Bug 64066 – DFG JIT does not implement op_construct

• Status: RESOLVED
• Resolution: FIXED
• Component: JavaScriptCore

[Filip Pizlo]

The DFG JIT does not implement the op_construct opcode.  This means that code containing object construction will not get compiled with DFG.  DFG JIT should implement op_construct.

[Filip Pizlo]

Created attachment 99946 [details]
the patch

[WebKit Review Bot]

Attachment 99946 [details] did not pass style-queue:

Failed to run "['Tools/Scripts/check-webkit-style', '--diff-files', u'Source/JavaScriptCore/ChangeLog', u'Source..." exit_code: 1

Source/JavaScriptCore/dfg/DFGRepatch.h:40:  The parameter name "kind" adds no information, so it should be removed.  [readability/parameter_name] [5]
Total errors found: 1 in 11 files


If any of these errors are false positives, please file a bug against check-webkit-style.

[Filip Pizlo]

Created attachment 99948 [details]
the patch (fix style)

[WebKit Review Bot]

Comment on attachment 99948 [details]
the patch (fix style)

Clearing flags on attachment: 99948

Committed r90602: <http://trac.webkit.org/changeset/90602>

[WebKit Review Bot]

All reviewed patches have been landed.  Closing bug.

[Gavin Barraclough]

Reopening - these changes were rolled out in r90673 due to https://bugs.webkit.org/show_bug.cgi?id=64181

[Filip Pizlo]

Created attachment 100583 [details]
the patch

This patch fixes two bugs in the previous patch, which, together with other patches since committed, makes this work with gmail, bankrate, facebook, reddit, and google (according to my testing), even with DFG restrictions turned off.

[Gavin Barraclough]

Comment on attachment 100583 [details]
the patch

Hey Filip, following on from our conversation earlier, looking at the two bugs this fixes since the original patch, one looks tougher to test for (error on unlinking calls), but the other does look testable (virtualFor always compiles for call).  To test we'd just have to have one function use new, call another, and have the constructor that is called do something that would prove whether it was invoked as a call or a constructor (as an example, returning a non-object value would do).  Unless there is any reason it won't be possible, let's add a LayoutTest.

[Filip Pizlo]

Created attachment 100592 [details]
the patch (fix review)

[WebKit Review Bot]

Comment on attachment 100592 [details]
the patch (fix review)

Clearing flags on attachment: 100592

Committed r90877: <http://trac.webkit.org/changeset/90877>

[WebKit Review Bot]

All reviewed patches have been landed.  Closing bug.