This commits introduces testdata changes, where NI behaviour strictly
improved, after several previous fixes.
For some tests, just WITH_NEW_INFERENCE directive was added. It
indicates, that some of previous commits first introduced error in that
test, and then some other commit fixed it (netting no overall testdata
change). It is preferrably to keep those annotations until we will
migrate to NI completely, to prevent unexpected regressions.
There are two different forms of types intestion:
1. Type parameters with multiple bounds
2. Smart casts
The problem was that when member scope of type intersection contained
effective duplicates and that lead to overload resolution ambiguity in
strange cases like `x.hashCode()`
For first type we do effectively the same thing as when building member
scope for class extending several interfaces: group all descriptors by
both-way-overridability relation and then choose most-specific in each
group.
For smart casts we do basically the same thing but with special
treatments:
1. From all descriptors that _equal_ to most specific we choose
the one that works without smartcast if possible (i.e. we choose first from candidates list)
2. If smart-cast value seems to be unstable we use only member scope
of receiver type + all descriptors from smart cast possible types
that has incompatible signature. If we'd include all of them and
choose one as more specific, and it would lead to false
SMART_CAST_IMPOSIBLE (see test unstableSmartCast.kt)
#KT-3996 Fixed
#KT-10315 Fixed
Dmitriy Novozhilov