This compiles to fewer bytecodes and uses built-in slots instead of
method dispatch, so it should be faster.

Benchmarks show a tiny improvement (using the same benchmarks as
issue #1233, so not exactly designed to test this):

RPS

+------+------+-------+
|Size  |master|+=     |
+------+------+-------+
|1     |5.60  | 5.63  |
+------+------+-------+
| 10   |5.49  |  5.53 |
+------+------+-------+
|  100 |4.99  |   5.00|
+------+------+-------+
|  1000|3.53  |   3.57|
+------+------+-------+