[Bug tree-optimization/28894] New: Optimizaition on AVR target breaks code.

[Bug tree-optimization/28894] New: Optimizaition on AVR target breaks code.

[ralf-engels at gmx dot de]

Attached code file is compiled.
Here is the critical passage:

inline void taskSignal(void) 
{
  unsigned char i;

  PORTB |= _BV(PB4);
  PORTB |= _BV(PB5);
  PORTB |= _BV(PB6);


  i=0;
  do 
    {
      i+=2;
    } 
  while( i<245 );

  i=0;
  do 
    {
      i+=3;
    } 
  while( i<245 );

  PORTB &= ~_BV(PB4);

  i=pos1;
  do 
    {
      i+=1;
    } 
  while( i<245 );


  PORTB &= ~_BV(PB5);
  PORTB &= ~_BV(PB6);
}  

When compiling without optimization I get:

  do 
    {
      i+=2;
 41e:   89 81           ldd     r24, Y+1        ; 0x01
 420:   8e 5f           subi    r24, 0xFE       ; 254
 422:   89 83           std     Y+1, r24        ; 0x01
    } 
 424:   89 81           ldd     r24, Y+1        ; 0x01
 426:   85 3f           cpi     r24, 0xF5       ; 245
 428:   08 f4           brcc    .+2             ; 0x42c <taskSignal+0x42>
 42a:   f9 cf           rjmp    .-14            ; 0x41e <taskSignal+0x34>
  while( i<245 );


But using -Os results in the following code:

 262:   80 e0           ldi     r24, 0x00       ; 0
  do 
    {
      i+=2;
 264:   8a 5f           subi    r24, 0xFA       ; 250
    } 
 266:   85 3f           cpi     r24, 0xF5       ; 245
 268:   e8 f3           brcs    .-6             ; 0x264 <taskSignal+0x8>
  while( i<245 );

Please note how the  substraction of -2 is optimized in something other.
The compiler said:
avr-gcc (GCC) 3.4.6 (Gentoo 3.4.6, ssp-3.4.5-1.0, pie-8.7.9)
Copyright (C) 2006 Free Software Foundation, Inc.

The line calling it was: 
avr-gcc -c -mmcu=attiny26 -I. -g -DF_CPU=8000000UL   -Os -Wall
-Wstrict-prototypes -Wa,-adhlns=tester.lst   -MD -MP -MF .dep/tester.o.d
tester.c -o tester.o


-- 
           Summary: Optimizaition on AVR target breaks code.
           Product: gcc
           Version: 3.4.6
            Status: UNCONFIRMED
          Severity: critical
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: ralf-engels at gmx dot de
GCC target triplet: avr


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28894

[Bug tree-optimization/28894] Optimizaition on AVR target breaks code.

[ralf-engels at gmx dot de]

------- Comment #1 from ralf-engels at gmx dot de  2006-08-29 22:30 -------
Created an attachment (id=12150)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=12150&action=view)
source code which is compiled strangely.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28894

[Bug target/28894] Optimizaition on AVR target breaks code.

[pinskia at gcc dot gnu dot org]

-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|critical                    |normal
          Component|tree-optimization           |target


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28894

[Bug target/28894] Optimizaition on AVR target breaks code.

[pinskia at gcc dot gnu dot org]

------- Comment #2 from pinskia at gcc dot gnu dot org  2006-08-29 22:34 -------
 264:   8a 5f           subi    r24, 0xFA       ; 250

That looks fine for me since it is r24 = r24 - -2.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28894

[Bug target/28894] Optimizaition on AVR target breaks code.

[eweddington at cso dot atmel dot com]

------- Comment #3 from eweddington at cso dot atmel dot com  2006-08-30 15:06 -------
The AVR does not have an Add Immediate instruction (addi), so this is normally
done using sbi with a negative number as Andrew correctly points out.

In Ralf's unoptimized output, it correctly shows a -2 (0xFE):

41e:   89 81           ldd     r24, Y+1        ; 0x01
420:   8e 5f           subi    r24, 0xFE       ; 254

The optimized output shows a -6 (0xFA)

264:   8a 5f           subi    r24, 0xFA       ; 250

Why is this? Well, the optimizer aggressively optimizes loops that do nothing,
like the code is here:

i=0;
do 
  {
    i+=2;
  } 
while( i<245 );

Note that you only have 'i' declared as unsigned char.

The first FAQ entry in the avr-libc user manual talks about using the keyword
'volatile':
http://www.nongnu.org/avr-libc/user-manual/FAQ.html#faq_volatile

If you define 'i' as volatile unsigned char, then the optimizer will not
agressively optimize your "do-nothing" loop. In fact, if you are trying to
achieve a delay using a "do-nothing" loop, then you are strongly suggested to
use the <util/delay.h> header that comes with avr-libc, which are "busy-wait
delay loops":
<http://www.nongnu.org/avr-libc/user-manual/group__util__delay.html>

GCC Bugmasters:
Please close this bug as invalid.

Eric Weddington


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28894

[Bug target/28894] Optimizaition on AVR target breaks code.

[pinskia at gcc dot gnu dot org]

------- Comment #4 from pinskia at gcc dot gnu dot org  2006-08-30 15:09 -------
This is just unrolling/removing empty loops so invalid.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |INVALID


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28894

[Bug target/28894] Optimizaition on AVR target breaks code.

[ralf-engels at gmx dot de]

------- Comment #5 from ralf-engels at gmx dot de  2006-08-30 20:43 -------
Wow, 
you mean that unrolling the loop three times but not removing it is the right
behaviour?

Anyway, 4.1.1 seems to do it right. It's removing the whole loop as expected.

delay.h is working.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28894