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 fmaxf = __fmaxf 31*25c28e83SPiotr Jasiukajtis 32*25c28e83SPiotr Jasiukajtis /* 33*25c28e83SPiotr Jasiukajtis * fmax(x,y) returns the larger of x and y. If just one of the 34*25c28e83SPiotr Jasiukajtis * arguments is NaN, fmax returns the other argument. If both 35*25c28e83SPiotr Jasiukajtis * arguments are NaN, fmax returns NaN (ideally, one of the 36*25c28e83SPiotr Jasiukajtis * argument NaNs). 37*25c28e83SPiotr Jasiukajtis * 38*25c28e83SPiotr Jasiukajtis * C99 does not require that fmax(-0,+0) = fmax(+0,-0) = +0, but 39*25c28e83SPiotr Jasiukajtis * ideally fmax should satisfy this. 40*25c28e83SPiotr Jasiukajtis * 41*25c28e83SPiotr Jasiukajtis * C99 makes no mention of exceptions for fmax. I suppose ideally 42*25c28e83SPiotr Jasiukajtis * either fmax never raises any exceptions or else it raises the 43*25c28e83SPiotr Jasiukajtis * invalid operation exception if and only if some argument is a 44*25c28e83SPiotr Jasiukajtis * signaling NaN. In the former case, fmax should always return 45*25c28e83SPiotr Jasiukajtis * one of its arguments. In the latter, fmax shouldn't return a 46*25c28e83SPiotr Jasiukajtis * signaling NaN, although when both arguments are signaling NaNs, 47*25c28e83SPiotr Jasiukajtis * this ideal is at odds with the stipulation that fmax should 48*25c28e83SPiotr Jasiukajtis * always return one of its arguments. 49*25c28e83SPiotr Jasiukajtis * 50*25c28e83SPiotr Jasiukajtis * Commutativity of fmax follows from the properties listed above 51*25c28e83SPiotr Jasiukajtis * except when both arguments are NaN. In that case, fmax may be 52*25c28e83SPiotr Jasiukajtis * declared commutative by fiat because there is no portable way 53*25c28e83SPiotr Jasiukajtis * to tell different NaNs apart. Ideally fmax would be truly com- 54*25c28e83SPiotr Jasiukajtis * mutative for all arguments. 55*25c28e83SPiotr Jasiukajtis * 56*25c28e83SPiotr Jasiukajtis * On SPARC V8, fmax must involve tests and branches. Ideally, 57*25c28e83SPiotr Jasiukajtis * an implementation on SPARC V9 should avoid branching, using 58*25c28e83SPiotr Jasiukajtis * conditional moves instead where necessary, and be as efficient 59*25c28e83SPiotr Jasiukajtis * as possible in its use of other resources. 60*25c28e83SPiotr Jasiukajtis * 61*25c28e83SPiotr Jasiukajtis * It appears to be impossible to attain all of the aforementioned 62*25c28e83SPiotr Jasiukajtis * ideals simultaneously. The implementation below satisfies the 63*25c28e83SPiotr Jasiukajtis * following (on SPARC): 64*25c28e83SPiotr Jasiukajtis * 65*25c28e83SPiotr Jasiukajtis * 1. fmax(x,y) returns the larger of x and y if neither x nor y 66*25c28e83SPiotr Jasiukajtis * is NaN and the non-NaN argument if just one of x or y is NaN. 67*25c28e83SPiotr Jasiukajtis * If both x and y are NaN, fmax(x,y) returns x unchanged. 68*25c28e83SPiotr Jasiukajtis * 2. fmax(-0,+0) = fmax(+0,-0) = +0. 69*25c28e83SPiotr Jasiukajtis * 3. If either argument is a signaling NaN, fmax raises the invalid 70*25c28e83SPiotr Jasiukajtis * operation exception. Otherwise, it raises no exceptions. 71*25c28e83SPiotr Jasiukajtis */ 72*25c28e83SPiotr Jasiukajtis 73*25c28e83SPiotr Jasiukajtis #include "libm.h" /* for isgreaterequal macro */ 74*25c28e83SPiotr Jasiukajtis 75*25c28e83SPiotr Jasiukajtis float 76*25c28e83SPiotr Jasiukajtis __fmaxf(float x, float y) { 77*25c28e83SPiotr Jasiukajtis /* 78*25c28e83SPiotr Jasiukajtis * On SPARC v8plus/v9, this could be implemented as follows 79*25c28e83SPiotr Jasiukajtis * (assuming %f0 = x, %f1 = y, return value left in %f0): 80*25c28e83SPiotr Jasiukajtis * 81*25c28e83SPiotr Jasiukajtis * fcmps %fcc0,%f1,%f1 82*25c28e83SPiotr Jasiukajtis * fmovsu %fcc0,%f0,%f1 83*25c28e83SPiotr Jasiukajtis * fcmps %fcc0,%f0,%f1 84*25c28e83SPiotr Jasiukajtis * fmovsul %fcc0,%f1,%f0 85*25c28e83SPiotr Jasiukajtis * st %f0,[x] 86*25c28e83SPiotr Jasiukajtis * st %f1,[y] 87*25c28e83SPiotr Jasiukajtis * ld [x],%l0 88*25c28e83SPiotr Jasiukajtis * ld [y],%l1 89*25c28e83SPiotr Jasiukajtis * and %l0,%l1,%l2 90*25c28e83SPiotr Jasiukajtis * sethi %hi(0x80000000),%l3 91*25c28e83SPiotr Jasiukajtis * andn %l3,%l2,%l2 92*25c28e83SPiotr Jasiukajtis * andn %l0,%l2,%l0 93*25c28e83SPiotr Jasiukajtis * st %l0,[x] 94*25c28e83SPiotr Jasiukajtis * ld [x],%f0 95*25c28e83SPiotr Jasiukajtis * 96*25c28e83SPiotr Jasiukajtis * If VIS instructions are available, use this code instead: 97*25c28e83SPiotr Jasiukajtis * 98*25c28e83SPiotr Jasiukajtis * fcmps %fcc0,%f1,%f1 99*25c28e83SPiotr Jasiukajtis * fmovsu %fcc0,%f0,%f1 100*25c28e83SPiotr Jasiukajtis * fcmps %fcc0,%f0,%f1 101*25c28e83SPiotr Jasiukajtis * fmovsul %fcc0,%f1,%f0 102*25c28e83SPiotr Jasiukajtis * fands %f0,%f1,%f2 103*25c28e83SPiotr Jasiukajtis * fzeros %f3 104*25c28e83SPiotr Jasiukajtis * fnegs %f3,%f3 105*25c28e83SPiotr Jasiukajtis * fandnot2s %f3,%f2,%f2 106*25c28e83SPiotr Jasiukajtis * fandnot2s %f0,%f2,%f0 107*25c28e83SPiotr Jasiukajtis * 108*25c28e83SPiotr Jasiukajtis * If VIS 3.0 instructions are available, use this: 109*25c28e83SPiotr Jasiukajtis * 110*25c28e83SPiotr Jasiukajtis * flcmps %fcc0,%f0,%f1 111*25c28e83SPiotr Jasiukajtis * fmovslg %fcc0,%f1,%f0 ! move if %fcc0 is 1 or 2 112*25c28e83SPiotr Jasiukajtis */ 113*25c28e83SPiotr Jasiukajtis 114*25c28e83SPiotr Jasiukajtis union { 115*25c28e83SPiotr Jasiukajtis unsigned i; 116*25c28e83SPiotr Jasiukajtis float f; 117*25c28e83SPiotr Jasiukajtis } xx, yy; 118*25c28e83SPiotr Jasiukajtis unsigned s; 119*25c28e83SPiotr Jasiukajtis 120*25c28e83SPiotr Jasiukajtis /* if y is nan, replace it by x */ 121*25c28e83SPiotr Jasiukajtis if (y != y) 122*25c28e83SPiotr Jasiukajtis y = x; 123*25c28e83SPiotr Jasiukajtis 124*25c28e83SPiotr Jasiukajtis /* if x is nan, replace it by y */ 125*25c28e83SPiotr Jasiukajtis if (x != x) 126*25c28e83SPiotr Jasiukajtis x = y; 127*25c28e83SPiotr Jasiukajtis 128*25c28e83SPiotr Jasiukajtis /* At this point, x and y are either both numeric, or both NaN */ 129*25c28e83SPiotr Jasiukajtis if (!isnan(x) && !isgreaterequal(x, y)) 130*25c28e83SPiotr Jasiukajtis x = y; 131*25c28e83SPiotr Jasiukajtis 132*25c28e83SPiotr Jasiukajtis /* 133*25c28e83SPiotr Jasiukajtis * clear the sign of the result if either x or y has its sign clear 134*25c28e83SPiotr Jasiukajtis */ 135*25c28e83SPiotr Jasiukajtis xx.f = x; 136*25c28e83SPiotr Jasiukajtis yy.f = y; 137*25c28e83SPiotr Jasiukajtis s = ~(xx.i & yy.i) & 0x80000000; 138*25c28e83SPiotr Jasiukajtis xx.i &= ~s; 139*25c28e83SPiotr Jasiukajtis 140*25c28e83SPiotr Jasiukajtis return (xx.f); 141*25c28e83SPiotr Jasiukajtis } 142