[avr-gcc-list] Follow up to AVR-GCC -Os issues

[Nicholas Vinen]

After further investigation I've noticed something odd.

Here is an example of a function which is much smaller with avr-gcc
4.1.2 than 4.3.2:

void calc_bin_values() {
  charge_start_voltage_bin  = decimal_to_binary(charge_start_voltage_dec,  
5000);
  charge_finish_voltage_bin = decimal_to_binary(charge_finish_voltage_dec, 
5000);
  charge_start_current_bin  = decimal_to_binary(charge_start_current_dec,  
1000);
  charge_finish_current_bin = decimal_to_binary(charge_finish_current_dec, 
1000);
}



This is what avr-gcc 4.1.2 produces with -Os -fno-inline-functions:

0000072c <calc_bin_values>:
 72c:   68 e8           ldi     r22, 0x88       ; 136
 72e:   73 e1           ldi     r23, 0x13       ; 19
 730:   80 91 96 01     lds     r24, 0x0196
 734:   90 91 97 01     lds     r25, 0x0197
 738:   14 df           rcall   .-472           ; 0x562 <decimal_to_binary>
 73a:   90 93 a1 01     sts     0x01A1, r25
 73e:   80 93 a0 01     sts     0x01A0, r24
 742:   68 e8           ldi     r22, 0x88       ; 136
 744:   73 e1           ldi     r23, 0x13       ; 19
 746:   80 91 9a 01     lds     r24, 0x019A
 74a:   90 91 9b 01     lds     r25, 0x019B
 74e:   09 df           rcall   .-494           ; 0x562 <decimal_to_binary>
 750:   90 93 9f 01     sts     0x019F, r25
 754:   80 93 9e 01     sts     0x019E, r24
 758:   68 ee           ldi     r22, 0xE8       ; 232
 75a:   73 e0           ldi     r23, 0x03       ; 3
 75c:   80 91 98 01     lds     r24, 0x0198
 760:   90 91 99 01     lds     r25, 0x0199
 764:   fe de           rcall   .-516           ; 0x562 <decimal_to_binary>
 766:   90 93 82 01     sts     0x0182, r25
 76a:   80 93 81 01     sts     0x0181, r24
 76e:   68 ee           ldi     r22, 0xE8       ; 232
 770:   73 e0           ldi     r23, 0x03       ; 3
 772:   80 91 a2 01     lds     r24, 0x01A2
 776:   90 91 a3 01     lds     r25, 0x01A3
 77a:   f3 de           rcall   .-538           ; 0x562 <decimal_to_binary>
 77c:   90 93 ab 01     sts     0x01AB, r25
 780:   80 93 aa 01     sts     0x01AA, r24
 784:   08 95           ret



This seems about what I'd expect. Here's what avr-gcc 4.3.2 products
with the same options:

00000750 <calc_bin_values>:
 750:   60 91 96 01     lds     r22, 0x0196
 754:   70 91 97 01     lds     r23, 0x0197
 758:   80 e0           ldi     r24, 0x00       ; 0
 75a:   90 e0           ldi     r25, 0x00       ; 0
 75c:   26 e0           ldi     r18, 0x06       ; 6
 75e:   34 e0           ldi     r19, 0x04       ; 4
 760:   40 e0           ldi     r20, 0x00       ; 0
 762:   50 e0           ldi     r21, 0x00       ; 0
 764:   b3 d3           rcall   .+1894          ; 0xecc <__mulsi3>
 766:   28 e8           ldi     r18, 0x88       ; 136
 768:   33 e1           ldi     r19, 0x13       ; 19
 76a:   40 e0           ldi     r20, 0x00       ; 0
 76c:   50 e0           ldi     r21, 0x00       ; 0
 76e:   ed d3           rcall   .+2010          ; 0xf4a <__udivmodsi4>
 770:   30 93 a1 01     sts     0x01A1, r19
 774:   20 93 a0 01     sts     0x01A0, r18
 778:   60 91 9a 01     lds     r22, 0x019A
 77c:   70 91 9b 01     lds     r23, 0x019B
 780:   80 e0           ldi     r24, 0x00       ; 0
 782:   90 e0           ldi     r25, 0x00       ; 0
 784:   26 e0           ldi     r18, 0x06       ; 6
 786:   34 e0           ldi     r19, 0x04       ; 4
 788:   40 e0           ldi     r20, 0x00       ; 0
 78a:   50 e0           ldi     r21, 0x00       ; 0
 78c:   9f d3           rcall   .+1854          ; 0xecc <__mulsi3>
 78e:   28 e8           ldi     r18, 0x88       ; 136
 790:   33 e1           ldi     r19, 0x13       ; 19
 792:   40 e0           ldi     r20, 0x00       ; 0
 794:   50 e0           ldi     r21, 0x00       ; 0
 796:   d9 d3           rcall   .+1970          ; 0xf4a <__udivmodsi4>
 798:   30 93 9f 01     sts     0x019F, r19
 79c:   20 93 9e 01     sts     0x019E, r18
 7a0:   60 91 98 01     lds     r22, 0x0198
 7a4:   70 91 99 01     lds     r23, 0x0199
 7a8:   80 e0           ldi     r24, 0x00       ; 0
 7aa:   90 e0           ldi     r25, 0x00       ; 0
 7ac:   26 e0           ldi     r18, 0x06       ; 6
 7ae:   34 e0           ldi     r19, 0x04       ; 4
 7b0:   40 e0           ldi     r20, 0x00       ; 0
 7b2:   50 e0           ldi     r21, 0x00       ; 0
 7b4:   8b d3           rcall   .+1814          ; 0xecc <__mulsi3>
 7b6:   28 ee           ldi     r18, 0xE8       ; 232
 7b8:   33 e0           ldi     r19, 0x03       ; 3
 7ba:   40 e0           ldi     r20, 0x00       ; 0
 7bc:   50 e0           ldi     r21, 0x00       ; 0
 7be:   c5 d3           rcall   .+1930          ; 0xf4a <__udivmodsi4>
 7c0:   30 93 82 01     sts     0x0182, r19
 7c4:   20 93 81 01     sts     0x0181, r18
 7c8:   60 91 a2 01     lds     r22, 0x01A2
 7cc:   70 91 a3 01     lds     r23, 0x01A3
 7d0:   80 e0           ldi     r24, 0x00       ; 0
 7d2:   90 e0           ldi     r25, 0x00       ; 0
 7d4:   26 e0           ldi     r18, 0x06       ; 6
 7d6:   34 e0           ldi     r19, 0x04       ; 4
 7d8:   40 e0           ldi     r20, 0x00       ; 0
 7da:   50 e0           ldi     r21, 0x00       ; 0
 7dc:   77 d3           rcall   .+1774          ; 0xecc <__mulsi3>
 7de:   28 ee           ldi     r18, 0xE8       ; 232
 7e0:   33 e0           ldi     r19, 0x03       ; 3
 7e2:   40 e0           ldi     r20, 0x00       ; 0
 7e4:   50 e0           ldi     r21, 0x00       ; 0
 7e6:   b1 d3           rcall   .+1890          ; 0xf4a <__udivmodsi4>
 7e8:   30 93 ab 01     sts     0x01AB, r19
 7ec:   20 93 aa 01     sts     0x01AA, r18
 7f0:   08 95           ret



Note that it appears to be inlining the subroutine despite
-fno-inline-functions and this is causing it to produce much larger
code. I think -Os should inline functions if doing so reduces code size
- e.g. they are declared inline and only called in one place. This
subroutine (decimal_to_binary) is called in many places and is not
declared inline. Why is avr-gcc 4.3.2 inlining it, especially with -Os?




Nicholas