Bug 64745 – JSC GC lazy sweep does not inline the common cases of cell destruction

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

[Filip Pizlo]

Most objects in the heap are instances of JSFinalObject, which has a relatively straight-forward destructor.  This destructor's common case is to do nothing: the property storage is not inline, so there is nothing to destroy.  However, the lazy sweep will currently always make a slow virtual call because it doesn't special-case the common cases.

[Filip Pizlo]

Created attachment 101197 [details]
the patch

[Oliver Hunt]

Comment on attachment 101197 [details]
the patch

View in context: https://bugs.webkit.org/attachment.cgi?id=101197&action=review

r- :(

> Source/JavaScriptCore/heap/MarkedBlock.cpp:96
> +            if (*reinterpret_cast<void**>(cell) == jsFinalObjectVPtr) {

use cell->vptr();

> Source/JavaScriptCore/heap/MarkedBlock.cpp:99
> +                if (!object->isUsingInlineStorage())
> +                    delete [] object->m_propertyStorage;

object->JSFinalObject::~JSFinalObject() didn't work?

[Filip Pizlo]

(In reply to comment #2)
> (From update of attachment 101197 [details])
> View in context: https://bugs.webkit.org/attachment.cgi?id=101197&action=review
> 
> r- :(
> 
> > Source/JavaScriptCore/heap/MarkedBlock.cpp:96
> > +            if (*reinterpret_cast<void**>(cell) == jsFinalObjectVPtr) {
> 
> use cell->vptr();

Ah, that's handy!

> > Source/JavaScriptCore/heap/MarkedBlock.cpp:99
> > +                if (!object->isUsingInlineStorage())
> > +                    delete [] object->m_propertyStorage;
> 
> object->JSFinalObject::~JSFinalObject() didn't work?

I tried object->~JSFinalObject() but that didn't give a speed-up.  I forgot the object->JSFinalObject::~JSFinalObject syntax.  That works great.  Doing some quick tests and will resubmit shortly...

[Filip Pizlo]

Created attachment 101205 [details]
the patch (fix review)

[Darin Adler]

Comment on attachment 101205 [details]
the patch (fix review)

View in context: https://bugs.webkit.org/attachment.cgi?id=101205&action=review

> Source/JavaScriptCore/heap/MarkedBlock.cpp:98
> +                JSFinalObject* object = reinterpret_cast<JSFinalObject*>(cell);
> +                object->JSFinalObject::~JSFinalObject();

I would have done this as a one liner:

    reinterpret_cast<JSFinalObject*>(cell)->object->JSFinalObject::~JSFinalObject()

I also think it deserves a “why” comment, stating this is a special case for speed.

[WebKit Review Bot]

Comment on attachment 101205 [details]
the patch (fix review)

Clearing flags on attachment: 101205

Committed r91218: <http://trac.webkit.org/changeset/91218>

[WebKit Review Bot]

All reviewed patches have been landed.  Closing bug.