Fast-codegen: Implementing try/finally on top of nesting context.

src/ia32/fast-codegen-ia32.cc

Index: src/ia32/fast-codegen-ia32.cc
diff --git a/src/ia32/fast-codegen-ia32.cc b/src/ia32/fast-codegen-ia32.cc
index b6e16f3229b6f88536517a376671d0d1f23563fc..4e179c59c2a4b0c434382ed4f2ed6725a0dd4749 100644
--- a/src/ia32/fast-codegen-ia32.cc
+++ b/src/ia32/fast-codegen-ia32.cc
@@ -1625,14 +1625,49 @@ void FastCodeGenerator::VisitCompareOperation(CompareOperation* expr) {
   // Convert current context to test context: End post-test code.
 }
 
-Register FastCodeGenerator::result_register() { return eax; }
 
 void FastCodeGenerator::VisitThisFunction(ThisFunction* expr) {
   __ mov(eax, Operand(ebp, JavaScriptFrameConstants::kFunctionOffset));
   Move(expr->context(), eax);
 }
 
-#undef __
 
+Register FastCodeGenerator::result_register() { return eax; }
+
+// ----------------------------------------------------------------------------
+// Non-local control flow support.
+
+void FastCodeGenerator::EnterFinallyBlock() {
+  // Cook return address on top of stack (smi encoded Code* delta)
+  __ mov(edx, Operand(esp, 0));

[Kevin Millikin (Chromium), 2009/12/10 13:51:12]

Assert result_register is not edx?

[Lasse Reichstein, 2009/12/16 08:43:30]

Good idea. Done.

+  __ sub(Operand(edx), Immediate(masm_->CodeObject()));
+  ASSERT_EQ(1, kSmiTagSize + kSmiShiftSize);
+  ASSERT_EQ(0, kSmiTag);
+  __ add(edx, Operand(edx));  // Convert to smi.
+  __ mov(Operand(esp, 0), edx);
+  // Store result register while executing finally block.
+  __ push(eax);

[Kevin Millikin (Chromium), 2009/12/10 13:51:12]

__ push(result_register())

[Lasse Reichstein, 2009/12/16 08:43:30]

Indeed. Half an abstractions is ... useless.

+}
+
+
+void FastCodeGenerator::ReturnFromFinallyBlock() {
+  // Restore result register from stack.
+  __ pop(eax);
+  // Uncook return address.
+  __ mov(edx, Operand(esp, 0));

[Kevin Millikin (Chromium), 2009/12/10 13:51:12]

Is this really better than:

__ pop(edx);
__ sar(edx, 1);
__ add(Operand(edx), Immediate(masm_->CodeObject()));
__ jmp(Operand(edx));

[Lasse Reichstein, 2009/12/16 08:43:30]

The CPU has a call/return stack that expects calls/returns to be properly
nested. If they aren't, the stack gets out of sync, and return becomes slower.

I don't know if messing with stack values break it anyway (I'm guess it won't,
which is why I cook and uncook without popping, so that the next expected return
value is never outside the stack).

We definitely get out of sync with the try-handler chain management, so maybe
it's not worth keeping it here.
I'll keep it for now, and try to do a little research on whether we break it,
and how much it matters.

+  __ sar(edx, 1);  // Convert smi to int.
+  __ add(Operand(edx), Immediate(masm_->CodeObject()));
+  __ mov(Operand(esp, 0), edx);
+  // And return.
+  __ ret(0);
+}
+
+
+void FastCodeGenerator::ThrowException() {
+  __ push(eax);
+  __ CallRuntime(Runtime::kThrow, 1);
+}
+
+#undef __
 
 } }  // namespace v8::internal