int
ar_ifdiv32 (AR_IEEE_32 *x,
const AR_IEEE_32 *a,
const AR_IEEE_32 *b,
int roundmode) {
int i, s;
int res = AR_STAT_OK;
unsigned long x_lbits, y_lbits, z_lbits, rbits, carry;
signed int x_expo;
AR_IEEE_32 y, y2, z, z2;
/* If either a or b is a NaN, it's the result. */
if (IS_IEEE32_NaN(a)) {
*x = *a;
return res | AR_STAT_UNDEFINED;
}
if (IS_IEEE32_NaN(b)) {
*x = *b;
return res | AR_STAT_UNDEFINED;
}
s = a->sign ^ b->sign;
/* Test for infinities and zeros */
if (b->expo > AR_IEEE32_MAX_EXPO)
if (a->expo > AR_IEEE32_MAX_EXPO) {
/* infinity / infinity = quiet NaN */
QNaNIEEE32 (x);
return res | AR_STAT_UNDEFINED;
} else {
/* anything / infinity = zero */
if (s)
res |= AR_STAT_NEGATIVE;
ZEROIEEE32 (*x);
x->sign = s;
return res | AR_STAT_UNDERFLOW | AR_STAT_ZERO;
}
if (a->expo > AR_IEEE32_MAX_EXPO) {
/* infinity / anything = infinity */
if (s)
res |= AR_STAT_NEGATIVE;
*x = *a;
x->sign = s;
return res | AR_STAT_OVERFLOW;
}
if (b->expo == 0 && !IS_IEEE32_NZ_COEFF(b)) {
if (a->expo == 0 && !IS_IEEE32_NZ_COEFF(a)) {
/* zero/zero = quiet NaN */
QNaNIEEE32 (x);
return res | AR_STAT_UNDEFINED;
} else {
/* anything / zero = infinity */
if (s)
res |= AR_STAT_NEGATIVE;
ZEROIEEE32 (*x);
x->sign = s;
x->expo = AR_IEEE32_MAX_EXPO + 1;
return res | AR_STAT_OVERFLOW;
}
}
/* Test for denorms (they have zero exponents) to determine the
* values of the implicit normalization bits; make them explicit.
*/
if (ar_state_register.ar_denorms_trap &&
((!a->expo && IS_IEEE32_NZ_COEFF(a)) ||
(!b->expo && IS_IEEE32_NZ_COEFF(b)))) {
/* operand is a denorm and denorms cause a trap */
x->expo = AR_IEEE32_MAX_EXPO + 1;
return res | AR_STAT_UNDEFINED;
}
y = *a;
y_lbits = !!a->expo;
ZEROIEEE32 (y2);
z = *b;
z_lbits = !!b->expo;
ZEROIEEE32 (z2);
x_lbits = 0;
x_expo = a->expo - b->expo + !a->expo - !b->expo + AR_IEEE32_EXPO_BIAS;
ZEROIEEE32 (*x);
x->sign = s;
rbits = 0;
/* Handle division by a denormalized value */
if (!b->expo) {
while (!z_lbits) {
z_lbits = z.coeff0 >> (AR_IEEE32_C0_BITS - 1);
SHLEFTIEEE32 (z);
x_expo++;
}
}
if (x_expo <= 0)
x_expo--;
/* Divide by repeated subtraction */
for (i = 0; i < AR_IEEE32_COEFF_BITS + AR_IEEE32_ROUND_BITS + 1; i++) {
x_lbits = x->coeff0 >> (AR_IEEE32_C0_BITS - 1);
SHLEFTIEEE32 (*x);
x->coeff1 |= rbits >> (AR_IEEE32_ROUND_BITS - 1);
rbits = (rbits << 1) & MASKR (AR_IEEE32_ROUND_BITS);
/* If scaled denominator <= numerator, subtract */
if (z_lbits < y_lbits ||
z_lbits == y_lbits &&
(z.coeff0 < y.coeff0 ||
z.coeff0 == y.coeff0 &&
(z.coeff1 < y.coeff1 ||
z.coeff1 == y.coeff1 &&
(z2.coeff0 < y2.coeff0 ||
z2.coeff0 == y2.coeff0 &&
z2.coeff1 <= y2.coeff1)))) {
y2.coeff1 = carry = y2.coeff1 + 1 +
(z2.coeff1 ^
MASKR (AR_IEEE32_C1_BITS));
carry >>= AR_IEEE32_C1_BITS;
y2.coeff0 = carry += y2.coeff0 +
(z2.coeff0 ^
MASKR (AR_IEEE32_C0_BITS));
carry >>= AR_IEEE32_C0_BITS;
y.coeff1 = carry += y.coeff1 +
(z.coeff1 ^
MASKR (AR_IEEE32_C1_BITS));
carry >>= AR_IEEE32_C1_BITS;
y.coeff0 = carry += y.coeff0 +
(z.coeff0 ^
MASKR (AR_IEEE32_C0_BITS));
carry >>= AR_IEEE32_C0_BITS;
y_lbits = (y_lbits + carry + (z_lbits ^ 1)) & 1;
rbits |= 1;
}
SHRIGHTIEEE32_2 (z, z2);
z.coeff0 |= z_lbits << (AR_IEEE32_C0_BITS - 1);
z_lbits = 0;
}
int
ar_ifmul32 (AR_IEEE_32 *x,
const AR_IEEE_32 *a,
const AR_IEEE_32 *b,
int roundmode)
{
int i, res = AR_STAT_OK;
unsigned long x_lbits, z_lbits, rbits, carry;
signed int x_expo;
AR_IEEE_32 x2, y, y2, z;
/* If either a or b is a NaN, it's the result. */
if (IS_IEEE32_NaN(a)) {
*x = *a;
return res | AR_STAT_UNDEFINED;
}
if (IS_IEEE32_NaN(b)) {
*x = *b;
return res | AR_STAT_UNDEFINED;
}
/* Test for infinities */
if (b->expo > AR_IEEE32_MAX_EXPO)
if (a->expo == 0 && !IS_IEEE32_NZ_COEFF(a)){
/* zero * inf = quiet NaN */
QNaNIEEE32 (x);
return res | AR_STAT_UNDEFINED;
} else {
/* anything * infinity = infinity */
if (i = a->sign ^ b->sign)
res |= AR_STAT_NEGATIVE;
*x = *b;
x->sign = i;
return res | AR_STAT_OVERFLOW;
}
if (a->expo > AR_IEEE32_MAX_EXPO)
if (b->expo == 0 && !IS_IEEE32_NZ_COEFF(b)){
/* infinity * zero = quiet NaN */
QNaNIEEE32 (x);
return res | AR_STAT_UNDEFINED;
} else {
/* infinity * anything = infinity */
if (i = a->sign ^ b->sign)
res |= AR_STAT_NEGATIVE;
*x = *a;
x->sign = i;
return res | AR_STAT_OVERFLOW;
}
/* Test for denorms (they have zero exponents) to determine the
* values of the implicit normalization bits; make them explicit.
*/
if (ar_state_register.ar_denorms_trap &&
((!a->expo && IS_IEEE32_NZ_COEFF(a)) ||
(!b->expo && IS_IEEE32_NZ_COEFF(b)))) {
/* operand is a denorm and denorms cause a trap */
x->expo = AR_IEEE32_MAX_EXPO + 1;
return res | AR_STAT_UNDEFINED;
}
ZEROIEEE32 (y);
y.coeff1 = !!a->expo;
y2 = *a;
z = *b;
z_lbits = !!b->expo;
x_expo = a->expo + b->expo + !a->expo + !b->expo - AR_IEEE32_EXPO_BIAS;
if (x_expo <= 0)
x_expo--;
i = a->sign ^ b->sign;
/* Sum the pyramid */
if (z.coeff1 & 1) {
x2 = *a;
*x = y;
}
else {
ZEROIEEE32 (*x);
ZEROIEEE32 (x2);
}
x->sign = i;
x_lbits = z_lbits & y.coeff1;
SHLEFTIEEE32_2 (y, y2);
SHRIGHTIEEE32 (z);
z.coeff0 |= z_lbits << (AR_IEEE32_C0_BITS - 1);
for (i = 1; i < AR_IEEE32_COEFF_BITS + 1; i++) {
if (z.coeff1 & 1) {
carry = 0;
ADDIEEE32 (x2, carry, x2, y2);
ADDIEEE32 (*x, carry, *x, y);
x_lbits += carry;
}
SHLEFTIEEE32_2 (y, y2);
SHRIGHTIEEE32 (z);
}
/* Extract rounding bits */
rbits = x2.coeff0 >> (AR_IEEE32_C0_BITS - AR_IEEE32_ROUND_BITS);
if (x2.coeff0 & MASKR (AR_IEEE32_C0_BITS - AR_IEEE32_ROUND_BITS) |
x2.coeff1)
rbits |= 1; /* sticky bit */
/* Normalize and round */
return ar_i32norm (x_expo, x_lbits, rbits, x, roundmode);
}
