xref: /openbmc/qemu/util/bufferiszero.c (revision a6caeee8)
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 
28 static bool
29 buffer_zero_int(const void *buf, size_t len)
30 {
31     if (unlikely(len < 8)) {
32         /* For a very small buffer, simply accumulate all the bytes.  */
33         const unsigned char *p = buf;
34         const unsigned char *e = buf + len;
35         unsigned char t = 0;
36 
37         do {
38             t |= *p++;
39         } while (p < e);
40 
41         return t == 0;
42     } else {
43         /* Otherwise, use the unaligned memory access functions to
44            handle the beginning and end of the buffer, with a couple
45            of loops handling the middle aligned section.  */
46         uint64_t t = ldq_he_p(buf);
47         const uint64_t *p = (uint64_t *)(((uintptr_t)buf + 8) & -8);
48         const uint64_t *e = (uint64_t *)(((uintptr_t)buf + len) & -8);
49 
50         for (; p + 8 <= e; p += 8) {
51             __builtin_prefetch(p + 8);
52             if (t) {
53                 return false;
54             }
55             t = p[0] | p[1] | p[2] | p[3] | p[4] | p[5] | p[6] | p[7];
56         }
57         while (p < e) {
58             t |= *p++;
59         }
60         t |= ldq_he_p(buf + len - 8);
61 
62         return t == 0;
63     }
64 }
65 
66 #if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT) || defined(__SSE2__)
67 /* Do not use push_options pragmas unnecessarily, because clang
68  * does not support them.
69  */
70 #if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT)
71 #pragma GCC push_options
72 #pragma GCC target("sse2")
73 #endif
74 #include <emmintrin.h>
75 
76 /* Note that each of these vectorized functions require len >= 64.  */
77 
78 static bool
79 buffer_zero_sse2(const void *buf, size_t len)
80 {
81     __m128i t = _mm_loadu_si128(buf);
82     __m128i *p = (__m128i *)(((uintptr_t)buf + 5 * 16) & -16);
83     __m128i *e = (__m128i *)(((uintptr_t)buf + len) & -16);
84     __m128i zero = _mm_setzero_si128();
85 
86     /* Loop over 16-byte aligned blocks of 64.  */
87     while (likely(p <= e)) {
88         __builtin_prefetch(p);
89         t = _mm_cmpeq_epi8(t, zero);
90         if (unlikely(_mm_movemask_epi8(t) != 0xFFFF)) {
91             return false;
92         }
93         t = p[-4] | p[-3] | p[-2] | p[-1];
94         p += 4;
95     }
96 
97     /* Finish the aligned tail.  */
98     t |= e[-3];
99     t |= e[-2];
100     t |= e[-1];
101 
102     /* Finish the unaligned tail.  */
103     t |= _mm_loadu_si128(buf + len - 16);
104 
105     return _mm_movemask_epi8(_mm_cmpeq_epi8(t, zero)) == 0xFFFF;
106 }
107 #if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT)
108 #pragma GCC pop_options
109 #endif
110 
111 #ifdef CONFIG_AVX2_OPT
112 /* Note that due to restrictions/bugs wrt __builtin functions in gcc <= 4.8,
113  * the includes have to be within the corresponding push_options region, and
114  * therefore the regions themselves have to be ordered with increasing ISA.
115  */
116 #pragma GCC push_options
117 #pragma GCC target("sse4")
118 #include <smmintrin.h>
119 
120 static bool
121 buffer_zero_sse4(const void *buf, size_t len)
122 {
123     __m128i t = _mm_loadu_si128(buf);
124     __m128i *p = (__m128i *)(((uintptr_t)buf + 5 * 16) & -16);
125     __m128i *e = (__m128i *)(((uintptr_t)buf + len) & -16);
126 
127     /* Loop over 16-byte aligned blocks of 64.  */
128     while (likely(p <= e)) {
129         __builtin_prefetch(p);
130         if (unlikely(!_mm_testz_si128(t, t))) {
131             return false;
132         }
133         t = p[-4] | p[-3] | p[-2] | p[-1];
134         p += 4;
135     }
136 
137     /* Finish the aligned tail.  */
138     t |= e[-3];
139     t |= e[-2];
140     t |= e[-1];
141 
142     /* Finish the unaligned tail.  */
143     t |= _mm_loadu_si128(buf + len - 16);
144 
145     return _mm_testz_si128(t, t);
146 }
147 
148 #pragma GCC pop_options
149 #pragma GCC push_options
150 #pragma GCC target("avx2")
151 #include <immintrin.h>
152 
153 static bool
154 buffer_zero_avx2(const void *buf, size_t len)
155 {
156     /* Begin with an unaligned head of 32 bytes.  */
157     __m256i t = _mm256_loadu_si256(buf);
158     __m256i *p = (__m256i *)(((uintptr_t)buf + 5 * 32) & -32);
159     __m256i *e = (__m256i *)(((uintptr_t)buf + len) & -32);
160 
161     /* Loop over 32-byte aligned blocks of 128.  */
162     while (p <= e) {
163         __builtin_prefetch(p);
164         if (unlikely(!_mm256_testz_si256(t, t))) {
165             return false;
166         }
167         t = p[-4] | p[-3] | p[-2] | p[-1];
168         p += 4;
169     } ;
170 
171     /* Finish the last block of 128 unaligned.  */
172     t |= _mm256_loadu_si256(buf + len - 4 * 32);
173     t |= _mm256_loadu_si256(buf + len - 3 * 32);
174     t |= _mm256_loadu_si256(buf + len - 2 * 32);
175     t |= _mm256_loadu_si256(buf + len - 1 * 32);
176 
177     return _mm256_testz_si256(t, t);
178 }
179 #pragma GCC pop_options
180 #endif /* CONFIG_AVX2_OPT */
181 
182 #ifdef CONFIG_AVX512F_OPT
183 #pragma GCC push_options
184 #pragma GCC target("avx512f")
185 #include <immintrin.h>
186 
187 static bool
188 buffer_zero_avx512(const void *buf, size_t len)
189 {
190     /* Begin with an unaligned head of 64 bytes.  */
191     __m512i t = _mm512_loadu_si512(buf);
192     __m512i *p = (__m512i *)(((uintptr_t)buf + 5 * 64) & -64);
193     __m512i *e = (__m512i *)(((uintptr_t)buf + len) & -64);
194 
195     /* Loop over 64-byte aligned blocks of 256.  */
196     while (p <= e) {
197         __builtin_prefetch(p);
198         if (unlikely(_mm512_test_epi64_mask(t, t))) {
199             return false;
200         }
201         t = p[-4] | p[-3] | p[-2] | p[-1];
202         p += 4;
203     }
204 
205     t |= _mm512_loadu_si512(buf + len - 4 * 64);
206     t |= _mm512_loadu_si512(buf + len - 3 * 64);
207     t |= _mm512_loadu_si512(buf + len - 2 * 64);
208     t |= _mm512_loadu_si512(buf + len - 1 * 64);
209 
210     return !_mm512_test_epi64_mask(t, t);
211 
212 }
213 #pragma GCC pop_options
214 #endif
215 
216 
217 /* Note that for test_buffer_is_zero_next_accel, the most preferred
218  * ISA must have the least significant bit.
219  */
220 #define CACHE_AVX512F 1
221 #define CACHE_AVX2    2
222 #define CACHE_SSE4    4
223 #define CACHE_SSE2    8
224 
225 /* Make sure that these variables are appropriately initialized when
226  * SSE2 is enabled on the compiler command-line, but the compiler is
227  * too old to support CONFIG_AVX2_OPT.
228  */
229 #if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT)
230 # define INIT_CACHE 0
231 # define INIT_ACCEL buffer_zero_int
232 #else
233 # ifndef __SSE2__
234 #  error "ISA selection confusion"
235 # endif
236 # define INIT_CACHE CACHE_SSE2
237 # define INIT_ACCEL buffer_zero_sse2
238 #endif
239 
240 static unsigned cpuid_cache = INIT_CACHE;
241 static bool (*buffer_accel)(const void *, size_t) = INIT_ACCEL;
242 static int length_to_accel = 64;
243 
244 static void init_accel(unsigned cache)
245 {
246     bool (*fn)(const void *, size_t) = buffer_zero_int;
247     if (cache & CACHE_SSE2) {
248         fn = buffer_zero_sse2;
249         length_to_accel = 64;
250     }
251 #ifdef CONFIG_AVX2_OPT
252     if (cache & CACHE_SSE4) {
253         fn = buffer_zero_sse4;
254         length_to_accel = 64;
255     }
256     if (cache & CACHE_AVX2) {
257         fn = buffer_zero_avx2;
258         length_to_accel = 128;
259     }
260 #endif
261 #ifdef CONFIG_AVX512F_OPT
262     if (cache & CACHE_AVX512F) {
263         fn = buffer_zero_avx512;
264         length_to_accel = 256;
265     }
266 #endif
267     buffer_accel = fn;
268 }
269 
270 #if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT)
271 #include "qemu/cpuid.h"
272 
273 static void __attribute__((constructor)) init_cpuid_cache(void)
274 {
275     unsigned max = __get_cpuid_max(0, NULL);
276     int a, b, c, d;
277     unsigned cache = 0;
278 
279     if (max >= 1) {
280         __cpuid(1, a, b, c, d);
281         if (d & bit_SSE2) {
282             cache |= CACHE_SSE2;
283         }
284         if (c & bit_SSE4_1) {
285             cache |= CACHE_SSE4;
286         }
287 
288         /* We must check that AVX is not just available, but usable.  */
289         if ((c & bit_OSXSAVE) && (c & bit_AVX) && max >= 7) {
290             int bv;
291             __asm("xgetbv" : "=a"(bv), "=d"(d) : "c"(0));
292             __cpuid_count(7, 0, a, b, c, d);
293             if ((bv & 0x6) == 0x6 && (b & bit_AVX2)) {
294                 cache |= CACHE_AVX2;
295             }
296             /* 0xe6:
297             *  XCR0[7:5] = 111b (OPMASK state, upper 256-bit of ZMM0-ZMM15
298             *                    and ZMM16-ZMM31 state are enabled by OS)
299             *  XCR0[2:1] = 11b (XMM state and YMM state are enabled by OS)
300             */
301             if ((bv & 0xe6) == 0xe6 && (b & bit_AVX512F)) {
302                 cache |= CACHE_AVX512F;
303             }
304         }
305     }
306     cpuid_cache = cache;
307     init_accel(cache);
308 }
309 #endif /* CONFIG_AVX2_OPT */
310 
311 bool test_buffer_is_zero_next_accel(void)
312 {
313     /* If no bits set, we just tested buffer_zero_int, and there
314        are no more acceleration options to test.  */
315     if (cpuid_cache == 0) {
316         return false;
317     }
318     /* Disable the accelerator we used before and select a new one.  */
319     cpuid_cache &= cpuid_cache - 1;
320     init_accel(cpuid_cache);
321     return true;
322 }
323 
324 static bool select_accel_fn(const void *buf, size_t len)
325 {
326     if (likely(len >= length_to_accel)) {
327         return buffer_accel(buf, len);
328     }
329     return buffer_zero_int(buf, len);
330 }
331 
332 #else
333 #define select_accel_fn  buffer_zero_int
334 bool test_buffer_is_zero_next_accel(void)
335 {
336     return false;
337 }
338 #endif
339 
340 /*
341  * Checks if a buffer is all zeroes
342  */
343 bool buffer_is_zero(const void *buf, size_t len)
344 {
345     if (unlikely(len == 0)) {
346         return true;
347     }
348 
349     /* Fetch the beginning of the buffer while we select the accelerator.  */
350     __builtin_prefetch(buf);
351 
352     /* Use an optimized zero check if possible.  Note that this also
353        includes a check for an unrolled loop over 64-bit integers.  */
354     return select_accel_fn(buf, len);
355 }
356