "memoryssa"
45using namespace llvm;
46STATISTIC(NumClobberCacheLookups, "Number of Memory SSA version cache lookups");
47STATISTIC(NumClobberCacheHits, "Number of Memory SSA version cache hits");
48STATISTIC(NumClobberCacheInserts, "Number of MemorySSA version cache inserts");
Geoff Berryb96d3b22016-06-01 21:30:40 +000049
50INITIALIZE_PASS_BEGIN(MemorySSAWrapperPass, "memoryssa", "Memory SSA", true,
51 true)
George Burgess IVe1100f52016-02-02 22:46:49 +000052INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
53INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
Geoff Berryb96d3b22016-06-01 21:30:40 +000054INITIALIZE_PASS_END(MemorySSAWrapperPass, "memoryssa", "Memory SSA", true, true)
George Burgess IVe1100f52016-02-02 22:46:49 +000055
56namespace llvm {
57
58/// \brief An assembly annotator class to print Memory SSA information in
59/// comments.
60class MemorySSAAnnotatedWriter : public AssemblyAnnotationWriter {
61 friend class MemorySSA;
62 const MemorySSA *MSSA;
63
64public:
65 MemorySSAAnnotatedWriter(const MemorySSA *M) : MSSA(M) {}
66
67 virtual void emitBasicBlockStartAnnot(const BasicBlock *BB,
68 formatted_raw_ostream &OS) {
69 if (MemoryAccess *MA = MSSA->getMemoryAccess(BB))
70 OS << "; " << *MA << "\n";
71 }
72
73 virtual void emitInstructionAnnot(const Instruction *I,
74 formatted_raw_ostream &OS) {
75 if (MemoryAccess *MA = MSSA->getMemoryAccess(I))
76 OS << "; " << *MA << "\n";
77 }
78};
79}
80
81namespace {
82struct RenamePassData {
83 DomTreeNode *DTN;
84 DomTreeNode::const_iterator ChildIt;
85 MemoryAccess *IncomingVal;
86
87 RenamePassData(DomTreeNode *D, DomTreeNode::const_iterator It,
88 MemoryAccess *M)
89 : DTN(D), ChildIt(It), IncomingVal(M) {}
90 void swap(RenamePassData &RHS) {
91 std::swap(DTN, RHS.DTN);
92 std::swap(ChildIt, RHS.ChildIt);
93 std::swap(IncomingVal, RHS.IncomingVal);
94 }
95};
96}
97
98namespace llvm {
99/// \brief Rename a single basic block into MemorySSA form.
100/// Uses the standard SSA renaming algorithm.
101/// \returns The new incoming value.
102MemoryAccess *MemorySSA::renameBlock(BasicBlock *BB,
103 MemoryAccess *IncomingVal) {
104 auto It = PerBlockAccesses.find(BB);
105 // Skip most processing if the list is empty.
106 if (It != PerBlockAccesses.end()) {
107 AccessListType *Accesses = It->second.get();
108 for (MemoryAccess &L : *Accesses) {
109 switch (L.getValueID()) {
110 case Value::MemoryUseVal:
111 cast<MemoryUse>(&L)->setDefiningAccess(IncomingVal);
112 break;
113 case Value::MemoryDefVal:
114 // We can't legally optimize defs, because we only allow single
115 // memory phis/uses on operations, and if we optimize these, we can
116 // end up with multiple reaching defs. Uses do not have this
117 // problem, since they do not produce a value
118 cast<MemoryDef>(&L)->setDefiningAccess(IncomingVal);
119 IncomingVal = &L;
120 break;
121 case Value::MemoryPhiVal:
122 IncomingVal = &L;
123 break;
124 }
125 }
126 }
127
128 // Pass through values to our successors
129 for (const BasicBlock *S : successors(BB)) {
130 auto It = PerBlockAccesses.find(S);
131 // Rename the phi nodes in our successor block
132 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front()))
133 continue;
134 AccessListType *Accesses = It->second.get();
135 auto *Phi = cast<MemoryPhi>(&Accesses->front());
136 assert(std::find(succ_begin(BB), succ_end(BB), S) != succ_end(BB) &&
137 "Must be at least one edge from Succ to BB!");
138 Phi->addIncoming(IncomingVal, BB);
139 }
140
141 return IncomingVal;
142}
143
144/// \brief This is the standard SSA renaming algorithm.
145///
146/// We walk the dominator tree in preorder, renaming accesses, and then filling
147/// in phi nodes in our successors.
148void MemorySSA::renamePass(DomTreeNode *Root, MemoryAccess *IncomingVal,
149 SmallPtrSet<BasicBlock *, 16> &Visited) {
150 SmallVector<RenamePassData, 32> WorkStack;
151 IncomingVal = renameBlock(Root->getBlock(), IncomingVal);
152 WorkStack.push_back({Root, Root->begin(), IncomingVal});
153 Visited.insert(Root->getBlock());
154
155 while (!WorkStack.empty()) {
156 DomTreeNode *Node = WorkStack.back().DTN;
157 DomTreeNode::const_iterator ChildIt = WorkStack.back().ChildIt;
158 IncomingVal = WorkStack.back().IncomingVal;
159
160 if (ChildIt == Node->end()) {
161 WorkStack.pop_back();
162 } else {
163 DomTreeNode *Child = *ChildIt;
164 ++WorkStack.back().ChildIt;
165 BasicBlock *BB = Child->getBlock();
166 Visited.insert(BB);
167 IncomingVal = renameBlock(BB, IncomingVal);
168 WorkStack.push_back({Child, Child->begin(), IncomingVal});
169 }
170 }
171}
172
173/// \brief Compute dominator levels, used by the phi insertion algorithm above.
174void MemorySSA::computeDomLevels(DenseMap<DomTreeNode *, unsigned> &DomLevels) {
175 for (auto DFI = df_begin(DT->getRootNode()), DFE = df_end(DT->getRootNode());
176 DFI != DFE; ++DFI)
177 DomLevels[*DFI] = DFI.getPathLength() - 1;
178}
179
180/// \brief This handles unreachable block acccesses by deleting phi nodes in
181/// unreachable blocks, and marking all other unreachable MemoryAccess's as
182/// being uses of the live on entry definition.
183void MemorySSA::markUnreachableAsLiveOnEntry(BasicBlock *BB) {
184 assert(!DT->isReachableFromEntry(BB) &&
185 "Reachable block found while handling unreachable blocks");
186
187 auto It = PerBlockAccesses.find(BB);
188 if (It == PerBlockAccesses.end())
189 return;
190
191 auto &Accesses = It->second;
192 for (auto AI = Accesses->begin(), AE = Accesses->end(); AI != AE;) {
193 auto Next = std::next(AI);
194 // If we have a phi, just remove it. We are going to replace all
195 // users with live on entry.
196 if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(AI))
197 UseOrDef->setDefiningAccess(LiveOnEntryDef.get());
198 else
199 Accesses->erase(AI);
200 AI = Next;
201 }
202}
203
Geoff Berryb96d3b22016-06-01 21:30:40 +0000204MemorySSA::MemorySSA(Function &Func, AliasAnalysis *AA, DominatorTree *DT)
205 : AA(AA), DT(DT), F(Func), LiveOnEntryDef(nullptr), Walker(nullptr),
206 NextID(0) {
207 getWalker(); // Ensure MemorySSA has been built.
208}
209
210MemorySSA::MemorySSA(MemorySSA &&MSSA)
211 : AA(MSSA.AA), DT(MSSA.DT), F(MSSA.F),
212 ValueToMemoryAccess(std::move(MSSA.ValueToMemoryAccess)),
213 PerBlockAccesses(std::move(MSSA.PerBlockAccesses)),
214 LiveOnEntryDef(std::move(MSSA.LiveOnEntryDef)),
215 Walker(std::move(MSSA.Walker)), NextID(MSSA.NextID) {
216 // Update the Walker MSSA pointer so it doesn't point to the moved-from MSSA
217 // object any more.
218 Walker->MSSA = this;
219}
George Burgess IVe1100f52016-02-02 22:46:49 +0000220
221MemorySSA::~MemorySSA() {
222 // Drop all our references
223 for (const auto &Pair : PerBlockAccesses)
224 for (MemoryAccess &MA : *Pair.second)
225 MA.dropAllReferences();
226}
227
228MemorySSA::AccessListType *MemorySSA::getOrCreateAccessList(BasicBlock *BB) {
229 auto Res = PerBlockAccesses.insert(std::make_pair(BB, nullptr));
230
231 if (Res.second)
232 Res.first->second = make_unique<AccessListType>();
233 return Res.first->second.get();
234}
235
Geoff Berryb96d3b22016-06-01 21:30:40 +0000236MemorySSAWalker *MemorySSA::getWalker() {
George Burgess IVe1100f52016-02-02 22:46:49 +0000237 if (Walker)
Geoff Berryb96d3b22016-06-01 21:30:40 +0000238 return Walker.get();
George Burgess IVe1100f52016-02-02 22:46:49 +0000239
Geoff Berryb96d3b22016-06-01 21:30:40 +0000240 Walker = make_unique<CachingMemorySSAWalker>(this, AA, DT);
George Burgess IVe1100f52016-02-02 22:46:49 +0000241
242 // We create an access to represent "live on entry", for things like
243 // arguments or users of globals, where the memory they use is defined before
244 // the beginning of the function. We do not actually insert it into the IR.
245 // We do not define a live on exit for the immediate uses, and thus our
246 // semantics do *not* imply that something with no immediate uses can simply
247 // be removed.
248 BasicBlock &StartingPoint = F.getEntryBlock();
249 LiveOnEntryDef = make_unique<MemoryDef>(F.getContext(), nullptr, nullptr,
250 &StartingPoint, NextID++);
251
252 // We maintain lists of memory accesses per-block, trading memory for time. We
253 // could just look up the memory access for every possible instruction in the
254 // stream.
255 SmallPtrSet<BasicBlock *, 32> DefiningBlocks;
Daniel Berlin1b51a292016-02-07 01:52:19 +0000256 SmallPtrSet<BasicBlock *, 32> DefUseBlocks;
George Burgess IVe1100f52016-02-02 22:46:49 +0000257 // Go through each block, figure out where defs occur, and chain together all
258 // the accesses.
259 for (BasicBlock &B : F) {
Daniel Berlin7898ca62016-02-07 01:52:15 +0000260 bool InsertIntoDef = false;
George Burgess IVe1100f52016-02-02 22:46:49 +0000261 AccessListType *Accesses = nullptr;
262 for (Instruction &I : B) {
Peter Collingbourneffecb142016-05-26 01:19:17 +0000263 MemoryUseOrDef *MUD = createNewAccess(&I);
George Burgess IVb42b7622016-03-11 19:34:03 +0000264 if (!MUD)
George Burgess IVe1100f52016-02-02 22:46:49 +0000265 continue;
George Burgess IV3887a412016-03-21 21:25:39 +0000266 InsertIntoDef |= isa<MemoryDef>(MUD);
Daniel Berlin1b51a292016-02-07 01:52:19 +0000267
George Burgess IVe1100f52016-02-02 22:46:49 +0000268 if (!Accesses)
269 Accesses = getOrCreateAccessList(&B);
George Burgess IVb42b7622016-03-11 19:34:03 +0000270 Accesses->push_back(MUD);
George Burgess IVe1100f52016-02-02 22:46:49 +0000271 }
Daniel Berlin7898ca62016-02-07 01:52:15 +0000272 if (InsertIntoDef)
273 DefiningBlocks.insert(&B);
George Burgess IV3887a412016-03-21 21:25:39 +0000274 if (Accesses)
Daniel Berlin1b51a292016-02-07 01:52:19 +0000275 DefUseBlocks.insert(&B);
276 }
277
278 // Compute live-in.
279 // Live in is normally defined as "all the blocks on the path from each def to
280 // each of it's uses".
281 // MemoryDef's are implicit uses of previous state, so they are also uses.
282 // This means we don't really have def-only instructions. The only
283 // MemoryDef's that are not really uses are those that are of the LiveOnEntry
284 // variable (because LiveOnEntry can reach anywhere, and every def is a
285 // must-kill of LiveOnEntry).
286 // In theory, you could precisely compute live-in by using alias-analysis to
287 // disambiguate defs and uses to see which really pair up with which.
288 // In practice, this would be really expensive and difficult. So we simply
289 // assume all defs are also uses that need to be kept live.
290 // Because of this, the end result of this live-in computation will be "the
291 // entire set of basic blocks that reach any use".
292
293 SmallPtrSet<BasicBlock *, 32> LiveInBlocks;
294 SmallVector<BasicBlock *, 64> LiveInBlockWorklist(DefUseBlocks.begin(),
295 DefUseBlocks.end());
296 // Now that we have a set of blocks where a value is live-in, recursively add
297 // predecessors until we find the full region the value is live.
298 while (!LiveInBlockWorklist.empty()) {
299 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
300
301 // The block really is live in here, insert it into the set. If already in
302 // the set, then it has already been processed.
303 if (!LiveInBlocks.insert(BB).second)
304 continue;
305
306 // Since the value is live into BB, it is either defined in a predecessor or
307 // live into it to.
308 LiveInBlockWorklist.append(pred_begin(BB), pred_end(BB));
George Burgess IVe1100f52016-02-02 22:46:49 +0000309 }
310
311 // Determine where our MemoryPhi's should go
Daniel Berlin77fa84e2016-04-19 06:13:28 +0000312 ForwardIDFCalculator IDFs(*DT);
George Burgess IVe1100f52016-02-02 22:46:49 +0000313 IDFs.setDefiningBlocks(DefiningBlocks);
Daniel Berlin1b51a292016-02-07 01:52:19 +0000314 IDFs.setLiveInBlocks(LiveInBlocks);
George Burgess IVe1100f52016-02-02 22:46:49 +0000315 SmallVector<BasicBlock *, 32> IDFBlocks;
316 IDFs.calculate(IDFBlocks);
317
318 // Now place MemoryPhi nodes.
319 for (auto &BB : IDFBlocks) {
320 // Insert phi node
321 AccessListType *Accesses = getOrCreateAccessList(BB);
322 MemoryPhi *Phi = new MemoryPhi(F.getContext(), BB, NextID++);
Daniel Berlinf6c9ae92016-02-10 17:41:25 +0000323 ValueToMemoryAccess.insert(std::make_pair(BB, Phi));
George Burgess IVe1100f52016-02-02 22:46:49 +0000324 // Phi's always are placed at the front of the block.
325 Accesses->push_front(Phi);
326 }
327
328 // Now do regular SSA renaming on the MemoryDef/MemoryUse. Visited will get
329 // filled in with all blocks.
330 SmallPtrSet<BasicBlock *, 16> Visited;
331 renamePass(DT->getRootNode(), LiveOnEntryDef.get(), Visited);
332
333 // Now optimize the MemoryUse's defining access to point to the nearest
334 // dominating clobbering def.
335 // This ensures that MemoryUse's that are killed by the same store are
336 // immediate users of that store, one of the invariants we guarantee.
337 for (auto DomNode : depth_first(DT)) {
338 BasicBlock *BB = DomNode->getBlock();
339 auto AI = PerBlockAccesses.find(BB);
340 if (AI == PerBlockAccesses.end())
341 continue;
342 AccessListType *Accesses = AI->second.get();
343 for (auto &MA : *Accesses) {
344 if (auto *MU = dyn_cast<MemoryUse>(&MA)) {
345 Instruction *Inst = MU->getMemoryInst();
Daniel Berlin64120022016-03-02 21:16:28 +0000346 MU->setDefiningAccess(Walker->getClobberingMemoryAccess(Inst));
George Burgess IVe1100f52016-02-02 22:46:49 +0000347 }
348 }
349 }
350
351 // Mark the uses in unreachable blocks as live on entry, so that they go
352 // somewhere.
353 for (auto &BB : F)
354 if (!Visited.count(&BB))
355 markUnreachableAsLiveOnEntry(&BB);
356
Geoff Berryb96d3b22016-06-01 21:30:40 +0000357 return Walker.get();
George Burgess IVe1100f52016-02-02 22:46:49 +0000358}
359
360/// \brief Helper function to create new memory accesses
Peter Collingbourneffecb142016-05-26 01:19:17 +0000361MemoryUseOrDef *MemorySSA::createNewAccess(Instruction *I) {
Peter Collingbourneb9aa1f42016-05-26 04:58:46 +0000362 // The assume intrinsic has a control dependency which we model by claiming
363 // that it writes arbitrarily. Ignore that fake memory dependency here.
364 // FIXME: Replace this special casing with a more accurate modelling of
365 // assume's control dependency.
366 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
367 if (II->getIntrinsicID() == Intrinsic::assume)
368 return nullptr;
369
George Burgess IVe1100f52016-02-02 22:46:49 +0000370 // Find out what affect this instruction has on memory.
371 ModRefInfo ModRef = AA->getModRefInfo(I);
372 bool Def = bool(ModRef & MRI_Mod);
373 bool Use = bool(ModRef & MRI_Ref);
374
375 // It's possible for an instruction to not modify memory at all. During
376 // construction, we ignore them.
Peter Collingbourneffecb142016-05-26 01:19:17 +0000377 if (!Def && !Use)
George Burgess IVe1100f52016-02-02 22:46:49 +0000378 return nullptr;
379
380 assert((Def || Use) &&
381 "Trying to create a memory access with a non-memory instruction");
382
George Burgess IVb42b7622016-03-11 19:34:03 +0000383 MemoryUseOrDef *MUD;
George Burgess IVe1100f52016-02-02 22:46:49 +0000384 if (Def)
George Burgess IVb42b7622016-03-11 19:34:03 +0000385 MUD = new MemoryDef(I->getContext(), nullptr, I, I->getParent(), NextID++);
George Burgess IVe1100f52016-02-02 22:46:49 +0000386 else
George Burgess IVb42b7622016-03-11 19:34:03 +0000387 MUD = new MemoryUse(I->getContext(), nullptr, I, I->getParent());
388 ValueToMemoryAccess.insert(std::make_pair(I, MUD));
389 return MUD;
George Burgess IVe1100f52016-02-02 22:46:49 +0000390}
391
392MemoryAccess *MemorySSA::findDominatingDef(BasicBlock *UseBlock,
393 enum InsertionPlace Where) {
394 // Handle the initial case
395 if (Where == Beginning)
396 // The only thing that could define us at the beginning is a phi node
397 if (MemoryPhi *Phi = getMemoryAccess(UseBlock))
398 return Phi;
399
400 DomTreeNode *CurrNode = DT->getNode(UseBlock);
401 // Need to be defined by our dominator
402 if (Where == Beginning)
403 CurrNode = CurrNode->getIDom();
404 Where = End;
405 while (CurrNode) {
406 auto It = PerBlockAccesses.find(CurrNode->getBlock());
407 if (It != PerBlockAccesses.end()) {
408 auto &Accesses = It->second;
409 for (auto RAI = Accesses->rbegin(), RAE = Accesses->rend(); RAI != RAE;
410 ++RAI) {
411 if (isa<MemoryDef>(*RAI) || isa<MemoryPhi>(*RAI))
412 return &*RAI;
413 }
414 }
415 CurrNode = CurrNode->getIDom();
416 }
417 return LiveOnEntryDef.get();
418}
419
420/// \brief Returns true if \p Replacer dominates \p Replacee .
421bool MemorySSA::dominatesUse(const MemoryAccess *Replacer,
422 const MemoryAccess *Replacee) const {
423 if (isa<MemoryUseOrDef>(Replacee))
424 return DT->dominates(Replacer->getBlock(), Replacee->getBlock());
425 const auto *MP = cast<MemoryPhi>(Replacee);
426 // For a phi node, the use occurs in the predecessor block of the phi node.
427 // Since we may occur multiple times in the phi node, we have to check each
428 // operand to ensure Replacer dominates each operand where Replacee occurs.
429 for (const Use &Arg : MP->operands()) {
George Burgess IVb5a229f2016-02-02 23:15:26 +0000430 if (Arg.get() != Replacee &&
George Burgess IVe1100f52016-02-02 22:46:49 +0000431 !DT->dominates(Replacer->getBlock(), MP->getIncomingBlock(Arg)))
432 return false;
433 }
434 return true;
435}
436
Daniel Berlin83fc77b2016-03-01 18:46:54 +0000437/// \brief If all arguments of a MemoryPHI are defined by the same incoming
438/// argument, return that argument.
439static MemoryAccess *onlySingleValue(MemoryPhi *MP) {
440 MemoryAccess *MA = nullptr;
441
442 for (auto &Arg : MP->operands()) {
443 if (!MA)
444 MA = cast<MemoryAccess>(Arg);
445 else if (MA != Arg)
446 return nullptr;
447 }
448 return MA;
449}
450
451/// \brief Properly remove \p MA from all of MemorySSA's lookup tables.
452///
453/// Because of the way the intrusive list and use lists work, it is important to
454/// do removal in the right order.
455void MemorySSA::removeFromLookups(MemoryAccess *MA) {
456 assert(MA->use_empty() &&
457 "Trying to remove memory access that still has uses");
458 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA))
459 MUD->setDefiningAccess(nullptr);
460 // Invalidate our walker's cache if necessary
461 if (!isa<MemoryUse>(MA))
462 Walker->invalidateInfo(MA);
463 // The call below to erase will destroy MA, so we can't change the order we
464 // are doing things here
465 Value *MemoryInst;
466 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA)) {
467 MemoryInst = MUD->getMemoryInst();
468 } else {
469 MemoryInst = MA->getBlock();
470 }
471 ValueToMemoryAccess.erase(MemoryInst);
472
George Burgess IVe0e6e482016-03-02 02:35:04 +0000473 auto AccessIt = PerBlockAccesses.find(MA->getBlock());
474 std::unique_ptr<AccessListType> &Accesses = AccessIt->second;
Daniel Berlin83fc77b2016-03-01 18:46:54 +0000475 Accesses->erase(MA);
George Burgess IVe0e6e482016-03-02 02:35:04 +0000476 if (Accesses->empty())
477 PerBlockAccesses.erase(AccessIt);
Daniel Berlin83fc77b2016-03-01 18:46:54 +0000478}
479
480void MemorySSA::removeMemoryAccess(MemoryAccess *MA) {
481 assert(!isLiveOnEntryDef(MA) && "Trying to remove the live on entry def");
482 // We can only delete phi nodes if they have no uses, or we can replace all
483 // uses with a single definition.
484 MemoryAccess *NewDefTarget = nullptr;
485 if (MemoryPhi *MP = dyn_cast<MemoryPhi>(MA)) {
486 // Note that it is sufficient to know that all edges of the phi node have
487 // the same argument. If they do, by the definition of dominance frontiers
488 // (which we used to place this phi), that argument must dominate this phi,
489 // and thus, must dominate the phi's uses, and so we will not hit the assert
490 // below.
491 NewDefTarget = onlySingleValue(MP);
492 assert((NewDefTarget || MP->use_empty()) &&
493 "We can't delete this memory phi");
494 } else {
495 NewDefTarget = cast<MemoryUseOrDef>(MA)->getDefiningAccess();
496 }
497
498 // Re-point the uses at our defining access
499 if (!MA->use_empty())
500 MA->replaceAllUsesWith(NewDefTarget);
501
502 // The call below to erase will destroy MA, so we can't change the order we
503 // are doing things here
504 removeFromLookups(MA);
505}
506
George Burgess IVe1100f52016-02-02 22:46:49 +0000507void MemorySSA::print(raw_ostream &OS) const {
508 MemorySSAAnnotatedWriter Writer(this);
509 F.print(OS, &Writer);
510}
511
512void MemorySSA::dump() const {
513 MemorySSAAnnotatedWriter Writer(this);
514 F.print(dbgs(), &Writer);
515}
516
Daniel Berlin932b4cb2016-02-10 17:39:43 +0000517void MemorySSA::verifyMemorySSA() const {
518 verifyDefUses(F);
519 verifyDomination(F);
520}
521
George Burgess IVe1100f52016-02-02 22:46:49 +0000522/// \brief Verify the domination properties of MemorySSA by checking that each
523/// definition dominates all of its uses.
Daniel Berlin932b4cb2016-02-10 17:39:43 +0000524void MemorySSA::verifyDomination(Function &F) const {
George Burgess IVe1100f52016-02-02 22:46:49 +0000525 for (BasicBlock &B : F) {
526 // Phi nodes are attached to basic blocks
527 if (MemoryPhi *MP = getMemoryAccess(&B)) {
528 for (User *U : MP->users()) {
529 BasicBlock *UseBlock;
530 // Phi operands are used on edges, we simulate the right domination by
531 // acting as if the use occurred at the end of the predecessor block.
532 if (MemoryPhi *P = dyn_cast<MemoryPhi>(U)) {
533 for (const auto &Arg : P->operands()) {
534 if (Arg == MP) {
535 UseBlock = P->getIncomingBlock(Arg);
536 break;
537 }
538 }
539 } else {
540 UseBlock = cast<MemoryAccess>(U)->getBlock();
541 }
George Burgess IV60adac42016-02-02 23:26:01 +0000542 (void)UseBlock;
George Burgess IVe1100f52016-02-02 22:46:49 +0000543 assert(DT->dominates(MP->getBlock(), UseBlock) &&
544 "Memory PHI does not dominate it's uses");
545 }
546 }
547
548 for (Instruction &I : B) {
549 MemoryAccess *MD = dyn_cast_or_null<MemoryDef>(getMemoryAccess(&I));
550 if (!MD)
551 continue;
552
Benjamin Kramer451f54c2016-02-22 13:11:58 +0000553 for (User *U : MD->users()) {
Mehdi Amini89038a12016-04-02 05:34:19 +0000554 BasicBlock *UseBlock; (void)UseBlock;
George Burgess IVe1100f52016-02-02 22:46:49 +0000555 // Things are allowed to flow to phi nodes over their predecessor edge.
556 if (auto *P = dyn_cast<MemoryPhi>(U)) {
557 for (const auto &Arg : P->operands()) {
558 if (Arg == MD) {
559 UseBlock = P->getIncomingBlock(Arg);
560 break;
561 }
562 }
563 } else {
564 UseBlock = cast<MemoryAccess>(U)->getBlock();
565 }
566 assert(DT->dominates(MD->getBlock(), UseBlock) &&
567 "Memory Def does not dominate it's uses");
568 }
569 }
570 }
571}
572
573/// \brief Verify the def-use lists in MemorySSA, by verifying that \p Use
574/// appears in the use list of \p Def.
575///
576/// llvm_unreachable is used instead of asserts because this may be called in
577/// a build without asserts. In that case, we don't want this to turn into a
578/// nop.
Daniel Berlin932b4cb2016-02-10 17:39:43 +0000579void MemorySSA::verifyUseInDefs(MemoryAccess *Def, MemoryAccess *Use) const {
George Burgess IVe1100f52016-02-02 22:46:49 +0000580 // The live on entry use may cause us to get a NULL def here
581 if (!Def) {
582 if (!isLiveOnEntryDef(Use))
583 llvm_unreachable("Null def but use not point to live on entry def");
584 } else if (std::find(Def->user_begin(), Def->user_end(), Use) ==
585 Def->user_end()) {
586 llvm_unreachable("Did not find use in def's use list");
587 }
588}
589
590/// \brief Verify the immediate use information, by walking all the memory
591/// accesses and verifying that, for each use, it appears in the
592/// appropriate def's use list
Daniel Berlin932b4cb2016-02-10 17:39:43 +0000593void MemorySSA::verifyDefUses(Function &F) const {
George Burgess IVe1100f52016-02-02 22:46:49 +0000594 for (BasicBlock &B : F) {
595 // Phi nodes are attached to basic blocks
596 if (MemoryPhi *Phi = getMemoryAccess(&B))
597 for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I)
598 verifyUseInDefs(Phi->getIncomingValue(I), Phi);
599
600 for (Instruction &I : B) {
601 if (MemoryAccess *MA = getMemoryAccess(&I)) {
602 assert(isa<MemoryUseOrDef>(MA) &&
603 "Found a phi node not attached to a bb");
604 verifyUseInDefs(cast<MemoryUseOrDef>(MA)->getDefiningAccess(), MA);
605 }
606 }
607 }
608}
609
610MemoryAccess *MemorySSA::getMemoryAccess(const Value *I) const {
Daniel Berlinf6c9ae92016-02-10 17:41:25 +0000611 return ValueToMemoryAccess.lookup(I);
George Burgess IVe1100f52016-02-02 22:46:49 +0000612}
613
614MemoryPhi *MemorySSA::getMemoryAccess(const BasicBlock *BB) const {
615 return cast_or_null<MemoryPhi>(getMemoryAccess((const Value *)BB));
616}
617
618/// \brief Determine, for two memory accesses in the same block,
619/// whether \p Dominator dominates \p Dominatee.
620/// \returns True if \p Dominator dominates \p Dominatee.
621bool MemorySSA::locallyDominates(const MemoryAccess *Dominator,
622 const MemoryAccess *Dominatee) const {
623
624 assert((Dominator->getBlock() == Dominatee->getBlock()) &&
625 "Asking for local domination when accesses are in different blocks!");
Sebastian Pope1f60b12016-06-10 21:36:41 +0000626
627 // A node dominates itself.
628 if (Dominatee == Dominator)
629 return true;
630
631 // When Dominatee is defined on function entry, it is not dominated by another
632 // memory access.
633 if (isLiveOnEntryDef(Dominatee))
634 return false;
635
636 // When Dominator is defined on function entry, it dominates the other memory
637 // access.
638 if (isLiveOnEntryDef(Dominator))
639 return true;
640
George Burgess IVe1100f52016-02-02 22:46:49 +0000641 // Get the access list for the block
642 const AccessListType *AccessList = getBlockAccesses(Dominator->getBlock());
643 AccessListType::const_reverse_iterator It(Dominator->getIterator());
644
645 // If we hit the beginning of the access list before we hit dominatee, we must
646 // dominate it
647 return std::none_of(It, AccessList->rend(),
648 [&](const MemoryAccess &MA) { return &MA == Dominatee; });
649}
650
651const static char LiveOnEntryStr[] = "liveOnEntry";
652
653void MemoryDef::print(raw_ostream &OS) const {
654 MemoryAccess *UO = getDefiningAccess();
655
656 OS << getID() << " = MemoryDef(";
657 if (UO && UO->getID())
658 OS << UO->getID();
659 else
660 OS << LiveOnEntryStr;
661 OS << ')';
662}
663
664void MemoryPhi::print(raw_ostream &OS) const {
665 bool First = true;
666 OS << getID() << " = MemoryPhi(";
667 for (const auto &Op : operands()) {
668 BasicBlock *BB = getIncomingBlock(Op);
669 MemoryAccess *MA = cast<MemoryAccess>(Op);
670 if (!First)
671 OS << ',';
672 else
673 First = false;
674
675 OS << '{';
676 if (BB->hasName())
677 OS << BB->getName();
678 else
679 BB->printAsOperand(OS, false);
680 OS << ',';
681 if (unsigned ID = MA->getID())
682 OS << ID;
683 else
684 OS << LiveOnEntryStr;
685 OS << '}';
686 }
687 OS << ')';
688}
689
690MemoryAccess::~MemoryAccess() {}
691
692void MemoryUse::print(raw_ostream &OS) const {
693 MemoryAccess *UO = getDefiningAccess();
694 OS << "MemoryUse(";
695 if (UO && UO->getID())
696 OS << UO->getID();
697 else
698 OS << LiveOnEntryStr;
699 OS << ')';
700}
701
702void MemoryAccess::dump() const {
703 print(dbgs());
704 dbgs() << "\n";
705}
706
Geoff Berryb96d3b22016-06-01 21:30:40 +0000707char MemorySSAAnalysis::PassID;
George Burgess IVe1100f52016-02-02 22:46:49 +0000708
Geoff Berryb96d3b22016-06-01 21:30:40 +0000709MemorySSA MemorySSAAnalysis::run(Function &F, AnalysisManager<Function> &AM) {
710 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
711 auto &AA = AM.getResult<AAManager>(F);
712 return MemorySSA(F, &AA, &DT);
George Burgess IVe1100f52016-02-02 22:46:49 +0000713}
714
Geoff Berryb96d3b22016-06-01 21:30:40 +0000715PreservedAnalyses MemorySSAPrinterPass::run(Function &F,
716 FunctionAnalysisManager &AM) {
717 OS << "MemorySSA for function: " << F.getName() << "\n";
718 AM.getResult<MemorySSAAnalysis>(F).print(OS);
719
720 return PreservedAnalyses::all();
George Burgess IVe1100f52016-02-02 22:46:49 +0000721}
722
Geoff Berryb96d3b22016-06-01 21:30:40 +0000723PreservedAnalyses MemorySSAVerifierPass::run(Function &F,
724 FunctionAnalysisManager &AM) {
725 AM.getResult<MemorySSAAnalysis>(F).verifyMemorySSA();
726
727 return PreservedAnalyses::all();
728}
729
730char MemorySSAWrapperPass::ID = 0;
731
732MemorySSAWrapperPass::MemorySSAWrapperPass() : FunctionPass(ID) {
733 initializeMemorySSAWrapperPassPass(*PassRegistry::getPassRegistry());
734}
735
736void MemorySSAWrapperPass::releaseMemory() { MSSA.reset(); }
737
738void MemorySSAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
George Burgess IVe1100f52016-02-02 22:46:49 +0000739 AU.setPreservesAll();
Geoff Berryb96d3b22016-06-01 21:30:40 +0000740 AU.addRequiredTransitive<DominatorTreeWrapperPass>();
741 AU.addRequiredTransitive<AAResultsWrapperPass>();
George Burgess IVe1100f52016-02-02 22:46:49 +0000742}
743
Geoff Berryb96d3b22016-06-01 21:30:40 +0000744bool MemorySSAWrapperPass::runOnFunction(Function &F) {
745 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
746 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
747 MSSA.reset(new MemorySSA(F, &AA, &DT));
George Burgess IVe1100f52016-02-02 22:46:49 +0000748 return false;
749}
750
Geoff Berryb96d3b22016-06-01 21:30:40 +0000751void MemorySSAWrapperPass::verifyAnalysis() const { MSSA->verifyMemorySSA(); }
George Burgess IVe1100f52016-02-02 22:46:49 +0000752
Geoff Berryb96d3b22016-06-01 21:30:40 +0000753void MemorySSAWrapperPass::print(raw_ostream &OS, const Module *M) const {
George Burgess IVe1100f52016-02-02 22:46:49 +0000754 MSSA->print(OS);
755}
756
George Burgess IVe1100f52016-02-02 22:46:49 +0000757MemorySSAWalker::MemorySSAWalker(MemorySSA *M) : MSSA(M) {}
758
759CachingMemorySSAWalker::CachingMemorySSAWalker(MemorySSA *M, AliasAnalysis *A,
760 DominatorTree *D)
761 : MemorySSAWalker(M), AA(A), DT(D) {}
762
763CachingMemorySSAWalker::~CachingMemorySSAWalker() {}
764
765struct CachingMemorySSAWalker::UpwardsMemoryQuery {
766 // True if we saw a phi whose predecessor was a backedge
767 bool SawBackedgePhi;
768 // True if our original query started off as a call
769 bool IsCall;
770 // The pointer location we started the query with. This will be empty if
771 // IsCall is true.
772 MemoryLocation StartingLoc;
773 // This is the instruction we were querying about.
774 const Instruction *Inst;
775 // Set of visited Instructions for this query.
776 DenseSet<MemoryAccessPair> Visited;
George Burgess IV49cad7d2016-03-30 03:12:08 +0000777 // Vector of visited call accesses for this query. This is separated out
778 // because you can always cache and lookup the result of call queries (IE when
779 // IsCall == true) for every call in the chain. The calls have no AA location
George Burgess IVe1100f52016-02-02 22:46:49 +0000780 // associated with them with them, and thus, no context dependence.
George Burgess IV49cad7d2016-03-30 03:12:08 +0000781 SmallVector<const MemoryAccess *, 32> VisitedCalls;
George Burgess IVe1100f52016-02-02 22:46:49 +0000782 // The MemoryAccess we actually got called with, used to test local domination
783 const MemoryAccess *OriginalAccess;
784 // The Datalayout for the module we started in
785 const DataLayout *DL;
786
787 UpwardsMemoryQuery()
788 : SawBackedgePhi(false), IsCall(false), Inst(nullptr),
789 OriginalAccess(nullptr), DL(nullptr) {}
790};
791
Daniel Berlin83fc77b2016-03-01 18:46:54 +0000792void CachingMemorySSAWalker::invalidateInfo(MemoryAccess *MA) {
793
794 // TODO: We can do much better cache invalidation with differently stored
795 // caches. For now, for MemoryUses, we simply remove them
796 // from the cache, and kill the entire call/non-call cache for everything
797 // else. The problem is for phis or defs, currently we'd need to follow use
798 // chains down and invalidate anything below us in the chain that currently
799 // terminates at this access.
800
801 // See if this is a MemoryUse, if so, just remove the cached info. MemoryUse
802 // is by definition never a barrier, so nothing in the cache could point to
803 // this use. In that case, we only need invalidate the info for the use
804 // itself.
805
806 if (MemoryUse *MU = dyn_cast<MemoryUse>(MA)) {
807 UpwardsMemoryQuery Q;
808 Instruction *I = MU->getMemoryInst();
809 Q.IsCall = bool(ImmutableCallSite(I));
810 Q.Inst = I;
811 if (!Q.IsCall)
812 Q.StartingLoc = MemoryLocation::get(I);
813 doCacheRemove(MA, Q, Q.StartingLoc);
Geoff Berry9fe26e62016-04-22 14:44:10 +0000814 } else {
815 // If it is not a use, the best we can do right now is destroy the cache.
Daniel Berlin83fc77b2016-03-01 18:46:54 +0000816 CachedUpwardsClobberingCall.clear();
Daniel Berlin83fc77b2016-03-01 18:46:54 +0000817 CachedUpwardsClobberingAccess.clear();
Geoff Berry9fe26e62016-04-22 14:44:10 +0000818 }
819
Filipe Cabecinhas0da99372016-04-29 15:22:48 +0000820#ifdef EXPENSIVE_CHECKS
Geoff Berry9fe26e62016-04-22 14:44:10 +0000821 // Run this only when expensive checks are enabled.
822 verifyRemoved(MA);
823#endif
Daniel Berlin83fc77b2016-03-01 18:46:54 +0000824}
825
George Burgess IVe1100f52016-02-02 22:46:49 +0000826void CachingMemorySSAWalker::doCacheRemove(const MemoryAccess *M,
827 const UpwardsMemoryQuery &Q,
828 const MemoryLocation &Loc) {
829 if (Q.IsCall)
830 CachedUpwardsClobberingCall.erase(M);
831 else
832 CachedUpwardsClobberingAccess.erase({M, Loc});
833}
834
835void CachingMemorySSAWalker::doCacheInsert(const MemoryAccess *M,
836 MemoryAccess *Result,
837 const UpwardsMemoryQuery &Q,
838 const MemoryLocation &Loc) {
George Burgess IV1b1fef32016-04-29 18:42:55 +0000839 // This is fine for Phis, since there are times where we can't optimize them.
840 // Making a def its own clobber is never correct, though.
841 assert((Result != M || isa<MemoryPhi>(M)) &&
842 "Something can't clobber itself!");
George Burgess IVe1100f52016-02-02 22:46:49 +0000843 ++NumClobberCacheInserts;
844 if (Q.IsCall)
845 CachedUpwardsClobberingCall[M] = Result;
846 else
847 CachedUpwardsClobberingAccess[{M, Loc}] = Result;
848}
849
850MemoryAccess *CachingMemorySSAWalker::doCacheLookup(const MemoryAccess *M,
851 const UpwardsMemoryQuery &Q,
852 const MemoryLocation &Loc) {
853 ++NumClobberCacheLookups;
854 MemoryAccess *Result = nullptr;
855
856 if (Q.IsCall)
857 Result = CachedUpwardsClobberingCall.lookup(M);
858 else
859 Result = CachedUpwardsClobberingAccess.lookup({M, Loc});
860
861 if (Result)
862 ++NumClobberCacheHits;
863 return Result;
864}
865
866bool CachingMemorySSAWalker::instructionClobbersQuery(
867 const MemoryDef *MD, UpwardsMemoryQuery &Q,
868 const MemoryLocation &Loc) const {
869 Instruction *DefMemoryInst = MD->getMemoryInst();
870 assert(DefMemoryInst && "Defining instruction not actually an instruction");
871
872 if (!Q.IsCall)
873 return AA->getModRefInfo(DefMemoryInst, Loc) & MRI_Mod;
874
875 // If this is a call, mark it for caching
876 if (ImmutableCallSite(DefMemoryInst))
George Burgess IV49cad7d2016-03-30 03:12:08 +0000877 Q.VisitedCalls.push_back(MD);
George Burgess IVe1100f52016-02-02 22:46:49 +0000878 ModRefInfo I = AA->getModRefInfo(DefMemoryInst, ImmutableCallSite(Q.Inst));
879 return I != MRI_NoModRef;
880}
881
882MemoryAccessPair CachingMemorySSAWalker::UpwardsDFSWalk(
883 MemoryAccess *StartingAccess, const MemoryLocation &Loc,
884 UpwardsMemoryQuery &Q, bool FollowingBackedge) {
885 MemoryAccess *ModifyingAccess = nullptr;
886
887 auto DFI = df_begin(StartingAccess);
888 for (auto DFE = df_end(StartingAccess); DFI != DFE;) {
889 MemoryAccess *CurrAccess = *DFI;
890 if (MSSA->isLiveOnEntryDef(CurrAccess))
891 return {CurrAccess, Loc};
George Burgess IV1b1fef32016-04-29 18:42:55 +0000892 // If this is a MemoryDef, check whether it clobbers our current query. This
893 // needs to be done before consulting the cache, because the cache reports
894 // the clobber for CurrAccess. If CurrAccess is a clobber for this query,
895 // and we ask the cache for information first, then we might skip this
896 // clobber, which is bad.
George Burgess IVe1100f52016-02-02 22:46:49 +0000897 if (auto *MD = dyn_cast<MemoryDef>(CurrAccess)) {
898 // If we hit the top, stop following this path.
899 // While we can do lookups, we can't sanely do inserts here unless we were
900 // to track everything we saw along the way, since we don't know where we
901 // will stop.
902 if (instructionClobbersQuery(MD, Q, Loc)) {
903 ModifyingAccess = CurrAccess;
904 break;
905 }
906 }
George Burgess IV1b1fef32016-04-29 18:42:55 +0000907 if (auto CacheResult = doCacheLookup(CurrAccess, Q, Loc))
908 return {CacheResult, Loc};
George Burgess IVe1100f52016-02-02 22:46:49 +0000909
910 // We need to know whether it is a phi so we can track backedges.
911 // Otherwise, walk all upward defs.
912 if (!isa<MemoryPhi>(CurrAccess)) {
913 ++DFI;
914 continue;
915 }
916
George Burgess IV0e489862016-03-23 18:31:55 +0000917#ifndef NDEBUG
918 // The loop below visits the phi's children for us. Because phis are the
919 // only things with multiple edges, skipping the children should always lead
920 // us to the end of the loop.
921 //
922 // Use a copy of DFI because skipChildren would kill our search stack, which
923 // would make caching anything on the way back impossible.
924 auto DFICopy = DFI;
925 assert(DFICopy.skipChildren() == DFE &&
926 "Skipping phi's children doesn't end the DFS?");
927#endif
928
George Burgess IV82ee9422016-03-30 00:26:26 +0000929 const MemoryAccessPair PHIPair(CurrAccess, Loc);
930
931 // Don't try to optimize this phi again if we've already tried to do so.
932 if (!Q.Visited.insert(PHIPair).second) {
933 ModifyingAccess = CurrAccess;
934 break;
935 }
936
George Burgess IV49cad7d2016-03-30 03:12:08 +0000937 std::size_t InitialVisitedCallSize = Q.VisitedCalls.size();
938
George Burgess IVe1100f52016-02-02 22:46:49 +0000939 // Recurse on PHI nodes, since we need to change locations.
940 // TODO: Allow graphtraits on pairs, which would turn this whole function
941 // into a normal single depth first walk.
942 MemoryAccess *FirstDef = nullptr;
George Burgess IVe1100f52016-02-02 22:46:49 +0000943 for (auto MPI = upward_defs_begin(PHIPair), MPE = upward_defs_end();
944 MPI != MPE; ++MPI) {
George Burgess IVe1100f52016-02-02 22:46:49 +0000945 bool Backedge =
946 !FollowingBackedge &&
947 DT->dominates(CurrAccess->getBlock(), MPI.getPhiArgBlock());
948
949 MemoryAccessPair CurrentPair =
950 UpwardsDFSWalk(MPI->first, MPI->second, Q, Backedge);
951 // All the phi arguments should reach the same point if we can bypass
952 // this phi. The alternative is that they hit this phi node, which
953 // means we can skip this argument.
954 if (FirstDef && CurrentPair.first != PHIPair.first &&
955 CurrentPair.first != FirstDef) {
956 ModifyingAccess = CurrAccess;
957 break;
958 }
959
960 if (!FirstDef)
961 FirstDef = CurrentPair.first;
George Burgess IVe1100f52016-02-02 22:46:49 +0000962 }
963
George Burgess IV0e489862016-03-23 18:31:55 +0000964 // If we exited the loop early, go with the result it gave us.
965 if (!ModifyingAccess) {
George Burgess IV82ee9422016-03-30 00:26:26 +0000966 assert(FirstDef && "Found a Phi with no upward defs?");
967 ModifyingAccess = FirstDef;
George Burgess IV49cad7d2016-03-30 03:12:08 +0000968 } else {
969 // If we can't optimize this Phi, then we can't safely cache any of the
970 // calls we visited when trying to optimize it. Wipe them out now.
971 Q.VisitedCalls.resize(InitialVisitedCallSize);
George Burgess IV0e489862016-03-23 18:31:55 +0000972 }
973 break;
George Burgess IVe1100f52016-02-02 22:46:49 +0000974 }
975
976 if (!ModifyingAccess)
977 return {MSSA->getLiveOnEntryDef(), Q.StartingLoc};
978
George Burgess IV0e489862016-03-23 18:31:55 +0000979 const BasicBlock *OriginalBlock = StartingAccess->getBlock();
George Burgess IV1b1fef32016-04-29 18:42:55 +0000980 assert(DFI.getPathLength() > 0 && "We dropped our path?");
George Burgess IVe1100f52016-02-02 22:46:49 +0000981 unsigned N = DFI.getPathLength();
George Burgess IV1b1fef32016-04-29 18:42:55 +0000982 // If we found a clobbering def, the last element in the path will be our
983 // clobber, so we don't want to cache that to itself. OTOH, if we optimized a
984 // phi, we can add the last thing in the path to the cache, since that won't
985 // be the result.
986 if (DFI.getPath(N - 1) == ModifyingAccess)
987 --N;
988 for (; N > 1; --N) {
George Burgess IV0e489862016-03-23 18:31:55 +0000989 MemoryAccess *CacheAccess = DFI.getPath(N - 1);
990 BasicBlock *CurrBlock = CacheAccess->getBlock();
George Burgess IVe1100f52016-02-02 22:46:49 +0000991 if (!FollowingBackedge)
George Burgess IV0e489862016-03-23 18:31:55 +0000992 doCacheInsert(CacheAccess, ModifyingAccess, Q, Loc);
George Burgess IVe1100f52016-02-02 22:46:49 +0000993 if (DT->dominates(CurrBlock, OriginalBlock) &&
994 (CurrBlock != OriginalBlock || !FollowingBackedge ||
George Burgess IV0e489862016-03-23 18:31:55 +0000995 MSSA->locallyDominates(CacheAccess, StartingAccess)))
George Burgess IVe1100f52016-02-02 22:46:49 +0000996 break;
997 }
998
999 // Cache everything else on the way back. The caller should cache
George Burgess IV1b1fef32016-04-29 18:42:55 +00001000 // StartingAccess for us.
1001 for (; N > 1; --N) {
George Burgess IVe1100f52016-02-02 22:46:49 +00001002 MemoryAccess *CacheAccess = DFI.getPath(N - 1);
1003 doCacheInsert(CacheAccess, ModifyingAccess, Q, Loc);
1004 }
1005 assert(Q.Visited.size() < 1000 && "Visited too much");
1006
1007 return {ModifyingAccess, Loc};
1008}
1009
1010/// \brief Walk the use-def chains starting at \p MA and find
1011/// the MemoryAccess that actually clobbers Loc.
1012///
1013/// \returns our clobbering memory access
1014MemoryAccess *
1015CachingMemorySSAWalker::getClobberingMemoryAccess(MemoryAccess *StartingAccess,
1016 UpwardsMemoryQuery &Q) {
1017 return UpwardsDFSWalk(StartingAccess, Q.StartingLoc, Q, false).first;
1018}
1019
1020MemoryAccess *
1021CachingMemorySSAWalker::getClobberingMemoryAccess(MemoryAccess *StartingAccess,
1022 MemoryLocation &Loc) {
1023 if (isa<MemoryPhi>(StartingAccess))
1024 return StartingAccess;
1025
1026 auto *StartingUseOrDef = cast<MemoryUseOrDef>(StartingAccess);
1027 if (MSSA->isLiveOnEntryDef(StartingUseOrDef))
1028 return StartingUseOrDef;
1029
1030 Instruction *I = StartingUseOrDef->getMemoryInst();
1031
1032 // Conservatively, fences are always clobbers, so don't perform the walk if we
1033 // hit a fence.
1034 if (isa<FenceInst>(I))
1035 return StartingUseOrDef;
1036
1037 UpwardsMemoryQuery Q;
1038 Q.OriginalAccess = StartingUseOrDef;
1039 Q.StartingLoc = Loc;
1040 Q.Inst = StartingUseOrDef->getMemoryInst();
1041 Q.IsCall = false;
1042 Q.DL = &Q.Inst->getModule()->getDataLayout();
1043
1044 if (auto CacheResult = doCacheLookup(StartingUseOrDef, Q, Q.StartingLoc))
1045 return CacheResult;
1046
1047 // Unlike the other function, do not walk to the def of a def, because we are
1048 // handed something we already believe is the clobbering access.
1049 MemoryAccess *DefiningAccess = isa<MemoryUse>(StartingUseOrDef)
1050 ? StartingUseOrDef->getDefiningAccess()
1051 : StartingUseOrDef;
1052
1053 MemoryAccess *Clobber = getClobberingMemoryAccess(DefiningAccess, Q);
George Burgess IV1b1fef32016-04-29 18:42:55 +00001054 // Only cache this if it wouldn't make Clobber point to itself.
1055 if (Clobber != StartingAccess)
1056 doCacheInsert(Q.OriginalAccess, Clobber, Q, Q.StartingLoc);
George Burgess IVe1100f52016-02-02 22:46:49 +00001057 DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
1058 DEBUG(dbgs() << *StartingUseOrDef << "\n");
1059 DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ");
1060 DEBUG(dbgs() << *Clobber << "\n");
1061 return Clobber;
1062}
1063
1064MemoryAccess *
1065CachingMemorySSAWalker::getClobberingMemoryAccess(const Instruction *I) {
1066 // There should be no way to lookup an instruction and get a phi as the
1067 // access, since we only map BB's to PHI's. So, this must be a use or def.
1068 auto *StartingAccess = cast<MemoryUseOrDef>(MSSA->getMemoryAccess(I));
1069
1070 // We can't sanely do anything with a FenceInst, they conservatively
1071 // clobber all memory, and have no locations to get pointers from to
1072 // try to disambiguate
1073 if (isa<FenceInst>(I))
1074 return StartingAccess;
1075
1076 UpwardsMemoryQuery Q;
1077 Q.OriginalAccess = StartingAccess;
1078 Q.IsCall = bool(ImmutableCallSite(I));
1079 if (!Q.IsCall)
1080 Q.StartingLoc = MemoryLocation::get(I);
1081 Q.Inst = I;
1082 Q.DL = &Q.Inst->getModule()->getDataLayout();
1083 if (auto CacheResult = doCacheLookup(StartingAccess, Q, Q.StartingLoc))
1084 return CacheResult;
1085
1086 // Start with the thing we already think clobbers this location
1087 MemoryAccess *DefiningAccess = StartingAccess->getDefiningAccess();
1088
1089 // At this point, DefiningAccess may be the live on entry def.
1090 // If it is, we will not get a better result.
1091 if (MSSA->isLiveOnEntryDef(DefiningAccess))
1092 return DefiningAccess;
1093
1094 MemoryAccess *Result = getClobberingMemoryAccess(DefiningAccess, Q);
George Burgess IV1b1fef32016-04-29 18:42:55 +00001095 // DFS won't cache a result for DefiningAccess. So, if DefiningAccess isn't
1096 // our clobber, be sure that it gets a cache entry, too.
1097 if (Result != DefiningAccess)
1098 doCacheInsert(DefiningAccess, Result, Q, Q.StartingLoc);
George Burgess IVe1100f52016-02-02 22:46:49 +00001099 doCacheInsert(Q.OriginalAccess, Result, Q, Q.StartingLoc);
1100 // TODO: When this implementation is more mature, we may want to figure out
1101 // what this additional caching buys us. It's most likely A Good Thing.
1102 if (Q.IsCall)
1103 for (const MemoryAccess *MA : Q.VisitedCalls)
George Burgess IV1b1fef32016-04-29 18:42:55 +00001104 if (MA != Result)
1105 doCacheInsert(MA, Result, Q, Q.StartingLoc);
George Burgess IVe1100f52016-02-02 22:46:49 +00001106
1107 DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ");
1108 DEBUG(dbgs() << *DefiningAccess << "\n");
1109 DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is ");
1110 DEBUG(dbgs() << *Result << "\n");
1111
1112 return Result;
1113}
1114
Geoff Berry9fe26e62016-04-22 14:44:10 +00001115// Verify that MA doesn't exist in any of the caches.
1116void CachingMemorySSAWalker::verifyRemoved(MemoryAccess *MA) {
1117#ifndef NDEBUG
1118 for (auto &P : CachedUpwardsClobberingAccess)
1119 assert(P.first.first != MA && P.second != MA &&
1120 "Found removed MemoryAccess in cache.");
1121 for (auto &P : CachedUpwardsClobberingCall)
1122 assert(P.first != MA && P.second != MA &&
1123 "Found removed MemoryAccess in cache.");
1124#endif // !NDEBUG
1125}
1126
George Burgess IVe1100f52016-02-02 22:46:49 +00001127MemoryAccess *
1128DoNothingMemorySSAWalker::getClobberingMemoryAccess(const Instruction *I) {
1129 MemoryAccess *MA = MSSA->getMemoryAccess(I);
1130 if (auto *Use = dyn_cast<MemoryUseOrDef>(MA))
1131 return Use->getDefiningAccess();
1132 return MA;
1133}
1134
1135MemoryAccess *DoNothingMemorySSAWalker::getClobberingMemoryAccess(
1136 MemoryAccess *StartingAccess, MemoryLocation &) {
1137 if (auto *Use = dyn_cast<MemoryUseOrDef>(StartingAccess))
1138 return Use->getDefiningAccess();
1139 return StartingAccess;
1140}
1141}