------- Comment #28 from jamborm at gcc dot gnu dot org 2009-09-18 15:52
-------
(In reply to comment #27)
> Today we regressed with the introduction of IPA-SRA at -O2 and -O3:
>
The problem is that I call compute_inline_parameters() whenever I
change a single call site, even when the caller is the same. The
following patch (which I am about to bootstrap and test) should fix
that:
2009-09-18 Martin Jambor <mjam...@suse.cz>
* tree-sra.c (convert_callers): Do not call
compute_inline_parameters on one caller more than once.
Index: mine/gcc/tree-sra.c
===================================================================
--- mine.orig/gcc/tree-sra.c 2009-09-18 14:53:07.000000000 +0200
+++ mine/gcc/tree-sra.c 2009-09-18 17:26:57.000000000 +0200
@@ -3637,6 +3637,7 @@ convert_callers (struct cgraph_node *nod
tree old_cur_fndecl = current_function_decl;
struct cgraph_edge *cs;
basic_block this_block;
+ bitmap recomputed_callers = BITMAP_ALLOC (NULL);
for (cs = node->callers; cs; cs = cs->next_caller)
{
@@ -3644,15 +3645,24 @@ convert_callers (struct cgraph_node *nod
push_cfun (DECL_STRUCT_FUNCTION (cs->caller->decl));
if (dump_file)
- fprintf (dump_file, "Adjusting call %s -> %s\n",
+ fprintf (dump_file, "Adjusting call (%i -> %i) %s -> %s\n",
+ cs->caller->uid, cs->callee->uid,
cgraph_node_name (cs->caller),
cgraph_node_name (cs->callee));
ipa_modify_call_arguments (cs, cs->call_stmt, adjustments);
- compute_inline_parameters (cs->caller);
pop_cfun ();
}
+
+ for (cs = node->callers; cs; cs = cs->next_caller)
+ if (!bitmap_bit_p (recomputed_callers, cs->caller->uid))
+ {
+ compute_inline_parameters (cs->caller);
+ bitmap_set_bit (recomputed_callers, cs->caller->uid);
+ }
+ BITMAP_FREE (recomputed_callers);
+
current_function_decl = old_cur_fndecl;
FOR_EACH_BB (this_block)
{
--
jamborm at gcc dot gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
CC|mjambor at suse dot cz |jamborm at gcc dot gnu dot
| |org
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=12392
