Document alignment & underlying size invariants for array types in GHC.Prim
Through a combination of reading some of the wiki entries on the RTS, some trac tickets, and experimentation I've inferred that the payloads for bytearray types will always be aligned to the machine word size, even after GC. I guess this is obvious in retrospect since otherwise non-pinned arrays would be pretty useless.
Another thing I've picked up is that the memory backing a ByteArray is always evenly divisible by the machine word size: http://haskell.1045720.n5.nabble.com/Is-it-safe-to-index-a-little-bit-out-of-bounds-td5872779.html
It would be great if these were documented, but more importantly (to me), I'd like to be able to rely on them as invariants in my own libraries. The first is obviously important (any code doing reads larger than single byte relies on it), but the second one is also really useful: e.g. I'm working on a hashing library that works on text
and reads machine words for efficiency, doing endian adjustments as needed; it's a little easier and more efficient to do word-aligned word-sized "dirty" reads that cover all the data and then to clean up either end.