xref: /openbmc/qemu/util/bufferiszero.c (revision 4921d0a7)
1 /*
2  * Simple C functions to supplement the C library
3  *
4  * Copyright (c) 2006 Fabrice Bellard
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to deal
8  * in the Software without restriction, including without limitation the rights
9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10  * copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22  * THE SOFTWARE.
23  */
24 #include "qemu/osdep.h"
25 #include "qemu/cutils.h"
26 #include "qemu/bswap.h"
27 #include "host/cpuinfo.h"
28 
29 static bool
30 buffer_zero_int(const void *buf, size_t len)
31 {
32     if (unlikely(len < 8)) {
33         /* For a very small buffer, simply accumulate all the bytes.  */
34         const unsigned char *p = buf;
35         const unsigned char *e = buf + len;
36         unsigned char t = 0;
37 
38         do {
39             t |= *p++;
40         } while (p < e);
41 
42         return t == 0;
43     } else {
44         /* Otherwise, use the unaligned memory access functions to
45            handle the beginning and end of the buffer, with a couple
46            of loops handling the middle aligned section.  */
47         uint64_t t = ldq_he_p(buf);
48         const uint64_t *p = (uint64_t *)(((uintptr_t)buf + 8) & -8);
49         const uint64_t *e = (uint64_t *)(((uintptr_t)buf + len) & -8);
50 
51         for (; p + 8 <= e; p += 8) {
52             __builtin_prefetch(p + 8);
53             if (t) {
54                 return false;
55             }
56             t = p[0] | p[1] | p[2] | p[3] | p[4] | p[5] | p[6] | p[7];
57         }
58         while (p < e) {
59             t |= *p++;
60         }
61         t |= ldq_he_p(buf + len - 8);
62 
63         return t == 0;
64     }
65 }
66 
67 #if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT) || defined(__SSE2__)
68 #include <immintrin.h>
69 
70 /* Note that each of these vectorized functions require len >= 64.  */
71 
72 static bool __attribute__((target("sse2")))
73 buffer_zero_sse2(const void *buf, size_t len)
74 {
75     __m128i t = _mm_loadu_si128(buf);
76     __m128i *p = (__m128i *)(((uintptr_t)buf + 5 * 16) & -16);
77     __m128i *e = (__m128i *)(((uintptr_t)buf + len) & -16);
78     __m128i zero = _mm_setzero_si128();
79 
80     /* Loop over 16-byte aligned blocks of 64.  */
81     while (likely(p <= e)) {
82         __builtin_prefetch(p);
83         t = _mm_cmpeq_epi8(t, zero);
84         if (unlikely(_mm_movemask_epi8(t) != 0xFFFF)) {
85             return false;
86         }
87         t = p[-4] | p[-3] | p[-2] | p[-1];
88         p += 4;
89     }
90 
91     /* Finish the aligned tail.  */
92     t |= e[-3];
93     t |= e[-2];
94     t |= e[-1];
95 
96     /* Finish the unaligned tail.  */
97     t |= _mm_loadu_si128(buf + len - 16);
98 
99     return _mm_movemask_epi8(_mm_cmpeq_epi8(t, zero)) == 0xFFFF;
100 }
101 
102 #ifdef CONFIG_AVX2_OPT
103 static bool __attribute__((target("sse4")))
104 buffer_zero_sse4(const void *buf, size_t len)
105 {
106     __m128i t = _mm_loadu_si128(buf);
107     __m128i *p = (__m128i *)(((uintptr_t)buf + 5 * 16) & -16);
108     __m128i *e = (__m128i *)(((uintptr_t)buf + len) & -16);
109 
110     /* Loop over 16-byte aligned blocks of 64.  */
111     while (likely(p <= e)) {
112         __builtin_prefetch(p);
113         if (unlikely(!_mm_testz_si128(t, t))) {
114             return false;
115         }
116         t = p[-4] | p[-3] | p[-2] | p[-1];
117         p += 4;
118     }
119 
120     /* Finish the aligned tail.  */
121     t |= e[-3];
122     t |= e[-2];
123     t |= e[-1];
124 
125     /* Finish the unaligned tail.  */
126     t |= _mm_loadu_si128(buf + len - 16);
127 
128     return _mm_testz_si128(t, t);
129 }
130 
131 static bool __attribute__((target("avx2")))
132 buffer_zero_avx2(const void *buf, size_t len)
133 {
134     /* Begin with an unaligned head of 32 bytes.  */
135     __m256i t = _mm256_loadu_si256(buf);
136     __m256i *p = (__m256i *)(((uintptr_t)buf + 5 * 32) & -32);
137     __m256i *e = (__m256i *)(((uintptr_t)buf + len) & -32);
138 
139     /* Loop over 32-byte aligned blocks of 128.  */
140     while (p <= e) {
141         __builtin_prefetch(p);
142         if (unlikely(!_mm256_testz_si256(t, t))) {
143             return false;
144         }
145         t = p[-4] | p[-3] | p[-2] | p[-1];
146         p += 4;
147     } ;
148 
149     /* Finish the last block of 128 unaligned.  */
150     t |= _mm256_loadu_si256(buf + len - 4 * 32);
151     t |= _mm256_loadu_si256(buf + len - 3 * 32);
152     t |= _mm256_loadu_si256(buf + len - 2 * 32);
153     t |= _mm256_loadu_si256(buf + len - 1 * 32);
154 
155     return _mm256_testz_si256(t, t);
156 }
157 #endif /* CONFIG_AVX2_OPT */
158 
159 #ifdef CONFIG_AVX512F_OPT
160 static bool __attribute__((target("avx512f")))
161 buffer_zero_avx512(const void *buf, size_t len)
162 {
163     /* Begin with an unaligned head of 64 bytes.  */
164     __m512i t = _mm512_loadu_si512(buf);
165     __m512i *p = (__m512i *)(((uintptr_t)buf + 5 * 64) & -64);
166     __m512i *e = (__m512i *)(((uintptr_t)buf + len) & -64);
167 
168     /* Loop over 64-byte aligned blocks of 256.  */
169     while (p <= e) {
170         __builtin_prefetch(p);
171         if (unlikely(_mm512_test_epi64_mask(t, t))) {
172             return false;
173         }
174         t = p[-4] | p[-3] | p[-2] | p[-1];
175         p += 4;
176     }
177 
178     t |= _mm512_loadu_si512(buf + len - 4 * 64);
179     t |= _mm512_loadu_si512(buf + len - 3 * 64);
180     t |= _mm512_loadu_si512(buf + len - 2 * 64);
181     t |= _mm512_loadu_si512(buf + len - 1 * 64);
182 
183     return !_mm512_test_epi64_mask(t, t);
184 
185 }
186 #endif /* CONFIG_AVX512F_OPT */
187 
188 /*
189  * Make sure that these variables are appropriately initialized when
190  * SSE2 is enabled on the compiler command-line, but the compiler is
191  * too old to support CONFIG_AVX2_OPT.
192  */
193 #if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT)
194 # define INIT_USED     0
195 # define INIT_LENGTH   0
196 # define INIT_ACCEL    buffer_zero_int
197 #else
198 # ifndef __SSE2__
199 #  error "ISA selection confusion"
200 # endif
201 # define INIT_USED     CPUINFO_SSE2
202 # define INIT_LENGTH   64
203 # define INIT_ACCEL    buffer_zero_sse2
204 #endif
205 
206 static unsigned used_accel = INIT_USED;
207 static unsigned length_to_accel = INIT_LENGTH;
208 static bool (*buffer_accel)(const void *, size_t) = INIT_ACCEL;
209 
210 static unsigned __attribute__((noinline))
211 select_accel_cpuinfo(unsigned info)
212 {
213     /* Array is sorted in order of algorithm preference. */
214     static const struct {
215         unsigned bit;
216         unsigned len;
217         bool (*fn)(const void *, size_t);
218     } all[] = {
219 #ifdef CONFIG_AVX512F_OPT
220         { CPUINFO_AVX512F, 256, buffer_zero_avx512 },
221 #endif
222 #ifdef CONFIG_AVX2_OPT
223         { CPUINFO_AVX2,    128, buffer_zero_avx2 },
224         { CPUINFO_SSE4,     64, buffer_zero_sse4 },
225 #endif
226         { CPUINFO_SSE2,     64, buffer_zero_sse2 },
227         { CPUINFO_ALWAYS,    0, buffer_zero_int },
228     };
229 
230     for (unsigned i = 0; i < ARRAY_SIZE(all); ++i) {
231         if (info & all[i].bit) {
232             length_to_accel = all[i].len;
233             buffer_accel = all[i].fn;
234             return all[i].bit;
235         }
236     }
237     return 0;
238 }
239 
240 #if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT)
241 static void __attribute__((constructor)) init_accel(void)
242 {
243     used_accel = select_accel_cpuinfo(cpuinfo_init());
244 }
245 #endif /* CONFIG_AVX2_OPT */
246 
247 bool test_buffer_is_zero_next_accel(void)
248 {
249     /*
250      * Accumulate the accelerators that we've already tested, and
251      * remove them from the set to test this round.  We'll get back
252      * a zero from select_accel_cpuinfo when there are no more.
253      */
254     unsigned used = select_accel_cpuinfo(cpuinfo & ~used_accel);
255     used_accel |= used;
256     return used;
257 }
258 
259 static bool select_accel_fn(const void *buf, size_t len)
260 {
261     if (likely(len >= length_to_accel)) {
262         return buffer_accel(buf, len);
263     }
264     return buffer_zero_int(buf, len);
265 }
266 
267 #else
268 #define select_accel_fn  buffer_zero_int
269 bool test_buffer_is_zero_next_accel(void)
270 {
271     return false;
272 }
273 #endif
274 
275 /*
276  * Checks if a buffer is all zeroes
277  */
278 bool buffer_is_zero(const void *buf, size_t len)
279 {
280     if (unlikely(len == 0)) {
281         return true;
282     }
283 
284     /* Fetch the beginning of the buffer while we select the accelerator.  */
285     __builtin_prefetch(buf);
286 
287     /* Use an optimized zero check if possible.  Note that this also
288        includes a check for an unrolled loop over 64-bit integers.  */
289     return select_accel_fn(buf, len);
290 }
291