Hi Bruno, Le Thu, 05 Jul 2018 23:17:26 +0200, Bruno Haible <br...@clisp.org> a écrit :
> Albert ARIBAUD wrote: > > I was under the impression that you wanted the > > 64-bit-time stuff to go in gnulib before it went in glibc, so I don't > > get what the "once glibc has such a macro" means. Can you elaborate on > > what you had in mind? > > I can't speak for Paul, but for me the sequence of steps that produces the > desired result with the least effort would be: > > 1) glibc implements the 'time_t' type that depends on the value _TIME_BITS > defined at preprocessor level, like you described. Including support for > 'gettimeofday', 'stat', 'fstat' and the like. > While doing so, pay attention that the implementation of mktime, > strftime, > strptime, etc. can be compiled with a 32-bit time_t or a 64-bit time_t. > > 2) gnulib modifies its year2038 module to define _TIME_BITS to 64 at > configure > time, on platforms where glibc supports it. > > 3) During the next source-code sync from glibc to gnulib, involving mktime.c > etc., the gnulib people (likely Paul) make sure that this source code can > still be used on non-glibc platforms with either 32-bit time_t or 64-bit > time_t. Usually this involves a couple of #ifs and conditional #includes. > These changes can then flow back into glibc. They won't have a functional > effect in glibc, therefore won't break the atomicity of step 1. > > AFAICS, steps 3 could also be executed before step 2. > > Bruno Thanks for the detailed answer(s, including your previous one). Do code syncs from glibc to gnulib happen on new glibc releases? If they do, then my chick-and-egg problem remains (unless glibc accepts pulling in only my glibc patch which add __time_64_t without changing the public API, but that would amount to dead code. Cordialement, Albert ARIBAUD 3ADEV
