[turbofan] Support inline receiver allocation for class constructors.

src/compiler/js-typed-lowering.cc

Index: src/compiler/js-typed-lowering.cc
diff --git a/src/compiler/js-typed-lowering.cc b/src/compiler/js-typed-lowering.cc
index e9a6fec08894e7a1f8204bd42787d11a181fd065..7724e0499fc85ea23a4ebe7c04a2a5dcd2e5176f 100644
--- a/src/compiler/js-typed-lowering.cc
+++ b/src/compiler/js-typed-lowering.cc
@@ -1411,8 +1411,9 @@ const int kMaxInlineInstanceSize = 64 * kPointerSize;
 
 // Checks whether allocation using the given constructor can be inlined.
 bool IsAllocationInlineable(Handle<JSFunction> constructor) {
-  // TODO(bmeurer): Support inlining of class constructors.
-  if (IsClassConstructor(constructor->shared()->kind())) return false;
+  // TODO(bmeurer): Further relax restrictions on inlining, i.e.
+  // instance type and maybe instance size (inobject properties
+  // are limited anyways by the runtime).

[Toon Verwaest, 2015/12/16 20:00:47]

DBC: Don't you only want to allow it for base classes with new.target == base?
We also store initial maps on subclasses, but the base class needs to get the
map from the new.target and verify it.

[Benedikt Meurer, 2015/12/16 20:29:07]

You already answered the question. I'm currently working on supporting
new.target here as well (essentially the case where target != new_target on the
Node* level), but I first had/still need to tweak some other parts of TurboFan
and the way we compile super calls to be able to actually trigger this case at
all.

   return constructor->has_initial_map() &&
          constructor->initial_map()->instance_type() == JS_OBJECT_TYPE &&
          constructor->initial_map()->instance_size() < kMaxInlineInstanceSize;