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.
4 
5    Contributed to the GNU project by Torbjorn Granlund.
6 
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.
10 
11 Copyright 2006, 2007, 2009, 2017 Free Software Foundation, Inc.
12 
13 This file is part of the GNU MP Library.
14 
15 The GNU MP Library is free software; you can redistribute it and/or modify
16 it under the terms of either:
17 
18   * the GNU Lesser General Public License as published by the Free
19     Software Foundation; either version 3 of the License, or (at your
20     option) any later version.
21 
22 or
23 
24   * the GNU General Public License as published by the Free Software
25     Foundation; either version 2 of the License, or (at your option) any
26     later version.
27 
28 or both in parallel, as here.
29 
30 The GNU MP Library is distributed in the hope that it will be useful, but
31 WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
32 or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
33 for more details.
34 
35 You should have received copies of the GNU General Public License and the
36 GNU Lesser General Public License along with the GNU MP Library.  If not,
37 see https://www.gnu.org/licenses/.  */
38 
39 
40 #include "gmp-impl.h"
41 #include "longlong.h"
42 
43 #if 1
44 void
mpn_divexact(mp_ptr qp,mp_srcptr np,mp_size_t nn,mp_srcptr dp,mp_size_t dn)45 mpn_divexact (mp_ptr qp,
46 	      mp_srcptr np, mp_size_t nn,
47 	      mp_srcptr dp, mp_size_t dn)
48 {
49   unsigned shift;
50   mp_size_t qn;
51   mp_ptr tp;
52   TMP_DECL;
53 
54   ASSERT (dn > 0);
55   ASSERT (nn >= dn);
56   ASSERT (dp[dn-1] > 0);
57 
58   while (dp[0] == 0)
59     {
60       ASSERT (np[0] == 0);
61       dp++;
62       np++;
63       dn--;
64       nn--;
65     }
66 
67   if (dn == 1)
68     {
69       MPN_DIVREM_OR_DIVEXACT_1 (qp, np, nn, dp[0]);
70       return;
71     }
72 
73   TMP_MARK;
74 
75   qn = nn + 1 - dn;
76   count_trailing_zeros (shift, dp[0]);
77 
78   if (shift > 0)
79     {
80       mp_ptr wp;
81       mp_size_t ss;
82       ss = (dn > qn) ? qn + 1 : dn;
83 
84       tp = TMP_ALLOC_LIMBS (ss);
85       mpn_rshift (tp, dp, ss, shift);
86       dp = tp;
87 
88       /* Since we have excluded dn == 1, we have nn > qn, and we need
89 	 to shift one limb beyond qn. */
90       wp = TMP_ALLOC_LIMBS (qn + 1);
91       mpn_rshift (wp, np, qn + 1, shift);
92       np = wp;
93     }
94 
95   if (dn > qn)
96     dn = qn;
97 
98   tp = TMP_ALLOC_LIMBS (mpn_bdiv_q_itch (qn, dn));
99   mpn_bdiv_q (qp, np, qn, dp, dn, tp);
100   TMP_FREE;
101 
102   /* Since bdiv_q computes -N/D (mod B^{qn}), we must negate now. */
103   mpn_neg (qp, qp, qn);
104 }
105 
106 #else
107 
108 /* We use the Jebelean's bidirectional exact division algorithm.  This is
109    somewhat naively implemented, with equal quotient parts done by 2-adic
110    division and truncating division.  Since 2-adic division is faster, it
111    should be used for a larger chunk.
112 
113    This code is horrendously ugly, in all sorts of ways.
114 
115    * It was hacked without much care or thought, but with a testing program.
116    * It handles scratch space frivolously, and furthermore the itch function
117      is broken.
118    * Doesn't provide any measures to deal with mu_divappr_q's +3 error.  We
119      have yet to provoke an error due to this, though.
120    * Algorithm selection leaves a lot to be desired.  In particular, the choice
121      between DC and MU isn't a point, but we treat it like one.
122    * It makes the msb part 1 or 2 limbs larger than the lsb part, in spite of
123      that the latter is faster.  We should at least reverse this, but perhaps
124      we should make the lsb part considerably larger.  (How do we tune this?)
125 */
126 
127 mp_size_t
mpn_divexact_itch(mp_size_t nn,mp_size_t dn)128 mpn_divexact_itch (mp_size_t nn, mp_size_t dn)
129 {
130   return nn + dn;		/* FIXME this is not right */
131 }
132 
133 void
mpn_divexact(mp_ptr qp,mp_srcptr np,mp_size_t nn,mp_srcptr dp,mp_size_t dn,mp_ptr scratch)134 mpn_divexact (mp_ptr qp,
135 	      mp_srcptr np, mp_size_t nn,
136 	      mp_srcptr dp, mp_size_t dn,
137 	      mp_ptr scratch)
138 {
139   mp_size_t qn;
140   mp_size_t nn0, qn0;
141   mp_size_t nn1, qn1;
142   mp_ptr tp;
143   mp_limb_t qml;
144   mp_limb_t qh;
145   int cnt;
146   mp_ptr xdp;
147   mp_limb_t di;
148   mp_limb_t cy;
149   gmp_pi1_t dinv;
150   TMP_DECL;
151 
152   TMP_MARK;
153 
154   qn = nn - dn + 1;
155 
156   /* For small divisors, and small quotients, don't use Jebelean's algorithm. */
157   if (dn < DIVEXACT_JEB_THRESHOLD || qn < DIVEXACT_JEB_THRESHOLD)
158     {
159       tp = scratch;
160       MPN_COPY (tp, np, qn);
161       binvert_limb (di, dp[0]);  di = -di;
162       dn = MIN (dn, qn);
163       mpn_sbpi1_bdiv_q (qp, tp, qn, dp, dn, di);
164       TMP_FREE;
165       return;
166     }
167 
168   qn0 = ((nn - dn) >> 1) + 1;	/* low quotient size */
169 
170   /* If quotient is much larger than the divisor, the bidirectional algorithm
171      does not work as currently implemented.  Fall back to plain bdiv.  */
172   if (qn0 > dn)
173     {
174       if (BELOW_THRESHOLD (dn, DC_BDIV_Q_THRESHOLD))
175 	{
176 	  tp = scratch;
177 	  MPN_COPY (tp, np, qn);
178 	  binvert_limb (di, dp[0]);  di = -di;
179 	  dn = MIN (dn, qn);
180 	  mpn_sbpi1_bdiv_q (qp, tp, qn, dp, dn, di);
181 	}
182       else if (BELOW_THRESHOLD (dn, MU_BDIV_Q_THRESHOLD))
183 	{
184 	  tp = scratch;
185 	  MPN_COPY (tp, np, qn);
186 	  binvert_limb (di, dp[0]);  di = -di;
187 	  mpn_dcpi1_bdiv_q (qp, tp, qn, dp, dn, di);
188 	}
189       else
190 	{
191 	  mpn_mu_bdiv_q (qp, np, qn, dp, dn, scratch);
192 	}
193       TMP_FREE;
194       return;
195     }
196 
197   nn0 = qn0 + qn0;
198 
199   nn1 = nn0 - 1 + ((nn-dn) & 1);
200   qn1 = qn0;
201   if (LIKELY (qn0 != dn))
202     {
203       nn1 = nn1 + 1;
204       qn1 = qn1 + 1;
205       if (UNLIKELY (dp[dn - 1] == 1 && qn1 != dn))
206 	{
207 	  /* If the leading divisor limb == 1, i.e. has just one bit, we have
208 	     to include an extra limb in order to get the needed overlap.  */
209 	  /* FIXME: Now with the mu_divappr_q function, we should really need
210 	     more overlap. That indicates one of two things: (1) The test code
211 	     is not good. (2) We actually overlap too much by default.  */
212 	  nn1 = nn1 + 1;
213 	  qn1 = qn1 + 1;
214 	}
215     }
216 
217   tp = TMP_ALLOC_LIMBS (nn1 + 1);
218 
219   count_leading_zeros (cnt, dp[dn - 1]);
220 
221   /* Normalize divisor, store into tmp area.  */
222   if (cnt != 0)
223     {
224       xdp = TMP_ALLOC_LIMBS (qn1);
225       mpn_lshift (xdp, dp + dn - qn1, qn1, cnt);
226     }
227   else
228     {
229       xdp = (mp_ptr) dp + dn - qn1;
230     }
231 
232   /* Shift dividend according to the divisor normalization.  */
233   /* FIXME: We compute too much here for XX_divappr_q, but these functions'
234      interfaces want a pointer to the imaginative least significant limb, not
235      to the least significant *used* limb.  Of course, we could leave nn1-qn1
236      rubbish limbs in the low part, to save some time.  */
237   if (cnt != 0)
238     {
239       cy = mpn_lshift (tp, np + nn - nn1, nn1, cnt);
240       if (cy != 0)
241 	{
242 	  tp[nn1] = cy;
243 	  nn1++;
244 	}
245     }
246   else
247     {
248       /* FIXME: This copy is not needed for mpn_mu_divappr_q, except when the
249 	 mpn_sub_n right before is executed.  */
250       MPN_COPY (tp, np + nn - nn1, nn1);
251     }
252 
253   invert_pi1 (dinv, xdp[qn1 - 1], xdp[qn1 - 2]);
254   if (BELOW_THRESHOLD (qn1, DC_DIVAPPR_Q_THRESHOLD))
255     {
256       qp[qn0 - 1 + nn1 - qn1] = mpn_sbpi1_divappr_q (qp + qn0 - 1, tp, nn1, xdp, qn1, dinv.inv32);
257     }
258   else if (BELOW_THRESHOLD (qn1, MU_DIVAPPR_Q_THRESHOLD))
259     {
260       qp[qn0 - 1 + nn1 - qn1] = mpn_dcpi1_divappr_q (qp + qn0 - 1, tp, nn1, xdp, qn1, &dinv);
261     }
262   else
263     {
264       /* FIXME: mpn_mu_divappr_q doesn't handle qh != 0.  Work around it with a
265 	 conditional subtraction here.  */
266       qh = mpn_cmp (tp + nn1 - qn1, xdp, qn1) >= 0;
267       if (qh)
268 	mpn_sub_n (tp + nn1 - qn1, tp + nn1 - qn1, xdp, qn1);
269       mpn_mu_divappr_q (qp + qn0 - 1, tp, nn1, xdp, qn1, scratch);
270       qp[qn0 - 1 + nn1 - qn1] = qh;
271     }
272   qml = qp[qn0 - 1];
273 
274   binvert_limb (di, dp[0]);  di = -di;
275 
276   if (BELOW_THRESHOLD (qn0, DC_BDIV_Q_THRESHOLD))
277     {
278       MPN_COPY (tp, np, qn0);
279       mpn_sbpi1_bdiv_q (qp, tp, qn0, dp, qn0, di);
280     }
281   else if (BELOW_THRESHOLD (qn0, MU_BDIV_Q_THRESHOLD))
282     {
283       MPN_COPY (tp, np, qn0);
284       mpn_dcpi1_bdiv_q (qp, tp, qn0, dp, qn0, di);
285     }
286   else
287     {
288       mpn_mu_bdiv_q (qp, np, qn0, dp, qn0, scratch);
289     }
290 
291   if (qml < qp[qn0 - 1])
292     mpn_decr_u (qp + qn0, 1);
293 
294   TMP_FREE;
295 }
296 #endif
297