ftoaryu.mx raw

   1  // Copyright 2021 The Go Authors. All rights reserved.
   2  // Use of this source code is governed by a BSD-style
   3  // license that can be found in the LICENSE file.
   4  
   5  package strconv
   6  
   7  import (
   8  	"math/bits"
   9  )
  10  
  11  // binary to decimal conversion using the Ryū algorithm.
  12  //
  13  // See Ulf Adams, "Ryū: Fast Float-to-String Conversion" (doi:10.1145/3192366.3192369)
  14  //
  15  // Fixed precision formatting is a variant of the original paper's
  16  // algorithm, where a single multiplication by 10^k is required,
  17  // sharing the same rounding guarantees.
  18  
  19  // ryuFtoaFixed32 formats mant*(2^exp) with prec decimal digits.
  20  func ryuFtoaFixed32(d *decimalSlice, mant uint32, exp int, prec int) {
  21  	if prec < 0 {
  22  		panic("ryuFtoaFixed32 called with negative prec")
  23  	}
  24  	if prec > 9 {
  25  		panic("ryuFtoaFixed32 called with prec > 9")
  26  	}
  27  	// Zero input.
  28  	if mant == 0 {
  29  		d.nd, d.dp = 0, 0
  30  		return
  31  	}
  32  	// Renormalize to a 25-bit mantissa.
  33  	e2 := exp
  34  	if b := bits.Len32(mant); b < 25 {
  35  		mant <<= uint(25 - b)
  36  		e2 += b - 25
  37  	}
  38  	// Choose an exponent such that rounded mant*(2^e2)*(10^q) has
  39  	// at least prec decimal digits, i.e
  40  	//     mant*(2^e2)*(10^q) >= 10^(prec-1)
  41  	// Because mant >= 2^24, it is enough to choose:
  42  	//     2^(e2+24) >= 10^(-q+prec-1)
  43  	// or q = -mulByLog2Log10(e2+24) + prec - 1
  44  	q := -mulByLog2Log10(e2+24) + prec - 1
  45  
  46  	// Now compute mant*(2^e2)*(10^q).
  47  	// Is it an exact computation?
  48  	// Only small positive powers of 10 are exact (5^28 has 66 bits).
  49  	exact := q <= 27 && q >= 0
  50  
  51  	di, dexp2, d0 := mult64bitPow10(mant, e2, q)
  52  	if dexp2 >= 0 {
  53  		panic("not enough significant bits after mult64bitPow10")
  54  	}
  55  	// As a special case, computation might still be exact, if exponent
  56  	// was negative and if it amounts to computing an exact division.
  57  	// In that case, we ignore all lower bits.
  58  	// Note that division by 10^11 cannot be exact as 5^11 has 26 bits.
  59  	if q < 0 && q >= -10 && divisibleByPower5(uint64(mant), -q) {
  60  		exact = true
  61  		d0 = true
  62  	}
  63  	// Remove extra lower bits and keep rounding info.
  64  	extra := uint(-dexp2)
  65  	extraMask := uint32(1<<extra - 1)
  66  
  67  	di, dfrac := di>>extra, di&extraMask
  68  	roundUp := false
  69  	if exact {
  70  		// If we computed an exact product, d + 1/2
  71  		// should round to d+1 if 'd' is odd.
  72  		roundUp = dfrac > 1<<(extra-1) ||
  73  			(dfrac == 1<<(extra-1) && !d0) ||
  74  			(dfrac == 1<<(extra-1) && d0 && di&1 == 1)
  75  	} else {
  76  		// otherwise, d+1/2 always rounds up because
  77  		// we truncated below.
  78  		roundUp = dfrac>>(extra-1) == 1
  79  	}
  80  	if dfrac != 0 {
  81  		d0 = false
  82  	}
  83  	// Proceed to the requested number of digits
  84  	formatDecimal(d, uint64(di), !d0, roundUp, prec)
  85  	// Adjust exponent
  86  	d.dp -= q
  87  }
  88  
  89  // ryuFtoaFixed64 formats mant*(2^exp) with prec decimal digits.
  90  func ryuFtoaFixed64(d *decimalSlice, mant uint64, exp int, prec int) {
  91  	if prec > 18 {
  92  		panic("ryuFtoaFixed64 called with prec > 18")
  93  	}
  94  	// Zero input.
  95  	if mant == 0 {
  96  		d.nd, d.dp = 0, 0
  97  		return
  98  	}
  99  	// Renormalize to a 55-bit mantissa.
 100  	e2 := exp
 101  	if b := bits.Len64(mant); b < 55 {
 102  		mant = mant << uint(55-b)
 103  		e2 += b - 55
 104  	}
 105  	// Choose an exponent such that rounded mant*(2^e2)*(10^q) has
 106  	// at least prec decimal digits, i.e
 107  	//     mant*(2^e2)*(10^q) >= 10^(prec-1)
 108  	// Because mant >= 2^54, it is enough to choose:
 109  	//     2^(e2+54) >= 10^(-q+prec-1)
 110  	// or q = -mulByLog2Log10(e2+54) + prec - 1
 111  	//
 112  	// The minimal required exponent is -mulByLog2Log10(1025)+18 = -291
 113  	// The maximal required exponent is mulByLog2Log10(1074)+18 = 342
 114  	q := -mulByLog2Log10(e2+54) + prec - 1
 115  
 116  	// Now compute mant*(2^e2)*(10^q).
 117  	// Is it an exact computation?
 118  	// Only small positive powers of 10 are exact (5^55 has 128 bits).
 119  	exact := q <= 55 && q >= 0
 120  
 121  	di, dexp2, d0 := mult128bitPow10(mant, e2, q)
 122  	if dexp2 >= 0 {
 123  		panic("not enough significant bits after mult128bitPow10")
 124  	}
 125  	// As a special case, computation might still be exact, if exponent
 126  	// was negative and if it amounts to computing an exact division.
 127  	// In that case, we ignore all lower bits.
 128  	// Note that division by 10^23 cannot be exact as 5^23 has 54 bits.
 129  	if q < 0 && q >= -22 && divisibleByPower5(mant, -q) {
 130  		exact = true
 131  		d0 = true
 132  	}
 133  	// Remove extra lower bits and keep rounding info.
 134  	extra := uint(-dexp2)
 135  	extraMask := uint64(1<<extra - 1)
 136  
 137  	di, dfrac := di>>extra, di&extraMask
 138  	roundUp := false
 139  	if exact {
 140  		// If we computed an exact product, d + 1/2
 141  		// should round to d+1 if 'd' is odd.
 142  		roundUp = dfrac > 1<<(extra-1) ||
 143  			(dfrac == 1<<(extra-1) && !d0) ||
 144  			(dfrac == 1<<(extra-1) && d0 && di&1 == 1)
 145  	} else {
 146  		// otherwise, d+1/2 always rounds up because
 147  		// we truncated below.
 148  		roundUp = dfrac>>(extra-1) == 1
 149  	}
 150  	if dfrac != 0 {
 151  		d0 = false
 152  	}
 153  	// Proceed to the requested number of digits
 154  	formatDecimal(d, di, !d0, roundUp, prec)
 155  	// Adjust exponent
 156  	d.dp -= q
 157  }
 158  
 159  var uint64pow10 = [...]uint64{
 160  	1, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6, 1e7, 1e8, 1e9,
 161  	1e10, 1e11, 1e12, 1e13, 1e14, 1e15, 1e16, 1e17, 1e18, 1e19,
 162  }
 163  
 164  // formatDecimal fills d with at most prec decimal digits
 165  // of mantissa m. The boolean trunc indicates whether m
 166  // is truncated compared to the original number being formatted.
 167  func formatDecimal(d *decimalSlice, m uint64, trunc bool, roundUp bool, prec int) {
 168  	max := uint64pow10[prec]
 169  	trimmed := 0
 170  	for m >= max {
 171  		a, b := m/10, m%10
 172  		m = a
 173  		trimmed++
 174  		if b > 5 {
 175  			roundUp = true
 176  		} else if b < 5 {
 177  			roundUp = false
 178  		} else { // b == 5
 179  			// round up if there are trailing digits,
 180  			// or if the new value of m is odd (round-to-even convention)
 181  			roundUp = trunc || m&1 == 1
 182  		}
 183  		if b != 0 {
 184  			trunc = true
 185  		}
 186  	}
 187  	if roundUp {
 188  		m++
 189  	}
 190  	if m >= max {
 191  		// Happens if di was originally 99999....xx
 192  		m /= 10
 193  		trimmed++
 194  	}
 195  	// render digits (similar to formatBits)
 196  	n := uint(prec)
 197  	d.nd = prec
 198  	v := m
 199  	for v >= 100 {
 200  		var v1, v2 uint64
 201  		if v>>32 == 0 {
 202  			v1, v2 = uint64(uint32(v)/100), uint64(uint32(v)%100)
 203  		} else {
 204  			v1, v2 = v/100, v%100
 205  		}
 206  		n -= 2
 207  		d.d[n+1] = smallsString[2*v2+1]
 208  		d.d[n+0] = smallsString[2*v2+0]
 209  		v = v1
 210  	}
 211  	if v > 0 {
 212  		n--
 213  		d.d[n] = smallsString[2*v+1]
 214  	}
 215  	if v >= 10 {
 216  		n--
 217  		d.d[n] = smallsString[2*v]
 218  	}
 219  	for d.d[d.nd-1] == '0' {
 220  		d.nd--
 221  		trimmed++
 222  	}
 223  	d.dp = d.nd + trimmed
 224  }
 225  
 226  // ryuFtoaShortest formats mant*2^exp with prec decimal digits.
 227  func ryuFtoaShortest(d *decimalSlice, mant uint64, exp int, flt *floatInfo) {
 228  	if mant == 0 {
 229  		d.nd, d.dp = 0, 0
 230  		return
 231  	}
 232  	// If input is an exact integer with fewer bits than the mantissa,
 233  	// the previous and next integer are not admissible representations.
 234  	if exp <= 0 && bits.TrailingZeros64(mant) >= -exp {
 235  		mant >>= uint(-exp)
 236  		ryuDigits(d, mant, mant, mant, true, false)
 237  		return
 238  	}
 239  	ml, mc, mu, e2 := computeBounds(mant, exp, flt)
 240  	if e2 == 0 {
 241  		ryuDigits(d, ml, mc, mu, true, false)
 242  		return
 243  	}
 244  	// Find 10^q *larger* than 2^-e2
 245  	q := mulByLog2Log10(-e2) + 1
 246  
 247  	// We are going to multiply by 10^q using 128-bit arithmetic.
 248  	// The exponent is the same for all 3 numbers.
 249  	var dl, dc, du uint64
 250  	var dl0, dc0, du0 bool
 251  	if flt == &float32info {
 252  		var dl32, dc32, du32 uint32
 253  		dl32, _, dl0 = mult64bitPow10(uint32(ml), e2, q)
 254  		dc32, _, dc0 = mult64bitPow10(uint32(mc), e2, q)
 255  		du32, e2, du0 = mult64bitPow10(uint32(mu), e2, q)
 256  		dl, dc, du = uint64(dl32), uint64(dc32), uint64(du32)
 257  	} else {
 258  		dl, _, dl0 = mult128bitPow10(ml, e2, q)
 259  		dc, _, dc0 = mult128bitPow10(mc, e2, q)
 260  		du, e2, du0 = mult128bitPow10(mu, e2, q)
 261  	}
 262  	if e2 >= 0 {
 263  		panic("not enough significant bits after mult128bitPow10")
 264  	}
 265  	// Is it an exact computation?
 266  	if q > 55 {
 267  		// Large positive powers of ten are not exact
 268  		dl0, dc0, du0 = false, false, false
 269  	}
 270  	if q < 0 && q >= -24 {
 271  		// Division by a power of ten may be exact.
 272  		// (note that 5^25 is a 59-bit number so division by 5^25 is never exact).
 273  		if divisibleByPower5(ml, -q) {
 274  			dl0 = true
 275  		}
 276  		if divisibleByPower5(mc, -q) {
 277  			dc0 = true
 278  		}
 279  		if divisibleByPower5(mu, -q) {
 280  			du0 = true
 281  		}
 282  	}
 283  	// Express the results (dl, dc, du)*2^e2 as integers.
 284  	// Extra bits must be removed and rounding hints computed.
 285  	extra := uint(-e2)
 286  	extraMask := uint64(1<<extra - 1)
 287  	// Now compute the floored, integral base 10 mantissas.
 288  	dl, fracl := dl>>extra, dl&extraMask
 289  	dc, fracc := dc>>extra, dc&extraMask
 290  	du, fracu := du>>extra, du&extraMask
 291  	// Is it allowed to use 'du' as a result?
 292  	// It is always allowed when it is truncated, but also
 293  	// if it is exact and the original binary mantissa is even
 294  	// When disallowed, we can subtract 1.
 295  	uok := !du0 || fracu > 0
 296  	if du0 && fracu == 0 {
 297  		uok = mant&1 == 0
 298  	}
 299  	if !uok {
 300  		du--
 301  	}
 302  	// Is 'dc' the correctly rounded base 10 mantissa?
 303  	// The correct rounding might be dc+1
 304  	cup := false // don't round up.
 305  	if dc0 {
 306  		// If we computed an exact product, the half integer
 307  		// should round to next (even) integer if 'dc' is odd.
 308  		cup = fracc > 1<<(extra-1) ||
 309  			(fracc == 1<<(extra-1) && dc&1 == 1)
 310  	} else {
 311  		// otherwise, the result is a lower truncation of the ideal
 312  		// result.
 313  		cup = fracc>>(extra-1) == 1
 314  	}
 315  	// Is 'dl' an allowed representation?
 316  	// Only if it is an exact value, and if the original binary mantissa
 317  	// was even.
 318  	lok := dl0 && fracl == 0 && (mant&1 == 0)
 319  	if !lok {
 320  		dl++
 321  	}
 322  	// We need to remember whether the trimmed digits of 'dc' are zero.
 323  	c0 := dc0 && fracc == 0
 324  	// render digits
 325  	ryuDigits(d, dl, dc, du, c0, cup)
 326  	d.dp -= q
 327  }
 328  
 329  // mulByLog2Log10 returns math.Floor(x * log(2)/log(10)) for an integer x in
 330  // the range -1600 <= x && x <= +1600.
 331  //
 332  // The range restriction lets us work in faster integer arithmetic instead of
 333  // slower floating point arithmetic. Correctness is verified by unit tests.
 334  func mulByLog2Log10(x int) int {
 335  	// log(2)/log(10) ≈ 0.30102999566 ≈ 78913 / 2^18
 336  	return (x * 78913) >> 18
 337  }
 338  
 339  // mulByLog10Log2 returns math.Floor(x * log(10)/log(2)) for an integer x in
 340  // the range -500 <= x && x <= +500.
 341  //
 342  // The range restriction lets us work in faster integer arithmetic instead of
 343  // slower floating point arithmetic. Correctness is verified by unit tests.
 344  func mulByLog10Log2(x int) int {
 345  	// log(10)/log(2) ≈ 3.32192809489 ≈ 108853 / 2^15
 346  	return (x * 108853) >> 15
 347  }
 348  
 349  // computeBounds returns a floating-point vector (l, c, u)×2^e2
 350  // where the mantissas are 55-bit (or 26-bit) integers, describing the interval
 351  // represented by the input float64 or float32.
 352  func computeBounds(mant uint64, exp int, flt *floatInfo) (lower, central, upper uint64, e2 int) {
 353  	if mant != 1<<flt.mantbits || exp == flt.bias+1-int(flt.mantbits) {
 354  		// regular case (or denormals)
 355  		lower, central, upper = 2*mant-1, 2*mant, 2*mant+1
 356  		e2 = exp - 1
 357  		return
 358  	} else {
 359  		// border of an exponent
 360  		lower, central, upper = 4*mant-1, 4*mant, 4*mant+2
 361  		e2 = exp - 2
 362  		return
 363  	}
 364  }
 365  
 366  func ryuDigits(d *decimalSlice, lower, central, upper uint64,
 367  	c0, cup bool) {
 368  	lhi, llo := divmod1e9(lower)
 369  	chi, clo := divmod1e9(central)
 370  	uhi, ulo := divmod1e9(upper)
 371  	if uhi == 0 {
 372  		// only low digits (for denormals)
 373  		ryuDigits32(d, llo, clo, ulo, c0, cup, 8)
 374  	} else if lhi < uhi {
 375  		// truncate 9 digits at once.
 376  		if llo != 0 {
 377  			lhi++
 378  		}
 379  		c0 = c0 && clo == 0
 380  		cup = (clo > 5e8) || (clo == 5e8 && cup)
 381  		ryuDigits32(d, lhi, chi, uhi, c0, cup, 8)
 382  		d.dp += 9
 383  	} else {
 384  		d.nd = 0
 385  		// emit high part
 386  		n := uint(9)
 387  		for v := chi; v > 0; {
 388  			v1, v2 := v/10, v%10
 389  			v = v1
 390  			n--
 391  			d.d[n] = byte(v2 + '0')
 392  		}
 393  		d.d = d.d[n:]
 394  		d.nd = int(9 - n)
 395  		// emit low part
 396  		ryuDigits32(d, llo, clo, ulo,
 397  			c0, cup, d.nd+8)
 398  	}
 399  	// trim trailing zeros
 400  	for d.nd > 0 && d.d[d.nd-1] == '0' {
 401  		d.nd--
 402  	}
 403  	// trim initial zeros
 404  	for d.nd > 0 && d.d[0] == '0' {
 405  		d.nd--
 406  		d.dp--
 407  		d.d = d.d[1:]
 408  	}
 409  }
 410  
 411  // ryuDigits32 emits decimal digits for a number less than 1e9.
 412  func ryuDigits32(d *decimalSlice, lower, central, upper uint32,
 413  	c0, cup bool, endindex int) {
 414  	if upper == 0 {
 415  		d.dp = endindex + 1
 416  		return
 417  	}
 418  	trimmed := 0
 419  	// Remember last trimmed digit to check for round-up.
 420  	// c0 will be used to remember zeroness of following digits.
 421  	cNextDigit := 0
 422  	for upper > 0 {
 423  		// Repeatedly compute:
 424  		// l = Ceil(lower / 10^k)
 425  		// c = Round(central / 10^k)
 426  		// u = Floor(upper / 10^k)
 427  		// and stop when c goes out of the (l, u) interval.
 428  		l := (lower + 9) / 10
 429  		c, cdigit := central/10, central%10
 430  		u := upper / 10
 431  		if l > u {
 432  			// don't trim the last digit as it is forbidden to go below l
 433  			// other, trim and exit now.
 434  			break
 435  		}
 436  		// Check that we didn't cross the lower boundary.
 437  		// The case where l < u but c == l-1 is essentially impossible,
 438  		// but may happen if:
 439  		//    lower   = ..11
 440  		//    central = ..19
 441  		//    upper   = ..31
 442  		// and means that 'central' is very close but less than
 443  		// an integer ending with many zeros, and usually
 444  		// the "round-up" logic hides the problem.
 445  		if l == c+1 && c < u {
 446  			c++
 447  			cdigit = 0
 448  			cup = false
 449  		}
 450  		trimmed++
 451  		// Remember trimmed digits of c
 452  		c0 = c0 && cNextDigit == 0
 453  		cNextDigit = int(cdigit)
 454  		lower, central, upper = l, c, u
 455  	}
 456  	// should we round up?
 457  	if trimmed > 0 {
 458  		cup = cNextDigit > 5 ||
 459  			(cNextDigit == 5 && !c0) ||
 460  			(cNextDigit == 5 && c0 && central&1 == 1)
 461  	}
 462  	if central < upper && cup {
 463  		central++
 464  	}
 465  	// We know where the number ends, fill directly
 466  	endindex -= trimmed
 467  	v := central
 468  	n := endindex
 469  	for n > d.nd {
 470  		v1, v2 := v/100, v%100
 471  		d.d[n] = smallsString[2*v2+1]
 472  		d.d[n-1] = smallsString[2*v2+0]
 473  		n -= 2
 474  		v = v1
 475  	}
 476  	if n == d.nd {
 477  		d.d[n] = byte(v + '0')
 478  	}
 479  	d.nd = endindex + 1
 480  	d.dp = d.nd + trimmed
 481  }
 482  
 483  // mult64bitPow10 takes a floating-point input with a 25-bit
 484  // mantissa and multiplies it with 10^q. The resulting mantissa
 485  // is m*P >> 57 where P is a 64-bit element of the detailedPowersOfTen tables.
 486  // It is typically 31 or 32-bit wide.
 487  // The returned boolean is true if all trimmed bits were zero.
 488  //
 489  // That is:
 490  //
 491  //	m*2^e2 * round(10^q) = resM * 2^resE + ε
 492  //	exact = ε == 0
 493  func mult64bitPow10(m uint32, e2, q int) (resM uint32, resE int, exact bool) {
 494  	if q == 0 {
 495  		// P == 1<<63
 496  		return m << 6, e2 - 6, true
 497  	}
 498  	if q < detailedPowersOfTenMinExp10 || detailedPowersOfTenMaxExp10 < q {
 499  		// This never happens due to the range of float32/float64 exponent
 500  		panic("mult64bitPow10: power of 10 is out of range")
 501  	}
 502  	pow := detailedPowersOfTen[q-detailedPowersOfTenMinExp10][1]
 503  	if q < 0 {
 504  		// Inverse powers of ten must be rounded up.
 505  		pow += 1
 506  	}
 507  	hi, lo := bits.Mul64(uint64(m), pow)
 508  	e2 += mulByLog10Log2(q) - 63 + 57
 509  	return uint32(hi<<7 | lo>>57), e2, lo<<7 == 0
 510  }
 511  
 512  // mult128bitPow10 takes a floating-point input with a 55-bit
 513  // mantissa and multiplies it with 10^q. The resulting mantissa
 514  // is m*P >> 119 where P is a 128-bit element of the detailedPowersOfTen tables.
 515  // It is typically 63 or 64-bit wide.
 516  // The returned boolean is true is all trimmed bits were zero.
 517  //
 518  // That is:
 519  //
 520  //	m*2^e2 * round(10^q) = resM * 2^resE + ε
 521  //	exact = ε == 0
 522  func mult128bitPow10(m uint64, e2, q int) (resM uint64, resE int, exact bool) {
 523  	if q == 0 {
 524  		// P == 1<<127
 525  		return m << 8, e2 - 8, true
 526  	}
 527  	if q < detailedPowersOfTenMinExp10 || detailedPowersOfTenMaxExp10 < q {
 528  		// This never happens due to the range of float32/float64 exponent
 529  		panic("mult128bitPow10: power of 10 is out of range")
 530  	}
 531  	pow := detailedPowersOfTen[q-detailedPowersOfTenMinExp10]
 532  	if q < 0 {
 533  		// Inverse powers of ten must be rounded up.
 534  		pow[0] += 1
 535  	}
 536  	e2 += mulByLog10Log2(q) - 127 + 119
 537  
 538  	// long multiplication
 539  	l1, l0 := bits.Mul64(m, pow[0])
 540  	h1, h0 := bits.Mul64(m, pow[1])
 541  	mid, carry := bits.Add64(l1, h0, 0)
 542  	h1 += carry
 543  	return h1<<9 | mid>>55, e2, mid<<9 == 0 && l0 == 0
 544  }
 545  
 546  func divisibleByPower5(m uint64, k int) bool {
 547  	if m == 0 {
 548  		return true
 549  	}
 550  	for i := 0; i < k; i++ {
 551  		if m%5 != 0 {
 552  			return false
 553  		}
 554  		m /= 5
 555  	}
 556  	return true
 557  }
 558  
 559  // divmod1e9 computes quotient and remainder of division by 1e9,
 560  // avoiding runtime uint64 division on 32-bit platforms.
 561  func divmod1e9(x uint64) (uint32, uint32) {
 562  	if !host32bit {
 563  		return uint32(x / 1e9), uint32(x % 1e9)
 564  	}
 565  	// Use the same sequence of operations as the amd64 compiler.
 566  	hi, _ := bits.Mul64(x>>1, 0x89705f4136b4a598) // binary digits of 1e-9
 567  	q := hi >> 28
 568  	return uint32(q), uint32(x - q*1e9)
 569  }
 570