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 #include <asm/export.h>
11 .set noat 10 .set noat
12 .set noreorder 11 .set noreorder
13 .text 12 .text
14 13
15 .align 4 14 .align 4
16 .globl memmove 15 .globl memmove
17 .ent memmove 16 .ent memmove
18 memmove: 17 memmove:
19 ldgp $29, 0($27) 18 ldgp $29, 0($27)
20 unop 19 unop
21 nop 20 nop
22 .prologue 1 21 .prologue 1
23 22
24 addq $16,$18,$4 23 addq $16,$18,$4
25 addq $17,$18,$5 24 addq $17,$18,$5
26 cmpule $4,$17,$1 /* de 25 cmpule $4,$17,$1 /* dest + n <= src */
27 cmpule $5,$16,$2 /* de 26 cmpule $5,$16,$2 /* dest >= src + n */
28 27
29 bis $1,$2,$1 28 bis $1,$2,$1
30 mov $16,$0 29 mov $16,$0
31 xor $16,$17,$2 30 xor $16,$17,$2
32 bne $1,memcpy !sameg 31 bne $1,memcpy !samegp
33 32
34 and $2,7,$2 /* Tes 33 and $2,7,$2 /* Test for src/dest co-alignment. */
35 and $16,7,$1 34 and $16,7,$1
36 cmpule $16,$17,$3 35 cmpule $16,$17,$3
37 bne $3,$memmove_up /* des 36 bne $3,$memmove_up /* dest < src */
38 37
39 and $4,7,$1 38 and $4,7,$1
40 bne $2,$misaligned_dn 39 bne $2,$misaligned_dn
41 unop 40 unop
42 beq $1,$skip_aligned_byte_loop_head_dn 41 beq $1,$skip_aligned_byte_loop_head_dn
43 42
44 $aligned_byte_loop_head_dn: 43 $aligned_byte_loop_head_dn:
45 lda $4,-1($4) 44 lda $4,-1($4)
46 lda $5,-1($5) 45 lda $5,-1($5)
47 unop 46 unop
48 ble $18,$egress 47 ble $18,$egress
49 48
50 ldq_u $3,0($5) 49 ldq_u $3,0($5)
51 ldq_u $2,0($4) 50 ldq_u $2,0($4)
52 lda $18,-1($18) 51 lda $18,-1($18)
53 extbl $3,$5,$1 52 extbl $3,$5,$1
54 53
55 insbl $1,$4,$1 54 insbl $1,$4,$1
56 mskbl $2,$4,$2 55 mskbl $2,$4,$2
57 bis $1,$2,$1 56 bis $1,$2,$1
58 and $4,7,$6 57 and $4,7,$6
59 58
60 stq_u $1,0($4) 59 stq_u $1,0($4)
61 bne $6,$aligned_byte_loop_head_dn 60 bne $6,$aligned_byte_loop_head_dn
62 61
63 $skip_aligned_byte_loop_head_dn: 62 $skip_aligned_byte_loop_head_dn:
64 lda $18,-8($18) 63 lda $18,-8($18)
65 blt $18,$skip_aligned_word_loop_dn 64 blt $18,$skip_aligned_word_loop_dn
66 65
67 $aligned_word_loop_dn: 66 $aligned_word_loop_dn:
68 ldq $1,-8($5) 67 ldq $1,-8($5)
69 nop 68 nop
70 lda $5,-8($5) 69 lda $5,-8($5)
71 lda $18,-8($18) 70 lda $18,-8($18)
72 71
73 stq $1,-8($4) 72 stq $1,-8($4)
74 nop 73 nop
75 lda $4,-8($4) 74 lda $4,-8($4)
76 bge $18,$aligned_word_loop_dn 75 bge $18,$aligned_word_loop_dn
77 76
78 $skip_aligned_word_loop_dn: 77 $skip_aligned_word_loop_dn:
79 lda $18,8($18) 78 lda $18,8($18)
80 bgt $18,$byte_loop_tail_dn 79 bgt $18,$byte_loop_tail_dn
81 unop 80 unop
82 ret $31,($26),1 81 ret $31,($26),1
83 82
84 .align 4 83 .align 4
85 $misaligned_dn: 84 $misaligned_dn:
86 nop 85 nop
87 fnop 86 fnop
88 unop 87 unop
89 beq $18,$egress 88 beq $18,$egress
90 89
91 $byte_loop_tail_dn: 90 $byte_loop_tail_dn:
92 ldq_u $3,-1($5) 91 ldq_u $3,-1($5)
93 ldq_u $2,-1($4) 92 ldq_u $2,-1($4)
94 lda $5,-1($5) 93 lda $5,-1($5)
95 lda $4,-1($4) 94 lda $4,-1($4)
96 95
97 lda $18,-1($18) 96 lda $18,-1($18)
98 extbl $3,$5,$1 97 extbl $3,$5,$1
99 insbl $1,$4,$1 98 insbl $1,$4,$1
100 mskbl $2,$4,$2 99 mskbl $2,$4,$2
101 100
102 bis $1,$2,$1 101 bis $1,$2,$1
103 stq_u $1,0($4) 102 stq_u $1,0($4)
104 bgt $18,$byte_loop_tail_dn 103 bgt $18,$byte_loop_tail_dn
105 br $egress 104 br $egress
106 105
107 $memmove_up: 106 $memmove_up:
108 mov $16,$4 107 mov $16,$4
109 mov $17,$5 108 mov $17,$5
110 bne $2,$misaligned_up 109 bne $2,$misaligned_up
111 beq $1,$skip_aligned_byte_loop_head_up 110 beq $1,$skip_aligned_byte_loop_head_up
112 111
113 $aligned_byte_loop_head_up: 112 $aligned_byte_loop_head_up:
114 unop 113 unop
115 ble $18,$egress 114 ble $18,$egress
116 ldq_u $3,0($5) 115 ldq_u $3,0($5)
117 ldq_u $2,0($4) 116 ldq_u $2,0($4)
118 117
119 lda $18,-1($18) 118 lda $18,-1($18)
120 extbl $3,$5,$1 119 extbl $3,$5,$1
121 insbl $1,$4,$1 120 insbl $1,$4,$1
122 mskbl $2,$4,$2 121 mskbl $2,$4,$2
123 122
124 bis $1,$2,$1 123 bis $1,$2,$1
125 lda $5,1($5) 124 lda $5,1($5)
126 stq_u $1,0($4) 125 stq_u $1,0($4)
127 lda $4,1($4) 126 lda $4,1($4)
128 127
129 and $4,7,$6 128 and $4,7,$6
130 bne $6,$aligned_byte_loop_head_up 129 bne $6,$aligned_byte_loop_head_up
131 130
132 $skip_aligned_byte_loop_head_up: 131 $skip_aligned_byte_loop_head_up:
133 lda $18,-8($18) 132 lda $18,-8($18)
134 blt $18,$skip_aligned_word_loop_up 133 blt $18,$skip_aligned_word_loop_up
135 134
136 $aligned_word_loop_up: 135 $aligned_word_loop_up:
137 ldq $1,0($5) 136 ldq $1,0($5)
138 nop 137 nop
139 lda $5,8($5) 138 lda $5,8($5)
140 lda $18,-8($18) 139 lda $18,-8($18)
141 140
142 stq $1,0($4) 141 stq $1,0($4)
143 nop 142 nop
144 lda $4,8($4) 143 lda $4,8($4)
145 bge $18,$aligned_word_loop_up 144 bge $18,$aligned_word_loop_up
146 145
147 $skip_aligned_word_loop_up: 146 $skip_aligned_word_loop_up:
148 lda $18,8($18) 147 lda $18,8($18)
149 bgt $18,$byte_loop_tail_up 148 bgt $18,$byte_loop_tail_up
150 unop 149 unop
151 ret $31,($26),1 150 ret $31,($26),1
152 151
153 .align 4 152 .align 4
154 $misaligned_up: 153 $misaligned_up:
155 nop 154 nop
156 fnop 155 fnop
157 unop 156 unop
158 beq $18,$egress 157 beq $18,$egress
159 158
160 $byte_loop_tail_up: 159 $byte_loop_tail_up:
161 ldq_u $3,0($5) 160 ldq_u $3,0($5)
162 ldq_u $2,0($4) 161 ldq_u $2,0($4)
163 lda $18,-1($18) 162 lda $18,-1($18)
164 extbl $3,$5,$1 163 extbl $3,$5,$1
165 164
166 insbl $1,$4,$1 165 insbl $1,$4,$1
167 mskbl $2,$4,$2 166 mskbl $2,$4,$2
168 bis $1,$2,$1 167 bis $1,$2,$1
169 stq_u $1,0($4) 168 stq_u $1,0($4)
170 169
171 lda $5,1($5) 170 lda $5,1($5)
172 lda $4,1($4) 171 lda $4,1($4)
173 nop 172 nop
174 bgt $18,$byte_loop_tail_up 173 bgt $18,$byte_loop_tail_up
175 174
176 $egress: 175 $egress:
177 ret $31,($26),1 176 ret $31,($26),1
178 nop 177 nop
179 nop 178 nop
180 nop 179 nop
181 180
182 .end memmove 181 .end memmove
183 EXPORT_SYMBOL(memmove) 182 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.