As of now, at least four of the compiler test programs are passing. I still need to work through a lot of stuff:
- Infix operator precedence rules are busted
- Type inference for certain expression forms is broken
- Overload resolution doesn't quite work all the time
- Parser and compiler diagnostic errors are terrible (or just plain non-existent)
- There's still lots of room for optimizations
- Much of the code needs cleanup
- And of course there are vast swaths of documentation to write
So it'll be a while yet before Release 12 ships, but we're getting there. The end is definitely in sight, and it feels great to be in a position where the remaining tasks are just some bullet points and bugs rather than "hey, go rewrite an entire compiler! Yayy!"
Probably the single biggest task will be getting decent diagnostic errors from the compiler. That's a huge chunk of territory I haven't even begun to mess with in this new compiler architecture, and it'll be time before I have a good sense for how much effort will be involved in getting good errors out. Until then, any broken program is likely to just fail in cryptic and unhelpful ways.
I'm pondering kicking R12 out the door without decent errors, but I don't think it's the best idea ever.
Speaking of ideas, I'm thinking of changing a couple things in the future (probably not R12 because I already have enough to do for that milestone).
Currently, Epoch programs look like this:some_function : (integer(param)) -> (string(return, ""))
{
real(some_variable, 3.14159)
}
What if they looked like this, instead?some_function: integer<'param'> -> string<'return'>("")
{
real<'some_variable'>(3.14159)
}
Function definition syntax loses some extraneous parentheses, and variable definitions get called out in two specific ways: angle brackets indicate compile-time code invocation, and single quotes indicate identifiers that have not yet been used in the program. This distinguishes the variable definition foo<'bar'> from the function call foo(bar) in a clean way.
Thoughts?
[CODE]
some_function : integer:param -> string:return("")
{
real:someVariable(3.14159)
}
[/CODE]
Or alternatively...
[CODE]
some_function param:integer param2:real -> return:string
{
return = cast(string, param) ; cast(string, param2)
}
[/CODE]
I don't really like making ' and " different in any way. In fact, I much prefer the PYTHON style of quotation, where you just have to match the appropriate ' or ". Making them behave different is just error prone.
Also in that realm is that <> is a lot of extra typing (shift and all that) for very little gain.
The latter example of those two above could be read almost as
"some function that takes a parameter of type integer and a parameter of type real and returns a string."
The first example isn't too hard to parse mentally either:
"some function that takes an integer parameter and return a string result."
I also find object construction to be a bit wonky as well...
[CODE]
SomeType(variableName, constructionParameter1, constructionParameter2)
//or .. depending on the TYPE that is SomeType (fundamental vs structural)
SomeType variableName(cp1, cp2)
[/CODE]
A preference should be given for a single, clear, object (of any type) initialization. My personal preference in this case would be:
[code]
variableName = SomeType(cp1, cp2)
myInt = integer(12)
myString = string("")
myBuffer = buffer(512)
//OR, to make the language parsing a bit easier...
variableName := SomeType(cp1, cp2)
myInt := integer(12)
myString := string("")
myStr2 := string('note the quotes')
myBuffer := buffer(512)
[/code]