Re: altivec triplets: comments?

[Aldy Hernandez <aldyh at redhat dot com>]

On Thu, 2001-11-08 at 11:30, Daniel Jacobowitz wrote:
> On Wed, Nov 07, 2001 at 09:36:43PM -0500, Aldy Hernandez wrote:
> > >>>>> "Geoff" == Geoff Keating <geoffk@geoffk.org> writes:
> > 
> >  > The other problems only happen if you have a call-saved altivec
> >  > register, or use them for argument passing.  There's nothing that
> >  > prevents you from making all the altivec registers call-clobbered, and
> >  > prohibiting passing arguments in altivec registers, and then
> >  > everything should Just Work, although perhaps a little less efficiently.
> > 
> > Vector registers are only used to hold arguments for -mabi=altivec,
> > otherwise they go in the stack.  Also, altivec registers are
> > call-saved only for altivec abi.  So theoretically it should just
> > work.
> 
> Oh, I see that we actually already have 128-bit alignment for our
> stack in SysV.  I didn't know that.
> 
> Just to clarify my understanding - do the ABI changes include
> anything that would disturb non-Altivec-aware assembly (such as the
> stubs in glibc)?  I suppose they could in the presence of
> __attribute__((aligned)) and hard-coded structure offsets, but
> hopefully we don't have that problem very often.
> 
> (Is -mabi=altivec orthogonal to -maltivec, or does it imply it?)

-maltivec means use altivec instructions, but no abi changes.

-mabi=altivec means enhance the current abi with altivec extensions, but
does not imply that the compiler will generate altivec instructions.

to actually get something useful you'll probably need both.

-maltivec alone will give you altivec instructions but without the abi
changes, so everything gets passed on the stack, etc.

> 
> -- 
> Daniel Jacobowitz                           Carnegie Mellon University
> MontaVista Software                         Debian GNU/Linux Developer
-- 
Aldy Hernandez			E-mail: aldyh@redhat.com
Professional Gypsy
Red Hat, Inc.