On Sat, Dec 30, 2017 at 12:11:50PM -0500, Daniel Micay wrote:
> On 30 December 2017 at 06:44, Otto Moerbeek <o...@drijf.net> wrote:
> > On Sat, Dec 30, 2017 at 06:53:44AM +0000, kshe wrote:
> >
> >> Hi,
> >>
> >> Looking at this diff and the previous one, I found some more possible
> >> cleanups for malloc.c (the patch below is to be applied after both of
> >> them, even if the second one has not been committed yet):
> >>
> >> 1. In malloc_bytes(), use ffs(3) instead of manual loops, which on many
> >> architectures boils down to merely one or two instructions (for example,
> >> see src/lib/libc/arch/amd64/string/ffs.S).
> >
> > I remember doing some measurements using ffs a long time ago, and it
> > turned out that is was slower in some cases. Likely due to the
> > function call overhead or maybe a non-optimal implementation. So I'm
> > only willing to consider this after seeing benchmarks on a handfull of
> > architectures.
>
> There's __builtin_ffs but Clang / GCC might not be smart enough to use
> it for ffs(3) automatically.
A quick test on amd64 (clang) and armv7 (gcc) shows that on both cases
the compiler generates inline code and no function call. So that is
promising.
-Otto
