GCC Bugzilla Bug 15721

Here is the test program:

template <typename T> class gnu_obj_2
{
  public:
    static int my_static;
};

template <typename T> int gnu_obj_2<T>::my_static = 117;

// instantiate templates explicitly so their static members will exist
template class gnu_obj_2 <int>;
template class gnu_obj_2 <long>;

int main ()
{
  gnu_obj_2 <int>  test_int;
  gnu_obj_2 <long> test_long;
  return 0;
}

I compile this program with "gcc -S -gstabs+".  Then I look in the generated
assembly code for the static data members gnu_obj_2<int>::my_static and
gnu_obj_2<long>::my_static.

(I don't think the -gstabs+ is important but I just happened to start with this
flag and have kept it all the way till now, sigh.  This code originally came
from the gdb test suite, gdb.cp/m-static.cc).

The explicit instantiations of gnu_obj_2<int> and gnu_obj_2<long> are supposed
to instantiate the static data members along with everything else, but they don't.

http://gcc.gnu.org/ml/gcc/2004-05/msg01394.html

The last working gcc was 2004-05-19 01:00:00 UTC and the first broken gcc was
2004-05-19 02:00:00 UTC.  The problem is with this patch:

http://gcc.gnu.org/ml/gcc-patches/2004-05/msg01157.html
IMA repairs: Don't use DECL_ASSEMBLER_NAME for a key in cgraph(take two)

I'll add three attachments: the test program; the last good assembly code; and
the first bad assembly code.

Created an attachment (id=6424) [edit]
Test program

Compile with: "gcc -S -gstabs+".

Look for gnu_obj_2<int>::my_static and gnu_obj_2<long>::my_static in the
generated code.

Created an attachment (id=6425) [edit]
assembly code, last good compiler, 2004-05-19 01:00:00 UTC

Generated assembly code with "gcc -S -gstabs+ template-static.cc".
This is with the last good compiler.

Created an attachment (id=6426) [edit]
Generated code, first bad compiler, 2004-05-19 02:00:00 UTC

Code generated with "gcc -S -gstabs+ template-static.cc".
This is with the first bad compiler.

Confirmed,  I will try to figure out what is going on but if Zack beats me, I
would not be made.

mark_decl_referenced is being called for the variable.

The problem is that decl is not finalized before being marked as needed.

I know that rest_of_decl_compilation is being called for the decl but not
assemble_variable for some 
reason.  I have not checked how it was done before Zack's patch.

The difference between before and after is that DECL_EXTERNAL (decl) in
rest_of_decl_compilation is 
true after but false before which causes it not be outputted.

Ignore that last comment, I was wrong, it does not even get to
rest_of_decl_compilation for that decl.

Okay, wrapup_global_declarations checks TREE_SYMBOL_REFERENCED of
DECL_ASSEMBLER_NAME of the 
decl which is where the problem is.

This should work (I have tested it on the testcase though), I have not
bootstrapped and tested it yet:
Index: toplev.c
===============================================================
====
RCS file: /cvs/gcc/gcc/gcc/toplev.c,v
retrieving revision 1.899
diff -u -p -r1.899 toplev.c
--- toplev.c    30 May 2004 18:32:27 -0000      1.899
+++ toplev.c    2 Jun 2004 00:41:09 -0000
@@ -1455,16 +1455,17 @@ wrapup_global_declarations (tree *vec, i

          if (TREE_CODE (decl) == VAR_DECL && TREE_STATIC (decl))
            {
+              struct cgraph_varpool_node *node;
              bool needed = 1;
+              node = cgraph_varpool_node (decl);

-             if (flag_unit_at_a_time
-                 && cgraph_varpool_node (decl)->finalized)
+             if (node && node->finalized)
                needed = 0;
              else if ((flag_unit_at_a_time && !cgraph_global_info_ready)
                       && (TREE_USED (decl)
                           || TREE_USED (DECL_ASSEMBLER_NAME (decl))))
                /* needed */;
-             else if (TREE_SYMBOL_REFERENCED (DECL_ASSEMBLER_NAME (decl)))
+             else if (node && node->needed)
                /* needed */;
              else if (DECL_COMDAT (decl))
                needed = 0;

Mine, I posted a patch:
<http://gcc.gnu.org/ml/gcc-patches/2004-06/msg00088.html>.

Andrew, this patch was approved by Mark, I assume you just forgot to apply and 
close the bug...

Subject: Bug 15721

CVSROOT:	/cvs/gcc
Module name:	gcc
Changes by:	pinskia@gcc.gnu.org	2004-06-19 16:17:01

Modified files:
	gcc            : ChangeLog toplev.c 

Log message:
	2004-06-19  Andrew Pinski  <apinski@apple.com>
	
	PR c++/15721
	* toplev.c (wrapup_global_declarations): Do not check
	TREE_SYMBOL_REFERENCED of the DECL_ASSEMBLER_NAME but check
	cgraph_varpool_node's needed field.

Patches:
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&r1=2.4040&r2=2.4041
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/toplev.c.diff?cvsroot=gcc&r1=1.905&r2=1.906

No I was just testing it again.

Fixed now.

Subject: Re:  [3.5 regression] template instantation omits static data members

It works for me.
Thanks, Michael C