This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [GSoC] generation of Gimple code from isl_ast_node_if

On 26/07/2014 10:59, Roman Gareev wrote:
If I'm not mistaken, the reason of this bug is incorrect creation of a
poly_bb_p for basic_block from the existing pbb in new_pbb_from_pbb
(It is located in graphite-sese-to-poly.c). I think, that we should
set a new id of pbb1->domain (instead of using the id of the pbb),
which contains pointer to the pbb1.
I found out this after dumping of an index of pbb in the user
statement S_3. Its index is 9. It is created in
rewrite_reduction_out_of_ssa using new_pbb_from_pbb and the pbb with
index 3. After that the user statement S_3 is removed in
build_scop_drs, but the id of the  pbb->domain and the
pbb->transformed point to the old pbb with index 3.

Interesting. I was not even aware that we did statement splitting for reductions. Very nice analysis.

I've attached the patch, which may fix this.

                                    Cheers, Roman Gareev.


Index: gcc/graphite-sese-to-poly.c
--- gcc/graphite-sese-to-poly.c	(revision 212995)
+++ gcc/graphite-sese-to-poly.c	(working copy)
@@ -2044,6 +2044,10 @@

    pbb1->domain = isl_set_copy (pbb->domain);
+  char name[50];
+  snprintf (name, sizeof (name), "S_%d", pbb_index (pbb1));
+  pbb1->domain = isl_set_set_tuple_id (pbb1->domain,
+				       isl_id_alloc (scop->ctx, name, pbb1));

Any reason you don't use isl_id_for_pbb() to create this isl_id?

Otherwise, the patch looks good to me. You can commit it if the graphite tests still pass and you add an appropriate ChangeLog entry.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]