Perhaps review .MAP and .LST, look at what's eating the most memory.
If there's repetitive stuff that could be a subroutine or otherwise factored.
Compilers are not good with intent or identifying algorithms.
A lot of register code is very inefficient, using a lot of individual RMW code that the compiler can't fold due to the volatile and in-order expectations constraining it. Understand the load-store architecture and use your overview to decide what can be folded and where the order is or is not critical.
