//
// TODO: adapt as implementation progresses.
//
+// TODO: document lingo (pair, subscript, index)
+//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "lda"
+#include "llvm/ADT/DenseSet.h"
+#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/LoopDependenceAnalysis.h"
#include "llvm/Analysis/LoopPass.h"
#include "llvm/Analysis/ScalarEvolution.h"
+#include "llvm/Analysis/ScalarEvolutionExpressions.h"
#include "llvm/Instructions.h"
+#include "llvm/Operator.h"
+#include "llvm/Support/Allocator.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetData.h"
using namespace llvm;
+STATISTIC(NumAnswered, "Number of dependence queries answered");
+STATISTIC(NumAnalysed, "Number of distinct dependence pairs analysed");
+STATISTIC(NumDependent, "Number of pairs with dependent accesses");
+STATISTIC(NumIndependent, "Number of pairs with independent accesses");
+STATISTIC(NumUnknown, "Number of pairs with unknown accesses");
+
LoopPass *llvm::createLoopDependenceAnalysisPass() {
return new LoopDependenceAnalysis();
}
return I && (I->mayReadFromMemory() || I->mayWriteToMemory());
}
-static void GetMemRefInstrs(
- const Loop *L, SmallVectorImpl<Instruction*> &memrefs) {
+static void GetMemRefInstrs(const Loop *L,
+ SmallVectorImpl<Instruction*> &Memrefs) {
for (Loop::block_iterator b = L->block_begin(), be = L->block_end();
- b != be; ++b)
+ b != be; ++b)
for (BasicBlock::iterator i = (*b)->begin(), ie = (*b)->end();
- i != ie; ++i)
+ i != ie; ++i)
if (IsMemRefInstr(i))
- memrefs.push_back(i);
+ Memrefs.push_back(i);
}
static bool IsLoadOrStoreInst(Value *I) {
return i->getPointerOperand();
if (StoreInst *i = dyn_cast<StoreInst>(I))
return i->getPointerOperand();
- LLVM_UNREACHABLE("Value is no load or store instruction!");
+ llvm_unreachable("Value is no load or store instruction!");
// Never reached.
return 0;
}
+static AliasAnalysis::AliasResult UnderlyingObjectsAlias(AliasAnalysis *AA,
+ const Value *A,
+ const Value *B) {
+ const Value *aObj = A->getUnderlyingObject();
+ const Value *bObj = B->getUnderlyingObject();
+ return AA->alias(aObj, AA->getTypeStoreSize(aObj->getType()),
+ bObj, AA->getTypeStoreSize(bObj->getType()));
+}
+
+static inline const SCEV *GetZeroSCEV(ScalarEvolution *SE) {
+ return SE->getConstant(Type::getInt32Ty(SE->getContext()), 0L);
+}
+
//===----------------------------------------------------------------------===//
// Dependence Testing
//===----------------------------------------------------------------------===//
-bool LoopDependenceAnalysis::isDependencePair(const Value *x,
- const Value *y) const {
- return IsMemRefInstr(x) &&
- IsMemRefInstr(y) &&
- (cast<const Instruction>(x)->mayWriteToMemory() ||
- cast<const Instruction>(y)->mayWriteToMemory());
+bool LoopDependenceAnalysis::isDependencePair(const Value *A,
+ const Value *B) const {
+ return IsMemRefInstr(A) &&
+ IsMemRefInstr(B) &&
+ (cast<const Instruction>(A)->mayWriteToMemory() ||
+ cast<const Instruction>(B)->mayWriteToMemory());
+}
+
+bool LoopDependenceAnalysis::findOrInsertDependencePair(Value *A,
+ Value *B,
+ DependencePair *&P) {
+ void *insertPos = 0;
+ FoldingSetNodeID id;
+ id.AddPointer(A);
+ id.AddPointer(B);
+
+ P = Pairs.FindNodeOrInsertPos(id, insertPos);
+ if (P) return true;
+
+ P = PairAllocator.Allocate<DependencePair>();
+ new (P) DependencePair(id, A, B);
+ Pairs.InsertNode(P, insertPos);
+ return false;
+}
+
+void LoopDependenceAnalysis::getLoops(const SCEV *S,
+ DenseSet<const Loop*>* Loops) const {
+ // Refactor this into an SCEVVisitor, if efficiency becomes a concern.
+ for (const Loop *L = this->L; L != 0; L = L->getParentLoop())
+ if (!S->isLoopInvariant(L))
+ Loops->insert(L);
+}
+
+bool LoopDependenceAnalysis::isLoopInvariant(const SCEV *S) const {
+ DenseSet<const Loop*> loops;
+ getLoops(S, &loops);
+ return loops.empty();
+}
+
+bool LoopDependenceAnalysis::isAffine(const SCEV *S) const {
+ const SCEVAddRecExpr *rec = dyn_cast<SCEVAddRecExpr>(S);
+ return isLoopInvariant(S) || (rec && rec->isAffine());
+}
+
+bool LoopDependenceAnalysis::isZIVPair(const SCEV *A, const SCEV *B) const {
+ return isLoopInvariant(A) && isLoopInvariant(B);
+}
+
+bool LoopDependenceAnalysis::isSIVPair(const SCEV *A, const SCEV *B) const {
+ DenseSet<const Loop*> loops;
+ getLoops(A, &loops);
+ getLoops(B, &loops);
+ return loops.size() == 1;
+}
+
+LoopDependenceAnalysis::DependenceResult
+LoopDependenceAnalysis::analyseZIV(const SCEV *A,
+ const SCEV *B,
+ Subscript *S) const {
+ assert(isZIVPair(A, B) && "Attempted to ZIV-test non-ZIV SCEVs!");
+ return A == B ? Dependent : Independent;
}
-bool LoopDependenceAnalysis::depends(Value *src, Value *dst) {
- assert(isDependencePair(src, dst) && "Values form no dependence pair!");
- DOUT << "== LDA test ==\n" << *src << *dst;
+LoopDependenceAnalysis::DependenceResult
+LoopDependenceAnalysis::analyseSIV(const SCEV *A,
+ const SCEV *B,
+ Subscript *S) const {
+ return Unknown; // TODO: Implement.
+}
+
+LoopDependenceAnalysis::DependenceResult
+LoopDependenceAnalysis::analyseMIV(const SCEV *A,
+ const SCEV *B,
+ Subscript *S) const {
+ return Unknown; // TODO: Implement.
+}
- // We only analyse loads and stores; for possible memory accesses by e.g.
- // free, call, or invoke instructions we conservatively assume dependence.
- if (!IsLoadOrStoreInst(src) || !IsLoadOrStoreInst(dst))
- return true;
+LoopDependenceAnalysis::DependenceResult
+LoopDependenceAnalysis::analyseSubscript(const SCEV *A,
+ const SCEV *B,
+ Subscript *S) const {
+ DEBUG(dbgs() << " Testing subscript: " << *A << ", " << *B << "\n");
- Value *srcPtr = GetPointerOperand(src);
- Value *dstPtr = GetPointerOperand(dst);
- const Value *srcObj = srcPtr->getUnderlyingObject();
- const Value *dstObj = dstPtr->getUnderlyingObject();
- AliasAnalysis::AliasResult alias = AA->alias(
- srcObj, AA->getTargetData().getTypeStoreSize(srcObj->getType()),
- dstObj, AA->getTargetData().getTypeStoreSize(dstObj->getType()));
+ if (A == B) {
+ DEBUG(dbgs() << " -> [D] same SCEV\n");
+ return Dependent;
+ }
- // If we don't know whether or not the two objects alias, assume dependence.
- if (alias == AliasAnalysis::MayAlias)
- return true;
+ if (!isAffine(A) || !isAffine(B)) {
+ DEBUG(dbgs() << " -> [?] not affine\n");
+ return Unknown;
+ }
- // If the objects noalias, they are distinct, accesses are independent.
- if (alias == AliasAnalysis::NoAlias)
- return false;
+ if (isZIVPair(A, B))
+ return analyseZIV(A, B, S);
- // TODO: the underlying objects MustAlias, test for dependence
+ if (isSIVPair(A, B))
+ return analyseSIV(A, B, S);
- // We couldn't establish a more precise result, so we have to conservatively
- // assume full dependence.
- return true;
+ return analyseMIV(A, B, S);
+}
+
+LoopDependenceAnalysis::DependenceResult
+LoopDependenceAnalysis::analysePair(DependencePair *P) const {
+ DEBUG(dbgs() << "Analysing:\n" << *P->A << "\n" << *P->B << "\n");
+
+ // We only analyse loads and stores but no possible memory accesses by e.g.
+ // free, call, or invoke instructions.
+ if (!IsLoadOrStoreInst(P->A) || !IsLoadOrStoreInst(P->B)) {
+ DEBUG(dbgs() << "--> [?] no load/store\n");
+ return Unknown;
+ }
+
+ Value *aPtr = GetPointerOperand(P->A);
+ Value *bPtr = GetPointerOperand(P->B);
+
+ switch (UnderlyingObjectsAlias(AA, aPtr, bPtr)) {
+ case AliasAnalysis::MayAlias:
+ // We can not analyse objects if we do not know about their aliasing.
+ DEBUG(dbgs() << "---> [?] may alias\n");
+ return Unknown;
+
+ case AliasAnalysis::NoAlias:
+ // If the objects noalias, they are distinct, accesses are independent.
+ DEBUG(dbgs() << "---> [I] no alias\n");
+ return Independent;
+
+ case AliasAnalysis::MustAlias:
+ break; // The underlying objects alias, test accesses for dependence.
+ }
+
+ const GEPOperator *aGEP = dyn_cast<GEPOperator>(aPtr);
+ const GEPOperator *bGEP = dyn_cast<GEPOperator>(bPtr);
+
+ if (!aGEP || !bGEP)
+ return Unknown;
+
+ // FIXME: Is filtering coupled subscripts necessary?
+
+ // Collect GEP operand pairs (FIXME: use GetGEPOperands from BasicAA), adding
+ // trailing zeroes to the smaller GEP, if needed.
+ typedef SmallVector<std::pair<const SCEV*, const SCEV*>, 4> GEPOpdPairsTy;
+ GEPOpdPairsTy opds;
+ for(GEPOperator::const_op_iterator aIdx = aGEP->idx_begin(),
+ aEnd = aGEP->idx_end(),
+ bIdx = bGEP->idx_begin(),
+ bEnd = bGEP->idx_end();
+ aIdx != aEnd && bIdx != bEnd;
+ aIdx += (aIdx != aEnd), bIdx += (bIdx != bEnd)) {
+ const SCEV* aSCEV = (aIdx != aEnd) ? SE->getSCEV(*aIdx) : GetZeroSCEV(SE);
+ const SCEV* bSCEV = (bIdx != bEnd) ? SE->getSCEV(*bIdx) : GetZeroSCEV(SE);
+ opds.push_back(std::make_pair(aSCEV, bSCEV));
+ }
+
+ if (!opds.empty() && opds[0].first != opds[0].second) {
+ // We cannot (yet) handle arbitrary GEP pointer offsets. By limiting
+ //
+ // TODO: this could be relaxed by adding the size of the underlying object
+ // to the first subscript. If we have e.g. (GEP x,0,i; GEP x,2,-i) and we
+ // know that x is a [100 x i8]*, we could modify the first subscript to be
+ // (i, 200-i) instead of (i, -i).
+ return Unknown;
+ }
+
+ // Now analyse the collected operand pairs (skipping the GEP ptr offsets).
+ for (GEPOpdPairsTy::const_iterator i = opds.begin() + 1, end = opds.end();
+ i != end; ++i) {
+ Subscript subscript;
+ DependenceResult result = analyseSubscript(i->first, i->second, &subscript);
+ if (result != Dependent) {
+ // We either proved independence or failed to analyse this subscript.
+ // Further subscripts will not improve the situation, so abort early.
+ return result;
+ }
+ P->Subscripts.push_back(subscript);
+ }
+ // We successfully analysed all subscripts but failed to prove independence.
+ return Dependent;
+}
+
+bool LoopDependenceAnalysis::depends(Value *A, Value *B) {
+ assert(isDependencePair(A, B) && "Values form no dependence pair!");
+ ++NumAnswered;
+
+ DependencePair *p;
+ if (!findOrInsertDependencePair(A, B, p)) {
+ // The pair is not cached, so analyse it.
+ ++NumAnalysed;
+ switch (p->Result = analysePair(p)) {
+ case Dependent: ++NumDependent; break;
+ case Independent: ++NumIndependent; break;
+ case Unknown: ++NumUnknown; break;
+ }
+ }
+ return p->Result != Independent;
}
//===----------------------------------------------------------------------===//
return false;
}
+void LoopDependenceAnalysis::releaseMemory() {
+ Pairs.clear();
+ PairAllocator.Reset();
+}
+
void LoopDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesAll();
AU.addRequiredTransitive<AliasAnalysis>();
AU.addRequiredTransitive<ScalarEvolution>();
}
-static void PrintLoopInfo(
- raw_ostream &OS, LoopDependenceAnalysis *LDA, const Loop *L) {
+static void PrintLoopInfo(raw_ostream &OS,
+ LoopDependenceAnalysis *LDA, const Loop *L) {
if (!L->empty()) return; // ignore non-innermost loops
SmallVector<Instruction*, 8> memrefs;
OS << " Load/store instructions: " << memrefs.size() << "\n";
for (SmallVector<Instruction*, 8>::const_iterator x = memrefs.begin(),
- end = memrefs.end(); x != end; ++x)
- OS << "\t" << (x - memrefs.begin()) << ": " << **x;
+ end = memrefs.end(); x != end; ++x)
+ OS << "\t" << (x - memrefs.begin()) << ": " << **x << "\n";
OS << " Pairwise dependence results:\n";
for (SmallVector<Instruction*, 8>::const_iterator x = memrefs.begin(),
- end = memrefs.end(); x != end; ++x)
+ end = memrefs.end(); x != end; ++x)
for (SmallVector<Instruction*, 8>::const_iterator y = x + 1;
- y != end; ++y)
+ y != end; ++y)
if (LDA->isDependencePair(*x, *y))
OS << "\t" << (x - memrefs.begin()) << "," << (y - memrefs.begin())
<< ": " << (LDA->depends(*x, *y) ? "dependent" : "independent")
// TODO: doc why const_cast is safe
PrintLoopInfo(OS, const_cast<LoopDependenceAnalysis*>(this), this->L);
}
-
-void LoopDependenceAnalysis::print(std::ostream &OS, const Module *M) const {
- raw_os_ostream os(OS);
- print(os, M);
-}