Created attachment 5195 [details]
patch 1

The following assertion in mktime.c:

   verify (long_int_year_and_yday_are_wide_enough,
          INT_MAX <= LONG_MAX / 2 || TIME_T_MAX <= UINT_MAX);

is not satisfiable on any ILP32 architecture with 64-bit time_t.

I propose the following patch to fix the problem.  Given the purpose of this assertion, I've split my patch in 3 to prove that it doesn't introduce any regression.  Here's my semi-formal proof:

First let A be (INT_MAX <= LONG_MAX / 2).

  - Patch #1: This works on the assumption that for any value of LONG_MAX, the whole function either works correctly or fails to compile. With my patch, if a certain condition A is met, every instance of `long' type is replaced with `time_t', and the assertion that checks for LONG_MAX now checks for TIME_T_MAX in its place. It must remain true that for any value of TIME_T_MAX, the whole function either works correctly or fails to compile, regardless of condition A.

  - Patch #2: If condition A is met, clearly the assertion will always suceed, so it can be removed.

  - Patch #3: If condition A is met, the second assertion is skipped, which is equivalent to making A one of its options.

(or in plain English, if longs are too short, time_ts can be used instead)