insn does not satisfy its constraints

[Boris Boesler]

Hi!

  I'm working on a gcc (4.2.1) back-end for a new architecture (no  
silicon yet). I started with the CRX back-end, because it was the  
smallest one. Btw, yes, I've read "Porting GCC for dunces" and some  
parts of the internals document ;-)

  Somehow this seems to be a FAQ, but I can't find the correct answer.

  The architecture has data and address registers (ala m68k) and for  
the C code
int c;
void f(void){c=4711;}

  I get the code:
         MV_WI32 R0, #_c         /* movsi Ax */
         MV_WI32 R1, #4711       /* movsi Ax */
         MV_WW   A1, R0  	/* movsi r */
         store_D 0(A1), R1       /* movsi */

  The very first line is wrong, because there is a register  
constraint for address registers:
(define_insn "mov<mode>"
   [(set (match_operand:ALLMT 0 "nonimmediate_operand" "=r,   Ax,    
Rx, m, r")
	(match_operand:ALLMT 1 "general_operand"      " r, <iF>, <iF>, r, m"))]
   ""
   "@
   MV_%M0%M1\t%0, %1 \t/* mov<mode> r */
   MV_%M0%M1\t%0, %1 \t/* mov<mode> Ax */
   MV_%M0%M1\t%0, %1 \t/* mov<mode> Rx */
   store_%M0\t%0, %1 \t/* mov<mode> */
   load_%M0\t%0, %1 \t/* mov<mode> */"
   [(set_attr "length" "4,4,4,4,4")]
)

  The register sets Ax and Rx do not share any registers! Hm, I can  
only think of placing the symbol in R0 with the third alternative and  
then move again with the first alternative?!

  This strange thing goes on:
int a, b, c;
void add_var(){c = a + b;}

  results in:

add_var.c: In function 'add_var':
add_var.c:20: error: insn does not satisfy its constraints:
(insn 8 24 9 (set (reg:SI 1 R1 [orig:103 a.0 ] [103])
         (mem/c/i:SI (reg:SI 2 R2) [0 a+0 S4 A32])) 79 {movsi} (nil)
     (nil))
add_var.c:20: internal compiler error: in final_scan_insn, at final.c: 
2382

  The error happens, because GO_IF_LEGITIMATE_ADDRESS (strict  
version) detects, that a data register is used like an address  
register - that's illegal! (BASE_REG_CLASS is defined to Ax_REGS)

  So, where can I start my search to resolve this?

Thanks in advance,
Boris Boesler