On Tue, Jun 10, 2014 at 12:08 AM, Sebastian Lackner sebastian@fds-team.de wrote:
Thanks for all the feedback @ Ken, Daniel, Bruno and all the other people I talked to.
Changes in try 3:
- Although the fd_cache_epoch was slightly faster in my tests (-> see
plots in mail for try 1) and should also work on PPC, its probably better to choose a method which is easier to verify, and where we can be sure, that it doesn't break with the next gcc release. ;) In this patch I switched back to the old approach using interlocked_cmpxchg64 to set/get both fd and associated attributes at the same time.
- Although used in some other places, its probably better to avoid
assert() for stuff, which should never be optimized away.
Just an update to my tests: http://influenza.blog.br/wine/testpatch3.pdf
As expected it' slightly slower (Try 3 column), but still much better than the current implementation and still under windows 7 measures in my machine.
Best wishes, Bruno
-
Bruno Jesus