18th March 2008

Attendees

• Adobe: Edwin, Jennifer, Tom, Scott, Brent, Erik, Rick, Jeff
• Mozilla: Graydon
• Intel: Moh, Shengnan, Marsha

Action Items

• ACTION: moh will look at GCbench test and suggest some optimizations.
• ACTION: tom will see if he can find additional GC/memset tests to recommend to moh and post to test directory. maybe GCheap?
• ACTION: graydon will investigate Google "Summer of Code" agenda

Updates

Moh (Vtune)

• hot loops investigation:
  • testsinglebyte function in stringobject.cpp creates a loop that is causing slow-down
  • shingha has submitted patch for this, which needs to be approved. looks like performance has improved by 50% already.
• rick asks: what are you seeing with memset?
  • moh can have intel look at places where memset appears to be used a lot.
  • per tom, flash player zeroes memory during initialization. if we can improve zeroing time, it improves pause times.
  • ACTION: moh will look at GCbench test and suggest some optimizations.
  • ACTION: tom will see if he can find additional GC/memset tests to recommend to moh and post to test directory. maybe GCheap?
  • moh saw "fastzero" in tamarin-tracing but hasn't had a chance to investigate yet.
  • great places for memory clearing optimizations:
    • GCheap when zeroing a block
    • GCalloc when zeroing a non-RC object
      • freeitem in GCalloc.cpp: RC objects need to be zeroed manually. non-RC objects are zeroed automatically by memset.
• moh asks: on tamarin-central, testsinglebyte slow-down does not appear. why?
  • UTF-8 is used in tamarin-tracing, which causes VM to search variable-length strings.

Jeff

• still working on ECMAScript 4 spec.
• chris peyer is automating the spidermonkey test running ESC on tamarin-central.
  • 96% pass rate initially, so looking good!
  • lars is fixing the biggest bugs
  • eval references would have to be removed to run in AS3, so we're not doing that.
  • chris will post automated tests to buildbot when done.

Tom

• noticed that some mozilla folks are participating in google's "summer of code" -- would be cool to get some tamarin stuff on the agenda.
• ACTION: graydon will investigate

Graydon

• further investigation of mini-tracing
  • running acceptance tests and will post patch this afternoon if all goes well
  • mini-tracing is about 1.5 times faster than doing nothing at all. superwords is a little slower than mini-tracing but a little faster than doing nothing at all.
  • need to define how much code is needed to generate how much speed-up. is it worth it?
  • would be great to have more diverse benchmarks, as it's hard to tell what's noise when performance results are so close.
    • flex apps?
    • a version of flash that loads tamarin-tracing dynamically?
    • if we can get ESC running, we can find testing content from the internet (gmail, etc)
    • these are all a lot of work.
    • run ESC on tamarin-tracing -- easier than the aforementioned ideas.
    • JSbench tests
    • set loop threshold higher