Re: [PATCH] fixing sendfile on 64bit architectures

[David Mosberger <davidm@napali.hpl.hp.com>]

>>>>> On Thu, 20 May 2004 15:25:10 -0700, Andrew Morton <akpm@osdl.org> said:

  Andrew> Jesse Barnes <jbarnes@engr.sgi.com> wrote:
  >>  On Thursday, May 20, 2004 5:56 pm, Andrew Morton wrote: > An
  >> alternative might be to remove all the ifdefs, build with >
  >> -ffunction-sections and let the linker drop any unreferenced
  >> code...

  >> That would probably be even more confusing than the #ifdefs.  At
  >> least with those you know that you need to check whether the
  >> current code will be called...

  Andrew> Me no understand Jesse.

  Andrew> Removing the ifdefs and letting the linker do the job has
  Andrew> the advantage that the compiler gets to check more code for
  Andrew> you.

I think he meant that it's easier to see who (what platform) is
relying on the code.

Another disadvantage of -ffunction-sections is that the compiler won't
be able to do proper inlining.  There are several cases where you
have:

static
do_the_work ()
{
  ...
}

asmlinkage long
sys_one_variant ()
{
  do_the_work(some args...);
}

asmlinkage long
sys_other_variant ()
{
  do_the_work(other args...);
}

If only one variant is needed, GCC 3.4 will automatically inline
"do_the_work()" (since there is only one call-site), which is exactly
what you want.  You won't get that with -ffunction-sections.

	--david