1 /* 1 /*
2 * M7memset.S: SPARC M7 optimized memset. 2 * M7memset.S: SPARC M7 optimized memset.
3 * 3 *
4 * Copyright (c) 2016, Oracle and/or its affil 4 * Copyright (c) 2016, Oracle and/or its affiliates. All rights reserved.
5 */ 5 */
6 6
7 /* 7 /*
8 * M7memset.S: M7 optimized memset. 8 * M7memset.S: M7 optimized memset.
9 * 9 *
10 * char *memset(sp, c, n) 10 * char *memset(sp, c, n)
11 * 11 *
12 * Set an array of n chars starting at sp to t 12 * Set an array of n chars starting at sp to the character c.
13 * Return sp. 13 * Return sp.
14 * 14 *
15 * Fast assembler language version of the foll 15 * Fast assembler language version of the following C-program for memset
16 * which represents the `standard' for the C-l 16 * which represents the `standard' for the C-library.
17 * 17 *
18 * void * 18 * void *
19 * memset(void *sp1, int c, size_t n) 19 * memset(void *sp1, int c, size_t n)
20 * { 20 * {
21 * if (n != 0) { 21 * if (n != 0) {
22 * char *sp = sp1; 22 * char *sp = sp1;
23 * do { 23 * do {
24 * *sp++ = (char)c; 24 * *sp++ = (char)c;
25 * } while (--n != 0); 25 * } while (--n != 0);
26 * } 26 * }
27 * return (sp1); 27 * return (sp1);
28 * } 28 * }
29 * 29 *
30 * The algorithm is as follows : 30 * The algorithm is as follows :
31 * 31 *
32 * For small 6 or fewer bytes stores, byt 32 * For small 6 or fewer bytes stores, bytes will be stored.
33 * 33 *
34 * For less than 32 bytes stores, align t 34 * For less than 32 bytes stores, align the address on 4 byte boundary.
35 * Then store as many 4-byte chunks, foll 35 * Then store as many 4-byte chunks, followed by trailing bytes.
36 * 36 *
37 * For sizes greater than 32 bytes, align 37 * For sizes greater than 32 bytes, align the address on 8 byte boundary.
38 * if (count >= 64) { 38 * if (count >= 64) {
39 * store 8-bytes chunks to align 39 * store 8-bytes chunks to align the address on 64 byte boundary
40 * if (value to be set is zero && 40 * if (value to be set is zero && count >= MIN_ZERO) {
41 * Using BIS stores, set 41 * Using BIS stores, set the first long word of each
42 * 64-byte cache line to 42 * 64-byte cache line to zero which will also clear the
43 * other seven long words 43 * other seven long words of the cache line.
44 * } 44 * }
45 * else if (count >= MIN_LOOP) { 45 * else if (count >= MIN_LOOP) {
46 * Using BIS stores, set 46 * Using BIS stores, set the first long word of each of
47 * ST_CHUNK cache lines ( 47 * ST_CHUNK cache lines (64 bytes each) before the main
48 * loop is entered. 48 * loop is entered.
49 * In the main loop, cont 49 * In the main loop, continue pre-setting the first long
50 * word of each cache lin 50 * word of each cache line ST_CHUNK lines in advance while
51 * setting the other seve 51 * setting the other seven long words (56 bytes) of each
52 * cache line until fewer 52 * cache line until fewer than ST_CHUNK*64 bytes remain.
53 * Then set the remaining 53 * Then set the remaining seven long words of each cache
54 * line that has already 54 * line that has already had its first long word set.
55 * } 55 * }
56 * store remaining data in 64-byt 56 * store remaining data in 64-byte chunks until less than
57 * 64 bytes remain. 57 * 64 bytes remain.
58 * } 58 * }
59 * Store as many 8-byte chunks, followed 59 * Store as many 8-byte chunks, followed by trailing bytes.
60 * 60 *
61 * BIS = Block Init Store 61 * BIS = Block Init Store
62 * Doing the advance store of the first elem 62 * Doing the advance store of the first element of the cache line
63 * initiates the displacement of a cache lin 63 * initiates the displacement of a cache line while only using a single
64 * instruction in the pipeline. That avoids 64 * instruction in the pipeline. That avoids various pipeline delays,
65 * such as filling the miss buffer. The perf 65 * such as filling the miss buffer. The performance effect is
66 * similar to prefetching for normal stores. 66 * similar to prefetching for normal stores.
67 * The special case for zero fills runs fast 67 * The special case for zero fills runs faster and uses fewer instruction
68 * cycles than the normal memset loop. 68 * cycles than the normal memset loop.
69 * 69 *
70 * We only use BIS for memset of greater than 70 * We only use BIS for memset of greater than MIN_LOOP bytes because a sequence
71 * BIS stores must be followed by a membar #St 71 * BIS stores must be followed by a membar #StoreStore. The benefit of
72 * the BIS store must be balanced against the 72 * the BIS store must be balanced against the cost of the membar operation.
73 */ 73 */
74 74
75 /* 75 /*
76 * ASI_STBI_P marks the cache line as "least r 76 * ASI_STBI_P marks the cache line as "least recently used"
77 * which means if many threads are active, it 77 * which means if many threads are active, it has a high chance
78 * of being pushed out of the cache between th 78 * of being pushed out of the cache between the first initializing
79 * store and the final stores. 79 * store and the final stores.
80 * Thus, we use ASI_STBIMRU_P which marks the 80 * Thus, we use ASI_STBIMRU_P which marks the cache line as
81 * "most recently used" for all but the last s 81 * "most recently used" for all but the last store to the cache line.
82 */ 82 */
83 83
84 #include <asm/asi.h> 84 #include <asm/asi.h>
85 #include <asm/page.h> 85 #include <asm/page.h>
86 86
87 #define ASI_STBI_P ASI_BLK_INIT_QUAD_LDD_ 87 #define ASI_STBI_P ASI_BLK_INIT_QUAD_LDD_P
88 #define ASI_STBIMRU_P ASI_ST_BLKINIT_MRU_P 88 #define ASI_STBIMRU_P ASI_ST_BLKINIT_MRU_P
89 89
90 90
91 #define ST_CHUNK 24 /* multiple of 4 91 #define ST_CHUNK 24 /* multiple of 4 due to loop unrolling */
92 #define MIN_LOOP 16320 92 #define MIN_LOOP 16320
93 #define MIN_ZERO 512 93 #define MIN_ZERO 512
94 94
95 .section ".text" 95 .section ".text"
96 .align 32 96 .align 32
97 97
98 /* 98 /*
99 * Define clear_page(dest) as memset(dest, 0, 99 * Define clear_page(dest) as memset(dest, 0, PAGE_SIZE)
100 * (can create a more optimized version later. 100 * (can create a more optimized version later.)
101 */ 101 */
102 .globl M7clear_page 102 .globl M7clear_page
103 .globl M7clear_user_page 103 .globl M7clear_user_page
104 M7clear_page: /* clear_page(dest) */ 104 M7clear_page: /* clear_page(dest) */
105 M7clear_user_page: 105 M7clear_user_page:
106 set PAGE_SIZE, %o1 106 set PAGE_SIZE, %o1
107 /* fall through into bzero code */ 107 /* fall through into bzero code */
108 108
109 .size M7clear_page,.-M7clear 109 .size M7clear_page,.-M7clear_page
110 .size M7clear_user_page,.-M7 110 .size M7clear_user_page,.-M7clear_user_page
111 111
112 /* 112 /*
113 * Define bzero(dest, n) as memset(dest, 0, n) 113 * Define bzero(dest, n) as memset(dest, 0, n)
114 * (can create a more optimized version later. 114 * (can create a more optimized version later.)
115 */ 115 */
116 .globl M7bzero 116 .globl M7bzero
117 M7bzero: /* bzero(dest, size) * 117 M7bzero: /* bzero(dest, size) */
118 mov %o1, %o2 118 mov %o1, %o2
119 mov 0, %o1 119 mov 0, %o1
120 /* fall through into memset code */ 120 /* fall through into memset code */
121 121
122 .size M7bzero,.-M7bzero 122 .size M7bzero,.-M7bzero
123 123
124 .global M7memset 124 .global M7memset
125 .type M7memset, #function 125 .type M7memset, #function
126 .register %g3, #scratch 126 .register %g3, #scratch
127 M7memset: 127 M7memset:
128 mov %o0, %o5 ! copy 128 mov %o0, %o5 ! copy sp1 before using it
129 cmp %o2, 7 ! if s 129 cmp %o2, 7 ! if small counts, just write bytes
130 bleu,pn %xcc, .wrchar 130 bleu,pn %xcc, .wrchar
131 and %o1, 0xff, %o1 ! o1 131 and %o1, 0xff, %o1 ! o1 is (char)c
132 132
133 sll %o1, 8, %o3 133 sll %o1, 8, %o3
134 or %o1, %o3, %o1 ! now 134 or %o1, %o3, %o1 ! now o1 has 2 bytes of c
135 sll %o1, 16, %o3 135 sll %o1, 16, %o3
136 cmp %o2, 32 136 cmp %o2, 32
137 blu,pn %xcc, .wdalign 137 blu,pn %xcc, .wdalign
138 or %o1, %o3, %o1 ! now 138 or %o1, %o3, %o1 ! now o1 has 4 bytes of c
139 139
140 sllx %o1, 32, %o3 140 sllx %o1, 32, %o3
141 or %o1, %o3, %o1 ! now 141 or %o1, %o3, %o1 ! now o1 has 8 bytes of c
142 142
143 .dbalign: 143 .dbalign:
144 andcc %o5, 7, %o3 ! is s 144 andcc %o5, 7, %o3 ! is sp1 aligned on a 8 byte bound?
145 bz,pt %xcc, .blkalign ! alre 145 bz,pt %xcc, .blkalign ! already long word aligned
146 sub %o3, 8, %o3 ! -(b 146 sub %o3, 8, %o3 ! -(bytes till long word aligned)
147 147
148 add %o2, %o3, %o2 ! upda 148 add %o2, %o3, %o2 ! update o2 with new count
149 ! Set -(%o3) bytes till sp1 long word 149 ! Set -(%o3) bytes till sp1 long word aligned
150 1: stb %o1, [%o5] ! ther 150 1: stb %o1, [%o5] ! there is at least 1 byte to set
151 inccc %o3 ! byte 151 inccc %o3 ! byte clearing loop
152 bl,pt %xcc, 1b 152 bl,pt %xcc, 1b
153 inc %o5 153 inc %o5
154 154
155 ! Now sp1 is long word aligned (sp1 is 155 ! Now sp1 is long word aligned (sp1 is found in %o5)
156 .blkalign: 156 .blkalign:
157 cmp %o2, 64 ! chec 157 cmp %o2, 64 ! check if there are 64 bytes to set
158 blu,pn %xcc, .wrshort 158 blu,pn %xcc, .wrshort
159 mov %o2, %o3 159 mov %o2, %o3
160 160
161 andcc %o5, 63, %o3 ! is s 161 andcc %o5, 63, %o3 ! is sp1 block aligned?
162 bz,pt %xcc, .blkwr ! now 162 bz,pt %xcc, .blkwr ! now block aligned
163 sub %o3, 64, %o3 ! o3 163 sub %o3, 64, %o3 ! o3 is -(bytes till block aligned)
164 add %o2, %o3, %o2 ! o2 i 164 add %o2, %o3, %o2 ! o2 is the remainder
165 165
166 ! Store -(%o3) bytes till dst is block 166 ! Store -(%o3) bytes till dst is block (64 byte) aligned.
167 ! Use long word stores. 167 ! Use long word stores.
168 ! Recall that dst is already long word 168 ! Recall that dst is already long word aligned
169 1: 169 1:
170 addcc %o3, 8, %o3 170 addcc %o3, 8, %o3
171 stx %o1, [%o5] 171 stx %o1, [%o5]
172 bl,pt %xcc, 1b 172 bl,pt %xcc, 1b
173 add %o5, 8, %o5 173 add %o5, 8, %o5
174 174
175 ! Now sp1 is block aligned 175 ! Now sp1 is block aligned
176 .blkwr: 176 .blkwr:
177 andn %o2, 63, %o4 ! calc 177 andn %o2, 63, %o4 ! calculate size of blocks in bytes
178 brz,pn %o1, .wrzero ! spec 178 brz,pn %o1, .wrzero ! special case if c == 0
179 and %o2, 63, %o3 ! %o3 179 and %o2, 63, %o3 ! %o3 = bytes left after blk stores.
180 180
181 set MIN_LOOP, %g1 181 set MIN_LOOP, %g1
182 cmp %o4, %g1 ! chec 182 cmp %o4, %g1 ! check there are enough bytes to set
183 blu,pn %xcc, .short_set ! to j 183 blu,pn %xcc, .short_set ! to justify cost of membar
184 ! must 184 ! must be > pre-cleared lines
185 nop 185 nop
186 186
187 ! initial cache-clearing stores 187 ! initial cache-clearing stores
188 ! get store pipeline moving 188 ! get store pipeline moving
189 rd %asi, %g3 ! save 189 rd %asi, %g3 ! save %asi to be restored later
190 wr %g0, ASI_STBIMRU_P, %asi 190 wr %g0, ASI_STBIMRU_P, %asi
191 191
192 ! Primary memset loop for large memset 192 ! Primary memset loop for large memsets
193 .wr_loop: 193 .wr_loop:
194 sub %o5, 8, %o5 ! adju 194 sub %o5, 8, %o5 ! adjust %o5 for ASI store alignment
195 mov ST_CHUNK, %g1 195 mov ST_CHUNK, %g1
196 .wr_loop_start: 196 .wr_loop_start:
197 stxa %o1, [%o5+8]%asi 197 stxa %o1, [%o5+8]%asi
198 subcc %g1, 4, %g1 198 subcc %g1, 4, %g1
199 stxa %o1, [%o5+8+64]%asi 199 stxa %o1, [%o5+8+64]%asi
200 add %o5, 256, %o5 200 add %o5, 256, %o5
201 stxa %o1, [%o5+8-128]%asi 201 stxa %o1, [%o5+8-128]%asi
202 bgu %xcc, .wr_loop_start 202 bgu %xcc, .wr_loop_start
203 stxa %o1, [%o5+8-64]%asi 203 stxa %o1, [%o5+8-64]%asi
204 204
205 sub %o5, ST_CHUNK*64, %o5 ! rese 205 sub %o5, ST_CHUNK*64, %o5 ! reset %o5
206 mov ST_CHUNK, %g1 206 mov ST_CHUNK, %g1
207 207
208 .wr_loop_rest: 208 .wr_loop_rest:
209 stxa %o1, [%o5+8+8]%asi 209 stxa %o1, [%o5+8+8]%asi
210 sub %o4, 64, %o4 210 sub %o4, 64, %o4
211 stxa %o1, [%o5+16+8]%asi 211 stxa %o1, [%o5+16+8]%asi
212 subcc %g1, 1, %g1 212 subcc %g1, 1, %g1
213 stxa %o1, [%o5+24+8]%asi 213 stxa %o1, [%o5+24+8]%asi
214 stxa %o1, [%o5+32+8]%asi 214 stxa %o1, [%o5+32+8]%asi
215 stxa %o1, [%o5+40+8]%asi 215 stxa %o1, [%o5+40+8]%asi
216 add %o5, 64, %o5 216 add %o5, 64, %o5
217 stxa %o1, [%o5-8]%asi 217 stxa %o1, [%o5-8]%asi
218 bgu %xcc, .wr_loop_rest 218 bgu %xcc, .wr_loop_rest
219 stxa %o1, [%o5]ASI_STBI_P 219 stxa %o1, [%o5]ASI_STBI_P
220 220
221 ! If more than ST_CHUNK*64 bytes remai 221 ! If more than ST_CHUNK*64 bytes remain to set, continue
222 ! setting the first long word of each 222 ! setting the first long word of each cache line in advance
223 ! to keep the store pipeline moving. 223 ! to keep the store pipeline moving.
224 224
225 cmp %o4, ST_CHUNK*64 225 cmp %o4, ST_CHUNK*64
226 bge,pt %xcc, .wr_loop_start 226 bge,pt %xcc, .wr_loop_start
227 mov ST_CHUNK, %g1 227 mov ST_CHUNK, %g1
228 228
229 brz,a,pn %o4, .asi_done 229 brz,a,pn %o4, .asi_done
230 add %o5, 8, %o5 ! res 230 add %o5, 8, %o5 ! restore %o5 offset
231 231
232 .wr_loop_small: 232 .wr_loop_small:
233 stxa %o1, [%o5+8]%asi 233 stxa %o1, [%o5+8]%asi
234 stxa %o1, [%o5+8+8]%asi 234 stxa %o1, [%o5+8+8]%asi
235 stxa %o1, [%o5+16+8]%asi 235 stxa %o1, [%o5+16+8]%asi
236 stxa %o1, [%o5+24+8]%asi 236 stxa %o1, [%o5+24+8]%asi
237 stxa %o1, [%o5+32+8]%asi 237 stxa %o1, [%o5+32+8]%asi
238 subcc %o4, 64, %o4 238 subcc %o4, 64, %o4
239 stxa %o1, [%o5+40+8]%asi 239 stxa %o1, [%o5+40+8]%asi
240 add %o5, 64, %o5 240 add %o5, 64, %o5
241 stxa %o1, [%o5-8]%asi 241 stxa %o1, [%o5-8]%asi
242 bgu,pt %xcc, .wr_loop_small 242 bgu,pt %xcc, .wr_loop_small
243 stxa %o1, [%o5]ASI_STBI_P 243 stxa %o1, [%o5]ASI_STBI_P
244 244
245 ba .asi_done 245 ba .asi_done
246 add %o5, 8, %o5 ! res 246 add %o5, 8, %o5 ! restore %o5 offset
247 247
248 ! Special case loop for zero fill mems 248 ! Special case loop for zero fill memsets
249 ! For each 64 byte cache line, single 249 ! For each 64 byte cache line, single STBI to first element
250 ! clears line 250 ! clears line
251 .wrzero: 251 .wrzero:
252 cmp %o4, MIN_ZERO ! chec 252 cmp %o4, MIN_ZERO ! check if enough bytes to set
253 ! to p 253 ! to pay %asi + membar cost
254 blu %xcc, .short_set 254 blu %xcc, .short_set
255 nop 255 nop
256 sub %o4, 256, %o4 256 sub %o4, 256, %o4
257 257
258 .wrzero_loop: 258 .wrzero_loop:
259 mov 64, %g3 259 mov 64, %g3
260 stxa %o1, [%o5]ASI_STBI_P 260 stxa %o1, [%o5]ASI_STBI_P
261 subcc %o4, 256, %o4 261 subcc %o4, 256, %o4
262 stxa %o1, [%o5+%g3]ASI_STBI_P 262 stxa %o1, [%o5+%g3]ASI_STBI_P
263 add %o5, 256, %o5 263 add %o5, 256, %o5
264 sub %g3, 192, %g3 264 sub %g3, 192, %g3
265 stxa %o1, [%o5+%g3]ASI_STBI_P 265 stxa %o1, [%o5+%g3]ASI_STBI_P
266 add %g3, 64, %g3 266 add %g3, 64, %g3
267 bge,pt %xcc, .wrzero_loop 267 bge,pt %xcc, .wrzero_loop
268 stxa %o1, [%o5+%g3]ASI_STBI_P 268 stxa %o1, [%o5+%g3]ASI_STBI_P
269 add %o4, 256, %o4 269 add %o4, 256, %o4
270 270
271 brz,pn %o4, .bsi_done 271 brz,pn %o4, .bsi_done
272 nop 272 nop
273 273
274 .wrzero_small: 274 .wrzero_small:
275 stxa %o1, [%o5]ASI_STBI_P 275 stxa %o1, [%o5]ASI_STBI_P
276 subcc %o4, 64, %o4 276 subcc %o4, 64, %o4
277 bgu,pt %xcc, .wrzero_small 277 bgu,pt %xcc, .wrzero_small
278 add %o5, 64, %o5 278 add %o5, 64, %o5
279 ba,a .bsi_done 279 ba,a .bsi_done
280 280
281 .asi_done: 281 .asi_done:
282 wr %g3, 0x0, %asi ! rest 282 wr %g3, 0x0, %asi ! restored saved %asi
283 .bsi_done: 283 .bsi_done:
284 membar #StoreStore ! requ 284 membar #StoreStore ! required by use of Block Store Init
285 285
286 .short_set: 286 .short_set:
287 cmp %o4, 64 ! chec 287 cmp %o4, 64 ! check if 64 bytes to set
288 blu %xcc, 5f 288 blu %xcc, 5f
289 nop 289 nop
290 4: ! set 290 4: ! set final blocks of 64 bytes
291 stx %o1, [%o5] 291 stx %o1, [%o5]
292 stx %o1, [%o5+8] 292 stx %o1, [%o5+8]
293 stx %o1, [%o5+16] 293 stx %o1, [%o5+16]
294 stx %o1, [%o5+24] 294 stx %o1, [%o5+24]
295 subcc %o4, 64, %o4 295 subcc %o4, 64, %o4
296 stx %o1, [%o5+32] 296 stx %o1, [%o5+32]
297 stx %o1, [%o5+40] 297 stx %o1, [%o5+40]
298 add %o5, 64, %o5 298 add %o5, 64, %o5
299 stx %o1, [%o5-16] 299 stx %o1, [%o5-16]
300 bgu,pt %xcc, 4b 300 bgu,pt %xcc, 4b
301 stx %o1, [%o5-8] 301 stx %o1, [%o5-8]
302 302
303 5: 303 5:
304 ! Set the remaining long words 304 ! Set the remaining long words
305 .wrshort: 305 .wrshort:
306 subcc %o3, 8, %o3 ! Can 306 subcc %o3, 8, %o3 ! Can we store any long words?
307 blu,pn %xcc, .wrchars 307 blu,pn %xcc, .wrchars
308 and %o2, 7, %o2 ! cal 308 and %o2, 7, %o2 ! calc bytes left after long words
309 6: 309 6:
310 subcc %o3, 8, %o3 310 subcc %o3, 8, %o3
311 stx %o1, [%o5] ! stor 311 stx %o1, [%o5] ! store the long words
312 bgeu,pt %xcc, 6b 312 bgeu,pt %xcc, 6b
313 add %o5, 8, %o5 313 add %o5, 8, %o5
314 314
315 .wrchars: ! chec 315 .wrchars: ! check for extra chars
316 brnz %o2, .wrfin 316 brnz %o2, .wrfin
317 nop 317 nop
318 retl 318 retl
319 nop 319 nop
320 320
321 .wdalign: 321 .wdalign:
322 andcc %o5, 3, %o3 ! is s 322 andcc %o5, 3, %o3 ! is sp1 aligned on a word boundary
323 bz,pn %xcc, .wrword 323 bz,pn %xcc, .wrword
324 andn %o2, 3, %o3 ! cre 324 andn %o2, 3, %o3 ! create word sized count in %o3
325 325
326 dec %o2 ! decr 326 dec %o2 ! decrement count
327 stb %o1, [%o5] ! clea 327 stb %o1, [%o5] ! clear a byte
328 b .wdalign 328 b .wdalign
329 inc %o5 ! nex 329 inc %o5 ! next byte
330 330
331 .wrword: 331 .wrword:
332 subcc %o3, 4, %o3 332 subcc %o3, 4, %o3
333 st %o1, [%o5] ! 4-by 333 st %o1, [%o5] ! 4-byte writing loop
334 bnz,pt %xcc, .wrword 334 bnz,pt %xcc, .wrword
335 add %o5, 4, %o5 335 add %o5, 4, %o5
336 336
337 and %o2, 3, %o2 ! left 337 and %o2, 3, %o2 ! leftover count, if any
338 338
339 .wrchar: 339 .wrchar:
340 ! Set the remaining bytes, if any 340 ! Set the remaining bytes, if any
341 brz %o2, .exit 341 brz %o2, .exit
342 nop 342 nop
343 .wrfin: 343 .wrfin:
344 deccc %o2 344 deccc %o2
345 stb %o1, [%o5] 345 stb %o1, [%o5]
346 bgu,pt %xcc, .wrfin 346 bgu,pt %xcc, .wrfin
347 inc %o5 347 inc %o5
348 .exit: 348 .exit:
349 retl ! %o0 349 retl ! %o0 was preserved
350 nop 350 nop
351 351
352 .size M7memset,.-M7memset 352 .size M7memset,.-M7memset
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.