15th April 2008

Attendees

ACTION: Benjamin needs tom to review bug 423692
ACTION: Edwin: raise the priority of fixing string implementation -- The TT string optimization bug is https://bugzilla.mozilla.org/show_bug.cgi?id=411163 DONE
ACTION: Graydon: Per "callgrind" (call-graph profiler derived from "valgrind") - as hotspots are found, log bugs for them with profile data

Moh

Working on performance opportunities w/ tracing. Have prototyped a routine for fast scanning of memory and skipping large blocks of zeros. This routine seems particularly useful for TT's ZCT::PinStackObjects hot loop on a number of benchmarks
- Is there any plan for major changes w/ garbage collection, etc., that would make these new changes irrelevant?
  - Ed says that Garbage collector is stable and they will continue to want to look into this (could go into central as well)
A similar, but slightly different, pattern in the hot loop of _pcre_valid_utf8 also benefits from this optimization and results in 15% to 25% performance gain in Sunspider/string-unpack-code
No person is yet assigned to strings, but Adobe/Mozilla will want to coordinate with Moh on this
- Jim from Mozilla will be working on this when he is back.
CPUID Instruction: Need a flag in AVM core: Ed thinks we should just create a patch and do it. No one is assigned to this yet.

Ed

ABC->Forth generation @runtime, preliminary #s. Ed is working on change to tracking, working on Forth codebase. Interpreter of ABC’s are compiled into IL. This gives us a place to put static optimizations. Thus, the interpreter runs faster as is only keeping track of one location and are optimizing the code before interpreting it.
- Because of this, is working on code for managing the IL cache. It’s a change in body plan for tamarin tracing
- Creates no real change to Forth language and is focused on single pass