[RunDevelopment]

This adds another method from std::simd: reverse_bits.

It's implemented as first reversing the bits within the u8 bytes of the vector and then using swap_bytes to reverse the byte order for the given integer type. Platforms only differ in how they do the 8-bit reversal:

• Arm Neon uses its vrbitq_u8 instruction.
• AVX512 uses the gf2p8affine family of intrinsics to do the bit reversal in a single instruction (plus loading a constant). See this blog post.
• All others use shifts and masking.

[Shnatsel]

A shuffle-based formulation models on llvm-mca as 2.5x faster than lots of shifts and bitwise ops, benchmarks with about the same improvements on my CPU. I've pushed it.

[Shnatsel]

Looks good to me but I'd appreciate if you could take a look at my changes

[Shnatsel]

And in case you're wondering, the scalar fallback on SSE2 autovectorizes into the same slow-ish sequence of SIMD shifts.

[RunDevelopment]

A shuffle-based formulation

Ohh, I read about exactly this use case before and completely forgot about it. Great catch!

Changes LGTM!
(I can't approve approve the PR, so take this as my approval :) )

And in case you're wondering, the scalar fallback on SSE2 autovectorizes into the same slow-ish sequence of SIMD shifts.

I was wondering that. Thanks!

[Shnatsel]

(I can't approve approve the PR, so take this as my approval :)

Github does let you. It doesn't let anyone merge the code with only your review, but you can still leave a review.

[RunDevelopment]

I can't approve my own PR :)

[Shnatsel]

Ah I see! I didn't think of that