Re: Various gp-related mips optimisations

[Richard Sandiford <rsandifo at redhat dot com>]

Eric Christopher <echristo@redhat.com> writes:
> > ! 	  /* compute_frame_size assumes that any function which uses the
> > ! 	     constant pool will need a gp.  However, all constant
> > ! 	     pool references could be eliminated, in which case
> > ! 	     it is OK for flow to delete the gp load as well.  */
> > ! 	  insn = get_last_insn ();
> > ! 	  for (i = 0; i < 3; i++)
> > ! 	    {
> > ! 	      REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD, const0_rtx,
> > ! 						    REG_NOTES (insn));
> > ! 	      insn = PREV_INSN (insn);
> > ! 	    }
> > ! 	}
> >       }
> 
> Definitely don't like this block, but I think that we need it since we
> can't accurately (or at all afaict) determine ahead of time whether or
> not we're going to be deleting all of the constant pool references.
> Anyone else see a way around this?

Do we need a way around it?  I think the block is correct in terms of
what it's doing.  (i.e. I don't see it as a hack.)

i386 for one does a similar thing:

      /* Even with accurate pre-reload life analysis, we can wind up
	 deleting all references to the pic register after reload.
	 Consider if cross-jumping unifies two sides of a branch
	 controlled by a comparison vs the only read from a global.
	 In which case, allow the set_got to be deleted, though we're
	 too late to do anything about the ebx save in the prologue.  */

Note that that comment was written before ix86_save_reg started
checking current_function_uses_const_pool.  I think the problem
described in my comment would now apply to i386 as well.

I agree that the code isn't pretty.  I'd like to be able to write
it in a way that doesn't use a loop like the above.  But however
it's written, I still think we need the notes.

Richard