[llvm] [SLP][REVEC] Make ShuffleCostEstimator and ShuffleInstructionBuilder can vectorize vector instructions. (PR #99606)
Han-Kuan Chen via llvm-commits
llvm-commits at lists.llvm.org
Tue Jul 23 02:26:51 PDT 2024
https://github.com/HanKuanChen updated https://github.com/llvm/llvm-project/pull/99606
>From c467b1bd407de4ee6e7d81d61a67ad2a991f4cb7 Mon Sep 17 00:00:00 2001
From: Han-Kuan Chen <hankuan.chen at sifive.com>
Date: Fri, 28 Jun 2024 01:19:03 -0700
Subject: [PATCH 1/4] [SLP][REVEC] NFC. Add
transformScalarShuffleIndiciesToVector.
---
llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp | 15 +++++++++++++++
1 file changed, 15 insertions(+)
diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index d8c3bae06e932..ec3ba6f9ef6a1 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -253,6 +253,21 @@ static FixedVectorType *getWidenedType(Type *ScalarTy, unsigned VF) {
VF * getNumElements(ScalarTy));
}
+static void transformScalarShuffleIndiciesToVector(unsigned VecTyNumElements,
+ SmallVectorImpl<int> &Mask) {
+ // The ShuffleBuilder implementation use shufflevector to splat an "element".
+ // But the element have different meaning for SLP (scalar) and REVEC
+ // (vector). We need to expand Mask into masks which shufflevector can use
+ // directly.
+ SmallVector<int> NewMask(Mask.size() * VecTyNumElements);
+ for (unsigned I : seq<unsigned>(Mask.size()))
+ for (auto [J, MaskV] : enumerate(MutableArrayRef(NewMask).slice(
+ I * VecTyNumElements, VecTyNumElements)))
+ MaskV = Mask[I] == PoisonMaskElem ? PoisonMaskElem
+ : Mask[I] * VecTyNumElements + J;
+ Mask.swap(NewMask);
+}
+
/// \returns True if the value is a constant (but not globals/constant
/// expressions).
static bool isConstant(Value *V) {
>From 78b5dc033cea579261422186fddbb66afee9f622 Mon Sep 17 00:00:00 2001
From: Han-Kuan Chen <hankuan.chen at sifive.com>
Date: Fri, 28 Jun 2024 01:22:42 -0700
Subject: [PATCH 2/4] [SLP][REVEC] Make ShuffleInstructionBuilder::finalize
support vector instructions.
---
llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp | 8 ++++++++
1 file changed, 8 insertions(+)
diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index ec3ba6f9ef6a1..0d5b3bdffcf20 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -12160,6 +12160,14 @@ class BoUpSLP::ShuffleInstructionBuilder final : public BaseShuffleAnalysis {
finalize(ArrayRef<int> ExtMask, unsigned VF = 0,
function_ref<void(Value *&, SmallVectorImpl<int> &)> Action = {}) {
IsFinalized = true;
+ SmallVector<int> NewExtMask(ExtMask);
+ if (auto *VecTy = dyn_cast<FixedVectorType>(ScalarTy)) {
+ transformScalarShuffleIndiciesToVector(VecTy->getNumElements(),
+ CommonMask);
+ transformScalarShuffleIndiciesToVector(VecTy->getNumElements(),
+ NewExtMask);
+ ExtMask = NewExtMask;
+ }
if (Action) {
Value *Vec = InVectors.front();
if (InVectors.size() == 2) {
>From e79c60901231d6e519435d7ba59947140184844c Mon Sep 17 00:00:00 2001
From: Han-Kuan Chen <hankuan.chen at sifive.com>
Date: Mon, 1 Jul 2024 11:46:06 -0700
Subject: [PATCH 3/4] [SLP][REVEC] Make ShuffleCostEstimator::add and
ShuffleInstructionBuilder::add support vector instructions.
---
llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index 0d5b3bdffcf20..ce3c142d8f3a6 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -9113,6 +9113,9 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
assert(!InVectors.empty() && !CommonMask.empty() &&
"Expected only tree entries from extracts/reused buildvectors.");
unsigned VF = cast<FixedVectorType>(V1->getType())->getNumElements();
+ if (auto *VecTy = dyn_cast<FixedVectorType>(ScalarTy))
+ // If ScalarTy is <4 x Ty> and V1 is <8 x Ty>, the VF is 2 instead of 8.
+ VF /= VecTy->getNumElements();
if (InVectors.size() == 2) {
Cost += createShuffle(InVectors.front(), InVectors.back(), CommonMask);
transformMaskAfterShuffle(CommonMask, CommonMask);
@@ -12138,6 +12141,8 @@ class BoUpSLP::ShuffleInstructionBuilder final : public BaseShuffleAnalysis {
int VF = CommonMask.size();
if (auto *FTy = dyn_cast<FixedVectorType>(V1->getType()))
VF = FTy->getNumElements();
+ if (auto *VecTy = dyn_cast<FixedVectorType>(ScalarTy))
+ VF /= VecTy->getNumElements();
for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx)
if (Mask[Idx] != PoisonMaskElem && CommonMask[Idx] == PoisonMaskElem)
CommonMask[Idx] = Mask[Idx] + (It == InVectors.begin() ? 0 : VF);
>From cf115555ef94ad84f42f847888a12963a59c8395 Mon Sep 17 00:00:00 2001
From: Han-Kuan Chen <hankuan.chen at sifive.com>
Date: Mon, 1 Jul 2024 10:44:57 -0700
Subject: [PATCH 4/4] [SLP][REVEC] Make ShuffleCostEstimator::createShuffle
support vector instructions.
The VF is relative to the number of elements in ScalarTy instead of the
size of mask.
---
.../Transforms/Vectorize/SLPVectorizer.cpp | 19 +++++++++++++------
1 file changed, 13 insertions(+), 6 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index ce3c142d8f3a6..edaeb6181ba02 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -8703,6 +8703,7 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
Value *V1 = P1.dyn_cast<Value *>(), *V2 = P2.dyn_cast<Value *>();
unsigned CommonVF = Mask.size();
InstructionCost ExtraCost = 0;
+ unsigned ScalarTyNumElements = getNumElements(ScalarTy);
auto GetNodeMinBWAffectedCost = [&](const TreeEntry &E,
unsigned VF) -> InstructionCost {
if (E.isGather() && allConstant(E.Scalars))
@@ -8743,6 +8744,10 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
}
return TTI::TCC_Free;
};
+ auto getVF = [&](Value *V) {
+ return cast<FixedVectorType>(V->getType())->getNumElements() /
+ ScalarTyNumElements;
+ };
if (!V1 && !V2 && !P2.isNull()) {
// Shuffle 2 entry nodes.
const TreeEntry *E = P1.get<const TreeEntry *>();
@@ -8814,14 +8819,14 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
} else if (V1 && P2.isNull()) {
// Shuffle single vector.
ExtraCost += GetValueMinBWAffectedCost(V1);
- CommonVF = cast<FixedVectorType>(V1->getType())->getNumElements();
+ CommonVF = getVF(V1);
assert(
all_of(Mask,
[=](int Idx) { return Idx < static_cast<int>(CommonVF); }) &&
"All elements in mask must be less than CommonVF.");
} else if (V1 && !V2) {
// Shuffle vector and tree node.
- unsigned VF = cast<FixedVectorType>(V1->getType())->getNumElements();
+ unsigned VF = getVF(V1);
const TreeEntry *E2 = P2.get<const TreeEntry *>();
CommonVF = std::max(VF, E2->getVectorFactor());
assert(all_of(Mask,
@@ -8847,7 +8852,7 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
V2 = getAllOnesValue(*R.DL, getWidenedType(ScalarTy, CommonVF));
} else if (!V1 && V2) {
// Shuffle vector and tree node.
- unsigned VF = cast<FixedVectorType>(V2->getType())->getNumElements();
+ unsigned VF = getVF(V2);
const TreeEntry *E1 = P1.get<const TreeEntry *>();
CommonVF = std::max(VF, E1->getVectorFactor());
assert(all_of(Mask,
@@ -8875,9 +8880,8 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
V2 = getAllOnesValue(*R.DL, getWidenedType(ScalarTy, CommonVF));
} else {
assert(V1 && V2 && "Expected both vectors.");
- unsigned VF = cast<FixedVectorType>(V1->getType())->getNumElements();
- CommonVF =
- std::max(VF, cast<FixedVectorType>(V2->getType())->getNumElements());
+ unsigned VF = getVF(V1);
+ CommonVF = std::max(VF, getVF(V2));
assert(all_of(Mask,
[=](int Idx) {
return Idx < 2 * static_cast<int>(CommonVF);
@@ -8895,6 +8899,9 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
V2 = getAllOnesValue(*R.DL, getWidenedType(ScalarTy, CommonVF));
}
}
+ if (auto *VecTy = dyn_cast<FixedVectorType>(ScalarTy))
+ transformScalarShuffleIndiciesToVector(VecTy->getNumElements(),
+ CommonMask);
InVectors.front() =
Constant::getNullValue(getWidenedType(ScalarTy, CommonMask.size()));
if (InVectors.size() == 2)
More information about the llvm-commits
mailing list