public RyuFloat64(bool sign, ulong ieeeMantissa, uint ieeeExponent)
{
// Subtract 2 more so that the bounds computation has 2 additional bits
const int DOUBLE_EXP_DIFF_P2 = DOUBLE_BIAS + DOUBLE_MANTISSA_BITS + 2;
int exponent;
ulong mantissa;
if (ieeeExponent == 0)
{
exponent = 1 - DOUBLE_EXP_DIFF_P2;
mantissa = ieeeMantissa;
}
else
{
exponent = (int)ieeeExponent - DOUBLE_EXP_DIFF_P2;
mantissa = (1UL << DOUBLE_MANTISSA_BITS) | ieeeMantissa;
}
// Step 2: Determine the interval of valid decimal representations
bool even = (mantissa & 1UL) == 0UL, acceptBounds = even, mmShift =
ieeeMantissa != 0U || ieeeExponent <= 1U;
ulong mv = mantissa << 2;
// Step 3: Convert to a decimal power base using 128-bit arithmetic
ulong vr, vm, vp;
int e10;
bool vmTrailingZeroes = false, vrTrailingZeroes = false;
if (exponent >= 0)
{
// Tried special-casing q == 0, but there was no effect on performance
// This expression is slightly faster than max(0, log10Pow2(e2) - 1)
int q = RyuUtils.Log10Pow2(exponent);
if (exponent > 3)
{
q--;
}
int k = DOUBLE_POW5_INV_BITCOUNT + RyuUtils.Pow5Bits(q) - 1, i = -exponent +
q + k;
ulong a = RyuTables.double_computeInvPow5(q, out ulong b);
e10 = q;
vr = RyuUtils.MulShiftAll(mantissa, a, b, i, out vp, out vm, mmShift);
if (q <= 21U)
{
// This should use q <= 22, but I think 21 is also safe. Smaller values
// may still be safe, but it's more difficult to reason about them.
// Only one of mp, mv, and mm can be a multiple of 5, if any
uint mvMod5 = (uint)mv % 5U;
if (mvMod5 == 0U)
{
vrTrailingZeroes = RyuUtils.IsMultipleOf5Power(mv, q);
}
else if (acceptBounds)
{
// Same as min(e2 + (~mm & 1), pow5Factor(mm)) >= q
// <=> e2 + (~mm & 1) >= q && pow5Factor(mm) >= q
// <=> true && pow5Factor(mm) >= q, since e2 >= q
vmTrailingZeroes = RyuUtils.IsMultipleOf5Power(mv - 1UL - (mmShift ?
1UL : 0UL), q);
}
else
{
// Same as min(e2 + 1, pow5Factor(mp)) >= q
vp -= RyuUtils.IsMultipleOf5Power(mv + 2UL, q) ? 1UL : 0UL;
}
}
}
else
{
// This expression is slightly faster than max(0, log10Pow5(-e2) - 1)
int q = RyuUtils.Log10Pow5(-exponent);
if (-exponent > 1)
{
q--;
}
int i = -exponent - q, k = RyuUtils.Pow5Bits(i) - DOUBLE_POW5_BITCOUNT,
j = q - k;
ulong a = RyuTables.double_computePow5(i, out ulong b);
e10 = q + exponent;
vr = RyuUtils.MulShiftAll(mantissa, a, b, j, out vp, out vm, mmShift);
if (q <= 1U)
{
// {vr,vp,vm} is trailing zeros if {mv,mp,mm} has at least q trailing 0
// bits; mv = 4 * m2, so it always has at least two trailing 0 bits
vrTrailingZeroes = true;
if (acceptBounds)
{
// mm = mv - 1 - mmShift, so it has 1 trailing 0 bit iff mmShift == 1
vmTrailingZeroes = mmShift;
}
else
{
// mp = mv + 2, so it always has at least one trailing 0 bit
--vp;
}
}
else if (q < 63U)
{
// We want to know if the fUL product has at least q trailing zeros
// We need to compute min(p2(mv), p5(mv) - e2) >= q
// <=> p2(mv) >= q && p5(mv) - e2 >= q
// <=> p2(mv) >= q (because -e2 >= q)
vrTrailingZeroes = RyuUtils.IsMultipleOf2Power(mv, q);
}
}
// Step 4: Find the shortest decimal representation in the interval of valid
// representations
int removed = 0;
uint removedDigit = 0U;
ulong output, p10, m10;
// On average, we remove ~2 digits
if (vmTrailingZeroes || vrTrailingZeroes)
{
// General case, which happens rarely (~0.7%)
while ((p10 = vp / 10UL) > (m10 = vm.DivMod(10U, out uint vmRem)))
{
if (vmRem != 0U)
{
vmTrailingZeroes = false;
}
if (removedDigit != 0U)
{
vrTrailingZeroes = false;
}
vr = vr.DivMod(10U, out removedDigit);
vp = p10;
vm = m10;
removed++;
}
if (vmTrailingZeroes)
{
while (((uint)vm - 10U * (uint)(m10 = vm / 10UL)) == 0U)
{
if (removedDigit != 0U)
{
vrTrailingZeroes = false;
}
vr = vr.DivMod(10U, out removedDigit);
vp /= 10UL;
vm = m10;
removed++;
}
}
if (vrTrailingZeroes && removedDigit == 5U && vr % 2U == 0U)
{
// Round even if the exact number is .....50..0
removedDigit = 4U;
}
// We need to take vr + 1 if vr is outside bounds or we need to round up
output = vr;
if ((vr == vm && (!acceptBounds || !vmTrailingZeroes)) || removedDigit >= 5U)
{
output++;
}
}
else
{
// Specialized for the common case (~99.3%); percentages below are relative to
// this
bool roundUp = false;
if (RyuUtils.DivCompare(ref vp, ref vm, 100UL))
{
// Optimization: remove two digits at a time (~86.2%)
vr = vr.DivMod(100U, out uint round100);
roundUp = round100 >= 50U;
removed += 2;
}
// Loop iterations below (approximately), without optimization above:
// 0: 0.03%, 1: 13.8%, 2: 70.6%, 3: 14.0%, 4: 1.40%, 5: 0.14%, 6+: 0.02%
// Loop iterations below (approximately), with optimization above:
// 0: 70.6%, 1: 27.8%, 2: 1.40%, 3: 0.14%, 4+: 0.02%
while (RyuUtils.DivCompare(ref vp, ref vm, 10UL))
{
vr = vr.DivMod(10U, out uint vrRem);
roundUp = vrRem >= 5U;
removed++;
}
// We need to take vr + 1 if vr is outside bounds or we need to round up
output = vr;
if (vr == vm || roundUp)
{
output++;
}
}
this.sign = sign;
this.exponent = e10 + removed;
this.mantissa = output;
}
/// <summary>
/// Creates a Ryu float from an IEEE 32 bit float.
/// </summary>
public RyuFloat32(bool fSign, uint ieeeMantissa, uint ieeeExponent)
{
const int FLOAT_EXP_DIFF_P2 = FLOAT_BIAS + FLOAT_MANTISSA_BITS + 2;
int exponent;
uint mantissa;
if (ieeeExponent == 0U)
{
// Subtract 2 so that the bounds computation has 2 additional bits
exponent = 1 - FLOAT_EXP_DIFF_P2;
mantissa = ieeeMantissa;
}
else
{
exponent = (int)ieeeExponent - FLOAT_EXP_DIFF_P2;
mantissa = (1U << FLOAT_MANTISSA_BITS) | ieeeMantissa;
}
bool even = (mantissa & 1U) == 0U, acceptBounds = even, mmShift =
ieeeMantissa != 0 || ieeeExponent <= 1;
// Step 2: Determine the interval of valid decimal representations
uint mv = mantissa << 2, mp = mv + 2U, mm = mv - 1U - (mmShift ? 1U : 0U);
// Step 3: Convert to a decimal power base using 64-bit arithmetic
uint vr, vp, vm;
int e10;
bool vmTrailingZeroes = false, vrTrailingZeroes = false;
uint removedDigit = 0U;
if (exponent >= 0)
{
int q = RyuUtils.Log10Pow2(exponent), k = FLOAT_POW5_INV_BITCOUNT + RyuUtils.
Pow5Bits(q) - 1, i = -exponent + q + k;
e10 = q;
vr = RyuTables.MulPow5InvDivPow2(mv, q, i);
vp = RyuTables.MulPow5InvDivPow2(mp, q, i);
vm = RyuTables.MulPow5InvDivPow2(mm, q, i);
if (q != 0U && (vp - 1U) / 10U <= vm / 10U)
{
// We need to know one removed digit even if we are not going to loop
// below. We could use q = X - 1 above, except that would require 33 bits
// for the result, and we've found that 32-bit arithmetic is faster even on
// 64-bit machines
int l = FLOAT_POW5_INV_BITCOUNT + RyuUtils.Pow5Bits(q - 1) - 1;
removedDigit = RyuTables.MulPow5InvDivPow2(mv, q - 1, -exponent +
q - 1 + l) % 10U;
}
if (q <= 9U)
{
// The largest power of 5 that fits in 24 bits is 5^10, but q <= 9 seems to
// be safe as well. Only one of mp, mv, and mm can be a multiple of 5, if
// any
if (mv % 5U == 0U)
{
vrTrailingZeroes = RyuUtils.IsMultipleOf5Power(mv, q);
}
else if (acceptBounds)
{
vmTrailingZeroes = RyuUtils.IsMultipleOf5Power(mm, q);
}
else if (RyuUtils.IsMultipleOf5Power(mp, q))
{
vp -= 1U;
}
}
}
else
{
int q = RyuUtils.Log10Pow5(-exponent), i = -exponent - q, k = RyuUtils.
Pow5Bits(i) - FLOAT_POW5_BITCOUNT, j = q - k;
e10 = q + exponent;
vr = RyuTables.MulPow5DivPow2(mv, i, j);
vp = RyuTables.MulPow5DivPow2(mp, i, j);
vm = RyuTables.MulPow5DivPow2(mm, i, j);
if (q != 0U && (vp - 1U) / 10U <= vm / 10U)
{
j = q - 1 - (RyuUtils.Pow5Bits(i + 1) - FLOAT_POW5_BITCOUNT);
removedDigit = RyuTables.MulPow5DivPow2(mv, i + 1, j) % 10U;
}
if (q <= 1U)
{
// {vr,vp,vm} is trailing zeros if {mv,mp,mm} has at least q trailing 0
// bits. mv = 4 * m2, so it always has at least two trailing 0 bits
vrTrailingZeroes = true;
if (acceptBounds)
{
// mm = mv - 1 - mmShift, so it has 1 trailing 0 bit iff mmShift == 1
vmTrailingZeroes = mmShift;
}
else
{
// mp = mv + 2, so it always has at least one trailing 0 bit
--vp;
}
}
else if (q < 31U)
{
vrTrailingZeroes = RyuUtils.IsMultipleOf2Power(mv, q - 1);
}
}
// Step 4: Find the shortest decimal representation in the interval of valid
// representations
int removed = 0;
uint output, m10, p10;
if (vmTrailingZeroes || vrTrailingZeroes)
{
// General case, which happens rarely (~4.0%)
while ((p10 = vp / 10U) > (m10 = vm.DivMod(10U, out uint vmRem)))
{
if (vmRem != 0U)
{
vmTrailingZeroes = false;
}
if (removedDigit != 0U)
{
vrTrailingZeroes = false;
}
vr = vr.DivMod(10U, out removedDigit);
vp = p10;
vm = m10;
removed++;
}
if (vmTrailingZeroes)
{
while ((vm - 10U * (m10 = vm / 10U)) == 0U)
{
if (removedDigit != 0U)
{
vrTrailingZeroes = false;
}
vr = vr.DivMod(10U, out removedDigit);
vp /= 10U;
vm = m10;
removed++;
}
}
if (vrTrailingZeroes && removedDigit == 5U && vr % 2U == 0U)
{
// Round even if the exact number is .....50..0
removedDigit = 4U;
}
// We need to take vr + 1 if vr is outside bounds or we need to round up
output = vr;
if ((vr == vm && (!acceptBounds || !vmTrailingZeroes)) || removedDigit >= 5U)
{
output++;
}
}
else
{
// Specialized for the common case (~96.0%). Percentages below are relative to
// this. Loop iterations below (approximately):
// 0: 13.6%, 1: 70.7%, 2: 14.1%, 3: 1.39%, 4: 0.14%, 5+: 0.01%
while (RyuUtils.DivCompare10(ref vp, ref vm))
{
vr = vr.DivMod(10U, out removedDigit);
removed++;
}
// We need to take vr + 1 if vr is outside bounds or we need to round up
output = vr;
if (vr == vm || removedDigit >= 5)
{
output++;
}
}
sign = fSign;
this.mantissa = output;
this.exponent = e10 + removed;
}
