1 /* mpn_dcpi1_divappr_q -- divide-and-conquer division, returning approximate
2 quotient. The quotient returned is either correct, or one too large.
4 Contributed to the GNU project by Torbjorn Granlund.
6 THE FUNCTIONS IN THIS FILE ARE INTERNAL WITH MUTABLE INTERFACES. IT IS ONLY
7 SAFE TO REACH THEM THROUGH DOCUMENTED INTERFACES. IN FACT, IT IS ALMOST
8 GUARANTEED THAT THEY WILL CHANGE OR DISAPPEAR IN A FUTURE GMP RELEASE.
10 Copyright 2006, 2007, 2009, 2010 Free Software Foundation, Inc.
12 This file is part of the GNU MP Library.
14 The GNU MP Library is free software; you can redistribute it and/or modify
15 it under the terms of either:
17 * the GNU Lesser General Public License as published by the Free
18 Software Foundation; either version 3 of the License, or (at your
19 option) any later version.
23 * the GNU General Public License as published by the Free Software
24 Foundation; either version 2 of the License, or (at your option) any
27 or both in parallel, as here.
29 The GNU MP Library is distributed in the hope that it will be useful, but
30 WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
31 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
34 You should have received copies of the GNU General Public License and the
35 GNU Lesser General Public License along with the GNU MP Library. If not,
36 see https://www.gnu.org/licenses/. */
44 mpn_dcpi1_divappr_q_n (mp_ptr qp
, mp_ptr np
, mp_srcptr dp
, mp_size_t n
,
45 gmp_pi1_t
*dinv
, mp_ptr tp
)
50 lo
= n
>> 1; /* floor(n/2) */
51 hi
= n
- lo
; /* ceil(n/2) */
53 if (BELOW_THRESHOLD (hi
, DC_DIV_QR_THRESHOLD
))
54 qh
= mpn_sbpi1_div_qr (qp
+ lo
, np
+ 2 * lo
, 2 * hi
, dp
+ lo
, hi
, dinv
->inv32
);
56 qh
= mpn_dcpi1_div_qr_n (qp
+ lo
, np
+ 2 * lo
, dp
+ lo
, hi
, dinv
, tp
);
58 mpn_mul (tp
, qp
+ lo
, hi
, dp
, lo
);
60 cy
= mpn_sub_n (np
+ lo
, np
+ lo
, tp
, n
);
62 cy
+= mpn_sub_n (np
+ n
, np
+ n
, dp
, lo
);
66 qh
-= mpn_sub_1 (qp
+ lo
, qp
+ lo
, hi
, 1);
67 cy
-= mpn_add_n (np
+ lo
, np
+ lo
, dp
, n
);
70 if (BELOW_THRESHOLD (lo
, DC_DIVAPPR_Q_THRESHOLD
))
71 ql
= mpn_sbpi1_divappr_q (qp
, np
+ hi
, 2 * lo
, dp
+ hi
, lo
, dinv
->inv32
);
73 ql
= mpn_dcpi1_divappr_q_n (qp
, np
+ hi
, dp
+ hi
, lo
, dinv
, tp
);
75 if (UNLIKELY (ql
!= 0))
78 for (i
= 0; i
< lo
; i
++)
79 qp
[i
] = GMP_NUMB_MASK
;
86 mpn_dcpi1_divappr_q (mp_ptr qp
, mp_ptr np
, mp_size_t nn
,
87 mp_srcptr dp
, mp_size_t dn
, gmp_pi1_t
*dinv
)
90 mp_limb_t qh
, cy
, qsave
;
98 ASSERT (dp
[dn
-1] & GMP_NUMB_HIGHBIT
);
107 qn
++; /* pretend we'll need an extra limb */
108 /* Reduce qn mod dn without division, optimizing small operations. */
113 qp
-= qn
; /* point at low limb of next quotient block */
114 np
-= qn
; /* point in the middle of partial remainder */
116 tp
= TMP_SALLOC_LIMBS (dn
);
118 /* Perform the typically smaller block first. */
121 mp_limb_t q
, n2
, n1
, n0
, d1
, d0
;
123 /* Handle qh up front, for simplicity. */
124 qh
= mpn_cmp (np
- dn
+ 1, dp
- dn
, dn
) >= 0;
126 ASSERT_NOCARRY (mpn_sub_n (np
- dn
+ 1, np
- dn
+ 1, dp
- dn
, dn
));
128 /* A single iteration of schoolbook: One 3/2 division,
129 followed by the bignum update and adjustment. */
136 ASSERT (n2
< d1
|| (n2
== d1
&& n1
<= d0
));
138 if (UNLIKELY (n2
== d1
) && n1
== d0
)
141 cy
= mpn_submul_1 (np
- dn
, dp
- dn
, dn
, q
);
146 udiv_qr_3by2 (q
, n1
, n0
, n2
, n1
, n0
, d1
, d0
, dinv
->inv32
);
151 cy
= mpn_submul_1 (np
- dn
, dp
- dn
, dn
- 2, q
);
154 n0
= (n0
- cy
) & GMP_NUMB_MASK
;
156 n1
= (n1
- cy1
) & GMP_NUMB_MASK
;
159 if (UNLIKELY (cy
!= 0))
161 n1
+= d1
+ mpn_add_n (np
- dn
, np
- dn
, dp
- dn
, dn
- 1);
163 q
= (q
- 1) & GMP_NUMB_MASK
;
176 qh
= mpn_divrem_2 (qp
, 0L, np
- 2, 4, dp
- 2);
177 else if (BELOW_THRESHOLD (qn
, DC_DIV_QR_THRESHOLD
))
178 qh
= mpn_sbpi1_div_qr (qp
, np
- qn
, 2 * qn
, dp
- qn
, qn
, dinv
->inv32
);
180 qh
= mpn_dcpi1_div_qr_n (qp
, np
- qn
, dp
- qn
, qn
, dinv
, tp
);
185 mpn_mul (tp
, qp
, qn
, dp
- dn
, dn
- qn
);
187 mpn_mul (tp
, dp
- dn
, dn
- qn
, qp
, qn
);
189 cy
= mpn_sub_n (np
- dn
, np
- dn
, tp
, dn
);
191 cy
+= mpn_sub_n (np
- dn
+ qn
, np
- dn
+ qn
, dp
- dn
, dn
- qn
);
195 qh
-= mpn_sub_1 (qp
, qp
, qn
, 1);
196 cy
-= mpn_add_n (np
- dn
, np
- dn
, dp
- dn
, dn
);
200 qn
= nn
- dn
- qn
+ 1;
205 mpn_dcpi1_div_qr_n (qp
, np
- dn
, dp
- dn
, dn
, dinv
, tp
);
209 /* Since we pretended we'd need an extra quotient limb before, we now
210 have made sure the code above left just dn-1=qn quotient limbs to
211 develop. Develop that plus a guard limb. */
216 mpn_dcpi1_divappr_q_n (qp
, np
- dn
, dp
- dn
, dn
, dinv
, tp
);
217 MPN_COPY_INCR (qp
, qp
+ 1, qn
);
223 #if 0 /* not possible since we demand nn > dn */
226 qh
= mpn_cmp (np
- dn
, dp
- dn
, dn
) >= 0;
228 mpn_sub_n (np
- dn
, np
- dn
, dp
- dn
, dn
);
234 qp
-= qn
; /* point at low limb of next quotient block */
235 np
-= qn
; /* point in the middle of partial remainder */
237 q2p
= TMP_SALLOC_LIMBS (qn
+ 1);
238 /* Should we at all check DC_DIVAPPR_Q_THRESHOLD here, or reply on
239 callers not to be silly? */
240 if (BELOW_THRESHOLD (qn
, DC_DIVAPPR_Q_THRESHOLD
))
242 qh
= mpn_sbpi1_divappr_q (q2p
, np
- qn
- 2, 2 * (qn
+ 1),
243 dp
- (qn
+ 1), qn
+ 1, dinv
->inv32
);
247 /* It is tempting to use qp for recursive scratch and put quotient in
248 tp, but the recursive scratch needs one limb too many. */
249 tp
= TMP_SALLOC_LIMBS (qn
+ 1);
250 qh
= mpn_dcpi1_divappr_q_n (q2p
, np
- qn
- 2, dp
- (qn
+ 1), qn
+ 1, dinv
, tp
);
252 MPN_COPY (qp
, q2p
+ 1, qn
);