[Bug debug/34037] New: [4.1/4.2/4.3 Regression] Bounds for VLAs not emitted into debuginfo

[jakub at gcc dot gnu dot org]

With -g -O0 -dA
void bar (char *, char *, char *, int size);
void foo (int size)
{
  char temp[size];
  char temp3[48];
  temp[size-1] = '\0';
  {
    char temp2[size];
    bar (temp, temp2, temp3, size);
  }
};

in 3.4.x we got:
...
        .uleb128 0x4    # (DIE (0x74) DW_TAG_variable)
        .byte   0x1     # DW_AT_artificial
        .long   0x104   # DW_AT_type
        .byte   0x6     # DW_AT_location
        .byte   0x91    # DW_OP_fbreg
        .sleb128 -20
        .byte   0x94    # DW_OP_deref_size
        .byte   0x4
        .byte   0x31    # DW_OP_lit1
        .byte   0x1c    # DW_OP_minus
...
        .uleb128 0x8    # (DIE (0xdc) DW_TAG_array_type)
        .long   0xef    # DW_AT_sibling
        .long   0x115   # DW_AT_type
        .uleb128 0x9    # (DIE (0xe5) DW_TAG_subrange_type)
        .long   0xef    # DW_AT_type
        .long   0x74    # DW_AT_upper_bound
        .byte   0x0     # end of children of DIE 0xdc

but 4.1/4.2/4.3 only have DW_AT_upper_bound for temp3 array where it is
constant.

There seem to be 2 problems.  When gimplifying, the vars are gimplified into
temp vars in gimplify_type_sizes and as those vars are DECL_ARTIFICIAL and
DECL_IGNORED_P, they are usually even at -O0 just deleted as trivially dead
e.g. by CSE.  This part could be fixed say by:
--- gcc/gimplify.c.jj   2007-10-30 11:46:29.000000000 +0100
+++ gcc/gimplify.c      2007-11-08 22:50:35.000000000 +0100
@@ -6171,6 +6171,18 @@ gimplify_type_sizes (tree type, tree *li
       /* These types may not have declarations, so handle them here.  */
       gimplify_type_sizes (TREE_TYPE (type), list_p);
       gimplify_type_sizes (TYPE_DOMAIN (type), list_p);
+      /* When not optimizing, ensure VLA bounds aren't removed.  */
+      if (!optimize
+         && TYPE_DOMAIN (type)
+         && INTEGRAL_TYPE_P (TYPE_DOMAIN (type)))
+       {
+         t = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
+         if (t && TREE_CODE (t) == VAR_DECL && DECL_ARTIFICIAL (t))
+           DECL_IGNORED_P (t) = 0;
+         t = TYPE_MAX_VALUE (TYPE_DOMAIN (type));
+         if (t && TREE_CODE (t) == VAR_DECL && DECL_ARTIFICIAL (t))
+           DECL_IGNORED_P (t) = 0;
+       }
       break;

     case RECORD_TYPE:
The next problem is on the dwarf2out.c side, in add_bound_info we have:
        dw_die_ref decl_die = lookup_decl_die (bound);

        /* ??? Can this happen, or should the variable have been bound
           first?  Probably it can, since I imagine that we try to create
           the types of parameters in the order in which they exist in
           the list, and won't have created a forward reference to a
           later parameter.  */
        if (decl_die != NULL)
          add_AT_die_ref (subrange_die, bound_attr, decl_die);

It seems most even non-artificial VAR_DECLs aren't even
equate_decl_number_to_die'ed, so lookup_decl_die won't find them anyway.
Ideas?  For -O1 and above, I guess we need to wait for Alex or some alternate
representation which would allow us to see how can they be computed even if
they were optimized out.


-- 
           Summary: [4.1/4.2/4.3 Regression] Bounds for VLAs not emitted
                    into debuginfo
           Product: gcc
           Version: 4.1.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: debug
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: jakub at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34037