Changeset 98937 in webkit for trunk/Source/JavaScriptCore/wtf/Atomics.h

Timestamp:

Oct 31, 2011, 11:43:37 PM (14 years ago)

Author:

[email protected]

Message:

The GC should be parallel
​https://bugs.webkit.org/show_bug.cgi?id=70995

Source/JavaScriptCore:

Reviewed by Geoff Garen.

Added parallel tracing to the GC. This works by having local mark
stacks per thread, and a global shared one. Threads sometimes
donate cells from the mark stack to the global one if the heuristics
tell them that it's affordable to do so. Threads that have depleted
their local mark stacks try to steal some from the shared one.

Marking is now done using an atomic weak relaxed CAS (compare-and-swap).

This is a 23% speed-up on V8-splay when I use 4 marking threads,
leading to a 3.5% speed-up on V8.

It also appears that this reduces GC pause times on real websites by
more than half.

• JavaScriptCore.exp:
• JavaScriptCore.vcproj/JavaScriptCore/JavaScriptCore.def:
• heap/Heap.cpp:

(JSC::Heap::Heap):
(JSC::Heap::~Heap):
(JSC::Heap::markRoots):

• heap/Heap.h:
• heap/MarkStack.cpp:

(JSC::MarkStackSegmentAllocator::MarkStackSegmentAllocator):
(JSC::MarkStackSegmentAllocator::~MarkStackSegmentAllocator):
(JSC::MarkStackSegmentAllocator::allocate):
(JSC::MarkStackSegmentAllocator::release):
(JSC::MarkStackSegmentAllocator::shrinkReserve):
(JSC::MarkStackArray::MarkStackArray):
(JSC::MarkStackArray::~MarkStackArray):
(JSC::MarkStackArray::expand):
(JSC::MarkStackArray::refill):
(JSC::MarkStackArray::donateSomeCellsTo):
(JSC::MarkStackArray::stealSomeCellsFrom):
(JSC::MarkStackThreadSharedData::markingThreadMain):
(JSC::MarkStackThreadSharedData::markingThreadStartFunc):
(JSC::MarkStackThreadSharedData::MarkStackThreadSharedData):
(JSC::MarkStackThreadSharedData::~MarkStackThreadSharedData):
(JSC::MarkStackThreadSharedData::reset):
(JSC::MarkStack::reset):
(JSC::SlotVisitor::donateSlow):
(JSC::SlotVisitor::drain):
(JSC::SlotVisitor::drainFromShared):
(JSC::MarkStack::mergeOpaqueRoots):
(JSC::SlotVisitor::harvestWeakReferences):

• heap/MarkStack.h:

(JSC::MarkStackSegment::data):
(JSC::MarkStackSegment::capacityFromSize):
(JSC::MarkStackSegment::sizeFromCapacity):
(JSC::MarkStackArray::postIncTop):
(JSC::MarkStackArray::preDecTop):
(JSC::MarkStackArray::setTopForFullSegment):
(JSC::MarkStackArray::setTopForEmptySegment):
(JSC::MarkStackArray::top):
(JSC::MarkStackArray::validatePrevious):
(JSC::MarkStack::addWeakReferenceHarvester):
(JSC::MarkStack::mergeOpaqueRootsIfNecessary):
(JSC::MarkStack::mergeOpaqueRootsIfProfitable):
(JSC::MarkStack::MarkStack):
(JSC::MarkStack::addOpaqueRoot):
(JSC::MarkStack::containsOpaqueRoot):
(JSC::MarkStack::opaqueRootCount):
(JSC::MarkStackArray::append):
(JSC::MarkStackArray::canRemoveLast):
(JSC::MarkStackArray::removeLast):
(JSC::MarkStackArray::isEmpty):
(JSC::MarkStackArray::canDonateSomeCells):
(JSC::MarkStackArray::size):
(JSC::ParallelModeEnabler::ParallelModeEnabler):
(JSC::ParallelModeEnabler::~ParallelModeEnabler):

• heap/MarkedBlock.h:

(JSC::MarkedBlock::testAndSetMarked):

• heap/SlotVisitor.h:

(JSC::SlotVisitor::donate):
(JSC::SlotVisitor::donateAndDrain):
(JSC::SlotVisitor::donateKnownParallel):
(JSC::SlotVisitor::SlotVisitor):

• heap/WeakReferenceHarvester.h:
• runtime/Heuristics.cpp:

(JSC::Heuristics::initializeHeuristics):

• runtime/Heuristics.h:
• wtf/Atomics.h:

(WTF::weakCompareAndSwap):

• wtf/Bitmap.h:

(WTF::::Bitmap):
(WTF::::get):
(WTF::::set):
(WTF::::testAndSet):
(WTF::::testAndClear):
(WTF::::concurrentTestAndSet):
(WTF::::concurrentTestAndClear):
(WTF::::clear):
(WTF::::clearAll):
(WTF::::nextPossiblyUnset):
(WTF::::findRunOfZeros):
(WTF::::count):
(WTF::::isEmpty):
(WTF::::isFull):

• wtf/MainThread.h:

(WTF::isMainThreadOrGCThread):

• wtf/Platform.h:
• wtf/ThreadSpecific.h:

(WTF::::isSet):

• wtf/mac/MainThreadMac.mm:

(WTF::initializeGCThreads):
(WTF::initializeMainThreadPlatform):
(WTF::initializeMainThreadToProcessMainThreadPlatform):
(WTF::registerGCThread):
(WTF::isMainThreadOrGCThread):

Source/WebCore:

Reviewed by Geoff Garen.

Added parallel tracing to the GC. This required loosening some assertions,
since some code may now be called from outside the main thread.

No new tests, since no behavior was changed.

• platform/TreeShared.h:

(WebCore::TreeShared::parent):

File:

• trunk/Source/JavaScriptCore/wtf/Atomics.h (modified) (2 diffs)

trunk/Source/JavaScriptCore/wtf/Atomics.h

@@ -61 +61 @@
 
 #include "Platform.h"
+#include "UnusedParam.h"
 
 #if OS(WINDOWS)
@@ -117 +118 @@
 #endif
 
+inline bool weakCompareAndSwap(unsigned* location, unsigned expected, unsigned newValue)
+{
+    // FIXME: Implement COMPARE_AND_SWAP on other architectures and compilers. Currently
+    // it only works on X86 or X86_64 with a GCC-style compiler.
+#if ENABLE(COMPARE_AND_SWAP)
+    bool result;
+    asm volatile(
+        "lock; cmpxchgl %3, %2\n\t"
+        "sete %1"
+        : "+a"(expected), "=r"(result), "+m"(*location)
+        : "r"(newValue)
+        : "memory"
+        );
+    return result;
+#else
+    UNUSED_PARAM(location);
+    UNUSED_PARAM(expected);
+    UNUSED_PARAM(newValue);
+    CRASH();
+    return 0;
+#endif
+}
+
 } // namespace WTF