Use a bitwise and instead of shifts

[AreaZR]

This is faster in some instances and is clearer in intent

[StephanTLavavej]

Thanks for looking into this!

The optimized codegen is identical: https://godbolt.org/z/K4envEYs5

I believe that the debug codegen is not important here - these are per-algorithm-call, and we generally don't worry about slightly suboptimal instruction sequences (whereas we occasionally worry about many layers of unnecessary function calls, e.g. in invoke()).

So the main question is clarity, which is a judgement call. I believe that the existing pattern of shifting back and forth by N more clearly expresses the intent to clear the low N bits. Marking as "decision needed" for the other maintainers to consider.

[MikeGitb]

Just a suggestion: If you want to use this notation, I'd use binary litterals instead of decimal numbers.

[AlexGuteniev]

My opinion:

• The clear way is * 16 / 16, though the debug codegen is expected to be awful (with actual division)
• I'd wrote & 0xF. Hex is a compact binary.
• There's no much point to make this clearer. Those who read SSE / AVX will not be confused by any of the opitions

[strega-nil-ms]

I strongly agree that & ~... is more clear than >> N << N; however, I would also appreciate @AlexGuteniev's heximal literal change (or even binary literal). I'm not against making this clearer.

[StephanTLavavej]

Ok, I've validated and pushed changes to:

• Use hex literals as suggested by @AlexGuteniev and @strega-nil-ms.
• Use ~size_t{VALUE} to start with a constant of the desired width and avoid having to think about signed-to-unsigned conversions. (size_t & ~31 worked because ~31 is a negative int that gets sign-extended when converted to size_t, which I found highly surprising when I realized it.)
• Also change >> 6 << 5 and >> 5 << 4.
  • I believe that this is a clarity improvement, as this makes the division by 2 (followed by masking) more obvious, so I am now in favor of this change.

Codegen is unaffected: https://godbolt.org/z/h7qE3vca1

Thanks again!

[StephanTLavavej]

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.

[StephanTLavavej]

Thanks for this code cleanup! 😸 😸 😸 😸 😸 😸 😸 😸