In D12090, the ExprValueMap was added to reuse existing value during SCEV expansion.
However, const folding and sext/zext distribution can make the reuse still difficult.
A simplified case is: suppose we know S1 expands to V1 in ExprValueMap, and
S1 = S2 + C_a
S3 = S2 + C_b
where C_a and C_b are different SCEVConstants. Then we'd like to expand S3 as
V1 - C_a + C_b instead of expanding S2 literally. It is helpful when S2 is a
complex SCEV expr and S2 has no entry in ExprValueMap, which is usually caused
by the fact that S3 is generated from S1 after const folding.
In order to do that, we represent ExprValueMap as a mapping from SCEV to
ValueOffsetPair. We will save both S1->{V1, 0} and S2->{V1, C_a} into the
ExprValueMap when we create SCEV for V1. When S3 is expanded, it will first
expand S2 to V1 - C_a because of S2->{V1, C_a} in the map, then expand S3 to
V1 - C_a + C_b.
Differential Revision: https://reviews.llvm.org/D21313
llvm-svn: 276136
/// The typedef for ExprValueMap.
///
- typedef DenseMap<const SCEV *, SetVector<Value *>> ExprValueMapType;
+ typedef std::pair<Value *, ConstantInt *> ValueOffsetPair;
+ typedef DenseMap<const SCEV *, SetVector<ValueOffsetPair>> ExprValueMapType;
/// ExprValueMap -- This map records the original values from which
/// the SCEV expr is generated from.
+ ///
+ /// We want to represent the mapping as SCEV -> ValueOffsetPair instead
+ /// of SCEV -> Value:
+ /// Suppose we know S1 expands to V1, and
+ /// S1 = S2 + C_a
+ /// S3 = S2 + C_b
+ /// where C_a and C_b are different SCEVConstants. Then we'd like to
+ /// expand S3 as V1 - C_a + C_b instead of expanding S2 literally.
+ /// It is helpful when S2 is a complex SCEV expr.
+ ///
+ /// In order to do that, we represent ExprValueMap as a mapping from
+ /// SCEV to ValueOffsetPair. We will save both S1->{V1, 0} and
+ /// S2->{V1, C_a} into the map when we create SCEV for V1. When S3
+ /// is expanded, it will first expand S2 to V1 - C_a because of
+ /// S2->{V1, C_a} in the map, then expand S3 to V1 - C_a + C_b.
+ ///
+ /// Note: S->{V, Offset} in the ExprValueMap means S can be expanded
+ /// to V - Offset.
ExprValueMapType ExprValueMap;
/// The typedef for ValueExprMap.
bool containsAddRecurrence(const SCEV *S);
/// Return the Value set from which the SCEV expr is generated.
- SetVector<Value *> *getSCEVValues(const SCEV *S);
+ SetVector<ValueOffsetPair> *getSCEVValues(const SCEV *S);
/// Erase Value from ValueExprMap and ExprValueMap.
void eraseValueFromMap(Value *V);
PointerType *PTy, Type *Ty, Value *V);
/// \brief Find a previous Value in ExprValueMap for expand.
- Value *FindValueInExprValueMap(const SCEV *S, const Instruction *InsertPt);
+ ScalarEvolution::ValueOffsetPair
+ FindValueInExprValueMap(const SCEV *S, const Instruction *InsertPt);
Value *expand(const SCEV *S);
return F.FoundOne;
}
-/// Return the Value set from S.
-SetVector<Value *> *ScalarEvolution::getSCEVValues(const SCEV *S) {
+/// Try to split a SCEVAddExpr into a pair of {SCEV, ConstantInt}.
+/// If \p S is a SCEVAddExpr and is composed of a sub SCEV S' and an
+/// offset I, then return {S', I}, else return {\p S, nullptr}.
+static std::pair<const SCEV *, ConstantInt *> splitAddExpr(const SCEV *S) {
+ const auto *Add = dyn_cast<SCEVAddExpr>(S);
+ if (!Add)
+ return {S, nullptr};
+
+ if (Add->getNumOperands() != 2)
+ return {S, nullptr};
+
+ auto *ConstOp = dyn_cast<SCEVConstant>(Add->getOperand(0));
+ if (!ConstOp)
+ return {S, nullptr};
+
+ return {Add->getOperand(1), ConstOp->getValue()};
+}
+
+/// Return the ValueOffsetPair set for \p S. \p S can be represented
+/// by the value and offset from any ValueOffsetPair in the set.
+SetVector<ScalarEvolution::ValueOffsetPair> *
+ScalarEvolution::getSCEVValues(const SCEV *S) {
ExprValueMapType::iterator SI = ExprValueMap.find_as(S);
if (SI == ExprValueMap.end())
return nullptr;
if (VerifySCEVMap) {
// Check there is no dangling Value in the set returned.
for (const auto &VE : SI->second)
- assert(ValueExprMap.count(VE));
+ assert(ValueExprMap.count(VE.first));
}
#endif
return &SI->second;
}
-/// Erase Value from ValueExprMap and ExprValueMap. If ValueExprMap.erase(V) is
-/// not used together with forgetMemoizedResults(S), eraseValueFromMap should be
-/// used instead to ensure whenever V->S is removed from ValueExprMap, V is also
-/// removed from the set of ExprValueMap[S].
+/// Erase Value from ValueExprMap and ExprValueMap. ValueExprMap.erase(V)
+/// cannot be used separately. eraseValueFromMap should be used to remove
+/// V from ValueExprMap and ExprValueMap at the same time.
void ScalarEvolution::eraseValueFromMap(Value *V) {
ValueExprMapType::iterator I = ValueExprMap.find_as(V);
if (I != ValueExprMap.end()) {
const SCEV *S = I->second;
- SetVector<Value *> *SV = getSCEVValues(S);
- // Remove V from the set of ExprValueMap[S]
- if (SV)
- SV->remove(V);
+ // Remove {V, 0} from the set of ExprValueMap[S]
+ if (SetVector<ValueOffsetPair> *SV = getSCEVValues(S))
+ SV->remove({V, nullptr});
+
+ // Remove {V, Offset} from the set of ExprValueMap[Stripped]
+ const SCEV *Stripped;
+ ConstantInt *Offset;
+ std::tie(Stripped, Offset) = splitAddExpr(S);
+ if (Offset != nullptr) {
+ if (SetVector<ValueOffsetPair> *SV = getSCEVValues(Stripped))
+ SV->remove({V, Offset});
+ }
ValueExprMap.erase(V);
}
}
S = createSCEV(V);
// During PHI resolution, it is possible to create two SCEVs for the same
// V, so it is needed to double check whether V->S is inserted into
- // ValueExprMap before insert S->V into ExprValueMap.
+ // ValueExprMap before insert S->{V, 0} into ExprValueMap.
std::pair<ValueExprMapType::iterator, bool> Pair =
ValueExprMap.insert({SCEVCallbackVH(V, this), S});
- if (Pair.second)
- ExprValueMap[S].insert(V);
+ if (Pair.second) {
+ ExprValueMap[S].insert({V, nullptr});
+
+ // If S == Stripped + Offset, add Stripped -> {V, Offset} into
+ // ExprValueMap.
+ const SCEV *Stripped = S;
+ ConstantInt *Offset = nullptr;
+ std::tie(Stripped, Offset) = splitAddExpr(S);
+ // If stripped is SCEVUnknown, don't bother to save
+ // Stripped -> {V, offset}. It doesn't simplify and sometimes even
+ // increase the complexity of the expansion code.
+ // If V is GetElementPtrInst, don't save Stripped -> {V, offset}
+ // because it may generate add/sub instead of GEP in SCEV expansion.
+ if (Offset != nullptr && !isa<SCEVUnknown>(Stripped) &&
+ !isa<GetElementPtrInst>(V))
+ ExprValueMap[Stripped].insert({V, Offset});
+ }
}
return S;
}
const SCEV *S = I->second;
if (checkValidity(S))
return S;
+ eraseValueFromMap(V);
forgetMemoizedResults(S);
- ValueExprMap.erase(I);
}
return nullptr;
}
if (!isa<PHINode>(I) ||
!isa<SCEVUnknown>(Old) ||
(I != PN && Old == SymName)) {
+ eraseValueFromMap(It->first);
forgetMemoizedResults(Old);
- ValueExprMap.erase(It);
}
}
// to create an AddRecExpr for this PHI node. We can not keep this temporary
// as it will prevent later (possibly simpler) SCEV expressions to be added
// to the ValueExprMap.
- ValueExprMap.erase(PN);
+ eraseValueFromMap(PN);
}
return nullptr;
// case, createNodeForPHI will perform the necessary updates on its
// own when it gets to that point.
if (!isa<PHINode>(I) || !isa<SCEVUnknown>(Old)) {
+ eraseValueFromMap(It->first);
forgetMemoizedResults(Old);
- ValueExprMap.erase(It);
}
if (PHINode *PN = dyn_cast<PHINode>(I))
ConstantEvolutionLoopExitValue.erase(PN);
ValueExprMapType::iterator It =
ValueExprMap.find_as(static_cast<Value *>(I));
if (It != ValueExprMap.end()) {
+ eraseValueFromMap(It->first);
forgetMemoizedResults(It->second);
- ValueExprMap.erase(It);
if (PHINode *PN = dyn_cast<PHINode>(I))
ConstantEvolutionLoopExitValue.erase(PN);
}
ValueExprMapType::iterator It =
ValueExprMap.find_as(static_cast<Value *>(I));
if (It != ValueExprMap.end()) {
+ eraseValueFromMap(It->first);
forgetMemoizedResults(It->second);
- ValueExprMap.erase(It);
if (PHINode *PN = dyn_cast<PHINode>(I))
ConstantEvolutionLoopExitValue.erase(PN);
}
return V;
}
-Value *SCEVExpander::FindValueInExprValueMap(const SCEV *S,
- const Instruction *InsertPt) {
- SetVector<Value *> *Set = SE.getSCEVValues(S);
+ScalarEvolution::ValueOffsetPair
+SCEVExpander::FindValueInExprValueMap(const SCEV *S,
+ const Instruction *InsertPt) {
+ SetVector<ScalarEvolution::ValueOffsetPair> *Set = SE.getSCEVValues(S);
// If the expansion is not in CanonicalMode, and the SCEV contains any
// sub scAddRecExpr type SCEV, it is required to expand the SCEV literally.
if (CanonicalMode || !SE.containsAddRecurrence(S)) {
// Choose a Value from the set which dominates the insertPt.
// insertPt should be inside the Value's parent loop so as not to break
// the LCSSA form.
- for (auto const &Ent : *Set) {
+ for (auto const &VOPair : *Set) {
+ Value *V = VOPair.first;
+ ConstantInt *Offset = VOPair.second;
Instruction *EntInst = nullptr;
- if (Ent && isa<Instruction>(Ent) &&
- (EntInst = cast<Instruction>(Ent)) &&
- S->getType() == Ent->getType() &&
+ if (V && isa<Instruction>(V) && (EntInst = cast<Instruction>(V)) &&
+ S->getType() == V->getType() &&
EntInst->getFunction() == InsertPt->getFunction() &&
SE.DT.dominates(EntInst, InsertPt) &&
(SE.LI.getLoopFor(EntInst->getParent()) == nullptr ||
- SE.LI.getLoopFor(EntInst->getParent())->contains(InsertPt))) {
- return Ent;
- }
+ SE.LI.getLoopFor(EntInst->getParent())->contains(InsertPt)))
+ return {V, Offset};
}
}
}
- return nullptr;
+ return {nullptr, nullptr};
}
// The expansion of SCEV will either reuse a previous Value in ExprValueMap,
Builder.SetInsertPoint(InsertPt);
// Expand the expression into instructions.
- Value *V = FindValueInExprValueMap(S, InsertPt);
+ ScalarEvolution::ValueOffsetPair VO = FindValueInExprValueMap(S, InsertPt);
+ Value *V = VO.first;
if (!V)
V = visit(S);
+ else if (VO.second) {
+ V = Builder.CreateSub(V, VO.second);
+ }
// Remember the expanded value for this SCEV at this location.
//
// Use expand's logic which is used for reusing a previous Value in
// ExprValueMap.
- if (Value *Val = FindValueInExprValueMap(S, At))
+ if (Value *Val = FindValueInExprValueMap(S, At).first)
return Val;
// There is potential to make this significantly smarter, but this simple
--- /dev/null
+; RUN: opt < %s -loop-vectorize -force-vector-interleave=1 -force-vector-width=4 -S |FileCheck %s
+; SCEV expansion uses existing value or value + offset to reduce duplicate code expansion so foo should only generate one select inst after loop vectorization.
+; CHECK-LABEL: @foo(
+; CHECK: select
+; CHECK-NOT: select
+
+@ySrcL = common global i8* null, align 8
+@smL = common global i32 0, align 4
+
+define void @foo(i32 %rwL, i32 %kL, i32 %xfL) {
+entry:
+ %sub = add nsw i32 %rwL, -1
+ %shr = ashr i32 %xfL, 6
+ %cmp.i = icmp slt i32 %sub, %shr
+ %cond.i = select i1 %cmp.i, i32 %sub, i32 %shr
+ %cmp6 = icmp sgt i32 %cond.i, %kL
+ br i1 %cmp6, label %for.body.lr.ph, label %for.end
+
+for.body.lr.ph: ; preds = %entry
+ %tmp = load i8*, i8** @ySrcL, align 8
+ %tmp1 = sext i32 %kL to i64
+ %tmp2 = sext i32 %cond.i to i64
+ br label %for.body
+
+for.body: ; preds = %for.body, %for.body.lr.ph
+ %indvars.iv = phi i64 [ %tmp1, %for.body.lr.ph ], [ %indvars.iv.next, %for.body ]
+ %reduct.07 = phi i32 [ 0, %for.body.lr.ph ], [ %add, %for.body ]
+ %arrayidx = getelementptr inbounds i8, i8* %tmp, i64 %indvars.iv
+ %tmp3 = load i8, i8* %arrayidx, align 1
+ %conv = zext i8 %tmp3 to i32
+ %add = add nsw i32 %conv, %reduct.07
+ %indvars.iv.next = add nsw i64 %indvars.iv, 1
+ %cmp = icmp slt i64 %indvars.iv.next, %tmp2
+ br i1 %cmp, label %for.body, label %for.end.loopexit
+
+for.end.loopexit: ; preds = %for.body
+ %add.lcssa = phi i32 [ %add, %for.body ]
+ br label %for.end
+
+for.end: ; preds = %for.end.loopexit, %entry
+ %reduct.0.lcssa = phi i32 [ 0, %entry ], [ %add.lcssa, %for.end.loopexit ]
+ store i32 %reduct.0.lcssa, i32* @smL, align 4
+ ret void
+}
+; RUN: opt < %s -loop-vectorize -force-vector-interleave=1 -force-vector-width=4 -S |FileCheck %s
+; SCEV expansion uses existing value or value + offset to reduce duplicate code expansion so foo should only generate one select inst after loop vectorization.
+; CHECK-LABEL: @foo(
+; CHECK: select
+; CHECK-NOT: select
+
+@ySrcL = common global i8* null, align 8
+@smL = common global i32 0, align 4
+
+define void @foo(i32 %rwL, i32 %kL, i32 %xfL) {
+entry:
+ %sub = add nsw i32 %rwL, -1
+ %shr = ashr i32 %xfL, 6
+ %cmp.i = icmp slt i32 %sub, %shr
+ %cond.i = select i1 %cmp.i, i32 %sub, i32 %shr
+ %cmp6 = icmp sgt i32 %cond.i, %kL
+ br i1 %cmp6, label %for.body.lr.ph, label %for.end
+
+for.body.lr.ph: ; preds = %entry
+ %tmp = load i8*, i8** @ySrcL, align 8
+ %tmp1 = sext i32 %kL to i64
+ %tmp2 = sext i32 %cond.i to i64
+ br label %for.body
+
+for.body: ; preds = %for.body, %for.body.lr.ph
+ %indvars.iv = phi i64 [ %tmp1, %for.body.lr.ph ], [ %indvars.iv.next, %for.body ]
+ %reduct.07 = phi i32 [ 0, %for.body.lr.ph ], [ %add, %for.body ]
+ %arrayidx = getelementptr inbounds i8, i8* %tmp, i64 %indvars.iv
+ %tmp3 = load i8, i8* %arrayidx, align 1
+ %conv = zext i8 %tmp3 to i32
+ %add = add nsw i32 %conv, %reduct.07
+ %indvars.iv.next = add nsw i64 %indvars.iv, 1
+ %cmp = icmp slt i64 %indvars.iv.next, %tmp2
+ br i1 %cmp, label %for.body, label %for.end.loopexit
+
+for.end.loopexit: ; preds = %for.body
+ %add.lcssa = phi i32 [ %add, %for.body ]
+ br label %for.end
+
+for.end: ; preds = %for.end.loopexit, %entry
+ %reduct.0.lcssa = phi i32 [ 0, %entry ], [ %add.lcssa, %for.end.loopexit ]
+ store i32 %reduct.0.lcssa, i32* @smL, align 4
+ ret void
+}