Re: AW: [avr-gcc-list] avr-gcc bug: global register variable isbroken

[Klaus Rudolph]

Bernard Fouché schrieb:

> Is it normal that 'y' being a long, it is given to f1() even for 'g' and 'h'
> parameters?
>  

It is normal and must automaticly casted to target size. This is not 
normal code which I use :-)))
This is only an example which shows that the global register definition 
is broken.

> Maybe there is a way to avoid using particular registers for your
> application? If the code that you show is like the one you use in your
> interrupt handler and if you want to use a specific register because of
> performance, calling a function from an interrupt handler (and a function
> with so much parameters than many registers will be touched or the stack
> frame will have to be calculatedd) will use code space and processing time.
>  
I would not use such functions in my interrupt handler. But it is 
possible that such functions appear in
the "normal" non interrupt code. And if this code breaks my registers it 
is simply wrong.


> If the problem is not performance but only to call some particular function
> requiring specific register usage, why not write a few assembly lines to
> save/restore the corresponding registers?
>
>  
This needs time! Lots of time for a really frequent called irq-handler. 
And in my case, the irq handler
do the most work and the "application level" must do nothing fast and it 
is acceptable that there are only
10 registers which are used in the "main" part of my software. OK, this 
is maybe a hard requirement, but
for other targets this is possible  and I think it makes no sense to 
discuss if this makes sense or not.
I think if the compiler give this options we should look that this 
options could be used. If not, we should
give an error message or remove this options at all.

I will not continue discussions about what else could be done. I simply 
hope that
it will be accepted that this is a bug and I hope this bug could be 
fixed and!!!! the fix goes
to the sources :-) If it will be accepted that this is a bug I am maybe 
able to look for it.
I fixed "my" binutils problem last week and maybe I also could fix this. 
But it make no sense
to spend a lot of time if the potential pact will never used anywhere.

regards
  Klaus