Discover/link positional [de]construction about haskell-indexer HOT 2 OPEN

I added some comments on the slightly related #19 .

A quick summary of the current state after quickly checking https://github.com/google/haskell-indexer/blob/master/kythe-verification/testdata/basic/RecordReadRef.hs - correct me if it is wrong:

1. now a non-literal deconstruction like bar (Foo a _) = ... will only emit a binding for a, and connect the usages of a to it. There isn't any edge going from either the binding span or the ref spans to the field. (this is why @blackgnezdo opened the issue).

2. explicitly named deconstruction like bar Foo { x = a } = ... will connect x to the field binding. Usages of a won't be listed in the backrefs for field x.

3. punned deconstruction like bar Foo {x} won't introduce a local binding (?), both the pattern and the usages will directly reference the field binding. Which is a bit odd highlight-wise and semantically (say, all x refs in the file are hilit even when hovering a local usage in a single function?), but at least brings the usages in the backref.

4. For non-record constructor data Bar = Bar Int, not sure - maybe we don't even emit a field decl at the span of Int (?), so implicit matches don't even have the target to reference.

IMO, a user generally wants to see how a field is used in a specific context when looking through the backreferences. The various cases are:

• current punning refs in (3) are not semantically nice, but bring the backref

• for (1), (2) and (4), adding an edge from the pattern-matched name to the field would bring them in the backrefs, but if the pattern name would differ from the field name, the listing could look odd (the UI can solve this by bolding the reference in the listed line to resolve ambiguity)

As a unified solution, my gut feeling is that a local usage of a deconstructed var should directly reference both the local binding and the generic field binding in the data definition.

from haskell-indexer.

See discussion in kythe/kythe#3934 - TLDR the local binding should bind a variable/import node (maybe the subkind is not important, could be left or be field as well?), which should have an aliases edge towards the generic field binding. The index preprocessor (with appropriate config) will merge the backrefs for the local usage into the field's usage.

/cc @zrlk to verify I understand correctly

from haskell-indexer.