|
| 1 | +"""Python implementations of some algorithms for use by longobject.c. |
| 2 | +The goal is to provide asymptotically faster algorithms that can be |
| 3 | +used for operations on integers with many digits. In those cases, the |
| 4 | +performance overhead of the Python implementation is not significant |
| 5 | +since the asymptotic behavior is what dominates runtime. Functions |
| 6 | +provided by this module should be considered private and not part of any |
| 7 | +public API. |
| 8 | +
|
| 9 | +Note: for ease of maintainability, please prefer clear code and avoid |
| 10 | +"micro-optimizations". This module will only be imported and used for |
| 11 | +integers with a huge number of digits. Saving a few microseconds with |
| 12 | +tricky or non-obvious code is not worth it. For people looking for |
| 13 | +maximum performance, they should use something like gmpy2.""" |
| 14 | + |
| 15 | +import re |
| 16 | +import decimal |
| 17 | +try: |
| 18 | + import _decimal |
| 19 | +except ImportError: |
| 20 | + _decimal = None |
| 21 | + |
| 22 | +# A number of functions have this form, where `w` is a desired number of |
| 23 | +# digits in base `base`: |
| 24 | +# |
| 25 | +# def inner(...w...): |
| 26 | +# if w <= LIMIT: |
| 27 | +# return something |
| 28 | +# lo = w >> 1 |
| 29 | +# hi = w - lo |
| 30 | +# something involving base**lo, inner(...lo...), j, and inner(...hi...) |
| 31 | +# figure out largest w needed |
| 32 | +# result = inner(w) |
| 33 | +# |
| 34 | +# They all had some on-the-fly scheme to cache `base**lo` results for reuse. |
| 35 | +# Power is costly. |
| 36 | +# |
| 37 | +# This routine aims to compute all amd only the needed powers in advance, as |
| 38 | +# efficiently as reasonably possible. This isn't trivial, and all the |
| 39 | +# on-the-fly methods did needless work in many cases. The driving code above |
| 40 | +# changes to: |
| 41 | +# |
| 42 | +# figure out largest w needed |
| 43 | +# mycache = compute_powers(w, base, LIMIT) |
| 44 | +# result = inner(w) |
| 45 | +# |
| 46 | +# and `mycache[lo]` replaces `base**lo` in the inner function. |
| 47 | +# |
| 48 | +# While this does give minor speedups (a few percent at best), the primary |
| 49 | +# intent is to simplify the functions using this, by eliminating the need for |
| 50 | +# them to craft their own ad-hoc caching schemes. |
| 51 | +def compute_powers(w, base, more_than, show=False): |
| 52 | + seen = set() |
| 53 | + need = set() |
| 54 | + ws = {w} |
| 55 | + while ws: |
| 56 | + w = ws.pop() # any element is fine to use next |
| 57 | + if w in seen or w <= more_than: |
| 58 | + continue |
| 59 | + seen.add(w) |
| 60 | + lo = w >> 1 |
| 61 | + # only _need_ lo here; some other path may, or may not, need hi |
| 62 | + need.add(lo) |
| 63 | + ws.add(lo) |
| 64 | + if w & 1: |
| 65 | + ws.add(lo + 1) |
| 66 | + |
| 67 | + d = {} |
| 68 | + if not need: |
| 69 | + return d |
| 70 | + it = iter(sorted(need)) |
| 71 | + first = next(it) |
| 72 | + if show: |
| 73 | + print("pow at", first) |
| 74 | + d[first] = base ** first |
| 75 | + for this in it: |
| 76 | + if this - 1 in d: |
| 77 | + if show: |
| 78 | + print("* base at", this) |
| 79 | + d[this] = d[this - 1] * base # cheap |
| 80 | + else: |
| 81 | + lo = this >> 1 |
| 82 | + hi = this - lo |
| 83 | + assert lo in d |
| 84 | + if show: |
| 85 | + print("square at", this) |
| 86 | + # Multiplying a bigint by itself (same object!) is about twice |
| 87 | + # as fast in CPython. |
| 88 | + sq = d[lo] * d[lo] |
| 89 | + if hi != lo: |
| 90 | + assert hi == lo + 1 |
| 91 | + if show: |
| 92 | + print(" and * base") |
| 93 | + sq *= base |
| 94 | + d[this] = sq |
| 95 | + return d |
| 96 | + |
| 97 | +_unbounded_dec_context = decimal.getcontext().copy() |
| 98 | +_unbounded_dec_context.prec = decimal.MAX_PREC |
| 99 | +_unbounded_dec_context.Emax = decimal.MAX_EMAX |
| 100 | +_unbounded_dec_context.Emin = decimal.MIN_EMIN |
| 101 | +_unbounded_dec_context.traps[decimal.Inexact] = 1 # sanity check |
| 102 | + |
| 103 | +def int_to_decimal(n): |
| 104 | + """Asymptotically fast conversion of an 'int' to Decimal.""" |
| 105 | + |
| 106 | + # Function due to Tim Peters. See GH issue #90716 for details. |
| 107 | + # https://github.com/python/cpython/issues/90716 |
| 108 | + # |
| 109 | + # The implementation in longobject.c of base conversion algorithms |
| 110 | + # between power-of-2 and non-power-of-2 bases are quadratic time. |
| 111 | + # This function implements a divide-and-conquer algorithm that is |
| 112 | + # faster for large numbers. Builds an equal decimal.Decimal in a |
| 113 | + # "clever" recursive way. If we want a string representation, we |
| 114 | + # apply str to _that_. |
| 115 | + |
| 116 | + from decimal import Decimal as D |
| 117 | + BITLIM = 200 |
| 118 | + |
| 119 | + # Don't bother caching the "lo" mask in this; the time to compute it is |
| 120 | + # tiny compared to the multiply. |
| 121 | + def inner(n, w): |
| 122 | + if w <= BITLIM: |
| 123 | + return D(n) |
| 124 | + w2 = w >> 1 |
| 125 | + hi = n >> w2 |
| 126 | + lo = n & ((1 << w2) - 1) |
| 127 | + return inner(lo, w2) + inner(hi, w - w2) * w2pow[w2] |
| 128 | + |
| 129 | + with decimal.localcontext(_unbounded_dec_context): |
| 130 | + nbits = n.bit_length() |
| 131 | + w2pow = compute_powers(nbits, D(2), BITLIM) |
| 132 | + if n < 0: |
| 133 | + negate = True |
| 134 | + n = -n |
| 135 | + else: |
| 136 | + negate = False |
| 137 | + result = inner(n, nbits) |
| 138 | + if negate: |
| 139 | + result = -result |
| 140 | + return result |
| 141 | + |
| 142 | +def int_to_decimal_string(n): |
| 143 | + """Asymptotically fast conversion of an 'int' to a decimal string.""" |
| 144 | + w = n.bit_length() |
| 145 | + if w > 450_000 and _decimal is not None: |
| 146 | + # It is only usable with the C decimal implementation. |
| 147 | + # _pydecimal.py calls str() on very large integers, which in its |
| 148 | + # turn calls int_to_decimal_string(), causing very deep recursion. |
| 149 | + return str(int_to_decimal(n)) |
| 150 | + |
| 151 | + # Fallback algorithm for the case when the C decimal module isn't |
| 152 | + # available. This algorithm is asymptotically worse than the algorithm |
| 153 | + # using the decimal module, but better than the quadratic time |
| 154 | + # implementation in longobject.c. |
| 155 | + |
| 156 | + DIGLIM = 1000 |
| 157 | + def inner(n, w): |
| 158 | + if w <= DIGLIM: |
| 159 | + return str(n) |
| 160 | + w2 = w >> 1 |
| 161 | + hi, lo = divmod(n, pow10[w2]) |
| 162 | + return inner(hi, w - w2) + inner(lo, w2).zfill(w2) |
| 163 | + |
| 164 | + # The estimation of the number of decimal digits. |
| 165 | + # There is no harm in small error. If we guess too large, there may |
| 166 | + # be leading 0's that need to be stripped. If we guess too small, we |
| 167 | + # may need to call str() recursively for the remaining highest digits, |
| 168 | + # which can still potentially be a large integer. This is manifested |
| 169 | + # only if the number has way more than 10**15 digits, that exceeds |
| 170 | + # the 52-bit physical address limit in both Intel64 and AMD64. |
| 171 | + w = int(w * 0.3010299956639812 + 1) # log10(2) |
| 172 | + pow10 = compute_powers(w, 5, DIGLIM) |
| 173 | + for k, v in pow10.items(): |
| 174 | + pow10[k] = v << k # 5**k << k == 5**k * 2**k == 10**k |
| 175 | + if n < 0: |
| 176 | + n = -n |
| 177 | + sign = '-' |
| 178 | + else: |
| 179 | + sign = '' |
| 180 | + s = inner(n, w) |
| 181 | + if s[0] == '0' and n: |
| 182 | + # If our guess of w is too large, there may be leading 0's that |
| 183 | + # need to be stripped. |
| 184 | + s = s.lstrip('0') |
| 185 | + return sign + s |
| 186 | + |
| 187 | +def _str_to_int_inner(s): |
| 188 | + """Asymptotically fast conversion of a 'str' to an 'int'.""" |
| 189 | + |
| 190 | + # Function due to Bjorn Martinsson. See GH issue #90716 for details. |
| 191 | + # https://github.com/python/cpython/issues/90716 |
| 192 | + # |
| 193 | + # The implementation in longobject.c of base conversion algorithms |
| 194 | + # between power-of-2 and non-power-of-2 bases are quadratic time. |
| 195 | + # This function implements a divide-and-conquer algorithm making use |
| 196 | + # of Python's built in big int multiplication. Since Python uses the |
| 197 | + # Karatsuba algorithm for multiplication, the time complexity |
| 198 | + # of this function is O(len(s)**1.58). |
| 199 | + |
| 200 | + DIGLIM = 2048 |
| 201 | + |
| 202 | + def inner(a, b): |
| 203 | + if b - a <= DIGLIM: |
| 204 | + return int(s[a:b]) |
| 205 | + mid = (a + b + 1) >> 1 |
| 206 | + return (inner(mid, b) |
| 207 | + + ((inner(a, mid) * w5pow[b - mid]) |
| 208 | + << (b - mid))) |
| 209 | + |
| 210 | + w5pow = compute_powers(len(s), 5, DIGLIM) |
| 211 | + return inner(0, len(s)) |
| 212 | + |
| 213 | + |
| 214 | +def int_from_string(s): |
| 215 | + """Asymptotically fast version of PyLong_FromString(), conversion |
| 216 | + of a string of decimal digits into an 'int'.""" |
| 217 | + # PyLong_FromString() has already removed leading +/-, checked for invalid |
| 218 | + # use of underscore characters, checked that string consists of only digits |
| 219 | + # and underscores, and stripped leading whitespace. The input can still |
| 220 | + # contain underscores and have trailing whitespace. |
| 221 | + s = s.rstrip().replace('_', '') |
| 222 | + return _str_to_int_inner(s) |
| 223 | + |
| 224 | +def str_to_int(s): |
| 225 | + """Asymptotically fast version of decimal string to 'int' conversion.""" |
| 226 | + # FIXME: this doesn't support the full syntax that int() supports. |
| 227 | + m = re.match(r'\s*([+-]?)([0-9_]+)\s*', s) |
| 228 | + if not m: |
| 229 | + raise ValueError('invalid literal for int() with base 10') |
| 230 | + v = int_from_string(m.group(2)) |
| 231 | + if m.group(1) == '-': |
| 232 | + v = -v |
| 233 | + return v |
| 234 | + |
| 235 | + |
| 236 | +# Fast integer division, based on code from Mark Dickinson, fast_div.py |
| 237 | +# GH-47701. Additional refinements and optimizations by Bjorn Martinsson. The |
| 238 | +# algorithm is due to Burnikel and Ziegler, in their paper "Fast Recursive |
| 239 | +# Division". |
| 240 | + |
| 241 | +_DIV_LIMIT = 4000 |
| 242 | + |
| 243 | + |
| 244 | +def _div2n1n(a, b, n): |
| 245 | + """Divide a 2n-bit nonnegative integer a by an n-bit positive integer |
| 246 | + b, using a recursive divide-and-conquer algorithm. |
| 247 | +
|
| 248 | + Inputs: |
| 249 | + n is a positive integer |
| 250 | + b is a positive integer with exactly n bits |
| 251 | + a is a nonnegative integer such that a < 2**n * b |
| 252 | +
|
| 253 | + Output: |
| 254 | + (q, r) such that a = b*q+r and 0 <= r < b. |
| 255 | +
|
| 256 | + """ |
| 257 | + if a.bit_length() - n <= _DIV_LIMIT: |
| 258 | + return divmod(a, b) |
| 259 | + pad = n & 1 |
| 260 | + if pad: |
| 261 | + a <<= 1 |
| 262 | + b <<= 1 |
| 263 | + n += 1 |
| 264 | + half_n = n >> 1 |
| 265 | + mask = (1 << half_n) - 1 |
| 266 | + b1, b2 = b >> half_n, b & mask |
| 267 | + q1, r = _div3n2n(a >> n, (a >> half_n) & mask, b, b1, b2, half_n) |
| 268 | + q2, r = _div3n2n(r, a & mask, b, b1, b2, half_n) |
| 269 | + if pad: |
| 270 | + r >>= 1 |
| 271 | + return q1 << half_n | q2, r |
| 272 | + |
| 273 | + |
| 274 | +def _div3n2n(a12, a3, b, b1, b2, n): |
| 275 | + """Helper function for _div2n1n; not intended to be called directly.""" |
| 276 | + if a12 >> n == b1: |
| 277 | + q, r = (1 << n) - 1, a12 - (b1 << n) + b1 |
| 278 | + else: |
| 279 | + q, r = _div2n1n(a12, b1, n) |
| 280 | + r = (r << n | a3) - q * b2 |
| 281 | + while r < 0: |
| 282 | + q -= 1 |
| 283 | + r += b |
| 284 | + return q, r |
| 285 | + |
| 286 | + |
| 287 | +def _int2digits(a, n): |
| 288 | + """Decompose non-negative int a into base 2**n |
| 289 | +
|
| 290 | + Input: |
| 291 | + a is a non-negative integer |
| 292 | +
|
| 293 | + Output: |
| 294 | + List of the digits of a in base 2**n in little-endian order, |
| 295 | + meaning the most significant digit is last. The most |
| 296 | + significant digit is guaranteed to be non-zero. |
| 297 | + If a is 0 then the output is an empty list. |
| 298 | +
|
| 299 | + """ |
| 300 | + a_digits = [0] * ((a.bit_length() + n - 1) // n) |
| 301 | + |
| 302 | + def inner(x, L, R): |
| 303 | + if L + 1 == R: |
| 304 | + a_digits[L] = x |
| 305 | + return |
| 306 | + mid = (L + R) >> 1 |
| 307 | + shift = (mid - L) * n |
| 308 | + upper = x >> shift |
| 309 | + lower = x ^ (upper << shift) |
| 310 | + inner(lower, L, mid) |
| 311 | + inner(upper, mid, R) |
| 312 | + |
| 313 | + if a: |
| 314 | + inner(a, 0, len(a_digits)) |
| 315 | + return a_digits |
| 316 | + |
| 317 | + |
| 318 | +def _digits2int(digits, n): |
| 319 | + """Combine base-2**n digits into an int. This function is the |
| 320 | + inverse of `_int2digits`. For more details, see _int2digits. |
| 321 | + """ |
| 322 | + |
| 323 | + def inner(L, R): |
| 324 | + if L + 1 == R: |
| 325 | + return digits[L] |
| 326 | + mid = (L + R) >> 1 |
| 327 | + shift = (mid - L) * n |
| 328 | + return (inner(mid, R) << shift) + inner(L, mid) |
| 329 | + |
| 330 | + return inner(0, len(digits)) if digits else 0 |
| 331 | + |
| 332 | + |
| 333 | +def _divmod_pos(a, b): |
| 334 | + """Divide a non-negative integer a by a positive integer b, giving |
| 335 | + quotient and remainder.""" |
| 336 | + # Use grade-school algorithm in base 2**n, n = nbits(b) |
| 337 | + n = b.bit_length() |
| 338 | + a_digits = _int2digits(a, n) |
| 339 | + |
| 340 | + r = 0 |
| 341 | + q_digits = [] |
| 342 | + for a_digit in reversed(a_digits): |
| 343 | + q_digit, r = _div2n1n((r << n) + a_digit, b, n) |
| 344 | + q_digits.append(q_digit) |
| 345 | + q_digits.reverse() |
| 346 | + q = _digits2int(q_digits, n) |
| 347 | + return q, r |
| 348 | + |
| 349 | + |
| 350 | +def int_divmod(a, b): |
| 351 | + """Asymptotically fast replacement for divmod, for 'int'. |
| 352 | + Its time complexity is O(n**1.58), where n = #bits(a) + #bits(b). |
| 353 | + """ |
| 354 | + if b == 0: |
| 355 | + raise ZeroDivisionError |
| 356 | + elif b < 0: |
| 357 | + q, r = int_divmod(-a, -b) |
| 358 | + return q, -r |
| 359 | + elif a < 0: |
| 360 | + q, r = int_divmod(~a, b) |
| 361 | + return ~q, b + ~r |
| 362 | + else: |
| 363 | + return _divmod_pos(a, b) |
0 commit comments