The offsetof macro

[Walter Murray]

Karl Heuer writes:

>In article <8402 at smoke.ARPA> gwyn at smoke.ARPA (Doug Gwyn) writes:
>>In article <16490006 at hpcllca.HP.COM> walter at hpcllca.HP.COM (Walter Murray) writes:
>>>So what is a valid way to define offsetof?
>>
>>There is no portable way to #define an offsetof() macro in C.
>>The usual method you already know will nonetheless work on a
>>large number of implementations.
>
>I don't think the question was about a *portable* way to define it.  The
>original poster was noting that the "usual method" involves a macro that
>involves some trickery with a null pointer constant, and claimed that since
>this violates a constraint, a conforming compiler is required to diagnose an
>error.  This would seem to imply that offsetof() *must* be defined as a
>compiler builtin, even on well-behaved architectures.
>
>But I don't believe this to be the case.  The relevant portion of 3.4 is under
>"Semantics", not "Constraints", and says "Cast operators in an integral
>constant expression shall only convert arithmetic types to integral types".  A
>violation of a "shall" clause outside a constraint is undefined behavior,
>which does not require diagnosis.  Thus, a conforming compiler can do the
>right thing with the null pointer hack, and use it in offsetof().
>
>If there really is a violated constraint involved, could someone post the
>reference?
>
>Karl W. Z. Heuer (ima!haddock!karl or karl at haddock.isc.com), The Walking Lint
>----------

My thinking is that 3.4 defines the term "integral constant expression".
It seems to me that it doesn't matter that that definition is in the
Semantics section:  An expression which uses the null pointer hack
is, by definition, NOT an "integral constant expression".

Now consider a switch statement.  According to a CONSTRAINT in 3.6.4.2,
"The expression of each case label shall be an integral constant
expression."  An expression which uses the null pointer hack is
not an "integral constant expression", and therefore would violate
this constraint and require a diagnostic, would it not?

Is a conforming implementation free to change its definition of
"integral constant expression" on the grounds that the definition
appears under Semantics instead of Constraints?  

Walter Murray