tock/notes.md
2024-06-23 15:58:17 -04:00

71 lines
3.3 KiB
Markdown

# type system
data types:
i64, f64 - numbers
bool - logical boolean values
{fn [type1 type2 ... -> rettype]} / {fn [type1 type2 ...]} => [fn type1 type2 ... rettype]
{array [type length]} => [array type length]
(struct name
^type1 member1
^type2 member2
^type3 member3)
^{tuple [^{name member1} type1 ^{name member2} type2 ^{name member3} type3]}
^{name name} [tuple ^{name member1} type1 ^{name member2} type2 ^{name member3} type3]
(enum name
(clause1 type1 type2 ...)
(clause2 type1 type2 ...)
clause3)
{variant [^{tuple [type1 type2 ...]} clause1
^{tuple [type1 type2 ...]} clause2
^void clause3]}
Variables have both a datatype and an isolation modifier. There are three possible isolation types:
* `const` - this is the default, if no isolation modifier is given. No in-place mutations are possible with `const` values.
* `val` - a `val` variable can be mutated in-place, but changes _only_ affect that variable. If it is assigned to any other
variable or passed as a parameter, it is copied if necessary and can be treated as a new, totally distinct value.
* `mut` - Only valid on function parameters. Denotes a value in which mutations to the parameter are visible from the calling
function. Any assignments of a `mut` value to other variables (except being further passed as a `mut` parameter)
makes a copy.
All datatypes can have the following modifiers:
* `ref` - a `ref` is analogous to a full pointer or object reference. Copies of the same `ref` can exist in multiple places,
and refer to the same object in memory. Changes to data mutated via a `ref` are immediately visible to any other
code that has the same `ref`. `ref`s are created with the `box` function. References that point in the middle of
a structure are not possible; use a tuple or an enum or something.
* `opt` - equivalent to `(enum [opt type] (some type) none)`, assuming we had generics, which atm we do not. no idea what
destructuring helpers make sense here yet.
* `array` - contiguously-allocated values of any type. size can be dynamically specified, bounds are checked on access.
Q: should `const` structures be able to modify `ref` members directly? leaning towards yes. `ref` seems to inherently imply
interior mutability - after all, the value a given `ref` points to is _not_ constant.
Interesting thought: we only need GC for `ref`s! local variables + parameters can live in a stack-based arena, and
globals are statically allocated.
# typed dynamic dispatch
(protocol name
(method name [^mut self ^type1 arg1 ^type2 arg2 -> rettype]))
(impl protocol type
(method name [^mut self ^type1 arg1 ^type2 arg2 -> rettype]
body...))
# type restrictions
* `scoped` - would be nice to be able to define destructors and have some kind of RAII, but it's probably overkill
# memory management
* "hot reload" implies "given the previous source code and a memory, I can reason about the types of everything in the memory"
* compacting garbage collection is simply the degenerate case of rearranging memory to be legible to new code!
* we have typed roots (globals), and we can follow typed references from there
* stack is not accessible from wasm, and GC / reload would only happen when wasm code returns to JS - no suspension