Suboptimal codegen around Vector128.Create(span.Slice)

[EgorBo]

When I replace unsafe vectorized code with safe, I quite often see non-contained addressing modes, example:

int Sum(int[] arr)
{
    Span<int> span = arr;
    Vector128<int> sum = default;
    while (span.Length >= Vector128<int>.Count * 4)
    {
        var v1 = Vector128.Create(span.Slice(Vector128<int>.Count * 0));
        var v2 = Vector128.Create(span.Slice(Vector128<int>.Count * 1));
        var v3 = Vector128.Create(span.Slice(Vector128<int>.Count * 2));
        var v4 = Vector128.Create(span.Slice(Vector128<int>.Count * 3));
        sum += v1 + v2 + v3 + v4;
        span = span.Slice(Vector128<int>.Count * 4);
    }
    return Vector128.Sum(sum);
}

loop codegen:

G_M19912_IG05:  ;; offset=0x0015
       vmovups  xmm1, xmmword ptr [rax]
       lea      rdx, bword ptr [rax+0x10] ;; <---
       vmovups  xmm2, xmmword ptr [rdx]
       lea      rdx, bword ptr [rax+0x20] ;; <---
       vmovups  xmm3, xmmword ptr [rdx]
       lea      rdx, bword ptr [rax+0x30] ;; <---
       vpaddd   xmm1, xmm1, xmm2
       vpaddd   xmm1, xmm1, xmm3
       vpaddd   xmm1, xmm1, xmmword ptr [rdx]
       vpaddd   xmm0, xmm1, xmm0
       add      rax, 64
       add      ecx, -16
       cmp      ecx, 16
       jge      SHORT G_M19912_IG05

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.