Using the NHC.Binary library

This document sketches the York NHC.Binary library. (See also the BinArray library for an example of the use of Binary to build other abstractions.) For fuller details, see this paper.

The NHC.Binary library

    module NHC.Binary where

    data BinPtr a = ...
    data BinLocation = Memory | File FilePath BinIOMode
    data BinIOMode = RO | RW | WO
    data BinHandle = ...

    stdmem     :: BinHandle
    openBin    :: BinLocation -> IO BinHandle
    freezeBin  :: BinHandle -> IO ()	-- changes BinIOMode to RO
    closeBin   :: BinHandle -> IO ()

    putBits    :: BinHandle -> Int -> Int -> IO (BinPtr a)
    getBits    :: BinHandle -> Int -> IO Int
    getBitsF   :: BinHandle -> Int -> BinPtr a -> (Int, BinPtr b)

    seekBin    :: BinHandle -> BinPtr a -> IO ()
    tellBin    :: BinHandle -> IO (BinPtr a)
    isEOFBin   :: BinHandle -> IO Bool

    copyBin    :: BinHandle -> BinLocation -> IO BinHandle
    copyBits   :: BinHandle -> BinPtr a -> BinHandle -> BinPtr a -> Int -> IO ()
    copyBytes  :: BinHandle -> BinHandle -> Int -> IO (BinPtr a)

    class Binary a where
        put    :: BinHandle -> a -> IO (BinPtr a)
        get    :: BinHandle -> IO a
        getF   :: BinHandle -> BinPtr a -> (a, BinPtr b)
        sizeOf :: a -> Int

    putAt  :: Binary a => BinHandle -> BinPtr a -> a -> IO ()
    getAt  :: Binary a => BinHandle -> BinPtr a -> IO a
    getFAt :: Binary a => BinHandle -> BinPtr a -> a

Programming model

Both in-heap data compression and binary I/O can be achieved using the York Binary library. The basic model is rather like file I/O: binary data resides in a separate space which is accessed only through a BinHandle acting like a buffering file descriptor. Each item of binary data lies at a particular position within the space, the position being denoted by a BinPtr. Data can be written and read sequentially just as with ordinary files. Also, like ordinary files, we allow random-access reading and writing. However, the particular beauty of this scheme is the ability to engage in pure, lazy, random-access reading when a BinHandle is in the appropriate RO (read-only) mode. (A BinHandle which is already open for writing can be changed to RO mode with the freezeBin call.)

BinHandles do not just denote files - they can also refer to areas of heap memory. One such area is available by default - called stdmem - but new areas can be opened in just the same way as files. They are opened in the default mode RW. Binary heap areas grow automatically to fit the data placed in them, and, like files, they are naturally garbage-collected when they are no longer in use. (The closeBin operation is an explicit means to close a file or discard some memory.)

There are in principle two layers to the library functions. At the lower level, functions like getBits and putBits deal with raw bounded integers in the bit-stream. At the higher level, a type class abstracts these operations across arbitrary datatypes, providing overloaded functions put and get.

Low-level raw bit-stream functions

Each BinHandle has a notion of its ``current'' position. This is the position at which a subsequent read or write operation will start. You can think of it as a bit-offset from the start of the file/memory. The function putBits writes some bits into the bit-stream. Its first argument is the number of bits to write, and the second is an int value representing those bits. (Hence there is a maximum of 32 bits that can be written or read in one operation.) The function getBits similarly reads a number of bits from the bit-stream, returning an int which represents their value. Both operations update the ``current'' position in the stream.

    putBits    :: BinHandle -> Int -> Int -> IO (BinPtr a)
    getBits    :: BinHandle -> Int -> IO Int
    getBitsF   :: BinHandle -> Int -> BinPtr a -> (Int, BinPtr b)

The pure lazy function getBitsF is slightly different - because its result depends only on its arguments, you must tell it what position to start reading from. (It also returns the position immediately following the value read as part of its result.)

In order to get full control of the bit-stream, there are various other operations available, to move the current position, to report the current position, and so on.

    seekBin    :: BinHandle -> BinPtr a -> IO ()
    tellBin    :: BinHandle -> IO (BinPtr a)
    isEOFBin   :: BinHandle -> IO Bool

Higher-level typed binary functions

The Binary class is derivable for any datatype defined in a program except functions. (Please note however that cyclic or infinite values will cause the compressing function to diverge.)

The class member functions and their derivatives come in two varieties, one for sequential access, the other for random access. A BinHandle contains a hidden state, including the current position in the file or memory. Understanding the notion of the current position is important for using the sequential operations correctly. put and get always start reading or writing from the current position. All operations including the random-access ones, when they return, set the current position to the end of the value which has just been read or written.
put :: BinHandle -> a -> IO (BinPtr a)
put bh x writes a binary representation of the ordinary value x sequentially at the current position, returning a pointer to the beginning of the value. Where later sequential reading is sufficient, the return value of put can be discarded. When random-access is required, the return value of put can be used as the positional argument of getAt and putAt.
get :: BinHandle -> IO a
get bh reads a binary representation sequentially from the current position, returning the ordinary representation of the value.
putAt :: BinHandle -> a -> BinPtr a -> IO ()
putAt bh p x writes a binary representation of the ordinary value x at the position p, returning nothing. The pointer p might have been obtained as the result of an earlier put operation, or it may been read from a binary stream via a get operation, or indeed it may have been calculated.
getAt :: BinHandle -> BinPtr a -> IO a
getAt bh p reads a binary representation from the position p, returning the ordinary representation of the value.
getFAt :: BinHandle -> BinPtr a -> a
getFAt bh p is a pure, lazy, version of the getAt method, which can only be used on "frozen" BinHandles.

Transferring bits in bulk

The easiest way to transfer bits in bulk is with the copyBin operation. It takes an active BinHandle and copies its entire contents into the given BinLocation, returning a fresh BinHandle denoting the copy. As an alternative, copyBytes copies just a section of a bit-stream from the current position in one BinHandle to the current position in another - the copied section must be entirely byte-aligned. Finally, the least efficient but most flexible bulk transfer operation is copyBits, which allows any number of bits to be copied without alignment constraints - it even allows the source and destination bitstreams to overlap within the same BinHandle.

Defining your own compression

If you want to play with defining your own instances of Binary, have a look at some of the instances for standard types like Int and Lists in src/prelude/Binary/Instances.hs to see how things work.

The lower-level tools used in defining instances are:

    getBits  :: BinHandle -> Int -> BinPtr a -> IO Int
    putBits  :: BinHandle -> Int -> Int      -> IO (BinPtr a)
    getBitsF :: BinHandle -> Int -> BinPtr a -> (Int, BinPtr b)
    (<<) :: ((a->b),c) -> (c->(a,d)) -> (b,d)

Read and write modes

A file BinHandle can be opened in one of three modes: read-only (RO), write-only (WO), or read-write (RW). A memory BinHandle is always opened in RW mode, but may be changed to RO mode by the freezeBin operation. These modes differ from those of ordinary textual files:

A binary operation never raises an I/O exception.
When in RO mode, the operations put and putAt will not fail, but nor will they alter the file/memory.
When in WO mode, the operations get and getAt will return odd values, not corresponding to the real file/memory.
Encountering EOF in RO mode does not raise an error - reading beyond the end of the file/memory will simply return zeros. However, the operation isEOFBin can be used to test the condition.
The getFAt operation will give a runtime error if the file/memory is not in RO mode, but since this error does not arise within the I/O monad. it cannot be trapped.
In RW mode, interleaving read and write operations is safe.
The semantics of RW mode is that the file/memory is overwritten. In other words, you can write just a single bit in the middle of the file/memory if you want to - everything else will stay the same. In particular, unlike WO mode, a file is not truncated when you open it.

The latest updates to these pages are available on the WWW from

York Functional Programming Group