Hi Mike
I agree that forcing someone to put
dfdl:alignment on the element when they really want it on the simple type
is not good - but the spec does not imply that today. Unless I am
missing something I can write:
<simpleType name="string" dfdl:lengthKind="prefixed"
dfdl:prefixLengthType="stdStrPrefixType" dfdl:alignment="2"
dfdl:alignmentUnits="bytes" >
<restriction base="xs:string" >
<maxLength value='65535'/>
</restriction>
</simpleType>
<simpleType name="stdStrPrefixType" dfdl:representation="binary">
<restriction base="xs:unsignedShort"/>
</simpleType>
The properties of simple type "string"
and the element are combined. This alignment is applied before
the prefix length is consumed, as per the grammar. This seems fine
to me. Taking a (hypothetical) example, if I was modelling a PL/1 var char
then I know that all var chars are aligned on half-word boundaries, so
I put the alignment on the type for the var char.
Have I misunderstood the problem?
Regards
Steve Hanson
Architect, Data Format Description Language (DFDL)
Co-Chair, OGF
DFDL Working Group
IBM SWG, Hursley, UK
smh@uk.ibm.com
tel:+44-1962-815848
| From:
| Mike Beckerle <mbeckerle.dfdl@gmail.com>
|
| To:
| Steve Hanson/UK/IBM@IBMGB
|
| Cc:
| dfdl-wg@ogf.org, Tim Kimber/UK/IBM@IBMGB
|
| Date:
| 26/10/2011 16:02
|
| Subject:
| Re: [DFDL-WG] lengthKind='prefixed'
clarification needed |
Oops, a nit: I missed dfdl:outputValueCalc which should
also be in the list of things that cause a schema def error.
On Wed, Oct 26, 2011 at 10:55 AM, Mike Beckerle <mbeckerle.dfdl@gmail.com>
wrote:
Ok, I have a smaller simpler proposal to fix this which
is really a very small spec change The total wording of this would
be these additional errata:
Errata N Section 12.3.4 dfdl:lengthKind 'prefixed'
Replace phrase: "It is a schema definition error if the xs:simpleType
specifies dfdl:lengthKind 'delimited' or 'endOfParent' or a dfdl:outputValueCalc"
with "It is a schema definition error if the xs:simpleType specifies
dfdl:lengthKind 'delimited', dfdl:lengthKind 'endOfParent', dfdl:initiator,
dfdl:terminator, dfdl:leadingSkip, or dfdl:trailingSkip.
Errata M Section 9.1 DFDL Data Syntax Grammar
Replace "PrefixLength = SimpleContent" with "PrefixLength
= LeadingAlignment SimpleContent"
Today's errata on nested prefix length types would remain as errata.
The first new errata is important clarification, and the schema def error
is the conservative thing to do for the future. We have followed this principle
many other places in the spec.
The second new errata fixes a below-described issue with alignment and
information hiding/composition properties. It is, in some sense a very
small change to the spec to fix a rather glaring composition problem. Here's
the rationale:
To me lengthKind='prefixed' ought to handle the variable-length string
case like this:
In the schema for the specific application's data:
...
<element name="lname" type="fmt:string"/>
// note use of string, but from different namespace
<element name="addr1" type="fmt:string"/>
<element name="addr2" type="fmt:string"/>
<element name="postal" type="fmt:string"/>
...
Then, over in a different schema file.... which defines the namespace bound
to fmt above as its target namespace....
<simpleType name="string" dfdl:lengthKind="prefixed"
dfdl:prefixLengthType="stdStrPrefixType">
<restriction base="xs:string" >
<maxLength value='65535'/>
</restriction>
</simpleType>
<simpleType name="stdStrPrefixType" dfdl:alignment="2"
dfdl:alignmentUnits="bytes" dfdl:representation="binary">
<restriction base="xs:unsignedShort"/>
</simpleType>
The highlight here is that the alignment restriction is most naturally
expressed on the simpleType, the details of which are not near the string
itself's element declaration. The spec specifically allows this placement
on the simpleType for the dfdl annotations right now. It's only the grammar
that implies it would be ignored, as would initiator, terminator, leading
and trailing skip all be ignored.
To me the above pattern/idiom, where the format annotation cruft is isolated
on type definitions, is generally to be encouraged. THere's a composition
property I'm trying to preserve here, which is that you can put two things
side by side without having to worry about whether you are meeting one
or the other's alignment requirements. The alignment goes in the package
with the definition of the thing.
I don't think you should have to write:
...
<element name="lname" type="fmt:string" dfdl:alignment='2'
/>
<element name="addr1" type="fmt:string" dfdl:alignment='2'/>
<element name="addr2" type="fmt:string" dfdl:alignment='2'/>
<element name="postal" type="fmt:string" dfdl:alignment='2'/>
...
To me that violates a valuable principle of information hiding. I can hide
the type, but not hide its alignment requirements? Anywhere you can abstract
and hide a SimpleContent item that can be binary numeric where alignment
is a common requirement, you need also to be able to hide its alignment
requirements along side it so that using it won't be in the error prone
situation where the user can misalign it. PrefixLength is the one place
(That i know of) where we're violating this principle. It's an easy omission,
and an easy fix.
...mikeb
On Wed, Oct 26, 2011 at 6:15 AM, Steve Hanson <smh@uk.ibm.com>
wrote:
Mike
Length kind 'prefixed' was intended to handle the case where the length
is tightly bound to the data, ie, there is nothing between the length and
the data. For example a PL/1 var char or ASN.1 BER. If the length
causes the length/data to be aligned then that has to be taken into account
on the element itself. Length kind 'prefixed' was not intended to
cover more complex cases where the length itself has independent alignment
or there are delimiters involved. For those you use length kind 'explicit'
and an expression. Otherwise the combinations become too complicated. If
we wish to extend 'prefixed' to include the more complex cases, I think
that is a post 1.0 thought and is best handled using a different length
kind enum.
You say that ignoring the alignment property on the simple type used for
the length is strange, but if you allow that there is no way to align the
element's actual data separately. I think that it is even stranger.
The ASN.1 BER description at http://en.wikipedia.org/wiki/Basic_Encoding_Rules
describes how the length itself can have a prefix (see sub-section 'Length').
Regards
Steve Hanson
Architect, Data Format Description Language (DFDL)
Co-Chair, OGF
DFDL Working Group
IBM SWG, Hursley, UK
smh@uk.ibm.com
tel:+44-1962-815848
Definintely need an agenda slot to discuss this matter.
I think we should redefine PrefixLength to allow it to have framing.
There is a significant issue which is that some prefixLengthTypes will
be multi-byte binary integers (typically 2 or 4 bytes), and these commonly
require alignment to a 2 or 4 byte boundary, as that's how the data structures
they live in would have been laid out.
The spec currently doesn't allow prefixLengthTypes to be aligned themselves,
because the grammar has them as SimpleContent, without the surrounding
ElementLeftFraming and RightFraming. This is also why they cannot have
lengthKind='delimited'. Because there are no initiator nor terminator regions
surrounding them. So the only way they can be aligned is if the elements
that have these prefixLengths are themselves aligned properly.
However, if you specify alignment on a simple integer type, use it as a
prefixLengthType, and then that alignment annotation is *ignored* that
would seem strange, and buggy/hard-to-diagnose.
However, scoping rules for properties don't provide any way for this alignment
to get "into the scope chain", and I'd hate to start messing
with the scoping rules because of the corner case of prefixLength. We'd
need to put another scoping rule in just to handle this. I'd rather not
go there. Lots of our examples in the spec would have to change as they
use alignment as the example property...
But, the spec is not self-consistent, as the dfdl:alignment property can
be placed on a simple type definition, as well as on an element. So it
would seem a prefixLengthType could reference an aligned simple integer
type, but neither the grammar nor the scoping rules allow for using this
alignment property to control anything. Similarly, you can put an
initiator on a simple integer type, use it as a prefixLengthType, and have
the initiator be ignored.... because there is no initiator region for a
PrefixLength.
We need to fix this inconsistency.
I think prefixLengthType needs to be alignable, and one should be able
to specify alignment on a type definition, not just on an element.
I also think we're better off with a uniform general fix here, than a handful
of special case rules around prefix lengths. (E.g., the prefixLengthType
cannot have alignment, cannot have initiator/terminator or lengthKind delimited
warning or SDE if it does, etc. etc.)
So I think the grammar is wrong. I think
PrefixLength = SimpleElement
(where SimpleElement = ElementLeftFraming SimpleContent RightFraming )
is the right definition.
In working through examples, I'm convinced the current spec is problematic.
In the current spec one must model a 4-byte aligned binary integer prefix
length as a separate element (so that you can align it), and use lengthKind='explicit'
on the thing it controls. This is a lot of hassle for a very common situation.
The whole point of dfdl:lengthKind='prefixed' is to provide an easier way
to model the common cases.
For the same reason there is no alignment, the definition of dfdl:prefixLengthType
says the named type cannot have lengthKind='delimited'. That is because
the DFDL grammar defines the prefixLength region to be SimpleContent which
is without any of the surrounding framing regions where delimiters are
found.
So, one cannot for example, put an initiator and terminator on the prefix
length type so as to have syntax separating it from the actual content.
Even if it is fixed length you can't do it - Like you cannot model this
data as 3 string elements using prefix length:
(11)9 Ocean Way(20)Southwest(SW) Harbor(02)ME
(Notice in the above the unescaped "(SW)", which is why this
is not a delimited format.)
You also cannot do:
11(9 Ocean Way)20(Southwest(SW) Harbor)02(ME)
because that puts the initiator of the string element itself after its
prefix length region, which is backwards from the way we have it in the
grammar currently. Both of the examples above require use of a separate
element and lengthKind="explicit" to pull off, even though they
seem like fairly natural ways to textualize a binary format.
Now consider
xx9 Ocean WayxxSouthwest(SW) HarborxxME
where the "xx" is a 16 bit (2 byte) binary integer holding the
lengths 11, 20, and 2 respectively.
Except....That is, so long as the "xx" doesn't need to be on
a 2-byte alignment, because in my example the first element occupies 13
bytes including the prefix itself, so the next "xx" starts on
an odd boundary. I could specify alignment on each of the 3 elements
of my sequence here, which is unmotivated/weird since they're string elements
and their type may be distant from where the elements are declared, so
the motivation for the alighment may not be clear....... the alignment
constraint really wants to be expressed on the prefixLengthType, and the
dfdl annotation syntax lets you specify alignment there, ... it just doesn't
use it.
If we just redefine PrefixLength as SimpleElement, now all the example
formats above are easily modeled in the obvious way, and even the combinations
of text and binary lengths can be done naturally, as a binary prefixLengthType
integer type can have all the usual constraints binary data likes to have,
like alignment.
Even the 2-level ASN.1 wierd case "prefix-length of the prefix-length"
(see errata 2.13) works because ElementLeftFraming itself includes PrefixLength.
I believe we should put an explicit depth limit of 2 on this however.
(Side note: I'd like to see an example of the ASN.1 format that supposedly
requires this nested prefix of the prefix situation.)
Changing the grammar in this way lets us drop the special case handling
around prefixLength where it can't have lengthKind="delimited"
and ignores initiators and terminators and alignment which is a bunch less
special cases to have to implement and test, and create special warnings
for (e.g., "Warning: prefixLengthType 'lenType' has alignment property
which will be ignored.")
If we want to be more minimal about the changes, just changing
PrefixLength = ElementLeftFraming SimpleContent RightFraming
is sufficient and achieves the fix of the real problem.
(This also eliminates the need for current errata 2.13 and 2.14, or rather
replaces those errata with this new stuff.)
...mikeb
--
dfdl-wg mailing list
dfdl-wg@ogf.org
http://www.ogf.org/mailman/listinfo/dfdl-wg
Hi Mike,
I have always assumed that it works like this:
The Prefix region includes leading alignment, leading skip and initiator
The Content region contains the data, and the lengthKind property describes
how to determine the content length
The Suffix region includes Terminator and trailing alignment.
The lengthKind property describes the content region, and is not examined
until the Content region is reached. So the element's iniitator, if defined,
is not included in the length described by the prefix length.
If you view the prefixed length as describing the length of the *element*
(i..e its entire representation ) then this definition is not intuitive.
But I have always viewed lengthKind='prefixed' as being like the other
lengthKinds - it describes the length of the element's *content*.
So it's a consistent definition, but is it useful? I think so. In my experience,
prefixed lengths tend to be applied to complex elements ( structures )
rather than simple values. In such cases, the content of the complex element
will always be either a sequence group or a choice group, and any initiator/terminator
can be located on that group..
regards,
Tim Kimber, Common Transformation Team,
Hursley, UK
Internet: kimbert@uk.ibm.com
Tel. 01962-816742
Internal tel. 246742
From: Mike
Beckerle <mbeckerle.dfdl@gmail.com>
To: dfdl-wg@ogf.org
Date: 22/10/2011
19:12
Subject: [DFDL-WG]
lengthKind='prefixed' clarification needed
Sent by: dfdl-wg-bounces@ogf.org
For agenda/issues list
With respect to lengthKind='prefixed'. I'm concerned that there's a complex
interaction with initiator/terminator.
Can we have a prefix length and an initiator and terminator as well? If
so which comes first, and if it's the prefix does the prefix length include
the length of the initiator and terminator?
The grammar as written in current draft of spec has the initiator first,
then the prefix, then the content, and then the terminator. I think this
is wrong. I mean we can make it work, but it's not a useful, nor intuitive
behavior.
If we're going to fix this, I think we should make prefixed an alternative
to initiator and terminator, so that you can't have both on the element.
The alternative is to change the order around. Because initiator and terminator
can each be lists of alternative choices, the only sensible composition
of prefixed with these has prefix length providing the length of a syntax
which includes static initiator and terminator fields, which are sort of
like static padding to be trimmed off the string before extracting the
value.
E.g., prefix length of 10 preceeding these characters: [[123456]]
<element name='x' type='int' dfdl:initiator="[[" dfdl:terminator="]]"
dfdl:lengthKind='prefixed' .../>
But,....this is obscure enough that I'd rather make prefix length exclusive
of initiator/terminator. I.e. Schema Def Error if both are specified.
Rationale: Even if such formats are possible, and even if they do exist
somewhere, it's possible to model this format differently with hidden fields,
lengthKind='explicit' etc., so it's not like removing this complex interaction
of prefix with initiator/terminator reduces DFDL's expressive power in
any way.
Summary: To reduce complexity, suggest that lengthKind='prefixed' is exclusive
of both initiator and terminator properties directly on the same element.
Schema Definition Error if both are specified.
--
Mike Beckerle | OGF DFDL WG Co-Chair
Tel: 781-330-0412
--
dfdl-wg mailing list
dfdl-wg@ogf.org
http://www.ogf.org/mailman/listinfo/dfdl-wg
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
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
--
Mike Beckerle | OGF DFDL WG Co-Chair
Tel: 781-330-0412
--
Mike Beckerle | OGF DFDL WG Co-Chair
Tel: 781-330-0412
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