REG_LIBCALL noop move problems

[John Wehle <john at feith dot com>]

Compiling gcc.c-torture/unsorted/BUG25.c on HP-PA with
-O2 yields the following bit of rtl after addressof runs:

(insn 11 23 12 (set (reg:SI 26 %r26)
        (reg:SI 26 %r26)) 68 {*pa.md:2088} (nil)
    (insn_list:REG_LIBCALL 14 (nil)))

(call_insn 12 11 14 (parallel[ 
            (set (reg:SI 28 %r28)
                (call (mem:SI (symbol_ref:SI ("@ffs")) 0)
                    (const_int 16 [0x10])))
            (clobber (reg:SI 2 %r2))
            (use (const_int 0 [0x0]))
        ] ) 275 {call_value_internal_symref} (nil)
    (expr_list:REG_EH_REGION (const_int -1 [0xffffffff])
        (nil))
    (expr_list (use (reg:SI 26 %r26))
        (nil)))

(insn 14 12 15 (set (reg:SI 96)
        (reg:SI 28 %r28)) 68 {*pa.md:2088} (nil)
    (insn_list:REG_RETVAL 11 (expr_list:REG_EQUAL (ffs:SI (reg:SI 26 %r26))
            (nil))))

Running life_analysis (in order to setup things need by VRP)
causes insn 11 to be deleted by delete_noop_moves which
orphans the REG_RETVAL note.  Things later splat in propagate_block
as it walks off the beginning of the function due to the REG_RETVAL
note pointing to a deleted insn which is now before basic block 0.

Should delete_noop_moves skip an insn with a REG_LIBCALL note,
simply move the note to the next real insn (adjusting the REG_RETVAL
note), or should the problem be solved in some other fashion?

-- John
-------------------------------------------------------------------------
|   Feith Systems  |   Voice: 1-215-646-8000  |  Email: john@feith.com  |
|    John Wehle    |     Fax: 1-215-540-5495  |                         |
-------------------------------------------------------------------------