Type-assignment in one pass

["K.Kawaguchi" <k-kawa@bigfoot.com>]

> I'm a little confused by this example. I would have thought that the
> validator had to look ahead anyway in this case.

No, without any lookahead, validator can know the document is valid.

> to see if you've got the last y, how can you pick the matching
> definition?

Let me show you. To simplify the pattern, I'll use pseudo-TREX pattern.


<element name="x" label="x">
  <zeroOrMore>
    <element name="y" label="y1">
      <attribute name="z" type="xsd:string" />
    </element>
  </zeroOrMore>
  <element name="y" label="y2">
    <attribute name="z" type="xsd:integer" />
  </element>
</element>

Consider validating the following fragment with this pattern:

<x> <y z="1" /><y z="2" /><y z="3" /> </x>


1. When the validator see startElement(<y z="1">), it can see the two
possibilities, as "y1" and "y2". So it remembers both.

"y1"
"y2"

2. When it sees endElement(</y>). Both of the choice work. So again keep
both of them.

"y1" -- "<endtag>"
   (no child)

"y2" -- "<endtag>"
   (no child)

3. When it sees startElement(<y z="2">), it knows that the second
interpretation doesn't work any more. but the first thread has two
choices. So remember the both choice.

"y1" -- "<endtag>" -+- "y1"
   (no child)       |
                    +- "y2"

"y2" -- "<endtag>" --- X
   (no child)

If you link these interpretations by backward references and you have
garbage collection, simply remember the last set of choice. Dead
interpretation will be automatically collected by GC. Of course
reference counting works, too.


4. ...

5. When it sees startElement(<y z="3">), again it abandons one
interpretation, and find two options.


"y1" -- "<endtag>" -+- "y1" -- "<endtag>" -+- "y1"
   (no child)       |                      |
                    |                      +- "y2"
                    |
                    +- "y2" -- "<endtag>" --- X

"y2" -- "<endtag>" --- X
   (no child)

6. ...

7. When it sees endElement(</x>), now it sees the first thread is dead.


"y1" -- "</>" -+- "y1" -- "</>" -+- "y1" -- </> -- X
    (nc)       |      (nc)       |      (nc)
               |                 +- "y2" -- </> -- </>
               |                        (nc)
               +- "y2" -- "</>" --- X
                      (nc)
"y2" -- "</>" --- X
    (no child)


So, after GC the only interpretation left is

"y1" -- "</>" -+- "y1" -- "</>" -+- "y2" -- </> -- </>

and hey, it's the interpretation that works.


All of this process has done without any lookaheads. In other words, if
you want depth-first search, you need lookaheads. If you choose
width-first search, you don't.


regards,
----------------------
K.Kawaguchi
E-Mail: k-kawa@bigfoot.com