stay away from LARGE, always use SMALL.
if it is pointed to use IDATA.
if it is accessed diretly use DATA.
if then you run out of variable storage, carefully select the most raraely accessed variables and move those to XDATA.

hardware that REALLY improve speed:
Sufficient internal memory or 8 ports allow you NOT to use MMIO
Use chips with HW drivers for what you need (UART, IIC, SPI, ...)
Obviously a one clocker is faster than a 12 clocker

Erik