DO NOT MERGE: experimental implementation of Entry API RFC

[cristicbz]

RFC: rust-lang/rfcs#1769

• Implements Entry::key and VacantEntry::key only when Q=K.
• Updates both BTreeMap and HashMap.
• The only inference breakage in the rustc codebase was due to Constraint does not match because where clause hides blanket impl #37138 .
• Uses the single trait no specialisation version with AsBorrowOf.
  • Turns out the lack of an IntoOwned<K> trait ends being annoying requiring a B parameter added to or_insert and or_insert_with.
  • Adding a <B> parameter to or_insert_with is, I think, a breaking change because writing or_insert_with<_> is not valid any more. In reality I imagine this never happens.
  • We may want to go back to the IntoOwned/AsBorrowOf separation.

[rust-highfive]

r? @sfackler

(rust_highfive has picked a reviewer for you, use r? to override)

[cristicbz]

r? @aturon

[brson]

Starting crater.

[cristicbz]

Travis failures are entry doctests for which inference fails now :(. I'll fix them, but shouldn't affect crater run.

[brson]

https://gist.github.com/anonymous/d16c1c231ea7dd737e6cd5606a686952

[cristicbz]

@brson Thanks a lot for doing the crater run!

Here's my analysis of the results. There 9 irreconcilable regressions caused by actual inference failures introduced by widening the allowed argument types of Entry, split about evenly between:

• Short lived local HashMap with reference key types whose type can no longer be inferred from .entry(some_reference).
• Usage of Into to convert from &str -> String in calls like string_map.entry(some_str.into()) (this is ambiguous since both the intended conversion and &str -> &str are valid).

Instances of #37164 and #37138 appear once each.

[alexcrichton]

I'm going to close this in favor of the RFC for now, just trying to clear out the PR queue. I'm fine reopening though once we decide on the RFC!

[cristicbz]

@alexcrichton Makes sense, in the meantime I'd love to get more feedback on the RFC though!