[avr-gcc-list] Re: On the subject of lean ISR's

[David Brown]

Dennis Clark wrote:
> Rather than hijack a thread I'll start my own.
>
> How does everyone lean out an ISR?  I too have been appalled at the
> preamble/postamble code GCC puts into the ISR.
>
> I always put ISR variables in global space to avoid stack issues, but the
> heap access registers still need push/popped.  I tried using register
> variables to avoid heap access and found that there is no way to know if a
> library call is using that register - The compiler does NOT guarantee
> exclusive access.  Can you actually get away with a "NAKID" attribute on
> the ISR?
>
> How have others dealt with ISR bloat reduction?
>
> Many thanks,
> DLC

Use local variables rather than global ones where you can - it is always 
more efficient.

The biggest cost to the ISR preamble and postamble comes if the ISR 
makes a function call - all registers that the function *might* use have 
to be stacked for safety.  So in general, avoid function calls (inlined 
functions are fine).

However, if you have a very complex ISR (which is generally a bad idea - 
it's better to keep ISR's short and simple) so that your preamble is 
already fairly long, then any rarely-executed sequences can be put in a 
function (perhaps declared "noinline"), so that any additional pushing 
and popping is only done when that function actually runs.