Register Allocator Problem (was: insn does not satisfy its constraints)

[Boris Boesler]

Hi!

  I have a problem with the register allocator:
1) can't spill
2) places addresses in data-register instead of address-register

  I checked the target macros for registers (section 15.7) and  
register classes (section 15.8) and I dare to say that everything is  
set as it should be. But I get the following error during compiling  
some libs:
../../gcc-4.2.1/gcc/libgcc2.c: In function '__addvsi3':
../../gcc-4.2.1/gcc/libgcc2.c:93: error: unable to find a register to  
spill in class 'Rx_REGS'
../../gcc-4.2.1/gcc/libgcc2.c:93: error: this is the insn:
(insn 12 11 13 2 ../../gcc-4.2.1/gcc/libgcc2.c:87 (parallel [
             (set (reg/v:SI 101 [ w ])
                 (plus:SI (reg/v:SI 105 [ b ])
                     (reg/v:SI 104 [ a ])))
             (clobber (reg:CC 49 CONDSEL))
         ]) 3 {addsi3} (insn_list:REG_DEP_TRUE 6  
(insn_list:REG_DEP_TRUE 7 (nil)))
     (expr_list:REG_UNUSED (reg:CC 49 CONDSEL)
         (nil)))
../../gcc-4.2.1/gcc/libgcc2.c:93: confused by earlier errors, bailing  
out


  I get the same error for int main(void) { return(42); }

main-return-42.c: In function 'main':
main-return-42.c:16: error: unable to find a register to spill in  
class 'Rx_REGS'
main-return-42.c:16: error: this is the insn:
(insn 7 5 8 3 (set (reg:SI 101 [ D.1405 ])
         (const_int 42 [0x2a])) 79 {movsi} (nil)
     (nil))
main-return-42.c:16: confused by earlier errors, bailing out

  So, this has something to do with spill an reload. Is the error  
caused by my "movm" pattern:
(define_insn "mov<mode>"
   [(set (match_operand:ALLMT 0 "nonimmediate_operand" "=   r, r, m, r")
	(match_operand:ALLMT 1 "general_operand"      " <iF>, r, r, m"))]
   ""
   "@
   MV_%M0%M1\t%0, %1
   MV_%M0%M1\t%0, %1
   store_%M1\t%0, %1
   load_%M1\t%0, %1"
   [(set_attr "length" "4,4,4,4")]
)

  Is the general_operand the culprit?

  BTW, if I set some used registers as unused, then this error does  
not happen?! But then addresses are written into data-registers  
instead of address-registers.

  @Rask: Are you the only one familiar with GCC's register allocator?

Thanks,
Boris