This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: PATCH - remove 6 pointer fields from C's lang_identifier
- From: Per Bothner <per at bothner dot com>
- To: Per Bothner <per at bothner dot com>
- Cc: Gcc Patch List <gcc-patches at gcc dot gnu dot org>
- Date: Tue, 07 Jan 2003 11:55:52 -0800
- Subject: Re: PATCH - remove 6 pointer fields from C's lang_identifier
- References: <3E19F315.1070808@bothner.com>
As a sanity check, I added the line:
printf("alloc '%s' ext new kind:%d old kind:%d\n",
IDENTIFIER_POINTER(id), value_index, cur_kind);
in set_identifier_value when it allocates an extention TREE_VEC,
and then recompiled expr.o c-common.o and builtins.o (three
random large files), and got reassuring results:
expr.c had 2 "conflicts" ("index" and "exp").
c-decl.c had 4 ("link", "exp","dup", and "var_labelno").
builtinc.s had 2 ("link" and "exp").
All except one was where the same identifier had both
IDENTIFIER_GLOBAL_VALUE and IDENTIFIER_LOCAL_VALUE non-null.
In one case ("var_labelno") both IDENTIFIER_LIMBO_VALUE
and IDENTIFIER_LOCAL_VALUE were set.
These very low numbers supports the intuition that the
overwhelming number of identifiers do not need to
allocate an extensions TREE_VEC. So the net effect is
a definite and significant space saving. The net time
effect is harder to pin down: The new macros require extra
instructions (to test the IDENTIFIER_FIELD_KIND), but because
of the space savings we do less gc, and locality should be
improved. My numbers, plus some results sent me privately
by Judd Jenne, suggest a minor but worthwhile improvement.
It might be possible to tweak things better, perhaps by
re-ordering the tests in IDENTIFIER_FIELD, or adding some
annotation a la _builtin_expect (?), but that's not sort of
micro-optimization isn't really my field.
--
--Per Bothner
per@bothner.com http://www.bothner.com/per/