This essentially undo parts of 307d3d0. While it seemed a like good
idea to over-modularize type classes to reduce header dependencies, I
think it was a mistake. What 307d3d0 did was basically split each of
the components of a type class into a single header (typeclass/operators.hpp,
typeclass/mcd_1.hpp, typeclass/mcd_2.hpp, ...).
At first, it resolved many weird header dependency glitches. However, it
also made everything more complex; creating even easy type classes was
sometimes much longer than it should have been, and using type classes
was tricky because you had to know exactly what to include. It also went
against the idea of implicit type class instances being provided whenever
that's possible, which I think is a nice feature of the library. Being
dissatisfied with this, I opted for a simpler header organization with
a fwd/ directory that contains forward declaration headers, and everything
else in the same directory.
A possible objection to this change would be that you are now forced
to include sometimes more than what you strictly need when e.g. defining
an instance or using only some instance(s) of a data type. My answer to
this is that Hana is a really small library and the parsing is not
going to have a huge impact on overall compilation time. My bet is that
the time that will be saved by programmers with a simple header hierarchy
outweights the parsing time by far.
- Split type class instances into separate files
- Instances provided automatically by a type class are actually MCDs
- Test each instance in a single file, not one file per method
- Refactor the operator system to fix the ADL-related bug.
This allows us to mechanically translate most type traits, which is
a good hint that we've found a way to merge type computation with
value computation. The problematic type traits will be tackled later.
