Re: [PATCH][updated] Force SDmode function args into FP registers per the ABI

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

Peter Bergner wrote:
> On Tue, 2007-06-19 at 09:04 +0200, Paolo Bonzini wrote:
> > This can be simplified to be just
> >
> > enum machine_mode mode = PSEUDO_REGNO_MODE (i);
> > unsigned int inherent_size = targetm.min_stack_slot_size_for_mode (mode);
> > unsigned int total_size = MAX (inherent_size, reg_max_ref_width[i]);
> [snip]
> > Sorry for the further nitpicking...  I can take care of this, but I 
> > wouldn't mind preapproval since this was what Mark mentioned here:
> >
> > > The
> > > default would be PSEUDO_REGNO_BYTES (REGNO (reg)).  Then, in alter reg,
> > > use the max of reg_max_ref_width and the value of the hook?
>
> I debated about setting inherent_size with the target hook like you
> did above, but decided against it given all of the other uses of
> inherent_size in alter_reg and the fact that by changing its value
> like this (ie, inherent_size will be less than total_size), we will
> cause it to go down different paths in this routine and pass different
> function args to assign_stack_local.  Maybe it would be ok, but errored
> on the side of caution.

You're right.  From reading the code more carefully, however, I think 
that my proposed change would be the right thing to do, and not just a 
cleanup.  The reason is this code:

  /* Reuse a stack slot if possible.  */
  else if (spill_stack_slot[from_reg] != 0
           && spill_stack_slot_width[from_reg] >= total_size
	   && (GET_MODE_SIZE (GET_MODE (spill_stack_slot[from_reg]))
               >= inherent_size)

If the register has already been spilled as a SFmode register, and now 
it is requested to spill as a SDmode register, your patch will reuse the 
1-word spill slot for the SDmode register.  I don't know what kind of 
failure this would lead to, because of course I have no testcase, but it 
seems wrong.

Paolo