aboutsummaryrefslogtreecommitdiffstats
path: root/arch/sparc64/lib/NGmemcpy.S
diff options
context:
space:
mode:
Diffstat (limited to 'arch/sparc64/lib/NGmemcpy.S')
-rw-r--r--arch/sparc64/lib/NGmemcpy.S364
1 files changed, 364 insertions, 0 deletions
diff --git a/arch/sparc64/lib/NGmemcpy.S b/arch/sparc64/lib/NGmemcpy.S
new file mode 100644
index 000000000000..a39aa3bd4345
--- /dev/null
+++ b/arch/sparc64/lib/NGmemcpy.S
@@ -0,0 +1,364 @@
1/* NGmemcpy.S: Niagara optimized memcpy.
2 *
3 * Copyright (C) 2006 David S. Miller (davem@davemloft.net)
4 */
5
6#ifdef __KERNEL__
7#include <asm/asi.h>
8#define GLOBAL_SPARE %g7
9#define RESTORE_ASI wr %g0, ASI_AIUS, %asi
10#else
11#define GLOBAL_SPARE %g5
12#define RESTORE_ASI
13#endif
14
15#ifndef STORE_ASI
16#define STORE_ASI ASI_BLK_INIT_QUAD_LDD_P
17#endif
18
19#ifndef EX_LD
20#define EX_LD(x) x
21#endif
22
23#ifndef EX_ST
24#define EX_ST(x) x
25#endif
26
27#ifndef EX_RETVAL
28#define EX_RETVAL(x) x
29#endif
30
31#ifndef LOAD
32#ifndef MEMCPY_DEBUG
33#define LOAD(type,addr,dest) type [addr], dest
34#else
35#define LOAD(type,addr,dest) type##a [addr] 0x80, dest
36#endif
37#endif
38
39#ifndef LOAD_TWIN
40#define LOAD_TWIN(addr_reg,dest0,dest1) \
41 ldda [addr_reg] ASI_BLK_INIT_QUAD_LDD_P, dest0
42#endif
43
44#ifndef STORE
45#define STORE(type,src,addr) type src, [addr]
46#endif
47
48#ifndef STORE_INIT
49#define STORE_INIT(src,addr) stxa src, [addr] %asi
50#endif
51
52#ifndef FUNC_NAME
53#define FUNC_NAME NGmemcpy
54#endif
55
56#ifndef PREAMBLE
57#define PREAMBLE
58#endif
59
60#ifndef XCC
61#define XCC xcc
62#endif
63
64 .register %g2,#scratch
65 .register %g3,#scratch
66
67 .text
68 .align 64
69
70 .globl FUNC_NAME
71 .type FUNC_NAME,#function
72FUNC_NAME: /* %o0=dst, %o1=src, %o2=len */
73 srlx %o2, 31, %g2
74 cmp %g2, 0
75 tne %xcc, 5
76 PREAMBLE
77 mov %o0, GLOBAL_SPARE
78 cmp %o2, 0
79 be,pn %XCC, 85f
80 or %o0, %o1, %o3
81 cmp %o2, 16
82 blu,a,pn %XCC, 80f
83 or %o3, %o2, %o3
84
85 /* 2 blocks (128 bytes) is the minimum we can do the block
86 * copy with. We need to ensure that we'll iterate at least
87 * once in the block copy loop. At worst we'll need to align
88 * the destination to a 64-byte boundary which can chew up
89 * to (64 - 1) bytes from the length before we perform the
90 * block copy loop.
91 */
92 cmp %o2, (2 * 64)
93 blu,pt %XCC, 70f
94 andcc %o3, 0x7, %g0
95
96 /* %o0: dst
97 * %o1: src
98 * %o2: len (known to be >= 128)
99 *
100 * The block copy loops will use %o4/%o5,%g2/%g3 as
101 * temporaries while copying the data.
102 */
103
104 LOAD(prefetch, %o1, #one_read)
105 wr %g0, STORE_ASI, %asi
106
107 /* Align destination on 64-byte boundary. */
108 andcc %o0, (64 - 1), %o4
109 be,pt %XCC, 2f
110 sub %o4, 64, %o4
111 sub %g0, %o4, %o4 ! bytes to align dst
112 sub %o2, %o4, %o2
1131: subcc %o4, 1, %o4
114 EX_LD(LOAD(ldub, %o1, %g1))
115 EX_ST(STORE(stb, %g1, %o0))
116 add %o1, 1, %o1
117 bne,pt %XCC, 1b
118 add %o0, 1, %o0
119
120 /* If the source is on a 16-byte boundary we can do
121 * the direct block copy loop. If it is 8-byte aligned
122 * we can do the 16-byte loads offset by -8 bytes and the
123 * init stores offset by one register.
124 *
125 * If the source is not even 8-byte aligned, we need to do
126 * shifting and masking (basically integer faligndata).
127 *
128 * The careful bit with init stores is that if we store
129 * to any part of the cache line we have to store the whole
130 * cacheline else we can end up with corrupt L2 cache line
131 * contents. Since the loop works on 64-bytes of 64-byte
132 * aligned store data at a time, this is easy to ensure.
133 */
1342:
135 andcc %o1, (16 - 1), %o4
136 andn %o2, (64 - 1), %g1 ! block copy loop iterator
137 sub %o2, %g1, %o2 ! final sub-block copy bytes
138 be,pt %XCC, 50f
139 cmp %o4, 8
140 be,a,pt %XCC, 10f
141 sub %o1, 0x8, %o1
142
143 /* Neither 8-byte nor 16-byte aligned, shift and mask. */
144 mov %g1, %o4
145 and %o1, 0x7, %g1
146 sll %g1, 3, %g1
147 mov 64, %o3
148 andn %o1, 0x7, %o1
149 EX_LD(LOAD(ldx, %o1, %g2))
150 sub %o3, %g1, %o3
151 sllx %g2, %g1, %g2
152
153#define SWIVEL_ONE_DWORD(SRC, TMP1, TMP2, PRE_VAL, PRE_SHIFT, POST_SHIFT, DST)\
154 EX_LD(LOAD(ldx, SRC, TMP1)); \
155 srlx TMP1, PRE_SHIFT, TMP2; \
156 or TMP2, PRE_VAL, TMP2; \
157 EX_ST(STORE_INIT(TMP2, DST)); \
158 sllx TMP1, POST_SHIFT, PRE_VAL;
159
1601: add %o1, 0x8, %o1
161 SWIVEL_ONE_DWORD(%o1, %g3, %o5, %g2, %o3, %g1, %o0 + 0x00)
162 add %o1, 0x8, %o1
163 SWIVEL_ONE_DWORD(%o1, %g3, %o5, %g2, %o3, %g1, %o0 + 0x08)
164 add %o1, 0x8, %o1
165 SWIVEL_ONE_DWORD(%o1, %g3, %o5, %g2, %o3, %g1, %o0 + 0x10)
166 add %o1, 0x8, %o1
167 SWIVEL_ONE_DWORD(%o1, %g3, %o5, %g2, %o3, %g1, %o0 + 0x18)
168 add %o1, 32, %o1
169 LOAD(prefetch, %o1, #one_read)
170 sub %o1, 32 - 8, %o1
171 SWIVEL_ONE_DWORD(%o1, %g3, %o5, %g2, %o3, %g1, %o0 + 0x20)
172 add %o1, 8, %o1
173 SWIVEL_ONE_DWORD(%o1, %g3, %o5, %g2, %o3, %g1, %o0 + 0x28)
174 add %o1, 8, %o1
175 SWIVEL_ONE_DWORD(%o1, %g3, %o5, %g2, %o3, %g1, %o0 + 0x30)
176 add %o1, 8, %o1
177 SWIVEL_ONE_DWORD(%o1, %g3, %o5, %g2, %o3, %g1, %o0 + 0x38)
178 subcc %o4, 64, %o4
179 bne,pt %XCC, 1b
180 add %o0, 64, %o0
181
182#undef SWIVEL_ONE_DWORD
183
184 srl %g1, 3, %g1
185 ba,pt %XCC, 60f
186 add %o1, %g1, %o1
187
18810: /* Destination is 64-byte aligned, source was only 8-byte
189 * aligned but it has been subtracted by 8 and we perform
190 * one twin load ahead, then add 8 back into source when
191 * we finish the loop.
192 */
193 EX_LD(LOAD_TWIN(%o1, %o4, %o5))
1941: add %o1, 16, %o1
195 EX_LD(LOAD_TWIN(%o1, %g2, %g3))
196 add %o1, 16 + 32, %o1
197 LOAD(prefetch, %o1, #one_read)
198 sub %o1, 32, %o1
199 EX_ST(STORE_INIT(%o5, %o0 + 0x00)) ! initializes cache line
200 EX_ST(STORE_INIT(%g2, %o0 + 0x08))
201 EX_LD(LOAD_TWIN(%o1, %o4, %o5))
202 add %o1, 16, %o1
203 EX_ST(STORE_INIT(%g3, %o0 + 0x10))
204 EX_ST(STORE_INIT(%o4, %o0 + 0x18))
205 EX_LD(LOAD_TWIN(%o1, %g2, %g3))
206 add %o1, 16, %o1
207 EX_ST(STORE_INIT(%o5, %o0 + 0x20))
208 EX_ST(STORE_INIT(%g2, %o0 + 0x28))
209 EX_LD(LOAD_TWIN(%o1, %o4, %o5))
210 EX_ST(STORE_INIT(%g3, %o0 + 0x30))
211 EX_ST(STORE_INIT(%o4, %o0 + 0x38))
212 subcc %g1, 64, %g1
213 bne,pt %XCC, 1b
214 add %o0, 64, %o0
215
216 ba,pt %XCC, 60f
217 add %o1, 0x8, %o1
218
21950: /* Destination is 64-byte aligned, and source is 16-byte
220 * aligned.
221 */
2221: EX_LD(LOAD_TWIN(%o1, %o4, %o5))
223 add %o1, 16, %o1
224 EX_LD(LOAD_TWIN(%o1, %g2, %g3))
225 add %o1, 16 + 32, %o1
226 LOAD(prefetch, %o1, #one_read)
227 sub %o1, 32, %o1
228 EX_ST(STORE_INIT(%o4, %o0 + 0x00)) ! initializes cache line
229 EX_ST(STORE_INIT(%o5, %o0 + 0x08))
230 EX_LD(LOAD_TWIN(%o1, %o4, %o5))
231 add %o1, 16, %o1
232 EX_ST(STORE_INIT(%g2, %o0 + 0x10))
233 EX_ST(STORE_INIT(%g3, %o0 + 0x18))
234 EX_LD(LOAD_TWIN(%o1, %g2, %g3))
235 add %o1, 16, %o1
236 EX_ST(STORE_INIT(%o4, %o0 + 0x20))
237 EX_ST(STORE_INIT(%o5, %o0 + 0x28))
238 EX_ST(STORE_INIT(%g2, %o0 + 0x30))
239 EX_ST(STORE_INIT(%g3, %o0 + 0x38))
240 subcc %g1, 64, %g1
241 bne,pt %XCC, 1b
242 add %o0, 64, %o0
243 /* fall through */
244
24560:
246 /* %o2 contains any final bytes still needed to be copied
247 * over. If anything is left, we copy it one byte at a time.
248 */
249 RESTORE_ASI
250 brz,pt %o2, 85f
251 sub %o0, %o1, %o3
252 ba,a,pt %XCC, 90f
253
254 .align 64
25570: /* 16 < len <= 64 */
256 bne,pn %XCC, 75f
257 sub %o0, %o1, %o3
258
25972:
260 andn %o2, 0xf, %o4
261 and %o2, 0xf, %o2
2621: subcc %o4, 0x10, %o4
263 EX_LD(LOAD(ldx, %o1, %o5))
264 add %o1, 0x08, %o1
265 EX_LD(LOAD(ldx, %o1, %g1))
266 sub %o1, 0x08, %o1
267 EX_ST(STORE(stx, %o5, %o1 + %o3))
268 add %o1, 0x8, %o1
269 EX_ST(STORE(stx, %g1, %o1 + %o3))
270 bgu,pt %XCC, 1b
271 add %o1, 0x8, %o1
27273: andcc %o2, 0x8, %g0
273 be,pt %XCC, 1f
274 nop
275 sub %o2, 0x8, %o2
276 EX_LD(LOAD(ldx, %o1, %o5))
277 EX_ST(STORE(stx, %o5, %o1 + %o3))
278 add %o1, 0x8, %o1
2791: andcc %o2, 0x4, %g0
280 be,pt %XCC, 1f
281 nop
282 sub %o2, 0x4, %o2
283 EX_LD(LOAD(lduw, %o1, %o5))
284 EX_ST(STORE(stw, %o5, %o1 + %o3))
285 add %o1, 0x4, %o1
2861: cmp %o2, 0
287 be,pt %XCC, 85f
288 nop
289 ba,pt %xcc, 90f
290 nop
291
29275:
293 andcc %o0, 0x7, %g1
294 sub %g1, 0x8, %g1
295 be,pn %icc, 2f
296 sub %g0, %g1, %g1
297 sub %o2, %g1, %o2
298
2991: subcc %g1, 1, %g1
300 EX_LD(LOAD(ldub, %o1, %o5))
301 EX_ST(STORE(stb, %o5, %o1 + %o3))
302 bgu,pt %icc, 1b
303 add %o1, 1, %o1
304
3052: add %o1, %o3, %o0
306 andcc %o1, 0x7, %g1
307 bne,pt %icc, 8f
308 sll %g1, 3, %g1
309
310 cmp %o2, 16
311 bgeu,pt %icc, 72b
312 nop
313 ba,a,pt %xcc, 73b
314
3158: mov 64, %o3
316 andn %o1, 0x7, %o1
317 EX_LD(LOAD(ldx, %o1, %g2))
318 sub %o3, %g1, %o3
319 andn %o2, 0x7, %o4
320 sllx %g2, %g1, %g2
3211: add %o1, 0x8, %o1
322 EX_LD(LOAD(ldx, %o1, %g3))
323 subcc %o4, 0x8, %o4
324 srlx %g3, %o3, %o5
325 or %o5, %g2, %o5
326 EX_ST(STORE(stx, %o5, %o0))
327 add %o0, 0x8, %o0
328 bgu,pt %icc, 1b
329 sllx %g3, %g1, %g2
330
331 srl %g1, 3, %g1
332 andcc %o2, 0x7, %o2
333 be,pn %icc, 85f
334 add %o1, %g1, %o1
335 ba,pt %xcc, 90f
336 sub %o0, %o1, %o3
337
338 .align 64
33980: /* 0 < len <= 16 */
340 andcc %o3, 0x3, %g0
341 bne,pn %XCC, 90f
342 sub %o0, %o1, %o3
343
3441:
345 subcc %o2, 4, %o2
346 EX_LD(LOAD(lduw, %o1, %g1))
347 EX_ST(STORE(stw, %g1, %o1 + %o3))
348 bgu,pt %XCC, 1b
349 add %o1, 4, %o1
350
35185: retl
352 mov EX_RETVAL(GLOBAL_SPARE), %o0
353
354 .align 32
35590:
356 subcc %o2, 1, %o2
357 EX_LD(LOAD(ldub, %o1, %g1))
358 EX_ST(STORE(stb, %g1, %o1 + %o3))
359 bgu,pt %XCC, 90b
360 add %o1, 1, %o1
361 retl
362 mov EX_RETVAL(GLOBAL_SPARE), %o0
363
364 .size FUNC_NAME, .-FUNC_NAME
generated by cgit v1.2.2 (git 2.25.0) at 2024-10-10 15:20:04 -0400