[Webkit-unassigned] [Bug 79899] DFG BasicBlocks should not require that their nodes have contiguous indices in the graph

[bugzilla-daemon at webkit.org]

https://bugs.webkit.org/show_bug.cgi?id=79899





--- Comment #12 from Yuqiang Xian <yuqiang.xian at intel.com>  2012-02-29 22:38:07 PST ---
(In reply to comment #10)
> (From update of attachment 129627 [details])
> View in context: https://bugs.webkit.org/attachment.cgi?id=129627&action=review
> 
> Awesome!
> 
> Question: does this mean that the start index performance hack is no longer in effect at all?  I.e. you're not even stopping the CSE when it gets to the earliest possible child of the node being replaced?  That's not a problem - in fact it's great if that's true.  I had added that optimization prematurely and it's somewhat comforting to know that it isn't necessary.
> 

According to the data it may be true. But yes you're right that it's possible to terminate the look-back process earlier. Maybe we can do this in a separated patch?

> R=me.
> 
> > Source/JavaScriptCore/dfg/DFGByteCodeParser.cpp:616
> > +        block->prepend(resultIndex);
> 
> Ha!  I find it to be quite awesome that we see no regression even though prepend is an O(n) operation.  Or am I missing something?

I originally had the same concern as yours. So I had another version which just kept the Phi nodes in a separated block (just as how they're originally handled) but later I found no obvious performance difference comparing to current approach. So I still throw out this version.

-- 
Configure bugmail: https://bugs.webkit.org/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are the assignee for the bug.