]>
Commit | Line | Data |
---|---|---|
344d0b54 | 1 | /* Optimized mempcpy implementation for POWER7. |
b168057a | 2 | Copyright (C) 2010-2015 Free Software Foundation, Inc. |
344d0b54 LM |
3 | Contributed by Luis Machado <luisgpm@br.ibm.com>. |
4 | This file is part of the GNU C Library. | |
5 | ||
6 | The GNU C Library is free software; you can redistribute it and/or | |
7 | modify it under the terms of the GNU Lesser General Public | |
8 | License as published by the Free Software Foundation; either | |
9 | version 2.1 of the License, or (at your option) any later version. | |
10 | ||
11 | The GNU C Library is distributed in the hope that it will be useful, | |
12 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
14 | Lesser General Public License for more details. | |
15 | ||
16 | You should have received a copy of the GNU Lesser General Public | |
59ba27a6 PE |
17 | License along with the GNU C Library; if not, see |
18 | <http://www.gnu.org/licenses/>. */ | |
344d0b54 LM |
19 | |
20 | #include <sysdep.h> | |
344d0b54 LM |
21 | |
22 | ||
23 | /* __ptr_t [r3] __mempcpy (__ptr_t dst [r3], __ptr_t src [r4], size_t len [r5]); | |
24 | Returns 'dst' + 'len'. */ | |
25 | ||
26 | .machine power7 | |
2d67d91a | 27 | EALIGN (__mempcpy, 5, 0) |
344d0b54 LM |
28 | CALL_MCOUNT 3 |
29 | ||
30 | cmpldi cr1,5,31 | |
31 | neg 0,3 | |
32 | std 3,-16(1) | |
33 | std 31,-8(1) | |
34 | cfi_offset(31,-8) | |
35 | ble cr1,L(copy_LT_32) /* If move < 32 bytes use short move | |
36 | code. */ | |
37 | ||
38 | andi. 11,3,7 /* Check alignment of DST. */ | |
39 | ||
40 | ||
41 | clrldi 10,4,61 /* Check alignment of SRC. */ | |
42 | cmpld cr6,10,11 /* SRC and DST alignments match? */ | |
43 | mr 12,4 | |
44 | mr 31,5 | |
45 | bne cr6,L(copy_GE_32_unaligned) | |
46 | ||
47 | srdi 9,5,3 /* Number of full quadwords remaining. */ | |
48 | ||
49 | beq L(copy_GE_32_aligned_cont) | |
50 | ||
51 | clrldi 0,0,61 | |
52 | mtcrf 0x01,0 | |
53 | subf 31,0,5 | |
54 | ||
55 | /* Get the SRC aligned to 8 bytes. */ | |
56 | ||
57 | 1: bf 31,2f | |
58 | lbz 6,0(12) | |
59 | addi 12,12,1 | |
60 | stb 6,0(3) | |
61 | addi 3,3,1 | |
62 | 2: bf 30,4f | |
63 | lhz 6,0(12) | |
64 | addi 12,12,2 | |
65 | sth 6,0(3) | |
66 | addi 3,3,2 | |
67 | 4: bf 29,0f | |
68 | lwz 6,0(12) | |
69 | addi 12,12,4 | |
70 | stw 6,0(3) | |
71 | addi 3,3,4 | |
72 | 0: | |
73 | clrldi 10,12,61 /* Check alignment of SRC again. */ | |
74 | srdi 9,31,3 /* Number of full doublewords remaining. */ | |
75 | ||
76 | L(copy_GE_32_aligned_cont): | |
77 | ||
78 | clrldi 11,31,61 | |
79 | mtcrf 0x01,9 | |
80 | ||
81 | srdi 8,31,5 | |
82 | cmpldi cr1,9,4 | |
83 | cmpldi cr6,11,0 | |
84 | mr 11,12 | |
85 | ||
86 | /* Copy 1~3 doublewords so the main loop starts | |
87 | at a multiple of 32 bytes. */ | |
88 | ||
89 | bf 30,1f | |
90 | ld 6,0(12) | |
91 | ld 7,8(12) | |
92 | addi 11,12,16 | |
93 | mtctr 8 | |
94 | std 6,0(3) | |
95 | std 7,8(3) | |
96 | addi 10,3,16 | |
97 | bf 31,4f | |
98 | ld 0,16(12) | |
99 | std 0,16(3) | |
100 | blt cr1,3f | |
101 | addi 11,12,24 | |
102 | addi 10,3,24 | |
103 | b 4f | |
104 | ||
105 | .align 4 | |
106 | 1: /* Copy 1 doubleword and set the counter. */ | |
107 | mr 10,3 | |
108 | mtctr 8 | |
109 | bf 31,4f | |
110 | ld 6,0(12) | |
111 | addi 11,12,8 | |
112 | std 6,0(3) | |
113 | addi 10,3,8 | |
114 | ||
115 | /* Main aligned copy loop. Copies 32-bytes at a time. */ | |
116 | .align 4 | |
117 | 4: | |
118 | ld 6,0(11) | |
119 | ld 7,8(11) | |
120 | ld 8,16(11) | |
121 | ld 0,24(11) | |
122 | addi 11,11,32 | |
123 | ||
124 | std 6,0(10) | |
125 | std 7,8(10) | |
126 | std 8,16(10) | |
127 | std 0,24(10) | |
128 | addi 10,10,32 | |
129 | bdnz 4b | |
130 | 3: | |
131 | ||
132 | /* Check for tail bytes. */ | |
133 | rldicr 0,31,0,60 | |
134 | mtcrf 0x01,31 | |
135 | beq cr6,0f | |
136 | ||
137 | .L9: | |
138 | add 3,3,0 | |
139 | add 12,12,0 | |
140 | ||
141 | /* At this point we have a tail of 0-7 bytes and we know that the | |
142 | destination is doubleword-aligned. */ | |
143 | 4: /* Copy 4 bytes. */ | |
144 | bf 29,2f | |
145 | ||
146 | lwz 6,0(12) | |
147 | addi 12,12,4 | |
148 | stw 6,0(3) | |
149 | addi 3,3,4 | |
150 | 2: /* Copy 2 bytes. */ | |
151 | bf 30,1f | |
152 | ||
153 | lhz 6,0(12) | |
154 | addi 12,12,2 | |
155 | sth 6,0(3) | |
156 | addi 3,3,2 | |
157 | 1: /* Copy 1 byte. */ | |
158 | bf 31,0f | |
159 | ||
160 | lbz 6,0(12) | |
161 | stb 6,0(3) | |
162 | 0: /* Return DST + LEN pointer. */ | |
163 | ld 31,-8(1) | |
164 | ld 3,-16(1) | |
165 | add 3,3,5 | |
166 | blr | |
167 | ||
168 | /* Handle copies of 0~31 bytes. */ | |
169 | .align 4 | |
170 | L(copy_LT_32): | |
171 | cmpldi cr6,5,8 | |
172 | mr 12,4 | |
173 | mtcrf 0x01,5 | |
174 | ble cr6,L(copy_LE_8) | |
175 | ||
176 | /* At least 9 bytes to go. */ | |
177 | neg 8,4 | |
178 | clrrdi 11,4,2 | |
179 | andi. 0,8,3 | |
180 | cmpldi cr1,5,16 | |
181 | mr 10,5 | |
182 | beq L(copy_LT_32_aligned) | |
183 | ||
184 | /* Force 4-bytes alignment for SRC. */ | |
185 | mtocrf 0x01,0 | |
186 | subf 10,0,5 | |
187 | 2: bf 30,1f | |
188 | ||
189 | lhz 6,0(12) | |
190 | addi 12,12,2 | |
191 | sth 6,0(3) | |
192 | addi 3,3,2 | |
193 | 1: bf 31,L(end_4bytes_alignment) | |
194 | ||
195 | lbz 6,0(12) | |
196 | addi 12,12,1 | |
197 | stb 6,0(3) | |
198 | addi 3,3,1 | |
199 | ||
200 | .align 4 | |
201 | L(end_4bytes_alignment): | |
202 | cmpldi cr1,10,16 | |
203 | mtcrf 0x01,10 | |
204 | ||
205 | L(copy_LT_32_aligned): | |
206 | /* At least 6 bytes to go, and SRC is word-aligned. */ | |
207 | blt cr1,8f | |
208 | ||
209 | /* Copy 16 bytes. */ | |
210 | lwz 6,0(12) | |
211 | lwz 7,4(12) | |
212 | stw 6,0(3) | |
213 | lwz 8,8(12) | |
214 | stw 7,4(3) | |
215 | lwz 6,12(12) | |
216 | addi 12,12,16 | |
217 | stw 8,8(3) | |
218 | stw 6,12(3) | |
219 | addi 3,3,16 | |
220 | 8: /* Copy 8 bytes. */ | |
221 | bf 28,4f | |
222 | ||
223 | lwz 6,0(12) | |
224 | lwz 7,4(12) | |
225 | addi 12,12,8 | |
226 | stw 6,0(3) | |
227 | stw 7,4(3) | |
228 | addi 3,3,8 | |
229 | 4: /* Copy 4 bytes. */ | |
230 | bf 29,2f | |
231 | ||
232 | lwz 6,0(12) | |
233 | addi 12,12,4 | |
234 | stw 6,0(3) | |
235 | addi 3,3,4 | |
236 | 2: /* Copy 2-3 bytes. */ | |
237 | bf 30,1f | |
238 | ||
239 | lhz 6,0(12) | |
240 | sth 6,0(3) | |
241 | bf 31,0f | |
242 | lbz 7,2(12) | |
243 | stb 7,2(3) | |
244 | ld 3,-16(1) | |
245 | add 3,3,5 | |
246 | blr | |
247 | ||
248 | .align 4 | |
249 | 1: /* Copy 1 byte. */ | |
250 | bf 31,0f | |
251 | ||
252 | lbz 6,0(12) | |
253 | stb 6,0(3) | |
254 | 0: /* Return DST + LEN pointer. */ | |
255 | ld 3,-16(1) | |
256 | add 3,3,5 | |
257 | blr | |
258 | ||
259 | /* Handles copies of 0~8 bytes. */ | |
260 | .align 4 | |
261 | L(copy_LE_8): | |
262 | bne cr6,4f | |
263 | ||
264 | /* Though we could've used ld/std here, they are still | |
265 | slow for unaligned cases. */ | |
266 | ||
267 | lwz 6,0(4) | |
268 | lwz 7,4(4) | |
269 | stw 6,0(3) | |
270 | stw 7,4(3) | |
271 | ld 3,-16(1) /* Return DST + LEN pointer. */ | |
272 | add 3,3,5 | |
273 | blr | |
274 | ||
275 | .align 4 | |
276 | 4: /* Copies 4~7 bytes. */ | |
277 | bf 29,2b | |
278 | ||
279 | lwz 6,0(4) | |
280 | stw 6,0(3) | |
281 | bf 30,5f | |
282 | lhz 7,4(4) | |
283 | sth 7,4(3) | |
284 | bf 31,0f | |
285 | lbz 8,6(4) | |
286 | stb 8,6(3) | |
287 | ld 3,-16(1) | |
288 | add 3,3,5 | |
289 | blr | |
290 | ||
291 | .align 4 | |
292 | 5: /* Copy 1 byte. */ | |
293 | bf 31,0f | |
294 | ||
295 | lbz 6,4(4) | |
296 | stb 6,4(3) | |
297 | ||
298 | 0: /* Return DST + LEN pointer. */ | |
299 | ld 3,-16(1) | |
300 | add 3,3,5 | |
301 | blr | |
302 | ||
303 | /* Handle copies of 32+ bytes where DST is aligned (to quadword) but | |
304 | SRC is not. Use aligned quadword loads from SRC, shifted to realign | |
305 | the data, allowing for aligned DST stores. */ | |
306 | .align 4 | |
307 | L(copy_GE_32_unaligned): | |
308 | clrldi 0,0,60 /* Number of bytes until the 1st | |
309 | quadword. */ | |
310 | andi. 11,3,15 /* Check alignment of DST (against | |
311 | quadwords). */ | |
312 | srdi 9,5,4 /* Number of full quadwords remaining. */ | |
313 | ||
314 | beq L(copy_GE_32_unaligned_cont) | |
315 | ||
316 | /* SRC is not quadword aligned, get it aligned. */ | |
317 | ||
318 | mtcrf 0x01,0 | |
319 | subf 31,0,5 | |
320 | ||
321 | /* Vector instructions work best when proper alignment (16-bytes) | |
322 | is present. Move 0~15 bytes as needed to get DST quadword-aligned. */ | |
323 | 1: /* Copy 1 byte. */ | |
324 | bf 31,2f | |
325 | ||
326 | lbz 6,0(12) | |
327 | addi 12,12,1 | |
328 | stb 6,0(3) | |
329 | addi 3,3,1 | |
330 | 2: /* Copy 2 bytes. */ | |
331 | bf 30,4f | |
332 | ||
333 | lhz 6,0(12) | |
334 | addi 12,12,2 | |
335 | sth 6,0(3) | |
336 | addi 3,3,2 | |
337 | 4: /* Copy 4 bytes. */ | |
338 | bf 29,8f | |
339 | ||
340 | lwz 6,0(12) | |
341 | addi 12,12,4 | |
342 | stw 6,0(3) | |
343 | addi 3,3,4 | |
344 | 8: /* Copy 8 bytes. */ | |
345 | bf 28,0f | |
346 | ||
347 | ld 6,0(12) | |
348 | addi 12,12,8 | |
349 | std 6,0(3) | |
350 | addi 3,3,8 | |
351 | 0: | |
352 | clrldi 10,12,60 /* Check alignment of SRC. */ | |
353 | srdi 9,31,4 /* Number of full quadwords remaining. */ | |
354 | ||
355 | /* The proper alignment is present, it is OK to copy the bytes now. */ | |
356 | L(copy_GE_32_unaligned_cont): | |
357 | ||
358 | /* Setup two indexes to speed up the indexed vector operations. */ | |
359 | clrldi 11,31,60 | |
360 | li 6,16 /* Index for 16-bytes offsets. */ | |
361 | li 7,32 /* Index for 32-bytes offsets. */ | |
362 | cmpldi cr1,11,0 | |
363 | srdi 8,31,5 /* Setup the loop counter. */ | |
364 | mr 10,3 | |
365 | mr 11,12 | |
366 | mtcrf 0x01,9 | |
367 | cmpldi cr6,9,1 | |
759cfef3 AM |
368 | #ifdef __LITTLE_ENDIAN__ |
369 | lvsr 5,0,12 | |
370 | #else | |
371 | lvsl 5,0,12 | |
372 | #endif | |
344d0b54 LM |
373 | lvx 3,0,12 |
374 | bf 31,L(setup_unaligned_loop) | |
375 | ||
376 | /* Copy another 16 bytes to align to 32-bytes due to the loop . */ | |
377 | lvx 4,12,6 | |
759cfef3 AM |
378 | #ifdef __LITTLE_ENDIAN__ |
379 | vperm 6,4,3,5 | |
380 | #else | |
381 | vperm 6,3,4,5 | |
382 | #endif | |
344d0b54 LM |
383 | addi 11,12,16 |
384 | addi 10,3,16 | |
385 | stvx 6,0,3 | |
386 | vor 3,4,4 | |
387 | ||
388 | L(setup_unaligned_loop): | |
389 | mtctr 8 | |
390 | ble cr6,L(end_unaligned_loop) | |
391 | ||
392 | /* Copy 32 bytes at a time using vector instructions. */ | |
393 | .align 4 | |
394 | L(unaligned_loop): | |
395 | ||
396 | /* Note: vr6/vr10 may contain data that was already copied, | |
397 | but in order to get proper alignment, we may have to copy | |
398 | some portions again. This is faster than having unaligned | |
399 | vector instructions though. */ | |
400 | ||
401 | lvx 4,11,6 /* vr4 = r11+16. */ | |
759cfef3 AM |
402 | #ifdef __LITTLE_ENDIAN__ |
403 | vperm 6,4,3,5 | |
404 | #else | |
405 | vperm 6,3,4,5 | |
406 | #endif | |
344d0b54 | 407 | lvx 3,11,7 /* vr3 = r11+32. */ |
759cfef3 AM |
408 | #ifdef __LITTLE_ENDIAN__ |
409 | vperm 10,3,4,5 | |
410 | #else | |
411 | vperm 10,4,3,5 | |
412 | #endif | |
344d0b54 LM |
413 | addi 11,11,32 |
414 | stvx 6,0,10 | |
415 | stvx 10,10,6 | |
416 | addi 10,10,32 | |
417 | ||
418 | bdnz L(unaligned_loop) | |
419 | ||
420 | .align 4 | |
421 | L(end_unaligned_loop): | |
422 | ||
423 | /* Check for tail bytes. */ | |
424 | rldicr 0,31,0,59 | |
425 | mtcrf 0x01,31 | |
426 | beq cr1,0f | |
427 | ||
428 | add 3,3,0 | |
429 | add 12,12,0 | |
430 | ||
431 | /* We have 1~15 tail bytes to copy, and DST is quadword aligned. */ | |
432 | 8: /* Copy 8 bytes. */ | |
433 | bf 28,4f | |
434 | ||
435 | lwz 6,0(12) | |
436 | lwz 7,4(12) | |
437 | addi 12,12,8 | |
438 | stw 6,0(3) | |
439 | stw 7,4(3) | |
440 | addi 3,3,8 | |
441 | 4: /* Copy 4 bytes. */ | |
442 | bf 29,2f | |
443 | ||
444 | lwz 6,0(12) | |
445 | addi 12,12,4 | |
446 | stw 6,0(3) | |
447 | addi 3,3,4 | |
448 | 2: /* Copy 2~3 bytes. */ | |
449 | bf 30,1f | |
450 | ||
451 | lhz 6,0(12) | |
452 | addi 12,12,2 | |
453 | sth 6,0(3) | |
454 | addi 3,3,2 | |
455 | 1: /* Copy 1 byte. */ | |
456 | bf 31,0f | |
457 | ||
458 | lbz 6,0(12) | |
459 | stb 6,0(3) | |
460 | 0: /* Return DST + LEN pointer. */ | |
461 | ld 31,-8(1) | |
462 | ld 3,-16(1) | |
463 | add 3,3,5 | |
464 | blr | |
465 | ||
2d67d91a JM |
466 | END_GEN_TB (__mempcpy,TB_TOCLESS) |
467 | libc_hidden_def (__mempcpy) | |
468 | weak_alias (__mempcpy, mempcpy) | |
344d0b54 | 469 | libc_hidden_builtin_def (mempcpy) |