[Bug 39521] New: Calling { SetEvent(h); ResetEvent(h); } takes 36000 machine cycles without contention

[wine-bugs at winehq.org]

https://bugs.winehq.org/show_bug.cgi?id=39521

            Bug ID: 39521
           Summary: Calling { SetEvent(h); ResetEvent(h); } takes 36000
                    machine cycles without contention
           Product: Wine
           Version: unspecified
          Hardware: x86-64
                OS: Linux
            Status: UNCONFIRMED
          Severity: major
          Priority: P2
         Component: kernel32
          Assignee: wine-bugs at winehq.org
          Reporter: chris.cochran.protoncore at gmail.com
      Distribution: ---

This bug may be related to Bug #36692, but this one does not involve any calls
to WaitForSingleObject( ), so I am reporting it as a new bug.

A Windows 32-bit bench test, built with VC++ 2010 in Windows 7, designed to
measure the machine cycles of a wide range of low-level CPU and system
operations, reliably demonstrates that the sequence { SetEvent(h); 
ResetEvent(h); } costs a minimum of about 36000 machine cycles, when run on
Wine 1.7.50 over 64-bit Linux Mint 17.2.  This is an isolated single-threaded
test, without other threads involved and no contention or waiting for the event
object.  It operates on a previously created Windows Event object, from an
earlier h = CreateEvent(0,1,0,0) call.  The CPUs running this test include both
i7 930 and i7 980.  All systems are running normally and without any observed
problems.  All the other bench tests show Wine behavior to be similar and on
par with normal Windows 7 results.

The bench test operates by calling the above sequence in multiple passes of
thousands of repetitions per pass, and selects the minimum time for its
reported result.  On Wine, the result for a single repetition is about 36000
machine cycles, but the exact same 32-bit binary run on 64-bit Windows 7 shows
about 1350 machine cycles for the same test.  This result is highly
reproducible with very little variance in both 32-bit and 64-bit builds.  

I control 100% of the C++ code running this test (one .EXE and one .DLL), but
the above result should be easy to reproduce solely from the details I have
provided here.  I can provide the binary upon request.  If this bug is related
to Bug #36692, the absence of any WaitForSingleObject( ) calls should make this
Bug easier to independently test and demonstrate than the #36692 prescription.

This bug fulfills the definition of a "major bug", because it most likely
adversely affects response latency in a large majority of asynchronous
event-driven processes developed for Windows (e.g. like frame-rates), and there
are no practical alternatives for avoiding this problem.

-- 
Do not reply to this email, post in Bugzilla using the
above URL to reply.
You are receiving this mail because:
You are watching all bug changes.