Re: [WIP v2 1/2] Adding a record-time-zone command option for commit

From: Junio C Hamano <gitster@pobox.com>
Date: Oct 22, 2020, 16:27 UTC
Message-ID: <xmqqd01ans4h.fsf@gitster.c.googlers.com>
In-Reply-To: <xmqqzh4f76jr.fsf@gitster.c.googlers.com>

Junio C Hamano <gitster@pobox.com> writes:

Show 12 quoted lines

> Yes, we could check it in datestamp(), but ... 
>
>> Initially, I thought this would be sufficient to show "-0000" in commit log
>> message. However, I found that the show_date function is used for "decoding";
>> converting timestamp and tz to more readable format. Then I realize the
>> function won't distinguish between +0 and -0 as it only takes in a tz as
>> argument. As a result,...
>
> ... I would have imagined that you do not have to deal with all
> those complications if you don't hook this to such a low level of
> the call graph.  That is why I wondered:
> ...

Let me answer some of my puzzlement myself; that is, I would have understood the change well if it were explained to me this way, and if that explanation matched what the patches did ;-)

The topic has two major parts.

The code that prepares the timestamp to be recorded for the current user, who wants to record an anonymous timezone "-0000", is one (and the easier) part. And this part could be done all inside ident_default_date() without touching anything in date.c; when we need to call datestamp(), we are getting the current time for the current user, so we can mask the timezone.

The other part is that we need to read the timestamp from existing records, and if we choose to distinguish between timestamp in UTC and timestamp with anonymous timezone, we'd need to devise a way to encode the anonymous timezone differently. It is where the extra bit that says "this bit does not usually mean anything but only when the offset (which is a signed integer whose valid range is set to between -2400 to +2400 by date.c::match_tz()) is zero, and this bit is set, the zone is anonymous" comes in.

	Side note.  I suspect the damage to the callchain can be
	limited much narrower if we didn't add this bit throughout
	the API.  What if we instead pick a number outside the valid
	range of offsets, say -10000, as a sentinel value and passed
	that throughout the code when we want an anonymous zone?

	The functions in the callchain that care about the timezone
	must understand how anonymous zone is encoded anyway, so to
	them it's a matter of using an int plus one bit or using an
	int that can have a special value.  But other functions in
	the callchain whose sole purpose (with respect to the
	timezone information) is to pass it between their caller and
	their callee as an opaque piece of data, using just a single
	integer is much less error prone---the patch does not have
	to touch them at all.

Thanks.