[
Lists Home |
Date Index |
Thread Index
]
----- Original Message -----
From: "Tim Bray" <tbray@textuality.com>
> I think I've decided that the namespace handling is wrong. So I
> propose adding a new call
>
> int genxDeclareNamespace(genxWriter w, utf8Byte * uri, utf8Byte *
> prefix);
>
> if prefix == NULL then genx will generate one. prefix == "" is not
> allowed.
What about default namespaces?
Maybe NULL should mean "default namespace", and you add a second function:
int genxDeclareAutoNamespace(genxWriter w, utf8Byte * uri);
which auto-generates a prefix.
> Then the calls with namespaces lose the separate prefix argument, and
> genx fills in the declarations as appropriate. This feels more like
> the way that people actually think about writing and reading XML docs.
> Does anyone see a reason not to do this?
If you use just an uri argument you may have a problem of
ambiguity, as multiple prefixes can map to the same uri.
You should rather loose the uri argument, but this is a problem
for auto-generated prefixes, unless you return them to the caller.
Another option could be to return a handle for the declared mapping,
sort of like you are suggesting for elements and attributes below.
> NEW: Making it fast
>
> It dawned on me that with all the checking and so on, genx might not be
> as fast as it could possibly be. The way to fix this is obvious, but
> it might amount to premature optimization. The idea is that you
> predeclare your elements and attributes and get handles to them, so
> that they only need to be namechecked and sorted once. Something like.
>
> genxElement genxDeclareElement(genxWriter w, utf8Byte * namespaceURI,
> utf8Byte * type);
> genxAttribute genxDeclareAttribute( ... same args ... );
>
> Then you have
>
> int genxFastStartElement(genxWriter w, genxElement element);
> int genxFastAttribute(genxWriter w, genxAttribute attribute, utf8Byte *
> value);
>
> I think that with a little bit of care in the code, this should
> generate guaranteed-WF canonical XML at speeds close enough to the most
> deranged pedal-to-the-metal custom C code to vanish in the static of
> any conceivable application.
I like it because alternative APIs I can think of (e.g. for when you know
that the input has already been checked somewhere else) still leave more
room for programmer error, but here you can only obtain a handle as
the result of a checking process.
> Would this be premature optimization?
No, because it requires an API design change.
Some optimizations are inherent to an interface design and have
to be thought of at the beginning, as there is no way to put them
in later.
Karl
|
