1 /* SPDX-License-Identifier: GPL-2.0 */ 1 /* SPDX-License-Identifier: GPL-2.0 */
2 /* NGmemcpy.S: Niagara optimized memcpy. 2 /* NGmemcpy.S: Niagara optimized memcpy.
3 * 3 *
4 * Copyright (C) 2006, 2007 David S. Miller (d 4 * Copyright (C) 2006, 2007 David S. Miller (davem@davemloft.net)
5 */ 5 */
6 6
7 #ifdef __KERNEL__ 7 #ifdef __KERNEL__
8 #include <linux/linkage.h> 8 #include <linux/linkage.h>
9 #include <asm/asi.h> 9 #include <asm/asi.h>
10 #include <asm/thread_info.h> 10 #include <asm/thread_info.h>
11 #define GLOBAL_SPARE %g7 11 #define GLOBAL_SPARE %g7
12 #define RESTORE_ASI(TMP) \ 12 #define RESTORE_ASI(TMP) \
13 wr %g0, ASI_AIUS, %asi 13 wr %g0, ASI_AIUS, %asi
14 #else 14 #else
15 #define GLOBAL_SPARE %g5 15 #define GLOBAL_SPARE %g5
16 #define RESTORE_ASI(TMP) \ 16 #define RESTORE_ASI(TMP) \
17 wr %g0, ASI_PNF, %asi 17 wr %g0, ASI_PNF, %asi
18 #endif 18 #endif
19 19
20 #ifdef __sparc_v9__ 20 #ifdef __sparc_v9__
21 #define SAVE_AMOUNT 128 21 #define SAVE_AMOUNT 128
22 #else 22 #else
23 #define SAVE_AMOUNT 64 23 #define SAVE_AMOUNT 64
24 #endif 24 #endif
25 25
26 #ifndef STORE_ASI 26 #ifndef STORE_ASI
27 #define STORE_ASI ASI_BLK_INIT_QUAD_LDD_ 27 #define STORE_ASI ASI_BLK_INIT_QUAD_LDD_P
28 #endif 28 #endif
29 29
30 #ifndef EX_LD 30 #ifndef EX_LD
31 #define EX_LD(x,y) x 31 #define EX_LD(x,y) x
32 #endif 32 #endif
33 33
34 #ifndef EX_ST 34 #ifndef EX_ST
35 #define EX_ST(x,y) x 35 #define EX_ST(x,y) x
36 #endif 36 #endif
37 37
38 #ifndef LOAD 38 #ifndef LOAD
39 #ifndef MEMCPY_DEBUG 39 #ifndef MEMCPY_DEBUG
40 #define LOAD(type,addr,dest) type [addr], d 40 #define LOAD(type,addr,dest) type [addr], dest
41 #else 41 #else
42 #define LOAD(type,addr,dest) type##a [addr] 42 #define LOAD(type,addr,dest) type##a [addr] 0x80, dest
43 #endif 43 #endif
44 #endif 44 #endif
45 45
46 #ifndef LOAD_TWIN 46 #ifndef LOAD_TWIN
47 #define LOAD_TWIN(addr_reg,dest0,dest1) \ 47 #define LOAD_TWIN(addr_reg,dest0,dest1) \
48 ldda [addr_reg] ASI_BLK_INIT_QUAD_LDD_ 48 ldda [addr_reg] ASI_BLK_INIT_QUAD_LDD_P, dest0
49 #endif 49 #endif
50 50
51 #ifndef STORE 51 #ifndef STORE
52 #define STORE(type,src,addr) type src, [add 52 #define STORE(type,src,addr) type src, [addr]
53 #endif 53 #endif
54 54
55 #ifndef STORE_INIT 55 #ifndef STORE_INIT
56 #ifndef SIMULATE_NIAGARA_ON_NON_NIAGARA 56 #ifndef SIMULATE_NIAGARA_ON_NON_NIAGARA
57 #define STORE_INIT(src,addr) stxa src, [add 57 #define STORE_INIT(src,addr) stxa src, [addr] %asi
58 #else 58 #else
59 #define STORE_INIT(src,addr) stx src, [addr 59 #define STORE_INIT(src,addr) stx src, [addr + 0x00]
60 #endif 60 #endif
61 #endif 61 #endif
62 62
63 #ifndef FUNC_NAME 63 #ifndef FUNC_NAME
64 #define FUNC_NAME NGmemcpy 64 #define FUNC_NAME NGmemcpy
65 #endif 65 #endif
66 66
67 #ifndef PREAMBLE 67 #ifndef PREAMBLE
68 #define PREAMBLE 68 #define PREAMBLE
69 #endif 69 #endif
70 70
71 #ifndef XCC 71 #ifndef XCC
72 #define XCC xcc 72 #define XCC xcc
73 #endif 73 #endif
74 74
75 .register %g2,#scratch 75 .register %g2,#scratch
76 .register %g3,#scratch 76 .register %g3,#scratch
77 77
78 .text 78 .text
79 #ifndef EX_RETVAL 79 #ifndef EX_RETVAL
80 #define EX_RETVAL(x) x 80 #define EX_RETVAL(x) x
81 __restore_asi: 81 __restore_asi:
82 ret 82 ret
83 wr %g0, ASI_AIUS, %asi 83 wr %g0, ASI_AIUS, %asi
84 restore 84 restore
85 ENTRY(NG_ret_i2_plus_i4_plus_1) 85 ENTRY(NG_ret_i2_plus_i4_plus_1)
86 ba,pt %xcc, __restore_asi 86 ba,pt %xcc, __restore_asi
87 add %i2, %i5, %i0 87 add %i2, %i5, %i0
88 ENDPROC(NG_ret_i2_plus_i4_plus_1) 88 ENDPROC(NG_ret_i2_plus_i4_plus_1)
89 ENTRY(NG_ret_i2_plus_g1) 89 ENTRY(NG_ret_i2_plus_g1)
90 ba,pt %xcc, __restore_asi 90 ba,pt %xcc, __restore_asi
91 add %i2, %g1, %i0 91 add %i2, %g1, %i0
92 ENDPROC(NG_ret_i2_plus_g1) 92 ENDPROC(NG_ret_i2_plus_g1)
93 ENTRY(NG_ret_i2_plus_g1_minus_8) 93 ENTRY(NG_ret_i2_plus_g1_minus_8)
94 sub %g1, 8, %g1 94 sub %g1, 8, %g1
95 ba,pt %xcc, __restore_asi 95 ba,pt %xcc, __restore_asi
96 add %i2, %g1, %i0 96 add %i2, %g1, %i0
97 ENDPROC(NG_ret_i2_plus_g1_minus_8) 97 ENDPROC(NG_ret_i2_plus_g1_minus_8)
98 ENTRY(NG_ret_i2_plus_g1_minus_16) 98 ENTRY(NG_ret_i2_plus_g1_minus_16)
99 sub %g1, 16, %g1 99 sub %g1, 16, %g1
100 ba,pt %xcc, __restore_asi 100 ba,pt %xcc, __restore_asi
101 add %i2, %g1, %i0 101 add %i2, %g1, %i0
102 ENDPROC(NG_ret_i2_plus_g1_minus_16) 102 ENDPROC(NG_ret_i2_plus_g1_minus_16)
103 ENTRY(NG_ret_i2_plus_g1_minus_24) 103 ENTRY(NG_ret_i2_plus_g1_minus_24)
104 sub %g1, 24, %g1 104 sub %g1, 24, %g1
105 ba,pt %xcc, __restore_asi 105 ba,pt %xcc, __restore_asi
106 add %i2, %g1, %i0 106 add %i2, %g1, %i0
107 ENDPROC(NG_ret_i2_plus_g1_minus_24) 107 ENDPROC(NG_ret_i2_plus_g1_minus_24)
108 ENTRY(NG_ret_i2_plus_g1_minus_32) 108 ENTRY(NG_ret_i2_plus_g1_minus_32)
109 sub %g1, 32, %g1 109 sub %g1, 32, %g1
110 ba,pt %xcc, __restore_asi 110 ba,pt %xcc, __restore_asi
111 add %i2, %g1, %i0 111 add %i2, %g1, %i0
112 ENDPROC(NG_ret_i2_plus_g1_minus_32) 112 ENDPROC(NG_ret_i2_plus_g1_minus_32)
113 ENTRY(NG_ret_i2_plus_g1_minus_40) 113 ENTRY(NG_ret_i2_plus_g1_minus_40)
114 sub %g1, 40, %g1 114 sub %g1, 40, %g1
115 ba,pt %xcc, __restore_asi 115 ba,pt %xcc, __restore_asi
116 add %i2, %g1, %i0 116 add %i2, %g1, %i0
117 ENDPROC(NG_ret_i2_plus_g1_minus_40) 117 ENDPROC(NG_ret_i2_plus_g1_minus_40)
118 ENTRY(NG_ret_i2_plus_g1_minus_48) 118 ENTRY(NG_ret_i2_plus_g1_minus_48)
119 sub %g1, 48, %g1 119 sub %g1, 48, %g1
120 ba,pt %xcc, __restore_asi 120 ba,pt %xcc, __restore_asi
121 add %i2, %g1, %i0 121 add %i2, %g1, %i0
122 ENDPROC(NG_ret_i2_plus_g1_minus_48) 122 ENDPROC(NG_ret_i2_plus_g1_minus_48)
123 ENTRY(NG_ret_i2_plus_g1_minus_56) 123 ENTRY(NG_ret_i2_plus_g1_minus_56)
124 sub %g1, 56, %g1 124 sub %g1, 56, %g1
125 ba,pt %xcc, __restore_asi 125 ba,pt %xcc, __restore_asi
126 add %i2, %g1, %i0 126 add %i2, %g1, %i0
127 ENDPROC(NG_ret_i2_plus_g1_minus_56) 127 ENDPROC(NG_ret_i2_plus_g1_minus_56)
128 ENTRY(NG_ret_i2_plus_i4) 128 ENTRY(NG_ret_i2_plus_i4)
129 ba,pt %xcc, __restore_asi 129 ba,pt %xcc, __restore_asi
130 add %i2, %i4, %i0 130 add %i2, %i4, %i0
131 ENDPROC(NG_ret_i2_plus_i4) 131 ENDPROC(NG_ret_i2_plus_i4)
132 ENTRY(NG_ret_i2_plus_i4_minus_8) 132 ENTRY(NG_ret_i2_plus_i4_minus_8)
133 sub %i4, 8, %i4 133 sub %i4, 8, %i4
134 ba,pt %xcc, __restore_asi 134 ba,pt %xcc, __restore_asi
135 add %i2, %i4, %i0 135 add %i2, %i4, %i0
136 ENDPROC(NG_ret_i2_plus_i4_minus_8) 136 ENDPROC(NG_ret_i2_plus_i4_minus_8)
137 ENTRY(NG_ret_i2_plus_8) 137 ENTRY(NG_ret_i2_plus_8)
138 ba,pt %xcc, __restore_asi 138 ba,pt %xcc, __restore_asi
139 add %i2, 8, %i0 139 add %i2, 8, %i0
140 ENDPROC(NG_ret_i2_plus_8) 140 ENDPROC(NG_ret_i2_plus_8)
141 ENTRY(NG_ret_i2_plus_4) 141 ENTRY(NG_ret_i2_plus_4)
142 ba,pt %xcc, __restore_asi 142 ba,pt %xcc, __restore_asi
143 add %i2, 4, %i0 143 add %i2, 4, %i0
144 ENDPROC(NG_ret_i2_plus_4) 144 ENDPROC(NG_ret_i2_plus_4)
145 ENTRY(NG_ret_i2_plus_1) 145 ENTRY(NG_ret_i2_plus_1)
146 ba,pt %xcc, __restore_asi 146 ba,pt %xcc, __restore_asi
147 add %i2, 1, %i0 147 add %i2, 1, %i0
148 ENDPROC(NG_ret_i2_plus_1) 148 ENDPROC(NG_ret_i2_plus_1)
149 ENTRY(NG_ret_i2_plus_g1_plus_1) 149 ENTRY(NG_ret_i2_plus_g1_plus_1)
150 add %g1, 1, %g1 150 add %g1, 1, %g1
151 ba,pt %xcc, __restore_asi 151 ba,pt %xcc, __restore_asi
152 add %i2, %g1, %i0 152 add %i2, %g1, %i0
153 ENDPROC(NG_ret_i2_plus_g1_plus_1) 153 ENDPROC(NG_ret_i2_plus_g1_plus_1)
154 ENTRY(NG_ret_i2) 154 ENTRY(NG_ret_i2)
155 ba,pt %xcc, __restore_asi 155 ba,pt %xcc, __restore_asi
156 mov %i2, %i0 156 mov %i2, %i0
157 ENDPROC(NG_ret_i2) 157 ENDPROC(NG_ret_i2)
158 ENTRY(NG_ret_i2_and_7_plus_i4) 158 ENTRY(NG_ret_i2_and_7_plus_i4)
159 and %i2, 7, %i2 159 and %i2, 7, %i2
160 ba,pt %xcc, __restore_asi 160 ba,pt %xcc, __restore_asi
161 add %i2, %i4, %i0 161 add %i2, %i4, %i0
162 ENDPROC(NG_ret_i2_and_7_plus_i4) 162 ENDPROC(NG_ret_i2_and_7_plus_i4)
163 #endif 163 #endif
164 164
165 .align 64 165 .align 64
166 166
167 .globl FUNC_NAME 167 .globl FUNC_NAME
168 .type FUNC_NAME,#function 168 .type FUNC_NAME,#function
169 FUNC_NAME: /* %i0=dst, %i1=src, %i2=len * 169 FUNC_NAME: /* %i0=dst, %i1=src, %i2=len */
170 PREAMBLE 170 PREAMBLE
171 save %sp, -SAVE_AMOUNT, %sp 171 save %sp, -SAVE_AMOUNT, %sp
172 srlx %i2, 31, %g2 172 srlx %i2, 31, %g2
173 cmp %g2, 0 173 cmp %g2, 0
174 tne %xcc, 5 174 tne %xcc, 5
175 mov %i0, %o0 175 mov %i0, %o0
176 cmp %i2, 0 176 cmp %i2, 0
177 be,pn %XCC, 85f 177 be,pn %XCC, 85f
178 or %o0, %i1, %i3 178 or %o0, %i1, %i3
179 cmp %i2, 16 179 cmp %i2, 16
180 blu,a,pn %XCC, 80f 180 blu,a,pn %XCC, 80f
181 or %i3, %i2, %i3 181 or %i3, %i2, %i3
182 182
183 /* 2 blocks (128 bytes) is the minimum 183 /* 2 blocks (128 bytes) is the minimum we can do the block
184 * copy with. We need to ensure that 184 * copy with. We need to ensure that we'll iterate at least
185 * once in the block copy loop. At wo 185 * once in the block copy loop. At worst we'll need to align
186 * the destination to a 64-byte bounda 186 * the destination to a 64-byte boundary which can chew up
187 * to (64 - 1) bytes from the length b 187 * to (64 - 1) bytes from the length before we perform the
188 * block copy loop. 188 * block copy loop.
189 */ 189 */
190 cmp %i2, (2 * 64) 190 cmp %i2, (2 * 64)
191 blu,pt %XCC, 70f 191 blu,pt %XCC, 70f
192 andcc %i3, 0x7, %g0 192 andcc %i3, 0x7, %g0
193 193
194 /* %o0: dst 194 /* %o0: dst
195 * %i1: src 195 * %i1: src
196 * %i2: len (known to be >= 128) 196 * %i2: len (known to be >= 128)
197 * 197 *
198 * The block copy loops will use %i4/% 198 * The block copy loops will use %i4/%i5,%g2/%g3 as
199 * temporaries while copying the data. 199 * temporaries while copying the data.
200 */ 200 */
201 201
202 LOAD(prefetch, %i1, #one_read) 202 LOAD(prefetch, %i1, #one_read)
203 wr %g0, STORE_ASI, %asi 203 wr %g0, STORE_ASI, %asi
204 204
205 /* Align destination on 64-byte bounda 205 /* Align destination on 64-byte boundary. */
206 andcc %o0, (64 - 1), %i4 206 andcc %o0, (64 - 1), %i4
207 be,pt %XCC, 2f 207 be,pt %XCC, 2f
208 sub %i4, 64, %i4 208 sub %i4, 64, %i4
209 sub %g0, %i4, %i4 ! byte 209 sub %g0, %i4, %i4 ! bytes to align dst
210 sub %i2, %i4, %i2 210 sub %i2, %i4, %i2
211 1: subcc %i4, 1, %i4 211 1: subcc %i4, 1, %i4
212 EX_LD(LOAD(ldub, %i1, %g1), NG_ret_i2_ 212 EX_LD(LOAD(ldub, %i1, %g1), NG_ret_i2_plus_i4_plus_1)
213 EX_ST(STORE(stb, %g1, %o0), NG_ret_i2_ 213 EX_ST(STORE(stb, %g1, %o0), NG_ret_i2_plus_i4_plus_1)
214 add %i1, 1, %i1 214 add %i1, 1, %i1
215 bne,pt %XCC, 1b 215 bne,pt %XCC, 1b
216 add %o0, 1, %o0 216 add %o0, 1, %o0
217 217
218 /* If the source is on a 16-byte bound 218 /* If the source is on a 16-byte boundary we can do
219 * the direct block copy loop. If it 219 * the direct block copy loop. If it is 8-byte aligned
220 * we can do the 16-byte loads offset 220 * we can do the 16-byte loads offset by -8 bytes and the
221 * init stores offset by one register. 221 * init stores offset by one register.
222 * 222 *
223 * If the source is not even 8-byte al 223 * If the source is not even 8-byte aligned, we need to do
224 * shifting and masking (basically int 224 * shifting and masking (basically integer faligndata).
225 * 225 *
226 * The careful bit with init stores is 226 * The careful bit with init stores is that if we store
227 * to any part of the cache line we ha 227 * to any part of the cache line we have to store the whole
228 * cacheline else we can end up with c 228 * cacheline else we can end up with corrupt L2 cache line
229 * contents. Since the loop works on 229 * contents. Since the loop works on 64-bytes of 64-byte
230 * aligned store data at a time, this 230 * aligned store data at a time, this is easy to ensure.
231 */ 231 */
232 2: 232 2:
233 andcc %i1, (16 - 1), %i4 233 andcc %i1, (16 - 1), %i4
234 andn %i2, (64 - 1), %g1 234 andn %i2, (64 - 1), %g1 ! block copy loop iterator
235 be,pt %XCC, 50f 235 be,pt %XCC, 50f
236 sub %i2, %g1, %i2 236 sub %i2, %g1, %i2 ! final sub-block copy bytes
237 237
238 cmp %i4, 8 238 cmp %i4, 8
239 be,pt %XCC, 10f 239 be,pt %XCC, 10f
240 sub %i1, %i4, %i1 240 sub %i1, %i4, %i1
241 241
242 /* Neither 8-byte nor 16-byte aligned, 242 /* Neither 8-byte nor 16-byte aligned, shift and mask. */
243 and %i4, 0x7, GLOBAL_SPARE 243 and %i4, 0x7, GLOBAL_SPARE
244 sll GLOBAL_SPARE, 3, GLOBA 244 sll GLOBAL_SPARE, 3, GLOBAL_SPARE
245 mov 64, %i5 245 mov 64, %i5
246 EX_LD(LOAD_TWIN(%i1, %g2, %g3), NG_ret 246 EX_LD(LOAD_TWIN(%i1, %g2, %g3), NG_ret_i2_plus_g1)
247 sub %i5, GLOBAL_SPARE, %i5 247 sub %i5, GLOBAL_SPARE, %i5
248 mov 16, %o4 248 mov 16, %o4
249 mov 32, %o5 249 mov 32, %o5
250 mov 48, %o7 250 mov 48, %o7
251 mov 64, %i3 251 mov 64, %i3
252 252
253 bg,pn %XCC, 9f 253 bg,pn %XCC, 9f
254 nop 254 nop
255 255
256 #define MIX_THREE_WORDS(WORD1, WORD2, WORD3, P 256 #define MIX_THREE_WORDS(WORD1, WORD2, WORD3, PRE_SHIFT, POST_SHIFT, TMP) \
257 sllx WORD1, POST_SHIFT, WOR 257 sllx WORD1, POST_SHIFT, WORD1; \
258 srlx WORD2, PRE_SHIFT, TMP; 258 srlx WORD2, PRE_SHIFT, TMP; \
259 sllx WORD2, POST_SHIFT, WOR 259 sllx WORD2, POST_SHIFT, WORD2; \
260 or WORD1, TMP, WORD1; \ 260 or WORD1, TMP, WORD1; \
261 srlx WORD3, PRE_SHIFT, TMP; 261 srlx WORD3, PRE_SHIFT, TMP; \
262 or WORD2, TMP, WORD2; 262 or WORD2, TMP, WORD2;
263 263
264 8: EX_LD(LOAD_TWIN(%i1 + %o4, %o2, %o3), 264 8: EX_LD(LOAD_TWIN(%i1 + %o4, %o2, %o3), NG_ret_i2_plus_g1)
265 MIX_THREE_WORDS(%g2, %g3, %o2, %i5, GL 265 MIX_THREE_WORDS(%g2, %g3, %o2, %i5, GLOBAL_SPARE, %o1)
266 LOAD(prefetch, %i1 + %i3, #one_read) 266 LOAD(prefetch, %i1 + %i3, #one_read)
267 267
268 EX_ST(STORE_INIT(%g2, %o0 + 0x00), NG_ 268 EX_ST(STORE_INIT(%g2, %o0 + 0x00), NG_ret_i2_plus_g1)
269 EX_ST(STORE_INIT(%g3, %o0 + 0x08), NG_ 269 EX_ST(STORE_INIT(%g3, %o0 + 0x08), NG_ret_i2_plus_g1_minus_8)
270 270
271 EX_LD(LOAD_TWIN(%i1 + %o5, %g2, %g3), 271 EX_LD(LOAD_TWIN(%i1 + %o5, %g2, %g3), NG_ret_i2_plus_g1_minus_16)
272 MIX_THREE_WORDS(%o2, %o3, %g2, %i5, GL 272 MIX_THREE_WORDS(%o2, %o3, %g2, %i5, GLOBAL_SPARE, %o1)
273 273
274 EX_ST(STORE_INIT(%o2, %o0 + 0x10), NG_ 274 EX_ST(STORE_INIT(%o2, %o0 + 0x10), NG_ret_i2_plus_g1_minus_16)
275 EX_ST(STORE_INIT(%o3, %o0 + 0x18), NG_ 275 EX_ST(STORE_INIT(%o3, %o0 + 0x18), NG_ret_i2_plus_g1_minus_24)
276 276
277 EX_LD(LOAD_TWIN(%i1 + %o7, %o2, %o3), 277 EX_LD(LOAD_TWIN(%i1 + %o7, %o2, %o3), NG_ret_i2_plus_g1_minus_32)
278 MIX_THREE_WORDS(%g2, %g3, %o2, %i5, GL 278 MIX_THREE_WORDS(%g2, %g3, %o2, %i5, GLOBAL_SPARE, %o1)
279 279
280 EX_ST(STORE_INIT(%g2, %o0 + 0x20), NG_ 280 EX_ST(STORE_INIT(%g2, %o0 + 0x20), NG_ret_i2_plus_g1_minus_32)
281 EX_ST(STORE_INIT(%g3, %o0 + 0x28), NG_ 281 EX_ST(STORE_INIT(%g3, %o0 + 0x28), NG_ret_i2_plus_g1_minus_40)
282 282
283 EX_LD(LOAD_TWIN(%i1 + %i3, %g2, %g3), 283 EX_LD(LOAD_TWIN(%i1 + %i3, %g2, %g3), NG_ret_i2_plus_g1_minus_48)
284 add %i1, 64, %i1 284 add %i1, 64, %i1
285 MIX_THREE_WORDS(%o2, %o3, %g2, %i5, GL 285 MIX_THREE_WORDS(%o2, %o3, %g2, %i5, GLOBAL_SPARE, %o1)
286 286
287 EX_ST(STORE_INIT(%o2, %o0 + 0x30), NG_ 287 EX_ST(STORE_INIT(%o2, %o0 + 0x30), NG_ret_i2_plus_g1_minus_48)
288 EX_ST(STORE_INIT(%o3, %o0 + 0x38), NG_ 288 EX_ST(STORE_INIT(%o3, %o0 + 0x38), NG_ret_i2_plus_g1_minus_56)
289 289
290 subcc %g1, 64, %g1 290 subcc %g1, 64, %g1
291 bne,pt %XCC, 8b 291 bne,pt %XCC, 8b
292 add %o0, 64, %o0 292 add %o0, 64, %o0
293 293
294 ba,pt %XCC, 60f 294 ba,pt %XCC, 60f
295 add %i1, %i4, %i1 295 add %i1, %i4, %i1
296 296
297 9: EX_LD(LOAD_TWIN(%i1 + %o4, %o2, %o3), 297 9: EX_LD(LOAD_TWIN(%i1 + %o4, %o2, %o3), NG_ret_i2_plus_g1)
298 MIX_THREE_WORDS(%g3, %o2, %o3, %i5, GL 298 MIX_THREE_WORDS(%g3, %o2, %o3, %i5, GLOBAL_SPARE, %o1)
299 LOAD(prefetch, %i1 + %i3, #one_read) 299 LOAD(prefetch, %i1 + %i3, #one_read)
300 300
301 EX_ST(STORE_INIT(%g3, %o0 + 0x00), NG_ 301 EX_ST(STORE_INIT(%g3, %o0 + 0x00), NG_ret_i2_plus_g1)
302 EX_ST(STORE_INIT(%o2, %o0 + 0x08), NG_ 302 EX_ST(STORE_INIT(%o2, %o0 + 0x08), NG_ret_i2_plus_g1_minus_8)
303 303
304 EX_LD(LOAD_TWIN(%i1 + %o5, %g2, %g3), 304 EX_LD(LOAD_TWIN(%i1 + %o5, %g2, %g3), NG_ret_i2_plus_g1_minus_16)
305 MIX_THREE_WORDS(%o3, %g2, %g3, %i5, GL 305 MIX_THREE_WORDS(%o3, %g2, %g3, %i5, GLOBAL_SPARE, %o1)
306 306
307 EX_ST(STORE_INIT(%o3, %o0 + 0x10), NG_ 307 EX_ST(STORE_INIT(%o3, %o0 + 0x10), NG_ret_i2_plus_g1_minus_16)
308 EX_ST(STORE_INIT(%g2, %o0 + 0x18), NG_ 308 EX_ST(STORE_INIT(%g2, %o0 + 0x18), NG_ret_i2_plus_g1_minus_24)
309 309
310 EX_LD(LOAD_TWIN(%i1 + %o7, %o2, %o3), 310 EX_LD(LOAD_TWIN(%i1 + %o7, %o2, %o3), NG_ret_i2_plus_g1_minus_32)
311 MIX_THREE_WORDS(%g3, %o2, %o3, %i5, GL 311 MIX_THREE_WORDS(%g3, %o2, %o3, %i5, GLOBAL_SPARE, %o1)
312 312
313 EX_ST(STORE_INIT(%g3, %o0 + 0x20), NG_ 313 EX_ST(STORE_INIT(%g3, %o0 + 0x20), NG_ret_i2_plus_g1_minus_32)
314 EX_ST(STORE_INIT(%o2, %o0 + 0x28), NG_ 314 EX_ST(STORE_INIT(%o2, %o0 + 0x28), NG_ret_i2_plus_g1_minus_40)
315 315
316 EX_LD(LOAD_TWIN(%i1 + %i3, %g2, %g3), 316 EX_LD(LOAD_TWIN(%i1 + %i3, %g2, %g3), NG_ret_i2_plus_g1_minus_48)
317 add %i1, 64, %i1 317 add %i1, 64, %i1
318 MIX_THREE_WORDS(%o3, %g2, %g3, %i5, GL 318 MIX_THREE_WORDS(%o3, %g2, %g3, %i5, GLOBAL_SPARE, %o1)
319 319
320 EX_ST(STORE_INIT(%o3, %o0 + 0x30), NG_ 320 EX_ST(STORE_INIT(%o3, %o0 + 0x30), NG_ret_i2_plus_g1_minus_48)
321 EX_ST(STORE_INIT(%g2, %o0 + 0x38), NG_ 321 EX_ST(STORE_INIT(%g2, %o0 + 0x38), NG_ret_i2_plus_g1_minus_56)
322 322
323 subcc %g1, 64, %g1 323 subcc %g1, 64, %g1
324 bne,pt %XCC, 9b 324 bne,pt %XCC, 9b
325 add %o0, 64, %o0 325 add %o0, 64, %o0
326 326
327 ba,pt %XCC, 60f 327 ba,pt %XCC, 60f
328 add %i1, %i4, %i1 328 add %i1, %i4, %i1
329 329
330 10: /* Destination is 64-byte aligned, sou 330 10: /* Destination is 64-byte aligned, source was only 8-byte
331 * aligned but it has been subtracted 331 * aligned but it has been subtracted by 8 and we perform
332 * one twin load ahead, then add 8 bac 332 * one twin load ahead, then add 8 back into source when
333 * we finish the loop. 333 * we finish the loop.
334 */ 334 */
335 EX_LD(LOAD_TWIN(%i1, %o4, %o5), NG_ret 335 EX_LD(LOAD_TWIN(%i1, %o4, %o5), NG_ret_i2_plus_g1)
336 mov 16, %o7 336 mov 16, %o7
337 mov 32, %g2 337 mov 32, %g2
338 mov 48, %g3 338 mov 48, %g3
339 mov 64, %o1 339 mov 64, %o1
340 1: EX_LD(LOAD_TWIN(%i1 + %o7, %o2, %o3), 340 1: EX_LD(LOAD_TWIN(%i1 + %o7, %o2, %o3), NG_ret_i2_plus_g1)
341 LOAD(prefetch, %i1 + %o1, #one_read) 341 LOAD(prefetch, %i1 + %o1, #one_read)
342 EX_ST(STORE_INIT(%o5, %o0 + 0x00), NG_ 342 EX_ST(STORE_INIT(%o5, %o0 + 0x00), NG_ret_i2_plus_g1) ! initializes cache line
343 EX_ST(STORE_INIT(%o2, %o0 + 0x08), NG_ 343 EX_ST(STORE_INIT(%o2, %o0 + 0x08), NG_ret_i2_plus_g1_minus_8)
344 EX_LD(LOAD_TWIN(%i1 + %g2, %o4, %o5), 344 EX_LD(LOAD_TWIN(%i1 + %g2, %o4, %o5), NG_ret_i2_plus_g1_minus_16)
345 EX_ST(STORE_INIT(%o3, %o0 + 0x10), NG_ 345 EX_ST(STORE_INIT(%o3, %o0 + 0x10), NG_ret_i2_plus_g1_minus_16)
346 EX_ST(STORE_INIT(%o4, %o0 + 0x18), NG_ 346 EX_ST(STORE_INIT(%o4, %o0 + 0x18), NG_ret_i2_plus_g1_minus_24)
347 EX_LD(LOAD_TWIN(%i1 + %g3, %o2, %o3), 347 EX_LD(LOAD_TWIN(%i1 + %g3, %o2, %o3), NG_ret_i2_plus_g1_minus_32)
348 EX_ST(STORE_INIT(%o5, %o0 + 0x20), NG_ 348 EX_ST(STORE_INIT(%o5, %o0 + 0x20), NG_ret_i2_plus_g1_minus_32)
349 EX_ST(STORE_INIT(%o2, %o0 + 0x28), NG_ 349 EX_ST(STORE_INIT(%o2, %o0 + 0x28), NG_ret_i2_plus_g1_minus_40)
350 EX_LD(LOAD_TWIN(%i1 + %o1, %o4, %o5), 350 EX_LD(LOAD_TWIN(%i1 + %o1, %o4, %o5), NG_ret_i2_plus_g1_minus_48)
351 add %i1, 64, %i1 351 add %i1, 64, %i1
352 EX_ST(STORE_INIT(%o3, %o0 + 0x30), NG_ 352 EX_ST(STORE_INIT(%o3, %o0 + 0x30), NG_ret_i2_plus_g1_minus_48)
353 EX_ST(STORE_INIT(%o4, %o0 + 0x38), NG_ 353 EX_ST(STORE_INIT(%o4, %o0 + 0x38), NG_ret_i2_plus_g1_minus_56)
354 subcc %g1, 64, %g1 354 subcc %g1, 64, %g1
355 bne,pt %XCC, 1b 355 bne,pt %XCC, 1b
356 add %o0, 64, %o0 356 add %o0, 64, %o0
357 357
358 ba,pt %XCC, 60f 358 ba,pt %XCC, 60f
359 add %i1, 0x8, %i1 359 add %i1, 0x8, %i1
360 360
361 50: /* Destination is 64-byte aligned, and 361 50: /* Destination is 64-byte aligned, and source is 16-byte
362 * aligned. 362 * aligned.
363 */ 363 */
364 mov 16, %o7 364 mov 16, %o7
365 mov 32, %g2 365 mov 32, %g2
366 mov 48, %g3 366 mov 48, %g3
367 mov 64, %o1 367 mov 64, %o1
368 1: EX_LD(LOAD_TWIN(%i1 + %g0, %o4, %o5), 368 1: EX_LD(LOAD_TWIN(%i1 + %g0, %o4, %o5), NG_ret_i2_plus_g1)
369 EX_LD(LOAD_TWIN(%i1 + %o7, %o2, %o3), 369 EX_LD(LOAD_TWIN(%i1 + %o7, %o2, %o3), NG_ret_i2_plus_g1)
370 LOAD(prefetch, %i1 + %o1, #one_read) 370 LOAD(prefetch, %i1 + %o1, #one_read)
371 EX_ST(STORE_INIT(%o4, %o0 + 0x00), NG_ 371 EX_ST(STORE_INIT(%o4, %o0 + 0x00), NG_ret_i2_plus_g1) ! initializes cache line
372 EX_ST(STORE_INIT(%o5, %o0 + 0x08), NG_ 372 EX_ST(STORE_INIT(%o5, %o0 + 0x08), NG_ret_i2_plus_g1_minus_8)
373 EX_LD(LOAD_TWIN(%i1 + %g2, %o4, %o5), 373 EX_LD(LOAD_TWIN(%i1 + %g2, %o4, %o5), NG_ret_i2_plus_g1_minus_16)
374 EX_ST(STORE_INIT(%o2, %o0 + 0x10), NG_ 374 EX_ST(STORE_INIT(%o2, %o0 + 0x10), NG_ret_i2_plus_g1_minus_16)
375 EX_ST(STORE_INIT(%o3, %o0 + 0x18), NG_ 375 EX_ST(STORE_INIT(%o3, %o0 + 0x18), NG_ret_i2_plus_g1_minus_24)
376 EX_LD(LOAD_TWIN(%i1 + %g3, %o2, %o3), 376 EX_LD(LOAD_TWIN(%i1 + %g3, %o2, %o3), NG_ret_i2_plus_g1_minus_32)
377 add %i1, 64, %i1 377 add %i1, 64, %i1
378 EX_ST(STORE_INIT(%o4, %o0 + 0x20), NG_ 378 EX_ST(STORE_INIT(%o4, %o0 + 0x20), NG_ret_i2_plus_g1_minus_32)
379 EX_ST(STORE_INIT(%o5, %o0 + 0x28), NG_ 379 EX_ST(STORE_INIT(%o5, %o0 + 0x28), NG_ret_i2_plus_g1_minus_40)
380 EX_ST(STORE_INIT(%o2, %o0 + 0x30), NG_ 380 EX_ST(STORE_INIT(%o2, %o0 + 0x30), NG_ret_i2_plus_g1_minus_48)
381 EX_ST(STORE_INIT(%o3, %o0 + 0x38), NG_ 381 EX_ST(STORE_INIT(%o3, %o0 + 0x38), NG_ret_i2_plus_g1_minus_56)
382 subcc %g1, 64, %g1 382 subcc %g1, 64, %g1
383 bne,pt %XCC, 1b 383 bne,pt %XCC, 1b
384 add %o0, 64, %o0 384 add %o0, 64, %o0
385 /* fall through */ 385 /* fall through */
386 386
387 60: 387 60:
388 membar #Sync 388 membar #Sync
389 389
390 /* %i2 contains any final bytes still 390 /* %i2 contains any final bytes still needed to be copied
391 * over. If anything is left, we copy 391 * over. If anything is left, we copy it one byte at a time.
392 */ 392 */
393 RESTORE_ASI(%i3) 393 RESTORE_ASI(%i3)
394 brz,pt %i2, 85f 394 brz,pt %i2, 85f
395 sub %o0, %i1, %i3 395 sub %o0, %i1, %i3
396 ba,a,pt %XCC, 90f 396 ba,a,pt %XCC, 90f
397 nop 397 nop
398 398
399 .align 64 399 .align 64
400 70: /* 16 < len <= 64 */ 400 70: /* 16 < len <= 64 */
401 bne,pn %XCC, 75f 401 bne,pn %XCC, 75f
402 sub %o0, %i1, %i3 402 sub %o0, %i1, %i3
403 403
404 72: 404 72:
405 andn %i2, 0xf, %i4 405 andn %i2, 0xf, %i4
406 and %i2, 0xf, %i2 406 and %i2, 0xf, %i2
407 1: subcc %i4, 0x10, %i4 407 1: subcc %i4, 0x10, %i4
408 EX_LD(LOAD(ldx, %i1, %o4), NG_ret_i2_p 408 EX_LD(LOAD(ldx, %i1, %o4), NG_ret_i2_plus_i4)
409 add %i1, 0x08, %i1 409 add %i1, 0x08, %i1
410 EX_LD(LOAD(ldx, %i1, %g1), NG_ret_i2_p 410 EX_LD(LOAD(ldx, %i1, %g1), NG_ret_i2_plus_i4)
411 sub %i1, 0x08, %i1 411 sub %i1, 0x08, %i1
412 EX_ST(STORE(stx, %o4, %i1 + %i3), NG_r 412 EX_ST(STORE(stx, %o4, %i1 + %i3), NG_ret_i2_plus_i4)
413 add %i1, 0x8, %i1 413 add %i1, 0x8, %i1
414 EX_ST(STORE(stx, %g1, %i1 + %i3), NG_r 414 EX_ST(STORE(stx, %g1, %i1 + %i3), NG_ret_i2_plus_i4_minus_8)
415 bgu,pt %XCC, 1b 415 bgu,pt %XCC, 1b
416 add %i1, 0x8, %i1 416 add %i1, 0x8, %i1
417 73: andcc %i2, 0x8, %g0 417 73: andcc %i2, 0x8, %g0
418 be,pt %XCC, 1f 418 be,pt %XCC, 1f
419 nop 419 nop
420 sub %i2, 0x8, %i2 420 sub %i2, 0x8, %i2
421 EX_LD(LOAD(ldx, %i1, %o4), NG_ret_i2_p 421 EX_LD(LOAD(ldx, %i1, %o4), NG_ret_i2_plus_8)
422 EX_ST(STORE(stx, %o4, %i1 + %i3), NG_r 422 EX_ST(STORE(stx, %o4, %i1 + %i3), NG_ret_i2_plus_8)
423 add %i1, 0x8, %i1 423 add %i1, 0x8, %i1
424 1: andcc %i2, 0x4, %g0 424 1: andcc %i2, 0x4, %g0
425 be,pt %XCC, 1f 425 be,pt %XCC, 1f
426 nop 426 nop
427 sub %i2, 0x4, %i2 427 sub %i2, 0x4, %i2
428 EX_LD(LOAD(lduw, %i1, %i5), NG_ret_i2_ 428 EX_LD(LOAD(lduw, %i1, %i5), NG_ret_i2_plus_4)
429 EX_ST(STORE(stw, %i5, %i1 + %i3), NG_r 429 EX_ST(STORE(stw, %i5, %i1 + %i3), NG_ret_i2_plus_4)
430 add %i1, 0x4, %i1 430 add %i1, 0x4, %i1
431 1: cmp %i2, 0 431 1: cmp %i2, 0
432 be,pt %XCC, 85f 432 be,pt %XCC, 85f
433 nop 433 nop
434 ba,pt %xcc, 90f 434 ba,pt %xcc, 90f
435 nop 435 nop
436 436
437 75: 437 75:
438 andcc %o0, 0x7, %g1 438 andcc %o0, 0x7, %g1
439 sub %g1, 0x8, %g1 439 sub %g1, 0x8, %g1
440 be,pn %icc, 2f 440 be,pn %icc, 2f
441 sub %g0, %g1, %g1 441 sub %g0, %g1, %g1
442 sub %i2, %g1, %i2 442 sub %i2, %g1, %i2
443 443
444 1: subcc %g1, 1, %g1 444 1: subcc %g1, 1, %g1
445 EX_LD(LOAD(ldub, %i1, %i5), NG_ret_i2_ 445 EX_LD(LOAD(ldub, %i1, %i5), NG_ret_i2_plus_g1_plus_1)
446 EX_ST(STORE(stb, %i5, %i1 + %i3), NG_r 446 EX_ST(STORE(stb, %i5, %i1 + %i3), NG_ret_i2_plus_g1_plus_1)
447 bgu,pt %icc, 1b 447 bgu,pt %icc, 1b
448 add %i1, 1, %i1 448 add %i1, 1, %i1
449 449
450 2: add %i1, %i3, %o0 450 2: add %i1, %i3, %o0
451 andcc %i1, 0x7, %g1 451 andcc %i1, 0x7, %g1
452 bne,pt %icc, 8f 452 bne,pt %icc, 8f
453 sll %g1, 3, %g1 453 sll %g1, 3, %g1
454 454
455 cmp %i2, 16 455 cmp %i2, 16
456 bgeu,pt %icc, 72b 456 bgeu,pt %icc, 72b
457 nop 457 nop
458 ba,a,pt %xcc, 73b 458 ba,a,pt %xcc, 73b
459 459
460 8: mov 64, %i3 460 8: mov 64, %i3
461 andn %i1, 0x7, %i1 461 andn %i1, 0x7, %i1
462 EX_LD(LOAD(ldx, %i1, %g2), NG_ret_i2) 462 EX_LD(LOAD(ldx, %i1, %g2), NG_ret_i2)
463 sub %i3, %g1, %i3 463 sub %i3, %g1, %i3
464 andn %i2, 0x7, %i4 464 andn %i2, 0x7, %i4
465 sllx %g2, %g1, %g2 465 sllx %g2, %g1, %g2
466 1: add %i1, 0x8, %i1 466 1: add %i1, 0x8, %i1
467 EX_LD(LOAD(ldx, %i1, %g3), NG_ret_i2_a 467 EX_LD(LOAD(ldx, %i1, %g3), NG_ret_i2_and_7_plus_i4)
468 subcc %i4, 0x8, %i4 468 subcc %i4, 0x8, %i4
469 srlx %g3, %i3, %i5 469 srlx %g3, %i3, %i5
470 or %i5, %g2, %i5 470 or %i5, %g2, %i5
471 EX_ST(STORE(stx, %i5, %o0), NG_ret_i2_ 471 EX_ST(STORE(stx, %i5, %o0), NG_ret_i2_and_7_plus_i4)
472 add %o0, 0x8, %o0 472 add %o0, 0x8, %o0
473 bgu,pt %icc, 1b 473 bgu,pt %icc, 1b
474 sllx %g3, %g1, %g2 474 sllx %g3, %g1, %g2
475 475
476 srl %g1, 3, %g1 476 srl %g1, 3, %g1
477 andcc %i2, 0x7, %i2 477 andcc %i2, 0x7, %i2
478 be,pn %icc, 85f 478 be,pn %icc, 85f
479 add %i1, %g1, %i1 479 add %i1, %g1, %i1
480 ba,pt %xcc, 90f 480 ba,pt %xcc, 90f
481 sub %o0, %i1, %i3 481 sub %o0, %i1, %i3
482 482
483 .align 64 483 .align 64
484 80: /* 0 < len <= 16 */ 484 80: /* 0 < len <= 16 */
485 andcc %i3, 0x3, %g0 485 andcc %i3, 0x3, %g0
486 bne,pn %XCC, 90f 486 bne,pn %XCC, 90f
487 sub %o0, %i1, %i3 487 sub %o0, %i1, %i3
488 488
489 1: 489 1:
490 subcc %i2, 4, %i2 490 subcc %i2, 4, %i2
491 EX_LD(LOAD(lduw, %i1, %g1), NG_ret_i2_ 491 EX_LD(LOAD(lduw, %i1, %g1), NG_ret_i2_plus_4)
492 EX_ST(STORE(stw, %g1, %i1 + %i3), NG_r 492 EX_ST(STORE(stw, %g1, %i1 + %i3), NG_ret_i2_plus_4)
493 bgu,pt %XCC, 1b 493 bgu,pt %XCC, 1b
494 add %i1, 4, %i1 494 add %i1, 4, %i1
495 495
496 85: ret 496 85: ret
497 restore EX_RETVAL(%i0), %g0, % 497 restore EX_RETVAL(%i0), %g0, %o0
498 498
499 .align 32 499 .align 32
500 90: 500 90:
501 subcc %i2, 1, %i2 501 subcc %i2, 1, %i2
502 EX_LD(LOAD(ldub, %i1, %g1), NG_ret_i2_ 502 EX_LD(LOAD(ldub, %i1, %g1), NG_ret_i2_plus_1)
503 EX_ST(STORE(stb, %g1, %i1 + %i3), NG_r 503 EX_ST(STORE(stb, %g1, %i1 + %i3), NG_ret_i2_plus_1)
504 bgu,pt %XCC, 90b 504 bgu,pt %XCC, 90b
505 add %i1, 1, %i1 505 add %i1, 1, %i1
506 ret 506 ret
507 restore EX_RETVAL(%i0), %g0, % 507 restore EX_RETVAL(%i0), %g0, %o0
508 508
509 .size FUNC_NAME, .-FUNC_NAME 509 .size FUNC_NAME, .-FUNC_NAME
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.