Unholy Grail

Unholy Grail is a new programming language which compiles to Python.

It is quite alpha at the moment, but most of the features work.

You can get a feel for the language with the examples in the examples/ folder. They can be executed with the command: ug eval <filename>. You can also execute bits of code with ug eval -s <code>.

Here's what a recursive implementation of fibonacci looks like in UG:

fib =
    [0] -> 0
    [1] -> 1
    [n] when (n > 1) ->
        fib[n - 1] + fib[n - 2]

print["The 10th fibonaccy number is", fib[10]]

Current features

• No distinction between expression and statement.
• Featureful pattern matching. It works almost everywhere: variable declaration, parameter declaration, loop variables, exceptions, deconstruction into dictionary keys and as a standalone match construct.
• Optional "typing", which is in fact part of the pattern matching syntax.
• Ad hoc structures: #point[1, 2] instantiates a kind of tuple of type #point, #point[x = 1, y = 2] an object with named fields. They can be pattern matched.
• Exception handling operator: result = (1/0 !! None) or result = (1/0 !! ZeroDivisionError e -> None). This uses the same pattern matcher as everything else.
• A very simple abstract syntax. The nodes produced by the parser are identifiers (strings), #value, #begin, #square, #curly, #oper and #juxt. That's it.
  • However, macros may return other kinds of nodes such as #declare, #assign or #splice.

Future features

• A well-integrated macro system over the aforementioned AST.
• Expression interpolation in strings.
• Expressions evaluated in macro scope.
• Syntax to tag expressions with their location in source code.

Differences with Python

Superficially, they look quite similar, but there are a few drastic differences beyond what I just mentioned. You might not like them, but I do, and I actually have reasons behind these changes.

The reasons pretty much all boil down to syntactic consistency: for instance, function calls use [] and not () simply because f(x) should be equivalent to temp = (x), f temp, and obviously in most current languages that is not the case. If it was, their syntax would have at least one less rule!

In UG, f[x] is equivalent to args = [x], f args, f[x, *args, y = z, **kwargs] is equivalent to args = [x, *args, y = z, **kwargs], f args, f.x is equivalent to method = .x, f method, and so on.

Function calls

• Function calls use [] and not (). You must therefore write open["file"] and not open("file"). open("file") is still legal in UG, but it parses the same as open "file", which is equivalent to open.file.
  • The ! operator can serve as shorthand for calls sometimes: open ! "file" is like open["file"] and f! is like f[], but it has very low priority, so it's not usable everywhere.
• Indexing is the ? operator. You must therefore write [1, 2, 3] ? 0 and not [1, 2, 3][0].

Declaration and assignment

• = declares and sets variables. := assigns to them.
• Indent and (...) both delimit fresh scopes.
• To minimize confusion, I am considering not allowing the redeclaration of a variable declared in an outer scope using =. It would be allowed using an as of yet hypothetical let construct.

The comma

• (1, 2, 3) is not a tuple. It evaluates to 3.
• The comma (,) replaces ; in all instances you might want to use it.
• Since we just freed up that character, comments start with ;, not #.
• The comma has lowest priority in all situations. a, b = 1, 2 therefore parses as a, (b = 1), 2. You must write [a, b] = [1, 2].
• Every line break is equivalent to a comma. Every single one (unless there is a line continuation).
  • Technically, it's an "indent comma" -- the only difference with a standard comma is that multiple consecutive commas insert Void objects in the holes, but indent commas do not.

Indent and whitespace

• Tabs are not legal anywhere.
• Any indented block is equivalent to appending ( to the line right before the block, and ) to the last line of the block. That's it. Any code using significant indent can thus be rewritten without any indent.
• Unless the block is located within brackets. All indent is ignored between (), [] or {}, but warnings may be given if it looks like they should be significant. I might allow tabs there.
• Short juxtaposition, e.g. a[b], binds tighter than wide juxtaposition, e.g. a [b]. a[b] c[d] therefore parses as (a[b])(c[d]).
• Whitespace disambiguates fixity. a - -b is (a - (-b)), but a- -b is (a-)(-b).

Data structures

• This is not done yet, but data structures will follow a more consistent pattern than they do in Python:
  • Ordered collections (square brackets)
    • Immutable
      • tuple: [1, 2, 3]
      • immutable ordered dict: [a = 1, b = 2] or ["a" => 1, "b" => 2]
    • Mutable (mut macro)
      • list: mut [1, 2, 3]
      • ordered dict: mut [a = 1, b = 2] or mut ["a" => 1, "b" => 2]
  • Unordered collections (curly brackets)
    • Immutable
      • frozenset: {1, 2, 3}
      • immutable dict: {a = 1, b = 2} or {"a" => 1, "b" => 2}
    • Mutable (mut macro)
      • set: mut {1, 2, 3}
      • dict: mut {a = 1, b = 2} or mut {"a" => 1, "b" => 2}
• You can create "hybrid" structures that have a tuple part and a dictionary part: [1, a = 2].

Function declaration

• There is no def. All functions are declared with the -> operator. I have a plan to automatically name things when assigned to a variable with =.
• Arguments are either positional or by keyword, not both, though it may be possible to make a patterh matcher that honors Python's behavior (I'd only need to add a little thing to the protocol).

Notes

Python's magical __call__, __getattr__, __setattr__, __getitem__ and __setitem__ methods are all collapsed into a single magic method in UG: __recv__. Defining a function with pattern matching essentially implements __recv__ in a singleton.

If you install Terminus, you can look at pretty versions of the tokenizer's output as well as the AST before and after the application of the operator macros. Syntax errors also require that terminal to show up nicely at the moment (that'll be fixed in due time, of course).