FFI

[evincarofautumn]

• Allow declaring foreign modules (dynamic libraries) and importing definitions
• Make foreign definitions transparently callable

[evincarofautumn]

I have a prototype which parses C headers and extracts declarations of functions and types. I have decided on the following mapping between C types and Kitten data types.

struct

Data types with one constructor have boxed struct layout.

data point:
  case point (int, int)

struct point {
  int field0;
  int field1;
};

enum

Data types with one or more constructors with no fields have unboxed enum layout.

data bool:
  case false
  case true

enum bool {
  false,
  true,
};

ADT

Data types with multiple constructors with fields have boxed union of struct layout:

data number:
  case int (int)
  case float (float)

union number {
  word tag;
  struct {
    word tag;
    int field0;
  } case0;
  struct {
    word tag;
    float field1;
  } case1;
};

union

Data types with anonymous constructors have union layout. They are constructed and deconstructed using casts. Unspecified casts are disallowed.

data number:
  case (int)
  case (float)

1.0 as (number) -> pun;
pun as (int) -> converted;
…

union number {
  int case0;
  float case1;
};

Integral and Floating-point Types

Kitten’s int type will be parameterised by a bit size and a signedness. Integer literals may be affixed with a size and/or signedness, which will default to 64 bits and signed, respectively. There is a mapping from Kitten integer types onto C99 integer types via stdint.h, but the inverse mapping is platform-defined and will require (compile-time) range checks. I would like to have an arbitrary-precision integer type available for correctness purposes, but I am loath to use GMP for licensing reasons. float will similarly take a bit size.