1/*
2 * Copyright (c) 2016 Cisco and/or its affiliates.
3 * Licensed under the Apache License, Version 2.0 (the "License");
4 * you may not use this file except in compliance with the License.
5 * You may obtain a copy of the License at:
6 *
7 *     http://www.apache.org/licenses/LICENSE-2.0
8 *
9 * Unless required by applicable law or agreed to in writing, software
10 * distributed under the License is distributed on an "AS IS" BASIS,
11 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 * See the License for the specific language governing permissions and
13 * limitations under the License.
14 */
15/*-
16 *   BSD LICENSE
17 *
18 *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
19 *   All rights reserved.
20 *
21 *   Redistribution and use in source and binary forms, with or without
22 *   modification, are permitted provided that the following conditions
23 *   are met:
24 *
25 *     * Redistributions of source code must retain the above copyright
26 *       notice, this list of conditions and the following disclaimer.
27 *     * Redistributions in binary form must reproduce the above copyright
28 *       notice, this list of conditions and the following disclaimer in
29 *       the documentation and/or other materials provided with the
30 *       distribution.
31 *     * Neither the name of Intel Corporation nor the names of its
32 *       contributors may be used to endorse or promote products derived
33 *       from this software without specific prior written permission.
34 *
35 *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
36 *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
37 *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
38 *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
39 *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
40 *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
41 *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
42 *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
43 *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
44 *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
45 *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
46 */
47
48#ifndef included_clib_memcpy_sse3_h
49#define included_clib_memcpy_sse3_h
50
51#include <stdint.h>
52#include <x86intrin.h>
53#include <vppinfra/warnings.h>
54
55/* *INDENT-OFF* */
56WARN_OFF (stringop-overflow)
57/* *INDENT-ON* */
58
59static inline void
60clib_mov16 (u8 * dst, const u8 * src)
61{
62  __m128i xmm0;
63
64  xmm0 = _mm_loadu_si128 ((const __m128i *) src);
65  _mm_storeu_si128 ((__m128i *) dst, xmm0);
66}
67
68static inline void
69clib_mov32 (u8 * dst, const u8 * src)
70{
71  clib_mov16 ((u8 *) dst + 0 * 16, (const u8 *) src + 0 * 16);
72  clib_mov16 ((u8 *) dst + 1 * 16, (const u8 *) src + 1 * 16);
73}
74
75static inline void
76clib_mov64 (u8 * dst, const u8 * src)
77{
78  clib_mov32 ((u8 *) dst + 0 * 32, (const u8 *) src + 0 * 32);
79  clib_mov32 ((u8 *) dst + 1 * 32, (const u8 *) src + 1 * 32);
80}
81
82static inline void
83clib_mov128 (u8 * dst, const u8 * src)
84{
85  clib_mov64 ((u8 *) dst + 0 * 64, (const u8 *) src + 0 * 64);
86  clib_mov64 ((u8 *) dst + 1 * 64, (const u8 *) src + 1 * 64);
87}
88
89static inline void
90clib_mov256 (u8 * dst, const u8 * src)
91{
92  clib_mov128 ((u8 *) dst + 0 * 128, (const u8 *) src + 0 * 128);
93  clib_mov128 ((u8 *) dst + 1 * 128, (const u8 *) src + 1 * 128);
94}
95
96/**
97 * Macro for copying unaligned block from one location to another with constant load offset,
98 * 47 bytes leftover maximum,
99 * locations should not overlap.
100 * Requirements:
101 * - Store is aligned
102 * - Load offset is <offset>, which must be immediate value within [1, 15]
103 * - For <src>, make sure <offset> bit backwards & <16 - offset> bit forwards are available for loading
104 * - <dst>, <src>, <len> must be variables
105 * - __m128i <xmm0> ~ <xmm8> must be pre-defined
106 */
107#define CLIB_MVUNALIGN_LEFT47_IMM(dst, src, len, offset)                                                    \
108({                                                                                                          \
109    int tmp;                                                                                                \
110    while (len >= 128 + 16 - offset) {                                                                      \
111        xmm0 = _mm_loadu_si128((const __m128i *)((const u8 *)src - offset + 0 * 16));                       \
112        len -= 128;                                                                                         \
113        xmm1 = _mm_loadu_si128((const __m128i *)((const u8 *)src - offset + 1 * 16));                       \
114        xmm2 = _mm_loadu_si128((const __m128i *)((const u8 *)src - offset + 2 * 16));                       \
115        xmm3 = _mm_loadu_si128((const __m128i *)((const u8 *)src - offset + 3 * 16));                       \
116        xmm4 = _mm_loadu_si128((const __m128i *)((const u8 *)src - offset + 4 * 16));                       \
117        xmm5 = _mm_loadu_si128((const __m128i *)((const u8 *)src - offset + 5 * 16));                       \
118        xmm6 = _mm_loadu_si128((const __m128i *)((const u8 *)src - offset + 6 * 16));                       \
119        xmm7 = _mm_loadu_si128((const __m128i *)((const u8 *)src - offset + 7 * 16));                       \
120        xmm8 = _mm_loadu_si128((const __m128i *)((const u8 *)src - offset + 8 * 16));                       \
121        src = (const u8 *)src + 128;                                                                        \
122        _mm_storeu_si128((__m128i *)((u8 *)dst + 0 * 16), _mm_alignr_epi8(xmm1, xmm0, offset));             \
123        _mm_storeu_si128((__m128i *)((u8 *)dst + 1 * 16), _mm_alignr_epi8(xmm2, xmm1, offset));             \
124        _mm_storeu_si128((__m128i *)((u8 *)dst + 2 * 16), _mm_alignr_epi8(xmm3, xmm2, offset));             \
125        _mm_storeu_si128((__m128i *)((u8 *)dst + 3 * 16), _mm_alignr_epi8(xmm4, xmm3, offset));             \
126        _mm_storeu_si128((__m128i *)((u8 *)dst + 4 * 16), _mm_alignr_epi8(xmm5, xmm4, offset));             \
127        _mm_storeu_si128((__m128i *)((u8 *)dst + 5 * 16), _mm_alignr_epi8(xmm6, xmm5, offset));             \
128        _mm_storeu_si128((__m128i *)((u8 *)dst + 6 * 16), _mm_alignr_epi8(xmm7, xmm6, offset));             \
129        _mm_storeu_si128((__m128i *)((u8 *)dst + 7 * 16), _mm_alignr_epi8(xmm8, xmm7, offset));             \
130        dst = (u8 *)dst + 128;                                                                              \
131    }                                                                                                       \
132    tmp = len;                                                                                              \
133    len = ((len - 16 + offset) & 127) + 16 - offset;                                                        \
134    tmp -= len;                                                                                             \
135    src = (const u8 *)src + tmp;                                                                            \
136    dst = (u8 *)dst + tmp;                                                                                  \
137    if (len >= 32 + 16 - offset) {                                                                          \
138        while (len >= 32 + 16 - offset) {                                                                   \
139            xmm0 = _mm_loadu_si128((const __m128i *)((const u8 *)src - offset + 0 * 16));                   \
140            len -= 32;                                                                                      \
141            xmm1 = _mm_loadu_si128((const __m128i *)((const u8 *)src - offset + 1 * 16));                   \
142            xmm2 = _mm_loadu_si128((const __m128i *)((const u8 *)src - offset + 2 * 16));                   \
143            src = (const u8 *)src + 32;                                                                     \
144            _mm_storeu_si128((__m128i *)((u8 *)dst + 0 * 16), _mm_alignr_epi8(xmm1, xmm0, offset));         \
145            _mm_storeu_si128((__m128i *)((u8 *)dst + 1 * 16), _mm_alignr_epi8(xmm2, xmm1, offset));         \
146            dst = (u8 *)dst + 32;                                                                           \
147        }                                                                                                   \
148        tmp = len;                                                                                          \
149        len = ((len - 16 + offset) & 31) + 16 - offset;                                                     \
150        tmp -= len;                                                                                         \
151        src = (const u8 *)src + tmp;                                                                        \
152        dst = (u8 *)dst + tmp;                                                                              \
153    }                                                                                                       \
154})
155
156/**
157 * Macro for copying unaligned block from one location to another,
158 * 47 bytes leftover maximum,
159 * locations should not overlap.
160 * Use switch here because the aligning instruction requires immediate value for shift count.
161 * Requirements:
162 * - Store is aligned
163 * - Load offset is <offset>, which must be within [1, 15]
164 * - For <src>, make sure <offset> bit backwards & <16 - offset> bit forwards are available for loading
165 * - <dst>, <src>, <len> must be variables
166 * - __m128i <xmm0> ~ <xmm8> used in CLIB_MVUNALIGN_LEFT47_IMM must be pre-defined
167 */
168#define CLIB_MVUNALIGN_LEFT47(dst, src, len, offset)                  \
169({                                                                    \
170    switch (offset) {                                                 \
171    case 0x01: CLIB_MVUNALIGN_LEFT47_IMM(dst, src, n, 0x01); break;   \
172    case 0x02: CLIB_MVUNALIGN_LEFT47_IMM(dst, src, n, 0x02); break;   \
173    case 0x03: CLIB_MVUNALIGN_LEFT47_IMM(dst, src, n, 0x03); break;   \
174    case 0x04: CLIB_MVUNALIGN_LEFT47_IMM(dst, src, n, 0x04); break;   \
175    case 0x05: CLIB_MVUNALIGN_LEFT47_IMM(dst, src, n, 0x05); break;   \
176    case 0x06: CLIB_MVUNALIGN_LEFT47_IMM(dst, src, n, 0x06); break;   \
177    case 0x07: CLIB_MVUNALIGN_LEFT47_IMM(dst, src, n, 0x07); break;   \
178    case 0x08: CLIB_MVUNALIGN_LEFT47_IMM(dst, src, n, 0x08); break;   \
179    case 0x09: CLIB_MVUNALIGN_LEFT47_IMM(dst, src, n, 0x09); break;   \
180    case 0x0A: CLIB_MVUNALIGN_LEFT47_IMM(dst, src, n, 0x0A); break;   \
181    case 0x0B: CLIB_MVUNALIGN_LEFT47_IMM(dst, src, n, 0x0B); break;   \
182    case 0x0C: CLIB_MVUNALIGN_LEFT47_IMM(dst, src, n, 0x0C); break;   \
183    case 0x0D: CLIB_MVUNALIGN_LEFT47_IMM(dst, src, n, 0x0D); break;   \
184    case 0x0E: CLIB_MVUNALIGN_LEFT47_IMM(dst, src, n, 0x0E); break;   \
185    case 0x0F: CLIB_MVUNALIGN_LEFT47_IMM(dst, src, n, 0x0F); break;   \
186    default:;                                                         \
187    }                                                                 \
188})
189
190static inline void *
191clib_memcpy_fast (void *dst, const void *src, size_t n)
192{
193  __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7, xmm8;
194  uword dstu = (uword) dst;
195  uword srcu = (uword) src;
196  void *ret = dst;
197  size_t dstofss;
198  size_t srcofs;
199
200	/**
201	 * Copy less than 16 bytes
202	 */
203  if (n < 16)
204    {
205      if (n & 0x01)
206	{
207	  *(u8 *) dstu = *(const u8 *) srcu;
208	  srcu = (uword) ((const u8 *) srcu + 1);
209	  dstu = (uword) ((u8 *) dstu + 1);
210	}
211      if (n & 0x02)
212	{
213	  *(u16 *) dstu = *(const u16 *) srcu;
214	  srcu = (uword) ((const u16 *) srcu + 1);
215	  dstu = (uword) ((u16 *) dstu + 1);
216	}
217      if (n & 0x04)
218	{
219	  *(u32 *) dstu = *(const u32 *) srcu;
220	  srcu = (uword) ((const u32 *) srcu + 1);
221	  dstu = (uword) ((u32 *) dstu + 1);
222	}
223      if (n & 0x08)
224	{
225	  *(u64 *) dstu = *(const u64 *) srcu;
226	}
227      return ret;
228    }
229
230  /**
231   * Fast way when copy size doesn't exceed 512 bytes
232   */
233  if (n <= 32)
234    {
235      clib_mov16 ((u8 *) dst, (const u8 *) src);
236      clib_mov16 ((u8 *) dst - 16 + n, (const u8 *) src - 16 + n);
237      return ret;
238    }
239  if (n <= 48)
240    {
241      clib_mov32 ((u8 *) dst, (const u8 *) src);
242      clib_mov16 ((u8 *) dst - 16 + n, (const u8 *) src - 16 + n);
243      return ret;
244    }
245  if (n <= 64)
246    {
247      clib_mov32 ((u8 *) dst, (const u8 *) src);
248      clib_mov16 ((u8 *) dst + 32, (const u8 *) src + 32);
249      clib_mov16 ((u8 *) dst - 16 + n, (const u8 *) src - 16 + n);
250      return ret;
251    }
252  if (n <= 128)
253    {
254      goto COPY_BLOCK_128_BACK15;
255    }
256  if (n <= 512)
257    {
258      if (n >= 256)
259	{
260	  n -= 256;
261	  clib_mov128 ((u8 *) dst, (const u8 *) src);
262	  clib_mov128 ((u8 *) dst + 128, (const u8 *) src + 128);
263	  src = (const u8 *) src + 256;
264	  dst = (u8 *) dst + 256;
265	}
266    COPY_BLOCK_255_BACK15:
267      if (n >= 128)
268	{
269	  n -= 128;
270	  clib_mov128 ((u8 *) dst, (const u8 *) src);
271	  src = (const u8 *) src + 128;
272	  dst = (u8 *) dst + 128;
273	}
274    COPY_BLOCK_128_BACK15:
275      if (n >= 64)
276	{
277	  n -= 64;
278	  clib_mov64 ((u8 *) dst, (const u8 *) src);
279	  src = (const u8 *) src + 64;
280	  dst = (u8 *) dst + 64;
281	}
282    COPY_BLOCK_64_BACK15:
283      if (n >= 32)
284	{
285	  n -= 32;
286	  clib_mov32 ((u8 *) dst, (const u8 *) src);
287	  src = (const u8 *) src + 32;
288	  dst = (u8 *) dst + 32;
289	}
290      if (n > 16)
291	{
292	  clib_mov16 ((u8 *) dst, (const u8 *) src);
293	  clib_mov16 ((u8 *) dst - 16 + n, (const u8 *) src - 16 + n);
294	  return ret;
295	}
296      if (n > 0)
297	{
298	  clib_mov16 ((u8 *) dst - 16 + n, (const u8 *) src - 16 + n);
299	}
300      return ret;
301    }
302
303  /**
304   * Make store aligned when copy size exceeds 512 bytes,
305   * and make sure the first 15 bytes are copied, because
306   * unaligned copy functions require up to 15 bytes
307   * backwards access.
308   */
309  dstofss = (uword) dst & 0x0F;
310  if (dstofss > 0)
311    {
312      dstofss = 16 - dstofss + 16;
313      n -= dstofss;
314      clib_mov32 ((u8 *) dst, (const u8 *) src);
315      src = (const u8 *) src + dstofss;
316      dst = (u8 *) dst + dstofss;
317    }
318  srcofs = ((uword) src & 0x0F);
319
320  /**
321   * For aligned copy
322   */
323  if (srcofs == 0)
324    {
325      /**
326       * Copy 256-byte blocks
327       */
328      for (; n >= 256; n -= 256)
329	{
330	  clib_mov256 ((u8 *) dst, (const u8 *) src);
331	  dst = (u8 *) dst + 256;
332	  src = (const u8 *) src + 256;
333	}
334
335      /**
336       * Copy whatever left
337       */
338      goto COPY_BLOCK_255_BACK15;
339    }
340
341  /**
342   * For copy with unaligned load
343   */
344  CLIB_MVUNALIGN_LEFT47 (dst, src, n, srcofs);
345
346  /**
347   * Copy whatever left
348   */
349  goto COPY_BLOCK_64_BACK15;
350}
351
352/* *INDENT-OFF* */
353WARN_ON (stringop-overflow)
354/* *INDENT-ON* */
355
356#undef CLIB_MVUNALIGN_LEFT47_IMM
357#undef CLIB_MVUNALIGN_LEFT47
358
359#endif /* included_clib_memcpy_sse3_h */
360
361
362/*
363 * fd.io coding-style-patch-verification: ON
364 *
365 * Local Variables:
366 * eval: (c-set-style "gnu")
367 * End:
368 */
369