Use singleton thunks for virtual calls.

[Mark Lam]

The pre-existing implementation emits a unique thunk for each virtual call site.  This turns out to be very wasteful in terms of memory.  For example, Speedometer2 ends up generating ~16M of virtualFor thunks.  Switching to re-useable singleton thunks, saves that 16M, and also appears to improve Speedometer2 performance by 1.012x.

[Radar WebKit Bug Importer]

<rdar://problem/78357604>

[Mark Lam]

Forgot to say, the 1.012x was measured on a M1 Mac.

[Mark Lam]

Created attachment 429445 [details]
proposed patch.

Let's try this on the EWS.

[Mark Lam]

Created attachment 429457 [details]
proposed patch.

[Yusuke Suzuki]

Comment on attachment 429457 [details]
proposed patch.

View in context: https://bugs.webkit.org/attachment.cgi?id=429457&action=review

r=me

> Source/JavaScriptCore/ChangeLog:13
> +        2. Introduce Options::useUniqueVirtualThunks() to allow unique thunks to be
> +           generated for testing and comparisons.  Options::useUniqueVirtualThunks() is
> +           false by default.

Maybe this flag is not necessary.

> Source/JavaScriptCore/jit/ThunkGenerators.cpp:275
> +    bool isTailCall = (mode == CallMode::Tail);

() is not necessary.

[Mark Lam]

Comment on attachment 429457 [details]
proposed patch.

View in context: https://bugs.webkit.org/attachment.cgi?id=429457&action=review

>> Source/JavaScriptCore/ChangeLog:13
>> +           false by default.
> 
> Maybe this flag is not necessary.

I suppose we can always add it later if a need arises.  I'll remove it.

>> Source/JavaScriptCore/jit/ThunkGenerators.cpp:275
>> +    bool isTailCall = (mode == CallMode::Tail);
> 
> () is not necessary.

Will fix.

[Mark Lam]

Thanks for the review.  Landed in r277929: <http://trac.webkit.org/r277929>.