New subject: binaryNumberRep limitations, xs:decimal and binaryDecimalVirtualPoint

8 Jun 2022

      Rather long email, apologies in advance.

We are encountering many issues where DFDL is too limited in its binary
number representations.

6 points are made in this email.

To not bury the lead, which is Point 6, I am proposing that we make
dfdl:binaryNumberRep an extensible enum, allowing QNames as values so that
implementations can extend the set of supported binary number
representations. E.g., dfdl:binaryNumberRep='dfdlx:onesComplement', or
dfdl:binaryNumberRep='daffodil:signedASN1BERVariableLengthInteger'

Point 1: I think this is just a needed clarification. But Points 1 and 2
drove the motivation for this whole discussion.

So the description of dfdl:binaryDecimalVirtualPoint says it is allowed on
types whose *base* is xs:decimal.

That term 'base' is confusing. It can either mean allowed on types derived
from xs:decimal, as all the signed and unsigned integer types are.
Or it can mean literally <restriction base="xs:decimal"> ....
</restriction>. That is, the base attribute of simple type restriction must
be xs:decimal, or the type="xs:decimal" directly on the element.

I believe the latter is the only thing that makes sense. If
dfdl:binaryDecimalVirtualPoint is positive, as this type below is nonsense,
as the type says it is an unsigned integer, but the
binaryDecimalVirtualPoint says to divide by 100, so that 36000 would become
360.00, a non-integer. The infoset <angle>360.00</angle> makes no sense for
an integer type, and would not validate treating the DFDL schema as an XSD.

<element name="angle" type="with2DecimalFractionDigits"/>

<simpleType name="with2DecimalFractionDigits"
dfdl:binaryDecimalVirtualPoint="2"> <!-- will divide by 100 -->
   <restriction base="xs:unsignedShort"> <!-- makes no sense. Has to be
xs:decimal -->
      <maxInclusive value="360"/>
   </restriction>
</simpleType>

So based on that reasoning, I think we need to improve the clarity and say
that dfdl:binaryDecimalVirtualPoint applies only to type xs:decimal alone.
(Removing the word 'base' which creates the subtype confusion.)

That brings me to the next related point.

Point 2: Need for binaryNumberRep='unsignedBinary'

We have a use case we cannot express.

An angle from 0 to 360 degrees is represented by an unsigned 16 bit binary
integer which if dfdl:binaryDecimalVirtualPoint works with that, can
represent 000.00 to 360.00.

<simpleType name="angle360" dfdl:lengthKind="explicit" dfdl:length="16"

    dfdl:binaryNumberRep="binary" dfdl:binaryDecimalVirtualPoint="2">

    <restriction base="xs:decimal">

      <minInclusive value="0"/>

      <maxInclusive value="360"/>

    </restriction>

  </simpleType>

However, we have no way to say that the above is to be unsigned binary
integer representation.
dfdl:binaryNumberRep='binary' means unsigned binary if the type is
unsigned, and means signed twos-complement if the type is signed, which
xs:decimal is.

So my definition of angle360 above is no good, as the maximum positive
value is 327.67, which is insufficient.

I think we need to revise dfdl:binaryNumberRep to allow for distinguishing
binary unsigned from twosComplement signed types, as well as the packed
types.

Note that the packed types allow 'bcd' which is an unsigned representation,
so there is sort of a precedent there for allowing the binaryNumberRep type
to be unsigned even if the type is signed-capable.

There is already a proposal to add "offsetBinary" as a signed binary
integer representation. https://github.com/OpenGridForum/DFDL/issues/7
So I'm suggesting adding "unsignedBinary' as well.

So the complete set (so far) would be dfdl:binaryNumberRep:

   - 'twosComplementBinary' (with legacy 1.0 name 'binary')
   - 'unsignedBinary'
   - 'offsetBinary'
   - 'packed'
   - 'bcd'
   - 'ibm4690Packed'

Point 3:  There are other binary integer representations that will be
needed.

This table comes from a format specification we use:

[image: image.png]
Ignore the 'Logical' column above, that's about enums. Ignore the "*" which
is just about when a value must be reserved as an in-band null indicator
which is the suggested such value.

What is called 'Mod Twos Complement' here is what our existing proposed
DFDL 2.0 feature calls 'offsetBinary'.

So this table suggests the need for 'unsignedBinary' (already mentioned),
but also two others: 'signPlusMagnitudeBinary', and 'onesComplementBinary'.

Point 4: Zig Zag Integer representation is getting popular

There's one other representation I know of which is more recent/modern
called zig-zag integers, popularized by google protocol buffers, but it's a
clever representation and seemingly used in many places now.

*Binary Value*  *Zig Zag*
000            0
001           -1
010            1
011           -2
100            2
101           -3
110            3
111           -4

Point 5: Variable Length Binary Integers

There are also variable-length integer formats that are not just strings of
bits. A common one I have seen is used by ASN.1 BER representation where
each byte if its MSB is 1 indicates that the integer extends an additional
byte, contributing 7 bits to the value. Unsigned integers are just the
concatenation of these bits.

Signed integers are handled after the bits are concatenated together. If
the first bit of the concatenation is 1, the value is twos complement
negative value. Hence, if a positive value would have a first bit of 1,
then an additional byte containing 10000000 must be used as the most
significant byte so that the first bit will not be 1.

There is no way in DFDL to represent such a variable length integer
representation and get an integer in the infoset. You have to use a
hexBinary byte array.

There is a need for a variable-length integer like this to support not only
explicit length (used by ASN.1 BER), but implicit length as well. In this
case the last byte of the variable length integer does not have the MSB
set. Hence, a single byte can represent signed -64 to +63, or unsigned 0 to
127. Outside that range multiple bytes must be used, each byte contributing
7 bits.

This suggests a need for several additional dfdl:binaryNumberRep enums.

Point 6: Extensibility by implementations is needed here

There are many other representations out there as well.

I think we should have a convention where there is a core set that all DFDL
representations must provide, and a convention by which DFDL
implementations can provide additional support.

To me, a good way to do this is to allow the enum values for
dfdl:binaryNumberRep to be not only regular enums (all of which are
reserved) but QName syntax, where the prefix can be for a namespace
recognized by an implementation for providing an extended set of binary
number representations. (Perhaps the dfdlx: prefix and namespace, or maybe
we just allow implementation specific namespaces?)

This means of extending enums for existing properties is not part of our
existing 'experimental features' conventions, but I propose that it should
be added.

To me, this is a good way to generally allow property enums to be extended
with experimental features in DFDL implementations, and applies to other
places such as dfdl:binaryCalendarRep, and numerous other properties where
we are finding a need for additional enums and want to add them as
experimental features.

That was long. Thanks for your consideration.

Mike Beckerle
Apache Daffodil PMC | daffodil.apache.org
OGF DFDL Workgroup Co-Chair | www.ogf.org/ogf/doku.php/standards/dfdl/dfdl
Owl Cyber Defense | www.owlcyberdefense.com

binaryNumberRep limitations, xs:decimal and binaryDecimalVirtualPoint

Mike Beckerle

Steve Hanson

Mike Beckerle

tags

participants (2)