Bug 118749 - fourthTier: DFG should do a high-level LICM before going to FTL
Summary: fourthTier: DFG should do a high-level LICM before going to FTL
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: 118750 118774 118778 118835 118866 118880 118910 118946 118956 118957
Blocks: 112840
  Show dependency tree / graph
 
Reported: 2013-07-16 14:51 PDT by Filip Pizlo
Modified: 2013-07-22 22:39 PDT (History)
7 users (show)

See Also:

Attachments
WRONG PATCH (13.13 KB, patch)
2013-07-16 21:31 PDT, Filip Pizlo 		no flags Details | Formatted Diff | Diff
work in progress (12.38 KB, patch)
2013-07-21 15:29 PDT, Filip Pizlo 		no flags Details | Formatted Diff | Diff
it might work (22.54 KB, patch)
2013-07-21 16:20 PDT, Filip Pizlo 		no flags Details | Formatted Diff | Diff
it just motioned some loop-invariant code (39.12 KB, patch)
2013-07-21 21:57 PDT, Filip Pizlo 		no flags Details | Formatted Diff | Diff
fixed some stuff (40.04 KB, patch)
2013-07-21 22:12 PDT, Filip Pizlo 		no flags Details | Formatted Diff | Diff
the patch (47.59 KB, patch)
2013-07-22 11:45 PDT, Filip Pizlo 		no flags Details | Formatted Diff | Diff
the patch (45.47 KB, patch)
2013-07-22 11:47 PDT, Filip Pizlo 		no flags Details | Formatted Diff | Diff
the patch (57.37 KB, patch)
2013-07-22 15:22 PDT, Filip Pizlo 		oliver: review+
Details | Formatted Diff | Diff
Show Obsolete (7) 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 2013-07-16 14:51:05 PDT 
Fun!
Comment 1 Filip Pizlo 2013-07-16 21:31:50 PDT 
Created attachment 206847 [details]
WRONG PATCH
Comment 2 Filip Pizlo 2013-07-19 18:48:19 PDT 
Some notes:

- We cannot hoist, or sink, anything that has effects. This means that the easiest way of indicating that something cannot be hoisted is to claim that it side-effects some miscellaneous thing.

- Whenever we hoist, we must replace the original node with a Phantom.

- We cannot hoist forward-exiting nodes without some additional effort.  I believe that what it comes down to is that forward-exiting generally have their NodeExitsForward cleared upon hoist, except for forward-exiting nodes that take bogus state as their input.  Those are substantially harder. :-/
Comment 3 Filip Pizlo 2013-07-21 15:29:25 PDT 
Created attachment 207215 [details]
work in progress
Comment 4 Filip Pizlo 2013-07-21 16:20:56 PDT 
Created attachment 207218 [details]
it might work

But I can't test it because the prerequisites haven't landed yet.
Comment 5 Filip Pizlo 2013-07-21 20:37:00 PDT 
Comment on attachment 207218 [details]
it might work

I didn't mean to set r?, since I haven't even compiled this yet.
Comment 6 Filip Pizlo 2013-07-21 21:57:03 PDT 
Created attachment 207225 [details]
it just motioned some loop-invariant code
Comment 7 Filip Pizlo 2013-07-21 22:04:50 PDT 
Comment on attachment 207225 [details]
it just motioned some loop-invariant code

Once again, I didn't mean to r?.  I haven't fully tested it yet.
Comment 8 Filip Pizlo 2013-07-21 22:12:44 PDT 
Created attachment 207226 [details]
fixed some stuff

I totally forgot about updating the state-at-tail.
Comment 9 Filip Pizlo 2013-07-22 11:45:48 PDT 
Created attachment 207259 [details]
the patch
Comment 10 Filip Pizlo 2013-07-22 11:47:14 PDT 
Created attachment 207260 [details]
the patch

Fix changelog.
Comment 11 Filip Pizlo 2013-07-22 15:22:05 PDT 
Created attachment 207286 [details]
the patch

Lifted the inner-most loop restriction.
Comment 12 Filip Pizlo 2013-07-22 22:39:03 PDT 
Landed in http://trac.webkit.org/changeset/153033