FloatUtils: Remove union type punning from ApproximateReciprocal functions
This form of type punning invokes undefined behavior in C++
This commit is contained in:
parent
46a4243d9a
commit
fe218ea3f6
|
@ -5,6 +5,7 @@
|
||||||
#include "Common/FloatUtils.h"
|
#include "Common/FloatUtils.h"
|
||||||
|
|
||||||
#include <cmath>
|
#include <cmath>
|
||||||
|
#include <cstring>
|
||||||
|
|
||||||
namespace Common
|
namespace Common
|
||||||
{
|
{
|
||||||
|
@ -99,20 +100,20 @@ const std::array<BaseAndDec, 32> frsqrte_expected = {{
|
||||||
|
|
||||||
double ApproximateReciprocalSquareRoot(double val)
|
double ApproximateReciprocalSquareRoot(double val)
|
||||||
{
|
{
|
||||||
union
|
s64 integral;
|
||||||
{
|
std::memcpy(&integral, &val, sizeof(integral));
|
||||||
double valf;
|
|
||||||
s64 vali;
|
s64 mantissa = integral & ((1LL << 52) - 1);
|
||||||
};
|
const s64 sign = integral & (1ULL << 63);
|
||||||
valf = val;
|
s64 exponent = integral & (0x7FFLL << 52);
|
||||||
s64 mantissa = vali & ((1LL << 52) - 1);
|
|
||||||
s64 sign = vali & (1ULL << 63);
|
|
||||||
s64 exponent = vali & (0x7FFLL << 52);
|
|
||||||
|
|
||||||
// Special case 0
|
// Special case 0
|
||||||
if (mantissa == 0 && exponent == 0)
|
if (mantissa == 0 && exponent == 0)
|
||||||
|
{
|
||||||
return sign ? -std::numeric_limits<double>::infinity() :
|
return sign ? -std::numeric_limits<double>::infinity() :
|
||||||
std::numeric_limits<double>::infinity();
|
std::numeric_limits<double>::infinity();
|
||||||
|
}
|
||||||
|
|
||||||
// Special case NaN-ish numbers
|
// Special case NaN-ish numbers
|
||||||
if (exponent == (0x7FFLL << 52))
|
if (exponent == (0x7FFLL << 52))
|
||||||
{
|
{
|
||||||
|
@ -124,7 +125,7 @@ double ApproximateReciprocalSquareRoot(double val)
|
||||||
return 0.0;
|
return 0.0;
|
||||||
}
|
}
|
||||||
|
|
||||||
return 0.0 + valf;
|
return 0.0 + val;
|
||||||
}
|
}
|
||||||
|
|
||||||
// Negative numbers return NaN
|
// Negative numbers return NaN
|
||||||
|
@ -143,15 +144,18 @@ double ApproximateReciprocalSquareRoot(double val)
|
||||||
exponent += 1LL << 52;
|
exponent += 1LL << 52;
|
||||||
}
|
}
|
||||||
|
|
||||||
bool odd_exponent = !(exponent & (1LL << 52));
|
const bool odd_exponent = !(exponent & (1LL << 52));
|
||||||
exponent = ((0x3FFLL << 52) - ((exponent - (0x3FELL << 52)) / 2)) & (0x7FFLL << 52);
|
exponent = ((0x3FFLL << 52) - ((exponent - (0x3FELL << 52)) / 2)) & (0x7FFLL << 52);
|
||||||
|
integral = sign | exponent;
|
||||||
|
|
||||||
int i = (int)(mantissa >> 37);
|
const int i = static_cast<int>(mantissa >> 37);
|
||||||
vali = sign | exponent;
|
const int index = i / 2048 + (odd_exponent ? 16 : 0);
|
||||||
int index = i / 2048 + (odd_exponent ? 16 : 0);
|
|
||||||
const auto& entry = frsqrte_expected[index];
|
const auto& entry = frsqrte_expected[index];
|
||||||
vali |= (s64)(entry.m_base - entry.m_dec * (i % 2048)) << 26;
|
integral |= static_cast<s64>(entry.m_base - entry.m_dec * (i % 2048)) << 26;
|
||||||
return valf;
|
|
||||||
|
double result;
|
||||||
|
std::memcpy(&result, &integral, sizeof(result));
|
||||||
|
return result;
|
||||||
}
|
}
|
||||||
|
|
||||||
const std::array<BaseAndDec, 32> fres_expected = {{
|
const std::array<BaseAndDec, 32> fres_expected = {{
|
||||||
|
@ -167,44 +171,43 @@ const std::array<BaseAndDec, 32> fres_expected = {{
|
||||||
// Used by fres and ps_res.
|
// Used by fres and ps_res.
|
||||||
double ApproximateReciprocal(double val)
|
double ApproximateReciprocal(double val)
|
||||||
{
|
{
|
||||||
union
|
s64 integral;
|
||||||
{
|
std::memcpy(&integral, &val, sizeof(integral));
|
||||||
double valf;
|
|
||||||
s64 vali;
|
|
||||||
};
|
|
||||||
|
|
||||||
valf = val;
|
const s64 mantissa = integral & ((1LL << 52) - 1);
|
||||||
s64 mantissa = vali & ((1LL << 52) - 1);
|
const s64 sign = integral & (1ULL << 63);
|
||||||
s64 sign = vali & (1ULL << 63);
|
s64 exponent = integral & (0x7FFLL << 52);
|
||||||
s64 exponent = vali & (0x7FFLL << 52);
|
|
||||||
|
|
||||||
// Special case 0
|
// Special case 0
|
||||||
if (mantissa == 0 && exponent == 0)
|
if (mantissa == 0 && exponent == 0)
|
||||||
return std::copysign(std::numeric_limits<double>::infinity(), valf);
|
return std::copysign(std::numeric_limits<double>::infinity(), val);
|
||||||
|
|
||||||
// Special case NaN-ish numbers
|
// Special case NaN-ish numbers
|
||||||
if (exponent == (0x7FFLL << 52))
|
if (exponent == (0x7FFLL << 52))
|
||||||
{
|
{
|
||||||
if (mantissa == 0)
|
if (mantissa == 0)
|
||||||
return std::copysign(0.0, valf);
|
return std::copysign(0.0, val);
|
||||||
return 0.0 + valf;
|
return 0.0 + val;
|
||||||
}
|
}
|
||||||
|
|
||||||
// Special case small inputs
|
// Special case small inputs
|
||||||
if (exponent < (895LL << 52))
|
if (exponent < (895LL << 52))
|
||||||
return std::copysign(std::numeric_limits<float>::max(), valf);
|
return std::copysign(std::numeric_limits<float>::max(), val);
|
||||||
|
|
||||||
// Special case large inputs
|
// Special case large inputs
|
||||||
if (exponent >= (1149LL << 52))
|
if (exponent >= (1149LL << 52))
|
||||||
return std::copysign(0.0, valf);
|
return std::copysign(0.0, val);
|
||||||
|
|
||||||
exponent = (0x7FDLL << 52) - exponent;
|
exponent = (0x7FDLL << 52) - exponent;
|
||||||
|
|
||||||
int i = (int)(mantissa >> 37);
|
const int i = static_cast<int>(mantissa >> 37);
|
||||||
const auto& entry = fres_expected[i / 1024];
|
const auto& entry = fres_expected[i / 1024];
|
||||||
vali = sign | exponent;
|
integral = sign | exponent;
|
||||||
vali |= (s64)(entry.m_base - (entry.m_dec * (i % 1024) + 1) / 2) << 29;
|
integral |= static_cast<s64>(entry.m_base - (entry.m_dec * (i % 1024) + 1) / 2) << 29;
|
||||||
return valf;
|
|
||||||
|
double result;
|
||||||
|
std::memcpy(&result, &integral, sizeof(result));
|
||||||
|
return result;
|
||||||
}
|
}
|
||||||
|
|
||||||
} // namespace Common
|
} // namespace Common
|
||||||
|
|
Loading…
Reference in New Issue