Re: PATCH: PR target/40470: unable to fin d a register to spill in class âSSE_FIRST_REG â

[Jeff Law <law at redhat dot com>]

H.J. Lu wrote:
> On Wed, Jun 17, 2009 at 7:22 PM, Jeff Law<law@redhat.com> wrote:
>   
> > Ian Lance Taylor wrote:
> >     
> > > Of course it would be nice if ra/reload handled this sort of thing
> > > better.
> > >
> > >       
> > The problem is primarily reload.  I'll try to keep this kind of thing in
> > mind for reload v2 -- in fact I think it would just work with the code I've
> > got right now.
> >
> >     
>
> This sounds very promising. What do we do with gcc 4.3 and 4.4?
> Here is an updated patch for 4.3 and 4.4. It fixes additional testcases
> by forcing into XMM0 during expand and changing operand constraint
> to reg_xmm0_operand so that other passes won't replace XMM0
> with a different register in such a way that RA/reload can't handle. Any
> comments?
>   
Someone needs to look at what it's going to take to fix combine.  I 
would recommend against your patch until we know precisely what changes 
we'd need to make to combine to fix this problem.

Jeff