[Webkit-unassigned] [Bug 158906] REGRESSION(concurrent baseline JIT): Kraken/ai-astar runs 20% slower

[bugzilla-daemon at webkit.org]

https://bugs.webkit.org/show_bug.cgi?id=158906

--- Comment #2 from Filip Pizlo <fpizlo at apple.com> ---
This is fascinating!  Here's what's happening: we enqueued a compile of <global>, which is a ginormous piece of global code, during the benchmark's unmeansured initialization phase.  Then at the time that we try to compile search#, the JIT thread is busy compiling <global>.  That takes a long time to compile and it generates a disgusting amount of code.  <global> takes 16ms to compile and generates 8MB!!

There are a bunch of solutions we could attempt:

1) If we want to compile something and the concurrent JIT thread is busy then compile on the main thread instead.

2) Launch multiple JIT threads.

3) Delay the compilation of <global> based on size heuristics the same way that we do for DFG and FTL tier-up.  This might be hard because I think that <global> actually spends a significant amount of time in its silly loop.

4) Don't compile all of <global> with the baseline JIT.  It tiers up because of a tiny loop at the end.

I think I'll attempt (1).  That seems like the easiest way of recovering this regression.  I think it would be great to consider the other solutions, too.

-- 
You are receiving this mail because:
You are the assignee for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.webkit.org/pipermail/webkit-unassigned/attachments/20160618/d015cc99/attachment.html>