Re: c++/9881: Incorrect address calculation for static class member

[Wolfgang Bangerth <bangerth@ticam.utexas.edu>]

The following reply was made to PR c++/9881; it has been noted by GNATS.

From: Wolfgang Bangerth <bangerth@ticam.utexas.edu>
To: "Peter A. Buhr" <pabuhr@plg2.math.uwaterloo.ca>
Cc: asharji@uwaterloo.ca, <gcc-bugs@gcc.gnu.org>, <gcc-gnats@gcc.gnu.org>
Subject: Re: c++/9881: Incorrect address calculation for static class member
Date: Thu, 27 Feb 2003 16:05:07 -0600 (CST)

 >        In this code,
 > 	 foo f; // print output
 > 
 > 	 double *module::b = &(((bar *)&module::storage)->p);
 > 	 double module::storage = 0.0;
 > 
 >        The constructor of foo is run before module::b is initialized.
 >        If you change this order, the output is as you expect.
 > 
 >        W.
 > 
 >    http://gcc.gnu.org/cgi-bin/gnatsweb.pl?cmd=view%20audit-trail&database=gcc&pr=9881
 > 
 > Your response is incorrect. First, the program generates the correct
 > result with gcc3.2. If you run the example with both gcc3.2 and gcc3.3,
 > you will see that the output is different.
 
 Just for reference: of course I did this. It is not a proof, however, 
 which one is right and which one is wrong. 
 
 
 > Hence, one can conclude that
 > either gcc3.2 or gcc3.3 is wrong. Second, when the example program is
 > run with the SUN compiler, it generates the same output as for gcc3.2,
 > indicating that the problem is probably in gcc3.3.
 
 One might guess, not conclude.
 
 
 > Third, the position of the constructor has nothing to do
 > with the assignment to a static variable. The expression to initialize the
 > static variable "module::b" *MUST* be evaluated at compiler time. All static
 > variables *MUST* be initialized before any constructor is run because a
 > constructor can refer to these variables.
 
 You are referring to 3.6.2.1, which says that POD type objects with 
 constant initializers need to be initialized before dynamic objects and 
 POD objects with non-constant initializers. Within each group, objects are 
 initialized in the order in which their definitions appear.
 
 5.19 explains what is a constant expression. 5.19.1 talks about constant 
 integral expressions, 5.19.2 about everything else, and refers to 5.19.4 
 for address constant expressions. 5.19.4 says that casts are allowed, 
 but that the value of an object shall not be accessed. I would claim that 
 this is what the expression
   ((bar *)&module::storage)->p
 is doing (note: it is not accessing p, of which we are only taking the 
 address, but it is using the result of the cast; p is not a static member 
 variable). If this claim should be correct, then the initializer would be 
 non-constant, and module::b would have to be initialized after "f", so the 
 constructor of f is run _before_ module::b is initialized.
 
 > Finally, your suggestion is not even
 > a work-around, because the original problem occurred in separate compilation
 > units, so the notion of moving the constructor in this case does not apply.
 
 At which point you are out of luck in any case, since the standard does 
 not give any guarantee about the order of initialization in case there are 
 more than one translation units.
 
 May other language lawyers decide this case, regards
   Wolfgang
 
 -------------------------------------------------------------------------
 Wolfgang Bangerth             email:            bangerth@ticam.utexas.edu
                               www: http://www.ticam.utexas.edu/~bangerth/