Bug 77797 – DFG's child references from one node to another should have room for type information

• Component: JavaScriptCore
• Status: RESOLVED FIXED

[Filip Pizlo]

This will allow us to reason about not just whether a node is used, but whether it's used for a particular purpose.

[Filip Pizlo]

Created attachment 125460 [details]
it begins.

Work in progress.  Three files down, 51 to go.

[Filip Pizlo]

Created attachment 125466 [details]
getting closer

[Filip Pizlo]

Created attachment 125517 [details]
64-bit works

Still need to do 32-64 and make sure I didn't screw up performance.

[Gyuyoung Kim]

Comment on attachment 125517 [details]
64-bit works

Attachment 125517 [details] did not pass efl-ews (efl):
Output: http://queues.webkit.org/results/11424834

[Early Warning System Bot]

Comment on attachment 125517 [details]
64-bit works

Attachment 125517 [details] did not pass qt-ews (qt):
Output: http://queues.webkit.org/results/11421879

[Gustavo Noronha (kov)]

Comment on attachment 125517 [details]
64-bit works

Attachment 125517 [details] did not pass gtk-ews (gtk):
Output: http://queues.webkit.org/results/11427085

[Filip Pizlo]

<rdar://problem/10810732>

[Filip Pizlo]

Created attachment 125537 [details]
the patch

Made both 32_64 and 64 work.

[Early Warning System Bot]

Comment on attachment 125537 [details]
the patch

Attachment 125537 [details] did not pass qt-ews (qt):
Output: http://queues.webkit.org/results/11427257

[Filip Pizlo]

Created attachment 125538 [details]
the patch

Fix a really bad operator precedence bug found by Qt EWS.  Instead of writing (a << b) - c, I had written a << b - c, which is totally different.

[Oliver Hunt]

Comment on attachment 125538 [details]
the patch

View in context: https://bugs.webkit.org/attachment.cgi?id=125538&action=review

Other than the comments included (below?) this looks fairly sensible to me.

> Source/JavaScriptCore/dfg/DFGByteCodeParser.cpp:173
> -                    nodeIndex = addToGraph(GetLocal, OpInfo(variableAccessData), nodePtr->child1());
> +                    nodeIndex = addToGraph(GetLocal, OpInfo(variableAccessData), nodePtr->child1().index());

It seems like it would be nicer to make addToGraph take a Node/NodeUse rather than a NodeIndex, which would remove some of the churn in this diff.

> Source/JavaScriptCore/dfg/DFGByteCodeParser.cpp:2273
> +                phiNode->children.child1() = NodeUse(valueInPredecessor);

I'd prefer setChild1(..) to returning a mutable reference.

[Filip Pizlo]

(In reply to comment #11)
> (From update of attachment 125538 [details])
> View in context: https://bugs.webkit.org/attachment.cgi?id=125538&action=review
> 
> Other than the comments included (below?) this looks fairly sensible to me.
> 
> > Source/JavaScriptCore/dfg/DFGByteCodeParser.cpp:173
> > -                    nodeIndex = addToGraph(GetLocal, OpInfo(variableAccessData), nodePtr->child1());
> > +                    nodeIndex = addToGraph(GetLocal, OpInfo(variableAccessData), nodePtr->child1().index());
> 
> It seems like it would be nicer to make addToGraph take a Node/NodeUse rather than a NodeIndex, which would remove some of the churn in this diff.

Do you really think that's a good idea?  There are 4 addToGraph() methods, and only 2 places where this patch needs to convert a NodeUse to NodeIndex for a call to addToGraph().  In general, ByteCodeParser almost never needs to create a node that references a child of another node - which is the only case where it would make sense for addToGraph() to take NodeUse.

> 
> > Source/JavaScriptCore/dfg/DFGByteCodeParser.cpp:2273
> > +                phiNode->children.child1() = NodeUse(valueInPredecessor);
> 
> I'd prefer setChild1(..) to returning a mutable reference.

I can change those things, but it is necessary for NodeReferenceBlob to return a mutable reference.  Otherwise we'd have no way of later doing a pass that ascribes types to edges.

[Filip Pizlo]

Created attachment 125545 [details]
the patch

Changed all present uses of childX() = foo with setChildX(foo).

I did not remove the childX() methods that returned mutable references, because that would defeat the entire purpose of this patch.

[Filip Pizlo]

Landed in http://trac.webkit.org/changeset/106775