Re: [PATCH] elf: Replace memcmp with __memcmpeq for variable size

["H.J. Lu" <hjl.tools@gmail.com>]

On Sun, Feb 6, 2022 at 2:19 PM Florian Weimer <fw@deneb-ng.enyo.de> wrote:
>
> * H. J. Lu via Libc-alpha:
>
> > diff --git a/elf/dl-cache.c b/elf/dl-cache.c
> > index 88bf78ad7c..8574d4ded1 100644
> > --- a/elf/dl-cache.c
> > +++ b/elf/dl-cache.c
> > @@ -72,7 +72,7 @@ glibc_hwcaps_compare (uint32_t left_index, struct dl_hwcaps_priority *right)
> >      to_compare = left_name_length;
> >    else
> >      to_compare = right->name_length;
> > -  int cmp = memcmp (left_name, right->name, to_compare);
> > +  int cmp = memcmp_eq (left_name, right->name, to_compare);
> >    if (cmp != 0)
> >      return cmp;
> >    if (left_name_length < right->name_length)
>
> This change is not correct.

Fixed.

> The x86-specific <memcmp-eq.h> optimization probably applies to other
> targets as well.  I also do not quite see where the performance

It is useful only if __memcmpeq isn't an alias of memcmp.  I certainly
don't mind moving x86-64 <memcmp-eq.h> to generic.

> benefits come from.  None of the changed spots look particularly hot
> to me.

These are noises.  Here are the new data:

The cycles to run "elf/tst-relsort1 --direct" which calls __memcmpeq
24 times in ld.so:

Before:

     62704:
     62704: runtime linker statistics:
     62704:   total startup time in dynamic loader: 130771 cycles
     62704:             time needed for relocation: 32153 cycles (24.5%)
     62704:                  number of relocations: 97
     62704:       number of relocations from cache: 3
     62704:         number of relative relocations: 1347
     62704:            time needed to load objects: 43704 cycles (33.4%)
     62704:
     62704: runtime linker statistics:
     62704:            final number of relocations: 131
     62704: final number of relocations from cache: 3

After:

     62705:
     62705: runtime linker statistics:
     62705:   total startup time in dynamic loader: 117103 cycles
     62705:             time needed for relocation: 28327 cycles (24.1%)
     62705:                  number of relocations: 97
     62705:       number of relocations from cache: 3
     62705:         number of relative relocations: 1347
     62705:            time needed to load objects: 39550 cycles (33.7%)
     62705:
     62705: runtime linker statistics:
     62705:            final number of relocations: 131
     62705: final number of relocations from cache: 3

These numbers change for each run.  __memcmpeq has the lower
cycles.

> And I thought the assumption was that GCC would perform this optimization?

Yes, but it will take a while to implement it in GCC.

-- 
H.J.