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/Instructions.h"
31 #include "llvm/Operator.h"
32 #include "llvm/Support/Allocator.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetData.h"
37 using namespace llvm;
38
39 STATISTIC(NumAnswered, "Number of dependence queries answered");
40 STATISTIC(NumAnalysed, "Number of distinct dependence pairs analysed");
41 STATISTIC(NumDependent, "Number of pairs with dependent accesses");
42 STATISTIC(NumIndependent, "Number of pairs with independent accesses");
43 STATISTIC(NumUnknown, "Number of pairs with unknown accesses");
44
createLoopDependenceAnalysisPass()45 LoopPass *llvm::createLoopDependenceAnalysisPass() {
46 return new LoopDependenceAnalysis();
47 }
48
49 INITIALIZE_PASS(LoopDependenceAnalysis, "lda",
50 "Loop Dependence Analysis", false, true);
51 char LoopDependenceAnalysis::ID = 0;
52
53 //===----------------------------------------------------------------------===//
54 // Utility Functions
55 //===----------------------------------------------------------------------===//
56
IsMemRefInstr(const Value * V)57 static inline bool IsMemRefInstr(const Value *V) {
58 const Instruction *I = dyn_cast<const Instruction>(V);
59 return I && (I->mayReadFromMemory() || I->mayWriteToMemory());
60 }
61
GetMemRefInstrs(const Loop * L,SmallVectorImpl<Instruction * > & Memrefs)62 static void GetMemRefInstrs(const Loop *L,
63 SmallVectorImpl<Instruction*> &Memrefs) {
64 for (Loop::block_iterator b = L->block_begin(), be = L->block_end();
65 b != be; ++b)
66 for (BasicBlock::iterator i = (*b)->begin(), ie = (*b)->end();
67 i != ie; ++i)
68 if (IsMemRefInstr(i))
69 Memrefs.push_back(i);
70 }
71
IsLoadOrStoreInst(Value * I)72 static bool IsLoadOrStoreInst(Value *I) {
73 return isa<LoadInst>(I) || isa<StoreInst>(I);
74 }
75
GetPointerOperand(Value * I)76 static Value *GetPointerOperand(Value *I) {
77 if (LoadInst *i = dyn_cast<LoadInst>(I))
78 return i->getPointerOperand();
79 if (StoreInst *i = dyn_cast<StoreInst>(I))
80 return i->getPointerOperand();
81 llvm_unreachable("Value is no load or store instruction!");
82 // Never reached.
83 return 0;
84 }
85
UnderlyingObjectsAlias(AliasAnalysis * AA,const Value * A,const Value * B)86 static AliasAnalysis::AliasResult UnderlyingObjectsAlias(AliasAnalysis *AA,
87 const Value *A,
88 const Value *B) {
89 const Value *aObj = A->getUnderlyingObject();
90 const Value *bObj = B->getUnderlyingObject();
91 return AA->alias(aObj, AA->getTypeStoreSize(aObj->getType()),
92 bObj, AA->getTypeStoreSize(bObj->getType()));
93 }
94
GetZeroSCEV(ScalarEvolution * SE)95 static inline const SCEV *GetZeroSCEV(ScalarEvolution *SE) {
96 return SE->getConstant(Type::getInt32Ty(SE->getContext()), 0L);
97 }
98
99 //===----------------------------------------------------------------------===//
100 // Dependence Testing
101 //===----------------------------------------------------------------------===//
102
isDependencePair(const Value * A,const Value * B) const103 bool LoopDependenceAnalysis::isDependencePair(const Value *A,
104 const Value *B) const {
105 return IsMemRefInstr(A) &&
106 IsMemRefInstr(B) &&
107 (cast<const Instruction>(A)->mayWriteToMemory() ||
108 cast<const Instruction>(B)->mayWriteToMemory());
109 }
110
findOrInsertDependencePair(Value * A,Value * B,DependencePair * & P)111 bool LoopDependenceAnalysis::findOrInsertDependencePair(Value *A,
112 Value *B,
113 DependencePair *&P) {
114 void *insertPos = 0;
115 FoldingSetNodeID id;
116 id.AddPointer(A);
117 id.AddPointer(B);
118
119 P = Pairs.FindNodeOrInsertPos(id, insertPos);
120 if (P) return true;
121
122 P = new (PairAllocator) DependencePair(id, A, B);
123 Pairs.InsertNode(P, insertPos);
124 return false;
125 }
126
getLoops(const SCEV * S,DenseSet<const Loop * > * Loops) const127 void LoopDependenceAnalysis::getLoops(const SCEV *S,
128 DenseSet<const Loop*>* Loops) const {
129 // Refactor this into an SCEVVisitor, if efficiency becomes a concern.
130 for (const Loop *L = this->L; L != 0; L = L->getParentLoop())
131 if (!S->isLoopInvariant(L))
132 Loops->insert(L);
133 }
134
isLoopInvariant(const SCEV * S) const135 bool LoopDependenceAnalysis::isLoopInvariant(const SCEV *S) const {
136 DenseSet<const Loop*> loops;
137 getLoops(S, &loops);
138 return loops.empty();
139 }
140
isAffine(const SCEV * S) const141 bool LoopDependenceAnalysis::isAffine(const SCEV *S) const {
142 const SCEVAddRecExpr *rec = dyn_cast<SCEVAddRecExpr>(S);
143 return isLoopInvariant(S) || (rec && rec->isAffine());
144 }
145
isZIVPair(const SCEV * A,const SCEV * B) const146 bool LoopDependenceAnalysis::isZIVPair(const SCEV *A, const SCEV *B) const {
147 return isLoopInvariant(A) && isLoopInvariant(B);
148 }
149
isSIVPair(const SCEV * A,const SCEV * B) const150 bool LoopDependenceAnalysis::isSIVPair(const SCEV *A, const SCEV *B) const {
151 DenseSet<const Loop*> loops;
152 getLoops(A, &loops);
153 getLoops(B, &loops);
154 return loops.size() == 1;
155 }
156
157 LoopDependenceAnalysis::DependenceResult
analyseZIV(const SCEV * A,const SCEV * B,Subscript * S) const158 LoopDependenceAnalysis::analyseZIV(const SCEV *A,
159 const SCEV *B,
160 Subscript *S) const {
161 assert(isZIVPair(A, B) && "Attempted to ZIV-test non-ZIV SCEVs!");
162 return A == B ? Dependent : Independent;
163 }
164
165 LoopDependenceAnalysis::DependenceResult
analyseSIV(const SCEV * A,const SCEV * B,Subscript * S) const166 LoopDependenceAnalysis::analyseSIV(const SCEV *A,
167 const SCEV *B,
168 Subscript *S) const {
169 return Unknown; // TODO: Implement.
170 }
171
172 LoopDependenceAnalysis::DependenceResult
analyseMIV(const SCEV * A,const SCEV * B,Subscript * S) const173 LoopDependenceAnalysis::analyseMIV(const SCEV *A,
174 const SCEV *B,
175 Subscript *S) const {
176 return Unknown; // TODO: Implement.
177 }
178
179 LoopDependenceAnalysis::DependenceResult
analyseSubscript(const SCEV * A,const SCEV * B,Subscript * S) const180 LoopDependenceAnalysis::analyseSubscript(const SCEV *A,
181 const SCEV *B,
182 Subscript *S) const {
183 DEBUG(dbgs() << " Testing subscript: " << *A << ", " << *B << "\n");
184
185 if (A == B) {
186 DEBUG(dbgs() << " -> [D] same SCEV\n");
187 return Dependent;
188 }
189
190 if (!isAffine(A) || !isAffine(B)) {
191 DEBUG(dbgs() << " -> [?] not affine\n");
192 return Unknown;
193 }
194
195 if (isZIVPair(A, B))
196 return analyseZIV(A, B, S);
197
198 if (isSIVPair(A, B))
199 return analyseSIV(A, B, S);
200
201 return analyseMIV(A, B, S);
202 }
203
204 LoopDependenceAnalysis::DependenceResult
analysePair(DependencePair * P) const205 LoopDependenceAnalysis::analysePair(DependencePair *P) const {
206 DEBUG(dbgs() << "Analysing:\n" << *P->A << "\n" << *P->B << "\n");
207
208 // We only analyse loads and stores but no possible memory accesses by e.g.
209 // free, call, or invoke instructions.
210 if (!IsLoadOrStoreInst(P->A) || !IsLoadOrStoreInst(P->B)) {
211 DEBUG(dbgs() << "--> [?] no load/store\n");
212 return Unknown;
213 }
214
215 Value *aPtr = GetPointerOperand(P->A);
216 Value *bPtr = GetPointerOperand(P->B);
217
218 switch (UnderlyingObjectsAlias(AA, aPtr, bPtr)) {
219 case AliasAnalysis::MayAlias:
220 // We can not analyse objects if we do not know about their aliasing.
221 DEBUG(dbgs() << "---> [?] may alias\n");
222 return Unknown;
223
224 case AliasAnalysis::NoAlias:
225 // If the objects noalias, they are distinct, accesses are independent.
226 DEBUG(dbgs() << "---> [I] no alias\n");
227 return Independent;
228
229 case AliasAnalysis::MustAlias:
230 break; // The underlying objects alias, test accesses for dependence.
231 }
232
233 const GEPOperator *aGEP = dyn_cast<GEPOperator>(aPtr);
234 const GEPOperator *bGEP = dyn_cast<GEPOperator>(bPtr);
235
236 if (!aGEP || !bGEP)
237 return Unknown;
238
239 // FIXME: Is filtering coupled subscripts necessary?
240
241 // Collect GEP operand pairs (FIXME: use GetGEPOperands from BasicAA), adding
242 // trailing zeroes to the smaller GEP, if needed.
243 typedef SmallVector<std::pair<const SCEV*, const SCEV*>, 4> GEPOpdPairsTy;
244 GEPOpdPairsTy opds;
245 for(GEPOperator::const_op_iterator aIdx = aGEP->idx_begin(),
246 aEnd = aGEP->idx_end(),
247 bIdx = bGEP->idx_begin(),
248 bEnd = bGEP->idx_end();
249 aIdx != aEnd && bIdx != bEnd;
250 aIdx += (aIdx != aEnd), bIdx += (bIdx != bEnd)) {
251 const SCEV* aSCEV = (aIdx != aEnd) ? SE->getSCEV(*aIdx) : GetZeroSCEV(SE);
252 const SCEV* bSCEV = (bIdx != bEnd) ? SE->getSCEV(*bIdx) : GetZeroSCEV(SE);
253 opds.push_back(std::make_pair(aSCEV, bSCEV));
254 }
255
256 if (!opds.empty() && opds[0].first != opds[0].second) {
257 // We cannot (yet) handle arbitrary GEP pointer offsets. By limiting
258 //
259 // TODO: this could be relaxed by adding the size of the underlying object
260 // to the first subscript. If we have e.g. (GEP x,0,i; GEP x,2,-i) and we
261 // know that x is a [100 x i8]*, we could modify the first subscript to be
262 // (i, 200-i) instead of (i, -i).
263 return Unknown;
264 }
265
266 // Now analyse the collected operand pairs (skipping the GEP ptr offsets).
267 for (GEPOpdPairsTy::const_iterator i = opds.begin() + 1, end = opds.end();
268 i != end; ++i) {
269 Subscript subscript;
270 DependenceResult result = analyseSubscript(i->first, i->second, &subscript);
271 if (result != Dependent) {
272 // We either proved independence or failed to analyse this subscript.
273 // Further subscripts will not improve the situation, so abort early.
274 return result;
275 }
276 P->Subscripts.push_back(subscript);
277 }
278 // We successfully analysed all subscripts but failed to prove independence.
279 return Dependent;
280 }
281
depends(Value * A,Value * B)282 bool LoopDependenceAnalysis::depends(Value *A, Value *B) {
283 assert(isDependencePair(A, B) && "Values form no dependence pair!");
284 ++NumAnswered;
285
286 DependencePair *p;
287 if (!findOrInsertDependencePair(A, B, p)) {
288 // The pair is not cached, so analyse it.
289 ++NumAnalysed;
290 switch (p->Result = analysePair(p)) {
291 case Dependent: ++NumDependent; break;
292 case Independent: ++NumIndependent; break;
293 case Unknown: ++NumUnknown; break;
294 }
295 }
296 return p->Result != Independent;
297 }
298
299 //===----------------------------------------------------------------------===//
300 // LoopDependenceAnalysis Implementation
301 //===----------------------------------------------------------------------===//
302
runOnLoop(Loop * L,LPPassManager &)303 bool LoopDependenceAnalysis::runOnLoop(Loop *L, LPPassManager &) {
304 this->L = L;
305 AA = &getAnalysis<AliasAnalysis>();
306 SE = &getAnalysis<ScalarEvolution>();
307 return false;
308 }
309
releaseMemory()310 void LoopDependenceAnalysis::releaseMemory() {
311 Pairs.clear();
312 PairAllocator.Reset();
313 }
314
getAnalysisUsage(AnalysisUsage & AU) const315 void LoopDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
316 AU.setPreservesAll();
317 AU.addRequiredTransitive<AliasAnalysis>();
318 AU.addRequiredTransitive<ScalarEvolution>();
319 }
320
PrintLoopInfo(raw_ostream & OS,LoopDependenceAnalysis * LDA,const Loop * L)321 static void PrintLoopInfo(raw_ostream &OS,
322 LoopDependenceAnalysis *LDA, const Loop *L) {
323 if (!L->empty()) return; // ignore non-innermost loops
324
325 SmallVector<Instruction*, 8> memrefs;
326 GetMemRefInstrs(L, memrefs);
327
328 OS << "Loop at depth " << L->getLoopDepth() << ", header block: ";
329 WriteAsOperand(OS, L->getHeader(), false);
330 OS << "\n";
331
332 OS << " Load/store instructions: " << memrefs.size() << "\n";
333 for (SmallVector<Instruction*, 8>::const_iterator x = memrefs.begin(),
334 end = memrefs.end(); x != end; ++x)
335 OS << "\t" << (x - memrefs.begin()) << ": " << **x << "\n";
336
337 OS << " Pairwise dependence results:\n";
338 for (SmallVector<Instruction*, 8>::const_iterator x = memrefs.begin(),
339 end = memrefs.end(); x != end; ++x)
340 for (SmallVector<Instruction*, 8>::const_iterator y = x + 1;
341 y != end; ++y)
342 if (LDA->isDependencePair(*x, *y))
343 OS << "\t" << (x - memrefs.begin()) << "," << (y - memrefs.begin())
344 << ": " << (LDA->depends(*x, *y) ? "dependent" : "independent")
345 << "\n";
346 }
347
print(raw_ostream & OS,const Module *) const348 void LoopDependenceAnalysis::print(raw_ostream &OS, const Module*) const {
349 // TODO: doc why const_cast is safe
350 PrintLoopInfo(OS, const_cast<LoopDependenceAnalysis*>(this), this->L);
351 }
352