This will make it easier to generate integer identifiers when
generating production code (but st...

The graph nodes are set up for cyclic values, tail calls, and effect
managers. This also tracks ...

If we know the frequency of each field, we can assign them short/unique
names. So the 53 most us...

So far it converts everything except for case expressions. Need to make
some choices about how c...

The plan is to rely on JS minifiers for inlining of single-use
variables.

Longer-term, we would...

The main goals here are:

1. Gather usage counts for each argument. This will make it easier to
...

This organizes all the constraints from the value declarations and the
ports/effects in the modu...

Instead of a custom Solver monad, the new Type.Solve code uses plain IO
and threads some state a...

They will always be prefixed, so it is silly to follow that Haskell
convention of repeating the ...

Rather than using a ErrorT/StateT/IO monad stack, the new code uses a
CPS’d version that only tr...

I noticed that every use of CLet had a list like [Scheme] where there
was only one scheme. So ra...

The old version builds Bags as you go, turning it into dictionary at
the end. That means adding ...

Create new “Expectation” type that holds context information about why
certain things should be ...

Use a new format for error reporting. It provides both the Context of
the constraint and the Cat...

This introduces the AST.Canonical.Annotation type that holds on to all
the free type variables i...

The code seems to be better if it is more explicit on these points.

Remove the AST.Type module and distribute it between AST.Source,
AST.Canonical, and AST.Utils.Type

All this logic moves to constraining expressions and patterns
specifically. This makes it easier...

Rather than having all the logic that uses the core code in these
modules, the special uses are ...

This change started because I could not remember if an Int was
zero-indexed or one-indexed. Now ...

Binops can only be used outside of the module they are defined in.

Binops need to work exactly ...

New parser for numbers allocates a bit extra. Seems kind of lame. Need
to test if providing func...

This revealed that it is pretty silly to allow floats in pattern
matches. They had that only bec...

This also gives me an indexedMap function that I have been missing.

All the type information needed is stored on the PCtor constructor
during the canonicalization p...

The error types are shuffling around a bit as more things are caught
and as the code for catchin...

Moved the type instantiation code to Type.Instantiate. I suspect there
will be a couple differen...

Also move arity check of unions and aliases out into Canonicalize.Type

This means we will not need to look any types up when generating type
constraints for patterns. ...

The old version allocated lists of booleans, then used List.or to check
them all. This meant all...

The new getVarNames implementation is just in IO, not in StateT. I
suspect this will be a decent...

This now matches the limits in AST.Expression.Canonical

Also try to reduce allocations a bit by referring to NOINLINE’d names
when doing comparable tests.

Goal is to avoid allocating new versions.

The new implementation has a tail recursive gatherFreeVars function.
The idea is that you can av...

The “home” of any particular variable or type is tracked in the AST
directly now.

Thanks to warning about unused variables.

The changes to the AST mean that this happens during validation. It is
definitely simpler now, a...

This one allows cyclic values in top-level declarations as long as the
cycle includes at least o...

Some highlights include:

- The “definition sorting” moves into expression canonicalization.
- C...

Mostly updating for changes in the AST code

Stop supporting object fields that are quoted or numbers. Just use
normal names.

Things that moved to canonicalization include:

- Checking for free variables in unions and alia...