Timestamp:
Nov 6, 2012, 4:13:54 PM (13 years ago)
Author:
[email protected]
Message:

Reduce parser overhead in JSC
​https://bugs.webkit.org/show_bug.cgi?id=101127

Reviewed by Filip Pizlo.

An exciting journey into the world of architecture in which our hero
adds yet another layer to JSC codegeneration.

This patch adds a marginally more compact form of bytecode that is
free from any data specific to a given execution context, and that
does store any data structures necessary for execution. To actually
execute this UnlinkedBytecode we still need to instantiate a real
CodeBlock, but this is a much faster linear time operation than any
of the earlier parsing or code generation passes.

As the unlinked code is context free we can then simply use a cache
from source to unlinked code mapping to completely avoid all of the
old parser overhead. The cache is currently very simple and memory
heavy, using the complete source text as a key (rather than SourceCode
or equivalent), and a random eviction policy.

This seems to produce a substantial win when loading identical content
in different contexts.

  • API/tests/testapi.c:

(main):

  • CMakeLists.txt:
  • GNUmakefile.list.am:
  • JavaScriptCore.vcproj/JavaScriptCore/JavaScriptCore.vcproj:
  • JavaScriptCore.xcodeproj/project.pbxproj:
  • bytecode/CodeBlock.cpp:
  • bytecode/CodeBlock.h:

Moved a number of fields, and a bunch of logic to UnlinkedCodeBlock.h/cpp

  • bytecode/Opcode.h:

Added a global const init no op instruction needed to get correct
behaviour without any associated semantics.

  • bytecode/UnlinkedCodeBlock.cpp: Added.
  • bytecode/UnlinkedCodeBlock.h: Added.

A fairly shallow, GC allocated version of the old CodeBlock
classes with a 32bit instruction size, and just metadata
size tracking.

  • bytecompiler/BytecodeGenerator.cpp:
  • bytecompiler/BytecodeGenerator.h:

Replace direct access to m_symbolTable with access through
symbolTable(). ProgramCode no longer has a symbol table at
all so some previously unconditional (and pointless) uses
of symbolTable get null checks.
A few other changes to deal with type changes due to us generating
unlinked code (eg. pointer free, so profile indices rather than
pointers).

  • dfg/DFGByteCodeParser.cpp:
  • dfg/DFGCapabilities.h:

Support global_init_nop

  • interpreter/Interpreter.cpp:

Now get the ProgramExecutable to initialise new global properties
before starting execution.

  • jit/JIT.cpp:
  • jit/JITDriver.h:
  • jit/JITStubs.cpp:
  • llint/LLIntData.cpp:
  • llint/LLIntSlowPaths.cpp:
  • llint/LowLevelInterpreter.asm:
  • llint/LowLevelInterpreter32_64.asm:
  • llint/LowLevelInterpreter64.asm:

Adding init_global_const_nop everywhere else

  • parser/Parser.h:
  • parser/ParserModes.h: Added.
  • parser/ParserTokens.h:

Parser no longer needs a global object or callframe to function

  • runtime/CodeCache.cpp: Added.
  • runtime/CodeCache.h: Added.

A simple, random eviction, Source->UnlinkedCode cache

  • runtime/Executable.cpp:
  • runtime/Executable.h:

Executables now reference their unlinked counterparts, and
request code specifically for the target global object.

  • runtime/JSGlobalData.cpp:
  • runtime/JSGlobalData.h:

GlobalData now owns a CodeCache and a set of new structures
for the unlinked code types.

  • runtime/JSGlobalObject.cpp:
  • runtime/JSGlobalObject.h:

Utility functions used by executables to perform compilation

  • runtime/JSType.h: Add new JSTypes for unlinked code
File:
1 added

Note: See TracChangeset for help on using the changeset viewer.