new -O0 jump optimization causes SH regressions

[Joern Rennecke <amylaar at cygnus dot co dot uk>]

This change:

        * jump.c (jump_optimize_1): Do simple jump optimizations and
        dead code elimination.

causes a segfault when compiling gcc.c-torture/compile/920917-1.c  -O0 for
the SH targets -m2, -m2 -ml, -m3e, -m3e -ml, -m4, -m4 -ml, -m4-single,
-m4-single -ml, -m4-single-only and -m4-single-only -ml .

Just looking at the -m2 case, before jump we have:

============================================================================
(jump_insn 31 30 32 (parallel[ 
            (set (pc)
                (plus:SI (reg:SI 58)
                    (label_ref 0)))
            (use (label_ref 33))
        ] ) -1 (nil)
    (nil))

(barrier 32 31 33)

(code_label 33 32 34 8 "" "" [num uses: 0])

(jump_insn 34 33 35 (addr_diff_vec:HI (label_ref:SI 33)
        [ 
            (label_ref:SI 16)
            (label_ref:SI 17)
            (label_ref:SI 15)
            (label_ref:SI 15)
            (label_ref:SI 17)
            (label_ref:SI 13)
        ] 
        (const_int 0 [0x0])
        (const_int 0 [0x0])) -1 (nil)
    (nil))

(barrier 35 34 10)

(note 10 35 13 ("/s/egcs-fpscr2/gcc/testsuite/gcc.c-torture/compile/920917-1.c") 1)

(code_label 13 10 14 4 "" "" [num uses: 0])

(code_label 14 13 15 5 "" "" [num uses: 0])

(code_label 15 14 16 6 "" "" [num uses: 0])

(code_label 16 15 17 7 "" "" [num uses: 0])

(code_label 17 16 36 9 "" "" [num uses: 0])

(code_label 36 17 38 3 "" "" [num uses: 0])
============================================================================

And this is changed into:

============================================================================
(jump_insn 31 30 32 (set (pc)
        (label_ref 36)) -1 (nil)
    (nil))

(barrier 32 31 33)

(code_label 33 32 34 8 "" "" [num uses: 1])

(jump_insn 34 33 35 (addr_diff_vec:HI (label_ref:SI 33)
        [ 
            (label_ref:SI 36)
            (label_ref:SI 36)
            (label_ref:SI 36)
            (label_ref:SI 36)
            (label_ref:SI 36)
            (label_ref:SI 36)
        ] 
        (const_int 0 [0x0])
        (const_int 0 [0x0])) -1 (nil)
    (nil))

(barrier 35 34 10)

(note 10 35 36 ("/s/egcs-fpscr2/gcc/testsuite/gcc.c-torture/compile/920917-1.c") 1)

(code_label 36 10 38 3 "" "" [num uses: 8])
============================================================================

Note that the insn using the addr_diff_vec gets changed to not use it any
longer, but the addr_diff_vec is kept around.

In redirect_tablejump, there is this code:

  if (--LABEL_NUSES (olabel) == 0)
    {
      delete_labelref_insn (jump, olabel, 0);
      delete_insn (olabel);
    }

Obviously, this doesn't delete addr_diff_vecs, since their label is
referenced by the addr_diff_vec itself.
Supposedly that was not a problem before because flow removes these
stale addr_diff_vecs, but now that jump optimizations are done even
at -O0, you can no longer rely on that.