At first I attempted to make the whole thing reinterpret_cast-able from
a blob of memory (i.e., truly zero-overhead), but while that sounded
cool and all, it moved the overhead to basically all other code -- each
function had to special-case access to attribute/vertex/index data as
the pointers were no longer pointers, the binary representation had
various weird unexplainable gaps ("here an array deleter is stored, set
that to null and don't ask"), release*() functions got more complicated
and when I got to issues with move construction/assignment I knew this
was not the right path.
Now the MeshData internals are packed to a much more compact
representation (with the first attempt it was 128 bytes, now it's just
64) and the serialization doesn't make everything else slower, more
complex or harder to test, which is a win.
Vladimír Vondruš