Kurt Stephens

Pointers to data structures are almost never actually used “as is”, they are almost always used with a offset into a structure starting at an address.

struct cons {
  value type; 
  value car;
  value cdr;
};

For 32-bit word-aligned pointer ptr, encoded with 0x03 tag, accessing the car field in the hypothetical cons struct above, becomes *((ptr - 3) + 4), which constant-folds to *(ptr + 1). AFIAK, there is no difference in instruction time between *(ptr + 4) and *(ptr + 1). However, on some RISC architectures such may not the case.

As others have pointed out, all tag schemes have trade-offs. A trade-off for using non-zero bits for pointers might mean extra tag removal when communicating pointers to low-level code, e.g. C libraries. Fixed-size tags would likely simplify code and allow additional values besides small integers to be immediate values.

I would encode Symbols like every other allocated object.

— Kurt