[JSC] Always track out-of-bounds array access explicitly instead of relying on the slow case

• Component: New Bugs
• Status: RESOLVED FIXED

[Benjamin Poulain]

[JSC] Always track out-of-bounds array access explicitly instead of relying on the slow case

[Benjamin Poulain]

Created attachment 254317 [details]
Patch

[Benjamin Poulain]

Created attachment 254319 [details]
Patch

[Geoffrey Garen]

Comment on attachment 254319 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=254319&action=review

r=me

How much of a speedup is this worth?

> Source/JavaScriptCore/ChangeLog:41
> +        There are a few important cases that are still not covered correctly:

When you say "correctly" I think you might be talking about a bug. I think you meant "optimally".

> Source/JavaScriptCore/bytecode/ArrayProfile.h:213
> +    void setOutOfBounds() { m_outOfBounds = true; }

When I first read this, I though "out of bounds" meant "past length". But now I see that it actually means "past length or a hole value". I wonder if we can use a clearer name for this. How about "sawHole"? Technically, things past length are holes, just like unfilled items within length.

> Source/JavaScriptCore/jit/JITOperations.cpp:1481
> +static bool canUseFastArgumentAccess(JSObject& object, uint32_t index)

I think I would call this "canAccessArgumentIndexQuickly" to match the function name in DirectArguments and ScopedArguments.

I agree that it's surprising that arguments objects have different notions of fast indexed access as compared to other objects, but I think it's probably OK.

> Source/JavaScriptCore/jit/JITOperations.cpp:1484
> +    // For Arguments, we don't have holes creating out-of-bounds access, but we have "overrodeThings()" that
> +    // moves us to slow path.

I think this comment is slightly inaccurate: see below.

> Source/JavaScriptCore/jit/JITOperations.cpp:1489
> +        if (!directArguments->overrodeThings() && index < directArguments->internalLength())

Can we just call DirectArguments::canAccessIndexQuickly here instead? Seems like better encapsulation that way.

> Source/JavaScriptCore/jit/JITOperations.cpp:1495
> +        if (!scopedArguments->overrodeThings() && index < scopedArguments->internalLength())

Same comment. In this case, from reading ScopedArguments::canAccessIndexQuickly, I think this test is slightly wrong from an optimization perspective. ScopeArguments much check m_table to verify that an argument is stored inside the ScopedArguments object. In the DFG, failing this check will cause an OSR exit. So, it would be good to use a generic byVal operation instead of optimized access in the case where the m_table entry was present.

[Benjamin Poulain]

Comment on attachment 254319 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=254319&action=review

>> Source/JavaScriptCore/jit/JITOperations.cpp:1489
>> +        if (!directArguments->overrodeThings() && index < directArguments->internalLength())
> 
> Can we just call DirectArguments::canAccessIndexQuickly here instead? Seems like better encapsulation that way.

After discussing with Geoff, looks like we should have Arguments::canAccessArgumentIndexQuicklyInDFG()

[Benjamin Poulain]

Committed r185240: <http://trac.webkit.org/changeset/185240>

[Csaba Osztrogonác]

(In reply to comment #5)
> Committed r185240: <http://trac.webkit.org/changeset/185240>

It broke the ARM Linux build, new bug report with the fix: bug145754 .