+#include <fenv.h>
+#include "libm.h"
+
+#if LDBL_MANT_DIG==64 && LDBL_MAX_EXP==16384
+union ld80 {
+ long double x;
+ struct {
+ uint64_t m;
+ uint16_t e : 15;
+ uint16_t s : 1;
+ uint16_t pad;
+ } bits;
+};
+
+/* exact add, assumes exponent_x >= exponent_y */
+static void add(long double *hi, long double *lo, long double x, long double y)
+{
+ long double r;
+
+ r = x + y;
+ *hi = r;
+ r -= x;
+ *lo = y - r;
+}
+
+/* exact mul, assumes no over/underflow */
+static void mul(long double *hi, long double *lo, long double x, long double y)
+{
+ static const long double c = 1.0 + 0x1p32L;
+ long double cx, xh, xl, cy, yh, yl;
+
+ cx = c*x;
+ xh = (x - cx) + cx;
+ xl = x - xh;
+ cy = c*y;
+ yh = (y - cy) + cy;
+ yl = y - yh;
+ *hi = x*y;
+ *lo = (xh*yh - *hi) + xh*yl + xl*yh + xl*yl;
+}
+
+/*
+assume (long double)(hi+lo) == hi
+return an adjusted hi so that rounding it to double (or less) precision is correct
+*/
+static long double adjust(long double hi, long double lo)
+{
+ union ld80 uhi, ulo;
+
+ if (lo == 0)
+ return hi;
+ uhi.x = hi;
+ if (uhi.bits.m & 0x3ff)
+ return hi;
+ ulo.x = lo;
+ if (uhi.bits.s == ulo.bits.s)
+ uhi.bits.m++;
+ else {
+ uhi.bits.m--;
+ /* handle underflow and take care of ld80 implicit msb */
+ if (uhi.bits.m == (uint64_t)-1/2) {
+ uhi.bits.m *= 2;
+ uhi.bits.e--;
+ }
+ }
+ return uhi.x;
+}
+
+/* adjusted add so the result is correct when rounded to double (or less) precision */
+static long double dadd(long double x, long double y)
+{
+ add(&x, &y, x, y);
+ return adjust(x, y);
+}
+
+/* adjusted mul so the result is correct when rounded to double (or less) precision */
+static long double dmul(long double x, long double y)
+{
+ mul(&x, &y, x, y);
+ return adjust(x, y);
+}
+
+static int getexp(long double x)
+{
+ union ld80 u;
+ u.x = x;
+ return u.bits.e;
+}
+
+double fma(double x, double y, double z)
+{
+ #pragma STDC FENV_ACCESS ON
+ long double hi, lo1, lo2, xy;
+ int round, ez, exy;
+
+ /* handle +-inf,nan */
+ if (!isfinite(x) || !isfinite(y))
+ return x*y + z;
+ if (!isfinite(z))
+ return z;
+ /* handle +-0 */
+ if (x == 0.0 || y == 0.0)
+ return x*y + z;
+ round = fegetround();
+ if (z == 0.0) {
+ if (round == FE_TONEAREST)
+ return dmul(x, y);
+ return x*y;
+ }
+
+ /* exact mul and add require nearest rounding */
+ /* spurious inexact exceptions may be raised */
+ fesetround(FE_TONEAREST);
+ mul(&xy, &lo1, x, y);
+ exy = getexp(xy);
+ ez = getexp(z);
+ if (ez > exy) {
+ add(&hi, &lo2, z, xy);
+ } else if (ez > exy - 12) {
+ add(&hi, &lo2, xy, z);
+ if (hi == 0) {
+ /*
+ xy + z is 0, but it should be calculated with the
+ original rounding mode so the sign is correct, if the
+ compiler does not support FENV_ACCESS ON it does not
+ know about the changed rounding mode and eliminates
+ the xy + z below without the volatile memory access
+ */
+ volatile double z_;
+ fesetround(round);
+ z_ = z;
+ return (xy + z_) + lo1;
+ }
+ } else {
+ /*
+ ez <= exy - 12
+ the 12 extra bits (1guard, 11round+sticky) are needed so with
+ lo = dadd(lo1, lo2)
+ elo <= ehi - 11, and we use the last 10 bits in adjust so
+ dadd(hi, lo)
+ gives correct result when rounded to double
+ */
+ hi = xy;
+ lo2 = z;
+ }
+ /*
+ the result is stored before return for correct precision and exceptions
+
+ one corner case is when the underflow flag should be raised because
+ the precise result is an inexact subnormal double, but the calculated
+ long double result is an exact subnormal double
+ (so rounding to double does not raise exceptions)
+
+ in nearest rounding mode dadd takes care of this: the last bit of the
+ result is adjusted so rounding sees an inexact value when it should
+
+ in non-nearest rounding mode fenv is used for the workaround
+ */
+ fesetround(round);
+ if (round == FE_TONEAREST)
+ z = dadd(hi, dadd(lo1, lo2));
+ else {
+#if defined(FE_INEXACT) && defined(FE_UNDERFLOW)
+ int e = fetestexcept(FE_INEXACT);
+ feclearexcept(FE_INEXACT);
+#endif
+ z = hi + (lo1 + lo2);
+#if defined(FE_INEXACT) && defined(FE_UNDERFLOW)
+ if (getexp(z) < 0x3fff-1022 && fetestexcept(FE_INEXACT))
+ feraiseexcept(FE_UNDERFLOW);
+ else if (e)
+ feraiseexcept(FE_INEXACT);
+#endif
+ }
+ return z;
+}
+#else