Changeset 36972 in webkit for trunk/JavaScriptCore/VM/CTI.cpp

Timestamp:

Sep 26, 2008, 3:40:26 PM (17 years ago)

Author:

[email protected]

Message:

2008-09-26 Geoffrey Garen <​[email protected]>

Reviewed by Maciej Stachowiak.

Removed dynamic check for whether the callee needs an activation object.
Replaced with callee code to create the activation object.

0.5% speedup on SunSpider.
No change on v8 benchmark. (Might be a speedup, but it's in range of the
variance.)

0.7% speedup on v8 benchmark in bytecode.
1.3% speedup on empty call benchmark in bytecode.

• VM/CTI.cpp: (JSC::CTI::privateCompileMainPass): Added support for op_init_activation, the new opcode that specifies that the callee's initialization should create an activation object. (JSC::CTI::privateCompile): Removed previous code that did a similar thing in an ad-hoc way.

• VM/CodeBlock.cpp: (JSC::CodeBlock::dump): Added a case for dumping op_init_activation.

• VM/CodeGenerator.cpp: (JSC::CodeGenerator::generate): Added fixup code to change op_init to op_init_activation if necessary. (With a better parser, we would know which to use from the beginning.)

• VM/Instruction.h: (JSC::Instruction::Instruction): (WTF::): Faster traits for the instruction vector. An earlier version of this patch relied on inserting at the beginning of the vector, and depended on this change for speed.

• VM/Machine.cpp: (JSC::Machine::execute): Removed clients of setScopeChain, the old abstraction for dynamically checking for whether an activation object needed to be created. (JSC::Machine::privateExecute): ditto

(JSC::Machine::cti_op_push_activation): Renamed this function from
cti_vm_updateScopeChain, and made it faster by removing the call to
setScopeChain.

• VM/Machine.h:

• VM/Opcode.h: Declared op_init_activation.

File:

• trunk/JavaScriptCore/VM/CTI.cpp (modified) (3 diffs)

trunk/JavaScriptCore/VM/CTI.cpp

@@ -1752 +1752 @@
             break;
         }
+        case op_init_activation: {
+            emitCall(i, Machine::cti_op_push_activation);
+
+            // Even though CTI doesn't use them, we initialize our constant
+            // registers to zap stale pointers, to avoid unnecessarily prolonging
+            // object lifetime and increasing GC pressure.
+            size_t count = m_codeBlock->numVars + m_codeBlock->constantRegisters.size();
+            for (size_t j = 0; j < count; ++j)
+                emitInitRegister(j);
+
+            i+= 1;
+            break;
+        }
         case op_get_array_length:
         case op_get_by_id_chain:
@@ -2271 +2284 @@
     emitPutToCallFrameHeader(X86::ecx, RegisterFile::ReturnPC);
 
-    // Lazy copy of the scopeChain
-    X86Assembler::JmpSrc callToUpdateScopeChain;
-    if ((m_codeBlock->codeType == FunctionCode) && m_codeBlock->needsFullScopeChain) {
-        m_jit.emitRestoreArgumentReference();
-        callToUpdateScopeChain = m_jit.emitCall();
-    }
-
     privateCompileMainPass();
     privateCompileLinkPass();
@@ -2323 +2329 @@
         m_codeBlock->ctiReturnAddressVPCMap.add(m_jit.getRelocatedAddress(code, iter->from), iter->opcodeIndex);
     }
-
-    if ((m_codeBlock->codeType == FunctionCode) && m_codeBlock->needsFullScopeChain)
-        X86Assembler::link(code, callToUpdateScopeChain, (void*)Machine::cti_vm_updateScopeChain);
 
     // Link absolute addresses for jsr