Re: ERB discussions and decisions
Michael Sperberg-McQueen <U35395@UICVM.UIC.EDU> wrote:
> The author's apologies to busy members of the WG who would prefer a
> shorter account of the decisions;
Not at all! Your account of the meeting is very much
appreciated -- thanks for taking the time to report it
at that level of detail.
[ ... ]
> * Discussed, once more, question C.10 (allow or prohibit
> non-deterministic content models). [...] Determinism
> is not particularly important to XML, but in Full SGML, the AND
> connector and the definition of start-tag omission interact with it
> and make it far more important. A minority asked what AND has to do
> with it,
AND groups and the ambiguity restriction interact in at least two ways:
1) The presence of AND groups makes checking content models
for ambiguity much more difficult; and
2) The ambiguity restriction makes it easier to parse instances
against content models with AND groups.
#2 requires some explanation. There are dozens of algorithms for
matching a sequence against a regular expression. The introduction of AND
groups makes many of these algorithms either intractable or unusable.
However, if a parser is allowed to assume that the content model is
deterministic, then some of them become tractable again. In particular,
many finite automaton-based techniques (e.g., Ken Thompson's familiar
construction) and the "parse-tree walking" algorithm (the one used by SGMLS
and, I think, SP) can be straightforwardly adapted to handle AND
groups (using the "set an auxilliary flag" method), but this only works
if the content model as a whole is deterministic.
This is not to say that the ambiguity restriction is _necessary_ to
make validation tractable: there are efficient algorithms for matching
against extended regular expressions, including nondeterministic ones
with the permutation operator (AND groups); these are not, however, as
far as I know, as widely known.
Nor do I mean to say that there are no automaton-based algorithms that
can handle both AND groups and nondeterminism efficiently, it's just that I
haven't been able to find any yet.
Anyway, the upshot of all this is: the ambiguity rule makes it harder
to validate DTDs. AND groups make it harder to validate instances.
The ambiguity rule makes it "easier" to validate instances, in the sense
that more algorithms are available to choose from.
> and suggested that all cases of start-tag omission allowed
> by the current rules would also be possible with nondeterministic
> rules, as experimentation with any LALR(1) parser generator should
I've done some research on this issue too. I don't have any
conclusive results, but I can say with certainty that:
1) The "obvious" approach [*] for handling start-tag omission
by LL- or LR- parsing techniques will not work. In particular:
a) using the "obvious" DTD->CFG construction, a LALR or LL parser
will predict start-tag omission in places that an SGML
parser will not, and an SGML parser will allow omission
in places that an LALR or LL parser will not.
b) the "obvious" DTD->CFG construction will not, in many
cases, yield an LL- or LR- grammar [**].
2) Converting content models to CFG productions induces a
worst-case exponential increase in size (the average case
for typical content models is linear, though).
3) LALR parsing is much more expensive than "conventional"
SGML parsing (i.e., "how SP does it"). In particular,
constructing LALR parsing tables requires time polynomial
in the size of the entire grammar.
[*] In case the "obvious" approach isn't so obvious:
<!ELEMENT X - - (...)>
<!ELEMENT Y - O (...)>
<!ELEMENT Z O O (...)>
becomes, given suitable transformations for ?-start-tag, ?-content-model
X ::= x-start-tag x-content-model x-end-tag
Y ::= y-start-tag y-content-model y-end
y-end ::= y-end-tag | nothing
Z ::= z-start z-content-model z-end
z-start ::= z-start-tag | nothing
z-end ::= z-end-tag | nothing
<!ELEMENT FOO - - (a?, b)>
<!ELEMENT (a|b) O O (X)>
<!ELEMENT X - O EMPTY>
The "obvious" equivalent CFG is ambiguous (let alone LL or LR).