981554 - Move allocator for optimized Baseline stubs from JitCompartment to JitZone

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, defect)

Description

[Jan de Mooij [:jandem]]

Every compartment with Baseline code has a LifoAlloc (with 4K default chunk size) for optimized stubs.

Although this allocator is purged on every code-discarding-GC, it'd be nice to move it from compartment to zone (as every allocator will have a chunk with some unused space, and there's also the size of the LifoAlloc itself).

Comment 1

[Jan de Mooij [:jandem]]

Attachment: Patch

This patch adds JitZone (like JitCompartment and JitRuntime), and moves the allocator to it.

~AutoDebugModeInvalidation only discards JIT code for some compartments, so we can no longer free the LifoAlloc there, but that should be okay as long as Zone::discardJitCode does this.

There's probably more stuff in JitCompartment that we can move to JitZone; I'll look into it.

Comment 2

[Nicholas Nethercote [inactive]]

Comment on attachment 8388474 [details] [diff] [review]
Patch

Review of attachment 8388474 [details] [diff] [review]:
-----------------------------------------------------------------

Looks good.

::: js/src/gc/Zone.h
@@ +18,5 @@
>  
>  namespace js {
>  
> +namespace jit {
> +    class JitZone;

Nit: I think we usually don't indent this case.

::: js/xpconnect/src/XPCJSRuntime.cpp
@@ +2076,5 @@
>      ZCREPORT_BYTES(cJSPathPrefix + NS_LITERAL_CSTRING("baseline/data"),
>          cStats.baselineData,
>          "The Baseline JIT's compilation data (BaselineScripts).");
>  
> +    ZCREPORT_BYTES(cJSPathPrefix + NS_LITERAL_CSTRING("baseline/fallback-stubs"),

Is there a reason you moved the optimized stubs but kept the fallback stubs?

Comment 3

[Jan de Mooij [:jandem]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/4197edad59d6

(In reply to Nicholas Nethercote [:njn] from comment #2)
> Is there a reason you moved the optimized stubs but kept the fallback stubs?

The fallback stubs are allocated per BaselineScript, so we can't move them as easily.

We could add a per-Zone allocator for them (would be more memory-efficient), but then we can only purge it when there's no Baseline code for this Zone on the stack, so I'm not sure if that's better.

Comment 4

[Jan de Mooij [:jandem]]

Oh and this was green on Linux x64 a few days ago:

https://tbpl.mozilla.org/?tree=Try&rev=7b3254153f6e

Comment 5

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/4197edad59d6