struct snippet yields idempotency error

[Niols]

As of 9b2de13,

struct
  type t = unit
end

gets formatted once as

structtype t =
  unit end

which is then rejected by the OCaml tree-sitter grammar.

Note that the former piece of code is not valid top-level OCaml. It does get accepted by the tree-sitter grammar, which probably is a feature so that we can highlight snippets? in which case it would make sense to also be able to format such snippets?

@Xophmeister> An opinion on this “formatting snippets” question?

[Xophmeister]

In general, I completely agree with you: de-contextualised snippets should be formattable, even if the result is not valid top-level code. The Tree-sitter grammar for all languages (AIUI) will always,¹ however, assume it's valid top-level code and the root node of the tree will be something like program or source (in OCaml's case, it is compilation_unit).

As such, the formatting queries we write should, in principle, accommodate snippets.

However, having said that, in this case, while Tree-sitter parses your snippet, the parse tree is different to one that is correct top-level OCaml.

Parse tree for snippet:

struct
  type t = unit
end

Parse tree for top-level version² (highlight mine):

module Foo = struct
  type t = unit
end

This is why the formatting is wrong, in this case: The snippet version gives a completely different parse tree to what's defined by the formatting queries. The type_definition node is close to what it should be, but even then that end token forces the parser to parse the RHS into something unexpected.

So, to extend my answer: the formatting queries we write should accommodate snippets, providing their parse tree doesn't change due to context. Otherwise, we'd have to potentially duplicate each set of queries for in-context and out-of-context, which would be a lot of work and may cause problematic interactions.

Footnotes

1. I suppose it might be possible to define multiple valid root nodes and for Tree-sitter to assign the right one. Maybe; IDK 🤷 ↩

2. According to ChatGPT. YMMV! ↩