s390x: missed optimization to vec_unpackl

[folkertdev]

https://godbolt.org/z/Wxc8x8Tax

This LLVM IR

define range(i32 -32768, 32768) <4 x i32> @unpackh(<8 x i16> %a) unnamed_addr {
start:
  %0 = shufflevector <8 x i16> %a, <8 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
  %1 = sext <4 x i16> %0 to <4 x i32>
  ret <4 x i32> %1
}

define range(i32 -32768, 32768) <4 x i32> @unpackl(<8 x i16> %a) unnamed_addr {
start:
  %0 = shufflevector <8 x i16> %a, <8 x i16> poison, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
  %1 = sext <4 x i16> %0 to <4 x i32>
  ret <4 x i32> %1
}

optimizes to

unpackh:
        vuphh   %v24, %v24
        br      %r14

unpackl:
        vmrlg   %v0, %v24, %v24
        vuphh   %v24, %v0
        br      %r14

this is already very good, and optimal for unpackh, but not for unpackl:

https://godbolt.org/z/xfobea8ee

vector signed int foo(vector signed short a) { 
    return vec_unpackh(a);
}

vector signed int unpack(vector signed short a) { 
    return vec_unpackl(a);
}

optimizes to

foo:
        vuphh   %v24, %v24
        br      %r14

unpack:
        vuplhw  %v24, %v24
        br      %r14

So it looks like a final step is missed where vmrlg + vuphh can be rewritten to just vuplhw (and similarly for the other vector types). Or maybe the shuffle vector should be recognized directly. Anyway, this seems achievable.

This came up in the context of the rust standard library

cc @uweigand