Bug 130791

Summary:

FTL should use stackmap register liveness

Product:

WebKit

Reporter:

Filip Pizlo <fpizlo>

Component:

JavaScriptCore

Assignee:

Filip Pizlo <fpizlo>

Status:

RESOLVED FIXED

Severity:

Normal

CC:

atrick, barraclough, ggaren, juergen, mark.lam, mhahnenberg, msaboff, nrotem, oliver, sam

Priority:

Version:

528+ (Nightly build)

Hardware:

All

OS:

All

Bug Depends on:

130792, 130801, 130821, 130924, 130948

Bug Blocks:

112840, 130885

Attachments:

Description	Flags
work in progress	none
more	none
the patch	ggaren: review+

Filip Pizlo

Reported 2014-03-26 12:36:05 PDT

Patch forthcoming.

Attachments
work in progress (5.26 KB, patch) 2014-03-27 21:03 PDT, Filip Pizlo	no flags	Details Formatted Diff Diff
more (24.09 KB, patch) 2014-03-28 14:36 PDT, Filip Pizlo	no flags	Details Formatted Diff Diff
the patch (6.81 KB, patch) 2014-04-12 11:29 PDT, Filip Pizlo	ggaren: review+	Details Formatted Diff Diff
Show Obsolete (2) View All Add attachment proposed patch, testcase, etc.

Andrew Trick

Comment 1 2014-03-27 10:30:11 PDT

The LLVM side of this is untested. We need a runtime mode that clobbers the dead regs. Ideally we could also have a mode to disable inline caches (cause them to fail) to test as many patchpoint liveouts as possible.

Filip Pizlo

Comment 2 2014-03-27 11:24:18 PDT

(In reply to comment #1) > The LLVM side of this is untested. We need a runtime mode that clobbers the dead regs. Ideally we could also have a mode to disable inline caches (cause them to fail) to test as many patchpoint liveouts as possible. That's a great idea and it should be easy to add such a mode. I'll investigate adding it as part of the patch that glues all of this together. Hopefully doing so won't slow the usual test cycle down too much. I wonder if we could also strengthen this by making the C++ code clobber all registers, ideally with a runtime switch, and even better if it doesn't involve too many hard-to-maintain hacks.

Andrew Trick

Comment 3 2014-03-27 14:51:56 PDT

Regarding LLVM-based clobbering, see http://llvm.org/pr19267.

Filip Pizlo

Comment 4 2014-03-27 21:03:21 PDT

Created attachment 228021 [details] work in progress This doesn't work yet.

Filip Pizlo

Comment 5 2014-03-27 21:08:44 PDT

(In reply to comment #4) > Created an attachment (id=228021) [details] > work in progress > > This doesn't work yet. deltablue is super unhappy, I will investigate why.

Filip Pizlo

Comment 6 2014-03-27 22:00:26 PDT

(In reply to comment #5) > (In reply to comment #4) > > Created an attachment (id=228021) [details] [details] > > work in progress > > > > This doesn't work yet. > > deltablue is super unhappy, I will investigate why. Heh, it's less likely to fail in a debug build. Because of course the C++ code will use fewer registers in debug mode. ;-)

Filip Pizlo

Comment 7 2014-03-28 14:36:15 PDT

Created attachment 228084 [details] more

Filip Pizlo

Comment 8 2014-04-12 11:29:46 PDT

Created attachment 229207 [details] the patch

Geoffrey Garen

Comment 9 2014-04-12 11:33:35 PDT

Comment on attachment 229207 [details] the patch r=me

Filip Pizlo

Comment 10 2014-04-12 13:42:19 PDT

Landed in http://trac.webkit.org/changeset/167187

Note You need to log in before you can comment on or make changes to this bug.