[pcsx_rearmed.git] / frontend / blit320.s

/*
 * (C) Gražvydas "notaz" Ignotas, 2011
 *
 * This work is licensed under the terms of any of these licenses
 * (at your option):
 *  - GNU GPL, version 2 or later.
 *  - GNU LGPL, version 2.1 or later.
 * See the COPYING file in the top-level directory.
 *
 * This code tries to make better use of pollux/arm926 store buffer
 * by fusing words instead of using strhs.
 */

.text
.align 2

.macro lhw_str rl rt
    lsl     \rl, #16
    lsr     \rl, #16
    orr     \rl, \rt, lsl #16
    str     \rl, [r0], #4
.endm

.global blit320_640
blit320_640:
    stmfd   sp!, {r4-r8,lr}
    mov     r12, #40
    bic     r1, r1, #3
0:
    ldmia   r1!, {r2-r8,lr}
    lhw_str r2, r3
    lhw_str r4, r5
    lhw_str r6, r7
    subs    r12, #1
    lhw_str r8, lr
    bgt     0b
    ldmfd   sp!, {r4-r8,pc}


.global blit320_512
blit320_512:
    stmfd   sp!, {r4-r8,lr}
    mov     r12, #32
    bic     r1, r1, #3
0:
    ldmia   r1!, {r2-r8,lr}
    lsl     r2, #16
    lsr     r2, #16
    orr     r2, r3, lsl #16
    str     r2, [r0], #4         @ 0,2
    lsr     r4, #16
    lsr     r3, #16
    orr     r3, r4, lsl #16
    str     r3, [r0], #4         @ 3,5
    lsr     r5, #16
    orr     r5, r6, lsl #16
    str     r5, [r0], #4         @ 7,8
    lsr     r8, #16
    lsr     lr, #16
    str     r7, [r0], #4         @ 10,11
    orr     r8, lr, lsl #16
    subs    r12, #1
    str     r8, [r0], #4         @ 13,15
    bgt     0b
    ldmfd   sp!, {r4-r8,pc}


.macro unaligned_str rl rt
    lsr     \rl, #16
    orr     \rl, \rt, lsl #16
    str     \rl, [r0], #4
.endm

.global blit320_368
blit320_368:
    stmfd   sp!, {r4-r8,lr}
    mov     r12, #23
    bic     r1, r1, #3
0:
    ldmia   r1!, {r2-r8,lr}
    unaligned_str r2, r3         @ 1,2
    unaligned_str r3, r4         @ 3,4
    unaligned_str r4, r5         @ 5,6
    subs    r12, #1
    stmia   r0!, {r6-r8,lr}      @ 8-15
    bgt     0b
    ldmfd   sp!, {r4-r8,pc}


@ vim:filetype=armasm
Commit	Line	Data
	1	/*
	2	* (C) Gražvydas "notaz" Ignotas, 2011
	3	*
	4	* This work is licensed under the terms of any of these licenses
	5	* (at your option):
	6	* - GNU GPL, version 2 or later.
	7	* - GNU LGPL, version 2.1 or later.
	8	* See the COPYING file in the top-level directory.
	9	*
	10	* This code tries to make better use of pollux/arm926 store buffer
	11	* by fusing words instead of using strhs.
	12	*/
	13
	14	.text
	15	.align 2
	16
	17	.macro lhw_str rl rt
	18	lsl \rl, #16
	19	lsr \rl, #16
	20	orr \rl, \rt, lsl #16
	21	str \rl, [r0], #4
	22	.endm
	23
	24	.global blit320_640
	25	blit320_640:
	26	stmfd sp!, {r4-r8,lr}
	27	mov r12, #40
	28	bic r1, r1, #3
	29	0:
	30	ldmia r1!, {r2-r8,lr}
	31	lhw_str r2, r3
	32	lhw_str r4, r5
	33	lhw_str r6, r7
	34	subs r12, #1
	35	lhw_str r8, lr
	36	bgt 0b
	37	ldmfd sp!, {r4-r8,pc}
	38
	39
	40	.global blit320_512
	41	blit320_512:
	42	stmfd sp!, {r4-r8,lr}
	43	mov r12, #32
	44	bic r1, r1, #3
	45	0:
	46	ldmia r1!, {r2-r8,lr}
	47	lsl r2, #16
	48	lsr r2, #16
	49	orr r2, r3, lsl #16
	50	str r2, [r0], #4 @ 0,2
	51	lsr r4, #16
	52	lsr r3, #16
	53	orr r3, r4, lsl #16
	54	str r3, [r0], #4 @ 3,5
	55	lsr r5, #16
	56	orr r5, r6, lsl #16
	57	str r5, [r0], #4 @ 7,8
	58	lsr r8, #16
	59	lsr lr, #16
	60	str r7, [r0], #4 @ 10,11
	61	orr r8, lr, lsl #16
	62	subs r12, #1
	63	str r8, [r0], #4 @ 13,15
	64	bgt 0b
	65	ldmfd sp!, {r4-r8,pc}
	66
	67
	68	.macro unaligned_str rl rt
	69	lsr \rl, #16
	70	orr \rl, \rt, lsl #16
	71	str \rl, [r0], #4
	72	.endm
	73
	74	.global blit320_368
	75	blit320_368:
	76	stmfd sp!, {r4-r8,lr}
	77	mov r12, #23
	78	bic r1, r1, #3
	79	0:
	80	ldmia r1!, {r2-r8,lr}
	81	unaligned_str r2, r3 @ 1,2
	82	unaligned_str r3, r4 @ 3,4
	83	unaligned_str r4, r5 @ 5,6
	84	subs r12, #1
	85	stmia r0!, {r6-r8,lr} @ 8-15
	86	bgt 0b
	87	ldmfd sp!, {r4-r8,pc}
	88
	89
	90	@ vim:filetype=armasm