Measure Arc<Locked<T>> fields properly.

[nnethercote]

Currently when we measure various Arc<Locked> fields we don't measure the T
itself, but only the descendants of the T. This patch fixes this.

This fix requires introducing a new trait, MallocUnconditionalShallowSizeOf,
which is implemented for servo_arc::Arc. A similar trait,
MallocConditionalShallowSizeOf, is also introduced, though it has no uses as
yet.

---

• ./mach build -d does not report any errors
• ./mach test-tidy does not report any errors
• These changes fix #__ (github issue number if applicable).

• There are tests for these changes OR
• These changes do not require tests because they are tested in Gecko.

---

This change is 

[highfive]

Heads up! This PR modifies the following files:

• @bholley: components/style/stylesheets/media_rule.rs, components/style/stylesheets/style_rule.rs, components/style/stylesheets/supports_rule.rs, components/style/stylesheets/page_rule.rs, components/style/stylesheets/stylesheet.rs and 2 more
• @canaltinova: components/style/stylesheets/media_rule.rs, components/style/stylesheets/style_rule.rs, components/style/stylesheets/supports_rule.rs, components/style/stylesheets/page_rule.rs, components/style/stylesheets/stylesheet.rs and 2 more
• @emilio: components/style/stylesheets/media_rule.rs, components/style/stylesheets/style_rule.rs, components/style/stylesheets/supports_rule.rs, components/style/stylesheets/page_rule.rs, components/style/stylesheets/stylesheet.rs and 2 more

[highfive]

Warning

• These commits modify style code, but no tests are modified. Please consider adding a test!

[nnethercote]

This is for https://bugzilla.mozilla.org/show_bug.cgi?id=1398854.

[nnethercote]

r? @jdm

[jdm]

@bors-servo: r+

[bors-servo]

📌 Commit 779fbda has been approved by jdm

[bors-servo]

⌛ Testing commit 779fbda with merge bffe158...

[bors-servo]

☀️ Test successful - android, arm32, arm64, linux-dev, linux-rel-css, linux-rel-wpt, mac-dev-unit, mac-rel-css1, mac-rel-css2, mac-rel-wpt1, mac-rel-wpt2, mac-rel-wpt3, mac-rel-wpt4, windows-msvc-dev
Approved by: jdm
Pushing bffe158 to master...

[bzbarsky]

So I'd been hoping we would do something where you could call foo.size_of() (possibly passing a guard) on an Arc<Locked<T>> and it would automatically do the right thing for you, instead of having to make two separate function calls to measure the size of the Locked<T> and the things hanging off the T. Seems like it would be a lot less error-prone....

[nnethercote]

So I'd been hoping we would do something where you could call foo.size_of() (possibly passing a guard) on an Arc<Locked> and it would automatically do the right thing for you

I tried that and it didn't work all that well. I created a new trait called MallocSizeOfWithGuard, because you have to pass in the guard so the normal MallocSizeOf trait doesn't work. But then, sometimes to measure the T you need to invoke MallocSizeOfWithGuard:

self.rules.unconditional_shallow_size_of(ops) + self.rules.read_with(guard).size_of(guard, ops)

but sometimes you need to invoke MallocSizeOf:

self.block.unconditional_shallow_size_of(ops) + self.block.read_with(guard).size_of(ops)

So you'd actually need two traits, or a single trait with two methods, or something.

If we end up measuring enough of these Arc<Locked<T>> things maybe it'll be worth doing that, but for the moment I decided it was simpler just to make the two separate function calls.

[bzbarsky]

So for what it's worth, I tried to implement my suggestion but using a single trait that is automatically implemented for Locked when T is MallocSizeOf, and ran into all sorts of issues with Rust's rules for traits. In particular, since MallocSizeOf is defined in one crate and MallocSizeOfWithGuard in another (so it can see the guard), Rust assumes that a single T can implement both. :(

And if I put MallocSizeOfWithGuard in the malloc_size_of crate, then making it possible for it to take a guard is rocket science. :(

[nnethercote]

I can believe it. One of the motivations for overhauling MallocSizeOf in https://bugzilla.mozilla.org/show_bug.cgi?id=1398737 was to make it more flexible in the cases where things don't fit neatly into traits. Hence the ability to do shallow measurements of types like Box, Arc, Vec, HashMap, etc. That's a big problem with the existing HeapSizeOf crate that Servo uses.

[ionutgoldan]

Looks like this merge brought some improvements:

== Change summary for alert #9405 (as of September 12 2017 14:45 UTC) ==

Improvements:

4% Heap Unclassified summary windows10-64 opt 48,812,667.91 -> 46,759,664.87
4% Heap Unclassified summary windows7-32 opt 41,291,712.81 -> 39,557,429.30
4% Heap Unclassified summary windows7-32 pgo 41,090,548.03 -> 39,442,801.26
3% Heap Unclassified summary linux64-stylo-sequential opt stylo-sequential56,742,017.56 -> 54,809,147.79
3% Heap Unclassified summary windows10-64 pgo 48,751,513.46 -> 47,116,870.64
3% Heap Unclassified summary linux64 opt 57,025,569.58 -> 55,134,734.22

For up to date results, see: https://treeherder.mozilla.org/perf.html#/alerts?id=9405