1 /* UltraSPARC 64 mpn_divexact_1 -- mpn by limb exact division. 2 3 THE FUNCTIONS IN THIS FILE ARE FOR INTERNAL USE ONLY. THEY'RE ALMOST 4 CERTAIN TO BE SUBJECT TO INCOMPATIBLE CHANGES OR DISAPPEAR COMPLETELY IN 5 FUTURE GNU MP RELEASES. 6 7 Copyright 2000, 2001, 2003 Free Software Foundation, Inc. 8 9 This file is part of the GNU MP Library. 10 11 The GNU MP Library is free software; you can redistribute it and/or modify 12 it under the terms of either: 13 14 * the GNU Lesser General Public License as published by the Free 15 Software Foundation; either version 3 of the License, or (at your 16 option) any later version. 17 18 or 19 20 * the GNU General Public License as published by the Free Software 21 Foundation; either version 2 of the License, or (at your option) any 22 later version. 23 24 or both in parallel, as here. 25 26 The GNU MP Library is distributed in the hope that it will be useful, but 27 WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY 28 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 29 for more details. 30 31 You should have received copies of the GNU General Public License and the 32 GNU Lesser General Public License along with the GNU MP Library. If not, 33 see https://www.gnu.org/licenses/. */ 34 35 #include "gmp.h" 36 #include "gmp-impl.h" 37 #include "longlong.h" 38 39 #include "mpn/sparc64/sparc64.h" 40 41 42 /* 64-bit divisor 32-bit divisor 43 cycles/limb cycles/limb 44 (approx) (approx) 45 Ultrasparc 2i: 110 70 46 */ 47 48 49 /* There are two key ideas here to reduce mulx's. Firstly when the divisor 50 is 32-bits the high of q*d can be calculated without the two 32x32->64 51 cross-products involving the high 32-bits of the divisor, that being zero 52 of course. Secondly umul_ppmm_lowequal and umul_ppmm_half_lowequal save 53 one mulx (each) knowing the low of q*d is equal to the input limb l. 54 55 For size==1, a simple udivx is used. This is faster than calculating an 56 inverse. 57 58 For a 32-bit divisor and small sizes, an attempt was made at a simple 59 udivx loop (two per 64-bit limb), but it turned out to be slower than 60 mul-by-inverse. At size==2 the inverse is about 260 cycles total 61 compared to a udivx at 291. Perhaps the latter would suit when size==2 62 but the high 32-bits of the second limb is zero (saving one udivx), but 63 it doesn't seem worth a special case just for that. */ 64 65 void 66 mpn_divexact_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, mp_limb_t divisor) 67 { 68 mp_limb_t inverse, s, s_next, c, l, ls, q; 69 unsigned rshift, lshift; 70 mp_limb_t lshift_mask; 71 mp_limb_t divisor_h; 72 73 ASSERT (size >= 1); 74 ASSERT (divisor != 0); 75 ASSERT (MPN_SAME_OR_SEPARATE_P (dst, src, size)); 76 ASSERT_MPN (src, size); 77 ASSERT_LIMB (divisor); 78 79 s = *src++; /* src low limb */ 80 size--; 81 if (size == 0) 82 { 83 *dst = s / divisor; 84 return; 85 } 86 87 if ((divisor & 1) == 0) 88 { 89 count_trailing_zeros (rshift, divisor); 90 divisor >>= rshift; 91 } 92 else 93 rshift = 0; 94 95 binvert_limb (inverse, divisor); 96 97 lshift = 64 - rshift; 98 99 /* lshift==64 means no shift, so must mask out other part in this case */ 100 lshift_mask = (rshift == 0 ? 0 : MP_LIMB_T_MAX); 101 102 c = 0; 103 divisor_h = HIGH32 (divisor); 104 105 if (divisor_h == 0) 106 { 107 /* 32-bit divisor */ 108 do 109 { 110 s_next = *src++; 111 ls = (s >> rshift) | ((s_next << lshift) & lshift_mask); 112 s = s_next; 113 114 SUBC_LIMB (c, l, ls, c); 115 116 q = l * inverse; 117 *dst++ = q; 118 119 umul_ppmm_half_lowequal (l, q, divisor, l); 120 c += l; 121 122 size--; 123 } 124 while (size != 0); 125 126 ls = s >> rshift; 127 l = ls - c; 128 q = l * inverse; 129 *dst = q; 130 } 131 else 132 { 133 /* 64-bit divisor */ 134 mp_limb_t divisor_l = LOW32 (divisor); 135 do 136 { 137 s_next = *src++; 138 ls = (s >> rshift) | ((s_next << lshift) & lshift_mask); 139 s = s_next; 140 141 SUBC_LIMB (c, l, ls, c); 142 143 q = l * inverse; 144 *dst++ = q; 145 146 umul_ppmm_lowequal (l, q, divisor, divisor_h, divisor_l, l); 147 c += l; 148 149 size--; 150 } 151 while (size != 0); 152 153 ls = s >> rshift; 154 l = ls - c; 155 q = l * inverse; 156 *dst = q; 157 } 158 } 159