register long sp asm("r1") incorrect

[Benjamin Herrenschmidt]

On Mon, 2010-02-15 at 21:28 +0100, Pavel Machek wrote:
> On Tue 2010-02-16 06:59:52, Benjamin Herrenschmidt wrote:
> > On Mon, 2010-02-15 at 08:34 +0100, Pavel Machek wrote:
> > > > On Tue, 2010-02-09 at 16:24 +0100, Pavel Machek wrote:
> > > > > ...according to gcc docs, sp should be global, or placement in
> > > > > register is not guaranteed (except at asm boundaries, but there
> > > are
> > > > > none).
> > > > 
> > > > Sorry I'm not sure I grok what you mean.
> > > 
> > > Well, according to gcc doscs and my experience, local "register int
> > > __asm()" variables only work by accident (or not at all). 
> > 
> > Hrm... we definitely rely on that for our thread_info() access, and so
> > far it has worked well for us, but I'll poke our gcc folks just in case.
> 
> Thanks, and let me know about any results.

All the gcc folks I talked to say something along the lines that there
is no way in hell it doesn't work :-)

It's true that most other use of it we have are global scope (local_paca
in r13, glibc use of r2/r13, etc...) afaik, but since r1 itself is the
stack pointer always, I think they pretty much guarantee it works.

I'm CCing a couple of experts just to be sure.

Cheers,
Ben.