1 /* SPDX-License-Identifier: GPL-2.0 */ 1 /* SPDX-License-Identifier: GPL-2.0 */
2 /* 2 /*
3 * arch/alpha/lib/memmove.S 3 * arch/alpha/lib/memmove.S
4 * 4 *
5 * Barely optimized memmove routine for Alpha 5 * Barely optimized memmove routine for Alpha EV5.
6 * 6 *
7 * This is hand-massaged output from the origi 7 * This is hand-massaged output from the original memcpy.c. We defer to
8 * memcpy whenever possible; the backwards cop 8 * memcpy whenever possible; the backwards copy loops are not unrolled.
9 */ 9 */
10 #include <linux/export.h> 10 #include <linux/export.h>
11 .set noat 11 .set noat
12 .set noreorder 12 .set noreorder
13 .text 13 .text
14 14
15 .align 4 15 .align 4
16 .globl memmove 16 .globl memmove
17 .ent memmove 17 .ent memmove
18 memmove: 18 memmove:
19 ldgp $29, 0($27) 19 ldgp $29, 0($27)
20 unop 20 unop
21 nop 21 nop
22 .prologue 1 22 .prologue 1
23 23
24 addq $16,$18,$4 24 addq $16,$18,$4
25 addq $17,$18,$5 25 addq $17,$18,$5
26 cmpule $4,$17,$1 /* de 26 cmpule $4,$17,$1 /* dest + n <= src */
27 cmpule $5,$16,$2 /* de 27 cmpule $5,$16,$2 /* dest >= src + n */
28 28
29 bis $1,$2,$1 29 bis $1,$2,$1
30 mov $16,$0 30 mov $16,$0
31 xor $16,$17,$2 31 xor $16,$17,$2
32 bne $1,memcpy !sameg 32 bne $1,memcpy !samegp
33 33
34 and $2,7,$2 /* Tes 34 and $2,7,$2 /* Test for src/dest co-alignment. */
35 and $16,7,$1 35 and $16,7,$1
36 cmpule $16,$17,$3 36 cmpule $16,$17,$3
37 bne $3,$memmove_up /* des 37 bne $3,$memmove_up /* dest < src */
38 38
39 and $4,7,$1 39 and $4,7,$1
40 bne $2,$misaligned_dn 40 bne $2,$misaligned_dn
41 unop 41 unop
42 beq $1,$skip_aligned_byte_loop_head_dn 42 beq $1,$skip_aligned_byte_loop_head_dn
43 43
44 $aligned_byte_loop_head_dn: 44 $aligned_byte_loop_head_dn:
45 lda $4,-1($4) 45 lda $4,-1($4)
46 lda $5,-1($5) 46 lda $5,-1($5)
47 unop 47 unop
48 ble $18,$egress 48 ble $18,$egress
49 49
50 ldq_u $3,0($5) 50 ldq_u $3,0($5)
51 ldq_u $2,0($4) 51 ldq_u $2,0($4)
52 lda $18,-1($18) 52 lda $18,-1($18)
53 extbl $3,$5,$1 53 extbl $3,$5,$1
54 54
55 insbl $1,$4,$1 55 insbl $1,$4,$1
56 mskbl $2,$4,$2 56 mskbl $2,$4,$2
57 bis $1,$2,$1 57 bis $1,$2,$1
58 and $4,7,$6 58 and $4,7,$6
59 59
60 stq_u $1,0($4) 60 stq_u $1,0($4)
61 bne $6,$aligned_byte_loop_head_dn 61 bne $6,$aligned_byte_loop_head_dn
62 62
63 $skip_aligned_byte_loop_head_dn: 63 $skip_aligned_byte_loop_head_dn:
64 lda $18,-8($18) 64 lda $18,-8($18)
65 blt $18,$skip_aligned_word_loop_dn 65 blt $18,$skip_aligned_word_loop_dn
66 66
67 $aligned_word_loop_dn: 67 $aligned_word_loop_dn:
68 ldq $1,-8($5) 68 ldq $1,-8($5)
69 nop 69 nop
70 lda $5,-8($5) 70 lda $5,-8($5)
71 lda $18,-8($18) 71 lda $18,-8($18)
72 72
73 stq $1,-8($4) 73 stq $1,-8($4)
74 nop 74 nop
75 lda $4,-8($4) 75 lda $4,-8($4)
76 bge $18,$aligned_word_loop_dn 76 bge $18,$aligned_word_loop_dn
77 77
78 $skip_aligned_word_loop_dn: 78 $skip_aligned_word_loop_dn:
79 lda $18,8($18) 79 lda $18,8($18)
80 bgt $18,$byte_loop_tail_dn 80 bgt $18,$byte_loop_tail_dn
81 unop 81 unop
82 ret $31,($26),1 82 ret $31,($26),1
83 83
84 .align 4 84 .align 4
85 $misaligned_dn: 85 $misaligned_dn:
86 nop 86 nop
87 fnop 87 fnop
88 unop 88 unop
89 beq $18,$egress 89 beq $18,$egress
90 90
91 $byte_loop_tail_dn: 91 $byte_loop_tail_dn:
92 ldq_u $3,-1($5) 92 ldq_u $3,-1($5)
93 ldq_u $2,-1($4) 93 ldq_u $2,-1($4)
94 lda $5,-1($5) 94 lda $5,-1($5)
95 lda $4,-1($4) 95 lda $4,-1($4)
96 96
97 lda $18,-1($18) 97 lda $18,-1($18)
98 extbl $3,$5,$1 98 extbl $3,$5,$1
99 insbl $1,$4,$1 99 insbl $1,$4,$1
100 mskbl $2,$4,$2 100 mskbl $2,$4,$2
101 101
102 bis $1,$2,$1 102 bis $1,$2,$1
103 stq_u $1,0($4) 103 stq_u $1,0($4)
104 bgt $18,$byte_loop_tail_dn 104 bgt $18,$byte_loop_tail_dn
105 br $egress 105 br $egress
106 106
107 $memmove_up: 107 $memmove_up:
108 mov $16,$4 108 mov $16,$4
109 mov $17,$5 109 mov $17,$5
110 bne $2,$misaligned_up 110 bne $2,$misaligned_up
111 beq $1,$skip_aligned_byte_loop_head_up 111 beq $1,$skip_aligned_byte_loop_head_up
112 112
113 $aligned_byte_loop_head_up: 113 $aligned_byte_loop_head_up:
114 unop 114 unop
115 ble $18,$egress 115 ble $18,$egress
116 ldq_u $3,0($5) 116 ldq_u $3,0($5)
117 ldq_u $2,0($4) 117 ldq_u $2,0($4)
118 118
119 lda $18,-1($18) 119 lda $18,-1($18)
120 extbl $3,$5,$1 120 extbl $3,$5,$1
121 insbl $1,$4,$1 121 insbl $1,$4,$1
122 mskbl $2,$4,$2 122 mskbl $2,$4,$2
123 123
124 bis $1,$2,$1 124 bis $1,$2,$1
125 lda $5,1($5) 125 lda $5,1($5)
126 stq_u $1,0($4) 126 stq_u $1,0($4)
127 lda $4,1($4) 127 lda $4,1($4)
128 128
129 and $4,7,$6 129 and $4,7,$6
130 bne $6,$aligned_byte_loop_head_up 130 bne $6,$aligned_byte_loop_head_up
131 131
132 $skip_aligned_byte_loop_head_up: 132 $skip_aligned_byte_loop_head_up:
133 lda $18,-8($18) 133 lda $18,-8($18)
134 blt $18,$skip_aligned_word_loop_up 134 blt $18,$skip_aligned_word_loop_up
135 135
136 $aligned_word_loop_up: 136 $aligned_word_loop_up:
137 ldq $1,0($5) 137 ldq $1,0($5)
138 nop 138 nop
139 lda $5,8($5) 139 lda $5,8($5)
140 lda $18,-8($18) 140 lda $18,-8($18)
141 141
142 stq $1,0($4) 142 stq $1,0($4)
143 nop 143 nop
144 lda $4,8($4) 144 lda $4,8($4)
145 bge $18,$aligned_word_loop_up 145 bge $18,$aligned_word_loop_up
146 146
147 $skip_aligned_word_loop_up: 147 $skip_aligned_word_loop_up:
148 lda $18,8($18) 148 lda $18,8($18)
149 bgt $18,$byte_loop_tail_up 149 bgt $18,$byte_loop_tail_up
150 unop 150 unop
151 ret $31,($26),1 151 ret $31,($26),1
152 152
153 .align 4 153 .align 4
154 $misaligned_up: 154 $misaligned_up:
155 nop 155 nop
156 fnop 156 fnop
157 unop 157 unop
158 beq $18,$egress 158 beq $18,$egress
159 159
160 $byte_loop_tail_up: 160 $byte_loop_tail_up:
161 ldq_u $3,0($5) 161 ldq_u $3,0($5)
162 ldq_u $2,0($4) 162 ldq_u $2,0($4)
163 lda $18,-1($18) 163 lda $18,-1($18)
164 extbl $3,$5,$1 164 extbl $3,$5,$1
165 165
166 insbl $1,$4,$1 166 insbl $1,$4,$1
167 mskbl $2,$4,$2 167 mskbl $2,$4,$2
168 bis $1,$2,$1 168 bis $1,$2,$1
169 stq_u $1,0($4) 169 stq_u $1,0($4)
170 170
171 lda $5,1($5) 171 lda $5,1($5)
172 lda $4,1($4) 172 lda $4,1($4)
173 nop 173 nop
174 bgt $18,$byte_loop_tail_up 174 bgt $18,$byte_loop_tail_up
175 175
176 $egress: 176 $egress:
177 ret $31,($26),1 177 ret $31,($26),1
178 nop 178 nop
179 nop 179 nop
180 nop 180 nop
181 181
182 .end memmove 182 .end memmove
183 EXPORT_SYMBOL(memmove) 183 EXPORT_SYMBOL(memmove)
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.