]>
Commit | Line | Data |
---|---|---|
344d0b54 LM |
1 | /* Optimized mempcpy implementation for POWER7. |
2 | Copyright (C) 2010 Free Software Foundation, Inc. | |
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 | |
17 | License along with the GNU C Library; if not, write to the Free | |
18 | Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston MA | |
19 | 02110-1301 USA. */ | |
20 | ||
21 | #include <sysdep.h> | |
22 | #include <bp-sym.h> | |
23 | #include <bp-asm.h> | |
24 | ||
25 | /* __ptr_t [r3] __mempcpy (__ptr_t dst [r3], __ptr_t src [r4], size_t len [r5]); | |
26 | Returns 'dst' + 'len'. */ | |
27 | ||
28 | .machine power7 | |
29 | EALIGN (BP_SYM (__mempcpy), 5, 0) | |
30 | CALL_MCOUNT | |
31 | ||
32 | stwu 1,-32(1) | |
33 | cfi_adjust_cfa_offset(32) | |
34 | stw 30,20(1) | |
35 | cfi_offset(30,(20-32)) | |
36 | stw 31,24(1) | |
37 | mr 30,3 | |
38 | cmplwi cr1,5,31 | |
39 | neg 0,3 | |
40 | cfi_offset(31,-8) | |
41 | ble cr1,L(copy_LT_32) /* If move < 32 bytes use short move | |
42 | code. */ | |
43 | ||
44 | andi. 11,3,7 /* Check alignment of DST. */ | |
45 | clrlwi 10,4,29 /* Check alignment of SRC. */ | |
46 | cmplw cr6,10,11 /* SRC and DST alignments match? */ | |
47 | mr 12,4 | |
48 | mr 31,5 | |
49 | bne cr6,L(copy_GE_32_unaligned) | |
50 | ||
51 | srwi 9,5,3 /* Number of full quadwords remaining. */ | |
52 | ||
53 | beq L(copy_GE_32_aligned_cont) | |
54 | ||
55 | clrlwi 0,0,29 | |
56 | mtcrf 0x01,0 | |
57 | subf 31,0,5 | |
58 | ||
59 | /* Get the SRC aligned to 8 bytes. */ | |
60 | ||
61 | 1: bf 31,2f | |
62 | lbz 6,0(12) | |
63 | addi 12,12,1 | |
64 | stb 6,0(3) | |
65 | addi 3,3,1 | |
66 | 2: bf 30,4f | |
67 | lhz 6,0(12) | |
68 | addi 12,12,2 | |
69 | sth 6,0(3) | |
70 | addi 3,3,2 | |
71 | 4: bf 29,0f | |
72 | lwz 6,0(12) | |
73 | addi 12,12,4 | |
74 | stw 6,0(3) | |
75 | addi 3,3,4 | |
76 | 0: | |
77 | clrlwi 10,12,29 /* Check alignment of SRC again. */ | |
78 | srwi 9,31,3 /* Number of full doublewords remaining. */ | |
79 | ||
80 | L(copy_GE_32_aligned_cont): | |
81 | ||
82 | clrlwi 11,31,29 | |
83 | mtcrf 0x01,9 | |
84 | ||
85 | srwi 8,31,5 | |
86 | cmplwi cr1,9,4 | |
87 | cmplwi cr6,11,0 | |
88 | mr 11,12 | |
89 | ||
90 | /* Copy 1~3 doublewords so the main loop starts | |
91 | at a multiple of 32 bytes. */ | |
92 | ||
93 | bf 30,1f | |
94 | lfd 6,0(12) | |
95 | lfd 7,8(12) | |
96 | addi 11,12,16 | |
97 | mtctr 8 | |
98 | stfd 6,0(3) | |
99 | stfd 7,8(3) | |
100 | addi 10,3,16 | |
101 | bf 31,4f | |
102 | lfd 0,16(12) | |
103 | stfd 0,16(3) | |
104 | blt cr1,3f | |
105 | addi 11,12,24 | |
106 | addi 10,3,24 | |
107 | b 4f | |
108 | ||
109 | .align 4 | |
110 | 1: /* Copy 1 doubleword and set the counter. */ | |
111 | mr 10,3 | |
112 | mtctr 8 | |
113 | bf 31,4f | |
114 | lfd 6,0(12) | |
115 | addi 11,12,8 | |
116 | stfd 6,0(3) | |
117 | addi 10,3,8 | |
118 | ||
119 | .align 4 | |
120 | 4: /* Main aligned copy loop. Copies 32-bytes at a time. */ | |
121 | lfd 6,0(11) | |
122 | lfd 7,8(11) | |
123 | lfd 8,16(11) | |
124 | lfd 0,24(11) | |
125 | addi 11,11,32 | |
126 | ||
127 | stfd 6,0(10) | |
128 | stfd 7,8(10) | |
129 | stfd 8,16(10) | |
130 | stfd 0,24(10) | |
131 | addi 10,10,32 | |
132 | bdnz 4b | |
133 | 3: | |
134 | ||
135 | /* Check for tail bytes. */ | |
136 | ||
137 | clrrwi 0,31,3 | |
138 | mtcrf 0x01,31 | |
139 | beq cr6,0f | |
140 | ||
141 | .L9: | |
142 | add 3,3,0 | |
143 | add 12,12,0 | |
144 | ||
145 | /* At this point we have a tail of 0-7 bytes and we know that the | |
146 | destination is doubleword-aligned. */ | |
147 | 4: /* Copy 4 bytes. */ | |
148 | bf 29,2f | |
149 | ||
150 | lwz 6,0(12) | |
151 | addi 12,12,4 | |
152 | stw 6,0(3) | |
153 | addi 3,3,4 | |
154 | 2: /* Copy 2 bytes. */ | |
155 | bf 30,1f | |
156 | ||
157 | lhz 6,0(12) | |
158 | addi 12,12,2 | |
159 | sth 6,0(3) | |
160 | addi 3,3,2 | |
161 | 1: /* Copy 1 byte. */ | |
162 | bf 31,0f | |
163 | ||
164 | lbz 6,0(12) | |
165 | stb 6,0(3) | |
166 | 0: /* Return DST + LEN pointer. */ | |
167 | add 3,30,5 | |
168 | lwz 30,20(1) | |
169 | lwz 31,24(1) | |
170 | addi 1,1,32 | |
171 | blr | |
172 | ||
173 | /* Handle copies of 0~31 bytes. */ | |
174 | .align 4 | |
175 | L(copy_LT_32): | |
176 | cmplwi cr6,5,8 | |
177 | mr 12,4 | |
178 | mtcrf 0x01,5 | |
179 | ble cr6,L(copy_LE_8) | |
180 | ||
181 | /* At least 9 bytes to go. */ | |
182 | neg 8,4 | |
183 | clrrwi 11,4,2 | |
184 | andi. 0,8,3 | |
185 | cmplwi cr1,5,16 | |
186 | mr 10,5 | |
187 | beq L(copy_LT_32_aligned) | |
188 | ||
189 | /* Force 4-bytes alignment for SRC. */ | |
190 | mtocrf 0x01,0 | |
191 | subf 10,0,5 | |
192 | 2: bf 30,1f | |
193 | ||
194 | lhz 6,0(12) | |
195 | addi 12,12,2 | |
196 | sth 6,0(3) | |
197 | addi 3,3,2 | |
198 | 1: bf 31,L(end_4bytes_alignment) | |
199 | ||
200 | lbz 6,0(12) | |
201 | addi 12,12,1 | |
202 | stb 6,0(3) | |
203 | addi 3,3,1 | |
204 | ||
205 | .align 4 | |
206 | L(end_4bytes_alignment): | |
207 | cmplwi cr1,10,16 | |
208 | mtcrf 0x01,10 | |
209 | ||
210 | L(copy_LT_32_aligned): | |
211 | /* At least 6 bytes to go, and SRC is word-aligned. */ | |
212 | blt cr1,8f | |
213 | ||
214 | /* Copy 16 bytes. */ | |
215 | lwz 6,0(12) | |
216 | lwz 7,4(12) | |
217 | stw 6,0(3) | |
218 | lwz 8,8(12) | |
219 | stw 7,4(3) | |
220 | lwz 6,12(12) | |
221 | addi 12,12,16 | |
222 | stw 8,8(3) | |
223 | stw 6,12(3) | |
224 | addi 3,3,16 | |
225 | 8: /* Copy 8 bytes. */ | |
226 | bf 28,4f | |
227 | ||
228 | lwz 6,0(12) | |
229 | lwz 7,4(12) | |
230 | addi 12,12,8 | |
231 | stw 6,0(3) | |
232 | stw 7,4(3) | |
233 | addi 3,3,8 | |
234 | 4: /* Copy 4 bytes. */ | |
235 | bf 29,2f | |
236 | ||
237 | lwz 6,0(12) | |
238 | addi 12,12,4 | |
239 | stw 6,0(3) | |
240 | addi 3,3,4 | |
241 | 2: /* Copy 2-3 bytes. */ | |
242 | bf 30,1f | |
243 | ||
244 | lhz 6,0(12) | |
245 | sth 6,0(3) | |
246 | bf 31,0f | |
247 | lbz 7,2(12) | |
248 | stb 7,2(3) | |
249 | ||
250 | /* Return DST + LEN pointer. */ | |
251 | add 3,30,5 | |
252 | lwz 30,20(1) | |
253 | addi 1,1,32 | |
254 | blr | |
255 | ||
256 | .align 4 | |
257 | 1: /* Copy 1 byte. */ | |
258 | bf 31,0f | |
259 | ||
260 | lbz 6,0(12) | |
261 | stb 6,0(3) | |
262 | 0: /* Return DST + LEN pointer. */ | |
263 | add 3,30,5 | |
264 | lwz 30,20(1) | |
265 | addi 1,1,32 | |
266 | blr | |
267 | ||
268 | /* Handles copies of 0~8 bytes. */ | |
269 | .align 4 | |
270 | L(copy_LE_8): | |
271 | bne cr6,4f | |
272 | ||
273 | /* Though we could've used lfd/stfd here, they are still | |
274 | slow for unaligned cases. */ | |
275 | ||
276 | lwz 6,0(4) | |
277 | lwz 7,4(4) | |
278 | stw 6,0(3) | |
279 | stw 7,4(3) | |
280 | ||
281 | /* Return DST + LEN pointer. */ | |
282 | add 3,30,5 | |
283 | lwz 30,20(1) | |
284 | addi 1,1,32 | |
285 | blr | |
286 | ||
287 | .align 4 | |
288 | 4: /* Copies 4~7 bytes. */ | |
289 | bf 29,2b | |
290 | ||
291 | lwz 6,0(4) | |
292 | stw 6,0(3) | |
293 | bf 30,5f | |
294 | lhz 7,4(4) | |
295 | sth 7,4(3) | |
296 | bf 31,0f | |
297 | lbz 8,6(4) | |
298 | stb 8,6(3) | |
299 | ||
300 | /* Return DST + LEN pointer. */ | |
301 | add 3,30,5 | |
302 | lwz 30,20(1) | |
303 | addi 1,1,32 | |
304 | blr | |
305 | ||
306 | .align 4 | |
307 | 5: /* Copy 1 byte. */ | |
308 | bf 31,0f | |
309 | ||
310 | lbz 6,4(4) | |
311 | stb 6,4(3) | |
312 | ||
313 | 0: /* Return DST + LEN pointer. */ | |
314 | add 3,30,5 | |
315 | lwz 30,20(1) | |
316 | addi 1,1,32 | |
317 | blr | |
318 | ||
319 | /* Handle copies of 32+ bytes where DST is aligned (to quadword) but | |
320 | SRC is not. Use aligned quadword loads from SRC, shifted to realign | |
321 | the data, allowing for aligned DST stores. */ | |
322 | .align 4 | |
323 | L(copy_GE_32_unaligned): | |
324 | andi. 11,3,15 /* Check alignment of DST. */ | |
325 | clrlwi 0,0,28 /* Number of bytes until the 1st | |
326 | quadword of DST. */ | |
327 | srwi 9,5,4 /* Number of full quadwords remaining. */ | |
328 | ||
329 | beq L(copy_GE_32_unaligned_cont) | |
330 | ||
331 | /* SRC is not quadword aligned, get it aligned. */ | |
332 | ||
333 | mtcrf 0x01,0 | |
334 | subf 31,0,5 | |
335 | ||
336 | /* Vector instructions work best when proper alignment (16-bytes) | |
337 | is present. Move 0~15 bytes as needed to get DST quadword-aligned. */ | |
338 | 1: /* Copy 1 byte. */ | |
339 | bf 31,2f | |
340 | ||
341 | lbz 6,0(12) | |
342 | addi 12,12,1 | |
343 | stb 6,0(3) | |
344 | addi 3,3,1 | |
345 | 2: /* Copy 2 bytes. */ | |
346 | bf 30,4f | |
347 | ||
348 | lhz 6,0(12) | |
349 | addi 12,12,2 | |
350 | sth 6,0(3) | |
351 | addi 3,3,2 | |
352 | 4: /* Copy 4 bytes. */ | |
353 | bf 29,8f | |
354 | ||
355 | lwz 6,0(12) | |
356 | addi 12,12,4 | |
357 | stw 6,0(3) | |
358 | addi 3,3,4 | |
359 | 8: /* Copy 8 bytes. */ | |
360 | bf 28,0f | |
361 | ||
362 | lfd 6,0(12) | |
363 | addi 12,12,8 | |
364 | stfd 6,0(3) | |
365 | addi 3,3,8 | |
366 | 0: | |
367 | clrlwi 10,12,28 /* Check alignment of SRC. */ | |
368 | srwi 9,31,4 /* Number of full quadwords remaining. */ | |
369 | ||
370 | /* The proper alignment is present, it is OK to copy the bytes now. */ | |
371 | L(copy_GE_32_unaligned_cont): | |
372 | ||
373 | /* Setup two indexes to speed up the indexed vector operations. */ | |
374 | clrlwi 11,31,28 | |
375 | li 6,16 /* Index for 16-bytes offsets. */ | |
376 | li 7,32 /* Index for 32-bytes offsets. */ | |
377 | cmplwi cr1,11,0 | |
378 | srwi 8,31,5 /* Setup the loop counter. */ | |
379 | mr 10,3 | |
380 | mr 11,12 | |
381 | mtcrf 0x01,9 | |
382 | cmplwi cr6,9,1 | |
383 | lvsl 5,0,12 | |
384 | lvx 3,0,12 | |
385 | bf 31,L(setup_unaligned_loop) | |
386 | ||
387 | /* Copy another 16 bytes to align to 32-bytes due to the loop . */ | |
388 | lvx 4,12,6 | |
389 | vperm 6,3,4,5 | |
390 | addi 11,12,16 | |
391 | addi 10,3,16 | |
392 | stvx 6,0,3 | |
393 | vor 3,4,4 | |
394 | ||
395 | L(setup_unaligned_loop): | |
396 | mtctr 8 | |
397 | ble cr6,L(end_unaligned_loop) | |
398 | ||
399 | /* Copy 32 bytes at a time using vector instructions. */ | |
400 | .align 4 | |
401 | L(unaligned_loop): | |
402 | ||
403 | /* Note: vr6/vr10 may contain data that was already copied, | |
404 | but in order to get proper alignment, we may have to copy | |
405 | some portions again. This is faster than having unaligned | |
406 | vector instructions though. */ | |
407 | ||
408 | lvx 4,11,6 /* vr4 = r11+16. */ | |
409 | vperm 6,3,4,5 /* Merge the correctly-aligned portions | |
410 | of vr3/vr4 into vr6. */ | |
411 | lvx 3,11,7 /* vr3 = r11+32. */ | |
412 | vperm 10,4,3,5 /* Merge the correctly-aligned portions | |
413 | of vr3/vr4 into vr10. */ | |
414 | addi 11,11,32 | |
415 | stvx 6,0,10 | |
416 | stvx 10,10,6 | |
417 | addi 10,10,32 | |
418 | ||
419 | bdnz L(unaligned_loop) | |
420 | ||
421 | .align 4 | |
422 | L(end_unaligned_loop): | |
423 | ||
424 | /* Check for tail bytes. */ | |
425 | clrrwi 0,31,4 | |
426 | mtcrf 0x01,31 | |
427 | beq cr1,0f | |
428 | ||
429 | add 3,3,0 | |
430 | add 12,12,0 | |
431 | ||
432 | /* We have 1~15 tail bytes to copy, and DST is quadword aligned. */ | |
433 | 8: /* Copy 8 bytes. */ | |
434 | bf 28,4f | |
435 | ||
436 | lwz 6,0(12) | |
437 | lwz 7,4(12) | |
438 | addi 12,12,8 | |
439 | stw 6,0(3) | |
440 | stw 7,4(3) | |
441 | addi 3,3,8 | |
442 | 4: /* Copy 4 bytes. */ | |
443 | bf 29,2f | |
444 | ||
445 | lwz 6,0(12) | |
446 | addi 12,12,4 | |
447 | stw 6,0(3) | |
448 | addi 3,3,4 | |
449 | 2: /* Copy 2~3 bytes. */ | |
450 | bf 30,1f | |
451 | ||
452 | lhz 6,0(12) | |
453 | addi 12,12,2 | |
454 | sth 6,0(3) | |
455 | addi 3,3,2 | |
456 | 1: /* Copy 1 byte. */ | |
457 | bf 31,0f | |
458 | ||
459 | lbz 6,0(12) | |
460 | stb 6,0(3) | |
461 | 0: /* Return DST + LEN pointer. */ | |
462 | add 3,30,5 | |
463 | lwz 30,20(1) | |
464 | lwz 31,24(1) | |
465 | addi 1,1,32 | |
466 | blr | |
467 | ||
468 | END (BP_SYM (__mempcpy)) | |
469 | libc_hidden_def (BP_SYM (__mempcpy)) | |
470 | weak_alias (BP_SYM (__mempcpy), BP_SYM (mempcpy)) | |
471 | libc_hidden_builtin_def (mempcpy) |