Bug 140743 – BytecodeGenerator::initializeCapturedVariable() sets a misleading value for the 5th operand of op_put_to_scope

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

[Mark Lam]

Steps to reproduce:
1. Go to https://bugreport.apple.com/problem/viewproblem?problemID=18702537
2. Log in.

BytecodeGenerator::initializeCapturedVariable() is setting the 5th operand to op_put_to_scope to an inappropriate value.  As a result, the execution of put_to_scope could store a wrong inferred value into the VariableWatchpointSet for which ever captured variable is at local index 0.  In practice, this turns out to be the local for the Arguments object.  In this example, the wrong inferred value written there is the boolean true.

Subsequently, DFG compilation occurs and CreateArguments is emitted to first do a check of the local for the Arguments object.  But because that local has a wrong inferred value, the check always discovers a non-null value and we never actually create the Arguments object.  Immediately after this, an OSR exit occurs leaving the Arguments object local uninitialized.  Later on at arguments tear off, we run into a boolean true where we had expected to find an Arguments object, which in turn, leads to the crash.

[Mark Lam]

<rdar://problem/18821173>

[Mark Lam]

Created attachment 245095 [details]
the patch.

I need to rerun perf numbers but this patch has already passed the jsc stress tests and layout tests.

[Geoffrey Garen]

Comment on attachment 245095 [details]
the patch.

View in context: https://bugs.webkit.org/attachment.cgi?id=245095&action=review

> LayoutTests/js/script-tests/dfg-osr-exit-between-create-and-tearoff-arguments.js:11
> +    // Loop ridiculously many times here to get this function to DFG compile.

Can you use the predicate that dfgShouldBe uses instead? 20000 is probably much more than is necessary.

[Mark Lam]

Created attachment 245106 [details]
patch 2: applied Geoff's feedback

Thanks for the feedback.  The test has been updated to use dfgCompiled() instead.

The perf numbers came back neutral.  I'll upload the results shortly.

[Mark Lam]

Created attachment 245107 [details]
perf results

[WebKit Commit Bot]

Comment on attachment 245106 [details]
patch 2: applied Geoff's feedback

Clearing flags on attachment: 245106

Committed r178926: <http://trac.webkit.org/changeset/178926>

[WebKit Commit Bot]

All reviewed patches have been landed.  Closing bug.