Re: [avr-gcc-list] Stack usage under heavy inlining

[Paulo Marques]

Hi, Eric

Weddington, Eric wrote:
> [...]
> That's funny. I agree with Linus:
> "Gcc inlining is a total and utter pile of shit."
> At least it seems that way for the AVR.
>
> GCC has so many compiler switches just for inlining, coupled with
> confusing heuristics, it makes it difficult to understand what is
> happening with inlining.

Yes, gcc's switches and heuristics can be confusing, but sometimes it 
can produce some amazing results. Just look at the function I usually 
use to initialize a serial port:

> void init_serial(uint32_t baud_rate)
> {
>         uint32_t ef_baud1, ef_baud2;
>         uint16_t baud_sel1, baud_sel2;
>         WORD baud_sel;
>
>         // this code should be all optimized away when compiling with "-combine 
> -fwhole-program"
>         baud_sel1 = ((uint16_t)(((double)F_CPU/((double)(baud_rate)*16l)-0.5)));
>         ef_baud1 = ((uint32_t)(((double)F_CPU/((double)(baud_sel1 + 
> 1)*16l)+0.5)));
>         baud_sel2 = ((uint16_t)(((double)F_CPU/((double)(baud_rate)*8l)-0.5)));
>         ef_baud2 = ((uint32_t)(((double)F_CPU/((double)(baud_sel2 + 
> 1)*8l)+0.5)));
>         // low-speed can never be _more_ accurate than high-speed. We select it 
> only
>         // when it has the _same_ accuracy to improve robustness
>         if (ef_baud1 == ef_baud2) {
>                 UCSR0A = 0;
>                 baud_sel.w = baud_sel1;
>         } else {
>                 UCSR0A = bit(U2X0);
>                 baud_sel.w = baud_sel2;
>         }
>
>         // enable RxD / TxD and ints
>         UCSR0B = bit(RXCIE0) | bit(RXEN0) | bit(TXEN0);
>         // set baud rate
>         UBRR0H = baud_sel.b.h & 0xF;
>         UBRR0L = baud_sel.b.l;
> }

Calling it with something like "init_serial(115200)" will 
"auto-magically" produce something like half a dozen avr instructions 
that initialize the serial port with the best divider / double speed 
setting, all calculated at compile time.

If you use "-combine -fwhole-program" the call site might even be on a 
different source file.

> > I can try with more recent versions, if someone suspects that this 
> > behavior is better in a more recent gcc, but I just wanted to 
> > warn other 
> > developers who might hit the same problem.
>
> I don't think the behavior has changed in recent versions.

That's what I was afraid, too :(

> Interestingly, the lkml post mentions a switch:
> -fno-default-inline
>
> I wonder how that would work out for the AVR. Maybe that should be
> considered for the AVR default specs, or at least for the WinAVR
> Makefile Template and the default AVR Studio Makefile.

I didn't try that one, but interestingly enough, if I compile the 
project with "-combine -fwhole-program -fno-inline", gcc produces a 
single version of the init_serial function optimized to the called baud 
rate _without_ inlining it.

--
Paulo Marques
Software Development Department - Grupo PIE, S.A.
Phone: +351 252 290600, Fax: +351 252 290601
Web: www.grupopie.com

"C++ : increment the value of C and use the old one"