[PATCH 1/1] powerpc: crtsaveres.o needed only when -Os flag is enabled

[Ram Pai]

On Wed, Apr 30, 2014 at 03:14:54PM +1000, Benjamin Herrenschmidt wrote:
> On Tue, 2014-04-29 at 10:38 -0500, Brian W Hart wrote:
> 
> > >  CHECKFLAGS	+= -m$(CONFIG_WORD_SIZE) -D__powerpc__ -D__powerpc$(CONFIG_WORD_SIZE)__
> > >  
> > > +ifdef CONFIG_CC_OPTIMIZE_FOR_SIZE
> > >  KBUILD_LDFLAGS_MODULE += arch/powerpc/lib/crtsavres.o
> > > +endif
> > > +
> > >  
> > >  # No AltiVec or VSX instructions when building kernel
> > >  KBUILD_CFLAGS += $(call cc-option,-mno-altivec)
> > 
> > I didn't try building a kernel or in-tree modules, but I confirmed
> > that it allows building of out-of-tree modules when crtsavres.o is
> > not present (e.g. as for a distro install where the kernel headers
> > are provided by package, rather than being manually prepared from
> > the sources).
> > 
> > Tested-by: Brian W Hart <hartb at linux.vnet.ibm.com>
> 
> I still don't like it. What guarantee do we have that gcc will never
> call into this with other optimisation settings ? It might decide
> one day that calling out for saving a large pile of registers
> is still more efficient than unrolling the whole lot, including
> for speed.

This patch operates on the assumption that arch/powerpc/lib/crtsavres.o 
is needed only if the code is compiled with -Os.  Are you saying
this assumption is wrong?

> 
> Besides that doesn't fix the root problem. We want to be able to
> build the kernel with CONFIG_CC_OPTIMIZE_FOR_SIZE and still have
> modules.

And this patch will not stop you from doing that. You can compile your
kernel with CONFIG_CC_OPTIMIZE_FOR_SIZE and modules will be built
because arch/powerpc/lib/crtsavres.o will be linked with the module.
Now, if the arch/powerpc/lib/crtsavres.o file does not exist, that
is a different problem and has to be fixed by the distros for 
out-of-tree modules.

> 
> So a better solution needs to be found. I don't know what that
> solution is (we might want to look at what other archs are doing
> maybe ?), could be to include crtsaveres.S in the build of every
> module (we really don't want to EXPORT_SYMBOL these guys), but
> that would mean having it installed somewhere with the kernel
> headers for out-of-tree modules...

Currently crtsaveres.o is expected to be in the build during
the linking stage of the module. You suggest instead have 
crtsaveres.S and get it compiled and linked?

> 
> If necessary, involve lkml, Rusty etc... but this patch is crap.

I dont see other archs having this problem. Possibly because there
linker have inbuilt capabilities to satisfy the missing symbols?

Alan Modra did mention that the ppc linker will soon have the
capability to handle -Os compiled code, without the help
of arch/powerpc/lib/crtsavres.o.

However this patch is not about having crtsavres.o or crtsaveres.S

Its about not needing crtsavres.o if the code is not compiled for
space optimization using -Os. If you say that the assumption
is wrong, than yes the code is crap :)


RP