1 1.1 cgd /* 2 1.1 cgd * Copyright (c) 1992, 1993 3 1.1 cgd * The Regents of the University of California. All rights reserved. 4 1.1 cgd * 5 1.1 cgd * This software was developed by the Computer Systems Engineering group 6 1.1 cgd * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and 7 1.1 cgd * contributed to Berkeley. 8 1.1 cgd * 9 1.1 cgd * Redistribution and use in source and binary forms, with or without 10 1.1 cgd * modification, are permitted provided that the following conditions 11 1.1 cgd * are met: 12 1.1 cgd * 1. Redistributions of source code must retain the above copyright 13 1.1 cgd * notice, this list of conditions and the following disclaimer. 14 1.1 cgd * 2. Redistributions in binary form must reproduce the above copyright 15 1.1 cgd * notice, this list of conditions and the following disclaimer in the 16 1.1 cgd * documentation and/or other materials provided with the distribution. 17 1.5 agc * 3. Neither the name of the University nor the names of its contributors 18 1.1 cgd * may be used to endorse or promote products derived from this software 19 1.1 cgd * without specific prior written permission. 20 1.1 cgd * 21 1.1 cgd * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 22 1.1 cgd * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 23 1.1 cgd * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 24 1.1 cgd * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 25 1.1 cgd * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 26 1.1 cgd * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 27 1.1 cgd * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 28 1.1 cgd * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 1.1 cgd * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 30 1.1 cgd * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 1.1 cgd * SUCH DAMAGE. 32 1.1 cgd * 33 1.1 cgd * from: Header: divrem.m4,v 1.4 92/06/25 13:23:57 torek Exp 34 1.6 martin * $NetBSD: divrem.m4,v 1.6 2011/03/23 20:54:35 martin Exp $ 35 1.1 cgd */ 36 1.1 cgd 37 1.1 cgd /* 38 1.1 cgd * Division and remainder, from Appendix E of the Sparc Version 8 39 1.1 cgd * Architecture Manual, with fixes from Gordon Irlam. 40 1.1 cgd */ 41 1.1 cgd 42 1.1 cgd #if defined(LIBC_SCCS) && !defined(lint) 43 1.1 cgd .asciz "@(#)divrem.m4 8.1 (Berkeley) 6/4/93" 44 1.1 cgd #endif /* LIBC_SCCS and not lint */ 45 1.1 cgd 46 1.1 cgd /* 47 1.1 cgd * Input: dividend and divisor in %o0 and %o1 respectively. 48 1.1 cgd * 49 1.1 cgd * m4 parameters: 50 1.1 cgd * NAME name of function to generate 51 1.1 cgd * OP OP=div => %o0 / %o1; OP=rem => %o0 % %o1 52 1.1 cgd * S S=true => signed; S=false => unsigned 53 1.1 cgd * 54 1.1 cgd * Algorithm parameters: 55 1.1 cgd * N how many bits per iteration we try to get (4) 56 1.1 cgd * WORDSIZE total number of bits (32) 57 1.1 cgd * 58 1.1 cgd * Derived constants: 59 1.1 cgd * TWOSUPN 2^N, for label generation (m4 exponentiation currently broken) 60 1.1 cgd * TOPBITS number of bits in the top `decade' of a number 61 1.1 cgd * 62 1.1 cgd * Important variables: 63 1.1 cgd * Q the partial quotient under development (initially 0) 64 1.1 cgd * R the remainder so far, initially the dividend 65 1.1 cgd * ITER number of main division loop iterations required; 66 1.1 cgd * equal to ceil(log2(quotient) / N). Note that this 67 1.1 cgd * is the log base (2^N) of the quotient. 68 1.1 cgd * V the current comparand, initially divisor*2^(ITER*N-1) 69 1.1 cgd * 70 1.1 cgd * Cost: 71 1.1 cgd * Current estimate for non-large dividend is 72 1.1 cgd * ceil(log2(quotient) / N) * (10 + 7N/2) + C 73 1.1 cgd * A large dividend is one greater than 2^(31-TOPBITS) and takes a 74 1.1 cgd * different path, as the upper bits of the quotient must be developed 75 1.1 cgd * one bit at a time. 76 1.1 cgd */ 77 1.1 cgd 78 1.1 cgd define(N, `4') 79 1.1 cgd define(TWOSUPN, `16') 80 1.1 cgd define(WORDSIZE, `32') 81 1.1 cgd define(TOPBITS, eval(WORDSIZE - N*((WORDSIZE-1)/N))) 82 1.1 cgd 83 1.1 cgd define(dividend, `%o0') 84 1.1 cgd define(divisor, `%o1') 85 1.1 cgd define(Q, `%o2') 86 1.1 cgd define(R, `%o3') 87 1.1 cgd define(ITER, `%o4') 88 1.1 cgd define(V, `%o5') 89 1.1 cgd 90 1.1 cgd /* m4 reminder: ifelse(a,b,c,d) => if a is b, then c, else d */ 91 1.1 cgd define(T, `%g1') 92 1.6 martin define(SC, `%g5') 93 1.1 cgd ifelse(S, `true', `define(SIGN, `%g6')') 94 1.1 cgd 95 1.1 cgd /* 96 1.1 cgd * This is the recursive definition for developing quotient digits. 97 1.1 cgd * 98 1.1 cgd * Parameters: 99 1.1 cgd * $1 the current depth, 1 <= $1 <= N 100 1.1 cgd * $2 the current accumulation of quotient bits 101 1.1 cgd * N max depth 102 1.1 cgd * 103 1.1 cgd * We add a new bit to $2 and either recurse or insert the bits in 104 1.1 cgd * the quotient. R, Q, and V are inputs and outputs as defined above; 105 1.1 cgd * the condition codes are expected to reflect the input R, and are 106 1.1 cgd * modified to reflect the output R. 107 1.1 cgd */ 108 1.1 cgd define(DEVELOP_QUOTIENT_BITS, 109 1.1 cgd ` ! depth $1, accumulated bits $2 110 1.1 cgd bl L.$1.eval(TWOSUPN+$2) 111 1.1 cgd srl V,1,V 112 1.1 cgd ! remainder is positive 113 1.1 cgd subcc R,V,R 114 1.1 cgd ifelse($1, N, 115 1.1 cgd ` b 9f 116 1.1 cgd add Q, ($2*2+1), Q 117 1.1 cgd ', ` DEVELOP_QUOTIENT_BITS(incr($1), `eval(2*$2+1)')') 118 1.1 cgd L.$1.eval(TWOSUPN+$2): 119 1.1 cgd ! remainder is negative 120 1.1 cgd addcc R,V,R 121 1.1 cgd ifelse($1, N, 122 1.1 cgd ` b 9f 123 1.1 cgd add Q, ($2*2-1), Q 124 1.1 cgd ', ` DEVELOP_QUOTIENT_BITS(incr($1), `eval(2*$2-1)')') 125 1.1 cgd ifelse($1, 1, `9:')') 126 1.1 cgd 127 1.3 mrg #include <machine/asm.h> 128 1.1 cgd #include <machine/trap.h> 129 1.1 cgd 130 1.1 cgd FUNC(NAME) 131 1.1 cgd ifelse(S, `true', 132 1.1 cgd ` ! compute sign of result; if neither is negative, no problem 133 1.1 cgd orcc divisor, dividend, %g0 ! either negative? 134 1.1 cgd bge 2f ! no, go do the divide 135 1.2 pk ifelse(OP, `div', 136 1.2 pk `xor divisor, dividend, SIGN', 137 1.2 pk `mov dividend, SIGN') ! compute sign in any case 138 1.1 cgd tst divisor 139 1.1 cgd bge 1f 140 1.1 cgd tst dividend 141 1.1 cgd ! divisor is definitely negative; dividend might also be negative 142 1.1 cgd bge 2f ! if dividend not negative... 143 1.1 cgd neg divisor ! in any case, make divisor nonneg 144 1.1 cgd 1: ! dividend is negative, divisor is nonnegative 145 1.1 cgd neg dividend ! make dividend nonnegative 146 1.1 cgd 2: 147 1.1 cgd ') 148 1.1 cgd ! Ready to divide. Compute size of quotient; scale comparand. 149 1.1 cgd orcc divisor, %g0, V 150 1.1 cgd bnz 1f 151 1.1 cgd mov dividend, R 152 1.1 cgd 153 1.1 cgd ! Divide by zero trap. If it returns, return 0 (about as 154 1.1 cgd ! wrong as possible, but that is what SunOS does...). 155 1.1 cgd t ST_DIV0 156 1.1 cgd retl 157 1.1 cgd clr %o0 158 1.1 cgd 159 1.1 cgd 1: 160 1.1 cgd cmp R, V ! if divisor exceeds dividend, done 161 1.1 cgd blu Lgot_result ! (and algorithm fails otherwise) 162 1.1 cgd clr Q 163 1.1 cgd sethi %hi(1 << (WORDSIZE - TOPBITS - 1)), T 164 1.1 cgd cmp R, T 165 1.1 cgd blu Lnot_really_big 166 1.1 cgd clr ITER 167 1.1 cgd 168 1.1 cgd ! `Here the dividend is >= 2^(31-N) or so. We must be careful here, 169 1.1 cgd ! as our usual N-at-a-shot divide step will cause overflow and havoc. 170 1.1 cgd ! The number of bits in the result here is N*ITER+SC, where SC <= N. 171 1.1 cgd ! Compute ITER in an unorthodox manner: know we need to shift V into 172 1.1 cgd ! the top decade: so do not even bother to compare to R.' 173 1.1 cgd 1: 174 1.1 cgd cmp V, T 175 1.1 cgd bgeu 3f 176 1.1 cgd mov 1, SC 177 1.1 cgd sll V, N, V 178 1.1 cgd b 1b 179 1.1 cgd inc ITER 180 1.1 cgd 181 1.1 cgd ! Now compute SC. 182 1.1 cgd 2: addcc V, V, V 183 1.1 cgd bcc Lnot_too_big 184 1.1 cgd inc SC 185 1.1 cgd 186 1.1 cgd ! We get here if the divisor overflowed while shifting. 187 1.1 cgd ! This means that R has the high-order bit set. 188 1.1 cgd ! Restore V and subtract from R. 189 1.1 cgd sll T, TOPBITS, T ! high order bit 190 1.1 cgd srl V, 1, V ! rest of V 191 1.1 cgd add V, T, V 192 1.1 cgd b Ldo_single_div 193 1.1 cgd dec SC 194 1.1 cgd 195 1.1 cgd Lnot_too_big: 196 1.1 cgd 3: cmp V, R 197 1.1 cgd blu 2b 198 1.1 cgd nop 199 1.1 cgd be Ldo_single_div 200 1.1 cgd nop 201 1.1 cgd /* NB: these are commented out in the V8-Sparc manual as well */ 202 1.1 cgd /* (I do not understand this) */ 203 1.1 cgd ! V > R: went too far: back up 1 step 204 1.1 cgd ! srl V, 1, V 205 1.1 cgd ! dec SC 206 1.1 cgd ! do single-bit divide steps 207 1.1 cgd ! 208 1.1 cgd ! We have to be careful here. We know that R >= V, so we can do the 209 1.1 cgd ! first divide step without thinking. BUT, the others are conditional, 210 1.1 cgd ! and are only done if R >= 0. Because both R and V may have the high- 211 1.1 cgd ! order bit set in the first step, just falling into the regular 212 1.1 cgd ! division loop will mess up the first time around. 213 1.1 cgd ! So we unroll slightly... 214 1.1 cgd Ldo_single_div: 215 1.1 cgd deccc SC 216 1.1 cgd bl Lend_regular_divide 217 1.1 cgd nop 218 1.1 cgd sub R, V, R 219 1.1 cgd mov 1, Q 220 1.1 cgd b Lend_single_divloop 221 1.1 cgd nop 222 1.1 cgd Lsingle_divloop: 223 1.1 cgd sll Q, 1, Q 224 1.1 cgd bl 1f 225 1.1 cgd srl V, 1, V 226 1.1 cgd ! R >= 0 227 1.1 cgd sub R, V, R 228 1.1 cgd b 2f 229 1.1 cgd inc Q 230 1.1 cgd 1: ! R < 0 231 1.1 cgd add R, V, R 232 1.1 cgd dec Q 233 1.1 cgd 2: 234 1.1 cgd Lend_single_divloop: 235 1.1 cgd deccc SC 236 1.1 cgd bge Lsingle_divloop 237 1.1 cgd tst R 238 1.1 cgd b,a Lend_regular_divide 239 1.1 cgd 240 1.1 cgd Lnot_really_big: 241 1.1 cgd 1: 242 1.1 cgd sll V, N, V 243 1.1 cgd cmp V, R 244 1.1 cgd bleu 1b 245 1.1 cgd inccc ITER 246 1.1 cgd be Lgot_result 247 1.1 cgd dec ITER 248 1.1 cgd 249 1.1 cgd tst R ! set up for initial iteration 250 1.1 cgd Ldivloop: 251 1.1 cgd sll Q, N, Q 252 1.1 cgd DEVELOP_QUOTIENT_BITS(1, 0) 253 1.1 cgd Lend_regular_divide: 254 1.1 cgd deccc ITER 255 1.1 cgd bge Ldivloop 256 1.1 cgd tst R 257 1.1 cgd bl,a Lgot_result 258 1.1 cgd ! non-restoring fixup here (one instruction only!) 259 1.1 cgd ifelse(OP, `div', 260 1.1 cgd ` dec Q 261 1.1 cgd ', ` add R, divisor, R 262 1.1 cgd ') 263 1.1 cgd 264 1.1 cgd Lgot_result: 265 1.1 cgd ifelse(S, `true', 266 1.1 cgd ` ! check to see if answer should be < 0 267 1.1 cgd tst SIGN 268 1.1 cgd bl,a 1f 269 1.1 cgd ifelse(OP, `div', `neg Q', `neg R') 270 1.1 cgd 1:') 271 1.1 cgd retl 272 1.1 cgd ifelse(OP, `div', `mov Q, %o0', `mov R, %o0') 273