[Bug regression/57551] [4.9 Regression]: g++.dg/ext/visibility/anon6.C scan-assembler 1BIiE1cE

[hubicka at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57551

Jan Hubicka <hubicka at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jason at redhat dot com

--- Comment #2 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
You are right, I added attribute used to the variable and forgot about it. The
testcase is bogus by expecting unused local static to not be optmized away.

Looking closer, I am however confused. While the attribute fixes the testcase,
i think something wrong is still going on hre.

We get the following symbol out of FE:
_ZN12_GLOBAL__N_11BIiE1cE/1 (const int {anonymous}::B<int>::c) @0x7ffff75141a0
  Type: variable definition analyzed
  Visibility: force_output comdat
  Aux: @0x1  References:
  Referring:
  Availability: not-ready
  Varpool flags: initialized read-only const-value-known

i.e. it is COMDAT but not public. It is the force_output flag that keeps
variable alive and that is there because of keying. The variable is output as
static object:
jh@gcc10:~/trunk/build3/gcc$ more anon6.s
        .file   "anon6.C"
        .section        .rodata
        .align 4
        .type   _ZN12_GLOBAL__N_11BIiE1cE, @object
        .size   _ZN12_GLOBAL__N_11BIiE1cE, 4
_ZN12_GLOBAL__N_11BIiE1cE:
        .zero   4
        .ident  "GCC: (GNU) 4.9.0 20130602 (experimental)"
        .section        .note.GNU-stack,"",@progbits

i.e. no comdat visibility. This flag stays through the compilation and seems to
be ignored by output routines.  

And now adding the attribute used:

Index: /home/jh/trunk/gcc/testsuite/g++.dg/ext/visibility/anon6.C
===================================================================
--- /home/jh/trunk/gcc/testsuite/g++.dg/ext/visibility/anon6.C  (revision
199698)
+++ /home/jh/trunk/gcc/testsuite/g++.dg/ext/visibility/anon6.C  (working copy)
@@ -18,7 +18,7 @@ namespace
   template <typename T>
   class B
   {
-    static const T c = 0;
+    __attribute__ ((__used__)) static const T c = 0;
   };

   template <typename T> const T B<T>::c;

changes object to:
_ZN12_GLOBAL__N_11BIiE1cE/1 (const int {anonymous}::B<int>::c) @0x7ffff75141a0
  Type: variable definition analyzed
  Visibility: force_output forced_by_abi public weak comdat
comdat_group:_ZN12_GLOBAL__N_11BIiE1cE one_only
  Aux: @0x7ffff7402850  References:
  Referring:
  Availability: not-ready
  Varpool flags: initialized read-only const-value-known

which changes assembly to:
jh@gcc10:~/trunk/build/gcc$ ./xgcc -B ./ -O2
/home/jh/trunk/gcc/testsuite/g++.dg/ext/visibility/anon6.C -S -std=c++98
-pedantic-errors -Wno-long-long -S -fdump-ipa-all-details -fdump-ipa-cgraph
jh@gcc10:~/trunk/build/gcc$ more anon6.s 
        .file   "anon6.C"
        .weak   _ZN12_GLOBAL__N_11BIiE1cE
        .section       
.rodata._ZN12_GLOBAL__N_11BIiE1cE,"aG",@progbits,_ZN12_GLOBAL__N_11BIiE1cE,comdat
        .align 4
        .type   _ZN12_GLOBAL__N_11BIiE1cE, @gnu_unique_object
        .size   _ZN12_GLOBAL__N_11BIiE1cE, 4
_ZN12_GLOBAL__N_11BIiE1cE:
        .zero   4
        .ident  "GCC: (GNU) 4.9.0 20130605 (experimental)"
        .section        .note.GNU-stack,"",@progbits

Here we now disagree with GCC 4.5 that still outputs this as static.

Jason, why C++ FE behaves this way now?

clang also optimizes out the variable in testcase and adding attribute used
makes the var to be static.

I think the testcase should be fixed by adding the used flag and by checking
that weaks/comdats are not involved in the assembly and C++ FE should be fixed.
or do I miss something?

Honza