Currently we take the AST and transform it into a semantic tree. To better support IDE-like features we will need to support more variations in transformation.
With input trees immutable we could have a syntax transformer which
uses the visitor pattern. Would then have a visit_ for each kind of
Expression and have them produce a new expression. New expressions
would have synthesised Span information.
Initial optimisations/transformations would be a desugaring pass over
this AST before lowering to the sem tree. Maybe even want to do the
type check pass at this level and pass a fully-formed type checked
tree to sem. If we do add a desugaring pass then the output of that
would be nice to have as an --emit value.
This would be useful to observe the desugaring of things like for
into while.
Further use of these would be to implement a language server in the future and some kind of code formatter.
Given initial support for syntax transformations we could have a visitor for the tree which writes the tree to the standard output to prove out the round-tripping. Add a new set of tests which just round-trips all of the code from the specs.
API surface could be similar to Minsk's Compilation::EmitTree. I'm
thinking we come up with a trait something like TreeSink and
implement it for IO::Write or similar. Can then have a similar
emit_tree which walks the tree and pushes it to the writer.
I'm thinking we might also want to have this tree emission part of a
higher level --emit= flag on the compiler front end. Other values
for now would be to emit the Debug version of the parsed tree and to
emit the compilation result as IR, an object file or a linked
executable.
dumpast and dumptreeThe current flags for dumping compilation state may be best grouped
under a single --dump=<type> flag. This could allow for different
outputs: