|
|
fixed #2316: No longer cache entire SearchResult when looking for implicits in the parts of the expected type. (patch by retronym -- see ticket)
A SearchResult may contain symbols local to the scope of the search that
were used as implicit parameters, so they are not safely cacheable. The
fix for #2101 does not suffice. That patch avoided bound symbols being
duplicated, but the problem is much worse.
The implicits for an expected type depend on more than just that type,
so we cannot cache them using the expected type as a key.
The neg/t2316 test illustrates this: T1 may provide two implicits, one
requires an implicit T2, another an implicit T3. If an implicit T1
is first required when only a T2 is in scope, the SearchResult will
describe the corresponding implicit. Now, if we enter an implicit value
of type T3 into scope, the search should fail (it is ambiguous), but the
cache does not take this new fact into account.
The patch replaces the erroneous aggressive caching with a more
conservative version that only caches ImplicitInfo's.
|
