xref: /dpdk/lib/eal/x86/include/rte_memcpy.h (revision daa02b5cddbb8e11b31d41e2bf7bb1ae64dcae2f)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2010-2014 Intel Corporation
3  */
4 
5 #ifndef _RTE_MEMCPY_X86_64_H_
6 #define _RTE_MEMCPY_X86_64_H_
7 
8 /**
9  * @file
10  *
11  * Functions for SSE/AVX/AVX2/AVX512 implementation of memcpy().
12  */
13 
14 #include <stdio.h>
15 #include <stdint.h>
16 #include <string.h>
17 #include <rte_vect.h>
18 #include <rte_common.h>
19 #include <rte_config.h>
20 
21 #ifdef __cplusplus
22 extern "C" {
23 #endif
24 
25 #if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION >= 100000)
26 #pragma GCC diagnostic push
27 #pragma GCC diagnostic ignored "-Wstringop-overflow"
28 #endif
29 
30 /**
31  * Copy bytes from one location to another. The locations must not overlap.
32  *
33  * @note This is implemented as a macro, so it's address should not be taken
34  * and care is needed as parameter expressions may be evaluated multiple times.
35  *
36  * @param dst
37  *   Pointer to the destination of the data.
38  * @param src
39  *   Pointer to the source data.
40  * @param n
41  *   Number of bytes to copy.
42  * @return
43  *   Pointer to the destination data.
44  */
45 static __rte_always_inline void *
46 rte_memcpy(void *dst, const void *src, size_t n);
47 
48 #if defined __AVX512F__ && defined RTE_MEMCPY_AVX512
49 
50 #define ALIGNMENT_MASK 0x3F
51 
52 /**
53  * AVX512 implementation below
54  */
55 
56 /**
57  * Copy 16 bytes from one location to another,
58  * locations should not overlap.
59  */
60 static __rte_always_inline void
61 rte_mov16(uint8_t *dst, const uint8_t *src)
62 {
63 	__m128i xmm0;
64 
65 	xmm0 = _mm_loadu_si128((const __m128i *)src);
66 	_mm_storeu_si128((__m128i *)dst, xmm0);
67 }
68 
69 /**
70  * Copy 32 bytes from one location to another,
71  * locations should not overlap.
72  */
73 static __rte_always_inline void
74 rte_mov32(uint8_t *dst, const uint8_t *src)
75 {
76 	__m256i ymm0;
77 
78 	ymm0 = _mm256_loadu_si256((const __m256i *)src);
79 	_mm256_storeu_si256((__m256i *)dst, ymm0);
80 }
81 
82 /**
83  * Copy 64 bytes from one location to another,
84  * locations should not overlap.
85  */
86 static __rte_always_inline void
87 rte_mov64(uint8_t *dst, const uint8_t *src)
88 {
89 	__m512i zmm0;
90 
91 	zmm0 = _mm512_loadu_si512((const void *)src);
92 	_mm512_storeu_si512((void *)dst, zmm0);
93 }
94 
95 /**
96  * Copy 128 bytes from one location to another,
97  * locations should not overlap.
98  */
99 static __rte_always_inline void
100 rte_mov128(uint8_t *dst, const uint8_t *src)
101 {
102 	rte_mov64(dst + 0 * 64, src + 0 * 64);
103 	rte_mov64(dst + 1 * 64, src + 1 * 64);
104 }
105 
106 /**
107  * Copy 256 bytes from one location to another,
108  * locations should not overlap.
109  */
110 static __rte_always_inline void
111 rte_mov256(uint8_t *dst, const uint8_t *src)
112 {
113 	rte_mov64(dst + 0 * 64, src + 0 * 64);
114 	rte_mov64(dst + 1 * 64, src + 1 * 64);
115 	rte_mov64(dst + 2 * 64, src + 2 * 64);
116 	rte_mov64(dst + 3 * 64, src + 3 * 64);
117 }
118 
119 /**
120  * Copy 128-byte blocks from one location to another,
121  * locations should not overlap.
122  */
123 static __rte_always_inline void
124 rte_mov128blocks(uint8_t *dst, const uint8_t *src, size_t n)
125 {
126 	__m512i zmm0, zmm1;
127 
128 	while (n >= 128) {
129 		zmm0 = _mm512_loadu_si512((const void *)(src + 0 * 64));
130 		n -= 128;
131 		zmm1 = _mm512_loadu_si512((const void *)(src + 1 * 64));
132 		src = src + 128;
133 		_mm512_storeu_si512((void *)(dst + 0 * 64), zmm0);
134 		_mm512_storeu_si512((void *)(dst + 1 * 64), zmm1);
135 		dst = dst + 128;
136 	}
137 }
138 
139 /**
140  * Copy 512-byte blocks from one location to another,
141  * locations should not overlap.
142  */
143 static inline void
144 rte_mov512blocks(uint8_t *dst, const uint8_t *src, size_t n)
145 {
146 	__m512i zmm0, zmm1, zmm2, zmm3, zmm4, zmm5, zmm6, zmm7;
147 
148 	while (n >= 512) {
149 		zmm0 = _mm512_loadu_si512((const void *)(src + 0 * 64));
150 		n -= 512;
151 		zmm1 = _mm512_loadu_si512((const void *)(src + 1 * 64));
152 		zmm2 = _mm512_loadu_si512((const void *)(src + 2 * 64));
153 		zmm3 = _mm512_loadu_si512((const void *)(src + 3 * 64));
154 		zmm4 = _mm512_loadu_si512((const void *)(src + 4 * 64));
155 		zmm5 = _mm512_loadu_si512((const void *)(src + 5 * 64));
156 		zmm6 = _mm512_loadu_si512((const void *)(src + 6 * 64));
157 		zmm7 = _mm512_loadu_si512((const void *)(src + 7 * 64));
158 		src = src + 512;
159 		_mm512_storeu_si512((void *)(dst + 0 * 64), zmm0);
160 		_mm512_storeu_si512((void *)(dst + 1 * 64), zmm1);
161 		_mm512_storeu_si512((void *)(dst + 2 * 64), zmm2);
162 		_mm512_storeu_si512((void *)(dst + 3 * 64), zmm3);
163 		_mm512_storeu_si512((void *)(dst + 4 * 64), zmm4);
164 		_mm512_storeu_si512((void *)(dst + 5 * 64), zmm5);
165 		_mm512_storeu_si512((void *)(dst + 6 * 64), zmm6);
166 		_mm512_storeu_si512((void *)(dst + 7 * 64), zmm7);
167 		dst = dst + 512;
168 	}
169 }
170 
171 static __rte_always_inline void *
172 rte_memcpy_generic(void *dst, const void *src, size_t n)
173 {
174 	uintptr_t dstu = (uintptr_t)dst;
175 	uintptr_t srcu = (uintptr_t)src;
176 	void *ret = dst;
177 	size_t dstofss;
178 	size_t bits;
179 
180 	/**
181 	 * Copy less than 16 bytes
182 	 */
183 	if (n < 16) {
184 		if (n & 0x01) {
185 			*(uint8_t *)dstu = *(const uint8_t *)srcu;
186 			srcu = (uintptr_t)((const uint8_t *)srcu + 1);
187 			dstu = (uintptr_t)((uint8_t *)dstu + 1);
188 		}
189 		if (n & 0x02) {
190 			*(uint16_t *)dstu = *(const uint16_t *)srcu;
191 			srcu = (uintptr_t)((const uint16_t *)srcu + 1);
192 			dstu = (uintptr_t)((uint16_t *)dstu + 1);
193 		}
194 		if (n & 0x04) {
195 			*(uint32_t *)dstu = *(const uint32_t *)srcu;
196 			srcu = (uintptr_t)((const uint32_t *)srcu + 1);
197 			dstu = (uintptr_t)((uint32_t *)dstu + 1);
198 		}
199 		if (n & 0x08)
200 			*(uint64_t *)dstu = *(const uint64_t *)srcu;
201 		return ret;
202 	}
203 
204 	/**
205 	 * Fast way when copy size doesn't exceed 512 bytes
206 	 */
207 	if (n <= 32) {
208 		rte_mov16((uint8_t *)dst, (const uint8_t *)src);
209 		rte_mov16((uint8_t *)dst - 16 + n,
210 				  (const uint8_t *)src - 16 + n);
211 		return ret;
212 	}
213 	if (n <= 64) {
214 		rte_mov32((uint8_t *)dst, (const uint8_t *)src);
215 		rte_mov32((uint8_t *)dst - 32 + n,
216 				  (const uint8_t *)src - 32 + n);
217 		return ret;
218 	}
219 	if (n <= 512) {
220 		if (n >= 256) {
221 			n -= 256;
222 			rte_mov256((uint8_t *)dst, (const uint8_t *)src);
223 			src = (const uint8_t *)src + 256;
224 			dst = (uint8_t *)dst + 256;
225 		}
226 		if (n >= 128) {
227 			n -= 128;
228 			rte_mov128((uint8_t *)dst, (const uint8_t *)src);
229 			src = (const uint8_t *)src + 128;
230 			dst = (uint8_t *)dst + 128;
231 		}
232 COPY_BLOCK_128_BACK63:
233 		if (n > 64) {
234 			rte_mov64((uint8_t *)dst, (const uint8_t *)src);
235 			rte_mov64((uint8_t *)dst - 64 + n,
236 					  (const uint8_t *)src - 64 + n);
237 			return ret;
238 		}
239 		if (n > 0)
240 			rte_mov64((uint8_t *)dst - 64 + n,
241 					  (const uint8_t *)src - 64 + n);
242 		return ret;
243 	}
244 
245 	/**
246 	 * Make store aligned when copy size exceeds 512 bytes
247 	 */
248 	dstofss = ((uintptr_t)dst & 0x3F);
249 	if (dstofss > 0) {
250 		dstofss = 64 - dstofss;
251 		n -= dstofss;
252 		rte_mov64((uint8_t *)dst, (const uint8_t *)src);
253 		src = (const uint8_t *)src + dstofss;
254 		dst = (uint8_t *)dst + dstofss;
255 	}
256 
257 	/**
258 	 * Copy 512-byte blocks.
259 	 * Use copy block function for better instruction order control,
260 	 * which is important when load is unaligned.
261 	 */
262 	rte_mov512blocks((uint8_t *)dst, (const uint8_t *)src, n);
263 	bits = n;
264 	n = n & 511;
265 	bits -= n;
266 	src = (const uint8_t *)src + bits;
267 	dst = (uint8_t *)dst + bits;
268 
269 	/**
270 	 * Copy 128-byte blocks.
271 	 * Use copy block function for better instruction order control,
272 	 * which is important when load is unaligned.
273 	 */
274 	if (n >= 128) {
275 		rte_mov128blocks((uint8_t *)dst, (const uint8_t *)src, n);
276 		bits = n;
277 		n = n & 127;
278 		bits -= n;
279 		src = (const uint8_t *)src + bits;
280 		dst = (uint8_t *)dst + bits;
281 	}
282 
283 	/**
284 	 * Copy whatever left
285 	 */
286 	goto COPY_BLOCK_128_BACK63;
287 }
288 
289 #elif defined __AVX2__
290 
291 #define ALIGNMENT_MASK 0x1F
292 
293 /**
294  * AVX2 implementation below
295  */
296 
297 /**
298  * Copy 16 bytes from one location to another,
299  * locations should not overlap.
300  */
301 static __rte_always_inline void
302 rte_mov16(uint8_t *dst, const uint8_t *src)
303 {
304 	__m128i xmm0;
305 
306 	xmm0 = _mm_loadu_si128((const __m128i *)src);
307 	_mm_storeu_si128((__m128i *)dst, xmm0);
308 }
309 
310 /**
311  * Copy 32 bytes from one location to another,
312  * locations should not overlap.
313  */
314 static __rte_always_inline void
315 rte_mov32(uint8_t *dst, const uint8_t *src)
316 {
317 	__m256i ymm0;
318 
319 	ymm0 = _mm256_loadu_si256((const __m256i *)src);
320 	_mm256_storeu_si256((__m256i *)dst, ymm0);
321 }
322 
323 /**
324  * Copy 64 bytes from one location to another,
325  * locations should not overlap.
326  */
327 static __rte_always_inline void
328 rte_mov64(uint8_t *dst, const uint8_t *src)
329 {
330 	rte_mov32((uint8_t *)dst + 0 * 32, (const uint8_t *)src + 0 * 32);
331 	rte_mov32((uint8_t *)dst + 1 * 32, (const uint8_t *)src + 1 * 32);
332 }
333 
334 /**
335  * Copy 128 bytes from one location to another,
336  * locations should not overlap.
337  */
338 static __rte_always_inline void
339 rte_mov128(uint8_t *dst, const uint8_t *src)
340 {
341 	rte_mov32((uint8_t *)dst + 0 * 32, (const uint8_t *)src + 0 * 32);
342 	rte_mov32((uint8_t *)dst + 1 * 32, (const uint8_t *)src + 1 * 32);
343 	rte_mov32((uint8_t *)dst + 2 * 32, (const uint8_t *)src + 2 * 32);
344 	rte_mov32((uint8_t *)dst + 3 * 32, (const uint8_t *)src + 3 * 32);
345 }
346 
347 /**
348  * Copy 128-byte blocks from one location to another,
349  * locations should not overlap.
350  */
351 static __rte_always_inline void
352 rte_mov128blocks(uint8_t *dst, const uint8_t *src, size_t n)
353 {
354 	__m256i ymm0, ymm1, ymm2, ymm3;
355 
356 	while (n >= 128) {
357 		ymm0 = _mm256_loadu_si256((const __m256i *)((const uint8_t *)src + 0 * 32));
358 		n -= 128;
359 		ymm1 = _mm256_loadu_si256((const __m256i *)((const uint8_t *)src + 1 * 32));
360 		ymm2 = _mm256_loadu_si256((const __m256i *)((const uint8_t *)src + 2 * 32));
361 		ymm3 = _mm256_loadu_si256((const __m256i *)((const uint8_t *)src + 3 * 32));
362 		src = (const uint8_t *)src + 128;
363 		_mm256_storeu_si256((__m256i *)((uint8_t *)dst + 0 * 32), ymm0);
364 		_mm256_storeu_si256((__m256i *)((uint8_t *)dst + 1 * 32), ymm1);
365 		_mm256_storeu_si256((__m256i *)((uint8_t *)dst + 2 * 32), ymm2);
366 		_mm256_storeu_si256((__m256i *)((uint8_t *)dst + 3 * 32), ymm3);
367 		dst = (uint8_t *)dst + 128;
368 	}
369 }
370 
371 static __rte_always_inline void *
372 rte_memcpy_generic(void *dst, const void *src, size_t n)
373 {
374 	uintptr_t dstu = (uintptr_t)dst;
375 	uintptr_t srcu = (uintptr_t)src;
376 	void *ret = dst;
377 	size_t dstofss;
378 	size_t bits;
379 
380 	/**
381 	 * Copy less than 16 bytes
382 	 */
383 	if (n < 16) {
384 		if (n & 0x01) {
385 			*(uint8_t *)dstu = *(const uint8_t *)srcu;
386 			srcu = (uintptr_t)((const uint8_t *)srcu + 1);
387 			dstu = (uintptr_t)((uint8_t *)dstu + 1);
388 		}
389 		if (n & 0x02) {
390 			*(uint16_t *)dstu = *(const uint16_t *)srcu;
391 			srcu = (uintptr_t)((const uint16_t *)srcu + 1);
392 			dstu = (uintptr_t)((uint16_t *)dstu + 1);
393 		}
394 		if (n & 0x04) {
395 			*(uint32_t *)dstu = *(const uint32_t *)srcu;
396 			srcu = (uintptr_t)((const uint32_t *)srcu + 1);
397 			dstu = (uintptr_t)((uint32_t *)dstu + 1);
398 		}
399 		if (n & 0x08) {
400 			*(uint64_t *)dstu = *(const uint64_t *)srcu;
401 		}
402 		return ret;
403 	}
404 
405 	/**
406 	 * Fast way when copy size doesn't exceed 256 bytes
407 	 */
408 	if (n <= 32) {
409 		rte_mov16((uint8_t *)dst, (const uint8_t *)src);
410 		rte_mov16((uint8_t *)dst - 16 + n,
411 				(const uint8_t *)src - 16 + n);
412 		return ret;
413 	}
414 	if (n <= 48) {
415 		rte_mov16((uint8_t *)dst, (const uint8_t *)src);
416 		rte_mov16((uint8_t *)dst + 16, (const uint8_t *)src + 16);
417 		rte_mov16((uint8_t *)dst - 16 + n,
418 				(const uint8_t *)src - 16 + n);
419 		return ret;
420 	}
421 	if (n <= 64) {
422 		rte_mov32((uint8_t *)dst, (const uint8_t *)src);
423 		rte_mov32((uint8_t *)dst - 32 + n,
424 				(const uint8_t *)src - 32 + n);
425 		return ret;
426 	}
427 	if (n <= 256) {
428 		if (n >= 128) {
429 			n -= 128;
430 			rte_mov128((uint8_t *)dst, (const uint8_t *)src);
431 			src = (const uint8_t *)src + 128;
432 			dst = (uint8_t *)dst + 128;
433 		}
434 COPY_BLOCK_128_BACK31:
435 		if (n >= 64) {
436 			n -= 64;
437 			rte_mov64((uint8_t *)dst, (const uint8_t *)src);
438 			src = (const uint8_t *)src + 64;
439 			dst = (uint8_t *)dst + 64;
440 		}
441 		if (n > 32) {
442 			rte_mov32((uint8_t *)dst, (const uint8_t *)src);
443 			rte_mov32((uint8_t *)dst - 32 + n,
444 					(const uint8_t *)src - 32 + n);
445 			return ret;
446 		}
447 		if (n > 0) {
448 			rte_mov32((uint8_t *)dst - 32 + n,
449 					(const uint8_t *)src - 32 + n);
450 		}
451 		return ret;
452 	}
453 
454 	/**
455 	 * Make store aligned when copy size exceeds 256 bytes
456 	 */
457 	dstofss = (uintptr_t)dst & 0x1F;
458 	if (dstofss > 0) {
459 		dstofss = 32 - dstofss;
460 		n -= dstofss;
461 		rte_mov32((uint8_t *)dst, (const uint8_t *)src);
462 		src = (const uint8_t *)src + dstofss;
463 		dst = (uint8_t *)dst + dstofss;
464 	}
465 
466 	/**
467 	 * Copy 128-byte blocks
468 	 */
469 	rte_mov128blocks((uint8_t *)dst, (const uint8_t *)src, n);
470 	bits = n;
471 	n = n & 127;
472 	bits -= n;
473 	src = (const uint8_t *)src + bits;
474 	dst = (uint8_t *)dst + bits;
475 
476 	/**
477 	 * Copy whatever left
478 	 */
479 	goto COPY_BLOCK_128_BACK31;
480 }
481 
482 #else /* __AVX512F__ */
483 
484 #define ALIGNMENT_MASK 0x0F
485 
486 /**
487  * SSE & AVX implementation below
488  */
489 
490 /**
491  * Copy 16 bytes from one location to another,
492  * locations should not overlap.
493  */
494 static __rte_always_inline void
495 rte_mov16(uint8_t *dst, const uint8_t *src)
496 {
497 	__m128i xmm0;
498 
499 	xmm0 = _mm_loadu_si128((const __m128i *)(const __m128i *)src);
500 	_mm_storeu_si128((__m128i *)dst, xmm0);
501 }
502 
503 /**
504  * Copy 32 bytes from one location to another,
505  * locations should not overlap.
506  */
507 static __rte_always_inline void
508 rte_mov32(uint8_t *dst, const uint8_t *src)
509 {
510 	rte_mov16((uint8_t *)dst + 0 * 16, (const uint8_t *)src + 0 * 16);
511 	rte_mov16((uint8_t *)dst + 1 * 16, (const uint8_t *)src + 1 * 16);
512 }
513 
514 /**
515  * Copy 64 bytes from one location to another,
516  * locations should not overlap.
517  */
518 static __rte_always_inline void
519 rte_mov64(uint8_t *dst, const uint8_t *src)
520 {
521 	rte_mov16((uint8_t *)dst + 0 * 16, (const uint8_t *)src + 0 * 16);
522 	rte_mov16((uint8_t *)dst + 1 * 16, (const uint8_t *)src + 1 * 16);
523 	rte_mov16((uint8_t *)dst + 2 * 16, (const uint8_t *)src + 2 * 16);
524 	rte_mov16((uint8_t *)dst + 3 * 16, (const uint8_t *)src + 3 * 16);
525 }
526 
527 /**
528  * Copy 128 bytes from one location to another,
529  * locations should not overlap.
530  */
531 static __rte_always_inline void
532 rte_mov128(uint8_t *dst, const uint8_t *src)
533 {
534 	rte_mov16((uint8_t *)dst + 0 * 16, (const uint8_t *)src + 0 * 16);
535 	rte_mov16((uint8_t *)dst + 1 * 16, (const uint8_t *)src + 1 * 16);
536 	rte_mov16((uint8_t *)dst + 2 * 16, (const uint8_t *)src + 2 * 16);
537 	rte_mov16((uint8_t *)dst + 3 * 16, (const uint8_t *)src + 3 * 16);
538 	rte_mov16((uint8_t *)dst + 4 * 16, (const uint8_t *)src + 4 * 16);
539 	rte_mov16((uint8_t *)dst + 5 * 16, (const uint8_t *)src + 5 * 16);
540 	rte_mov16((uint8_t *)dst + 6 * 16, (const uint8_t *)src + 6 * 16);
541 	rte_mov16((uint8_t *)dst + 7 * 16, (const uint8_t *)src + 7 * 16);
542 }
543 
544 /**
545  * Copy 256 bytes from one location to another,
546  * locations should not overlap.
547  */
548 static inline void
549 rte_mov256(uint8_t *dst, const uint8_t *src)
550 {
551 	rte_mov16((uint8_t *)dst + 0 * 16, (const uint8_t *)src + 0 * 16);
552 	rte_mov16((uint8_t *)dst + 1 * 16, (const uint8_t *)src + 1 * 16);
553 	rte_mov16((uint8_t *)dst + 2 * 16, (const uint8_t *)src + 2 * 16);
554 	rte_mov16((uint8_t *)dst + 3 * 16, (const uint8_t *)src + 3 * 16);
555 	rte_mov16((uint8_t *)dst + 4 * 16, (const uint8_t *)src + 4 * 16);
556 	rte_mov16((uint8_t *)dst + 5 * 16, (const uint8_t *)src + 5 * 16);
557 	rte_mov16((uint8_t *)dst + 6 * 16, (const uint8_t *)src + 6 * 16);
558 	rte_mov16((uint8_t *)dst + 7 * 16, (const uint8_t *)src + 7 * 16);
559 	rte_mov16((uint8_t *)dst + 8 * 16, (const uint8_t *)src + 8 * 16);
560 	rte_mov16((uint8_t *)dst + 9 * 16, (const uint8_t *)src + 9 * 16);
561 	rte_mov16((uint8_t *)dst + 10 * 16, (const uint8_t *)src + 10 * 16);
562 	rte_mov16((uint8_t *)dst + 11 * 16, (const uint8_t *)src + 11 * 16);
563 	rte_mov16((uint8_t *)dst + 12 * 16, (const uint8_t *)src + 12 * 16);
564 	rte_mov16((uint8_t *)dst + 13 * 16, (const uint8_t *)src + 13 * 16);
565 	rte_mov16((uint8_t *)dst + 14 * 16, (const uint8_t *)src + 14 * 16);
566 	rte_mov16((uint8_t *)dst + 15 * 16, (const uint8_t *)src + 15 * 16);
567 }
568 
569 /**
570  * Macro for copying unaligned block from one location to another with constant load offset,
571  * 47 bytes leftover maximum,
572  * locations should not overlap.
573  * Requirements:
574  * - Store is aligned
575  * - Load offset is <offset>, which must be immediate value within [1, 15]
576  * - For <src>, make sure <offset> bit backwards & <16 - offset> bit forwards are available for loading
577  * - <dst>, <src>, <len> must be variables
578  * - __m128i <xmm0> ~ <xmm8> must be pre-defined
579  */
580 #define MOVEUNALIGNED_LEFT47_IMM(dst, src, len, offset)                                                     \
581 __extension__ ({                                                                                            \
582     size_t tmp;                                                                                                \
583     while (len >= 128 + 16 - offset) {                                                                      \
584         xmm0 = _mm_loadu_si128((const __m128i *)((const uint8_t *)src - offset + 0 * 16));                  \
585         len -= 128;                                                                                         \
586         xmm1 = _mm_loadu_si128((const __m128i *)((const uint8_t *)src - offset + 1 * 16));                  \
587         xmm2 = _mm_loadu_si128((const __m128i *)((const uint8_t *)src - offset + 2 * 16));                  \
588         xmm3 = _mm_loadu_si128((const __m128i *)((const uint8_t *)src - offset + 3 * 16));                  \
589         xmm4 = _mm_loadu_si128((const __m128i *)((const uint8_t *)src - offset + 4 * 16));                  \
590         xmm5 = _mm_loadu_si128((const __m128i *)((const uint8_t *)src - offset + 5 * 16));                  \
591         xmm6 = _mm_loadu_si128((const __m128i *)((const uint8_t *)src - offset + 6 * 16));                  \
592         xmm7 = _mm_loadu_si128((const __m128i *)((const uint8_t *)src - offset + 7 * 16));                  \
593         xmm8 = _mm_loadu_si128((const __m128i *)((const uint8_t *)src - offset + 8 * 16));                  \
594         src = (const uint8_t *)src + 128;                                                                   \
595         _mm_storeu_si128((__m128i *)((uint8_t *)dst + 0 * 16), _mm_alignr_epi8(xmm1, xmm0, offset));        \
596         _mm_storeu_si128((__m128i *)((uint8_t *)dst + 1 * 16), _mm_alignr_epi8(xmm2, xmm1, offset));        \
597         _mm_storeu_si128((__m128i *)((uint8_t *)dst + 2 * 16), _mm_alignr_epi8(xmm3, xmm2, offset));        \
598         _mm_storeu_si128((__m128i *)((uint8_t *)dst + 3 * 16), _mm_alignr_epi8(xmm4, xmm3, offset));        \
599         _mm_storeu_si128((__m128i *)((uint8_t *)dst + 4 * 16), _mm_alignr_epi8(xmm5, xmm4, offset));        \
600         _mm_storeu_si128((__m128i *)((uint8_t *)dst + 5 * 16), _mm_alignr_epi8(xmm6, xmm5, offset));        \
601         _mm_storeu_si128((__m128i *)((uint8_t *)dst + 6 * 16), _mm_alignr_epi8(xmm7, xmm6, offset));        \
602         _mm_storeu_si128((__m128i *)((uint8_t *)dst + 7 * 16), _mm_alignr_epi8(xmm8, xmm7, offset));        \
603         dst = (uint8_t *)dst + 128;                                                                         \
604     }                                                                                                       \
605     tmp = len;                                                                                              \
606     len = ((len - 16 + offset) & 127) + 16 - offset;                                                        \
607     tmp -= len;                                                                                             \
608     src = (const uint8_t *)src + tmp;                                                                       \
609     dst = (uint8_t *)dst + tmp;                                                                             \
610     if (len >= 32 + 16 - offset) {                                                                          \
611         while (len >= 32 + 16 - offset) {                                                                   \
612             xmm0 = _mm_loadu_si128((const __m128i *)((const uint8_t *)src - offset + 0 * 16));              \
613             len -= 32;                                                                                      \
614             xmm1 = _mm_loadu_si128((const __m128i *)((const uint8_t *)src - offset + 1 * 16));              \
615             xmm2 = _mm_loadu_si128((const __m128i *)((const uint8_t *)src - offset + 2 * 16));              \
616             src = (const uint8_t *)src + 32;                                                                \
617             _mm_storeu_si128((__m128i *)((uint8_t *)dst + 0 * 16), _mm_alignr_epi8(xmm1, xmm0, offset));    \
618             _mm_storeu_si128((__m128i *)((uint8_t *)dst + 1 * 16), _mm_alignr_epi8(xmm2, xmm1, offset));    \
619             dst = (uint8_t *)dst + 32;                                                                      \
620         }                                                                                                   \
621         tmp = len;                                                                                          \
622         len = ((len - 16 + offset) & 31) + 16 - offset;                                                     \
623         tmp -= len;                                                                                         \
624         src = (const uint8_t *)src + tmp;                                                                   \
625         dst = (uint8_t *)dst + tmp;                                                                         \
626     }                                                                                                       \
627 })
628 
629 /**
630  * Macro for copying unaligned block from one location to another,
631  * 47 bytes leftover maximum,
632  * locations should not overlap.
633  * Use switch here because the aligning instruction requires immediate value for shift count.
634  * Requirements:
635  * - Store is aligned
636  * - Load offset is <offset>, which must be within [1, 15]
637  * - For <src>, make sure <offset> bit backwards & <16 - offset> bit forwards are available for loading
638  * - <dst>, <src>, <len> must be variables
639  * - __m128i <xmm0> ~ <xmm8> used in MOVEUNALIGNED_LEFT47_IMM must be pre-defined
640  */
641 #define MOVEUNALIGNED_LEFT47(dst, src, len, offset)                   \
642 __extension__ ({                                                      \
643     switch (offset) {                                                 \
644     case 0x01: MOVEUNALIGNED_LEFT47_IMM(dst, src, n, 0x01); break;    \
645     case 0x02: MOVEUNALIGNED_LEFT47_IMM(dst, src, n, 0x02); break;    \
646     case 0x03: MOVEUNALIGNED_LEFT47_IMM(dst, src, n, 0x03); break;    \
647     case 0x04: MOVEUNALIGNED_LEFT47_IMM(dst, src, n, 0x04); break;    \
648     case 0x05: MOVEUNALIGNED_LEFT47_IMM(dst, src, n, 0x05); break;    \
649     case 0x06: MOVEUNALIGNED_LEFT47_IMM(dst, src, n, 0x06); break;    \
650     case 0x07: MOVEUNALIGNED_LEFT47_IMM(dst, src, n, 0x07); break;    \
651     case 0x08: MOVEUNALIGNED_LEFT47_IMM(dst, src, n, 0x08); break;    \
652     case 0x09: MOVEUNALIGNED_LEFT47_IMM(dst, src, n, 0x09); break;    \
653     case 0x0A: MOVEUNALIGNED_LEFT47_IMM(dst, src, n, 0x0A); break;    \
654     case 0x0B: MOVEUNALIGNED_LEFT47_IMM(dst, src, n, 0x0B); break;    \
655     case 0x0C: MOVEUNALIGNED_LEFT47_IMM(dst, src, n, 0x0C); break;    \
656     case 0x0D: MOVEUNALIGNED_LEFT47_IMM(dst, src, n, 0x0D); break;    \
657     case 0x0E: MOVEUNALIGNED_LEFT47_IMM(dst, src, n, 0x0E); break;    \
658     case 0x0F: MOVEUNALIGNED_LEFT47_IMM(dst, src, n, 0x0F); break;    \
659     default:;                                                         \
660     }                                                                 \
661 })
662 
663 static __rte_always_inline void *
664 rte_memcpy_generic(void *dst, const void *src, size_t n)
665 {
666 	__m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7, xmm8;
667 	uintptr_t dstu = (uintptr_t)dst;
668 	uintptr_t srcu = (uintptr_t)src;
669 	void *ret = dst;
670 	size_t dstofss;
671 	size_t srcofs;
672 
673 	/**
674 	 * Copy less than 16 bytes
675 	 */
676 	if (n < 16) {
677 		if (n & 0x01) {
678 			*(uint8_t *)dstu = *(const uint8_t *)srcu;
679 			srcu = (uintptr_t)((const uint8_t *)srcu + 1);
680 			dstu = (uintptr_t)((uint8_t *)dstu + 1);
681 		}
682 		if (n & 0x02) {
683 			*(uint16_t *)dstu = *(const uint16_t *)srcu;
684 			srcu = (uintptr_t)((const uint16_t *)srcu + 1);
685 			dstu = (uintptr_t)((uint16_t *)dstu + 1);
686 		}
687 		if (n & 0x04) {
688 			*(uint32_t *)dstu = *(const uint32_t *)srcu;
689 			srcu = (uintptr_t)((const uint32_t *)srcu + 1);
690 			dstu = (uintptr_t)((uint32_t *)dstu + 1);
691 		}
692 		if (n & 0x08) {
693 			*(uint64_t *)dstu = *(const uint64_t *)srcu;
694 		}
695 		return ret;
696 	}
697 
698 	/**
699 	 * Fast way when copy size doesn't exceed 512 bytes
700 	 */
701 	if (n <= 32) {
702 		rte_mov16((uint8_t *)dst, (const uint8_t *)src);
703 		rte_mov16((uint8_t *)dst - 16 + n, (const uint8_t *)src - 16 + n);
704 		return ret;
705 	}
706 	if (n <= 48) {
707 		rte_mov32((uint8_t *)dst, (const uint8_t *)src);
708 		rte_mov16((uint8_t *)dst - 16 + n, (const uint8_t *)src - 16 + n);
709 		return ret;
710 	}
711 	if (n <= 64) {
712 		rte_mov32((uint8_t *)dst, (const uint8_t *)src);
713 		rte_mov16((uint8_t *)dst + 32, (const uint8_t *)src + 32);
714 		rte_mov16((uint8_t *)dst - 16 + n, (const uint8_t *)src - 16 + n);
715 		return ret;
716 	}
717 	if (n <= 128) {
718 		goto COPY_BLOCK_128_BACK15;
719 	}
720 	if (n <= 512) {
721 		if (n >= 256) {
722 			n -= 256;
723 			rte_mov128((uint8_t *)dst, (const uint8_t *)src);
724 			rte_mov128((uint8_t *)dst + 128, (const uint8_t *)src + 128);
725 			src = (const uint8_t *)src + 256;
726 			dst = (uint8_t *)dst + 256;
727 		}
728 COPY_BLOCK_255_BACK15:
729 		if (n >= 128) {
730 			n -= 128;
731 			rte_mov128((uint8_t *)dst, (const uint8_t *)src);
732 			src = (const uint8_t *)src + 128;
733 			dst = (uint8_t *)dst + 128;
734 		}
735 COPY_BLOCK_128_BACK15:
736 		if (n >= 64) {
737 			n -= 64;
738 			rte_mov64((uint8_t *)dst, (const uint8_t *)src);
739 			src = (const uint8_t *)src + 64;
740 			dst = (uint8_t *)dst + 64;
741 		}
742 COPY_BLOCK_64_BACK15:
743 		if (n >= 32) {
744 			n -= 32;
745 			rte_mov32((uint8_t *)dst, (const uint8_t *)src);
746 			src = (const uint8_t *)src + 32;
747 			dst = (uint8_t *)dst + 32;
748 		}
749 		if (n > 16) {
750 			rte_mov16((uint8_t *)dst, (const uint8_t *)src);
751 			rte_mov16((uint8_t *)dst - 16 + n, (const uint8_t *)src - 16 + n);
752 			return ret;
753 		}
754 		if (n > 0) {
755 			rte_mov16((uint8_t *)dst - 16 + n, (const uint8_t *)src - 16 + n);
756 		}
757 		return ret;
758 	}
759 
760 	/**
761 	 * Make store aligned when copy size exceeds 512 bytes,
762 	 * and make sure the first 15 bytes are copied, because
763 	 * unaligned copy functions require up to 15 bytes
764 	 * backwards access.
765 	 */
766 	dstofss = (uintptr_t)dst & 0x0F;
767 	if (dstofss > 0) {
768 		dstofss = 16 - dstofss + 16;
769 		n -= dstofss;
770 		rte_mov32((uint8_t *)dst, (const uint8_t *)src);
771 		src = (const uint8_t *)src + dstofss;
772 		dst = (uint8_t *)dst + dstofss;
773 	}
774 	srcofs = ((uintptr_t)src & 0x0F);
775 
776 	/**
777 	 * For aligned copy
778 	 */
779 	if (srcofs == 0) {
780 		/**
781 		 * Copy 256-byte blocks
782 		 */
783 		for (; n >= 256; n -= 256) {
784 			rte_mov256((uint8_t *)dst, (const uint8_t *)src);
785 			dst = (uint8_t *)dst + 256;
786 			src = (const uint8_t *)src + 256;
787 		}
788 
789 		/**
790 		 * Copy whatever left
791 		 */
792 		goto COPY_BLOCK_255_BACK15;
793 	}
794 
795 	/**
796 	 * For copy with unaligned load
797 	 */
798 	MOVEUNALIGNED_LEFT47(dst, src, n, srcofs);
799 
800 	/**
801 	 * Copy whatever left
802 	 */
803 	goto COPY_BLOCK_64_BACK15;
804 }
805 
806 #endif /* __AVX512F__ */
807 
808 static __rte_always_inline void *
809 rte_memcpy_aligned(void *dst, const void *src, size_t n)
810 {
811 	void *ret = dst;
812 
813 	/* Copy size <= 16 bytes */
814 	if (n < 16) {
815 		if (n & 0x01) {
816 			*(uint8_t *)dst = *(const uint8_t *)src;
817 			src = (const uint8_t *)src + 1;
818 			dst = (uint8_t *)dst + 1;
819 		}
820 		if (n & 0x02) {
821 			*(uint16_t *)dst = *(const uint16_t *)src;
822 			src = (const uint16_t *)src + 1;
823 			dst = (uint16_t *)dst + 1;
824 		}
825 		if (n & 0x04) {
826 			*(uint32_t *)dst = *(const uint32_t *)src;
827 			src = (const uint32_t *)src + 1;
828 			dst = (uint32_t *)dst + 1;
829 		}
830 		if (n & 0x08)
831 			*(uint64_t *)dst = *(const uint64_t *)src;
832 
833 		return ret;
834 	}
835 
836 	/* Copy 16 <= size <= 32 bytes */
837 	if (n <= 32) {
838 		rte_mov16((uint8_t *)dst, (const uint8_t *)src);
839 		rte_mov16((uint8_t *)dst - 16 + n,
840 				(const uint8_t *)src - 16 + n);
841 
842 		return ret;
843 	}
844 
845 	/* Copy 32 < size <= 64 bytes */
846 	if (n <= 64) {
847 		rte_mov32((uint8_t *)dst, (const uint8_t *)src);
848 		rte_mov32((uint8_t *)dst - 32 + n,
849 				(const uint8_t *)src - 32 + n);
850 
851 		return ret;
852 	}
853 
854 	/* Copy 64 bytes blocks */
855 	for (; n >= 64; n -= 64) {
856 		rte_mov64((uint8_t *)dst, (const uint8_t *)src);
857 		dst = (uint8_t *)dst + 64;
858 		src = (const uint8_t *)src + 64;
859 	}
860 
861 	/* Copy whatever left */
862 	rte_mov64((uint8_t *)dst - 64 + n,
863 			(const uint8_t *)src - 64 + n);
864 
865 	return ret;
866 }
867 
868 static __rte_always_inline void *
869 rte_memcpy(void *dst, const void *src, size_t n)
870 {
871 	if (!(((uintptr_t)dst | (uintptr_t)src) & ALIGNMENT_MASK))
872 		return rte_memcpy_aligned(dst, src, n);
873 	else
874 		return rte_memcpy_generic(dst, src, n);
875 }
876 
877 #if defined(RTE_TOOLCHAIN_GCC) && (GCC_VERSION >= 100000)
878 #pragma GCC diagnostic pop
879 #endif
880 
881 #ifdef __cplusplus
882 }
883 #endif
884 
885 #endif /* _RTE_MEMCPY_X86_64_H_ */
886