[avr-gcc-list] Code Optimisation question re: volatile

[Reza Naima]

This really isn't a critical question, but I'm curious... I want a 2
pins (0, 1) to toggle on PORTB every time an interrupt it called.  This
is the most optimised code I've come up with : 

pwm_state = (PORTB = pwm_state) ^ _b00000011;

now, with -Os (size) optimisation, I get : 

 1325 053c 8091 0000            lds r24,pwm_state
 1326 0540 88BB                 out 56-0x20,r24
 1327 0542 88B3                 in r24,56-0x20
 1328 0544 93E0                 ldi r25,lo8(3)
 1329 0546 8927                 eor r24,r25
 1330 0548 8093 0000            sts pwm_state,r24

with -O3 optimisation, I get : 

 959 03c2 4091 0000             lds r20,pwm_state
 960 03c6 48BB                  out 56-0x20,r20
 961 03c8 28B3                  in r18,56-0x20
 962 03ca 33E0                  ldi r19,lo8(3)
 963 03cc 2327                  eor r18,r19
 964 03ce 2093 0000             sts pwm_state,r18

It seems to me that the ideal code would be :

 1325 053c 8091 0000            lds r24,pwm_state
 1326 0540 88BB                 out 56-0x20,r24
 1328 0544 93E0                 ldi r25,lo8(3)
 1329 0546 8927                 eor r24,r25
 1330 0548 8093 0000            sts pwm_state,r24

But it's not doing that because PORTB is defined as volatile.  Is there
any way to temporarily get gcc to ignore the volatile flag for one
subroutine?  I'm going to increase the xtal speed and see if that fixes
my problem :/  If it don't work, use a bigger hammer.  I might try to do
inline assembly if that fails... Is there any more efficient way of
doing it?  This is on an atmega16, btw.. Oh, also, the prolog and
epilogue are pretty big - would inline assembly get rid of that?

thanks,
Reza