162306a36Sopenharmony_ci/* SPDX-License-Identifier: GPL-2.0 */
262306a36Sopenharmony_ci/*
362306a36Sopenharmony_ci * arch/alpha/lib/memmove.S
462306a36Sopenharmony_ci *
562306a36Sopenharmony_ci * Barely optimized memmove routine for Alpha EV5.
662306a36Sopenharmony_ci *
762306a36Sopenharmony_ci * This is hand-massaged output from the original memcpy.c.  We defer to
862306a36Sopenharmony_ci * memcpy whenever possible; the backwards copy loops are not unrolled.
962306a36Sopenharmony_ci */
1062306a36Sopenharmony_ci#include <linux/export.h>
1162306a36Sopenharmony_ci	.set noat
1262306a36Sopenharmony_ci	.set noreorder
1362306a36Sopenharmony_ci	.text
1462306a36Sopenharmony_ci
1562306a36Sopenharmony_ci	.align 4
1662306a36Sopenharmony_ci	.globl memmove
1762306a36Sopenharmony_ci	.ent memmove
1862306a36Sopenharmony_cimemmove:
1962306a36Sopenharmony_ci	ldgp $29, 0($27)
2062306a36Sopenharmony_ci	unop
2162306a36Sopenharmony_ci	nop
2262306a36Sopenharmony_ci	.prologue 1
2362306a36Sopenharmony_ci
2462306a36Sopenharmony_ci	addq $16,$18,$4
2562306a36Sopenharmony_ci	addq $17,$18,$5
2662306a36Sopenharmony_ci	cmpule $4,$17,$1		/*  dest + n <= src  */
2762306a36Sopenharmony_ci	cmpule $5,$16,$2		/*  dest >= src + n  */
2862306a36Sopenharmony_ci
2962306a36Sopenharmony_ci	bis $1,$2,$1
3062306a36Sopenharmony_ci	mov $16,$0
3162306a36Sopenharmony_ci	xor $16,$17,$2
3262306a36Sopenharmony_ci	bne $1,memcpy			!samegp
3362306a36Sopenharmony_ci
3462306a36Sopenharmony_ci	and $2,7,$2			/* Test for src/dest co-alignment.  */
3562306a36Sopenharmony_ci	and $16,7,$1
3662306a36Sopenharmony_ci	cmpule $16,$17,$3
3762306a36Sopenharmony_ci	bne $3,$memmove_up		/* dest < src */
3862306a36Sopenharmony_ci
3962306a36Sopenharmony_ci	and $4,7,$1
4062306a36Sopenharmony_ci	bne $2,$misaligned_dn
4162306a36Sopenharmony_ci	unop
4262306a36Sopenharmony_ci	beq $1,$skip_aligned_byte_loop_head_dn
4362306a36Sopenharmony_ci
4462306a36Sopenharmony_ci$aligned_byte_loop_head_dn:
4562306a36Sopenharmony_ci	lda $4,-1($4)
4662306a36Sopenharmony_ci	lda $5,-1($5)
4762306a36Sopenharmony_ci	unop
4862306a36Sopenharmony_ci	ble $18,$egress
4962306a36Sopenharmony_ci
5062306a36Sopenharmony_ci	ldq_u $3,0($5)
5162306a36Sopenharmony_ci	ldq_u $2,0($4)
5262306a36Sopenharmony_ci	lda $18,-1($18)
5362306a36Sopenharmony_ci	extbl $3,$5,$1
5462306a36Sopenharmony_ci
5562306a36Sopenharmony_ci	insbl $1,$4,$1
5662306a36Sopenharmony_ci	mskbl $2,$4,$2
5762306a36Sopenharmony_ci	bis $1,$2,$1
5862306a36Sopenharmony_ci	and $4,7,$6
5962306a36Sopenharmony_ci
6062306a36Sopenharmony_ci	stq_u $1,0($4)
6162306a36Sopenharmony_ci	bne $6,$aligned_byte_loop_head_dn
6262306a36Sopenharmony_ci
6362306a36Sopenharmony_ci$skip_aligned_byte_loop_head_dn:
6462306a36Sopenharmony_ci	lda $18,-8($18)
6562306a36Sopenharmony_ci	blt $18,$skip_aligned_word_loop_dn
6662306a36Sopenharmony_ci
6762306a36Sopenharmony_ci$aligned_word_loop_dn:
6862306a36Sopenharmony_ci	ldq $1,-8($5)
6962306a36Sopenharmony_ci	nop
7062306a36Sopenharmony_ci	lda $5,-8($5)
7162306a36Sopenharmony_ci	lda $18,-8($18)
7262306a36Sopenharmony_ci
7362306a36Sopenharmony_ci	stq $1,-8($4)
7462306a36Sopenharmony_ci	nop
7562306a36Sopenharmony_ci	lda $4,-8($4)
7662306a36Sopenharmony_ci	bge $18,$aligned_word_loop_dn
7762306a36Sopenharmony_ci
7862306a36Sopenharmony_ci$skip_aligned_word_loop_dn:
7962306a36Sopenharmony_ci	lda $18,8($18)
8062306a36Sopenharmony_ci	bgt $18,$byte_loop_tail_dn
8162306a36Sopenharmony_ci	unop
8262306a36Sopenharmony_ci	ret $31,($26),1
8362306a36Sopenharmony_ci
8462306a36Sopenharmony_ci	.align 4
8562306a36Sopenharmony_ci$misaligned_dn:
8662306a36Sopenharmony_ci	nop
8762306a36Sopenharmony_ci	fnop
8862306a36Sopenharmony_ci	unop
8962306a36Sopenharmony_ci	beq $18,$egress
9062306a36Sopenharmony_ci
9162306a36Sopenharmony_ci$byte_loop_tail_dn:
9262306a36Sopenharmony_ci	ldq_u $3,-1($5)
9362306a36Sopenharmony_ci	ldq_u $2,-1($4)
9462306a36Sopenharmony_ci	lda $5,-1($5)
9562306a36Sopenharmony_ci	lda $4,-1($4)
9662306a36Sopenharmony_ci
9762306a36Sopenharmony_ci	lda $18,-1($18)
9862306a36Sopenharmony_ci	extbl $3,$5,$1
9962306a36Sopenharmony_ci	insbl $1,$4,$1
10062306a36Sopenharmony_ci	mskbl $2,$4,$2
10162306a36Sopenharmony_ci
10262306a36Sopenharmony_ci	bis $1,$2,$1
10362306a36Sopenharmony_ci	stq_u $1,0($4)
10462306a36Sopenharmony_ci	bgt $18,$byte_loop_tail_dn
10562306a36Sopenharmony_ci	br $egress
10662306a36Sopenharmony_ci
10762306a36Sopenharmony_ci$memmove_up:
10862306a36Sopenharmony_ci	mov $16,$4
10962306a36Sopenharmony_ci	mov $17,$5
11062306a36Sopenharmony_ci	bne $2,$misaligned_up
11162306a36Sopenharmony_ci	beq $1,$skip_aligned_byte_loop_head_up
11262306a36Sopenharmony_ci
11362306a36Sopenharmony_ci$aligned_byte_loop_head_up:
11462306a36Sopenharmony_ci	unop
11562306a36Sopenharmony_ci	ble $18,$egress
11662306a36Sopenharmony_ci	ldq_u $3,0($5)
11762306a36Sopenharmony_ci	ldq_u $2,0($4)
11862306a36Sopenharmony_ci
11962306a36Sopenharmony_ci	lda $18,-1($18)
12062306a36Sopenharmony_ci	extbl $3,$5,$1
12162306a36Sopenharmony_ci	insbl $1,$4,$1
12262306a36Sopenharmony_ci	mskbl $2,$4,$2
12362306a36Sopenharmony_ci
12462306a36Sopenharmony_ci	bis $1,$2,$1
12562306a36Sopenharmony_ci	lda $5,1($5)
12662306a36Sopenharmony_ci	stq_u $1,0($4)
12762306a36Sopenharmony_ci	lda $4,1($4)
12862306a36Sopenharmony_ci
12962306a36Sopenharmony_ci	and $4,7,$6
13062306a36Sopenharmony_ci	bne $6,$aligned_byte_loop_head_up
13162306a36Sopenharmony_ci
13262306a36Sopenharmony_ci$skip_aligned_byte_loop_head_up:
13362306a36Sopenharmony_ci	lda $18,-8($18)
13462306a36Sopenharmony_ci	blt $18,$skip_aligned_word_loop_up
13562306a36Sopenharmony_ci
13662306a36Sopenharmony_ci$aligned_word_loop_up:
13762306a36Sopenharmony_ci	ldq $1,0($5)
13862306a36Sopenharmony_ci	nop
13962306a36Sopenharmony_ci	lda $5,8($5)
14062306a36Sopenharmony_ci	lda $18,-8($18)
14162306a36Sopenharmony_ci
14262306a36Sopenharmony_ci	stq $1,0($4)
14362306a36Sopenharmony_ci	nop
14462306a36Sopenharmony_ci	lda $4,8($4)
14562306a36Sopenharmony_ci	bge $18,$aligned_word_loop_up
14662306a36Sopenharmony_ci
14762306a36Sopenharmony_ci$skip_aligned_word_loop_up:
14862306a36Sopenharmony_ci	lda $18,8($18)
14962306a36Sopenharmony_ci	bgt $18,$byte_loop_tail_up
15062306a36Sopenharmony_ci	unop
15162306a36Sopenharmony_ci	ret $31,($26),1
15262306a36Sopenharmony_ci
15362306a36Sopenharmony_ci	.align 4
15462306a36Sopenharmony_ci$misaligned_up:
15562306a36Sopenharmony_ci	nop
15662306a36Sopenharmony_ci	fnop
15762306a36Sopenharmony_ci	unop
15862306a36Sopenharmony_ci	beq $18,$egress
15962306a36Sopenharmony_ci
16062306a36Sopenharmony_ci$byte_loop_tail_up:
16162306a36Sopenharmony_ci	ldq_u $3,0($5)
16262306a36Sopenharmony_ci	ldq_u $2,0($4)
16362306a36Sopenharmony_ci	lda $18,-1($18)
16462306a36Sopenharmony_ci	extbl $3,$5,$1
16562306a36Sopenharmony_ci
16662306a36Sopenharmony_ci	insbl $1,$4,$1
16762306a36Sopenharmony_ci	mskbl $2,$4,$2
16862306a36Sopenharmony_ci	bis $1,$2,$1
16962306a36Sopenharmony_ci	stq_u $1,0($4)
17062306a36Sopenharmony_ci
17162306a36Sopenharmony_ci	lda $5,1($5)
17262306a36Sopenharmony_ci	lda $4,1($4)
17362306a36Sopenharmony_ci	nop
17462306a36Sopenharmony_ci	bgt $18,$byte_loop_tail_up
17562306a36Sopenharmony_ci
17662306a36Sopenharmony_ci$egress:
17762306a36Sopenharmony_ci	ret $31,($26),1
17862306a36Sopenharmony_ci	nop
17962306a36Sopenharmony_ci	nop
18062306a36Sopenharmony_ci	nop
18162306a36Sopenharmony_ci
18262306a36Sopenharmony_ci	.end memmove
18362306a36Sopenharmony_ci	EXPORT_SYMBOL(memmove)
184