Re: [Gnucap-devel] Feedback needed, especially non-Linux

[David Fang]

> On Thursday 22 February 2007 17:23, David Fang wrote:
> > ????????Ok, it didn't take long to find various comments
> > about link ordering: src/Makefile.am, globals.cc, etc... :)
> > ????????Link-ordering and global initialization ordering
> > between translation units are not related; you cannot depend
> > on link ordering to influence the other. ?When it happens to
> > work, consider yourself lucky!
>
> Actually, it is not guaranteed but is deterministic.  Such
> system dependencies are undesirable but exist more often than
> most of us are willing to admit.  One example is the assignment
> of device names in a unix-type operating system.  Suppose you
> have two ethernet ports.  Which is which?  What order are the
> file systems mounted?
>
> Based on claims of magic, that is one of the things I would
> expect autoconf to take care of.  It does a great job at
> working around the junk of pre-Ansi C, but we still have that
> baggage, and there isn't much there for C++.  It doesn't easily
> handle most C++ issues.

Al,

Correct, unspecified but determinstic.

Global initialization ordering is still beyond the build-system's control,
other than exhaustively trying combinations that work.  This isn't
autoconf's problem, or libtool's problem.  The holy standards of C and C++
simply never laid down the law about this ordering.  Forunately, there are
methods for enforcing proper ordering -- one can ensure correct ordering
for any non-cyclic dependence graph (of globals).

> It has to do with
>
> 1. A container obviously must be initialized before putting
> anything in it.
>
> 2. The objects in that container vary depending on
> configuration.
>
> The container is in globals.cc.  The objects in it are in
> assorted other files which may or may not actually be used.
> You can remove something, just leave it out of the Makefile.
> You can add your own modules, just add it to the Makefile.  No
> other changes are allowed.
>
> If you leave it out, you can add it at run time (like a kernel
> module).  There is no difference between a module that is
> linked at compile time and one that is added later, other than
> whether it is in the Makefile.
>
> Look at the file "Make1" .. It separates out the files that
> apply here.
>
> You will see....
> RAW_SRCS = \
> $(D_SRCS) \
> $(BM_SRCS) \
> $(C_SRCS) \
> md.cc \
> (etc.)
>
> If you remove those SRCS lines, omitting those files, the
> program will still link and run.  It won't do anything, but
> that's ok for now.  A couple of others with the link order
> problem are still there, so it won't fix the link order issue.
> I can deal with it on a small scale.
>
> > I'd advise wrapping access to
> > global variables/structures through some init-once interface,
> > like a function, rather than relying on ordering. ?If you
> > need to be precise about global object lifetime, consider
> > using a reference-count pointer to manage orderly
> > destruction. Does this help?
>
> Not sure ...   Remember, it is a requirement that there is no
> difference between a compiled in module and one attached later
> at run time.  It is also a requirement that no tables need to
> be kept in sync.  You are allowed to list it one place, only.
>
> I am leaning toward putting the commands and models in a .so
> file.  The program can start, which initializes the container,
> then load the .so.  The problem with this is that not all
> systems can do dynamic linking.  On these, the plugin feature
> won't work anyway, but it would be nice to still be able to
> build a static linked version.
>
> For an interesting sideline ....  How did they solve this
> problem for iostream?  It isn't documented anywhere, but look
> at the source, you will see.   uh...  Actually you will
> probably still wonder.  The issue on this one is that "cout"
> and friends must work in static constructors, therefore must be
> initialized before any other static constructor.
>
> Speaking of which ...
>
> Consider this program....
> ================
> #include <iostream>
> class foo {
> public:
>   foo() {std::cout << "Hello world\n";}
> } x;
>
> int main() {}
> ================

Solution:

If you need iostream facilities during global initialization (pre-main),
you can guarantee it is set up by inserting:

static const std::ios_base::Init _init;

before uses of std::cout, etc...

> Does it work?
> How?

By the same token, it also makes sure the stream facilities stay alive
long enough during global destruction, (some reference-count like
mechanism) since destruction happens in reverse order (within each
translation unit).



David Fang
Computer Systems Laboratory
Electrical & Computer Engineering
Cornell University
http://www.csl.cornell.edu/~fang/
        -- (2400 baud? Netscape 3.0?? lynx??? No problem!)