1 //===- LoopDependenceAnalysis.cpp - LDA Implementation ----------*- C++ -*-===//
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 is the (beginning) of an implementation of a loop dependence analysis
11 // framework, which is used to detect dependences in memory accesses in loops.
12 //
13 // Please note that this is work in progress and the interface is subject to
14 // change.
15 //
16 // TODO: adapt as implementation progresses.
17 //
18 // TODO: document lingo (pair, subscript, index)
19 //
20 //===----------------------------------------------------------------------===//
21
22 #define DEBUG_TYPE "lda"
23 #include "llvm/ADT/DenseSet.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/Analysis/AliasAnalysis.h"
26 #include "llvm/Analysis/LoopDependenceAnalysis.h"
27 #include "llvm/Analysis/LoopPass.h"
28 #include "llvm/Analysis/ScalarEvolution.h"
29 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/Assembly/Writer.h"
32 #include "llvm/Instructions.h"
33 #include "llvm/Operator.h"
34 #include "llvm/Support/Allocator.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/raw_ostream.h"
38 #include "llvm/Target/TargetData.h"
39 using namespace llvm;
40
41 STATISTIC(NumAnswered, "Number of dependence queries answered");
42 STATISTIC(NumAnalysed, "Number of distinct dependence pairs analysed");
43 STATISTIC(NumDependent, "Number of pairs with dependent accesses");
44 STATISTIC(NumIndependent, "Number of pairs with independent accesses");
45 STATISTIC(NumUnknown, "Number of pairs with unknown accesses");
46
createLoopDependenceAnalysisPass()47 LoopPass *llvm::createLoopDependenceAnalysisPass() {
48 return new LoopDependenceAnalysis();
49 }
50
51 INITIALIZE_PASS_BEGIN(LoopDependenceAnalysis, "lda",
52 "Loop Dependence Analysis", false, true)
53 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
54 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
55 INITIALIZE_PASS_END(LoopDependenceAnalysis, "lda",
56 "Loop Dependence Analysis", false, true)
57 char LoopDependenceAnalysis::ID = 0;
58
59 //===----------------------------------------------------------------------===//
60 // Utility Functions
61 //===----------------------------------------------------------------------===//
62
IsMemRefInstr(const Value * V)63 static inline bool IsMemRefInstr(const Value *V) {
64 const Instruction *I = dyn_cast<const Instruction>(V);
65 return I && (I->mayReadFromMemory() || I->mayWriteToMemory());
66 }
67
GetMemRefInstrs(const Loop * L,SmallVectorImpl<Instruction * > & Memrefs)68 static void GetMemRefInstrs(const Loop *L,
69 SmallVectorImpl<Instruction*> &Memrefs) {
70 for (Loop::block_iterator b = L->block_begin(), be = L->block_end();
71 b != be; ++b)
72 for (BasicBlock::iterator i = (*b)->begin(), ie = (*b)->end();
73 i != ie; ++i)
74 if (IsMemRefInstr(i))
75 Memrefs.push_back(i);
76 }
77
IsLoadOrStoreInst(Value * I)78 static bool IsLoadOrStoreInst(Value *I) {
79 // Returns true if the load or store can be analyzed. Atomic and volatile
80 // operations have properties which this analysis does not understand.
81 if (LoadInst *LI = dyn_cast<LoadInst>(I))
82 return LI->isUnordered();
83 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
84 return SI->isUnordered();
85 return false;
86 }
87
GetPointerOperand(Value * I)88 static Value *GetPointerOperand(Value *I) {
89 if (LoadInst *i = dyn_cast<LoadInst>(I))
90 return i->getPointerOperand();
91 if (StoreInst *i = dyn_cast<StoreInst>(I))
92 return i->getPointerOperand();
93 llvm_unreachable("Value is no load or store instruction!");
94 }
95
UnderlyingObjectsAlias(AliasAnalysis * AA,const Value * A,const Value * B)96 static AliasAnalysis::AliasResult UnderlyingObjectsAlias(AliasAnalysis *AA,
97 const Value *A,
98 const Value *B) {
99 const Value *aObj = GetUnderlyingObject(A);
100 const Value *bObj = GetUnderlyingObject(B);
101 return AA->alias(aObj, AA->getTypeStoreSize(aObj->getType()),
102 bObj, AA->getTypeStoreSize(bObj->getType()));
103 }
104
GetZeroSCEV(ScalarEvolution * SE)105 static inline const SCEV *GetZeroSCEV(ScalarEvolution *SE) {
106 return SE->getConstant(Type::getInt32Ty(SE->getContext()), 0L);
107 }
108
109 //===----------------------------------------------------------------------===//
110 // Dependence Testing
111 //===----------------------------------------------------------------------===//
112
isDependencePair(const Value * A,const Value * B) const113 bool LoopDependenceAnalysis::isDependencePair(const Value *A,
114 const Value *B) const {
115 return IsMemRefInstr(A) &&
116 IsMemRefInstr(B) &&
117 (cast<const Instruction>(A)->mayWriteToMemory() ||
118 cast<const Instruction>(B)->mayWriteToMemory());
119 }
120
findOrInsertDependencePair(Value * A,Value * B,DependencePair * & P)121 bool LoopDependenceAnalysis::findOrInsertDependencePair(Value *A,
122 Value *B,
123 DependencePair *&P) {
124 void *insertPos = 0;
125 FoldingSetNodeID id;
126 id.AddPointer(A);
127 id.AddPointer(B);
128
129 P = Pairs.FindNodeOrInsertPos(id, insertPos);
130 if (P) return true;
131
132 P = new (PairAllocator) DependencePair(id, A, B);
133 Pairs.InsertNode(P, insertPos);
134 return false;
135 }
136
getLoops(const SCEV * S,DenseSet<const Loop * > * Loops) const137 void LoopDependenceAnalysis::getLoops(const SCEV *S,
138 DenseSet<const Loop*>* Loops) const {
139 // Refactor this into an SCEVVisitor, if efficiency becomes a concern.
140 for (const Loop *L = this->L; L != 0; L = L->getParentLoop())
141 if (!SE->isLoopInvariant(S, L))
142 Loops->insert(L);
143 }
144
isLoopInvariant(const SCEV * S) const145 bool LoopDependenceAnalysis::isLoopInvariant(const SCEV *S) const {
146 DenseSet<const Loop*> loops;
147 getLoops(S, &loops);
148 return loops.empty();
149 }
150
isAffine(const SCEV * S) const151 bool LoopDependenceAnalysis::isAffine(const SCEV *S) const {
152 const SCEVAddRecExpr *rec = dyn_cast<SCEVAddRecExpr>(S);
153 return isLoopInvariant(S) || (rec && rec->isAffine());
154 }
155
isZIVPair(const SCEV * A,const SCEV * B) const156 bool LoopDependenceAnalysis::isZIVPair(const SCEV *A, const SCEV *B) const {
157 return isLoopInvariant(A) && isLoopInvariant(B);
158 }
159
isSIVPair(const SCEV * A,const SCEV * B) const160 bool LoopDependenceAnalysis::isSIVPair(const SCEV *A, const SCEV *B) const {
161 DenseSet<const Loop*> loops;
162 getLoops(A, &loops);
163 getLoops(B, &loops);
164 return loops.size() == 1;
165 }
166
167 LoopDependenceAnalysis::DependenceResult
analyseZIV(const SCEV * A,const SCEV * B,Subscript * S) const168 LoopDependenceAnalysis::analyseZIV(const SCEV *A,
169 const SCEV *B,
170 Subscript *S) const {
171 assert(isZIVPair(A, B) && "Attempted to ZIV-test non-ZIV SCEVs!");
172 return A == B ? Dependent : Independent;
173 }
174
175 LoopDependenceAnalysis::DependenceResult
analyseSIV(const SCEV * A,const SCEV * B,Subscript * S) const176 LoopDependenceAnalysis::analyseSIV(const SCEV *A,
177 const SCEV *B,
178 Subscript *S) const {
179 return Unknown; // TODO: Implement.
180 }
181
182 LoopDependenceAnalysis::DependenceResult
analyseMIV(const SCEV * A,const SCEV * B,Subscript * S) const183 LoopDependenceAnalysis::analyseMIV(const SCEV *A,
184 const SCEV *B,
185 Subscript *S) const {
186 return Unknown; // TODO: Implement.
187 }
188
189 LoopDependenceAnalysis::DependenceResult
analyseSubscript(const SCEV * A,const SCEV * B,Subscript * S) const190 LoopDependenceAnalysis::analyseSubscript(const SCEV *A,
191 const SCEV *B,
192 Subscript *S) const {
193 DEBUG(dbgs() << " Testing subscript: " << *A << ", " << *B << "\n");
194
195 if (A == B) {
196 DEBUG(dbgs() << " -> [D] same SCEV\n");
197 return Dependent;
198 }
199
200 if (!isAffine(A) || !isAffine(B)) {
201 DEBUG(dbgs() << " -> [?] not affine\n");
202 return Unknown;
203 }
204
205 if (isZIVPair(A, B))
206 return analyseZIV(A, B, S);
207
208 if (isSIVPair(A, B))
209 return analyseSIV(A, B, S);
210
211 return analyseMIV(A, B, S);
212 }
213
214 LoopDependenceAnalysis::DependenceResult
analysePair(DependencePair * P) const215 LoopDependenceAnalysis::analysePair(DependencePair *P) const {
216 DEBUG(dbgs() << "Analysing:\n" << *P->A << "\n" << *P->B << "\n");
217
218 // We only analyse loads and stores but no possible memory accesses by e.g.
219 // free, call, or invoke instructions.
220 if (!IsLoadOrStoreInst(P->A) || !IsLoadOrStoreInst(P->B)) {
221 DEBUG(dbgs() << "--> [?] no load/store\n");
222 return Unknown;
223 }
224
225 Value *aPtr = GetPointerOperand(P->A);
226 Value *bPtr = GetPointerOperand(P->B);
227
228 switch (UnderlyingObjectsAlias(AA, aPtr, bPtr)) {
229 case AliasAnalysis::MayAlias:
230 case AliasAnalysis::PartialAlias:
231 // We can not analyse objects if we do not know about their aliasing.
232 DEBUG(dbgs() << "---> [?] may alias\n");
233 return Unknown;
234
235 case AliasAnalysis::NoAlias:
236 // If the objects noalias, they are distinct, accesses are independent.
237 DEBUG(dbgs() << "---> [I] no alias\n");
238 return Independent;
239
240 case AliasAnalysis::MustAlias:
241 break; // The underlying objects alias, test accesses for dependence.
242 }
243
244 const GEPOperator *aGEP = dyn_cast<GEPOperator>(aPtr);
245 const GEPOperator *bGEP = dyn_cast<GEPOperator>(bPtr);
246
247 if (!aGEP || !bGEP)
248 return Unknown;
249
250 // FIXME: Is filtering coupled subscripts necessary?
251
252 // Collect GEP operand pairs (FIXME: use GetGEPOperands from BasicAA), adding
253 // trailing zeroes to the smaller GEP, if needed.
254 typedef SmallVector<std::pair<const SCEV*, const SCEV*>, 4> GEPOpdPairsTy;
255 GEPOpdPairsTy opds;
256 for(GEPOperator::const_op_iterator aIdx = aGEP->idx_begin(),
257 aEnd = aGEP->idx_end(),
258 bIdx = bGEP->idx_begin(),
259 bEnd = bGEP->idx_end();
260 aIdx != aEnd && bIdx != bEnd;
261 aIdx += (aIdx != aEnd), bIdx += (bIdx != bEnd)) {
262 const SCEV* aSCEV = (aIdx != aEnd) ? SE->getSCEV(*aIdx) : GetZeroSCEV(SE);
263 const SCEV* bSCEV = (bIdx != bEnd) ? SE->getSCEV(*bIdx) : GetZeroSCEV(SE);
264 opds.push_back(std::make_pair(aSCEV, bSCEV));
265 }
266
267 if (!opds.empty() && opds[0].first != opds[0].second) {
268 // We cannot (yet) handle arbitrary GEP pointer offsets. By limiting
269 //
270 // TODO: this could be relaxed by adding the size of the underlying object
271 // to the first subscript. If we have e.g. (GEP x,0,i; GEP x,2,-i) and we
272 // know that x is a [100 x i8]*, we could modify the first subscript to be
273 // (i, 200-i) instead of (i, -i).
274 return Unknown;
275 }
276
277 // Now analyse the collected operand pairs (skipping the GEP ptr offsets).
278 for (GEPOpdPairsTy::const_iterator i = opds.begin() + 1, end = opds.end();
279 i != end; ++i) {
280 Subscript subscript;
281 DependenceResult result = analyseSubscript(i->first, i->second, &subscript);
282 if (result != Dependent) {
283 // We either proved independence or failed to analyse this subscript.
284 // Further subscripts will not improve the situation, so abort early.
285 return result;
286 }
287 P->Subscripts.push_back(subscript);
288 }
289 // We successfully analysed all subscripts but failed to prove independence.
290 return Dependent;
291 }
292
depends(Value * A,Value * B)293 bool LoopDependenceAnalysis::depends(Value *A, Value *B) {
294 assert(isDependencePair(A, B) && "Values form no dependence pair!");
295 ++NumAnswered;
296
297 DependencePair *p;
298 if (!findOrInsertDependencePair(A, B, p)) {
299 // The pair is not cached, so analyse it.
300 ++NumAnalysed;
301 switch (p->Result = analysePair(p)) {
302 case Dependent: ++NumDependent; break;
303 case Independent: ++NumIndependent; break;
304 case Unknown: ++NumUnknown; break;
305 }
306 }
307 return p->Result != Independent;
308 }
309
310 //===----------------------------------------------------------------------===//
311 // LoopDependenceAnalysis Implementation
312 //===----------------------------------------------------------------------===//
313
runOnLoop(Loop * L,LPPassManager &)314 bool LoopDependenceAnalysis::runOnLoop(Loop *L, LPPassManager &) {
315 this->L = L;
316 AA = &getAnalysis<AliasAnalysis>();
317 SE = &getAnalysis<ScalarEvolution>();
318 return false;
319 }
320
releaseMemory()321 void LoopDependenceAnalysis::releaseMemory() {
322 Pairs.clear();
323 PairAllocator.Reset();
324 }
325
getAnalysisUsage(AnalysisUsage & AU) const326 void LoopDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
327 AU.setPreservesAll();
328 AU.addRequiredTransitive<AliasAnalysis>();
329 AU.addRequiredTransitive<ScalarEvolution>();
330 }
331
PrintLoopInfo(raw_ostream & OS,LoopDependenceAnalysis * LDA,const Loop * L)332 static void PrintLoopInfo(raw_ostream &OS,
333 LoopDependenceAnalysis *LDA, const Loop *L) {
334 if (!L->empty()) return; // ignore non-innermost loops
335
336 SmallVector<Instruction*, 8> memrefs;
337 GetMemRefInstrs(L, memrefs);
338
339 OS << "Loop at depth " << L->getLoopDepth() << ", header block: ";
340 WriteAsOperand(OS, L->getHeader(), false);
341 OS << "\n";
342
343 OS << " Load/store instructions: " << memrefs.size() << "\n";
344 for (SmallVector<Instruction*, 8>::const_iterator x = memrefs.begin(),
345 end = memrefs.end(); x != end; ++x)
346 OS << "\t" << (x - memrefs.begin()) << ": " << **x << "\n";
347
348 OS << " Pairwise dependence results:\n";
349 for (SmallVector<Instruction*, 8>::const_iterator x = memrefs.begin(),
350 end = memrefs.end(); x != end; ++x)
351 for (SmallVector<Instruction*, 8>::const_iterator y = x + 1;
352 y != end; ++y)
353 if (LDA->isDependencePair(*x, *y))
354 OS << "\t" << (x - memrefs.begin()) << "," << (y - memrefs.begin())
355 << ": " << (LDA->depends(*x, *y) ? "dependent" : "independent")
356 << "\n";
357 }
358
print(raw_ostream & OS,const Module *) const359 void LoopDependenceAnalysis::print(raw_ostream &OS, const Module*) const {
360 // TODO: doc why const_cast is safe
361 PrintLoopInfo(OS, const_cast<LoopDependenceAnalysis*>(this), this->L);
362 }
363