internal static int GetCachedPower(int e, int alpha, int gamma, DiyFp c_mk)
{
int kQ = DiyFp.kSignificandSize;
double k = System.Math.Ceiling((alpha - e + kQ - 1) * kD_1_LOG2_10);
int index = (GRISU_CACHE_OFFSET + (int)k - 1) / CACHED_POWERS_SPACING + 1;
CachedPowers.CachedPower cachedPower = CACHED_POWERS[index];
c_mk.SetF(cachedPower.significand);
c_mk.SetE(cachedPower.binaryExponent);
System.Diagnostics.Debug.Assert(((alpha <= c_mk.E() + e) && (c_mk.E() + e <= gamma)));
return cachedPower.decimalExponent;
}
// Returns the two boundaries of first argument.
// The bigger boundary (m_plus) is normalized. The lower boundary has the same
// exponent as m_plus.
internal static void NormalizedBoundaries(long d64, DiyFp m_minus, DiyFp m_plus)
{
DiyFp v = AsDiyFp(d64);
bool significand_is_zero = (v.F() == kHiddenBit);
m_plus.SetF((v.F() << 1) + 1);
m_plus.SetE(v.E() - 1);
m_plus.Normalize();
if (significand_is_zero && v.E() != kDenormalExponent)
{
// The boundary is closer. Think of v = 1000e10 and v- = 9999e9.
// Then the boundary (== (v - v-)/2) is not just at a distance of 1e9 but
// at a distance of 1e8.
// The only exception is for the smallest normal: the largest denormal is
// at the same distance as its successor.
// Note: denormals have the same exponent as the smallest normals.
m_minus.SetF((v.F() << 2) - 1);
m_minus.SetE(v.E() - 2);
}
else
{
m_minus.SetF((v.F() << 1) - 1);
m_minus.SetE(v.E() - 1);
}
m_minus.SetF(m_minus.F() << (m_minus.E() - m_plus.E()));
m_minus.SetE(m_plus.E());
}
// Provides a decimal representation of v.
// Returns true if it succeeds, otherwise the result cannot be trusted.
// There will be *length digits inside the buffer (not null-terminated).
// If the function returns true then
// v == (double) (buffer * 10^decimal_exponent).
// The digits in the buffer are the shortest representation possible: no
// 0.09999999999999999 instead of 0.1. The shorter representation will even be
// chosen even if the longer one would be closer to v.
// The last digit will be closest to the actual v. That is, even if several
// digits might correctly yield 'v' when read again, the closest will be
// computed.
internal static bool Grisu3(double v, FastDtoaBuilder buffer)
{
long bits = System.BitConverter.DoubleToInt64Bits(v);
DiyFp w = DoubleHelper.AsNormalizedDiyFp(bits);
// boundary_minus and boundary_plus are the boundaries between v and its
// closest floating-point neighbors. Any number strictly between
// boundary_minus and boundary_plus will round to v when convert to a double.
// Grisu3 will never output representations that lie exactly on a boundary.
DiyFp boundary_minus = new DiyFp();
DiyFp boundary_plus = new DiyFp();
DoubleHelper.NormalizedBoundaries(bits, boundary_minus, boundary_plus);
System.Diagnostics.Debug.Assert((boundary_plus.E() == w.E()));
DiyFp ten_mk = new DiyFp();
// Cached power of ten: 10^-k
int mk = CachedPowers.GetCachedPower(w.E() + DiyFp.kSignificandSize, minimal_target_exponent, maximal_target_exponent, ten_mk);
System.Diagnostics.Debug.Assert((minimal_target_exponent <= w.E() + ten_mk.E() + DiyFp.kSignificandSize && maximal_target_exponent >= w.E() + ten_mk.E() + DiyFp.kSignificandSize));
// Note that ten_mk is only an approximation of 10^-k. A DiyFp only contains a
// 64 bit significand and ten_mk is thus only precise up to 64 bits.
// The DiyFp::Times procedure rounds its result, and ten_mk is approximated
// too. The variable scaled_w (as well as scaled_boundary_minus/plus) are now
// off by a small amount.
// In fact: scaled_w - w*10^k < 1ulp (unit in the last place) of scaled_w.
// In other words: let f = scaled_w.f() and e = scaled_w.e(), then
// (f-1) * 2^e < w*10^k < (f+1) * 2^e
DiyFp scaled_w = DiyFp.Times(w, ten_mk);
System.Diagnostics.Debug.Assert((scaled_w.E() == boundary_plus.E() + ten_mk.E() + DiyFp.kSignificandSize));
// In theory it would be possible to avoid some recomputations by computing
// the difference between w and boundary_minus/plus (a power of 2) and to
// compute scaled_boundary_minus/plus by subtracting/adding from
// scaled_w. However the code becomes much less readable and the speed
// enhancements are not terriffic.
DiyFp scaled_boundary_minus = DiyFp.Times(boundary_minus, ten_mk);
DiyFp scaled_boundary_plus = DiyFp.Times(boundary_plus, ten_mk);
// DigitGen will generate the digits of scaled_w. Therefore we have
// v == (double) (scaled_w * 10^-mk).
// Set decimal_exponent == -mk and pass it to DigitGen. If scaled_w is not an
// integer than it will be updated. For instance if scaled_w == 1.23 then
// the buffer will be filled with "123" und the decimal_exponent will be
// decreased by 2.
return DigitGen(scaled_boundary_minus, scaled_w, scaled_boundary_plus, buffer, mk);
}
// Generates the digits of input number w.
// w is a floating-point number (DiyFp), consisting of a significand and an
// exponent. Its exponent is bounded by minimal_target_exponent and
// maximal_target_exponent.
// Hence -60 <= w.e() <= -32.
//
// Returns false if it fails, in which case the generated digits in the buffer
// should not be used.
// Preconditions:
// * low, w and high are correct up to 1 ulp (unit in the last place). That
// is, their error must be less that a unit of their last digits.
// * low.e() == w.e() == high.e()
// * low < w < high, and taking into account their error: low~ <= high~
// * minimal_target_exponent <= w.e() <= maximal_target_exponent
// Postconditions: returns false if procedure fails.
// otherwise:
// * buffer is not null-terminated, but len contains the number of digits.
// * buffer contains the shortest possible decimal digit-sequence
// such that LOW < buffer * 10^kappa < HIGH, where LOW and HIGH are the
// correct values of low and high (without their error).
// * if more than one decimal representation gives the minimal number of
// decimal digits then the one closest to W (where W is the correct value
// of w) is chosen.
// Remark: this procedure takes into account the imprecision of its input
// numbers. If the precision is not enough to guarantee all the postconditions
// then false is returned. This usually happens rarely (~0.5%).
//
// Say, for the sake of example, that
// w.e() == -48, and w.f() == 0x1234567890abcdef
// w's value can be computed by w.f() * 2^w.e()
// We can obtain w's integral digits by simply shifting w.f() by -w.e().
// -> w's integral part is 0x1234
// w's fractional part is therefore 0x567890abcdef.
// Printing w's integral part is easy (simply print 0x1234 in decimal).
// In order to print its fraction we repeatedly multiply the fraction by 10 and
// get each digit. Example the first digit after the point would be computed by
// (0x567890abcdef * 10) >> 48. -> 3
// The whole thing becomes slightly more complicated because we want to stop
// once we have enough digits. That is, once the digits inside the buffer
// represent 'w' we can stop. Everything inside the interval low - high
// represents w. However we have to pay attention to low, high and w's
// imprecision.
internal static bool DigitGen(DiyFp low, DiyFp w, DiyFp high, FastDtoaBuilder buffer, int mk)
{
System.Diagnostics.Debug.Assert((low.E() == w.E() && w.E() == high.E()));
System.Diagnostics.Debug.Assert(Uint64_lte(low.F() + 1, high.F() - 1));
System.Diagnostics.Debug.Assert((minimal_target_exponent <= w.E() && w.E() <= maximal_target_exponent));
// low, w and high are imprecise, but by less than one ulp (unit in the last
// place).
// If we remove (resp. add) 1 ulp from low (resp. high) we are certain that
// the new numbers are outside of the interval we want the final
// representation to lie in.
// Inversely adding (resp. removing) 1 ulp from low (resp. high) would yield
// numbers that are certain to lie in the interval. We will use this fact
// later on.
// We will now start by generating the digits within the uncertain
// interval. Later we will weed out representations that lie outside the safe
// interval and thus _might_ lie outside the correct interval.
long unit = 1;
DiyFp too_low = new DiyFp(low.F() - unit, low.E());
DiyFp too_high = new DiyFp(high.F() + unit, high.E());
// too_low and too_high are guaranteed to lie outside the interval we want the
// generated number in.
DiyFp unsafe_interval = DiyFp.Minus(too_high, too_low);
// We now cut the input number into two parts: the integral digits and the
// fractionals. We will not write any decimal separator though, but adapt
// kappa instead.
// Reminder: we are currently computing the digits (stored inside the buffer)
// such that: too_low < buffer * 10^kappa < too_high
// We use too_high for the digit_generation and stop as soon as possible.
// If we stop early we effectively round down.
DiyFp one = new DiyFp(1l << -w.E(), w.E());
// Division by one is a shift.
int integrals = (int)(((long)(((ulong)too_high.F()) >> -one.E())) & unchecked((long)(0xffffffffL)));
// Modulo by one is an and.
long fractionals = too_high.F() & (one.F() - 1);
long result = BiggestPowerTen(integrals, DiyFp.kSignificandSize - (-one.E()));
int divider = (int)(((long)(((ulong)result) >> 32)) & unchecked((long)(0xffffffffL)));
int divider_exponent = (int)(result & unchecked((long)(0xffffffffL)));
int kappa = divider_exponent + 1;
// Loop invariant: buffer = too_high / 10^kappa (integer division)
// The invariant holds for the first iteration: kappa has been initialized
// with the divider exponent + 1. And the divider is the biggest power of ten
// that is smaller than integrals.
while (kappa > 0)
{
int digit = integrals / divider;
buffer.Append((char)('0' + digit));
integrals %= divider;
kappa--;
// Note that kappa now equals the exponent of the divider and that the
// invariant thus holds again.
long rest = ((long)integrals << -one.E()) + fractionals;
// Invariant: too_high = buffer * 10^kappa + DiyFp(rest, one.e())
// Reminder: unsafe_interval.e() == one.e()
if (rest < unsafe_interval.F())
{
// Rounding down (by not emitting the remaining digits) yields a number
// that lies within the unsafe interval.
buffer.point = buffer.end - mk + kappa;
return RoundWeed(buffer, DiyFp.Minus(too_high, w).F(), unsafe_interval.F(), rest, (long)divider << -one.E(), unit);
}
divider /= 10;
}
// The integrals have been generated. We are at the point of the decimal
// separator. In the following loop we simply multiply the remaining digits by
// 10 and divide by one. We just need to pay attention to multiply associated
// data (like the interval or 'unit'), too.
// Instead of multiplying by 10 we multiply by 5 (cheaper operation) and
// increase its (imaginary) exponent. At the same time we decrease the
// divider's (one's) exponent and shift its significand.
// Basically, if fractionals was a DiyFp (with fractionals.e == one.e):
// fractionals.f *= 10;
// fractionals.f >>= 1; fractionals.e++; // value remains unchanged.
// one.f >>= 1; one.e++; // value remains unchanged.
// and we have again fractionals.e == one.e which allows us to divide
// fractionals.f() by one.f()
// We simply combine the *= 10 and the >>= 1.
while (true)
{
fractionals *= 5;
unit *= 5;
unsafe_interval.SetF(unsafe_interval.F() * 5);
unsafe_interval.SetE(unsafe_interval.E() + 1);
// Will be optimized out.
one.SetF((long)(((ulong)one.F()) >> 1));
one.SetE(one.E() + 1);
// Integer division by one.
int digit = (int)(((long)(((ulong)fractionals) >> -one.E())) & unchecked((long)(0xffffffffL)));
buffer.Append((char)('0' + digit));
fractionals &= one.F() - 1;
// Modulo by one.
kappa--;
if (fractionals < unsafe_interval.F())
{
buffer.point = buffer.end - mk + kappa;
return RoundWeed(buffer, DiyFp.Minus(too_high, w).F() * unit, unsafe_interval.F(), fractionals, one.F(), unit);
}
}
}
