orbitcpp-list

[orbitcpp-list] warnings

[Laszlo M. <las...@et...>]

Hi Folks!

Some days ago I tried to investigate why gcc gives tons of warnings
during compilation of orbitc++ generated stuff. After I found the main
cause (those const/non const cast operators) I looked at some other
orbs how they solve this problem. I found special ifdefs in both mico
and omniorb, unfortunately none of those solutions were 100% perfect.
Anyway when I do something similar for orbitc++, I was not able to
compile everything :(

And unfortunately gcc 2.96 issues these warnings too :(

But there is another kind of warnings, which is caused by this
stuff in pass_stubs.cc:

void IDLPassStubs::doInterface(IDLInterface &iface)

[...]

       // make cast operators for all mi base classes
        IDLInterface::BaseList::const_iterator
        first = iface.m_all_mi_bases.begin(),last = iface.m_all_mi_bases.end();

        while (first != last) {
                m_header
                << indent << "operator " << (*first)->getQualifiedCPPIdentifier() << " &() {" << endl;
                m_header
                << ++indent << "return *" << (*first)->getQualifiedCPPCast("this") << ';' << endl;
                m_header
                << --indent << '}' << endl;
                first++;
        }

Here gcc says, that there is no point to provide

   operator Base&()

for the derived classes, because it won't execute them, but do the
right thing itself ;)

So I guess, this code could be safely removed.

Laszlo

Re: [orbitcpp-list] warnings

[Phil D. <ph...@us...>]

Laszlo Molnar writes:
 > Hi Folks!

Hi Laslo,


 > 
 > Some days ago I tried to investigate why gcc gives tons of warnings
 > during compilation of orbitc++ generated stuff. After I found the main
 > cause (those const/non const cast operators) I looked at some other
 > orbs how they solve this problem. I found special ifdefs in both mico
 > and omniorb, unfortunately none of those solutions were 100% perfect.
 > Anyway when I do something similar for orbitc++, I was not able to
 > compile everything :(
 > 
 > And unfortunately gcc 2.96 issues these warnings too :(
 > 

I need to do some work in this area. Any info on what sort of hacks
mico and omniorb use (or where to look for them) would be very
helpful.


 > But there is another kind of warnings, which is caused by this
 > stuff in pass_stubs.cc:
 > 
    [example snipped]

 > Here gcc says, that there is no point to provide
 > 
 >    operator Base&()
 > 
 > for the derived classes, because it won't execute them, but do the
 > right thing itself ;)
 > 
 > So I guess, this code could be safely removed.
 > 

Hmmm - I suspect not. The thing is that we aren't using real C++
objects, we're casting ORBit C objects into C++ ones, and using these
cast operators to insure that the compiler doesn't try to do smart
thunking in the case of multiple inheritance downcasts.  This provides
a good speed increase over wrapping the C objects in C++ ones.

Does the 'everything' test in orbitcpp CVS work with gcc 2.96? 

Thanks for your work,

Cheers,

Phil.

Re: [orbitcpp-list] warnings

[Laszlo M. <las...@et...>]

> I need to do some work in this area. Any info on what sort of hacks
> mico and omniorb use (or where to look for them) would be very
> helpful.

OK, mico does this in include/var.h:

template<class T>
class ObjVar
{
  // [...]
// g++ const overload problem
#ifdef HAVE_CONST_OVERLOAD
  operator T*() const
  {
      return _ptr;
  }
  operator T*&()
  {
      check();
      return _ptr;
  }
#else
  operator T*&() const
  {
      ((ObjVar<T>*)this)->check();
      return (T*&)_ptr;
  }
#endif
}

and here is the test from its aclocal.m4

AC_DEFUN(AC_CONST_OVERLOAD,
[AC_MSG_CHECKING(for const overload)
AC_CACHE_VAL(ac_cv_have_const_overload,
[ac_cv_tmp_old_cxxflags="$CXXFLAGS"
if test X"$GXX" = Xyes; then
  CXXFLAGS="-Werror $CXXFLAGS"
fi
AC_TRY_COMPILE_GLOBAL(,[
  struct S {
    int *i;
    operator const int *() const { return i; }
    operator int *&() { return i; }
    operator const int &() const { return *i; }
    operator int &() { return *i; }
  };

  void bar (const int &, int &, const int *, int *, int *&);

  void foo (S &s, const S &cs)
  {
     int i = s; i = cs; i = *s; i = *cs;
     bar (s, s, s, s, s);
     bar (cs, s, cs, s, s);
  }
],
eval "ac_cv_have_const_overload=yes",
eval "ac_cv_have_const_overload=no",)
CXXFLAGS=$ac_cv_tmp_old_cxxflags
])


>  >    operator Base&()
>  > So I guess, this code could be safely removed.
> Hmmm - I suspect not. The thing is that we aren't using real C++
> objects, we're casting ORBit C objects into C++ ones, and using these
> cast operators to insure that the compiler doesn't try to do smart
> thunking in the case of multiple inheritance downcasts.  This provides
> a good speed increase over wrapping the C objects in C++ ones.

The problem is that as the warning says, gcc will NOT USE that operator
when it does the casting. Consider this example:

struct Base
{};

Base dummy;

struct Derived  : public Base
{
    operator Base &() { return dummy; }
};

gcc gives:
x.cc:8: warning: conversion to a reference to a base class will never
use a type conversion operator

And gcc is right, because it's totally silly what I'm doing.
I have no idea, why this is only a warning, IMHO it should be an error.

> Does the 'everything' test in orbitcpp CVS work with gcc 2.96?

Well, I just checked whether this const/non const warning is still
there, and didn't compile anything useful (I doubt I could - 2.96
is not binary compatible with my libraries).

Laszlo