[Bug fortran/55395] [4.8 Regression] libgfortran bootstrap failure on powerpc-linux-gnu and arm-linux-gnueabi

["jakub at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55395

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2012-12-04
                 CC|                            |hubicka at gcc dot gnu.org,
                   |                            |jakub at gcc dot gnu.org
     Ever Confirmed|0                           |1

--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-12-04 10:13:06 UTC ---
Reduced testcase:
/* { dg-do compile } */
/* { dg-options "-fdata-sections -g -O2" } */
/* { dg-additional-options "-fstack-protector" { target fstack_protector } } */

extern inline __attribute__ ((__always_inline__))
void *
memcpy (void *__restrict d, const void *__restrict s, __SIZE_TYPE__ l)
{
  return __builtin___memcpy_chk (d, s, l, __builtin_object_size (d, 0));
}

void
foo (char *p)
{
  static const char q[] = "\n";
  memcpy (p, &q, 1);
}

The reason why this fails is IMHO bogus change from
http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=187631

void
varpool_remove_node (struct varpool_node *node)
{
  symtab_unregister_node ((symtab_node)node);
  if (DECL_INITIAL (node->symbol.decl)
      && !DECL_IN_CONSTANT_POOL (node->symbol.decl)
      /* Keep vtables for BINFO folding.  */
      && !DECL_VIRTUAL_P (node->symbol.decl)
      /* dbxout output constant initializers for readonly vars.  */
      && (!host_integerp (DECL_INITIAL (node->symbol.decl), 0)
          || !TREE_READONLY (node->symbol.decl)))
    DECL_INITIAL (node->symbol.decl) = error_mark_node;
  ggc_free (node);
}

(the DECL_INITIAL setting to error_mark_node).  I can understand the aim at
saving compile time memory, but this is a wrong thing to do.  dwarf2out.c uses
DECL_INITIAL heavily to emit debug info even for optimized away variables.

The exact reason for the error is that we have get_named_section called twice,
once before the varpool_remove_node call, once after it.
The first one is:
#1  0x0000000000cc64dc in get_named_section (decl=0x7ffff19a5850,
name=0x7ffff17fdbf4 ".rodata.q.4121", reloc=0) at ../../gcc/varasm.c:411
#2  0x0000000000cc7d6c in get_variable_section (decl=0x7ffff19a5850,
prefer_noswitch_p=true) at ../../gcc/varasm.c:1030
#3  0x0000000000cc7f88 in get_block_for_decl (decl=0x7ffff19a5850) at
../../gcc/varasm.c:1076
#4  0x0000000000cc8a88 in make_decl_rtl (decl=0x7ffff19a5850) at
../../gcc/varasm.c:1295
#5  0x0000000000cc8e4f in make_decl_rtl_for_debug (decl=0x7ffff19a5850) at
../../gcc/varasm.c:1350
#6  0x0000000000654bea in expand_debug_expr (exp=0x7ffff19a5850) at
../../gcc/cfgexpand.c:2777
#7  0x0000000000657357 in expand_debug_expr (exp=0x7ffff17ed460) at
../../gcc/cfgexpand.c:3358
#8  0x0000000000658b37 in expand_debug_locations () at
../../gcc/cfgexpand.c:3739
#9  0x000000000065b3c5 in gimple_expand_cfg () at ../../gcc/cfgexpand.c:4606
when creating DEBUG_IMPLICIT_PTR for the parameter s.  Then varpool_remove_node
is called:
#0  varpool_remove_node (node=0x7ffff19a6410) at ../../gcc/varpool.c:61
#1  0x0000000000cdb98a in varpool_remove_unreferenced_decls () at
../../gcc/varpool.c:406
#2  0x0000000000cdbb07 in varpool_output_variables () at
../../gcc/varpool.c:440
#3  0x0000000000686a00 in compile () at ../../gcc/cgraphunit.c:2044
#4  0x0000000000686b7a in finalize_compilation_unit () at
../../gcc/cgraphunit.c:2120
and finally get_named_section again:
#1  0x0000000000cc64dc in get_named_section (decl=0x7ffff19a5850,
name=0x7ffff17fdbf4 ".rodata.q.4121", reloc=0) at ../../gcc/varasm.c:411
#2  0x0000000000cc7d6c in get_variable_section (decl=0x7ffff19a5850,
prefer_noswitch_p=true) at ../../gcc/varasm.c:1030
#3  0x0000000000cc7f88 in get_block_for_decl (decl=0x7ffff19a5850) at
../../gcc/varasm.c:1076
#4  0x0000000000cc8a88 in make_decl_rtl (decl=0x7ffff19a5850) at
../../gcc/varasm.c:1295
#5  0x0000000000cc8e4f in make_decl_rtl_for_debug (decl=0x7ffff19a5850) at
../../gcc/varasm.c:1350
#6  0x00000000006f4d33 in rtl_for_decl_location (decl=0x7ffff19a5850) at
../../gcc/dwarf2out.c:15150
#7  0x00000000006f503f in add_location_or_const_value_attribute
(die=0x7ffff1805be0, decl=0x7ffff19a5850, cache_p=false, attr=DW_AT_location)
    at ../../gcc/dwarf2out.c:15244
#8  0x0000000000709d9b in dwarf2out_finish (filename=0x7fffffffe58b
"pr55395.i") at ../../gcc/dwarf2out.c:23218
#9  0x0000000000a3e2ce in compile_file () at ../../gcc/toplev.c:600

In the first case, DECL_INITIAL is not NULL, nor error_mark_node, nor zero
initializer, so rodata.q.4121 is assumed to be a read-only section, in the
second case DECL_INITIAL of q is already error_mark_node and thus rodata.q.4121
is assumed to be a bss section, therefore a section flags conflict.  In the
case of array initialized with a constant string right now the initializer
won't be useful, but I guess we could emit DW_OP_GNU_implicit_pointer in that
case, pointing to a DW_OP_implicit_value initialized artificial object.  But
there are tons of cases where DECL_INITIAL is even successfully used at
dwarf2out_finish time, e.g. during the deferred location processing.  So, if
you want to save memory, please do that solely if not emitting debug info.
So, I'd suggest replacing the
      /* dbxout output constant initializers for readonly vars.  */
      && (!host_integerp (DECL_INITIAL (node->symbol.decl), 0)
          || !TREE_READONLY (node->symbol.decl)))
part with
      && write_symbols == NO_DEBUG)
or ammending it with
      /* dwarf2out.c uses DECL_INITIAL even on optimized away decls,
         very late during compilation.  */
      && (write_symbols != DWARF2_DEBUG && write_symbols !=
VMS_AND_DWARF2_DEBUG)