| 1 |
#include <math.h> |
| 2 |
#include <stdint.h> |
| 3 |
|
| 4 |
double fmod(double x, double y) |
| 5 |
{ |
| 6 |
union {double f; uint64_t i;} ux = {x}, uy = {y}; |
| 7 |
int ex = ux.i>>52 & 0x7ff; |
| 8 |
int ey = uy.i>>52 & 0x7ff; |
| 9 |
int sx = ux.i>>63; |
| 10 |
uint64_t i; |
| 11 |
|
| 12 |
/* in the followings uxi should be ux.i, but then gcc wrongly adds */ |
| 13 |
/* float load/store to inner loops ruining performance and code size */ |
| 14 |
uint64_t uxi = ux.i; |
| 15 |
|
| 16 |
if (uy.i<<1 == 0 || isnan(y) || ex == 0x7ff) |
| 17 |
return (x*y)/(x*y); |
| 18 |
if (uxi<<1 <= uy.i<<1) { |
| 19 |
if (uxi<<1 == uy.i<<1) |
| 20 |
return 0*x; |
| 21 |
return x; |
| 22 |
} |
| 23 |
|
| 24 |
/* normalize x and y */ |
| 25 |
if (!ex) { |
| 26 |
for (i = uxi<<12; i>>63 == 0; ex--, i <<= 1); |
| 27 |
uxi <<= -ex + 1; |
| 28 |
} else { |
| 29 |
uxi &= -1ULL >> 12; |
| 30 |
uxi |= 1ULL << 52; |
| 31 |
} |
| 32 |
if (!ey) { |
| 33 |
for (i = uy.i<<12; i>>63 == 0; ey--, i <<= 1); |
| 34 |
uy.i <<= -ey + 1; |
| 35 |
} else { |
| 36 |
uy.i &= -1ULL >> 12; |
| 37 |
uy.i |= 1ULL << 52; |
| 38 |
} |
| 39 |
|
| 40 |
/* x mod y */ |
| 41 |
for (; ex > ey; ex--) { |
| 42 |
i = uxi - uy.i; |
| 43 |
if (i >> 63 == 0) { |
| 44 |
if (i == 0) |
| 45 |
return 0*x; |
| 46 |
uxi = i; |
| 47 |
} |
| 48 |
uxi <<= 1; |
| 49 |
} |
| 50 |
i = uxi - uy.i; |
| 51 |
if (i >> 63 == 0) { |
| 52 |
if (i == 0) |
| 53 |
return 0*x; |
| 54 |
uxi = i; |
| 55 |
} |
| 56 |
for (; uxi>>52 == 0; uxi <<= 1, ex--); |
| 57 |
|
| 58 |
/* scale result */ |
| 59 |
if (ex > 0) { |
| 60 |
uxi -= 1ULL << 52; |
| 61 |
uxi |= (uint64_t)ex << 52; |
| 62 |
} else { |
| 63 |
uxi >>= -ex + 1; |
| 64 |
} |
| 65 |
uxi |= (uint64_t)sx << 63; |
| 66 |
ux.i = uxi; |
| 67 |
return ux.f; |
| 68 |
} |