[picodrive.git] / pico / 32x / draw_arm.s

@*
@* PicoDrive
@* (C) notaz, 2010
@*
@* This work is licensed under the terms of MAME license.
@* See COPYING file in the top-level directory.
@*

.extern Pico32x
.extern PicoDraw2FB
.extern HighPal

.equiv P32XV_PRI,  (1<< 7)

.bss
.align 2
.global Pico32xNativePal
Pico32xNativePal:
    .word 0

.text
.align 2


.macro call_scan_prep cond
.if \cond
    ldr     r4, =PicoScan32xBegin
    ldr     r5, =PicoScan32xEnd
    ldr     r6, =DrawLineDest
    ldr     r4, [r4]
    ldr     r5, [r5]
    stmfd   sp!, {r4,r5,r6}
.endif
.endm

.macro call_scan_fin_ge cond
.if \cond
    addge   sp, sp, #4*3
.endif
.endm

.macro call_scan_begin cond
.if \cond
    stmfd   sp!, {r1-r3}
    and     r0, r2, #0xff
    add     r0, r0, r4
    mov     lr, pc
    ldr     pc, [sp, #(3+0)*4]
    ldr     r0, [sp, #(3+2)*4] @ &DrawLineDest
    ldmfd   sp!, {r1-r3}
    ldr     r0, [r0]
.endif
.endm

.macro call_scan_end cond
.if \cond
    stmfd   sp!, {r0-r3}
    and     r0, r2, #0xff
    add     r0, r0, r4
    mov     lr, pc
    ldr     pc, [sp, #(4+1)*4]
    ldmfd   sp!, {r0-r3}
.endif
.endm

@ direct color
@ unsigned short *dst, unsigned short *dram, int lines_sft_offs, int mdbg
.macro make_do_loop_dc name call_scan do_md
.global \name
\name:
    stmfd   sp!, {r4-r11,lr}

    ldr     r10,=Pico32x
    ldr     r11,=PicoDraw2FB
    ldr     r10,[r10, #0x40] @ Pico32x.vdp_regs[0]
    ldr     r11,[r11]
    ldr     r9, =HighPal     @ palmd
    and     r4, r2, #0xff
    mov     r5, #328
    lsl     r3, #26          @ mdbg << 26
    mla     r11,r4,r5,r11    @ r11 = pmd = PicoDraw2FB + offs*328: md data
    tst     r10,#P32XV_PRI
    moveq   r10,#0
    movne   r10,#0x8000      @ r10 = inv_bit
    call_scan_prep \call_scan

    mov     r4, #0           @ line
    b       1f @ loop_outer_entry

0: @ loop_outer:
    call_scan_end \call_scan
    add     r4, r4, #1
    sub     r11,r11,#1       @ adjust for prev read
    cmp     r4, r2, lsr #16
    call_scan_fin_ge \call_scan
    ldmgefd sp!, {r4-r11,pc}

1: @ loop_outer_entry:
    call_scan_begin \call_scan
    mov     r12,r4, lsl #1
    ldrh    r12,[r1, r12]
    add     r11,r11,#8
    mov     r6, #320
    add     r5, r1, r12, lsl #1 @ p32x = dram + dram[l]

2: @ loop_inner:
    ldrb    r7, [r11], #1    @ MD pixel
    subs    r6, r6, #1
    blt     0b @ loop_outer
    ldrh    r8, [r5], #2     @ 32x pixel
    cmp     r3, r7, lsl #26  @ MD has bg pixel?
    beq     3f @ draw32x
    eor     r12,r8, r10
    ands    r12,r12,#0x8000  @ !((t ^ inv) & 0x8000)
.if \do_md
    mov     r7, r7, lsl #1
    ldreqh  r12,[r9, r7]
    streqh  r12,[r0], #2     @ *dst++ = palmd[*pmd]
.endif
    beq     2b @ loop_inner

3: @ draw32x:
    and     r12,r8, #0x03e0
    mov     r8, r8, lsl #11
    orr     r8, r8, r8, lsr #(10+11)
    orr     r8, r8, r12,lsl #1
    bic     r8, r8, #0x0020  @ kill prio bit
    strh    r8, [r0], #2     @ *dst++ = bgr2rgb(*p32x++)
    b       2b @ loop_inner
.endm


@ packed pixel
@ note: this may read a few bytes over the end of PicoDraw2FB and dram,
@       so those should have a bit more alloc'ed than really needed.
@ unsigned short *dst, unsigned short *dram, int lines_sft_offs, int mdbg
.macro make_do_loop_pp name call_scan do_md
.global \name
\name:
    stmfd   sp!, {r4-r11,lr}

    ldr     r11,=PicoDraw2FB
    ldr     r10,=Pico32xNativePal
    ldr     r11,[r11]
    ldr     r10,[r10]
    ldr     r9, =HighPal     @ palmd
    and     r4, r2, #0xff
    mov     r5, #328
    lsl     r3, #26          @ mdbg << 26
    mla     r11,r4,r5,r11    @ r11 = pmd = PicoDraw2FB + offs*328: md data
    call_scan_prep \call_scan

    mov     r4, #0           @ line
    b       1f @ loop_outer_entry

0: @ loop_outer:
    call_scan_end \call_scan
    add     r4, r4, #1
    cmp     r4, r2, lsr #16
    call_scan_fin_ge \call_scan
    ldmgefd sp!, {r4-r11,pc}

1: @ loop_outer_entry:
    call_scan_begin \call_scan
    mov     r12,r4, lsl #1
    ldrh    r12,[r1, r12]
    add     r11,r11,#8
    mov     r6, #320/2
    add     r5, r1, r12, lsl #1 @ p32x = dram + dram[l]
    and     r12,r2, #0x100      @ shift
    add     r5, r5, r12,lsr #8

2: @ loop_inner:
@ r4,r6 - counters; r5 - 32x data; r9,r10 - md,32x pal; r11 - md data
@ r7,r8,r12,lr - temp
    tst     r5, #1
    ldreqb  r8, [r5], #2
    ldrb    r7, [r5, #-1]
    ldrneb  r8, [r5, #2]!    @ r7,r8 - pixel 0,1 index
    subs    r6, r6, #1
    blt     0b @ loop_outer
    cmp     r7, r8
    beq     5f @ check_fill @ +8

3: @ no_fill:
    mov     r12,r7, lsl #1
    mov     lr, r8, lsl #1
    ldrh    r7, [r10,r12]
    ldrh    r8, [r10,lr]
    add     r11,r11,#2

    eor     r12,r7, #0x20
    tst     r12,#0x20
    ldrneb  r12,[r11,#-2]    @ MD pixel 0
    eor     lr, r8, #0x20
    cmpne   r3, r12, lsl #26 @ MD has bg pixel?
.if \do_md
    mov     r12,r12,lsl #1
    ldrneh  r7, [r9, r12]    @ t = palmd[pmd[0]]
    tst     lr, #0x20
    ldrneb  lr, [r11,#-1]    @ MD pixel 1
    strh    r7, [r0], #2
    cmpne   r3, lr, lsl #26  @ MD has bg pixel?
    mov     lr, lr, lsl #1
    ldrneh  r8, [r9, lr]     @ t = palmd[pmd[1]]
    strh    r8, [r0], #2
.else
    streqh  r7, [r0]
    tst     lr, #0x20
    ldrneb  lr, [r11,#-1]    @ MD pixel 1
    add     r0, r0, #4
    cmpne   r3, lr, lsl #26  @ MD has bg pixel?
    streqh  r8, [r0, #-2]
.endif
    b       2b @ loop_inner

5: @ check_fill
    @ count pixels, align if needed
    bic     r12,r5, #1
    ldrh    r12,[r12]
    orr     lr, r7, r7, lsl #8
    cmp     r12,lr
    bne     3b @ no_fill

    tst     r5, #1
    sub     lr, r5, #2       @ starting r5 (32x render data start)
    addeq   r5, r5, #2
    addne   r5, r5, #1       @ add for the check above
    add     r6, r6, #1       @ restore from dec
    orr     r7, r7, r7, lsl #8
6:
    sub     r12,r5, lr
    ldrh    r8, [r5], #2
    cmp     r12,r6, lsl #1
    ldrh    r12,[r5], #2
    bge     7f @ count_done
    cmp     r8, r7
    cmpeq   r12,r7
    beq     6b

7: @ count_done
    sub     r5, r5, #4      @ undo readahead

    @ fix alignment and check type
    sub     r8, r5, lr
    tst     r8, #1
    subne   r5, r5, #1
    subne   r8, r8, #1

    and     r7, r7, #0xff
    cmp     r8, r6, lsl #1
    mov     r7, r7, lsl #1
    movgt   r8, r6, lsl #1  @ r8=count
    ldrh    r7, [r10,r7]
    sub     r6, r6, r8, lsr #1 @ adjust counter
    tst     r7, #0x20
    beq     9f @ bg_mode

    add     r11,r11,r8
8:
    subs    r8, r8, #2
    strgeh  r7, [r0], #2
    strgeh  r7, [r0], #2
    bgt     8b
    b       2b @ loop_inner

9: @ bg_mode:
    ldrb    r12,[r11],#1     @ MD pixel
    ldrb    lr, [r11],#1
    cmp     r3, lr, lsl #26  @ MD has bg pixel?
.if \do_md
    mov     r12,r12,lsl #1
    ldrneh  r12,[r9, r12]    @ t = palmd[*pmd]
    moveq   r12,r7
    cmp     r3, lr, lsl #26
    mov     lr, lr, lsl #1
    ldrneh  lr, [r9, lr]
    moveq   lr, r7
    strh    r12,[r0], #2
    strh    lr, [r0], #2
.else
    streqh  r7, [r0]
    cmp     r3, lr, lsl #26
    streqh  r7, [r0, #2]
    add     r0, r0, #4
.endif
    subs    r8, r8, #2
    bgt     9b @ bg_mode
    b       2b @ loop_inner
.endm


@ run length
@ unsigned short *dst, unsigned short *dram, int lines_sft_offs, int mdbg
.macro make_do_loop_rl name call_scan do_md
.global \name
\name:
    stmfd   sp!, {r4-r11,lr}

    ldr     r11,=PicoDraw2FB
    ldr     r10,=Pico32xNativePal
    ldr     r11,[r11]
    ldr     r10,[r10]
    ldr     r9, =HighPal     @ palmd
    and     r4, r2, #0xff
    mov     r5, #328
    lsl     r3, #26          @ mdbg << 26
    mla     r11,r4,r5,r11    @ r11 = pmd = PicoDraw2FB + offs*328: md data
    call_scan_prep \call_scan

    mov     r4, #0           @ line
    b       1f @ loop_outer_entry

0: @ loop_outer:
    call_scan_end \call_scan
    add     r4, r4, #1
    sub     r11,r11,#1       @ adjust for prev read
    cmp     r4, r2, lsr #16
    call_scan_fin_ge \call_scan
    ldmgefd sp!, {r4-r11,pc}

1: @ loop_outer_entry:
    call_scan_begin \call_scan
    mov     r12,r4, lsl #1
    ldrh    r12,[r1, r12]
    add     r11,r11,#8
    mov     r6, #320
    add     r5, r1, r12, lsl #1 @ p32x = dram + dram[l]

2: @ loop_inner:
    ldrh    r8, [r5], #2     @ control word
    and     r12,r8, #0xff
    mov     r12,r12,lsl #1
    ldrh    lr, [r10,r12]    @ t = 32x pixel
    eor     lr, lr, #0x20

3: @ loop_innermost:
    ldrb    r7, [r11], #1    @ MD pixel
    subs    r6, r6, #1
    blt     0b @ loop_outer
    cmp     r3, r7, lsl #26  @ MD has bg pixel?
    mov     r7, r7, lsl #1
    tstne   lr, #0x20
.if \do_md
    ldrneh  r12,[r9, r7]     @ t = palmd[*pmd]
    streqh  lr, [r0], #2
    strneh  r12,[r0], #2     @ *dst++ = t
.else
    streqh  lr, [r0]
    add     r0, r0, #2
.endif
    subs    r8, r8, #0x100
    bge     3b @ loop_innermost
    b       2b @ loop_inner
.endm


make_do_loop_dc do_loop_dc,         0, 0
make_do_loop_dc do_loop_dc_md,      0, 1
make_do_loop_dc do_loop_dc_scan,    1, 0
make_do_loop_dc do_loop_dc_scan_md, 1, 1

make_do_loop_pp do_loop_pp,         0, 0
make_do_loop_pp do_loop_pp_md,      0, 1
make_do_loop_pp do_loop_pp_scan,    1, 0
make_do_loop_pp do_loop_pp_scan_md, 1, 1

make_do_loop_rl do_loop_rl,         0, 0
make_do_loop_rl do_loop_rl_md,      0, 1
make_do_loop_rl do_loop_rl_scan,    1, 0
make_do_loop_rl do_loop_rl_scan_md, 1, 1

@ vim:filetype=armasm
Commit	Line	Data
	1	@*
	2	@* PicoDrive
	3	@* (C) notaz, 2010
	4	@*
	5	@* This work is licensed under the terms of MAME license.
	6	@* See COPYING file in the top-level directory.
	7	@*
	8
	9	.extern Pico32x
	10	.extern PicoDraw2FB
	11	.extern HighPal
	12
	13	.equiv P32XV_PRI, (1<< 7)
	14
	15	.bss
	16	.align 2
	17	.global Pico32xNativePal
	18	Pico32xNativePal:
	19	.word 0
	20
	21	.text
	22	.align 2
	23
	24
	25	.macro call_scan_prep cond
	26	.if \cond
	27	ldr r4, =PicoScan32xBegin
	28	ldr r5, =PicoScan32xEnd
	29	ldr r6, =DrawLineDest
	30	ldr r4, [r4]
	31	ldr r5, [r5]
	32	stmfd sp!, {r4,r5,r6}
	33	.endif
	34	.endm
	35
	36	.macro call_scan_fin_ge cond
	37	.if \cond
	38	addge sp, sp, #4*3
	39	.endif
	40	.endm
	41
	42	.macro call_scan_begin cond
	43	.if \cond
	44	stmfd sp!, {r1-r3}
	45	and r0, r2, #0xff
	46	add r0, r0, r4
	47	mov lr, pc
	48	ldr pc, [sp, #(3+0)*4]
	49	ldr r0, [sp, #(3+2)*4] @ &DrawLineDest
	50	ldmfd sp!, {r1-r3}
	51	ldr r0, [r0]
	52	.endif
	53	.endm
	54
	55	.macro call_scan_end cond
	56	.if \cond
	57	stmfd sp!, {r0-r3}
	58	and r0, r2, #0xff
	59	add r0, r0, r4
	60	mov lr, pc
	61	ldr pc, [sp, #(4+1)*4]
	62	ldmfd sp!, {r0-r3}
	63	.endif
	64	.endm
	65
	66	@ direct color
	67	@ unsigned short dst, unsigned short dram, int lines_sft_offs, int mdbg
	68	.macro make_do_loop_dc name call_scan do_md
	69	.global \name
	70	\name:
	71	stmfd sp!, {r4-r11,lr}
	72
	73	ldr r10,=Pico32x
	74	ldr r11,=PicoDraw2FB
	75	ldr r10,[r10, #0x40] @ Pico32x.vdp_regs[0]
	76	ldr r11,[r11]
	77	ldr r9, =HighPal @ palmd
	78	and r4, r2, #0xff
	79	mov r5, #328
	80	lsl r3, #26 @ mdbg << 26
	81	mla r11,r4,r5,r11 @ r11 = pmd = PicoDraw2FB + offs*328: md data
	82	tst r10,#P32XV_PRI
	83	moveq r10,#0
	84	movne r10,#0x8000 @ r10 = inv_bit
	85	call_scan_prep \call_scan
	86
	87	mov r4, #0 @ line
	88	b 1f @ loop_outer_entry
	89
	90	0: @ loop_outer:
	91	call_scan_end \call_scan
	92	add r4, r4, #1
	93	sub r11,r11,#1 @ adjust for prev read
	94	cmp r4, r2, lsr #16
	95	call_scan_fin_ge \call_scan
	96	ldmgefd sp!, {r4-r11,pc}
	97
	98	1: @ loop_outer_entry:
	99	call_scan_begin \call_scan
	100	mov r12,r4, lsl #1
	101	ldrh r12,[r1, r12]
	102	add r11,r11,#8
	103	mov r6, #320
	104	add r5, r1, r12, lsl #1 @ p32x = dram + dram[l]
	105
	106	2: @ loop_inner:
	107	ldrb r7, [r11], #1 @ MD pixel
	108	subs r6, r6, #1
	109	blt 0b @ loop_outer
	110	ldrh r8, [r5], #2 @ 32x pixel
	111	cmp r3, r7, lsl #26 @ MD has bg pixel?
	112	beq 3f @ draw32x
	113	eor r12,r8, r10
	114	ands r12,r12,#0x8000 @ !((t ^ inv) & 0x8000)
	115	.if \do_md
	116	mov r7, r7, lsl #1
	117	ldreqh r12,[r9, r7]
	118	streqh r12,[r0], #2 @ dst++ = palmd[pmd]
	119	.endif
	120	beq 2b @ loop_inner
	121
	122	3: @ draw32x:
	123	and r12,r8, #0x03e0
	124	mov r8, r8, lsl #11
	125	orr r8, r8, r8, lsr #(10+11)
	126	orr r8, r8, r12,lsl #1
	127	bic r8, r8, #0x0020 @ kill prio bit
	128	strh r8, [r0], #2 @ dst++ = bgr2rgb(p32x++)
	129	b 2b @ loop_inner
	130	.endm
	131
	132
	133	@ packed pixel
	134	@ note: this may read a few bytes over the end of PicoDraw2FB and dram,
	135	@ so those should have a bit more alloc'ed than really needed.
	136	@ unsigned short dst, unsigned short dram, int lines_sft_offs, int mdbg
	137	.macro make_do_loop_pp name call_scan do_md
	138	.global \name
	139	\name:
	140	stmfd sp!, {r4-r11,lr}
	141
	142	ldr r11,=PicoDraw2FB
	143	ldr r10,=Pico32xNativePal
	144	ldr r11,[r11]
	145	ldr r10,[r10]
	146	ldr r9, =HighPal @ palmd
	147	and r4, r2, #0xff
	148	mov r5, #328
	149	lsl r3, #26 @ mdbg << 26
	150	mla r11,r4,r5,r11 @ r11 = pmd = PicoDraw2FB + offs*328: md data
	151	call_scan_prep \call_scan
	152
	153	mov r4, #0 @ line
	154	b 1f @ loop_outer_entry
	155
	156	0: @ loop_outer:
	157	call_scan_end \call_scan
	158	add r4, r4, #1
	159	cmp r4, r2, lsr #16
	160	call_scan_fin_ge \call_scan
	161	ldmgefd sp!, {r4-r11,pc}
	162
	163	1: @ loop_outer_entry:
	164	call_scan_begin \call_scan
	165	mov r12,r4, lsl #1
	166	ldrh r12,[r1, r12]
	167	add r11,r11,#8
	168	mov r6, #320/2
	169	add r5, r1, r12, lsl #1 @ p32x = dram + dram[l]
	170	and r12,r2, #0x100 @ shift
	171	add r5, r5, r12,lsr #8
	172
	173	2: @ loop_inner:
	174	@ r4,r6 - counters; r5 - 32x data; r9,r10 - md,32x pal; r11 - md data
	175	@ r7,r8,r12,lr - temp
	176	tst r5, #1
	177	ldreqb r8, [r5], #2
	178	ldrb r7, [r5, #-1]
	179	ldrneb r8, [r5, #2]! @ r7,r8 - pixel 0,1 index
	180	subs r6, r6, #1
	181	blt 0b @ loop_outer
	182	cmp r7, r8
	183	beq 5f @ check_fill @ +8
	184
	185	3: @ no_fill:
	186	mov r12,r7, lsl #1
	187	mov lr, r8, lsl #1
	188	ldrh r7, [r10,r12]
	189	ldrh r8, [r10,lr]
	190	add r11,r11,#2
	191
	192	eor r12,r7, #0x20
	193	tst r12,#0x20
	194	ldrneb r12,[r11,#-2] @ MD pixel 0
	195	eor lr, r8, #0x20
	196	cmpne r3, r12, lsl #26 @ MD has bg pixel?
	197	.if \do_md
	198	mov r12,r12,lsl #1
	199	ldrneh r7, [r9, r12] @ t = palmd[pmd[0]]
	200	tst lr, #0x20
	201	ldrneb lr, [r11,#-1] @ MD pixel 1
	202	strh r7, [r0], #2
	203	cmpne r3, lr, lsl #26 @ MD has bg pixel?
	204	mov lr, lr, lsl #1
	205	ldrneh r8, [r9, lr] @ t = palmd[pmd[1]]
	206	strh r8, [r0], #2
	207	.else
	208	streqh r7, [r0]
	209	tst lr, #0x20
	210	ldrneb lr, [r11,#-1] @ MD pixel 1
	211	add r0, r0, #4
	212	cmpne r3, lr, lsl #26 @ MD has bg pixel?
	213	streqh r8, [r0, #-2]
	214	.endif
	215	b 2b @ loop_inner
	216
	217	5: @ check_fill
	218	@ count pixels, align if needed
	219	bic r12,r5, #1
	220	ldrh r12,[r12]
	221	orr lr, r7, r7, lsl #8
	222	cmp r12,lr
	223	bne 3b @ no_fill
	224
	225	tst r5, #1
	226	sub lr, r5, #2 @ starting r5 (32x render data start)
	227	addeq r5, r5, #2
	228	addne r5, r5, #1 @ add for the check above
	229	add r6, r6, #1 @ restore from dec
	230	orr r7, r7, r7, lsl #8
	231	6:
	232	sub r12,r5, lr
	233	ldrh r8, [r5], #2
	234	cmp r12,r6, lsl #1
	235	ldrh r12,[r5], #2
	236	bge 7f @ count_done
	237	cmp r8, r7
	238	cmpeq r12,r7
	239	beq 6b
	240
	241	7: @ count_done
	242	sub r5, r5, #4 @ undo readahead
	243
	244	@ fix alignment and check type
	245	sub r8, r5, lr
	246	tst r8, #1
	247	subne r5, r5, #1
	248	subne r8, r8, #1
	249
	250	and r7, r7, #0xff
	251	cmp r8, r6, lsl #1
	252	mov r7, r7, lsl #1
	253	movgt r8, r6, lsl #1 @ r8=count
	254	ldrh r7, [r10,r7]
	255	sub r6, r6, r8, lsr #1 @ adjust counter
	256	tst r7, #0x20
	257	beq 9f @ bg_mode
	258
	259	add r11,r11,r8
	260	8:
	261	subs r8, r8, #2
	262	strgeh r7, [r0], #2
	263	strgeh r7, [r0], #2
	264	bgt 8b
	265	b 2b @ loop_inner
	266
	267	9: @ bg_mode:
	268	ldrb r12,[r11],#1 @ MD pixel
	269	ldrb lr, [r11],#1
	270	cmp r3, lr, lsl #26 @ MD has bg pixel?
	271	.if \do_md
	272	mov r12,r12,lsl #1
	273	ldrneh r12,[r9, r12] @ t = palmd[*pmd]
	274	moveq r12,r7
	275	cmp r3, lr, lsl #26
	276	mov lr, lr, lsl #1
	277	ldrneh lr, [r9, lr]
	278	moveq lr, r7
	279	strh r12,[r0], #2
	280	strh lr, [r0], #2
	281	.else
	282	streqh r7, [r0]
	283	cmp r3, lr, lsl #26
	284	streqh r7, [r0, #2]
	285	add r0, r0, #4
	286	.endif
	287	subs r8, r8, #2
	288	bgt 9b @ bg_mode
	289	b 2b @ loop_inner
	290	.endm
	291
	292
	293	@ run length
	294	@ unsigned short dst, unsigned short dram, int lines_sft_offs, int mdbg
	295	.macro make_do_loop_rl name call_scan do_md
	296	.global \name
	297	\name:
	298	stmfd sp!, {r4-r11,lr}
	299
	300	ldr r11,=PicoDraw2FB
	301	ldr r10,=Pico32xNativePal
	302	ldr r11,[r11]
	303	ldr r10,[r10]
	304	ldr r9, =HighPal @ palmd
	305	and r4, r2, #0xff
	306	mov r5, #328
	307	lsl r3, #26 @ mdbg << 26
	308	mla r11,r4,r5,r11 @ r11 = pmd = PicoDraw2FB + offs*328: md data
	309	call_scan_prep \call_scan
	310
	311	mov r4, #0 @ line
	312	b 1f @ loop_outer_entry
	313
	314	0: @ loop_outer:
	315	call_scan_end \call_scan
	316	add r4, r4, #1
	317	sub r11,r11,#1 @ adjust for prev read
	318	cmp r4, r2, lsr #16
	319	call_scan_fin_ge \call_scan
	320	ldmgefd sp!, {r4-r11,pc}
	321
	322	1: @ loop_outer_entry:
	323	call_scan_begin \call_scan
	324	mov r12,r4, lsl #1
	325	ldrh r12,[r1, r12]
	326	add r11,r11,#8
	327	mov r6, #320
	328	add r5, r1, r12, lsl #1 @ p32x = dram + dram[l]
	329
	330	2: @ loop_inner:
	331	ldrh r8, [r5], #2 @ control word
	332	and r12,r8, #0xff
	333	mov r12,r12,lsl #1
	334	ldrh lr, [r10,r12] @ t = 32x pixel
	335	eor lr, lr, #0x20
	336
	337	3: @ loop_innermost:
	338	ldrb r7, [r11], #1 @ MD pixel
	339	subs r6, r6, #1
	340	blt 0b @ loop_outer
	341	cmp r3, r7, lsl #26 @ MD has bg pixel?
	342	mov r7, r7, lsl #1
	343	tstne lr, #0x20
	344	.if \do_md
	345	ldrneh r12,[r9, r7] @ t = palmd[*pmd]
	346	streqh lr, [r0], #2
	347	strneh r12,[r0], #2 @ *dst++ = t
	348	.else
	349	streqh lr, [r0]
	350	add r0, r0, #2
	351	.endif
	352	subs r8, r8, #0x100
	353	bge 3b @ loop_innermost
	354	b 2b @ loop_inner
	355	.endm
	356
	357
	358	make_do_loop_dc do_loop_dc, 0, 0
	359	make_do_loop_dc do_loop_dc_md, 0, 1
	360	make_do_loop_dc do_loop_dc_scan, 1, 0
	361	make_do_loop_dc do_loop_dc_scan_md, 1, 1
	362
	363	make_do_loop_pp do_loop_pp, 0, 0
	364	make_do_loop_pp do_loop_pp_md, 0, 1
	365	make_do_loop_pp do_loop_pp_scan, 1, 0
	366	make_do_loop_pp do_loop_pp_scan_md, 1, 1
	367
	368	make_do_loop_rl do_loop_rl, 0, 0
	369	make_do_loop_rl do_loop_rl_md, 0, 1
	370	make_do_loop_rl do_loop_rl_scan, 1, 0
	371	make_do_loop_rl do_loop_rl_scan_md, 1, 1
	372
	373	@ vim:filetype=armasm