[Bug middle-end/22051] [4.0/4.1 regression] Wrong code for function pointer comparision during optimization

["danglin at gcc dot gnu dot org" <gcc-bugzilla at gcc dot gnu dot org>]

------- Additional Comments From danglin at gcc dot gnu dot org  2005-06-14 00:14 -------
Here's what we have at the test to see if function pointer canonicalization
should be done in dojump.c:

Breakpoint 1, do_compare_and_jump (exp=0x400ec708, signed_code=NE,
    unsigned_code=NE, if_false_label=0x400ecd70, if_true_label=0x0)
    at ../../gcc/gcc/dojump.c:928
928       if (HAVE_canonicalize_funcptr_for_compare
(gdb) p debug_tree (exp)
 <ne_expr 0x400ec708
    type <boolean_type 0x4000c620 _Bool public unsigned QI
        size <integer_cst 0x400030f0 constant invariant 8>
        unit size <integer_cst 0x40003108 constant invariant 1>
        align 8 symtab 0 alias set -1 precision 1 min <integer_cst 0x400034e0 0> 
max <integer_cst 0x40003510 1>>

    arg 0 <var_decl 0x400ef850 p
        type <pointer_type 0x400174d0 type <void_type 0x40017460 void>
            sizes-gimplified public unsigned SI
            size <integer_cst 0x400032e8 constant invariant 32>
            unit size <integer_cst 0x40003078 constant invariant 4>
            align 32 symtab 0 alias set -1
            pointer_to_this <pointer_type 0x4007fbd0>>
        used unsigned SI file cffc-bug.c line 8 size <integer_cst 0x400032e8 32> 
unit size <integer_cst 0x40003078 4>
        align 32 context <function_decl 0x400ef700 main>
        (reg/v/f:SI 94 [ p ])
        chain <var_decl 0x400ef8c0 result type <integer_type 0x4000c380 int>
            SI file cffc-bug.c line 9 size <integer_cst 0x400032e8 32> unit size 
<integer_cst 0x40003078 4>
            align 32 context <function_decl 0x400ef700 main>>>
    arg 1 <addr_expr 0x400eb820
        type <pointer_type 0x400f3070 type <function_type 0x400ef690>
            unsigned SI size <integer_cst 0x400032e8 32> unit size <integer_cst 
0x40003078 4>
            align 32 symtab 0 alias set -1>
        constant invariant
        arg 0 <function_decl 0x400ef700 main type <function_type 0x400ef690>
            addressable used public static decl_5 SI file cffc-bug.c line 7 
arguments <parm_decl 0x400ef540 argc> result <result_decl 0x400ef770 D.2328> 
initial <block 0x400f0b98>
            (mem:SI (symbol_ref/v:SI ("@main") <function_decl 0x400ef700 main>) 
[0 S4 A32])
            saved-insns 0x400de800>>>
$3 = void

The "(int (*)(int, char **))" cast appears to have been dropped from the
comparison.

I also believe that casts to int types are also being dropped when applied
to function pointer causing the opposite problem (unnecessary canonicalization).

In cffc-bug.c.t19.dce1, we have:

<L1>:;
  p___23_4 = (int (*<T4ef>) (int, char * *)) p_3;
  if (p___23_4 != main) goto <L2>; else goto <L3>;

In cffc-bug.c.t20.dom1, we have:

<L1>:;
  p___23_4 = (int (*<T4ef>) (int, char * *)) p_2;
  if (p_2 != main) goto <L2>; else goto <L3>;


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22051