The choice between full_masks.rs and bitmask.rs seems to be made at compile time (

) based on compile-time presence of the avx512f target feature.

When compiling for general x86-64 targets but doing runtime feature detection for avx512f, the current approach will generate suboptimal code for avx512, because it will use the full_masks.rs implementation instead of the bitmask.rs implementation.

I'd like to be able to use bitmasks in avx512 even when using runtime feature detection. It's probably out of scope (and, I think, undesirable) for std::simd to be responsible for runtime feature detection, so one option would be for std::simd to be refactored to expose both Mask and Bitmask as two separate types -- perhaps both implementing the same set of traits. This way, callers could choose between a full-mask implementation and a bitmask implementation based on whatever logic they choose, including e.g. runtime feature detection.