Re: [RFC] PR/19653: fld used for reloads in sse mode

[Paolo Bonzini <paolo dot bonzini at lu dot unisi dot ch>]

> > >     (push_reload): Allow PREFERRED_*_RELOAD_CLASS to liberally
> > >     return NO_REGS.
> >
> > Eh, why is this necessary?  Can't you just return the old class instead?
>
> It makes sense to return NO_REGS for optional output reloads, but you 
> don't know that inside the target macro.

I think I should expand more on it...

Also, find_reloads is checking if PREFERRED_RELOAD_CLASS returns NO_REGS 
and, if so, it is equivalent to adding a `?' to the constraint (but in a 
way that may be dependent on the target options).  With my patch 
find_reloads also uses PREFERRED_OUTPUT_RELOAD_CLASS, but the failure in 
push_reload can happen with PREFERRED_RELOAD_CLASS as well.

The find_reload code is incompatible with the assert in push_reload, 
even without my patch.  So I made push_reload understand a return value 
of `NO_REGS' as "ok, you'd not like that I used this alternative, but 
the choice was made and we have to live with it; let's proceed with the 
original regclass".

After bootstrapping and regtesting the patch, I will repost it in a more 
polished form, including documentation for the behavior of find_reloads.

Paolo