xref: /illumos-gate/usr/src/lib/libm/common/m9x/fmaxf.c (revision 4c28a617e3922d92a58e813a5b955eb526b9c386)
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 fmaxf = __fmaxf
31 
32 /*
33  * fmax(x,y) returns the larger of x and y.  If just one of the
34  * arguments is NaN, fmax returns the other argument.  If both
35  * arguments are NaN, fmax returns NaN (ideally, one of the
36  * argument NaNs).
37  *
38  * C99 does not require that fmax(-0,+0) = fmax(+0,-0) = +0, but
39  * ideally fmax should satisfy this.
40  *
41  * C99 makes no mention of exceptions for fmax.  I suppose ideally
42  * either fmax never raises any exceptions or else it raises the
43  * invalid operation exception if and only if some argument is a
44  * signaling NaN.  In the former case, fmax should always return
45  * one of its arguments.  In the latter, fmax shouldn't return a
46  * signaling NaN, although when both arguments are signaling NaNs,
47  * this ideal is at odds with the stipulation that fmax should
48  * always return one of its arguments.
49  *
50  * Commutativity of fmax follows from the properties listed above
51  * except when both arguments are NaN.  In that case, fmax may be
52  * declared commutative by fiat because there is no portable way
53  * to tell different NaNs apart.  Ideally fmax would be truly com-
54  * mutative for all arguments.
55  *
56  * On SPARC V8, fmax must involve tests and branches.  Ideally,
57  * an implementation on SPARC V9 should avoid branching, using
58  * conditional moves instead where necessary, and be as efficient
59  * as possible in its use of other resources.
60  *
61  * It appears to be impossible to attain all of the aforementioned
62  * ideals simultaneously.  The implementation below satisfies the
63  * following (on SPARC):
64  *
65  * 1. fmax(x,y) returns the larger of x and y if neither x nor y
66  *    is NaN and the non-NaN argument if just one of x or y is NaN.
67  *    If both x and y are NaN, fmax(x,y) returns x unchanged.
68  * 2. fmax(-0,+0) = fmax(+0,-0) = +0.
69  * 3. If either argument is a signaling NaN, fmax raises the invalid
70  *    operation exception.  Otherwise, it raises no exceptions.
71  */
72 
73 #include "libm.h"	/* for isgreaterequal macro */
74 
75 float
76 __fmaxf(float x, float y) {
77 	/*
78 	 * On SPARC v8plus/v9, this could be implemented as follows
79 	 * (assuming %f0 = x, %f1 = y, return value left in %f0):
80 	 *
81 	 * fcmps	%fcc0,%f1,%f1
82 	 * fmovsu	%fcc0,%f0,%f1
83 	 * fcmps	%fcc0,%f0,%f1
84 	 * fmovsul	%fcc0,%f1,%f0
85 	 * st		%f0,[x]
86 	 * st		%f1,[y]
87 	 * ld		[x],%l0
88 	 * ld		[y],%l1
89 	 * and		%l0,%l1,%l2
90 	 * sethi	%hi(0x80000000),%l3
91 	 * andn		%l3,%l2,%l2
92 	 * andn		%l0,%l2,%l0
93 	 * st		%l0,[x]
94 	 * ld		[x],%f0
95 	 *
96 	 * If VIS instructions are available, use this code instead:
97 	 *
98 	 * fcmps	%fcc0,%f1,%f1
99 	 * fmovsu	%fcc0,%f0,%f1
100 	 * fcmps	%fcc0,%f0,%f1
101 	 * fmovsul	%fcc0,%f1,%f0
102 	 * fands	%f0,%f1,%f2
103 	 * fzeros	%f3
104 	 * fnegs	%f3,%f3
105 	 * fandnot2s %f3,%f2,%f2
106 	 * fandnot2s %f0,%f2,%f0
107 	 *
108 	 * If VIS 3.0 instructions are available, use this:
109 	 *
110 	 * flcmps	%fcc0,%f0,%f1
111 	 * fmovslg	%fcc0,%f1,%f0	! move if %fcc0 is 1 or 2
112 	 */
113 
114 	union {
115 		unsigned i;
116 		float f;
117 	} xx, yy;
118 	unsigned s;
119 
120 	/* if y is nan, replace it by x */
121 	if (y != y)
122 		y = x;
123 
124 	/* if x is nan, replace it by y */
125 	if (x != x)
126 		x = y;
127 
128 	/* At this point, x and y are either both numeric, or both NaN */
129 	if (!isnan(x) && !isgreaterequal(x, y))
130 		x = y;
131 
132 	/*
133 	 * clear the sign of the result if either x or y has its sign clear
134 	 */
135 	xx.f = x;
136 	yy.f = y;
137 	s = ~(xx.i & yy.i) & 0x80000000;
138 	xx.i &= ~s;
139 
140 	return (xx.f);
141 }
142