Minor cleanup code in GenerateStoreField.

src/stub-cache-ia32.cc

Index: src/stub-cache-ia32.cc
===================================================================
--- src/stub-cache-ia32.cc	(revision 787)
+++ src/stub-cache-ia32.cc	(working copy)
@@ -401,16 +401,6 @@
     return;
   }
 
-  // Adjust for the number of properties stored in the object. Even in the
-  // face of a transition we can use the old map here because the size of the
-  // object and the number of in-object properties is not going to change.
-  index -= object->map()->inobject_properties();
-
-  if (index >= 0) {
-    // Get the properties array (optimistically).
-    __ mov(scratch, FieldOperand(receiver_reg, JSObject::kPropertiesOffset));
-  }
-
   if (transition != NULL) {
     // Update the map of the object; no write barrier updating is
     // needed because the map is never in new space.
@@ -418,6 +408,11 @@
            Immediate(Handle<Map>(transition)));
   }
 
+  // Adjust for the number of properties stored in the object. Even in the
+  // face of a transition we can use the old map here because the size of the
+  // object and the number of in-object properties is not going to change.
+  index -= object->map()->inobject_properties();
+
   if (index < 0) {
     // Set the property straight into the object.
     int offset = object->map()->instance_size() + (index * kPointerSize);
@@ -430,6 +425,8 @@
   } else {
     // Write to the properties array.
     int offset = index * kPointerSize + Array::kHeaderSize;
+    // Get the properties array (optimistically).
+    __ mov(scratch, FieldOperand(receiver_reg, JSObject::kPropertiesOffset));

[iposva, 2008/11/20 01:10:57]

Moving the load of the properties pointer down here only has a potential
performance impact when this store also happens to have a transition. This
should be negligible, but we should verify this with a benchmark run.

[Feng Qian, 2008/11/20 16:30:47]

Running V8 bench suite (Richards, DeltaBlue, Crypto, Raytrace, and EarleyBoyer),
the differences are tiny on Windows:

3166 vs 3168 (after change).

On 2008/11/20 01:10:57, iposva wrote:

     __ mov(FieldOperand(scratch, offset), eax);
 
     // Update the write barrier for the array address.