Re: [PATCH] Fix builtin attributes on powerpc

[Michael Meissner <meissner@linux.vnet.ibm.com>]

On Tue, Oct 13, 2009 at 09:30:28AM -0400, David Edelsohn wrote:
> On Thu, Oct 1, 2009 at 2:57 PM, Michael Meissner
> <meissner@linux.vnet.ibm.com> wrote:
> > I tried to get this in by yesterday, but I had some server issues, and I
> > couldn't get the tests run until today.
> >
> > In working with Jakub on a backport of the power7 patches to the Red Hat 4.4
> > branch, we found some problems in running spec.  The problem turned out to that
> > the vectorized math functions had different declaration attributes than the
> > scalar versions of the functions.  In doing further investigation, it turned
> > out that we weren't setting any attributes for any builtin functions.  This
> > patch fixes the attributes for the Altivec (VMX) and VSX builtins.  I have not
> > done the spe and paired builtins in this patch, but that can be easily done
> > later.
> >
> > I did the usual boostrap and make check with no regressions.
> >
> > Is this acceptable to checkin under stage 3 as a backend change?
> 
> If this is the design that the community wants to use in all backends,
> it is fine with me.

It is too late for GCC 4.5 to do things in a common fashion.  The real solution
is to move the backend builtins into the main builtins.  I had some patches to
allow backends to optionally do this, and when 4.6 opens up we can explore
this.

However, I'm trying to fix the problem right now, and so this particular
solution is limited to just to tie the definition of the builtin enum and flags
together in one place, rather than having a separate switch statement that is
problematical, because you have to edit stuff in multiple places.  We already
have the problem between the rs6000.h, rs6000.c, and rs6000-c.c files that each
have some knowledge of builtins (and overloaded builtins make it even worse).

> I think MAX_RS6000_BUILTINS is more consistent with the naming
> convention in the rest of GCC than RS6000_BUILTIN_COUNT.

I agree MAX_RS6000_BUILTINS is more consistant, but RS6000_BUILTIN_COUNT is
what was originally used, so I didn't change that.

> The patch is okay.

Ok.

-- 
Michael Meissner, IBM
4 Technology Place Drive, MS 2203A, Westford, MA, 01886, USA
meissner@linux.vnet.ibm.com