On Wed, 23 Dec 2009 09:54:20 -0500, grauzone <[email protected]> wrote:
Andrei Alexandrescu wrote:
grauzone wrote:
Steven Schveighoffer wrote:
All,
I created a new patched druntime that prevents array stomping and at
the same time increases append performance for thread-local array
appending.
The patch is attached to bugzilla entry 3637. (
http://d.puremagic.com/issues/show_bug.cgi?id=3637 )
Nobody cares about it? (Except someone commenting in bugzilla, but
where's Andrei and Walter?)
The entire Phobos team discussed this change very thoroughly. We are
all very excited about it.
Did anyone try it and make initial tests?
I did (obviously). I tried both a simple "append to N arrays in a loop"
test and also the affinity test posted by dsimcha in an earlier post. The
new code beats the old runtime in both tests *and* performs better on the
affinity test with multiple cores than it does with a single core. Tests
were done on linux. I realize testing my own code doesn't mean much, but
I think the patch is close to being assimilated, and everyone can try it
out at that point. I'd be very interested to see if others can find holes
in the design.
I don't have the numbers in front of me right now, but next week, I'll try
to post some.
AFAIK, only Sean tried to get it working, but had trouble. I was mostly
concerned with if the patch *worked* on Windows and OSX, since I didn't
have a mac or a Windows box to try it out. Since then, I've successfully
built and tested both OSX and Windows.
-Steve
