util/bufferiszero.c

   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 #include <immintrin.h>
  68
  69 /* Note that each of these vectorized functions require len >= 64.  */
  70
  71 static bool __attribute__((target("sse2")))
  72 buffer_zero_sse2(const void *buf, size_t len)
  73 {
  74     __m128i t = _mm_loadu_si128(buf);
  75     __m128i *p = (__m128i *)(((uintptr_t)buf + 5 * 16) & -16);
  76     __m128i *e = (__m128i *)(((uintptr_t)buf + len) & -16);
  77     __m128i zero = _mm_setzero_si128();
  78
  79     /* Loop over 16-byte aligned blocks of 64.  */
  80     while (likely(p <= e)) {
  81         __builtin_prefetch(p);
  82         t = _mm_cmpeq_epi8(t, zero);
  83         if (unlikely(_mm_movemask_epi8(t) != 0xFFFF)) {
  84             return false;
  85         }
  86         t = p[-4] | p[-3] | p[-2] | p[-1];
  87         p += 4;
  88     }
  89
  90     /* Finish the aligned tail.  */
  91     t |= e[-3];
  92     t |= e[-2];
  93     t |= e[-1];
  94
  95     /* Finish the unaligned tail.  */
  96     t |= _mm_loadu_si128(buf + len - 16);
  97
  98     return _mm_movemask_epi8(_mm_cmpeq_epi8(t, zero)) == 0xFFFF;
  99 }
 100
 101 #ifdef CONFIG_AVX2_OPT
 102 static bool __attribute__((target("sse4")))
 103 buffer_zero_sse4(const void *buf, size_t len)
 104 {
 105     __m128i t = _mm_loadu_si128(buf);
 106     __m128i *p = (__m128i *)(((uintptr_t)buf + 5 * 16) & -16);
 107     __m128i *e = (__m128i *)(((uintptr_t)buf + len) & -16);
 108
 109     /* Loop over 16-byte aligned blocks of 64.  */
 110     while (likely(p <= e)) {
 111         __builtin_prefetch(p);
 112         if (unlikely(!_mm_testz_si128(t, t))) {
 113             return false;
 114         }
 115         t = p[-4] | p[-3] | p[-2] | p[-1];
 116         p += 4;
 117     }
 118
 119     /* Finish the aligned tail.  */
 120     t |= e[-3];
 121     t |= e[-2];
 122     t |= e[-1];
 123
 124     /* Finish the unaligned tail.  */
 125     t |= _mm_loadu_si128(buf + len - 16);
 126
 127     return _mm_testz_si128(t, t);
 128 }
 129
 130 static bool __attribute__((target("avx2")))
 131 buffer_zero_avx2(const void *buf, size_t len)
 132 {
 133     /* Begin with an unaligned head of 32 bytes.  */
 134     __m256i t = _mm256_loadu_si256(buf);
 135     __m256i *p = (__m256i *)(((uintptr_t)buf + 5 * 32) & -32);
 136     __m256i *e = (__m256i *)(((uintptr_t)buf + len) & -32);
 137
 138     /* Loop over 32-byte aligned blocks of 128.  */
 139     while (p <= e) {
 140         __builtin_prefetch(p);
 141         if (unlikely(!_mm256_testz_si256(t, t))) {
 142             return false;
 143         }
 144         t = p[-4] | p[-3] | p[-2] | p[-1];
 145         p += 4;
 146     } ;
 147
 148     /* Finish the last block of 128 unaligned.  */
 149     t |= _mm256_loadu_si256(buf + len - 4 * 32);
 150     t |= _mm256_loadu_si256(buf + len - 3 * 32);
 151     t |= _mm256_loadu_si256(buf + len - 2 * 32);
 152     t |= _mm256_loadu_si256(buf + len - 1 * 32);
 153
 154     return _mm256_testz_si256(t, t);
 155 }
 156 #endif /* CONFIG_AVX2_OPT */
 157
 158 #ifdef CONFIG_AVX512F_OPT
 159 static bool __attribute__((target("avx512f")))
 160 buffer_zero_avx512(const void *buf, size_t len)
 161 {
 162     /* Begin with an unaligned head of 64 bytes.  */
 163     __m512i t = _mm512_loadu_si512(buf);
 164     __m512i *p = (__m512i *)(((uintptr_t)buf + 5 * 64) & -64);
 165     __m512i *e = (__m512i *)(((uintptr_t)buf + len) & -64);
 166
 167     /* Loop over 64-byte aligned blocks of 256.  */
 168     while (p <= e) {
 169         __builtin_prefetch(p);
 170         if (unlikely(_mm512_test_epi64_mask(t, t))) {
 171             return false;
 172         }
 173         t = p[-4] | p[-3] | p[-2] | p[-1];
 174         p += 4;
 175     }
 176
 177     t |= _mm512_loadu_si512(buf + len - 4 * 64);
 178     t |= _mm512_loadu_si512(buf + len - 3 * 64);
 179     t |= _mm512_loadu_si512(buf + len - 2 * 64);
 180     t |= _mm512_loadu_si512(buf + len - 1 * 64);
 181
 182     return !_mm512_test_epi64_mask(t, t);
 183
 184 }
 185 #endif /* CONFIG_AVX512F_OPT */
 186
 187
 188 /* Note that for test_buffer_is_zero_next_accel, the most preferred
 189  * ISA must have the least significant bit.
 190  */
 191 #define CACHE_AVX512F 1
 192 #define CACHE_AVX2    2
 193 #define CACHE_SSE4    4
 194 #define CACHE_SSE2    8
 195
 196 /* Make sure that these variables are appropriately initialized when
 197  * SSE2 is enabled on the compiler command-line, but the compiler is
 198  * too old to support CONFIG_AVX2_OPT.
 199  */
 200 #if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT)
 201 # define INIT_CACHE 0
 202 # define INIT_ACCEL buffer_zero_int
 203 #else
 204 # ifndef __SSE2__
 205 #  error "ISA selection confusion"
 206 # endif
 207 # define INIT_CACHE CACHE_SSE2
 208 # define INIT_ACCEL buffer_zero_sse2
 209 #endif
 210
 211 static unsigned cpuid_cache = INIT_CACHE;
 212 static bool (*buffer_accel)(const void *, size_t) = INIT_ACCEL;
 213 static int length_to_accel = 64;
 214
 215 static void init_accel(unsigned cache)
 216 {
 217     bool (*fn)(const void *, size_t) = buffer_zero_int;
 218     if (cache & CACHE_SSE2) {
 219         fn = buffer_zero_sse2;
 220         length_to_accel = 64;
 221     }
 222 #ifdef CONFIG_AVX2_OPT
 223     if (cache & CACHE_SSE4) {
 224         fn = buffer_zero_sse4;
 225         length_to_accel = 64;
 226     }
 227     if (cache & CACHE_AVX2) {
 228         fn = buffer_zero_avx2;
 229         length_to_accel = 128;
 230     }
 231 #endif
 232 #ifdef CONFIG_AVX512F_OPT
 233     if (cache & CACHE_AVX512F) {
 234         fn = buffer_zero_avx512;
 235         length_to_accel = 256;
 236     }
 237 #endif
 238     buffer_accel = fn;
 239 }
 240
 241 #if defined(CONFIG_AVX512F_OPT) || defined(CONFIG_AVX2_OPT)
 242 #include "qemu/cpuid.h"
 243
 244 static void __attribute__((constructor)) init_cpuid_cache(void)
 245 {
 246     unsigned max = __get_cpuid_max(0, NULL);
 247     int a, b, c, d;
 248     unsigned cache = 0;
 249
 250     if (max >= 1) {
 251         __cpuid(1, a, b, c, d);
 252         if (d & bit_SSE2) {
 253             cache |= CACHE_SSE2;
 254         }
 255         if (c & bit_SSE4_1) {
 256             cache |= CACHE_SSE4;
 257         }
 258
 259         /* We must check that AVX is not just available, but usable.  */
 260         if ((c & bit_OSXSAVE) && (c & bit_AVX) && max >= 7) {
 261             int bv;
 262             __asm("xgetbv" : "=a"(bv), "=d"(d) : "c"(0));
 263             __cpuid_count(7, 0, a, b, c, d);
 264             if ((bv & 0x6) == 0x6 && (b & bit_AVX2)) {
 265                 cache |= CACHE_AVX2;
 266             }
 267             /* 0xe6:
 268             *  XCR0[7:5] = 111b (OPMASK state, upper 256-bit of ZMM0-ZMM15
 269             *                    and ZMM16-ZMM31 state are enabled by OS)
 270             *  XCR0[2:1] = 11b (XMM state and YMM state are enabled by OS)
 271             */
 272             if ((bv & 0xe6) == 0xe6 && (b & bit_AVX512F)) {
 273                 cache |= CACHE_AVX512F;
 274             }
 275         }
 276     }
 277     cpuid_cache = cache;
 278     init_accel(cache);
 279 }
 280 #endif /* CONFIG_AVX2_OPT */
 281
 282 bool test_buffer_is_zero_next_accel(void)
 283 {
 284     /* If no bits set, we just tested buffer_zero_int, and there
 285        are no more acceleration options to test.  */
 286     if (cpuid_cache == 0) {
 287         return false;
 288     }
 289     /* Disable the accelerator we used before and select a new one.  */
 290     cpuid_cache &= cpuid_cache - 1;
 291     init_accel(cpuid_cache);
 292     return true;
 293 }
 294
 295 static bool select_accel_fn(const void *buf, size_t len)
 296 {
 297     if (likely(len >= length_to_accel)) {
 298         return buffer_accel(buf, len);
 299     }
 300     return buffer_zero_int(buf, len);
 301 }
 302
 303 #else
 304 #define select_accel_fn  buffer_zero_int
 305 bool test_buffer_is_zero_next_accel(void)
 306 {
 307     return false;
 308 }
 309 #endif
 310
 311 /*
 312  * Checks if a buffer is all zeroes
 313  */
 314 bool buffer_is_zero(const void *buf, size_t len)
 315 {
 316     if (unlikely(len == 0)) {
 317         return true;
 318     }
 319
 320     /* Fetch the beginning of the buffer while we select the accelerator.  */
 321     __builtin_prefetch(buf);
 322
 323     /* Use an optimized zero check if possible.  Note that this also
 324        includes a check for an unrolled loop over 64-bit integers.  */
 325     return select_accel_fn(buf, len);
 326 }