add all_lane_counts feature to enable non-power-of-2 lane counts <= 64

[programmerjake]

as discussed here: https://rust-lang.zulipchat.com/#narrow/stream/257879-project-portable-simd/topic/experimental.20feature.20to.20allow.20non-power-of-2.20vector.20lengths

[calebzulawski]

I'm curious if this will dramatically slow down CI. I remember some sort of issue in the past with testing too many lane counts (though maybe that's not an issue anymore?)

[programmerjake]

I'm curious if this will dramatically slow down CI. I remember some sort of issue in the past with testing too many lane counts (though maybe that's not an issue anymore?)

the only ci jobs with the feature enabled finished waay faster than a lot of other jobs, i'd expect no slowdown.

[calebzulawski]

Some sort of const generics would be really nice here, I guess we need bounds first. This is pretty clean though.

[programmerjake]

we could just implement them for all usize values using a const generic impl, but iirc @workingjubilee specifically wanted to avoid 0.

[calebzulawski]

Yeah, I'm on the fence about 0. But more than that, LLVM has had issues with larger values in the past, and rustc also has an upper limit on #[repr(simd)], so we need an upper limit anyway.

[calebzulawski]

Looks good to me, assuming we're okay with the premise