I uncommented append_iter and made it compile. I hope it wasn't
horribly flawed in some other way... but if so, there was no comment
explaining how.

There was a FIXME noting that ty::enum_variants and typeck::check::
check_enum_variants both call eval_const_expr. I tried refactoring the
code so that check_enum_variants does all the work and enum_variants
just looks up cached results, but this turned out not to be easy because
several ty functions call enum_variants and it might get invoked on an
enum before that enum item has been typechecked. Instead, I just made
check_enum_variants update the enum_var_cache so that enum_variants
will never call eval_const_expr twice on the same const.

This reverts commit 31f4b63d.

It was a little hard for me to believe, but it seems that re-exporting
an impl doesn't work at a, because encoder::encode_info_for_mod requires
that all the impls in the current module's impl map be local (that is,
bound to a value in the current crate's item map). Fixed it.

Closes #2414.

Issue 2526 showed a test case where a library exported only a type
that was a synonym for a class. Because the class's destructor wasn't
getting marked as reachable, its linkage was wrongly getting set to
"internal". The solution is for reachability to traverse types.

Closes #2526.

This is useful on OS X in order to handle the .dSYM files.

Incoming