Re: [PATCH 1/25] Remove nested functions: crypt/md5-crypt.c

["Joseph S. Myers" <joseph at codesourcery dot com>]

On Tue, 20 May 2014, Siddhesh Poyarekar wrote:

> On Tue, May 20, 2014 at 06:23:40PM +0400, Konstantin Serebryany wrote:
> > Do you think such a benchmarks is possible at all?
> > I've made a simple test (attached) and profiled it with "perf":
> >     85.26%    a.out  libcrypt-2.15.so   [.] _ufc_doit_r
> >      6.13%    a.out  libcrypt-2.15.so   [.] _ufc_mk_keytab_r
> >      2.40%    a.out  libcrypt-2.15.so   [.] _ufc_setup_salt_r
> >      1.56%    a.out  libcrypt-2.15.so   [.] __crypt_r
> >      1.51%    a.out  libcrypt-2.15.so   [.] _ufc_output_conversion_r
> >      1.35%    a.out  libcrypt-2.15.so   [.] crypt
> > 
> > As you can see, crypt_r itself takes a tiny fraction of time,
> > most of it is spent in _ufc_* which are defined in another module.
> > Any changes in crypt itself (unless you do something insane) will not
> > be observable in profile.
> 
> You'll need to choose inputs so that __md5_crypt_r is called -
> crypt/crypt-entry.c should help you with that.  if __md5_crypt_r takes
> a tiny fraction of time again, then this change should be safe, but I
> would prefer that the inputs get fed into the benchtests anyway so
> that we can automatically track performance of crypt for those inputs
> in future.

Another thing that can be done is to look at the code generated for the 
relevant file before and after the patch - I'd expect very little change.

-- 
Joseph S. Myers
joseph@codesourcery.com