Hi,
the following patch sets BINFOs of prevailing but non-main-variant
types to what the main variant has, so that for a type t with a binfo
the following is always true:
TYPE_BINFO (t) == TYPE_BINFO (TYPE_MAIN_VARIANT (t))
as I I believe in non-LTO world it always is. Moreover, the BINFO
trees of non-prevailing type nodes now always get garbage collected
(at least as far as my experiments with 483.xalancbmk go). The memory
savings are negligible, but still I think we should do it.
Bootstrapped and tested on x86_64-linux, also tested by LTOing Mozilla
Firefox and 483.xalancbmk.
OK for trunk?
Thanks,
Martin
2011-08-03 Martin Jambor <mjam...@suse.cz>
* lto.c (uniquify_nodes): Use main variant's BINFO too.
Index: src/gcc/lto/lto.c
===================================================================
--- src.orig/gcc/lto/lto.c
+++ src/gcc/lto/lto.c
@@ -720,6 +720,8 @@ uniquify_nodes (struct data_in *data_in,
{
TYPE_NEXT_VARIANT (t) = TYPE_NEXT_VARIANT (mv);
TYPE_NEXT_VARIANT (mv) = t;
+ if (RECORD_OR_UNION_TYPE_P (t))
+ TYPE_BINFO (t) = TYPE_BINFO (mv);
}
/* Finally adjust our main variant and fix it up. */
