Improve on support for ... and ..i in odd places
#157
Conversation
It doesn't seem to be respected by VS Code otherwise
They break `tree-sitter test --update` in weird ways
| field("name", $.identifier), | ||
| $._parameter_name, | ||
| "=", | ||
| optional(field("default", $._expression)) | ||
| ), | ||
|
|
||
| _parameter_without_default: $ => field("name", choice($.identifier, $.dots)), | ||
| _parameter_without_default: $ => $._parameter_name, | ||
|
|
||
| _parameter_name: $ => field("name", $._identifier_or_dots_or_dot_dot_i), |
There was a problem hiding this comment.
Can now collapse the parameter "name" field for with/without default cases into a singular _parameter_name, since the definitions have converged
| // Identifier-ish, but useful enough to be their own nodes | ||
| dots: $ => "...", | ||
| dot_dot_i: $ => token(withPrec(PREC.DOT_DOT_I, /[.][.]\d+/)), |
There was a problem hiding this comment.
Moved these to be very close to identifier, since they are indentifier-ish and always get lumped in with identifier now
| // NOTE: Technically R allows `...` and `..1` anywhere we want an `$.identifier`, | ||
| // but practically it can be useful for downstream consumers to have separate | ||
| // nodes for these particular constructs. Our compromise is to keep these as separate | ||
| // nodes, but then use this in most places we want an identifier. | ||
| _identifier_or_dots_or_dot_dot_i: $ => choice( | ||
| $.identifier, | ||
| $.dots, | ||
| $.dot_dot_i | ||
| ), | ||
|
|
||
| // NOTE: Having this as an actual node (rather than inlining the `choice()`) somehow | ||
| // ends up allowing better error recovery in a few cases | ||
| _string_or_identifier_or_dots_or_dot_dot_i: $ => choice( | ||
| $.string, | ||
| $.identifier, | ||
| $.dots, | ||
| $.dot_dot_i | ||
| ), |
There was a problem hiding this comment.
I played with making these names shorter but ultimately I think it made things more obscure.
It's easier to read the call sites with these long explicit names that tell you exactly what can go there.
| // NOTE: This is exactly `_string_or_identifier_or_dots_or_dot_dot_i` but with | ||
| // a precedence of 1. It seems like we have to set the `prec(1, )` on the `choice()` | ||
| // directly, we can't reuse `_string_or_identifier_or_dots_or_dot_dot_i` here | ||
| // otherwise `tree-sitter generate` throws an unresolved conflict error. | ||
| // | ||
| // This is only for use in `_argument_named`. | ||
| // | ||
| // Since `_argument_unnamed` can be an arbitrary `_expression` (with precedence 0) | ||
| // which includes `string`, `identifier`, `dots`, and `dot_dot_i`, there is an | ||
| // ambiguity between: | ||
| // - Starting the `value` of an `_argument_unnamed` | ||
| // - Starting the `name` of an `_argument_named` | ||
| // | ||
| // We set a higher precedence here to try and match `_argument_named` first. | ||
| _argument_name_string_or_identifier_or_dots_or_dot_dot_i: $ => prec(1, choice( | ||
| $.string, | ||
| $.identifier, | ||
| $.dots, | ||
| $.dot_dot_i | ||
| )), |
There was a problem hiding this comment.
Slightly tricky case where the "name" field of _argument_named is technically _string_or_identifier_or_dots_or_dot_dot_i, but we need to bump the precedence there to disambiguate it from _argument_unnamed.
This replaces the prec.right(1, ) that has been removed from _argument_named
Closes #148
This is motivated by needing support for
x$..., which is the most common case of this weirdness. But I've managed to tighten up quite a few of these extremely obscure cases while making things consistent. All of these are valid R code:x$...andx@...x$..1and[email protected]foo::...and...::foo(and:::versions)foo::..iand..i:::foo(and:::versions)for (... in 1:2) print(get("..."))for (..1 in 1:2) print(get("..1"))function(..1) get("..1")function(..1 = 2) get("..1")fn(..1 = 2)fn(... = 2)I went with the approach we talked about of keeping
$.identifier,$.dots, and$.dot_dot_ias separate nodes (because they are useful to downstream consumers), and making a union type of$._identifier_or_dots_or_dot_dot_ithat gets used internally.The commits are technically backwards. I wrote the impl but didn't commit it until I committed all the updated tests first, just to make sure it was correct. Probably easiest to read each commit individually.