varargs historical regression question

[Frank Griffin]

I'm having a hard time googling/searching for anything that would 
explain this, and I was hoping that the description might stir a memory 
in some of those with some history here.

It involves varargs-related code which should never have worked, and in 
fact does not work in 4.7.2, but appears to have worked in older GCC 
versions as well as current (VS2008+) MS Visual C/C++.

The code example is something like this:

void  *f1(  int x, ... )
{
   va_list args;
   va_start( args, x );

   f2( x, args );
   f3( x, args );

   var_end( args );

   return;
}

void  *f2( int x, va_list args )
{
   va_arg( args, int );
   return;
}

void  *f3( int x, va_list args )
{
   va_arg( args, int );
   return;
}

The idea is that f1() creates a va_list from its variable-length 
parmlist, and passes it to f2() which advances the internal va_list 
pointer by referencing the first parameter via va_arg().  Then f3() 
references a parameter via va_arg() which should get it the second 
parameter in the variable list.

Except in the older versions we've used to build our product (2.9.9 
definitely, 3.4.6 pretty sure), it doesn't --- f3() sees the first 
va_list parameter, not the second.  Our code mistakenly assumed this to 
be the case, and has been working happily.  Of course, what it should 
have done if f3() wanted to see the first parameter  was something like:


void  *f1(  int x, ... )
{
   va_list args;
   va_list args2;
   va_start( args, x );

   va_copy( args2, args );

   f2( x, args );
   f3( x, args2 );

   var_end( args );
   var_end( args2 );

Can anyone confirm this or shed any light thereupon ?  I'll have to do 
the testing to verify that the corrected code works when built with the 
older compilers and so forth, but I'd be interested to know if this was 
a known problem and at what point it was corrected.

   return;
}