ARCHIVE/rippled
1
0
Fork 0
mirror of https://github.com/XRPLF/rippled.git synced 2026-06-03 08:46:46 +00:00
Code Issues Packages Projects Releases Wiki Activity

Significant rewrite

Browse Source 
- Simplify Number::operator/= to use more constexpr values, and fewer
  variations.
  - Most significantly, rounding up doesn't need more precision, it only
    needs to know if there's a remainder after the current precision work
    is done. Tracked similarly Guard::xbit_.
- Build a constexpr lookup array for powers of 10. Only a handful of
  values are used, but since it's built at compile time, and constexpr,
  unused values do not affect memory or performance.
This commit is contained in:
Ed Hennis
2026-05-27 00:07:48 -04:00
parent 12670b0c3f
commit 5abecb9fcb
3 changed files with 194 additions and 66 deletions
Download Patch File Download Diff File Expand all files Collapse all files

169
src/libxrpl/basics/Number.cpp
View File

@@ -178,6 +178,10 @@ public:
setPositive() noexcept;
void
setNegative() noexcept;
// Should only be called by doNormalize, and then only for division
// operations with remainders.
void
setDropped() noexcept;
[[nodiscard]] bool
isNegative() const noexcept;
@@ -250,6 +254,12 @@ Number::Guard::setNegative() noexcept
sbit_ = 1;
}
inline void
Number::Guard::setDropped() noexcept
{
xbit_ = 1;
}
inline bool
Number::Guard::isNegative() const noexcept
{
@@ -512,8 +522,6 @@ Number::one()
return Number{false, range.min, -range.log, Number::Unchecked{}};
}
// Use the member names in this static function for now so the diff is cleaner
// TODO: Rename the function parameters to get rid of the "_" suffix
template <class T>
void
doNormalize(
@@ -522,7 +530,8 @@ doNormalize(
int& exponent,
MantissaRange::rep const& minMantissa,
MantissaRange::rep const& maxMantissa,
MantissaRange::CuspRoundingFix cuspRoundingFixEnabled)
MantissaRange::CuspRoundingFix cuspRoundingFixEnabled,
bool dropped)
{
static constexpr auto kMinExponent = Number::kMinExponent;
static constexpr auto kMaxExponent = Number::kMaxExponent;
@@ -547,6 +556,8 @@ doNormalize(
Guard g;
if (negative)
g.setNegative();
if (dropped)
g.setDropped();
while (m > maxMantissa)
{
if (exponent >= kMaxExponent)
@@ -611,7 +622,8 @@ Number::normalize<uint128_t>(
internalrep const& maxMantissa,
MantissaRange::CuspRoundingFix cuspRoundingFixEnabled)
{
doNormalize(negative, mantissa, exponent, minMantissa, maxMantissa, cuspRoundingFixEnabled);
doNormalize(
negative, mantissa, exponent, minMantissa, maxMantissa, cuspRoundingFixEnabled, false);
}
template <>
@@ -624,7 +636,8 @@ Number::normalize<unsigned long long>(
internalrep const& maxMantissa,
MantissaRange::CuspRoundingFix cuspRoundingFixEnabled)
{
doNormalize(negative, mantissa, exponent, minMantissa, maxMantissa, cuspRoundingFixEnabled);
doNormalize(
negative, mantissa, exponent, minMantissa, maxMantissa, cuspRoundingFixEnabled, false);
}
template <>
@@ -637,7 +650,8 @@ Number::normalize<unsigned long>(
internalrep const& maxMantissa,
MantissaRange::CuspRoundingFix cuspRoundingFixEnabled)
{
doNormalize(negative, mantissa, exponent, minMantissa, maxMantissa, cuspRoundingFixEnabled);
doNormalize(
negative, mantissa, exponent, minMantissa, maxMantissa, cuspRoundingFixEnabled, false);
}
void
@@ -858,32 +872,62 @@ Number::operator/=(Number const& y)
auto const& maxMantissa = range.max;
auto const cuspRoundingFixEnabled = range.cuspRoundingFixEnabled;
// Cache power of 10 computations to not waste time on subsequent calls
auto getPower10 = [](int exponent) {
static std::unordered_map<int, uint128_t> powersOf10;
if (!powersOf10.contains(exponent))
{
uint128_t result = 1;
for (int i = 0; i < exponent; ++i)
result *= 10;
powersOf10[exponent] = result;
}
return powersOf10.at(exponent);
};
// Shift by 10^17 gives greatest precision while not overflowing
// uint128_t or the cast back to int64_t for small mantissas.
// Division operates on two large integers (16-digit for small
// mantissas, 19-digit for large) using integer math. If the values
// were just divided directly, the result would be at most 9.
// Introduce a power-of-ten multiplication factor for the numerator
// which will ensure the result has a meaningful number of digits.
//
// Consider numbers with a 2-digit mantissa:
// * Assume both numbers have an exponent of 0, using "ToNearest" rounding
// * 23 / 67 = 0
// * Use a factor of 10^4
// * 230'000 / 67 = 3432 with an exponent of -4
// * The normalized result will be 34, exponent -2
//
// The most extreme results are 10/99 and 99/10
// * 100'000 / 99 = 1'010e-4 = 10e-2
// * 990'000 / 10 = 99'000e-4 = 99e-1
//
// Note that the computations give 2 or 3 digits after the
// decimal point to determine which way to round for most scenarios.
//
// For small mantissas (where the MantissaRange.log == 15), shifting by 10^17 gives sufficient
// precision while not overflowing uint128_t or the cast back to int64_t. (This is legacy
// behavior, which must not be changed.)
//
// For large mantissas (where the MantissaRange.log == 18), a shift by 10^20 would be optimal
// for most scenarios. However, larger mantissa values would overflow 2^128 (log(2^128,10)
// ~ 38.5).
//
// For large mantissas:
// * log(2^128,10) ~ 38.5
// * largeRange.log = 18, fits in 10^19
// * The expanded numerator must fit in 10^38
// * f can be up to 10^(38-19) = 10^19 safely
bool const small = range.scale == MantissaRange::MantissaScale::Small;
auto const factorExponent = small ? 17 : 19;
// * f not be more than 10^(38-19) = 10^19 safely
//
// So, we do the division into stages:
//
// Stage 1: Use the same factor of 10^17, for the initial division. This
// will frequently not result in a whole number quotient.
//
// Stage 2: If there is a remainder from the first step, repeat the
// process with a "correction" factor of 10^5. Shift the
// result of Stage 1 over by 5 places, and add the second result to it.
// This is equivalent to if we had used an initial factor of 10^22,
// a couple digits more than we actually need.
//
// Stage 3: If there is still a remainder, and the CuspRoundingFix
// is enabled, pass a flag indicating such to doNormalize. The Guard
// in doNormalize will treat that flag as if non-zero digits had
// been dropped from the mantissa when shrinking it into range.
// This is only relevant when rounding away from zero (Upward for
// positive numbers, Downward for negative), or if the "regular"
// remainder is exactly 0.5 for "ToNearest". This will give the
// rounding the most accurate result possible, as if infinite
// precision was used in the initial calculation.
auto constexpr factorExponent = 17;
uint128_t const f = getPower10(factorExponent);
XRPL_ASSERT_PARTS(f >= minMantissa * 10, "Number::operator/=", "factor expected size");
uint128_t constexpr f = kPowerOfTen[factorExponent];
// unsigned denominator
auto const dmu = static_cast<uint128_t>(dm);
@@ -892,21 +936,21 @@ Number::operator/=(Number const& y)
auto zm = numerator / dmu;
auto ze = ne - de - factorExponent;
bool zn = (ns * ds) < 0;
if (!small)
{
// zm may now hold up to 20 digits. The correctionFactor must be small enough to not cause
// overflow, but as large as possible otherwise
// * zm fits in 10^21
// * correctionFactor can be up to 10^(38-21) = 10^17 safely
bool const useBigCorrection =
cuspRoundingFixEnabled == MantissaRange::CuspRoundingFix::Enabled;
auto const correctionExponent = useBigCorrection ? 17 : 3;
uint128_t const correctionFactor = getPower10(correctionExponent);
// dropped is used in the same way as Guard::xbit_. In the case of
// division, it indicates if there's any remainder left over after
// we have been as precise as reasonable. If there is, it would be as
// if we were using infinite precision math, and a non-zero digit
// had been shifted off the end of the result when normalizing.
bool dropped = false;
// Virtually multiply numerator by correctionFactor. Since that would
// overflow in the existing uint128_t, we'll do that part separately.
if (range.scale != MantissaRange::MantissaScale::Small)
{
// Stage 2
//
// If there is a remainder, treat is as a secondary numerator.
// Multiply by correctionFactor separately from stage 1.
// The math for this would work for small mantissas, but we need to
// preserve existing behavior.
// preserve legacy behavior.
//
// Consider:
// ((numerator * correctionFactor) / dmu) / correctionFactor
@@ -917,24 +961,47 @@ Number::operator/=(Number const& y)
//
// = ((numerator / dmu * correctionFactor)
// + ((numerator % dmu) * correctionFactor) / dmu) / correctionFactor
// = ((zm * correctionFactor)
// + (remainder * correctionFactor) / dmu) / correctionFactor
//
// We have already set `mantissa_ = numerator / dmu`. Now we
// compute `remainder = numerator % dmu`, and if it is
// nonzero, we do the rest of the arithmetic. If it's zero, we can skip
// it.
// The trick is that multiplication by correctionFactor is done on the mantissa, but
// division by correctionFactor is done by modifying the exponent, so no precision is lost
// until we normalize.
//
// If remainder is zero, we can skip this stage entirely because
// the first stage gave an exact answer.
auto constexpr correctionExponent = 5;
uint128_t constexpr correctionFactor = kPowerOfTen[correctionExponent];
static_assert(factorExponent + correctionExponent == 22);
auto const remainder = (numerator % dmu);
if (remainder != 0)
{
zm *= correctionFactor;
// divide by the correctionFactor by moving the exponent, so we don't lose the
// integer value we just computed
ze -= correctionExponent;
auto const partialNumerator = remainder * correctionFactor;
auto const correction = partialNumerator / dmu;
auto const correction = (remainder * correctionFactor) / dmu;
zm += correction;
if (correction != 0)
{
zm *= correctionFactor;
// divide by the correctionFactor by moving the exponent, so we don't lose the
// integer value we just computed
ze -= correctionExponent;
zm += correction;
}
// Stage 3: If there's still anything left, and the cusp
// rounding fix is enabled, flag if there is still
// a remainder from stage 2.
bool const useTrailingRemainder =
cuspRoundingFixEnabled == MantissaRange::CuspRoundingFix::Enabled;
if (useTrailingRemainder)
{
dropped = partialNumerator % dmu != 0;
}
}
}
normalize(zn, zm, ze, minMantissa, maxMantissa, cuspRoundingFixEnabled);
doNormalize(zn, zm, ze, minMantissa, maxMantissa, cuspRoundingFixEnabled, dropped);
negative_ = zn;
mantissa_ = static_cast<internalrep>(zm);
exponent_ = ze;

4
src/test/basics/Number_test.cpp
View File

@@ -50,11 +50,15 @@ class Number_test : public beast::unit_test::Suite
{
T p = 1;
if (n >= 0)
{
for (int i = 0; i < n; ++i)
p *= 10;
}
else
{
for (int i = 0; i < -n; ++i)
p /= 10;
}
return p;
}