[Python-checkins] r46146 - sandbox/trunk/rjsh-pybench/pybench.py

[Tim Peters]

[Tim]
...
>> time.clock() has sub-microsecond resolution (better than a millionth
>> of a second) on all flavors of Windows.

[Martin v. Löwis]
> Are you sure about that? It has a *scale* of sub-microseconds (100ns);
> the resolution is unspecified.

"Most" Windows boxes run on HW with a capable real-time counter.

> clock() uses GetSystemTimeAsFileTime(). This likely reads some
> kernel variable, and I very much doubt that this kernel variable
> is updated every 100ns. For example, in ...

Windoes clock() does, yes, but Python's time.clock() on Windows does
not.  See the source ;-)

> ...
> To get a finer time measurement, you should use QueryPerformanceCounter.
> On x86 processors that support it, this uses rdtsc to fetch the
> time-stamping counter:

That's exactly what Python uses for time.clock() on Windows ...
although I see it doesn't on Win64:

#if defined(MS_WINDOWS) && !defined(MS_WIN64) && !defined(__BORLANDC__)
/* Win32 has better clock replacement
   XXX Win64 does not yet, but might when the platform matures. */
#undef HAVE_CLOCK /* We have our own version down below */
#endif /* MS_WINDOWS && !MS_WIN64 */

Do you know whether Win64 still has that limitation?