Ketil Z. Malde wrote (snipped)
Would it be possible to separate binary I/O in two layers, the binary layer converting to/from (U?)Arrays of Word8, and another, the I/O layer, reading and writing such arrays?
I think the proposal I posted is better because it allows you to mix up both single Word8's and arrays. In my experience, structures to be exported typically contain both. For example, I pack integers as a sequence of bytes (so that integers between -64 and 63 can be represented by one byte, for example). Constructing singleton arrays for each of these bytes would be tiresome, and presumably not as efficient as writing them directly into the target memory area. Here are two examples of cases where I want to write binary data to memory, not to a Handle: (1) Suppose I maintain a dictionary indexed by Strings (say) whose values belong to different types with binary representations, and I want this dictionary too to have a binary representation. Whenever I look something up in the dictionary I know what type I'm expecting, but I don't know what types to expect when decoding the dictionary. Then the most practical encoding seems to be as [(String,Bytes)], where each Bytes contains an encoded entry in the dictionary. (2) If I want to process binary data in some way before sending it to the outside world, for example by compressing it using zlib (the API to the GZIP algorithm), or by encrypting it.