[turbofan] Optimize loads from the global object in JSTypeFeedbackSpecializer.

src/compiler/pipeline.cc

Index: src/compiler/pipeline.cc
diff --git a/src/compiler/pipeline.cc b/src/compiler/pipeline.cc
index 5d8135e811e333ebf2a065f4168579aa8a1203e5..ebf9b7c4651e5335970dacbea749a8fdcd097f32 100644
--- a/src/compiler/pipeline.cc
+++ b/src/compiler/pipeline.cc
@@ -519,8 +519,14 @@ struct JSTypeFeedbackPhase {
                               data->info()->unoptimized_code(),
                               data->info()->feedback_vector(), native_context);
     GraphReducer graph_reducer(data->graph(), temp_zone);
-    JSTypeFeedbackSpecializer specializer(data->jsgraph(),
-                                          data->js_type_feedback(), &oracle);
+    Handle<GlobalObject> global_object = Handle<GlobalObject>::null();
+    if (data->info()->has_global_object()) {

[Michael Starzinger, 2015/04/21 12:34:01]

This is only valid if we don't share TurboFan code accross native contexts. We
are starting to add more and more assumptions like this. I think it's time to
introduce some sort of "sharing mode" that let's the pipeline (and the different
reducers) know what level of sharing is expected. WDYT?

Can we leave TODO in that regard here and introduce the "sharing mode" in a
follow-up CL?

[titzer, 2015/04/21 12:56:52]

Acknowledge, agreed. I have sketched out a specialization mode struct in another
CL and dropped in a TODO.

+      global_object =
+          Handle<GlobalObject>(data->info()->global_object(), data->isolate());
+    }
+    JSTypeFeedbackSpecializer specializer(
+        data->jsgraph(), data->js_type_feedback(), &oracle, global_object,
+        data->info()->dependencies());
     AddReducer(data, &graph_reducer, &specializer);
     graph_reducer.ReduceGraph();
   }