Re: [PATCH v2 1/3] help: redirect to aliased commands for "git cmd --help"

[Rasmus Villemoes <rv@rasmusvillemoes.dk>]

On 2018-10-03 04:13, Jeff King wrote:
>> +		/*
>> +		 * If we were invoked as "git help cmd", or cmd is an
>> +		 * alias for a shell command, we inform the user what
>> +		 * cmd is an alias for and do nothing else.
>> +		 */
>> +		if (!exclude_guides || alias[0] == '!') {
>> +			printf_ln(_("'%s' is aliased to '%s'"), cmd, alias);
>> +			free(alias);
>> +			exit(0);
>> +		}
> 
> I'm not sure I understand why exclude_guides is relevant. We check it
> below when we know that we _don't_ have an alias. Hrm. I guess you're
> using it here as a proxy for "git foo --help" being used instead of "git
> help foo".

Exactly. Perhaps it's abusing the existing machinery, but I didn't know
how else to distinguish the two cases, and didn't feel like introducing
another way of passing on the exact same information.

> The comment probably needs to spell out that exclude_guides
> is the same as your "we were invoked as...".

Will do. That will also make the string --exclude-guides (i.e., with a
dash) appear in the comment, making it more likely to be found should
anyone change when and how --exclude-guides is implied.

> I wonder if we could change the name of that option. It is an
> undocumented, hidden option that we use internally, so it should be OK
> to do so (or we could always add another one). That might prevent
> somebody in the future from using --exclude-guides in more places and
> breaking your assumption here.

Perhaps, but I think that's better left for a separate patch, if really
necessary even with the expanded comment.

>> +		count = split_cmdline(alias, &argv);
>> +		if (count < 0)
>> +			die(_("bad alias.%s string: %s"), cmd,
>> +			    split_cmdline_strerror(count));
>> +		return alias;
> 
> So we split only to find argv[0] here. But then we don't return it. That
> works because the split is done in place, meaning we must have inserted
> a NUL in alias. That's sufficiently subtle that it might be worth
> spelling it out in a comment.

OK, I actually had precisely

+		/*
+		 * We use split_cmdline() to get the first word of the
+		 * alias, to ensure that we use the same rules as when
+		 * the alias is actually used. split_cmdline()
+		 * modifies alias in-place.
+		 */

in v1, but thought it might be overly verbose. I'll put it back in.

> We don't need to free alias here as we do above, because we're passing
> it back. We should free argv, though, I think (not its elements, just
> the array itself).

Yeah, I thought about this, and removing free(argv) was the last thing I
did before sending v1 - because we were going to leak alias anyway. I'm
happy to put it back in, along with...

> Unfortunately the caller is going to leak our returned "alias", [...] I think it may be OK to overlook
> that and just UNLEAK() it in cmd_help().

...this. Except I'd rather do the UNLEAK in check_git_cmd (the
documentation does say "only from cmd_* functions or their direct
helpers") to make it a more targeted annotation.

Thanks,
Rasmus