On 9/25/06, Tomasz Zielonka
On Mon, Sep 25, 2006 at 05:09:13PM +0900, Robert Marlow wrote:
I think HostAddresses are probably good how they are now. It seems a bit too much work to convert a HostAddresses endianness everytime it's used.
Do you mean too much work for the programmer or for the computer?
If the latter, then I propose to make HostAddresses an abstract newtype, which doesn't allow to observe the endianness of the platform.
I'm surprised it isn't one already... I believe port numbers are.
Internally it could still keep the address in network order, so no conversions would be neccessary. However, the easiest way to implement eg. comparisons for such a datatype would be to flip endianness before comparing. So you will have to pay the cost of conversion for some operations. I would prefer if this cost was associated with network operations (connect, sentto, recvfrom, ...), because they already have a relatively big cost and you don't call them so often.
How often do you need to do comparisons on addresses for anything but equality (which doesn't require flipping to test)?
If the problem is how it appears (which along with PortAddress can be confusing) perhaps it just needs a different show implementation.
The problem is everything that gives different results on platforms with different byte-orders. Show is one example, but there are also arithmetic operations, Ord, etc.
... How often do you need to do arithmatic? Oh, btw, there are ntoh*/hton* functions for converting between host and network byte orders regardless of how crazy the host byte-order is...