trait item metadata encoding for supertraits is messy and ugly

[bblum]

As part of #8562 I have added the notion that traits have a disjoint set of super traits and super kinds. In the typeck data structures, these are represented separately, with the former in the tcx.supertraits map, and the latter as a field of ty::TraitDef.

However, when encoded in the metadata, I just recycled the old code which would encode super-builtin-kinds as though they were regular supertraits (fixme in metadata/encoder.rs), and on the other side, repartition them into the two distinct sets when reading cross-crate trait info (fixmes in metadata/decoder.rs).

A consequence of this is that in typeck/collect.rs, I had to make sure even the builtin supertraits were inserted into the tcx.trait_refs map so they could be encoded using the same old metadata code.

It would be better to encode a TraitDef's builtin bounds like the separate field that it is. This would also require, in the item_trait case in encoder.rs, iterating over the tcx.supertraits map's value instead of iterating over what the AST thinks the supertraits are.

[emberian]

I spent a few minutes looking at this and got lost somewhere in the depths of metadata. This is still relevant, but isn't that easy, unless you're already familiar with metadata.

Not a priority, though.

[nrc]

A note, that this will also allow us to cleanup the way we handle the unsized/sized bound/generalisation marker in encode/decode too.

[nikomatsakis]

I'm fixing this in one of my branches.

[jroesch]

This seems related to the clean up Niko started on the where clause branch for #5527. I've been working around all this code the past few days. cc me.

[arielb1]

This is not an issue in new rustc - all supertraits are stored in the "super_predicates" field.