[llvm] [VectorCombine] Fix the type used in foldShuffleOfIntrinsics Cost. (PR #138419)

[Han-Kuan Chen via llvm-commits]

================
@@ -84,6 +98,27 @@ entry:
   ret <8 x float> %6
 }
 
+define <2 x float> @test6(<4 x float> %a1, <4 x float> %b1, <4 x float> %c1, <4 x float> %a2, <4 x float> %b2, <4 x float> %c2) {
+; SSE-LABEL: @test6(
+; SSE-NEXT:    [[TMP1:%.*]] = shufflevector <4 x float> [[A1:%.*]], <4 x float> [[A2:%.*]], <2 x i32> <i32 0, i32 4>
+; SSE-NEXT:    [[TMP2:%.*]] = shufflevector <4 x float> [[B1:%.*]], <4 x float> [[B2:%.*]], <2 x i32> <i32 0, i32 4>
+; SSE-NEXT:    [[TMP3:%.*]] = shufflevector <4 x float> [[C1:%.*]], <4 x float> [[C2:%.*]], <2 x i32> <i32 0, i32 4>
+; SSE-NEXT:    [[S:%.*]] = call <2 x float> @llvm.fma.v2f32(<2 x float> [[TMP1]], <2 x float> [[TMP2]], <2 x float> [[TMP3]])
+; SSE-NEXT:    ret <2 x float> [[S]]
+;
+; AVX-LABEL: @test6(
+; AVX-NEXT:    [[F1:%.*]] = call <4 x float> @llvm.fma.v4f32(<4 x float> [[A1:%.*]], <4 x float> [[B1:%.*]], <4 x float> [[C1:%.*]])
+; AVX-NEXT:    [[F2:%.*]] = call <4 x float> @llvm.fma.v4f32(<4 x float> [[A2:%.*]], <4 x float> [[B2:%.*]], <4 x float> [[C2:%.*]])
+; AVX-NEXT:    [[S:%.*]] = shufflevector <4 x float> [[F1]], <4 x float> [[F2]], <2 x i32> <i32 0, i32 4>
+; AVX-NEXT:    ret <2 x float> [[S]]
+;
+  %f1 = call <4 x float> @llvm.fma.v4f32(<4 x float> %a1, <4 x float> %b1, <4 x float> %c1)
+  %f2 = call <4 x float> @llvm.fma.v4f32(<4 x float> %a2, <4 x float> %b2, <4 x float> %c2)
+  %s = shufflevector <4 x float> %f1, <4 x float> %f2, <2 x i32> <i32 0, i32 4>
+  ret <2 x float> %s
+}
+
+
----------------
HanKuanChen wrote:

remove 1 empty line

https://github.com/llvm/llvm-project/pull/138419