#define NLWA(a) (((u32_t)(a)) & 3)
#define NWA(a) (((u32_t)(a)) & 1)


#define LD32B(d,p)           \
    d  = *((p) + 3); d <<= 8;\
    d |= *((p) + 2); d <<= 8;\
    d |= *((p) + 1); d <<= 8;\
    d |= *((p) + 0); p  += 4

#define LD32W(d,p)            \
    d  = *((p) + 1); d <<= 16;\
    d |= *((p) + 0); p  += 2





/******************************************************************************
 * DESCRIPTION:
 * 
 *****************************************************************************/
void * sa_memcpy(void *dst,void *src,u32_t sz) {
 REG_t u32_t *p32s;
 REG_t u32_t *p32d;
 REG_t u16_t *p16s; 
 REG_t u32_t  lw  ;
 REG_t  u8_t *p8s = (u8_t*)src;
 REG_t  u8_t *p8d = (u8_t*)dst;
      vu32_t  i   ;
 

    if ((p8s == NULL) || (p8d == NULL) || !sz) {
       return p8d;
    }
    /////////////////////////    
    //    выравниваем dst  //
    /////////////////////////    
 	while (NLWA(p8d) && sz) {
 	                    sz--;
 	   	       *p8d++ = *p8s++;
 	}
    /////////////////////////
    // копируем по 4 байта //
    /////////////////////////    
    p32d = (u32_t*)p8d;
    i    = sz >> 2;
    sz  &= 3;
    
    if (i) {
	    if (NLWA(p8s)) {
	       /*** src != %4 ***/ 
	       if (NWA(p8s)) {
	          /*** src != %2 ***/
	          do { LD32B(lw,p8s);
	               *p32d++ = lw ;
	          } while(--i);
	       } else {
	          /*** src = %2 ***/
	                        p16s = (u16_t*)p8s;
	          do { LD32W(lw,p16s);
	               *p32d++ = lw;
	          } while(--i);
	          p8s = (u8_t*)p16s;
	       }
	    } else {
	       /*** src  = %4 ***/
	                       p32s = (u32_t*)p8s;
	       do { *p32d++ = *p32s++;
           } while(--i);
	       p8s = (u8_t*)p32s;
	    }
    }  	
    /////////////////////////
    // копируем остаток    //
    /////////////////////////    
    p8d = (u8_t*)p32d;
    while(sz--) {
     *p8d++ = *p8s++;
    }
    return p8d;
}