// macro to enter a stack with gcc

// macro to utilize lpm with inc

.macro  rpm     reg
#if defined  (__AVR_ENHANCED__)
        lpm     \reg, Z+
#else
        lpm
        mov     \reg, r0
        adiw    ZL, 1
#endif
.endm

// macro to enter a stack with gcc

.macro ENTER unused_regs stack_locals

ENTER:
        clr     XH
        ldi     XL, \stack_locals
        ldi     ZH, hi8(ENTRY)      // there is no way to rescale an address by 0.5
        ldi     ZL, lo8(ENTRY)      // thus we have to right shift the address with additional code
        lsr     ZH
        ror     ZL
        xjmp    (__prologue_saves__ + 2 * \unused_regs)
ENTRY:
.endm


// macro to leave a stack entered with ENTER

.macro LEAVE unused_regs stack_locals ptr_changed

LEAVE:
        ldi     ZL, 18 - \unused_regs
        .if (\ptr_changed)
           in      YL, _SFR_IO_ADDR(SPL)
           in      YH, _SFR_IO_ADDR(SPH)
        .endif
        .if (\stack_locals)
           adiw    YL, \stack_locals
        .endif
        xjmp    (__epilogue_restores__ + 2 * (\unused_regs))
.endm


// macro unsigned 8x8 multiplication

.macro  MUL8x8  resH resL A B temp
        clr        \resH
        ldi     \temp, 8
        lsr        \A
        brcc    .+2
        add     \resH, \B
        ror     \resH
        ror     \resL
        dec     \temp
        brne    .-14
.endm


// movw replacement, support words on odd addresses, overlapping, hardware dependend
// and avoid dummy moves on same registers, to get this macro working we have to reassign
// register defines, such as above

.macro  _movw dst src
    .if (\src - \dst)
        .if ((\src | \dst) & 0x01)
            .if ((\src - \dst) & 0x80)
                mov     \dst +1, \src +1
                mov     \dst   , \src
            .else
                mov     \dst   , \src
                mov     \dst +1, \src +1
            .endif
        .else
#if __AVR_ENHANCED__
                movw    \dst   , \src
#else
                mov     \dst   , \src
                mov     \dst +1, \src +1
#endif
        .endif
    .endif
.endm

typedef union {
    uint8_t All;
    struct {
    uint8_t AutoLineFeed:1,    
        FixedFont:1,
        Clipping:1;              
    };
} glcdFlags_t;

typedef union {
    uint8_t All;
    struct {
    uint8_t AutoLineFeed:1,    
        FixedFont:1,
        Clipping:1;              
    } s;
} glcdFlags_t;

#define __AVR_ATmega16__

#include <mega16.h>
#include <stdio.h>
#include <glcd.h>



void main ()

{
glcdWait(500);
}

extern void glcdWait(uint16_t ms)
{
#asm
glcdWait:
        sbiw    P0, 1                                   // 2
        brcc    glcdWait1                               // 2
        ret                                             // 2 -2 = 0 for no branch in both loops
glcdWait1:
        nop
        ldi     T0, lo8(DELAY_ONE_MILLISECOND -4)       // 1
        ldi     T1, hi8(DELAY_ONE_MILLISECOND -4)       // 1
glcdWait2:
        sbiw    T0, 1                                   // 2
        brne    glcdWait2                               // 2 = 4, = 3 for last loop
        rjmp    glcdWait                                // 2 = 8

#endasm
}


Как выгялдел этот фрагмент в glcd_init.asm:

[code]

// void glcdWait(uint16_t ms);          
// wait some milliseconds, XTAL must be correct defined

        .global glcdWait                // used T0,T1 stack 0
glcdWait:
        sbiw    P0, 1                                   // 2
        brcc    glcdWait1                               // 2
        ret                                             // 2 -2 = 0 for no branch in both loops
glcdWait1:
        nop
        ldi     T0, lo8(DELAY_ONE_MILLISECOND -4)       // 1
        ldi     T1, hi8(DELAY_ONE_MILLISECOND -4)       // 1
glcdWait2:
        sbiw    T0, 1                                   // 2
        brne    glcdWait2                               // 2 = 4, = 3 for last loop
        rjmp    glcdWait                                // 2 = 8


;************************************************

;     176         ret                                             // 2 -2 = 0 for no branch in both loops
        ret                                             // 2 -2 = 0 for no branch in both loops
;     177 glcdWait1:
glcdWait1:
;     178         nop
        nop
;     179         ldi     T0, lo8(DELAY_ONE_MILLISECOND -4)       // 1
        ldi     T0, lo8(DELAY_ONE_MILLISECOND -4)       // 1       // ------СТРОКА 1257----
;     180         ldi     T1, hi8(DELAY_ONE_MILLISECOND -4)       // 1
        ldi     T1, hi8(DELAY_ONE_MILLISECOND -4)       // 1