Re: [ros-dev] __attribute__((weak)) - in mingw

[Danny Smith]

Boaz Harrosh wrote:
> > Danny Smith wrote:
> > > Boaz Harrosh wrote:
> > > > [Q] I'm (well ATL is) using  __attribute__((weak)) (translated from
> > > > __declspec( selectany) ) for instantiation of members and variables in
> > > > headers. I had no problem with it On GCC in Linux (gcc 3.2.2). On MinGW
> > >
> > > IMAGE_COMDAT_SELECT_ANY is not quite same as PECOFF version of "weak" , but
> > > AFAICT is equivalent to the GCC section flag ".linkonce discard".  I don't
> > > think there is a way for the user to specify that for data using an
> > > attribute, but it could be done with asm statements.
> > >
> > > .weak directive is partially supported in current binutils CVS.
> > > __attribute__((weak)) is not supported by GCC-3.4.x but will be in the next
> > > major GCC release (4.0.0).

Sorry,  that was unclear.  I should have qualified with "on windows
targets"
> > > The semantics of weak for PECOFF differ from that on Linux.See the PECOFF60
> > > specs (Microsoft Portable Executable and Common Object File Format
> > > Specification)  section on weak externals
> > >
> > > Danny
> >
> > (binutils at sources dot redhat dot com please also cc me as I'm not on
> > the list)
> >
> > Attached is a proof (See fooInt.h) that gcc (gcc version 3.2 (mingw
> > special 20020817-1)) has support for weak symbols. Just not with the
> > regular syntax.
> > But when templates are used duplicate symbols are merged by the linker.
> >
> > What would be the assembler magic to cram into the __WEAK__ definition
> > that would make this project link?

I am currently testing a patch to GCC to add an __attribute__ ((selectany)) that
would work the way it is discribed in MS docs and  in the spec of
IMAGE_COMDAT_SELECT_ANY  in PECOFF doc.  Basically it puts the symbol into its
own section with ".linkonce discard"  charcteristics.

__declspec (selectany) int foo = 1;

becomes:

.globl _foo
 .section .data$foo,"w"
 .linkonce discard
 .align 4
_foo:
 .long 1

So far the only problem I've seen is that, although it works for global objects
with non-trivial constructors

eg:

struct X {
  static int s;
  int m;
  X(int _i) : m(_i) {
    m++;
  }
};

__declspec(selectany) int X::s = -1;  // OK
__declspec(selectany) X x(1); // This is OK but app will be bloated with
                              // duplicate initialization code.
.
I haven't figured out how to get rid of the duplicated inititialization code

I will submit the patch for comment  once  GCC trunk branches.

Danny
> > Free Life
> > Boaz