svg/animations/svglength-element-removed-crash.svg is a flaky failure but is becoming more consistent in failing.
Dashboard shows the failures happening on iOS Simulator Release:
I was able to reproduce the issue locally with:
run-webkit-tests svg/animations/svglength-element-removed-crash.svg --iterations=100 --no-retry-failure --ios-simulator
I was also able to narrow down the failure to Revision r217944 using same testing method
@@ -1 +1 @@
-PASS x = 10
+FAIL: -20 extra live node(s) x = 10
This is affecting EWS results, so we should take care of this quickly or roll out the offending change.
Ryan, can you disable the test for now? That fix was definitely addressing a JSC bug and test flakiness, while this may be just a bad test.
But it is somewhat surprising.
(In reply to Alexey Proskuryakov from comment #2)
> Ryan, can you disable the test for now? That fix was definitely addressing a
> JSC bug and test flakiness, while this may be just a bad test.
> But it is somewhat surprising.
Maybe this SVG test was inadvertently taking advantage of LLInt's old wrong behavior. It would flake, because the wrong behavior was in a LLInt optimization, so sometimes you wouldn't get the wrong behavior.
Specifically, what happens if you have both a custom property and an element called "foo". What does document.foo do? Some what the property, some want the element. Correct behavior is to give the element, LLInt would sometimes give the property.
I agree that it's worth skipping the SVG test. It's probably not too hard to fix, if I'm right about what's going on.
This should be fixed by removing the SVG tear-off object. https://bugs.webkit.org/show_bug.cgi?id=191237
*** This bug has been marked as a duplicate of bug 191237 ***