commit dafef97fdd79246aa2458e22a94e1da116c02dcf
author ysuzuki@apple.com <ysuzuki@apple.com@268f45cc-cd09-0410-ab3c-d52691b4dbfc> Sun May 12 22:50:21 2019 +0000
committer ysuzuki@apple.com <ysuzuki@apple.com@268f45cc-cd09-0410-ab3c-d52691b4dbfc> Sun May 12 22:50:21 2019 +0000
tree 8a5c45a043877eef74c018cab5f39f055378d75f
parent 2d44a80731d84e045d78724b6c9fb69fa596a4a7

[JSC] Compress Watchpoint size by using enum type and Packed<> data structure
https://bugs.webkit.org/show_bug.cgi?id=197730

Reviewed by Filip Pizlo.

Source/JavaScriptCore:

Watchpoint takes 5~ MB memory in Gmail (total memory starts with 400 - 500 MB), so 1~%. Since it is allocated massively,
reducing each size of Watchpoint reduces memory footprint significantly.

As a first step, this patch uses Packed<> and enum to reduce the size of Watchpoint.

1. Watchpoint should have enum type and should not use vtable. vtable takes one pointer, and it is too costly for such a
   memory sensitive objects. We perform downcast and dispatch the method of the derived classes based on this enum. Since
   the # of derived Watchpoint classes are limited (Only 8), we can list up them easily. One unfortunate thing is that
   we cannot do this for destructor so long as we use "delete" for deleting objects. If we dispatch the destructor of derived
   class in the destructor of the base class, we call the destructor of the base class multiple times. delete operator override
   does not help since custom delete operator is called after the destructor is called. While we can fix this issue by always
   using custom deleter, currently we do not since all the watchpoints do not have members which have non trivial destructor.
   Once it is strongly required, we can start using custom deleter, but for now, we do not need to do this.

2. We use Packed<> to compact pointers in Watchpoint. Since Watchpoint is a node of doubly linked list, each one has two
   pointers for prev and next. This is also too costly. PackedPtr reduces the size and makes alignment 1.S

3. We use PackedCellPtr<> for JSCells in Watchpoint. This leverages alignment information and makes pointers smaller in
   Darwin ARM64. One important thing to note here is that since this pointer is packed, it cannot be found by conservative
   GC scan. It is OK for watchpoint since they are allocated in the heap anyway.

We applied this change to Watchpoint and get the following memory reduction. The highlight is that CodeBlockJettisoningWatchpoint in
ARM64 only takes 2 pointers size.

                                                                      ORIGINAL    X86_64   ARM64
    WatchpointSet:                                                    40          32       28
    CodeBlockJettisoningWatchpoint:                                   32          19       15
    StructureStubClearingWatchpoint:                                  56          48       40
    AdaptiveInferredPropertyValueWatchpointBase::StructureWatchpoint: 24          13       11
    AdaptiveInferredPropertyValueWatchpointBase::PropertyWatchpoint:  24          13       11
    FunctionRareData::AllocationProfileClearingWatchpoint:            32          19       15
    ObjectToStringAdaptiveStructureWatchpoint:                        56          48       40
    LLIntPrototypeLoadAdaptiveStructureWatchpoint:                    64          48       48
    DFG::AdaptiveStructureWatchpoint:                                 56          48       40

While we will re-architect the mechanism of Watchpoint, anyway Packed<> mechanism and enum types will be used too.

* CMakeLists.txt:
* JavaScriptCore.xcodeproj/project.pbxproj:
* Sources.txt:
* bytecode/AdaptiveInferredPropertyValueWatchpointBase.h:
* bytecode/CodeBlockJettisoningWatchpoint.h:
* bytecode/CodeOrigin.h:
* bytecode/LLIntPrototypeLoadAdaptiveStructureWatchpoint.cpp:
(JSC::LLIntPrototypeLoadAdaptiveStructureWatchpoint::LLIntPrototypeLoadAdaptiveStructureWatchpoint):
(JSC::LLIntPrototypeLoadAdaptiveStructureWatchpoint::fireInternal):
* bytecode/LLIntPrototypeLoadAdaptiveStructureWatchpoint.h:
* bytecode/StructureStubClearingWatchpoint.cpp:
(JSC::StructureStubClearingWatchpoint::fireInternal):
* bytecode/StructureStubClearingWatchpoint.h:
* bytecode/Watchpoint.cpp:
(JSC::Watchpoint::fire):
* bytecode/Watchpoint.h:
(JSC::Watchpoint::Watchpoint):
* dfg/DFGAdaptiveStructureWatchpoint.cpp:
(JSC::DFG::AdaptiveStructureWatchpoint::AdaptiveStructureWatchpoint):
* dfg/DFGAdaptiveStructureWatchpoint.h:
* heap/PackedCellPtr.h: Added.
* runtime/FunctionRareData.h:
* runtime/ObjectToStringAdaptiveStructureWatchpoint.cpp: Added.
(JSC::ObjectToStringAdaptiveStructureWatchpoint::ObjectToStringAdaptiveStructureWatchpoint):
(JSC::ObjectToStringAdaptiveStructureWatchpoint::install):
(JSC::ObjectToStringAdaptiveStructureWatchpoint::fireInternal):
* runtime/ObjectToStringAdaptiveStructureWatchpoint.h: Added.
* runtime/StructureRareData.cpp:
(JSC::StructureRareData::clearObjectToStringValue):
(JSC::ObjectToStringAdaptiveStructureWatchpoint::ObjectToStringAdaptiveStructureWatchpoint): Deleted.
(JSC::ObjectToStringAdaptiveStructureWatchpoint::install): Deleted.
(JSC::ObjectToStringAdaptiveStructureWatchpoint::fireInternal): Deleted.
* runtime/StructureRareData.h:

Source/WTF:

This patch introduces a new data structures, WTF::Packed, WTF::PackedPtr, and WTF::PackedAlignedPtr.

- WTF::Packed

    WTF::Packed is data storage. We can read and write trivial (in C++ term [1]) data to this storage. The difference to
    the usual storage is that the alignment of this storage is always 1. We access the underlying data by using unalignedLoad/unalignedStore.
    This class offers alignment = 1 data structure instead of missing the following characteristics.

        1. Load / Store are non atomic even if the data size is within a pointer width. We should not use this for a member which can be accessed
           in a racy way. (e.g. fields accessed optimistically from the concurrent compilers).

        2. We cannot take reference / pointer to the underlying storage since they are unaligned.

        3. Access to this storage is unaligned access. The code is using memcpy, and the compiler will convert to an appropriate unaligned access
           in certain architectures (x86_64 / ARM64). It could be slow. So use it for non performance sensitive & memory sensitive places.

- WTF::PackedPtr

    WTF::PackedPtr is a specialization of WTF::Packed<T*>. And it is basically WTF::PackedAlignedPtr with alignment = 1. We further compact
    the pointer by leveraging the platform specific knowledge. In 64bit architectures, the effective width of pointers are less than 64 bit.
    In x86_64, it is 48 bits. And Darwin ARM64 is further smaller, 36 bits. This information allows us to compact the pointer to 6 bytes in
    x86_64 and 5 bytes in Darwin ARM64.

- WTF::PackedAlignedPtr

    WTF::PackedAlignedPtr is the WTF::PackedPtr with alignment information of the T. If we use this alignment information, we could reduce the
    size of packed pointer further in some cases. For example, since we guarantee that JSCells are 16 byte aligned, low 4 bits are empty. Leveraging
    this information in Darwin ARM64 platform allows us to make packed JSCell pointer 4 bytes (36 - 4 bits). We do not use passed alignment
    information if it is not profitable.

We also have PackedPtrTraits. This is new PtrTraits and use it for various data structures such as Bag<>.

[1]: https://en.cppreference.com/w/cpp/types/is_trivial

* WTF.xcodeproj/project.pbxproj:
* wtf/Bag.h:
(WTF::Bag::clear):
(WTF::Bag::iterator::operator++):
* wtf/CMakeLists.txt:
* wtf/DumbPtrTraits.h:
* wtf/DumbValueTraits.h:
* wtf/MathExtras.h:
(WTF::clzConstexpr):
(WTF::clz):
(WTF::ctzConstexpr):
(WTF::ctz):
(WTF::getLSBSetConstexpr):
(WTF::getMSBSetConstexpr):
* wtf/Packed.h: Added.
(WTF::Packed::Packed):
(WTF::Packed::get const):
(WTF::Packed::set):
(WTF::Packed::operator=):
(WTF::Packed::exchange):
(WTF::Packed::swap):
(WTF::alignof):
(WTF::PackedPtrTraits::exchange):
(WTF::PackedPtrTraits::swap):
(WTF::PackedPtrTraits::unwrap):
* wtf/Platform.h:
* wtf/SentinelLinkedList.h:
(WTF::BasicRawSentinelNode::BasicRawSentinelNode):
(WTF::BasicRawSentinelNode::prev):
(WTF::BasicRawSentinelNode::next):
(WTF::PtrTraits>::remove):
(WTF::PtrTraits>::prepend):
(WTF::PtrTraits>::append):
(WTF::RawNode>::SentinelLinkedList):
(WTF::RawNode>::remove):
(WTF::BasicRawSentinelNode<T>::remove): Deleted.
(WTF::BasicRawSentinelNode<T>::prepend): Deleted.
(WTF::BasicRawSentinelNode<T>::append): Deleted.
* wtf/StdLibExtras.h:
(WTF::roundUpToMultipleOfImpl):
(WTF::roundUpToMultipleOfImpl0): Deleted.
* wtf/UnalignedAccess.h:
(WTF::unalignedLoad):
(WTF::unalignedStore):

Tools:

* TestWebKitAPI/CMakeLists.txt:
* TestWebKitAPI/TestWebKitAPI.xcodeproj/project.pbxproj:
* TestWebKitAPI/Tests/WTF/MathExtras.cpp:
(TestWebKitAPI::TEST):
* TestWebKitAPI/Tests/WTF/Packed.cpp: Added.
(TestWebKitAPI::TEST):

git-svn-id: http://svn.webkit.org/repository/webkit/trunk@245214 268f45cc-cd09-0410-ab3c-d52691b4dbfc