Extracted some loop-invariants in replace_many().

It's now ~82% as fast as the serial case in the benchmark (it used to be ~53%).

Rev: src/builtin_functions.c:1.613