Shouldn't processors have a special instruction for this operation?

mananaysiempre · on Dec 7, 2022

Rotate through carry works for that. (Maybe not on RISC-V, it’s kind of picky about things you can’t easily access in high-level languages.) Unsigned floored, x86:

  add eax, ebx
  rcr eax, 1

ARM:

  adds r0, r0, r1
  rrx r0, r0

I’d need to think a bit more to come up with a signed version.

rep_lodsb · on Dec 8, 2022

>I’d need to think a bit more to come up with a signed version.

Invert the high bits, turning two's complement into biased format (0 = lowest negative number, 0xFF...F = highest positive number). Then do the add+rotate and convert the result back.

josefx · on Dec 7, 2022

Those four integer instructions are generally quite fast already. So unless this operation is needed extremely often there is probably no point in dedicating hardware to it.