[TUHS] Date madness

Random832 random832 at fastmail.com
Thu Dec 14 13:07:35 AEST 2017

On Wed, Dec 13, 2017, at 14:12, Doug McIlroy wrote:
> >  Flipping it to unsigned int was the quickest way out to kick the can until Sun Feb 6 06:28:15 2106.  If you have source it’s incredibly trivial to change, and nothing changes size wise.
> Easy, but perhaps unwise. The sign bit is a potential escape hatch for
> times, just as it is for characters in UTF-8.
> Doug

I'm not sure what you're suggesting - UTF-8 works because it's a
variable length encoding - a variable length timestamp might be
interesting as an academic exercise, but there's no way for time() to
specify how much space is needed, or be informed of how much space is

Anyway, making it use an "unsigned" 32-bit value is a bit less trivial
on V7 (where AIUI there's an unsigned type and a long type, but not
both), but the following works - not tested on actual V7 yet, but works
with the values as int16_t/int32_t on a modern system

         * break initial number into days
        hms = *tim % 86400;
        day = *tim / 86400;
        if(*tim < 0) {
            hms += 109696;
            day += 49709 + hms / 86400;
            hms %= 86400;

47909 days + 109696 seconds = 2^32 seconds. The "extra day" is to ensure
hms is positive.

This is instead of the existing hms < 0 clause

Making V6 do it (no long *or* unsigned type) is a bit more involved, but
I'm close to something viable. Incidentally, one gotcha I've run into is
that the code as-is passes the year minus 1900 to dysize, whereas modern
dysize implementations that respect the 100 and 400 year rules require
the full year.

My current attempt to extend it to more bits without resorting to
floating point relies on the fact that 86400=675*128 to eke out an extra
seven bits without implementing a general purpose division algorithm,
which gets it as far as the year ~10680 (~19391 unsigned). Though
without a long for the day count, V6 can only get to 2149.

More information about the TUHS mailing list