Re: Reducing duplication of bit operation code.

[Paul Eggert]

On 2024-12-17 14:56, Bruno Haible via Gnulib discussion list wrote:
> Hi Collin,
>
> > Is there any reason not to use stdbit.h for ffs, ffsl, ffsll,
> > count-leading-zeros, count-one-bits, count-trailing-zeros?
>
> Before doing so, how about making sure that stdbit.in.h works
> at least as well as the old code w.r.t. inlining (assuming -O1 or higher)?
> (Recall that these functions do very little work, therefore an extra
> function call cost is significant.) Maybe it already does, but I haven't
> verified it; could you verify it with reasonably modern gcc and clang?
>
> Also, I see that MSVC is not well covered by this default in stdbit.in.h:
>
> #else
> /* This platform is so old that it lacks typeof, so _Generic is likely
>     missing or unreliable.  The C23 standard seems to allow yielding B
>     (which is always unsigned long long int), so do that.  */
> # define _GL_STDBIT_TYPEOF_CAST(a, b) (b)
> #endif
>
> Can it be improved?

Although both of these questions are reasonable if we have a lot of 
time, I suspect they're not worth the effort of answering. With modern 
GCC or Clang I don't think you'll see much performance difference, and 
for ancient GCC/Clang as long as it works correctly we should be OK.

Even MSVC should be fine if it's new enough; i.e., I don't think there 
will be significant performance differences in real applications.

That being said, I don't have access to MSVC and didn't check its 
performance. Also, I looked at GCC performance with -O2 not -O1, as -O2 
is the usual build default.