Optimize #at_css and #css initialization #14
Conversation
zyc9012
force-pushed
the
optimize-nl-node-find-init
branch
5 times, most recently
from
21f99b6 to
a3257b6
Compare
zyc9012
force-pushed
the
optimize-nl-node-find-init
branch
from
a3257b6 to
44b899f
Compare
There was a problem hiding this comment.
Thanks @zyc9012 - I like the look of that micro-benchmark!
Some preliminary tests with our scraper parsers shows an average improvement of about 3%, and a small improvement to memory consumption.
I'm not sure we really need these pthread_key_t variables - they can be shared across objects, they can surely also be shared across threads provided they're protected by a mutex, which we have implicitly in the GIL. Can we drop all the pthread stuff and just have:
static lxb_html_parser_t *html_parser = NULL;static lxb_css_parser_t *css_parser = NULL;
static lxb_selectors_t *selectors = NULL;And allocate them on initialization or first use? Letting them leak seems less important if they're just a one-time process-wide allocation.
ext/nokolexbor/nl_node.c
Outdated
| pthread_key_create(&p_key_css_parser, free_css_parser); | ||
| pthread_key_create(&p_key_selectors, free_selectors); |
There was a problem hiding this comment.
These functions are fallible and can fail to allocate a key, which would lead to undefined behaviour in later calls to pthread_get/setspecific.
ext/nokolexbor/nl_node.c
Outdated
| css_parser = lxb_css_parser_create(); | ||
| status = lxb_css_parser_init(css_parser, NULL, NULL); | ||
| if (status != LXB_STATUS_OK) { | ||
| css_parser = NULL; |
There was a problem hiding this comment.
Unless the error was specifically LXB_STATUS_ERROR_OBJECT_IS_NULL this will leak the initial allocation plus any others that happened to succeed within the init call.
ext/nokolexbor/nl_node.c
Outdated
| selectors = lxb_selectors_create(); | ||
| status = lxb_selectors_init(selectors); | ||
| if (status != LXB_STATUS_OK) { | ||
| selectors = NULL; |
There was a problem hiding this comment.
Similarly, if the error was anything other than LXB_STATUS_ERROR_INCOMPLETE_OBJECT, the initial allocation will leak, plus any of the dobject allocations that did succeed within the init function.
ext/nokolexbor/nl_node.c
Outdated
| /* Destroy Selectors object. */ | ||
| (void)lxb_selectors_destroy(selectors, true); | ||
|
|
||
| /* Destroy resources for CSS Parser. */ | ||
| (void)lxb_css_parser_destroy(parser, true); | ||
|
|
There was a problem hiding this comment.
These should be kept for the error handling cases, and also for the #ifndef HAVE_PTHREAD_H case where the destructor call is not done for us at thread exit.
| lxb_status_t status = lxb_html_parser_init(g_parser); | ||
| if (html_parser == NULL) { | ||
| html_parser = lxb_html_parser_create(); | ||
| lxb_status_t status = lxb_html_parser_init(html_parser); | ||
| if (status != LXB_STATUS_OK) { | ||
| nl_raise_lexbor_error(status); |
There was a problem hiding this comment.
Leaks the initial allocation unless status was LXB_STATUS_ERROR_OBJECT_IS_NULL. Otherwise it also leaks other allocations performed by lxb_html_parser_init. The parser object also leaks in the non-HAVE_PTHREAD_H case due to reliance on it calling the destructor.
|
Thanks @Freaky. I've removed all pthread stuff and made sure that it's thread-safe on truffleruby. |
Before
After