🚀 go-pugleaf

Stan Ryckman (sgr@ae-sun01.alden.com) wrote:
: In article <43ugdt$ihe@newshost.lanl.gov>,
: Tanmoy Bhattacharya <tanmoy@qcd.lanl.gov> wrote:
: >In article <43s7k4$j2v@gatekeeper.atml.co.uk>, jrr@atml.co.uk (John Rickard)
: >writes: 
: >|> Stan Ryckman (sgr@ae-sun01.alden.com) wrote:
: >|> : All pointers-to-struct must have the same "shape"
: >|> 
: >|> Is that so?  I've often written code that makes that assumption, but I
: >|> can't find any justification for it in the standard.  Can you give
: >|> references?
: 
: It was somewhat beaten to death in this group a while ago, I think,
: as derivable from the standard (though not stated there as such)
: in order to make pointers to incomplete structures work,

(The article by Tanmoy Bhattacharya hasn't arrived here yet, so I may
be missing something.)

It is indeed with pointers to incomplete structures that I have made
the assumption.  But where does the standard say -- or even suggest --
that such pointers must `work'?

So far as I can tell from the standard, if struct S is an incomplete
structure type that is never completed in the translation unit, there
are no requirements for the type struct S * to be compatible with, or
have the same representation as, any other type.  If (say) char * and
int * have different representations, then the implementation can
(again, so far as I can tell) choose to give struct S * the same
implementation as either of those, or even put absurdly strong
alignment restrictions on never-completed structure types and
represent struct S * in a single byte as the top 8 bits of the
address.

Or is the argument that one can write, say,

  file1.c:                file2.c:
     struct S1;              struct S2 { int i; };

knowing secretly, and not telling the compiler, that struct S1 really
has a single member, called `i', of type int?  If this is so, then
6.1.2.6 says that struct S1 and struct S2 are compatible types, and so
(by 6.5.4.1) struct S1 * and struct S2 * are compatible types.  That
does not seem the most plausible reading of the standard to me (I'd
say that struct S1 had no members), but I suppose it is possible.

-- 
John Rickard