Re: [new-ra] DF_REF_REG_CLOBBER related bugfix

[Denis Chertykov <denisc at overta dot ru>]

Michael Matz <matz@suse.de> writes:

> Hi Denis,
> 
> On Fri, 24 Oct 2003, Denis Chertykov wrote:
> 
> > Current `insert_stores' handle (clobber REG) as def. It's wrong.
> > The slot for REG must be deleted from `slots' hash list.
> 
> Indeed.  Please commit.
> 
> > 	(coalesce_spill_slot): Handle webs without conflicts.
> >
> > --- 2345,2355 ----
> >       }
> >     else
> >       return 0;
> > !
> > !   /*  Webs without conflicts must be rejected.  */
> > !   if (!s->conflict_list || !t->conflict_list)
> > !     return 0;
> > !
> 
> For what is this needed?  Could you extend the comment?

After you have added a support of webs with "xcall" attribute you have
enabled webs without usable_regs and without conflict_list.
This is a spill-webs which will be definitely in stack memory.
coalesce_spill_slot can't handle such webs because they havn't
conflicts with 'corresponding webs'.
IE:

w1 - web without conflict_list (now w1 got stack slot)
w2 - 'corresponding web' (web which connected to w1 by move) and
     havn't a conflict with w1 because of w1->usable_regs is empty

i1 def w1
....
i2 move w2 <= w1
i3 def w2, use w2
i4 use w2
i5 use w1

coalesce_spill_slot will try to coalesce w1 with w2. 

The following condition isn't work
(it's condition from `non_conflicting_for_combine' and similar
condition exists in coalesce_spill_slot):

  if (!TEST_BIT (sup_igraph, w1->id * num_webs + w2->id)
      && !TEST_BIT (sup_igraph, w2->id * num_webs + w1->id))
    return 1;

So, the result:

i1 def w1-slot
....
INSN_DELETED i2 move w2 <= w1
i3 def w1-slot, use w1-slot
i4 use w1-slot
i5 use w1-slot

It's a buggy result because w2 has a def in i3.

May be non_conflicting_for_combine will have a similar problem ?

Also, Right now I have understood the following thing:
By my idea coalesce_spill_slot must remove all unnecessary moves
generated by rewrite_program2 if the web will be in stack but in
described example coalesce_spill_slot tries to delete move which
wasn't generated by rewrite_program2. It's not a bug. It's a bit
wrong realization ;(

Denis.