libast/tm: DST offset and contemporaneous time zone fixes.

[pghvlaans]

RE #976, 98800e9

Pre-u+m code in tminit.c sets the standard (ST) and daylight savings time (DST) offsets by checking times over the previous year of the time requested to compare DST status and offsets from UTC. If no offset change is detected, the DST offset is set equal to the ST offset.

Although assigning the DST offset in this way is counter-intuitive, it does not result in incorrect times being reported so long as two implicit assumptions are met:

• If the requested time is DST, a change from ST occurred in the previous year.
• If the offset changed over the previous year, this was the result of a DST transition.

The first assumption fails on musl Linux systems (which would not appear until well after the code in question was written). musl interprets POSIX timezone specifications with a DST name but without rules as being DST year-round. tminit.c therefore calculates both the DST and ST offsets as equal to the overall UTC offset, which is applied twice when reporting times. Because musl apparently disregards the top zoneinfo directory, even common aliases such as PST8PDT return wildly inaccurate results (rather than the one-hour errors that /bin/date would suggest on musl systems).

The second assumption is broken at all times and places between a historical timezone change and a DST shift or one year later, whichever comes first. Although Pacific/Apia between 2011-12-31 and 2012-04-01 yields the most striking results, there are many other examples.

Fixing these problems involves relatively minor changes to tminit.c and related tests:

• In tminit.c, save and restore the UTC offset and DST status of the requested time; only break offset calculations for a DST shift.
• In attributes.sh, use full POSIX timezone specifications in lieu of EST5EDT and PST8PDT. This is necessary because musl applies DST for the entire year and the time of the test is ST.
• In printf.sh, add several TZ tests to account for different combinations of DST and ST designations across historical timezone changes. Also add an EST5EDT test in DST.

In addition, although 98800e9 introduced a fix for looking up times in zones with historical changes, it applied only to searches using seconds from the epoch. Other lookup formats still fail to change to contemporaneous zones. A convenient reproducer is:

TZ=Europe/London printf "%(%s)T\n" "1970-01-01 01:00:00"

The output should be 0 and not 3600.

• In tmxdate.c, after running through tmxdate for the first time, compare the values of west, dst and isdst to ensure that a proper zone has been chosen. If the zones differ, this means that a historical change has occurred. Apply an offset and run again, repeating as necessary.
• In printf.sh, add 'reverse' tests looking up epoch seconds by date.

[pghvlaans]

Tested on Slackware64 -current (glibc), Void (musl), FreeBSD, OpenBSD, omnios (attributes and printf only) and NetBSD. I don't currently have a macOS VM.

[pghvlaans]

8b003db broadens the scope a bit. I can revert it and make a different PR after this one is merged if that would be more convenient.

[McDutchie]

Many thanks for this solidly executed PR.

I don't think there's a real need to split the PR; your commit message does a nice job of explaining every distinct change. If it becomes necessary, we can always go back to the commit history of this PR to pick it apart further.

Unfortunately I don't have the time and inclination to become expert enough at the tm(3) code myself to give this a proper review, but your explanations do make sense to me. I've tested the PR and everything seems to work as described. There is good regression test coverage.

So, in it goes.

One thing I will edit first, though, is some of those one-letter variable names, to change them to more descriptive names. I'm really not a fan of that aspect of Glenn Fowler's coding style; it badly reduces the legibility of the code.

[McDutchie]

Changed a couple of things:

• tminit.c: Rename one-letter variables for legibility.
• tmxdate.c: Move the check for historical time zone changes to the end of tmxdate. This allows eliminating the first and good_check variables.

Does that look OK to you?

[pghvlaans]

Yes, very sensible.

[pghvlaans]

Actually, one very small point about the NEWS entry. The contemporaneous timezone issue wasn't musl-only.

[McDutchie]

Fixed, thanks.