xref: /dpdk/lib/eal/arm/include/rte_memcpy_32.h (revision 719834a6849e1daf4a70ff7742bbcc3ae7e25607)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2015 RehiveTech. All rights reserved.
3  */
4 
5 #ifndef _RTE_MEMCPY_ARM32_H_
6 #define _RTE_MEMCPY_ARM32_H_
7 
8 #include <stdint.h>
9 #include <string.h>
10 
11 #include "generic/rte_memcpy.h"
12 
13 #ifdef RTE_ARCH_ARM_NEON_MEMCPY
14 
15 #ifndef __ARM_NEON
16 #error "Cannot optimize memcpy by NEON as the CPU seems to not support this"
17 #endif
18 
19 /* ARM NEON Intrinsics are used to copy data */
20 #include <arm_neon.h>
21 
22 #ifdef __cplusplus
23 extern "C" {
24 #endif
25 
26 static inline void
27 rte_mov16(uint8_t *dst, const uint8_t *src)
28 {
29 	vst1q_u8(dst, vld1q_u8(src));
30 }
31 
32 static inline void
33 rte_mov32(uint8_t *dst, const uint8_t *src)
34 {
35 	asm volatile (
36 		"vld1.8 {d0-d3}, [%0]\n\t"
37 		"vst1.8 {d0-d3}, [%1]\n\t"
38 		: "+r" (src), "+r" (dst)
39 		: : "memory", "d0", "d1", "d2", "d3");
40 }
41 
42 static inline void
43 rte_mov48(uint8_t *dst, const uint8_t *src)
44 {
45 	asm volatile (
46 		"vld1.8 {d0-d3}, [%0]!\n\t"
47 		"vld1.8 {d4-d5}, [%0]\n\t"
48 		"vst1.8 {d0-d3}, [%1]!\n\t"
49 		"vst1.8 {d4-d5}, [%1]\n\t"
50 		: "+r" (src), "+r" (dst)
51 		:
52 		: "memory", "d0", "d1", "d2", "d3", "d4", "d5");
53 }
54 
55 static inline void
56 rte_mov64(uint8_t *dst, const uint8_t *src)
57 {
58 	asm volatile (
59 		"vld1.8 {d0-d3}, [%0]!\n\t"
60 		"vld1.8 {d4-d7}, [%0]\n\t"
61 		"vst1.8 {d0-d3}, [%1]!\n\t"
62 		"vst1.8 {d4-d7}, [%1]\n\t"
63 		: "+r" (src), "+r" (dst)
64 		:
65 		: "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7");
66 }
67 
68 static inline void
69 rte_mov128(uint8_t *dst, const uint8_t *src)
70 {
71 	asm volatile ("pld [%0, #64]" : : "r" (src));
72 	asm volatile (
73 		"vld1.8 {d0-d3},   [%0]!\n\t"
74 		"vld1.8 {d4-d7},   [%0]!\n\t"
75 		"vld1.8 {d8-d11},  [%0]!\n\t"
76 		"vld1.8 {d12-d15}, [%0]\n\t"
77 		"vst1.8 {d0-d3},   [%1]!\n\t"
78 		"vst1.8 {d4-d7},   [%1]!\n\t"
79 		"vst1.8 {d8-d11},  [%1]!\n\t"
80 		"vst1.8 {d12-d15}, [%1]\n\t"
81 		: "+r" (src), "+r" (dst)
82 		:
83 		: "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7",
84 		"d8", "d9", "d10", "d11", "d12", "d13", "d14", "d15");
85 }
86 
87 static inline void
88 rte_mov256(uint8_t *dst, const uint8_t *src)
89 {
90 	asm volatile ("pld [%0,  #64]" : : "r" (src));
91 	asm volatile ("pld [%0, #128]" : : "r" (src));
92 	asm volatile ("pld [%0, #192]" : : "r" (src));
93 	asm volatile ("pld [%0, #256]" : : "r" (src));
94 	asm volatile ("pld [%0, #320]" : : "r" (src));
95 	asm volatile ("pld [%0, #384]" : : "r" (src));
96 	asm volatile ("pld [%0, #448]" : : "r" (src));
97 	asm volatile (
98 		"vld1.8 {d0-d3},   [%0]!\n\t"
99 		"vld1.8 {d4-d7},   [%0]!\n\t"
100 		"vld1.8 {d8-d11},  [%0]!\n\t"
101 		"vld1.8 {d12-d15}, [%0]!\n\t"
102 		"vld1.8 {d16-d19}, [%0]!\n\t"
103 		"vld1.8 {d20-d23}, [%0]!\n\t"
104 		"vld1.8 {d24-d27}, [%0]!\n\t"
105 		"vld1.8 {d28-d31}, [%0]\n\t"
106 		"vst1.8 {d0-d3},   [%1]!\n\t"
107 		"vst1.8 {d4-d7},   [%1]!\n\t"
108 		"vst1.8 {d8-d11},  [%1]!\n\t"
109 		"vst1.8 {d12-d15}, [%1]!\n\t"
110 		"vst1.8 {d16-d19}, [%1]!\n\t"
111 		"vst1.8 {d20-d23}, [%1]!\n\t"
112 		"vst1.8 {d24-d27}, [%1]!\n\t"
113 		"vst1.8 {d28-d31}, [%1]!\n\t"
114 		: "+r" (src), "+r" (dst)
115 		:
116 		: "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7",
117 		"d8", "d9", "d10", "d11", "d12", "d13", "d14", "d15",
118 		"d16", "d17", "d18", "d19", "d20", "d21", "d22", "d23",
119 		"d24", "d25", "d26", "d27", "d28", "d29", "d30", "d31");
120 }
121 
122 #define rte_memcpy(dst, src, n)              \
123 	__extension__ ({                     \
124 	(__builtin_constant_p(n)) ?          \
125 	memcpy((dst), (src), (n)) :          \
126 	rte_memcpy_func((dst), (src), (n)); })
127 
128 static inline void *
129 rte_memcpy_func(void *dst, const void *src, size_t n)
130 {
131 	void *ret = dst;
132 
133 	/* We can't copy < 16 bytes using XMM registers so do it manually. */
134 	if (n < 16) {
135 		if (n & 0x01) {
136 			*(uint8_t *)dst = *(const uint8_t *)src;
137 			dst = (uint8_t *)dst + 1;
138 			src = (const uint8_t *)src + 1;
139 		}
140 		if (n & 0x02) {
141 			*(uint16_t *)dst = *(const uint16_t *)src;
142 			dst = (uint16_t *)dst + 1;
143 			src = (const uint16_t *)src + 1;
144 		}
145 		if (n & 0x04) {
146 			*(uint32_t *)dst = *(const uint32_t *)src;
147 			dst = (uint32_t *)dst + 1;
148 			src = (const uint32_t *)src + 1;
149 		}
150 		if (n & 0x08) {
151 			/* ARMv7 can not handle unaligned access to long long
152 			 * (uint64_t). Therefore two uint32_t operations are
153 			 * used.
154 			 */
155 			*(uint32_t *)dst = *(const uint32_t *)src;
156 			dst = (uint32_t *)dst + 1;
157 			src = (const uint32_t *)src + 1;
158 			*(uint32_t *)dst = *(const uint32_t *)src;
159 		}
160 		return ret;
161 	}
162 
163 	/* Special fast cases for <= 128 bytes */
164 	if (n <= 32) {
165 		rte_mov16((uint8_t *)dst, (const uint8_t *)src);
166 		rte_mov16((uint8_t *)dst - 16 + n,
167 			(const uint8_t *)src - 16 + n);
168 		return ret;
169 	}
170 
171 	if (n <= 64) {
172 		rte_mov32((uint8_t *)dst, (const uint8_t *)src);
173 		rte_mov32((uint8_t *)dst - 32 + n,
174 			(const uint8_t *)src - 32 + n);
175 		return ret;
176 	}
177 
178 	if (n <= 128) {
179 		rte_mov64((uint8_t *)dst, (const uint8_t *)src);
180 		rte_mov64((uint8_t *)dst - 64 + n,
181 			(const uint8_t *)src - 64 + n);
182 		return ret;
183 	}
184 
185 	/*
186 	 * For large copies > 128 bytes. This combination of 256, 64 and 16 byte
187 	 * copies was found to be faster than doing 128 and 32 byte copies as
188 	 * well.
189 	 */
190 	for ( ; n >= 256; n -= 256) {
191 		rte_mov256((uint8_t *)dst, (const uint8_t *)src);
192 		dst = (uint8_t *)dst + 256;
193 		src = (const uint8_t *)src + 256;
194 	}
195 
196 	/*
197 	 * We split the remaining bytes (which will be less than 256) into
198 	 * 64byte (2^6) chunks.
199 	 * Using incrementing integers in the case labels of a switch statement
200 	 * encourages the compiler to use a jump table. To get incrementing
201 	 * integers, we shift the 2 relevant bits to the LSB position to first
202 	 * get decrementing integers, and then subtract.
203 	 */
204 	switch (3 - (n >> 6)) {
205 	case 0x00:
206 		rte_mov64((uint8_t *)dst, (const uint8_t *)src);
207 		n -= 64;
208 		dst = (uint8_t *)dst + 64;
209 		src = (const uint8_t *)src + 64;      /* fallthrough */
210 	case 0x01:
211 		rte_mov64((uint8_t *)dst, (const uint8_t *)src);
212 		n -= 64;
213 		dst = (uint8_t *)dst + 64;
214 		src = (const uint8_t *)src + 64;      /* fallthrough */
215 	case 0x02:
216 		rte_mov64((uint8_t *)dst, (const uint8_t *)src);
217 		n -= 64;
218 		dst = (uint8_t *)dst + 64;
219 		src = (const uint8_t *)src + 64;      /* fallthrough */
220 	default:
221 		break;
222 	}
223 
224 	/*
225 	 * We split the remaining bytes (which will be less than 64) into
226 	 * 16byte (2^4) chunks, using the same switch structure as above.
227 	 */
228 	switch (3 - (n >> 4)) {
229 	case 0x00:
230 		rte_mov16((uint8_t *)dst, (const uint8_t *)src);
231 		n -= 16;
232 		dst = (uint8_t *)dst + 16;
233 		src = (const uint8_t *)src + 16;      /* fallthrough */
234 	case 0x01:
235 		rte_mov16((uint8_t *)dst, (const uint8_t *)src);
236 		n -= 16;
237 		dst = (uint8_t *)dst + 16;
238 		src = (const uint8_t *)src + 16;      /* fallthrough */
239 	case 0x02:
240 		rte_mov16((uint8_t *)dst, (const uint8_t *)src);
241 		n -= 16;
242 		dst = (uint8_t *)dst + 16;
243 		src = (const uint8_t *)src + 16;      /* fallthrough */
244 	default:
245 		break;
246 	}
247 
248 	/* Copy any remaining bytes, without going beyond end of buffers */
249 	if (n != 0)
250 		rte_mov16((uint8_t *)dst - 16 + n,
251 			(const uint8_t *)src - 16 + n);
252 	return ret;
253 }
254 
255 #else
256 
257 static inline void
258 rte_mov16(uint8_t *dst, const uint8_t *src)
259 {
260 	memcpy(dst, src, 16);
261 }
262 
263 static inline void
264 rte_mov32(uint8_t *dst, const uint8_t *src)
265 {
266 	memcpy(dst, src, 32);
267 }
268 
269 static inline void
270 rte_mov48(uint8_t *dst, const uint8_t *src)
271 {
272 	memcpy(dst, src, 48);
273 }
274 
275 static inline void
276 rte_mov64(uint8_t *dst, const uint8_t *src)
277 {
278 	memcpy(dst, src, 64);
279 }
280 
281 static inline void
282 rte_mov128(uint8_t *dst, const uint8_t *src)
283 {
284 	memcpy(dst, src, 128);
285 }
286 
287 static inline void
288 rte_mov256(uint8_t *dst, const uint8_t *src)
289 {
290 	memcpy(dst, src, 256);
291 }
292 
293 static inline void *
294 rte_memcpy(void *dst, const void *src, size_t n)
295 {
296 	return memcpy(dst, src, n);
297 }
298 
299 #endif /* RTE_ARCH_ARM_NEON_MEMCPY */
300 
301 #ifdef __cplusplus
302 }
303 #endif
304 
305 #endif /* _RTE_MEMCPY_ARM32_H_ */
306