DFG should have a precise view of jump targets

• Component: JavaScriptCore
• Status: RESOLVED FIXED

[Filip Pizlo]

This will reduce the pressure on the CFG simplifier.

[Filip Pizlo]

Created attachment 186476 [details]
the patch

[Oliver Hunt]

Comment on attachment 186476 [details]
the patch

View in context: https://bugs.webkit.org/attachment.cgi?id=186476&action=review

Looks fine, my only concern is whether std::sort is guaranteed to not be a stupid quick sort

> Source/JavaScriptCore/bytecode/PreciseJumpTargets.cpp:116
> +    std::sort(out.begin(), out.end());

How well does std::sort work on mostly ordered data?

[Filip Pizlo]

(In reply to comment #2)
> (From update of attachment 186476 [details])
> View in context: https://bugs.webkit.org/attachment.cgi?id=186476&action=review
> 
> Looks fine, my only concern is whether std::sort is guaranteed to not be a stupid quick sort
> 
> > Source/JavaScriptCore/bytecode/PreciseJumpTargets.cpp:116
> > +    std::sort(out.begin(), out.end());
> 
> How well does std::sort work on mostly ordered data?

I don't think I have a guarantee there.  My recollection is that they don't do the rookie use-first-element-as-pivot mistake but I haven't read the code.  But the performance numbers are looking consistently good...

As well, even if std::sort were to deteriorate to O(n^2), it's O(n^2) over basic blocks and not instructions.  The downside of _not_ running the precise jump target calculator is that we deteriorate to O(n^2) over instructions in the CFG simplifier - so much much worse.

It would be great to try to optimize this code more, though, if we see it showing up as a hot spot.

[Filip Pizlo]

Oops, I'll have to do a pretty big rebase apparently.

[Filip Pizlo]

Created attachment 186526 [details]
patch for landing

[Filip Pizlo]

Landed in http://trac.webkit.org/changeset/141931