ftoa.c 8.9 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301
  1. #include "mupdf/fitz.h"
  2. #include <assert.h>
  3. /*
  4. Convert IEEE single precision numbers into decimal ASCII strings, while
  5. satisfying the following two properties:
  6. 1) Calling strtof or '(float) strtod' on the result must produce the
  7. original float, independent of the rounding mode used by strtof/strtod.
  8. 2) Minimize the number of produced decimal digits. E.g. the float 0.7f
  9. should convert to "0.7", not "0.69999999".
  10. To solve this we use a dedicated single precision version of
  11. Florian Loitsch's Grisu2 algorithm. See
  12. http://florian.loitsch.com/publications/dtoa-pldi2010.pdf?attredirects=0
  13. The code below is derived from Loitsch's C code, which
  14. implements the same algorithm for IEEE double precision. See
  15. http://florian.loitsch.com/publications/bench.tar.gz?attredirects=0
  16. */
  17. /*
  18. Copyright (c) 2009 Florian Loitsch
  19. Permission is hereby granted, free of charge, to any person
  20. obtaining a copy of this software and associated documentation
  21. files (the "Software"), to deal in the Software without
  22. restriction, including without limitation the rights to use,
  23. copy, modify, merge, publish, distribute, sublicense, and/or sell
  24. copies of the Software, and to permit persons to whom the
  25. Software is furnished to do so, subject to the following
  26. conditions:
  27. The above copyright notice and this permission notice shall be
  28. included in all copies or substantial portions of the Software.
  29. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  30. EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
  31. OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  32. NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
  33. HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
  34. WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  35. FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  36. OTHER DEALINGS IN THE SOFTWARE.
  37. */
  38. static uint32_t
  39. float_to_uint32(float d)
  40. {
  41. union
  42. {
  43. float d;
  44. uint32_t n;
  45. } tmp;
  46. tmp.d = d;
  47. return tmp.n;
  48. }
  49. typedef struct
  50. {
  51. uint64_t f;
  52. int e;
  53. } diy_fp_t;
  54. #define DIY_SIGNIFICAND_SIZE 64
  55. #define DIY_LEADING_BIT ((uint64_t) 1 << (DIY_SIGNIFICAND_SIZE - 1))
  56. static diy_fp_t
  57. minus(diy_fp_t x, diy_fp_t y)
  58. {
  59. diy_fp_t result = {x.f - y.f, x.e};
  60. assert(x.e == y.e && x.f >= y.f);
  61. return result;
  62. }
  63. static diy_fp_t
  64. multiply(diy_fp_t x, diy_fp_t y)
  65. {
  66. uint64_t a, b, c, d, ac, bc, ad, bd, tmp;
  67. int half = DIY_SIGNIFICAND_SIZE / 2;
  68. diy_fp_t r; uint64_t mask = ((uint64_t) 1 << half) - 1;
  69. a = x.f >> half; b = x.f & mask;
  70. c = y.f >> half; d = y.f & mask;
  71. ac = a * c; bc = b * c; ad = a * d; bd = b * d;
  72. tmp = (bd >> half) + (ad & mask) + (bc & mask);
  73. tmp += ((uint64_t)1U) << (half - 1); /* Round. */
  74. r.f = ac + (ad >> half) + (bc >> half) + (tmp >> half);
  75. r.e = x.e + y.e + half * 2;
  76. return r;
  77. }
  78. #define SP_SIGNIFICAND_SIZE 23
  79. #define SP_EXPONENT_BIAS (127 + SP_SIGNIFICAND_SIZE)
  80. #define SP_MIN_EXPONENT (-SP_EXPONENT_BIAS)
  81. #define SP_EXPONENT_MASK 0x7f800000
  82. #define SP_SIGNIFICAND_MASK 0x7fffff
  83. #define SP_HIDDEN_BIT 0x800000 /* 2^23 */
  84. /* Does not normalize the result. */
  85. static diy_fp_t
  86. float2diy_fp(float d)
  87. {
  88. uint32_t d32 = float_to_uint32(d);
  89. int biased_e = (d32 & SP_EXPONENT_MASK) >> SP_SIGNIFICAND_SIZE;
  90. uint32_t significand = d32 & SP_SIGNIFICAND_MASK;
  91. diy_fp_t res;
  92. if (biased_e != 0)
  93. {
  94. res.f = significand + SP_HIDDEN_BIT;
  95. res.e = biased_e - SP_EXPONENT_BIAS;
  96. }
  97. else
  98. {
  99. res.f = significand;
  100. res.e = SP_MIN_EXPONENT + 1;
  101. }
  102. return res;
  103. }
  104. static diy_fp_t
  105. normalize_boundary(diy_fp_t in)
  106. {
  107. diy_fp_t res = in;
  108. /* The original number could have been a denormal. */
  109. while (! (res.f & (SP_HIDDEN_BIT << 1)))
  110. {
  111. res.f <<= 1;
  112. res.e--;
  113. }
  114. /* Do the final shifts in one go. */
  115. res.f <<= (DIY_SIGNIFICAND_SIZE - SP_SIGNIFICAND_SIZE - 2);
  116. res.e = res.e - (DIY_SIGNIFICAND_SIZE - SP_SIGNIFICAND_SIZE - 2);
  117. return res;
  118. }
  119. static void
  120. normalized_boundaries(float f, diy_fp_t* lower_ptr, diy_fp_t* upper_ptr)
  121. {
  122. diy_fp_t v = float2diy_fp(f);
  123. diy_fp_t upper, lower;
  124. int significand_is_zero = v.f == SP_HIDDEN_BIT;
  125. upper.f = (v.f << 1) + 1; upper.e = v.e - 1;
  126. upper = normalize_boundary(upper);
  127. if (significand_is_zero)
  128. {
  129. lower.f = (v.f << 2) - 1;
  130. lower.e = v.e - 2;
  131. }
  132. else
  133. {
  134. lower.f = (v.f << 1) - 1;
  135. lower.e = v.e - 1;
  136. }
  137. lower.f <<= lower.e - upper.e;
  138. lower.e = upper.e;
  139. /* Adjust to double boundaries, so that we can also read the numbers with '(float) strtod'. */
  140. upper.f -= 1 << 10;
  141. lower.f += 1 << 10;
  142. *upper_ptr = upper;
  143. *lower_ptr = lower;
  144. }
  145. static int
  146. k_comp(int n)
  147. {
  148. /* Avoid ceil and floating point multiplication for better
  149. * performance and portability. Instead use the approximation
  150. * log10(2) ~ 1233/(2^12). Tests show that this gives the correct
  151. * result for all values of n in the range -500..500. */
  152. int tmp = n + DIY_SIGNIFICAND_SIZE - 1;
  153. int k = (tmp * 1233) / (1 << 12);
  154. return tmp > 0 ? k + 1 : k;
  155. }
  156. /* Cached powers of ten from 10**-37..10**46. Produced using GNU MPFR's mpfr_pow_si. */
  157. /* Significands. */
  158. static uint64_t powers_ten[84] = {
  159. 0x881cea14545c7575ull, 0xaa242499697392d3ull, 0xd4ad2dbfc3d07788ull,
  160. 0x84ec3c97da624ab5ull, 0xa6274bbdd0fadd62ull, 0xcfb11ead453994baull,
  161. 0x81ceb32c4b43fcf5ull, 0xa2425ff75e14fc32ull, 0xcad2f7f5359a3b3eull,
  162. 0xfd87b5f28300ca0eull, 0x9e74d1b791e07e48ull, 0xc612062576589ddbull,
  163. 0xf79687aed3eec551ull, 0x9abe14cd44753b53ull, 0xc16d9a0095928a27ull,
  164. 0xf1c90080baf72cb1ull, 0x971da05074da7befull, 0xbce5086492111aebull,
  165. 0xec1e4a7db69561a5ull, 0x9392ee8e921d5d07ull, 0xb877aa3236a4b449ull,
  166. 0xe69594bec44de15bull, 0x901d7cf73ab0acd9ull, 0xb424dc35095cd80full,
  167. 0xe12e13424bb40e13ull, 0x8cbccc096f5088ccull, 0xafebff0bcb24aaffull,
  168. 0xdbe6fecebdedd5bfull, 0x89705f4136b4a597ull, 0xabcc77118461cefdull,
  169. 0xd6bf94d5e57a42bcull, 0x8637bd05af6c69b6ull, 0xa7c5ac471b478423ull,
  170. 0xd1b71758e219652cull, 0x83126e978d4fdf3bull, 0xa3d70a3d70a3d70aull,
  171. 0xcccccccccccccccdull, 0x8000000000000000ull, 0xa000000000000000ull,
  172. 0xc800000000000000ull, 0xfa00000000000000ull, 0x9c40000000000000ull,
  173. 0xc350000000000000ull, 0xf424000000000000ull, 0x9896800000000000ull,
  174. 0xbebc200000000000ull, 0xee6b280000000000ull, 0x9502f90000000000ull,
  175. 0xba43b74000000000ull, 0xe8d4a51000000000ull, 0x9184e72a00000000ull,
  176. 0xb5e620f480000000ull, 0xe35fa931a0000000ull, 0x8e1bc9bf04000000ull,
  177. 0xb1a2bc2ec5000000ull, 0xde0b6b3a76400000ull, 0x8ac7230489e80000ull,
  178. 0xad78ebc5ac620000ull, 0xd8d726b7177a8000ull, 0x878678326eac9000ull,
  179. 0xa968163f0a57b400ull, 0xd3c21bcecceda100ull, 0x84595161401484a0ull,
  180. 0xa56fa5b99019a5c8ull, 0xcecb8f27f4200f3aull, 0x813f3978f8940984ull,
  181. 0xa18f07d736b90be5ull, 0xc9f2c9cd04674edfull, 0xfc6f7c4045812296ull,
  182. 0x9dc5ada82b70b59eull, 0xc5371912364ce305ull, 0xf684df56c3e01bc7ull,
  183. 0x9a130b963a6c115cull, 0xc097ce7bc90715b3ull, 0xf0bdc21abb48db20ull,
  184. 0x96769950b50d88f4ull, 0xbc143fa4e250eb31ull, 0xeb194f8e1ae525fdull,
  185. 0x92efd1b8d0cf37beull, 0xb7abc627050305aeull, 0xe596b7b0c643c719ull,
  186. 0x8f7e32ce7bea5c70ull, 0xb35dbf821ae4f38cull, 0xe0352f62a19e306full,
  187. };
  188. /* Exponents. */
  189. static int powers_ten_e[84] = {
  190. -186, -183, -180, -176, -173, -170, -166, -163, -160, -157, -153,
  191. -150, -147, -143, -140, -137, -133, -130, -127, -123, -120, -117,
  192. -113, -110, -107, -103, -100, -97, -93, -90, -87, -83, -80,
  193. -77, -73, -70, -67, -63, -60, -57, -54, -50, -47, -44,
  194. -40, -37, -34, -30, -27, -24, -20, -17, -14, -10, -7,
  195. -4, 0, 3, 6, 10, 13, 16, 20, 23, 26, 30,
  196. 33, 36, 39, 43, 46, 49, 53, 56, 59, 63, 66,
  197. 69, 73, 76, 79, 83, 86, 89
  198. };
  199. static diy_fp_t
  200. cached_power(int i)
  201. {
  202. diy_fp_t result;
  203. assert (i >= -37 && i <= 46);
  204. result.f = powers_ten[i + 37];
  205. result.e = powers_ten_e[i + 37];
  206. return result;
  207. }
  208. /* Returns buffer length. */
  209. static int
  210. digit_gen_mix_grisu2(diy_fp_t D_upper, diy_fp_t delta, char* buffer, int* K)
  211. {
  212. int kappa;
  213. diy_fp_t one = {(uint64_t) 1 << -D_upper.e, D_upper.e};
  214. unsigned char p1 = D_upper.f >> -one.e;
  215. uint64_t p2 = D_upper.f & (one.f - 1);
  216. unsigned char div = 10;
  217. uint64_t mask = one.f - 1;
  218. int len = 0;
  219. for (kappa = 2; kappa > 0; --kappa)
  220. {
  221. unsigned char digit = p1 / div;
  222. if (digit || len)
  223. buffer[len++] = '0' + digit;
  224. p1 %= div; div /= 10;
  225. if ((((uint64_t) p1) << -one.e) + p2 <= delta.f)
  226. {
  227. *K += kappa - 1;
  228. return len;
  229. }
  230. }
  231. do
  232. {
  233. p2 *= 10;
  234. buffer[len++] = '0' + (p2 >> -one.e);
  235. p2 &= mask;
  236. kappa--;
  237. delta.f *= 10;
  238. }
  239. while (p2 > delta.f);
  240. *K += kappa;
  241. return len;
  242. }
  243. /*
  244. Compute decimal integer m, exp such that:
  245. f = m * 10^exp
  246. m is as short as possible without losing exactness
  247. Assumes special cases (0, NaN, +Inf, -Inf) have been handled.
  248. */
  249. int
  250. fz_grisu(float v, char* buffer, int* K)
  251. {
  252. diy_fp_t w_lower, w_upper, D_upper, D_lower, c_mk, delta;
  253. int length, mk, alpha = -DIY_SIGNIFICAND_SIZE + 4;
  254. normalized_boundaries(v, &w_lower, &w_upper);
  255. mk = k_comp(alpha - w_upper.e - DIY_SIGNIFICAND_SIZE);
  256. c_mk = cached_power(mk);
  257. D_upper = multiply(w_upper, c_mk);
  258. D_lower = multiply(w_lower, c_mk);
  259. D_upper.f--;
  260. D_lower.f++;
  261. delta = minus(D_upper, D_lower);
  262. *K = -mk;
  263. length = digit_gen_mix_grisu2(D_upper, delta, buffer, K);
  264. buffer[length] = 0;
  265. return length;
  266. }