Ignore:
Timestamp:
Oct 12, 2015, 3:41:01 PM (10 years ago)
Author:
[email protected]
Message:

GC should have a Baker barrier for concurrent copying
https://bugs.webkit.org/show_bug.cgi?id=149852

Reviewed by Geoffrey Garen.

This adds a Baker-style read barrier [1] to copied space accesses. This barrier incurs some
overhead (0%-2% depending on benchmark suite), but what it buys is the ability to make the GC copy
phase concurrent.

The barrier relies on copied space pointers having two "space bits" in the low pointer bits. The
space bits indicate whether the backing store is being copied right now or not, and if it is being
copied, what stage of copying it's in. Two barrier variants are supported:

Read only barrier: if you load a backing store and immediately load from it without doing anything
else, you can just mask off the bits. In the worst case, you'll get the old backing store while
some copying thread is already allocating and populating the new version of the backing store. But
in that case, forwarding to the new backing store will not enable you to load a more up-to-date
value from the backing store. So, just masking the bits is enough. The read-only barrier is only
used in ICs where we know that we are only reading, and opportunistically within the DFG and FTL
thanks to the CopyBarrierOptimizationPhase. We never explicitly emit a read-only barrier in those
compilers; instead the phase will turn a GetButterfly into GetButterflyReadOnly if it proves that a
bunch of requirements are met.

Normal barrier: if the space bits are non-zero, call a slow path. The slow path will either do
nothing (if the copy phase hasn't started yet), or it will copy the backing store and update the
pointer (if the copy phase hasn't gotten around to copying this particular backing store), or it
will wait for the copying thread to finish (if some thread is copying this backing store right
now), or it will do nothing (if by the time we called into the slow path the backing store was
already copied). This is just like Baker's CAR/CDR barrier, but with a lock thrown in to handle
concurrent execution.

This is a 1% slow-down on SunSpider, a 1.5% slow-down on Octane, a 1.5% slow-down on Kraken, and a
0% slow-down on AsmBench. Note that the Octane slow-down is excluding the SplayLatency benchmark.
That benchmark will eventually speed up a lot once we finish doing all of this stuff. Probably, the
JetStream splay-latency will see an even larger speed-up, since our version of the latency tests do
a better job of punishing bad worst-case behavior.

[1] http://dspace.mit.edu/bitstream/handle/1721.1/41976/AI_WP_139.pdf, look for the CAR and CDR
procedures on page 9.

(JSC::AccessCase::generate):

  • dfg/DFGAbstractInterpreterInlines.h:

(JSC::DFG::AbstractInterpreter<AbstractStateType>::executeEffects):

  • dfg/DFGArgumentsEliminationPhase.cpp:
  • dfg/DFGClobberize.h:

(JSC::DFG::clobberize):

  • dfg/DFGCopyBarrierOptimizationPhase.cpp: Added.

(JSC::DFG::performCopyBarrierOptimization):

  • dfg/DFGCopyBarrierOptimizationPhase.h: Added.
  • dfg/DFGDoesGC.cpp:

(JSC::DFG::doesGC):

  • dfg/DFGFixupPhase.cpp:

(JSC::DFG::FixupPhase::fixupNode):

  • dfg/DFGHeapLocation.cpp:

(WTF::printInternal):

  • dfg/DFGHeapLocation.h:
  • dfg/DFGLICMPhase.cpp:

(JSC::DFG::LICMPhase::run):

  • dfg/DFGNodeType.h:
  • dfg/DFGOperations.cpp:
  • dfg/DFGOperations.h:
  • dfg/DFGPlan.cpp:

(JSC::DFG::Plan::compileInThreadImpl):

  • dfg/DFGPredictionPropagationPhase.cpp:

(JSC::DFG::PredictionPropagationPhase::propagate):

  • dfg/DFGSafeToExecute.h:

(JSC::DFG::safeToExecute):

  • dfg/DFGSpeculativeJIT.cpp:

(JSC::DFG::SpeculativeJIT::compileGetIndexedPropertyStorage):
(JSC::DFG::SpeculativeJIT::compileGetTypedArrayByteOffset):
(JSC::DFG::SpeculativeJIT::compileReallocatePropertyStorage):
(JSC::DFG::SpeculativeJIT::compileGetButterfly):
(JSC::DFG::SpeculativeJIT::temporaryRegisterForPutByVal):

  • dfg/DFGSpeculativeJIT.h:
  • dfg/DFGSpeculativeJIT32_64.cpp:

(JSC::DFG::SpeculativeJIT::compile):

  • dfg/DFGSpeculativeJIT64.cpp:

(JSC::DFG::SpeculativeJIT::compile):

  • dfg/DFGTypeCheckHoistingPhase.cpp:

(JSC::DFG::TypeCheckHoistingPhase::identifyRedundantStructureChecks):
(JSC::DFG::TypeCheckHoistingPhase::identifyRedundantArrayChecks):

  • ftl/FTLCapabilities.cpp:

(JSC::FTL::canCompile):

  • ftl/FTLLowerDFGToLLVM.cpp:

(JSC::FTL::DFG::LowerDFGToLLVM::compileNode):
(JSC::FTL::DFG::LowerDFGToLLVM::compileGetButterfly):
(JSC::FTL::DFG::LowerDFGToLLVM::compileGetButterflyReadOnly):
(JSC::FTL::DFG::LowerDFGToLLVM::compileConstantStoragePointer):
(JSC::FTL::DFG::LowerDFGToLLVM::compileGetIndexedPropertyStorage):
(JSC::FTL::DFG::LowerDFGToLLVM::compileCheckArray):
(JSC::FTL::DFG::LowerDFGToLLVM::compileGetTypedArrayByteOffset):
(JSC::FTL::DFG::LowerDFGToLLVM::compileMultiGetByOffset):
(JSC::FTL::DFG::LowerDFGToLLVM::compileMultiPutByOffset):
(JSC::FTL::DFG::LowerDFGToLLVM::compileGetDirectPname):
(JSC::FTL::DFG::LowerDFGToLLVM::storageForTransition):
(JSC::FTL::DFG::LowerDFGToLLVM::getById):
(JSC::FTL::DFG::LowerDFGToLLVM::loadButterflyWithBarrier):
(JSC::FTL::DFG::LowerDFGToLLVM::loadVectorWithBarrier):
(JSC::FTL::DFG::LowerDFGToLLVM::copyBarrier):
(JSC::FTL::DFG::LowerDFGToLLVM::loadButterflyReadOnly):
(JSC::FTL::DFG::LowerDFGToLLVM::loadVectorReadOnly):
(JSC::FTL::DFG::LowerDFGToLLVM::removeSpaceBits):
(JSC::FTL::DFG::LowerDFGToLLVM::baseIndex):

  • ftl/FTLOperations.cpp:

(JSC::FTL::operationNewObjectWithButterfly):
(JSC::FTL::operationPopulateObjectInOSR):

  • ftl/FTLOutput.h:

(JSC::FTL::Output::testNonZero32):
(JSC::FTL::Output::testIsZero64):
(JSC::FTL::Output::testNonZero64):
(JSC::FTL::Output::testIsZeroPtr):
(JSC::FTL::Output::testNonZeroPtr):
(JSC::FTL::Output::select):
(JSC::FTL::Output::extractValue):

  • heap/CopyBarrier.h: Copied from Source/JavaScriptCore/heap/CopyWriteBarrier.h.

(JSC::CopyBarrierBase::CopyBarrierBase):
(JSC::CopyBarrierBase::operator!):
(JSC::CopyBarrierBase::operator bool):
(JSC::CopyBarrierBase::getWithoutBarrier):
(JSC::CopyBarrierBase::get):
(JSC::CopyBarrierBase::copyState):
(JSC::CopyBarrierBase::setCopyState):
(JSC::CopyBarrierBase::clear):
(JSC::CopyBarrierBase::set):
(JSC::CopyBarrierBase::setWithoutBarrier):
(JSC::CopyBarrierBase::weakCASWithoutBarrier):
(JSC::CopyBarrier::CopyBarrier):
(JSC::CopyBarrier::getWithoutBarrier):
(JSC::CopyBarrier::get):
(JSC::CopyBarrier::set):
(JSC::CopyBarrier::setWithoutBarrier):
(JSC::CopyBarrier::weakCASWithoutBarrier):
(JSC::CopyWriteBarrier::CopyWriteBarrier): Deleted.
(JSC::CopyWriteBarrier::operator!): Deleted.
(JSC::CopyWriteBarrier::operator bool): Deleted.
(JSC::CopyWriteBarrier::get): Deleted.
(JSC::CopyWriteBarrier::operator*): Deleted.
(JSC::CopyWriteBarrier::operator->): Deleted.
(JSC::CopyWriteBarrier::set): Deleted.
(JSC::CopyWriteBarrier::setWithoutWriteBarrier): Deleted.
(JSC::CopyWriteBarrier::clear): Deleted.

  • heap/CopyVisitorInlines.h:

(JSC::CopyVisitor::checkIfShouldCopy):

  • heap/CopyWriteBarrier.h: Removed.
  • heap/Heap.cpp:

(JSC::Heap::addToRememberedSet):
(JSC::Heap::copyBarrier):
(JSC::Heap::collectAndSweep):

  • heap/Heap.h:

(JSC::Heap::writeBarrierBuffer):

  • heap/HeapInlines.h:
  • jit/AssemblyHelpers.h:

(JSC::AssemblyHelpers::branchStructure):
(JSC::AssemblyHelpers::branchIfNotToSpace):
(JSC::AssemblyHelpers::removeSpaceBits):
(JSC::AssemblyHelpers::addressForByteOffset):

  • jit/JIT.cpp:

(JSC::JIT::privateCompileMainPass):
(JSC::JIT::privateCompileSlowCases):

  • jit/JITOpcodes.cpp:

(JSC::JIT::emitSlow_op_has_indexed_property):
(JSC::JIT::emit_op_get_direct_pname):
(JSC::JIT::emitSlow_op_get_direct_pname):

  • jit/JITOpcodes32_64.cpp:

(JSC::JIT::emit_op_get_direct_pname):
(JSC::JIT::emitSlow_op_get_direct_pname):

  • jit/JITPropertyAccess.cpp:

(JSC::JIT::emitDoubleLoad):
(JSC::JIT::emitContiguousLoad):
(JSC::JIT::emitArrayStorageLoad):
(JSC::JIT::emitSlow_op_get_by_val):
(JSC::JIT::emitGenericContiguousPutByVal):
(JSC::JIT::emitArrayStoragePutByVal):
(JSC::JIT::emitSlow_op_put_by_val):
(JSC::JIT::emit_op_get_from_scope):
(JSC::JIT::emitSlow_op_get_from_scope):
(JSC::JIT::emit_op_put_to_scope):
(JSC::JIT::emitSlow_op_put_to_scope):
(JSC::JIT::emitIntTypedArrayGetByVal):
(JSC::JIT::emitFloatTypedArrayGetByVal):
(JSC::JIT::emitIntTypedArrayPutByVal):
(JSC::JIT::emitFloatTypedArrayPutByVal):

  • llint/LowLevelInterpreter.asm:
  • llint/LowLevelInterpreter64.asm:
  • runtime/DirectArguments.cpp:

(JSC::DirectArguments::visitChildren):
(JSC::DirectArguments::copyBackingStore):
(JSC::DirectArguments::overrideThings):
(JSC::DirectArguments::overrideThingsIfNecessary):
(JSC::DirectArguments::overrideArgument):
(JSC::DirectArguments::copyToArguments):

  • runtime/DirectArguments.h:

(JSC::DirectArguments::canAccessIndexQuickly):
(JSC::DirectArguments::canAccessArgumentIndexQuicklyInDFG):

  • runtime/JSArray.cpp:

(JSC::JSArray::setLength):
(JSC::JSArray::pop):
(JSC::JSArray::push):
(JSC::JSArray::fastSlice):
(JSC::JSArray::fastConcatWith):
(JSC::JSArray::shiftCountWithArrayStorage):
(JSC::JSArray::shiftCountWithAnyIndexingType):
(JSC::JSArray::unshiftCountWithAnyIndexingType):
(JSC::JSArray::fillArgList):
(JSC::JSArray::copyToArguments):

  • runtime/JSArrayBufferView.cpp:

(JSC::JSArrayBufferView::ConstructionContext::ConstructionContext):
(JSC::JSArrayBufferView::JSArrayBufferView):
(JSC::JSArrayBufferView::finishCreation):
(JSC::JSArrayBufferView::finalize):

  • runtime/JSArrayBufferView.h:

(JSC::JSArrayBufferView::vector):
(JSC::JSArrayBufferView::length):

  • runtime/JSArrayBufferViewInlines.h:

(JSC::JSArrayBufferView::neuter):
(JSC::JSArrayBufferView::byteOffset):

  • runtime/JSGenericTypedArrayView.h:

(JSC::JSGenericTypedArrayView::typedVector):

  • runtime/JSGenericTypedArrayViewInlines.h:

(JSC::JSGenericTypedArrayView<Adaptor>::visitChildren):
(JSC::JSGenericTypedArrayView<Adaptor>::copyBackingStore):
(JSC::JSGenericTypedArrayView<Adaptor>::slowDownAndWasteMemory):

  • runtime/JSMap.h:

(JSC::JSMap::JSMap):

  • runtime/JSObject.cpp:

(JSC::JSObject::copyButterfly):
(JSC::JSObject::visitChildren):
(JSC::JSObject::copyBackingStore):
(JSC::JSObject::getOwnPropertySlotByIndex):
(JSC::JSObject::putByIndex):
(JSC::JSObject::enterDictionaryIndexingMode):
(JSC::JSObject::createInitialIndexedStorage):
(JSC::JSObject::createArrayStorage):
(JSC::JSObject::convertUndecidedToInt32):
(JSC::JSObject::convertUndecidedToDouble):
(JSC::JSObject::convertUndecidedToContiguous):
(JSC::JSObject::constructConvertedArrayStorageWithoutCopyingElements):
(JSC::JSObject::convertUndecidedToArrayStorage):
(JSC::JSObject::convertInt32ToDouble):
(JSC::JSObject::convertInt32ToContiguous):
(JSC::JSObject::convertInt32ToArrayStorage):
(JSC::JSObject::convertDoubleToContiguous):
(JSC::JSObject::convertDoubleToArrayStorage):
(JSC::JSObject::convertContiguousToArrayStorage):
(JSC::JSObject::setIndexQuicklyToUndecided):
(JSC::JSObject::ensureArrayStorageExistsAndEnterDictionaryIndexingMode):
(JSC::JSObject::deletePropertyByIndex):
(JSC::JSObject::getOwnPropertyNames):
(JSC::JSObject::putIndexedDescriptor):
(JSC::JSObject::defineOwnIndexedProperty):
(JSC::JSObject::putByIndexBeyondVectorLengthWithoutAttributes):
(JSC::JSObject::putDirectIndexBeyondVectorLength):
(JSC::JSObject::getNewVectorLength):
(JSC::JSObject::ensureLengthSlow):
(JSC::JSObject::reallocateAndShrinkButterfly):
(JSC::JSObject::growOutOfLineStorage):
(JSC::JSObject::getOwnPropertyDescriptor):
(JSC::JSObject::getEnumerableLength):

  • runtime/JSObject.h:

(JSC::JSObject::getArrayLength):
(JSC::JSObject::getVectorLength):
(JSC::JSObject::canGetIndexQuickly):
(JSC::JSObject::getIndexQuickly):
(JSC::JSObject::tryGetIndexQuickly):
(JSC::JSObject::canSetIndexQuickly):
(JSC::JSObject::canSetIndexQuicklyForPutDirect):
(JSC::JSObject::setIndexQuickly):
(JSC::JSObject::initializeIndex):
(JSC::JSObject::hasSparseMap):
(JSC::JSObject::inSparseIndexingMode):
(JSC::JSObject::inlineStorage):
(JSC::JSObject::butterfly):
(JSC::JSObject::outOfLineStorage):
(JSC::JSObject::locationForOffset):
(JSC::JSObject::ensureInt32):
(JSC::JSObject::ensureDouble):
(JSC::JSObject::ensureContiguous):
(JSC::JSObject::ensureArrayStorage):
(JSC::JSObject::arrayStorage):
(JSC::JSObject::arrayStorageOrNull):
(JSC::JSObject::ensureLength):
(JSC::JSObject::putDirectWithoutTransition):

  • runtime/JSSet.h:

(JSC::JSSet::JSSet):

  • runtime/MapData.h:

(JSC::JSIterator>::MapDataImpl):
(JSC::JSIterator>::IteratorData::next):
(JSC::JSIterator>::IteratorData::refreshCursor):

  • runtime/MapDataInlines.h:

(JSC::JSIterator>::clear):
(JSC::JSIterator>::find):
(JSC::JSIterator>::add):
(JSC::JSIterator>::remove):
(JSC::JSIterator>::replaceAndPackBackingStore):
(JSC::JSIterator>::replaceBackingStore):
(JSC::JSIterator>::ensureSpaceForAppend):
(JSC::JSIterator>::visitChildren):
(JSC::JSIterator>::copyBackingStore):

  • runtime/Options.h:
File:
1 edited

Legend:

Unmodified
Added
Removed
  • trunk/Source/JavaScriptCore/runtime/DirectArguments.cpp

    r190563 r190896  
    9898        visitor.copyLater(
    9999            thisObject, DirectArgumentsOverridesCopyToken,
    100             thisObject->m_overrides.get(), thisObject->overridesSize());
     100            thisObject->m_overrides.getWithoutBarrier(), thisObject->overridesSize());
    101101    }
    102102}
     
    109109    RELEASE_ASSERT(token == DirectArgumentsOverridesCopyToken);
    110110   
    111     bool* oldOverrides = thisObject->m_overrides.get();
    112     if (!oldOverrides)
    113         return;
    114    
     111    void* oldOverrides = thisObject->m_overrides.getWithoutBarrier();
    115112    if (visitor.checkIfShouldCopy(oldOverrides)) {
    116113        bool* newOverrides = static_cast<bool*>(visitor.allocateNewSpace(thisObject->overridesSize()));
    117114        memcpy(newOverrides, oldOverrides, thisObject->m_length);
    118         thisObject->m_overrides.setWithoutWriteBarrier(newOverrides);
     115        thisObject->m_overrides.setWithoutBarrier(newOverrides);
    119116        visitor.didCopy(oldOverrides, thisObject->overridesSize());
    120117    }
     
    136133    void* backingStore;
    137134    RELEASE_ASSERT(vm.heap.tryAllocateStorage(this, overridesSize(), &backingStore));
    138     m_overrides.set(vm, this, static_cast<bool*>(backingStore));
     135    bool* overrides = static_cast<bool*>(backingStore);
     136    m_overrides.set(vm, this, overrides);
    139137    for (unsigned i = m_length; i--;)
    140         m_overrides.get()[i] = false;
     138        overrides[i] = false;
    141139}
    142140
     
    150148{
    151149    overrideThingsIfNecessary(vm);
    152     m_overrides.get()[index] = true;
     150    m_overrides.get(this)[index] = true;
    153151}
    154152
Note: See TracChangeset for help on using the changeset viewer.