19990502-0.f fails at -O2 on x86; reload bug

[Zack Weinberg <zack at rabi dot columbia dot edu>]

19990502-0.f is failing on x86 at -O2 and no other optimization
level.  We abort in print_operand:

$ f771 19990502-0.f -O2 -w -o /tmp/ccxldaaa.s
19990502-0.f: In subroutine `trf2f2':
19990502-0.f:350: Internal compiler error in `print_operand', 
	at config/i386/i386.c:3405

because we tried to emit a hard register reference to %argp.  There is
no such register; reload is supposed to eliminate %argp to %fp or %sp
always.

The offending insn, just before reload, is

(insn 1659 1653 1677 (set (reg:SI 447)
        (const_int 0 [0x0])) 48 {movsi+2} (nil)
    (nil))

Looks perfectly harmless, ne?  But after reload we have

(insn 1659 2138 2141 (set (reg:SI 16 %argp)
        (const_int 0 [0x0])) 48 {movsi+2} (nil)
    (nil))

(insn 2141 1659 1677 (set (mem:SI (plus:SI (reg:SI 6 %ebp)
                (const_int -196 [0xffffff3c])) 0)
        (reg:SI 16 %argp)) 48 {movsi+2} (nil)
    (nil))

The register allocation matrix indicates that reg 447 wound up on the
stack - presumably at (mem:SI (plus:SI (reg:SI 6 %ebp) (const_int -196)). 
Reload has apparently decided that it can use %argp as a scratch
register to copy (const_int 0) onto the stack.

;; Register dispositions:
22 in 0    23 in 0    40 in 11   41 in 13   43 in 14   45 in 12  
47 in 5    51 in 0    52 in 3    68 in 8    79 in 8    91 in 9  
95 in 10   99 in 8    105 in 8   109 in 8   117 in 4   123 in 0  
124 in 9   134 in 8   135 in 8   136 in 9   147 in 8   157 in 8  
160 in 9   164 in 10  168 in 8   174 in 8   178 in 8   187 in 8  
188 in 10  190 in 8   193 in 9   194 in 8   197 in 8   200 in 8  
205 in 8   209 in 8   211 in 8   215 in 0   216 in 3   232 in 8  
243 in 8   262 in 8   266 in 8   267 in 10  270 in 9   271 in 8  
273 in 8   291 in 10  292 in 11  294 in 15  297 in 9   298 in 8  
314 in 0   315 in 9   325 in 8   326 in 13  327 in 9   338 in 8  
348 in 11  350 in 8   358 in 8   362 in 10  363 in 10  366 in 9  
367 in 8   369 in 9   370 in 8   388 in 10  389 in 11  394 in 9  
395 in 8   410 in 0   417 in 0   418 in 4   423 in 0   424 in 4  
425 in 0   426 in 4   428 in 0   433 in 3   434 in 3   435 in 3  
438 in 3   439 in 4   442 in 5   453 in 5   464 in 3   466 in 4  
469 in 4   471 in 2   472 in 1   476 in 3   479 in 4   480 in 2  
481 in 1   482 in 0  

;; Hard regs used:  0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

The docs imply that hard registers marked in FIXED_REGISTERS may not
be used by register allocation (which I take to include reload), and
argp is so marked, so I'm guessing the bug is there.

zw