MinGW localtime_r works in one time zone, fails in another

Question

I have the following file test.c:

 #define _POSIX_THREAD_SAFE_FUNCTIONS
 #include <time.h>
 #include <stdio.h>
 #include <string.h>
 #include <stdint.h>
 #include <inttypes.h>

 int main(int argc,char**argv) {
    struct tm t1, t2, t3;
    time_t w1, w2, w3;
    memset(&t1,0,sizeof(struct tm));
    memset(&t2,0,sizeof(struct tm));
    memset(&t3,0,sizeof(struct tm));
    w1 = 0;
    errno = 0;
    localtime_r(&w1,&t1);
    printf("localtime_r: errno=%d\n",errno);
    errno = 0;
    w2 = mktime(&t1);
    printf("mktime: errno=%d result=%" PRId64 "\n",errno,((int64_t)w2));
    errno = 0;
    localtime_r(&w2,&t2);
    printf("localtime_r: errno=%d\n",errno);
    errno = 0;
    w3 = mktime(&t2);
    printf("mktime: errno=%d result=%" PRId64 "\n",errno,((int64_t)w3));
    errno = 0;
    localtime_r(&w3,&t3);
    printf("localtime_r: errno=%d\n",errno);
    printf("sizeof(time_t)=%" PRId64 "\n", ((int64_t)sizeof(time_t)));
    printf("W1=%" PRId64 " W2=%" PRId64 " W3=%" PRId64 "\n",((int64_t)w1),((int64_t)w2),((int64_t)w3));
    printf("Y1=%d Y2=%d Y3=%d\n",t1.tm_year,t2.tm_year,t3.tm_year);
    return 0;
 }

I compile it like this:

i686-w64-mingw32-gcc -D__MINGW_USE_VC2005_COMPAT=1 -o test.exe test.c

Note, i686-w64-mingw32-gcc --version reports 8.3-win32 20190406 This is running in a Docker image of Ubuntu 19.04, using the MinGW version that comes with Ubuntu 19.04 (it says version 6.0.0-3).

I have a Windows 10 VM (Version 1809 OS Build 17763.379). By default, time zone is set to US Pacific Time (UTC-8). I copy test.exe to this VM and run it there.

It prints:

localtime_r: errno=0
mktime: errno=0 result=0
localtime_r: errno=0
mktime: errno=0 result=0
localtime_r: errno=0
sizeof(time_t)=8
W1=0 W2=0 W3=0
Y1=69 Y2=69 Y3=69

That's the expected result. (At UTC midnight on 1 Jan 1970, it was still 1969 in UTC-8.)

I change the Windows time zone to UTC+10 (Canberra, Melbourne, Sydney). Run it again. It prints:

localtime_r: errno=0
mktime: errno=0 result=47244640256
localtime_r: errno=22
mktime: errno=22 result=4294967295
localtime_r: errno=0
sizeof(time_t)=8
W1=0 W2=47244640256 W3=4294967295
Y1=70 Y2=-1 Y3=206

It seems the mktime() call is returning an invalid value in UTC+10 time zone, but returns the correct value of 0 in UTC-8 time zone.

Why does this code work in one timezone break in another?

Note, this is only a problem with -D__MINGW_USE_VC2005_COMPAT=1 to enable 64-bit time_t. If I leave that out, which means 32-bit time_t, then the code works in both timezones. (But, 32-bit time_t is not a good idea, because it breaks in the year 2038, and that's less than twenty years away now.)

Answer 1

I worked out the cause of the problem. Sander De Dycker's suggestion, that mktime is returning a 32-bit value, is correct.

The problem is basically this: the MSVCRT defines three mktime functions: _mktime32 for 32-bit time_t, _mktime64 for 64-bit time_t, and _mktime which is a legacy alias for _mktime32.

_mingw.h does a #define _USE_32BIT_TIME_T in 32-bit code unless you #define __MINGW_USE_VC2005_COMPAT to disable that. Once you have #define __MINGW_USE_VC2005_COMPAT, then localtime_s is defined as an inline function which calls _localtime64_s. And #define _POSIX_THREAD_SAFE_FUNCTIONS defines localtime_r as an inline function which calls localtime_s. However, mktime is still 32-bit. To get 64-bit mktime, you need to also #define __MSVCRT_VERSION__ 0x1400 (or higher). Once you do that, mktime becomes an inline function which calls _mktime64. Before that, mktime is a normal function declaration which is linked to the legacy 32-bit mktime.

So #define __MINGW_USE_VC2005_COMPAT 1 without #define __MSVCRT_VERSION__ 0x1400 (or -D equivalent) gives you a localtime_r with 64-bit time_t, but a mktime with 32-bit time_t, which obviously won't work. Even worse than that, the actual implementation of the mktime symbol is returning a 32-bit time_t, but the function declaration is for a 64-bit time_t, which is what causes the junk in the upper 32-bits.

As to the difference behaviour in different time zones, I don't have a complete explanation for that, but I think the reason is likely as follows: when you have a function which actually returns a 32-bit value but is incorrectly being defined to return a 64-bit value, the upper 32-bits of the return value will hold random junk data left over from previous calculations. So, any difference in the previous calculations, or slightly different code paths, may result in different random junk. With a UTC-8 timezone, for whatever reason, the random junk is coincidentally zero, so the code (despite its incorrectness) actually works. With a UTC+10 timezone, the random junk turns out to be non-zero, which causes the rest of the code to stop working.