1 /* mpn_divexact(qp,np,nn,dp,dn,tp) -- Divide N = {np,nn} by D = {dp,dn} storing
2 the result in Q = {qp,nn-dn+1} expecting no remainder. Overlap allowed
3 between Q and N; all other overlap disallowed.
5 Contributed to the GNU project by Torbjorn Granlund.
7 THE FUNCTIONS IN THIS FILE ARE INTERNAL WITH MUTABLE INTERFACES. IT IS ONLY
8 SAFE TO REACH THEM THROUGH DOCUMENTED INTERFACES. IN FACT, IT IS ALMOST
9 GUARANTEED THAT THEY WILL CHANGE OR DISAPPEAR IN A FUTURE GMP RELEASE.
11 Copyright 2006, 2007, 2009 Free Software Foundation, Inc.
13 This file is part of the GNU MP Library.
15 The GNU MP Library is free software; you can redistribute it and/or modify
16 it under the terms of the GNU Lesser General Public License as published by
17 the Free Software Foundation; either version 3 of the License, or (at your
18 option) any later version.
20 The GNU MP Library is distributed in the hope that it will be useful, but
21 WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
22 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public
23 License for more details.
25 You should have received a copy of the GNU Lesser General Public License
26 along with the GNU MP Library. If not, see http://www.gnu.org/licenses/. */
35 mpn_divexact (mp_ptr qp,
36 mp_srcptr np, mp_size_t nn,
37 mp_srcptr dp, mp_size_t dn)
46 ASSERT (dp[dn-1] > 0);
59 MPN_DIVREM_OR_DIVEXACT_1 (qp, np, nn, dp[0]);
66 count_trailing_zeros (shift, dp[0]);
70 mp_size_t ss = (dn > qn) ? qn + 1 : dn;
72 tp = TMP_ALLOC_LIMBS (ss);
73 mpn_rshift (tp, dp, ss, shift);
76 /* Since we have excluded dn == 1, we have nn > qn, and we need
77 to shift one limb beyond qn. */
78 wp = TMP_ALLOC_LIMBS (qn + 1);
79 mpn_rshift (wp, np, qn + 1, shift);
83 wp = TMP_ALLOC_LIMBS (qn);
84 MPN_COPY (wp, np, qn);
90 tp = TMP_ALLOC_LIMBS (mpn_bdiv_q_itch (qn, dn));
91 mpn_bdiv_q (qp, wp, qn, dp, dn, tp);
97 /* We use the Jebelean's bidirectional exact division algorithm. This is
98 somewhat naively implemented, with equal quotient parts done by 2-adic
99 division and truncating division. Since 2-adic division is faster, it
100 should be used for a larger chunk.
102 This code is horrendously ugly, in all sorts of ways.
104 * It was hacked without much care or thought, but with a testing program.
105 * It handles scratch space frivolously, and furthermore the itch function
107 * Doesn't provide any measures to deal with mu_divappr_q's +3 error. We
108 have yet to provoke an error due to this, though.
109 * Algorithm selection leaves a lot to be desired. In particular, the choice
110 between DC and MU isn't a point, but we treat it like one.
111 * It makes the msb part 1 or 2 limbs larger than the lsb part, in spite of
112 that the latter is faster. We should at least reverse this, but perhaps
113 we should make the lsb part considerably larger. (How do we tune this?)
117 mpn_divexact_itch (mp_size_t nn, mp_size_t dn)
119 return nn + dn; /* FIXME this is not right */
123 mpn_divexact (mp_ptr qp,
124 mp_srcptr np, mp_size_t nn,
125 mp_srcptr dp, mp_size_t dn,
145 /* For small divisors, and small quotients, don't use Jebelean's algorithm. */
146 if (dn < DIVEXACT_JEB_THRESHOLD || qn < DIVEXACT_JEB_THRESHOLD)
149 MPN_COPY (tp, np, qn);
150 binvert_limb (di, dp[0]); di = -di;
152 mpn_sbpi1_bdiv_q (qp, tp, qn, dp, dn, di);
157 qn0 = ((nn - dn) >> 1) + 1; /* low quotient size */
159 /* If quotient is much larger than the divisor, the bidirectional algorithm
160 does not work as currently implemented. Fall back to plain bdiv. */
163 if (BELOW_THRESHOLD (dn, DC_BDIV_Q_THRESHOLD))
166 MPN_COPY (tp, np, qn);
167 binvert_limb (di, dp[0]); di = -di;
169 mpn_sbpi1_bdiv_q (qp, tp, qn, dp, dn, di);
171 else if (BELOW_THRESHOLD (dn, MU_BDIV_Q_THRESHOLD))
174 MPN_COPY (tp, np, qn);
175 binvert_limb (di, dp[0]); di = -di;
176 mpn_dcpi1_bdiv_q (qp, tp, qn, dp, dn, di);
180 mpn_mu_bdiv_q (qp, np, qn, dp, dn, scratch);
188 nn1 = nn0 - 1 + ((nn-dn) & 1);
190 if (LIKELY (qn0 != dn))
194 if (UNLIKELY (dp[dn - 1] == 1 && qn1 != dn))
196 /* If the leading divisor limb == 1, i.e. has just one bit, we have
197 to include an extra limb in order to get the needed overlap. */
198 /* FIXME: Now with the mu_divappr_q function, we should really need
199 more overlap. That indicates one of two things: (1) The test code
200 is not good. (2) We actually overlap too much by default. */
206 tp = TMP_ALLOC_LIMBS (nn1 + 1);
208 count_leading_zeros (cnt, dp[dn - 1]);
210 /* Normalize divisor, store into tmp area. */
213 xdp = TMP_ALLOC_LIMBS (qn1);
214 mpn_lshift (xdp, dp + dn - qn1, qn1, cnt);
218 xdp = (mp_ptr) dp + dn - qn1;
221 /* Shift dividend according to the divisor normalization. */
222 /* FIXME: We compute too much here for XX_divappr_q, but these functions'
223 interfaces want a pointer to the imaginative least significant limb, not
224 to the least significant *used* limb. Of course, we could leave nn1-qn1
225 rubbish limbs in the low part, to save some time. */
228 cy = mpn_lshift (tp, np + nn - nn1, nn1, cnt);
237 /* FIXME: This copy is not needed for mpn_mu_divappr_q, except when the
238 mpn_sub_n right before is executed. */
239 MPN_COPY (tp, np + nn - nn1, nn1);
242 invert_pi1 (dinv, xdp[qn1 - 1], xdp[qn1 - 2]);
243 if (BELOW_THRESHOLD (qn1, DC_DIVAPPR_Q_THRESHOLD))
245 qp[qn0 - 1 + nn1 - qn1] = mpn_sbpi1_divappr_q (qp + qn0 - 1, tp, nn1, xdp, qn1, dinv.inv32);
247 else if (BELOW_THRESHOLD (qn1, MU_DIVAPPR_Q_THRESHOLD))
249 qp[qn0 - 1 + nn1 - qn1] = mpn_dcpi1_divappr_q (qp + qn0 - 1, tp, nn1, xdp, qn1, &dinv);
253 /* FIXME: mpn_mu_divappr_q doesn't handle qh != 0. Work around it with a
254 conditional subtraction here. */
255 qh = mpn_cmp (tp + nn1 - qn1, xdp, qn1) >= 0;
257 mpn_sub_n (tp + nn1 - qn1, tp + nn1 - qn1, xdp, qn1);
258 mpn_mu_divappr_q (qp + qn0 - 1, tp, nn1, xdp, qn1, scratch);
259 qp[qn0 - 1 + nn1 - qn1] = qh;
263 binvert_limb (di, dp[0]); di = -di;
265 if (BELOW_THRESHOLD (qn0, DC_BDIV_Q_THRESHOLD))
267 MPN_COPY (tp, np, qn0);
268 mpn_sbpi1_bdiv_q (qp, tp, qn0, dp, qn0, di);
270 else if (BELOW_THRESHOLD (qn0, MU_BDIV_Q_THRESHOLD))
272 MPN_COPY (tp, np, qn0);
273 mpn_dcpi1_bdiv_q (qp, tp, qn0, dp, qn0, di);
277 mpn_mu_bdiv_q (qp, np, qn0, dp, qn0, scratch);
280 if (qml < qp[qn0 - 1])
281 mpn_decr_u (qp + qn0, 1);