Signed and unsigned don't quite capture the essence of what the different
combinations of x86 flags are doing for SubX. The crucial distinction is
that one set of comparison operators is for integers and the second is
for addresses.
When I created it I was conflating two things:
a) needing to refer to just the start, rather than the whole, and
b) counting indirections.
Both are kinda ill-posed. Now Mu will have just `addr` and `handle` types.
Normal types will translate implicitly to `addr` types, while `handle`
will always require explicit handling.
I can now run this program:
fn main -> result/ebx: int {
result <- do-add 3 4
}
fn do-add a: int, b: int -> result/ebx: int {
result <- copy a
result <- add b
}
We still can't define local variables, but can write any programs involving
ints by passing in enough arguments for temporaries.
We haven't implemented it yet, but there's now a design for how we check
whether a function has written its output correctly. Functions must write
to each output at the top level at least once, and never overwrite an output
register in the top-level once it's been defined.
This is conservative (it can be perfectly reasonable for functions to write
the output, reuse the register for a temporary, and then write the output
again) but easy to check.
One test failing. It took enough debugging just to get to the expected
failure that it seems worth saving this snapshot.
Saw some signs that I have to remember to zero-out allocated memory. We
need a scalable solution for this.
I think parse-var-with-type needs to be rewritten. Just added a test and
a hacky fix for now.