Remember that the Lofi compiler is a simple single pass compiler that does not perform optimizations on its own, it compiles source code as-is.
So a few things to keep in mind:
Minimize computations, pre-compute if possible. Ex: a=9; is faster than a=4+5;
Calling functions requires arguments to be pushed to the stack during call, than pop from the stack in the function execution. Sometimes it's better to just inline stuff (performance wise), even if it is less 'pretty'.
Custom types are particularly heavy to pass around. If a function just uses a single member, it is better to just pass the member beforehand for instance. Being able to pass custom types as references would be better but it is what it is at the moment.
Back to the scrolling topic, time dilution is a powerful tool that can reduce dramatically the cost of copying the tile buffer. Thanks to the double buffered tile buffer and the hardware scrolling, say you move 2 pixels/frame, you can copy 1/8th of the draw buffer each frame and only swap the buffers once complete, dividing the cost by 8. The slower the scrolling, the more you can divide, making it pretty much costless.
