Re: [avr-gcc-list] switch/case optimisation in TWI ISR

[Paulo Marques]

James Washer wrote:
> In GCC, it's legal to have an array of gotos, so this should work...
> and save a lot of space (and execution too, I would think)
>
> void *jumptable[]={ &&label0, &&label1, &&label31};
>
> goto *jumptable[TWSR>>3];
>
> label0:
>         //code for state 0
>         return;
> label1:
>         //code for state 1
>         return;
> label31:
>         //code for state31
>         return;

This is some manual optimization that the compiler should do for you.

I've just tried this in one of the latest gcc 4.2 and it indeed does 
this as expected.

This code fragment:

> void test(unsigned char a)
> {
>         switch (a) {
>                 case 0: PORTB = 12; break;
>                 case 1: PORTB = 72; break;
>                 case 2: PORTB = 25; break;
>                 case 3: PORTB = 17; break;
>                 case 4: PORTB = 19; break;
>                 case 5: PORTB = 29; break;
>                 case 6: PORTB = 54; break;
>                 case 7: PORTB = 23; break;
>         }
> }

is translated into:

> 00000026 <__ctors_end>:
>   26:   2a c0           rjmp    .+84            ; 0x7c <test+0x10>
>   28:   2b c0           rjmp    .+86            ; 0x80 <test+0x14>
>   2a:   2c c0           rjmp    .+88            ; 0x84 <test+0x18>
>   2c:   2d c0           rjmp    .+90            ; 0x88 <test+0x1c>
>   2e:   2e c0           rjmp    .+92            ; 0x8c <test+0x20>
>   30:   2f c0           rjmp    .+94            ; 0x90 <test+0x24>
>   32:   30 c0           rjmp    .+96            ; 0x94 <test+0x28>
>   34:   32 c0           rjmp    .+100           ; 0x9a <test+0x2e>

> 0000006c <test>:
>
> void test(unsigned char a)
> {
>         switch (a) {
>   6c:   e8 2f           mov     r30, r24
>   6e:   ff 27           eor     r31, r31
>   70:   e8 30           cpi     r30, 0x08       ; 8
>   72:   f1 05           cpc     r31, r1
>   74:   a0 f4           brcc    .+40            ; 0x9e <test+0x32>
>   76:   ed 5e           subi    r30, 0xED       ; 237
>   78:   ff 4f           sbci    r31, 0xFF       ; 255
>   7a:   09 94           ijmp
>                 case 0: PORTB = 12; break;
>   7c:   8c e0           ldi     r24, 0x0C       ; 12
>   7e:   0b c0           rjmp    .+22            ; 0x96 <test+0x2a>
>                 case 1: PORTB = 72; break;
>   80:   88 e4           ldi     r24, 0x48       ; 72
>   82:   09 c0           rjmp    .+18            ; 0x96 <test+0x2a>
>                 case 2: PORTB = 25; break;
>   84:   89 e1           ldi     r24, 0x19       ; 25
>   86:   07 c0           rjmp    .+14            ; 0x96 <test+0x2a>
>                 case 3: PORTB = 17; break;
>   88:   81 e1           ldi     r24, 0x11       ; 17
>   8a:   05 c0           rjmp    .+10            ; 0x96 <test+0x2a>
>                 case 4: PORTB = 19; break;
>   8c:   83 e1           ldi     r24, 0x13       ; 19
>   8e:   03 c0           rjmp    .+6             ; 0x96 <test+0x2a>
>                 case 5: PORTB = 29; break;
>   90:   8d e1           ldi     r24, 0x1D       ; 29
>   92:   01 c0           rjmp    .+2             ; 0x96 <test+0x2a>
>                 case 6: PORTB = 54; break;
>   94:   86 e3           ldi     r24, 0x36       ; 54
>   96:   88 bb           out     0x18, r24       ; 24
>   98:   08 95           ret
>                 case 7: PORTB = 23; break;
>   9a:   87 e1           ldi     r24, 0x17       ; 23
>   9c:   88 bb           out     0x18, r24       ; 24
>   9e:   08 95           ret


The compiler builds the jump table for you as it should.

If the values are 0,8,16,...,56, then the compiler does a sort of binary 
search conditions to go to the value, bute the tests are done using just 
8 bits, unlike what you've shown in your original mail.

Maybe you should try doing "switch (state >> 3)" so that the targets are 
better aligned. Upgrading the compiler might be a good idea too, though 
there are still situations in which gcc4 behaves worse than gcc3... :(

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

Pointy-Haired Boss: I don't see anything that could stand in our way.
           Dilbert: Sanity? Reality? The laws of physics?