[Webkit-unassigned] [Bug 178389] Enable tier-up in loops created by recursive tail call optimizations.

[bugzilla-daemon at webkit.org]

https://bugs.webkit.org/show_bug.cgi?id=178389

Robin Morisset <rmorisset at apple.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|webkit-unassigned at lists.web |rmorisset at apple.com
                   |kit.org                     |

--- Comment #1 from Robin Morisset <rmorisset at apple.com> ---
I ran some tests between no optimization, optimization in both DFG and FTL, and optimization only in FTL:
                         Disabled                   Normal                   FTLOnly               FTLOnly v. Disabled    

n-body              679.4277+-12.1793    ^    608.0302+-17.0313    ^    556.8847+-9.6934        ^ definitely 1.2201x faster
merge-sort         1056.9295+-13.0481        1025.3372+-19.3403    ^    764.4767+-28.6580       ^ definitely 1.3826x faster

<geometric>         847.3528+-10.6054    ^    789.3578+-12.1466    ^    652.3523+-16.1727       ^ definitely 1.2989x faster

Two conclusions:
- the win between no optimization and unconditional optimization is a lot smaller than what I had measured earlier (I had got about 20% in a less rigorous benchmark, before I got TailBench supported into the run-jsc-benchmarks script)
- there is clearly a huge gain in only running it in FTL mode, which confirms our suspicion that the optimization is preventing tier-up from DFG to FTL.

-- 
You are receiving this mail because:
You are the assignee for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.webkit.org/pipermail/webkit-unassigned/attachments/20171023/3fa47384/attachment.html>