Performance/Status Meetings/2008-January-16

Gecko 1.9 Performance and Leaks Meeting Details

Leaks

Lots of leaks with extensions.
- Many look like shut-down leaks. Would still be good to get these fixed so that we know for sure that they are shut-down leaks, and to make it easier to find other leaks. Have been giving these P3 priority.
- Some do not look like shut-down leaks, have been marking these P1 or P2.
- Will test which cause leaks that grow over time and make sure these are given top priority.
Tomcats testing procedure documented here: [1] and leak testing plan here [2]
Peterv has a patch to fix bug in cycle collector that causes many random crashes bug 407034, and a patch to fix one source of hidden-window-leaks which is likely source of many extension leaks bug 412491.
More progress on making DEBUG_CC more useful by dbaron bug 410116. Seems mostly fixed.
We need better tools in non-debug builds that extension authors can use for leak testing. Leak gauge should already work?

AI:sicking Look into enabling trace-refcount, without the tracing part in nightlies.
AI:cbarrett Look into OS leaks on Mac.
AI:dbaron Make assertions fatal on new windows leak box.
AI:sicking Test leaked handles on windows manually using purify.
AI:sayrer Run leak buster while looking for trace-malloc leaks.
AI:sicking to go through and prioritize leak bugs

MMX/SSE/SSE2 patches floating around
- Michael Moy has patch for JPEG color conversion and some other cleanup there

Txul platform differences according to mac mini machines
- ~~mac: slowest on 1.8, slower still on 1.9 (30% regression)~~ Looks even or better than branch
- windows: fastest on 1.8, now in second place (30% regression) -- improving, still slower
- linux: middle on 1.8, now fastest on 1.9 (30% improvement)

massive Ts regression in 2006
- Related to thread manager landing
- Darin will comment in bug with thoughts

Firefox front-end perf
- sayer to put list of tasks on wiki
  - focus on measurements on tabbrowser reduction
  - Performance:FrontEnd
- dolske will do some profiling work

On radar (unowned stuff that people have mentioned have mentioned we should look at at some point)
- Theme optimization
- roc's scrollbar issues
- startup performance (fastload perf, loading libXUL and other files, etc.) (unowned)
- more work to speed up onSecurityChange, but further work will require SSL work and reducing the number of notifications

Profile Guided Optimization
- WPO turned on for windows (bug 361343) -- nice performance win on XP
- Spidermonkey seems to respond well to PGO under GCC (bug 410994)
- sayrer has entire browser building with -fprofile-generate on Linux. no results yet.
- dbaron: shipped fx 1.0 with PGO on linux as well, was a 10% perf win... but compiler upgrade + build system breakage stopped us. We need to restart this on linux (and see about OSX).

bug 407459 -- Hook up jemalloc to our build
- ported to windows, mac and linux. integrated to build system on mac and linux. windows work continuing. Hope to have numbers by end of the week.
bug 403830 -- Content arena
- bug 408720 -- Fallout
- New arenas
Other bugs (should have perf and footprint keywords)
- bug 403354 -- Get rid of nsCStringArray ParseString usage (maybe fixed?)
- bug 407442 -- nsEventListenerManager allocates a lot [patch]
- bug 408113 -- JS reallocate context->stackPool arenas way too often [patch]
- bug 408301 -- XPConnect wrappers w/o XPConnect proto don't share JS scopes with their protos. [patch]
- bug 410250 -- nsID/nsJSID stupid mallocs [patch]
- bug 410357 -- use nsAutoTArray for nsCSSDeclaration::mOrder [patch]
- bug 410392 -- stack allocate gfx objects [patch, needs work]
- bug 408921 -- Arenas size+header are too big for size classification