[Webkit-unassigned] [Bug 162124] op_get_by_id_with_this should use inline caching

[bugzilla-daemon at webkit.org]

https://bugs.webkit.org/show_bug.cgi?id=162124

--- Comment #30 from Caio Lima <ticaiolima at gmail.com> ---
(In reply to comment #28)
> (In reply to comment #20)
> > (In reply to comment #19)
> > > Why would we do this?
> > 
> > If I didn't miss any context *_with_this were created to solve the bug on
> > (https://bugs.webkit.org/show_bug.cgi?id=147064). The current
> > implementationis just a call to slow paths "operation*WithThis", meaning
> > that we don't use IC or PIC on "super.foo" for example.
> > Merging get_by_id_with_this with op_get_by_id we enable its optimization
> > including IC and PIC. In terms of performance, now it means 20% faster over
> > op_get_by_id_with_this version.
> 
> Then you should make get_by_id_with_this use inline caches. 
> 
> You shouldn't change get_by_id. 
> 
> > 
> > > It's definitely good for lower tiers that get_by_id does not take an extra
> > > this operand.  It's also easier to parse.  I don't understand why we would
> > > remove this optimization.
> > 
> > I didn't understand why it is easier to parse.
> > Also, About good parts of get_by_id does not take an extra this operand, do
> > you mind explain me why? I can think that we use extra memory to store the
> > this operand, which has a big impact because get_by_id is one of the most
> > used op_codes.
> 
> Compiler analyses always have an O(E) component, where E is the number of
> data flow edges. One such analysis is the OSR likeness analysis in the DFG
> and FTL. It's a very expensive analysis and we don't want to make it worse.
> You're adding a useless "this" data flow edge to every property access. I
> don't think we want that. 

Thank you for this explanation. I implemented a logic to avoid create an edge if ```base == this``` just thinking in memory usage, but it is interesting know that it also affects analyses performance.

> I just want to be very clear: I'm strongly against uniting get_by_id with
> get_by_id_with_this. Please change your approach to making the unification
> happen at the PolymorphicAxcess/StructureStubInfo level and please keep
> these opcodes sepaate.

Agreed. I don't like the idea of an operand for ```this``` in all property access on op_get_by_id since it is useless ~95% of time. 

(In reply to comment #29)
> I agree with Fil. I think we should just abstract the IC machinery to allow
> for get_by_id_with_this semantics. It probably requires modifying
> PolymorphicAccess, a new IC generator for the baseline, and proper handling
> in the DFG/FTL. I suspect you'll be able to abstract certain code in the
> DFG/FTL to handle both nodes. Also, I suspect the heart of the work you
> already did was in PolymorphicAccess and StructureStubInfo, so all of that
> code should not really be changed regardless of if the opcodes are united or
> not.

Cool. I am going to work on that.

-- 
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/20161009/84948066/attachment-0001.html>