On Tue, May 29, 2012 at 12:03 PM, Steve Hanson
<smh@uk.ibm.com> wrote:
These were on the agenda for today's cancelled
call.
2. Is parent/child alignment rule
too strict?
Spec property description for
alignment says "The alignment of
a child component must be less than
or equal the alignment of its parent
element, sequence or choice".
Experiments with creating DFDL schema
from C structures have shown that
this rule is violated. It seems unnecessary. Specifically, C ensures
that the alignment
of objects in a repeating structure
is preserved by rounding up the length of the structure, rather than aligning
it.
I think the point of this restriction was to insure that we didin't have conflicting alignments. Where the begining of the structure was aligned, say 64, but the first element was aligned at some incompatible boundary. Like 128. It would almost certainly be an error if the first element in a structure has a larger alignment constraint than the overall structure, and the user should pick where to place the constraint, and not constrain both, except when the constraints on one are implied by the constraints of another.
Example: I have a 4k page aligned structure. The first element is some type of double float coming from a global simple type def. The global simple type definition that says my specific kind of double floats are always 8-byte aligned. The 4k page is obviously 4096 byte aligned. There should be no conflict here. The page alignment "Wins" and the begining of the structure is 4k aligned, and that satisifies the 8-byte alignment of my specialized double float.
The point of having any restriction at all, is to catch conflicts. Eg., what if I had something that was 64-byte aligned, and I put it inside a structure that has only 2-byte alignment....., this is probably an error.
If this isn't going to catch any errors, then this restriction is unnecessary.
3. Does DFDL need to support
signed integer types with lengthUnits 'bits'?
Spec only allows unsigned integers
to have lengthUnits 'bits', but it is possible in C structures to delare
bit fields for signed integers.
They behave like two's complement. It looks like the spec is being
too restrictive in its types for
'bits'.
I missed this over restriction. Signed are definitely needed.