Paul Eggert wrote:
> > What shall we do with this one? Override gettime_res() to return 1 ns
> > instead of 5 ns? Or adapt the unit test?
>
> Perhaps overriding is best. I'm a bit puzzled why the current code
> doesn't work, though.
It doesn't work because it takes only 32 samples of current_timespec ().
In order to work reliably it would need 10 million samples. Which is,
of course, not feasible :-)
I'm committing this override:
2024-03-26 Bruno Haible <br...@clisp.org>
gettime-res: Fix test failure on Solaris 11.4/SPARC.
* lib/gettime-res.c (gettime_res): On Solaris/SPARC, just return 1 ns.
diff --git a/lib/gettime-res.c b/lib/gettime-res.c
index afc828a5fb..8bde4122a7 100644
--- a/lib/gettime-res.c
+++ b/lib/gettime-res.c
@@ -38,16 +38,28 @@ gcd (long int a, long int b)
long int
gettime_res (void)
{
+#if defined __sun && defined __sparc
+ /* On Solaris 11.4/SPARC (with a SPARC-M8 CPU)
+ clock_getres (CLOCK_REALTIME, ...) returns a resolution of 1000000 ns,
+ and clock_gettime (CLOCK_REALTIME) returns a multiple of 5 ns with a
+ probability of ca. 1 - 1/300000. Thus the heuristic below with the
+ 32 samples guesses a resolution of 5 ns with a probability of ca.
+ 1 - 1/10000. But occasionally clock_gettime (CLOCK_REALTIME) returns
+ only a multiple of 1 ns.
+ Simple guesswork is not enough to cope with this irregular behaviour.
+ Therefore, here is an override. */
+ return 1;
+#else
struct timespec res;
-#if defined CLOCK_REALTIME && HAVE_CLOCK_GETRES
+# if defined CLOCK_REALTIME && HAVE_CLOCK_GETRES
clock_getres (CLOCK_REALTIME, &res);
-#elif defined HAVE_TIMESPEC_GETRES
+# elif defined HAVE_TIMESPEC_GETRES
timespec_getres (&res, TIME_UTC);
-#else
+# else
/* Guess high and let the later code deduce better. */
res.tv_sec = 1;
res.tv_nsec = 0;
-#endif
+# endif
/* On all Gnulib platforms the following calculations do not overflow. */
@@ -83,6 +95,7 @@ gettime_res (void)
}
return r;
+#endif
}
/*
