new easy6502 blit routine

define r16al  $f0
define r16ah  $f1
define r16bl  $f2
define r16bh  $f3
define r16cl  $f4
define r16ch  $f5
define arg0   $f6
define arg1   $f7

define tmp0   $fc
define tmp1   $fd

define random $fe
define key    $ff


init:
  lda #$00
  ldx #<amoog
  ldy #>amoog
  sta pos
loop:
  lda random
  jsr blit
  jmp loop


amoog:
  dcb  0,10,10,10
  dcb 10,10, 3, 3
  dcb 10,10,10,10
  dcb  0,10, 0,10



;(automatically loads pos into A)
pos: dcb 0
pos_loadInc:
  lda pos
  sta tmp0
  inc pos
  lda pos
  and #%00001000
  beq pos_loadInc_noYInc

  lda pos
  and #%01110111
  clc
  adc #16
  sta pos

  pos_loadInc_noYInc:
  lda tmp0
  rts



;blits a 4x4 color tile
 ;(restores A,X,Y)
;A = position (as #$YX)
;X = color (0 -> 15)
blit_color:
  ;(X is read-only here; storing is unnecessary)
  sta arg0 ;store arguments
  sty arg1  ;^^
  jsr blit_load_dst ;load dst addr to r16a

  ;completely unrolled loop
  ldy #0        ;init index to 0
  txa           ;pixel 0
  sta (r16al),y  ;^^
  iny            ;^^
  sta (r16al),y ;pixel 1
  iny            ;^^
  sta (r16al),y ;pixel 2
  iny            ;^^
  sta (r16al),y ;pixel 3
  tya            ;^^
  clc            ;^^
  adc #29        ;^^
  tay            ;^^
  txa           ;pixel 4
  sta (r16al),y  ;^^
  iny            ;^^
  sta (r16al),y ;pixel 5
  iny            ;^^
  sta (r16al),y ;pixel 6
  iny            ;^^
  sta (r16al),y ;pixel 7
  tya            ;^^
  clc            ;^^
  adc #29        ;^^
  tay            ;^^
  txa           ;pixel 8
  sta (r16al),y  ;^^
  iny            ;^^
  sta (r16al),y ;pixel 9
  iny            ;^^
  sta (r16al),y ;pixel 10
  iny            ;^^
  sta (r16al),y ;pixel 11
  tya            ;^^
  clc            ;^^
  adc #29        ;^^
  tay            ;^^
  txa           ;pixel 12
  sta (r16al),y  ;^^
  iny            ;^^
  sta (r16al),y ;pixel 13
  iny            ;^^
  sta (r16al),y ;pixel 14
  iny            ;^^
  sta (r16al),y ;pixel 15

  lda arg0 ;restore registers
  ldy arg1  ;^^
  rts


;blits a 4x4 bitmap tile
 ;(restores A,X,Y)
;A = position (as #$YX)
;X = bitmap addr lo byte
;Y = bitmap addr hi byte
blit:
  sta arg0          ;store arguments
  stx r16bl          ;^^
  sty r16bh          ;^^
  jsr blit_load_dst ;load dst addr to r16a

  ;completely unrolled loop
  ldy #0        ;pixel 0,0 (0,0)
  lda (r16bl),y  ;^^
  sta (r16al),y  ;^^
  iny            ;^^
  lda (r16bl),y ;pixel 1,0 (1,1)
  sta (r16al),y  ;^^
  iny            ;^^
  lda (r16bl),y ;pixel 2,0 (2,2)
  sta (r16al),y  ;^^
  iny            ;^^
  lda (r16bl),y ;pixel 3,0 (3,3)
  sta (r16al),y  ;^^
  iny            ;^^
  sty tmp0       ;^^
  tya            ;^^
  clc            ;^^
  adc #28        ;^^
  sta tmp1       ;^^

  lda (r16bl),y ;pixel 0,1 (4,32)
  ldy tmp1       ;^^
  sta (r16al),y  ;^^
  inc tmp0       ;^^
  iny            ;^^
  sty tmp1       ;^^
  ldy tmp0      ;pixel 1,1 (5,33)
  lda (r16bl),y  ;^^
  ldy tmp1       ;^^
  sta (r16al),y  ;^^
  inc tmp0       ;^^
  iny            ;^^
  sty tmp1       ;^^
  ldy tmp0      ;pixel 2,1 (6,34)
  lda (r16bl),y  ;^^
  ldy tmp1       ;^^
  sta (r16al),y  ;^^
  inc tmp0       ;^^
  iny            ;^^
  sty tmp1       ;^^
  ldy tmp0      ;pixel 3,1 (7,35)
  lda (r16bl),y  ;^^
  ldy tmp1       ;^^
  sta (r16al),y  ;^^
  inc tmp0       ;^^
  tya            ;^^
  clc            ;^^
  adc #29        ;^^
  sta tmp1       ;^^

  ldy tmp0      ;pixel 0,2 (4,64)
  lda (r16bl),y  ;^^
  ldy tmp1       ;^^
  sta (r16al),y  ;^^
  inc tmp0       ;^^
  iny            ;^^
  sty tmp1       ;^^
  ldy tmp0      ;pixel 1,2 (5,65)
  lda (r16bl),y  ;^^
  ldy tmp1       ;^^
  sta (r16al),y  ;^^
  inc tmp0       ;^^
  iny            ;^^
  sty tmp1       ;^^
  ldy tmp0      ;pixel 2,2 (6,66)
  lda (r16bl),y  ;^^
  ldy tmp1       ;^^
  sta (r16al),y  ;^^
  inc tmp0       ;^^
  iny            ;^^
  sty tmp1       ;^^
  ldy tmp0      ;pixel 3,2 (7,67)
  lda (r16bl),y  ;^^
  ldy tmp1       ;^^
  sta (r16al),y  ;^^
  inc tmp0       ;^^
  tya            ;^^
  clc            ;^^
  adc #29        ;^^
  sta tmp1       ;^^

  ldy tmp0      ;pixel 0,3 (4,96)
  lda (r16bl),y  ;^^
  ldy tmp1       ;^^
  sta (r16al),y  ;^^
  inc tmp0       ;^^
  iny            ;^^
  sty tmp1       ;^^
  ldy tmp0      ;pixel 1,3 (5,97)
  lda (r16bl),y  ;^^
  ldy tmp1       ;^^
  sta (r16al),y  ;^^
  inc tmp0       ;^^
  iny            ;^^
  sty tmp1       ;^^
  ldy tmp0      ;pixel 2,3 (6,98)
  lda (r16bl),y  ;^^
  ldy tmp1       ;^^
  sta (r16al),y  ;^^
  inc tmp0       ;^^
  iny            ;^^
  sty tmp1       ;^^
  ldy tmp0      ;pixel 3,3 (7,99)
  lda (r16bl),y  ;^^
  ldy tmp1       ;^^
  sta (r16al),y  ;^^

  lda arg0  ;restore registers
  ldx r16bl  ;^^
  ldy r16bh  ;^^
  rts


;private routine that loads tile's
 ;destination address into r16a
blit_load_dst:
  tay               ;lookup table offset into Y
  and #%00000111     ;^^
  sta tmp0           ;^^
  tya                ;^^
  lsr                ;^^
  and #%00111000     ;^^
  ora tmp0           ;^^
  asl                ;^^
  tay                ;^^
  lda #<blit_lookup ;lookup table addr. into r16c
  sta r16cl          ;^^
  lda #>blit_lookup  ;^^
  sta r16ch          ;^^
  lda (r16cl),y     ;destination addr. into r16a
  sta r16al          ;^^
  iny                ;^^
  lda (r16cl),y      ;^^
  sta r16ah          ;^^
  rts


blit_lookup:
dcb $00,$02,$04,$02,$08,$02,$0c,$02,$10,$02,$14,$02,$18,$02,$1c,$02
dcb $80,$02,$84,$02,$88,$02,$8c,$02,$90,$02,$94,$02,$98,$02,$9c,$02
dcb $00,$03,$04,$03,$08,$03,$0c,$03,$10,$03,$14,$03,$18,$03,$1c,$03
dcb $80,$03,$84,$03,$88,$03,$8c,$03,$90,$03,$94,$03,$98,$03,$9c,$03
dcb $00,$04,$04,$04,$08,$04,$0c,$04,$10,$04,$14,$04,$18,$04,$1c,$04
dcb $80,$04,$84,$04,$88,$04,$8c,$04,$90,$04,$94,$04,$98,$04,$9c,$04
dcb $00,$05,$04,$05,$08,$05,$0c,$05,$10,$05,$14,$05,$18,$05,$1c,$05
dcb $80,$05,$84,$05,$88,$05,$8c,$05,$90,$05,$94,$05,$98,$05,$9c,$05