Re: [avr-gcc-list] a newbie's question

[TeeN]

On 7 Jun 2002 at 9:06, Bruce D. Lightner wrote:

> Dear ???:
> 
> > Hi Guru,
> > 
> > I have used the following CFLAGS to make the gcctest8.c:
> > 
> > CFLAGS =  -O1 -Wall -Wstrict-prototypes -Wa,-ahlms=$(<:.c=.lst)
> > 
> > the original C code:
> > 
> > for (j=0; j<255; j++)
> >     k++;
> > 
> > generates:
> > 
> >                  ldi     r24, 0          ; for (j=0;
> >  loc_5F:                                 ;
> >                  subi    r24, -1         ; ;j++)
> >                  cpi     r24, -1
> >                  brcs    loc_5F          ; r24 == 0x00-0xFE : loop
> >                                            r24 == 0xFF : fall thru
> > 
> > 1. k++ line is optimised. it is ok.
> > 2. since the "subi r24,0xFF" has increased r24 before the
> > comparision, the loop actually executed 253 times instead of 254
> > times.
> > 
> > if i use -O3 to optimise it, the disassembled code is even funny. Is
> > it something wrong with my system ?
> > thanks!
> 
> This is a common misunderstanding of "avr-gcc" and it's optimizer...
> 
> You didn't provide the entire program, but I'll bet that you don't have "k"
> declared as:
> 
>    volatile int k;
> 
> and "k" is not a returned parameter.
> 
> If you do *not* declare "k" as volatile, then the "gcc" optimizer is free to
> do whatever it wants with your for-loop and what is inside it.  What is
> usually does in this case seldom makes sense to us humans! :-)
> 
> Best regards,
> 
> Bruce
> 
> P.S.- Next time provide the entire program...and your name! :-)
> 
> -- 
>  Bruce D. Lightner
>  La Jolla, California
>  Email: address@hidden
>  URL: http://www.lightner.net/lightner/bruce/

Hi Bruce,

Thanks for you tips.

1. that was my first post to this mailing list so i have forgotten to leave my 
name too :p

2. my question is not on the k++ line being optimised. My problem is on the for 
loops.

Here i have included another code fragment below to show my problem. I would 
appreciate if any guru can shed some lights on it.

C code:

typedef unsigned char u08;

u08 func1( void )
{
   u08 i, j, k;

   k = 0;
   for (i=0; i<0xfe; i++)   /* outer delay loop */
           for(j=0; j<0xfe;j++) /* inner delay loop */
               k++;            /* just do something */
   return k;
}

disassembled by avr-objdump -D

00000096 <func1>:
  96:   80 e0           ldi     r24, 0x00                 ; k=0;
  98:   98 2f           mov     r25, r24          ; for (i=0;
  9a:   20 e0           ldi     r18, 0x00                 ;      for (j=0;
  9c:   8e 5f           subi    r24, 0xFE            ;           k += 0x02; 
<--*1
  9e:   2e 5f           subi    r18, 0xFE            ;           j += 0x02; 
<--*2
  a0:   2e 3f           cpi     r18, 0xFE                 ;    if (j<0xfe) loop 
<-*3
  a2:   e0 f3           brcs    0x9c
  a4:   9f 5f           subi    r25, 0xFF           ;  i++;
  a6:   9e 3f           cpi     r25, 0xFE                 ; if (i<0xfe) loop  
<--*4      
  a8:   c0 f3           brcs    0x9a
  aa:   99 27           eor     r25, r25 
  ac:   08 95           ret

problem *1: k++  is compiled to k += 0x02;
problem *2: j++ is compiled to j += 0x02;
problem *3 and *4: in the for (i=0; i<0xfe; i++) code, gcc compiled it into an 
assembled code that first do the increment BEFORE the 
comparision. This creates OBOB (off-by-one bug) -- the loop actually executed 1 
time less than expected (assume that problem *1 
and *2 is not taken into account).

In my last post, i quoted:
> > 2. since the "subi r24,0xFF" has increased r24 before the
> > comparision, the loop actually executed 253 times instead of 254
> > times.
It  should be interpreted as:
>>2. since the "subi r24,0xFF" has increased r24 before the
>>omparision, the loop actually executed "254" times instead of "255"
>>times.
This is the OBOB I am talking about.

my gcc version is: 3.0.2

Any comment will be appreciated. Thanks!
TeeN (This time i remember to leave my name :)

_________________________________________________________
Do You Yahoo!?
Get your free @yahoo.com address at http://mail.yahoo.com

avr-gcc-list at http://avr1.org