Re: [avr-gcc-list] dev and mod may not be optimized

[Paulo Marques]

Paulo Marques wrote:
> [...]
>
> a = (c * 65534) >> 16;
> b = c - (a * 10);
>
> leaving you with no divisions to perform.
>
> The main problem is that the first multiplication should be done with 
> some hand optimized assembly to perform a 8x16 bit multiplication, 
> leaving just the upper 8 bits of a 24 bit result, but it should be 
> doable with just 2 mul instructions and one 16 bit addition.

Actually after a few more digging, we can do this all in plain C, by 
using a smaller reciprocal multiplier:

unsigned char a,b,c;
unsigned int r;

r = c * 205;
a = r / 2048;
b = c - (a * 10);

generates this code with avr-gcc 4.2.1:

        r = c * 205;
        a = r / 2048;
  d8:   2d ec           ldi     r18, 0xCD       ; 205
  da:   82 9f           mul     r24, r18
  dc:   90 01           movw    r18, r0
  de:   11 24           eor     r1, r1
  e0:   23 2f           mov     r18, r19
  e2:   33 27           eor     r19, r19
  e4:   26 95           lsr     r18
  e6:   26 95           lsr     r18
  e8:   26 95           lsr     r18
        b = c - (a * 10);
  ea:   9a e0           ldi     r25, 0x0A       ; 10
  ec:   92 9f           mul     r25, r18
  ee:   90 2d           mov     r25, r0
  f0:   11 24           eor     r1, r1
  f2:   89 1b           sub     r24, r25

This is slightly larger than 22 bytes (28 bytes), but it should execute 
in just 16 cycles. We could certainly improve the size a little with 
hand optimized assembly, though.

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

"God is real, unless declared integer."