Re: [PATCH v4 2/5] notes.c: replace `git_config()` with `git_config_get_value_multi()`

[Matthieu Moy <Matthieu.Moy@grenoble-inp.fr>]

Tanay Abhra <tanayabh@gmail.com> writes:

> On 7/30/2014 7:43 PM, Matthieu Moy wrote:
>> Tanay Abhra <tanayabh@gmail.com> writes:
>> 
>>> -	git_config(notes_display_config, &load_config_refs);
>>> +	if (load_config_refs) {
>>> +		values = git_config_get_value_multi("notes.displayref");
>>> +		if (values) {
>>> +			for (i = 0; i < values->nr; i++) {
>>> +				if (!values->items[i].string)
>>> +					config_error_nonbool("notes.displayref");
>>> +				else
>>> +					string_list_add_refs_by_glob(&display_notes_refs,
>>> +								     values->items[i].string);
>>> +			}
>>> +		}
>>> +	}
>> 
>> It seems to me that you're doing a lot here that should have been done
>> once in the config API:
>> 
>> * if (values) {
>>           for (i = 0; i < values->nr
>> 
>>   => We could avoid the "if" statement if git_config_get_value_multi was
>>   always returning a string_list, possibly empty (values->nr == 0
>>   instead of values == NULL).
>>
>
> or we can do something like,
>
> 	if (!git_config_get_value_multi("notes.displayref", &values)) {
> 		/* return 0 if there is a value_list for the key */
>
>>   Not as obvious as it seems, because you normally return a pointer to
>>   the string_list that is already in the hashmap, so you can't just
>>   malloc() an empty one if you don't want to leak it.
>> 
>>   Another option would be to provide an iterator that would call a
>>   function on each value of the list, and do nothing when there's no
>>   list at all (back to the callback-style API, but you would iterate
>>   only through the values for the right key).
>>
>
> This is also a good idea, but still we are back to the callback API,
> and what we are gaining is fewer loop iterations than git_config().

Regardless of performance, the code would also be a bit shorter, since
the callback just gets the values for the right key, so it doesn't need
to re-test that the key is the right one.

Here, the callback would basically be the body of the for loop above.

> Which way do you prefer, a reroll is easy but Junio might have been sick
> of replacing the patches in pu by now. :)

No need to replace anything, you can add new helpers on top of the
existing.

Do it the way you feel is better, I'm just giving ideas.

>> * if (!values->items[i].string)
>>           config_error_nonbool(
>> 
>>   => This check could be done once and for all in a function, say
>>   git_config_get_value_multi_nonbool, a trivial wrapper around
>>   git_config_get_value_multi like
>> 
>> const struct string_list *git_configset_get_value_multi_nonbool(struct config_set *cs, const char *key)
>> {
>> 	struct string_list l = git_configset_get_value_multi(cs, key);
>>         // possibly if(l) depending on the point above.
>> 	for (i = 0; i < values->nr; i++) {
>> 		if (!values->items[i].string)
>> 			git_config_die(key);
>> 	}
>> 	return l;
>> }
>>
>
> Not worth it, most the multi value calls do not die on a nonbool.

Can you cite some multi-value variables that can be nonbool? I can't
find many multi-valued variables, and I can't find any which would allow
bool and nonbool.

>> const struct string_list *git_config_get_value_multi_nonbool(const char *key)
>> {
>> 	git_config_check_init();
>> 	return git_configset_get_value_multi_nonbool(&the_config_set, key);
>> }
>> 
>> 
>>   (totally untested)
>> 
>>   BTW, is it intentional that you call config_error_nonbool() without
>>   die-ing?
>>
>
> Yup, it's intentional, original code didn't die for empty values, and it seemed
> logical to me to emulate that over to the rewrite.

The old code was doing

	if (*load_refs && !strcmp(k, "notes.displayref")) {
		if (!v)
			config_error_nonbool(k);
		string_list_add_refs_by_glob(&display_notes_refs, v);

It seems that the intent of the programmer was

	if (*load_refs && !strcmp(k, "notes.displayref")) {
		if (!v)
			return config_error_nonbool(k); // <---------------
		string_list_add_refs_by_glob(&display_notes_refs, v);

At least, that would explain why the code uses v even after testing that
it is a NULL pointer.

You're already fixing a bug in your patch by not using NULL values, but
then I don't see any reason to keep the old weird behavior (display an
error but do not die).

-- 
Matthieu Moy
http://www-verimag.imag.fr/~moy/