1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3 * arch/alpha/lib/ev6-copy_user.S
4 *
5 * 21264 version contributed by Rick Gorton <ri
6 *
7 * Copy to/from user space, handling exception
8 * isn't exactly pretty.
9 *
10 * This is essentially the same as "memcpy()",
11 * Notably, we have to make sure that $0 is al
12 * contains the right "bytes left to copy" val
13 * only _after_ a successful copy). There is a
14 * exception setup stuff..
15 *
16 * Much of the information about 21264 schedul
17 * Compiler Writer's Guide for the Alpha
18 * abbreviated as 'CWG' in other comments
19 * ftp.digital.com/pub/Digital/info/semic
20 * Scheduling notation:
21 * E - either cluster
22 * U - upper subcluster; U0 - subcl
23 * L - lower subcluster; L0 - subcl
24 */
25
26 #include <linux/export.h>
27 /* Allow an exception for an insn; exit if we
28 #define EXI(x,y...) \
29 99: x,##y; \
30 .section __ex_table,"a"; \
31 .long 99b - .; \
32 lda $31, $exitin-99b($31); \
33 .previous
34
35 #define EXO(x,y...) \
36 99: x,##y; \
37 .section __ex_table,"a"; \
38 .long 99b - .; \
39 lda $31, $exitout-99b($31); \
40 .previous
41
42 .set noat
43 .align 4
44 .globl __copy_user
45 .ent __copy_user
46 # Pipeline inf
47 __copy_user:
48 .prologue 0
49 mov $18, $0 # .. .. .. E
50 subq $18, 32, $1 # .. .. E. ..
51 nop # .. E .. ..
52 beq $18, $zerolength # U .. .. ..
53
54 and $16,7,$3 # .. .. .. E
55 ble $1, $onebyteloop # .. .. U ..
56 beq $3, $destaligned # .. U .. ..
57 subq $3, 8, $3 # E .. .. ..
58 /*
59 * The fetcher stall also hides the 1 cycle cr
60 * This loop aligns the destination a byte at
61 * We know we have at least one trip through t
62 */
63 $aligndest:
64 EXI( ldbu $1,0($17) ) # .. .. .. L
65 addq $16,1,$16 # .. .. E ..
66 addq $3,1,$3 # .. E .. ..
67 nop # E .. .. ..
68
69 /*
70 * the -1 is to compensate for the inc($16) do
71 * which allows us zero dependencies within ei
72 */
73 EXO( stb $1,-1($16) ) # .. .. .. L
74 addq $17,1,$17 # .. .. E ..
75 subq $0,1,$0 # .. E .. ..
76 bne $3, $aligndest # U .. .. ..
77
78 /*
79 * If we fell through into here, we have a min
80 * If we arrived via branch, we have a minimum
81 */
82 $destaligned:
83 and $17,7,$1 # .. .. .. E
84 bic $0,7,$4 # .. .. E ..
85 EXI( ldq_u $3,0($17) ) # .. L .. ..
86 beq $1,$quadaligned # U .. .. ..
87
88 /*
89 * In the worst case, we've just executed an l
90 * and we'll repeat it once if we take the bra
91 */
92
93 /* Misaligned quadword loop - not unrolled. L
94 $misquad:
95 EXI( ldq_u $2,8($17) ) # .. .. .. L
96 subq $4,8,$4 # .. .. E ..
97 extql $3,$17,$3 # .. U .. ..
98 extqh $2,$17,$1 # U .. .. ..
99
100 bis $3,$1,$1 # .. .. .. E
101 EXO( stq $1,0($16) ) # .. .. L ..
102 addq $17,8,$17 # .. E .. ..
103 subq $0,8,$0 # E .. .. ..
104
105 addq $16,8,$16 # .. .. .. E
106 bis $2,$2,$3 # .. .. E ..
107 nop # .. E .. ..
108 bne $4,$misquad # U .. .. ..
109
110 nop # .. .. .. E
111 nop # .. .. E ..
112 nop # .. E .. ..
113 beq $0,$zerolength # U .. .. ..
114
115 /* We know we have at least one trip through t
116 EXI ( ldbu $2,0($17) ) # .. .. .. L
117 addq $16,1,$16 # .. .. E ..
118 nop # .. E .. ..
119 br $31, $dirtyentry # L0 .. .. ..
120 /* Do the trailing byte loop load, then hop in
121
122 /*
123 * A minimum of (33 - 7) bytes to do a quad at
124 * Based upon the usage context, it's worth th
125 * $0 - number of bytes to be moved
126 * $4 - number of bytes to move as quadwords
127 * $16 is current destination address
128 * $17 is current source address
129 */
130 $quadaligned:
131 subq $4, 32, $2 # .. .. .. E
132 nop # .. .. E ..
133 nop # .. E .. ..
134 blt $2, $onequad # U .. .. ..
135
136 /*
137 * There is a significant assumption here that
138 * addresses differ by more than 32 bytes. In
139 * sparsity of registers further bounds this t
140 * But if this isn't met, then the output resu
141 * Furthermore, due to a lack of available reg
142 * unroll this to be an 8x loop (which would e
143 * instruction memory hint instruction).
144 */
145 $unroll4:
146 EXI( ldq $1,0($17) ) # .. .. .. L
147 EXI( ldq $2,8($17) ) # .. .. L ..
148 subq $4,32,$4 # .. E .. ..
149 nop # E .. .. ..
150
151 addq $17,16,$17 # .. .. .. E
152 EXO( stq $1,0($16) ) # .. .. L ..
153 EXO( stq $2,8($16) ) # .. L .. ..
154 subq $0,16,$0 # E .. .. ..
155
156 addq $16,16,$16 # .. .. .. E
157 EXI( ldq $1,0($17) ) # .. .. L ..
158 EXI( ldq $2,8($17) ) # .. L .. ..
159 subq $4, 32, $3 # E .. .. ..
160
161 EXO( stq $1,0($16) ) # .. .. .. L
162 EXO( stq $2,8($16) ) # .. .. L ..
163 subq $0,16,$0 # .. E .. ..
164 addq $17,16,$17 # E .. .. ..
165
166 nop # .. .. .. E
167 nop # .. .. E ..
168 addq $16,16,$16 # .. E .. ..
169 bgt $3,$unroll4 # U .. .. ..
170
171 nop
172 nop
173 nop
174 beq $4, $noquads
175
176 $onequad:
177 EXI( ldq $1,0($17) )
178 subq $4,8,$4
179 addq $17,8,$17
180 nop
181
182 EXO( stq $1,0($16) )
183 subq $0,8,$0
184 addq $16,8,$16
185 bne $4,$onequad
186
187 $noquads:
188 nop
189 nop
190 nop
191 beq $0,$zerolength
192
193 /*
194 * For small copies (or the tail of a larger c
195 * There's no point in doing a lot of complex
196 * to quadword stuff for a small amount of dat
197 * $0 - remaining number of bytes left to
198 * $16 - current dest addr
199 * $17 - current source addr
200 */
201
202 $onebyteloop:
203 EXI ( ldbu $2,0($17) ) # .. .. .. L
204 addq $16,1,$16 # .. .. E ..
205 nop # .. E .. ..
206 nop # E .. .. ..
207
208 $dirtyentry:
209 /*
210 * the -1 is to compensate for the inc($16) do
211 * which allows us zero dependencies within ei
212 */
213 EXO ( stb $2,-1($16) ) # .. .. .. L
214 addq $17,1,$17 # .. .. E ..
215 subq $0,1,$0 # .. E .. ..
216 bgt $0,$onebyteloop # U .. .. ..
217
218 $zerolength:
219 $exitin:
220 $exitout: # Destination
221 nop # .. .. .. E
222 nop # .. .. E ..
223 nop # .. E .. ..
224 ret $31,($26),1 # L0 .. .. ..
225
226 .end __copy_user
227 EXPORT_SYMBOL(__copy_user)
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.