The submillisecond test was broken in at least two ways. First, the...

The submillisecond test was broken in at least two ways.  First, the
high resolution clock is intentionally disabled on some systems (old AMDs). 
If QueryPerformanceCounter doesn't work on this system, we shouldn't run the
test.

Second, however, if the time between two HighResNow() calls is *always* 0us,
then this test would fail.  Due to speedstep technology with intentionally
underclocked QPC (at the windows level), this is quite possible.

BUG=42850
TEST=TimeTicks.SubMillisecondTimers

[Paweł Hajdan Jr., 2010-09-22 09:35:31]

http://codereview.chromium.org/3387011/diff/1/4
File base/time_win_unittest.cc (right):

http://codereview.chromium.org/3387011/diff/1/4#newcode120
base/time_win_unittest.cc:120: // Run kRetries attempts to see a sub-millisecond
timer.
This test has still potential for flakiness. What if we tried to gather some
samples (1000? 10000?) and see if their statistical properties match
expectations (are within limits)? I think that should make the test more solid,
but I'm not really sure. What do you think?

http://codereview.chromium.org/3387011/diff/1/4#newcode124
base/time_win_unittest.cc:124: // Spin until the clock changes a little
nit: Dot at the end, and explain why.

[Mike Belshe, 2010-09-22 21:58:48]

http://codereview.chromium.org/3387011/diff/1/4
File base/time_win_unittest.cc (right):

http://codereview.chromium.org/3387011/diff/1/4#newcode120
base/time_win_unittest.cc:120: // Run kRetries attempts to see a sub-millisecond
timer.
On 2010/09/22 09:35:32, Paweł Hajdan Jr. wrote:
> This test has still potential for flakiness. What if we tried to gather some
> samples (1000? 10000?) and see if their statistical properties match
> expectations (are within limits)? I think that should make the test more
solid,
> but I'm not really sure. What do you think?

Unfortunately, I don't think it helps.  In this case, we run 1000 samples
looking for just *one* that clocks in sub-millisecond.  If we don't get one
within 1000 tries, we give up.  I can't think of a way that statistical sampling
would work when just needing one doesn't.

http://codereview.chromium.org/3387011/diff/1/4#newcode124
base/time_win_unittest.cc:124: // Spin until the clock changes a little
On 2010/09/22 09:35:32, Paweł Hajdan Jr. wrote:
> nit: Dot at the end, and explain why.

Done.

[Paweł Hajdan Jr., 2010-09-23 11:09:29]

LGTM

Please upload an updated patchset so I can also review the updated test comment.
I'm not letting the review depending on that because of the TZ difference.