Re: Fix PR 46297

[Jeff Law <law at redhat dot com>]

On 11/17/10 03:51, Bernd Schmidt wrote:
> On 11/06/2010 04:53 PM, Jeff Law wrote:
> >   The fundamental problem is some ports have embedded auto-inc addressing
> > modes without REG_INC nodes.  The improved reload optimizations were not
> > accounting for the possibility of embedded auto-inc addressing,
> > particularly for argument pushing.
> > +
> > +   if (MEM_P (dst))
> > +     {
> > +       dst = XEXP (dst, 0);
> > +       if (GET_CODE (dst) == PRE_INC || GET_CODE (dst) == POST_INC
> > + 	  || GET_CODE (dst) == PRE_DEC || GET_CODE (dst) == POST_DEC)
> > + 	{
> > + 	  for (i = hard_regno_nregs[regno][mode] - 1 + regno; i>= regno; i--)
> I don't think regno is set correctly here. Otherwise it looks sensible.
It certainly looks odd.  It's a cut-n-paste from the 
ZERO_EXTRACT/STRICT_LOW_PART slightly later in 
reload_combine_note_store.   ISTM inserting


          regno = REGNO (XEXP (dst, 0));
          mode = GET_MODE (XEXP (dst, 0));


just before the loop should address this problem.  Note we want the mode 
of the REG, not the mode of the memory reference when determining how 
many regs to mark within the loop.

I'll add the changes & retest.

Jeff