[SLP][REVEC] Fix CommonMask is transformed into vector form but used outside finalize. #120952

HanKuanChen · 2024-12-23T09:19:51Z

No description provided.

outside finalize.

llvmbot · 2024-12-23T09:20:23Z

@llvm/pr-subscribers-vectorizers

@llvm/pr-subscribers-llvm-transforms

Author: Han-Kuan Chen (HanKuanChen)

Changes

Full diff: https://github.com/llvm/llvm-project/pull/120952.diff

2 Files Affected:

(modified) llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp (+31)
(modified) llvm/test/Transforms/SLPVectorizer/revec-shufflevector.ll (+114)

diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index d967813075bb9f..b0f30abf56ed74 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -311,6 +311,26 @@ static void transformScalarShuffleIndiciesToVector(unsigned VecTyNumElements,
   Mask.swap(NewMask);
 }
 
+static void transformVectorShuffleIndiciesToScalar(unsigned VecTyNumElements,
+                                                   SmallVectorImpl<int> &Mask) {
+  // This is a restored process of transformScalarShuffleIndiciesToVector.
+  SmallVector<int> NewMask(Mask.size() / VecTyNumElements);
+  for (unsigned I : seq<unsigned>(NewMask.size())) {
+    ArrayRef<int> Elements =
+        ArrayRef(Mask).slice(I * VecTyNumElements, VecTyNumElements);
+    if (all_of(Elements, [](int E) { return E == PoisonMaskElem; })) {
+      NewMask[I] = PoisonMaskElem;
+      continue;
+    }
+    assert((Elements[0] % VecTyNumElements == 0) &&
+           equal(Elements,
+                 createSequentialMask(Elements[0], VecTyNumElements, 0)) &&
+           "Not a valid mask from transformScalarShuffleIndiciesToVector.");
+    NewMask[I] = Elements[0] / VecTyNumElements;
+  }
+  Mask.swap(NewMask);
+}
+
 /// \returns the number of groups of shufflevector
 /// A group has the following features
 /// 1. All of value in a group are shufflevector.
@@ -14286,7 +14306,18 @@ class BoUpSLP::ShuffleInstructionBuilder final : public BaseShuffleAnalysis {
         std::iota(ResizeMask.begin(), std::next(ResizeMask.begin(), VecVF), 0);
         Vec = createShuffle(Vec, nullptr, ResizeMask);
       }
+      // We need to transform CommonMask into scalar form because the Action
+      // (TryPackScalars) interpret the mask by the number of elements in
+      // Scalars.
+      if (ScalarTyNumElements != 1) {
+        assert(SLPReVec && "FixedVectorType is not expected.");
+        transformVectorShuffleIndiciesToScalar(ScalarTyNumElements, CommonMask);
+      }
       Action(Vec, CommonMask);
+      if (ScalarTyNumElements != 1) {
+        assert(SLPReVec && "FixedVectorType is not expected.");
+        transformScalarShuffleIndiciesToVector(ScalarTyNumElements, CommonMask);
+      }
       InVectors.front() = Vec;
     }
     if (!SubVectors.empty()) {
diff --git a/llvm/test/Transforms/SLPVectorizer/revec-shufflevector.ll b/llvm/test/Transforms/SLPVectorizer/revec-shufflevector.ll
index a2673d81068d8d..d6c09bc224a7d1 100644
--- a/llvm/test/Transforms/SLPVectorizer/revec-shufflevector.ll
+++ b/llvm/test/Transforms/SLPVectorizer/revec-shufflevector.ll
@@ -121,3 +121,117 @@ entry:
   store <4 x i32> %1, ptr %3, align 4
   ret void
 }
+
+define void @test6(ptr %in0, ptr %in1, ptr %in2) {
+; CHECK-LABEL: @test6(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[GEP1:%.*]] = getelementptr inbounds nuw i8, ptr [[IN0:%.*]], i64 32
+; CHECK-NEXT:    [[LOAD2:%.*]] = load <4 x float>, ptr [[GEP1]], align 16
+; CHECK-NEXT:    [[TMP0:%.*]] = load <8 x float>, ptr [[IN0]], align 16
+; CHECK-NEXT:    [[TMP1:%.*]] = load <32 x i8>, ptr [[IN1:%.*]], align 1
+; CHECK-NEXT:    [[TMP2:%.*]] = uitofp <32 x i8> [[TMP1]] to <32 x float>
+; CHECK-NEXT:    [[TMP14:%.*]] = shufflevector <8 x float> [[TMP0]], <8 x float> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
+; CHECK-NEXT:    [[TMP15:%.*]] = shufflevector <4 x float> [[LOAD2]], <4 x float> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
+; CHECK-NEXT:    [[TMP16:%.*]] = shufflevector <16 x float> [[TMP14]], <16 x float> [[TMP15]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 16, i32 17, i32 18, i32 19, i32 poison, i32 poison, i32 poison, i32 poison>
+; CHECK-NEXT:    [[TMP3:%.*]] = shufflevector <16 x float> [[TMP16]], <16 x float> poison, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+; CHECK-NEXT:    [[TMP4:%.*]] = fmul <32 x float> [[TMP3]], [[TMP2]]
+; CHECK-NEXT:    store <32 x float> [[TMP4]], ptr [[IN2:%.*]], align 16
+; CHECK-NEXT:    [[GEP10:%.*]] = getelementptr inbounds nuw i8, ptr [[IN1]], i64 32
+; CHECK-NEXT:    [[LOAD5:%.*]] = load <16 x i8>, ptr [[GEP10]], align 1
+; CHECK-NEXT:    [[TMP5:%.*]] = shufflevector <8 x float> [[TMP0]], <8 x float> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+; CHECK-NEXT:    [[GEP11:%.*]] = getelementptr inbounds nuw i8, ptr [[IN2]], i64 128
+; CHECK-NEXT:    [[TMP6:%.*]] = uitofp <16 x i8> [[LOAD5]] to <16 x float>
+; CHECK-NEXT:    [[TMP7:%.*]] = shufflevector <4 x float> [[LOAD2]], <4 x float> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
+; CHECK-NEXT:    [[TMP8:%.*]] = shufflevector <8 x float> [[TMP0]], <8 x float> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
+; CHECK-NEXT:    [[TMP9:%.*]] = shufflevector <16 x float> [[TMP7]], <16 x float> [[TMP8]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 16, i32 17, i32 18, i32 19, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
+; CHECK-NEXT:    [[TMP10:%.*]] = shufflevector <4 x float> [[TMP5]], <4 x float> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
+; CHECK-NEXT:    [[TMP11:%.*]] = shufflevector <16 x float> [[TMP9]], <16 x float> [[TMP10]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 16, i32 17, i32 18, i32 19, i32 poison, i32 poison, i32 poison, i32 poison>
+; CHECK-NEXT:    [[TMP12:%.*]] = shufflevector <16 x float> [[TMP11]], <16 x float> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 0, i32 1, i32 2, i32 3>
+; CHECK-NEXT:    [[TMP13:%.*]] = fmul <16 x float> [[TMP12]], [[TMP6]]
+; CHECK-NEXT:    store <16 x float> [[TMP13]], ptr [[GEP11]], align 16
+; CHECK-NEXT:    ret void
+;
+entry:
+  %gep0 = getelementptr inbounds i8, ptr %in0, i64 16
+  %gep1 = getelementptr inbounds i8, ptr %in0, i64 32
+  %load0 = load <4 x float>, ptr %in0, align 16
+  %load1 = load <4 x float>, ptr %gep0, align 16
+  %load2 = load <4 x float>, ptr %gep1, align 16
+  %gep2 = getelementptr inbounds i8, ptr %in1, i64 16
+  %load3 = load <16 x i8>, ptr %in1, align 1
+  %load4 = load <16 x i8>, ptr %gep2, align 1
+  %shufflevector0 = shufflevector <16 x i8> %load3, <16 x i8> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %shufflevector1 = shufflevector <16 x i8> %load3, <16 x i8> poison, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %shufflevector2 = shufflevector <16 x i8> %load4, <16 x i8> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %shufflevector3 = shufflevector <16 x i8> %load4, <16 x i8> poison, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %zext0 = zext <8 x i8> %shufflevector0 to <8 x i16>
+  %zext1 = zext <8 x i8> %shufflevector1 to <8 x i16>
+  %zext2 = zext <8 x i8> %shufflevector2 to <8 x i16>
+  %zext3 = zext <8 x i8> %shufflevector3 to <8 x i16>
+  %shufflevector4 = shufflevector <8 x i16> %zext0, <8 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %shufflevector5 = shufflevector <8 x i16> %zext0, <8 x i16> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %shufflevector6 = shufflevector <8 x i16> %zext1, <8 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %shufflevector7 = shufflevector <8 x i16> %zext1, <8 x i16> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %shufflevector8 = shufflevector <8 x i16> %zext2, <8 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %shufflevector9 = shufflevector <8 x i16> %zext2, <8 x i16> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %shufflevector10 = shufflevector <8 x i16> %zext3, <8 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %shufflevector11 = shufflevector <8 x i16> %zext3, <8 x i16> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %uitofp0 = uitofp nneg <4 x i16> %shufflevector4 to <4 x float>
+  %uitofp1 = uitofp nneg <4 x i16> %shufflevector5 to <4 x float>
+  %uitofp2 = uitofp nneg <4 x i16> %shufflevector6 to <4 x float>
+  %uitofp3 = uitofp nneg <4 x i16> %shufflevector7 to <4 x float>
+  %uitofp4 = uitofp nneg <4 x i16> %shufflevector8 to <4 x float>
+  %uitofp5 = uitofp nneg <4 x i16> %shufflevector9 to <4 x float>
+  %uitofp6 = uitofp nneg <4 x i16> %shufflevector10 to <4 x float>
+  %uitofp7 = uitofp nneg <4 x i16> %shufflevector11 to <4 x float>
+  %fmul0 = fmul <4 x float> %load0, %uitofp0
+  %fmul1 = fmul <4 x float> %load1, %uitofp1
+  %fmul2 = fmul <4 x float> %load2, %uitofp2
+  %fmul3 = fmul <4 x float> %load0, %uitofp3
+  %fmul4 = fmul <4 x float> %load1, %uitofp4
+  %fmul5 = fmul <4 x float> %load2, %uitofp5
+  %fmul6 = fmul <4 x float> %load0, %uitofp6
+  %fmul7 = fmul <4 x float> %load1, %uitofp7
+  %gep3 = getelementptr inbounds i8, ptr %in2, i64 16
+  %gep4 = getelementptr inbounds i8, ptr %in2, i64 32
+  %gep5 = getelementptr inbounds i8, ptr %in2, i64 48
+  %gep6 = getelementptr inbounds i8, ptr %in2, i64 64
+  %gep7 = getelementptr inbounds i8, ptr %in2, i64 80
+  %gep8 = getelementptr inbounds i8, ptr %in2, i64 96
+  %gep9 = getelementptr inbounds i8, ptr %in2, i64 112
+  store <4 x float> %fmul0, ptr %in2, align 16
+  store <4 x float> %fmul1, ptr %gep3, align 16
+  store <4 x float> %fmul2, ptr %gep4, align 16
+  store <4 x float> %fmul3, ptr %gep5, align 16
+  store <4 x float> %fmul4, ptr %gep6, align 16
+  store <4 x float> %fmul5, ptr %gep7, align 16
+  store <4 x float> %fmul6, ptr %gep8, align 16
+  store <4 x float> %fmul7, ptr %gep9, align 16
+  %gep10 = getelementptr inbounds i8, ptr %in1, i64 32
+  %load5 = load <16 x i8>, ptr %gep10, align 1
+  %shufflevector12 = shufflevector <16 x i8> %load5, <16 x i8> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %shufflevector13 = shufflevector <16 x i8> %load5, <16 x i8> poison, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %zext4 = zext <8 x i8> %shufflevector12 to <8 x i16>
+  %zext5 = zext <8 x i8> %shufflevector13 to <8 x i16>
+  %shufflevector14 = shufflevector <8 x i16> %zext4, <8 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %shufflevector15 = shufflevector <8 x i16> %zext4, <8 x i16> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %shufflevector16 = shufflevector <8 x i16> %zext5, <8 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %shufflevector17 = shufflevector <8 x i16> %zext5, <8 x i16> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %uitofp8 = uitofp nneg <4 x i16> %shufflevector14 to <4 x float>
+  %uitofp9 = uitofp nneg <4 x i16> %shufflevector15 to <4 x float>
+  %uitofp10 = uitofp nneg <4 x i16> %shufflevector16 to <4 x float>
+  %uitofp11 = uitofp nneg <4 x i16> %shufflevector17 to <4 x float>
+  %fmul8 = fmul <4 x float> %load2, %uitofp8
+  %fmul9 = fmul <4 x float> %load0, %uitofp9
+  %fmul10 = fmul <4 x float> %load1, %uitofp10
+  %fmul11 = fmul <4 x float> %load2, %uitofp11
+  %gep11 = getelementptr inbounds i8, ptr %in2, i64 128
+  %gep12 = getelementptr inbounds i8, ptr %in2, i64 144
+  %gep13 = getelementptr inbounds i8, ptr %in2, i64 160
+  %gep14 = getelementptr inbounds i8, ptr %in2, i64 176
+  store <4 x float> %fmul8, ptr %gep11, align 16
+  store <4 x float> %fmul9, ptr %gep12, align 16
+  store <4 x float> %fmul10, ptr %gep13, align 16
+  store <4 x float> %fmul11, ptr %gep14, align 16
+  ret void
+}

alexey-bataev

I don't like this approach. We'd better to operate with actual-sized masks/shuffles, otherwise at least cost is not estimated correctly

HanKuanChen · 2024-12-24T06:24:35Z

What do you mean "actual-sized"? Determined by unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements();
If that is true, it will modify a lot of code. Especially we don't have test to make sure it is true.

A code like unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements(); should be divided by the number of elements of ScalarTy. Just like getVF. e.g.,

    unsigned ScalarTyNumElements = getNumElements(ScalarTy);
    unsigned VNumElements =
        cast<FixedVectorType>(V->getType())->getNumElements();
    assert(VNumElements > ScalarTyNumElements &&
           "the number of elements of V is not large enough");
    assert(VNumElements % ScalarTyNumElements == 0 &&
           "the number of elements of V is not a vectorized value");
    return VNumElements / ScalarTyNumElements;

alexey-bataev · 2024-12-24T15:48:20Z

What do you mean "actual-sized"? Determined by unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements(); If that is true, it will modify a lot of code. Especially we don't have test to make sure it is true.

It is not a problem. The main problem - to make the code completely correct

A code like unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements(); should be divided by the number of elements of ScalarTy. Just like getVF. e.g.,

    unsigned ScalarTyNumElements = getNumElements(ScalarTy);
    unsigned VNumElements =
        cast<FixedVectorType>(V->getType())->getNumElements();
    assert(VNumElements > ScalarTyNumElements &&
           "the number of elements of V is not large enough");
    assert(VNumElements % ScalarTyNumElements == 0 &&
           "the number of elements of V is not a vectorized value");
    return VNumElements / ScalarTyNumElements;

No, it should not be. You should operate with the full masks from the very beginning

HanKuanChen · 2024-12-25T09:37:31Z

Take the following code as an example.

    if (Action) {
      Value *Vec = InVectors.front();
      if (InVectors.size() == 2) {
        Vec = createShuffle(Vec, InVectors.back(), CommonMask);
        InVectors.pop_back();
      } else {
        Vec = createShuffle(Vec, nullptr, CommonMask);
      }
      for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx)
        if (CommonMask[Idx] != PoisonMaskElem)
          CommonMask[Idx] = Idx;
      assert(VF > 0 &&
             "Expected vector length for the final value before action.");
      unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements();
      if (VecVF < VF) {
        SmallVector<int> ResizeMask(VF, PoisonMaskElem);
        std::iota(ResizeMask.begin(), std::next(ResizeMask.begin(), VecVF), 0);
        Vec = createShuffle(Vec, nullptr, ResizeMask);
      }
      Action(Vec, CommonMask);
      InVectors.front() = Vec;
    }

How do you know <4 x Ty> from unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements(); is two <2 x Ty> (so VF is 2) instead of four Ty (so VF is 4)?

alexey-bataev · 2024-12-26T16:34:54Z

Take the following code as an example.

    if (Action) {
      Value *Vec = InVectors.front();
      if (InVectors.size() == 2) {
        Vec = createShuffle(Vec, InVectors.back(), CommonMask);
        InVectors.pop_back();
      } else {
        Vec = createShuffle(Vec, nullptr, CommonMask);
      }
      for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx)
        if (CommonMask[Idx] != PoisonMaskElem)
          CommonMask[Idx] = Idx;
      assert(VF > 0 &&
             "Expected vector length for the final value before action.");
      unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements();
      if (VecVF < VF) {
        SmallVector<int> ResizeMask(VF, PoisonMaskElem);
        std::iota(ResizeMask.begin(), std::next(ResizeMask.begin(), VecVF), 0);
        Vec = createShuffle(Vec, nullptr, ResizeMask);
      }
      Action(Vec, CommonMask);
      InVectors.front() = Vec;
    }

How do you know <4 x Ty> from unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements(); is two <2 x Ty> (so VF is 2) instead of four Ty (so VF is 4)?

It should be considered as four Ty, we should operate on actual types here

HanKuanChen · 2024-12-26T17:42:47Z

Take the following code as an example.

    if (Action) {
      Value *Vec = InVectors.front();
      if (InVectors.size() == 2) {
        Vec = createShuffle(Vec, InVectors.back(), CommonMask);
        InVectors.pop_back();
      } else {
        Vec = createShuffle(Vec, nullptr, CommonMask);
      }
      for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx)
        if (CommonMask[Idx] != PoisonMaskElem)
          CommonMask[Idx] = Idx;
      assert(VF > 0 &&
             "Expected vector length for the final value before action.");
      unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements();
      if (VecVF < VF) {
        SmallVector<int> ResizeMask(VF, PoisonMaskElem);
        std::iota(ResizeMask.begin(), std::next(ResizeMask.begin(), VecVF), 0);
        Vec = createShuffle(Vec, nullptr, ResizeMask);
      }
      Action(Vec, CommonMask);
      InVectors.front() = Vec;
    }

How do you know <4 x Ty> from unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements(); is two <2 x Ty> (so VF is 2) instead of four Ty (so VF is 4)?

It should be considered as four Ty, we should operate on actual types here

If we have VL with two <2 x Ty>.

      Res = ShuffleBuilder.finalize(
          E->ReuseShuffleIndices, SubVectors, SubVectorsMask, E->Scalars.size(),
          [&](Value *&Vec, SmallVectorImpl<int> &Mask) {
            TryPackScalars(NonConstants, Mask, /*IsRootPoison=*/false);
            Vec = ShuffleBuilder.gather(NonConstants, Mask.size(), Vec);
          });

The VF it expects is 2. But unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements(); here is 4? And we compare VecVF (4) with the incoming VF (2)?

alexey-bataev · 2024-12-26T17:50:25Z

Take the following code as an example.

    if (Action) {
      Value *Vec = InVectors.front();
      if (InVectors.size() == 2) {
        Vec = createShuffle(Vec, InVectors.back(), CommonMask);
        InVectors.pop_back();
      } else {
        Vec = createShuffle(Vec, nullptr, CommonMask);
      }
      for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx)
        if (CommonMask[Idx] != PoisonMaskElem)
          CommonMask[Idx] = Idx;
      assert(VF > 0 &&
             "Expected vector length for the final value before action.");
      unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements();
      if (VecVF < VF) {
        SmallVector<int> ResizeMask(VF, PoisonMaskElem);
        std::iota(ResizeMask.begin(), std::next(ResizeMask.begin(), VecVF), 0);
        Vec = createShuffle(Vec, nullptr, ResizeMask);
      }
      Action(Vec, CommonMask);
      InVectors.front() = Vec;
    }

How do you know <4 x Ty> from unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements(); is two <2 x Ty> (so VF is 2) instead of four Ty (so VF is 4)?

It should be considered as four Ty, we should operate on actual types here

If we have VL with two <2 x Ty>.

      Res = ShuffleBuilder.finalize(
          E->ReuseShuffleIndices, SubVectors, SubVectorsMask, E->Scalars.size(),
          [&](Value *&Vec, SmallVectorImpl<int> &Mask) {
            TryPackScalars(NonConstants, Mask, /*IsRootPoison=*/false);
            Vec = ShuffleBuilder.gather(NonConstants, Mask.size(), Vec);
          });

The VF it expects is 2. But unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements(); here is 4? And we compare VecVF (4) with the incoming VF (2)?

The VF here should be expanded to 4 too

HanKuanChen · 2024-12-27T07:34:48Z

The current VF definition is based on the number of VL instead of the number of elements. If we modify the VF definition to the number of elements, it will require modifying a lot of code, and I don't see any clear benefits here. For example, what is the VF you expect when we vectorize FixedVectorType in the following code?

  unsigned Sz = R.getVectorElementSize(I0);
  unsigned MinVF = R.getMinVF(Sz);
  unsigned MaxVF = std::max<unsigned>(llvm::bit_floor(VL.size()), MinVF);
  MaxVF = std::min(R.getMaximumVF(Sz, S.getOpcode()), MaxVF);

I don't think partially modifying the definition in the finalize is a good idea either.

ShuffleCostEstimator::createShuffle can handle FixedVectorType now. It expands CommonMask into a correct mask before calling BaseShuffleAnalysis::createShuffle. We just need to make GetNodeMinBWAffectedCost and GetValueMinBWAffectedCost support FixedVectorType.

ShuffleInstructionBuilder::finalize can handle FixedVectorType now. Any development that modifies the mask can be easily identified and may cause a crash. At that point, we can quickly fix the bug.

I think the real solution is to wrap the mask into a class. This class can distinguish whether we are vectorizing a scalar to a vector or a vector to a vector. Additionally, we can add a cost model and an instruction builder within the class.

alexey-bataev · 2024-12-30T15:07:21Z

The current VF definition is based on the number of VL instead of the number of elements. If we modify the VF definition to the number of elements, it will require modifying a lot of code, and I don't see any clear benefits here.

Modifying 'a lot of code' is not a problem here, correct cost estimation and vectorization factor is the actual problem. The benefit is that all possible combinations are processed in a uniform way, which correctly allows to estimate cost and emit code.

For example, what is the VF you expect when we vectorize FixedVectorType in the following code?
  unsigned Sz = R.getVectorElementSize(I0);
  unsigned MinVF = R.getMinVF(Sz);
  unsigned MaxVF = std::max<unsigned>(llvm::bit_floor(VL.size()), MinVF);
  MaxVF = std::min(R.getMaximumVF(Sz, S.getOpcode()), MaxVF);
I don't think partially modifying the definition in the finalize is a good idea either.

Maybe, anyway you should operate on a full masks here. The user should not know if it operates on scalars or vectors, it just shall operate on the whole masks. Who expands these masks, the finalize function or other functions, is another question

ShuffleCostEstimator::createShuffle can handle FixedVectorType now. It expands CommonMask into a correct mask before calling BaseShuffleAnalysis::createShuffle. We just need to make GetNodeMinBWAffectedCost and GetValueMinBWAffectedCost support FixedVectorType.

ShuffleInstructionBuilder::finalize can handle FixedVectorType now. Any development that modifies the mask can be easily identified and may cause a crash. At that point, we can quickly fix the bug.

I think the real solution is to wrap the mask into a class. This class can distinguish whether we are vectorizing a scalar to a vector or a vector to a vector. Additionally, we can add a cost model and an instruction builder within the class.

Bad decision. As I said, these classes should not know anything about high level abstraction, they should operate on actual masks/data

HanKuanChen · 2024-12-31T09:08:42Z

ShuffleCostEstimator::createShuffle can handle FixedVectorType now. It expands CommonMask into a correct mask before calling BaseShuffleAnalysis::createShuffle. We just need to make GetNodeMinBWAffectedCost and GetValueMinBWAffectedCost support FixedVectorType.
ShuffleInstructionBuilder::finalize can handle FixedVectorType now. Any development that modifies the mask can be easily identified and may cause a crash. At that point, we can quickly fix the bug.
I think the real solution is to wrap the mask into a class. This class can distinguish whether we are vectorizing a scalar to a vector or a vector to a vector. Additionally, we can add a cost model and an instruction builder within the class.

Bad decision. As I said, these classes should not know anything about high level abstraction, they should operate on actual masks/data

I think we have some misunderstanding here. They already operate the actual mask.

alexey-bataev · 2025-01-03T13:50:53Z

ShuffleBuilders should operate on full masks, that's by the design. Otherwise, it is a potential source of bugs and further code complications

HanKuanChen · 2025-01-28T04:07:01Z

#124244 fixes the issue.

HanKuanChen added 2 commits December 23, 2024 01:17

[SLP][REVEC] Pre-commit test.

115f3f3

[SLP][REVEC] Fix CommonMask is transformed into vector form but used

7abc885

outside finalize.

HanKuanChen requested a review from alexey-bataev December 23, 2024 09:19

llvmbot added vectorizers llvm:transforms labels Dec 23, 2024

alexey-bataev reviewed Dec 23, 2024

View reviewed changes

HanKuanChen closed this Jan 28, 2025

HanKuanChen deleted the slp-revec-finalize-action branch January 28, 2025 04:08

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[SLP][REVEC] Fix CommonMask is transformed into vector form but used outside finalize. #120952

[SLP][REVEC] Fix CommonMask is transformed into vector form but used outside finalize. #120952

Uh oh!

HanKuanChen commented Dec 23, 2024

Uh oh!

llvmbot commented Dec 23, 2024 •

edited

Loading

Uh oh!

alexey-bataev left a comment

Uh oh!

HanKuanChen commented Dec 24, 2024

Uh oh!

alexey-bataev commented Dec 24, 2024

Uh oh!

HanKuanChen commented Dec 25, 2024

Uh oh!

alexey-bataev commented Dec 26, 2024

Uh oh!

HanKuanChen commented Dec 26, 2024

Uh oh!

alexey-bataev commented Dec 26, 2024

Uh oh!

HanKuanChen commented Dec 27, 2024

Uh oh!

alexey-bataev commented Dec 30, 2024

Uh oh!

HanKuanChen commented Dec 31, 2024

Uh oh!

alexey-bataev commented Jan 3, 2025

Uh oh!

HanKuanChen commented Jan 28, 2025

Uh oh!

Uh oh!

[SLP][REVEC] Fix CommonMask is transformed into vector form but used outside finalize. #120952

[SLP][REVEC] Fix CommonMask is transformed into vector form but used outside finalize. #120952

Uh oh!

Conversation

HanKuanChen commented Dec 23, 2024

Uh oh!

llvmbot commented Dec 23, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

alexey-bataev left a comment

Choose a reason for hiding this comment

Uh oh!

HanKuanChen commented Dec 24, 2024

Uh oh!

alexey-bataev commented Dec 24, 2024

Uh oh!

HanKuanChen commented Dec 25, 2024

Uh oh!

alexey-bataev commented Dec 26, 2024

Uh oh!

HanKuanChen commented Dec 26, 2024

Uh oh!

alexey-bataev commented Dec 26, 2024

Uh oh!

HanKuanChen commented Dec 27, 2024

Uh oh!

alexey-bataev commented Dec 30, 2024

Uh oh!

HanKuanChen commented Dec 31, 2024

Uh oh!

alexey-bataev commented Jan 3, 2025

Uh oh!

HanKuanChen commented Jan 28, 2025

Uh oh!

Uh oh!

llvmbot commented Dec 23, 2024 •

edited

Loading