#197 STM8. Not very good code generation

[Vladimir Antonenko]

STM8. Not very good code generation.

Source code:

static void _delay(unsigned char ticks)
{
    while(--ticks);
}

Generated code:

__delay:
00101$:
    dec a
    tnz a
    jrne    00101$
    ret

TNZ instruction is not needed because DEC already sets N flag.
I am surprised that such fairly old compiler generates ineffective code for such fairly old CPU.

[Vladimir Antonenko]

I mean Z flag.

[Philipp Klaus Krause]

In current trunk, I see codegen generate the tnz, but the peephole optimizer optimizes it out, so the final result is:

__delay:
;   test.c: 75: while(--ticks);
00101$:
    dec a
    jrne    00101$
;   test.c: 76: }
    ret

Which version of sdcc did you use?

[Vladimir Antonenko]

Hi,
I used sdcc-win64 4.4.0 rc3. Options: --opt-code-speed or --opt-code-size

[Vladimir Antonenko]

Another one:
SDCC : mcs51/z80/z180/r2k/r2ka/r3ka/sm83/tlcs90/ez80_z80/z80n/r800/ds390/pic16/pic14/TININative/ds400/hc08/s08/stm8/pdk13/pdk14/pdk15/mos6502/mos65c02 TD- 4.4.0 #14620 (Linux)
sdcc -mstm8 --opt-code-speed test01.c

[Philipp Klaus Krause]

Looks fine in current trunk to me, too.
There were some stm8 peephole optimizer fixes in early March this year. Maybe one of them helped here.

[Vladimir Antonenko]

Another interesting case:

static void memcpy1(unsigned char* dst, unsigned char* src, unsigned char size)
{
    do { *dst++ = *src++; } while(--size);
}

static void memcpy2(unsigned char* dst, unsigned char* src, unsigned char size)
{
    if (!size) return;
    do { *dst++ = *src++; } while(--size);
}  

the difference of loop code between these two functions.

[Philipp Klaus Krause]

When using stronger optimization (I tried with --max-allocs-per-node 100000), I get this using sdcc from trunk:

_memcpy1:
    push    a
    ldw y, (0x04, sp)
    ld  a, (0x06, sp)
    ld  (0x01, sp), a
00101$:
    ld  a, (y)
    incw    y
    ld  (x), a
    incw    x
    dec (0x01, sp)
    jrne    00101$
    ldw x, (2, sp)
    addw    sp, #6
    jp  (x)

_memcpy2:
    push    a
    tnz (0x06, sp)
    jreq    00106$
    ldw y, (0x04, sp)
    ld  a, (0x06, sp)
    ld  (0x01, sp), a
00103$:
    ld  a, (y)
    incw    y
    ld  (x), a
    incw    x
    dec (0x01, sp)
    jrne    00103$
00106$:
    ldw x, (2, sp)
    addw    sp, #6
    jp  (x)

[Vladimir Antonenko]

OK, thanks.

ld a, (0x06, sp)
ld (0x01, sp), a - saving copy of <size> on stack looks unnecessary to me.</size>

[Philipp Klaus Krause]

Yes. I just introduced an optimization for this in [r14867]. It is still very basic (only works for some simple cases and only for stm8), for your first function, I now get:

_memcpy1:
    ldw y, (0x03, sp)
00101$:
    ld  a, (y)
    incw    y
    ld  (x), a
    incw    x
    dec (0x05, sp)
    jrne    00101$
    ldw x, (1, sp)
    addw    sp, #5
    jp  (x)

[Philipp Klaus Krause]

Looks like current trunk can generate good enough code for this, so I'm closing the ticket.