Comments inline

On Fri, Apr 11, 2014 at 6:22 AM, Mark Frost <FROSTMAR@uk.ibm.com> wrote:
When we were implementing unordered sequences, this raised some questions around evaluating relative paths in expressions, for elements in a choice or unordered sequence :

DFDL spec: (gwdrp-dfdl-v1.0.4 section 15)
"When processing a choice group the parser validates
any contained path expressions. If a path
expression contained inside a choice branch refers
to any other branch of the choice, then it is a
schema definition error."
1.        I'm not clear what benefit this restriction on path expressions gives.
It seems redundant since in any single instance of a choice group, if the branch  being processed exists, then by definition none of it's sibling branches exist. Any expression path referring to a non-existent branch  would correctly return <empty sequence>

Typically in XPath, such paths would just be empty-sequence at runtime. Making it an SDE hoists the error to (hopefully) compile time, and making it SDE (non-recoverable) changes the way one must write expressions. You can't write utter nonsense paths and have them be runnable.
 
If the choice group is inside a repeating structure, then expressions referring to choice branches within other instances of the choice could be useful.
Should an expression referring to branches in other instances of a choice cause a schemadef error?

Should be no issue if you are looking at say, position() - n. If you reach to something that doesn't exist, then you'll get empty sequence.

My experience so far with XPath is that this notion that non-existance returns empty sequence is painful at best and a nightmare at worst. Expressions that are utterly nonsense are accepted executed, and silently fail by returning empty sequence.  The most common mistake is writing /a/b/c when you needed /ns1:a/ns2:b/ns3:c.
 

Example
expression on el_b could be { fn:count(../../el_choice/el_a) }

- parent
 [sequence]
   - el_choice [minOccurs=5 maxOccurs=5]
     [choice]
       - el_a
       - el_b
 

2.        Should an expression that potentially refers to branches in the choice cause a schemadef error?

Example
identically named elements in and out of a choice
expression on el_c could be { fn:count(../el_a) }

- parent
 [sequence]
   - el_a
   - el_b
   - [embedded choice group]
      - el_a
      - el_c


I'd love to restrict this, because we're looking at having to create a DFDL expression language implementation for performance reasons, and complex things like this require a very complex implementation tantamount to a query-engine.

I would claim that these two el_a elements are different, and we could choose to restrict a DFDL path expression to return only nodes described by the same schema component, with "same schema component" meaning same path from document element to the schema component where an element or group or type reference counts as part of that path. So two different element references to the same global element would be two different schema components.

But I suspect that this is too restrictive, and implementations are just going to have to be sophisticated enough to execute queries like this one, and a good implementation will optimize simpler cases for faster execution.
 
...mikeb