11 Mar
2012
11 Mar
'12
9:20 p.m.
Maybe we should also think a little bit more if 'Type' really is the
right interface to use. Both Guice and CDI defines 'TypeLiteral'
classes. Maybe we should do the same?
Best Regards,
/Andreas
2012-03-11 21:05, Andreas Jonsson skrev:
Hi,
We must consider how we define the equality of RoleHints. Different
implementations of the interface ParameterizedType do not need to
implement equals and hashValue consistently. For instance, this lookup
doesn't work at the moment:
Utils.getComponent(new ParameterizedType() {
@Override
public Type [] getActualTypeArguments() {
Type [] args = {
String.class
};
return args;
}
@Override
public Type getOwnerType() {
return null;
}
@Override
public Type getRawType() {
return DocumentReferenceResolver.class;
}
});
At the same time, there is the backwards compliancy issue. This might be
ok for RoleHint equality:
/**
* Defines the equality of types.
*
* @param t1 operand for comparison.
* @param t2 operand for comparison.
* @return {\code true} if the roles should be considered equal.
*/
private boolean typeEquals(Type t1, Type t2) {
if (t1 instanceof ParameterizedType && t2 instanceof
ParameterizedType) {
ParameterizedType pt1 = (ParameterizedType) t1;
ParameterizedType pt2 = (ParameterizedType) t2;
Type [] args1 = pt1.getActualTypeArguments();
Type [] args2 = pt2.getActualTypeArguments();
if (args1.length != args2.length) {
return false;
}
for (int i = 0; i < args1.length; i++) {
if (!typeEquals(args1[i], args2[i])) {
return false;
}
}
} else if ((t1 instanceof Class && !(t2 instanceof Class))
|| (t2 instanceof Class && !(t1 instanceof Class))) {
/*
* For compliancy with legacy code.
*
* Looking up components using raw classes only
* matches components registered as raw classes, and a
* component registered as raw class can only be
* looked up using a raw class.
*/
return false;
}
Class c1 = ReflectionUtils.getTypeClass(t1);
Class c2 = ReflectionUtils.getTypeClass(t2);
return c1.equals(c2);
}
@Override
@SuppressWarnings("unchecked")
public boolean equals(Object obj)
{
if (this == obj) {
return true;
}
if ((obj == null) || (obj.getClass() != this.getClass())) {
return false;
}
RoleHint<T> rolehint = (RoleHint<T>) obj;
return typeEquals(getRoleType(), rolehint.getRoleType())
&& (getHint() == rolehint.getHint() || (getHint() != null &&
getHint().equals(rolehint.getHint())));
}
@Override
public int hashCode()
{
int hash = 8;
hash = 31 * hash + (null == getRoleType() ? 0 :
ReflectionUtils.getTypeClass(getRoleType()).hashCode());
hash = 31 * hash + (null == getHint() ? 0 : getHint().hashCode());
return hash;
}
/Andreas
2012-03-06 14:51, Thomas Mortagne skrev:
It's committed.
Would be nice to have comments on the new APIs (suggest better naming,
etc...). It's a very important API and we really need to have
something we are sure of before 4.0 final.
On Mon, Mar 5, 2012 at 6:53 PM, Thomas Mortagne
<thomas.mortagne(a)xwiki.com> wrote:
Hi guys,
I'm almost done but there is one little detail left: for easier
retro-compatibility I would like to introduce a new @Role annotation
and deprecate the @ComponentRole annotation. The idea is that Role
will "officially" take into account generic parameters while
ComponentRole does not. That way no existing component role behavior
will be broken or require crappy retro-compatibility like registering
automatically both generic and non generic roles (what I started to do
at first).
On Thu, Mar 1, 2012 at 6:20 PM, Thomas Mortagne
<thomas.mortagne(a)xwiki.com> wrote:
Hi devs,
I recently modified component manager internals and changed the role
used as key in the map from Class to Type so that it properly make a
difference between Provider<Toto> and Provider<Titi> without the need
to generate a special hint.
For the one not aware of what Type is lets say it's what Class is
extending. Among other things Type is also extended by
ParameterizedType which is interesting here because it basically
contain a Class with it's generic parameters (it also indicate if the
type in included in another one but we don't care in our case ;)).
To summarize:
* Provider is a Class
* Provider<String> is ParameterizedType
and both are Type
That said I would like to change the whole ComponentManager and
related descriptors APIs to use Type as role instead of Class.
= What it bring
What it means is that it makes possible to cleanly represent a
Provider in which the generic type is meaningful right in the
ComponentDescriptor (right now the generic type is extract from the
implementation class when the provider is registered which is pretty
crappy) and simplify a lot the current implementation.
But it also mean we could use this new "feature" for other components.
For example right now we have things like:
@Component
public class DefaultStringDocumentReferenceResolver implements
DocumentReferenceResolver<String>
and
@Component
@Named("default/reference")
public class DefaultReferenceDocumentReferenceResolver implements
DocumentReferenceResolver<EntityReference>
and injected by
@Inject
DocumentReferenceResolver<String> resolver;
and
@Inject
@Named("default/reference")
DocumentReferenceResolver<EntityReference> resolver;
Having Type used as role means that we could have:
@Inject
DocumentReferenceResolver<String> resolver;
and
@Inject
DocumentReferenceResolver<EntityReference> resolver;
= What it breaks
So what's the catch ?
Basically it would break any code that currently programmatically
register or unregister Provider since It would remove the current hack
to discover the Provider parameter (assuming that the
ComponentDescriptor now provide the complete role). But that's OK IMO
given that Provider support is very recent and I doubt it's a common
use case.
We would obviously keep and deprecate all the existing APIs based on
Class which are pretty easy to implement since Class extends Type.
Even if removing them would not cause any issue when you build it's
not binary compatible since it's not the same method signature.
So WDYT ?
Here is my +1 and since we are starting a new major version it sounds
a good occasion so introduce such change.
--
Thomas Mortagne
--
Thomas Mortagne