[PATCH] regex performance improvement

[Isamu Hasegawa]

Hi,

From: Ulrich Drepper <drepper@redhat.com>

> As for this specific patch: in general it looks good.  I'd like to see a
> few more comments, especially before function definitions explaining the
> purpose of functions.
> 
> There are also some more critical problems.  You're using memory
> allocation results without checking for NULL results.

Attached patch is a revised version.
It contains more comments and a bit more optimization.  I think the problems
of checking malloc failure are also fixed.  How about this version?

If it get a passing mark, I'd like to work on the homework from Jakub.


2002-11-27  Isamu Hasegawa  <isamu@yamato.ibm.com>

	* posix/regcomp.c (parse_expression): Set the bit since the back
	reference is used in the regular expression.
	* posix/regex_internal.c (re_node_set_init_1): Make it clean in case
	of malloc failure.
	(re_node_set_init_copy): Likewise.
	* posix/regex_internal.h (state_array_t): New structure.
	(re_sub_match_last_t): Likewise.
	(re_sub_match_top_t): Likewise.
	(re_match_context_t): Add new members.
	(re_dfa_t): Likewise.
	* posix/regexec.c (re_search_internal): Invoke prune_impossible_nodes
	to check the matching is really correct, and retry if failed.
	Move the routin pruning the impossible nodes from here, ...
	(prune_impossible_nodes): To this function.
	(check_matching): Invoke check_subexp_matching_top, and replace
	redundant checking with transit_state_bkref invocation.
	(proceed_next_node): Replace strncmp with memcmp.  Reported by
	Paolo Bonzini  <bonzini@gnu.org>.
	(update_cur_sifted_state): Remove search_subexp invocation.
	(search_subexp): Remove this function.
	(check_dst_limits_calc_pos): Use search_cur_bkref_entry for
	optimization.
	(sift_states_bkref): Use search_cur_bkref_entry for optimization.
	Remove unused invocation of match_ctx_add_entry.
	(transit_state): Invoke check_subexp_matching_top.
	(check_subexp_matching_top): New function.
	(transit_state_bkref): Remove unused array.
	Merge transit_state_bkref_loop.
	(transit_state_bkref_loop): Use get_subexp instead of
	sift_states_backward.  Use search_cur_bkref_entry for optimization.
	Merge this function to transit_state_bkref.
	(get_subexp): New function.
	(get_subexp_sub): Likewise.
	(find_subexp_node): Likewise.
	(check_arrival): Likewise.
	(check_arrival_expand_ecl): Likewise.
	(check_arrival_expand_ecl_sub): Likewise.
	(expand_bkref_cache): Likewise.
	(match_ctx_init): Initialize new members.
	(match_ctx_clean): New function.
	(match_ctx_free): Release new members.
	(match_ctx_free_subtops): New function.
	(match_ctx_add_entry): Fix indent.
	(search_cur_bkref_entry): New function.
	(match_ctx_add_subtop): Likewise.
	(match_ctx_add_sublast): Likewise.

Thanks,
-- 
Isamu Hasegawa
IBM Japan, Ltd.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: patch.021127.gz
Type: application/octet-stream
Size: 12707 bytes
Desc: not available
URL: <http://sourceware.org/pipermail/libc-alpha/attachments/20021127/6d8f1e6b/attachment.obj>