1 /* SPDX-License-Identifier: GPL-2.0 */ <<
2 /* 1 /*
3 * arch/alpha/lib/memmove.S 2 * arch/alpha/lib/memmove.S
4 * 3 *
5 * Barely optimized memmove routine for Alpha 4 * Barely optimized memmove routine for Alpha EV5.
6 * 5 *
7 * This is hand-massaged output from the origi 6 * This is hand-massaged output from the original memcpy.c. We defer to
8 * memcpy whenever possible; the backwards cop 7 * memcpy whenever possible; the backwards copy loops are not unrolled.
9 */ 8 */
10 #include <linux/export.h> !! 9
11 .set noat 10 .set noat
12 .set noreorder 11 .set noreorder
13 .text 12 .text
14 13
15 .align 4 14 .align 4
>> 15 .globl bcopy
>> 16 .ent bcopy
>> 17 bcopy:
>> 18 ldgp $29, 0($27)
>> 19 .prologue 1
>> 20 mov $16,$0
>> 21 mov $17,$16
>> 22 mov $0,$17
>> 23 br $31, memmove !samegp
>> 24 .end bcopy
>> 25
>> 26 .align 4
16 .globl memmove 27 .globl memmove
17 .ent memmove 28 .ent memmove
18 memmove: 29 memmove:
19 ldgp $29, 0($27) 30 ldgp $29, 0($27)
20 unop 31 unop
21 nop 32 nop
22 .prologue 1 33 .prologue 1
23 34
24 addq $16,$18,$4 35 addq $16,$18,$4
25 addq $17,$18,$5 36 addq $17,$18,$5
26 cmpule $4,$17,$1 /* de 37 cmpule $4,$17,$1 /* dest + n <= src */
27 cmpule $5,$16,$2 /* de 38 cmpule $5,$16,$2 /* dest >= src + n */
28 39
29 bis $1,$2,$1 40 bis $1,$2,$1
30 mov $16,$0 41 mov $16,$0
31 xor $16,$17,$2 42 xor $16,$17,$2
32 bne $1,memcpy !sameg 43 bne $1,memcpy !samegp
33 44
34 and $2,7,$2 /* Tes 45 and $2,7,$2 /* Test for src/dest co-alignment. */
35 and $16,7,$1 46 and $16,7,$1
36 cmpule $16,$17,$3 47 cmpule $16,$17,$3
37 bne $3,$memmove_up /* des 48 bne $3,$memmove_up /* dest < src */
38 49
39 and $4,7,$1 50 and $4,7,$1
40 bne $2,$misaligned_dn 51 bne $2,$misaligned_dn
41 unop 52 unop
42 beq $1,$skip_aligned_byte_loop_head_dn 53 beq $1,$skip_aligned_byte_loop_head_dn
43 54
44 $aligned_byte_loop_head_dn: 55 $aligned_byte_loop_head_dn:
45 lda $4,-1($4) 56 lda $4,-1($4)
46 lda $5,-1($5) 57 lda $5,-1($5)
47 unop 58 unop
48 ble $18,$egress 59 ble $18,$egress
49 60
50 ldq_u $3,0($5) 61 ldq_u $3,0($5)
51 ldq_u $2,0($4) 62 ldq_u $2,0($4)
52 lda $18,-1($18) 63 lda $18,-1($18)
53 extbl $3,$5,$1 64 extbl $3,$5,$1
54 65
55 insbl $1,$4,$1 66 insbl $1,$4,$1
56 mskbl $2,$4,$2 67 mskbl $2,$4,$2
57 bis $1,$2,$1 68 bis $1,$2,$1
58 and $4,7,$6 69 and $4,7,$6
59 70
60 stq_u $1,0($4) 71 stq_u $1,0($4)
61 bne $6,$aligned_byte_loop_head_dn 72 bne $6,$aligned_byte_loop_head_dn
62 73
63 $skip_aligned_byte_loop_head_dn: 74 $skip_aligned_byte_loop_head_dn:
64 lda $18,-8($18) 75 lda $18,-8($18)
65 blt $18,$skip_aligned_word_loop_dn 76 blt $18,$skip_aligned_word_loop_dn
66 77
67 $aligned_word_loop_dn: 78 $aligned_word_loop_dn:
68 ldq $1,-8($5) 79 ldq $1,-8($5)
69 nop 80 nop
70 lda $5,-8($5) 81 lda $5,-8($5)
71 lda $18,-8($18) 82 lda $18,-8($18)
72 83
73 stq $1,-8($4) 84 stq $1,-8($4)
74 nop 85 nop
75 lda $4,-8($4) 86 lda $4,-8($4)
76 bge $18,$aligned_word_loop_dn 87 bge $18,$aligned_word_loop_dn
77 88
78 $skip_aligned_word_loop_dn: 89 $skip_aligned_word_loop_dn:
79 lda $18,8($18) 90 lda $18,8($18)
80 bgt $18,$byte_loop_tail_dn 91 bgt $18,$byte_loop_tail_dn
81 unop 92 unop
82 ret $31,($26),1 93 ret $31,($26),1
83 94
84 .align 4 95 .align 4
85 $misaligned_dn: 96 $misaligned_dn:
86 nop 97 nop
87 fnop 98 fnop
88 unop 99 unop
89 beq $18,$egress 100 beq $18,$egress
90 101
91 $byte_loop_tail_dn: 102 $byte_loop_tail_dn:
92 ldq_u $3,-1($5) 103 ldq_u $3,-1($5)
93 ldq_u $2,-1($4) 104 ldq_u $2,-1($4)
94 lda $5,-1($5) 105 lda $5,-1($5)
95 lda $4,-1($4) 106 lda $4,-1($4)
96 107
97 lda $18,-1($18) 108 lda $18,-1($18)
98 extbl $3,$5,$1 109 extbl $3,$5,$1
99 insbl $1,$4,$1 110 insbl $1,$4,$1
100 mskbl $2,$4,$2 111 mskbl $2,$4,$2
101 112
102 bis $1,$2,$1 113 bis $1,$2,$1
103 stq_u $1,0($4) 114 stq_u $1,0($4)
104 bgt $18,$byte_loop_tail_dn 115 bgt $18,$byte_loop_tail_dn
105 br $egress 116 br $egress
106 117
107 $memmove_up: 118 $memmove_up:
108 mov $16,$4 119 mov $16,$4
109 mov $17,$5 120 mov $17,$5
110 bne $2,$misaligned_up 121 bne $2,$misaligned_up
111 beq $1,$skip_aligned_byte_loop_head_up 122 beq $1,$skip_aligned_byte_loop_head_up
112 123
113 $aligned_byte_loop_head_up: 124 $aligned_byte_loop_head_up:
114 unop 125 unop
115 ble $18,$egress 126 ble $18,$egress
116 ldq_u $3,0($5) 127 ldq_u $3,0($5)
117 ldq_u $2,0($4) 128 ldq_u $2,0($4)
118 129
119 lda $18,-1($18) 130 lda $18,-1($18)
120 extbl $3,$5,$1 131 extbl $3,$5,$1
121 insbl $1,$4,$1 132 insbl $1,$4,$1
122 mskbl $2,$4,$2 133 mskbl $2,$4,$2
123 134
124 bis $1,$2,$1 135 bis $1,$2,$1
125 lda $5,1($5) 136 lda $5,1($5)
126 stq_u $1,0($4) 137 stq_u $1,0($4)
127 lda $4,1($4) 138 lda $4,1($4)
128 139
129 and $4,7,$6 140 and $4,7,$6
130 bne $6,$aligned_byte_loop_head_up 141 bne $6,$aligned_byte_loop_head_up
131 142
132 $skip_aligned_byte_loop_head_up: 143 $skip_aligned_byte_loop_head_up:
133 lda $18,-8($18) 144 lda $18,-8($18)
134 blt $18,$skip_aligned_word_loop_up 145 blt $18,$skip_aligned_word_loop_up
135 146
136 $aligned_word_loop_up: 147 $aligned_word_loop_up:
137 ldq $1,0($5) 148 ldq $1,0($5)
138 nop 149 nop
139 lda $5,8($5) 150 lda $5,8($5)
140 lda $18,-8($18) 151 lda $18,-8($18)
141 152
142 stq $1,0($4) 153 stq $1,0($4)
143 nop 154 nop
144 lda $4,8($4) 155 lda $4,8($4)
145 bge $18,$aligned_word_loop_up 156 bge $18,$aligned_word_loop_up
146 157
147 $skip_aligned_word_loop_up: 158 $skip_aligned_word_loop_up:
148 lda $18,8($18) 159 lda $18,8($18)
149 bgt $18,$byte_loop_tail_up 160 bgt $18,$byte_loop_tail_up
150 unop 161 unop
151 ret $31,($26),1 162 ret $31,($26),1
152 163
153 .align 4 164 .align 4
154 $misaligned_up: 165 $misaligned_up:
155 nop 166 nop
156 fnop 167 fnop
157 unop 168 unop
158 beq $18,$egress 169 beq $18,$egress
159 170
160 $byte_loop_tail_up: 171 $byte_loop_tail_up:
161 ldq_u $3,0($5) 172 ldq_u $3,0($5)
162 ldq_u $2,0($4) 173 ldq_u $2,0($4)
163 lda $18,-1($18) 174 lda $18,-1($18)
164 extbl $3,$5,$1 175 extbl $3,$5,$1
165 176
166 insbl $1,$4,$1 177 insbl $1,$4,$1
167 mskbl $2,$4,$2 178 mskbl $2,$4,$2
168 bis $1,$2,$1 179 bis $1,$2,$1
169 stq_u $1,0($4) 180 stq_u $1,0($4)
170 181
171 lda $5,1($5) 182 lda $5,1($5)
172 lda $4,1($4) 183 lda $4,1($4)
173 nop 184 nop
174 bgt $18,$byte_loop_tail_up 185 bgt $18,$byte_loop_tail_up
175 186
176 $egress: 187 $egress:
177 ret $31,($26),1 188 ret $31,($26),1
178 nop 189 nop
179 nop 190 nop
180 nop 191 nop
181 192
182 .end memmove 193 .end 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.