Problem with GP on IA64

[kenner at vlsi1 dot ultra dot nyu dot edu (Richard Kenner)]

When compiling the GNAT file make.adb with -gnatpga -O2, the following is 
in .sibling:

(call_insn 2695 2694 2697 (parallel[ 
            (set (reg:TI 8 r8)
                (call (mem:DI (symbol_ref/v:DI ("osint__read_library_info")) [0 S8 A64])
                    (const_int 2 [0x2])))
            (use (unspec[ 
                        (reg:DI 1 r1)
                    ]  9))
            (clobber (reg:DI 320 b0))
        ] ) -1 (nil)
    (nil)
    (expr_list (use (reg:QI 121 out1))
        (expr_list (use (reg:SI 120 out0))
            (nil))))

(insn 2697 2695 2699 (set (reg:DI 1 r1)
        (reg:DI 458)) -1 (nil)
    (nil))

...

(insn 2705 2702 2707 (parallel[ 
            (set (reg/f:DI 1160)
                (symbol_ref:DI ("opt__check_object_consistency")))
            (clobber (reg:DI 1161))
            (use (reg:DI 1 r1))
        ] ) -1 (nil)
    (expr_list:REG_EQUAL (symbol_ref:DI ("opt__check_object_consistency"))
        (nil)))

(insn 2707 2705 2708 (set (mem/f:QI (reg/f:DI 1160) [2 opt__check_object_consistency+0 S1 A8])
        (subreg/s/u:QI (reg/v:SI 1133) 0)) -1 (nil)
    (nil))


After CSE, that last insn is:

(insn 2707 2702 2708 (set (mem/f:QI (reg/f:DI 1134) [2 opt__check_object_consistency+0 S1 A8])
        (subreg/s/u:QI (reg/v:SI 1133) 0)) 2 {*movqi_internal} (nil)
    (nil))

Later, it gets a REG_EQUIV note for the MEM and is scheduled before 2697
since there is no reason to do otherwise.

Now reg 1134 doesn't get a pseudo, so reload tries to do the reload
for it by loading the address via GP.  But GP hasn't been restored yet!

So it blows up.

How is this supposed to work?