This patch provides thread-safe write barriers and extends GC::MarkQueueAndStack to mark all mutator's stacks. Thread-safe write-barriers are trivial, given that GC::IsMarkedThenMakeQueued has been made thread-safe with atomic operations in bug 623618. All else that this patch has to add is: - Protect slow-path mutator access to m_barrierWork and m_incrementalWork with critical sections. - Ensure that GC::marking cannot change under the feet of a mutator due to a safepoint task. I.e. we need to ensure the following: GC::foo() { if (marking) { // Uninterruptable by safepoint tasks that may change the value of 'marking' } } GC::MarkQueueAndStack is extended to mark all mutator's stacks by iterating over all SafepointRecords known to the GC's SafepointManager (except the GC-owner thread's), and adding their stack extents as GCWorkItems. Note that as GC::MarkQueueAndStack is called from within a safepoint task, then all mutators will be safe.

Slight tweak: a SCOPE_LOCK_SP should have been a MMGC_SCOPE_LOCK

Made GCPolicyManager::barrierStageLastCollection thread-safe (via an AtomicCounter32) (This had found its way into another patch by mistake)

The switch of the type for barrierStageLastCollection introduces somewhat large overhead on allocation heavy benchmarks. Here are asmicro/alloc-* performance results comparing non-atomic and atomic counter types. non-atomic-ctrs: ../../objdir-rel32/shell/avmshell.isolatemutpaths-031disable-atomic-barrierstagecounts version: cyclone atomic-ctrs: ../../objdir-rel32/shell/avmshell.isolatemutpaths-021disable-wbrc-asserts version: cyclone iterations: 25 non-atomic-ctrs atomic-ctrs test best avg best avg %dBst %dAvg Metric: iterations/second Dir: asmicro/ alloc-1 41.9 41.8 40.5 40.5 -3.3 -3.3 - alloc-10 14.9 14.8 14.3 14.3 -3.7 -3.8 - alloc-11 15.3 15.2 14.9 14.9 -2.2 -2.1 - alloc-12 15.3 15.3 15.0 14.9 -2.4 -2.3 - alloc-13 86.3 86.1 84.5 84.4 -2.1 -2.0 - alloc-14 72.9 72.4 71.1 70.7 -2.5 -2.4 - alloc-2 20.4 20.4 20.0 20.0 -2.2 -2.1 - alloc-3 17.3 17.3 16.6 16.4 -4.2 -5.0 - alloc-4 46.8 46.8 45.5 45.4 -2.9 -2.9 - alloc-5 33.9 33.9 33.8 33.7 -0.5 -0.5 alloc-6 70.2 70.0 68.7 68.6 -2.0 -2.0 - alloc-7 39.1 39.1 37.8 37.8 -3.4 -3.4 - alloc-8 15.9 15.9 15.3 15.3 -3.9 -4.0 - alloc-9 15.9 15.9 15.3 15.3 -4.0 -4.0 -

(In reply to comment #3) > The switch of the type for barrierStageLastCollection introduces somewhat large > overhead on allocation heavy benchmarks. Here are asmicro/alloc-* performance > results comparing non-atomic and atomic counter types. Clarification: The overhead described would only be introduced into builds with MMGC_POLICY_PROFILING turned on. From what I understand, this means includes the AVM shell, but not the Flash player. That could mean that we should turn that feature switch off when doing performance evaluation for AsymmGC; or it could mean that we still need to work around this issue some other way.

Comment typo: 'with in' -> 'within'

Retargeting to Dolores.

(unassigning self; we may or may not land this, but we have other fish to fry in the meantime.)

Tamarin isn't maintained anymore. WONTFIX remaining bugs.