static void fsqrt_(FPU_REG *st0_ptr, u_char st0_tag)
{
int expon;
clear_C1();
if (st0_tag == TAG_Valid) {
u_char tag;
if (signnegative(st0_ptr)) {
arith_invalid(0); /* sqrt(negative) is invalid */
return;
}
/* make st(0) in [1.0 .. 4.0) */
expon = exponent(st0_ptr);
denormal_arg:
setexponent16(st0_ptr, (expon & 1));
/* Do the computation, the sign of the result will be positive. */
tag = wm_sqrt(st0_ptr, 0, 0, control_word, SIGN_POS);
addexponent(st0_ptr, expon >> 1);
FPU_settag0(tag);
return;
}
static void fsqrt_(FPU_REG *st0_ptr, u_char st0_tag)
{
int expon;
clear_C1();
if (st0_tag == TAG_Valid) {
u_char tag;
if (signnegative(st0_ptr)) {
arith_invalid(0);
return;
}
expon = exponent(st0_ptr);
denormal_arg:
setexponent16(st0_ptr, (expon & 1));
tag = wm_sqrt(st0_ptr, 0, 0, control_word, SIGN_POS);
addexponent(st0_ptr, expon >> 1);
FPU_settag0(tag);
return;
}
static void f2xm1(FPU_REG *st0_ptr, u_char tag)
{
FPU_REG a;
clear_C1();
if ( tag == TAG_Valid )
{
/* For an 80486 FPU, the result is undefined if the arg is >= 1.0 */
if ( exponent(st0_ptr) < 0 )
{
denormal_arg:
FPU_to_exp16(st0_ptr, &a);
/* poly_2xm1(x) requires 0 < st(0) < 1. */
poly_2xm1(getsign(st0_ptr), &a, st0_ptr);
}
set_precision_flag_up(); /* 80486 appears to always do this */
return;
}
if ( tag == TAG_Zero )
return;
if ( tag == TAG_Special )
tag = FPU_Special(st0_ptr);
switch ( tag )
{
case TW_Denormal:
if ( denormal_operand() < 0 )
return;
goto denormal_arg;
case TW_Infinity:
if ( signnegative(st0_ptr) )
{
/* -infinity gives -1 (p16-10) */
FPU_copy_to_reg0(&CONST_1, TAG_Valid);
setnegative(st0_ptr);
}
return;
default:
single_arg_error(st0_ptr, tag);
}
}
static void f2xm1(FPU_REG *st0_ptr, u_char tag)
{
FPU_REG a;
clear_C1();
if (tag == TAG_Valid) {
if (exponent(st0_ptr) < 0) {
denormal_arg:
FPU_to_exp16(st0_ptr, &a);
poly_2xm1(getsign(st0_ptr), &a, st0_ptr);
}
set_precision_flag_up();
return;
}
if (tag == TAG_Zero)
return;
if (tag == TAG_Special)
tag = FPU_Special(st0_ptr);
switch (tag) {
case TW_Denormal:
if (denormal_operand() < 0)
return;
goto denormal_arg;
case TW_Infinity:
if (signnegative(st0_ptr)) {
FPU_copy_to_reg0(&CONST_1, TAG_Valid);
setnegative(st0_ptr);
}
return;
default:
single_arg_error(st0_ptr, tag);
}
}
static int trig_arg(FPU_REG *st0_ptr, int even)
{
FPU_REG tmp;
u_char tmptag;
unsigned long long q;
int old_cw = control_word, saved_status = partial_status;
int tag, st0_tag = TAG_Valid;
if (exponent(st0_ptr) >= 63) {
partial_status |= SW_C2;
return -1;
}
control_word &= ~CW_RC;
control_word |= RC_CHOP;
setpositive(st0_ptr);
tag = FPU_u_div(st0_ptr, &CONST_PI2, &tmp, PR_64_BITS | RC_CHOP | 0x3f,
SIGN_POS);
FPU_round_to_int(&tmp, tag);
q = significand(&tmp);
if (q) {
rem_kernel(significand(st0_ptr),
&significand(&tmp),
significand(&CONST_PI2),
q, exponent(st0_ptr) - exponent(&CONST_PI2));
setexponent16(&tmp, exponent(&CONST_PI2));
st0_tag = FPU_normalize(&tmp);
FPU_copy_to_reg0(&tmp, st0_tag);
}
if ((even && !(q & 1)) || (!even && (q & 1))) {
st0_tag =
FPU_sub(REV | LOADED | TAG_Valid, (int)&CONST_PI2,
FULL_PRECISION);
#ifdef BETTER_THAN_486
if ((exponent(st0_ptr) <= exponent(&CONST_PI2extra) + 64)
|| (q > 1)) {
significand(&tmp) = q + 1;
setexponent16(&tmp, 63);
FPU_normalize(&tmp);
tmptag =
FPU_u_mul(&CONST_PI2extra, &tmp, &tmp,
FULL_PRECISION, SIGN_POS,
exponent(&CONST_PI2extra) +
exponent(&tmp));
setsign(&tmp, getsign(&CONST_PI2extra));
st0_tag = FPU_add(&tmp, tmptag, 0, FULL_PRECISION);
if (signnegative(st0_ptr)) {
setpositive(st0_ptr);
q++;
}
}
#endif
}
#ifdef BETTER_THAN_486
else {
if (((q > 0)
&& (exponent(st0_ptr) <= exponent(&CONST_PI2extra) + 64))
|| (q > 1)) {
significand(&tmp) = q;
setexponent16(&tmp, 63);
FPU_normalize(&tmp);
tmptag =
FPU_u_mul(&CONST_PI2extra, &tmp, &tmp,
FULL_PRECISION, SIGN_POS,
exponent(&CONST_PI2extra) +
exponent(&tmp));
setsign(&tmp, getsign(&CONST_PI2extra));
st0_tag = FPU_sub(LOADED | (tmptag & 0x0f), (int)&tmp,
FULL_PRECISION);
if ((exponent(st0_ptr) == exponent(&CONST_PI2)) &&
((st0_ptr->sigh > CONST_PI2.sigh)
|| ((st0_ptr->sigh == CONST_PI2.sigh)
&& (st0_ptr->sigl > CONST_PI2.sigl)))) {
st0_tag =
FPU_sub(REV | LOADED | TAG_Valid,
(int)&CONST_PI2, FULL_PRECISION);
q++;
}
}
}
#endif
FPU_settag0(st0_tag);
control_word = old_cw;
partial_status = saved_status & ~SW_C2;
return (q & 3) | even;
}
/*--- poly_tan() ------------------------------------------------------------+
| |
+---------------------------------------------------------------------------*/
void poly_tan(FPU_REG *st0_ptr)
{
long int exponent;
int invert;
Xsig argSq, argSqSq, accumulatoro, accumulatore, accum,
argSignif, fix_up;
unsigned long adj;
exponent = exponent(st0_ptr);
#ifdef PARANOID
if ( signnegative(st0_ptr) ) /* Can't hack a number < 0.0 */
{ arith_invalid(0); return; } /* Need a positive number */
#endif PARANOID
/* Split the problem into two domains, smaller and larger than pi/4 */
if ( (exponent == 0) || ((exponent == -1) && (st0_ptr->sigh > 0xc90fdaa2)) )
{
/* The argument is greater than (approx) pi/4 */
invert = 1;
accum.lsw = 0;
XSIG_LL(accum) = significand(st0_ptr);
if ( exponent == 0 )
{
/* The argument is >= 1.0 */
/* Put the binary point at the left. */
XSIG_LL(accum) <<= 1;
}
/* pi/2 in hex is: 1.921fb54442d18469 898CC51701B839A2 52049C1 */
XSIG_LL(accum) = 0x921fb54442d18469LL - XSIG_LL(accum);
/* This is a special case which arises due to rounding. */
if ( XSIG_LL(accum) == 0xffffffffffffffffLL )
{
FPU_settag0(TAG_Valid);
significand(st0_ptr) = 0x8a51e04daabda360LL;
setexponent16(st0_ptr, (0x41 + EXTENDED_Ebias) | SIGN_Negative);
return;
}
argSignif.lsw = accum.lsw;
XSIG_LL(argSignif) = XSIG_LL(accum);
exponent = -1 + norm_Xsig(&argSignif);
}
else
{
invert = 0;
argSignif.lsw = 0;
XSIG_LL(accum) = XSIG_LL(argSignif) = significand(st0_ptr);
if ( exponent < -1 )
{
/* shift the argument right by the required places */
if ( FPU_shrx(&XSIG_LL(accum), -1-exponent) >= 0x80000000U )
XSIG_LL(accum) ++; /* round up */
}
}
XSIG_LL(argSq) = XSIG_LL(accum); argSq.lsw = accum.lsw;
mul_Xsig_Xsig(&argSq, &argSq);
XSIG_LL(argSqSq) = XSIG_LL(argSq); argSqSq.lsw = argSq.lsw;
mul_Xsig_Xsig(&argSqSq, &argSqSq);
/* Compute the negative terms for the numerator polynomial */
accumulatoro.msw = accumulatoro.midw = accumulatoro.lsw = 0;
polynomial_Xsig(&accumulatoro, &XSIG_LL(argSqSq), oddnegterm, HiPOWERon-1);
mul_Xsig_Xsig(&accumulatoro, &argSq);
negate_Xsig(&accumulatoro);
/* Add the positive terms */
polynomial_Xsig(&accumulatoro, &XSIG_LL(argSqSq), oddplterm, HiPOWERop-1);
/* Compute the positive terms for the denominator polynomial */
accumulatore.msw = accumulatore.midw = accumulatore.lsw = 0;
polynomial_Xsig(&accumulatore, &XSIG_LL(argSqSq), evenplterm, HiPOWERep-1);
mul_Xsig_Xsig(&accumulatore, &argSq);
negate_Xsig(&accumulatore);
/* Add the negative terms */
polynomial_Xsig(&accumulatore, &XSIG_LL(argSqSq), evennegterm, HiPOWERen-1);
/* Multiply by arg^2 */
mul64_Xsig(&accumulatore, &XSIG_LL(argSignif));
mul64_Xsig(&accumulatore, &XSIG_LL(argSignif));
/* de-normalize and divide by 2 */
shr_Xsig(&accumulatore, -2*(1+exponent) + 1);
negate_Xsig(&accumulatore); /* This does 1 - accumulator */
/* Now find the ratio. */
if ( accumulatore.msw == 0 )
{
/* accumulatoro must contain 1.0 here, (actually, 0) but it
really doesn't matter what value we use because it will
have negligible effect in later calculations
*/
XSIG_LL(accum) = 0x8000000000000000LL;
accum.lsw = 0;
}
else
{
div_Xsig(&accumulatoro, &accumulatore, &accum);
}
/* Multiply by 1/3 * arg^3 */
mul64_Xsig(&accum, &XSIG_LL(argSignif));
mul64_Xsig(&accum, &XSIG_LL(argSignif));
mul64_Xsig(&accum, &XSIG_LL(argSignif));
mul64_Xsig(&accum, &twothirds);
shr_Xsig(&accum, -2*(exponent+1));
/* tan(arg) = arg + accum */
add_two_Xsig(&accum, &argSignif, &exponent);
if ( invert )
{
/* We now have the value of tan(pi_2 - arg) where pi_2 is an
approximation for pi/2
*/
/* The next step is to fix the answer to compensate for the
error due to the approximation used for pi/2
*/
/* This is (approx) delta, the error in our approx for pi/2
(see above). It has an exponent of -65
*/
XSIG_LL(fix_up) = 0x898cc51701b839a2LL;
fix_up.lsw = 0;
if ( exponent == 0 )
adj = 0xffffffff; /* We want approx 1.0 here, but
this is close enough. */
else if ( exponent > -30 )
{
adj = accum.msw >> -(exponent+1); /* tan */
adj = mul_32_32(adj, adj); /* tan^2 */
}
else
/* Limited measurements show no results worse than 64 bit precision
except for the results for arguments close to 2^63, where the
precision of the result sometimes degrades to about 63.9 bits */
static int trig_arg(FPU_REG *st0_ptr, int even)
{
FPU_REG tmp;
u_char tmptag;
unsigned long long q;
int old_cw = control_word, saved_status = partial_status;
int tag, st0_tag = TAG_Valid;
if (exponent(st0_ptr) >= 63) {
partial_status |= SW_C2; /* Reduction incomplete. */
return -1;
}
control_word &= ~CW_RC;
control_word |= RC_CHOP;
setpositive(st0_ptr);
tag = FPU_u_div(st0_ptr, &CONST_PI2, &tmp, PR_64_BITS | RC_CHOP | 0x3f,
SIGN_POS);
FPU_round_to_int(&tmp, tag); /* Fortunately, this can't overflow
to 2^64 */
q = significand(&tmp);
if (q) {
rem_kernel(significand(st0_ptr),
&significand(&tmp),
significand(&CONST_PI2),
q, exponent(st0_ptr) - exponent(&CONST_PI2));
setexponent16(&tmp, exponent(&CONST_PI2));
st0_tag = FPU_normalize(&tmp);
FPU_copy_to_reg0(&tmp, st0_tag);
}
if ((even && !(q & 1)) || (!even && (q & 1))) {
st0_tag =
FPU_sub(REV | LOADED | TAG_Valid, (int)&CONST_PI2,
FULL_PRECISION);
#ifdef BETTER_THAN_486
/* So far, the results are exact but based upon a 64 bit
precision approximation to pi/2. The technique used
now is equivalent to using an approximation to pi/2 which
is accurate to about 128 bits. */
if ((exponent(st0_ptr) <= exponent(&CONST_PI2extra) + 64)
|| (q > 1)) {
/* This code gives the effect of having pi/2 to better than
128 bits precision. */
significand(&tmp) = q + 1;
setexponent16(&tmp, 63);
FPU_normalize(&tmp);
tmptag =
FPU_u_mul(&CONST_PI2extra, &tmp, &tmp,
FULL_PRECISION, SIGN_POS,
exponent(&CONST_PI2extra) +
exponent(&tmp));
setsign(&tmp, getsign(&CONST_PI2extra));
st0_tag = FPU_add(&tmp, tmptag, 0, FULL_PRECISION);
if (signnegative(st0_ptr)) {
/* CONST_PI2extra is negative, so the result of the addition
can be negative. This means that the argument is actually
in a different quadrant. The correction is always < pi/2,
so it can't overflow into yet another quadrant. */
setpositive(st0_ptr);
q++;
}
}
#endif /* BETTER_THAN_486 */
}
#ifdef BETTER_THAN_486
else {
/* So far, the results are exact but based upon a 64 bit
precision approximation to pi/2. The technique used
now is equivalent to using an approximation to pi/2 which
is accurate to about 128 bits. */
if (((q > 0)
&& (exponent(st0_ptr) <= exponent(&CONST_PI2extra) + 64))
|| (q > 1)) {
/* This code gives the effect of having p/2 to better than
128 bits precision. */
significand(&tmp) = q;
setexponent16(&tmp, 63);
FPU_normalize(&tmp); /* This must return TAG_Valid */
tmptag =
FPU_u_mul(&CONST_PI2extra, &tmp, &tmp,
FULL_PRECISION, SIGN_POS,
exponent(&CONST_PI2extra) +
exponent(&tmp));
setsign(&tmp, getsign(&CONST_PI2extra));
st0_tag = FPU_sub(LOADED | (tmptag & 0x0f), (int)&tmp,
FULL_PRECISION);
if ((exponent(st0_ptr) == exponent(&CONST_PI2)) &&
((st0_ptr->sigh > CONST_PI2.sigh)
|| ((st0_ptr->sigh == CONST_PI2.sigh)
&& (st0_ptr->sigl > CONST_PI2.sigl)))) {
/* CONST_PI2extra is negative, so the result of the
subtraction can be larger than pi/2. This means
that the argument is actually in a different quadrant.
The correction is always < pi/2, so it can't overflow
into yet another quadrant. */
st0_tag =
FPU_sub(REV | LOADED | TAG_Valid,
(int)&CONST_PI2, FULL_PRECISION);
q++;
}
}
}
#endif /* BETTER_THAN_486 */
FPU_settag0(st0_tag);
control_word = old_cw;
partial_status = saved_status & ~SW_C2; /* Reduction complete. */
return (q & 3) | even;
}