Am Wed, 10 Oct 2012 15:02:40 +0200 schrieb Marco Leise <[email protected]>:
> Am Wed, 10 Oct 2012 12:55:59 +0200 > schrieb Piotr Szturmaj <[email protected]>: > > > And replace them with global reset function: > > > > void reset(T)(ref T digest) > > if (isDigest!T) > > { > > digest = T.init; > > } > > > > Current usage: > > > > SHA1 sha1; > > sha1.start(); > > ... calculate hash > > sha1.start(); // start again > > ... calculate hash > > There's makeDigest which calls start for you, so you can write code like this: auto sha1 = makeDigest!SHA1(); //calculate hash sha1.start(); //start again //calculate hash... But the digests still have to implement a start method, even if it's only a "this = typeof(this).init;". Maybe we could avoid the start method by introducing a fallback UFCS start method... > Unlike classes, structs don't mandate a ctor call, so in > general you need to disguise the constructor as 'start()' > 'Create()' or similar, where T.init is not sufficient. If on > the other hand, you use a ".isInitialized = false" field > instead of start(), then you'd have to unnecessarily check > that every time data is added to the digest, instead of doing > it once. (That might be personal taste though :) ) > > I think the constraints are: > * no start() method ever takes arguments > * no digest needs to allocate to start > * no digest depends on global/static variables > Yes, that's the main reason why we have 'start', but it also must work as a reset function. The difficult question is whether there actually is a hash algorithm (or whether there will be one in the future) which needs such a complex start method. No-allocation often means no C wrapper (although a C API could allow to use stack objects, but often they just provide opaque structs), so this is a little controversial. But if we allowed allocation we'd have to worry about deallocation as well and then things get tricky. Reference counting is probably still the easiest solution for those cases.
