937674 - Don't instantiate types when adding type object guards around type write barriers

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, defect)

Description

[Brian Hackett [Laid off!]]

Attachment: patch

This is a use of JSContext through GetIonContext() that hasn't been eliminated yet.

Comment 1

[Jan de Mooij [:jandem]]

Comment on attachment 830880 [details] [diff] [review]
patch

Review of attachment 830880 [details] [diff] [review]:
-----------------------------------------------------------------

::: js/src/jit/Lowering.cpp
@@ +2833,5 @@
>  bool
> +LIRGenerator::visitGuardObjectIdentity(MGuardObjectIdentity *ins)
> +{
> +    LGuardObjectIdentity *guard = new LGuardObjectIdentity(useRegister(ins->obj()));
> +    return assignSnapshot(guard) && add(guard, ins);

MGuardObjectIdentity redefines its operand, so I think this should be:

return assignSnapshot(guard) && redefine(ins, ins->obj()) && add(guard, ins);

::: js/src/jit/MIR.h
@@ +6892,5 @@
> +class MGuardObjectIdentity
> +  : public MUnaryInstruction,
> +    public SingleObjectPolicy
> +{
> +    CompilerRoot<JSObject *> singleObject_;

Just curious, do we still need this in the no-GC-during-compilation world?

Comment 2

[Brian Hackett [Laid off!]]

Currently, the only thing I think the compiler roots help with is asserting that we don't end up with nursery pointers in the Ion data structures in GGC builds.

https://hg.mozilla.org/integration/mozilla-inbound/rev/2eaab43b65d4

Comment 3

[Wes Kocher (:KWierso) (Not reading bugmail; email directly if needed)]

https://hg.mozilla.org/mozilla-central/rev/2eaab43b65d4