Cleanup of class id loading sequences.

runtime/vm/assembler_arm64.cc

Index: runtime/vm/assembler_arm64.cc
===================================================================
--- runtime/vm/assembler_arm64.cc	(revision 38317)
+++ runtime/vm/assembler_arm64.cc	(working copy)
@@ -1095,18 +1095,13 @@
 
 
 void Assembler::LoadTaggedClassIdMayBeSmi(Register result, Register object) {
-  ASSERT(result != TMP);
-  ASSERT(object != TMP);
-
-  // Make a copy of object since result and object can be the same register.
-  mov(TMP, object);
   // Load up a null object. We only need it so we can use LoadClassId on it in
   // the case that object is a Smi..
-  LoadObject(result, Object::null_object(), PP);
+  LoadObject(TMP, Object::null_object(), PP);

[Vyacheslav Egorov (Google), 2014/07/16 23:57:15]

Loading null requires either memory read from constant pool or movw/movt pair. 

We could apply the same idea here that we use on x86: pool a value into constant
pool that is kSmiCid shifted by class id shift then put into TMP an address of
ConstantPool + IndexOfThatValue - kHeapObjectTag.

[zra, 2014/07/17 15:50:10]

That will work in LoadClassIdInstr, but we can't add arbitrary entries to an
object pool in a stub, in general, so it wouldn't work in the check inline cache
stub. (null is fine since null is guaranteed to be in every object pool.)
Loading a 64-bit immediate not from the constant pool on arm64 is up to 4
instructions.

Anyway, I'll try your suggestion in LoadClassIdInstr, but leave this sequence
alone.

   // Check if the object is a Smi.
-  tsti(TMP, kSmiTagMask);
-  // If the object *is* a Smi, load the null object into tmp. o/w leave alone.
-  csel(TMP, result, TMP, EQ);
+  tsti(object, kSmiTagMask);
+  // If the object *is* a Smi, use the null object instead. o/w leave alone.
+  csel(TMP, TMP, object, EQ);
   // Loads either the cid of the object if it isn't a Smi, or the cid of null
   // if it is a Smi, which will be ignored.
   LoadClassId(result, TMP, PP);