Graphite middle-end parts review

[Sebastian Pop]

Hi,

Attached patch fixes most of the changes asked, except the factoring
of update_dominators_in_loop code and the following two parts:

On Thu, Aug 7, 2008 at 7:13 AM, Richard Guenther <rguenther@suse.de> wrote:
> +  /* Construct IV code in loop.  */
> +  initial_value = force_gimple_operand (initial_value, &stmts, true, iv);
> +  if (stmts)
> +    {
> +      gsi_insert_seq_on_edge (loop_preheader_edge (loop), stmts);
> +      gsi_commit_edge_inserts ();
> +    }
> +
> +  add_referenced_var (iv);
>
> The caller should have done this.

No, Jan did the right thing here.  I also think that we should have
more systematic functions that take care of everything: creating the
variable, etc.

> Index: gcc/passes.c
> ===================================================================
> --- gcc/passes.c	(.../trunk)	(revision 138275)
> +++ gcc/passes.c	(.../branches/graphite)	(revision 138569)
> @@ -661,6 +661,7 @@ init_optimization_passes (void)
>  	  NEXT_PASS (pass_check_data_deps);
>  	  NEXT_PASS (pass_loop_distribution);
>  	  NEXT_PASS (pass_linear_transform);
> +	  NEXT_PASS (pass_graphite_transforms);
>
> So we still have loop-distribution and interchange before graphite?

Yes for the moment this is not changed.  It will change in the future.

Sebastian Pop
--
AMD - GNU Tools
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 1147_graphite_middle-end.diff
Type: text/x-diff
Size: 12413 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20080813/3ce8e3f0/attachment.bin>