Changeset 192531 in webkit for trunk/Source/JavaScriptCore/dfg/DFGSpeculativeJIT64.cpp

Timestamp:

Nov 17, 2015, 1:55:33 PM (10 years ago)

Author:

[email protected]

Message:

Use the JITAddGenerator snippet in the DFG.
​https://bugs.webkit.org/show_bug.cgi?id=151266

Reviewed by Geoffrey Garen.

No tests added because the op_add.js stress test already tests for correctness
(using the LLINT as a reference).

Performance-wise, the difference from the pre-existing DFG implementation is
insignificant (at least as measured on x86 and x86_64). We're moving forward
with adopting this implementation because it unifies the 32-bit and 64-bit
implementations, as well as lays ground work for a repatching inline cache
implementation of op_add later.

• dfg/DFGAbstractValue.cpp:

(JSC::DFG::AbstractValue::resultType):

• Made an assertion less restrictive. For ValueAdd operands, the DFG thinks that the operand can also be empty (though we should never see this in practice).

• dfg/DFGFixupPhase.cpp:

(JSC::DFG::FixupPhase::fixupNode):

• Add fallback to unused type operands.

• dfg/DFGSpeculativeJIT.cpp:

(JSC::DFG::SpeculativeJIT::compileValueAdd):

• Introduce a common function to compile the ValueAdd node.

• dfg/DFGSpeculativeJIT.h:

(JSC::DFG::JSValueOperand::JSValueOperand):

• Add the forwarding constructor so that we can use Optional<JSValueOperand>.

• dfg/DFGSpeculativeJIT32_64.cpp:

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

• dfg/DFGSpeculativeJIT64.cpp:

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

• Changed to use the common compileValueAdd().

• jit/AssemblyHelpers.h:

(JSC::AssemblyHelpers::moveValue):

• Similar to moveTrustedValue() but used for untrusted constants.

• jit/JITAddGenerator.cpp:

(JSC::JITAddGenerator::generateFastPath):

• jit/JITAddGenerator.h:

(JSC::JITAddGenerator::JITAddGenerator):

• Updated to take the left or right operand as a constant. This is necessary because the client should not be making assumptions about whether the snippet will determine the operation to be commutative or not. Instead, the client should just pass in the operands and let the snippet do any operand order swapping if necessary.

• jit/JITArithmetic.cpp:

(JSC::JIT::emit_op_add):

• Updated to use the new JITAddGenerator interface.

• tests/stress/op_add.js:

(stringifyIfNeeded):
(runTest):

• Made test output more clear about when string results are expected.

File:

• trunk/Source/JavaScriptCore/dfg/DFGSpeculativeJIT64.cpp (modified) (1 diff)

trunk/Source/JavaScriptCore/dfg/DFGSpeculativeJIT64.cpp

@@ -2269 +2269 @@
         break;
     }
-        
-    case ValueAdd: {
-        JSValueOperand op1(this, node->child1());
-        JSValueOperand op2(this, node->child2());
-        
-        GPRReg op1GPR = op1.gpr();
-        GPRReg op2GPR = op2.gpr();
-        
-        flushRegisters();
-        
-        GPRFlushedCallResult result(this);
-        if (isKnownNotNumber(node->child1().node()) || isKnownNotNumber(node->child2().node()))
-            callOperation(operationValueAddNotNumber, result.gpr(), op1GPR, op2GPR);
-        else
-            callOperation(operationValueAdd, result.gpr(), op1GPR, op2GPR);
-        m_jit.exceptionCheck();
-        
-        jsValueResult(result.gpr(), node);
-        break;
-    }
-        
+
+    case ValueAdd:
+        compileValueAdd(node);
+        break;
+
     case StrCat: {
         JSValueOperand op1(this, node->child1(), ManualOperandSpeculation);