24 bit integer types, supposed to work as uint24_t / int24_t could be useful:
1) For use in hardware-specific code.
2) For automatic use by the compiler for temporaries (e.g. instead of long / unsigned long) once we have better data-flow analysis. This would reduce register pressure, runtime and stack space usage.
Since our targets are mostly 8-bit anyway, such a type could be implemented esily and efficiently. In fact, most of the backend code seems to be already there.