837014 - IonMonkey: Polymorphic inline for deltablue Plan.execute()

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Sean Stangl [:sstangl]]

]Deltablue is divided into independent halves, chainTest() and projectionTest(), each of which is roughly a SunSpider-style benchmark. We are about equal with v8 on projectionTest() (52000 versus 53000), but slower on chainTest() (40000 versus 55000).

On the chainTest() half, we miss an important polymorphic inlining opportunity.

This is the main loop of chainTest():
> Plan.prototype.execute = function () {
>   for (var i = 0; i < this.size(); i++) {
>     var c = this.constraintAt(i);
>     c.execute();
>   }
> }

The constraint objects in play are one of StayConstraint, EditConstraint, and EqualityConstraint. Their .execute() functions are defined as follows:

> StayConstraint.prototype.execute = function () {
>   // Stay constraints do nothing
> }

> EditConstraint.prototype.execute = function () {
>   // Edit constraints do nothing
> }

> EqualityConstraint.prototype.execute = function () {
>   this.output().value = this.input().value;
> }

The benchmark sets up these objects into a chain such that there are N EqualityConstraints but only 1 StayConstraint at the end and a few interspersed EditConstraints. Because of this, when we go to inline, we find that the three given targets have the following useCounts:

> StayConstraint: 1
> EditConstraint: 122 (saved by js_IonOptions.inlineUseCountRatio)
> EqualityConstraint: 10187

And so the Oracle vetos the inlining due to the one low useCount, and the main loop winds up using visitCallGeneric(). As you are noticing, this is very silly, because that function doesn't do anything.

There are a few options to solve this in a more generic way, but the easiest solution is to just check the script length and ignore the useCount entirely if the script is trivial (length 1).

There's also an MGetPropertyCache that shouldn't be there after inlining, for follow-up work.

Comment 1

[Sean Stangl [:sstangl]]

Attachment: Always inline the empty script.

chaintest(): 40000 -> 56000 (v8 55000)
deltablue: ~20150 -> ~20500 (v8 23571)

The benchmark is divided into halves: we are now faster than v8 on the former and equally fast on the latter, but slower on the whole. Possibly some bad information is bleeding through the barrier between them.

Comment 2

[Sean Stangl [:sstangl]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/0e7e657e7739

Comment 3

[Phil Ringnalda (:philor)]

https://hg.mozilla.org/mozilla-central/rev/0e7e657e7739