The current implementation does not reliably initialize t0 once.
Additionally the initialization requires two calls to _gettimeofday().
Let's sacrifice a byte to keep the initialization status
and reduce the maximum number of calls to _gettimeofday().
This has caused issues in an application that invokes clock().
The problematic situation is as follows:
1) The program calls clock() which calls _times().
2) _gettimeofday(&t0, 0) puts 0 in t0.tv_usec (because less than 1 us has
elapsed since the beginning of time).
3) _gettimeofday(&t, 0) puts 1 in t.tv_usec (since now more than 1 us has
elapsed since the beginning of time).
4) That call to clock() returns 1 (the value from step 3 minus the value in
step 2).
5) The program does a second call to clock().
6) The code above still sees 0 in t0 so it tries to update t0 again and
_gettimeofday(&t0, 0) puts 1 in t0.tv_usec.
7) The _gettimeofday(&t, 0) puts 1 in t.tv_usec (since less than 1us has
elapsed since step 3).
8) clock() returns 0 (step 7 minus step 6) and indicates that time is
moving backwards.
Signed-off-by: Christoph Muellner <cmuellner@gcc.gnu.org>
clock_t
_times(struct tms *buf)
{
- // when called for the first time, initialize t0
+ static char initialized;
static struct timeval t0;
- if (t0.tv_sec == 0 && t0.tv_usec == 0)
- _gettimeofday (&t0, 0);
-
struct timeval t;
+
_gettimeofday (&t, 0);
+ // when called for the first time, initialize t0
+ if (!initialized) {
+ t0.tv_sec = t.tv_sec;
+ t0.tv_usec = t.tv_usec;
+ initialized = 1;
+ }
+
long long utime = (t.tv_sec - t0.tv_sec) * 1000000 + (t.tv_usec - t0.tv_usec);
buf->tms_utime = utime * CLOCKS_PER_SEC / 1000000;
buf->tms_stime = buf->tms_cstime = buf->tms_cutime = 0;