Remaining 037 review issues for today WG call (20 Jan)
I have answered most of the issues and comments raised by Steve and Mike but some need further discussion. Issues from Steve H General. Although dfdl:encoding enums are case insensitive, we should stick to UC throughout in examples. 2. I agree with the existing comment that the RFC2119 key words should be upper case. 14.3.4. There are type/rep combinations where lengthKind="implicit" is not allowed - so saying that 'pattern' is replaced by 'implicit' on unparsing does not work. TBD 16.2. I'm not sure that scannability in this constant encoding sense is necessary for patterns. I can create a regular expression that extracts all characters up to hex value xXX or all characters up to xYY, thereby treating the content as an encoding in-sensitive black box. Issues from Mike B · Tracker issue: codepoints outside BMP, as literals and in data. · If I put in a value that requires use of a high/low surrogate pair, is that an error, does it require me to put in two separate %#...; thingys, one for each of the surrogates (in which case these are not really code points in ISO10646). If I put in a codepoint for one of the supplemental characters and the schema itself is written in UTF-16 then that has to translate into literal surrogate pair. Ok, but I?m very uncertain about all this stuffTracker Issue: illegal character encodings for parsing and unparsing. TBD: how do these make it into the infoset or are they replaced, and if so how TBD: can one represent these in the infoset for output? Ideally not, but? · Tracker Issue: Processing-time Schema Definition Errors This section (2.3.1 in this draft), is problematic as we?re trying to allow simple DFDL implementations to not do a bunch of static checking, yet if implementations differ on when Schema Definition errors are detected, then the second paragraph says they are converted to processing errors. This lets different implementations do very different things in terms of how the speculative parsing back-tracks around. Grammar ambiguity is a very tricky case. Unless a DFDL implementation can prove a grammar to be unambiguous, then it is very hard to say that any particular combinatino of delimiters make up a legal DFDL schema definition. If the parser simply fails because the grammar was ambiguous, there?s no way to tell the difference between this and just broken data without proving the grammar is unambiguous. In general it is formally undecidable whether a grammar is ambiguous or unambiguous. ( http://books.google.com/books?id=lIuu53IcKWoC&pg=PT217&lpg=PT217&dq=proving+a+grammar+is+unambiguous&source=bl&ots=wie8TAt-MT&sig=ZSD7tIwnXZIT8Ic91BWMH2H2dKg&hl=en&ei=hAQ5S5vPOIri7APc37CKBg&sa=X&oi=book_result&ct=result&resnum=10&ved=0CDAQ6AEwCQ#v=onepage&q=proving%20a%20grammar%20is%20unambiguous&f=false ) Since DFDL v1.0 doesn?t allow recursive declarations/definitions, it may be possible to provide the ambiguity or unambiguity of a DFDL schema (or rather, the data syntax grammar described by it ? if you want to bother to distinguish the two), but recursion isn?t something we want to rule out for the future, so Type checking is decidable in DFDL?s expression language, so we could always detect type safety before run time; however, if we allow a simplistic DFDL implementation to just check types at run time then this would, by the definition in this section (2.3.1), issue processing errors when it detects these at run time, thereby allowing backtracking of the speculative parser to be driven off of type-checks in the expression language. It seems to me that we need to find a way to put this problem back into the hands of the user, and say that a schema where this actually matters (one where a type error causes a backtrack, which ultimately causes a successful parse) are illegal but implementations are allowed to not detect this particular illegality. It seems to me we need to put this problem back into the hands of the user. · Tracker Issue: "round trip" for infoset. Should we omit the whole point? · Tracker Issue: [schema] is an absolute or relative SCD. Why bother allowing absolute? · Tracker Issue: Glossary as the place for centralized definitions, or should they be repeated there, but also introduced at point of first use, or should we put the definitions only at the places where they are discussed, and xref from the glossary? · TBD: Issue - semantics of expressions containing relative paths that are inherited via ref to a dfdl:defineFormat. (also section 10.3) · TBD: Issue - XPath term - we are not consistent about using the term XPath, or "expression" when referring to our expression language. I prefer to call it our expression language, and then in the section that defines it state that it is a strict subset of XPath 2.0. · TBD: Issue - fn:position is unclear given that we've just said we don't support sequences in the expression language. · TBD: Issue - order of sections. Scoping rules section should come before variables section, which uses these concepts. TBD: Issue: Case sensitivity of enum names - did we say whether this is case sensitive or not? I believe it should be case sensitive. · Issue: dfdl:representation - Strings in binary rep. I see no reason why elements of type xs:string will examine dfdl:representation. They shouldn?t' care what it is, they are always "text". I should be able to specify a bunch of inter-mixed binary number and string elements without having to specify dfdl:representation="text' just to avoid an error on the string type elements. I believe xs:string type ignores dfdl:representation (always behaves as if dfdl:representation is 'text').(If we change this then the property precedence section for simpletypes changes slightly as representation="text" is implied if type is string.) That will make it impossible to introduce a binary representation of text later textStringPadCharacter textNumberPadCharacter - did we agree that this character must be a "minimum width" character if the char set encoding is variable width? (i.e., the pad char must be 1 byte if the encoding is UTF-8. numberInfinityRep numberNanRep - Is this applicable only to xs:double and xs:float? Also, what I've seen requires a distinction of sign. I.e., there are positive and negative infinities often printing as -inf and +inf. · TBD: Issue - \n in regular expressions - clarify relationship of this to entities like NL entity. Also, if I include an entity like WSP* in a regular expression (can I?) does it then match accordingly? It appears that some of our multi-valued entities like WSP+ create conditional "matching" behavior without having to use regular expressions, e.g., when WSP+ is used as a separator. But can you use entities like WSP+ in a regular expression? It seems you should be able to use regular "single valued" entities in a regular expression, its these multi-valued ones that have tricky semantics. Added Unicode values to /n, /t,/r. Disallow DFDL entities in regular expressions. 14.1 Alignment - TBD: Issue - zero-based thinking here. But all the bits stuff and everything else in DFDL uses 1-based reasoning. Need to revisit to make this sensible for 1 based world. Added implicit alignment table. TBD zero-based finalTerminatorCanBeMissing - spec is not clear. Also is there a finalSeparatorCanBeMissing Chaned to finalDocumentTerminatorCanBeMissing and finalDocumentSeparatorCanBeMissing. Not sure where finalDocumentSeparatorCanBeMissing should be specified. Looks odd on 'distinguished root'. These properties operate differently from other properties as they are defined on the 'distinguished root' but affect some lower down element. Effectively they are put in scope by a different mechanism Alan Powell MP 211, IBM UK Labs, Hursley, Winchester, SO21 2JN, England Notes Id: Alan Powell/UK/IBM email: alan_powell@uk.ibm.com Tel: +44 (0)1962 815073 Fax: +44 (0)1962 816898 Unless stated otherwise above: IBM United Kingdom Limited - Registered in England and Wales with number 741598. Registered office: PO Box 41, North Harbour, Portsmouth, Hampshire PO6 3AU
A few comments in-line below On Wed, Jan 20, 2010 at 7:01 AM, Alan Powell <alan_powell@uk.ibm.com> wrote:
I have answered most of the issues and comments raised by Steve and Mike but some need further discussion.
*Issues from Steve H*
*General*. Although dfdl:encoding enums are case insensitive, we should stick to UC throughout in examples.
*2*. I agree with the existing comment that the RFC2119 key words should be upper case.
*14.3.4.* There are type/rep combinations where lengthKind="implicit" is not allowed - so saying that 'pattern' is replaced by 'implicit' on unparsing does not work. TBD
We covered this on the most recent wg call.
*16.2*. I'm not sure that scannability in this constant encoding sense is necessary for patterns. I can create a regular expression that extracts all characters up to hex value xXX or all characters up to xYY, thereby treating the content as an encoding in-sensitive black box.
If your byte pattern happens to be a legal part of a multi-byte character sequence, then you'll get a false recognition, or you won't get what you expect. Example: You are searching for byte 0xAA, but that can legally appear as byte 3 of a 3-byte UTF-8 encoded character. When you say you are looking for hex AA in a string, DFDL is currently defined to mean you are looking for the character reprsented by that raw byte. If the encoding is UTF-8, that isn't a legal character encoding sequence even, so the decoder should cause an error or something. Even for a fixed length single byte character set, you have to have no unused code that have no mapping to ISO 10646, because our infoset is defined in terms of translations into that. I think we need encoding="none" or encoding="bytes" or something if you really want to scan bytes without encoding causing problems.
*Issues from Mike B*
· Tracker issue: codepoints outside BMP, as literals and in data.
· If I put in a value that requires use of a high/low surrogate pair, is that an error, does it require me to put in two separate %#...; thingys, one for each of the surrogates (in which case these are not really code points in ISO10646). If I put in a codepoint for one of the supplemental characters and the schema itself is written in UTF-16 then that has to translate into literal surrogate pair. Ok, but I’m very uncertain about all this stuff
The above item had two issues glomed together. There really are two separate issues. The above is about these crazy codepoints that use surrogate pairs. That's a minor corner case given the amount of use those get. The bigger issue is the one below, which is about things that either are in strings and are broken character encodings, but we still need to be able to process the data. There's also the matter of recovery from errors in decoding, and what we put out when the infoset contains a character code where there is no valid encoding, or just a character code which isn't even in ISO 10646 (e.g., character code 0xFFFFFFFF, which is not a valid character at all.
Tracker Issue: illegal character encodings for parsing and unparsing. TBD: how do these make it into the infoset or are they replaced, and if so how TBD: can one represent these in the infoset for output? Ideally not, but…
· Tracker Issue: Processing-time Schema Definition Errors
This section (2.3.1 in this draft), is problematic as we’re trying to allow simple DFDL implementations to not do a bunch of static checking, yet if implementations differ on when Schema Definition errors are detected, then the second paragraph says they are converted to processing errors. This lets different implementations do very different things in terms of how the speculative parsing back-tracks around.
Grammar ambiguity is a very tricky case. Unless a DFDL implementation can prove a grammar to be unambiguous, then it is very hard to say that any particular combinatino of delimiters make up a legal DFDL schema definition. If the parser simply fails because the grammar was ambiguous, there’s no way to tell the difference between this and just broken data without proving the grammar is unambiguous. In general it is formally undecidable whether a grammar is ambiguous or unambiguous. (* http://books.google.com/books?id=lIuu53IcKWoC&pg=PT217&lpg=PT217&dq=proving+a+grammar+is+unambiguous&source=bl&ots=wie8TAt-MT&sig=ZSD7tIwnXZIT8Ic91BWMH2H2dKg&hl=en&ei=hAQ5S5vPOIri7APc37CKBg&sa=X&oi=book_result&ct=result&resnum=10&ved=0CDAQ6AEwCQ#v=onepage&q=proving%20a%20grammar%20is%20unambiguous&f=false *<http://books.google.com/books?id=lIuu53IcKWoC&pg=PT217&lpg=PT217&dq=proving+a+grammar+is+unambiguous&source=bl&ots=wie8TAt-MT&sig=ZSD7tIwnXZIT8Ic91BWMH2H2dKg&hl=en&ei=hAQ5S5vPOIri7APc37CKBg&sa=X&oi=book_result&ct=result&resnum=10&ved=0CDAQ6AEwCQ%5Clv=onepage&q=proving%20a%20grammar%20is%20unambiguous&f=false%5Ct>)
Since DFDL v1.0 doesn’t allow recursive declarations/definitions, it may be possible to provide the ambiguity or unambiguity of a DFDL schema (or rather, the data syntax grammar described by it – if you want to bother to distinguish the two), but recursion isn’t something we want to rule out for the future, so
Type checking is decidable in DFDL’s expression language, so we could always detect type safety before run time; however, if we allow a simplistic DFDL implementation to just check types at run time then this would, by the definition in this section (2.3.1), issue processing errors when it detects these at run time, thereby allowing backtracking of the speculative parser to be driven off of type-checks in the expression language. It seems to me that we need to find a way to put this problem back into the hands of the user, and say that a schema where this actually matters (one where a type error causes a backtrack, which ultimately causes a successful parse) are illegal but implementations are allowed to not detect this particular illegality.
It seems to me we need to put this problem back into the hands of the user.
· Tracker Issue: "round trip" for infoset. Should we omit the whole point?
· Tracker Issue: [schema] is an absolute or relative SCD. Why bother allowing absolute?
· Tracker Issue: Glossary as the place for centralized definitions, or should they be repeated there, but also introduced at point of first use, or should we put the definitions only at the places where they are discussed, and xref from the glossary?
· TBD: Issue - semantics of expressions containing relative paths that are inherited via ref to a dfdl:defineFormat. (also section 10.3)
· TBD: Issue - XPath term - we are not consistent about using the term XPath, or "expression" when referring to our expression language. I prefer to call it our expression language, and then in the section that defines it state that it is a strict subset of XPath 2.0.
· TBD: Issue - fn:position is unclear given that we've just said we don't support sequences in the expression language.
· TBD: Issue - order of sections. Scoping rules section should come before variables section, which uses these concepts.
- TBD: Issue: Case sensitivity of enum names - did we say whether this is case sensitive or not? I believe it should be case sensitive.
· Issue: dfdl:representation - Strings in binary rep. I see no reason why elements of type xs:string will examine dfdl:representation. They shouldn’t' care what it is, they are always "text". I should be able to specify a bunch of inter-mixed binary number and string elements without having to specify dfdl:representation="text' just to avoid an error on the string type elements. I believe xs:string type ignores dfdl:representation (always behaves as if dfdl:representation is 'text').(If we change this then the property precedence section for simpletypes changes slightly as representation="text" is implied if type is string.) That will make it impossible to introduce a binary representation of text later
What is "a binary representation of text"? Is there a real issue here. This is a primary convenience and clarity issue for me. I do not want to have to change to representation="text" for every string inside a cobol structure, which is ultimately a binary representation object. To me type="string" is enough. I want to put in the file scope level of the schema a representation="binary", and then decorate the elements with the specifics of their types, but I do not expect to have to put representation="text" on anything. I do not understand what you are trying to achieve by requiring representation="text" for things that are already textual based on the type. The rest of the issues below I think we need to discuss on calls.
- textStringPadCharacter textNumberPadCharacter - did we agree that this character must be a "minimum width" character if the char set encoding is variable width? (i.e., the pad char must be 1 byte if the encoding is UTF-8.
- numberInfinityRep numberNanRep - Is this applicable only to xs:double and xs:float? Also, what I've seen requires a distinction of sign. I.e., there are positive and negative infinities often printing as -inf and +inf.
· TBD: Issue - \n in regular expressions - clarify relationship of this to entities like NL entity. Also, if I include an entity like WSP* in a regular expression (can I?) does it then match accordingly?
It appears that some of our multi-valued entities like WSP+ create conditional "matching" behavior without having to use regular expressions, e.g., when WSP+ is used as a separator. But can you use entities like WSP+ in a regular expression? It seems you should be able to use regular "single valued" entities in a regular expression, its these multi-valued ones that have tricky semantics. Added Unicode values to /n, /t,/r. Disallow DFDL entities in regular expressions. - 14.1 Alignment - TBD: Issue - zero-based thinking here. But all the bits stuff and everything else in DFDL uses 1-based reasoning. Need to revisit to make this sensible for 1 based world. Added implicit alignment table. TBD zero-based
- finalTerminatorCanBeMissing - spec is not clear. Also is there a finalSeparatorCanBeMissing Chaned to finalDocumentTerminatorCanBeMissing and finalDocumentSeparatorCanBeMissing. Not sure where finalDocumentSeparatorCanBeMissing should be specified. Looks odd on 'distinguished root'. These properties operate differently from other properties as they are defined on the 'distinguished root' but affect some lower down element. Effectively they are put in scope by a different mechanism
Alan Powell
MP 211, IBM UK Labs, Hursley, Winchester, SO21 2JN, England Notes Id: Alan Powell/UK/IBM email: alan_powell@uk.ibm.com Tel: +44 (0)1962 815073 Fax: +44 (0)1962 816898
------------------------------
* *
*Unless stated otherwise above: IBM United Kingdom Limited - Registered in England and Wales with number 741598. Registered office: PO Box 41, North Harbour, Portsmouth, Hampshire PO6 3AU *
-- dfdl-wg mailing list dfdl-wg@ogf.org http://www.ogf.org/mailman/listinfo/dfdl-wg
participants (2)
-
Alan Powell -
Mike Beckerle