It appears that InlineGetProp in js/src/methodjit/StubCalls.cpp is not being inlined on Mac x64 builds. When I profile part of Dromaeo DOM using Instruments (the XCode 4 equivalent of Shark), InlineGetProp, called by js::mjit::stubs::GetProp, shows up in the profile. I presume from the name and its usage that we want this function to be inlined? Should be easy enough to force, at least on gcc.
(In reply to comment #0) > I presume from the name and its usage that we want this function to be > inlined? Should be easy enough to force, at least on gcc. In general it's hard to predict whether inlining a big function will matter. But it's worth a try to see if it helps.
It seems to help a bit on the Dromaeo DOM microbenchmark element.property, which does: var ret; var elem = document.getElementById('some-element'); time this operation in a loop: ret = elem.id; Currently scores: 469 467 473 480 472 471 With inlining: 493 497 491 492 490 336 (Assuming the 336 is an outlier...)
Created attachment 536374 [details] [diff] [review] Patch v1
The units on the numbers from comment 2 are whatever Dromaeo uses -- I think it's runs/s.
I'm not familiar with the process for landing JS changes. Do I need to land on TM, or can I land directly on MC?
(In reply to comment #5) > I'm not familiar with the process for landing JS changes. Do I need to land > on TM, or can I land directly on MC? Either is acceptable, but TM is generally preferred for JS changes.
This was pushed in some TM merge; closing.