New register allocator

[Geoff Keating]

> Date: Tue, 9 Jul 2002 23:05:03 +0200 (CEST)
> From: Michael Matz <matz@suse.de>
> Cc: <gcc-patches@gcc.gnu.org>
> 
> Hi Geoff,
> 
> I'm back, so:
> 
> On Wed, 3 Jul 2002, Geoff Keating wrote:
> 
> > wanted reviewed and would commit.  I looked at:
> >
> > <http://user.cs.tu-berlin.de/~matzmich/ra/ra_all.diff.gz>
> 
> That's the correct one.  Besides the changes to one .texi file equal to
> what I send you directly.

Did you post that?  There are others on gcc-patches that are better
than I at reviewing documentation patches...

> I addressed your comments, except the below to which I've questions:
> 
> > - Please don't add any new invocations of ggc_add_root or
> >   ggc_mark_rtx.  Both these routines are obsolete and will be deleted
> >   Real Soon Now.  Use GTY markers instead.
> 
> Hmm.  But the one place where I need a new root, the exact content of the
> variable is only created dynamically.  Does it still work for a static
> variable that is used in this way?  (I believe Mark once had a similar
> problem a few weeks ago).  I.e. I create a "template" insn into a static
> var (originally initialized to NULL_RTX), which I then reuse for the rest
> of the compile run, only slightly changing it each time, to not have to
> reconstruct it again and again. I.e. basically this pattern:
> 
>   void bla() {
>     static ptr the_var = 0;
>     if (!the_var)
>       the_var = init_me ();
>   }
> 
> Can I simply mark the_var with GTY() and it works as expected?

Yes, so long as it's globally visible, that is you have to say

static GTY(()) rtx the_var;
void bla()
{
  if (the_var == NULL)
    the_var = ...
...
}

There are lots of examples in GCC's source code of how to do this, for
instance, look at lookup_cl in java/parse.y.

It doesn't matter that sometimes the_var is NULL.  The GC code knows
about NULL pointers.

-- 
- Geoffrey Keating <geoffk@geoffk.org> <geoffk@redhat.com>