1156533 - Simplify how we trace Shapes for marking

Status: RESOLVED FIXED

(Core :: JavaScript: GC, defect)

Description

[Terrence Cole [:terrence]]

Because why use a while loop if a goto will suffice. Considering this particular over-use, I'd be tempted to say that someone thought goto was magic fast sauce.

A couple of things in this patch deserve comment. First, the ascii-art diagram is still a bit of a lie -- a few edges need to be trimmed to fit that view still. I hope to make it not a lie before closing bug 1154959.

In particular, I've added markAndTraceChildren and /removed/ the default implementation of traverse. This makes using traverse on a non-layout type a linking error. Hence, I've added several static_cast to manually coerce derived types into their base types in the eager and mark-stack paths. This seems like a clearer solution to me than ConvertToBase, given that this is all in the optimized path and I don't want to have an additional internal shim layer.

The new test was to verify for myself that the getter/setter isTenured() checks are actually required. Sadly, they are. The pre-barrier does a standard DoMarking, which for Shapes eagerly traverses children. Unfortunate. We are very close to being able to easily add a custom pre-barrier marker that uses a mark stack for all marking instead.

Comment 1

[Terrence Cole [:terrence]]

Attachment: 10.1_simplify_shape_tracing-v0.diff

Comment 2

[Steve Fink [:sfink] [:s:]]

Comment on attachment 8595005 [details] [diff] [review]
10.1_simplify_shape_tracing-v0.diff

Review of attachment 8595005 [details] [diff] [review]:
-----------------------------------------------------------------

::: js/src/gc/Marking.cpp
@@ +55,5 @@
> +// Callback
> +// --------
> +//
> +// The secondary JSTracer is the CallbackTracer. This simply invokes a callback
> +// on each edge in a child.

Ok, *now* that all makes sense. I was hung up on the distinction between visiting a node vs tracing its children, but the distinction the naming is based on is just marking vs generic callbacks. Thanks!

@@ +92,5 @@
> +//      |       '----------------------'     '----------------'    ||                           //
> +//      |                  |                         ||            ||                           //
> +//      |                  |                         ||            ||                           //
> +//      |                  |                         ||            ||                           //
> +//      o<-----------------o<========================OO============Oo                           //

Nice diagram!

@@ +567,1 @@
>  // to visit the child edges. In the absense of other traversal mechanisms, this

*absence

"fully generic traceChild virtual method", maybe? Or perhaps "fully generic" is redundant with "virtual method"?

@@ +1051,2 @@
>  
> +        // When triggered onbetween slices on belhalf of a barrier, these

"When triggered between slices on behalf..."

Comment 3

[Terrence Cole [:terrence]]

https://treeherder.mozilla.org/#/jobs?repo=try&revision=9d4c64043ab2

Comment 4

[Pulsebot]

https://hg.mozilla.org/integration/mozilla-inbound/rev/333aaffb8ad7

Comment 5

[Carsten Book [:Tomcat]]

https://hg.mozilla.org/mozilla-central/rev/333aaffb8ad7