Bug 129920 – MarkedBlocks that are "full enough" shouldn't be swept after EdenCollections

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

[Mark Hahnenberg]

Imagine a Heap this is completely full after an EdenCollection. The next allocation could potentially walk every atom of every MarkedBlock that its MarkedAllocator has. To avoid this situation we should ignore MarkedBlocks that are "full enough" so we can avoid this linear cost sweep phase.

[Mark Hahnenberg]

Created attachment 226449 [details]
Patch

[Mark Hahnenberg]

Created attachment 226458 [details]
Patch

[Geoffrey Garen]

Comment on attachment 226458 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=226458&action=review

r=me

> Source/JavaScriptCore/heap/MarkedSpace.cpp:-317
> -#ifndef NDEBUG
> -    VerifyNewlyAllocated verifyFunctor;
> -    forEachBlock(verifyFunctor);
> -#endif

Why no verify?

[Mark Hahnenberg]

(In reply to comment #3)
> (From update of attachment 226458 [details])
> View in context: https://bugs.webkit.org/attachment.cgi?id=226458&action=review
> 
> r=me
> 
> > Source/JavaScriptCore/heap/MarkedSpace.cpp:-317
> > -#ifndef NDEBUG
> > -    VerifyNewlyAllocated verifyFunctor;
> > -    forEachBlock(verifyFunctor);
> > -#endif
> 
> Why no verify?

I can undo that one. In an older version of the patch I was using m_newlyAllocated but no longer.

[Mark Hahnenberg]

Committed r165458: <http://trac.webkit.org/changeset/165458>

[Jer Noble]

This commit may have caused the fast/workers/worker-gc.html test to begin crashing <https://bugs.webkit.org/show_bug.cgi?id=130135>