If the number of warnings is huge the '--'/2 operator is slow.

The Maps implementation handles ?unit in more cases.

Exactly when t_is_none_or_unit() is to be called is not clear to me.
The added cases are about a map type being ?unit, but the key or the
value of an association can also be ?unit, but that is not always
checked.

As of commit 12b379 many of the reasons displayed after
"Analysis failed with error:"
erroneously include a call stack. The bug is now corrected.

The bug was pointed out in
http://erlang.org/pipermail/erlang-questions/2017-October/093838.html.

~tw and new string functions are new since OTP-20 (stdlib-3.4)

The implementation of OTP-14218 (commit 6d3b38a) has a weakness: only
a very limited part of the type form is checked. This is now fixed:
types not used by specs are checked equally well as types used by
specs.

The new function erl_types:t_from_form_check_remote() checks usage of
remote types. It does not expand used local types, and has (almost) no
limits on depth and size.

This commit removes the call to ets:safe_fixtable/2 introduced in
commit 9458495. ETS memory can now be reclaimed while merging PLTs.

A missing call to ets:safe_fixtable/2 meant that information
was sometimes lost when merging PLT:s.

Memory consumption is marginally reduced.

Instead of two records, #plt{} with dict:s and #mini_plt{} with ETS
tables, one record is used for representing PLT:s in RAM.

The #mini_plt{} is the one now used throughout analyses, but it is
called #plt{}.

When writing the PLT to file, another record is used, #file_plt{} (as
before). When creating #file_plt{}, the #plt{} is deleted (it cannot
be used for further analyses).