[avr-gcc-list] Wrong excution order in 4.1.1, but not 3.4.5, regression?

[Bob Paddock]

I'm seeing something in 4.1.1 that is not making
sense.  The code produced is clearly the wrong
order, in 4.1.1.  Code from 3.4.5 is the correct order.

Optimization in both cases is set for -Os;
I'm using the same source code/Makfile and
pointing at different versions of WINAVR,
between 4.1.1 and 3.4.5.

I've looked at the known problem list, but
I could not related this to any of those.
http://www.nongnu.org/avr-libc/bugs.html

I can't seem to get a good cut down example
that produces the error.  The code
effectively looks like this:

static volatile PGM_UP  pattern_flash_ptr_u8_v; /* IRQ updates PORTB */
static volatile PGM_UP  pattern_flash_ptr_reload_u8_v;
[PGM_UP is unsigned version of PGM_P.]

      do{
        pattern_flash_ptr_u8_v = pattern_flash_ptr_reload_u8_v = (PGM_UP)  
pgm_read_word( &pattern_table[ 0 ] );
      }while( mode_button_press_count_u8 < max_button_presses_u8 );

In the real code 'mode_button_press_count_u8' actually could
change its value inside of the do{}while(),
and the pattern_table[ 0 ] index would
be a variable.  I made these changes to the real code,
and the problem persists.

For:

 pattern_flash_ptr_u8_v = pattern_flash_ptr_reload_u8_v = (PGM_UP)  
pgm_read_word( &pattern_table[ 0 ] );

 4.1.1 produces:

 3ca:   e6 e8           ldi     r30, 0x86       ; 134 [Address of pattern_table]
 3cc:   f0 e0           ldi     r31, 0x00       ; 0
 3ce:   c8 95           lpm
 3d0:   c0 2d           mov     r28, r0
 3d2:   31 96           adiw    r30, 0x01       ; 1
 3d4:   c8 95           lpm
 3d6:   d0 2d           mov     r29, r0
 3d8:	0c 2f       	mov	r16, r28  [Missed optimization, but that is  
separate matter for an other time.  r16/r17 was loaded above only to be  
overwritten here, and not needed below.]
 3da:   1d 2f           mov     r17, r29
 3dc:   ff 20           and     r15, r15
 3de:   c9 f6           brne    .-78            ; 0x392 <main+0x74> [Jump to 
do{]
 3e0:	10 93 7b 00 	sts	0x007B, r17  [This is outside of the do{}while() so  
the code fails.]
 3e4:   00 93 7a 00     sts     0x007A, r16
 3e8:   d0 93 7d 00     sts     0x007D, r29
 3ec:   c0 93 7c 00     sts     0x007C, r28


3.4.5 produces this, correct code:

            pattern_flash_ptr_u8_v = pattern_flash_ptr_reload_u8_v =  
(PGM_UP) pgm_read_word( &pattern_table[ 0 ] );
 3ec:   ef e3           ldi     r30, 0x3F       ; 63  [Address of pattern_table]
 3ee:   f0 e0           ldi     r31, 0x00       ; 0
 3f0:   c8 95           lpm
 3f2:   80 2d           mov     r24, r0
 3f4:   31 96           adiw    r30, 0x01       ; 1
 3f6:   c8 95           lpm
 3f8:   90 2d           mov     r25, r0
 3fa:	90 93 80 00 	sts	0x0080, r25 [This is inside the do{}while(), code  
works.  Doesn't have optimization issues either.]
 3fe:   80 93 7f 00     sts     0x007F, r24
 402:   90 93 7e 00     sts     0x007E, r25
 406:   80 93 7d 00     sts     0x007D, r24
          }

      }while( mode_button_press_count_u8 < max_button_presses_u8 );
 40a:   90 91 79 00     lds     r25, 0x0079
 40e:   80 91 60 00     lds     r24, 0x0060
 412:   98 17           cp      r25, r24
 414:   08 f4           brcc    .+2             ; 0x418 <main+0xdc> [Exit 
do{}while]
 416:   c5 cf           rjmp    .-118           ; 0x3a2 <main+0x66> [Jump to 
do{]