1466458 - Refactor Realm::enterRealmDepth_ a bit

Status: RESOLVED FIXED

(Core :: JavaScript Engine, enhancement)

Description

[Jan de Mooij [:jandem]]

The JITs will be able to switch to a different same-compartment realm before a function call. We want this to be just a single |cx->realm_ = realm;| store, but AutoRealm currently also increments Realm::enterRealmDepth_ - we don't have to do that in JIT code but it's subtle and needs some comments and minor refactoring to make this more obvious.

Comment 1

[Jan de Mooij [:jandem]]

Attachment: Part 1 - Remove JSContext::enterRealmDepth_

Comment 2

[Jan de Mooij [:jandem]]

Attachment: Part 2 - Refactor Realm::enterRealmDepth_ to account for Realms entered from JIT code

This renames Realm::enterRealmDepth_ to Realm::enterRealmDepthIgnoringJit_, to make it more clear that it's only bumped for C++ AutoRealm and not for JIT code Realm entering.

Fortunately, that's all we need:

* enterRealmDepthIgnoringJit_ is sufficient for Realm::shouldTraceGlobal - when the JIT enters a different realm, there must be a frame/script/function on the stack for that realm and this will ensure the realm's global is traced anyway.

* We don't relazify functions in realms that are active. Here I added a JSCompartment::hasEnteredRealm flag we set during GC - key here is that the JIT can only enter a *same-compartment* realm, so there must always be a realm for that compartment that has been entered from C++/AutoRealm.

It's subtle, but I hope the comments and refactoring make it a bit clearer.

Comment 3

[Luke Wagner [:luke]]

Comment on attachment 8982990 [details] [diff] [review]
Part 2 - Refactor Realm::enterRealmDepth_ to account for Realms entered from JIT code

Review of attachment 8982990 [details] [diff] [review]:
-----------------------------------------------------------------

Great analysis comments and factoring.  Agreed.

::: js/src/vm/JSCompartment.h
@@ +584,5 @@
> +    // compartment. Note that (without a stack walk) we don't know exactly
> +    // *which* realms, because Realm::enterRealmDepthIgnoringJit_ does not
> +    // account for cross-Realm calls in JIT code updating cx->realm_. See also
> +    // the enterRealmDepthIgnoringJit_ comment.
> +    bool hasEnteredRealm = false;

semi-pre-existing: I was wondering how a non-counter bool was maintained in the presence of reentrance but I see that this is computed per-GC.  I was wondering if this field and the above two could be commented together in a block that says: these are only for temporary GC use and only meaningful at particular points in the GC (e.g., it seems like running a slice would make their meaning stale).

(In theory, these fields could be removed, using an external HashSet or something, but that's a separate question.)

Comment 4

[Jan de Mooij [:jandem]]

(In reply to Luke Wagner [:luke] from comment #3)
> semi-pre-existing: I was wondering how a non-counter bool was maintained in
> the presence of reentrance but I see that this is computed per-GC.  I was
> wondering if this field and the above two could be commented together in a
> block that says: these are only for temporary GC use and only meaningful at
> particular points in the GC (e.g., it seems like running a slice would make
> their meaning stale).

What do you think about grouping them in a struct? Something like:

class JSCompartment
{
    ...
  public:
    struct {
        bool maybeLive = true;
        ...
        bool hasEnteredRealm = false;
    } flagsSetByGC;
};

comp->flagsSetByGC.hasEnteredRealm might make it clearer these are set by the GC.

Other names are gcFlags, gcState...

Comment 5

[Luke Wagner [:luke]]

I really like that idea.  gcState sgtm (I was going to suggest gcTemporalState, but the comments could speak to the temporality clearly enough).

Comment 6

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/mozilla-inbound/rev/b968ba27d88c
part 1 - Remove JSContext::enterRealmDepth_. r=luke
https://hg.mozilla.org/integration/mozilla-inbound/rev/60828108f4b8
part 2 - Refactor Realm::enterRealmDepth_ to account for Realms entered from JIT code. r=luke

Comment 7

[Andrei Ciure[:aciure]]

https://hg.mozilla.org/mozilla-central/rev/b968ba27d88c
https://hg.mozilla.org/mozilla-central/rev/60828108f4b8