*=$C000

; Generic memcpy optimized for speed. No overlapping 
; fcatrin@gmail.com
;
; Copy operation is divided in two parts.  N is the total length 
; * memcpyLong  for N / 256 blocks
; * memcpyShort for N % 256 remaining bytes
 
memcpySrc = $E0
memcpyDst = $E2
memcpyLen = $E4
 

; set for characters
     lda #$95
     sta memcpySrc
     lda #$35
     sta memcpySrc+1

     lda #$00
     sta memcpyDst
     lda #$04
     sta memcpyDst+1

     lda #$E8
     sta memcpyLen
     lda #$03
     sta memcpyLen+1

     jsr memcpy

; set for attributes
     lda #$95
     sta memcpySrc
     lda #$39
     sta memcpySrc+1

     lda #$00
     sta memcpyDst
     lda #$d8
     sta memcpyDst+1

     lda #$E8
     sta memcpyLen
     lda #$03
     sta memcpyLen+1

     jsr memcpy
     rts


memcpy
     ldy #0
     ldx memcpyLen+1
     beq memcpyShort           ; We need only the short version for <1 pages
 
memcpyLoopLong                 ; Copy X pages
     lda (memcpySrc),y         ; Loop unrolling can be done with confidence here
     sta (memcpyDst),y         ; any power of 2 will work
     iny
     bne memcpyLoopLong
     dex
     beq memcpyShort
     inc memcpySrc+1           ; Go to the next page
     inc memcpyDst+1
     jmp memcpyLoopLong
 
memcpyShort                    ; Copy remaining bytes
     ldx memcpyLen
     beq memcpyEnd
 
memcpyLoopShort                ; Copy X bytes
     lda (memcpySrc),y
     sta (memcpyDst),y
     iny
     dex
     bne memcpyLoopShort
 
memcpyEnd
     rts