1 1.7 joerg /* $NetBSD: modf.S,v 1.7 2013/09/12 15:36:16 joerg Exp $ */ 2 1.5 christos 3 1.1 cgd /* 4 1.1 cgd * Copyright (c) 1992, 1993 5 1.1 cgd * The Regents of the University of California. All rights reserved. 6 1.1 cgd * 7 1.1 cgd * This software was developed by the Computer Systems Engineering group 8 1.1 cgd * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and 9 1.1 cgd * contributed to Berkeley. 10 1.1 cgd * 11 1.1 cgd * Redistribution and use in source and binary forms, with or without 12 1.1 cgd * modification, are permitted provided that the following conditions 13 1.1 cgd * are met: 14 1.1 cgd * 1. Redistributions of source code must retain the above copyright 15 1.1 cgd * notice, this list of conditions and the following disclaimer. 16 1.1 cgd * 2. Redistributions in binary form must reproduce the above copyright 17 1.1 cgd * notice, this list of conditions and the following disclaimer in the 18 1.1 cgd * documentation and/or other materials provided with the distribution. 19 1.6 agc * 3. Neither the name of the University nor the names of its contributors 20 1.1 cgd * may be used to endorse or promote products derived from this software 21 1.1 cgd * without specific prior written permission. 22 1.1 cgd * 23 1.1 cgd * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 24 1.1 cgd * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 25 1.1 cgd * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 26 1.1 cgd * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 27 1.1 cgd * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 28 1.1 cgd * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 29 1.1 cgd * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 30 1.1 cgd * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 31 1.1 cgd * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 32 1.1 cgd * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 33 1.1 cgd * SUCH DAMAGE. 34 1.1 cgd * 35 1.1 cgd * from: Header: modf.s,v 1.3 92/06/20 00:00:54 torek Exp 36 1.1 cgd */ 37 1.1 cgd 38 1.5 christos #include <machine/asm.h> 39 1.1 cgd #if defined(LIBC_SCCS) && !defined(lint) 40 1.5 christos #if 0 41 1.1 cgd .asciz "@(#)modf.s 8.1 (Berkeley) 6/4/93" 42 1.5 christos #else 43 1.7 joerg RCSID("$NetBSD: modf.S,v 1.7 2013/09/12 15:36:16 joerg Exp $") 44 1.5 christos #endif 45 1.1 cgd #endif /* LIBC_SCCS and not lint */ 46 1.1 cgd 47 1.1 cgd #include <machine/fsr.h> 48 1.1 cgd 49 1.1 cgd /* 50 1.1 cgd * double modf(double val, double *iptr) 51 1.1 cgd * 52 1.1 cgd * Returns the fractional part of `val', storing the integer part of 53 1.1 cgd * `val' in *iptr. Both *iptr and the return value have the same sign 54 1.1 cgd * as `val'. 55 1.1 cgd * 56 1.1 cgd * Method: 57 1.1 cgd * 58 1.1 cgd * We use the fpu's normalization hardware to compute the integer portion 59 1.1 cgd * of the double precision argument. Sun IEEE double precision numbers 60 1.1 cgd * have 52 bits of mantissa, 11 bits of exponent, and one bit of sign, 61 1.1 cgd * with the sign occupying bit 31 of word 0, and the exponent bits 30:20 62 1.1 cgd * of word 0. Thus, values >= 2^52 are by definition integers. 63 1.1 cgd * 64 1.1 cgd * If we take a value that is in the range [+0..2^52) and add 2^52, all 65 1.1 cgd * of the fractional bits fall out and all of the integer bits are summed 66 1.1 cgd * with 2^52. If we then subtract 2^52, we get those integer bits back. 67 1.1 cgd * This must be done with rounding set to `towards 0' or `towards -inf'. 68 1.1 cgd * `Toward -inf' fails when the value is 0 (we get -0 back).... 69 1.1 cgd * 70 1.1 cgd * Note that this method will work anywhere, but is machine dependent in 71 1.1 cgd * various aspects. 72 1.1 cgd * 73 1.1 cgd * Stack usage: 74 1.1 cgd * 4@[%fp - 4] saved %fsr 75 1.1 cgd * 4@[%fp - 8] new %fsr with rounding set to `towards 0' 76 1.1 cgd * 8@[%fp - 16] space for moving between %i and %f registers 77 1.1 cgd * Register usage: 78 1.1 cgd * %i0%i1 double val; 79 1.1 cgd * %l0 scratch 80 1.1 cgd * %l1 sign bit (0x80000000) 81 1.1 cgd * %i2 double *iptr; 82 1.1 cgd * %f2:f3 `magic number' 2^52, in fpu registers 83 1.1 cgd * %f4:f5 double v, in fpu registers 84 1.1 cgd */ 85 1.1 cgd 86 1.1 cgd .align 8 87 1.1 cgd Lmagic: 88 1.1 cgd .word 0x43300000 ! sign = 0, exponent = 52 + 1023, mantissa = 0 89 1.1 cgd .word 0 ! (i.e., .double 0r4503599627370496e+00) 90 1.1 cgd 91 1.1 cgd L0: 92 1.1 cgd .word 0 ! 0.0 93 1.1 cgd .word 0 94 1.1 cgd 95 1.1 cgd ENTRY(modf) 96 1.1 cgd save %sp, -64-16, %sp 97 1.1 cgd 98 1.1 cgd /* 99 1.1 cgd * First, compute v = abs(val) by clearing sign bit, 100 1.1 cgd * and then set up the fpu registers. This would be 101 1.1 cgd * much easier if we could do alu operations on fpu registers! 102 1.1 cgd */ 103 1.2 pk sethi %hi(0x80000000), %l1 ! sign bit 104 1.1 cgd andn %i0, %l1, %l0 105 1.1 cgd st %l0, [%fp - 16] 106 1.7 joerg #ifdef __PIC__ 107 1.3 pk PICCY_SET(Lmagic, %l0, %o7) 108 1.3 pk ldd [%l0], %f2 109 1.3 pk #else 110 1.1 cgd sethi %hi(Lmagic), %l0 111 1.1 cgd ldd [%l0 + %lo(Lmagic)], %f2 112 1.3 pk #endif 113 1.1 cgd st %i1, [%fp - 12] 114 1.1 cgd ldd [%fp - 16], %f4 ! %f4:f5 = v 115 1.1 cgd 116 1.1 cgd /* 117 1.1 cgd * Is %f4:f5 >= %f2:f3 ? If so, it is all integer bits. 118 1.1 cgd * It is probably less, though. 119 1.1 cgd */ 120 1.1 cgd fcmped %f4, %f2 121 1.1 cgd nop ! fpop2 delay 122 1.1 cgd fbuge Lbig ! if >= (or unordered), go out 123 1.1 cgd nop 124 1.1 cgd 125 1.1 cgd /* 126 1.1 cgd * v < 2^52, so add 2^52, then subtract 2^52, but do it all 127 1.1 cgd * with rounding set towards zero. We leave any enabled 128 1.1 cgd * traps enabled, but change the rounding mode. This might 129 1.1 cgd * not be so good. Oh well.... 130 1.1 cgd */ 131 1.1 cgd st %fsr, [%fp - 4] ! %l5 = current FSR mode 132 1.1 cgd set FSR_RD, %l3 ! %l3 = rounding direction mask 133 1.1 cgd ld [%fp - 4], %l5 134 1.1 cgd set FSR_RD_RZ << FSR_RD_SHIFT, %l4 135 1.1 cgd andn %l5, %l3, %l6 136 1.1 cgd or %l6, %l4, %l6 ! round towards zero, please 137 1.1 cgd and %l5, %l3, %l5 ! save original rounding mode 138 1.1 cgd st %l6, [%fp - 8] 139 1.1 cgd ld [%fp - 8], %fsr 140 1.1 cgd 141 1.1 cgd faddd %f4, %f2, %f4 ! %f4:f5 += 2^52 142 1.1 cgd fsubd %f4, %f2, %f4 ! %f4:f5 -= 2^52 143 1.1 cgd 144 1.1 cgd /* 145 1.1 cgd * Restore %fsr, but leave exceptions accrued. 146 1.1 cgd */ 147 1.1 cgd st %fsr, [%fp - 4] 148 1.1 cgd ld [%fp - 4], %l6 149 1.1 cgd andn %l6, %l3, %l6 ! %l6 = %fsr & ~FSR_RD; 150 1.1 cgd or %l5, %l6, %l5 ! %l5 |= %l6; 151 1.1 cgd st %l5, [%fp - 4] 152 1.1 cgd ld [%fp - 4], %fsr ! restore %fsr, leaving accrued stuff 153 1.1 cgd 154 1.1 cgd /* 155 1.1 cgd * Now insert the original sign in %f4:f5. 156 1.1 cgd * This is a lot of work, so it is conditional here. 157 1.1 cgd */ 158 1.1 cgd btst %l1, %i0 159 1.1 cgd be 1f 160 1.1 cgd nop 161 1.1 cgd st %f4, [%fp - 16] 162 1.1 cgd ld [%fp - 16], %g1 163 1.1 cgd or %l1, %g1, %g1 164 1.1 cgd st %g1, [%fp - 16] 165 1.1 cgd ld [%fp - 16], %f4 166 1.1 cgd 1: 167 1.1 cgd 168 1.1 cgd /* 169 1.1 cgd * The value in %f4:f5 is now the integer portion of the original 170 1.1 cgd * argument. We need to store this in *ival (%i2), subtract it 171 1.1 cgd * from the original value argument (%i0:i1), and return the result. 172 1.1 cgd */ 173 1.1 cgd std %f4, [%i2] ! *ival = %f4:f5; 174 1.1 cgd std %i0, [%fp - 16] 175 1.1 cgd ldd [%fp - 16], %f0 ! %f0:f1 = val; 176 1.1 cgd fsubd %f0, %f4, %f0 ! %f0:f1 -= %f4:f5; 177 1.1 cgd ret 178 1.1 cgd restore 179 1.1 cgd 180 1.1 cgd Lbig: 181 1.1 cgd /* 182 1.1 cgd * We get here if the original comparison of %f4:f5 (v) to 183 1.1 cgd * %f2:f3 (2^52) came out `greater or unordered'. In this 184 1.1 cgd * case the integer part is the original value, and the 185 1.1 cgd * fractional part is 0. 186 1.1 cgd */ 187 1.7 joerg #ifdef __PIC__ 188 1.3 pk PICCY_SET(L0, %l0, %o7) 189 1.3 pk std %f0, [%i2] ! *ival = val; 190 1.3 pk ldd [%l0], %f0 ! return 0.0; 191 1.3 pk #else 192 1.1 cgd sethi %hi(L0), %l0 193 1.1 cgd std %f0, [%i2] ! *ival = val; 194 1.1 cgd ldd [%l0 + %lo(L0)], %f0 ! return 0.0; 195 1.3 pk #endif 196 1.1 cgd ret 197 1.1 cgd restore 198