1118604 - Move activation stack to JSRuntime

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Shu-yu Guo [:shu]]

Without PJS, there's no reason for the activation stack to be allocated per thread.

Comment 1

[Lars T Hansen [:lth]]

I assume this means moving the activation-related members and methods from PerThreadData to JSRuntime.

Comment 2

[Shu-yu Guo [:shu]]

Yes, that was my intention. It's not clear it's useful to have PTD around still, but it's fairly well entrenched in parts of the code and I don't think it hurts to keep it around.

Comment 3

[Jan de Mooij [:jandem]]

(In reply to Shu-yu Guo [:shu] from comment #2)
> It's not clear it's useful to have PTD around
> still, but it's fairly well entrenched in parts of the code and I don't
> think it hurts to keep it around.

The helper threads can also have a PTD.

It seems we can also move these fields to JSRuntime: jitTop, jitJSContext, jitStackLimit_, regexpStack, traceLogger.

Comment 4

[Jan de Mooij [:jandem]]

(In reply to Jan de Mooij [:jandem] from comment #3)
> jitTop, jitJSContext, jitStackLimit_

Hm, it might also be nice to have these in JitRuntime instead of JSRuntime, but that's not very important and can wait.

Comment 5

[Jan de Mooij [:jandem]]

(In reply to Jan de Mooij [:jandem] from comment #3)
> traceLogger

Or maybe this one needs to be in PerThreadData for off-thread compilation. Sorry for the bug spam.

Comment 6

[Lars T Hansen [:lth]]

Attachment: WIP

This is probably roughly right -- it passes all --tbpl test so far -- but I'm still running tests and fixing misc problems.  I know this patch does not compile yet with the ARM and MIPS simulators, for example.

Comment 7

[Lars T Hansen [:lth]]

Attachment: WIP v2

This is probably clean.
Testing now: https://treeherder.mozilla.org/#/jobs?repo=try&revision=2f3d97d6c5f0

Comment 8

[Lars T Hansen [:lth]]

After this patch lands there will still be some types and methods related to activation iteration that take a PerThreadData* instead of a JSRuntime*.  We should investigate whether that's as it needs to be or if more cleanup is possible.  I suspect the latter.

Comment 9

[Lars T Hansen [:lth]]

Attachment: Move activation fields

Large patch but most of this is mechanical.

Full try run: https://treeherder.mozilla.org/#/jobs?repo=try&revision=50d40016139b

Comment 10

[Jan de Mooij [:jandem]]

Comment on attachment 8554526 [details] [diff] [review]
Move activation fields

Review of attachment 8554526 [details] [diff] [review]:
-----------------------------------------------------------------

Nice!

::: js/src/asmjs/AsmJSValidate.cpp
@@ +8509,5 @@
>          //   act.prevJitTop_ = cx->mainThread().jitTop;
>          //   act.prevJitJSContext_ = cx->mainThread().jitJSContext;
>          //   cx->mainThread().jitJSContext = cx;
>          //   act.prevJitActivation_ = cx->mainThread().jitActivation;
>          //   cx->mainThread().jitActivation = act;

Nit: please update these lines as well.

::: js/src/jit/arm/Simulator-arm.cpp
@@ +4438,5 @@
>  
> +void
> +js::PerThreadData::setSimulator(js::jit::Simulator *sim)
> +{
> +    runtime_->setSimulator(sim);

We can now simplify the simulator(s) by merging Simulator and SimulatorRuntime. I think we can just move everything in SimulatorRuntime to Simulator. Maybe we can then rename CreateSimulatorRuntime to CreateSimulator, that way we can assume we always have a non-null simulator in Simulator::Current().

We can also remove AutoLockSimulatorRuntime, might be nice for performance.

Follow-up bug :)

::: js/src/vm/Stack.cpp
@@ +1693,5 @@
>  {
>      settle();
>  }
>  
>  ActivationIterator::ActivationIterator(PerThreadData *perThreadData)

Nit: is this constructor still necessary?

Comment 11

[Lars T Hansen [:lth]]

(In reply to Jan de Mooij [:jandem] from comment #10)
> Comment on attachment 8554526 [details] [diff] [review]
> Move activation fields
> 
> We can now simplify the simulator(s) by merging Simulator and SimulatorRuntime.
> ...
> We can also remove AutoLockSimulatorRuntime, might be nice for performance.
> 
> Follow-up bug :)

Thank you for that clarification :)  I will file it.

> ::: js/src/vm/Stack.cpp
> @@ +1693,5 @@
> >  {
> >      settle();
> >  }
> >  
> >  ActivationIterator::ActivationIterator(PerThreadData *perThreadData)
> 
> Nit: is this constructor still necessary?

I don't know.  It is still being referenced from the GC, but whether it is ultimately necessary to have it depends on some details of the GC that I haven't investigated yet.  (I could have sworn I had filed a bug about that but I have not, so I will now.  Thanks.)

Comment 12

[Lars T Hansen [:lth]]

Pushed with updates to AsmJSValidate.cpp:
https://hg.mozilla.org/integration/mozilla-inbound/rev/d590c1f472b2

Comment 13

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/d590c1f472b2