[SLP] Compute a shuffle mask for SK_InsertSubvector

[preames]

This is the third of a series of small patches to compute shuffle masks for the couple of cases where we call getShuffleCost without one. My goal is to add an invariant that all calls to getShuffleCost for fixed length vectors have a mask.

After this change, there is one SK_InsertSubvector case left. I excluded it from this patch just because I thought it worthy of individual attention and review.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Philip Reames (preames)

Changes

This is the third of a series of small patches to compute shuffle masks for the couple of cases where we call getShuffleCost without one. My goal is to add an invariant that all calls to getShuffleCost for fixed length vectors have a mask.

After this change, there is one SK_InsertSubvector case left. I excluded it from this patch just because I thought it worthy of individual attention and review.

---

Full diff: https://github.com/llvm/llvm-project/pull/85408.diff

1 Files Affected:

• (modified) llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp (+11-4)

diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index b4cce680e2876f..b6868fb3f3ca3f 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -4328,9 +4328,12 @@ BoUpSLP::LoadsState BoUpSLP::canVectorizeLoads(
               llvm_unreachable(
                   "Expected only consecutive, strided or masked gather loads.");
             }
+            SmallVector<int> ShuffleMask(VL.size());
+            for (int i = 0; i < VL.size(); i++)
+              ShuffleMask[i] = i / VF == I ? VL.size() + i % VF : i;
             VecLdCost +=
                 TTI.getShuffleCost(TTI ::SK_InsertSubvector, VecTy,
-                                   std::nullopt, CostKind, I * VF, SubVecTy);
+                                   ShuffleMask, CostKind, I * VF, SubVecTy);
           }
           // If masked gather cost is higher - better to vectorize, so
           // consider it as a gather node. It will be better estimated
@@ -7454,7 +7457,7 @@ getShuffleCost(const TargetTransformInfo &TTI, TTI::ShuffleKind Kind,
         Index + NumSrcElts <= static_cast<int>(Mask.size()))
       return TTI.getShuffleCost(
           TTI::SK_InsertSubvector,
-          FixedVectorType::get(Tp->getElementType(), Mask.size()), std::nullopt,
+          FixedVectorType::get(Tp->getElementType(), Mask.size()), Mask,
           TTI::TCK_RecipThroughput, Index, Tp);
   }
   return TTI.getShuffleCost(Kind, Tp, Mask, CostKind, Index, SubTp, Args);
@@ -7727,9 +7730,13 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
         }
         if (NeedInsertSubvectorAnalysis) {
           // Add the cost for the subvectors insert.
-          for (int I = VF, E = VL.size(); I < E; I += VF)
+          SmallVector<int> ShuffleMask(VL.size());
+          for (int I = VF, E = VL.size(); I < E; I += VF) {
+            for (int i = 0; i < E; i++)
+              ShuffleMask[i] = i / VF == I ? E + i % VF : i;
             GatherCost += TTI.getShuffleCost(TTI::SK_InsertSubvector, VecTy,
-                                             std::nullopt, CostKind, I, LoadTy);
+                                             ShuffleMask, CostKind, I, LoadTy);
+          }
         }
         GatherCost -= ScalarsCost;
       }

[github-actions]

⚠️ C/C++ code formatter, clang-format found issues in your code. ⚠️

You can test this locally with the following command:

git-clang-format --diff 33960c90258ed78b9b877b1a43e219d1cbc2efce fdb6519fed34446194bec7d23155248302aed6ad -- llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

View the diff from clang-format here.

diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index d850c0ad78..50af772e30 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -4332,8 +4332,8 @@ BoUpSLP::LoadsState BoUpSLP::canVectorizeLoads(
             for (int Idx : seq<int>(0, VL.size()))
               ShuffleMask[Idx] = Idx / VF == I ? VL.size() + Idx % VF : Idx;
             VecLdCost +=
-                TTI.getShuffleCost(TTI ::SK_InsertSubvector, VecTy,
-                                   ShuffleMask, CostKind, I * VF, SubVecTy);
+                TTI.getShuffleCost(TTI ::SK_InsertSubvector, VecTy, ShuffleMask,
+                                   CostKind, I * VF, SubVecTy);
           }
           // If masked gather cost is higher - better to vectorize, so
           // consider it as a gather node. It will be better estimated

[alexey-bataev]

All is must be replaced too

[alexey-bataev]

LG