Re: Breakage in flow with table jumps

[Franz Sirl <Franz dot Sirl-kernel at lauterbach dot com>]

At 20:04 20.06.2001, Andrew Haley wrote:
>I've been trying to sort out a problem in ARM gcc, with a testcase
>that looks like this:
>
>int SQLSetConnectOption( void * hDrvDbc,
>                         unsigned nOption,
>                         unsigned long vParam)
>{
>
>   if ( hDrvDbc == 0 )
>     return (-2);
>
>   switch ( nOption )
>     {
>     case 106:
>     case 107:
>
>     case 110:
>     case 104:
>       switch ( vParam )
>         {
>         case 1UL:
>         case 0UL:
>         default:
>           ;
>         }
>       break;
>
>     case 105:
>     case 101:
>     case 102:
>     default:
>       ;
>     }
>
>
>   printf( "SQLSetConnectOption.c", "SQLSetConnectOption.c", 49, 1, 1, 
> "SQL_SUCCESS_WITH_INFO Function not fully implemented" );
>   return 1;
>}
>
>When compiled with -fpic -O2, the compiler aborts in flow when trying
>to merge a block that ends like this:
>
>(jump_insn 52 51 53 (parallel[
>             (set (pc)
>                 (if_then_else (leu (reg:SI 3 r3 [37])
>                         (const_int 9 [0x9]))
>                     (mem:SI (plus:SI (mult:SI (reg:SI 3 r3 [37])
>                                 (const_int 4 [0x4]))
>                             (label_ref 53)) 0)
>                     (label_ref 56)))
>             (clobber (reg:CC 24 cc))
>             (use (label_ref 53))
>         ] ) 263 {casesi_internal} (insn_list 51 (nil))
>     (nil))
>
>and a table like this:
>
>(jump_insn 54 15 55 (addr_diff_vec:SI (label_ref:SI 53)
>         [
>             (label_ref:SI 56)
>             (label_ref:SI 56)
>             (label_ref:SI 56)
>             (label_ref:SI 56)
>             (label_ref:SI 56)
>             (label_ref:SI 56)
>             (label_ref:SI 56)
>             (label_ref:SI 56)
>             (label_ref:SI 56)
>             (label_ref:SI 56)
>         ]
>         (const_int 0 [0x0])
>         (const_int 0 [0x0])) -1 (nil)
>     (nil))
>
>with a block that begins like this:
>
>(code_label 56 43 101 3 "" "" [11 uses])
>
>The compiler quite properly notices that the basic blocks have only a
>single edge; the jump table and the if-then-else in insn 52 all have
>label 56 as their destination.  However, the merge fails when we get
>to merge_blocks_move_predecessor_nojumps:
>
>   barrier = next_nonnote_insn (end);
>   if (GET_CODE (barrier) != BARRIER)
>     abort ();
>
>and the compiler aborts.
>
>I could just replace
>
>   if (GET_CODE (barrier) != BARRIER)
>     abort ();
>
>with
>
>   if (GET_CODE (barrier) != BARRIER)
>     return 0;
>
>which seems to work OK.
>
>Thoughts?

Is this code from the unixODBC package? It seems I stumbled over the same 
failure yesterday on ppc-linux then.

Franz.