Re: [PATCH] [RFC] x86/cpu: rework instruction set selection

["H. Peter Anvin" <hpa@zytor.com>]

On April 27, 2025 12:34:46 PM PDT, Linus Torvalds <torvalds@linux-foundation.org> wrote:
>On Sun, 27 Apr 2025 at 12:17, Andrew Cooper <andrew.cooper3@citrix.com> wrote:
>>
>> ffs/fls are commonly found inside loops where x is the loop condition
>> too.  Therefore, using statically_true() to provide a form without the
>> zero compatibility turns out to be a win.
>
>We already have the version without the zero capability - it's just
>called "__ffs()" and "__fls()", and performance-critical code uses
>those.
>
>So fls/ffs are the "standard" library functions that have to handle
>zero, and add that stupid "+1" because that interface was designed by
>some Pascal person who doesn't understand that we start counting from
>0.
>
>Standards bodies: "companies aren't sending their best people".
>
>But it's silly that we then spend effort on magic cmov in inline asm
>on those things when it's literally the "don't use this version unless
>you don't actually care about performance" case.
>
>I don't think it would be wrong to just make the x86-32 code just do
>the check against zero ahead of time - in C.
>
>And yes, that will generate some extra code - you'll test for zero
>before, and then the caller might also test for a zero result that
>then results in another test for zero that can't actually happen (but
>the compiler doesn't know that). But I suspect that on the whole, it
>is likely to generate better code anyway just because the compiler
>sees that first test and can DTRT.
>
>UNTESTED patch applied in case somebody wants to play with this. It
>removes 10 lines of silly code, and along with them that 'cmov' use.
>
>Anybody?
>
>              Linus

It's noteworthy that hardware implementations are now invariably using a different interface, which makes sense for the LSB case (tzcnt/ctz) but has its own drain bramage on for the MSB case (lzcnt/clz)...