Changeset 38322 in webkit for trunk/JavaScriptCore/parser/Nodes.h

Timestamp:

Nov 11, 2008, 4:32:38 PM (17 years ago)

Author:

[email protected]

Message:

2008-11-11 Geoffrey Garen <​[email protected]>

Reviewed by Darin Adler.

Fixed ​https://bugs.webkit.org/show_bug.cgi?id=22174
Simplified op_call by nixing its responsibility for moving the value of
"this" into the first argument slot.

Instead, the caller emits an explicit load or mov instruction, or relies
on implicit knowledge that "this" is already in the first argument slot.
As a result, two operands to op_call are gone: firstArg and thisVal.

SunSpider and v8 tests show no change in bytecode or CTI.

• VM/CTI.cpp: (JSC::CTI::compileOpCallSetupArgs): (JSC::CTI::compileOpCallEvalSetupArgs): (JSC::CTI::compileOpConstructSetupArgs): Split apart these three versions of setting up arguments to op_call, because they're more different than they are the same -- even more so with this patch.

(JSC::CTI::compileOpCall): Updated for the fact that op_construct doesn't
match op_call anymore.

(JSC::CTI::privateCompileMainPass):
(JSC::CTI::privateCompileSlowCases): Merged a few call cases. Updated
for changes mentioned above.

• VM/CTI.h:

• VM/CodeBlock.cpp: (JSC::CodeBlock::dump): Updated for new bytecode format of call / construct.

• VM/Machine.cpp: (JSC::Machine::callEval): Updated for new bytecode format of call / construct.

(JSC::Machine::dumpCallFrame):
(JSC::Machine::dumpRegisters): Simplified these debugging functions,
taking advantage of the new call frame layout.

(JSC::Machine::execute): Fixed up the eval version of execute to be
friendlier to calls in the new format.

(JSC::Machine::privateExecute): Implemented the new call format in
bytecode.

(JSC::Machine::cti_op_call_NotJSFunction):
(JSC::Machine::cti_op_construct_JSConstruct):
(JSC::Machine::cti_op_construct_NotJSConstruct):
(JSC::Machine::cti_op_call_eval): Updated CTI helpers to match the new
call format.

Fixed a latent bug in stack overflow checking that is now hit because
the register layout has changed a bit -- namely: when throwing a stack
overflow exception inside an op_call helper, we need to account for the
fact that the current call frame is only half-constructed, and use the
parent call frame instead.

• VM/Machine.h:

• bytecompiler/CodeGenerator.cpp: (JSC::CodeGenerator::emitCall): (JSC::CodeGenerator::emitCallEval): (JSC::CodeGenerator::emitConstruct):
• bytecompiler/CodeGenerator.h: Updated codegen to match the new call format.

• parser/Nodes.cpp: (JSC::EvalFunctionCallNode::emitCode): (JSC::FunctionCallValueNode::emitCode): (JSC::FunctionCallResolveNode::emitCode): (JSC::FunctionCallBracketNode::emitCode): (JSC::FunctionCallDotNode::emitCode):
• parser/Nodes.h: (JSC::ScopeNode::neededConstants): ditto

2008-11-10 Geoffrey Garen <​[email protected]>

Reviewed by Darin Adler.

Updated a test after fixing ​https://bugs.webkit.org/show_bug.cgi?id=22174
Simplified op_call by nixing its responsibility for moving the value of
"this" into the first argument slot.

• fast/js/global-recursion-on-full-stack-expected.txt: This test passes a little differently now, because the register layout has changed. Specifically, the stack overflow now happens in the call to f() instead of the initiation of the <script> tag, so it is caught, and it does not log an exception to the console.

File:

• trunk/JavaScriptCore/parser/Nodes.h (modified) (1 diff)

trunk/JavaScriptCore/parser/Nodes.h

@@ -2079 +2079 @@
         int neededConstants()
         {
-            // We may need 1 more constant than the count given by the parser,
-            // because of the various uses of jsUndefined().
-            return m_numConstants + 1;
+            // We may need 2 more constants than the count given by the parser,
+            // because of the various uses of jsUndefined() and jsNull().
+            return m_numConstants + 2;
         }