GCC Bugzilla Bug 17828

--- gcc/testsuite/g++.old-deja/g++.other/comdat5.C.jj   2004-10-04
11:20:08.335577980 +0200
+++ gcc/testsuite/g++.old-deja/g++.other/comdat5.C      2004-10-04
11:19:49.362989353 +0200
@@ -0,0 +1,5 @@
+// { dg-do link }
+// { dg-additional-sources " comdat5-aux.cc" }
+// { dg-options "-O2 -fPIC" }
+
+#include "comdat4.C"
--- gcc/testsuite/g++.old-deja/g++.other/comdat5-aux.cc.jj      2004-10-04
11:20:03.496448065 +0200
+++ gcc/testsuite/g++.old-deja/g++.other/comdat5-aux.cc 2004-10-04
11:19:59.249211735 +0200
@@ -0,0 +1,1 @@
+#include "comdat4-aux.cc"

fails on powerpc-redhat-linux.
The local label at the start of jump table is spilled into .got2 section.

GCC 3.3.x has the same problem.
Not sure if we want to backport COMDAT patches to GCC 3.3 and 3.4.
Other possible alternative would be to:
a) change ppc32 linker script, so that .got2 section includes both .got2
   and .gnu.linkonce.got2.* sections
b) change rs6000 backend, so that it uses per-function .got2 sections
   for linkonce functions
Or change ld to not error on this problem for another 3 years until all compilers
use COMDAT.

Well, the COMDAT changes on HEAD don't solve this problem either.
So I guess we need either per-function .got2 for comdat functions, 
or revert the binutils warnings (or at least turn them off by default).

If we turn off this ld warning, then we lose sight of the fact that gcc does
have a bug here.  gcc ought to be using linkonce for all sections emitted for a
given function if any section is linkonce (waiver granted for debug and other
sections that the linker edits).

Confirmed.

Is there a workaround (some --foo option for gcc/ld) for these link errors?

Created an attachment (id=9368) [edit]
broken patch

I'm actually not really sure if COMDAT .got2/.got1/.toc1 etc. (as implemented
in this patch) is the way to go.  The reason is that then suddenly we can't
have
one constant pool but need per-function constant pools
(e.g. with this patch pool_allocator.cc -m32 -O2 -fPIC fails to assemble).

Can't we instead do link editing of these ppc sections?

http://gcc.gnu.org/ml/gcc-patches/2005-07/msg01757.html

Fixed by linker patch
http://sources.redhat.com/ml/binutils-cvs/2005-07/msg00201.html

Do we want to check the comdat5* testcase in?

Subject: Bug 17828

Author: jakub
Date: Mon Nov 21 09:52:20 2005
New Revision: 107297

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=107297
Log:
        PR target/17828
        * g++.old-deja/g++.other/comdat5.C: New test.
        * g++.old-deja/g++.other/comdat5-aux.cc: New file.

Added:
   
branches/gcc-4_0-rhl-branch/gcc/testsuite/g++.old-deja/g++.other/comdat5-aux.cc
    branches/gcc-4_0-rhl-branch/gcc/testsuite/g++.old-deja/g++.other/comdat5.C
Modified:
    branches/gcc-4_0-rhl-branch/gcc/testsuite/ChangeLog

Subject: Bug 17828

Author: jakub
Date: Thu Dec  8 21:47:10 2005
New Revision: 108245

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=108245
Log:
        PR target/17828
        * g++.old-deja/g++.other/comdat5.C: New test.
        * g++.old-deja/g++.other/comdat5-aux.cc: New file.

Added:
    trunk/gcc/testsuite/g++.old-deja/g++.other/comdat5-aux.cc
    trunk/gcc/testsuite/g++.old-deja/g++.other/comdat5.C
Modified:
    trunk/gcc/testsuite/ChangeLog

Subject: Bug 17828

Author: jakub
Date: Thu Dec  8 21:53:59 2005
New Revision: 108251

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=108251
Log:
        PR target/17828
        * g++.old-deja/g++.other/comdat5.C: New test.
        * g++.old-deja/g++.other/comdat5-aux.cc: New file.

Added:
    branches/gcc-4_1-branch/gcc/testsuite/g++.old-deja/g++.other/comdat5-aux.cc
    branches/gcc-4_1-branch/gcc/testsuite/g++.old-deja/g++.other/comdat5.C
Modified:
    branches/gcc-4_1-branch/gcc/testsuite/ChangeLog