Re: [Patch 1/4] Allow arch-specific cleanup before breakpoint unregistration

[Frederic Weisbecker <fweisbec@gmail.com>]

On Wed, May 26, 2010 at 10:47:42PM +0530, K.Prasad wrote:
> On Wed, May 26, 2010 at 10:54:41AM +0100, David Howells wrote:
> > K.Prasad <prasad@linux.vnet.ibm.com> wrote:
> > 
> > > > My understanding is weak function definitions must appear in a different C
> > > > file than their call sites to work on some toolchains.
> > > > 
> > > 
> > > Atleast, there are quite a few precedents inside the Linux kernel for
> > > __weak functions being invoked from the file in which they are defined
> > > (arch_hwblk_init, arch_enable_nonboot_cpus_begin and hw_perf_disable to
> > > name a few).
> > > Moreover the online GCC docs haven't any such constraints mentioned.
> > 
> > I've seen problems in this area.  gcc sometimes inlines a weak function that's
> > in the same file as the call point.
> > 
> 
> We've seen such behaviour even otherwise....even with noinline attribute
> in place. I'm not sure if this gcc fix
> (http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16922) helped correct the
> behaviour, but the lesson has been to not trust a function to be
> inlined/remain non-inline consistently.


If we can't put the call to the function in the same file of its weak
definition, then perf is totally screwed.

And in fact it makes __weak basically useless and unusable. I guess
that happened in old gcc versions that have been fixed now.

Anyway, I'm personally fine with this patch (you can put my hack
if you want).

Thanks.