Home | History | Annotate | Line # | Download | only in fpu
fpu_subr.c revision 1.2
      1 /*	$NetBSD: fpu_subr.c,v 1.2 2003/07/15 02:54:43 lukem Exp $ */
      2 
      3 /*
      4  * Copyright (c) 1992, 1993
      5  *	The Regents of the University of California.  All rights reserved.
      6  *
      7  * This software was developed by the Computer Systems Engineering group
      8  * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
      9  * contributed to Berkeley.
     10  *
     11  * All advertising materials mentioning features or use of this software
     12  * must display the following acknowledgement:
     13  *	This product includes software developed by the University of
     14  *	California, Lawrence Berkeley Laboratory.
     15  *
     16  * Redistribution and use in source and binary forms, with or without
     17  * modification, are permitted provided that the following conditions
     18  * are met:
     19  * 1. Redistributions of source code must retain the above copyright
     20  *    notice, this list of conditions and the following disclaimer.
     21  * 2. Redistributions in binary form must reproduce the above copyright
     22  *    notice, this list of conditions and the following disclaimer in the
     23  *    documentation and/or other materials provided with the distribution.
     24  * 3. All advertising materials mentioning features or use of this software
     25  *    must display the following acknowledgement:
     26  *	This product includes software developed by the University of
     27  *	California, Berkeley and its contributors.
     28  * 4. Neither the name of the University nor the names of its contributors
     29  *    may be used to endorse or promote products derived from this software
     30  *    without specific prior written permission.
     31  *
     32  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
     33  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     34  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     35  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
     36  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     37  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     38  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     39  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     40  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     41  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     42  * SUCH DAMAGE.
     43  *
     44  *	@(#)fpu_subr.c	8.1 (Berkeley) 6/11/93
     45  */
     46 
     47 /*
     48  * FPU subroutines.
     49  */
     50 
     51 #include <sys/cdefs.h>
     52 __KERNEL_RCSID(0, "$NetBSD: fpu_subr.c,v 1.2 2003/07/15 02:54:43 lukem Exp $");
     53 
     54 #include <sys/types.h>
     55 #if defined(DIAGNOSTIC)||defined(DEBUG)
     56 #include <sys/systm.h>
     57 #endif
     58 
     59 #include <machine/reg.h>
     60 #include <powerpc/instr.h>
     61 #include <machine/fpu.h>
     62 
     63 #include <powerpc/fpu/fpu_arith.h>
     64 #include <powerpc/fpu/fpu_emu.h>
     65 #include <powerpc/fpu/fpu_extern.h>
     66 
     67 /*
     68  * Shift the given number right rsh bits.  Any bits that `fall off' will get
     69  * shoved into the sticky field; we return the resulting sticky.  Note that
     70  * shifting NaNs is legal (this will never shift all bits out); a NaN's
     71  * sticky field is ignored anyway.
     72  */
     73 int
     74 fpu_shr(struct fpn *fp, int rsh)
     75 {
     76 	u_int m0, m1, m2, m3, s;
     77 	int lsh;
     78 
     79 #ifdef DIAGNOSTIC
     80 	if (rsh <= 0 || (fp->fp_class != FPC_NUM && !ISNAN(fp)))
     81 		panic("fpu_rightshift 1");
     82 #endif
     83 
     84 	m0 = fp->fp_mant[0];
     85 	m1 = fp->fp_mant[1];
     86 	m2 = fp->fp_mant[2];
     87 	m3 = fp->fp_mant[3];
     88 
     89 	/* If shifting all the bits out, take a shortcut. */
     90 	if (rsh >= FP_NMANT) {
     91 #ifdef DIAGNOSTIC
     92 		if ((m0 | m1 | m2 | m3) == 0)
     93 			panic("fpu_rightshift 2");
     94 #endif
     95 		fp->fp_mant[0] = 0;
     96 		fp->fp_mant[1] = 0;
     97 		fp->fp_mant[2] = 0;
     98 		fp->fp_mant[3] = 0;
     99 #ifdef notdef
    100 		if ((m0 | m1 | m2 | m3) == 0)
    101 			fp->fp_class = FPC_ZERO;
    102 		else
    103 #endif
    104 			fp->fp_sticky = 1;
    105 		return (1);
    106 	}
    107 
    108 	/* Squish out full words. */
    109 	s = fp->fp_sticky;
    110 	if (rsh >= 32 * 3) {
    111 		s |= m3 | m2 | m1;
    112 		m3 = m0, m2 = 0, m1 = 0, m0 = 0;
    113 	} else if (rsh >= 32 * 2) {
    114 		s |= m3 | m2;
    115 		m3 = m1, m2 = m0, m1 = 0, m0 = 0;
    116 	} else if (rsh >= 32) {
    117 		s |= m3;
    118 		m3 = m2, m2 = m1, m1 = m0, m0 = 0;
    119 	}
    120 
    121 	/* Handle any remaining partial word. */
    122 	if ((rsh &= 31) != 0) {
    123 		lsh = 32 - rsh;
    124 		s |= m3 << lsh;
    125 		m3 = (m3 >> rsh) | (m2 << lsh);
    126 		m2 = (m2 >> rsh) | (m1 << lsh);
    127 		m1 = (m1 >> rsh) | (m0 << lsh);
    128 		m0 >>= rsh;
    129 	}
    130 	fp->fp_mant[0] = m0;
    131 	fp->fp_mant[1] = m1;
    132 	fp->fp_mant[2] = m2;
    133 	fp->fp_mant[3] = m3;
    134 	fp->fp_sticky = s;
    135 	return (s);
    136 }
    137 
    138 /*
    139  * Force a number to be normal, i.e., make its fraction have all zero
    140  * bits before FP_1, then FP_1, then all 1 bits.  This is used for denorms
    141  * and (sometimes) for intermediate results.
    142  *
    143  * Internally, this may use a `supernormal' -- a number whose fp_mant
    144  * is greater than or equal to 2.0 -- so as a side effect you can hand it
    145  * a supernormal and it will fix it (provided fp->fp_mant[3] == 0).
    146  */
    147 void
    148 fpu_norm(struct fpn *fp)
    149 {
    150 	u_int m0, m1, m2, m3, top, sup, nrm;
    151 	int lsh, rsh, exp;
    152 
    153 	exp = fp->fp_exp;
    154 	m0 = fp->fp_mant[0];
    155 	m1 = fp->fp_mant[1];
    156 	m2 = fp->fp_mant[2];
    157 	m3 = fp->fp_mant[3];
    158 
    159 	/* Handle severe subnormals with 32-bit moves. */
    160 	if (m0 == 0) {
    161 		if (m1)
    162 			m0 = m1, m1 = m2, m2 = m3, m3 = 0, exp -= 32;
    163 		else if (m2)
    164 			m0 = m2, m1 = m3, m2 = 0, m3 = 0, exp -= 2 * 32;
    165 		else if (m3)
    166 			m0 = m3, m1 = 0, m2 = 0, m3 = 0, exp -= 3 * 32;
    167 		else {
    168 			fp->fp_class = FPC_ZERO;
    169 			return;
    170 		}
    171 	}
    172 
    173 	/* Now fix any supernormal or remaining subnormal. */
    174 	nrm = FP_1;
    175 	sup = nrm << 1;
    176 	if (m0 >= sup) {
    177 		/*
    178 		 * We have a supernormal number.  We need to shift it right.
    179 		 * We may assume m3==0.
    180 		 */
    181 		for (rsh = 1, top = m0 >> 1; top >= sup; rsh++)	/* XXX slow */
    182 			top >>= 1;
    183 		exp += rsh;
    184 		lsh = 32 - rsh;
    185 		m3 = m2 << lsh;
    186 		m2 = (m2 >> rsh) | (m1 << lsh);
    187 		m1 = (m1 >> rsh) | (m0 << lsh);
    188 		m0 = top;
    189 	} else if (m0 < nrm) {
    190 		/*
    191 		 * We have a regular denorm (a subnormal number), and need
    192 		 * to shift it left.
    193 		 */
    194 		for (lsh = 1, top = m0 << 1; top < nrm; lsh++)	/* XXX slow */
    195 			top <<= 1;
    196 		exp -= lsh;
    197 		rsh = 32 - lsh;
    198 		m0 = top | (m1 >> rsh);
    199 		m1 = (m1 << lsh) | (m2 >> rsh);
    200 		m2 = (m2 << lsh) | (m3 >> rsh);
    201 		m3 <<= lsh;
    202 	}
    203 
    204 	fp->fp_exp = exp;
    205 	fp->fp_mant[0] = m0;
    206 	fp->fp_mant[1] = m1;
    207 	fp->fp_mant[2] = m2;
    208 	fp->fp_mant[3] = m3;
    209 }
    210 
    211 /*
    212  * Concoct a `fresh' Quiet NaN per Appendix N.
    213  * As a side effect, we set NV (invalid) for the current exceptions.
    214  */
    215 struct fpn *
    216 fpu_newnan(struct fpemu *fe)
    217 {
    218 	struct fpn *fp;
    219 
    220 	fe->fe_cx |= FPSCR_VXSNAN;
    221 	fp = &fe->fe_f3;
    222 	fp->fp_class = FPC_QNAN;
    223 	fp->fp_sign = 0;
    224 	fp->fp_mant[0] = FP_1 - 1;
    225 	fp->fp_mant[1] = fp->fp_mant[2] = fp->fp_mant[3] = ~0;
    226 	DUMPFPN(FPE_REG, fp);
    227 	return (fp);
    228 }
    229