[VPlan] Support VPReverseVectorPointer in DataWithEVL vectorization #113667
+48
−32
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@llvm/pr-subscribers-vectorizers @llvm/pr-subscribers-backend-risc-v Author: Shih-Po Hung (arcbbb) ChangesThis patch replaces VF with EVL for VPReverseVectorPointer in DataWithEVL Tail-Folding VPReverseVectorPointer relies on runtime VF, so it’s necessary to replace VF with EVL in DataWithEVL tail-folding, as EVL can be less than VF at runtime. Since VPReverseVectorPointer does not depend on the header mask and is created only for the widen load/store operations, the replacement is applied while the widen load/store recipe is matched. This depends on #113410 Patch is 113.90 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/113667.diff 29 Files Affected:
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 6e082b1c134dee..220625908dff9e 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -4442,6 +4442,7 @@ static bool willGenerateVectors(VPlan &Plan, ElementCount VF,
case VPDef::VPInstructionSC:
case VPDef::VPCanonicalIVPHISC:
case VPDef::VPVectorPointerSC:
+ case VPDef::VPReverseVectorPointerSC:
case VPDef::VPExpandSCEVSC:
case VPDef::VPEVLBasedIVPHISC:
case VPDef::VPPredInstPHISC:
@@ -8160,9 +8161,15 @@ VPRecipeBuilder::tryToWidenMemory(Instruction *I, ArrayRef<VPValue *> Operands,
if (Consecutive) {
auto *GEP = dyn_cast<GetElementPtrInst>(
Ptr->getUnderlyingValue()->stripPointerCasts());
- auto *VectorPtr = new VPVectorPointerRecipe(
- Ptr, getLoadStoreType(I), Reverse, GEP ? GEP->isInBounds() : false,
- I->getDebugLoc());
+ VPSingleDefRecipe *VectorPtr;
+ if (Reverse)
+ VectorPtr = new VPReverseVectorPointerRecipe(
+ Ptr, &Plan.getVF(), getLoadStoreType(I),
+ GEP ? GEP->isInBounds() : false, I->getDebugLoc());
+ else
+ VectorPtr = new VPVectorPointerRecipe(Ptr, getLoadStoreType(I),
+ GEP ? GEP->isInBounds() : false,
+ I->getDebugLoc());
Builder.getInsertBlock()->appendRecipe(VectorPtr);
Ptr = VectorPtr;
}
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index 594492344d43ce..8df63b5796c739 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -882,6 +882,7 @@ class VPSingleDefRecipe : public VPRecipeBase, public VPValue {
case VPRecipeBase::VPReplicateSC:
case VPRecipeBase::VPScalarIVStepsSC:
case VPRecipeBase::VPVectorPointerSC:
+ case VPRecipeBase::VPReverseVectorPointerSC:
case VPRecipeBase::VPWidenCallSC:
case VPRecipeBase::VPWidenCanonicalIVSC:
case VPRecipeBase::VPWidenCastSC:
@@ -1078,6 +1079,7 @@ class VPRecipeWithIRFlags : public VPSingleDefRecipe {
R->getVPDefID() == VPRecipeBase::VPWidenGEPSC ||
R->getVPDefID() == VPRecipeBase::VPWidenCastSC ||
R->getVPDefID() == VPRecipeBase::VPReplicateSC ||
+ R->getVPDefID() == VPRecipeBase::VPReverseVectorPointerSC ||
R->getVPDefID() == VPRecipeBase::VPVectorPointerSC;
}
@@ -1785,20 +1787,64 @@ class VPWidenGEPRecipe : public VPRecipeWithIRFlags {
#endif
};
-/// A recipe to compute the pointers for widened memory accesses of IndexTy for
-/// all parts. If IsReverse is true, compute pointers for accessing the input in
-/// reverse order per part.
+/// A recipe to compute the pointers for widened memory accesses of IndexTy
+/// in reverse order.
+class VPReverseVectorPointerRecipe : public VPRecipeWithIRFlags,
+ public VPUnrollPartAccessor<2> {
+ Type *IndexedTy;
+
+public:
+ VPReverseVectorPointerRecipe(VPValue *Ptr, VPValue *VF, Type *IndexedTy,
+ bool IsInBounds, DebugLoc DL)
+ : VPRecipeWithIRFlags(VPDef::VPReverseVectorPointerSC,
+ ArrayRef<VPValue *>({Ptr, VF}),
+ GEPFlagsTy(IsInBounds), DL),
+ IndexedTy(IndexedTy) {}
+
+ VP_CLASSOF_IMPL(VPDef::VPReverseVectorPointerSC)
+
+ VPValue *getVFValue() { return getOperand(1); }
+ const VPValue *getVFValue() const { return getOperand(1); }
+
+ void execute(VPTransformState &State) override;
+
+ bool onlyFirstLaneUsed(const VPValue *Op) const override {
+ assert(is_contained(operands(), Op) &&
+ "Op must be an operand of the recipe");
+ return true;
+ }
+
+ /// Returns true if the recipe only uses the first part of operand \p Op.
+ bool onlyFirstPartUsed(const VPValue *Op) const override {
+ assert(is_contained(operands(), Op) &&
+ "Op must be an operand of the recipe");
+ assert(getNumOperands() <= 2 && "must have at most two operands");
+ return true;
+ }
+
+ VPReverseVectorPointerRecipe *clone() override {
+ return new VPReverseVectorPointerRecipe(
+ getOperand(0), getVFValue(), IndexedTy, isInBounds(), getDebugLoc());
+ }
+
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
+ /// Print the recipe.
+ void print(raw_ostream &O, const Twine &Indent,
+ VPSlotTracker &SlotTracker) const override;
+#endif
+};
+
+/// A recipe to compute the pointers for widened memory accesses of IndexTy.
class VPVectorPointerRecipe : public VPRecipeWithIRFlags,
public VPUnrollPartAccessor<1> {
Type *IndexedTy;
- bool IsReverse;
public:
- VPVectorPointerRecipe(VPValue *Ptr, Type *IndexedTy, bool IsReverse,
- bool IsInBounds, DebugLoc DL)
+ VPVectorPointerRecipe(VPValue *Ptr, Type *IndexedTy, bool IsInBounds,
+ DebugLoc DL)
: VPRecipeWithIRFlags(VPDef::VPVectorPointerSC, ArrayRef<VPValue *>(Ptr),
GEPFlagsTy(IsInBounds), DL),
- IndexedTy(IndexedTy), IsReverse(IsReverse) {}
+ IndexedTy(IndexedTy) {}
VP_CLASSOF_IMPL(VPDef::VPVectorPointerSC)
@@ -1819,8 +1865,8 @@ class VPVectorPointerRecipe : public VPRecipeWithIRFlags,
}
VPVectorPointerRecipe *clone() override {
- return new VPVectorPointerRecipe(getOperand(0), IndexedTy, IsReverse,
- isInBounds(), getDebugLoc());
+ return new VPVectorPointerRecipe(getOperand(0), IndexedTy, isInBounds(),
+ getDebugLoc());
}
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
diff --git a/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp b/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp
index 277df0637372d8..7bfbcc3dff0e1f 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp
@@ -261,9 +261,10 @@ Type *VPTypeAnalysis::inferScalarType(const VPValue *V) {
[](const auto *R) { return R->getScalarType(); })
.Case<VPReductionRecipe, VPPredInstPHIRecipe, VPWidenPHIRecipe,
VPScalarIVStepsRecipe, VPWidenGEPRecipe, VPVectorPointerRecipe,
- VPWidenCanonicalIVRecipe>([this](const VPRecipeBase *R) {
- return inferScalarType(R->getOperand(0));
- })
+ VPReverseVectorPointerRecipe, VPWidenCanonicalIVRecipe>(
+ [this](const VPRecipeBase *R) {
+ return inferScalarType(R->getOperand(0));
+ })
.Case<VPBlendRecipe, VPInstruction, VPWidenRecipe, VPWidenEVLRecipe,
VPReplicateRecipe, VPWidenCallRecipe, VPWidenMemoryRecipe,
VPWidenSelectRecipe>(
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index 0d092b9c10acc8..1b715a28604e33 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -154,6 +154,8 @@ bool VPRecipeBase::mayHaveSideEffects() const {
case VPDerivedIVSC:
case VPPredInstPHISC:
case VPScalarCastSC:
+ case VPVectorPointerSC:
+ case VPReverseVectorPointerSC:
return false;
case VPInstructionSC:
return mayWriteToMemory();
@@ -1813,38 +1815,63 @@ void VPWidenGEPRecipe::print(raw_ostream &O, const Twine &Indent,
}
#endif
-void VPVectorPointerRecipe ::execute(VPTransformState &State) {
- auto &Builder = State.Builder;
- State.setDebugLocFrom(getDebugLoc());
- unsigned CurrentPart = getUnrollPart(*this);
+static Type *getGEPIndexTy(bool IsScalable, bool IsReverse,
+ unsigned CurrentPart, IRBuilderBase &Builder) {
// Use i32 for the gep index type when the value is constant,
// or query DataLayout for a more suitable index type otherwise.
const DataLayout &DL = Builder.GetInsertBlock()->getDataLayout();
- Type *IndexTy = State.VF.isScalable() && (IsReverse || CurrentPart > 0)
- ? DL.getIndexType(Builder.getPtrTy(0))
- : Builder.getInt32Ty();
+ return IsScalable && (IsReverse || CurrentPart > 0)
+ ? DL.getIndexType(Builder.getPtrTy(0))
+ : Builder.getInt32Ty();
+}
+
+void VPReverseVectorPointerRecipe::execute(VPTransformState &State) {
+ auto &Builder = State.Builder;
+ State.setDebugLocFrom(getDebugLoc());
+ unsigned CurrentPart = getUnrollPart(*this);
+ Type *IndexTy = getGEPIndexTy(State.VF.isScalable(), /*IsReverse*/ true,
+ CurrentPart, Builder);
+
+ // The wide store needs to start at the last vector element.
+ Value *RunTimeVF = State.get(getVFValue(), VPLane(0));
+ if (IndexTy != RunTimeVF->getType())
+ RunTimeVF = Builder.CreateZExtOrTrunc(RunTimeVF, IndexTy);
+ // NumElt = -CurrentPart * RunTimeVF
+ Value *NumElt = Builder.CreateMul(
+ ConstantInt::get(IndexTy, -(int64_t)CurrentPart), RunTimeVF);
+ // LastLane = 1 - RunTimeVF
+ Value *LastLane = Builder.CreateSub(ConstantInt::get(IndexTy, 1), RunTimeVF);
Value *Ptr = State.get(getOperand(0), VPLane(0));
bool InBounds = isInBounds();
+ Value *ResultPtr = Builder.CreateGEP(IndexedTy, Ptr, NumElt, "", InBounds);
+ ResultPtr = Builder.CreateGEP(IndexedTy, ResultPtr, LastLane, "", InBounds);
- Value *ResultPtr = nullptr;
- if (IsReverse) {
- // If the address is consecutive but reversed, then the
- // wide store needs to start at the last vector element.
- // RunTimeVF = VScale * VF.getKnownMinValue()
- // For fixed-width VScale is 1, then RunTimeVF = VF.getKnownMinValue()
- Value *RunTimeVF = getRuntimeVF(Builder, IndexTy, State.VF);
- // NumElt = -CurrentPart * RunTimeVF
- Value *NumElt = Builder.CreateMul(
- ConstantInt::get(IndexTy, -(int64_t)CurrentPart), RunTimeVF);
- // LastLane = 1 - RunTimeVF
- Value *LastLane =
- Builder.CreateSub(ConstantInt::get(IndexTy, 1), RunTimeVF);
- ResultPtr = Builder.CreateGEP(IndexedTy, Ptr, NumElt, "", InBounds);
- ResultPtr = Builder.CreateGEP(IndexedTy, ResultPtr, LastLane, "", InBounds);
- } else {
- Value *Increment = createStepForVF(Builder, IndexTy, State.VF, CurrentPart);
- ResultPtr = Builder.CreateGEP(IndexedTy, Ptr, Increment, "", InBounds);
- }
+ State.set(this, ResultPtr, /*IsScalar*/ true);
+}
+
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
+void VPReverseVectorPointerRecipe::print(raw_ostream &O, const Twine &Indent,
+ VPSlotTracker &SlotTracker) const {
+ O << Indent;
+ printAsOperand(O, SlotTracker);
+ O << " = reverse-vector-pointer ";
+ if (isInBounds())
+ O << "inbounds ";
+ printOperands(O, SlotTracker);
+}
+#endif
+
+void VPVectorPointerRecipe::execute(VPTransformState &State) {
+ auto &Builder = State.Builder;
+ State.setDebugLocFrom(getDebugLoc());
+ unsigned CurrentPart = getUnrollPart(*this);
+ Type *IndexTy = getGEPIndexTy(State.VF.isScalable(), /*IsReverse*/ false,
+ CurrentPart, Builder);
+ Value *Ptr = State.get(getOperand(0), VPLane(0));
+ bool InBounds = isInBounds();
+
+ Value *Increment = createStepForVF(Builder, IndexTy, State.VF, CurrentPart);
+ Value *ResultPtr = Builder.CreateGEP(IndexedTy, Ptr, Increment, "", InBounds);
State.set(this, ResultPtr, /*IsScalar*/ true);
}
@@ -1855,8 +1882,9 @@ void VPVectorPointerRecipe::print(raw_ostream &O, const Twine &Indent,
O << Indent;
printAsOperand(O, SlotTracker);
O << " = vector-pointer ";
- if (IsReverse)
- O << "(reverse) ";
+
+ if (isInBounds())
+ O << "inbounds ";
printOperands(O, SlotTracker);
}
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index a878613c4ba483..9a3b8939751f53 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -1361,15 +1361,22 @@ static void transformRecipestoEVLRecipes(VPlan &Plan, VPValue &EVL) {
assert(OrigMask && "Unmasked recipe when folding tail");
return HeaderMask == OrigMask ? nullptr : OrigMask;
};
+ auto SetEVLForReversePointer = [&EVL](VPValue *V) -> void {
+ if (auto R =
+ dyn_cast<VPReverseVectorPointerRecipe>(V->getDefiningRecipe()))
+ R->setOperand(1, &EVL);
+ };
VPRecipeBase *NewRecipe =
TypeSwitch<VPRecipeBase *, VPRecipeBase *>(CurRecipe)
.Case<VPWidenLoadRecipe>([&](VPWidenLoadRecipe *L) {
VPValue *NewMask = GetNewMask(L->getMask());
+ SetEVLForReversePointer(L->getOperand(0));
return new VPWidenLoadEVLRecipe(*L, EVL, NewMask);
})
.Case<VPWidenStoreRecipe>([&](VPWidenStoreRecipe *S) {
VPValue *NewMask = GetNewMask(S->getMask());
+ SetEVLForReversePointer(S->getOperand(0));
return new VPWidenStoreEVLRecipe(*S, EVL, NewMask);
})
.Case<VPWidenRecipe>([&](VPWidenRecipe *W) -> VPRecipeBase * {
diff --git a/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp b/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
index ca78f32506ef71..1e32865e8ee576 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
@@ -316,12 +316,12 @@ void UnrollState::unrollRecipeByUF(VPRecipeBase &R) {
// Add operand indicating the part to generate code for, to recipes still
// requiring it.
if (isa<VPScalarIVStepsRecipe, VPWidenCanonicalIVRecipe,
- VPVectorPointerRecipe>(Copy) ||
+ VPVectorPointerRecipe, VPReverseVectorPointerRecipe>(Copy) ||
match(Copy, m_VPInstruction<VPInstruction::CanonicalIVIncrementForPart>(
m_VPValue())))
Copy->addOperand(getConstantVPV(Part));
- if (isa<VPVectorPointerRecipe>(R))
+ if (isa<VPVectorPointerRecipe, VPReverseVectorPointerRecipe>(R))
Copy->setOperand(0, R.getOperand(0));
}
}
diff --git a/llvm/lib/Transforms/Vectorize/VPlanValue.h b/llvm/lib/Transforms/Vectorize/VPlanValue.h
index 4c383244f96f1a..d7626e437db338 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanValue.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanValue.h
@@ -346,6 +346,7 @@ class VPDef {
VPScalarCastSC,
VPScalarIVStepsSC,
VPVectorPointerSC,
+ VPReverseVectorPointerSC,
VPWidenCallSC,
VPWidenCanonicalIVSC,
VPWidenCastSC,
diff --git a/llvm/lib/Transforms/Vectorize/VPlanVerifier.cpp b/llvm/lib/Transforms/Vectorize/VPlanVerifier.cpp
index 99bc4c38a3c3cd..ea9794bcd59530 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanVerifier.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanVerifier.cpp
@@ -153,6 +153,10 @@ bool VPlanVerifier::verifyEVLRecipe(const VPInstruction &EVL) const {
})
.Case<VPScalarCastRecipe>(
[&](const VPScalarCastRecipe *S) { return true; })
+ .Case<VPReverseVectorPointerRecipe>(
+ [&](const VPReverseVectorPointerRecipe *R) {
+ return VerifyEVLUse(*R, 1);
+ })
.Case<VPInstruction>([&](const VPInstruction *I) {
if (I->getOpcode() != Instruction::Add) {
errs()
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/induction-costs.ll b/llvm/test/Transforms/LoopVectorize/AArch64/induction-costs.ll
index 8080c3a9ba0a7d..36eee8d0c98cea 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/induction-costs.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/induction-costs.ll
@@ -110,7 +110,6 @@ define i64 @pointer_induction_only(ptr %start, ptr %end) {
; CHECK-NEXT: [[OFFSET_IDX:%.*]] = mul i64 [[INDEX]], 4
; CHECK-NEXT: [[TMP4:%.*]] = add i64 [[OFFSET_IDX]], 0
; CHECK-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[START]], i64 [[TMP4]]
-; CHECK-NEXT: [[TMP6:%.*]] = getelementptr i32, ptr [[NEXT_GEP]], i32 0
; CHECK-NEXT: [[TMP7:%.*]] = getelementptr i32, ptr [[NEXT_GEP]], i32 2
; CHECK-NEXT: [[WIDE_LOAD4:%.*]] = load <2 x i32>, ptr [[TMP7]], align 1
; CHECK-NEXT: [[TMP9]] = zext <2 x i32> [[WIDE_LOAD4]] to <2 x i64>
@@ -169,7 +168,6 @@ define i64 @int_and_pointer_iv(ptr %start, i32 %N) {
; CHECK-NEXT: [[OFFSET_IDX:%.*]] = mul i64 [[INDEX]], 4
; CHECK-NEXT: [[TMP0:%.*]] = add i64 [[OFFSET_IDX]], 0
; CHECK-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[START]], i64 [[TMP0]]
-; CHECK-NEXT: [[TMP2:%.*]] = getelementptr i32, ptr [[NEXT_GEP]], i32 0
; CHECK-NEXT: [[TMP3:%.*]] = getelementptr i32, ptr [[NEXT_GEP]], i32 4
; CHECK-NEXT: [[WIDE_LOAD3:%.*]] = load <4 x i32>, ptr [[TMP3]], align 4
; CHECK-NEXT: [[TMP5]] = zext <4 x i32> [[WIDE_LOAD3]] to <4 x i64>
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/sve-vector-reverse.ll b/llvm/test/Transforms/LoopVectorize/AArch64/sve-vector-reverse.ll
index 81121019efe767..76562e80fbc4a1 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/sve-vector-reverse.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/sve-vector-reverse.ll
@@ -24,43 +24,36 @@ define void @vector_reverse_f64(i64 %N, ptr noalias %a, ptr noalias %b) #0{
; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], [[TMP3]]
; CHECK-NEXT: [[N_VEC:%.*]] = sub nsw i64 [[N]], [[N_MOD_VF]]
; CHECK-NEXT: [[TMP4:%.*]] = call i64 @llvm.vscale.i64()
-; CHECK-NEXT: [[TMP5:%.*]] = shl i64 [[TMP4]], 4
+; CHECK-NEXT: [[TMP5:%.*]] = shl i64 [[TMP4]], 3
+; CHECK-NEXT: [[TMP6:%.*]] = shl i64 [[TMP4]], 4
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
; CHECK: vector.body:
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT: [[TMP6:%.*]] = xor i64 [[INDEX]], -1
-; CHECK-NEXT: [[TMP7:%.*]] = add i64 [[N]], [[TMP6]]
-; CHECK-NEXT: [[TMP8:%.*]] = getelementptr inbounds double, ptr [[B:%.*]], i64 [[TMP7]]
-; CHECK-NEXT: [[TMP9:%.*]] = call i64 @llvm.vscale.i64()
-; CHECK-NEXT: [[TMP10:%.*]] = shl i64 [[TMP9]], 3
-; CHECK-NEXT: [[TMP11:%.*]] = sub i64 1, [[TMP10]]
-; CHECK-NEXT: [[TMP12:%.*]] = getelementptr inbounds double, ptr [[TMP8]], i64 [[TMP11]]
-; CHECK-NEXT: [[TMP13:%.*]] = call i64 @llvm.vscale.i64()
-; CHECK-NEXT: [[TMP14:%.*]] = shl i64 [[TMP13]], 3
-; CHECK-NEXT: [[TMP15:%.*]] = sub i64 0, [[TMP14]]
-; CHECK-NEXT: [[TMP16:%.*]] = sub i64 1, [[TMP14]]
-; CHECK-NEXT: [[TMP17:%.*]] = getelementptr inbounds double, ptr [[TMP8]], i64 [[TMP15]]
-; CHECK-NEXT: [[TMP18:%.*]] = getelementptr inbounds double, ptr [[TMP17]], i64 [[TMP16]]
-; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 8 x double>, ptr [[TMP12]], align 8
-; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = load <vscale x 8 x double>, ptr [[TMP18]], align 8
-; CHECK-NEXT: [[TMP19:%.*]] = fadd <vscale x 8 x double> [[WIDE_LOAD]], shufflevector (<vscale x 8 x double> insertelement (<vscale x 8 x double> poison, double 1.000000e+00, i64 0), <vscale x 8 x double> poison, <vscale x 8 x i32> zeroinitializer)
-; CHECK-NEXT: [[TMP20:%.*]] = fadd <vscale x 8 x double> [[WIDE_LOAD1]], shufflevector (<vscale x 8 x double> insertelement (<vscale x 8 x double> poison, double 1.000000e+00, i64 0), <vscale x 8 x double> poison, <vscale x 8 x i32> zeroinitializer)
-; CHECK-NEXT: [[TMP21:%.*]] = getelementptr inbounds double, ptr [[A:%.*]], i64 [[TMP7]]
-; CHECK-NEXT: [[TMP22:%.*]] = call i64 @llvm.vscale.i64()
-; CHECK-NEXT: [[TMP23:%.*]] = shl i64 [[TMP22]], 3
-; CHECK-NEXT: [[TMP24:%.*]] = sub i64 1, [[TMP23]]
-; CHECK-NEXT: [[TMP25:%.*]] = getelementptr inbounds double, ptr [[TMP21]], i64 [[TMP24]]
-; CHECK-NEXT: [[TMP26:%.*]] = call i64 @llvm.vscale.i64()
-; CHECK-NEXT: [[TMP27:%.*]] = shl i64 [[TMP26]], 3
-; CHECK-NEXT: [[TMP28:%.*]] = sub i64 0, [[TMP27]]
-; CHECK-NEXT: [[TMP29:%.*]] = sub i64 1, [[TMP27]]
-; CHECK-NEXT: [[TMP30:%.*]] = getelementptr inbounds double, ptr [[TMP21]], i64 [[TMP28]]
-; CHECK-NEXT: [[TMP31:%.*]] = getelementptr inbounds double, ptr [[TMP30]], i64 [[TMP29]]
-; CHECK-NEXT: store <vscale x 8 x double> [[TMP19]], ptr [[TMP25]], align 8
-; CHECK-NEXT: store <vscale x 8 x double> [[TMP20]], ptr [[TMP31]], align 8
-; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP5]]
-; CHECK-NEXT: [[TMP3...
[truncated]
|
|
rebase after #110974 is merged. |
Mel-Chen
reviewed
Oct 29, 2024
| @@ -1462,15 +1462,22 @@ static void transformRecipestoEVLRecipes(VPlan &Plan, VPValue &EVL) { | |||
| assert(OrigMask && "Unmasked recipe when folding tail"); | |||
| return HeaderMask == OrigMask ? nullptr : OrigMask; | |||
| }; | |||
| auto SetEVLForReversePointer = [&EVL](VPValue *V) -> void { | |||
| if (auto R = | |||
| dyn_cast<VPReverseVectorPointerRecipe>(V->getDefiningRecipe())) | |||
There was a problem hiding this comment.
Is it possible for V->getDefiningRecipe() to return nullptr?
There was a problem hiding this comment.
Nice catch. I think this can occur when V is loop-invariant. Fixed.
There was a problem hiding this comment.
Using dyn_cast_if_present<VPReverseVectorPointerRecipe>(...) might be better.
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This patch replaces VF with EVL for VPReverseVectorPointer in DataWithEVL Tail-Folding
VPReverseVectorPointer relies on runtime VF, so it’s necessary to replace VF with EVL in DataWithEVL tail-folding, as EVL can be less than VF at runtime. Since VPReverseVectorPointer does not depend on the header mask and is created only for the widen load/store operations, the replacement is applied while the widen load/store recipe is matched.