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-common.h"
  26 #include "qemu/cutils.h"
  27 #include "qemu/bswap.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_AVX2_OPT) || defined(__SSE2__)
  68 /* Do not use push_options pragmas unnecessarily, because clang
  69  * does not support them.
  70  */
  71 #ifdef CONFIG_AVX2_OPT
  72 #pragma GCC push_options
  73 #pragma GCC target("sse2")
  74 #endif
  75 #include <emmintrin.h>
  76
  77 /* Note that each of these vectorized functions require len >= 64.  */
  78
  79 static bool
  80 buffer_zero_sse2(const void *buf, size_t len)
  81 {
  82     __m128i t = _mm_loadu_si128(buf);
  83     __m128i *p = (__m128i *)(((uintptr_t)buf + 5 * 16) & -16);
  84     __m128i *e = (__m128i *)(((uintptr_t)buf + len) & -16);
  85     __m128i zero = _mm_setzero_si128();
  86
  87     /* Loop over 16-byte aligned blocks of 64.  */
  88     while (likely(p <= e)) {
  89         __builtin_prefetch(p);
  90         t = _mm_cmpeq_epi8(t, zero);
  91         if (unlikely(_mm_movemask_epi8(t) != 0xFFFF)) {
  92             return false;
  93         }
  94         t = p[-4] | p[-3] | p[-2] | p[-1];
  95         p += 4;
  96     }
  97
  98     /* Finish the aligned tail.  */
  99     t |= e[-3];
 100     t |= e[-2];
 101     t |= e[-1];
 102
 103     /* Finish the unaligned tail.  */
 104     t |= _mm_loadu_si128(buf + len - 16);
 105
 106     return _mm_movemask_epi8(_mm_cmpeq_epi8(t, zero)) == 0xFFFF;
 107 }
 108 #ifdef CONFIG_AVX2_OPT
 109 #pragma GCC pop_options
 110 #endif
 111
 112 #ifdef CONFIG_AVX2_OPT
 113 /* Note that due to restrictions/bugs wrt __builtin functions in gcc <= 4.8,
 114  * the includes have to be within the corresponding push_options region, and
 115  * therefore the regions themselves have to be ordered with increasing ISA.
 116  */
 117 #pragma GCC push_options
 118 #pragma GCC target("sse4")
 119 #include <smmintrin.h>
 120
 121 static bool
 122 buffer_zero_sse4(const void *buf, size_t len)
 123 {
 124     __m128i t = _mm_loadu_si128(buf);
 125     __m128i *p = (__m128i *)(((uintptr_t)buf + 5 * 16) & -16);
 126     __m128i *e = (__m128i *)(((uintptr_t)buf + len) & -16);
 127
 128     /* Loop over 16-byte aligned blocks of 64.  */
 129     while (likely(p <= e)) {
 130         __builtin_prefetch(p);
 131         if (unlikely(!_mm_testz_si128(t, t))) {
 132             return false;
 133         }
 134         t = p[-4] | p[-3] | p[-2] | p[-1];
 135         p += 4;
 136     }
 137
 138     /* Finish the aligned tail.  */
 139     t |= e[-3];
 140     t |= e[-2];
 141     t |= e[-1];
 142
 143     /* Finish the unaligned tail.  */
 144     t |= _mm_loadu_si128(buf + len - 16);
 145
 146     return _mm_testz_si128(t, t);
 147 }
 148
 149 #pragma GCC pop_options
 150 #pragma GCC push_options
 151 #pragma GCC target("avx2")
 152 #include <immintrin.h>
 153
 154 static bool
 155 buffer_zero_avx2(const void *buf, size_t len)
 156 {
 157     /* Begin with an unaligned head of 32 bytes.  */
 158     __m256i t = _mm256_loadu_si256(buf);
 159     __m256i *p = (__m256i *)(((uintptr_t)buf + 5 * 32) & -32);
 160     __m256i *e = (__m256i *)(((uintptr_t)buf + len) & -32);
 161
 162     if (likely(p <= e)) {
 163         /* Loop over 32-byte aligned blocks of 128.  */
 164         do {
 165             __builtin_prefetch(p);
 166             if (unlikely(!_mm256_testz_si256(t, t))) {
 167                 return false;
 168             }
 169             t = p[-4] | p[-3] | p[-2] | p[-1];
 170             p += 4;
 171         } while (p <= e);
 172     } else {
 173         t |= _mm256_loadu_si256(buf + 32);
 174         if (len <= 128) {
 175             goto last2;
 176         }
 177     }
 178
 179     /* Finish the last block of 128 unaligned.  */
 180     t |= _mm256_loadu_si256(buf + len - 4 * 32);
 181     t |= _mm256_loadu_si256(buf + len - 3 * 32);
 182  last2:
 183     t |= _mm256_loadu_si256(buf + len - 2 * 32);
 184     t |= _mm256_loadu_si256(buf + len - 1 * 32);
 185
 186     return _mm256_testz_si256(t, t);
 187 }
 188 #pragma GCC pop_options
 189 #endif /* CONFIG_AVX2_OPT */
 190
 191 /* Note that for test_buffer_is_zero_next_accel, the most preferred
 192  * ISA must have the least significant bit.
 193  */
 194 #define CACHE_AVX2    1
 195 #define CACHE_SSE4    2
 196 #define CACHE_SSE2    4
 197
 198 /* Make sure that these variables are appropriately initialized when
 199  * SSE2 is enabled on the compiler command-line, but the compiler is
 200  * too old to support CONFIG_AVX2_OPT.
 201  */
 202 #ifdef CONFIG_AVX2_OPT
 203 # define INIT_CACHE 0
 204 # define INIT_ACCEL buffer_zero_int
 205 #else
 206 # ifndef __SSE2__
 207 #  error "ISA selection confusion"
 208 # endif
 209 # define INIT_CACHE CACHE_SSE2
 210 # define INIT_ACCEL buffer_zero_sse2
 211 #endif
 212
 213 static unsigned cpuid_cache = INIT_CACHE;
 214 static bool (*buffer_accel)(const void *, size_t) = INIT_ACCEL;
 215
 216 static void init_accel(unsigned cache)
 217 {
 218     bool (*fn)(const void *, size_t) = buffer_zero_int;
 219     if (cache & CACHE_SSE2) {
 220         fn = buffer_zero_sse2;
 221     }
 222 #ifdef CONFIG_AVX2_OPT
 223     if (cache & CACHE_SSE4) {
 224         fn = buffer_zero_sse4;
 225     }
 226     if (cache & CACHE_AVX2) {
 227         fn = buffer_zero_avx2;
 228     }
 229 #endif
 230     buffer_accel = fn;
 231 }
 232
 233 #ifdef CONFIG_AVX2_OPT
 234 #include "qemu/cpuid.h"
 235
 236 static void __attribute__((constructor)) init_cpuid_cache(void)
 237 {
 238     int max = __get_cpuid_max(0, NULL);
 239     int a, b, c, d;
 240     unsigned cache = 0;
 241
 242     if (max >= 1) {
 243         __cpuid(1, a, b, c, d);
 244         if (d & bit_SSE2) {
 245             cache |= CACHE_SSE2;
 246         }
 247         if (c & bit_SSE4_1) {
 248             cache |= CACHE_SSE4;
 249         }
 250
 251         /* We must check that AVX is not just available, but usable.  */
 252         if ((c & bit_OSXSAVE) && (c & bit_AVX) && max >= 7) {
 253             int bv;
 254             __asm("xgetbv" : "=a"(bv), "=d"(d) : "c"(0));
 255             __cpuid_count(7, 0, a, b, c, d);
 256             if ((bv & 6) == 6 && (b & bit_AVX2)) {
 257                 cache |= CACHE_AVX2;
 258             }
 259         }
 260     }
 261     cpuid_cache = cache;
 262     init_accel(cache);
 263 }
 264 #endif /* CONFIG_AVX2_OPT */
 265
 266 bool test_buffer_is_zero_next_accel(void)
 267 {
 268     /* If no bits set, we just tested buffer_zero_int, and there
 269        are no more acceleration options to test.  */
 270     if (cpuid_cache == 0) {
 271         return false;
 272     }
 273     /* Disable the accelerator we used before and select a new one.  */
 274     cpuid_cache &= cpuid_cache - 1;
 275     init_accel(cpuid_cache);
 276     return true;
 277 }
 278
 279 static bool select_accel_fn(const void *buf, size_t len)
 280 {
 281     if (likely(len >= 64)) {
 282         return buffer_accel(buf, len);
 283     }
 284     return buffer_zero_int(buf, len);
 285 }
 286
 287 #else
 288 #define select_accel_fn  buffer_zero_int
 289 bool test_buffer_is_zero_next_accel(void)
 290 {
 291     return false;
 292 }
 293 #endif
 294
 295 /*
 296  * Checks if a buffer is all zeroes
 297  */
 298 bool buffer_is_zero(const void *buf, size_t len)
 299 {
 300     if (unlikely(len == 0)) {
 301         return true;
 302     }
 303
 304     /* Fetch the beginning of the buffer while we select the accelerator.  */
 305     __builtin_prefetch(buf);
 306
 307     /* Use an optimized zero check if possible.  Note that this also
 308        includes a check for an unrolled loop over 64-bit integers.  */
 309     return select_accel_fn(buf, len);
 310 }