fill screen 3ds asm

#define VRAM_BASE     (0x18000000)
#define LCD_FB_PDC0           (0x10400400)
#define LCD_FB_PDC1           (0x10400500)
#define LCD_FB_A_ADDR_OFFSET  (0x68)
#define LCD_FB_B_ADDR_OFFSET  (0x94)
#define FB_TOP_SIZE           (400*240*3)
#define FB_BOT_SIZE           (320*240*3)
#define FB_TOP_LEFT1          (VRAM_BASE)
#define FB_TOP_LEFT2          (FB_TOP_LEFT1  + FB_TOP_SIZE)
#define FB_TOP_RIGHT1         (FB_TOP_LEFT2  + FB_TOP_SIZE)
#define FB_TOP_RIGHT2         (FB_TOP_RIGHT1 + FB_TOP_SIZE)
#define FB_BOT_1              (FB_TOP_RIGHT2 + FB_TOP_SIZE)
#define FB_BOT_2              (FB_BOT_1      + FB_BOT_SIZE)


MLG:
    mov r0, #0x0000FF
    bl fill_screen
    b MLG

@ r0 = color
fill_screen:
    mov r9, r0

    and r3, r0, #0xFF
    lsr r4, r0, #8
    and r4, r4, #0xFF
    lsr r5, r0, #16
    and r5, r5, #0xFF

    ldr r0, =(FB_TOP_SIZE)
    ldr r1, =(FB_TOP_LEFT1)
    ldr r7, =(FB_TOP_SIZE)
    add r2, r1, r0 @limit_addr
    mov r6, r1 @ ptr
_fill_for_1:
    strb r3, [r6, #0]
    strb r4, [r6, #1]
    strb r5, [r6, #2]
    add r8, r6, r7
    strb r3, [r8, #0]
    strb r4, [r8, #1]
    strb r5, [r8, #2]
    add r6, r6, #3
    cmp r6, r2
    blt _fill_for_1

    mov r0, r9

    and r3, r0, #0xFF
    lsr r4, r0, #8
    and r4, r4, #0xFF
    lsr r5, r0, #16
    and r5, r5, #0xFF

    ldr r0, =(FB_TOP_SIZE)
    ldr r1, =(FB_TOP_RIGHT1)
    ldr r7, =(FB_TOP_SIZE)
    add r2, r1, r0 @limit_addr
    mov r6, r1 @ ptr
_fill_for_2:
    strb r3, [r6, #0]
    strb r4, [r6, #1]
    strb r5, [r6, #2]
    add r8, r6, r7
    strb r3, [r8, #0]
    strb r4, [r8, #1]
    strb r5, [r8, #2]
    add r6, r6, #3
    cmp r6, r2
    blt _fill_for_2

    bx lr