[Webkit-unassigned] [Bug 156147] [JSC] implement async functions proposal

[bugzilla-daemon at webkit.org]

https://bugs.webkit.org/show_bug.cgi?id=156147

--- Comment #112 from Yusuke Suzuki <utatane.tea at gmail.com> ---
Comment on attachment 280020
  --> https://bugs.webkit.org/attachment.cgi?id=280020
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=280020&action=review

>>>>> Source/JavaScriptCore/parser/Parser.cpp:624
>>>>> +        if (UNLIKELY(m_runtimeFlags.isAsyncAwaitEnabled() && matchContextualKeyword(m_vm->propertyNames->async))) {
>>>> 
>>>> There is a chance that this may be faster if the LHS and RHS were swapped because it might aid the register allocator. 
>>>> Same with the similar compare elsewhere.
>>>> Is this what's slowing down the parser?
>>>> Have you tried making ASYNC a key word? That may be faster than checking the IDENT's StringImpl
>>> 
>>> When executing `matchContextualKeyword(m_vm->propertyNames->async)`, it becomes `Identifier == Identifier` check and it should be pointer comparison because Identifier is already interned.
>> 
>> I was just thinking that we may get almost identical codegen (or we would get slightly worse codegen but the machine's branch/indirect jump prediction would make up for it) for the 'switch' statement if we made ASYNC a keyword, and we wouldn't have this branch at each IDENT. But maybe this would cause slow downs elsewhere. I'm not sure. It's hard to predict without implementing it and measuring perf.
> 
> Actually, a good starting point is to see if these branches are actually what the remaining slow down is. Because octane/code-load doesn't use async/await, we can probably get away with removing the branch and seeing if the entire regression goes away. If it doesn't, we know there is slow down elsewhere.

Sounds reasonable.

-- 
You are receiving this mail because:
You are the assignee for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.webkit.org/pipermail/webkit-unassigned/attachments/20160528/d5e5db68/attachment.html>