[avr-gcc-list] bug with bitfields in (volatile) registers?

[Ned Konz]

I was just bit by this surprise with registers suddenly being treated as if 
they were not volatile (in avr-gcc 3.4.3). I'm using a chip with no RAM, so I 
would like the registers to work properly.

// avr-gcc -Os -mmcu=avr2 -mcall-prologues -mint8 -mtiny-stack -save-temps 
-Wall -g -Wa,-almshd=registerbug.lst -Wl,--oformat=ihex -o registerbug.hex 
registerbug.c
// avr-objdump -h -S registerbug.o  > registerbug.lss

typedef struct
{
    unsigned char uartState:4;
    unsigned char unused:4;
} state_t;

register unsigned char c __asm__("r16");
register state_t state __asm__("r3");  // actually volatile...
register unsigned char state2 __asm__("r17");
register unsigned char state3 __asm__("r9");

int main(void)
{
   0: c0 e0        ldi r28, 0x00 ; 0
   2: d0 e0        ldi r29, 0x00 ; 0
   4: de bf        out 0x3e, r29 ; 62
   6: cd bf        out 0x3d, r28 ; 61

// look what's happening here: r3 is getting copied to r24...
   8: 83 2d        mov r24, r3
   a: 8f 70        andi r24, 0x0F ; 15
 for (;;)
 {
// and then the comparison is being done on r24.
// however, meanwhile an interrupt handler is changing r3, and we
// get stuck in this loop forever:
  while (state.uartState == 10)
   c: 8a 30        cpi r24, 0x0A ; 10
   e: f1 f3        breq .-4       ; 0xc
// that should be a branch back to 0x8 instead of 0xc!
   ;

  c = 'A';
  10: 01 e4        ldi r16, 0x41 ; 65

// here everything is fine because we can do a cpi:
  while (state2 == 10)
  12: 1a 30        cpi r17, 0x0A ; 10
  14: f1 f3        breq .-4       ; 0x12
   ;

// and here everything is fine because the generated code
// branches to the right place:
  while (state3 == 10)
  16: 9a e0        ldi r25, 0x0A ; 10
  18: 99 16        cp r9, r25
  1a: c1 f7        brne .-16      ; 0xc
  1c: 9a e0        ldi r25, 0x0A ; 10
  1e: 99 16        cp r9, r25
  20: d1 f3        breq .-12      ; 0x16
  22: f4 cf        rjmp .-24      ; 0xc
   ;
       }
}


Are registers supposed to be treated as if they were volatile?

Is this a bug?

Thanks,
-- 
Ned Konz
http://bike-nomad.com