Deleting properties poisons objects

• Component: JavaScriptCore
• Status: REOPENED ---

[Mark Hahnenberg]

A rather vitriolic title, but it's sort of true. Deleting a property on an object causes it to transition to an uncacheable dictionary structure. We then gather no information on that object in the baseline JIT's inline caches, which then causes the higher tier JITs to do fully generic GetById/PutById operations on these objects which is very slow.

The idea behind uncacheable dictionaries is that objects whose properties are deleted are being used like a dictionary to create and remove properties. We'd like to avoid creating tons of Structures for these objects. Additionally, re-use of previously deleted property offsets would wreak havoc on our inline caches. However, for objects where a small set of properties are added or removed, we should try to avoid dooming them to an eternity of slowness.

Currently the first delete transitions to the uncacheable dictionary mode, but we could increase that limit to allow objects that are only involved in a small number of deletes to continue to function normally.

[Mark Hahnenberg]

Created attachment 229144 [details]
Patch

[Geoffrey Garen]

Comment on attachment 229144 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=229144&action=review

r=me

> Source/JavaScriptCore/runtime/Structure.cpp:472
> +    if (structure->m_forgivenDeletes < s_maxForgivenDeletes) {
> +        Structure* transition = create(vm, structure);

Another interesting way to forgive a delete would be to notice that you were deleting the last property added, and tradition backward. I believe we already have all the information needed to do that, since each Structure points to its previous, and holds the string that it added to the set of strings in the property table (m_nameInPrevious). You would also need to verify that going backward did not violate other invariants. For example, if going forward grew your backing store, it might not be valid to go backward anymore.

[Geoffrey Garen]

Can you add a JSRegress test for this?

[Build Bot]

Comment on attachment 229144 [details]
Patch

Attachment 229144 [details] did not pass mac-wk2-ews (mac-wk2):
Output: http://webkit-queues.appspot.com/results/6447102225809408

New failing tests:
jquery/manipulation.html
js/pic/undictionary.html
js/named-function-expression.html
storage/indexeddb/cursor-value.html
js/preventExtensions.html
js/array-functions-non-arrays.html
jquery/attributes.html
js/dom/constructor-attributes.html
jquery/deferred.html
jquery/data.html
js/mozilla/strict/8.12.7.html
jquery/event.html
js/dictionary-prototype-caching.html

[Build Bot]

Created attachment 229155 [details]
Archive of layout-test-results from webkit-ews-13 for mac-mountainlion-wk2

The attached test failures were seen while running run-webkit-tests on the mac-wk2-ews.
Bot: webkit-ews-13  Port: mac-mountainlion-wk2  Platform: Mac OS X 10.8.5

[Mark Hahnenberg]

Yeah, this borked some stuff. Investigating...

[Build Bot]

Comment on attachment 229144 [details]
Patch

Attachment 229144 [details] did not pass mac-ews (mac):
Output: http://webkit-queues.appspot.com/results/6682139479244800

New failing tests:
jquery/manipulation.html
js/pic/undictionary.html
js/named-function-expression.html
js/preventExtensions.html
js/array-functions-non-arrays.html
jquery/attributes.html
js/dom/constructor-attributes.html
jquery/deferred.html
jquery/data.html
js/mozilla/strict/8.12.7.html
jquery/event.html
js/dictionary-prototype-caching.html

[Build Bot]

Created attachment 229158 [details]
Archive of layout-test-results from webkit-ews-07 for mac-mountainlion

The attached test failures were seen while running run-webkit-tests on the mac-ews.
Bot: webkit-ews-07  Port: mac-mountainlion  Platform: Mac OS X 10.8.5

[Geoffrey Garen]

Comment on attachment 229144 [details]
Patch

You should fix these failing tests by pinning a property table when doing a delete transition. Our code only knows how to reconstruct a property table through insertions, not deletions, so we have to hang onto a property table that indicates deletions.

[Mark Hahnenberg]

(In reply to comment #9)
> (From update of attachment 229144 [details])
> You should fix these failing tests by pinning a property table when doing a delete transition. Our code only knows how to reconstruct a property table through insertions, not deletions, so we have to hang onto a property table that indicates deletions.

Unfortunately pinning doesn't quite fix this issue. When we pin a property table we also clear its m_nameInPrevious which is used as a key in the previous Structure's m_transitionTable. If we want to be able to cache these transitions then we need to "pin" in such a way that we can't go back farther in the transition chain when doing materializePropertyMap (e.g. previousID should be null), but that we can still use in the previous Structure's transition table. Or we could just give up on caching delete transitions, but allow future transitions from the deleted Structure to be cached (which pin()-ing would give us).

[Mark Hahnenberg]

Created attachment 229620 [details]
Patch

[Mark Hahnenberg]

Comment on attachment 229620 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=229620&action=review

> Source/JavaScriptCore/ChangeLog:8
> +        This is ~6% progression on Dromaeo.

Correction. This is a 6% speedup on the jslib portion of Dromaeo. I will do a full run as well.

[Build Bot]

Comment on attachment 229620 [details]
Patch

Attachment 229620 [details] did not pass mac-wk2-ews (mac-wk2):
Output: http://webkit-queues.appspot.com/results/6102370081570816

New failing tests:
jquery/data.html

[Build Bot]

Created attachment 229628 [details]
Archive of layout-test-results from webkit-ews-16 for mac-mountainlion-wk2

The attached test failures were seen while running run-webkit-tests on the mac-wk2-ews.
Bot: webkit-ews-16  Port: mac-mountainlion-wk2  Platform: Mac OS X 10.8.5

[Build Bot]

Comment on attachment 229620 [details]
Patch

Attachment 229620 [details] did not pass mac-ews (mac):
Output: http://webkit-queues.appspot.com/results/6178124664406016

New failing tests:
jquery/data.html

[Build Bot]

Created attachment 229630 [details]
Archive of layout-test-results from webkit-ews-02 for mac-mountainlion

The attached test failures were seen while running run-webkit-tests on the mac-ews.
Bot: webkit-ews-02  Port: mac-mountainlion  Platform: Mac OS X 10.8.5

[Build Bot]

Comment on attachment 229620 [details]
Patch

Attachment 229620 [details] did not pass mac-ews (mac):
Output: http://webkit-queues.appspot.com/results/5751994165755904

New failing tests:
jquery/data.html

[Build Bot]

Created attachment 229634 [details]
Archive of layout-test-results from webkit-ews-07 for mac-mountainlion

The attached test failures were seen while running run-webkit-tests on the mac-ews.
Bot: webkit-ews-07  Port: mac-mountainlion  Platform: Mac OS X 10.8.5

[Geoffrey Garen]

Comment on attachment 229620 [details]
Patch

EWS test failure looks real.

[Mark Hahnenberg]

(In reply to comment #19)
> (From update of attachment 229620 [details])
> EWS test failure looks real.

Yep :-(

[Mark Hahnenberg]

The test that's failing is the following:

var obj = { foo: "bar" };
jQuery(obj).data("foo", "baz");

dataObj = jQuery.extend(true, {}, jQuery(obj).data());

// TODO: Remove this hack which was introduced for 1.5.1
delete dataObj.toJSON;

deepEqual( dataObj, { foo: "baz" }, "Retrieve data object from a wrapped JS object (#7524)" );

Somehow the delete operation is screwing up dataObj to where it's no longer deepEqual to {foo: "baz"}. Still investigating why this would happen.

[Mark Hahnenberg]

(In reply to comment #21)
> The test that's failing is the following:
> 
> var obj = { foo: "bar" };
> jQuery(obj).data("foo", "baz");
> 
> dataObj = jQuery.extend(true, {}, jQuery(obj).data());
> 
> // TODO: Remove this hack which was introduced for 1.5.1
> delete dataObj.toJSON;
> 
> deepEqual( dataObj, { foo: "baz" }, "Retrieve data object from a wrapped JS object (#7524)" );
> 
> Somehow the delete operation is screwing up dataObj to where it's no longer deepEqual to {foo: "baz"}. Still investigating why this would happen.

deepEqual() uses for-in, so maybe JSPropertyNameIterator is confused about things.

[Mark Hahnenberg]

> deepEqual() uses for-in, so maybe JSPropertyNameIterator is confused about things.

Seems like that was the issue. Re-running test and benchmarks.

[Mark Hahnenberg]

Created attachment 229659 [details]
Patch

[Mark Hahnenberg]

Created attachment 229660 [details]
Patch

[Geoffrey Garen]

Comment on attachment 229660 [details]
Patch

r=me

[Geoffrey Garen]

Comment on attachment 229660 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=229660&action=review

> Source/JavaScriptCore/runtime/JSPropertyNameIterator.cpp:67
> +    if (o->structure()->isDictionary() || o->structure()->hasDeletedOffsets())
>          return jsPropertyNameIterator;

You should adda comment explaining that the iterator doesn't know how to skip deleted buckets.

[Mark Hahnenberg]

Committed r167501: <http://trac.webkit.org/changeset/167501>

[Andreas Kling]

(In reply to comment #28)
> Committed r167501: <http://trac.webkit.org/changeset/167501>

I think this may be causing DoYouEvenBench to crash reliably on bawts: http://build.webkit.org/builders/Apple%20MountainLion%20Release%20%28Perf%29?numbuilds=50

[WebKit Commit Bot]

Re-opened since this is blocked by bug 131913

(In reply to comment #28)
> Committed r167501: <http://trac.webkit.org/changeset/167501>

This change also broke the Find My iPhone web interface at icloud.com, but since it’s already been reverted, I didn’t file a bug about that.

[Mark Hahnenberg]

Created attachment 230122 [details]
Fixed some issues

[Build Bot]

Comment on attachment 230122 [details]
Fixed some issues

Attachment 230122 [details] did not pass mac-ews (mac):
Output: http://webkit-queues.appspot.com/results/5381299967623168

New failing tests:
js/regress/delete-a-few-properties-then-get-by-id.html

[Build Bot]

Created attachment 230130 [details]
Archive of layout-test-results from webkit-ews-05 for mac-mountainlion

The attached test failures were seen while running run-webkit-tests on the mac-ews.
Bot: webkit-ews-05  Port: mac-mountainlion  Platform: Mac OS X 10.8.5

[Build Bot]

Comment on attachment 230122 [details]
Fixed some issues

Attachment 230122 [details] did not pass mac-ews (mac):
Output: http://webkit-queues.appspot.com/results/5737659846623232

New failing tests:
js/regress/delete-a-few-properties-then-get-by-id.html

[Build Bot]

Created attachment 230132 [details]
Archive of layout-test-results from webkit-ews-04 for mac-mountainlion

The attached test failures were seen while running run-webkit-tests on the mac-ews.
Bot: webkit-ews-04  Port: mac-mountainlion  Platform: Mac OS X 10.8.5

[Build Bot]

Comment on attachment 230122 [details]
Fixed some issues

Attachment 230122 [details] did not pass mac-wk2-ews (mac-wk2):
Output: http://webkit-queues.appspot.com/results/5348661437399040

New failing tests:
js/regress/delete-a-few-properties-then-get-by-id.html

[Build Bot]

Created attachment 230136 [details]
Archive of layout-test-results from webkit-ews-11 for mac-mountainlion-wk2

The attached test failures were seen while running run-webkit-tests on the mac-wk2-ews.
Bot: webkit-ews-11  Port: mac-mountainlion-wk2  Platform: Mac OS X 10.8.5

[Mark Hahnenberg]

Created attachment 230183 [details]
Patch

[Oliver Hunt]

Comment on attachment 230183 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=230183&action=review

> Source/JavaScriptCore/ChangeLog:15
> +        (JSC::PropertyTable::hadDeletedOffset): If we ever had deleted properties we can no longer cache offsets when 
> +        iterating properties because we're required to iterate properties in insertion order.

yay javascript!!!

> Source/JavaScriptCore/ChangeLog:21
> +        (JSC::Structure::removePropertyTransition): We allow up to 5 deletes for a particular path through the tree of 
> +        Structure transitions. After that, we convert to an uncacheable dictionary like we used to. We don't cache 
> +        delete transitions, but we allow transitioning from them.

how did we choose 5? Is it possible to use 2^n (±1?) so it looks deliberate? :D

[Mark Hahnenberg]

> > Source/JavaScriptCore/ChangeLog:21
> > +        (JSC::Structure::removePropertyTransition): We allow up to 5 deletes for a particular path through the tree of 
> > +        Structure transitions. After that, we convert to an uncacheable dictionary like we used to. We don't cache 
> > +        delete transitions, but we allow transitioning from them.
> 
> how did we choose 5? Is it possible to use 2^n (±1?) so it looks deliberate? :D

I picked the smallest number that achieved the maximal speedup on a benchmark affected by this optimization. 1-4 produced speedups, but not as much as 5, and anything more than 5 had no effect. Super scientific!

[WebKit Commit Bot]

Comment on attachment 230183 [details]
Patch

Clearing flags on attachment: 230183

Committed r167889: <http://trac.webkit.org/changeset/167889>

[WebKit Commit Bot]

All reviewed patches have been landed.  Closing bug.

[Ryosuke Niwa]

This patch appears to have caused ~1.5% regression on DYEBench:
https://perf.webkit.org/#mode=charts&chartList=%5B%5B%22mac-mavericks%22%2C%22DoYouEvenBench%2FFull%3ATime%3ATotal%22%5D%2C%5B%22mac-mountainlion%22%2C%22DoYouEvenBench%2FFull%3ATime%3ATotal%22%5D%5D&days=17

[Mark Hahnenberg]

This broke things. Reopening for a rollout. We're going to reevaluate how we do Structure transitions in general as part of bug 132618.

[Mark Hahnenberg]

Rolled out in http://trac.webkit.org/changeset/168373.