snappy 0.0.12-java → 0.1.0-java
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +5 -5
- data/.travis.yml +28 -1
- data/Gemfile +6 -1
- data/README.md +28 -4
- data/Rakefile +1 -0
- data/ext/extconf.rb +21 -24
- data/lib/snappy.rb +3 -1
- data/lib/snappy/hadoop.rb +22 -0
- data/lib/snappy/hadoop/reader.rb +58 -0
- data/lib/snappy/hadoop/writer.rb +51 -0
- data/lib/snappy/reader.rb +11 -7
- data/lib/snappy/shim.rb +30 -0
- data/lib/snappy/version.rb +3 -1
- data/lib/snappy/writer.rb +14 -9
- data/smoke.sh +8 -0
- data/snappy.gemspec +6 -30
- data/test/hadoop/test-snappy-hadoop-reader.rb +103 -0
- data/test/hadoop/test-snappy-hadoop-writer.rb +48 -0
- data/test/test-snappy-hadoop.rb +22 -0
- data/vendor/snappy/CMakeLists.txt +174 -0
- data/vendor/snappy/CONTRIBUTING.md +26 -0
- data/vendor/snappy/COPYING +1 -1
- data/vendor/snappy/NEWS +52 -0
- data/vendor/snappy/{README → README.md} +23 -9
- data/vendor/snappy/cmake/SnappyConfig.cmake +1 -0
- data/vendor/snappy/cmake/config.h.in +62 -0
- data/vendor/snappy/snappy-c.h +3 -3
- data/vendor/snappy/snappy-internal.h +101 -27
- data/vendor/snappy/snappy-sinksource.cc +33 -0
- data/vendor/snappy/snappy-sinksource.h +51 -6
- data/vendor/snappy/snappy-stubs-internal.h +107 -37
- data/vendor/snappy/snappy-stubs-public.h.in +16 -20
- data/vendor/snappy/snappy-test.cc +15 -9
- data/vendor/snappy/snappy-test.h +34 -43
- data/vendor/snappy/snappy.cc +529 -320
- data/vendor/snappy/snappy.h +23 -4
- data/vendor/snappy/snappy_unittest.cc +240 -185
- metadata +27 -74
- data/vendor/snappy/ChangeLog +0 -1916
- data/vendor/snappy/Makefile.am +0 -23
- data/vendor/snappy/autogen.sh +0 -7
- data/vendor/snappy/configure.ac +0 -133
- data/vendor/snappy/m4/gtest.m4 +0 -74
- data/vendor/snappy/testdata/alice29.txt +0 -3609
- data/vendor/snappy/testdata/asyoulik.txt +0 -4122
- data/vendor/snappy/testdata/baddata1.snappy +0 -0
- data/vendor/snappy/testdata/baddata2.snappy +0 -0
- data/vendor/snappy/testdata/baddata3.snappy +0 -0
- data/vendor/snappy/testdata/fireworks.jpeg +0 -0
- data/vendor/snappy/testdata/geo.protodata +0 -0
- data/vendor/snappy/testdata/html +0 -1
- data/vendor/snappy/testdata/html_x_4 +0 -1
- data/vendor/snappy/testdata/kppkn.gtb +0 -0
- data/vendor/snappy/testdata/lcet10.txt +0 -7519
- data/vendor/snappy/testdata/paper-100k.pdf +2 -600
- data/vendor/snappy/testdata/plrabn12.txt +0 -10699
- data/vendor/snappy/testdata/urls.10K +0 -10000
data/vendor/snappy/snappy.cc
CHANGED
@@ -30,6 +30,18 @@
|
|
30
30
|
#include "snappy-internal.h"
|
31
31
|
#include "snappy-sinksource.h"
|
32
32
|
|
33
|
+
#ifndef SNAPPY_HAVE_SSE2
|
34
|
+
#if defined(__SSE2__) || defined(_M_X64) || \
|
35
|
+
(defined(_M_IX86_FP) && _M_IX86_FP >= 2)
|
36
|
+
#define SNAPPY_HAVE_SSE2 1
|
37
|
+
#else
|
38
|
+
#define SNAPPY_HAVE_SSE2 0
|
39
|
+
#endif
|
40
|
+
#endif
|
41
|
+
|
42
|
+
#if SNAPPY_HAVE_SSE2
|
43
|
+
#include <emmintrin.h>
|
44
|
+
#endif
|
33
45
|
#include <stdio.h>
|
34
46
|
|
35
47
|
#include <algorithm>
|
@@ -39,6 +51,12 @@
|
|
39
51
|
|
40
52
|
namespace snappy {
|
41
53
|
|
54
|
+
using internal::COPY_1_BYTE_OFFSET;
|
55
|
+
using internal::COPY_2_BYTE_OFFSET;
|
56
|
+
using internal::LITERAL;
|
57
|
+
using internal::char_table;
|
58
|
+
using internal::kMaximumTagLength;
|
59
|
+
|
42
60
|
// Any hash function will produce a valid compressed bitstream, but a good
|
43
61
|
// hash function reduces the number of collisions and thus yields better
|
44
62
|
// compression for compressible input, and more speed for incompressible
|
@@ -76,79 +94,129 @@ size_t MaxCompressedLength(size_t source_len) {
|
|
76
94
|
return 32 + source_len + source_len/6;
|
77
95
|
}
|
78
96
|
|
79
|
-
|
80
|
-
|
81
|
-
|
82
|
-
|
83
|
-
|
84
|
-
|
85
|
-
static const int kMaximumTagLength = 5; // COPY_4_BYTE_OFFSET plus the actual offset.
|
86
|
-
|
87
|
-
// Copy "len" bytes from "src" to "op", one byte at a time. Used for
|
88
|
-
// handling COPY operations where the input and output regions may
|
89
|
-
// overlap. For example, suppose:
|
90
|
-
// src == "ab"
|
91
|
-
// op == src + 2
|
92
|
-
// len == 20
|
93
|
-
// After IncrementalCopy(src, op, len), the result will have
|
94
|
-
// eleven copies of "ab"
|
95
|
-
// ababababababababababab
|
96
|
-
// Note that this does not match the semantics of either memcpy()
|
97
|
-
// or memmove().
|
98
|
-
static inline void IncrementalCopy(const char* src, char* op, ssize_t len) {
|
99
|
-
assert(len > 0);
|
100
|
-
do {
|
101
|
-
*op++ = *src++;
|
102
|
-
} while (--len > 0);
|
97
|
+
namespace {
|
98
|
+
|
99
|
+
void UnalignedCopy64(const void* src, void* dst) {
|
100
|
+
char tmp[8];
|
101
|
+
memcpy(tmp, src, 8);
|
102
|
+
memcpy(dst, tmp, 8);
|
103
103
|
}
|
104
104
|
|
105
|
-
|
106
|
-
//
|
107
|
-
//
|
108
|
-
|
109
|
-
|
110
|
-
|
111
|
-
|
112
|
-
|
113
|
-
|
114
|
-
|
115
|
-
|
116
|
-
|
117
|
-
// [------] op
|
118
|
-
//
|
119
|
-
// a single eight-byte copy from <src> to <op> will repeat the pattern once,
|
120
|
-
// after which we can move <op> two bytes without moving <src>:
|
121
|
-
//
|
122
|
-
// ababxxxxxxxxxx
|
123
|
-
// [------] src
|
124
|
-
// [------] op
|
125
|
-
//
|
126
|
-
// and repeat the exercise until the two no longer overlap.
|
127
|
-
//
|
128
|
-
// This allows us to do very well in the special case of one single byte
|
129
|
-
// repeated many times, without taking a big hit for more general cases.
|
130
|
-
//
|
131
|
-
// The worst case of extra writing past the end of the match occurs when
|
132
|
-
// op - src == 1 and len == 1; the last copy will read from byte positions
|
133
|
-
// [0..7] and write to [4..11], whereas it was only supposed to write to
|
134
|
-
// position 1. Thus, ten excess bytes.
|
105
|
+
void UnalignedCopy128(const void* src, void* dst) {
|
106
|
+
// TODO(alkis): Remove this when we upgrade to a recent compiler that emits
|
107
|
+
// SSE2 moves for memcpy(dst, src, 16).
|
108
|
+
#if SNAPPY_HAVE_SSE2
|
109
|
+
__m128i x = _mm_loadu_si128(static_cast<const __m128i*>(src));
|
110
|
+
_mm_storeu_si128(static_cast<__m128i*>(dst), x);
|
111
|
+
#else
|
112
|
+
char tmp[16];
|
113
|
+
memcpy(tmp, src, 16);
|
114
|
+
memcpy(dst, tmp, 16);
|
115
|
+
#endif
|
116
|
+
}
|
135
117
|
|
136
|
-
|
118
|
+
// Copy [src, src+(op_limit-op)) to [op, (op_limit-op)) a byte at a time. Used
|
119
|
+
// for handling COPY operations where the input and output regions may overlap.
|
120
|
+
// For example, suppose:
|
121
|
+
// src == "ab"
|
122
|
+
// op == src + 2
|
123
|
+
// op_limit == op + 20
|
124
|
+
// After IncrementalCopySlow(src, op, op_limit), the result will have eleven
|
125
|
+
// copies of "ab"
|
126
|
+
// ababababababababababab
|
127
|
+
// Note that this does not match the semantics of either memcpy() or memmove().
|
128
|
+
inline char* IncrementalCopySlow(const char* src, char* op,
|
129
|
+
char* const op_limit) {
|
130
|
+
while (op < op_limit) {
|
131
|
+
*op++ = *src++;
|
132
|
+
}
|
133
|
+
return op_limit;
|
134
|
+
}
|
137
135
|
|
138
|
-
|
136
|
+
// Copy [src, src+(op_limit-op)) to [op, (op_limit-op)) but faster than
|
137
|
+
// IncrementalCopySlow. buf_limit is the address past the end of the writable
|
138
|
+
// region of the buffer.
|
139
|
+
inline char* IncrementalCopy(const char* src, char* op, char* const op_limit,
|
140
|
+
char* const buf_limit) {
|
141
|
+
// Terminology:
|
142
|
+
//
|
143
|
+
// slop = buf_limit - op
|
144
|
+
// pat = op - src
|
145
|
+
// len = limit - op
|
146
|
+
assert(src < op);
|
147
|
+
assert(op_limit <= buf_limit);
|
148
|
+
// NOTE: The compressor always emits 4 <= len <= 64. It is ok to assume that
|
149
|
+
// to optimize this function but we have to also handle these cases in case
|
150
|
+
// the input does not satisfy these conditions.
|
151
|
+
|
152
|
+
size_t pattern_size = op - src;
|
153
|
+
// The cases are split into different branches to allow the branch predictor,
|
154
|
+
// FDO, and static prediction hints to work better. For each input we list the
|
155
|
+
// ratio of invocations that match each condition.
|
156
|
+
//
|
157
|
+
// input slop < 16 pat < 8 len > 16
|
158
|
+
// ------------------------------------------
|
159
|
+
// html|html4|cp 0% 1.01% 27.73%
|
160
|
+
// urls 0% 0.88% 14.79%
|
161
|
+
// jpg 0% 64.29% 7.14%
|
162
|
+
// pdf 0% 2.56% 58.06%
|
163
|
+
// txt[1-4] 0% 0.23% 0.97%
|
164
|
+
// pb 0% 0.96% 13.88%
|
165
|
+
// bin 0.01% 22.27% 41.17%
|
166
|
+
//
|
167
|
+
// It is very rare that we don't have enough slop for doing block copies. It
|
168
|
+
// is also rare that we need to expand a pattern. Small patterns are common
|
169
|
+
// for incompressible formats and for those we are plenty fast already.
|
170
|
+
// Lengths are normally not greater than 16 but they vary depending on the
|
171
|
+
// input. In general if we always predict len <= 16 it would be an ok
|
172
|
+
// prediction.
|
173
|
+
//
|
174
|
+
// In order to be fast we want a pattern >= 8 bytes and an unrolled loop
|
175
|
+
// copying 2x 8 bytes at a time.
|
176
|
+
|
177
|
+
// Handle the uncommon case where pattern is less than 8 bytes.
|
178
|
+
if (SNAPPY_PREDICT_FALSE(pattern_size < 8)) {
|
179
|
+
// Expand pattern to at least 8 bytes. The worse case scenario in terms of
|
180
|
+
// buffer usage is when the pattern is size 3. ^ is the original position
|
181
|
+
// of op. x are irrelevant bytes copied by the last UnalignedCopy64.
|
182
|
+
//
|
183
|
+
// abc
|
184
|
+
// abcabcxxxxx
|
185
|
+
// abcabcabcabcxxxxx
|
186
|
+
// ^
|
187
|
+
// The last x is 14 bytes after ^.
|
188
|
+
if (SNAPPY_PREDICT_TRUE(op <= buf_limit - 14)) {
|
189
|
+
while (pattern_size < 8) {
|
190
|
+
UnalignedCopy64(src, op);
|
191
|
+
op += pattern_size;
|
192
|
+
pattern_size *= 2;
|
193
|
+
}
|
194
|
+
if (SNAPPY_PREDICT_TRUE(op >= op_limit)) return op_limit;
|
195
|
+
} else {
|
196
|
+
return IncrementalCopySlow(src, op, op_limit);
|
197
|
+
}
|
198
|
+
}
|
199
|
+
assert(pattern_size >= 8);
|
139
200
|
|
140
|
-
|
141
|
-
|
201
|
+
// Copy 2x 8 bytes at a time. Because op - src can be < 16, a single
|
202
|
+
// UnalignedCopy128 might overwrite data in op. UnalignedCopy64 is safe
|
203
|
+
// because expanding the pattern to at least 8 bytes guarantees that
|
204
|
+
// op - src >= 8.
|
205
|
+
while (op <= buf_limit - 16) {
|
142
206
|
UnalignedCopy64(src, op);
|
143
|
-
|
144
|
-
|
207
|
+
UnalignedCopy64(src + 8, op + 8);
|
208
|
+
src += 16;
|
209
|
+
op += 16;
|
210
|
+
if (SNAPPY_PREDICT_TRUE(op >= op_limit)) return op_limit;
|
145
211
|
}
|
146
|
-
|
212
|
+
// We only take this branch if we didn't have enough slop and we can do a
|
213
|
+
// single 8 byte copy.
|
214
|
+
if (SNAPPY_PREDICT_FALSE(op <= buf_limit - 8)) {
|
147
215
|
UnalignedCopy64(src, op);
|
148
216
|
src += 8;
|
149
217
|
op += 8;
|
150
|
-
len -= 8;
|
151
218
|
}
|
219
|
+
return IncrementalCopySlow(src, op, op_limit);
|
152
220
|
}
|
153
221
|
|
154
222
|
} // namespace
|
@@ -157,26 +225,29 @@ static inline char* EmitLiteral(char* op,
|
|
157
225
|
const char* literal,
|
158
226
|
int len,
|
159
227
|
bool allow_fast_path) {
|
160
|
-
|
161
|
-
|
228
|
+
// The vast majority of copies are below 16 bytes, for which a
|
229
|
+
// call to memcpy is overkill. This fast path can sometimes
|
230
|
+
// copy up to 15 bytes too much, but that is okay in the
|
231
|
+
// main loop, since we have a bit to go on for both sides:
|
232
|
+
//
|
233
|
+
// - The input will always have kInputMarginBytes = 15 extra
|
234
|
+
// available bytes, as long as we're in the main loop, and
|
235
|
+
// if not, allow_fast_path = false.
|
236
|
+
// - The output will always have 32 spare bytes (see
|
237
|
+
// MaxCompressedLength).
|
238
|
+
assert(len > 0); // Zero-length literals are disallowed
|
239
|
+
int n = len - 1;
|
240
|
+
if (allow_fast_path && len <= 16) {
|
162
241
|
// Fits in tag byte
|
163
242
|
*op++ = LITERAL | (n << 2);
|
164
243
|
|
165
|
-
|
166
|
-
|
167
|
-
|
168
|
-
|
169
|
-
|
170
|
-
//
|
171
|
-
|
172
|
-
// if not, allow_fast_path = false.
|
173
|
-
// - The output will always have 32 spare bytes (see
|
174
|
-
// MaxCompressedLength).
|
175
|
-
if (allow_fast_path && len <= 16) {
|
176
|
-
UnalignedCopy64(literal, op);
|
177
|
-
UnalignedCopy64(literal + 8, op + 8);
|
178
|
-
return op + len;
|
179
|
-
}
|
244
|
+
UnalignedCopy128(literal, op);
|
245
|
+
return op + len;
|
246
|
+
}
|
247
|
+
|
248
|
+
if (n < 60) {
|
249
|
+
// Fits in tag byte
|
250
|
+
*op++ = LITERAL | (n << 2);
|
180
251
|
} else {
|
181
252
|
// Encode in upcoming bytes
|
182
253
|
char* base = op;
|
@@ -195,42 +266,54 @@ static inline char* EmitLiteral(char* op,
|
|
195
266
|
return op + len;
|
196
267
|
}
|
197
268
|
|
198
|
-
static inline char*
|
269
|
+
static inline char* EmitCopyAtMost64(char* op, size_t offset, size_t len,
|
270
|
+
bool len_less_than_12) {
|
199
271
|
assert(len <= 64);
|
200
272
|
assert(len >= 4);
|
201
273
|
assert(offset < 65536);
|
274
|
+
assert(len_less_than_12 == (len < 12));
|
202
275
|
|
203
|
-
if (
|
204
|
-
|
205
|
-
|
206
|
-
*op++ = COPY_1_BYTE_OFFSET + ((
|
276
|
+
if (len_less_than_12 && SNAPPY_PREDICT_TRUE(offset < 2048)) {
|
277
|
+
// offset fits in 11 bits. The 3 highest go in the top of the first byte,
|
278
|
+
// and the rest go in the second byte.
|
279
|
+
*op++ = COPY_1_BYTE_OFFSET + ((len - 4) << 2) + ((offset >> 3) & 0xe0);
|
207
280
|
*op++ = offset & 0xff;
|
208
281
|
} else {
|
209
|
-
|
210
|
-
|
211
|
-
|
282
|
+
// Write 4 bytes, though we only care about 3 of them. The output buffer
|
283
|
+
// is required to have some slack, so the extra byte won't overrun it.
|
284
|
+
uint32 u = COPY_2_BYTE_OFFSET + ((len - 1) << 2) + (offset << 8);
|
285
|
+
LittleEndian::Store32(op, u);
|
286
|
+
op += 3;
|
212
287
|
}
|
213
288
|
return op;
|
214
289
|
}
|
215
290
|
|
216
|
-
static inline char* EmitCopy(char* op, size_t offset,
|
217
|
-
|
218
|
-
|
219
|
-
|
220
|
-
len
|
221
|
-
}
|
291
|
+
static inline char* EmitCopy(char* op, size_t offset, size_t len,
|
292
|
+
bool len_less_than_12) {
|
293
|
+
assert(len_less_than_12 == (len < 12));
|
294
|
+
if (len_less_than_12) {
|
295
|
+
return EmitCopyAtMost64(op, offset, len, true);
|
296
|
+
} else {
|
297
|
+
// A special case for len <= 64 might help, but so far measurements suggest
|
298
|
+
// it's in the noise.
|
222
299
|
|
223
|
-
|
224
|
-
|
225
|
-
|
226
|
-
|
227
|
-
|
300
|
+
// Emit 64 byte copies but make sure to keep at least four bytes reserved.
|
301
|
+
while (SNAPPY_PREDICT_FALSE(len >= 68)) {
|
302
|
+
op = EmitCopyAtMost64(op, offset, 64, false);
|
303
|
+
len -= 64;
|
304
|
+
}
|
228
305
|
|
229
|
-
|
230
|
-
|
231
|
-
|
232
|
-
|
306
|
+
// One or two copies will now finish the job.
|
307
|
+
if (len > 64) {
|
308
|
+
op = EmitCopyAtMost64(op, offset, 60, false);
|
309
|
+
len -= 60;
|
310
|
+
}
|
233
311
|
|
312
|
+
// Emit remainder.
|
313
|
+
op = EmitCopyAtMost64(op, offset, len, len < 12);
|
314
|
+
return op;
|
315
|
+
}
|
316
|
+
}
|
234
317
|
|
235
318
|
bool GetUncompressedLength(const char* start, size_t n, size_t* result) {
|
236
319
|
uint32 v = 0;
|
@@ -344,7 +427,7 @@ char* CompressFragment(const char* input,
|
|
344
427
|
const char* next_emit = ip;
|
345
428
|
|
346
429
|
const size_t kInputMarginBytes = 15;
|
347
|
-
if (
|
430
|
+
if (SNAPPY_PREDICT_TRUE(input_size >= kInputMarginBytes)) {
|
348
431
|
const char* ip_limit = input + input_size - kInputMarginBytes;
|
349
432
|
|
350
433
|
for (uint32 next_hash = Hash(++ip, shift); ; ) {
|
@@ -364,9 +447,9 @@ char* CompressFragment(const char* input,
|
|
364
447
|
//
|
365
448
|
// Heuristic match skipping: If 32 bytes are scanned with no matches
|
366
449
|
// found, start looking only at every other byte. If 32 more bytes are
|
367
|
-
// scanned, look at every third byte, etc.. When a match is
|
368
|
-
// immediately go back to looking at every byte. This is a small
|
369
|
-
// (~5% performance, ~0.1% density) for compressible data due to more
|
450
|
+
// scanned (or skipped), look at every third byte, etc.. When a match is
|
451
|
+
// found, immediately go back to looking at every byte. This is a small
|
452
|
+
// loss (~5% performance, ~0.1% density) for compressible data due to more
|
370
453
|
// bookkeeping, but for non-compressible data (such as JPEG) it's a huge
|
371
454
|
// win since the compressor quickly "realizes" the data is incompressible
|
372
455
|
// and doesn't bother looking for matches everywhere.
|
@@ -382,9 +465,10 @@ char* CompressFragment(const char* input,
|
|
382
465
|
ip = next_ip;
|
383
466
|
uint32 hash = next_hash;
|
384
467
|
assert(hash == Hash(ip, shift));
|
385
|
-
uint32 bytes_between_hash_lookups = skip
|
468
|
+
uint32 bytes_between_hash_lookups = skip >> 5;
|
469
|
+
skip += bytes_between_hash_lookups;
|
386
470
|
next_ip = ip + bytes_between_hash_lookups;
|
387
|
-
if (
|
471
|
+
if (SNAPPY_PREDICT_FALSE(next_ip > ip_limit)) {
|
388
472
|
goto emit_remainder;
|
389
473
|
}
|
390
474
|
next_hash = Hash(next_ip, shift);
|
@@ -393,8 +477,8 @@ char* CompressFragment(const char* input,
|
|
393
477
|
assert(candidate < ip);
|
394
478
|
|
395
479
|
table[hash] = ip - base_ip;
|
396
|
-
} while (
|
397
|
-
|
480
|
+
} while (SNAPPY_PREDICT_TRUE(UNALIGNED_LOAD32(ip) !=
|
481
|
+
UNALIGNED_LOAD32(candidate)));
|
398
482
|
|
399
483
|
// Step 2: A 4-byte match has been found. We'll later see if more
|
400
484
|
// than 4 bytes match. But, prior to the match, input
|
@@ -417,19 +501,21 @@ char* CompressFragment(const char* input,
|
|
417
501
|
// We have a 4-byte match at ip, and no need to emit any
|
418
502
|
// "literal bytes" prior to ip.
|
419
503
|
const char* base = ip;
|
420
|
-
|
504
|
+
std::pair<size_t, bool> p =
|
505
|
+
FindMatchLength(candidate + 4, ip + 4, ip_end);
|
506
|
+
size_t matched = 4 + p.first;
|
421
507
|
ip += matched;
|
422
508
|
size_t offset = base - candidate;
|
423
509
|
assert(0 == memcmp(base, candidate, matched));
|
424
|
-
op = EmitCopy(op, offset, matched);
|
425
|
-
// We could immediately start working at ip now, but to improve
|
426
|
-
// compression we first update table[Hash(ip - 1, ...)].
|
427
|
-
const char* insert_tail = ip - 1;
|
510
|
+
op = EmitCopy(op, offset, matched, p.second);
|
428
511
|
next_emit = ip;
|
429
|
-
if (
|
512
|
+
if (SNAPPY_PREDICT_FALSE(ip >= ip_limit)) {
|
430
513
|
goto emit_remainder;
|
431
514
|
}
|
432
|
-
|
515
|
+
// We are now looking for a 4-byte match again. We read
|
516
|
+
// table[Hash(ip, shift)] for that. To improve compression,
|
517
|
+
// we also update table[Hash(ip - 1, shift)] and table[Hash(ip, shift)].
|
518
|
+
input_bytes = GetEightBytesAt(ip - 1);
|
433
519
|
uint32 prev_hash = HashBytes(GetUint32AtOffset(input_bytes, 0), shift);
|
434
520
|
table[prev_hash] = ip - base_ip - 1;
|
435
521
|
uint32 cur_hash = HashBytes(GetUint32AtOffset(input_bytes, 1), shift);
|
@@ -453,6 +539,10 @@ char* CompressFragment(const char* input,
|
|
453
539
|
}
|
454
540
|
} // end namespace internal
|
455
541
|
|
542
|
+
// Called back at avery compression call to trace parameters and sizes.
|
543
|
+
static inline void Report(const char *algorithm, size_t compressed_size,
|
544
|
+
size_t uncompressed_size) {}
|
545
|
+
|
456
546
|
// Signature of output types needed by decompression code.
|
457
547
|
// The decompression code is templatized on a type that obeys this
|
458
548
|
// signature so that we do not pay virtual function call overhead in
|
@@ -493,162 +583,14 @@ char* CompressFragment(const char* input,
|
|
493
583
|
// bool TryFastAppend(const char* ip, size_t available, size_t length);
|
494
584
|
// };
|
495
585
|
|
496
|
-
|
497
|
-
// Lookup table for decompression code. Generated by ComputeTable() below.
|
498
|
-
// -----------------------------------------------------------------------
|
586
|
+
namespace internal {
|
499
587
|
|
500
588
|
// Mapping from i in range [0,4] to a mask to extract the bottom 8*i bits
|
501
589
|
static const uint32 wordmask[] = {
|
502
590
|
0u, 0xffu, 0xffffu, 0xffffffu, 0xffffffffu
|
503
591
|
};
|
504
592
|
|
505
|
-
//
|
506
|
-
// Range Bits-used Description
|
507
|
-
// ------------------------------------
|
508
|
-
// 1..64 0..7 Literal/copy length encoded in opcode byte
|
509
|
-
// 0..7 8..10 Copy offset encoded in opcode byte / 256
|
510
|
-
// 0..4 11..13 Extra bytes after opcode
|
511
|
-
//
|
512
|
-
// We use eight bits for the length even though 7 would have sufficed
|
513
|
-
// because of efficiency reasons:
|
514
|
-
// (1) Extracting a byte is faster than a bit-field
|
515
|
-
// (2) It properly aligns copy offset so we do not need a <<8
|
516
|
-
static const uint16 char_table[256] = {
|
517
|
-
0x0001, 0x0804, 0x1001, 0x2001, 0x0002, 0x0805, 0x1002, 0x2002,
|
518
|
-
0x0003, 0x0806, 0x1003, 0x2003, 0x0004, 0x0807, 0x1004, 0x2004,
|
519
|
-
0x0005, 0x0808, 0x1005, 0x2005, 0x0006, 0x0809, 0x1006, 0x2006,
|
520
|
-
0x0007, 0x080a, 0x1007, 0x2007, 0x0008, 0x080b, 0x1008, 0x2008,
|
521
|
-
0x0009, 0x0904, 0x1009, 0x2009, 0x000a, 0x0905, 0x100a, 0x200a,
|
522
|
-
0x000b, 0x0906, 0x100b, 0x200b, 0x000c, 0x0907, 0x100c, 0x200c,
|
523
|
-
0x000d, 0x0908, 0x100d, 0x200d, 0x000e, 0x0909, 0x100e, 0x200e,
|
524
|
-
0x000f, 0x090a, 0x100f, 0x200f, 0x0010, 0x090b, 0x1010, 0x2010,
|
525
|
-
0x0011, 0x0a04, 0x1011, 0x2011, 0x0012, 0x0a05, 0x1012, 0x2012,
|
526
|
-
0x0013, 0x0a06, 0x1013, 0x2013, 0x0014, 0x0a07, 0x1014, 0x2014,
|
527
|
-
0x0015, 0x0a08, 0x1015, 0x2015, 0x0016, 0x0a09, 0x1016, 0x2016,
|
528
|
-
0x0017, 0x0a0a, 0x1017, 0x2017, 0x0018, 0x0a0b, 0x1018, 0x2018,
|
529
|
-
0x0019, 0x0b04, 0x1019, 0x2019, 0x001a, 0x0b05, 0x101a, 0x201a,
|
530
|
-
0x001b, 0x0b06, 0x101b, 0x201b, 0x001c, 0x0b07, 0x101c, 0x201c,
|
531
|
-
0x001d, 0x0b08, 0x101d, 0x201d, 0x001e, 0x0b09, 0x101e, 0x201e,
|
532
|
-
0x001f, 0x0b0a, 0x101f, 0x201f, 0x0020, 0x0b0b, 0x1020, 0x2020,
|
533
|
-
0x0021, 0x0c04, 0x1021, 0x2021, 0x0022, 0x0c05, 0x1022, 0x2022,
|
534
|
-
0x0023, 0x0c06, 0x1023, 0x2023, 0x0024, 0x0c07, 0x1024, 0x2024,
|
535
|
-
0x0025, 0x0c08, 0x1025, 0x2025, 0x0026, 0x0c09, 0x1026, 0x2026,
|
536
|
-
0x0027, 0x0c0a, 0x1027, 0x2027, 0x0028, 0x0c0b, 0x1028, 0x2028,
|
537
|
-
0x0029, 0x0d04, 0x1029, 0x2029, 0x002a, 0x0d05, 0x102a, 0x202a,
|
538
|
-
0x002b, 0x0d06, 0x102b, 0x202b, 0x002c, 0x0d07, 0x102c, 0x202c,
|
539
|
-
0x002d, 0x0d08, 0x102d, 0x202d, 0x002e, 0x0d09, 0x102e, 0x202e,
|
540
|
-
0x002f, 0x0d0a, 0x102f, 0x202f, 0x0030, 0x0d0b, 0x1030, 0x2030,
|
541
|
-
0x0031, 0x0e04, 0x1031, 0x2031, 0x0032, 0x0e05, 0x1032, 0x2032,
|
542
|
-
0x0033, 0x0e06, 0x1033, 0x2033, 0x0034, 0x0e07, 0x1034, 0x2034,
|
543
|
-
0x0035, 0x0e08, 0x1035, 0x2035, 0x0036, 0x0e09, 0x1036, 0x2036,
|
544
|
-
0x0037, 0x0e0a, 0x1037, 0x2037, 0x0038, 0x0e0b, 0x1038, 0x2038,
|
545
|
-
0x0039, 0x0f04, 0x1039, 0x2039, 0x003a, 0x0f05, 0x103a, 0x203a,
|
546
|
-
0x003b, 0x0f06, 0x103b, 0x203b, 0x003c, 0x0f07, 0x103c, 0x203c,
|
547
|
-
0x0801, 0x0f08, 0x103d, 0x203d, 0x1001, 0x0f09, 0x103e, 0x203e,
|
548
|
-
0x1801, 0x0f0a, 0x103f, 0x203f, 0x2001, 0x0f0b, 0x1040, 0x2040
|
549
|
-
};
|
550
|
-
|
551
|
-
// In debug mode, allow optional computation of the table at startup.
|
552
|
-
// Also, check that the decompression table is correct.
|
553
|
-
#ifndef NDEBUG
|
554
|
-
DEFINE_bool(snappy_dump_decompression_table, false,
|
555
|
-
"If true, we print the decompression table at startup.");
|
556
|
-
|
557
|
-
static uint16 MakeEntry(unsigned int extra,
|
558
|
-
unsigned int len,
|
559
|
-
unsigned int copy_offset) {
|
560
|
-
// Check that all of the fields fit within the allocated space
|
561
|
-
assert(extra == (extra & 0x7)); // At most 3 bits
|
562
|
-
assert(copy_offset == (copy_offset & 0x7)); // At most 3 bits
|
563
|
-
assert(len == (len & 0x7f)); // At most 7 bits
|
564
|
-
return len | (copy_offset << 8) | (extra << 11);
|
565
|
-
}
|
566
|
-
|
567
|
-
static void ComputeTable() {
|
568
|
-
uint16 dst[256];
|
569
|
-
|
570
|
-
// Place invalid entries in all places to detect missing initialization
|
571
|
-
int assigned = 0;
|
572
|
-
for (int i = 0; i < 256; i++) {
|
573
|
-
dst[i] = 0xffff;
|
574
|
-
}
|
575
|
-
|
576
|
-
// Small LITERAL entries. We store (len-1) in the top 6 bits.
|
577
|
-
for (unsigned int len = 1; len <= 60; len++) {
|
578
|
-
dst[LITERAL | ((len-1) << 2)] = MakeEntry(0, len, 0);
|
579
|
-
assigned++;
|
580
|
-
}
|
581
|
-
|
582
|
-
// Large LITERAL entries. We use 60..63 in the high 6 bits to
|
583
|
-
// encode the number of bytes of length info that follow the opcode.
|
584
|
-
for (unsigned int extra_bytes = 1; extra_bytes <= 4; extra_bytes++) {
|
585
|
-
// We set the length field in the lookup table to 1 because extra
|
586
|
-
// bytes encode len-1.
|
587
|
-
dst[LITERAL | ((extra_bytes+59) << 2)] = MakeEntry(extra_bytes, 1, 0);
|
588
|
-
assigned++;
|
589
|
-
}
|
590
|
-
|
591
|
-
// COPY_1_BYTE_OFFSET.
|
592
|
-
//
|
593
|
-
// The tag byte in the compressed data stores len-4 in 3 bits, and
|
594
|
-
// offset/256 in 5 bits. offset%256 is stored in the next byte.
|
595
|
-
//
|
596
|
-
// This format is used for length in range [4..11] and offset in
|
597
|
-
// range [0..2047]
|
598
|
-
for (unsigned int len = 4; len < 12; len++) {
|
599
|
-
for (unsigned int offset = 0; offset < 2048; offset += 256) {
|
600
|
-
dst[COPY_1_BYTE_OFFSET | ((len-4)<<2) | ((offset>>8)<<5)] =
|
601
|
-
MakeEntry(1, len, offset>>8);
|
602
|
-
assigned++;
|
603
|
-
}
|
604
|
-
}
|
605
|
-
|
606
|
-
// COPY_2_BYTE_OFFSET.
|
607
|
-
// Tag contains len-1 in top 6 bits, and offset in next two bytes.
|
608
|
-
for (unsigned int len = 1; len <= 64; len++) {
|
609
|
-
dst[COPY_2_BYTE_OFFSET | ((len-1)<<2)] = MakeEntry(2, len, 0);
|
610
|
-
assigned++;
|
611
|
-
}
|
612
|
-
|
613
|
-
// COPY_4_BYTE_OFFSET.
|
614
|
-
// Tag contents len-1 in top 6 bits, and offset in next four bytes.
|
615
|
-
for (unsigned int len = 1; len <= 64; len++) {
|
616
|
-
dst[COPY_4_BYTE_OFFSET | ((len-1)<<2)] = MakeEntry(4, len, 0);
|
617
|
-
assigned++;
|
618
|
-
}
|
619
|
-
|
620
|
-
// Check that each entry was initialized exactly once.
|
621
|
-
if (assigned != 256) {
|
622
|
-
fprintf(stderr, "ComputeTable: assigned only %d of 256\n", assigned);
|
623
|
-
abort();
|
624
|
-
}
|
625
|
-
for (int i = 0; i < 256; i++) {
|
626
|
-
if (dst[i] == 0xffff) {
|
627
|
-
fprintf(stderr, "ComputeTable: did not assign byte %d\n", i);
|
628
|
-
abort();
|
629
|
-
}
|
630
|
-
}
|
631
|
-
|
632
|
-
if (FLAGS_snappy_dump_decompression_table) {
|
633
|
-
printf("static const uint16 char_table[256] = {\n ");
|
634
|
-
for (int i = 0; i < 256; i++) {
|
635
|
-
printf("0x%04x%s",
|
636
|
-
dst[i],
|
637
|
-
((i == 255) ? "\n" : (((i%8) == 7) ? ",\n " : ", ")));
|
638
|
-
}
|
639
|
-
printf("};\n");
|
640
|
-
}
|
641
|
-
|
642
|
-
// Check that computed table matched recorded table
|
643
|
-
for (int i = 0; i < 256; i++) {
|
644
|
-
if (dst[i] != char_table[i]) {
|
645
|
-
fprintf(stderr, "ComputeTable: byte %d: computed (%x), expect (%x)\n",
|
646
|
-
i, static_cast<int>(dst[i]), static_cast<int>(char_table[i]));
|
647
|
-
abort();
|
648
|
-
}
|
649
|
-
}
|
650
|
-
}
|
651
|
-
#endif /* !NDEBUG */
|
593
|
+
} // end namespace internal
|
652
594
|
|
653
595
|
// Helper class for decompression
|
654
596
|
class SnappyDecompressor {
|
@@ -701,7 +643,9 @@ class SnappyDecompressor {
|
|
701
643
|
if (n == 0) return false;
|
702
644
|
const unsigned char c = *(reinterpret_cast<const unsigned char*>(ip));
|
703
645
|
reader_->Skip(1);
|
704
|
-
|
646
|
+
uint32 val = c & 0x7f;
|
647
|
+
if (((val << shift) >> shift) != val) return false;
|
648
|
+
*result |= val << shift;
|
705
649
|
if (c < 128) {
|
706
650
|
break;
|
707
651
|
}
|
@@ -715,6 +659,19 @@ class SnappyDecompressor {
|
|
715
659
|
template <class Writer>
|
716
660
|
void DecompressAllTags(Writer* writer) {
|
717
661
|
const char* ip = ip_;
|
662
|
+
// For position-independent executables, accessing global arrays can be
|
663
|
+
// slow. Move wordmask array onto the stack to mitigate this.
|
664
|
+
uint32 wordmask[sizeof(internal::wordmask)/sizeof(uint32)];
|
665
|
+
// Do not use memcpy to copy internal::wordmask to
|
666
|
+
// wordmask. LLVM converts stack arrays to global arrays if it detects
|
667
|
+
// const stack arrays and this hurts the performance of position
|
668
|
+
// independent code. This change is temporary and can be reverted when
|
669
|
+
// https://reviews.llvm.org/D30759 is approved.
|
670
|
+
wordmask[0] = internal::wordmask[0];
|
671
|
+
wordmask[1] = internal::wordmask[1];
|
672
|
+
wordmask[2] = internal::wordmask[2];
|
673
|
+
wordmask[3] = internal::wordmask[3];
|
674
|
+
wordmask[4] = internal::wordmask[4];
|
718
675
|
|
719
676
|
// We could have put this refill fragment only at the beginning of the loop.
|
720
677
|
// However, duplicating it at the end of each branch gives the compiler more
|
@@ -728,10 +685,29 @@ class SnappyDecompressor {
|
|
728
685
|
}
|
729
686
|
|
730
687
|
MAYBE_REFILL();
|
688
|
+
// Add loop alignment directive. Without this directive, we observed
|
689
|
+
// significant performance degradation on several intel architectures
|
690
|
+
// in snappy benchmark built with LLVM. The degradation was caused by
|
691
|
+
// increased branch miss prediction.
|
692
|
+
#if defined(__clang__) && defined(__x86_64__)
|
693
|
+
asm volatile (".p2align 5");
|
694
|
+
#endif
|
731
695
|
for ( ;; ) {
|
732
696
|
const unsigned char c = *(reinterpret_cast<const unsigned char*>(ip++));
|
733
697
|
|
734
|
-
|
698
|
+
// Ratio of iterations that have LITERAL vs non-LITERAL for different
|
699
|
+
// inputs.
|
700
|
+
//
|
701
|
+
// input LITERAL NON_LITERAL
|
702
|
+
// -----------------------------------
|
703
|
+
// html|html4|cp 23% 77%
|
704
|
+
// urls 36% 64%
|
705
|
+
// jpg 47% 53%
|
706
|
+
// pdf 19% 81%
|
707
|
+
// txt[1-4] 25% 75%
|
708
|
+
// pb 24% 76%
|
709
|
+
// bin 24% 76%
|
710
|
+
if (SNAPPY_PREDICT_FALSE((c & 0x3) == LITERAL)) {
|
735
711
|
size_t literal_length = (c >> 2) + 1u;
|
736
712
|
if (writer->TryFastAppend(ip, ip_limit_ - ip, literal_length)) {
|
737
713
|
assert(literal_length < 61);
|
@@ -741,7 +717,7 @@ class SnappyDecompressor {
|
|
741
717
|
// bytes in addition to the literal.
|
742
718
|
continue;
|
743
719
|
}
|
744
|
-
if (
|
720
|
+
if (SNAPPY_PREDICT_FALSE(literal_length >= 61)) {
|
745
721
|
// Long literal.
|
746
722
|
const size_t literal_length_length = literal_length - 60;
|
747
723
|
literal_length =
|
@@ -767,15 +743,15 @@ class SnappyDecompressor {
|
|
767
743
|
ip += literal_length;
|
768
744
|
MAYBE_REFILL();
|
769
745
|
} else {
|
770
|
-
const
|
771
|
-
const
|
772
|
-
const
|
746
|
+
const size_t entry = char_table[c];
|
747
|
+
const size_t trailer = LittleEndian::Load32(ip) & wordmask[entry >> 11];
|
748
|
+
const size_t length = entry & 0xff;
|
773
749
|
ip += entry >> 11;
|
774
750
|
|
775
751
|
// copy_offset/256 is encoded in bits 8..10. By just fetching
|
776
752
|
// those bits, we get copy_offset (since the bit-field starts at
|
777
753
|
// bit 8).
|
778
|
-
const
|
754
|
+
const size_t copy_offset = entry & 0x700;
|
779
755
|
if (!writer->AppendFromSelf(copy_offset + trailer, length)) {
|
780
756
|
return;
|
781
757
|
}
|
@@ -795,10 +771,8 @@ bool SnappyDecompressor::RefillTag() {
|
|
795
771
|
size_t n;
|
796
772
|
ip = reader_->Peek(&n);
|
797
773
|
peeked_ = n;
|
798
|
-
|
799
|
-
|
800
|
-
return false;
|
801
|
-
}
|
774
|
+
eof_ = (n == 0);
|
775
|
+
if (eof_) return false;
|
802
776
|
ip_limit_ = ip + n;
|
803
777
|
}
|
804
778
|
|
@@ -823,7 +797,7 @@ bool SnappyDecompressor::RefillTag() {
|
|
823
797
|
size_t length;
|
824
798
|
const char* src = reader_->Peek(&length);
|
825
799
|
if (length == 0) return false;
|
826
|
-
uint32 to_add = min<uint32>(needed - nbuf, length);
|
800
|
+
uint32 to_add = std::min<uint32>(needed - nbuf, length);
|
827
801
|
memcpy(scratch_ + nbuf, src, to_add);
|
828
802
|
nbuf += to_add;
|
829
803
|
reader_->Skip(to_add);
|
@@ -852,17 +826,23 @@ static bool InternalUncompress(Source* r, Writer* writer) {
|
|
852
826
|
SnappyDecompressor decompressor(r);
|
853
827
|
uint32 uncompressed_len = 0;
|
854
828
|
if (!decompressor.ReadUncompressedLength(&uncompressed_len)) return false;
|
855
|
-
|
829
|
+
|
830
|
+
return InternalUncompressAllTags(&decompressor, writer, r->Available(),
|
831
|
+
uncompressed_len);
|
856
832
|
}
|
857
833
|
|
858
834
|
template <typename Writer>
|
859
835
|
static bool InternalUncompressAllTags(SnappyDecompressor* decompressor,
|
860
836
|
Writer* writer,
|
837
|
+
uint32 compressed_len,
|
861
838
|
uint32 uncompressed_len) {
|
839
|
+
Report("snappy_uncompress", compressed_len, uncompressed_len);
|
840
|
+
|
862
841
|
writer->SetExpectedLength(uncompressed_len);
|
863
842
|
|
864
843
|
// Process the entire input
|
865
844
|
decompressor->DecompressAllTags(writer);
|
845
|
+
writer->Flush();
|
866
846
|
return (decompressor->eof() && writer->CheckLength());
|
867
847
|
}
|
868
848
|
|
@@ -874,6 +854,7 @@ bool GetUncompressedLength(Source* source, uint32* result) {
|
|
874
854
|
size_t Compress(Source* reader, Sink* writer) {
|
875
855
|
size_t written = 0;
|
876
856
|
size_t N = reader->Available();
|
857
|
+
const size_t uncompressed_size = N;
|
877
858
|
char ulength[Varint::kMax32];
|
878
859
|
char* p = Varint::Encode32(ulength, N);
|
879
860
|
writer->Append(ulength, p-ulength);
|
@@ -888,7 +869,7 @@ size_t Compress(Source* reader, Sink* writer) {
|
|
888
869
|
size_t fragment_size;
|
889
870
|
const char* fragment = reader->Peek(&fragment_size);
|
890
871
|
assert(fragment_size != 0); // premature end of input
|
891
|
-
const size_t num_to_read = min(N, kBlockSize);
|
872
|
+
const size_t num_to_read = std::min(N, kBlockSize);
|
892
873
|
size_t bytes_read = fragment_size;
|
893
874
|
|
894
875
|
size_t pending_advance = 0;
|
@@ -909,7 +890,7 @@ size_t Compress(Source* reader, Sink* writer) {
|
|
909
890
|
|
910
891
|
while (bytes_read < num_to_read) {
|
911
892
|
fragment = reader->Peek(&fragment_size);
|
912
|
-
size_t n = min<size_t>(fragment_size, num_to_read - bytes_read);
|
893
|
+
size_t n = std::min<size_t>(fragment_size, num_to_read - bytes_read);
|
913
894
|
memcpy(scratch + bytes_read, fragment, n);
|
914
895
|
bytes_read += n;
|
915
896
|
reader->Skip(n);
|
@@ -946,6 +927,8 @@ size_t Compress(Source* reader, Sink* writer) {
|
|
946
927
|
reader->Skip(pending_advance);
|
947
928
|
}
|
948
929
|
|
930
|
+
Report("snappy_compress", written, uncompressed_size);
|
931
|
+
|
949
932
|
delete[] scratch;
|
950
933
|
delete[] scratch_output;
|
951
934
|
|
@@ -965,7 +948,7 @@ class SnappyIOVecWriter {
|
|
965
948
|
const size_t output_iov_count_;
|
966
949
|
|
967
950
|
// We are currently writing into output_iov_[curr_iov_index_].
|
968
|
-
|
951
|
+
size_t curr_iov_index_;
|
969
952
|
|
970
953
|
// Bytes written to output_iov_[curr_iov_index_] so far.
|
971
954
|
size_t curr_iov_written_;
|
@@ -976,7 +959,7 @@ class SnappyIOVecWriter {
|
|
976
959
|
// Maximum number of bytes that will be decompressed into output_iov_.
|
977
960
|
size_t output_limit_;
|
978
961
|
|
979
|
-
inline char* GetIOVecPointer(
|
962
|
+
inline char* GetIOVecPointer(size_t index, size_t offset) {
|
980
963
|
return reinterpret_cast<char*>(output_iov_[index].iov_base) +
|
981
964
|
offset;
|
982
965
|
}
|
@@ -1037,8 +1020,7 @@ class SnappyIOVecWriter {
|
|
1037
1020
|
output_iov_[curr_iov_index_].iov_len - curr_iov_written_ >= 16) {
|
1038
1021
|
// Fast path, used for the majority (about 95%) of invocations.
|
1039
1022
|
char* ptr = GetIOVecPointer(curr_iov_index_, curr_iov_written_);
|
1040
|
-
|
1041
|
-
UnalignedCopy64(ip + 8, ptr + 8);
|
1023
|
+
UnalignedCopy128(ip, ptr);
|
1042
1024
|
curr_iov_written_ += len;
|
1043
1025
|
total_written_ += len;
|
1044
1026
|
return true;
|
@@ -1057,7 +1039,7 @@ class SnappyIOVecWriter {
|
|
1057
1039
|
}
|
1058
1040
|
|
1059
1041
|
// Locate the iovec from which we need to start the copy.
|
1060
|
-
|
1042
|
+
size_t from_iov_index = curr_iov_index_;
|
1061
1043
|
size_t from_iov_offset = curr_iov_written_;
|
1062
1044
|
while (offset > 0) {
|
1063
1045
|
if (from_iov_offset >= offset) {
|
@@ -1066,8 +1048,8 @@ class SnappyIOVecWriter {
|
|
1066
1048
|
}
|
1067
1049
|
|
1068
1050
|
offset -= from_iov_offset;
|
1051
|
+
assert(from_iov_index > 0);
|
1069
1052
|
--from_iov_index;
|
1070
|
-
assert(from_iov_index >= 0);
|
1071
1053
|
from_iov_offset = output_iov_[from_iov_index].iov_len;
|
1072
1054
|
}
|
1073
1055
|
|
@@ -1102,9 +1084,10 @@ class SnappyIOVecWriter {
|
|
1102
1084
|
if (to_copy > len) {
|
1103
1085
|
to_copy = len;
|
1104
1086
|
}
|
1105
|
-
|
1106
|
-
|
1107
|
-
|
1087
|
+
IncrementalCopySlow(
|
1088
|
+
GetIOVecPointer(from_iov_index, from_iov_offset),
|
1089
|
+
GetIOVecPointer(curr_iov_index_, curr_iov_written_),
|
1090
|
+
GetIOVecPointer(curr_iov_index_, curr_iov_written_) + to_copy);
|
1108
1091
|
curr_iov_written_ += to_copy;
|
1109
1092
|
from_iov_offset += to_copy;
|
1110
1093
|
total_written_ += to_copy;
|
@@ -1115,6 +1098,7 @@ class SnappyIOVecWriter {
|
|
1115
1098
|
return true;
|
1116
1099
|
}
|
1117
1100
|
|
1101
|
+
inline void Flush() {}
|
1118
1102
|
};
|
1119
1103
|
|
1120
1104
|
bool RawUncompressToIOVec(const char* compressed, size_t compressed_length,
|
@@ -1145,7 +1129,8 @@ class SnappyArrayWriter {
|
|
1145
1129
|
public:
|
1146
1130
|
inline explicit SnappyArrayWriter(char* dst)
|
1147
1131
|
: base_(dst),
|
1148
|
-
op_(dst)
|
1132
|
+
op_(dst),
|
1133
|
+
op_limit_(dst) {
|
1149
1134
|
}
|
1150
1135
|
|
1151
1136
|
inline void SetExpectedLength(size_t len) {
|
@@ -1172,8 +1157,7 @@ class SnappyArrayWriter {
|
|
1172
1157
|
const size_t space_left = op_limit_ - op;
|
1173
1158
|
if (len <= 16 && available >= 16 + kMaximumTagLength && space_left >= 16) {
|
1174
1159
|
// Fast path, used for the majority (about 95%) of invocations.
|
1175
|
-
|
1176
|
-
UnalignedCopy64(ip + 8, op + 8);
|
1160
|
+
UnalignedCopy128(ip, op);
|
1177
1161
|
op_ = op + len;
|
1178
1162
|
return true;
|
1179
1163
|
} else {
|
@@ -1182,8 +1166,7 @@ class SnappyArrayWriter {
|
|
1182
1166
|
}
|
1183
1167
|
|
1184
1168
|
inline bool AppendFromSelf(size_t offset, size_t len) {
|
1185
|
-
char*
|
1186
|
-
const size_t space_left = op_limit_ - op;
|
1169
|
+
char* const op_end = op_ + len;
|
1187
1170
|
|
1188
1171
|
// Check if we try to append from before the start of the buffer.
|
1189
1172
|
// Normally this would just be a check for "produced < offset",
|
@@ -1192,29 +1175,16 @@ class SnappyArrayWriter {
|
|
1192
1175
|
// to a very big number. This is convenient, as offset==0 is another
|
1193
1176
|
// invalid case that we also want to catch, so that we do not go
|
1194
1177
|
// into an infinite loop.
|
1195
|
-
|
1196
|
-
|
1197
|
-
if (produced <= offset - 1u) {
|
1198
|
-
return false;
|
1199
|
-
}
|
1200
|
-
if (len <= 16 && offset >= 8 && space_left >= 16) {
|
1201
|
-
// Fast path, used for the majority (70-80%) of dynamic invocations.
|
1202
|
-
UnalignedCopy64(op - offset, op);
|
1203
|
-
UnalignedCopy64(op - offset + 8, op + 8);
|
1204
|
-
} else {
|
1205
|
-
if (space_left >= len + kMaxIncrementCopyOverflow) {
|
1206
|
-
IncrementalCopyFastPath(op - offset, op, len);
|
1207
|
-
} else {
|
1208
|
-
if (space_left < len) {
|
1209
|
-
return false;
|
1210
|
-
}
|
1211
|
-
IncrementalCopy(op - offset, op, len);
|
1212
|
-
}
|
1213
|
-
}
|
1178
|
+
if (Produced() <= offset - 1u || op_end > op_limit_) return false;
|
1179
|
+
op_ = IncrementalCopy(op_ - offset, op_, op_end, op_limit_);
|
1214
1180
|
|
1215
|
-
op_ = op + len;
|
1216
1181
|
return true;
|
1217
1182
|
}
|
1183
|
+
inline size_t Produced() const {
|
1184
|
+
assert(op_ >= base_);
|
1185
|
+
return op_ - base_;
|
1186
|
+
}
|
1187
|
+
inline void Flush() {}
|
1218
1188
|
};
|
1219
1189
|
|
1220
1190
|
bool RawUncompress(const char* compressed, size_t n, char* uncompressed) {
|
@@ -1241,7 +1211,6 @@ bool Uncompress(const char* compressed, size_t n, string* uncompressed) {
|
|
1241
1211
|
return RawUncompress(compressed, n, string_as_array(uncompressed));
|
1242
1212
|
}
|
1243
1213
|
|
1244
|
-
|
1245
1214
|
// A Writer that drops everything on the floor and just does validation
|
1246
1215
|
class SnappyDecompressionValidator {
|
1247
1216
|
private:
|
@@ -1249,7 +1218,7 @@ class SnappyDecompressionValidator {
|
|
1249
1218
|
size_t produced_;
|
1250
1219
|
|
1251
1220
|
public:
|
1252
|
-
inline SnappyDecompressionValidator() : produced_(0) { }
|
1221
|
+
inline SnappyDecompressionValidator() : expected_(0), produced_(0) { }
|
1253
1222
|
inline void SetExpectedLength(size_t len) {
|
1254
1223
|
expected_ = len;
|
1255
1224
|
}
|
@@ -1270,6 +1239,7 @@ class SnappyDecompressionValidator {
|
|
1270
1239
|
produced_ += len;
|
1271
1240
|
return produced_ <= expected_;
|
1272
1241
|
}
|
1242
|
+
inline void Flush() {}
|
1273
1243
|
};
|
1274
1244
|
|
1275
1245
|
bool IsValidCompressedBuffer(const char* compressed, size_t n) {
|
@@ -1278,6 +1248,11 @@ bool IsValidCompressedBuffer(const char* compressed, size_t n) {
|
|
1278
1248
|
return InternalUncompress(&reader, &writer);
|
1279
1249
|
}
|
1280
1250
|
|
1251
|
+
bool IsValidCompressed(Source* compressed) {
|
1252
|
+
SnappyDecompressionValidator writer;
|
1253
|
+
return InternalUncompress(compressed, &writer);
|
1254
|
+
}
|
1255
|
+
|
1281
1256
|
void RawCompress(const char* input,
|
1282
1257
|
size_t input_length,
|
1283
1258
|
char* compressed,
|
@@ -1292,7 +1267,7 @@ void RawCompress(const char* input,
|
|
1292
1267
|
|
1293
1268
|
size_t Compress(const char* input, size_t input_length, string* compressed) {
|
1294
1269
|
// Pre-grow the buffer to the max length of the compressed output
|
1295
|
-
compressed
|
1270
|
+
STLStringResizeUninitialized(compressed, MaxCompressedLength(input_length));
|
1296
1271
|
|
1297
1272
|
size_t compressed_length;
|
1298
1273
|
RawCompress(input, input_length, string_as_array(compressed),
|
@@ -1301,6 +1276,240 @@ size_t Compress(const char* input, size_t input_length, string* compressed) {
|
|
1301
1276
|
return compressed_length;
|
1302
1277
|
}
|
1303
1278
|
|
1279
|
+
// -----------------------------------------------------------------------
|
1280
|
+
// Sink interface
|
1281
|
+
// -----------------------------------------------------------------------
|
1304
1282
|
|
1305
|
-
|
1283
|
+
// A type that decompresses into a Sink. The template parameter
|
1284
|
+
// Allocator must export one method "char* Allocate(int size);", which
|
1285
|
+
// allocates a buffer of "size" and appends that to the destination.
|
1286
|
+
template <typename Allocator>
|
1287
|
+
class SnappyScatteredWriter {
|
1288
|
+
Allocator allocator_;
|
1289
|
+
|
1290
|
+
// We need random access into the data generated so far. Therefore
|
1291
|
+
// we keep track of all of the generated data as an array of blocks.
|
1292
|
+
// All of the blocks except the last have length kBlockSize.
|
1293
|
+
std::vector<char*> blocks_;
|
1294
|
+
size_t expected_;
|
1295
|
+
|
1296
|
+
// Total size of all fully generated blocks so far
|
1297
|
+
size_t full_size_;
|
1298
|
+
|
1299
|
+
// Pointer into current output block
|
1300
|
+
char* op_base_; // Base of output block
|
1301
|
+
char* op_ptr_; // Pointer to next unfilled byte in block
|
1302
|
+
char* op_limit_; // Pointer just past block
|
1303
|
+
|
1304
|
+
inline size_t Size() const {
|
1305
|
+
return full_size_ + (op_ptr_ - op_base_);
|
1306
|
+
}
|
1307
|
+
|
1308
|
+
bool SlowAppend(const char* ip, size_t len);
|
1309
|
+
bool SlowAppendFromSelf(size_t offset, size_t len);
|
1310
|
+
|
1311
|
+
public:
|
1312
|
+
inline explicit SnappyScatteredWriter(const Allocator& allocator)
|
1313
|
+
: allocator_(allocator),
|
1314
|
+
full_size_(0),
|
1315
|
+
op_base_(NULL),
|
1316
|
+
op_ptr_(NULL),
|
1317
|
+
op_limit_(NULL) {
|
1318
|
+
}
|
1319
|
+
|
1320
|
+
inline void SetExpectedLength(size_t len) {
|
1321
|
+
assert(blocks_.empty());
|
1322
|
+
expected_ = len;
|
1323
|
+
}
|
1324
|
+
|
1325
|
+
inline bool CheckLength() const {
|
1326
|
+
return Size() == expected_;
|
1327
|
+
}
|
1328
|
+
|
1329
|
+
// Return the number of bytes actually uncompressed so far
|
1330
|
+
inline size_t Produced() const {
|
1331
|
+
return Size();
|
1332
|
+
}
|
1333
|
+
|
1334
|
+
inline bool Append(const char* ip, size_t len) {
|
1335
|
+
size_t avail = op_limit_ - op_ptr_;
|
1336
|
+
if (len <= avail) {
|
1337
|
+
// Fast path
|
1338
|
+
memcpy(op_ptr_, ip, len);
|
1339
|
+
op_ptr_ += len;
|
1340
|
+
return true;
|
1341
|
+
} else {
|
1342
|
+
return SlowAppend(ip, len);
|
1343
|
+
}
|
1344
|
+
}
|
1345
|
+
|
1346
|
+
inline bool TryFastAppend(const char* ip, size_t available, size_t length) {
|
1347
|
+
char* op = op_ptr_;
|
1348
|
+
const int space_left = op_limit_ - op;
|
1349
|
+
if (length <= 16 && available >= 16 + kMaximumTagLength &&
|
1350
|
+
space_left >= 16) {
|
1351
|
+
// Fast path, used for the majority (about 95%) of invocations.
|
1352
|
+
UnalignedCopy128(ip, op);
|
1353
|
+
op_ptr_ = op + length;
|
1354
|
+
return true;
|
1355
|
+
} else {
|
1356
|
+
return false;
|
1357
|
+
}
|
1358
|
+
}
|
1359
|
+
|
1360
|
+
inline bool AppendFromSelf(size_t offset, size_t len) {
|
1361
|
+
char* const op_end = op_ptr_ + len;
|
1362
|
+
// See SnappyArrayWriter::AppendFromSelf for an explanation of
|
1363
|
+
// the "offset - 1u" trick.
|
1364
|
+
if (SNAPPY_PREDICT_TRUE(offset - 1u < op_ptr_ - op_base_ &&
|
1365
|
+
op_end <= op_limit_)) {
|
1366
|
+
// Fast path: src and dst in current block.
|
1367
|
+
op_ptr_ = IncrementalCopy(op_ptr_ - offset, op_ptr_, op_end, op_limit_);
|
1368
|
+
return true;
|
1369
|
+
}
|
1370
|
+
return SlowAppendFromSelf(offset, len);
|
1371
|
+
}
|
1372
|
+
|
1373
|
+
// Called at the end of the decompress. We ask the allocator
|
1374
|
+
// write all blocks to the sink.
|
1375
|
+
inline void Flush() { allocator_.Flush(Produced()); }
|
1376
|
+
};
|
1377
|
+
|
1378
|
+
template<typename Allocator>
|
1379
|
+
bool SnappyScatteredWriter<Allocator>::SlowAppend(const char* ip, size_t len) {
|
1380
|
+
size_t avail = op_limit_ - op_ptr_;
|
1381
|
+
while (len > avail) {
|
1382
|
+
// Completely fill this block
|
1383
|
+
memcpy(op_ptr_, ip, avail);
|
1384
|
+
op_ptr_ += avail;
|
1385
|
+
assert(op_limit_ - op_ptr_ == 0);
|
1386
|
+
full_size_ += (op_ptr_ - op_base_);
|
1387
|
+
len -= avail;
|
1388
|
+
ip += avail;
|
1389
|
+
|
1390
|
+
// Bounds check
|
1391
|
+
if (full_size_ + len > expected_) {
|
1392
|
+
return false;
|
1393
|
+
}
|
1394
|
+
|
1395
|
+
// Make new block
|
1396
|
+
size_t bsize = std::min<size_t>(kBlockSize, expected_ - full_size_);
|
1397
|
+
op_base_ = allocator_.Allocate(bsize);
|
1398
|
+
op_ptr_ = op_base_;
|
1399
|
+
op_limit_ = op_base_ + bsize;
|
1400
|
+
blocks_.push_back(op_base_);
|
1401
|
+
avail = bsize;
|
1402
|
+
}
|
1403
|
+
|
1404
|
+
memcpy(op_ptr_, ip, len);
|
1405
|
+
op_ptr_ += len;
|
1406
|
+
return true;
|
1407
|
+
}
|
1408
|
+
|
1409
|
+
template<typename Allocator>
|
1410
|
+
bool SnappyScatteredWriter<Allocator>::SlowAppendFromSelf(size_t offset,
|
1411
|
+
size_t len) {
|
1412
|
+
// Overflow check
|
1413
|
+
// See SnappyArrayWriter::AppendFromSelf for an explanation of
|
1414
|
+
// the "offset - 1u" trick.
|
1415
|
+
const size_t cur = Size();
|
1416
|
+
if (offset - 1u >= cur) return false;
|
1417
|
+
if (expected_ - cur < len) return false;
|
1418
|
+
|
1419
|
+
// Currently we shouldn't ever hit this path because Compress() chops the
|
1420
|
+
// input into blocks and does not create cross-block copies. However, it is
|
1421
|
+
// nice if we do not rely on that, since we can get better compression if we
|
1422
|
+
// allow cross-block copies and thus might want to change the compressor in
|
1423
|
+
// the future.
|
1424
|
+
size_t src = cur - offset;
|
1425
|
+
while (len-- > 0) {
|
1426
|
+
char c = blocks_[src >> kBlockLog][src & (kBlockSize-1)];
|
1427
|
+
Append(&c, 1);
|
1428
|
+
src++;
|
1429
|
+
}
|
1430
|
+
return true;
|
1431
|
+
}
|
1432
|
+
|
1433
|
+
class SnappySinkAllocator {
|
1434
|
+
public:
|
1435
|
+
explicit SnappySinkAllocator(Sink* dest): dest_(dest) {}
|
1436
|
+
~SnappySinkAllocator() {}
|
1437
|
+
|
1438
|
+
char* Allocate(int size) {
|
1439
|
+
Datablock block(new char[size], size);
|
1440
|
+
blocks_.push_back(block);
|
1441
|
+
return block.data;
|
1442
|
+
}
|
1443
|
+
|
1444
|
+
// We flush only at the end, because the writer wants
|
1445
|
+
// random access to the blocks and once we hand the
|
1446
|
+
// block over to the sink, we can't access it anymore.
|
1447
|
+
// Also we don't write more than has been actually written
|
1448
|
+
// to the blocks.
|
1449
|
+
void Flush(size_t size) {
|
1450
|
+
size_t size_written = 0;
|
1451
|
+
size_t block_size;
|
1452
|
+
for (int i = 0; i < blocks_.size(); ++i) {
|
1453
|
+
block_size = std::min<size_t>(blocks_[i].size, size - size_written);
|
1454
|
+
dest_->AppendAndTakeOwnership(blocks_[i].data, block_size,
|
1455
|
+
&SnappySinkAllocator::Deleter, NULL);
|
1456
|
+
size_written += block_size;
|
1457
|
+
}
|
1458
|
+
blocks_.clear();
|
1459
|
+
}
|
1460
|
+
|
1461
|
+
private:
|
1462
|
+
struct Datablock {
|
1463
|
+
char* data;
|
1464
|
+
size_t size;
|
1465
|
+
Datablock(char* p, size_t s) : data(p), size(s) {}
|
1466
|
+
};
|
1467
|
+
|
1468
|
+
static void Deleter(void* arg, const char* bytes, size_t size) {
|
1469
|
+
delete[] bytes;
|
1470
|
+
}
|
1471
|
+
|
1472
|
+
Sink* dest_;
|
1473
|
+
std::vector<Datablock> blocks_;
|
1474
|
+
|
1475
|
+
// Note: copying this object is allowed
|
1476
|
+
};
|
1306
1477
|
|
1478
|
+
size_t UncompressAsMuchAsPossible(Source* compressed, Sink* uncompressed) {
|
1479
|
+
SnappySinkAllocator allocator(uncompressed);
|
1480
|
+
SnappyScatteredWriter<SnappySinkAllocator> writer(allocator);
|
1481
|
+
InternalUncompress(compressed, &writer);
|
1482
|
+
return writer.Produced();
|
1483
|
+
}
|
1484
|
+
|
1485
|
+
bool Uncompress(Source* compressed, Sink* uncompressed) {
|
1486
|
+
// Read the uncompressed length from the front of the compressed input
|
1487
|
+
SnappyDecompressor decompressor(compressed);
|
1488
|
+
uint32 uncompressed_len = 0;
|
1489
|
+
if (!decompressor.ReadUncompressedLength(&uncompressed_len)) {
|
1490
|
+
return false;
|
1491
|
+
}
|
1492
|
+
|
1493
|
+
char c;
|
1494
|
+
size_t allocated_size;
|
1495
|
+
char* buf = uncompressed->GetAppendBufferVariable(
|
1496
|
+
1, uncompressed_len, &c, 1, &allocated_size);
|
1497
|
+
|
1498
|
+
const size_t compressed_len = compressed->Available();
|
1499
|
+
// If we can get a flat buffer, then use it, otherwise do block by block
|
1500
|
+
// uncompression
|
1501
|
+
if (allocated_size >= uncompressed_len) {
|
1502
|
+
SnappyArrayWriter writer(buf);
|
1503
|
+
bool result = InternalUncompressAllTags(&decompressor, &writer,
|
1504
|
+
compressed_len, uncompressed_len);
|
1505
|
+
uncompressed->Append(buf, writer.Produced());
|
1506
|
+
return result;
|
1507
|
+
} else {
|
1508
|
+
SnappySinkAllocator allocator(uncompressed);
|
1509
|
+
SnappyScatteredWriter<SnappySinkAllocator> writer(allocator);
|
1510
|
+
return InternalUncompressAllTags(&decompressor, &writer, compressed_len,
|
1511
|
+
uncompressed_len);
|
1512
|
+
}
|
1513
|
+
}
|
1514
|
+
|
1515
|
+
} // end namespace snappy
|