Re: [PATCH, ARM] Use a pseudo for the PIC base

[Richard Earnshaw <Richard dot Earnshaw at buzzard dot freeserve dot co dot uk>]

On Wed, 18 Jan 2006 01:01:08 GMT, Paul Brook wrote:
> On Wednesday 18 January 2006 00:51, Richard Earnshaw wrote:
> > On Tue, 17 Jan 2006 16:04:24 PST, Richard Henderson <rth@redhat.com>
> >
> > wrote:
> > > On Tue, Jan 17, 2006 at 08:29:48PM +0000, Richard Earnshaw wrote:
> > > > It seems that contrary to traditional belief no further changes to the
> > > > mid-end code are needed.
> > >
> > > How do you handle constants being spilled to memory during reload,
> > > in a block for which the pic pseudo register was not previously live?
> >
> > I've never seen it happen in any testing, and I've seen a fair number of
> > cases where reload was affecting the code.  If the pic base can't have a
> > hard register it just gets split and reloaded as a normal pseudo.  The sum
> > of the pic base plus one of the offsets, isn't recorded as a constant, so
> > I guess reload would never just throw it away.
> 
> Could it be because of the way the Arm backend does its own constant pool 
> generation? ie. reload never tries to spill constants to a PIC-base relative 
> address.
> 
> Paul
> 

I don't think so. 

AFAICT reload will only throw a value away if it is in reg_equiv_constant. 
 But that is only set if the value is accepted by LEGITIMATE_CONSTANT_P, 
and that doesn't include anything that includes the PIC base, since that's 
not wrapped in (const (...))

R.