...

Text file src/github.com/bytedance/sonic/native/atof_native.c

Documentation: github.com/bytedance/sonic/native

     1/*
     2 * Copyright 2021 ByteDance Inc.
     3 *
     4 * Licensed under the Apache License, Version 2.0 (the "License");
     5 * you may not use this file except in compliance with the License.
     6 * You may obtain a copy of the License at
     7 *
     8 *     http://www.apache.org/licenses/LICENSE-2.0
     9 *
    10 * Unless required by applicable law or agreed to in writing, software
    11 * distributed under the License is distributed on an "AS IS" BASIS,
    12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    13 * See the License for the specific language governing permissions and
    14 * limitations under the License.
    15 */
    16
    17#include "native.h"
    18
    19/* decimical shift witout overflow, e.g. 9 << 61 overflow */
    20#define MAX_SHIFT 60
    21
    22/* Decimal represent the integer or float
    23 * example 1: 1.1   {"11", 2, 1, 0}
    24 * example 2: -0.1  {"1", 1, 0, 1}
    25 * example 3: 999   {"999", 3, 3, 0}
    26 */
    27typedef struct Decimal {
    28    char*  d;
    29    size_t cap;
    30    int    nd;
    31    int    dp;
    32    int    neg;
    33    int    trunc;
    34} Decimal;
    35
    36/* decimal power of ten to binary power of two.
    37 * For example: POW_TAB[1]: 10 ** 1 ~ 2 ** 3
    38 */
    39static const int POW_TAB[9] = {1, 3, 6, 9, 13, 16, 19, 23, 26};
    40
    41/* Left shift information for decimal.
    42 * For example, {2, "625"}.  That means that it will add 2 digits to the new decimal
    43 * when the prefix of decimal is from "625" to "999", and 1 digit from "0" to "624".
    44 */
    45typedef struct lshift_cheat  {
    46    int   delta;                             // number of added digits when left shift
    47    const char  cutoff[100];                 // minus one digit if under the half(cutoff).
    48} lshift_cheat;
    49
    50/* Look up for the decimal shift information by binary shift bits.
    51 * idx is shift bits for binary.
    52 * value is the shift information for decimal.
    53 * For example, idx is 4, the value is {2, "625"}.
    54 * That means the binary shift 4 bits left, will cause add 2 digits to the decimal
    55 * if the prefix of decimal is under "625".
    56 */
    57const static lshift_cheat LSHIFT_TAB[61];
    58
    59static inline void decimal_init(Decimal *d, char *dbuf, size_t cap) {
    60    d->d = dbuf;
    61    d->cap = cap;
    62    for (int i = 0; i < d->cap; ++i) {
    63        d->d[i] = 0;
    64    }
    65    d->dp    = 0;
    66    d->nd    = 0;
    67    d->neg   = 0;
    68    d->trunc = 0;
    69}
    70
    71static inline void decimal_set(Decimal *d, const char *s, ssize_t len, char *dbuf, ssize_t cap) {
    72    int i = 0;
    73
    74    decimal_init(d, dbuf, cap);
    75    if (s[i] == '-') {
    76        i++;
    77        d->neg = 1;
    78    }
    79
    80    int saw_dot = 0;
    81    for (; i < len; i++) {
    82        if ('0' <= s[i] && s[i] <= '9') {
    83            if (s[i] == '0' && d->nd == 0) { // ignore leading zeros
    84                d->dp--;
    85                continue;
    86            }
    87            if (d->nd < d->cap) {
    88                d->d[d->nd] = s[i];
    89                d->nd++;
    90            } else if (s[i] != '0') {
    91                /* truncat the remaining digits */
    92                d->trunc = 1;
    93            }
    94        } else if (s[i] == '.') {
    95            saw_dot = 1;
    96            d->dp = d->nd;
    97        } else {
    98            break;
    99        }
   100    }
   101
   102    /* integer */
   103    if (saw_dot == 0) {
   104        d->dp = d->nd;
   105    }
   106
   107    /* exponent */
   108    if (i < len && (s[i] == 'e' || s[i] == 'E')) {
   109        int exp = 0;
   110        int esgn = 1;
   111
   112        i++;
   113        if (s[i] == '+') {
   114            i++;
   115        } else if (s[i] == '-') {
   116            i++;
   117            esgn = -1;
   118        }
   119
   120        for (; i < len && ('0' <= s[i] && s[i] <= '9') && exp < 10000; i++) {
   121                exp = exp * 10 + (s[i] - '0');
   122        }
   123        d->dp += exp * esgn;
   124    }
   125
   126    return;
   127}
   128
   129/* trim trailing zeros from number */
   130static inline void trim(Decimal *d) {
   131    while (d->nd > 0 && d->d[d->nd - 1] == '0') {
   132        d->nd--;
   133    }
   134    if (d->nd == 0) {
   135        d->dp = 0;
   136    }
   137}
   138
   139/* Binary shift right (/ 2) by k bits.  k <= maxShift to avoid overflow */
   140static inline void right_shift(Decimal *d, uint32_t k) {
   141    int      r = 0; // read pointer
   142    int      w = 0; // write pointer
   143    uint64_t n = 0;
   144
   145    /* Pick up enough leading digits to cover first shift */
   146    for (; n >> k == 0; r++) {
   147        if (r >= d->nd) {
   148            if (n == 0) {
   149                d->nd = 0; // no digits for this num
   150                return;
   151            }
   152            /* until n has enough bits for right shift */
   153            while (n >> k == 0) {
   154                n *= 10;
   155                r++;
   156            }
   157            break;
   158        }
   159        n = n * 10 + d->d[r] - '0'; // read the value from d.d
   160    }
   161    d->dp -= r - 1; // point shift left
   162
   163    uint64_t mask = (1ull << k) - 1;
   164    uint64_t dig = 0;
   165
   166    /* Pick up a digit, put down a digit */
   167    for (; r < d->nd; r++) {
   168        dig = n >> k;
   169        n &= mask;
   170        d->d[w++] = (char)(dig + '0');
   171        n = n * 10 + d->d[r] - '0';
   172    }
   173
   174    /* Put down extra digits */
   175    while (n > 0) {
   176        dig = n >> k;
   177        n &= mask;
   178        if (w < d->cap) {
   179            d->d[w] = (char)(dig + '0');
   180            w++;
   181        } else if (dig > 0) {
   182            /* truncated */
   183            d->trunc = 1;
   184        }
   185        n *= 10;
   186    }
   187
   188    d->nd = w;
   189    trim(d);
   190}
   191
   192/* Compare the leading prefix, if b is lexicographically less, return 0 */
   193static inline bool prefix_is_less(const char *b, const char *s, uint64_t bn) {
   194    int i = 0;
   195    for (; i < bn; i++) {
   196        if (s[i] == '\0') {
   197            return false;
   198        }
   199        if (b[i] != s[i]) {
   200            return b[i] < s[i];
   201        }
   202    }
   203    return s[i] != '\0';
   204}
   205
   206/* Binary shift left (* 2) by k bits.  k <= maxShift to avoid overflow */
   207static inline void left_shift(Decimal *d, uint32_t k) {
   208    int delta = LSHIFT_TAB[k].delta;
   209
   210    if (prefix_is_less(d->d, LSHIFT_TAB[k].cutoff, d->nd)){
   211        delta--;
   212    }
   213
   214    int r = d->nd;         // read index
   215    int w = d->nd + delta; // write index
   216    uint64_t n = 0;
   217    uint64_t quo = 0;
   218    uint64_t rem = 0;
   219
   220    /* Pick up a digit, put down a digit */
   221    for (r--; r >= 0; r--) {
   222        n += (uint64_t)(d->d[r] - '0') << k;
   223        quo = n / 10;
   224        rem = n - 10 * quo;
   225        w--;
   226        if (w < d->cap) {
   227            d->d[w] = (char)(rem + '0');
   228        } else if (rem != 0) {
   229            /* truncated */
   230            d->trunc = 1;
   231        }
   232        n = quo;
   233    }
   234
   235    /* Put down extra digits */
   236    while (n > 0) {
   237        quo = n / 10;
   238        rem = n - 10 * quo;
   239        w--;
   240        if (w < d->cap) {
   241            d->d[w] = (char)(rem + '0');
   242        } else if (rem != 0) {
   243            /* truncated */
   244            d->trunc = 1;
   245        }
   246        n = quo;
   247    }
   248
   249    d->nd += delta;
   250    if (d->nd >= d->cap) {
   251        d->nd = d->cap;
   252    }
   253    d->dp += delta;
   254    trim(d);
   255}
   256
   257static inline void decimal_shift(Decimal *d, int k) {
   258    if (d->nd == 0 || k == 0) {
   259        return;
   260    }
   261
   262    if (k > 0) {
   263        while (k > MAX_SHIFT) {
   264            left_shift(d, MAX_SHIFT);
   265            k -= MAX_SHIFT;
   266        }
   267        if (k) {
   268            left_shift(d, k);
   269        }
   270    }
   271
   272    if (k < 0) {
   273        while (k < -MAX_SHIFT) {
   274            right_shift(d, MAX_SHIFT);
   275            k += MAX_SHIFT;
   276        }
   277        if (k) {
   278            right_shift(d, -k);
   279        }
   280    }
   281
   282}
   283
   284static inline int should_roundup(Decimal *d, int nd) {
   285    if (nd < 0 || nd >= d->nd) {
   286        return 0;
   287    }
   288
   289    /* Exactly halfway - round to even */
   290    if (d->d[nd] == '5' && nd+1 == d->nd) {
   291        if (d->trunc) {
   292            return 1;
   293        }
   294        return nd > 0 && (d->d[nd-1]-'0')%2 != 0;
   295    }
   296
   297    /* not halfway - round to the nearest */
   298    return d->d[nd] >= '5';
   299}
   300
   301/* Extract integer part, rounded appropriately */
   302static inline uint64_t rounded_integer(Decimal *d) {
   303    if (d->dp > 20) { // overflow
   304        return 0xFFFFFFFFFFFFFFFF; //64 bits
   305    }
   306
   307    int i = 0;
   308    uint64_t n = 0;
   309    for (i = 0; i < d->dp && i < d->nd; i++) {
   310        n = n * 10 + (d->d[i] - '0');
   311    }
   312    for (; i < d->dp; i++) {
   313        n *= 10;
   314    }
   315    if (should_roundup(d, d->dp)) {
   316        n++;
   317    }
   318    return n;
   319}
   320
   321int decimal_to_f64(Decimal *d, double *val) {
   322    int exp2 = 0;
   323    uint64_t mant = 0;
   324    uint64_t bits = 0;
   325
   326    /* d is zero */
   327    if (d->nd == 0) {
   328        mant = 0;
   329        exp2 = -1023;
   330        goto out;
   331    }
   332
   333    /* Overflow, return inf/INF */
   334    if (d->dp > 310) {
   335        goto overflow;
   336    }
   337    /* Underflow, return zero */
   338    if (d->dp < -330) {
   339        mant = 0;
   340        exp2 = -1023;
   341        goto out;
   342    }
   343
   344    /* Scale by powers of two until in range [0.5, 1.0) */
   345    int n = 0;
   346    while (d->dp > 0) { // d >= 1
   347        if (d->dp >= 9) {
   348            n = 27;
   349        } else {
   350            n = POW_TAB[d->dp];
   351        }
   352        decimal_shift(d, -n); // shift right
   353        exp2 += n;
   354    }
   355    while ((d->dp < 0) || ((d->dp == 0) && (d->d[0] < '5'))) { // d < 0.5
   356        if (-d->dp >= 9) {
   357            n = 27;
   358        } else {
   359            n = POW_TAB[-d->dp];
   360        }
   361        decimal_shift(d, n); // shift left
   362        exp2 -= n;
   363    }
   364
   365    /* Our range is [0.5,1) but floating point range is [1,2) */
   366    exp2 --;
   367
   368    /* Minimum exp2 for doulbe is -1022.
   369     * If the exponent is smaller, move it up and
   370     * adjust d accordingly.
   371     */
   372    if (exp2 < -1022) {
   373        n = -1022 - exp2;
   374        decimal_shift(d, -n); // shift right
   375        exp2 += n;
   376    }
   377
   378    /* Exp2 too large */
   379    if ((exp2 + 1023) >= 0x7FF) {
   380        goto overflow;
   381    }
   382
   383    /* Extract 53 bits. */
   384    decimal_shift(d, 53);  // shift left
   385    mant = rounded_integer(d);
   386
   387    /* Rounding might have added a bit; shift down. */
   388    if (mant == (2ull << 52)) { // mant has 54 bits
   389        mant >>= 1;
   390        exp2 ++;
   391        if ((exp2 + 1023) >= 0x7FF) {
   392            goto overflow;
   393        }
   394    }
   395
   396    /* Denormalized? */
   397    if ((mant & (1ull << 52)) == 0) {
   398        exp2 = -1023;
   399    }
   400    goto out;
   401
   402overflow:
   403    /* ±INF/inf */
   404    mant = 0;
   405    exp2 = 0x7FF - 1023;
   406
   407out:
   408    /* Assemble bits. */
   409    bits = mant & 0x000FFFFFFFFFFFFF;
   410    bits |= (uint64_t)((exp2 + 1023) & 0x7FF) << 52;
   411    if (d->neg) {
   412        bits |= 1ull << 63;
   413    }
   414    *(uint64_t*)val = bits;
   415    return 0;
   416}
   417
   418double atof_native(const char *sp, ssize_t nb, char* dbuf, ssize_t cap) {
   419    Decimal d;
   420    double val = 0;
   421    decimal_set(&d, sp, nb, dbuf, cap);
   422    decimal_to_f64(&d, &val);
   423    return val;
   424}
   425
   426#undef MAX_SHIFT
   427
   428const static lshift_cheat LSHIFT_TAB[61] = {
   429    // Leading digits of 1/2^i = 5^i.
   430    // 5^23 is not an exact 64-bit floating point number,
   431    // so have to use bc for the math.
   432    // Go up to 60 to be large enough for 32bit and 64bit platforms.
   433    /*
   434        seq 60 | sed 's/^/5^/' | bc |
   435        awk 'BEGIN{ print "\t{ 0, \"\" }," }
   436        {
   437            log2 = log(2)/log(10)
   438            printf("\t{ %d, \"%s\" },\t// * %d\n",
   439                int(log2*NR+1), $0, 2**NR)
   440        }'
   441    */
   442    {0, ""},
   443    {1, "5"},                                           // * 2
   444    {1, "25"},                                          // * 4
   445    {1, "125"},                                         // * 8
   446    {2, "625"},                                         // * 16
   447    {2, "3125"},                                        // * 32
   448    {2, "15625"},                                       // * 64
   449    {3, "78125"},                                       // * 128
   450    {3, "390625"},                                      // * 256
   451    {3, "1953125"},                                     // * 512
   452    {4, "9765625"},                                     // * 1024
   453    {4, "48828125"},                                    // * 2048
   454    {4, "244140625"},                                   // * 4096
   455    {4, "1220703125"},                                  // * 8192
   456    {5, "6103515625"},                                  // * 16384
   457    {5, "30517578125"},                                 // * 32768
   458    {5, "152587890625"},                                // * 65536
   459    {6, "762939453125"},                                // * 131072
   460    {6, "3814697265625"},                               // * 262144
   461    {6, "19073486328125"},                              // * 524288
   462    {7, "95367431640625"},                              // * 1048576
   463    {7, "476837158203125"},                             // * 2097152
   464    {7, "2384185791015625"},                            // * 4194304
   465    {7, "11920928955078125"},                           // * 8388608
   466    {8, "59604644775390625"},                           // * 16777216
   467    {8, "298023223876953125"},                          // * 33554432
   468    {8, "1490116119384765625"},                         // * 67108864
   469    {9, "7450580596923828125"},                         // * 134217728
   470    {9, "37252902984619140625"},                        // * 268435456
   471    {9, "186264514923095703125"},                       // * 536870912
   472    {10, "931322574615478515625"},                      // * 1073741824
   473    {10, "4656612873077392578125"},                     // * 2147483648
   474    {10, "23283064365386962890625"},                    // * 4294967296
   475    {10, "116415321826934814453125"},                   // * 8589934592
   476    {11, "582076609134674072265625"},                   // * 17179869184
   477    {11, "2910383045673370361328125"},                  // * 34359738368
   478    {11, "14551915228366851806640625"},                 // * 68719476736
   479    {12, "72759576141834259033203125"},                 // * 137438953472
   480    {12, "363797880709171295166015625"},                // * 274877906944
   481    {12, "1818989403545856475830078125"},               // * 549755813888
   482    {13, "9094947017729282379150390625"},               // * 1099511627776
   483    {13, "45474735088646411895751953125"},              // * 2199023255552
   484    {13, "227373675443232059478759765625"},             // * 4398046511104
   485    {13, "1136868377216160297393798828125"},            // * 8796093022208
   486    {14, "5684341886080801486968994140625"},            // * 17592186044416
   487    {14, "28421709430404007434844970703125"},           // * 35184372088832
   488    {14, "142108547152020037174224853515625"},          // * 70368744177664
   489    {15, "710542735760100185871124267578125"},          // * 140737488355328
   490    {15, "3552713678800500929355621337890625"},         // * 281474976710656
   491    {15, "17763568394002504646778106689453125"},        // * 562949953421312
   492    {16, "88817841970012523233890533447265625"},        // * 1125899906842624
   493    {16, "444089209850062616169452667236328125"},       // * 2251799813685248
   494    {16, "2220446049250313080847263336181640625"},      // * 4503599627370496
   495    {16, "11102230246251565404236316680908203125"},     // * 9007199254740992
   496    {17, "55511151231257827021181583404541015625"},     // * 18014398509481984
   497    {17, "277555756156289135105907917022705078125"},    // * 36028797018963968
   498    {17, "1387778780781445675529539585113525390625"},   // * 72057594037927936
   499    {18, "6938893903907228377647697925567626953125"},   // * 144115188075855872
   500    {18, "34694469519536141888238489627838134765625"},  // * 288230376151711744
   501    {18, "173472347597680709441192448139190673828125"}, // * 576460752303423488
   502    {19, "867361737988403547205962240695953369140625"}, // * 1152921504606846976
   503};

View as plain text