Currently we do all arithmetic in 16-bit, even when the operands and destination are 8-bit (i.e. char types). This is a very single-pass-compiler and C thing to do (surely this is what the usual arithmetic conversions were designed for), but since uxn has native 8-bit operations and a limited stack size, it's neither efficient nor makes for particularly æsthetically-pleasing assembly. This will get worse if we start doing sign extension when promoting char to int (see #9).

So, it would be nice if (char)(some_char * 2 + 3) could be codegen'd as #02 MUL #03 ADD rather than #0002 MUL2 #0003 ADD2. As I see it there's two ways this could be done: in a “single-pass” fashion by changing the codegen step, or with some sort of later optimisation pass.

I am optimistic about the former approach. I think we could do it by propagating cast/conversion information downwards when doing codegen for expressions.

Currently the codegen behaviour is something like:

• When encountering (char)(some_char * 2 + 3):
  • Recurse to generate (some_char * 2 + 3)
    • Recurse to generate some_char * 2
      • Recurse to generate some_char
        • Output code for loading some_char
        • Output code to extend to int (something like 00 SWP)
      • Recurse to generate 2
        • Output 0002
      • Output MUL2
    • Recurse to generate 3
      • Output 0003
    • Output ADD2
  • Output code to truncate to char (something like NIP)
  • Output code to extend to int (something like 00 SWP)

In the new system there would be a new flag used in expression codegen, something like truncate_to_byte. Now the behaviour would look something like:

• When encountering (char)(some_char * 2 + 3):
  • Recurse to generate (some_char * 2 + 3) with truncate_to_byte set
    • Recurse to generate some_char * 2 with truncate_to_byte set
      • Recurse to generate some_char with truncate_to_byte set
        • Output code for loading some_char
        • Output code to extend to int (something like 00 SWP)
      • Recurse to generate 2 with truncate_to_byte set
        • Output 02 0002
      • Output MUL MUL2
    • Recurse to generate 3 with truncate_to_byte set
      • Output 03 0003
    • Output ADD ADD2
  • Output code to truncate to char (something like NIP)
  • Output code to extend to int (something like 00 SWP) but only if truncate_to_byte is not set

A more tricky case is something like (char)((1 << 8) >> 8), where we can't use 8-bit operations all the way down. That can be handled by not propagating truncate_to_byte when dealing with such operators.

An important thing to note here is that it's strictly a codegen optimisation: the type information isn't affected. I don't think it's possible to do the same trick during type assignment instead, it would break things.