Bug 113624 - FTL should support double variables
Summary: FTL should support double variables
Status: RESOLVED FIXED
Alias: None
Product: WebKit
Classification: Unclassified
Component: JavaScriptCore (show other bugs)
Version: 528+ (Nightly build)
Hardware: All All
: P2 Normal
Assignee: Filip Pizlo
URL:
Keywords:
Depends on:
Blocks: 112840
  Show dependency tree / graph
 
Reported: 2013-03-29 16:15 PDT by Filip Pizlo
Modified: 2013-04-28 18:04 PDT (History)
7 users (show)

See Also:

Attachments
the patch (32.82 KB, patch)
2013-04-27 15:50 PDT, Filip Pizlo 		no flags Details | Formatted Diff | Diff
the patch (32.98 KB, patch)
2013-04-27 16:12 PDT, Filip Pizlo 		ggaren: review+
Details | Formatted Diff | Diff
latest MCJIT patch (18.60 KB, patch)
2013-04-28 17:37 PDT, Filip Pizlo 		no flags Details | Formatted Diff | Diff
Show Obsolete (1) View All Add an attachment (proposed patch, testcase, etc.)

Note You need to log in before you can comment on or make changes to this bug.
Description Filip Pizlo 2013-03-29 16:15:52 PDT 
FTL should support double variables
Comment 1 Filip Pizlo 2013-04-27 15:50:51 PDT 
Created attachment 199919 [details]
the patch

Still need to test it more, so not ready for review.
Comment 2 Filip Pizlo 2013-04-27 15:59:03 PDT 
Heh.  Looping over an array to sum up some doubles just got 50% faster.


[pizlo@dethklok OpenSource] DYLD_FRAMEWORK_PATH=WebKitBuild/Release/ WebKitBuild/Release/jsc --useExperimentalFTL=false test10.js
Warmup: 1000 iterations took 3.915071487426758 ms.
Warmup: 10000 iterations took 33.653974533081055 ms.
Measurement: 100000 iterations took 329.6809196472168 ms.
[pizlo@dethklok OpenSource] DYLD_FRAMEWORK_PATH=WebKitBuild/Release/ WebKitBuild/Release/jsc --useExperimentalFTL=true test10.js
Warmup: 1000 iterations took 8.111000061035156 ms.
Warmup: 10000 iterations took 22.084951400756836 ms.
Measurement: 100000 iterations took 220.31879425048828 ms.
Comment 3 Filip Pizlo 2013-04-27 16:12:08 PDT 
Created attachment 199921 [details]
the patch

Note, for this to work on Darwin, I'll need to fix LLVM's MCJIT on Darwin.  Lol.
Comment 4 Geoffrey Garen 2013-04-27 17:44:01 PDT 
Comment on attachment 199921 [details]
the patch

r=me

Now all we need is for Darwin to become a supported platform for LLVM...
Comment 5 Filip Pizlo 2013-04-28 00:07:58 PDT 
(In reply to comment #4)
> (From update of attachment 199921 [details])
> r=me
> 
> Now all we need is for Darwin to become a supported platform for LLVM...

Lol.

I will wait with landing this until I have that figured out.  I've got a backlog of concurrent DFG stuff to do anyway.
Comment 6 Filip Pizlo 2013-04-28 17:36:27 PDT 
(In reply to comment #5)
> (In reply to comment #4)
> > (From update of attachment 199921 [details] [details])
> > r=me
> > 
> > Now all we need is for Darwin to become a supported platform for LLVM...
> 
> Lol.
> 
> I will wait with landing this until I have that figured out.  I've got a backlog of concurrent DFG stuff to do anyway.

And I "fixed" it.  All that is needed is to switch the MCJIT to CodeModel::Small.  This required exposing more things via the C API; I'll attach the patch here and try to land it in LLVM trunk soon.
Comment 7 Filip Pizlo 2013-04-28 17:37:49 PDT 
Created attachment 199976 [details]
latest MCJIT patch

This includes my previous MCJIT patch since it's still in a land-rollout pattern. ;-)
Comment 8 Filip Pizlo 2013-04-28 18:04:15 PDT 
Landed in http://trac.webkit.org/changeset/149264