2N/A/*
2N/A * CDDL HEADER START
2N/A *
2N/A * The contents of this file are subject to the terms of the
2N/A * Common Development and Distribution License, Version 1.0 only
2N/A * (the "License"). You may not use this file except in compliance
2N/A * with the License.
2N/A *
2N/A * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
2N/A * or http://www.opensolaris.org/os/licensing.
2N/A * See the License for the specific language governing permissions
2N/A * and limitations under the License.
2N/A *
2N/A * When distributing Covered Code, include this CDDL HEADER in each
2N/A * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
2N/A * If applicable, add the following below this CDDL HEADER, with the
2N/A * fields enclosed by brackets "[]" replaced with your own identifying
2N/A * information: Portions Copyright [yyyy] [name of copyright owner]
2N/A *
2N/A * CDDL HEADER END
2N/A */
2N/A/*
2N/A * Copyright 2003 Sun Microsystems, Inc. All rights reserved.
2N/A * Use is subject to license terms.
2N/A */
2N/A
2N/A#pragma ident "%Z%%M% %I% %E% SMI"
2N/A
2N/A/*
2N/A * _D_cplx_div(z, w) returns z / w with infinities handled according
2N/A * to C99.
2N/A *
2N/A * If z and w are both finite and w is nonzero, _D_cplx_div(z, w)
2N/A * delivers the complex quotient q according to the usual formula:
2N/A * let a = Re(z), b = Im(z), c = Re(w), and d = Im(w); then q = x +
2N/A * I * y where x = (a * c + b * d) / r and y = (b * c - a * d) / r
2N/A * with r = c * c + d * d. This implementation scales to avoid
2N/A * premature underflow or overflow.
2N/A *
2N/A * If z is neither NaN nor zero and w is zero, or if z is infinite
2N/A * and w is finite and nonzero, _D_cplx_div delivers an infinite
2N/A * result. If z is finite and w is infinite, _D_cplx_div delivers
2N/A * a zero result.
2N/A *
2N/A * If z and w are both zero or both infinite, or if either z or w is
2N/A * a complex NaN, _D_cplx_div delivers NaN + I * NaN. C99 doesn't
2N/A * specify these cases.
2N/A *
2N/A * This implementation can raise spurious underflow, overflow, in-
2N/A * valid operation, inexact, and division-by-zero exceptions. C99
2N/A * allows this.
2N/A *
2N/A * Warning: Do not attempt to "optimize" this code by removing multi-
2N/A * plications by zero.
2N/A */
2N/A
2N/A#if !defined(sparc) && !defined(__sparc)
2N/A#error This code is for SPARC only
2N/A#endif
2N/A
2N/Astatic union {
2N/A int i[2];
2N/A double d;
2N/A} inf = {
2N/A 0x7ff00000, 0
2N/A};
2N/A
2N/A/*
2N/A * Return +1 if x is +Inf, -1 if x is -Inf, and 0 otherwise
2N/A */
2N/Astatic int
2N/Atestinf(double x)
2N/A{
2N/A union {
2N/A int i[2];
2N/A double d;
2N/A } xx;
2N/A
2N/A xx.d = x;
2N/A return (((((xx.i[0] << 1) - 0xffe00000) | xx.i[1]) == 0)?
2N/A (1 | (xx.i[0] >> 31)) : 0);
2N/A}
2N/A
2N/Adouble _Complex
2N/A_D_cplx_div(double _Complex z, double _Complex w)
2N/A{
2N/A double _Complex v;
2N/A union {
2N/A int i[2];
2N/A double d;
2N/A } aa, bb, cc, dd, ss;
2N/A double a, b, c, d, r;
2N/A int ha, hb, hc, hd, hz, hw, hs, i, j;
2N/A
2N/A /*
2N/A * The following is equivalent to
2N/A *
2N/A * a = creal(z); b = cimag(z);
2N/A * c = creal(w); d = cimag(w);
2N/A */
2N/A a = ((double *)&z)[0];
2N/A b = ((double *)&z)[1];
2N/A c = ((double *)&w)[0];
2N/A d = ((double *)&w)[1];
2N/A
2N/A /* extract high-order words to estimate |z| and |w| */
2N/A aa.d = a;
2N/A bb.d = b;
2N/A ha = aa.i[0] & ~0x80000000;
2N/A hb = bb.i[0] & ~0x80000000;
2N/A hz = (ha > hb)? ha : hb;
2N/A
2N/A cc.d = c;
2N/A dd.d = d;
2N/A hc = cc.i[0] & ~0x80000000;
2N/A hd = dd.i[0] & ~0x80000000;
2N/A hw = (hc > hd)? hc : hd;
2N/A
2N/A /* check for special cases */
2N/A if (hw >= 0x7ff00000) { /* w is inf or nan */
2N/A r = 0.0;
2N/A i = testinf(c);
2N/A j = testinf(d);
2N/A if (i | j) { /* w is infinite */
2N/A /*
2N/A * "factor out" infinity, being careful to preserve
2N/A * signs of finite values
2N/A */
2N/A c = i? i : ((cc.i[0] < 0)? -0.0 : 0.0);
2N/A d = j? j : ((dd.i[0] < 0)? -0.0 : 0.0);
2N/A if (hz >= 0x7fe00000) {
2N/A /* scale to avoid overflow below */
2N/A c *= 0.5;
2N/A d *= 0.5;
2N/A }
2N/A }
2N/A ((double *)&v)[0] = (a * c + b * d) * r;
2N/A ((double *)&v)[1] = (b * c - a * d) * r;
2N/A return (v);
2N/A }
2N/A
2N/A if (hw < 0x00100000) {
2N/A /*
2N/A * This nonsense is needed to work around some SPARC
2N/A * implementations of nonstandard mode; if both parts
2N/A * of w are subnormal, multiply them by one to force
2N/A * them to be flushed to zero when nonstandard mode
2N/A * is enabled. Sheesh.
2N/A */
2N/A cc.d = c = c * 1.0;
2N/A dd.d = d = d * 1.0;
2N/A hc = cc.i[0] & ~0x80000000;
2N/A hd = dd.i[0] & ~0x80000000;
2N/A hw = (hc > hd)? hc : hd;
2N/A }
2N/A
2N/A if (hw == 0 && (cc.i[1] | dd.i[1]) == 0) {
2N/A /* w is zero; multiply z by 1/Re(w) - I * Im(w) */
2N/A c = 1.0 / c;
2N/A i = testinf(a);
2N/A j = testinf(b);
2N/A if (i | j) { /* z is infinite */
2N/A a = i;
2N/A b = j;
2N/A }
2N/A ((double *)&v)[0] = a * c + b * d;
2N/A ((double *)&v)[1] = b * c - a * d;
2N/A return (v);
2N/A }
2N/A
2N/A if (hz >= 0x7ff00000) { /* z is inf or nan */
2N/A r = 1.0;
2N/A i = testinf(a);
2N/A j = testinf(b);
2N/A if (i | j) { /* z is infinite */
2N/A a = i;
2N/A b = j;
2N/A r = inf.d;
2N/A }
2N/A ((double *)&v)[0] = (a * c + b * d) * r;
2N/A ((double *)&v)[1] = (b * c - a * d) * r;
2N/A return (v);
2N/A }
2N/A
2N/A /*
2N/A * Scale c and d to compute 1/|w|^2 and the real and imaginary
2N/A * parts of the quotient.
2N/A *
2N/A * Note that for any s, if we let c' = sc, d' = sd, c'' = sc',
2N/A * and d'' = sd', then
2N/A *
2N/A * (ac'' + bd'') / (c'^2 + d'^2) = (ac + bd) / (c^2 + d^2)
2N/A *
2N/A * and similarly for the imaginary part of the quotient. We want
2N/A * to choose s such that (i) r := 1/(c'^2 + d'^2) can be computed
2N/A * without overflow or harmful underflow, and (ii) (ac'' + bd'')
2N/A * and (bc'' - ad'') can be computed without spurious overflow or
2N/A * harmful underflow. To avoid unnecessary rounding, we restrict
2N/A * s to a power of two.
2N/A *
2N/A * To satisfy (i), we need to choose s such that max(|c'|,|d'|)
2N/A * is not too far from one. To satisfy (ii), we need to choose
2N/A * s such that max(|c''|,|d''|) is also not too far from one.
2N/A * There is some leeway in our choice, but to keep the logic
2N/A * from getting overly complicated, we simply attempt to roughly
2N/A * balance these constraints by choosing s so as to make r about
2N/A * the same size as max(|c''|,|d''|). This corresponds to choos-
2N/A * ing s to be a power of two near |w|^(-3/4).
2N/A *
2N/A * Regarding overflow, observe that if max(|c''|,|d''|) <= 1/2,
2N/A * then the computation of (ac'' + bd'') and (bc'' - ad'') can-
2N/A * not overflow; otherwise, the computation of either of these
2N/A * values can only incur overflow if the true result would be
2N/A * within a factor of two of the overflow threshold. In other
2N/A * words, if we bias the choice of s such that at least one of
2N/A *
2N/A * max(|c''|,|d''|) <= 1/2 or r >= 2
2N/A *
2N/A * always holds, then no undeserved overflow can occur.
2N/A *
2N/A * To cope with underflow, note that if r < 2^-53, then any
2N/A * intermediate results that underflow are insignificant; either
2N/A * they will be added to normal results, rendering the under-
2N/A * flow no worse than ordinary roundoff, or they will contribute
2N/A * to a final result that is smaller than the smallest subnormal
2N/A * number. Therefore, we need only modify the preceding logic
2N/A * when z is very small and w is not too far from one. In that
2N/A * case, we can reduce the effect of any intermediate underflow
2N/A * to no worse than ordinary roundoff error by choosing s so as
2N/A * to make max(|c''|,|d''|) large enough that at least one of
2N/A * (ac'' + bd'') or (bc'' - ad'') is normal.
2N/A */
2N/A hs = (((hw >> 2) - hw) + 0x6fd7ffff) & 0xfff00000;
2N/A if (hz < 0x07200000) { /* |z| < 2^-909 */
2N/A if (((hw - 0x32800000) | (0x47100000 - hw)) >= 0)
2N/A hs = (((0x47100000 - hw) >> 1) & 0xfff00000)
2N/A + 0x3ff00000;
2N/A }
2N/A ss.i[0] = hs;
2N/A ss.i[1] = 0;
2N/A
2N/A c *= ss.d;
2N/A d *= ss.d;
2N/A r = 1.0 / (c * c + d * d);
2N/A
2N/A c *= ss.d;
2N/A d *= ss.d;
2N/A ((double *)&v)[0] = (a * c + b * d) * r;
2N/A ((double *)&v)[1] = (b * c - a * d) * r;
2N/A return (v);
2N/A}