s390x: widening multiplication does not optimize

[folkertdev]

this LLVM

https://godbolt.org/z/cx8adPc9f

define range(i32 0, -131070) <4 x i32> @manual_mule(<8 x i16> %a, <8 x i16> %b) unnamed_addr {
start:
  %0 = shufflevector <8 x i16> %a, <8 x i16> poison, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
  %1 = zext <4 x i16> %0 to <4 x i32>
  %2 = shufflevector <8 x i16> %b, <8 x i16> poison, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
  %3 = zext <4 x i16> %2 to <4 x i32>
  %4 = mul nuw <4 x i32> %3, %1
  ret <4 x i32> %4
}

does not optimize to the expected output of vec_mule, a single vmleh instruction. The same is true for the other multiplication flavors (low, high, odd).

[folkertdev]

by extension I think the widening multiplication should also merge into subsequent additions, so that vec_mule(a, b) + c is equal to vec_meadd(a, b, c). That doesn't seem to be happening today either.

https://godbolt.org/z/n5cvaqfb5