RE: Bogus 'pure virtual method called' abort in all g++ versions

[Jan Reimers <janr at molienergy dot com>]

Your storing a reference to a temporary which gets destroyed by the time you
call c.g().  When compiling with -Wall the compiler (2.95.2) still gave no
warning.  I think it should.

JR


> -----Original Message-----
> From: James A. Bednar [mailto:jbednar@cs.utexas.edu]
> Sent: Tuesday, September 19, 2000 12:43 PM
> To: bug-gcc@gnu.org
> Subject: Bogus 'pure virtual method called' abort in all g++ versions
> 
> 
> 
> The following code aborts with a message saying 'pure virtual method
> called' (see output attached), even though C++ compilers on other
> platforms do not complain, and the code looks valid to me:
> 
>   struct A {
>     virtual ~A() { }
>     virtual void f() const=0;
>   };
>   
>   strict B : public A {
>     virtual ~B() { }
>     virtual void f() const { }
>   };
>   
>   struct C {
>     const A& a;
>     
>     C(const A& a_init=B()) : a(a_init) {
                 ^^^^^^


>       a.f();
>     }
>   
>     void g() const {
>       a.f(); 
>     }
>   };
>   
>   
>   int main(int argc, char **argv)
>   {
>     const C c;
>     c.g();
>   }
> 
> The program runs fine under g++ if the second "a.f()" call is
> commented out.  Presumably, the type B default argument to C's
> constructor is somehow being sliced to its base type A before that
> call.  Oddly enough, that slicing appears to happen *after* the
> constructor completes, since the first "a.f()" call does not cause any
> problems.
> 
> Perhaps I'm misunderstanding some obscure part of the C++ standard
> with respect to default arguments, but this behavior certainly *seems*
> incorrect.  The same behavior was observed in egcs 1.1.2, GCC 2.95.2,
> and the 20000919 CVS snapshot at
> http://www.codesourcery.com/gcc-compile.shtml.
> 
> By the way, debugging things like this would be a lot easier if we got
> a little more debugging information than just "pure virtual method
> called", at least when compiled with -g :-).
>  
> In case this code proves useful as a test case, I hereby place it into
> the public domain for anyone to use for any purpose whatsoever.
> 
> Jim Bednar
> 
> --------------------------------------------------------------
> -----------------
> pale:~> g++ -v
> Reading specs from /usr/lib/gcc-lib/i386-redhat-linux/2.95.2/specs
> gcc version 2.95.2 19991024 (release)
> pale:~> cat bug16.c
> struct A {
>   virtual ~A() { }
>   virtual void f() const=0;
> };
> 
> struct B : public A {
>   virtual ~B() { }
>   virtual void f() const { }
> };
> 
> struct C {
>   const A& a;
>   
>   C(const A& a_init=B()) : a(a_init) {
>     a.f();
>   }
> 
>   void g() const {
>     a.f(); 
>   }
> };
> 
> 
> int main(int argc, char **argv)
> {
>   const C c;
>   c.g();
> }
>   
> 
> pale:~> g++ bug16.c -o bug16
> pale:~> ./bug16 
> pure virtual method called
> Abort (core dumped)
> pale:~> diff -u bug16.c bug16a.c
> --- bug16.c     Tue Sep 19 14:30:11 2000
> +++ bug16a.c    Tue Sep 19 14:31:49 2000
> @@ -16,7 +16,7 @@
>    }
>  
>    void g() const {
> -    a.f(); 
> +    //    a.f(); 
>    }
>  };
>  
> pale:~> g++ bug16a.c -o bug16a
> pale:~> ./bug16a
> pale:~> 
> 
>