1// Copyright 2005 Google Inc. All Rights Reserved.
2//
3// Redistribution and use in source and binary forms, with or without
4// modification, are permitted provided that the following conditions are
5// met:
6//
7// * Redistributions of source code must retain the above copyright
8// notice, this list of conditions and the following disclaimer.
9// * Redistributions in binary form must reproduce the above
10// copyright notice, this list of conditions and the following disclaimer
11// in the documentation and/or other materials provided with the
12// distribution.
13// * Neither the name of Google Inc. nor the names of its
14// contributors may be used to endorse or promote products derived from
15// this software without specific prior written permission.
16//
17// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
18// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
19// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
20// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
21// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
22// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
23// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
24// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
27// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28
29#include "snappy.h"
30#include "snappy-internal.h"
31#include "snappy-sinksource.h"
32
33#if !defined(SNAPPY_HAVE_SSSE3)
34// __SSSE3__ is defined by GCC and Clang. Visual Studio doesn't target SIMD
35// support between SSE2 and AVX (so SSSE3 instructions require AVX support), and
36// defines __AVX__ when AVX support is available.
37#if defined(__SSSE3__) || defined(__AVX__)
38#define SNAPPY_HAVE_SSSE3 1
39#else
40#define SNAPPY_HAVE_SSSE3 0
41#endif
42#endif // !defined(SNAPPY_HAVE_SSSE3)
43
44#if !defined(SNAPPY_HAVE_BMI2)
45// __BMI2__ is defined by GCC and Clang. Visual Studio doesn't target BMI2
46// specifically, but it does define __AVX2__ when AVX2 support is available.
47// Fortunately, AVX2 was introduced in Haswell, just like BMI2.
48//
49// BMI2 is not defined as a subset of AVX2 (unlike SSSE3 and AVX above). So,
50// GCC and Clang can build code with AVX2 enabled but BMI2 disabled, in which
51// case issuing BMI2 instructions results in a compiler error.
52#if defined(__BMI2__) || (defined(_MSC_VER) && defined(__AVX2__))
53#define SNAPPY_HAVE_BMI2 1
54#else
55#define SNAPPY_HAVE_BMI2 0
56#endif
57#endif // !defined(SNAPPY_HAVE_BMI2)
58
59#if SNAPPY_HAVE_SSSE3
60// Please do not replace with <x86intrin.h>. or with headers that assume more
61// advanced SSE versions without checking with all the OWNERS.
62#include <tmmintrin.h>
63#endif
64
65#if SNAPPY_HAVE_BMI2
66// Please do not replace with <x86intrin.h>. or with headers that assume more
67// advanced SSE versions without checking with all the OWNERS.
68#include <immintrin.h>
69#endif
70
71#include <stdio.h>
72
73#include <algorithm>
74#include <string>
75#include <vector>
76
77
78namespace snappy {
79
80using internal::COPY_1_BYTE_OFFSET;
81using internal::COPY_2_BYTE_OFFSET;
82using internal::LITERAL;
83using internal::char_table;
84using internal::kMaximumTagLength;
85
86// Any hash function will produce a valid compressed bitstream, but a good
87// hash function reduces the number of collisions and thus yields better
88// compression for compressible input, and more speed for incompressible
89// input. Of course, it doesn't hurt if the hash function is reasonably fast
90// either, as it gets called a lot.
91static inline uint32 HashBytes(uint32 bytes, int shift) {
92 uint32 kMul = 0x1e35a7bd;
93 return (bytes * kMul) >> shift;
94}
95static inline uint32 Hash(const char* p, int shift) {
96 return HashBytes(UNALIGNED_LOAD32(p), shift);
97}
98
99size_t MaxCompressedLength(size_t source_len) {
100 // Compressed data can be defined as:
101 // compressed := item* literal*
102 // item := literal* copy
103 //
104 // The trailing literal sequence has a space blowup of at most 62/60
105 // since a literal of length 60 needs one tag byte + one extra byte
106 // for length information.
107 //
108 // Item blowup is trickier to measure. Suppose the "copy" op copies
109 // 4 bytes of data. Because of a special check in the encoding code,
110 // we produce a 4-byte copy only if the offset is < 65536. Therefore
111 // the copy op takes 3 bytes to encode, and this type of item leads
112 // to at most the 62/60 blowup for representing literals.
113 //
114 // Suppose the "copy" op copies 5 bytes of data. If the offset is big
115 // enough, it will take 5 bytes to encode the copy op. Therefore the
116 // worst case here is a one-byte literal followed by a five-byte copy.
117 // I.e., 6 bytes of input turn into 7 bytes of "compressed" data.
118 //
119 // This last factor dominates the blowup, so the final estimate is:
120 return 32 + source_len + source_len/6;
121}
122
123namespace {
124
125void UnalignedCopy64(const void* src, void* dst) {
126 char tmp[8];
127 memcpy(tmp, src, 8);
128 memcpy(dst, tmp, 8);
129}
130
131void UnalignedCopy128(const void* src, void* dst) {
132 // memcpy gets vectorized when the appropriate compiler options are used.
133 // For example, x86 compilers targeting SSE2+ will optimize to an SSE2 load
134 // and store.
135 char tmp[16];
136 memcpy(tmp, src, 16);
137 memcpy(dst, tmp, 16);
138}
139
140// Copy [src, src+(op_limit-op)) to [op, (op_limit-op)) a byte at a time. Used
141// for handling COPY operations where the input and output regions may overlap.
142// For example, suppose:
143// src == "ab"
144// op == src + 2
145// op_limit == op + 20
146// After IncrementalCopySlow(src, op, op_limit), the result will have eleven
147// copies of "ab"
148// ababababababababababab
149// Note that this does not match the semantics of either memcpy() or memmove().
150inline char* IncrementalCopySlow(const char* src, char* op,
151 char* const op_limit) {
152 // TODO: Remove pragma when LLVM is aware this function is only called in
153 // cold regions and when cold regions don't get vectorized or unrolled.
154#ifdef __clang__
155#pragma clang loop unroll(disable)
156#endif
157 while (op < op_limit) {
158 *op++ = *src++;
159 }
160 return op_limit;
161}
162
163#if SNAPPY_HAVE_SSSE3
164
165// This is a table of shuffle control masks that can be used as the source
166// operand for PSHUFB to permute the contents of the destination XMM register
167// into a repeating byte pattern.
168alignas(16) const char pshufb_fill_patterns[7][16] = {
169 {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
170 {0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1},
171 {0, 1, 2, 0, 1, 2, 0, 1, 2, 0, 1, 2, 0, 1, 2, 0},
172 {0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3},
173 {0, 1, 2, 3, 4, 0, 1, 2, 3, 4, 0, 1, 2, 3, 4, 0},
174 {0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3},
175 {0, 1, 2, 3, 4, 5, 6, 0, 1, 2, 3, 4, 5, 6, 0, 1},
176};
177
178#endif // SNAPPY_HAVE_SSSE3
179
180// Copy [src, src+(op_limit-op)) to [op, (op_limit-op)) but faster than
181// IncrementalCopySlow. buf_limit is the address past the end of the writable
182// region of the buffer.
183inline char* IncrementalCopy(const char* src, char* op, char* const op_limit,
184 char* const buf_limit) {
185 // Terminology:
186 //
187 // slop = buf_limit - op
188 // pat = op - src
189 // len = limit - op
190 assert(src < op);
191 assert(op <= op_limit);
192 assert(op_limit <= buf_limit);
193 // NOTE: The compressor always emits 4 <= len <= 64. It is ok to assume that
194 // to optimize this function but we have to also handle other cases in case
195 // the input does not satisfy these conditions.
196
197 size_t pattern_size = op - src;
198 // The cases are split into different branches to allow the branch predictor,
199 // FDO, and static prediction hints to work better. For each input we list the
200 // ratio of invocations that match each condition.
201 //
202 // input slop < 16 pat < 8 len > 16
203 // ------------------------------------------
204 // html|html4|cp 0% 1.01% 27.73%
205 // urls 0% 0.88% 14.79%
206 // jpg 0% 64.29% 7.14%
207 // pdf 0% 2.56% 58.06%
208 // txt[1-4] 0% 0.23% 0.97%
209 // pb 0% 0.96% 13.88%
210 // bin 0.01% 22.27% 41.17%
211 //
212 // It is very rare that we don't have enough slop for doing block copies. It
213 // is also rare that we need to expand a pattern. Small patterns are common
214 // for incompressible formats and for those we are plenty fast already.
215 // Lengths are normally not greater than 16 but they vary depending on the
216 // input. In general if we always predict len <= 16 it would be an ok
217 // prediction.
218 //
219 // In order to be fast we want a pattern >= 8 bytes and an unrolled loop
220 // copying 2x 8 bytes at a time.
221
222 // Handle the uncommon case where pattern is less than 8 bytes.
223 if (SNAPPY_PREDICT_FALSE(pattern_size < 8)) {
224#if SNAPPY_HAVE_SSSE3
225 // Load the first eight bytes into an 128-bit XMM register, then use PSHUFB
226 // to permute the register's contents in-place into a repeating sequence of
227 // the first "pattern_size" bytes.
228 // For example, suppose:
229 // src == "abc"
230 // op == op + 3
231 // After _mm_shuffle_epi8(), "pattern" will have five copies of "abc"
232 // followed by one byte of slop: abcabcabcabcabca.
233 //
234 // The non-SSE fallback implementation suffers from store-forwarding stalls
235 // because its loads and stores partly overlap. By expanding the pattern
236 // in-place, we avoid the penalty.
237 if (SNAPPY_PREDICT_TRUE(op <= buf_limit - 16)) {
238 const __m128i shuffle_mask = _mm_load_si128(
239 reinterpret_cast<const __m128i*>(pshufb_fill_patterns)
240 + pattern_size - 1);
241 const __m128i pattern = _mm_shuffle_epi8(
242 _mm_loadl_epi64(reinterpret_cast<const __m128i*>(src)), shuffle_mask);
243 // Uninitialized bytes are masked out by the shuffle mask.
244 SNAPPY_ANNOTATE_MEMORY_IS_INITIALIZED(&pattern, sizeof(pattern));
245 pattern_size *= 16 / pattern_size;
246 char* op_end = std::min(op_limit, buf_limit - 15);
247 while (op < op_end) {
248 _mm_storeu_si128(reinterpret_cast<__m128i*>(op), pattern);
249 op += pattern_size;
250 }
251 if (SNAPPY_PREDICT_TRUE(op >= op_limit)) return op_limit;
252 }
253 return IncrementalCopySlow(src, op, op_limit);
254#else // !SNAPPY_HAVE_SSSE3
255 // If plenty of buffer space remains, expand the pattern to at least 8
256 // bytes. The way the following loop is written, we need 8 bytes of buffer
257 // space if pattern_size >= 4, 11 bytes if pattern_size is 1 or 3, and 10
258 // bytes if pattern_size is 2. Precisely encoding that is probably not
259 // worthwhile; instead, invoke the slow path if we cannot write 11 bytes
260 // (because 11 are required in the worst case).
261 if (SNAPPY_PREDICT_TRUE(op <= buf_limit - 11)) {
262 while (pattern_size < 8) {
263 UnalignedCopy64(src, op);
264 op += pattern_size;
265 pattern_size *= 2;
266 }
267 if (SNAPPY_PREDICT_TRUE(op >= op_limit)) return op_limit;
268 } else {
269 return IncrementalCopySlow(src, op, op_limit);
270 }
271#endif // SNAPPY_HAVE_SSSE3
272 }
273 assert(pattern_size >= 8);
274
275 // Copy 2x 8 bytes at a time. Because op - src can be < 16, a single
276 // UnalignedCopy128 might overwrite data in op. UnalignedCopy64 is safe
277 // because expanding the pattern to at least 8 bytes guarantees that
278 // op - src >= 8.
279 //
280 // Typically, the op_limit is the gating factor so try to simplify the loop
281 // based on that.
282 if (SNAPPY_PREDICT_TRUE(op_limit <= buf_limit - 16)) {
283 // Factor the displacement from op to the source into a variable. This helps
284 // simplify the loop below by only varying the op pointer which we need to
285 // test for the end. Note that this was done after carefully examining the
286 // generated code to allow the addressing modes in the loop below to
287 // maximize micro-op fusion where possible on modern Intel processors. The
288 // generated code should be checked carefully for new processors or with
289 // major changes to the compiler.
290 // TODO: Simplify this code when the compiler reliably produces the correct
291 // x86 instruction sequence.
292 ptrdiff_t op_to_src = src - op;
293
294 // The trip count of this loop is not large and so unrolling will only hurt
295 // code size without helping performance.
296 //
297 // TODO: Replace with loop trip count hint.
298#ifdef __clang__
299#pragma clang loop unroll(disable)
300#endif
301 do {
302 UnalignedCopy64(op + op_to_src, op);
303 UnalignedCopy64(op + op_to_src + 8, op + 8);
304 op += 16;
305 } while (op < op_limit);
306 return op_limit;
307 }
308
309 // Fall back to doing as much as we can with the available slop in the
310 // buffer. This code path is relatively cold however so we save code size by
311 // avoiding unrolling and vectorizing.
312 //
313 // TODO: Remove pragma when when cold regions don't get vectorized or
314 // unrolled.
315#ifdef __clang__
316#pragma clang loop unroll(disable)
317#endif
318 for (char *op_end = buf_limit - 16; op < op_end; op += 16, src += 16) {
319 UnalignedCopy64(src, op);
320 UnalignedCopy64(src + 8, op + 8);
321 }
322 if (op >= op_limit)
323 return op_limit;
324
325 // We only take this branch if we didn't have enough slop and we can do a
326 // single 8 byte copy.
327 if (SNAPPY_PREDICT_FALSE(op <= buf_limit - 8)) {
328 UnalignedCopy64(src, op);
329 src += 8;
330 op += 8;
331 }
332 return IncrementalCopySlow(src, op, op_limit);
333}
334
335} // namespace
336
337template <bool allow_fast_path>
338static inline char* EmitLiteral(char* op,
339 const char* literal,
340 int len) {
341 // The vast majority of copies are below 16 bytes, for which a
342 // call to memcpy is overkill. This fast path can sometimes
343 // copy up to 15 bytes too much, but that is okay in the
344 // main loop, since we have a bit to go on for both sides:
345 //
346 // - The input will always have kInputMarginBytes = 15 extra
347 // available bytes, as long as we're in the main loop, and
348 // if not, allow_fast_path = false.
349 // - The output will always have 32 spare bytes (see
350 // MaxCompressedLength).
351 assert(len > 0); // Zero-length literals are disallowed
352 int n = len - 1;
353 if (allow_fast_path && len <= 16) {
354 // Fits in tag byte
355 *op++ = LITERAL | (n << 2);
356
357 UnalignedCopy128(literal, op);
358 return op + len;
359 }
360
361 if (n < 60) {
362 // Fits in tag byte
363 *op++ = LITERAL | (n << 2);
364 } else {
365 // Encode in upcoming bytes
366 char* base = op;
367 int count = 0;
368 op++;
369 while (n > 0) {
370 *op++ = n & 0xff;
371 n >>= 8;
372 count++;
373 }
374 assert(count >= 1);
375 assert(count <= 4);
376 *base = LITERAL | ((59+count) << 2);
377 }
378 memcpy(op, literal, len);
379 return op + len;
380}
381
382template <bool len_less_than_12>
383static inline char* EmitCopyAtMost64(char* op, size_t offset, size_t len) {
384 assert(len <= 64);
385 assert(len >= 4);
386 assert(offset < 65536);
387 assert(len_less_than_12 == (len < 12));
388
389 if (len_less_than_12 && SNAPPY_PREDICT_TRUE(offset < 2048)) {
390 // offset fits in 11 bits. The 3 highest go in the top of the first byte,
391 // and the rest go in the second byte.
392 *op++ = COPY_1_BYTE_OFFSET + ((len - 4) << 2) + ((offset >> 3) & 0xe0);
393 *op++ = offset & 0xff;
394 } else {
395 // Write 4 bytes, though we only care about 3 of them. The output buffer
396 // is required to have some slack, so the extra byte won't overrun it.
397 uint32 u = COPY_2_BYTE_OFFSET + ((len - 1) << 2) + (offset << 8);
398 LittleEndian::Store32(op, u);
399 op += 3;
400 }
401 return op;
402}
403
404template <bool len_less_than_12>
405static inline char* EmitCopy(char* op, size_t offset, size_t len) {
406 assert(len_less_than_12 == (len < 12));
407 if (len_less_than_12) {
408 return EmitCopyAtMost64</*len_less_than_12=*/true>(op, offset, len);
409 } else {
410 // A special case for len <= 64 might help, but so far measurements suggest
411 // it's in the noise.
412
413 // Emit 64 byte copies but make sure to keep at least four bytes reserved.
414 while (SNAPPY_PREDICT_FALSE(len >= 68)) {
415 op = EmitCopyAtMost64</*len_less_than_12=*/false>(op, offset, 64);
416 len -= 64;
417 }
418
419 // One or two copies will now finish the job.
420 if (len > 64) {
421 op = EmitCopyAtMost64</*len_less_than_12=*/false>(op, offset, 60);
422 len -= 60;
423 }
424
425 // Emit remainder.
426 if (len < 12) {
427 op = EmitCopyAtMost64</*len_less_than_12=*/true>(op, offset, len);
428 } else {
429 op = EmitCopyAtMost64</*len_less_than_12=*/false>(op, offset, len);
430 }
431 return op;
432 }
433}
434
435bool GetUncompressedLength(const char* start, size_t n, size_t* result) {
436 uint32 v = 0;
437 const char* limit = start + n;
438 if (Varint::Parse32WithLimit(start, limit, &v) != NULL) {
439 *result = v;
440 return true;
441 } else {
442 return false;
443 }
444}
445
446namespace {
447uint32 CalculateTableSize(uint32 input_size) {
448 assert(kMaxHashTableSize >= 256);
449 if (input_size > kMaxHashTableSize) {
450 return kMaxHashTableSize;
451 }
452 if (input_size < 256) {
453 return 256;
454 }
455 // This is equivalent to Log2Ceiling(input_size), assuming input_size > 1.
456 // 2 << Log2Floor(x - 1) is equivalent to 1 << (1 + Log2Floor(x - 1)).
457 return 2u << Bits::Log2Floor(input_size - 1);
458}
459} // namespace
460
461namespace internal {
462WorkingMemory::WorkingMemory(size_t input_size) {
463 const size_t max_fragment_size = std::min(input_size, kBlockSize);
464 const size_t table_size = CalculateTableSize(max_fragment_size);
465 size_ = table_size * sizeof(*table_) + max_fragment_size +
466 MaxCompressedLength(max_fragment_size);
467 mem_ = std::allocator<char>().allocate(size_);
468 table_ = reinterpret_cast<uint16*>(mem_);
469 input_ = mem_ + table_size * sizeof(*table_);
470 output_ = input_ + max_fragment_size;
471}
472
473WorkingMemory::~WorkingMemory() {
474 std::allocator<char>().deallocate(mem_, size_);
475}
476
477uint16* WorkingMemory::GetHashTable(size_t fragment_size,
478 int* table_size) const {
479 const size_t htsize = CalculateTableSize(fragment_size);
480 memset(table_, 0, htsize * sizeof(*table_));
481 *table_size = htsize;
482 return table_;
483}
484} // end namespace internal
485
486// For 0 <= offset <= 4, GetUint32AtOffset(GetEightBytesAt(p), offset) will
487// equal UNALIGNED_LOAD32(p + offset). Motivation: On x86-64 hardware we have
488// empirically found that overlapping loads such as
489// UNALIGNED_LOAD32(p) ... UNALIGNED_LOAD32(p+1) ... UNALIGNED_LOAD32(p+2)
490// are slower than UNALIGNED_LOAD64(p) followed by shifts and casts to uint32.
491//
492// We have different versions for 64- and 32-bit; ideally we would avoid the
493// two functions and just inline the UNALIGNED_LOAD64 call into
494// GetUint32AtOffset, but GCC (at least not as of 4.6) is seemingly not clever
495// enough to avoid loading the value multiple times then. For 64-bit, the load
496// is done when GetEightBytesAt() is called, whereas for 32-bit, the load is
497// done at GetUint32AtOffset() time.
498
499#ifdef ARCH_K8
500
501typedef uint64 EightBytesReference;
502
503static inline EightBytesReference GetEightBytesAt(const char* ptr) {
504 return UNALIGNED_LOAD64(ptr);
505}
506
507static inline uint32 GetUint32AtOffset(uint64 v, int offset) {
508 assert(offset >= 0);
509 assert(offset <= 4);
510 return v >> (LittleEndian::IsLittleEndian() ? 8 * offset : 32 - 8 * offset);
511}
512
513#else
514
515typedef const char* EightBytesReference;
516
517static inline EightBytesReference GetEightBytesAt(const char* ptr) {
518 return ptr;
519}
520
521static inline uint32 GetUint32AtOffset(const char* v, int offset) {
522 assert(offset >= 0);
523 assert(offset <= 4);
524 return UNALIGNED_LOAD32(v + offset);
525}
526
527#endif
528
529// Flat array compression that does not emit the "uncompressed length"
530// prefix. Compresses "input" string to the "*op" buffer.
531//
532// REQUIRES: "input" is at most "kBlockSize" bytes long.
533// REQUIRES: "op" points to an array of memory that is at least
534// "MaxCompressedLength(input.size())" in size.
535// REQUIRES: All elements in "table[0..table_size-1]" are initialized to zero.
536// REQUIRES: "table_size" is a power of two
537//
538// Returns an "end" pointer into "op" buffer.
539// "end - op" is the compressed size of "input".
540namespace internal {
541char* CompressFragment(const char* input,
542 size_t input_size,
543 char* op,
544 uint16* table,
545 const int table_size) {
546 // "ip" is the input pointer, and "op" is the output pointer.
547 const char* ip = input;
548 assert(input_size <= kBlockSize);
549 assert((table_size & (table_size - 1)) == 0); // table must be power of two
550 const int shift = 32 - Bits::Log2Floor(table_size);
551 assert(static_cast<int>(kuint32max >> shift) == table_size - 1);
552 const char* ip_end = input + input_size;
553 const char* base_ip = ip;
554 // Bytes in [next_emit, ip) will be emitted as literal bytes. Or
555 // [next_emit, ip_end) after the main loop.
556 const char* next_emit = ip;
557
558 const size_t kInputMarginBytes = 15;
559 if (SNAPPY_PREDICT_TRUE(input_size >= kInputMarginBytes)) {
560 const char* ip_limit = input + input_size - kInputMarginBytes;
561
562 for (uint32 next_hash = Hash(++ip, shift); ; ) {
563 assert(next_emit < ip);
564 // The body of this loop calls EmitLiteral once and then EmitCopy one or
565 // more times. (The exception is that when we're close to exhausting
566 // the input we goto emit_remainder.)
567 //
568 // In the first iteration of this loop we're just starting, so
569 // there's nothing to copy, so calling EmitLiteral once is
570 // necessary. And we only start a new iteration when the
571 // current iteration has determined that a call to EmitLiteral will
572 // precede the next call to EmitCopy (if any).
573 //
574 // Step 1: Scan forward in the input looking for a 4-byte-long match.
575 // If we get close to exhausting the input then goto emit_remainder.
576 //
577 // Heuristic match skipping: If 32 bytes are scanned with no matches
578 // found, start looking only at every other byte. If 32 more bytes are
579 // scanned (or skipped), look at every third byte, etc.. When a match is
580 // found, immediately go back to looking at every byte. This is a small
581 // loss (~5% performance, ~0.1% density) for compressible data due to more
582 // bookkeeping, but for non-compressible data (such as JPEG) it's a huge
583 // win since the compressor quickly "realizes" the data is incompressible
584 // and doesn't bother looking for matches everywhere.
585 //
586 // The "skip" variable keeps track of how many bytes there are since the
587 // last match; dividing it by 32 (ie. right-shifting by five) gives the
588 // number of bytes to move ahead for each iteration.
589 uint32 skip = 32;
590
591 const char* next_ip = ip;
592 const char* candidate;
593 do {
594 ip = next_ip;
595 uint32 hash = next_hash;
596 assert(hash == Hash(ip, shift));
597 uint32 bytes_between_hash_lookups = skip >> 5;
598 skip += bytes_between_hash_lookups;
599 next_ip = ip + bytes_between_hash_lookups;
600 if (SNAPPY_PREDICT_FALSE(next_ip > ip_limit)) {
601 goto emit_remainder;
602 }
603 next_hash = Hash(next_ip, shift);
604 candidate = base_ip + table[hash];
605 assert(candidate >= base_ip);
606 assert(candidate < ip);
607
608 table[hash] = ip - base_ip;
609 } while (SNAPPY_PREDICT_TRUE(UNALIGNED_LOAD32(ip) !=
610 UNALIGNED_LOAD32(candidate)));
611
612 // Step 2: A 4-byte match has been found. We'll later see if more
613 // than 4 bytes match. But, prior to the match, input
614 // bytes [next_emit, ip) are unmatched. Emit them as "literal bytes."
615 assert(next_emit + 16 <= ip_end);
616 op = EmitLiteral</*allow_fast_path=*/true>(op, next_emit, ip - next_emit);
617
618 // Step 3: Call EmitCopy, and then see if another EmitCopy could
619 // be our next move. Repeat until we find no match for the
620 // input immediately after what was consumed by the last EmitCopy call.
621 //
622 // If we exit this loop normally then we need to call EmitLiteral next,
623 // though we don't yet know how big the literal will be. We handle that
624 // by proceeding to the next iteration of the main loop. We also can exit
625 // this loop via goto if we get close to exhausting the input.
626 EightBytesReference input_bytes;
627 uint32 candidate_bytes = 0;
628
629 do {
630 // We have a 4-byte match at ip, and no need to emit any
631 // "literal bytes" prior to ip.
632 const char* base = ip;
633 std::pair<size_t, bool> p =
634 FindMatchLength(candidate + 4, ip + 4, ip_end);
635 size_t matched = 4 + p.first;
636 ip += matched;
637 size_t offset = base - candidate;
638 assert(0 == memcmp(base, candidate, matched));
639 if (p.second) {
640 op = EmitCopy</*len_less_than_12=*/true>(op, offset, matched);
641 } else {
642 op = EmitCopy</*len_less_than_12=*/false>(op, offset, matched);
643 }
644 next_emit = ip;
645 if (SNAPPY_PREDICT_FALSE(ip >= ip_limit)) {
646 goto emit_remainder;
647 }
648 // We are now looking for a 4-byte match again. We read
649 // table[Hash(ip, shift)] for that. To improve compression,
650 // we also update table[Hash(ip - 1, shift)] and table[Hash(ip, shift)].
651 input_bytes = GetEightBytesAt(ip - 1);
652 uint32 prev_hash = HashBytes(GetUint32AtOffset(input_bytes, 0), shift);
653 table[prev_hash] = ip - base_ip - 1;
654 uint32 cur_hash = HashBytes(GetUint32AtOffset(input_bytes, 1), shift);
655 candidate = base_ip + table[cur_hash];
656 candidate_bytes = UNALIGNED_LOAD32(candidate);
657 table[cur_hash] = ip - base_ip;
658 } while (GetUint32AtOffset(input_bytes, 1) == candidate_bytes);
659
660 next_hash = HashBytes(GetUint32AtOffset(input_bytes, 2), shift);
661 ++ip;
662 }
663 }
664
665 emit_remainder:
666 // Emit the remaining bytes as a literal
667 if (next_emit < ip_end) {
668 op = EmitLiteral</*allow_fast_path=*/false>(op, next_emit,
669 ip_end - next_emit);
670 }
671
672 return op;
673}
674} // end namespace internal
675
676// Called back at avery compression call to trace parameters and sizes.
677static inline void Report(const char *algorithm, size_t compressed_size,
678 size_t uncompressed_size) {}
679
680// Signature of output types needed by decompression code.
681// The decompression code is templatized on a type that obeys this
682// signature so that we do not pay virtual function call overhead in
683// the middle of a tight decompression loop.
684//
685// class DecompressionWriter {
686// public:
687// // Called before decompression
688// void SetExpectedLength(size_t length);
689//
690// // Called after decompression
691// bool CheckLength() const;
692//
693// // Called repeatedly during decompression
694// bool Append(const char* ip, size_t length);
695// bool AppendFromSelf(uint32 offset, size_t length);
696//
697// // The rules for how TryFastAppend differs from Append are somewhat
698// // convoluted:
699// //
700// // - TryFastAppend is allowed to decline (return false) at any
701// // time, for any reason -- just "return false" would be
702// // a perfectly legal implementation of TryFastAppend.
703// // The intention is for TryFastAppend to allow a fast path
704// // in the common case of a small append.
705// // - TryFastAppend is allowed to read up to <available> bytes
706// // from the input buffer, whereas Append is allowed to read
707// // <length>. However, if it returns true, it must leave
708// // at least five (kMaximumTagLength) bytes in the input buffer
709// // afterwards, so that there is always enough space to read the
710// // next tag without checking for a refill.
711// // - TryFastAppend must always return decline (return false)
712// // if <length> is 61 or more, as in this case the literal length is not
713// // decoded fully. In practice, this should not be a big problem,
714// // as it is unlikely that one would implement a fast path accepting
715// // this much data.
716// //
717// bool TryFastAppend(const char* ip, size_t available, size_t length);
718// };
719
720static inline uint32 ExtractLowBytes(uint32 v, int n) {
721 assert(n >= 0);
722 assert(n <= 4);
723 // TODO(b/121042345): Remove !defined(MEMORY_SANITIZER) once MSan
724 // handles _bzhi_u32() correctly.
725#if SNAPPY_HAVE_BMI2 && !defined(MEMORY_SANITIZER)
726 return _bzhi_u32(v, 8 * n);
727#else
728 // This needs to be wider than uint32 otherwise `mask << 32` will be
729 // undefined.
730 uint64 mask = 0xffffffff;
731 return v & ~(mask << (8 * n));
732#endif
733}
734
735static inline bool LeftShiftOverflows(uint8 value, uint32 shift) {
736 assert(shift < 32);
737 static const uint8 masks[] = {
738 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, //
739 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, //
740 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, //
741 0x00, 0x80, 0xc0, 0xe0, 0xf0, 0xf8, 0xfc, 0xfe};
742 return (value & masks[shift]) != 0;
743}
744
745// Helper class for decompression
746class SnappyDecompressor {
747 private:
748 Source* reader_; // Underlying source of bytes to decompress
749 const char* ip_; // Points to next buffered byte
750 const char* ip_limit_; // Points just past buffered bytes
751 uint32 peeked_; // Bytes peeked from reader (need to skip)
752 bool eof_; // Hit end of input without an error?
753 char scratch_[kMaximumTagLength]; // See RefillTag().
754
755 // Ensure that all of the tag metadata for the next tag is available
756 // in [ip_..ip_limit_-1]. Also ensures that [ip,ip+4] is readable even
757 // if (ip_limit_ - ip_ < 5).
758 //
759 // Returns true on success, false on error or end of input.
760 bool RefillTag();
761
762 public:
763 explicit SnappyDecompressor(Source* reader)
764 : reader_(reader),
765 ip_(NULL),
766 ip_limit_(NULL),
767 peeked_(0),
768 eof_(false) {
769 }
770
771 ~SnappyDecompressor() {
772 // Advance past any bytes we peeked at from the reader
773 reader_->Skip(peeked_);
774 }
775
776 // Returns true iff we have hit the end of the input without an error.
777 bool eof() const {
778 return eof_;
779 }
780
781 // Read the uncompressed length stored at the start of the compressed data.
782 // On success, stores the length in *result and returns true.
783 // On failure, returns false.
784 bool ReadUncompressedLength(uint32* result) {
785 assert(ip_ == NULL); // Must not have read anything yet
786 // Length is encoded in 1..5 bytes
787 *result = 0;
788 uint32 shift = 0;
789 while (true) {
790 if (shift >= 32) return false;
791 size_t n;
792 const char* ip = reader_->Peek(&n);
793 if (n == 0) return false;
794 const unsigned char c = *(reinterpret_cast<const unsigned char*>(ip));
795 reader_->Skip(1);
796 uint32 val = c & 0x7f;
797 if (LeftShiftOverflows(static_cast<uint8>(val), shift)) return false;
798 *result |= val << shift;
799 if (c < 128) {
800 break;
801 }
802 shift += 7;
803 }
804 return true;
805 }
806
807 // Process the next item found in the input.
808 // Returns true if successful, false on error or end of input.
809 template <class Writer>
810#if defined(__GNUC__) && defined(__x86_64__)
811 __attribute__((aligned(32)))
812#endif
813 void DecompressAllTags(Writer* writer) {
814 // In x86, pad the function body to start 16 bytes later. This function has
815 // a couple of hotspots that are highly sensitive to alignment: we have
816 // observed regressions by more than 20% in some metrics just by moving the
817 // exact same code to a different position in the benchmark binary.
818 //
819 // Putting this code on a 32-byte-aligned boundary + 16 bytes makes us hit
820 // the "lucky" case consistently. Unfortunately, this is a very brittle
821 // workaround, and future differences in code generation may reintroduce
822 // this regression. If you experience a big, difficult to explain, benchmark
823 // performance regression here, first try removing this hack.
824#if defined(__GNUC__) && defined(__x86_64__)
825 // Two 8-byte "NOP DWORD ptr [EAX + EAX*1 + 00000000H]" instructions.
826 asm(".byte 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00");
827 asm(".byte 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00");
828#endif
829
830 const char* ip = ip_;
831 // We could have put this refill fragment only at the beginning of the loop.
832 // However, duplicating it at the end of each branch gives the compiler more
833 // scope to optimize the <ip_limit_ - ip> expression based on the local
834 // context, which overall increases speed.
835 #define MAYBE_REFILL() \
836 if (ip_limit_ - ip < kMaximumTagLength) { \
837 ip_ = ip; \
838 if (!RefillTag()) return; \
839 ip = ip_; \
840 }
841
842 MAYBE_REFILL();
843 for ( ;; ) {
844 const unsigned char c = *(reinterpret_cast<const unsigned char*>(ip++));
845
846 // Ratio of iterations that have LITERAL vs non-LITERAL for different
847 // inputs.
848 //
849 // input LITERAL NON_LITERAL
850 // -----------------------------------
851 // html|html4|cp 23% 77%
852 // urls 36% 64%
853 // jpg 47% 53%
854 // pdf 19% 81%
855 // txt[1-4] 25% 75%
856 // pb 24% 76%
857 // bin 24% 76%
858 if (SNAPPY_PREDICT_FALSE((c & 0x3) == LITERAL)) {
859 size_t literal_length = (c >> 2) + 1u;
860 if (writer->TryFastAppend(ip, ip_limit_ - ip, literal_length)) {
861 assert(literal_length < 61);
862 ip += literal_length;
863 // NOTE(user): There is no MAYBE_REFILL() here, as TryFastAppend()
864 // will not return true unless there's already at least five spare
865 // bytes in addition to the literal.
866 continue;
867 }
868 if (SNAPPY_PREDICT_FALSE(literal_length >= 61)) {
869 // Long literal.
870 const size_t literal_length_length = literal_length - 60;
871 literal_length =
872 ExtractLowBytes(LittleEndian::Load32(ip), literal_length_length) +
873 1;
874 ip += literal_length_length;
875 }
876
877 size_t avail = ip_limit_ - ip;
878 while (avail < literal_length) {
879 if (!writer->Append(ip, avail)) return;
880 literal_length -= avail;
881 reader_->Skip(peeked_);
882 size_t n;
883 ip = reader_->Peek(&n);
884 avail = n;
885 peeked_ = avail;
886 if (avail == 0) return; // Premature end of input
887 ip_limit_ = ip + avail;
888 }
889 if (!writer->Append(ip, literal_length)) {
890 return;
891 }
892 ip += literal_length;
893 MAYBE_REFILL();
894 } else {
895 const size_t entry = char_table[c];
896 const size_t trailer =
897 ExtractLowBytes(LittleEndian::Load32(ip), entry >> 11);
898 const size_t length = entry & 0xff;
899 ip += entry >> 11;
900
901 // copy_offset/256 is encoded in bits 8..10. By just fetching
902 // those bits, we get copy_offset (since the bit-field starts at
903 // bit 8).
904 const size_t copy_offset = entry & 0x700;
905 if (!writer->AppendFromSelf(copy_offset + trailer, length)) {
906 return;
907 }
908 MAYBE_REFILL();
909 }
910 }
911
912#undef MAYBE_REFILL
913 }
914};
915
916bool SnappyDecompressor::RefillTag() {
917 const char* ip = ip_;
918 if (ip == ip_limit_) {
919 // Fetch a new fragment from the reader
920 reader_->Skip(peeked_); // All peeked bytes are used up
921 size_t n;
922 ip = reader_->Peek(&n);
923 peeked_ = n;
924 eof_ = (n == 0);
925 if (eof_) return false;
926 ip_limit_ = ip + n;
927 }
928
929 // Read the tag character
930 assert(ip < ip_limit_);
931 const unsigned char c = *(reinterpret_cast<const unsigned char*>(ip));
932 const uint32 entry = char_table[c];
933 const uint32 needed = (entry >> 11) + 1; // +1 byte for 'c'
934 assert(needed <= sizeof(scratch_));
935
936 // Read more bytes from reader if needed
937 uint32 nbuf = ip_limit_ - ip;
938 if (nbuf < needed) {
939 // Stitch together bytes from ip and reader to form the word
940 // contents. We store the needed bytes in "scratch_". They
941 // will be consumed immediately by the caller since we do not
942 // read more than we need.
943 memmove(scratch_, ip, nbuf);
944 reader_->Skip(peeked_); // All peeked bytes are used up
945 peeked_ = 0;
946 while (nbuf < needed) {
947 size_t length;
948 const char* src = reader_->Peek(&length);
949 if (length == 0) return false;
950 uint32 to_add = std::min<uint32>(needed - nbuf, length);
951 memcpy(scratch_ + nbuf, src, to_add);
952 nbuf += to_add;
953 reader_->Skip(to_add);
954 }
955 assert(nbuf == needed);
956 ip_ = scratch_;
957 ip_limit_ = scratch_ + needed;
958 } else if (nbuf < kMaximumTagLength) {
959 // Have enough bytes, but move into scratch_ so that we do not
960 // read past end of input
961 memmove(scratch_, ip, nbuf);
962 reader_->Skip(peeked_); // All peeked bytes are used up
963 peeked_ = 0;
964 ip_ = scratch_;
965 ip_limit_ = scratch_ + nbuf;
966 } else {
967 // Pass pointer to buffer returned by reader_.
968 ip_ = ip;
969 }
970 return true;
971}
972
973template <typename Writer>
974static bool InternalUncompress(Source* r, Writer* writer) {
975 // Read the uncompressed length from the front of the compressed input
976 SnappyDecompressor decompressor(r);
977 uint32 uncompressed_len = 0;
978 if (!decompressor.ReadUncompressedLength(&uncompressed_len)) return false;
979
980 return InternalUncompressAllTags(&decompressor, writer, r->Available(),
981 uncompressed_len);
982}
983
984template <typename Writer>
985static bool InternalUncompressAllTags(SnappyDecompressor* decompressor,
986 Writer* writer,
987 uint32 compressed_len,
988 uint32 uncompressed_len) {
989 Report("snappy_uncompress", compressed_len, uncompressed_len);
990
991 writer->SetExpectedLength(uncompressed_len);
992
993 // Process the entire input
994 decompressor->DecompressAllTags(writer);
995 writer->Flush();
996 return (decompressor->eof() && writer->CheckLength());
997}
998
999bool GetUncompressedLength(Source* source, uint32* result) {
1000 SnappyDecompressor decompressor(source);
1001 return decompressor.ReadUncompressedLength(result);
1002}
1003
1004size_t Compress(Source* reader, Sink* writer) {
1005 size_t written = 0;
1006 size_t N = reader->Available();
1007 const size_t uncompressed_size = N;
1008 char ulength[Varint::kMax32];
1009 char* p = Varint::Encode32(ulength, N);
1010 writer->Append(ulength, p-ulength);
1011 written += (p - ulength);
1012
1013 internal::WorkingMemory wmem(N);
1014
1015 while (N > 0) {
1016 // Get next block to compress (without copying if possible)
1017 size_t fragment_size;
1018 const char* fragment = reader->Peek(&fragment_size);
1019 assert(fragment_size != 0); // premature end of input
1020 const size_t num_to_read = std::min(N, kBlockSize);
1021 size_t bytes_read = fragment_size;
1022
1023 size_t pending_advance = 0;
1024 if (bytes_read >= num_to_read) {
1025 // Buffer returned by reader is large enough
1026 pending_advance = num_to_read;
1027 fragment_size = num_to_read;
1028 } else {
1029 char* scratch = wmem.GetScratchInput();
1030 memcpy(scratch, fragment, bytes_read);
1031 reader->Skip(bytes_read);
1032
1033 while (bytes_read < num_to_read) {
1034 fragment = reader->Peek(&fragment_size);
1035 size_t n = std::min<size_t>(fragment_size, num_to_read - bytes_read);
1036 memcpy(scratch + bytes_read, fragment, n);
1037 bytes_read += n;
1038 reader->Skip(n);
1039 }
1040 assert(bytes_read == num_to_read);
1041 fragment = scratch;
1042 fragment_size = num_to_read;
1043 }
1044 assert(fragment_size == num_to_read);
1045
1046 // Get encoding table for compression
1047 int table_size;
1048 uint16* table = wmem.GetHashTable(num_to_read, &table_size);
1049
1050 // Compress input_fragment and append to dest
1051 const int max_output = MaxCompressedLength(num_to_read);
1052
1053 // Need a scratch buffer for the output, in case the byte sink doesn't
1054 // have room for us directly.
1055
1056 // Since we encode kBlockSize regions followed by a region
1057 // which is <= kBlockSize in length, a previously allocated
1058 // scratch_output[] region is big enough for this iteration.
1059 char* dest = writer->GetAppendBuffer(max_output, wmem.GetScratchOutput());
1060 char* end = internal::CompressFragment(fragment, fragment_size, dest, table,
1061 table_size);
1062 writer->Append(dest, end - dest);
1063 written += (end - dest);
1064
1065 N -= num_to_read;
1066 reader->Skip(pending_advance);
1067 }
1068
1069 Report("snappy_compress", written, uncompressed_size);
1070
1071 return written;
1072}
1073
1074// -----------------------------------------------------------------------
1075// IOVec interfaces
1076// -----------------------------------------------------------------------
1077
1078// A type that writes to an iovec.
1079// Note that this is not a "ByteSink", but a type that matches the
1080// Writer template argument to SnappyDecompressor::DecompressAllTags().
1081class SnappyIOVecWriter {
1082 private:
1083 // output_iov_end_ is set to iov + count and used to determine when
1084 // the end of the iovs is reached.
1085 const struct iovec* output_iov_end_;
1086
1087#if !defined(NDEBUG)
1088 const struct iovec* output_iov_;
1089#endif // !defined(NDEBUG)
1090
1091 // Current iov that is being written into.
1092 const struct iovec* curr_iov_;
1093
1094 // Pointer to current iov's write location.
1095 char* curr_iov_output_;
1096
1097 // Remaining bytes to write into curr_iov_output.
1098 size_t curr_iov_remaining_;
1099
1100 // Total bytes decompressed into output_iov_ so far.
1101 size_t total_written_;
1102
1103 // Maximum number of bytes that will be decompressed into output_iov_.
1104 size_t output_limit_;
1105
1106 static inline char* GetIOVecPointer(const struct iovec* iov, size_t offset) {
1107 return reinterpret_cast<char*>(iov->iov_base) + offset;
1108 }
1109
1110 public:
1111 // Does not take ownership of iov. iov must be valid during the
1112 // entire lifetime of the SnappyIOVecWriter.
1113 inline SnappyIOVecWriter(const struct iovec* iov, size_t iov_count)
1114 : output_iov_end_(iov + iov_count),
1115#if !defined(NDEBUG)
1116 output_iov_(iov),
1117#endif // !defined(NDEBUG)
1118 curr_iov_(iov),
1119 curr_iov_output_(iov_count ? reinterpret_cast<char*>(iov->iov_base)
1120 : nullptr),
1121 curr_iov_remaining_(iov_count ? iov->iov_len : 0),
1122 total_written_(0),
1123 output_limit_(-1) {}
1124
1125 inline void SetExpectedLength(size_t len) {
1126 output_limit_ = len;
1127 }
1128
1129 inline bool CheckLength() const {
1130 return total_written_ == output_limit_;
1131 }
1132
1133 inline bool Append(const char* ip, size_t len) {
1134 if (total_written_ + len > output_limit_) {
1135 return false;
1136 }
1137
1138 return AppendNoCheck(ip, len);
1139 }
1140
1141 inline bool AppendNoCheck(const char* ip, size_t len) {
1142 while (len > 0) {
1143 if (curr_iov_remaining_ == 0) {
1144 // This iovec is full. Go to the next one.
1145 if (curr_iov_ + 1 >= output_iov_end_) {
1146 return false;
1147 }
1148 ++curr_iov_;
1149 curr_iov_output_ = reinterpret_cast<char*>(curr_iov_->iov_base);
1150 curr_iov_remaining_ = curr_iov_->iov_len;
1151 }
1152
1153 const size_t to_write = std::min(len, curr_iov_remaining_);
1154 memcpy(curr_iov_output_, ip, to_write);
1155 curr_iov_output_ += to_write;
1156 curr_iov_remaining_ -= to_write;
1157 total_written_ += to_write;
1158 ip += to_write;
1159 len -= to_write;
1160 }
1161
1162 return true;
1163 }
1164
1165 inline bool TryFastAppend(const char* ip, size_t available, size_t len) {
1166 const size_t space_left = output_limit_ - total_written_;
1167 if (len <= 16 && available >= 16 + kMaximumTagLength && space_left >= 16 &&
1168 curr_iov_remaining_ >= 16) {
1169 // Fast path, used for the majority (about 95%) of invocations.
1170 UnalignedCopy128(ip, curr_iov_output_);
1171 curr_iov_output_ += len;
1172 curr_iov_remaining_ -= len;
1173 total_written_ += len;
1174 return true;
1175 }
1176
1177 return false;
1178 }
1179
1180 inline bool AppendFromSelf(size_t offset, size_t len) {
1181 // See SnappyArrayWriter::AppendFromSelf for an explanation of
1182 // the "offset - 1u" trick.
1183 if (offset - 1u >= total_written_) {
1184 return false;
1185 }
1186 const size_t space_left = output_limit_ - total_written_;
1187 if (len > space_left) {
1188 return false;
1189 }
1190
1191 // Locate the iovec from which we need to start the copy.
1192 const iovec* from_iov = curr_iov_;
1193 size_t from_iov_offset = curr_iov_->iov_len - curr_iov_remaining_;
1194 while (offset > 0) {
1195 if (from_iov_offset >= offset) {
1196 from_iov_offset -= offset;
1197 break;
1198 }
1199
1200 offset -= from_iov_offset;
1201 --from_iov;
1202#if !defined(NDEBUG)
1203 assert(from_iov >= output_iov_);
1204#endif // !defined(NDEBUG)
1205 from_iov_offset = from_iov->iov_len;
1206 }
1207
1208 // Copy <len> bytes starting from the iovec pointed to by from_iov_index to
1209 // the current iovec.
1210 while (len > 0) {
1211 assert(from_iov <= curr_iov_);
1212 if (from_iov != curr_iov_) {
1213 const size_t to_copy =
1214 std::min(from_iov->iov_len - from_iov_offset, len);
1215 AppendNoCheck(GetIOVecPointer(from_iov, from_iov_offset), to_copy);
1216 len -= to_copy;
1217 if (len > 0) {
1218 ++from_iov;
1219 from_iov_offset = 0;
1220 }
1221 } else {
1222 size_t to_copy = curr_iov_remaining_;
1223 if (to_copy == 0) {
1224 // This iovec is full. Go to the next one.
1225 if (curr_iov_ + 1 >= output_iov_end_) {
1226 return false;
1227 }
1228 ++curr_iov_;
1229 curr_iov_output_ = reinterpret_cast<char*>(curr_iov_->iov_base);
1230 curr_iov_remaining_ = curr_iov_->iov_len;
1231 continue;
1232 }
1233 if (to_copy > len) {
1234 to_copy = len;
1235 }
1236
1237 IncrementalCopy(GetIOVecPointer(from_iov, from_iov_offset),
1238 curr_iov_output_, curr_iov_output_ + to_copy,
1239 curr_iov_output_ + curr_iov_remaining_);
1240 curr_iov_output_ += to_copy;
1241 curr_iov_remaining_ -= to_copy;
1242 from_iov_offset += to_copy;
1243 total_written_ += to_copy;
1244 len -= to_copy;
1245 }
1246 }
1247
1248 return true;
1249 }
1250
1251 inline void Flush() {}
1252};
1253
1254bool RawUncompressToIOVec(const char* compressed, size_t compressed_length,
1255 const struct iovec* iov, size_t iov_cnt) {
1256 ByteArraySource reader(compressed, compressed_length);
1257 return RawUncompressToIOVec(&reader, iov, iov_cnt);
1258}
1259
1260bool RawUncompressToIOVec(Source* compressed, const struct iovec* iov,
1261 size_t iov_cnt) {
1262 SnappyIOVecWriter output(iov, iov_cnt);
1263 return InternalUncompress(compressed, &output);
1264}
1265
1266// -----------------------------------------------------------------------
1267// Flat array interfaces
1268// -----------------------------------------------------------------------
1269
1270// A type that writes to a flat array.
1271// Note that this is not a "ByteSink", but a type that matches the
1272// Writer template argument to SnappyDecompressor::DecompressAllTags().
1273class SnappyArrayWriter {
1274 private:
1275 char* base_;
1276 char* op_;
1277 char* op_limit_;
1278
1279 public:
1280 inline explicit SnappyArrayWriter(char* dst)
1281 : base_(dst),
1282 op_(dst),
1283 op_limit_(dst) {
1284 }
1285
1286 inline void SetExpectedLength(size_t len) {
1287 op_limit_ = op_ + len;
1288 }
1289
1290 inline bool CheckLength() const {
1291 return op_ == op_limit_;
1292 }
1293
1294 inline bool Append(const char* ip, size_t len) {
1295 char* op = op_;
1296 const size_t space_left = op_limit_ - op;
1297 if (space_left < len) {
1298 return false;
1299 }
1300 memcpy(op, ip, len);
1301 op_ = op + len;
1302 return true;
1303 }
1304
1305 inline bool TryFastAppend(const char* ip, size_t available, size_t len) {
1306 char* op = op_;
1307 const size_t space_left = op_limit_ - op;
1308 if (len <= 16 && available >= 16 + kMaximumTagLength && space_left >= 16) {
1309 // Fast path, used for the majority (about 95%) of invocations.
1310 UnalignedCopy128(ip, op);
1311 op_ = op + len;
1312 return true;
1313 } else {
1314 return false;
1315 }
1316 }
1317
1318 inline bool AppendFromSelf(size_t offset, size_t len) {
1319 char* const op_end = op_ + len;
1320
1321 // Check if we try to append from before the start of the buffer.
1322 // Normally this would just be a check for "produced < offset",
1323 // but "produced <= offset - 1u" is equivalent for every case
1324 // except the one where offset==0, where the right side will wrap around
1325 // to a very big number. This is convenient, as offset==0 is another
1326 // invalid case that we also want to catch, so that we do not go
1327 // into an infinite loop.
1328 if (Produced() <= offset - 1u || op_end > op_limit_) return false;
1329 op_ = IncrementalCopy(op_ - offset, op_, op_end, op_limit_);
1330
1331 return true;
1332 }
1333 inline size_t Produced() const {
1334 assert(op_ >= base_);
1335 return op_ - base_;
1336 }
1337 inline void Flush() {}
1338};
1339
1340bool RawUncompress(const char* compressed, size_t n, char* uncompressed) {
1341 ByteArraySource reader(compressed, n);
1342 return RawUncompress(&reader, uncompressed);
1343}
1344
1345bool RawUncompress(Source* compressed, char* uncompressed) {
1346 SnappyArrayWriter output(uncompressed);
1347 return InternalUncompress(compressed, &output);
1348}
1349
1350bool Uncompress(const char* compressed, size_t n, string* uncompressed) {
1351 size_t ulength;
1352 if (!GetUncompressedLength(compressed, n, &ulength)) {
1353 return false;
1354 }
1355 // On 32-bit builds: max_size() < kuint32max. Check for that instead
1356 // of crashing (e.g., consider externally specified compressed data).
1357 if (ulength > uncompressed->max_size()) {
1358 return false;
1359 }
1360 STLStringResizeUninitialized(uncompressed, ulength);
1361 return RawUncompress(compressed, n, string_as_array(uncompressed));
1362}
1363
1364// A Writer that drops everything on the floor and just does validation
1365class SnappyDecompressionValidator {
1366 private:
1367 size_t expected_;
1368 size_t produced_;
1369
1370 public:
1371 inline SnappyDecompressionValidator() : expected_(0), produced_(0) { }
1372 inline void SetExpectedLength(size_t len) {
1373 expected_ = len;
1374 }
1375 inline bool CheckLength() const {
1376 return expected_ == produced_;
1377 }
1378 inline bool Append(const char* ip, size_t len) {
1379 produced_ += len;
1380 return produced_ <= expected_;
1381 }
1382 inline bool TryFastAppend(const char* ip, size_t available, size_t length) {
1383 return false;
1384 }
1385 inline bool AppendFromSelf(size_t offset, size_t len) {
1386 // See SnappyArrayWriter::AppendFromSelf for an explanation of
1387 // the "offset - 1u" trick.
1388 if (produced_ <= offset - 1u) return false;
1389 produced_ += len;
1390 return produced_ <= expected_;
1391 }
1392 inline void Flush() {}
1393};
1394
1395bool IsValidCompressedBuffer(const char* compressed, size_t n) {
1396 ByteArraySource reader(compressed, n);
1397 SnappyDecompressionValidator writer;
1398 return InternalUncompress(&reader, &writer);
1399}
1400
1401bool IsValidCompressed(Source* compressed) {
1402 SnappyDecompressionValidator writer;
1403 return InternalUncompress(compressed, &writer);
1404}
1405
1406void RawCompress(const char* input,
1407 size_t input_length,
1408 char* compressed,
1409 size_t* compressed_length) {
1410 ByteArraySource reader(input, input_length);
1411 UncheckedByteArraySink writer(compressed);
1412 Compress(&reader, &writer);
1413
1414 // Compute how many bytes were added
1415 *compressed_length = (writer.CurrentDestination() - compressed);
1416}
1417
1418size_t Compress(const char* input, size_t input_length, string* compressed) {
1419 // Pre-grow the buffer to the max length of the compressed output
1420 STLStringResizeUninitialized(compressed, MaxCompressedLength(input_length));
1421
1422 size_t compressed_length;
1423 RawCompress(input, input_length, string_as_array(compressed),
1424 &compressed_length);
1425 compressed->resize(compressed_length);
1426 return compressed_length;
1427}
1428
1429// -----------------------------------------------------------------------
1430// Sink interface
1431// -----------------------------------------------------------------------
1432
1433// A type that decompresses into a Sink. The template parameter
1434// Allocator must export one method "char* Allocate(int size);", which
1435// allocates a buffer of "size" and appends that to the destination.
1436template <typename Allocator>
1437class SnappyScatteredWriter {
1438 Allocator allocator_;
1439
1440 // We need random access into the data generated so far. Therefore
1441 // we keep track of all of the generated data as an array of blocks.
1442 // All of the blocks except the last have length kBlockSize.
1443 std::vector<char*> blocks_;
1444 size_t expected_;
1445
1446 // Total size of all fully generated blocks so far
1447 size_t full_size_;
1448
1449 // Pointer into current output block
1450 char* op_base_; // Base of output block
1451 char* op_ptr_; // Pointer to next unfilled byte in block
1452 char* op_limit_; // Pointer just past block
1453
1454 inline size_t Size() const {
1455 return full_size_ + (op_ptr_ - op_base_);
1456 }
1457
1458 bool SlowAppend(const char* ip, size_t len);
1459 bool SlowAppendFromSelf(size_t offset, size_t len);
1460
1461 public:
1462 inline explicit SnappyScatteredWriter(const Allocator& allocator)
1463 : allocator_(allocator),
1464 full_size_(0),
1465 op_base_(NULL),
1466 op_ptr_(NULL),
1467 op_limit_(NULL) {
1468 }
1469
1470 inline void SetExpectedLength(size_t len) {
1471 assert(blocks_.empty());
1472 expected_ = len;
1473 }
1474
1475 inline bool CheckLength() const {
1476 return Size() == expected_;
1477 }
1478
1479 // Return the number of bytes actually uncompressed so far
1480 inline size_t Produced() const {
1481 return Size();
1482 }
1483
1484 inline bool Append(const char* ip, size_t len) {
1485 size_t avail = op_limit_ - op_ptr_;
1486 if (len <= avail) {
1487 // Fast path
1488 memcpy(op_ptr_, ip, len);
1489 op_ptr_ += len;
1490 return true;
1491 } else {
1492 return SlowAppend(ip, len);
1493 }
1494 }
1495
1496 inline bool TryFastAppend(const char* ip, size_t available, size_t length) {
1497 char* op = op_ptr_;
1498 const int space_left = op_limit_ - op;
1499 if (length <= 16 && available >= 16 + kMaximumTagLength &&
1500 space_left >= 16) {
1501 // Fast path, used for the majority (about 95%) of invocations.
1502 UnalignedCopy128(ip, op);
1503 op_ptr_ = op + length;
1504 return true;
1505 } else {
1506 return false;
1507 }
1508 }
1509
1510 inline bool AppendFromSelf(size_t offset, size_t len) {
1511 char* const op_end = op_ptr_ + len;
1512 // See SnappyArrayWriter::AppendFromSelf for an explanation of
1513 // the "offset - 1u" trick.
1514 if (SNAPPY_PREDICT_TRUE(offset - 1u < op_ptr_ - op_base_ &&
1515 op_end <= op_limit_)) {
1516 // Fast path: src and dst in current block.
1517 op_ptr_ = IncrementalCopy(op_ptr_ - offset, op_ptr_, op_end, op_limit_);
1518 return true;
1519 }
1520 return SlowAppendFromSelf(offset, len);
1521 }
1522
1523 // Called at the end of the decompress. We ask the allocator
1524 // write all blocks to the sink.
1525 inline void Flush() { allocator_.Flush(Produced()); }
1526};
1527
1528template<typename Allocator>
1529bool SnappyScatteredWriter<Allocator>::SlowAppend(const char* ip, size_t len) {
1530 size_t avail = op_limit_ - op_ptr_;
1531 while (len > avail) {
1532 // Completely fill this block
1533 memcpy(op_ptr_, ip, avail);
1534 op_ptr_ += avail;
1535 assert(op_limit_ - op_ptr_ == 0);
1536 full_size_ += (op_ptr_ - op_base_);
1537 len -= avail;
1538 ip += avail;
1539
1540 // Bounds check
1541 if (full_size_ + len > expected_) {
1542 return false;
1543 }
1544
1545 // Make new block
1546 size_t bsize = std::min<size_t>(kBlockSize, expected_ - full_size_);
1547 op_base_ = allocator_.Allocate(bsize);
1548 op_ptr_ = op_base_;
1549 op_limit_ = op_base_ + bsize;
1550 blocks_.push_back(op_base_);
1551 avail = bsize;
1552 }
1553
1554 memcpy(op_ptr_, ip, len);
1555 op_ptr_ += len;
1556 return true;
1557}
1558
1559template<typename Allocator>
1560bool SnappyScatteredWriter<Allocator>::SlowAppendFromSelf(size_t offset,
1561 size_t len) {
1562 // Overflow check
1563 // See SnappyArrayWriter::AppendFromSelf for an explanation of
1564 // the "offset - 1u" trick.
1565 const size_t cur = Size();
1566 if (offset - 1u >= cur) return false;
1567 if (expected_ - cur < len) return false;
1568
1569 // Currently we shouldn't ever hit this path because Compress() chops the
1570 // input into blocks and does not create cross-block copies. However, it is
1571 // nice if we do not rely on that, since we can get better compression if we
1572 // allow cross-block copies and thus might want to change the compressor in
1573 // the future.
1574 size_t src = cur - offset;
1575 while (len-- > 0) {
1576 char c = blocks_[src >> kBlockLog][src & (kBlockSize-1)];
1577 Append(&c, 1);
1578 src++;
1579 }
1580 return true;
1581}
1582
1583class SnappySinkAllocator {
1584 public:
1585 explicit SnappySinkAllocator(Sink* dest): dest_(dest) {}
1586 ~SnappySinkAllocator() {}
1587
1588 char* Allocate(int size) {
1589 Datablock block(new char[size], size);
1590 blocks_.push_back(block);
1591 return block.data;
1592 }
1593
1594 // We flush only at the end, because the writer wants
1595 // random access to the blocks and once we hand the
1596 // block over to the sink, we can't access it anymore.
1597 // Also we don't write more than has been actually written
1598 // to the blocks.
1599 void Flush(size_t size) {
1600 size_t size_written = 0;
1601 size_t block_size;
1602 for (int i = 0; i < blocks_.size(); ++i) {
1603 block_size = std::min<size_t>(blocks_[i].size, size - size_written);
1604 dest_->AppendAndTakeOwnership(blocks_[i].data, block_size,
1605 &SnappySinkAllocator::Deleter, NULL);
1606 size_written += block_size;
1607 }
1608 blocks_.clear();
1609 }
1610
1611 private:
1612 struct Datablock {
1613 char* data;
1614 size_t size;
1615 Datablock(char* p, size_t s) : data(p), size(s) {}
1616 };
1617
1618 static void Deleter(void* arg, const char* bytes, size_t size) {
1619 delete[] bytes;
1620 }
1621
1622 Sink* dest_;
1623 std::vector<Datablock> blocks_;
1624
1625 // Note: copying this object is allowed
1626};
1627
1628size_t UncompressAsMuchAsPossible(Source* compressed, Sink* uncompressed) {
1629 SnappySinkAllocator allocator(uncompressed);
1630 SnappyScatteredWriter<SnappySinkAllocator> writer(allocator);
1631 InternalUncompress(compressed, &writer);
1632 return writer.Produced();
1633}
1634
1635bool Uncompress(Source* compressed, Sink* uncompressed) {
1636 // Read the uncompressed length from the front of the compressed input
1637 SnappyDecompressor decompressor(compressed);
1638 uint32 uncompressed_len = 0;
1639 if (!decompressor.ReadUncompressedLength(&uncompressed_len)) {
1640 return false;
1641 }
1642
1643 char c;
1644 size_t allocated_size;
1645 char* buf = uncompressed->GetAppendBufferVariable(
1646 1, uncompressed_len, &c, 1, &allocated_size);
1647
1648 const size_t compressed_len = compressed->Available();
1649 // If we can get a flat buffer, then use it, otherwise do block by block
1650 // uncompression
1651 if (allocated_size >= uncompressed_len) {
1652 SnappyArrayWriter writer(buf);
1653 bool result = InternalUncompressAllTags(&decompressor, &writer,
1654 compressed_len, uncompressed_len);
1655 uncompressed->Append(buf, writer.Produced());
1656 return result;
1657 } else {
1658 SnappySinkAllocator allocator(uncompressed);
1659 SnappyScatteredWriter<SnappySinkAllocator> writer(allocator);
1660 return InternalUncompressAllTags(&decompressor, &writer, compressed_len,
1661 uncompressed_len);
1662 }
1663}
1664
1665} // namespace snappy
1666