Re: [avr-gcc-list] switch/case optimisation in TWI ISR

[Vincent Trouilliez]

On Fri, 2005-12-23 at 17:17 +0100, David Bourgeois wrote:
> I have a TWI interrupt routine inspired from the proycon avrlib and which  
> handles multi-master transmissions in an atmega48.
> It's basically a sequence of cases for every value of the twi status.
> 
> Right now, my twi code is 780 bytes long but 292 are for the SWITCH alone,  
> while only 266 are for what I'm actually doing in the cases.
> 
> Is there a better way to write this? I've been told that a if/else if  
> sequence could be more efficient but can't really understand why. I guess  
> this also depends on the compiler. What about AVR-GCC in such a case?
> And even if I have to write all this ISR in assembler, could I be able to  
> reduce this code significantly?


If you are willing to write assembly, then there is room for significant
improvement I think:


PUSH TWI_STATUS_REG
RETURN

Two instructions, hard to beat ;-)
That was for an Intel 8051 but surely it can be adapted for an AVR ???

In the 8051 the I2C status register's lower 3 bits are unused, if in the
above code, so you have 8 bytes to deal with each status case.

Basically it's a look-up table with the status byte as the index. Stated
like this, you can now translate it into C statements: write separate
functions for all the various cases, put there address in a table of
pointers, and use the TWI status byte as the table index to retrieve the
matching function pointer, then jump to the required function. Done.
Obviously, the more cases you have to process, the more
efficient/interesting the look-up table becomes. Doing it for only two
cases would be a tad overkill ;-P

Sounds complex but it should still occupy much less than 292 I would
think !!


HTH,


--
Vince