• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* insert_string.h
2  *
3  * Copyright 2019 The Chromium Authors. All rights reserved.
4  * Use of this source code is governed by a BSD-style license that can be
5  * found in the Chromium source repository LICENSE file.
6  */
7 #ifdef _MSC_VER
8 #define INLINE __inline
9 #else
10 #define INLINE inline
11 #endif
12 
13 /* Optimized insert_string block */
14 #if defined(CRC32_SIMD_SSE42_PCLMUL) || defined(CRC32_ARMV8_CRC32)
15 #define TARGET_CPU_WITH_CRC
16 // clang-format off
17 #if defined(CRC32_SIMD_SSE42_PCLMUL)
18   /* Required to make MSVC bot build pass. */
19   #include <smmintrin.h>
20   #if defined(__GNUC__) || defined(__clang__)
21     #undef TARGET_CPU_WITH_CRC
22     #define TARGET_CPU_WITH_CRC __attribute__((target("sse4.2")))
23   #endif
24 
25   #define _cpu_crc32_u32 _mm_crc32_u32
26 
27 #elif defined(CRC32_ARMV8_CRC32)
28   #include "arm_features.h"
29   #if defined(__clang__)
30     #undef TARGET_CPU_WITH_CRC
31     #define __crc32cw __builtin_arm_crc32cw
32   #endif
33 
34   #define _cpu_crc32_u32 __crc32cw
35 
36   #if defined(__aarch64__)
37     #define TARGET_CPU_WITH_CRC __attribute__((target("crc")))
38   #else  // !defined(__aarch64__)
39     #define TARGET_CPU_WITH_CRC __attribute__((target("armv8-a,crc")))
40   #endif  // defined(__aarch64__)
41 #endif
42 // clang-format on
43 TARGET_CPU_WITH_CRC
insert_string_optimized(deflate_state * const s,const Pos str)44 local INLINE Pos insert_string_optimized(deflate_state* const s,
45                                          const Pos str) {
46   Pos ret;
47   unsigned *ip, val, h = 0;
48 
49   ip = (unsigned*)&s->window[str];
50   val = *ip;
51 
52   if (s->level >= 6)
53     val &= 0xFFFFFF;
54 
55   /* Unlike the case of data integrity checks for GZIP format where the
56    * polynomial used is defined (https://tools.ietf.org/html/rfc1952#page-11),
57    * here it is just a hash function for the hash table used while
58    * performing compression.
59    */
60   h = _cpu_crc32_u32(h, val);
61 
62   ret = s->head[h & s->hash_mask];
63   s->head[h & s->hash_mask] = str;
64   s->prev[str & s->w_mask] = ret;
65   return ret;
66 }
67 #endif /* Optimized insert_string block */
68 
69 /* ===========================================================================
70  * Update a hash value with the given input byte
71  * IN  assertion: all calls to UPDATE_HASH are made with consecutive input
72  *    characters, so that a running hash key can be computed from the previous
73  *    key instead of complete recalculation each time.
74  */
75 #define UPDATE_HASH(s, h, c) (h = (((h) << s->hash_shift) ^ (c)) & s->hash_mask)
76 
77 /* ===========================================================================
78  * Insert string str in the dictionary and set match_head to the previous head
79  * of the hash chain (the most recent string with same hash key). Return
80  * the previous length of the hash chain.
81  * If this file is compiled with -DFASTEST, the compression level is forced
82  * to 1, and no hash chains are maintained.
83  * IN  assertion: all calls to INSERT_STRING are made with consecutive input
84  *    characters and the first MIN_MATCH bytes of str are valid (except for
85  *    the last MIN_MATCH-1 bytes of the input file).
86  */
insert_string_c(deflate_state * const s,const Pos str)87 local INLINE Pos insert_string_c(deflate_state* const s, const Pos str) {
88   Pos ret;
89 
90   UPDATE_HASH(s, s->ins_h, s->window[(str) + (MIN_MATCH - 1)]);
91 #ifdef FASTEST
92   ret = s->head[s->ins_h];
93 #else
94   ret = s->prev[str & s->w_mask] = s->head[s->ins_h];
95 #endif
96   s->head[s->ins_h] = str;
97 
98   return ret;
99 }
100 
insert_string(deflate_state * const s,const Pos str)101 local INLINE Pos insert_string(deflate_state* const s, const Pos str) {
102 /* String dictionary insertion: faster symbol hashing has a positive impact
103  * on data compression speeds (around 20% on Intel and 36% on Arm Cortex big
104  * cores).
105  * A misfeature is that the generated compressed output will differ from
106  * vanilla zlib (even though it is still valid 'DEFLATE-d' content).
107  *
108  * We offer here a way to disable the optimization if there is the expectation
109  * that compressed content should match when compared to vanilla zlib.
110  */
111 #if !defined(CHROMIUM_ZLIB_NO_CASTAGNOLI)
112   /* TODO(cavalcantii): unify CPU features code. */
113 #if defined(CRC32_ARMV8_CRC32)
114   if (arm_cpu_enable_crc32)
115     return insert_string_optimized(s, str);
116 #elif defined(CRC32_SIMD_SSE42_PCLMUL)
117   if (x86_cpu_enable_simd)
118     return insert_string_optimized(s, str);
119 #endif
120 #endif
121   return insert_string_c(s, str);
122 }
123