#ifndef __FASTMEM_HPP
#define __FASTMEM_HPP

#ifdef __cplusplus
extern "C" {
#endif
#ifndef _BIG_INLINE
void __far memset8(void __far *p, unsigned char value, unsigned bytes);
#pragma aux memset8                 \
        parm [es di] [ax] [cx]      \
        modify exact [di ax cx dx]

void __far memset16(void __far *p, unsigned value, unsigned words);
#pragma aux memset16                \
        parm [es di] [ax] [cx]      \
        modify exact [di ax cx dx]

void __far memset32(void __far *p, unsigned long value, unsigned dwords);
#pragma aux memset32                \
        parm [es di] [dx ax] [cx]   \
        modify exact [di ax cx dx]

void __far memcpy8(void __far *d, const void __far *s, unsigned bytes);
#pragma aux memcpy8                 \
        parm [es di] [dx si] [cx]   \
        modify exact [di si cx dx]

#else /*_BIG_INLINE*/
void memset8(void __far *p, unsigned char value, unsigned bytes);
#pragma aux memset8 =               \
"               .386p              "\
"               jcxz thatsIt       "\
"                                  "\
"               mov ah,al          "\
"               mov dx,ax          "\
"               shl eax,16         "\
"               mov ax,dx          "\
"                                  "\
"               cld                "\
"               test di,1          "\
"               jz aligned16       "\
"               stosb              "\
"               dec cx             "\
"               jz thatsIt         "\
"                                  "\
"aligned16:     test di,2          "\
"               jz aligned32       "\
"               stosb              "\
"               dec cx             "\
"               jz thatsIt         "\
"               stosb              "\
"               dec cx             "\
"                                  "\
"aligned32:     mov dx,cx          "\
"               shr cx,2           "\
"               rep stosd          "\
"               mov cx,dx          "\
"               and cx,3           "\
"               rep stosb          "\
"thatsIt:                          "\
        parm [es di] [ax] [cx]      \
        modify exact [di ax cx dx]

void memset16(void __far *p, unsigned value, unsigned words);
#pragma aux memset16 =              \
".386p                             "\
"               jcxz thatsIt       "\
"                                  "\
"               cld                "\
"                                  "\
"               mov bx,ax          "\
"               shl eax,16         "\
"               mov ax,bx          "\
"                                  "\
"               mov dx,di          "\
"                                  "\
"               test di,1          "\
"               jz aligned16       "\
"               stosb              "\
"               ror eax,8          "\
"               dec cx             "\
"               jz lastByte        "\
"                                  "\
"aligned16:     test di,2          "\
"               jz aligned32       "\
"               stosw              "\
"               dec cx             "\
"                                  "\
"aligned32:     shr cx,1           "\
"               rep stosd          "\
"               jnc lastByte       "\
"               stosw              "\
"                                  "\
"lastByte:      test dx,1          "\
"               jz thatsIt         "\
"               stosb              "\
"                                  "\
"thatsIt:                          "\
        parm [es di] [ax] [cx]      \
        modify exact [di ax cx dx]

void memset32(void __far *p, unsigned long value, unsigned dwords);
#pragma aux memset32 =              \
".386p                             "\
"               jcxz thatsIt       "\
"                                  "\
"               cld                "\
"                                  "\
"               xchg dx,ax         "\
"               shl eax,16         "\
"               mov ax,dx          "\
"                                  "\
"               mov dx,di          "\
"                                  "\
"               test di,1          "\
"               jz aligned16       "\
"               stosb              "\
"               ror eax,8          "\
"               dec cx             "\
"               jz doRest          "\
"               test di,2          "\
"               jz mainCopy        "\
"               stosw              "\
"               ror eax,16         "\
"               jmp mainCopy       "\
"                                  "\
"aligned16:     test di,2          "\
"               jz mainCopy        "\
"               stosw              "\
"               ror eax,16         "\
"               dec cx             "\
"               jz doRest          "\
"                                  "\
"mainCopy:      rep stosd          "\
"                                  "\
"doRest:        test dx,2          "\
"               jz lastByte        "\
"               stosw              "\
"               ror eax,16         "\
"lastByte:      test dx,1          "\
"               jz thatsIt         "\
"               stosb              "\
"                                  "\
"thatsIt:                          "\
        parm [es di] [dx ax] [cx]   \
        modify exact [di dx ax cx]


void memcpy8(void __far *d, const void __far *s, unsigned bytes);
#pragma aux memcpy8 =                 \
"               push ds              "\
"               mov ds,dx            "\
"               jcxz thatsIt         "\
"                                    "\
"               cld                  "\
"                                    "\
"               test si,1            "\
"               jnz si_word_unaligned"\
"               test di,1            "\
"               jz sidi_word_aligned "\
"               jmp maincopy         "\
"si_word_unaligned:                  "\
"               test di,1            "\
"               jz maincopy          "\
"                                    "\
"sidi_word_unaligned:                "\
"               movsb                "\
"               dec cx               "\
"               jz thatsIt           "\
"                                    "\
"sidi_word_aligned:                  "\
"               test di,2            "\
"               jz maincopy          "\
"               test si,2            "\
"               jz maincopy          "\
"               movsw                "\
"               sub cx,2             "\
"               jz thatsIt           "\
"                                    "\
"maincopy:                           "\
"               mov dx,cx            "\
"               shr cx,2             "\
"               rep movsd            "\
"               mov cx,dx            "\
"               and cx,3             "\
"               rep movsb            "\
"                                    "\
"thatsIt:                            "\
"               pop ds               "\
        parm [es di] [dx si] [cx]     \
        modify exact [di si cx dx]

#endif


#ifdef __cplusplus
};
#endif

#endif /*__FASTMEM_HPP*/