[Bug optimization/12260] [3.4 Regression] ICE in output_operand: invalid expression as operand

[falk at debian dot org]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=12260


falk at debian dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|                            |1


------- Additional Comments From falk at debian dot org  2003-09-23 21:11 -------
The problem is caused by output_operand getting passed this rtx:

(neg:DI (const:DI (plus:DI (symbol_ref:DI ("buf")
					  [flags 0x2]
					  <var_decl 0x200002a6680 buf>)
			   (const_int -1 [0xffffffffffffffff]))))

which it doesn't know how to handle. It originates from gcse, which adds to this
insn:

(insn 34 33 35 2
      (set (reg:DI 79)
	   (minus:DI (reg:DI 80)
		     (reg/v/f:DI 72 [ op ]))) 28 {subdi3} (nil)
		     (nil))

this note:

(neg:DI (const:DI (plus:DI (symbol_ref:DI ("buf")
					  [flags 0x2]
					  <var_decl 0x200002a6680 buf>)
			   (const_int -1 [0xffffffffffffffff]))))

It seems that gcc decided (op - buf) + 1 would be better calculated as 
op + (-buf - 1). Phil's regression hunter says:

Search converges between 2003-09-11-trunk (#398) and 2003-09-12-trunk (#399).

I suspect this patch exposed the problem:
2003-09-10  Roger Sayle  <roger@eyesopen.com>

	* combine.c (combine_simplify_rtx): Move several NOT and NEG
	optimizations from here...
	* simplify-rtx.c (simplify_unary_operation): to here.  Recursively
	simplify expressions using simplify_gen_*ary instead of gen_rtx_*.

The transformation seems valid, so I guess a simple solution would be to
have output_operand handle (neg x) just like (0 - x). I don't know whether
all assemblers would like that, though...