xref: /titanic_44/usr/src/lib/libm/common/m9x/nearbyintf.c (revision ddc0e0b53c661f6e439e3b7072b3ef353eadb4af)
1*25c28e83SPiotr Jasiukajtis /*
2*25c28e83SPiotr Jasiukajtis  * CDDL HEADER START
3*25c28e83SPiotr Jasiukajtis  *
4*25c28e83SPiotr Jasiukajtis  * The contents of this file are subject to the terms of the
5*25c28e83SPiotr Jasiukajtis  * Common Development and Distribution License (the "License").
6*25c28e83SPiotr Jasiukajtis  * You may not use this file except in compliance with the License.
7*25c28e83SPiotr Jasiukajtis  *
8*25c28e83SPiotr Jasiukajtis  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9*25c28e83SPiotr Jasiukajtis  * or http://www.opensolaris.org/os/licensing.
10*25c28e83SPiotr Jasiukajtis  * See the License for the specific language governing permissions
11*25c28e83SPiotr Jasiukajtis  * and limitations under the License.
12*25c28e83SPiotr Jasiukajtis  *
13*25c28e83SPiotr Jasiukajtis  * When distributing Covered Code, include this CDDL HEADER in each
14*25c28e83SPiotr Jasiukajtis  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15*25c28e83SPiotr Jasiukajtis  * If applicable, add the following below this CDDL HEADER, with the
16*25c28e83SPiotr Jasiukajtis  * fields enclosed by brackets "[]" replaced with your own identifying
17*25c28e83SPiotr Jasiukajtis  * information: Portions Copyright [yyyy] [name of copyright owner]
18*25c28e83SPiotr Jasiukajtis  *
19*25c28e83SPiotr Jasiukajtis  * CDDL HEADER END
20*25c28e83SPiotr Jasiukajtis  */
21*25c28e83SPiotr Jasiukajtis 
22*25c28e83SPiotr Jasiukajtis /*
23*25c28e83SPiotr Jasiukajtis  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
24*25c28e83SPiotr Jasiukajtis  */
25*25c28e83SPiotr Jasiukajtis /*
26*25c28e83SPiotr Jasiukajtis  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
27*25c28e83SPiotr Jasiukajtis  * Use is subject to license terms.
28*25c28e83SPiotr Jasiukajtis  */
29*25c28e83SPiotr Jasiukajtis 
30*25c28e83SPiotr Jasiukajtis #pragma weak nearbyintf = __nearbyintf
31*25c28e83SPiotr Jasiukajtis 
32*25c28e83SPiotr Jasiukajtis #include "libm.h"
33*25c28e83SPiotr Jasiukajtis #include <fenv.h>
34*25c28e83SPiotr Jasiukajtis 
35*25c28e83SPiotr Jasiukajtis float
__nearbyintf(float x)36*25c28e83SPiotr Jasiukajtis __nearbyintf(float x) {
37*25c28e83SPiotr Jasiukajtis 	union {
38*25c28e83SPiotr Jasiukajtis 		unsigned i;
39*25c28e83SPiotr Jasiukajtis 		float f;
40*25c28e83SPiotr Jasiukajtis 	} xx;
41*25c28e83SPiotr Jasiukajtis 	unsigned hx, sx, i, frac;
42*25c28e83SPiotr Jasiukajtis 	int rm;
43*25c28e83SPiotr Jasiukajtis 
44*25c28e83SPiotr Jasiukajtis 	xx.f = x;
45*25c28e83SPiotr Jasiukajtis 	sx = xx.i & 0x80000000;
46*25c28e83SPiotr Jasiukajtis 	hx = xx.i & ~0x80000000;
47*25c28e83SPiotr Jasiukajtis 
48*25c28e83SPiotr Jasiukajtis 	/* handle trivial cases */
49*25c28e83SPiotr Jasiukajtis 	if (hx >= 0x4b000000) {	/* x is nan, inf, or already integral */
50*25c28e83SPiotr Jasiukajtis 		if (hx > 0x7f800000)	/* x is nan */
51*25c28e83SPiotr Jasiukajtis 			return (x * x);		/* + -> * for Cheetah */
52*25c28e83SPiotr Jasiukajtis 		return (x);
53*25c28e83SPiotr Jasiukajtis 	} else if (hx == 0)		/* x is zero */
54*25c28e83SPiotr Jasiukajtis 		return (x);
55*25c28e83SPiotr Jasiukajtis 
56*25c28e83SPiotr Jasiukajtis 	/* get the rounding mode */
57*25c28e83SPiotr Jasiukajtis 	rm = fegetround();
58*25c28e83SPiotr Jasiukajtis 
59*25c28e83SPiotr Jasiukajtis 	/* flip the sense of directed roundings if x is negative */
60*25c28e83SPiotr Jasiukajtis 	if (sx && (rm == FE_UPWARD || rm == FE_DOWNWARD))
61*25c28e83SPiotr Jasiukajtis 		rm = (FE_UPWARD + FE_DOWNWARD) - rm;
62*25c28e83SPiotr Jasiukajtis 
63*25c28e83SPiotr Jasiukajtis 	/* handle |x| < 1 */
64*25c28e83SPiotr Jasiukajtis 	if (hx < 0x3f800000) {
65*25c28e83SPiotr Jasiukajtis 		if (rm == FE_UPWARD || (rm == FE_TONEAREST && hx > 0x3f000000))
66*25c28e83SPiotr Jasiukajtis 			xx.i = sx | 0x3f800000;
67*25c28e83SPiotr Jasiukajtis 		else
68*25c28e83SPiotr Jasiukajtis 			xx.i = sx;
69*25c28e83SPiotr Jasiukajtis 		return (xx.f);
70*25c28e83SPiotr Jasiukajtis 	}
71*25c28e83SPiotr Jasiukajtis 
72*25c28e83SPiotr Jasiukajtis 	/* round x at the integer bit */
73*25c28e83SPiotr Jasiukajtis 	i = 1 << (0x96 - (hx >> 23));
74*25c28e83SPiotr Jasiukajtis 	frac = hx & (i - 1);
75*25c28e83SPiotr Jasiukajtis 	if (!frac)
76*25c28e83SPiotr Jasiukajtis 		return (x);
77*25c28e83SPiotr Jasiukajtis 
78*25c28e83SPiotr Jasiukajtis 	hx &= ~(i - 1);
79*25c28e83SPiotr Jasiukajtis 	if (rm == FE_UPWARD || (rm == FE_TONEAREST && (frac > (i >> 1) ||
80*25c28e83SPiotr Jasiukajtis 		((frac == (i >> 1)) && (hx & i)))))
81*25c28e83SPiotr Jasiukajtis 		xx.i = sx | (hx + i);
82*25c28e83SPiotr Jasiukajtis 	else
83*25c28e83SPiotr Jasiukajtis 		xx.i = sx | hx;
84*25c28e83SPiotr Jasiukajtis 	return (xx.f);
85*25c28e83SPiotr Jasiukajtis }
86*25c28e83SPiotr Jasiukajtis 
87*25c28e83SPiotr Jasiukajtis #if 0
88*25c28e83SPiotr Jasiukajtis 
89*25c28e83SPiotr Jasiukajtis /*
90*25c28e83SPiotr Jasiukajtis  * Alternate implementations for SPARC, x86, using fp ops.  These may
91*25c28e83SPiotr Jasiukajtis  * be faster depending on how expensive saving and restoring the fp
92*25c28e83SPiotr Jasiukajtis  * modes and status flags is.
93*25c28e83SPiotr Jasiukajtis  */
94*25c28e83SPiotr Jasiukajtis 
95*25c28e83SPiotr Jasiukajtis #include "libm.h"
96*25c28e83SPiotr Jasiukajtis #include "fma.h"
97*25c28e83SPiotr Jasiukajtis 
98*25c28e83SPiotr Jasiukajtis #if defined(__sparc)
99*25c28e83SPiotr Jasiukajtis 
100*25c28e83SPiotr Jasiukajtis float
101*25c28e83SPiotr Jasiukajtis __nearbyintf(float x) {
102*25c28e83SPiotr Jasiukajtis 	union {
103*25c28e83SPiotr Jasiukajtis 		unsigned i;
104*25c28e83SPiotr Jasiukajtis 		float f;
105*25c28e83SPiotr Jasiukajtis 	} xx, yy;
106*25c28e83SPiotr Jasiukajtis 	float z;
107*25c28e83SPiotr Jasiukajtis 	unsigned hx, sx, fsr, oldfsr;
108*25c28e83SPiotr Jasiukajtis 	int rm;
109*25c28e83SPiotr Jasiukajtis 
110*25c28e83SPiotr Jasiukajtis 	xx.f = x;
111*25c28e83SPiotr Jasiukajtis 	sx = xx.i & 0x80000000;
112*25c28e83SPiotr Jasiukajtis 	hx = xx.i & ~0x80000000;
113*25c28e83SPiotr Jasiukajtis 
114*25c28e83SPiotr Jasiukajtis 	/* handle trivial cases */
115*25c28e83SPiotr Jasiukajtis 	if (hx >= 0x4b000000)	/* x is nan, inf, or already integral */
116*25c28e83SPiotr Jasiukajtis 		return (x + 0.0f);
117*25c28e83SPiotr Jasiukajtis 	else if (hx == 0)	/* x is zero */
118*25c28e83SPiotr Jasiukajtis 		return (x);
119*25c28e83SPiotr Jasiukajtis 
120*25c28e83SPiotr Jasiukajtis 	/* save the fsr */
121*25c28e83SPiotr Jasiukajtis 	__fenv_getfsr(&oldfsr);
122*25c28e83SPiotr Jasiukajtis 
123*25c28e83SPiotr Jasiukajtis 	/* handle |x| < 1 */
124*25c28e83SPiotr Jasiukajtis 	if (hx < 0x3f800000) {
125*25c28e83SPiotr Jasiukajtis 		/* flip the sense of directed roundings if x is negative */
126*25c28e83SPiotr Jasiukajtis 		rm = oldfsr >> 30;
127*25c28e83SPiotr Jasiukajtis 		if (sx)
128*25c28e83SPiotr Jasiukajtis 			rm ^= rm >> 1;
129*25c28e83SPiotr Jasiukajtis 		if (rm == FSR_RP || (rm == FSR_RN && hx > 0x3f000000))
130*25c28e83SPiotr Jasiukajtis 			xx.i = sx | 0x3f800000;
131*25c28e83SPiotr Jasiukajtis 		else
132*25c28e83SPiotr Jasiukajtis 			xx.i = sx;
133*25c28e83SPiotr Jasiukajtis 		return (xx.f);
134*25c28e83SPiotr Jasiukajtis 	}
135*25c28e83SPiotr Jasiukajtis 
136*25c28e83SPiotr Jasiukajtis 	/* clear the inexact trap */
137*25c28e83SPiotr Jasiukajtis 	fsr = oldfsr & ~FSR_NXM;
138*25c28e83SPiotr Jasiukajtis 	__fenv_setfsr(&fsr);
139*25c28e83SPiotr Jasiukajtis 
140*25c28e83SPiotr Jasiukajtis 	/* round x at the integer bit */
141*25c28e83SPiotr Jasiukajtis 	yy.i = sx | 0x4b000000;
142*25c28e83SPiotr Jasiukajtis 	z = (x + yy.f) - yy.f;
143*25c28e83SPiotr Jasiukajtis 
144*25c28e83SPiotr Jasiukajtis 	/* restore the old fsr */
145*25c28e83SPiotr Jasiukajtis 	__fenv_setfsr(&oldfsr);
146*25c28e83SPiotr Jasiukajtis 
147*25c28e83SPiotr Jasiukajtis 	return (z);
148*25c28e83SPiotr Jasiukajtis }
149*25c28e83SPiotr Jasiukajtis 
150*25c28e83SPiotr Jasiukajtis #elif defined(__x86)
151*25c28e83SPiotr Jasiukajtis 
152*25c28e83SPiotr Jasiukajtis /* inline template */
153*25c28e83SPiotr Jasiukajtis extern long double frndint(long double);
154*25c28e83SPiotr Jasiukajtis 
155*25c28e83SPiotr Jasiukajtis float
156*25c28e83SPiotr Jasiukajtis __nearbyintf(float x) {
157*25c28e83SPiotr Jasiukajtis 	long double z;
158*25c28e83SPiotr Jasiukajtis 	unsigned oldcwsw, cwsw;
159*25c28e83SPiotr Jasiukajtis 
160*25c28e83SPiotr Jasiukajtis 	/* save the control and status words, mask the inexact exception */
161*25c28e83SPiotr Jasiukajtis 	__fenv_getcwsw(&oldcwsw);
162*25c28e83SPiotr Jasiukajtis 	cwsw = oldcwsw | 0x00200000;
163*25c28e83SPiotr Jasiukajtis 	__fenv_setcwsw(&cwsw);
164*25c28e83SPiotr Jasiukajtis 
165*25c28e83SPiotr Jasiukajtis 	z = frndint((long double) x);
166*25c28e83SPiotr Jasiukajtis 
167*25c28e83SPiotr Jasiukajtis 	/*
168*25c28e83SPiotr Jasiukajtis 	 * restore the control and status words, preserving all but the
169*25c28e83SPiotr Jasiukajtis 	 * inexact flag
170*25c28e83SPiotr Jasiukajtis 	 */
171*25c28e83SPiotr Jasiukajtis 	__fenv_getcwsw(&cwsw);
172*25c28e83SPiotr Jasiukajtis 	oldcwsw |= (cwsw & 0x1f);
173*25c28e83SPiotr Jasiukajtis 	__fenv_setcwsw(&oldcwsw);
174*25c28e83SPiotr Jasiukajtis 
175*25c28e83SPiotr Jasiukajtis 	/* note: the value of z is representable in single precision */
176*25c28e83SPiotr Jasiukajtis 	return (z);
177*25c28e83SPiotr Jasiukajtis }
178*25c28e83SPiotr Jasiukajtis 
179*25c28e83SPiotr Jasiukajtis #else
180*25c28e83SPiotr Jasiukajtis #error Unknown architecture
181*25c28e83SPiotr Jasiukajtis #endif
182*25c28e83SPiotr Jasiukajtis 
183*25c28e83SPiotr Jasiukajtis #endif
184