[Fedora-packaging] Re: ocaml signature hashing: really neccessary?

[Axel Thimm]

On Fri, Jun 15, 2007 at 12:29:13PM +0100, Richard W.M. Jones wrote:
>  Axel Thimm wrote:
>  >I wonder whether this is maybe overdesigned. AFAIU this signature
>  >hashing was made because ocaml is not considered stable enough to
>  >carry over signatures from release to release.
>  >
>  >Same could be told about hundreds of C libraries, wouldn't the
>  >neccessity in ocaml then imply a neccessity to hash C-library APIs as
>  >well? Maybe it's something we will consider to do someday, but the
>  >order would be to cater for C/C++/Fortran/etc libraries first and then
>  >for niche languages like ocaml.
>  >
>  >I think it's a bit too much, or did I miss something important (I'm
>  >not a real ocaml user, there is just this one application that even
>  >justifies ocaml's existance ;)
>  
>  No, it's really necessary and has nothing to do with stability or 
>  otherwise of OCaml (which is a very mature language that has been around 
>  in one form or another since the mid 80s).
>  
>  When OCaml compiles a library A, it takes a hash over the whole 
>  interface -- every single function, every argument to every function, 
>  and some of the internals, are just some of the things included in this 
>  hash.
>  
>  When OCaml compiles library B which depends on library A, it encodes the 
>  hash of A into B.
>  
>  Now we come to link a program against library B (and hence against 
>  library A).  The hashes are checked and the linking will fail if, for 
>  example, the hash of A has changed since B was compiled.
>  
>  C has only weak checking in comparison.  Sure, you can change a library, 
>  but you'd better hope for example that some struct in that library 
>  didn't change the size of one of its fields.  If it did your program 
>  will still link, but will fail in interesting ways at runtime.
>  
>  OCaml's checking has the big downside, which is that it goes above and 
>  beyond what is necessary for just checking compatibility.  For example, 
>  you can't add more functions to library A, even though such a change is 
>  probably safe.  Nevertheless, RPM hashes are just enforcing what the 
>  OCaml linker enforces, and without them you'd be able to install 
>  incompatible OCaml RPMs which won't actually work together. 
Thanks Rich for dusting off an old mind. I agree, if this is an
upstream mechanism that makes even conventional rpm packaging fail,
then we need that. I thought it was something put on top of ocaml,
e.g. a pure packaging level hashing. Standing corrected and in
agreement now. ;)
-- 
Axel.Thimm at ATrpms.net