Make --always-opt also optimize toplevel code.

src/arm64/lithium-codegen-arm64.cc

Index: src/arm64/lithium-codegen-arm64.cc
diff --git a/src/arm64/lithium-codegen-arm64.cc b/src/arm64/lithium-codegen-arm64.cc
index 3f6be920af459a8b1d9aff48f7e0a2076faacf24..d444d1c1a86ddaac8d693d12cef130f6c064c3b2 100644
--- a/src/arm64/lithium-codegen-arm64.cc
+++ b/src/arm64/lithium-codegen-arm64.cc
@@ -689,7 +689,11 @@ bool LCodeGen::GeneratePrologue() {
     Comment(";;; Allocate local context");
     bool need_write_barrier = true;
     // Argument to NewContext is the function, which is in x1.
-    if (heap_slots <= FastNewContextStub::kMaximumSlots) {
+    if (FLAG_harmony_scoping && info()->scope()->is_global_scope()) {
+      __ Push(x1);
+      __ Push(info()->scope()->GetScopeInfo());

[jbramley, 2014/07/28 16:00:32]

I think it would be slightly better load the scope info explicity so we can
merge the pushes:

UseScratchRegisterScope temps(masm);
Register scope_info = temps.AcquireX();
__ Mov(scope_info, Operand(info()->scope()->GetScopeInfo()));
__ Push(x1, scope_info);

Each 'Push' has an overhead, and Push(Handle<ScopeInfo>) will just load the
handle into a scratch register anyway so the 'Mov' here doesn't have any extra
cost.

[Michael Starzinger, 2014/07/28 16:15:53]

Done.

+      __ CallRuntime(Runtime::kNewGlobalContext, 2);
+    } else if (heap_slots <= FastNewContextStub::kMaximumSlots) {
       FastNewContextStub stub(isolate(), heap_slots);
       __ CallStub(&stub);
       // Result of FastNewContextStub is always in new space.
@@ -5034,8 +5038,6 @@ void LCodeGen::DoDeclareGlobals(LDeclareGlobals* instr) {
   Register scratch1 = x5;
   Register scratch2 = x6;
   ASSERT(instr->IsMarkedAsCall());
-
-  ASM_UNIMPLEMENTED_BREAK("DoDeclareGlobals");

[jbramley, 2014/07/28 16:00:32]

We used ASM_UNIMPLEMENTED_BREAK during development. I thought they were all
gone, to be honest. This one probably remains because we couldn't test this
code. (That is, it's probably safe to remove it.)

[Michael Starzinger, 2014/07/28 16:15:53]

Acknowledged. This actually was the last use. Would you be fine if I removed the
macro (and the related flag) in a follow-up CL or do you intend to keep this
around?

[jbramley, 2014/07/28 16:28:11]

I see no reason to keep it. We can recreate it fairly easily if we ever find
that we need it again, but it seems unlikely.

   // TODO(all): if Mov could handle object in new space then it could be used
   // here.
   __ LoadHeapObject(scratch1, instr->hydrogen()->pairs());