1 ///////////////////////////////////////////////////////////////////////////////
2 //
3 /// \file crc64.c
4 /// \brief CRC64 calculation
5 ///
6 /// There are two methods in this file. crc64_generic uses the
7 /// the slice-by-four algorithm. This is the same idea that is
8 /// used in crc32_fast.c, but for CRC64 we use only four tables
9 /// instead of eight to avoid increasing CPU cache usage.
10 ///
11 /// crc64_clmul uses 32/64-bit x86 SSSE3, SSE4.1, and CLMUL instructions.
12 /// It was derived from
13 /// https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/fast-crc-computation-generic-polynomials-pclmulqdq-paper.pdf
14 /// and the public domain code from https://github.com/rawrunprotected/crc
15 /// (URLs were checked on 2022-11-07).
16 ///
17 /// FIXME: Builds for 32-bit x86 use crc64_x86.S by default instead
18 /// of this file and thus CLMUL version isn't available on 32-bit x86
19 /// unless configured with --disable-assembler. Even then the lookup table
20 /// isn't omitted in crc64_table.c since it doesn't know that assembly
21 /// code has been disabled.
22 //
23 // Authors: Lasse Collin
24 // Ilya Kurdyukov
25 //
26 // This file has been put into the public domain.
27 // You can do whatever you want with this file.
28 //
29 ///////////////////////////////////////////////////////////////////////////////
30
31 #include "check.h"
32
33 #undef CRC_GENERIC
34 #undef CRC_CLMUL
35 #undef CRC_USE_GENERIC_FOR_SMALL_INPUTS
36
37 // If CLMUL cannot be used then only the generic slice-by-four is built.
38 #if !defined(HAVE_USABLE_CLMUL)
39 # define CRC_GENERIC 1
40
41 // If CLMUL is allowed unconditionally in the compiler options then the
42 // generic version can be omitted. Note that this doesn't work with MSVC
43 // as I don't know how to detect the features here.
44 //
45 // NOTE: Keep this this in sync with crc64_table.c.
46 #elif (defined(__SSSE3__) && defined(__SSE4_1__) && defined(__PCLMUL__)) \
47 || (defined(__e2k__) && __iset__ >= 6)
48 # define CRC_CLMUL 1
49
50 // Otherwise build both and detect at runtime which version to use.
51 #else
52 # define CRC_GENERIC 1
53 # define CRC_CLMUL 1
54
55 /*
56 // The generic code is much faster with 1-8-byte inputs and has
57 // similar performance up to 16 bytes at least in microbenchmarks
58 // (it depends on input buffer alignment too). If both versions are
59 // built, this #define will use the generic version for inputs up to
60 // 16 bytes and CLMUL for bigger inputs. It saves a little in code
61 // size since the special cases for 0-16-byte inputs will be omitted
62 // from the CLMUL code.
63 # define CRC_USE_GENERIC_FOR_SMALL_INPUTS 1
64 */
65
66 # if defined(_MSC_VER)
67 # include <intrin.h>
68 # elif defined(HAVE_CPUID_H)
69 # include <cpuid.h>
70 # endif
71 #endif
72
73
74 /////////////////////////////////
75 // Generic slice-by-four CRC64 //
76 /////////////////////////////////
77
78 #ifdef CRC_GENERIC
79
80 #include "crc_macros.h"
81
82
83 #ifdef WORDS_BIGENDIAN
84 # define A1(x) ((x) >> 56)
85 #else
86 # define A1 A
87 #endif
88
89
90 // See the comments in crc32_fast.c. They aren't duplicated here.
91 static uint64_t
92 crc64_generic(const uint8_t *buf, size_t size, uint64_t crc)
93 {
94 crc = ~crc;
95
96 #ifdef WORDS_BIGENDIAN
97 crc = bswap64(crc);
98 #endif
99
100 if (size > 4) {
101 while ((uintptr_t)(buf) & 3) {
102 crc = lzma_crc64_table[0][*buf++ ^ A1(crc)] ^ S8(crc);
103 --size;
104 }
105
106 const uint8_t *const limit = buf + (size & ~(size_t)(3));
107 size &= (size_t)(3);
108
109 while (buf < limit) {
110 #ifdef WORDS_BIGENDIAN
111 const uint32_t tmp = (uint32_t)(crc >> 32)
112 ^ aligned_read32ne(buf);
113 #else
114 const uint32_t tmp = (uint32_t)crc
115 ^ aligned_read32ne(buf);
116 #endif
117 buf += 4;
118
119 crc = lzma_crc64_table[3][A(tmp)]
120 ^ lzma_crc64_table[2][B(tmp)]
121 ^ S32(crc)
122 ^ lzma_crc64_table[1][C(tmp)]
123 ^ lzma_crc64_table[0][D(tmp)];
124 }
125 }
126
127 while (size-- != 0)
128 crc = lzma_crc64_table[0][*buf++ ^ A1(crc)] ^ S8(crc);
129
130 #ifdef WORDS_BIGENDIAN
131 crc = bswap64(crc);
132 #endif
133
134 return ~crc;
135 }
136 #endif
137
138
139 /////////////////////
140 // x86 CLMUL CRC64 //
141 /////////////////////
142
143 #ifdef CRC_CLMUL
144
145 #include <immintrin.h>
146
147
148 /*
149 // These functions were used to generate the constants
150 // at the top of crc64_clmul().
151 static uint64_t
152 calc_lo(uint64_t poly)
153 {
154 uint64_t a = poly;
155 uint64_t b = 0;
156
157 for (unsigned i = 0; i < 64; ++i) {
158 b = (b >> 1) | (a << 63);
159 a = (a >> 1) ^ (a & 1 ? poly : 0);
160 }
161
162 return b;
163 }
164
165 static uint64_t
166 calc_hi(uint64_t poly, uint64_t a)
167 {
168 for (unsigned i = 0; i < 64; ++i)
169 a = (a >> 1) ^ (a & 1 ? poly : 0);
170
171 return a;
172 }
173 */
174
175
176 #define MASK_L(in, mask, r) \
177 r = _mm_shuffle_epi8(in, mask)
178
179 #define MASK_H(in, mask, r) \
180 r = _mm_shuffle_epi8(in, _mm_xor_si128(mask, vsign))
181
182 #define MASK_LH(in, mask, low, high) \
183 MASK_L(in, mask, low); \
184 MASK_H(in, mask, high)
185
186
187 // MSVC (VS2015 - VS2022) produces bad 32-bit x86 code from the CLMUL CRC
188 // code when optimizations are enabled (release build). According to the bug
189 // report, the ebx register is corrupted and the calculated result is wrong.
190 // Trying to workaround the problem with "__asm mov ebx, ebx" didn't help.
191 // The following pragma works and performance is still good. x86-64 builds
192 // aren't affected by this problem.
193 //
194 // NOTE: Another pragma after the function restores the optimizations.
195 // If the #if condition here is updated, the other one must be updated too.
196 #if defined(_MSC_VER) && !defined(__INTEL_COMPILER) && !defined(__clang__) \
197 && defined(_M_IX86)
198 # pragma optimize("g", off)
199 #endif
200
201 // EDG-based compilers (Intel's classic compiler and compiler for E2K) can
202 // define __GNUC__ but the attribute must not be used with them.
203 // The new Clang-based ICX needs the attribute.
204 //
205 // NOTE: Build systems check for this too, keep them in sync with this.
206 #if (defined(__GNUC__) || defined(__clang__)) && !defined(__EDG__)
207 __attribute__((__target__("ssse3,sse4.1,pclmul")))
208 #endif
209 // The intrinsics use 16-byte-aligned reads from buf, thus they may read
210 // up to 15 bytes before or after the buffer (depending on the alignment
211 // of the buf argument). The values of the extra bytes are ignored.
212 // This unavoidably trips -fsanitize=address so address sanitizier has
213 // to be disabled for this function.
214 #if lzma_has_attribute(__no_sanitize_address__)
215 __attribute__((__no_sanitize_address__))
216 #endif
217 static uint64_t
218 crc64_clmul(const uint8_t *buf, size_t size, uint64_t crc)
219 {
220 // The prototypes of the intrinsics use signed types while most of
221 // the values are treated as unsigned here. These warnings in this
222 // function have been checked and found to be harmless so silence them.
223 #if TUKLIB_GNUC_REQ(4, 6) || defined(__clang__)
224 # pragma GCC diagnostic push
225 # pragma GCC diagnostic ignored "-Wsign-conversion"
226 # pragma GCC diagnostic ignored "-Wconversion"
227 #endif
228
229 #ifndef CRC_USE_GENERIC_FOR_SMALL_INPUTS
230 // The code assumes that there is at least one byte of input.
231 if (size == 0)
232 return crc;
233 #endif
234
235 // const uint64_t poly = 0xc96c5795d7870f42; // CRC polynomial
236 const uint64_t p = 0x92d8af2baf0e1e85; // (poly << 1) | 1
237 const uint64_t mu = 0x9c3e466c172963d5; // (calc_lo(poly) << 1) | 1
238 const uint64_t k2 = 0xdabe95afc7875f40; // calc_hi(poly, 1)
239 const uint64_t k1 = 0xe05dd497ca393ae4; // calc_hi(poly, k2)
240 const __m128i vfold0 = _mm_set_epi64x(p, mu);
241 const __m128i vfold1 = _mm_set_epi64x(k2, k1);
242
243 // Create a vector with 8-bit values 0 to 15. This is used to
244 // construct control masks for _mm_blendv_epi8 and _mm_shuffle_epi8.
245 const __m128i vramp = _mm_setr_epi32(
246 0x03020100, 0x07060504, 0x0b0a0908, 0x0f0e0d0c);
247
248 // This is used to inverse the control mask of _mm_shuffle_epi8
249 // so that bytes that wouldn't be picked with the original mask
250 // will be picked and vice versa.
251 const __m128i vsign = _mm_set1_epi8(0x80);
252
253 // Memory addresses A to D and the distances between them:
254 //
255 // A B C D
256 // [skip_start][size][skip_end]
257 // [ size2 ]
258 //
259 // A and D are 16-byte aligned. B and C are 1-byte aligned.
260 // skip_start and skip_end are 0-15 bytes. size is at least 1 byte.
261 //
262 // A = aligned_buf will initially point to this address.
263 // B = The address pointed by the caller-supplied buf.
264 // C = buf + size == aligned_buf + size2
265 // D = buf + size + skip_end == aligned_buf + size2 + skip_end
266 const size_t skip_start = (size_t)((uintptr_t)buf & 15);
267 const size_t skip_end = (size_t)((0U - (uintptr_t)(buf + size)) & 15);
268 const __m128i *aligned_buf = (const __m128i *)(
269 (uintptr_t)buf & ~(uintptr_t)15);
270
271 // If size2 <= 16 then the whole input fits into a single 16-byte
272 // vector. If size2 > 16 then at least two 16-byte vectors must
273 // be processed. If size2 > 16 && size <= 16 then there is only
274 // one 16-byte vector's worth of input but it is unaligned in memory.
275 //
276 // NOTE: There is no integer overflow here if the arguments are valid.
277 // If this overflowed, buf + size would too.
278 size_t size2 = skip_start + size;
279
280 // Masks to be used with _mm_blendv_epi8 and _mm_shuffle_epi8:
281 // The first skip_start or skip_end bytes in the vectors will have
282 // the high bit (0x80) set. _mm_blendv_epi8 and _mm_shuffle_epi8
283 // will produce zeros for these positions. (Bitwise-xor of these
284 // masks with vsign will produce the opposite behavior.)
285 const __m128i mask_start
286 = _mm_sub_epi8(vramp, _mm_set1_epi8(skip_start));
287 const __m128i mask_end = _mm_sub_epi8(vramp, _mm_set1_epi8(skip_end));
288
289 // Get the first 1-16 bytes into data0. If loading less than 16 bytes,
290 // the bytes are loaded to the high bits of the vector and the least
291 // significant positions are filled with zeros.
292 const __m128i data0 = _mm_blendv_epi8(_mm_load_si128(aligned_buf),
293 _mm_setzero_si128(), mask_start);
294 ++aligned_buf;
295
296 #if defined(__i386__) || defined(_M_IX86)
297 const __m128i initial_crc = _mm_set_epi64x(0, ~crc);
298 #else
299 // GCC and Clang would produce good code with _mm_set_epi64x
300 // but MSVC needs _mm_cvtsi64_si128 on x86-64.
301 const __m128i initial_crc = _mm_cvtsi64_si128(~crc);
302 #endif
303
304 __m128i v0, v1, v2, v3;
305
306 #ifndef CRC_USE_GENERIC_FOR_SMALL_INPUTS
307 if (size <= 16) {
308 // Right-shift initial_crc by 1-16 bytes based on "size"
309 // and store the result in v1 (high bytes) and v0 (low bytes).
310 //
311 // NOTE: The highest 8 bytes of initial_crc are zeros so
312 // v1 will be filled with zeros if size >= 8. The highest 8
313 // bytes of v1 will always become zeros.
314 //
315 // [ v1 ][ v0 ]
316 // [ initial_crc ] size == 1
317 // [ initial_crc ] size == 2
318 // [ initial_crc ] size == 15
319 // [ initial_crc ] size == 16 (all in v0)
320 const __m128i mask_low = _mm_add_epi8(
321 vramp, _mm_set1_epi8(size - 16));
322 MASK_LH(initial_crc, mask_low, v0, v1);
323
324 if (size2 <= 16) {
325 // There are 1-16 bytes of input and it is all
326 // in data0. Copy the input bytes to v3. If there
327 // are fewer than 16 bytes, the low bytes in v3
328 // will be filled with zeros. That is, the input
329 // bytes are stored to the same position as
330 // (part of) initial_crc is in v0.
331 MASK_L(data0, mask_end, v3);
332 } else {
333 // There are 2-16 bytes of input but not all bytes
334 // are in data0.
335 const __m128i data1 = _mm_load_si128(aligned_buf);
336
337 // Collect the 2-16 input bytes from data0 and data1
338 // to v2 and v3, and bitwise-xor them with the
339 // low bits of initial_crc in v0. Note that the
340 // the second xor is below this else-block as it
341 // is shared with the other branch.
342 MASK_H(data0, mask_end, v2);
343 MASK_L(data1, mask_end, v3);
344 v0 = _mm_xor_si128(v0, v2);
345 }
346
347 v0 = _mm_xor_si128(v0, v3);
348 v1 = _mm_alignr_epi8(v1, v0, 8);
349 } else
350 #endif
351 {
352 const __m128i data1 = _mm_load_si128(aligned_buf);
353 MASK_LH(initial_crc, mask_start, v0, v1);
354 v0 = _mm_xor_si128(v0, data0);
355 v1 = _mm_xor_si128(v1, data1);
356
357 #define FOLD \
358 v1 = _mm_xor_si128(v1, _mm_clmulepi64_si128(v0, vfold1, 0x00)); \
359 v0 = _mm_xor_si128(v1, _mm_clmulepi64_si128(v0, vfold1, 0x11));
360
361 while (size2 > 32) {
362 ++aligned_buf;
363 size2 -= 16;
364 FOLD
365 v1 = _mm_load_si128(aligned_buf);
366 }
367
368 if (size2 < 32) {
369 MASK_H(v0, mask_end, v2);
370 MASK_L(v0, mask_end, v0);
371 MASK_L(v1, mask_end, v3);
372 v1 = _mm_or_si128(v2, v3);
373 }
374
375 FOLD
376 v1 = _mm_srli_si128(v0, 8);
377 #undef FOLD
378 }
379
380 v1 = _mm_xor_si128(_mm_clmulepi64_si128(v0, vfold1, 0x10), v1);
381 v0 = _mm_clmulepi64_si128(v1, vfold0, 0x00);
382 v2 = _mm_clmulepi64_si128(v0, vfold0, 0x10);
383 v0 = _mm_xor_si128(_mm_xor_si128(v2, _mm_slli_si128(v0, 8)), v1);
384
385 #if defined(__i386__) || defined(_M_IX86)
386 return ~(((uint64_t)(uint32_t)_mm_extract_epi32(v0, 3) << 32) |
387 (uint64_t)(uint32_t)_mm_extract_epi32(v0, 2));
388 #else
389 return ~(uint64_t)_mm_extract_epi64(v0, 1);
390 #endif
391
392 #if TUKLIB_GNUC_REQ(4, 6) || defined(__clang__)
393 # pragma GCC diagnostic pop
394 #endif
395 }
396 #if defined(_MSC_VER) && !defined(__INTEL_COMPILER) && !defined(__clang__) \
397 && defined(_M_IX86)
398 # pragma optimize("", on)
399 #endif
400 #endif
401
402
403 ////////////////////////
404 // Detect CPU support //
405 ////////////////////////
406
407 #if defined(CRC_GENERIC) && defined(CRC_CLMUL)
408 static inline bool
409 is_clmul_supported(void)
410 {
411 int success = 1;
412 uint32_t r[4]; // eax, ebx, ecx, edx
413
414 #if defined(_MSC_VER)
415 // This needs <intrin.h> with MSVC. ICC has it as a built-in
416 // on all platforms.
417 __cpuid(r, 1);
418 #elif defined(HAVE_CPUID_H)
419 // Compared to just using __asm__ to run CPUID, this also checks
420 // that CPUID is supported and saves and restores ebx as that is
421 // needed with GCC < 5 with position-independent code (PIC).
422 success = __get_cpuid(1, &r[0], &r[1], &r[2], &r[3]);
423 #else
424 // Just a fallback that shouldn't be needed.
425 __asm__("cpuid\n\t"
426 : "=a"(r[0]), "=b"(r[1]), "=c"(r[2]), "=d"(r[3])
427 : "a"(1), "c"(0));
428 #endif
429
430 // Returns true if these are supported:
431 // CLMUL (bit 1 in ecx)
432 // SSSE3 (bit 9 in ecx)
433 // SSE4.1 (bit 19 in ecx)
434 const uint32_t ecx_mask = (1 << 1) | (1 << 9) | (1 << 19);
435 return success && (r[2] & ecx_mask) == ecx_mask;
436
437 // Alternative methods that weren't used:
438 // - ICC's _may_i_use_cpu_feature: the other methods should work too.
439 // - GCC >= 6 / Clang / ICX __builtin_cpu_supports("pclmul")
440 //
441 // CPUID decding is needed with MSVC anyway and older GCC. This keeps
442 // the feature checks in the build system simpler too. The nice thing
443 // about __builtin_cpu_supports would be that it generates very short
444 // code as is it only reads a variable set at startup but a few bytes
445 // doesn't matter here.
446 }
447
448
449 #ifdef HAVE_FUNC_ATTRIBUTE_CONSTRUCTOR
450 # define CRC64_FUNC_INIT
451 # define CRC64_SET_FUNC_ATTR __attribute__((__constructor__))
452 #else
453 # define CRC64_FUNC_INIT = &crc64_dispatch
454 # define CRC64_SET_FUNC_ATTR
455 static uint64_t crc64_dispatch(const uint8_t *buf, size_t size, uint64_t crc);
456 #endif
457
458
459 // Pointer to the the selected CRC64 method.
460 static uint64_t (*crc64_func)(const uint8_t *buf, size_t size, uint64_t crc)
461 CRC64_FUNC_INIT;
462
463
464 CRC64_SET_FUNC_ATTR
465 static void
466 crc64_set_func(void)
467 {
468 crc64_func = is_clmul_supported() ? &crc64_clmul : &crc64_generic;
469 return;
470 }
471
472
473 #ifndef HAVE_FUNC_ATTRIBUTE_CONSTRUCTOR
474 static uint64_t
475 crc64_dispatch(const uint8_t *buf, size_t size, uint64_t crc)
476 {
477 // When __attribute__((__constructor__)) isn't supported, set the
478 // function pointer without any locking. If multiple threads run
479 // the detection code in parallel, they will all end up setting
480 // the pointer to the same value. This avoids the use of
481 // mythread_once() on every call to lzma_crc64() but this likely
482 // isn't strictly standards compliant. Let's change it if it breaks.
483 crc64_set_func();
484 return crc64_func(buf, size, crc);
485 }
486 #endif
487 #endif
488
489
490 extern LZMA_API(uint64_t)
491 lzma_crc64(const uint8_t *buf, size_t size, uint64_t crc)
492 {
493 #if defined(CRC_GENERIC) && defined(CRC_CLMUL)
494 // If CLMUL is available, it is the best for non-tiny inputs,
495 // being over twice as fast as the generic slice-by-four version.
496 // However, for size <= 16 it's different. In the extreme case
497 // of size == 1 the generic version can be five times faster.
498 // At size >= 8 the CLMUL starts to become reasonable. It
499 // varies depending on the alignment of buf too.
500 //
501 // The above doesn't include the overhead of mythread_once().
502 // At least on x86-64 GNU/Linux, pthread_once() is very fast but
503 // it still makes lzma_crc64(buf, 1, crc) 50-100 % slower. When
504 // size reaches 12-16 bytes the overhead becomes negligible.
505 //
506 // So using the generic version for size <= 16 may give better
507 // performance with tiny inputs but if such inputs happen rarely
508 // it's not so obvious because then the lookup table of the
509 // generic version may not be in the processor cache.
510 #ifdef CRC_USE_GENERIC_FOR_SMALL_INPUTS
511 if (size <= 16)
512 return crc64_generic(buf, size, crc);
513 #endif
514
515 /*
516 #ifndef HAVE_FUNC_ATTRIBUTE_CONSTRUCTOR
517 // See crc64_dispatch(). This would be the alternative which uses
518 // locking and doesn't use crc64_dispatch(). Note that on Windows
519 // this method needs Vista threads.
520 mythread_once(crc64_set_func);
521 #endif
522 */
523
524 return crc64_func(buf, size, crc);
525
526 #elif defined(CRC_CLMUL)
527 // If CLMUL is used unconditionally without runtime CPU detection
528 // then omitting the generic version and its 8 KiB lookup table
529 // makes the library smaller.
530 //
531 // FIXME: Lookup table isn't currently omitted on 32-bit x86,
532 // see crc64_table.c.
533 return crc64_clmul(buf, size, crc);
534
535 #else
536 return crc64_generic(buf, size, crc);
537 #endif
538 }