xref: /illumos-gate/usr/src/lib/libm/common/m9x/llrintl.c (revision 25c28e83)
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 #if defined(ELFOBJ)
31 #pragma weak llrintl = __llrintl
32 #if defined(__sparcv9) || defined(__amd64)
33 #pragma weak lrintl = __llrintl
34 #pragma weak __lrintl = __llrintl
35 #endif
36 #endif
37 
38 #include "libm.h"
39 
40 #if defined(__sparc)
41 
42 #include "fma.h"
43 #include "fenv_inlines.h"
44 
45 long long
46 llrintl(long double x) {
47 	union {
48 		unsigned i[4];
49 		long double q;
50 	} xx;
51 	union {
52 		unsigned i[2];
53 		long long l;
54 	} zz;
55 	union {
56 		unsigned i;
57 		float f;
58 	} tt;
59 	unsigned int hx, sx, frac, fsr;
60 	int rm, j;
61 	volatile float dummy;
62 
63 	xx.q = x;
64 	sx = xx.i[0] & 0x80000000;
65 	hx = xx.i[0] & ~0x80000000;
66 
67 	/* handle trivial cases */
68 	if (hx > 0x403e0000) { /* |x| > 2^63 + ... or x is nan */
69 		/* convert an out-of-range float */
70 		tt.i = sx | 0x7f000000;
71 		return ((long long) tt.f);
72 	} else if ((hx | xx.i[1] | xx.i[2] | xx.i[3]) == 0) /* x is zero */
73 		return (0LL);
74 
75 	/* get the rounding mode */
76 	__fenv_getfsr32(&fsr);
77 	rm = fsr >> 30;
78 
79 	/* flip the sense of directed roundings if x is negative */
80 	if (sx)
81 		rm ^= rm >> 1;
82 
83 	/* handle |x| < 1 */
84 	if (hx < 0x3fff0000) {
85 		dummy = 1.0e30f; /* x is nonzero, so raise inexact */
86 		dummy += 1.0e-30f;
87 		if (rm == FSR_RP || (rm == FSR_RN && (hx >= 0x3ffe0000 &&
88 			((hx & 0xffff) | xx.i[1] | xx.i[2] | xx.i[3]))))
89 			return (sx ? -1LL : 1LL);
90 		return (0LL);
91 	}
92 
93 	/* extract the integer and fractional parts of x */
94 	j = 0x406f - (hx >> 16);
95 	xx.i[0] = 0x10000 | (xx.i[0] & 0xffff);
96 	if (j >= 96) {
97 		zz.i[0] = 0;
98 		zz.i[1] = xx.i[0] >> (j - 96);
99 		frac = ((xx.i[0] << 1) << (127 - j)) | (xx.i[1] >> (j - 96));
100 		if (((xx.i[1] << 1) << (127 - j)) | xx.i[2] | xx.i[3])
101 			frac |= 1;
102 	} else if (j >= 64) {
103 		zz.i[0] = xx.i[0] >> (j - 64);
104 		zz.i[1] = ((xx.i[0] << 1) << (95 - j)) | (xx.i[1] >> (j - 64));
105 		frac = ((xx.i[1] << 1) << (95 - j)) | (xx.i[2] >> (j - 64));
106 		if (((xx.i[2] << 1) << (95 - j)) | xx.i[3])
107 			frac |= 1;
108 	} else {
109 		zz.i[0] = ((xx.i[0] << 1) << (63 - j)) | (xx.i[1] >> (j - 32));
110 		zz.i[1] = ((xx.i[1] << 1) << (63 - j)) | (xx.i[2] >> (j - 32));
111 		frac = ((xx.i[2] << 1) << (63 - j)) | (xx.i[3] >> (j - 32));
112 		if ((xx.i[3] << 1) << (63 - j))
113 			frac |= 1;
114 	}
115 
116 	/* round */
117 	if (frac && (rm == FSR_RP || (rm == FSR_RN && (frac > 0x80000000u ||
118 		(frac == 0x80000000 && (zz.i[1] & 1)))))) {
119 		if (++zz.i[1] == 0)
120 			zz.i[0]++;
121 	}
122 
123 	/* check for result out of range (note that z is |x| at this point) */
124 	if (zz.i[0] > 0x80000000u || (zz.i[0] == 0x80000000 && (zz.i[1] ||
125 		!sx))) {
126 		tt.i = sx | 0x7f000000;
127 		return ((long long) tt.f);
128 	}
129 
130 	/* raise inexact if need be */
131 	if (frac) {
132 		dummy = 1.0e30F;
133 		dummy += 1.0e-30F;
134 	}
135 
136 	/* negate result if need be */
137 	if (sx) {
138 		zz.i[0] = ~zz.i[0];
139 		zz.i[1] = -zz.i[1];
140 		if (zz.i[1] == 0)
141 			zz.i[0]++;
142 	}
143 	return (zz.l);
144 }
145 #elif defined(__x86)
146 long long
147 llrintl(long double x) {
148 	/*
149 	 * Note: The following code works on x86 (in the default rounding
150 	 * precision mode), but one ought to just use the fistpll instruction
151 	 * instead.
152 	 */
153 	union {
154 		unsigned i[3];
155 		long double e;
156 	} xx, yy;
157 	int ex;
158 
159 	xx.e = x;
160 	ex = xx.i[2] & 0x7fff;
161 
162 	if (ex < 0x403e) { /* |x| < 2^63 */
163 		/* add and subtract a power of two to round x to an integer */
164 		yy.i[2] = (xx.i[2] & 0x8000) | 0x403e;
165 		yy.i[1] = 0x80000000;
166 		yy.i[0] = 0;
167 		x = (x + yy.e) - yy.e;
168 	}
169 
170 	/* now x is nan, inf, or integral */
171 	return ((long long) x);
172 }
173 #else
174 #error Unknown architecture
175 #endif
176