Previously, gorename rejected all method renamings if it would
change the assignability relation.
Now, so long as the renaming was initiated at an abstract
method, the renaming proceeds, changing concrete methods (and
possibly other abstract methods) as needed. The user
intention is clear.
The intention of a renaming initiated at a concrete method is
less clear, so we still reject it if it would change the
assignability relation. The diagnostic advises the user to
rename the abstract method if that was the intention.
Additional safety checks are required: for each
satisfy.Constraint that couples a concrete type C and an
interface type I, we must treat it just like a set of implicit
selections C.f, one per abstract method f of I, and ensure the
selections' meanings are unchanged.
The satisfy package no longer canonicalizes types, since this
substitutes one interface for another (equivalent) one, which
is sound, but makes the type names random and the error
messages confusing.
Also, fixed a bug in 'satisfy' relating to map keys.
+ Lots more tests.
LGTM=sameer
R=sameer
CC=golang-codereviews
https://golang.org/cl/173430043
Rewrite performed with this command:
sed -i '' 's_code.google.com/p/go\._golang.org/x/_g' \
$(grep -lr 'code.google.com/p/go.' *)
LGTM=rsc
R=rsc
CC=golang-codereviews
https://golang.org/cl/170920043
The type checker does not currently report the set of pairs of
types that are tested for assignability (though gri and I
agree that it should). This information is useful for many
applications. For example, refactoring tool needs to know the
minimal set of interface satisfaction constraints that must be
preserved during a refactoring.
This package is a stopgap measure to deduce the same
information using another pass. Unlike go/types, it is not
robust against inputs containing type errors.
LGTM=sameer
R=gri, sameer
CC=golang-codereviews
https://golang.org/cl/136470043
