Created attachment 14097 [details]
preprocessed

I get this too:
/home/apinski/src/local/gcc/libstdc++-v3/src/system_error.cc:67: error: std::system_category causes a section type conflict

http://gcc.gnu.org/ml/gcc-patches/2007-08/msg01566.html
Broke this.

Testcase:
namespace std {
  struct gnu_error_category 
  {
    virtual const int name() const {}
  };
  const gnu_error_category gnu_category;
}
namespace std {
  const gnu_error_category& system_category = gnu_category;
}


------- CUT ----
Compile with -O1 -fdata-section.

A patch was posted:
http://gcc.gnu.org/ml/gcc-patches/2007-08/msg01683.html

This is related to PR 31490

Subject: Bug 33168

Author: dje
Date: Wed Aug 29 17:01:35 2007
New Revision: 127910

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=127910
Log:
2007-08-29  Paolo Bonzini  <bonzini@gnu.org>

        PR target/33168
        * config/rs6000/rs6000.c (compare_section_name): New function.
        (rs6000_elf_in_small_data_p): Compare section prefixes instead
        of full name.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/rs6000/rs6000.c

Subject: Bug 33168

Author: jakub
Date: Thu Aug 30 16:43:19 2007
New Revision: 127928

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=127928
Log:
	PR target/33168
	* config/rs6000/rs6000.c (rs6000_elf_in_small_data_p): Return
	true if any of the compare_section_name calls returned true,
	rather than if any returned false.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/rs6000/rs6000.c

Fixed.

This still fails in 64-bit mode bootstrap.

We first choose a section here, when decl readonly_flag is false:

#0  get_section (name=0x40000cca824 ".data._ZSt15system_category", flags=512, 
    decl=0x40000e44000) at /src/gcc-current/gcc/varasm.c:527
#1  0x000000001067bd70 in get_block_for_decl (decl=0x40000e44000)
    at /src/gcc-current/gcc/varasm.c:1196
#2  0x000000001067c588 in make_decl_rtl (decl=0x40000e44000)
    at /src/gcc-current/gcc/varasm.c:1395
#3  0x000000001067d3d8 in notice_global_symbol (decl=0x40000e44000)
    at /src/gcc-current/gcc/varasm.c:1558
#4  0x00000000106dd04c in varpool_mark_needed_node (node=0x40000f12500)
    at /src/gcc-current/gcc/varpool.c:192
#5  0x00000000106dda78 in varpool_finalize_decl (decl=0x40000e44000)
    at /src/gcc-current/gcc/varpool.c:297
#6  0x0000000010403fac in rest_of_decl_compilation (decl=0x40000e44000, 
    top_level=1, at_end=0) at /src/gcc-current/gcc/passes.c:192
#7  0x000000001002c624 in make_rtl_for_nonlocal_decl (decl=0x40000e44000, 
    init=0x0, asmspec=0x10b4f398 "") at /src/gcc-current/gcc/cp/decl.c:5136
#8  0x000000001003e1ec in cp_finish_decl (decl=0x40000e44000, init=0x0, 
    init_const_expr_p=0 '\0', asmspec_tree=<value optimized out>, flags=8)
    at /src/gcc-current/gcc/cp/decl.c:5511

Note flags=512 is SECTION_WRITE in get_section call.
 
We set readonly near the end of cp_finish_decl.  This change in the readonly flag is the cause of later trouble when assemble_variable is called, as targetm.section_type_flags (called from get_named_section, from get_variable_section, from assemble_variable) now returns 0.  get_section is then called with flags=0, triggering the error.

Segher has a patch for related PR 31490 that fixed the original 32-bit bootstrap problem.  Yesterday I confirmed that it also fixes the 64-bit powerpc64-linux bootstrap for all languages but Ada.  The only test differences between a 64-bit build with his patch and a 32-bit build without it are gcc.dg/20051207-3.c, which he earlier said is expected to behave differently with this patch, and cc1 segfaults with some of the gcc.c-torture/compile/limits-*.c tests, which also fail with a C-only 64-bit bootstrap without the patch.

Here's the patch I tested:

Index: gcc/varasm.c
===================================================================
--- gcc/varasm.c        (revision 129361)
+++ gcc/varasm.c        (working copy)
@@ -536,10 +536,20 @@
   else
     {
       sect = *slot;
+
+      /* Sanity check user variables for flag changes.  */
       if ((sect->common.flags & ~SECTION_DECLARED) != flags
          && ((sect->common.flags | flags) & SECTION_OVERRIDE) == 0)
        {
-         /* Sanity check user variables for flag changes.  */
+         /* Allow mixed writable and read-only objects in named sections.  */
+         if ((sect->common.flags & SECTION_NAMED) != 0
+              && ((sect->common.flags ^ flags) & ~SECTION_DECLARED)
+              == SECTION_WRITE)
+           {
+             sect->common.flags |= SECTION_WRITE;
+             return sect;
+           }
+
          if (decl == 0)
            decl = sect->named.decl;
          gcc_assert (decl);

When I first tested this patch a month ago (with a 32-bit bootstrap) there was a discussion between Segher and Dinar Temirbulatov about the best way to fix this; I've copied both of them.

I disagree with this patch (at least for this bug).  The bug is on cp_finish_decl
side.  Working on it.

Smaller testcase:
// PR target/33168
// { dg-do compile }
// { dg-options "-O2 -fdata-sections" }

extern const int& foo;
namespace
{
  const int bar = 16;
}
const int &foo = bar;

Subject: Bug 33168

Author: jakub
Date: Tue Nov  6 08:33:25 2007
New Revision: 129922

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=129922
Log:
	PR target/33168
	* decl.c (cp_finish_decl): Call make_rtl_for_nonlocal_decl already
	with the final TREE_READONLY flag in place.  processing_template_decl
	is known to be 0 in this part of function.

	* g++.dg/other/datasec1.C: New test.

Added:
    trunk/gcc/testsuite/g++.dg/other/datasec1.C
Modified:
    trunk/gcc/cp/ChangeLog
    trunk/gcc/cp/decl.c
    trunk/gcc/testsuite/ChangeLog

Fixed, ppc64-linux bootstraps just fine now even without --with-cpu=default32.

Subject: Bug 33168

Author: aoliva
Date: Thu Jan  3 05:15:42 2008
New Revision: 131276

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=131276
Log:
gcc/cp/ChangeLog:
PR target/33168
* decl.c (cp_finish_decl): Call make_rtl_for_nonlocal_decl already
with the final TREE_READONLY flag in place.  processing_template_decl
is known to be 0 in this part of function.
gcc/testsuite/ChangeLog:
PR target/33168
* g++.dg/other/datasec1.C: New test.

Added:
    branches/var-tracking-assignments-branch/gcc/testsuite/g++.dg/other/datasec1.C
Modified:
    branches/var-tracking-assignments-branch/gcc/cp/ChangeLog
    branches/var-tracking-assignments-branch/gcc/cp/decl.c
    branches/var-tracking-assignments-branch/gcc/testsuite/ChangeLog