Here's the first JS team update newsletter, as suggested by jorendorff.
This first one covers roughly the previous 2 weeks. It seems pretty long
so I'll try for a weekly update.
If I missed any big news, please post it as a reply. Also, in the future
feel free to email me about anything you'd like included in the next
newsletter.
*New Hire*
Marty Rosenberg has joined the JS team to work on ARM compiler
performance. He's starting out by implementing typed array ICs for ARM
and finding opportunities to improve the JM ARM backend. Welcome, Marty!
https://bugzilla.mozilla.org/show_bug.cgi?id=649202
https://bugzilla.mozilla.org/show_bug.cgi?id=674274
*Changes to SpiderMonkey that might affect you*
- Chris Leary redid the option parser (bug 668095). He removed some of
the little-used options while he was at it. If you needed one of them,
file a bug to get it back.
*Project Status Updates*
IonMonkey
- David Anderson updated the IonMonkey feature page with detailed
planning info, including major chunks of work to be done in the near
future.
https://wiki.mozilla.org/Platform/Features/IonMonkey
- Thanks in part to trampolines written by Andy Scheff (bugs 670819 and
670820) , IM recently ran its first program, a tiny program with a &
(bug 670816).
- David Anderson is making progress on bailouts from optimized native
code (bug 670827). Optimized native code is compiled for a certain set
of types for each value. If those assumptions don't hold (e.g., an
integer + overflows when it was predicted not to), then the code must
jump to a handler that can continue execution with the actual types, and
possibly also trigger recompilation. Basic bailouts will take us to the
interpreter.
Debug Object
- Jim Blandy and Jason Orendorff are getting close to landing. Most of
the patches have now been reviewed. Jason blogged about the prompt reviews:
http://blog.mozilla.com/jorendorff/2011/07/25/why-fast-reviews-happen/
GC
- Bill McCloskey is getting close on Incremental GC: write barrier
performance is now good in the benchmarks, and it can be tuned down to
pause times of 1ms or less.
- Bill has also been landing diagnostics for the GC topcrashes (bug
670702), which are among the most common JS topcrashes and are very hard
to debug. The most promising lead so far has been the discovery that
some scripts are getting double-destroyed. Bug 673625 is trying to
collect more data on that.
*Other News*
- Dave Mandelin gave a reprise of the "Know Your Engines" talk at Meebo
last Monday. There are a few people there interested in contributing to
performance tools--JS performance and also DOM performance are major
concerns to them. A couple of new ideas came out of our discussion: an
IC profiler, to show which property access sites are getting IC'd and
why or why not; and a debugging tool to show whether an array is dense
or sparse, because denseness is so important for performance but
currently not directly visible to the user in any way.
- Nick Nethercote continues to improve about:memory, with even more
detail now on JS. He also discovered that often as little as 30% of the
JS heap is live data--the rest is fragmentation (see bug 669611). There
are a few bugs on file to try simple things to improve this (bugs 671702
and 669245), but generational GC and compacting GC are the best
candidates for a definitive fix.
*Key landings not mentioned above*
title: Workers: Make one OS thread and JS runtime per worker, and lose
XPConnect
author: Ben Turner
link: https://bugzilla.mozilla.org/show_bug.cgi?id=649537
notes: I'm not an XPConnect expert, but I guess this greatly
simplifies DOM<->SpiderMonkey communication and removes the need for
XPConnect there. This will help us greatly simplify things in SM as
well: We can now make JSRuntime single-threaded, and that in turn lets
us do compartment-per-global.
title: Update Unicode Table
author: Tom Schuster
link: https://bugzilla.mozilla.org/show_bug.cgi?id=652771
notes: Excellent to see this. Our old Unicode support was derived from
an old JDK, so it was about 10 years old by this point, had some errors,
and didn't support some key concepts like middle-casing. This patch
redid all the Unicode support and (almost completely) fixes our Unicode
compatibility.
title: XPCWrappedNative keys can disappear from WeakMaps
author: Andrew McCreight
link: https://bugzilla.mozilla.org/show_bug.cgi?id=655297
notes: This was blocking work on DOM tree editing in devtools, so it's
great to see it cleared. This bug solves only part of the
problem--followup work will be in bug 673468.
title: Reimplement watchpoints using a JSObject bit
author: Jason Orendorff
link: https://bugzilla.mozilla.org/show_bug.cgi?id=637985
notes: The old watchpoint implementation was tricky and a common source
of regressions, so the simplification here is much appreciated.
title: JM: getelement pic for arguments
author: Tom Schuster
link: https://bugzilla.mozilla.org/show_bug.cgi?id=616744
notes: Makes args[i] fast in JM; no data on speedups, but might help us
a bit with web frameworks (jQuery et. al.), which tend to use |arguments|.
title: Make it more obvious which bits in JSObject::flags are up for grabs
author: Jeff Walden
link: https://bugzilla.mozilla.org/show_bug.cgi?id=673647
notes: jsobj.h now makes it clear exactly which flag bits are in use.
