[SLP][REVEC] getScalarizationOverhead should not be used when ScalarTy is FixedVectorType. (#117536)

Author: HanKuanChen

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

@@ -9614,8 +9614,20 @@ void BoUpSLP::reorderGatherNode(TreeEntry &TE) {
     Cost += ::getShuffleCost(*TTI, TTI::SK_InsertSubvector, VecTy, {}, CostKind,
                              Idx, getWidenedType(ScalarTy, Sz));
   }
-  Cost += TTI->getScalarizationOverhead(VecTy, DemandedElts, /*Insert=*/true,
-                                        /*Extract=*/false, CostKind);
+  if (auto *FTy = dyn_cast<FixedVectorType>(ScalarTy)) {
+    assert(SLPReVec && "Only supported by REVEC.");
+    // If ScalarTy is FixedVectorType, we should use CreateInsertVector instead
+    // of CreateInsertElement.
+    unsigned ScalarTyNumElements = getNumElements(ScalarTy);
+    for (unsigned I : seq<unsigned>(TE.Scalars.size()))
+      if (DemandedElts[I])
+        Cost +=
+            TTI->getShuffleCost(TTI::SK_InsertSubvector, VecTy, std::nullopt,
+                                CostKind, I * ScalarTyNumElements, FTy);
+  } else {
+    Cost += TTI->getScalarizationOverhead(VecTy, DemandedElts, /*Insert=*/true,
+                                          /*Extract=*/false, CostKind);
+  }
   int Sz = TE.Scalars.size();
   SmallVector<int> ReorderMask(TE.ReorderIndices.begin(),
                                TE.ReorderIndices.end());

llvm/test/Transforms/SLPVectorizer/SystemZ/revec-fix-117393.ll

@@ -0,0 +1,30 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -mtriple=systemz-unknown -mcpu=z15 -passes=slp-vectorizer -S -slp-revec %s | FileCheck %s
+
+define void @h() {
+; CHECK-LABEL: @h(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = shl <4 x i32> zeroinitializer, zeroinitializer
+; CHECK-NEXT:    [[TMP1:%.*]] = or <4 x i32> [[TMP0]], zeroinitializer
+; CHECK-NEXT:    [[TMP2:%.*]] = or <4 x i32> splat (i32 1), zeroinitializer
+; CHECK-NEXT:    [[TMP3:%.*]] = shl <4 x i32> zeroinitializer, zeroinitializer
+; CHECK-NEXT:    [[TMP4:%.*]] = or <4 x i32> [[TMP3]], zeroinitializer
+; CHECK-NEXT:    [[TMP5:%.*]] = and <4 x i32> [[TMP2]], [[TMP1]]
+; CHECK-NEXT:    [[TMP6:%.*]] = and <4 x i32> zeroinitializer, [[TMP5]]
+; CHECK-NEXT:    [[TMP7:%.*]] = and <4 x i32> [[TMP4]], [[TMP6]]
+; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @llvm.vector.reduce.and.v4i32(<4 x i32> [[TMP7]])
+; CHECK-NEXT:    ret void
+;
+entry:
+  %0 = shl <4 x i32> zeroinitializer, zeroinitializer
+  %1 = or <4 x i32> %0, zeroinitializer
+  %2 = or <4 x i32> splat (i32 1), zeroinitializer
+  %3 = or <4 x i32> zeroinitializer, zeroinitializer
+  %4 = shl <4 x i32> zeroinitializer, zeroinitializer
+  %5 = or <4 x i32> %4, zeroinitializer
+  %6 = and <4 x i32> %2, %1
+  %7 = and <4 x i32> %3, %6
+  %8 = and <4 x i32> %5, %7
+  %9 = call i32 @llvm.vector.reduce.and.v4i32(<4 x i32> %8)
+  ret void
+}