xref: /illumos-gate/usr/src/lib/libm/common/m9x/nearbyintf.c (revision 1ec68d336ba97cd53f46053ac10401d16014d075)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
24  */
25 /*
26  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
27  * Use is subject to license terms.
28  */
29 
30 #pragma weak nearbyintf = __nearbyintf
31 
32 #include "libm.h"
33 #include "fenv_synonyms.h"
34 #include <fenv.h>
35 
36 float
37 __nearbyintf(float x) {
38 	union {
39 		unsigned i;
40 		float f;
41 	} xx;
42 	unsigned hx, sx, i, frac;
43 	int rm;
44 
45 	xx.f = x;
46 	sx = xx.i & 0x80000000;
47 	hx = xx.i & ~0x80000000;
48 
49 	/* handle trivial cases */
50 	if (hx >= 0x4b000000) {	/* x is nan, inf, or already integral */
51 		if (hx > 0x7f800000)	/* x is nan */
52 			return (x * x);		/* + -> * for Cheetah */
53 		return (x);
54 	} else if (hx == 0)		/* x is zero */
55 		return (x);
56 
57 	/* get the rounding mode */
58 	rm = fegetround();
59 
60 	/* flip the sense of directed roundings if x is negative */
61 	if (sx && (rm == FE_UPWARD || rm == FE_DOWNWARD))
62 		rm = (FE_UPWARD + FE_DOWNWARD) - rm;
63 
64 	/* handle |x| < 1 */
65 	if (hx < 0x3f800000) {
66 		if (rm == FE_UPWARD || (rm == FE_TONEAREST && hx > 0x3f000000))
67 			xx.i = sx | 0x3f800000;
68 		else
69 			xx.i = sx;
70 		return (xx.f);
71 	}
72 
73 	/* round x at the integer bit */
74 	i = 1 << (0x96 - (hx >> 23));
75 	frac = hx & (i - 1);
76 	if (!frac)
77 		return (x);
78 
79 	hx &= ~(i - 1);
80 	if (rm == FE_UPWARD || (rm == FE_TONEAREST && (frac > (i >> 1) ||
81 		((frac == (i >> 1)) && (hx & i)))))
82 		xx.i = sx | (hx + i);
83 	else
84 		xx.i = sx | hx;
85 	return (xx.f);
86 }
87 
88 #if 0
89 
90 /*
91  * Alternate implementations for SPARC, x86, using fp ops.  These may
92  * be faster depending on how expensive saving and restoring the fp
93  * modes and status flags is.
94  */
95 
96 #include "libm.h"
97 #include "fma.h"
98 
99 #if defined(__sparc)
100 
101 float
102 __nearbyintf(float x) {
103 	union {
104 		unsigned i;
105 		float f;
106 	} xx, yy;
107 	float z;
108 	unsigned hx, sx, fsr, oldfsr;
109 	int rm;
110 
111 	xx.f = x;
112 	sx = xx.i & 0x80000000;
113 	hx = xx.i & ~0x80000000;
114 
115 	/* handle trivial cases */
116 	if (hx >= 0x4b000000)	/* x is nan, inf, or already integral */
117 		return (x + 0.0f);
118 	else if (hx == 0)	/* x is zero */
119 		return (x);
120 
121 	/* save the fsr */
122 	__fenv_getfsr(&oldfsr);
123 
124 	/* handle |x| < 1 */
125 	if (hx < 0x3f800000) {
126 		/* flip the sense of directed roundings if x is negative */
127 		rm = oldfsr >> 30;
128 		if (sx)
129 			rm ^= rm >> 1;
130 		if (rm == FSR_RP || (rm == FSR_RN && hx > 0x3f000000))
131 			xx.i = sx | 0x3f800000;
132 		else
133 			xx.i = sx;
134 		return (xx.f);
135 	}
136 
137 	/* clear the inexact trap */
138 	fsr = oldfsr & ~FSR_NXM;
139 	__fenv_setfsr(&fsr);
140 
141 	/* round x at the integer bit */
142 	yy.i = sx | 0x4b000000;
143 	z = (x + yy.f) - yy.f;
144 
145 	/* restore the old fsr */
146 	__fenv_setfsr(&oldfsr);
147 
148 	return (z);
149 }
150 
151 #elif defined(__x86)
152 
153 /* inline template */
154 extern long double frndint(long double);
155 
156 float
157 __nearbyintf(float x) {
158 	long double z;
159 	unsigned oldcwsw, cwsw;
160 
161 	/* save the control and status words, mask the inexact exception */
162 	__fenv_getcwsw(&oldcwsw);
163 	cwsw = oldcwsw | 0x00200000;
164 	__fenv_setcwsw(&cwsw);
165 
166 	z = frndint((long double) x);
167 
168 	/*
169 	 * restore the control and status words, preserving all but the
170 	 * inexact flag
171 	 */
172 	__fenv_getcwsw(&cwsw);
173 	oldcwsw |= (cwsw & 0x1f);
174 	__fenv_setcwsw(&oldcwsw);
175 
176 	/* note: the value of z is representable in single precision */
177 	return (z);
178 }
179 
180 #else
181 #error Unknown architecture
182 #endif
183 
184 #endif
185