Abstracting pointer arithmetic

[Lars Brinkhoff]

kenner@vlsi1.ultra.nyu.edu (Richard Kenner) writes:
>     I'm writing a back end for a word-adressed machine.  For this machine,
>     pointer arithmetic doesn't quite work like integer arithmetic.
>     Pointers to sub-word objects have a word address in the low bits, and
>     a byte selector in the top bits, and are incremented using a special
>     instruction.  Subtracting or comparing these pointers takes several
>     instructions.  Pointers to words are incremented by adding 1*<number
>     of words>, not 4*<number of words>, but otherwise behave as integers.
> For *those* machines you can define Pmode as a different class.  I thought
> you were proposing doing it for *all* machines in order to support those.

No, I didn't mean to do that.

So with the clarification that my suggestion was only to apply for
machines with exotic pointer aritmetic (say, #define EXOTIC_POINTERS
in the target description), was my suggestion reasonable?

Or would something at the tree level still make more sense?