Changeset 99898 in webkit for trunk/Source/JavaScriptCore/dfg/DFGByteCodeParser.cpp

Timestamp:

Nov 10, 2011, 1:59:39 PM (14 years ago)

Author:

[email protected]

Message:

DFG should not reparse code that was just parsed
​https://bugs.webkit.org/show_bug.cgi?id=71977

Reviewed by Geoff Garen.

The instruction stream of a code block is now kept around until
the next GC. When doing either an optimizing compilation of an
executable, or inlining of an executable, we now try to find the
already preexisting bytecode. If we find it, we don't have to parse.
If we don't find it, we parse as before. Inlining takes the extra
step of caching code blocks, so if the same executable gets inlined
multiple times into the same caller, then we parse it at most once
even if prior to inlining that executable did not have any code
blocks with an instruction stream.

Also fixed a silly bug where the strict mode for various operations
was being determined by looking at the machine code block rather
than the inlinee.

To enable the delete-on-next-GC policy, I introduced the notion
of an ultra weak finalizer, which anyone can register during
tracing. This is thread-safe (for parallel GC) and
stop-the-world-safe (so calls to free() are postponed until the
world is resumed). This required reusing some facilities previously
created for WeakReferenceHarvester, so I created a common utility
class. I also retweaked the handling of WeakReferenceHarvesters,
since they should be executed during stop-the-world since in the
future we may want to allow them to call drain().

2% win on SunSpider. 2% win on V8, when run in my harness. Neutral
elsewhere.

• JavaScriptCore.xcodeproj/project.pbxproj:
• bytecode/CodeBlock.cpp:

(JSC::CodeBlock::CodeBlock):
(JSC::CodeBlock::visitAggregate):
(JSC::CodeBlock::copyPostParseDataFrom):
(JSC::CodeBlock::copyPostParseDataFromAlternative):
(JSC::CodeBlock::finalizeUnconditionally):

• bytecode/CodeBlock.h:

(JSC::CodeBlock::canProduceCopyWithBytecode):
(JSC::CodeBlock::discardBytecodeLater):
(JSC::CodeBlock::handleBytecodeDiscardingOpportunity):
(JSC::GlobalCodeBlock::GlobalCodeBlock):
(JSC::ProgramCodeBlock::ProgramCodeBlock):
(JSC::EvalCodeBlock::EvalCodeBlock):
(JSC::FunctionCodeBlock::FunctionCodeBlock):
(JSC::BytecodeDestructionBlocker::BytecodeDestructionBlocker):
(JSC::BytecodeDestructionBlocker::~BytecodeDestructionBlocker):

• dfg/DFGAssemblyHelpers.h:

(JSC::DFG::AssemblyHelpers::strictModeFor):

• dfg/DFGByteCodeCache.h: Added.

(JSC::DFG::CodeBlockKey::CodeBlockKey):
(JSC::DFG::CodeBlockKey::operator==):
(JSC::DFG::CodeBlockKey::hash):
(JSC::DFG::CodeBlockKey::executable):
(JSC::DFG::CodeBlockKey::kind):
(JSC::DFG::CodeBlockKey::isHashTableDeletedValue):
(JSC::DFG::CodeBlockKeyHash::hash):
(JSC::DFG::CodeBlockKeyHash::equal):
(JSC::DFG::ByteCodeCache::ByteCodeCache):
(JSC::DFG::ByteCodeCache::~ByteCodeCache):
(JSC::DFG::ByteCodeCache::get):

• dfg/DFGByteCodeParser.cpp:

(JSC::DFG::ByteCodeParser::handleInlining):

• dfg/DFGJITCodeGenerator32_64.cpp:

(JSC::DFG::JITCodeGenerator::cachedPutById):

• dfg/DFGJITCodeGenerator64.cpp:

(JSC::DFG::JITCodeGenerator::cachedPutById):

• dfg/DFGSpeculativeJIT64.cpp:

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

• heap/Heap.cpp:

(JSC::Heap::finalizeUnconditionally):
(JSC::Heap::markRoots):
(JSC::Heap::collect):

• heap/Heap.h:
• heap/ListableHandler.h: Added.

(JSC::ListableHandler::ListableHandler):
(JSC::ListableHandler::~ListableHandler):
(JSC::ListableHandler::List::List):
(JSC::ListableHandler::List::addNotThreadSafe):
(JSC::ListableHandler::List::addThreadSafe):
(JSC::ListableHandler::List::hasNext):
(JSC::ListableHandler::List::removeNext):

• heap/MarkStack.cpp:

(JSC::MarkStackThreadSharedData::MarkStackThreadSharedData):
(JSC::SlotVisitor::harvestWeakReferences):
(JSC::SlotVisitor::finalizeUnconditionally):

• heap/MarkStack.h:

(JSC::MarkStack::addWeakReferenceHarvester):
(JSC::MarkStack::addUnconditionalFinalizer):

• heap/SlotVisitor.h:
• heap/UnconditionalFinalizer.h: Added.

(JSC::UnconditionalFinalizer::~UnconditionalFinalizer):

• heap/WeakReferenceHarvester.h:

(JSC::WeakReferenceHarvester::WeakReferenceHarvester):
(JSC::WeakReferenceHarvester::~WeakReferenceHarvester):

• runtime/Executable.cpp:

(JSC::EvalExecutable::compileInternal):
(JSC::ProgramExecutable::compileInternal):
(JSC::FunctionExecutable::baselineCodeBlockFor):
(JSC::FunctionExecutable::codeBlockWithBytecodeFor):
(JSC::FunctionExecutable::produceCodeBlockFor):
(JSC::FunctionExecutable::compileForCallInternal):
(JSC::FunctionExecutable::compileForConstructInternal):

• runtime/Executable.h:

(JSC::FunctionExecutable::profiledCodeBlockFor):

File:

• trunk/Source/JavaScriptCore/dfg/DFGByteCodeParser.cpp (modified) (4 diffs)

trunk/Source/JavaScriptCore/dfg/DFGByteCodeParser.cpp

@@ -29 +29 @@
 #if ENABLE(DFG_JIT)
 
+#include "DFGByteCodeCache.h"
 #include "DFGCapabilities.h"
 #include "CodeBlock.h"
@@ -842 +843 @@
     // Mapping between values and constant numbers.
     JSValueMap m_jsValueMap;
+    
+    // Cache of code blocks that we've generated bytecode for.
+    ByteCodeCache<canInlineFunctionFor> m_codeBlockCache;
 };
 
@@ -957 +961 @@
     // If we get here then it looks like we should definitely inline this code. Proceed
     // with parsing the code to get bytecode, so that we can then parse the bytecode.
-    // Note that the code block we get here is intended to die after handleInlining()
-    // returns.
-    JSObject* exception;
-    OwnPtr<CodeBlock> codeBlock = executable->produceCodeBlockFor(expectedFunction->scope(), OptimizingCompilation, kind, exception);
+    CodeBlock* codeBlock = m_codeBlockCache.get(CodeBlockKey(executable, kind), expectedFunction->scope());
     if (!codeBlock)
         return false;
-    ASSERT(!exception);
-    
-    // Now that we have the bytecode, check if we really can inline it. This may fail
-    // if the code block contains some nasty opcodes.
-    if (!canInlineFunctionFor(codeBlock.get(), kind))
-        return false;
+    
+    ASSERT(canInlineFunctionFor(codeBlock, kind));
 
 #if DFG_ENABLE(DEBUG_VERBOSE)
@@ -999 +996 @@
     }
 
-    InlineStackEntry inlineStackEntry(this, codeBlock.get(), profiledBlock, m_graph.m_blocks.size() - 1, (VirtualRegister)m_inlineStackTop->remapOperand(callTarget), expectedFunction, (VirtualRegister)m_inlineStackTop->remapOperand(usesResult ? resultOperand : InvalidVirtualRegister), (VirtualRegister)inlineCallFrameStart, kind);
+    InlineStackEntry inlineStackEntry(this, codeBlock, profiledBlock, m_graph.m_blocks.size() - 1, (VirtualRegister)m_inlineStackTop->remapOperand(callTarget), expectedFunction, (VirtualRegister)m_inlineStackTop->remapOperand(usesResult ? resultOperand : InvalidVirtualRegister), (VirtualRegister)inlineCallFrameStart, kind);
     
     // This is where the actual inlining really happens.