SV: First pass at generated schema for DOM 1 + HTML

comments inlined

would you like a helping hand with the Schema? If so, how would you like us
to split the work?

-----Ursprungligt meddelande-----
Från: Arnold, Curt [mailto:Curt.Arnold@hyprotech.com]
Skickat: den 31 maj 2001 19:47
Till: 'www-dom-ts@w3.org'
Ämne: RE: First pass at generated schema for DOM 1 + HTML


I'm not sure that I understand your shorthand.  So point by point:

>     * Use IDL for attribute / method names

Do you think the recent schemas do not use IDL for attribute method names?

>     * Specify var's, parameters, and returnTypes according to 
> the spec:
>                 Spec always --> required in the schema
>                 Spec sometimes --> optional in the schema
>                 Spec never --> should not appear in the schema

returnType's do not need to be in the tests since they are fixed by the DOM
spec.
I believe that anything that is fixed by the DOM spec should not be
replicated 
in the instance document or implied to be in the instance
document by fixed or defaulted attributes in the DTD.  Any code generation
can get the needed information either from the DOM spec or from
annotations in the schema.

[dd] I don't see your point. If the instance document needs to contain
information on a returned type to generate a particular type of code that in
turns points to expected type of return in order to make a comparison,
wouldn't it be preferrable _not_ to roundtrip to the schema or DOM spec or
have the intelligence in the generating XSLT? If we do roundtripping, we
presuppose that people will validate while writing tests, which we cannot
demand. Therefore it seems to me we cannot leave out any relevant
information from the instance document.

I would require a var attribute for any read-only accessor or any non-void
function.
I think it is highly unlikely that you would want to invoke a function or
access a property where it would not be beneficial to make an assertion
on the return value or use it in a later call.  If there is a method or
property
where the return value is meaningless, then the overhead of declaring a
variable to hold it is trivial.  Making it optional when it is used 99% of 
the type is more likely to cause errors.  Read-write properties have both
var and value optional, however omitting both out is a no operation.

I've had a change of heart on parameters.  In my manual schema, 
parameters that could be null were optional.  However that information is
not
in the xml source for the DOM spec and I don't think we want to introduce
any supplimentary information.  So if the parameter is required, how do
you specify that it is null.  One option would be to make allow "null" as
a special value in the argument.  Unfortunately, that could seriously 
complicate the code generation for C++.  It is a little more awkward
in the test, but it could greatly simplify the C++ code generation, if
null parameters are passed by passing in declared but uninitialized
variables, such as:

[dd] I'm not a C++ specialist, how does it make the generation of C++ code
more difficult compared to, say, Java?

<declare var="refNode" type="Node"/>
<declare var="newNode" type="Node"/>
...
<createElement var="newNew" obj="doc" tagName="address"/>
<insertBefore obj="employee" refNode="refNode" insertedNode="newNode"/>

I would explicitly declare all variables.  For languages where their
is no common ancestor class, then it would be critical and potentially
complicated to infer the appropriate type.

[dd] OK, would the stylesheet that crunches the instance document need to be
particulalry intelligent to write out the right thing?

It is safer to code the tests with explicit declaration now and
find out later that it was unnecessary than to depend on the 
ability to infer types from context and find out later that it
was difficult to do.

[dd] definitely

>     * Interface name -- should be inferable from somewhere, or defined
>        as an attribute on the method name
>                 - defaulted as an attribute in cases where it 
> is unambiguous
>                 - one of a list if it is available in more 
> than one place.

The interface attribute has to be there when the same method or
property is introduced in multiple interfaces 
(ProcessingInstruction.data and CharacterData.data for example).

If not required, then we need to choose whether the interface 
attribute is:

a) not present in the DTD or schema
b) an optional attribute whose only acceptible value is the 
interface that defined the method.
c) an optional attribute whose acceptible values are the
defining interface and any derived interfaces.  Specifying
a derived interface adds an implicit instanceOf assertion.

I think that using an explcit instanceOf assertion is
more appropriate than implying one by specifying a derived
type.  That is, I would prefer:

<assertInstanceOf obj="node" type="Comment" id="nodeIsComment"/>
<nodeValue var="val" obj="node"/>

to

<nodeValue var="val" obj="node" interface="Comment"/>

Option b would allow tests to anticipate an method or
attribute being multiple defined.


>     * Exceptions -- correspond according to the spec -- ie, 
> only be able
>        to specify the particular exceptions that can be 
> thrown on a given
>        attribute or method.
> 
>       any others ...

In my recent schemas, <assertDOMException code="INDEX_SIZE_ERR"> 
can be used to assert that the enclosed statement should throw
a DOM Exception with code=INDEX_SIZE_ERR.  Only statements that
have been declared to throw DOMException can appear within
<assertDOMException> however there is nothing that constraints
the code to be an appropriate value for the method.

I had originally had an expectException="" attribute on the
statement which could constrain the code values to the declared 
list.  However that had the disadvantages of either requiring
all invocations to have ID so that the exception assertions
could be identified or risking that exception assertions 
would not have an ID.  Also, it would require mixing the
production of the try {}/catch {}/fail production in a lot of 
places.

If it is compelling to make sure that the codes are in synch,
then I would introduce specific assertions for each code, i.e.

<assertINDEX_SIZE_ERR id="AssertionIDgoesHere">
	<substring obj="str" start="-1" end="5"/>
</assertINDEX_SIZE_ERR>

but that seems overkill to me.

Received on Thursday, 31 May 2001 14:26:52 UTC