1 | /* |
1 | /* |
2 | * libecb - http://software.schmorp.de/pkg/libecb |
2 | * libecb - http://software.schmorp.de/pkg/libecb |
3 | * |
3 | * |
4 | * Copyright (©) 2009-2015,2018-2020 Marc Alexander Lehmann <libecb@schmorp.de> |
4 | * Copyright (©) 2009-2015,2018-2021 Marc Alexander Lehmann <libecb@schmorp.de> |
5 | * Copyright (©) 2011 Emanuele Giaquinta |
5 | * Copyright (©) 2011 Emanuele Giaquinta |
6 | * All rights reserved. |
6 | * All rights reserved. |
7 | * |
7 | * |
8 | * Redistribution and use in source and binary forms, with or without modifica- |
8 | * Redistribution and use in source and binary forms, with or without modifica- |
9 | * tion, are permitted provided that the following conditions are met: |
9 | * tion, are permitted provided that the following conditions are met: |
… | |
… | |
40 | |
40 | |
41 | #ifndef ECB_H |
41 | #ifndef ECB_H |
42 | #define ECB_H |
42 | #define ECB_H |
43 | |
43 | |
44 | /* 16 bits major, 16 bits minor */ |
44 | /* 16 bits major, 16 bits minor */ |
45 | #define ECB_VERSION 0x00010008 |
45 | #define ECB_VERSION 0x00010009 |
46 | |
46 | |
47 | #include <string.h> /* for memcpy */ |
47 | #include <string.h> /* for memcpy */ |
48 | |
48 | |
49 | #if defined (_WIN32) && !defined (__MINGW32__) |
49 | #if defined (_WIN32) && !defined (__MINGW32__) |
50 | typedef signed char int8_t; |
50 | typedef signed char int8_t; |
… | |
… | |
102 | #if _ILP32 |
102 | #if _ILP32 |
103 | #define ECB_AMD64_X32 1 |
103 | #define ECB_AMD64_X32 1 |
104 | #else |
104 | #else |
105 | #define ECB_AMD64 1 |
105 | #define ECB_AMD64 1 |
106 | #endif |
106 | #endif |
|
|
107 | #endif |
|
|
108 | |
|
|
109 | #if ECB_PTRSIZE >= 8 || ECB_AMD64_X32 |
|
|
110 | #define ECB_64BIT_NATIVE 1 |
|
|
111 | #else |
|
|
112 | #define ECB_64BIT_NATIVE 0 |
107 | #endif |
113 | #endif |
108 | |
114 | |
109 | /* many compilers define _GNUC_ to some versions but then only implement |
115 | /* many compilers define _GNUC_ to some versions but then only implement |
110 | * what their idiot authors think are the "more important" extensions, |
116 | * what their idiot authors think are the "more important" extensions, |
111 | * causing enormous grief in return for some better fake benchmark numbers. |
117 | * causing enormous grief in return for some better fake benchmark numbers. |
… | |
… | |
242 | #if ECB_GCC_VERSION(4,7) |
248 | #if ECB_GCC_VERSION(4,7) |
243 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
249 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
244 | #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) |
250 | #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) |
245 | #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) |
251 | #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) |
246 | #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) |
252 | #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) |
|
|
253 | #undef ECB_MEMORY_FENCE_RELAXED |
247 | #define ECB_MEMORY_FENCE_RELAXED __atomic_thread_fence (__ATOMIC_RELAXED) |
254 | #define ECB_MEMORY_FENCE_RELAXED __atomic_thread_fence (__ATOMIC_RELAXED) |
248 | |
255 | |
249 | #elif ECB_CLANG_EXTENSION(c_atomic) |
256 | #elif ECB_CLANG_EXTENSION(c_atomic) |
250 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
257 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
251 | #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) |
258 | #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) |
252 | #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) |
259 | #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) |
253 | #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) |
260 | #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) |
|
|
261 | #undef ECB_MEMORY_FENCE_RELAXED |
254 | #define ECB_MEMORY_FENCE_RELAXED __c11_atomic_thread_fence (__ATOMIC_RELAXED) |
262 | #define ECB_MEMORY_FENCE_RELAXED __c11_atomic_thread_fence (__ATOMIC_RELAXED) |
255 | |
263 | |
256 | #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
264 | #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
257 | #define ECB_MEMORY_FENCE __sync_synchronize () |
265 | #define ECB_MEMORY_FENCE __sync_synchronize () |
258 | #elif _MSC_VER >= 1500 /* VC++ 2008 */ |
266 | #elif _MSC_VER >= 1500 /* VC++ 2008 */ |
… | |
… | |
938 | |
946 | |
939 | return s | 0x7c00 | m | !m; |
947 | return s | 0x7c00 | m | !m; |
940 | } |
948 | } |
941 | |
949 | |
942 | /*******************************************************************************/ |
950 | /*******************************************************************************/ |
|
|
951 | /* fast integer to ascii */ |
|
|
952 | |
|
|
953 | /* |
|
|
954 | * This code is pretty complicated because it is general. The idea behind it, |
|
|
955 | * however, is pretty simple: first, the number is multiplied with a scaling |
|
|
956 | * factor (2**bits / 10**(digits-1)) to convert the integer into a fixed-point |
|
|
957 | * number with the first digit in the upper bits. |
|
|
958 | * Then this digit is converted to text and masked out. The resulting number |
|
|
959 | * is then multiplied by 10, by multiplying the fixed point representation |
|
|
960 | * by 5 and shifting the (binary) decimal point one to the right, so a 4.28 |
|
|
961 | * format becomes 5.27, 6.26 and so on. |
|
|
962 | * The rest involves only advancing the pointer if we already generated a |
|
|
963 | * non-zero digit, so leading zeroes are overwritten. |
|
|
964 | */ |
|
|
965 | |
|
|
966 | // simply return a mask with "bits" bits set |
|
|
967 | #define ecb_i2a_mask(type,bits) ((((type)1) << (bits)) - 1) |
|
|
968 | |
|
|
969 | // oputput a single digit. maskvalue is 10**digitidx |
|
|
970 | #define ecb_i2a_digit(type,bits,digitmask,maskvalue,digitidx) \ |
|
|
971 | if (digitmask >= maskvalue) /* constant, used to decide how many digits to generate */ \ |
|
|
972 | { \ |
|
|
973 | char digit = x >> (bits - digitidx); /* calculate the topmost digit */ \ |
|
|
974 | *ptr = digit + '0'; /* output it */ \ |
|
|
975 | nz = (digitmask == maskvalue) || nz || digit; /* first term == always output last digit */ \ |
|
|
976 | ptr += nz; /* output digit only if non-zero digit seen */ \ |
|
|
977 | x = (x & ecb_i2a_mask (type, bits - digitidx)) * 5; /* *10, but shift decimal point right */ \ |
|
|
978 | } |
|
|
979 | |
|
|
980 | // convert integer to fixed point format and multiply out digits, highest first |
|
|
981 | // requires magic constants: max. digits and number of bits after the decimal point |
|
|
982 | #define ecb_i2a_def(suffix,ptr,v,type,bits,digitmask,lz) \ |
|
|
983 | ecb_inline char *ecb_i2a_ ## suffix (char *ptr, uint32_t u) \ |
|
|
984 | { \ |
|
|
985 | char nz = lz; /* non-zero digit seen? */ \ |
|
|
986 | /* convert to x.bits fixed-point */ \ |
|
|
987 | type x = u * ((ecb_i2a_mask (type, bits) + digitmask) / digitmask); \ |
|
|
988 | /* output up to 10 digits */ \ |
|
|
989 | ecb_i2a_digit (type,bits,digitmask, 1, 0); \ |
|
|
990 | ecb_i2a_digit (type,bits,digitmask, 10, 1); \ |
|
|
991 | ecb_i2a_digit (type,bits,digitmask, 100, 2); \ |
|
|
992 | ecb_i2a_digit (type,bits,digitmask, 1000, 3); \ |
|
|
993 | ecb_i2a_digit (type,bits,digitmask, 10000, 4); \ |
|
|
994 | ecb_i2a_digit (type,bits,digitmask, 100000, 5); \ |
|
|
995 | ecb_i2a_digit (type,bits,digitmask, 1000000, 6); \ |
|
|
996 | ecb_i2a_digit (type,bits,digitmask, 10000000, 7); \ |
|
|
997 | ecb_i2a_digit (type,bits,digitmask, 100000000, 8); \ |
|
|
998 | ecb_i2a_digit (type,bits,digitmask, 1000000000, 9); \ |
|
|
999 | return ptr; \ |
|
|
1000 | } |
|
|
1001 | |
|
|
1002 | // predefined versions of the above, for various digits |
|
|
1003 | // ecb_i2a_xN = almost N digits, limit defined by macro |
|
|
1004 | // ecb_i2a_N = up to N digits, leading zeroes suppressed |
|
|
1005 | // ecb_i2a_0N = exactly N digits, including leading zeroes |
|
|
1006 | |
|
|
1007 | // non-leading-zero versions, limited range |
|
|
1008 | #define ECB_I2A_MAX_X5 59074 // limit for ecb_i2a_x5 |
|
|
1009 | #define ECB_I2A_MAX_X10 2932500665 // limit for ecb_i2a_x10 |
|
|
1010 | ecb_i2a_def ( x5, ptr, v, uint32_t, 26, 10000, 0) |
|
|
1011 | ecb_i2a_def (x10, ptr, v, uint64_t, 60, 1000000000, 0) |
|
|
1012 | |
|
|
1013 | // non-leading zero versions, all digits, 4 and 9 are optimal for 32/64 bit |
|
|
1014 | ecb_i2a_def ( 2, ptr, v, uint32_t, 10, 10, 0) |
|
|
1015 | ecb_i2a_def ( 3, ptr, v, uint32_t, 12, 100, 0) |
|
|
1016 | ecb_i2a_def ( 4, ptr, v, uint32_t, 26, 1000, 0) |
|
|
1017 | ecb_i2a_def ( 5, ptr, v, uint64_t, 30, 10000, 0) |
|
|
1018 | ecb_i2a_def ( 6, ptr, v, uint64_t, 36, 100000, 0) |
|
|
1019 | ecb_i2a_def ( 7, ptr, v, uint64_t, 44, 1000000, 0) |
|
|
1020 | ecb_i2a_def ( 8, ptr, v, uint64_t, 50, 10000000, 0) |
|
|
1021 | ecb_i2a_def ( 9, ptr, v, uint64_t, 56, 100000000, 0) |
|
|
1022 | |
|
|
1023 | // leading-zero versions, all digits, 04 and 09 are optimal for 32/64 bit |
|
|
1024 | ecb_i2a_def (02, ptr, v, uint32_t, 10, 10, 1) |
|
|
1025 | ecb_i2a_def (03, ptr, v, uint32_t, 12, 100, 1) |
|
|
1026 | ecb_i2a_def (04, ptr, v, uint32_t, 26, 1000, 1) |
|
|
1027 | ecb_i2a_def (05, ptr, v, uint64_t, 30, 10000, 1) |
|
|
1028 | ecb_i2a_def (06, ptr, v, uint64_t, 36, 100000, 1) |
|
|
1029 | ecb_i2a_def (07, ptr, v, uint64_t, 44, 1000000, 1) |
|
|
1030 | ecb_i2a_def (08, ptr, v, uint64_t, 50, 10000000, 1) |
|
|
1031 | ecb_i2a_def (09, ptr, v, uint64_t, 56, 100000000, 1) |
|
|
1032 | |
|
|
1033 | #define ECB_I2A_I32_DIGITS 11 |
|
|
1034 | #define ECB_I2A_U32_DIGITS 10 |
|
|
1035 | #define ECB_I2A_I64_DIGITS 20 |
|
|
1036 | #define ECB_I2A_U64_DIGITS 21 |
|
|
1037 | #define ECB_I2A_MAX_DIGITS 21 |
|
|
1038 | |
|
|
1039 | ecb_inline char * |
|
|
1040 | ecb_i2a_u32 (char *ptr, uint32_t u) |
|
|
1041 | { |
|
|
1042 | #if ECB_64BIT_NATIVE |
|
|
1043 | if (ecb_expect_true (u <= ECB_I2A_MAX_X10)) |
|
|
1044 | ptr = ecb_i2a_x10 (ptr, u); |
|
|
1045 | else // x10 almost, but not fully, covers 32 bit |
|
|
1046 | { |
|
|
1047 | uint32_t u1 = u % 1000000000; |
|
|
1048 | uint32_t u2 = u / 1000000000; |
|
|
1049 | |
|
|
1050 | *ptr++ = u2 + '0'; |
|
|
1051 | ptr = ecb_i2a_09 (ptr, u1); |
|
|
1052 | } |
|
|
1053 | #else |
|
|
1054 | if (ecb_expect_true (u <= ECB_I2A_MAX_X5)) |
|
|
1055 | ecb_i2a_x5 (ptr, u); |
|
|
1056 | else if (ecb_expect_true (u <= ECB_I2A_MAX_X5 * 10000)) |
|
|
1057 | { |
|
|
1058 | uint32_t u1 = u % 10000; |
|
|
1059 | uint32_t u2 = u / 10000; |
|
|
1060 | |
|
|
1061 | ptr = ecb_i2a_x5 (ptr, u2); |
|
|
1062 | ptr = ecb_i2a_04 (ptr, u1); |
|
|
1063 | } |
|
|
1064 | else |
|
|
1065 | { |
|
|
1066 | uint32_t u1 = u % 10000; |
|
|
1067 | uint32_t ua = u / 10000; |
|
|
1068 | uint32_t u2 = ua % 10000; |
|
|
1069 | uint32_t u3 = ua / 10000; |
|
|
1070 | |
|
|
1071 | ptr = ecb_i2a_2 (ptr, u3); |
|
|
1072 | ptr = ecb_i2a_04 (ptr, u2); |
|
|
1073 | ptr = ecb_i2a_04 (ptr, u1); |
|
|
1074 | } |
|
|
1075 | #endif |
|
|
1076 | |
|
|
1077 | return ptr; |
|
|
1078 | } |
|
|
1079 | |
|
|
1080 | ecb_inline char * |
|
|
1081 | ecb_i2a_i32 (char *ptr, int32_t v) |
|
|
1082 | { |
|
|
1083 | *ptr = '-'; ptr += v < 0; |
|
|
1084 | uint32_t u = v < 0 ? -(uint32_t)v : v; |
|
|
1085 | |
|
|
1086 | #if ECB_64BIT_NATIVE |
|
|
1087 | ptr = ecb_i2a_x10 (ptr, u); // x10 fully covers 31 bit |
|
|
1088 | #else |
|
|
1089 | ptr = ecb_i2a_u32 (ptr, u); |
|
|
1090 | #endif |
|
|
1091 | |
|
|
1092 | return ptr; |
|
|
1093 | } |
|
|
1094 | |
|
|
1095 | ecb_inline char * |
|
|
1096 | ecb_i2a_u64 (char *ptr, uint64_t u) |
|
|
1097 | { |
|
|
1098 | #if ECB_64BIT_NATIVE |
|
|
1099 | if (ecb_expect_true (u <= ECB_I2A_MAX_X10)) |
|
|
1100 | ptr = ecb_i2a_x10 (ptr, u); |
|
|
1101 | else if (ecb_expect_false (u <= ECB_I2A_MAX_X10 * 1000000000)) |
|
|
1102 | { |
|
|
1103 | uint64_t u1 = u % 1000000000; |
|
|
1104 | uint64_t u2 = u / 1000000000; |
|
|
1105 | |
|
|
1106 | ptr = ecb_i2a_x10 (ptr, u2); |
|
|
1107 | ptr = ecb_i2a_09 (ptr, u1); |
|
|
1108 | } |
|
|
1109 | else |
|
|
1110 | { |
|
|
1111 | uint64_t u1 = u % 1000000000; |
|
|
1112 | uint64_t ua = u / 1000000000; |
|
|
1113 | uint64_t u2 = ua % 1000000000; |
|
|
1114 | uint64_t u3 = ua / 1000000000; |
|
|
1115 | |
|
|
1116 | ptr = ecb_i2a_2 (ptr, u3); |
|
|
1117 | ptr = ecb_i2a_09 (ptr, u2); |
|
|
1118 | ptr = ecb_i2a_09 (ptr, u1); |
|
|
1119 | } |
|
|
1120 | #else |
|
|
1121 | if (ecb_expect_true (u <= ECB_I2A_MAX_X5)) |
|
|
1122 | ptr = ecb_i2a_x5 (ptr, u); |
|
|
1123 | else |
|
|
1124 | { |
|
|
1125 | uint64_t u1 = u % 10000; |
|
|
1126 | uint64_t u2 = u / 10000; |
|
|
1127 | |
|
|
1128 | ptr = ecb_i2a_u64 (ptr, u2); |
|
|
1129 | ptr = ecb_i2a_04 (ptr, u1); |
|
|
1130 | } |
|
|
1131 | #endif |
|
|
1132 | |
|
|
1133 | return ptr; |
|
|
1134 | } |
|
|
1135 | |
|
|
1136 | ecb_inline char * |
|
|
1137 | ecb_i2a_i64 (char *ptr, int64_t v) |
|
|
1138 | { |
|
|
1139 | *ptr = '-'; ptr += v < 0; |
|
|
1140 | uint64_t u = v < 0 ? -(uint64_t)v : v; |
|
|
1141 | |
|
|
1142 | #if ECB_64BIT_NATIVE |
|
|
1143 | if (ecb_expect_true (u <= ECB_I2A_MAX_X10)) |
|
|
1144 | ptr = ecb_i2a_x10 (ptr, u); |
|
|
1145 | else if (ecb_expect_false (u <= ECB_I2A_MAX_X10 * 1000000000)) |
|
|
1146 | { |
|
|
1147 | uint64_t u1 = u % 1000000000; |
|
|
1148 | uint64_t u2 = u / 1000000000; |
|
|
1149 | |
|
|
1150 | ptr = ecb_i2a_x10 (ptr, u2); |
|
|
1151 | ptr = ecb_i2a_09 (ptr, u1); |
|
|
1152 | } |
|
|
1153 | else |
|
|
1154 | { |
|
|
1155 | uint64_t u1 = u % 1000000000; |
|
|
1156 | uint64_t ua = u / 1000000000; |
|
|
1157 | uint64_t u2 = ua % 1000000000; |
|
|
1158 | uint64_t u3 = ua / 1000000000; |
|
|
1159 | |
|
|
1160 | // 2**31 is 19 digits, so the top is exactly one digit |
|
|
1161 | *ptr++ = u3 + '0'; |
|
|
1162 | ptr = ecb_i2a_09 (ptr, u2); |
|
|
1163 | ptr = ecb_i2a_09 (ptr, u1); |
|
|
1164 | } |
|
|
1165 | #else |
|
|
1166 | ptr = ecb_i2a_u64 (ptr, u); |
|
|
1167 | #endif |
|
|
1168 | |
|
|
1169 | return ptr; |
|
|
1170 | } |
|
|
1171 | |
|
|
1172 | /*******************************************************************************/ |
943 | /* floating point stuff, can be disabled by defining ECB_NO_LIBM */ |
1173 | /* floating point stuff, can be disabled by defining ECB_NO_LIBM */ |
944 | |
1174 | |
945 | /* basically, everything uses "ieee pure-endian" floating point numbers */ |
1175 | /* basically, everything uses "ieee pure-endian" floating point numbers */ |
946 | /* the only noteworthy exception is ancient armle, which uses order 43218765 */ |
1176 | /* the only noteworthy exception is ancient armle, which uses order 43218765 */ |
947 | #if 0 \ |
1177 | #if 0 \ |