diff options
Diffstat (limited to 'Modules/_decimal/libmpdec')
42 files changed, 17294 insertions, 0 deletions
diff --git a/Modules/_decimal/libmpdec/README.txt b/Modules/_decimal/libmpdec/README.txt new file mode 100644 index 0000000000..ad8f88c80c --- /dev/null +++ b/Modules/_decimal/libmpdec/README.txt @@ -0,0 +1,90 @@ + + +libmpdec +======== + +libmpdec is a fast C/C++ library for correctly-rounded arbitrary precision +decimal floating point arithmetic. It is a complete implementation of +Mike Cowlishaw/IBM's General Decimal Arithmetic Specification. + + +Files required for the Python _decimal module +============================================= + + Core files for small and medium precision arithmetic + ---------------------------------------------------- + + basearith.{c,h} -> Core arithmetic in base 10**9 or 10**19. + bits.h -> Portable detection of least/most significant one-bit. + constants.{c,h} -> Constants that are used in multiple files. + context.c -> Context functions. + io.{c,h} -> Conversions between mpd_t and ASCII strings, + mpd_t formatting (allows UTF-8 fill character). + memory.{c,h} -> Allocation handlers with overflow detection + and functions for switching between static + and dynamic mpd_t. + mpdecimal.{c,h} -> All (quiet) functions of the specification. + typearith.h -> Fast primitives for double word multiplication, + division etc. + + Visual Studio only: + ~~~~~~~~~~~~~~~~~~~ + vccompat.h -> snprintf <==> sprintf_s and similar things. + vcstdint.h -> stdint.h (included in VS 2010 but not in VS 2008). + vcdiv64.asm -> Double word division used in typearith.h. VS 2008 does + not allow inline asm for x64. Also, it does not provide + an intrinsic for double word division. + + Files for bignum arithmetic: + ---------------------------- + + The following files implement the Fast Number Theoretic Transform + used for multiplying coefficients with more than 1024 words (see + mpdecimal.c: _mpd_fntmul()). + + umodarith.h -> Fast low level routines for unsigned modular arithmetic. + numbertheory.{c,h} -> Routines for setting up the Number Theoretic Transform. + difradix2.{c,h} -> Decimation in frequency transform, used as the + "base case" by the following three files: + + fnt.{c,h} -> Transform arrays up to 4096 words. + sixstep.{c,h} -> Transform larger arrays of length 2**n. + fourstep.{c,h} -> Transform larger arrays of length 3 * 2**n. + + convolute.{c,h} -> Fast convolution using one of the three transform + functions. + transpose.{c,h} -> Transpositions needed for the sixstep algorithm. + crt.{c,h} -> Chinese Remainder Theorem: use information from three + transforms modulo three different primes to get the + final result. + + +Pointers to literature, proofs and more +======================================= + + literature/ + ----------- + + REFERENCES.txt -> List of relevant papers. + bignum.txt -> Explanation of the Fast Number Theoretic Transform (FNT). + fnt.py -> Verify constants used in the FNT; Python demo for the + O(N**2) discrete transform. + + matrix-transform.txt -> Proof for the Matrix Fourier Transform used in + fourstep.c. + six-step.txt -> Show that the algorithm used in sixstep.c is + a variant of the Matrix Fourier Transform. + mulmod-64.txt -> Proof for the mulmod64 algorithm from + umodarith.h. + mulmod-ppro.txt -> Proof for the x87 FPU modular multiplication + from umodarith.h. + umodarith.lisp -> ACL2 proofs for many functions from umodarith.h. + + +Library Author +============== + + Stefan Krah <skrah@bytereef.org> + + + diff --git a/Modules/_decimal/libmpdec/basearith.c b/Modules/_decimal/libmpdec/basearith.c new file mode 100644 index 0000000000..e9d5024fda --- /dev/null +++ b/Modules/_decimal/libmpdec/basearith.c @@ -0,0 +1,635 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#include "mpdecimal.h" +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <assert.h> +#include "constants.h" +#include "memory.h" +#include "typearith.h" +#include "basearith.h" + + +/*********************************************************************/ +/* Calculations in base MPD_RADIX */ +/*********************************************************************/ + + +/* + * Knuth, TAOCP, Volume 2, 4.3.1: + * w := sum of u (len m) and v (len n) + * n > 0 and m >= n + * The calling function has to handle a possible final carry. + */ +mpd_uint_t +_mpd_baseadd(mpd_uint_t *w, const mpd_uint_t *u, const mpd_uint_t *v, + mpd_size_t m, mpd_size_t n) +{ + mpd_uint_t s; + mpd_uint_t carry = 0; + mpd_size_t i; + + assert(n > 0 && m >= n); + + /* add n members of u and v */ + for (i = 0; i < n; i++) { + s = u[i] + (v[i] + carry); + carry = (s < u[i]) | (s >= MPD_RADIX); + w[i] = carry ? s-MPD_RADIX : s; + } + /* if there is a carry, propagate it */ + for (; carry && i < m; i++) { + s = u[i] + carry; + carry = (s == MPD_RADIX); + w[i] = carry ? 0 : s; + } + /* copy the rest of u */ + for (; i < m; i++) { + w[i] = u[i]; + } + + return carry; +} + +/* + * Add the contents of u to w. Carries are propagated further. The caller + * has to make sure that w is big enough. + */ +void +_mpd_baseaddto(mpd_uint_t *w, const mpd_uint_t *u, mpd_size_t n) +{ + mpd_uint_t s; + mpd_uint_t carry = 0; + mpd_size_t i; + + if (n == 0) return; + + /* add n members of u to w */ + for (i = 0; i < n; i++) { + s = w[i] + (u[i] + carry); + carry = (s < w[i]) | (s >= MPD_RADIX); + w[i] = carry ? s-MPD_RADIX : s; + } + /* if there is a carry, propagate it */ + for (; carry; i++) { + s = w[i] + carry; + carry = (s == MPD_RADIX); + w[i] = carry ? 0 : s; + } +} + +/* + * Add v to w (len m). The calling function has to handle a possible + * final carry. Assumption: m > 0. + */ +mpd_uint_t +_mpd_shortadd(mpd_uint_t *w, mpd_size_t m, mpd_uint_t v) +{ + mpd_uint_t s; + mpd_uint_t carry; + mpd_size_t i; + + assert(m > 0); + + /* add v to w */ + s = w[0] + v; + carry = (s < v) | (s >= MPD_RADIX); + w[0] = carry ? s-MPD_RADIX : s; + + /* if there is a carry, propagate it */ + for (i = 1; carry && i < m; i++) { + s = w[i] + carry; + carry = (s == MPD_RADIX); + w[i] = carry ? 0 : s; + } + + return carry; +} + +/* Increment u. The calling function has to handle a possible carry. */ +mpd_uint_t +_mpd_baseincr(mpd_uint_t *u, mpd_size_t n) +{ + mpd_uint_t s; + mpd_uint_t carry = 1; + mpd_size_t i; + + assert(n > 0); + + /* if there is a carry, propagate it */ + for (i = 0; carry && i < n; i++) { + s = u[i] + carry; + carry = (s == MPD_RADIX); + u[i] = carry ? 0 : s; + } + + return carry; +} + +/* + * Knuth, TAOCP, Volume 2, 4.3.1: + * w := difference of u (len m) and v (len n). + * number in u >= number in v; + */ +void +_mpd_basesub(mpd_uint_t *w, const mpd_uint_t *u, const mpd_uint_t *v, + mpd_size_t m, mpd_size_t n) +{ + mpd_uint_t d; + mpd_uint_t borrow = 0; + mpd_size_t i; + + assert(m > 0 && n > 0); + + /* subtract n members of v from u */ + for (i = 0; i < n; i++) { + d = u[i] - (v[i] + borrow); + borrow = (u[i] < d); + w[i] = borrow ? d + MPD_RADIX : d; + } + /* if there is a borrow, propagate it */ + for (; borrow && i < m; i++) { + d = u[i] - borrow; + borrow = (u[i] == 0); + w[i] = borrow ? MPD_RADIX-1 : d; + } + /* copy the rest of u */ + for (; i < m; i++) { + w[i] = u[i]; + } +} + +/* + * Subtract the contents of u from w. w is larger than u. Borrows are + * propagated further, but eventually w can absorb the final borrow. + */ +void +_mpd_basesubfrom(mpd_uint_t *w, const mpd_uint_t *u, mpd_size_t n) +{ + mpd_uint_t d; + mpd_uint_t borrow = 0; + mpd_size_t i; + + if (n == 0) return; + + /* subtract n members of u from w */ + for (i = 0; i < n; i++) { + d = w[i] - (u[i] + borrow); + borrow = (w[i] < d); + w[i] = borrow ? d + MPD_RADIX : d; + } + /* if there is a borrow, propagate it */ + for (; borrow; i++) { + d = w[i] - borrow; + borrow = (w[i] == 0); + w[i] = borrow ? MPD_RADIX-1 : d; + } +} + +/* w := product of u (len n) and v (single word) */ +void +_mpd_shortmul(mpd_uint_t *w, const mpd_uint_t *u, mpd_size_t n, mpd_uint_t v) +{ + mpd_uint_t hi, lo; + mpd_uint_t carry = 0; + mpd_size_t i; + + assert(n > 0); + + for (i=0; i < n; i++) { + + _mpd_mul_words(&hi, &lo, u[i], v); + lo = carry + lo; + if (lo < carry) hi++; + + _mpd_div_words_r(&carry, &w[i], hi, lo); + } + w[i] = carry; +} + +/* + * Knuth, TAOCP, Volume 2, 4.3.1: + * w := product of u (len m) and v (len n) + * w must be initialized to zero + */ +void +_mpd_basemul(mpd_uint_t *w, const mpd_uint_t *u, const mpd_uint_t *v, + mpd_size_t m, mpd_size_t n) +{ + mpd_uint_t hi, lo; + mpd_uint_t carry; + mpd_size_t i, j; + + assert(m > 0 && n > 0); + + for (j=0; j < n; j++) { + carry = 0; + for (i=0; i < m; i++) { + + _mpd_mul_words(&hi, &lo, u[i], v[j]); + lo = w[i+j] + lo; + if (lo < w[i+j]) hi++; + lo = carry + lo; + if (lo < carry) hi++; + + _mpd_div_words_r(&carry, &w[i+j], hi, lo); + } + w[j+m] = carry; + } +} + +/* + * Knuth, TAOCP Volume 2, 4.3.1, exercise 16: + * w := quotient of u (len n) divided by a single word v + */ +mpd_uint_t +_mpd_shortdiv(mpd_uint_t *w, const mpd_uint_t *u, mpd_size_t n, mpd_uint_t v) +{ + mpd_uint_t hi, lo; + mpd_uint_t rem = 0; + mpd_size_t i; + + assert(n > 0); + + for (i=n-1; i != MPD_SIZE_MAX; i--) { + + _mpd_mul_words(&hi, &lo, rem, MPD_RADIX); + lo = u[i] + lo; + if (lo < u[i]) hi++; + + _mpd_div_words(&w[i], &rem, hi, lo, v); + } + + return rem; +} + +/* + * Knuth, TAOCP Volume 2, 4.3.1: + * q, r := quotient and remainder of uconst (len nplusm) + * divided by vconst (len n) + * nplusm >= n + * + * If r is not NULL, r will contain the remainder. If r is NULL, the + * return value indicates if there is a remainder: 1 for true, 0 for + * false. A return value of -1 indicates an error. + */ +int +_mpd_basedivmod(mpd_uint_t *q, mpd_uint_t *r, + const mpd_uint_t *uconst, const mpd_uint_t *vconst, + mpd_size_t nplusm, mpd_size_t n) +{ + mpd_uint_t ustatic[MPD_MINALLOC_MAX]; + mpd_uint_t vstatic[MPD_MINALLOC_MAX]; + mpd_uint_t *u = ustatic; + mpd_uint_t *v = vstatic; + mpd_uint_t d, qhat, rhat, w2[2]; + mpd_uint_t hi, lo, x; + mpd_uint_t carry; + mpd_size_t i, j, m; + int retval = 0; + + assert(n > 1 && nplusm >= n); + m = sub_size_t(nplusm, n); + + /* D1: normalize */ + d = MPD_RADIX / (vconst[n-1] + 1); + + if (nplusm >= MPD_MINALLOC_MAX) { + if ((u = mpd_alloc(nplusm+1, sizeof *u)) == NULL) { + return -1; + } + } + if (n >= MPD_MINALLOC_MAX) { + if ((v = mpd_alloc(n+1, sizeof *v)) == NULL) { + mpd_free(u); + return -1; + } + } + + _mpd_shortmul(u, uconst, nplusm, d); + _mpd_shortmul(v, vconst, n, d); + + /* D2: loop */ + for (j=m; j != MPD_SIZE_MAX; j--) { + + /* D3: calculate qhat and rhat */ + rhat = _mpd_shortdiv(w2, u+j+n-1, 2, v[n-1]); + qhat = w2[1] * MPD_RADIX + w2[0]; + + while (1) { + if (qhat < MPD_RADIX) { + _mpd_singlemul(w2, qhat, v[n-2]); + if (w2[1] <= rhat) { + if (w2[1] != rhat || w2[0] <= u[j+n-2]) { + break; + } + } + } + qhat -= 1; + rhat += v[n-1]; + if (rhat < v[n-1] || rhat >= MPD_RADIX) { + break; + } + } + /* D4: multiply and subtract */ + carry = 0; + for (i=0; i <= n; i++) { + + _mpd_mul_words(&hi, &lo, qhat, v[i]); + + lo = carry + lo; + if (lo < carry) hi++; + + _mpd_div_words_r(&hi, &lo, hi, lo); + + x = u[i+j] - lo; + carry = (u[i+j] < x); + u[i+j] = carry ? x+MPD_RADIX : x; + carry += hi; + } + q[j] = qhat; + /* D5: test remainder */ + if (carry) { + q[j] -= 1; + /* D6: add back */ + (void)_mpd_baseadd(u+j, u+j, v, n+1, n); + } + } + + /* D8: unnormalize */ + if (r != NULL) { + _mpd_shortdiv(r, u, n, d); + /* we are not interested in the return value here */ + retval = 0; + } + else { + retval = !_mpd_isallzero(u, n); + } + + +if (u != ustatic) mpd_free(u); +if (v != vstatic) mpd_free(v); +return retval; +} + +/* + * Left shift of src by 'shift' digits; src may equal dest. + * + * dest := area of n mpd_uint_t with space for srcdigits+shift digits. + * src := coefficient with length m. + * + * The case splits in the function are non-obvious. The following + * equations might help: + * + * Let msdigits denote the number of digits in the most significant + * word of src. Then 1 <= msdigits <= rdigits. + * + * 1) shift = q * rdigits + r + * 2) srcdigits = qsrc * rdigits + msdigits + * 3) destdigits = shift + srcdigits + * = q * rdigits + r + qsrc * rdigits + msdigits + * = q * rdigits + (qsrc * rdigits + (r + msdigits)) + * + * The result has q zero words, followed by the coefficient that + * is left-shifted by r. The case r == 0 is trivial. For r > 0, it + * is important to keep in mind that we always read m source words, + * but write m+1 destination words if r + msdigits > rdigits, m words + * otherwise. + */ +void +_mpd_baseshiftl(mpd_uint_t *dest, mpd_uint_t *src, mpd_size_t n, mpd_size_t m, + mpd_size_t shift) +{ +#if defined(__GNUC__) && !defined(__INTEL_COMPILER) && !defined(__clang__) + /* spurious uninitialized warnings */ + mpd_uint_t l=l, lprev=lprev, h=h; +#else + mpd_uint_t l, lprev, h; +#endif + mpd_uint_t q, r; + mpd_uint_t ph; + + assert(m > 0 && n >= m); + + _mpd_div_word(&q, &r, (mpd_uint_t)shift, MPD_RDIGITS); + + if (r != 0) { + + ph = mpd_pow10[r]; + + --m; --n; + _mpd_divmod_pow10(&h, &lprev, src[m--], MPD_RDIGITS-r); + if (h != 0) { /* r + msdigits > rdigits <==> h != 0 */ + dest[n--] = h; + } + /* write m-1 shifted words */ + for (; m != MPD_SIZE_MAX; m--,n--) { + _mpd_divmod_pow10(&h, &l, src[m], MPD_RDIGITS-r); + dest[n] = ph * lprev + h; + lprev = l; + } + /* write least significant word */ + dest[q] = ph * lprev; + } + else { + while (--m != MPD_SIZE_MAX) { + dest[m+q] = src[m]; + } + } + + mpd_uint_zero(dest, q); +} + +/* + * Right shift of src by 'shift' digits; src may equal dest. + * Assumption: srcdigits-shift > 0. + * + * dest := area with space for srcdigits-shift digits. + * src := coefficient with length 'slen'. + * + * The case splits in the function rely on the following equations: + * + * Let msdigits denote the number of digits in the most significant + * word of src. Then 1 <= msdigits <= rdigits. + * + * 1) shift = q * rdigits + r + * 2) srcdigits = qsrc * rdigits + msdigits + * 3) destdigits = srcdigits - shift + * = qsrc * rdigits + msdigits - (q * rdigits + r) + * = (qsrc - q) * rdigits + msdigits - r + * + * Since destdigits > 0 and 1 <= msdigits <= rdigits: + * + * 4) qsrc >= q + * 5) qsrc == q ==> msdigits > r + * + * The result has slen-q words if msdigits > r, slen-q-1 words otherwise. + */ +mpd_uint_t +_mpd_baseshiftr(mpd_uint_t *dest, mpd_uint_t *src, mpd_size_t slen, + mpd_size_t shift) +{ +#if defined(__GNUC__) && !defined(__INTEL_COMPILER) && !defined(__clang__) + /* spurious uninitialized warnings */ + mpd_uint_t l=l, h=h, hprev=hprev; /* low, high, previous high */ +#else + mpd_uint_t l, h, hprev; /* low, high, previous high */ +#endif + mpd_uint_t rnd, rest; /* rounding digit, rest */ + mpd_uint_t q, r; + mpd_size_t i, j; + mpd_uint_t ph; + + assert(slen > 0); + + _mpd_div_word(&q, &r, (mpd_uint_t)shift, MPD_RDIGITS); + + rnd = rest = 0; + if (r != 0) { + + ph = mpd_pow10[MPD_RDIGITS-r]; + + _mpd_divmod_pow10(&hprev, &rest, src[q], r); + _mpd_divmod_pow10(&rnd, &rest, rest, r-1); + + if (rest == 0 && q > 0) { + rest = !_mpd_isallzero(src, q); + } + /* write slen-q-1 words */ + for (j=0,i=q+1; i<slen; i++,j++) { + _mpd_divmod_pow10(&h, &l, src[i], r); + dest[j] = ph * l + hprev; + hprev = h; + } + /* write most significant word */ + if (hprev != 0) { /* always the case if slen==q-1 */ + dest[j] = hprev; + } + } + else { + if (q > 0) { + _mpd_divmod_pow10(&rnd, &rest, src[q-1], MPD_RDIGITS-1); + /* is there any non-zero digit below rnd? */ + if (rest == 0) rest = !_mpd_isallzero(src, q-1); + } + for (j = 0; j < slen-q; j++) { + dest[j] = src[q+j]; + } + } + + /* 0-4 ==> rnd+rest < 0.5 */ + /* 5 ==> rnd+rest == 0.5 */ + /* 6-9 ==> rnd+rest > 0.5 */ + return (rnd == 0 || rnd == 5) ? rnd + !!rest : rnd; +} + + +/*********************************************************************/ +/* Calculations in base b */ +/*********************************************************************/ + +/* + * Add v to w (len m). The calling function has to handle a possible + * final carry. Assumption: m > 0. + */ +mpd_uint_t +_mpd_shortadd_b(mpd_uint_t *w, mpd_size_t m, mpd_uint_t v, mpd_uint_t b) +{ + mpd_uint_t s; + mpd_uint_t carry; + mpd_size_t i; + + assert(m > 0); + + /* add v to w */ + s = w[0] + v; + carry = (s < v) | (s >= b); + w[0] = carry ? s-b : s; + + /* if there is a carry, propagate it */ + for (i = 1; carry && i < m; i++) { + s = w[i] + carry; + carry = (s == b); + w[i] = carry ? 0 : s; + } + + return carry; +} + +/* w := product of u (len n) and v (single word) */ +void +_mpd_shortmul_b(mpd_uint_t *w, const mpd_uint_t *u, mpd_size_t n, + mpd_uint_t v, mpd_uint_t b) +{ + mpd_uint_t hi, lo; + mpd_uint_t carry = 0; + mpd_size_t i; + + assert(n > 0); + + for (i=0; i < n; i++) { + + _mpd_mul_words(&hi, &lo, u[i], v); + lo = carry + lo; + if (lo < carry) hi++; + + _mpd_div_words(&carry, &w[i], hi, lo, b); + } + w[i] = carry; +} + +/* + * Knuth, TAOCP Volume 2, 4.3.1, exercise 16: + * w := quotient of u (len n) divided by a single word v + */ +mpd_uint_t +_mpd_shortdiv_b(mpd_uint_t *w, const mpd_uint_t *u, mpd_size_t n, + mpd_uint_t v, mpd_uint_t b) +{ + mpd_uint_t hi, lo; + mpd_uint_t rem = 0; + mpd_size_t i; + + assert(n > 0); + + for (i=n-1; i != MPD_SIZE_MAX; i--) { + + _mpd_mul_words(&hi, &lo, rem, b); + lo = u[i] + lo; + if (lo < u[i]) hi++; + + _mpd_div_words(&w[i], &rem, hi, lo, v); + } + + return rem; +} + + + diff --git a/Modules/_decimal/libmpdec/basearith.h b/Modules/_decimal/libmpdec/basearith.h new file mode 100644 index 0000000000..94de862b6c --- /dev/null +++ b/Modules/_decimal/libmpdec/basearith.h @@ -0,0 +1,213 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#ifndef BASEARITH_H +#define BASEARITH_H + + +#include "mpdecimal.h" +#include <stdio.h> +#include "typearith.h" + + +mpd_uint_t _mpd_baseadd(mpd_uint_t *w, const mpd_uint_t *u, const mpd_uint_t *v, + mpd_size_t m, mpd_size_t n); +void _mpd_baseaddto(mpd_uint_t *w, const mpd_uint_t *u, mpd_size_t n); +mpd_uint_t _mpd_shortadd(mpd_uint_t *w, mpd_size_t m, mpd_uint_t v); +mpd_uint_t _mpd_shortadd_b(mpd_uint_t *w, mpd_size_t m, mpd_uint_t v, + mpd_uint_t b); +mpd_uint_t _mpd_baseincr(mpd_uint_t *u, mpd_size_t n); +void _mpd_basesub(mpd_uint_t *w, const mpd_uint_t *u, const mpd_uint_t *v, + mpd_size_t m, mpd_size_t n); +void _mpd_basesubfrom(mpd_uint_t *w, const mpd_uint_t *u, mpd_size_t n); +void _mpd_basemul(mpd_uint_t *w, const mpd_uint_t *u, const mpd_uint_t *v, + mpd_size_t m, mpd_size_t n); +void _mpd_shortmul(mpd_uint_t *w, const mpd_uint_t *u, mpd_size_t n, + mpd_uint_t v); +void _mpd_shortmul_b(mpd_uint_t *w, const mpd_uint_t *u, mpd_size_t n, + mpd_uint_t v, mpd_uint_t b); +mpd_uint_t _mpd_shortdiv(mpd_uint_t *w, const mpd_uint_t *u, mpd_size_t n, + mpd_uint_t v); +mpd_uint_t _mpd_shortdiv_b(mpd_uint_t *w, const mpd_uint_t *u, mpd_size_t n, + mpd_uint_t v, mpd_uint_t b); +int _mpd_basedivmod(mpd_uint_t *q, mpd_uint_t *r, const mpd_uint_t *uconst, + const mpd_uint_t *vconst, mpd_size_t nplusm, mpd_size_t n); +void _mpd_baseshiftl(mpd_uint_t *dest, mpd_uint_t *src, mpd_size_t n, + mpd_size_t m, mpd_size_t shift); +mpd_uint_t _mpd_baseshiftr(mpd_uint_t *dest, mpd_uint_t *src, mpd_size_t slen, + mpd_size_t shift); + + + +#ifdef CONFIG_64 +extern const mpd_uint_t mprime_rdx; + +/* + * Algorithm from: Division by Invariant Integers using Multiplication, + * T. Granlund and P. L. Montgomery, Proceedings of the SIGPLAN '94 + * Conference on Programming Language Design and Implementation. + * + * http://gmplib.org/~tege/divcnst-pldi94.pdf + * + * Variables from the paper and their translations (See section 8): + * + * N := 64 + * d := MPD_RADIX + * l := 64 + * m' := floor((2**(64+64) - 1)/MPD_RADIX) - 2**64 + * + * Since N-l == 0: + * + * dnorm := d + * n2 := hi + * n10 := lo + * + * ACL2 proof: mpd-div-words-r-correct + */ +static inline void +_mpd_div_words_r(mpd_uint_t *q, mpd_uint_t *r, mpd_uint_t hi, mpd_uint_t lo) +{ + mpd_uint_t n_adj, h, l, t; + mpd_uint_t n1_neg; + + /* n1_neg = if lo >= 2**63 then MPD_UINT_MAX else 0 */ + n1_neg = (lo & (1ULL<<63)) ? MPD_UINT_MAX : 0; + /* n_adj = if lo >= 2**63 then lo+MPD_RADIX else lo */ + n_adj = lo + (n1_neg & MPD_RADIX); + + /* (h, l) = if lo >= 2**63 then m'*(hi+1) else m'*hi */ + _mpd_mul_words(&h, &l, mprime_rdx, hi-n1_neg); + l = l + n_adj; + if (l < n_adj) h++; + t = h + hi; + /* At this point t == qest, with q == qest or q == qest+1: + * 1) 0 <= 2**64*hi + lo - qest*MPD_RADIX < 2*MPD_RADIX + */ + + /* t = 2**64-1 - qest = 2**64 - (qest+1) */ + t = MPD_UINT_MAX - t; + + /* (h, l) = 2**64*MPD_RADIX - (qest+1)*MPD_RADIX */ + _mpd_mul_words(&h, &l, t, MPD_RADIX); + l = l + lo; + if (l < lo) h++; + h += hi; + h -= MPD_RADIX; + /* (h, l) = 2**64*hi + lo - (qest+1)*MPD_RADIX (mod 2**128) + * Case q == qest+1: + * a) h == 0, l == r + * b) q := h - t == qest+1 + * c) r := l + * Case q == qest: + * a) h == MPD_UINT_MAX, l == 2**64-(MPD_RADIX-r) + * b) q := h - t == qest + * c) r := l + MPD_RADIX = r + */ + + *q = (h - t); + *r = l + (MPD_RADIX & h); +} +#else +static inline void +_mpd_div_words_r(mpd_uint_t *q, mpd_uint_t *r, mpd_uint_t hi, mpd_uint_t lo) +{ + _mpd_div_words(q, r, hi, lo, MPD_RADIX); +} +#endif + + +/* Multiply two single base MPD_RADIX words, store result in array w[2]. */ +static inline void +_mpd_singlemul(mpd_uint_t w[2], mpd_uint_t u, mpd_uint_t v) +{ + mpd_uint_t hi, lo; + + _mpd_mul_words(&hi, &lo, u, v); + _mpd_div_words_r(&w[1], &w[0], hi, lo); +} + +/* Multiply u (len 2) and v (len m, 1 <= m <= 2). */ +static inline void +_mpd_mul_2_le2(mpd_uint_t w[4], mpd_uint_t u[2], mpd_uint_t v[2], mpd_ssize_t m) +{ + mpd_uint_t hi, lo; + + _mpd_mul_words(&hi, &lo, u[0], v[0]); + _mpd_div_words_r(&w[1], &w[0], hi, lo); + + _mpd_mul_words(&hi, &lo, u[1], v[0]); + lo = w[1] + lo; + if (lo < w[1]) hi++; + _mpd_div_words_r(&w[2], &w[1], hi, lo); + if (m == 1) return; + + _mpd_mul_words(&hi, &lo, u[0], v[1]); + lo = w[1] + lo; + if (lo < w[1]) hi++; + _mpd_div_words_r(&w[3], &w[1], hi, lo); + + _mpd_mul_words(&hi, &lo, u[1], v[1]); + lo = w[2] + lo; + if (lo < w[2]) hi++; + lo = w[3] + lo; + if (lo < w[3]) hi++; + _mpd_div_words_r(&w[3], &w[2], hi, lo); +} + + +/* + * Test if all words from data[len-1] to data[0] are zero. If len is 0, nothing + * is tested and the coefficient is regarded as "all zero". + */ +static inline int +_mpd_isallzero(const mpd_uint_t *data, mpd_ssize_t len) +{ + while (--len >= 0) { + if (data[len] != 0) return 0; + } + return 1; +} + +/* + * Test if all full words from data[len-1] to data[0] are MPD_RADIX-1 + * (all nines). Return true if len == 0. + */ +static inline int +_mpd_isallnine(const mpd_uint_t *data, mpd_ssize_t len) +{ + while (--len >= 0) { + if (data[len] != MPD_RADIX-1) return 0; + } + return 1; +} + + +#endif /* BASEARITH_H */ + + + diff --git a/Modules/_decimal/libmpdec/bits.h b/Modules/_decimal/libmpdec/bits.h new file mode 100644 index 0000000000..949ec944ca --- /dev/null +++ b/Modules/_decimal/libmpdec/bits.h @@ -0,0 +1,192 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#ifndef BITS_H +#define BITS_H + + +#include "mpdecimal.h" +#include <stdio.h> + + +/* Check if n is a power of 2. */ +static inline int +ispower2(mpd_size_t n) +{ + return n != 0 && (n & (n-1)) == 0; +} + +#if defined(ANSI) +/* + * Return the most significant bit position of n from 0 to 31 (63). + * Assumptions: n != 0. + */ +static inline int +mpd_bsr(mpd_size_t n) +{ + int pos = 0; + mpd_size_t tmp; + +#ifdef CONFIG_64 + tmp = n >> 32; + if (tmp != 0) { n = tmp; pos += 32; } +#endif + tmp = n >> 16; + if (tmp != 0) { n = tmp; pos += 16; } + tmp = n >> 8; + if (tmp != 0) { n = tmp; pos += 8; } + tmp = n >> 4; + if (tmp != 0) { n = tmp; pos += 4; } + tmp = n >> 2; + if (tmp != 0) { n = tmp; pos += 2; } + tmp = n >> 1; + if (tmp != 0) { n = tmp; pos += 1; } + + return pos + (int)n - 1; +} + +/* + * Return the least significant bit position of n from 0 to 31 (63). + * Assumptions: n != 0. + */ +static inline int +mpd_bsf(mpd_size_t n) +{ + int pos; + +#ifdef CONFIG_64 + pos = 63; + if (n & 0x00000000FFFFFFFFULL) { pos -= 32; } else { n >>= 32; } + if (n & 0x000000000000FFFFULL) { pos -= 16; } else { n >>= 16; } + if (n & 0x00000000000000FFULL) { pos -= 8; } else { n >>= 8; } + if (n & 0x000000000000000FULL) { pos -= 4; } else { n >>= 4; } + if (n & 0x0000000000000003ULL) { pos -= 2; } else { n >>= 2; } + if (n & 0x0000000000000001ULL) { pos -= 1; } +#else + pos = 31; + if (n & 0x000000000000FFFFUL) { pos -= 16; } else { n >>= 16; } + if (n & 0x00000000000000FFUL) { pos -= 8; } else { n >>= 8; } + if (n & 0x000000000000000FUL) { pos -= 4; } else { n >>= 4; } + if (n & 0x0000000000000003UL) { pos -= 2; } else { n >>= 2; } + if (n & 0x0000000000000001UL) { pos -= 1; } +#endif + return pos; +} +/* END ANSI */ + +#elif defined(ASM) +/* + * Bit scan reverse. Assumptions: a != 0. + */ +static inline int +mpd_bsr(mpd_size_t a) +{ + mpd_size_t retval; + + __asm__ ( +#ifdef CONFIG_64 + "bsrq %1, %0\n\t" +#else + "bsr %1, %0\n\t" +#endif + :"=r" (retval) + :"r" (a) + :"cc" + ); + + return (int)retval; +} + +/* + * Bit scan forward. Assumptions: a != 0. + */ +static inline int +mpd_bsf(mpd_size_t a) +{ + mpd_size_t retval; + + __asm__ ( +#ifdef CONFIG_64 + "bsfq %1, %0\n\t" +#else + "bsf %1, %0\n\t" +#endif + :"=r" (retval) + :"r" (a) + :"cc" + ); + + return (int)retval; +} +/* END ASM */ + +#elif defined(MASM) +#include <intrin.h> +/* + * Bit scan reverse. Assumptions: a != 0. + */ +static inline int __cdecl +mpd_bsr(mpd_size_t a) +{ + unsigned long retval; + +#ifdef CONFIG_64 + _BitScanReverse64(&retval, a); +#else + _BitScanReverse(&retval, a); +#endif + + return (int)retval; +} + +/* + * Bit scan forward. Assumptions: a != 0. + */ +static inline int __cdecl +mpd_bsf(mpd_size_t a) +{ + unsigned long retval; + +#ifdef CONFIG_64 + _BitScanForward64(&retval, a); +#else + _BitScanForward(&retval, a); +#endif + + return (int)retval; +} +/* END MASM (_MSC_VER) */ +#else + #error "missing preprocessor definitions" +#endif /* BSR/BSF */ + + +#endif /* BITS_H */ + + + diff --git a/Modules/_decimal/libmpdec/constants.c b/Modules/_decimal/libmpdec/constants.c new file mode 100644 index 0000000000..92f5891b56 --- /dev/null +++ b/Modules/_decimal/libmpdec/constants.c @@ -0,0 +1,132 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#include "mpdecimal.h" +#include <stdio.h> +#include "constants.h" + + +#if defined(CONFIG_64) + + /* number-theory.c */ + const mpd_uint_t mpd_moduli[3] = { + 18446744069414584321ULL, 18446744056529682433ULL, 18446742974197923841ULL + }; + const mpd_uint_t mpd_roots[3] = {7ULL, 10ULL, 19ULL}; + + /* crt.c */ + const mpd_uint_t INV_P1_MOD_P2 = 18446744055098026669ULL; + const mpd_uint_t INV_P1P2_MOD_P3 = 287064143708160ULL; + const mpd_uint_t LH_P1P2 = 18446744052234715137ULL; /* (P1*P2) % 2^64 */ + const mpd_uint_t UH_P1P2 = 18446744052234715141ULL; /* (P1*P2) / 2^64 */ + + /* transpose.c */ + const mpd_size_t mpd_bits[64] = { + 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384, + 32768, 65536, 131072, 262144, 524288, 1048576, 2097152, 4194304, 8388608, + 16777216, 33554432, 67108864, 134217728, 268435456, 536870912, 1073741824, + 2147483648ULL, 4294967296ULL, 8589934592ULL, 17179869184ULL, 34359738368ULL, + 68719476736ULL, 137438953472ULL, 274877906944ULL, 549755813888ULL, + 1099511627776ULL, 2199023255552ULL, 4398046511104, 8796093022208ULL, + 17592186044416ULL, 35184372088832ULL, 70368744177664ULL, 140737488355328ULL, + 281474976710656ULL, 562949953421312ULL, 1125899906842624ULL, + 2251799813685248ULL, 4503599627370496ULL, 9007199254740992ULL, + 18014398509481984ULL, 36028797018963968ULL, 72057594037927936ULL, + 144115188075855872ULL, 288230376151711744ULL, 576460752303423488ULL, + 1152921504606846976ULL, 2305843009213693952ULL, 4611686018427387904ULL, + 9223372036854775808ULL + }; + + /* mpdecimal.c */ + const mpd_uint_t mpd_pow10[MPD_RDIGITS+1] = { + 1,10,100,1000,10000,100000,1000000,10000000,100000000,1000000000, + 10000000000ULL,100000000000ULL,1000000000000ULL,10000000000000ULL, + 100000000000000ULL,1000000000000000ULL,10000000000000000ULL, + 100000000000000000ULL,1000000000000000000ULL,10000000000000000000ULL + }; + + /* magic number for constant division by MPD_RADIX */ + const mpd_uint_t mprime_rdx = 15581492618384294730ULL; + +#elif defined(CONFIG_32) + + /* number-theory.c */ + const mpd_uint_t mpd_moduli[3] = {2113929217UL, 2013265921UL, 1811939329UL}; + const mpd_uint_t mpd_roots[3] = {5UL, 31UL, 13UL}; + + /* PentiumPro modular multiplication: These constants have to be loaded as + * 80 bit long doubles, which are not supported by certain compilers. */ + const uint32_t mpd_invmoduli[3][3] = { + {4293885170U, 2181570688U, 16352U}, /* ((long double) 1 / 2113929217UL) */ + {1698898177U, 2290649223U, 16352U}, /* ((long double) 1 / 2013265921UL) */ + {2716021846U, 2545165803U, 16352U} /* ((long double) 1 / 1811939329UL) */ + }; + + const float MPD_TWO63 = 9223372036854775808.0; /* 2^63 */ + + /* crt.c */ + const mpd_uint_t INV_P1_MOD_P2 = 2013265901UL; + const mpd_uint_t INV_P1P2_MOD_P3 = 54UL; + const mpd_uint_t LH_P1P2 = 4127195137UL; /* (P1*P2) % 2^32 */ + const mpd_uint_t UH_P1P2 = 990904320UL; /* (P1*P2) / 2^32 */ + + /* transpose.c */ + const mpd_size_t mpd_bits[32] = { + 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384, + 32768, 65536, 131072, 262144, 524288, 1048576, 2097152, 4194304, 8388608, + 16777216, 33554432, 67108864, 134217728, 268435456, 536870912, 1073741824, + 2147483648UL + }; + + /* mpdecimal.c */ + const mpd_uint_t mpd_pow10[MPD_RDIGITS+1] = { + 1,10,100,1000,10000,100000,1000000,10000000,100000000,1000000000 + }; + +#else + #error "CONFIG_64 or CONFIG_32 must be defined." +#endif + +const char *mpd_round_string[MPD_ROUND_GUARD] = { + "ROUND_UP", /* round away from 0 */ + "ROUND_DOWN", /* round toward 0 (truncate) */ + "ROUND_CEILING", /* round toward +infinity */ + "ROUND_FLOOR", /* round toward -infinity */ + "ROUND_HALF_UP", /* 0.5 is rounded up */ + "ROUND_HALF_DOWN", /* 0.5 is rounded down */ + "ROUND_HALF_EVEN", /* 0.5 is rounded to even */ + "ROUND_05UP", /* round zero or five away from 0 */ + "ROUND_TRUNC", /* truncate, but set infinity */ +}; + +const char *mpd_clamp_string[MPD_CLAMP_GUARD] = { + "CLAMP_DEFAULT", + "CLAMP_IEEE_754" +}; + + diff --git a/Modules/_decimal/libmpdec/constants.h b/Modules/_decimal/libmpdec/constants.h new file mode 100644 index 0000000000..2d63d7e3c8 --- /dev/null +++ b/Modules/_decimal/libmpdec/constants.h @@ -0,0 +1,83 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#ifndef CONSTANTS_H +#define CONSTANTS_H + + +#include "mpdecimal.h" + + +/* choice of optimized functions */ +#if defined(CONFIG_64) +/* x64 */ + #define MULMOD(a, b) x64_mulmod(a, b, umod) + #define MULMOD2C(a0, a1, w) x64_mulmod2c(a0, a1, w, umod) + #define MULMOD2(a0, b0, a1, b1) x64_mulmod2(a0, b0, a1, b1, umod) + #define POWMOD(base, exp) x64_powmod(base, exp, umod) + #define SETMODULUS(modnum) std_setmodulus(modnum, &umod) + #define SIZE3_NTT(x0, x1, x2, w3table) std_size3_ntt(x0, x1, x2, w3table, umod) +#elif defined(PPRO) +/* PentiumPro (or later) gcc inline asm */ + #define MULMOD(a, b) ppro_mulmod(a, b, &dmod, dinvmod) + #define MULMOD2C(a0, a1, w) ppro_mulmod2c(a0, a1, w, &dmod, dinvmod) + #define MULMOD2(a0, b0, a1, b1) ppro_mulmod2(a0, b0, a1, b1, &dmod, dinvmod) + #define POWMOD(base, exp) ppro_powmod(base, exp, &dmod, dinvmod) + #define SETMODULUS(modnum) ppro_setmodulus(modnum, &umod, &dmod, dinvmod) + #define SIZE3_NTT(x0, x1, x2, w3table) ppro_size3_ntt(x0, x1, x2, w3table, umod, &dmod, dinvmod) +#else + /* ANSI C99 */ + #define MULMOD(a, b) std_mulmod(a, b, umod) + #define MULMOD2C(a0, a1, w) std_mulmod2c(a0, a1, w, umod) + #define MULMOD2(a0, b0, a1, b1) std_mulmod2(a0, b0, a1, b1, umod) + #define POWMOD(base, exp) std_powmod(base, exp, umod) + #define SETMODULUS(modnum) std_setmodulus(modnum, &umod) + #define SIZE3_NTT(x0, x1, x2, w3table) std_size3_ntt(x0, x1, x2, w3table, umod) +#endif + +/* PentiumPro (or later) gcc inline asm */ +extern const float MPD_TWO63; +extern const uint32_t mpd_invmoduli[3][3]; + +enum {P1, P2, P3}; + +extern const mpd_uint_t mpd_moduli[]; +extern const mpd_uint_t mpd_roots[]; +extern const mpd_size_t mpd_bits[]; +extern const mpd_uint_t mpd_pow10[]; + +extern const mpd_uint_t INV_P1_MOD_P2; +extern const mpd_uint_t INV_P1P2_MOD_P3; +extern const mpd_uint_t LH_P1P2; +extern const mpd_uint_t UH_P1P2; + + +#endif /* CONSTANTS_H */ + + + diff --git a/Modules/_decimal/libmpdec/context.c b/Modules/_decimal/libmpdec/context.c new file mode 100644 index 0000000000..159f88c339 --- /dev/null +++ b/Modules/_decimal/libmpdec/context.c @@ -0,0 +1,286 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#include "mpdecimal.h" +#include <stdio.h> +#include <string.h> +#include <signal.h> + + +void +mpd_dflt_traphandler(mpd_context_t *ctx UNUSED) +{ + raise(SIGFPE); +} + +void (* mpd_traphandler)(mpd_context_t *) = mpd_dflt_traphandler; + + +/* Set guaranteed minimum number of coefficient words. The function may + be used once at program start. Setting MPD_MINALLOC to out-of-bounds + values is a catastrophic error, so in that case the function exits rather + than relying on the user to check a return value. */ +void +mpd_setminalloc(mpd_ssize_t n) +{ + static int minalloc_is_set = 0; + + if (minalloc_is_set) { + mpd_err_warn("mpd_setminalloc: ignoring request to set " + "MPD_MINALLOC a second time\n"); + return; + } + if (n < MPD_MINALLOC_MIN || n > MPD_MINALLOC_MAX) { + mpd_err_fatal("illegal value for MPD_MINALLOC"); /* GCOV_NOT_REACHED */ + } + MPD_MINALLOC = n; + minalloc_is_set = 1; +} + +void +mpd_init(mpd_context_t *ctx, mpd_ssize_t prec) +{ + mpd_ssize_t ideal_minalloc; + + mpd_defaultcontext(ctx); + + if (!mpd_qsetprec(ctx, prec)) { + mpd_addstatus_raise(ctx, MPD_Invalid_context); + return; + } + + ideal_minalloc = 2 * ((prec+MPD_RDIGITS-1) / MPD_RDIGITS); + if (ideal_minalloc < MPD_MINALLOC_MIN) ideal_minalloc = MPD_MINALLOC_MIN; + if (ideal_minalloc > MPD_MINALLOC_MAX) ideal_minalloc = MPD_MINALLOC_MAX; + + mpd_setminalloc(ideal_minalloc); +} + +void +mpd_maxcontext(mpd_context_t *ctx) +{ + ctx->prec=MPD_MAX_PREC; + ctx->emax=MPD_MAX_EMAX; + ctx->emin=MPD_MIN_EMIN; + ctx->round=MPD_ROUND_HALF_EVEN; + ctx->traps=MPD_Traps; + ctx->status=0; + ctx->newtrap=0; + ctx->clamp=0; + ctx->allcr=1; +} + +void +mpd_defaultcontext(mpd_context_t *ctx) +{ + ctx->prec=2*MPD_RDIGITS; + ctx->emax=MPD_MAX_EMAX; + ctx->emin=MPD_MIN_EMIN; + ctx->round=MPD_ROUND_HALF_UP; + ctx->traps=MPD_Traps; + ctx->status=0; + ctx->newtrap=0; + ctx->clamp=0; + ctx->allcr=1; +} + +void +mpd_basiccontext(mpd_context_t *ctx) +{ + ctx->prec=9; + ctx->emax=MPD_MAX_EMAX; + ctx->emin=MPD_MIN_EMIN; + ctx->round=MPD_ROUND_HALF_UP; + ctx->traps=MPD_Traps|MPD_Clamped; + ctx->status=0; + ctx->newtrap=0; + ctx->clamp=0; + ctx->allcr=1; +} + +int +mpd_ieee_context(mpd_context_t *ctx, int bits) +{ + if (bits <= 0 || bits > MPD_IEEE_CONTEXT_MAX_BITS || bits % 32) { + return -1; + } + + ctx->prec = 9 * (bits/32) - 2; + ctx->emax = 3 * ((mpd_ssize_t)1<<(bits/16+3)); + ctx->emin = 1 - ctx->emax; + ctx->round=MPD_ROUND_HALF_EVEN; + ctx->traps=0; + ctx->status=0; + ctx->newtrap=0; + ctx->clamp=1; + ctx->allcr=1; + + return 0; +} + +mpd_ssize_t +mpd_getprec(const mpd_context_t *ctx) +{ + return ctx->prec; +} + +mpd_ssize_t +mpd_getemax(const mpd_context_t *ctx) +{ + return ctx->emax; +} + +mpd_ssize_t +mpd_getemin(const mpd_context_t *ctx) +{ + return ctx->emin; +} + +int +mpd_getround(const mpd_context_t *ctx) +{ + return ctx->round; +} + +uint32_t +mpd_gettraps(const mpd_context_t *ctx) +{ + return ctx->traps; +} + +uint32_t +mpd_getstatus(const mpd_context_t *ctx) +{ + return ctx->status; +} + +int +mpd_getclamp(const mpd_context_t *ctx) +{ + return ctx->clamp; +} + +int +mpd_getcr(const mpd_context_t *ctx) +{ + return ctx->allcr; +} + + +int +mpd_qsetprec(mpd_context_t *ctx, mpd_ssize_t prec) +{ + if (prec <= 0 || prec > MPD_MAX_PREC) { + return 0; + } + ctx->prec = prec; + return 1; +} + +int +mpd_qsetemax(mpd_context_t *ctx, mpd_ssize_t emax) +{ + if (emax < 0 || emax > MPD_MAX_EMAX) { + return 0; + } + ctx->emax = emax; + return 1; +} + +int +mpd_qsetemin(mpd_context_t *ctx, mpd_ssize_t emin) +{ + if (emin > 0 || emin < MPD_MIN_EMIN) { + return 0; + } + ctx->emin = emin; + return 1; +} + +int +mpd_qsetround(mpd_context_t *ctx, int round) +{ + if (!(0 <= round && round < MPD_ROUND_GUARD)) { + return 0; + } + ctx->round = round; + return 1; +} + +int +mpd_qsettraps(mpd_context_t *ctx, uint32_t traps) +{ + if (traps > MPD_Max_status) { + return 0; + } + ctx->traps = traps; + return 1; +} + +int +mpd_qsetstatus(mpd_context_t *ctx, uint32_t flags) +{ + if (flags > MPD_Max_status) { + return 0; + } + ctx->status = flags; + return 1; +} + +int +mpd_qsetclamp(mpd_context_t *ctx, int c) +{ + if (c != 0 && c != 1) { + return 0; + } + ctx->clamp = c; + return 1; +} + +int +mpd_qsetcr(mpd_context_t *ctx, int c) +{ + if (c != 0 && c != 1) { + return 0; + } + ctx->allcr = c; + return 1; +} + + +void +mpd_addstatus_raise(mpd_context_t *ctx, uint32_t flags) +{ + ctx->status |= flags; + if (flags&ctx->traps) { + ctx->newtrap = (flags&ctx->traps); + mpd_traphandler(ctx); + } +} + + diff --git a/Modules/_decimal/libmpdec/convolute.c b/Modules/_decimal/libmpdec/convolute.c new file mode 100644 index 0000000000..b5fe131b07 --- /dev/null +++ b/Modules/_decimal/libmpdec/convolute.c @@ -0,0 +1,174 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#include "mpdecimal.h" +#include <stdio.h> +#include "bits.h" +#include "constants.h" +#include "fnt.h" +#include "fourstep.h" +#include "numbertheory.h" +#include "sixstep.h" +#include "umodarith.h" +#include "convolute.h" + + +/* Bignum: Fast convolution using the Number Theoretic Transform. Used for + the multiplication of very large coefficients. */ + + +/* Convolute the data in c1 and c2. Result is in c1. */ +int +fnt_convolute(mpd_uint_t *c1, mpd_uint_t *c2, mpd_size_t n, int modnum) +{ + int (*fnt)(mpd_uint_t *, mpd_size_t, int); + int (*inv_fnt)(mpd_uint_t *, mpd_size_t, int); +#ifdef PPRO + double dmod; + uint32_t dinvmod[3]; +#endif + mpd_uint_t n_inv, umod; + mpd_size_t i; + + + SETMODULUS(modnum); + n_inv = POWMOD(n, (umod-2)); + + if (ispower2(n)) { + if (n > SIX_STEP_THRESHOLD) { + fnt = six_step_fnt; + inv_fnt = inv_six_step_fnt; + } + else { + fnt = std_fnt; + inv_fnt = std_inv_fnt; + } + } + else { + fnt = four_step_fnt; + inv_fnt = inv_four_step_fnt; + } + + if (!fnt(c1, n, modnum)) { + return 0; + } + if (!fnt(c2, n, modnum)) { + return 0; + } + for (i = 0; i < n-1; i += 2) { + mpd_uint_t x0 = c1[i]; + mpd_uint_t y0 = c2[i]; + mpd_uint_t x1 = c1[i+1]; + mpd_uint_t y1 = c2[i+1]; + MULMOD2(&x0, y0, &x1, y1); + c1[i] = x0; + c1[i+1] = x1; + } + + if (!inv_fnt(c1, n, modnum)) { + return 0; + } + for (i = 0; i < n-3; i += 4) { + mpd_uint_t x0 = c1[i]; + mpd_uint_t x1 = c1[i+1]; + mpd_uint_t x2 = c1[i+2]; + mpd_uint_t x3 = c1[i+3]; + MULMOD2C(&x0, &x1, n_inv); + MULMOD2C(&x2, &x3, n_inv); + c1[i] = x0; + c1[i+1] = x1; + c1[i+2] = x2; + c1[i+3] = x3; + } + + return 1; +} + +/* Autoconvolute the data in c1. Result is in c1. */ +int +fnt_autoconvolute(mpd_uint_t *c1, mpd_size_t n, int modnum) +{ + int (*fnt)(mpd_uint_t *, mpd_size_t, int); + int (*inv_fnt)(mpd_uint_t *, mpd_size_t, int); +#ifdef PPRO + double dmod; + uint32_t dinvmod[3]; +#endif + mpd_uint_t n_inv, umod; + mpd_size_t i; + + + SETMODULUS(modnum); + n_inv = POWMOD(n, (umod-2)); + + if (ispower2(n)) { + if (n > SIX_STEP_THRESHOLD) { + fnt = six_step_fnt; + inv_fnt = inv_six_step_fnt; + } + else { + fnt = std_fnt; + inv_fnt = std_inv_fnt; + } + } + else { + fnt = four_step_fnt; + inv_fnt = inv_four_step_fnt; + } + + if (!fnt(c1, n, modnum)) { + return 0; + } + for (i = 0; i < n-1; i += 2) { + mpd_uint_t x0 = c1[i]; + mpd_uint_t x1 = c1[i+1]; + MULMOD2(&x0, x0, &x1, x1); + c1[i] = x0; + c1[i+1] = x1; + } + + if (!inv_fnt(c1, n, modnum)) { + return 0; + } + for (i = 0; i < n-3; i += 4) { + mpd_uint_t x0 = c1[i]; + mpd_uint_t x1 = c1[i+1]; + mpd_uint_t x2 = c1[i+2]; + mpd_uint_t x3 = c1[i+3]; + MULMOD2C(&x0, &x1, n_inv); + MULMOD2C(&x2, &x3, n_inv); + c1[i] = x0; + c1[i+1] = x1; + c1[i+2] = x2; + c1[i+3] = x3; + } + + return 1; +} + + diff --git a/Modules/_decimal/libmpdec/convolute.h b/Modules/_decimal/libmpdec/convolute.h new file mode 100644 index 0000000000..2f8d6d831b --- /dev/null +++ b/Modules/_decimal/libmpdec/convolute.h @@ -0,0 +1,43 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#ifndef CONVOLUTE_H +#define CONVOLUTE_H + + +#include "mpdecimal.h" +#include <stdio.h> + +#define SIX_STEP_THRESHOLD 4096 + + +int fnt_convolute(mpd_uint_t *c1, mpd_uint_t *c2, mpd_size_t n, int modnum); +int fnt_autoconvolute(mpd_uint_t *c1, mpd_size_t n, int modnum); + + +#endif diff --git a/Modules/_decimal/libmpdec/crt.c b/Modules/_decimal/libmpdec/crt.c new file mode 100644 index 0000000000..c71c4ee8f8 --- /dev/null +++ b/Modules/_decimal/libmpdec/crt.c @@ -0,0 +1,179 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#include "mpdecimal.h" +#include <stdio.h> +#include <assert.h> +#include "numbertheory.h" +#include "umodarith.h" +#include "crt.h" + + +/* Bignum: Chinese Remainder Theorem, extends the maximum transform length. */ + + +/* Multiply P1P2 by v, store result in w. */ +static inline void +_crt_mulP1P2_3(mpd_uint_t w[3], mpd_uint_t v) +{ + mpd_uint_t hi1, hi2, lo; + + _mpd_mul_words(&hi1, &lo, LH_P1P2, v); + w[0] = lo; + + _mpd_mul_words(&hi2, &lo, UH_P1P2, v); + lo = hi1 + lo; + if (lo < hi1) hi2++; + + w[1] = lo; + w[2] = hi2; +} + +/* Add 3 words from v to w. The result is known to fit in w. */ +static inline void +_crt_add3(mpd_uint_t w[3], mpd_uint_t v[3]) +{ + mpd_uint_t carry; + mpd_uint_t s; + + s = w[0] + v[0]; + carry = (s < w[0]); + w[0] = s; + + s = w[1] + (v[1] + carry); + carry = (s < w[1]); + w[1] = s; + + w[2] = w[2] + (v[2] + carry); +} + +/* Divide 3 words in u by v, store result in w, return remainder. */ +static inline mpd_uint_t +_crt_div3(mpd_uint_t *w, const mpd_uint_t *u, mpd_uint_t v) +{ + mpd_uint_t r1 = u[2]; + mpd_uint_t r2; + + if (r1 < v) { + w[2] = 0; + } + else { + _mpd_div_word(&w[2], &r1, u[2], v); /* GCOV_NOT_REACHED */ + } + + _mpd_div_words(&w[1], &r2, r1, u[1], v); + _mpd_div_words(&w[0], &r1, r2, u[0], v); + + return r1; +} + + +/* + * Chinese Remainder Theorem: + * Algorithm from Joerg Arndt, "Matters Computational", + * Chapter 37.4.1 [http://www.jjj.de/fxt/] + * + * See also Knuth, TAOCP, Volume 2, 4.3.2, exercise 7. + */ + +/* + * CRT with carry: x1, x2, x3 contain numbers modulo p1, p2, p3. For each + * triple of members of the arrays, find the unique z modulo p1*p2*p3, with + * zmax = p1*p2*p3 - 1. + * + * In each iteration of the loop, split z into result[i] = z % MPD_RADIX + * and carry = z / MPD_RADIX. Let N be the size of carry[] and cmax the + * maximum carry. + * + * Limits for the 32-bit build: + * + * N = 2**96 + * cmax = 7711435591312380274 + * + * Limits for the 64 bit build: + * + * N = 2**192 + * cmax = 627710135393475385904124401220046371710 + * + * The following statements hold for both versions: + * + * 1) cmax + zmax < N, so the addition does not overflow. + * + * 2) (cmax + zmax) / MPD_RADIX == cmax. + * + * 3) If c <= cmax, then c_next = (c + zmax) / MPD_RADIX <= cmax. + */ +void +crt3(mpd_uint_t *x1, mpd_uint_t *x2, mpd_uint_t *x3, mpd_size_t rsize) +{ + mpd_uint_t p1 = mpd_moduli[P1]; + mpd_uint_t umod; +#ifdef PPRO + double dmod; + uint32_t dinvmod[3]; +#endif + mpd_uint_t a1, a2, a3; + mpd_uint_t s; + mpd_uint_t z[3], t[3]; + mpd_uint_t carry[3] = {0,0,0}; + mpd_uint_t hi, lo; + mpd_size_t i; + + for (i = 0; i < rsize; i++) { + + a1 = x1[i]; + a2 = x2[i]; + a3 = x3[i]; + + SETMODULUS(P2); + s = ext_submod(a2, a1, umod); + s = MULMOD(s, INV_P1_MOD_P2); + + _mpd_mul_words(&hi, &lo, s, p1); + lo = lo + a1; + if (lo < a1) hi++; + + SETMODULUS(P3); + s = dw_submod(a3, hi, lo, umod); + s = MULMOD(s, INV_P1P2_MOD_P3); + + z[0] = lo; + z[1] = hi; + z[2] = 0; + + _crt_mulP1P2_3(t, s); + _crt_add3(z, t); + _crt_add3(carry, z); + + x1[i] = _crt_div3(carry, carry, MPD_RADIX); + } + + assert(carry[0] == 0 && carry[1] == 0 && carry[2] == 0); +} + + diff --git a/Modules/_decimal/libmpdec/crt.h b/Modules/_decimal/libmpdec/crt.h new file mode 100644 index 0000000000..0e03e5d077 --- /dev/null +++ b/Modules/_decimal/libmpdec/crt.h @@ -0,0 +1,40 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#ifndef CRT_H +#define CRT_H + + +#include "mpdecimal.h" +#include <stdio.h> + + +void crt3(mpd_uint_t *x1, mpd_uint_t *x2, mpd_uint_t *x3, mpd_size_t nmemb); + + +#endif diff --git a/Modules/_decimal/libmpdec/difradix2.c b/Modules/_decimal/libmpdec/difradix2.c new file mode 100644 index 0000000000..4ebb0b54b0 --- /dev/null +++ b/Modules/_decimal/libmpdec/difradix2.c @@ -0,0 +1,173 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#include "mpdecimal.h" +#include <stdio.h> +#include <assert.h> +#include "bits.h" +#include "numbertheory.h" +#include "umodarith.h" +#include "difradix2.h" + + +/* Bignum: The actual transform routine (decimation in frequency). */ + + +/* + * Generate index pairs (x, bitreverse(x)) and carry out the permutation. + * n must be a power of two. + * Algorithm due to Brent/Lehmann, see Joerg Arndt, "Matters Computational", + * Chapter 1.14.4. [http://www.jjj.de/fxt/] + */ +static inline void +bitreverse_permute(mpd_uint_t a[], mpd_size_t n) +{ + mpd_size_t x = 0; + mpd_size_t r = 0; + mpd_uint_t t; + + do { /* Invariant: r = bitreverse(x) */ + if (r > x) { + t = a[x]; + a[x] = a[r]; + a[r] = t; + } + /* Flip trailing consecutive 1 bits and the first zero bit + * that absorbs a possible carry. */ + x += 1; + /* Mirror the operation on r: Flip n_trailing_zeros(x)+1 + high bits of r. */ + r ^= (n - (n >> (mpd_bsf(x)+1))); + /* The loop invariant is preserved. */ + } while (x < n); +} + + +/* Fast Number Theoretic Transform, decimation in frequency. */ +void +fnt_dif2(mpd_uint_t a[], mpd_size_t n, struct fnt_params *tparams) +{ + mpd_uint_t *wtable = tparams->wtable; + mpd_uint_t umod; +#ifdef PPRO + double dmod; + uint32_t dinvmod[3]; +#endif + mpd_uint_t u0, u1, v0, v1; + mpd_uint_t w, w0, w1, wstep; + mpd_size_t m, mhalf; + mpd_size_t j, r; + + + assert(ispower2(n)); + assert(n >= 4); + + SETMODULUS(tparams->modnum); + + /* m == n */ + mhalf = n / 2; + for (j = 0; j < mhalf; j += 2) { + + w0 = wtable[j]; + w1 = wtable[j+1]; + + u0 = a[j]; + v0 = a[j+mhalf]; + + u1 = a[j+1]; + v1 = a[j+1+mhalf]; + + a[j] = addmod(u0, v0, umod); + v0 = submod(u0, v0, umod); + + a[j+1] = addmod(u1, v1, umod); + v1 = submod(u1, v1, umod); + + MULMOD2(&v0, w0, &v1, w1); + + a[j+mhalf] = v0; + a[j+1+mhalf] = v1; + + } + + wstep = 2; + for (m = n/2; m >= 2; m>>=1, wstep<<=1) { + + mhalf = m / 2; + + /* j == 0 */ + for (r = 0; r < n; r += 2*m) { + + u0 = a[r]; + v0 = a[r+mhalf]; + + u1 = a[m+r]; + v1 = a[m+r+mhalf]; + + a[r] = addmod(u0, v0, umod); + v0 = submod(u0, v0, umod); + + a[m+r] = addmod(u1, v1, umod); + v1 = submod(u1, v1, umod); + + a[r+mhalf] = v0; + a[m+r+mhalf] = v1; + } + + for (j = 1; j < mhalf; j++) { + + w = wtable[j*wstep]; + + for (r = 0; r < n; r += 2*m) { + + u0 = a[r+j]; + v0 = a[r+j+mhalf]; + + u1 = a[m+r+j]; + v1 = a[m+r+j+mhalf]; + + a[r+j] = addmod(u0, v0, umod); + v0 = submod(u0, v0, umod); + + a[m+r+j] = addmod(u1, v1, umod); + v1 = submod(u1, v1, umod); + + MULMOD2C(&v0, &v1, w); + + a[r+j+mhalf] = v0; + a[m+r+j+mhalf] = v1; + } + + } + + } + + bitreverse_permute(a, n); +} + + diff --git a/Modules/_decimal/libmpdec/difradix2.h b/Modules/_decimal/libmpdec/difradix2.h new file mode 100644 index 0000000000..759442a2bc --- /dev/null +++ b/Modules/_decimal/libmpdec/difradix2.h @@ -0,0 +1,41 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#ifndef DIF_RADIX2_H +#define DIF_RADIX2_H + + +#include "mpdecimal.h" +#include <stdio.h> +#include "numbertheory.h" + + +void fnt_dif2(mpd_uint_t a[], mpd_size_t n, struct fnt_params *tparams); + + +#endif diff --git a/Modules/_decimal/libmpdec/fnt.c b/Modules/_decimal/libmpdec/fnt.c new file mode 100644 index 0000000000..93116539b9 --- /dev/null +++ b/Modules/_decimal/libmpdec/fnt.c @@ -0,0 +1,81 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#include "mpdecimal.h" +#include <stdio.h> +#include <stdlib.h> +#include <assert.h> +#include "bits.h" +#include "difradix2.h" +#include "numbertheory.h" +#include "fnt.h" + + +/* Bignum: Fast transform for medium-sized coefficients. */ + + +/* forward transform, sign = -1 */ +int +std_fnt(mpd_uint_t *a, mpd_size_t n, int modnum) +{ + struct fnt_params *tparams; + + assert(ispower2(n)); + assert(n >= 4); + assert(n <= 3*MPD_MAXTRANSFORM_2N); + + if ((tparams = _mpd_init_fnt_params(n, -1, modnum)) == NULL) { + return 0; + } + fnt_dif2(a, n, tparams); + + mpd_free(tparams); + return 1; +} + +/* reverse transform, sign = 1 */ +int +std_inv_fnt(mpd_uint_t *a, mpd_size_t n, int modnum) +{ + struct fnt_params *tparams; + + assert(ispower2(n)); + assert(n >= 4); + assert(n <= 3*MPD_MAXTRANSFORM_2N); + + if ((tparams = _mpd_init_fnt_params(n, 1, modnum)) == NULL) { + return 0; + } + fnt_dif2(a, n, tparams); + + mpd_free(tparams); + return 1; +} + + + diff --git a/Modules/_decimal/libmpdec/fnt.h b/Modules/_decimal/libmpdec/fnt.h new file mode 100644 index 0000000000..2d701b6c8a --- /dev/null +++ b/Modules/_decimal/libmpdec/fnt.h @@ -0,0 +1,42 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#ifndef FNT_H +#define FNT_H + + +#include "mpdecimal.h" +#include <stdio.h> + + +int std_fnt(mpd_uint_t a[], mpd_size_t n, int modnum); +int std_inv_fnt(mpd_uint_t a[], mpd_size_t n, int modnum); + + +#endif + diff --git a/Modules/_decimal/libmpdec/fourstep.c b/Modules/_decimal/libmpdec/fourstep.c new file mode 100644 index 0000000000..aa32c0d5cf --- /dev/null +++ b/Modules/_decimal/libmpdec/fourstep.c @@ -0,0 +1,257 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#include "mpdecimal.h" +#include <assert.h> +#include "numbertheory.h" +#include "sixstep.h" +#include "transpose.h" +#include "umodarith.h" +#include "fourstep.h" + + +/* Bignum: Cache efficient Matrix Fourier Transform for arrays of the + form 3 * 2**n (See literature/matrix-transform.txt). */ + + +#ifndef PPRO +static inline void +std_size3_ntt(mpd_uint_t *x1, mpd_uint_t *x2, mpd_uint_t *x3, + mpd_uint_t w3table[3], mpd_uint_t umod) +{ + mpd_uint_t r1, r2; + mpd_uint_t w; + mpd_uint_t s, tmp; + + + /* k = 0 -> w = 1 */ + s = *x1; + s = addmod(s, *x2, umod); + s = addmod(s, *x3, umod); + + r1 = s; + + /* k = 1 */ + s = *x1; + + w = w3table[1]; + tmp = MULMOD(*x2, w); + s = addmod(s, tmp, umod); + + w = w3table[2]; + tmp = MULMOD(*x3, w); + s = addmod(s, tmp, umod); + + r2 = s; + + /* k = 2 */ + s = *x1; + + w = w3table[2]; + tmp = MULMOD(*x2, w); + s = addmod(s, tmp, umod); + + w = w3table[1]; + tmp = MULMOD(*x3, w); + s = addmod(s, tmp, umod); + + *x3 = s; + *x2 = r2; + *x1 = r1; +} +#else /* PPRO */ +static inline void +ppro_size3_ntt(mpd_uint_t *x1, mpd_uint_t *x2, mpd_uint_t *x3, mpd_uint_t w3table[3], + mpd_uint_t umod, double *dmod, uint32_t dinvmod[3]) +{ + mpd_uint_t r1, r2; + mpd_uint_t w; + mpd_uint_t s, tmp; + + + /* k = 0 -> w = 1 */ + s = *x1; + s = addmod(s, *x2, umod); + s = addmod(s, *x3, umod); + + r1 = s; + + /* k = 1 */ + s = *x1; + + w = w3table[1]; + tmp = ppro_mulmod(*x2, w, dmod, dinvmod); + s = addmod(s, tmp, umod); + + w = w3table[2]; + tmp = ppro_mulmod(*x3, w, dmod, dinvmod); + s = addmod(s, tmp, umod); + + r2 = s; + + /* k = 2 */ + s = *x1; + + w = w3table[2]; + tmp = ppro_mulmod(*x2, w, dmod, dinvmod); + s = addmod(s, tmp, umod); + + w = w3table[1]; + tmp = ppro_mulmod(*x3, w, dmod, dinvmod); + s = addmod(s, tmp, umod); + + *x3 = s; + *x2 = r2; + *x1 = r1; +} +#endif + + +/* forward transform, sign = -1; transform length = 3 * 2**n */ +int +four_step_fnt(mpd_uint_t *a, mpd_size_t n, int modnum) +{ + mpd_size_t R = 3; /* number of rows */ + mpd_size_t C = n / 3; /* number of columns */ + mpd_uint_t w3table[3]; + mpd_uint_t kernel, w0, w1, wstep; + mpd_uint_t *s, *p0, *p1, *p2; + mpd_uint_t umod; +#ifdef PPRO + double dmod; + uint32_t dinvmod[3]; +#endif + mpd_size_t i, k; + + + assert(n >= 48); + assert(n <= 3*MPD_MAXTRANSFORM_2N); + + + /* Length R transform on the columns. */ + SETMODULUS(modnum); + _mpd_init_w3table(w3table, -1, modnum); + for (p0=a, p1=p0+C, p2=p0+2*C; p0<a+C; p0++,p1++,p2++) { + + SIZE3_NTT(p0, p1, p2, w3table); + } + + /* Multiply each matrix element (addressed by i*C+k) by r**(i*k). */ + kernel = _mpd_getkernel(n, -1, modnum); + for (i = 1; i < R; i++) { + w0 = 1; /* r**(i*0): initial value for k=0 */ + w1 = POWMOD(kernel, i); /* r**(i*1): initial value for k=1 */ + wstep = MULMOD(w1, w1); /* r**(2*i) */ + for (k = 0; k < C-1; k += 2) { + mpd_uint_t x0 = a[i*C+k]; + mpd_uint_t x1 = a[i*C+k+1]; + MULMOD2(&x0, w0, &x1, w1); + MULMOD2C(&w0, &w1, wstep); /* r**(i*(k+2)) = r**(i*k) * r**(2*i) */ + a[i*C+k] = x0; + a[i*C+k+1] = x1; + } + } + + /* Length C transform on the rows. */ + for (s = a; s < a+n; s += C) { + if (!six_step_fnt(s, C, modnum)) { + return 0; + } + } + +#if 0 + /* An unordered transform is sufficient for convolution. */ + /* Transpose the matrix. */ + transpose_3xpow2(a, R, C); +#endif + + return 1; +} + +/* backward transform, sign = 1; transform length = 3 * 2**n */ +int +inv_four_step_fnt(mpd_uint_t *a, mpd_size_t n, int modnum) +{ + mpd_size_t R = 3; /* number of rows */ + mpd_size_t C = n / 3; /* number of columns */ + mpd_uint_t w3table[3]; + mpd_uint_t kernel, w0, w1, wstep; + mpd_uint_t *s, *p0, *p1, *p2; + mpd_uint_t umod; +#ifdef PPRO + double dmod; + uint32_t dinvmod[3]; +#endif + mpd_size_t i, k; + + + assert(n >= 48); + assert(n <= 3*MPD_MAXTRANSFORM_2N); + + +#if 0 + /* An unordered transform is sufficient for convolution. */ + /* Transpose the matrix, producing an R*C matrix. */ + transpose_3xpow2(a, C, R); +#endif + + /* Length C transform on the rows. */ + for (s = a; s < a+n; s += C) { + if (!inv_six_step_fnt(s, C, modnum)) { + return 0; + } + } + + /* Multiply each matrix element (addressed by i*C+k) by r**(i*k). */ + SETMODULUS(modnum); + kernel = _mpd_getkernel(n, 1, modnum); + for (i = 1; i < R; i++) { + w0 = 1; + w1 = POWMOD(kernel, i); + wstep = MULMOD(w1, w1); + for (k = 0; k < C; k += 2) { + mpd_uint_t x0 = a[i*C+k]; + mpd_uint_t x1 = a[i*C+k+1]; + MULMOD2(&x0, w0, &x1, w1); + MULMOD2C(&w0, &w1, wstep); + a[i*C+k] = x0; + a[i*C+k+1] = x1; + } + } + + /* Length R transform on the columns. */ + _mpd_init_w3table(w3table, 1, modnum); + for (p0=a, p1=p0+C, p2=p0+2*C; p0<a+C; p0++,p1++,p2++) { + + SIZE3_NTT(p0, p1, p2, w3table); + } + + return 1; +} + + diff --git a/Modules/_decimal/libmpdec/fourstep.h b/Modules/_decimal/libmpdec/fourstep.h new file mode 100644 index 0000000000..61d9d6a71a --- /dev/null +++ b/Modules/_decimal/libmpdec/fourstep.h @@ -0,0 +1,41 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#ifndef FOUR_STEP_H +#define FOUR_STEP_H + + +#include "mpdecimal.h" +#include <stdio.h> + + +int four_step_fnt(mpd_uint_t *a, mpd_size_t n, int modnum); +int inv_four_step_fnt(mpd_uint_t *a, mpd_size_t n, int modnum); + + +#endif diff --git a/Modules/_decimal/libmpdec/io.c b/Modules/_decimal/libmpdec/io.c new file mode 100644 index 0000000000..2648135bb6 --- /dev/null +++ b/Modules/_decimal/libmpdec/io.c @@ -0,0 +1,1575 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#include "mpdecimal.h" +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <ctype.h> +#include <limits.h> +#include <assert.h> +#include <errno.h> +#include <locale.h> +#include "bits.h" +#include "constants.h" +#include "memory.h" +#include "typearith.h" +#include "io.h" + + +/* This file contains functions for decimal <-> string conversions, including + PEP-3101 formatting for numeric types. */ + + +/* + * Work around the behavior of tolower() and strcasecmp() in certain + * locales. For example, in tr_TR.utf8: + * + * tolower((unsigned char)'I') == 'I' + * + * u is the exact uppercase version of l; n is strlen(l) or strlen(l)+1 + */ +static inline int +_mpd_strneq(const char *s, const char *l, const char *u, size_t n) +{ + while (--n != SIZE_MAX) { + if (*s != *l && *s != *u) { + return 0; + } + s++; u++; l++; + } + + return 1; +} + +static mpd_ssize_t +strtoexp(const char *s) +{ + char *end; + mpd_ssize_t retval; + + errno = 0; + retval = mpd_strtossize(s, &end, 10); + if (errno == 0 && !(*s != '\0' && *end == '\0')) + errno = EINVAL; + + return retval; +} + +/* + * Scan 'len' words. The most significant word contains 'r' digits, + * the remaining words are full words. Skip dpoint. The string 's' must + * consist of digits and an optional single decimal point at 'dpoint'. + */ +static void +string_to_coeff(mpd_uint_t *data, const char *s, const char *dpoint, int r, + size_t len) +{ + int j; + + if (r > 0) { + data[--len] = 0; + for (j = 0; j < r; j++, s++) { + if (s == dpoint) s++; + data[len] = 10 * data[len] + (*s - '0'); + } + } + + while (--len != SIZE_MAX) { + data[len] = 0; + for (j = 0; j < MPD_RDIGITS; j++, s++) { + if (s == dpoint) s++; + data[len] = 10 * data[len] + (*s - '0'); + } + } +} + +/* + * Partially verify a numeric string of the form: + * + * [cdigits][.][cdigits][eE][+-][edigits] + * + * If successful, return a pointer to the location of the first + * relevant coefficient digit. This digit is either non-zero or + * part of one of the following patterns: + * + * ["0\x00", "0.\x00", "0.E", "0.e", "0E", "0e"] + * + * The locations of a single optional dot or indicator are stored + * in 'dpoint' and 'exp'. + * + * The end of the string is stored in 'end'. If an indicator [eE] + * occurs without trailing [edigits], the condition is caught + * later by strtoexp(). + */ +static const char * +scan_dpoint_exp(const char *s, const char **dpoint, const char **exp, + const char **end) +{ + const char *coeff = NULL; + + *dpoint = NULL; + *exp = NULL; + for (; *s != '\0'; s++) { + switch (*s) { + case '.': + if (*dpoint != NULL || *exp != NULL) + return NULL; + *dpoint = s; + break; + case 'E': case 'e': + if (*exp != NULL) + return NULL; + *exp = s; + if (*(s+1) == '+' || *(s+1) == '-') + s++; + break; + default: + if (!isdigit((uchar)*s)) + return NULL; + if (coeff == NULL && *exp == NULL) { + if (*s == '0') { + if (!isdigit((uchar)*(s+1))) + if (!(*(s+1) == '.' && + isdigit((uchar)*(s+2)))) + coeff = s; + } + else { + coeff = s; + } + } + break; + + } + } + + *end = s; + return coeff; +} + +/* scan the payload of a NaN */ +static const char * +scan_payload(const char *s, const char **end) +{ + const char *coeff; + + while (*s == '0') + s++; + coeff = s; + + while (isdigit((uchar)*s)) + s++; + *end = s; + + return (*s == '\0') ? coeff : NULL; +} + +/* convert a character string to a decimal */ +void +mpd_qset_string(mpd_t *dec, const char *s, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_ssize_t q, r, len; + const char *coeff, *end; + const char *dpoint = NULL, *exp = NULL; + size_t digits; + uint8_t sign = MPD_POS; + + mpd_set_flags(dec, 0); + dec->len = 0; + dec->exp = 0; + + /* sign */ + if (*s == '+') { + s++; + } + else if (*s == '-') { + mpd_set_negative(dec); + sign = MPD_NEG; + s++; + } + + if (_mpd_strneq(s, "nan", "NAN", 3)) { /* NaN */ + s += 3; + mpd_setspecial(dec, sign, MPD_NAN); + if (*s == '\0') + return; + /* validate payload: digits only */ + if ((coeff = scan_payload(s, &end)) == NULL) + goto conversion_error; + /* payload consists entirely of zeros */ + if (*coeff == '\0') + return; + digits = end - coeff; + /* prec >= 1, clamp is 0 or 1 */ + if (digits > (size_t)(ctx->prec-ctx->clamp)) + goto conversion_error; + } /* sNaN */ + else if (_mpd_strneq(s, "snan", "SNAN", 4)) { + s += 4; + mpd_setspecial(dec, sign, MPD_SNAN); + if (*s == '\0') + return; + /* validate payload: digits only */ + if ((coeff = scan_payload(s, &end)) == NULL) + goto conversion_error; + /* payload consists entirely of zeros */ + if (*coeff == '\0') + return; + digits = end - coeff; + if (digits > (size_t)(ctx->prec-ctx->clamp)) + goto conversion_error; + } + else if (_mpd_strneq(s, "inf", "INF", 3)) { + s += 3; + if (*s == '\0' || _mpd_strneq(s, "inity", "INITY", 6)) { + /* numeric-value: infinity */ + mpd_setspecial(dec, sign, MPD_INF); + return; + } + goto conversion_error; + } + else { + /* scan for start of coefficient, decimal point, indicator, end */ + if ((coeff = scan_dpoint_exp(s, &dpoint, &exp, &end)) == NULL) + goto conversion_error; + + /* numeric-value: [exponent-part] */ + if (exp) { + /* exponent-part */ + end = exp; exp++; + dec->exp = strtoexp(exp); + if (errno) { + if (!(errno == ERANGE && + (dec->exp == MPD_SSIZE_MAX || + dec->exp == MPD_SSIZE_MIN))) + goto conversion_error; + } + } + + digits = end - coeff; + if (dpoint) { + size_t fracdigits = end-dpoint-1; + if (dpoint > coeff) digits--; + + if (fracdigits > MPD_MAX_PREC) { + goto conversion_error; + } + if (dec->exp < MPD_SSIZE_MIN+(mpd_ssize_t)fracdigits) { + dec->exp = MPD_SSIZE_MIN; + } + else { + dec->exp -= (mpd_ssize_t)fracdigits; + } + } + if (digits > MPD_MAX_PREC) { + goto conversion_error; + } + if (dec->exp > MPD_EXP_INF) { + dec->exp = MPD_EXP_INF; + } + if (dec->exp == MPD_SSIZE_MIN) { + dec->exp = MPD_SSIZE_MIN+1; + } + } + + _mpd_idiv_word(&q, &r, (mpd_ssize_t)digits, MPD_RDIGITS); + + len = (r == 0) ? q : q+1; + if (len == 0) { + goto conversion_error; /* GCOV_NOT_REACHED */ + } + if (!mpd_qresize(dec, len, status)) { + mpd_seterror(dec, MPD_Malloc_error, status); + return; + } + dec->len = len; + + string_to_coeff(dec->data, coeff, dpoint, (int)r, len); + + mpd_setdigits(dec); + mpd_qfinalize(dec, ctx, status); + return; + +conversion_error: + /* standard wants a positive NaN */ + mpd_seterror(dec, MPD_Conversion_syntax, status); +} + +/* Print word x with n decimal digits to string s. dot is either NULL + or the location of a decimal point. */ +#define EXTRACT_DIGIT(s, x, d, dot) \ + if (s == dot) *s++ = '.'; *s++ = '0' + (char)(x / d); x %= d +static inline char * +word_to_string(char *s, mpd_uint_t x, int n, char *dot) +{ + switch(n) { +#ifdef CONFIG_64 + case 20: EXTRACT_DIGIT(s, x, 10000000000000000000ULL, dot); /* GCOV_NOT_REACHED */ + case 19: EXTRACT_DIGIT(s, x, 1000000000000000000ULL, dot); + case 18: EXTRACT_DIGIT(s, x, 100000000000000000ULL, dot); + case 17: EXTRACT_DIGIT(s, x, 10000000000000000ULL, dot); + case 16: EXTRACT_DIGIT(s, x, 1000000000000000ULL, dot); + case 15: EXTRACT_DIGIT(s, x, 100000000000000ULL, dot); + case 14: EXTRACT_DIGIT(s, x, 10000000000000ULL, dot); + case 13: EXTRACT_DIGIT(s, x, 1000000000000ULL, dot); + case 12: EXTRACT_DIGIT(s, x, 100000000000ULL, dot); + case 11: EXTRACT_DIGIT(s, x, 10000000000ULL, dot); +#endif + case 10: EXTRACT_DIGIT(s, x, 1000000000UL, dot); + case 9: EXTRACT_DIGIT(s, x, 100000000UL, dot); + case 8: EXTRACT_DIGIT(s, x, 10000000UL, dot); + case 7: EXTRACT_DIGIT(s, x, 1000000UL, dot); + case 6: EXTRACT_DIGIT(s, x, 100000UL, dot); + case 5: EXTRACT_DIGIT(s, x, 10000UL, dot); + case 4: EXTRACT_DIGIT(s, x, 1000UL, dot); + case 3: EXTRACT_DIGIT(s, x, 100UL, dot); + case 2: EXTRACT_DIGIT(s, x, 10UL, dot); + default: if (s == dot) *s++ = '.'; *s++ = '0' + (char)x; + } + + *s = '\0'; + return s; +} + +/* Print exponent x to string s. Undefined for MPD_SSIZE_MIN. */ +static inline char * +exp_to_string(char *s, mpd_ssize_t x) +{ + char sign = '+'; + + if (x < 0) { + sign = '-'; + x = -x; + } + *s++ = sign; + + return word_to_string(s, x, mpd_word_digits(x), NULL); +} + +/* Print the coefficient of dec to string s. len(dec) > 0. */ +static inline char * +coeff_to_string(char *s, const mpd_t *dec) +{ + mpd_uint_t x; + mpd_ssize_t i; + + /* most significant word */ + x = mpd_msword(dec); + s = word_to_string(s, x, mpd_word_digits(x), NULL); + + /* remaining full words */ + for (i=dec->len-2; i >= 0; --i) { + x = dec->data[i]; + s = word_to_string(s, x, MPD_RDIGITS, NULL); + } + + return s; +} + +/* Print the coefficient of dec to string s. len(dec) > 0. dot is either + NULL or a pointer to the location of a decimal point. */ +static inline char * +coeff_to_string_dot(char *s, char *dot, const mpd_t *dec) +{ + mpd_uint_t x; + mpd_ssize_t i; + + /* most significant word */ + x = mpd_msword(dec); + s = word_to_string(s, x, mpd_word_digits(x), dot); + + /* remaining full words */ + for (i=dec->len-2; i >= 0; --i) { + x = dec->data[i]; + s = word_to_string(s, x, MPD_RDIGITS, dot); + } + + return s; +} + +/* Format type */ +#define MPD_FMT_LOWER 0x00000000 +#define MPD_FMT_UPPER 0x00000001 +#define MPD_FMT_TOSCI 0x00000002 +#define MPD_FMT_TOENG 0x00000004 +#define MPD_FMT_EXP 0x00000008 +#define MPD_FMT_FIXED 0x00000010 +#define MPD_FMT_PERCENT 0x00000020 +#define MPD_FMT_SIGN_SPACE 0x00000040 +#define MPD_FMT_SIGN_PLUS 0x00000080 + +/* Default place of the decimal point for MPD_FMT_TOSCI, MPD_FMT_EXP */ +#define MPD_DEFAULT_DOTPLACE 1 + +/* + * Set *result to the string representation of a decimal. Return the length + * of *result, not including the terminating '\0' character. + * + * Formatting is done according to 'flags'. A return value of -1 with *result + * set to NULL indicates MPD_Malloc_error. + * + * 'dplace' is the default place of the decimal point. It is always set to + * MPD_DEFAULT_DOTPLACE except for zeros in combination with MPD_FMT_EXP. + */ +static mpd_ssize_t +_mpd_to_string(char **result, const mpd_t *dec, int flags, mpd_ssize_t dplace) +{ + char *decstring = NULL, *cp = NULL; + mpd_ssize_t ldigits; + mpd_ssize_t mem = 0, k; + + if (mpd_isspecial(dec)) { + + mem = sizeof "-Infinity"; + if (mpd_isnan(dec) && dec->len > 0) { + /* diagnostic code */ + mem += dec->digits; + } + cp = decstring = mpd_alloc(mem, sizeof *decstring); + if (cp == NULL) { + *result = NULL; + return -1; + } + + if (mpd_isnegative(dec)) { + *cp++ = '-'; + } + else if (flags&MPD_FMT_SIGN_SPACE) { + *cp++ = ' '; + } + else if (flags&MPD_FMT_SIGN_PLUS) { + *cp++ = '+'; + } + + if (mpd_isnan(dec)) { + if (mpd_isqnan(dec)) { + strcpy(cp, "NaN"); + cp += 3; + } + else { + strcpy(cp, "sNaN"); + cp += 4; + } + if (dec->len > 0) { /* diagnostic code */ + cp = coeff_to_string(cp, dec); + } + } + else if (mpd_isinfinite(dec)) { + strcpy(cp, "Infinity"); + cp += 8; + } + else { /* debug */ + abort(); /* GCOV_NOT_REACHED */ + } + } + else { + assert(dec->len > 0); + + /* + * For easier manipulation of the decimal point's location + * and the exponent that is finally printed, the number is + * rescaled to a virtual representation with exp = 0. Here + * ldigits denotes the number of decimal digits to the left + * of the decimal point and remains constant once initialized. + * + * dplace is the location of the decimal point relative to + * the start of the coefficient. Note that 3) always holds + * when dplace is shifted. + * + * 1) ldigits := dec->digits - dec->exp + * 2) dplace := ldigits (initially) + * 3) exp := ldigits - dplace (initially exp = 0) + * + * 0.00000_.____._____000000. + * ^ ^ ^ ^ + * | | | | + * | | | `- dplace >= digits + * | | `- dplace in the middle of the coefficient + * | ` dplace = 1 (after the first coefficient digit) + * `- dplace <= 0 + */ + + ldigits = dec->digits + dec->exp; + + if (flags&MPD_FMT_EXP) { + ; + } + else if (flags&MPD_FMT_FIXED || (dec->exp <= 0 && ldigits > -6)) { + /* MPD_FMT_FIXED: always use fixed point notation. + * MPD_FMT_TOSCI, MPD_FMT_TOENG: for a certain range, + * override exponent notation. */ + dplace = ldigits; + } + else if (flags&MPD_FMT_TOENG) { + if (mpd_iszero(dec)) { + /* If the exponent is divisible by three, + * dplace = 1. Otherwise, move dplace one + * or two places to the left. */ + dplace = -1 + mod_mpd_ssize_t(dec->exp+2, 3); + } + else { /* ldigits-1 is the adjusted exponent, which + * should be divisible by three. If not, move + * dplace one or two places to the right. */ + dplace += mod_mpd_ssize_t(ldigits-1, 3); + } + } + + /* + * Basic space requirements: + * + * [-][.][coeffdigits][E][-][expdigits+1][%]['\0'] + * + * If the decimal point lies outside of the coefficient digits, + * space is adjusted accordingly. + */ + if (dplace <= 0) { + mem = -dplace + dec->digits + 2; + } + else if (dplace >= dec->digits) { + mem = dplace; + } + else { + mem = dec->digits; + } + mem += (MPD_EXPDIGITS+1+6); + + cp = decstring = mpd_alloc(mem, sizeof *decstring); + if (cp == NULL) { + *result = NULL; + return -1; + } + + + if (mpd_isnegative(dec)) { + *cp++ = '-'; + } + else if (flags&MPD_FMT_SIGN_SPACE) { + *cp++ = ' '; + } + else if (flags&MPD_FMT_SIGN_PLUS) { + *cp++ = '+'; + } + + if (dplace <= 0) { + /* space: -dplace+dec->digits+2 */ + *cp++ = '0'; + *cp++ = '.'; + for (k = 0; k < -dplace; k++) { + *cp++ = '0'; + } + cp = coeff_to_string(cp, dec); + } + else if (dplace >= dec->digits) { + /* space: dplace */ + cp = coeff_to_string(cp, dec); + for (k = 0; k < dplace-dec->digits; k++) { + *cp++ = '0'; + } + } + else { + /* space: dec->digits+1 */ + cp = coeff_to_string_dot(cp, cp+dplace, dec); + } + + /* + * Conditions for printing an exponent: + * + * MPD_FMT_TOSCI, MPD_FMT_TOENG: only if ldigits != dplace + * MPD_FMT_FIXED: never (ldigits == dplace) + * MPD_FMT_EXP: always + */ + if (ldigits != dplace || flags&MPD_FMT_EXP) { + /* space: expdigits+2 */ + *cp++ = (flags&MPD_FMT_UPPER) ? 'E' : 'e'; + cp = exp_to_string(cp, ldigits-dplace); + } + + if (flags&MPD_FMT_PERCENT) { + *cp++ = '%'; + } + } + + assert(cp < decstring+mem); + assert(cp-decstring < MPD_SSIZE_MAX); + + *cp = '\0'; + *result = decstring; + return (mpd_ssize_t)(cp-decstring); +} + +char * +mpd_to_sci(const mpd_t *dec, int fmt) +{ + char *res; + int flags = MPD_FMT_TOSCI; + + flags |= fmt ? MPD_FMT_UPPER : MPD_FMT_LOWER; + (void)_mpd_to_string(&res, dec, flags, MPD_DEFAULT_DOTPLACE); + return res; +} + +char * +mpd_to_eng(const mpd_t *dec, int fmt) +{ + char *res; + int flags = MPD_FMT_TOENG; + + flags |= fmt ? MPD_FMT_UPPER : MPD_FMT_LOWER; + (void)_mpd_to_string(&res, dec, flags, MPD_DEFAULT_DOTPLACE); + return res; +} + +mpd_ssize_t +mpd_to_sci_size(char **res, const mpd_t *dec, int fmt) +{ + int flags = MPD_FMT_TOSCI; + + flags |= fmt ? MPD_FMT_UPPER : MPD_FMT_LOWER; + return _mpd_to_string(res, dec, flags, MPD_DEFAULT_DOTPLACE); +} + +mpd_ssize_t +mpd_to_eng_size(char **res, const mpd_t *dec, int fmt) +{ + int flags = MPD_FMT_TOENG; + + flags |= fmt ? MPD_FMT_UPPER : MPD_FMT_LOWER; + return _mpd_to_string(res, dec, flags, MPD_DEFAULT_DOTPLACE); +} + +/* Copy a single UTF-8 char to dest. See: The Unicode Standard, version 5.2, + chapter 3.9: Well-formed UTF-8 byte sequences. */ +static int +_mpd_copy_utf8(char dest[5], const char *s) +{ + const uchar *cp = (const uchar *)s; + uchar lb, ub; + int count, i; + + + if (*cp == 0) { + /* empty string */ + dest[0] = '\0'; + return 0; + } + else if (*cp <= 0x7f) { + /* ascii */ + dest[0] = *cp; + dest[1] = '\0'; + return 1; + } + else if (0xc2 <= *cp && *cp <= 0xdf) { + lb = 0x80; ub = 0xbf; + count = 2; + } + else if (*cp == 0xe0) { + lb = 0xa0; ub = 0xbf; + count = 3; + } + else if (*cp <= 0xec) { + lb = 0x80; ub = 0xbf; + count = 3; + } + else if (*cp == 0xed) { + lb = 0x80; ub = 0x9f; + count = 3; + } + else if (*cp <= 0xef) { + lb = 0x80; ub = 0xbf; + count = 3; + } + else if (*cp == 0xf0) { + lb = 0x90; ub = 0xbf; + count = 4; + } + else if (*cp <= 0xf3) { + lb = 0x80; ub = 0xbf; + count = 4; + } + else if (*cp == 0xf4) { + lb = 0x80; ub = 0x8f; + count = 4; + } + else { + /* invalid */ + goto error; + } + + dest[0] = *cp++; + if (*cp < lb || ub < *cp) { + goto error; + } + dest[1] = *cp++; + for (i = 2; i < count; i++) { + if (*cp < 0x80 || 0xbf < *cp) { + goto error; + } + dest[i] = *cp++; + } + dest[i] = '\0'; + + return count; + +error: + dest[0] = '\0'; + return -1; +} + +int +mpd_validate_lconv(mpd_spec_t *spec) +{ + size_t n; +#if CHAR_MAX == SCHAR_MAX + const char *cp = spec->grouping; + while (*cp != '\0') { + if (*cp++ < 0) { + return -1; + } + } +#endif + n = strlen(spec->dot); + if (n == 0 || n > 4) { + return -1; + } + if (strlen(spec->sep) > 4) { + return -1; + } + + return 0; +} + +int +mpd_parse_fmt_str(mpd_spec_t *spec, const char *fmt, int caps) +{ + char *cp = (char *)fmt; + int have_align = 0, n; + + /* defaults */ + spec->min_width = 0; + spec->prec = -1; + spec->type = caps ? 'G' : 'g'; + spec->align = '>'; + spec->sign = '-'; + spec->dot = ""; + spec->sep = ""; + spec->grouping = ""; + + + /* presume that the first character is a UTF-8 fill character */ + if ((n = _mpd_copy_utf8(spec->fill, cp)) < 0) { + return 0; + } + + /* alignment directive, prefixed by a fill character */ + if (*cp && (*(cp+n) == '<' || *(cp+n) == '>' || + *(cp+n) == '=' || *(cp+n) == '^')) { + cp += n; + spec->align = *cp++; + have_align = 1; + } /* alignment directive */ + else { + /* default fill character */ + spec->fill[0] = ' '; + spec->fill[1] = '\0'; + if (*cp == '<' || *cp == '>' || + *cp == '=' || *cp == '^') { + spec->align = *cp++; + have_align = 1; + } + } + + /* sign formatting */ + if (*cp == '+' || *cp == '-' || *cp == ' ') { + spec->sign = *cp++; + } + + /* zero padding */ + if (*cp == '0') { + /* zero padding implies alignment, which should not be + * specified twice. */ + if (have_align) { + return 0; + } + spec->align = 'z'; + spec->fill[0] = *cp++; + spec->fill[1] = '\0'; + } + + /* minimum width */ + if (isdigit((uchar)*cp)) { + if (*cp == '0') { + return 0; + } + errno = 0; + spec->min_width = mpd_strtossize(cp, &cp, 10); + if (errno == ERANGE || errno == EINVAL) { + return 0; + } + } + + /* thousands separator */ + if (*cp == ',') { + spec->dot = "."; + spec->sep = ","; + spec->grouping = "\003\003"; + cp++; + } + + /* fraction digits or significant digits */ + if (*cp == '.') { + cp++; + if (!isdigit((uchar)*cp)) { + return 0; + } + errno = 0; + spec->prec = mpd_strtossize(cp, &cp, 10); + if (errno == ERANGE || errno == EINVAL) { + return 0; + } + } + + /* type */ + if (*cp == 'E' || *cp == 'e' || *cp == 'F' || *cp == 'f' || + *cp == 'G' || *cp == 'g' || *cp == '%') { + spec->type = *cp++; + } + else if (*cp == 'N' || *cp == 'n') { + /* locale specific conversion */ + struct lconv *lc; + /* separator has already been specified */ + if (*spec->sep) { + return 0; + } + spec->type = *cp++; + spec->type = (spec->type == 'N') ? 'G' : 'g'; + lc = localeconv(); + spec->dot = lc->decimal_point; + spec->sep = lc->thousands_sep; + spec->grouping = lc->grouping; + if (mpd_validate_lconv(spec) < 0) { + return 0; /* GCOV_NOT_REACHED */ + } + } + + /* check correctness */ + if (*cp != '\0') { + return 0; + } + + return 1; +} + +/* + * The following functions assume that spec->min_width <= MPD_MAX_PREC, which + * is made sure in mpd_qformat_spec. Then, even with a spec that inserts a + * four-byte separator after each digit, nbytes in the following struct + * cannot overflow. + */ + +/* Multibyte string */ +typedef struct { + mpd_ssize_t nbytes; /* length in bytes */ + mpd_ssize_t nchars; /* length in chars */ + mpd_ssize_t cur; /* current write index */ + char *data; +} mpd_mbstr_t; + +static inline void +_mpd_bcopy(char *dest, const char *src, mpd_ssize_t n) +{ + while (--n >= 0) { + dest[n] = src[n]; + } +} + +static inline void +_mbstr_copy_char(mpd_mbstr_t *dest, const char *src, mpd_ssize_t n) +{ + dest->nbytes += n; + dest->nchars += (n > 0 ? 1 : 0); + dest->cur -= n; + + if (dest->data != NULL) { + _mpd_bcopy(dest->data+dest->cur, src, n); + } +} + +static inline void +_mbstr_copy_ascii(mpd_mbstr_t *dest, const char *src, mpd_ssize_t n) +{ + dest->nbytes += n; + dest->nchars += n; + dest->cur -= n; + + if (dest->data != NULL) { + _mpd_bcopy(dest->data+dest->cur, src, n); + } +} + +static inline void +_mbstr_copy_pad(mpd_mbstr_t *dest, mpd_ssize_t n) +{ + dest->nbytes += n; + dest->nchars += n; + dest->cur -= n; + + if (dest->data != NULL) { + char *cp = dest->data + dest->cur; + while (--n >= 0) { + cp[n] = '0'; + } + } +} + +/* + * Copy a numeric string to dest->data, adding separators in the integer + * part according to spec->grouping. If leading zero padding is enabled + * and the result is smaller than spec->min_width, continue adding zeros + * and separators until the minimum width is reached. + * + * The final length of dest->data is stored in dest->nbytes. The number + * of UTF-8 characters is stored in dest->nchars. + * + * First run (dest->data == NULL): determine the length of the result + * string and store it in dest->nbytes. + * + * Second run (write to dest->data): data is written in chunks and in + * reverse order, starting with the rest of the numeric string. + */ +static void +_mpd_add_sep_dot(mpd_mbstr_t *dest, + const char *sign, /* location of optional sign */ + const char *src, mpd_ssize_t n_src, /* integer part and length */ + const char *dot, /* location of optional decimal point */ + const char *rest, mpd_ssize_t n_rest, /* remaining part and length */ + const mpd_spec_t *spec) +{ + mpd_ssize_t n_sep, n_sign, consume; + const char *g; + int pad = 0; + + n_sign = sign ? 1 : 0; + n_sep = (mpd_ssize_t)strlen(spec->sep); + /* Initial write index: set to location of '\0' in the output string. + * Irrelevant for the first run. */ + dest->cur = dest->nbytes; + dest->nbytes = dest->nchars = 0; + + _mbstr_copy_ascii(dest, rest, n_rest); + + if (dot) { + _mbstr_copy_char(dest, dot, (mpd_ssize_t)strlen(dot)); + } + + g = spec->grouping; + consume = *g; + while (1) { + /* If the group length is 0 or CHAR_MAX or greater than the + * number of source bytes, consume all remaining bytes. */ + if (*g == 0 || *g == CHAR_MAX || consume > n_src) { + consume = n_src; + } + n_src -= consume; + if (pad) { + _mbstr_copy_pad(dest, consume); + } + else { + _mbstr_copy_ascii(dest, src+n_src, consume); + } + + if (n_src == 0) { + /* Either the real source of intpart digits or the virtual + * source of padding zeros is exhausted. */ + if (spec->align == 'z' && + dest->nchars + n_sign < spec->min_width) { + /* Zero padding is set and length < min_width: + * Generate n_src additional characters. */ + n_src = spec->min_width - (dest->nchars + n_sign); + /* Next iteration: + * case *g == 0 || *g == CHAR_MAX: + * consume all padding characters + * case consume < g*: + * fill remainder of current group + * case consume == g* + * copying is a no-op */ + consume = *g - consume; + /* Switch on virtual source of zeros. */ + pad = 1; + continue; + } + break; + } + + if (n_sep > 0) { + /* If padding is switched on, separators are counted + * as padding characters. This rule does not apply if + * the separator would be the first character of the + * result string. */ + if (pad && n_src > 1) n_src -= 1; + _mbstr_copy_char(dest, spec->sep, n_sep); + } + + /* If non-NUL, use the next value for grouping. */ + if (*g && *(g+1)) g++; + consume = *g; + } + + if (sign) { + _mbstr_copy_ascii(dest, sign, 1); + } + + if (dest->data) { + dest->data[dest->nbytes] = '\0'; + } +} + +/* + * Convert a numeric-string to its locale-specific appearance. + * The string must have one of these forms: + * + * 1) [sign] digits [exponent-part] + * 2) [sign] digits '.' [digits] [exponent-part] + * + * Not allowed, since _mpd_to_string() never returns this form: + * + * 3) [sign] '.' digits [exponent-part] + * + * Input: result->data := original numeric string (ASCII) + * result->bytes := strlen(result->data) + * result->nchars := strlen(result->data) + * + * Output: result->data := modified or original string + * result->bytes := strlen(result->data) + * result->nchars := number of characters (possibly UTF-8) + */ +static int +_mpd_apply_lconv(mpd_mbstr_t *result, const mpd_spec_t *spec, uint32_t *status) +{ + const char *sign = NULL, *intpart = NULL, *dot = NULL; + const char *rest, *dp; + char *decstring; + mpd_ssize_t n_int, n_rest; + + /* original numeric string */ + dp = result->data; + + /* sign */ + if (*dp == '+' || *dp == '-' || *dp == ' ') { + sign = dp++; + } + /* integer part */ + assert(isdigit((uchar)*dp)); + intpart = dp++; + while (isdigit((uchar)*dp)) { + dp++; + } + n_int = (mpd_ssize_t)(dp-intpart); + /* decimal point */ + if (*dp == '.') { + dp++; dot = spec->dot; + } + /* rest */ + rest = dp; + n_rest = result->nbytes - (mpd_ssize_t)(dp-result->data); + + if (dot == NULL && (*spec->sep == '\0' || *spec->grouping == '\0')) { + /* _mpd_add_sep_dot() would not change anything */ + return 1; + } + + /* Determine the size of the new decimal string after inserting the + * decimal point, optional separators and optional padding. */ + decstring = result->data; + result->data = NULL; + _mpd_add_sep_dot(result, sign, intpart, n_int, dot, + rest, n_rest, spec); + + result->data = mpd_alloc(result->nbytes+1, 1); + if (result->data == NULL) { + *status |= MPD_Malloc_error; + mpd_free(decstring); + return 0; + } + + /* Perform actual writes. */ + _mpd_add_sep_dot(result, sign, intpart, n_int, dot, + rest, n_rest, spec); + + mpd_free(decstring); + return 1; +} + +/* Add padding to the formatted string if necessary. */ +static int +_mpd_add_pad(mpd_mbstr_t *result, const mpd_spec_t *spec, uint32_t *status) +{ + if (result->nchars < spec->min_width) { + mpd_ssize_t add_chars, add_bytes; + size_t lpad = 0, rpad = 0; + size_t n_fill, len, i, j; + char align = spec->align; + uint8_t err = 0; + char *cp; + + n_fill = strlen(spec->fill); + add_chars = (spec->min_width - result->nchars); + /* max value: MPD_MAX_PREC * 4 */ + add_bytes = add_chars * (mpd_ssize_t)n_fill; + + cp = result->data = mpd_realloc(result->data, + result->nbytes+add_bytes+1, + sizeof *result->data, &err); + if (err) { + *status |= MPD_Malloc_error; + mpd_free(result->data); + return 0; + } + + if (align == 'z') { + align = '='; + } + + if (align == '<') { + rpad = add_chars; + } + else if (align == '>' || align == '=') { + lpad = add_chars; + } + else { /* align == '^' */ + lpad = add_chars/2; + rpad = add_chars-lpad; + } + + len = result->nbytes; + if (align == '=' && (*cp == '-' || *cp == '+' || *cp == ' ')) { + /* leave sign in the leading position */ + cp++; len--; + } + + memmove(cp+n_fill*lpad, cp, len); + for (i = 0; i < lpad; i++) { + for (j = 0; j < n_fill; j++) { + cp[i*n_fill+j] = spec->fill[j]; + } + } + cp += (n_fill*lpad + len); + for (i = 0; i < rpad; i++) { + for (j = 0; j < n_fill; j++) { + cp[i*n_fill+j] = spec->fill[j]; + } + } + + result->nbytes += add_bytes; + result->nchars += add_chars; + result->data[result->nbytes] = '\0'; + } + + return 1; +} + +/* Round a number to prec digits. The adjusted exponent stays the same + or increases by one if rounding up crosses a power of ten boundary. + If result->digits would exceed MPD_MAX_PREC+1, MPD_Invalid_operation + is set and the result is NaN. */ +static inline void +_mpd_round(mpd_t *result, const mpd_t *a, mpd_ssize_t prec, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_ssize_t exp = a->exp + a->digits - prec; + + if (prec <= 0) { + mpd_seterror(result, MPD_Invalid_operation, status); /* GCOV_NOT_REACHED */ + return; /* GCOV_NOT_REACHED */ + } + if (mpd_isspecial(a) || mpd_iszero(a)) { + mpd_qcopy(result, a, status); /* GCOV_NOT_REACHED */ + return; /* GCOV_NOT_REACHED */ + } + + mpd_qrescale_fmt(result, a, exp, ctx, status); + if (result->digits > prec) { + mpd_qrescale_fmt(result, result, exp+1, ctx, status); + } +} + +/* + * Return the string representation of an mpd_t, formatted according to 'spec'. + * The format specification is assumed to be valid. Memory errors are indicated + * as usual. This function is quiet. + */ +char * +mpd_qformat_spec(const mpd_t *dec, const mpd_spec_t *spec, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_uint_t dt[MPD_MINALLOC_MAX]; + mpd_t tmp = {MPD_STATIC|MPD_STATIC_DATA,0,0,0,MPD_MINALLOC_MAX,dt}; + mpd_ssize_t dplace = MPD_DEFAULT_DOTPLACE; + mpd_mbstr_t result; + mpd_spec_t stackspec; + char type = spec->type; + int flags = 0; + + + if (spec->min_width > MPD_MAX_PREC) { + *status |= MPD_Invalid_operation; + return NULL; + } + + if (isupper((uchar)type)) { + type = tolower((uchar)type); + flags |= MPD_FMT_UPPER; + } + if (spec->sign == ' ') { + flags |= MPD_FMT_SIGN_SPACE; + } + else if (spec->sign == '+') { + flags |= MPD_FMT_SIGN_PLUS; + } + + if (mpd_isspecial(dec)) { + if (spec->align == 'z') { + stackspec = *spec; + stackspec.fill[0] = ' '; + stackspec.fill[1] = '\0'; + stackspec.align = '>'; + spec = &stackspec; + } + } + else { + uint32_t workstatus = 0; + mpd_ssize_t prec; + + switch (type) { + case 'g': flags |= MPD_FMT_TOSCI; break; + case 'e': flags |= MPD_FMT_EXP; break; + case '%': flags |= MPD_FMT_PERCENT; + if (!mpd_qcopy(&tmp, dec, status)) { + return NULL; + } + tmp.exp += 2; + dec = &tmp; + type = 'f'; /* fall through */ + case 'f': flags |= MPD_FMT_FIXED; break; + default: abort(); /* debug: GCOV_NOT_REACHED */ + } + + if (spec->prec >= 0) { + if (spec->prec > MPD_MAX_PREC) { + *status |= MPD_Invalid_operation; + goto error; + } + + switch (type) { + case 'g': + prec = (spec->prec == 0) ? 1 : spec->prec; + if (dec->digits > prec) { + _mpd_round(&tmp, dec, prec, ctx, + &workstatus); + dec = &tmp; + } + break; + case 'e': + if (mpd_iszero(dec)) { + dplace = 1-spec->prec; + } + else { + _mpd_round(&tmp, dec, spec->prec+1, ctx, + &workstatus); + dec = &tmp; + } + break; + case 'f': + mpd_qrescale(&tmp, dec, -spec->prec, ctx, + &workstatus); + dec = &tmp; + break; + } + } + + if (type == 'f') { + if (mpd_iszero(dec) && dec->exp > 0) { + mpd_qrescale(&tmp, dec, 0, ctx, &workstatus); + dec = &tmp; + } + } + + if (workstatus&MPD_Errors) { + *status |= (workstatus&MPD_Errors); + goto error; + } + } + + /* + * At this point, for all scaled or non-scaled decimals: + * 1) 1 <= digits <= MAX_PREC+1 + * 2) adjexp(scaled) = adjexp(orig) [+1] + * 3) case 'g': MIN_ETINY <= exp <= MAX_EMAX+1 + * case 'e': MIN_ETINY-MAX_PREC <= exp <= MAX_EMAX+1 + * case 'f': MIN_ETINY <= exp <= MAX_EMAX+1 + * 4) max memory alloc in _mpd_to_string: + * case 'g': MAX_PREC+36 + * case 'e': MAX_PREC+36 + * case 'f': 2*MPD_MAX_PREC+30 + */ + result.nbytes = _mpd_to_string(&result.data, dec, flags, dplace); + result.nchars = result.nbytes; + if (result.nbytes < 0) { + *status |= MPD_Malloc_error; + goto error; + } + + if (*spec->dot != '\0' && !mpd_isspecial(dec)) { + if (result.nchars > MPD_MAX_PREC+36) { + /* Since a group length of one is not explicitly + * disallowed, ensure that it is always possible to + * insert a four byte separator after each digit. */ + *status |= MPD_Invalid_operation; + mpd_free(result.data); + goto error; + } + if (!_mpd_apply_lconv(&result, spec, status)) { + goto error; + } + } + + if (spec->min_width) { + if (!_mpd_add_pad(&result, spec, status)) { + goto error; + } + } + + mpd_del(&tmp); + return result.data; + +error: + mpd_del(&tmp); + return NULL; +} + +char * +mpd_qformat(const mpd_t *dec, const char *fmt, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_spec_t spec; + + if (!mpd_parse_fmt_str(&spec, fmt, 1)) { + *status |= MPD_Invalid_operation; + return NULL; + } + + return mpd_qformat_spec(dec, &spec, ctx, status); +} + +/* + * The specification has a *condition* called Invalid_operation and an + * IEEE *signal* called Invalid_operation. The former corresponds to + * MPD_Invalid_operation, the latter to MPD_IEEE_Invalid_operation. + * MPD_IEEE_Invalid_operation comprises the following conditions: + * + * [MPD_Conversion_syntax, MPD_Division_impossible, MPD_Division_undefined, + * MPD_Fpu_error, MPD_Invalid_context, MPD_Invalid_operation, + * MPD_Malloc_error] + * + * In the following functions, 'flag' denotes the condition, 'signal' + * denotes the IEEE signal. + */ + +static const char *mpd_flag_string[MPD_NUM_FLAGS] = { + "Clamped", + "Conversion_syntax", + "Division_by_zero", + "Division_impossible", + "Division_undefined", + "Fpu_error", + "Inexact", + "Invalid_context", + "Invalid_operation", + "Malloc_error", + "Not_implemented", + "Overflow", + "Rounded", + "Subnormal", + "Underflow", +}; + +static const char *mpd_signal_string[MPD_NUM_FLAGS] = { + "Clamped", + "IEEE_Invalid_operation", + "Division_by_zero", + "IEEE_Invalid_operation", + "IEEE_Invalid_operation", + "IEEE_Invalid_operation", + "Inexact", + "IEEE_Invalid_operation", + "IEEE_Invalid_operation", + "IEEE_Invalid_operation", + "Not_implemented", + "Overflow", + "Rounded", + "Subnormal", + "Underflow", +}; + +/* print conditions to buffer, separated by spaces */ +int +mpd_snprint_flags(char *dest, int nmemb, uint32_t flags) +{ + char *cp; + int n, j; + + assert(nmemb >= MPD_MAX_FLAG_STRING); + + *dest = '\0'; cp = dest; + for (j = 0; j < MPD_NUM_FLAGS; j++) { + if (flags & (1U<<j)) { + n = snprintf(cp, nmemb, "%s ", mpd_flag_string[j]); + if (n < 0 || n >= nmemb) return -1; + cp += n; nmemb -= n; + } + } + + if (cp != dest) { + *(--cp) = '\0'; + } + + return (int)(cp-dest); +} + +/* print conditions to buffer, in list form */ +int +mpd_lsnprint_flags(char *dest, int nmemb, uint32_t flags, const char *flag_string[]) +{ + char *cp; + int n, j; + + assert(nmemb >= MPD_MAX_FLAG_LIST); + if (flag_string == NULL) { + flag_string = mpd_flag_string; + } + + *dest = '['; + *(dest+1) = '\0'; + cp = dest+1; + --nmemb; + + for (j = 0; j < MPD_NUM_FLAGS; j++) { + if (flags & (1U<<j)) { + n = snprintf(cp, nmemb, "%s, ", flag_string[j]); + if (n < 0 || n >= nmemb) return -1; + cp += n; nmemb -= n; + } + } + + /* erase the last ", " */ + if (cp != dest+1) { + cp -= 2; + } + + *cp++ = ']'; + *cp = '\0'; + + return (int)(cp-dest); /* strlen, without NUL terminator */ +} + +/* print signals to buffer, in list form */ +int +mpd_lsnprint_signals(char *dest, int nmemb, uint32_t flags, const char *signal_string[]) +{ + char *cp; + int n, j; + int ieee_invalid_done = 0; + + assert(nmemb >= MPD_MAX_SIGNAL_LIST); + if (signal_string == NULL) { + signal_string = mpd_signal_string; + } + + *dest = '['; + *(dest+1) = '\0'; + cp = dest+1; + --nmemb; + + for (j = 0; j < MPD_NUM_FLAGS; j++) { + uint32_t f = flags & (1U<<j); + if (f) { + if (f&MPD_IEEE_Invalid_operation) { + if (ieee_invalid_done) { + continue; + } + ieee_invalid_done = 1; + } + n = snprintf(cp, nmemb, "%s, ", signal_string[j]); + if (n < 0 || n >= nmemb) return -1; + cp += n; nmemb -= n; + } + } + + /* erase the last ", " */ + if (cp != dest+1) { + cp -= 2; + } + + *cp++ = ']'; + *cp = '\0'; + + return (int)(cp-dest); /* strlen, without NUL terminator */ +} + +/* The following two functions are mainly intended for debugging. */ +void +mpd_fprint(FILE *file, const mpd_t *dec) +{ + char *decstring; + + decstring = mpd_to_sci(dec, 1); + if (decstring != NULL) { + fprintf(file, "%s\n", decstring); + mpd_free(decstring); + } + else { + fputs("mpd_fprint: output error\n", file); /* GCOV_NOT_REACHED */ + } +} + +void +mpd_print(const mpd_t *dec) +{ + char *decstring; + + decstring = mpd_to_sci(dec, 1); + if (decstring != NULL) { + printf("%s\n", decstring); + mpd_free(decstring); + } + else { + fputs("mpd_fprint: output error\n", stderr); /* GCOV_NOT_REACHED */ + } +} + + diff --git a/Modules/_decimal/libmpdec/io.h b/Modules/_decimal/libmpdec/io.h new file mode 100644 index 0000000000..3dfce732aa --- /dev/null +++ b/Modules/_decimal/libmpdec/io.h @@ -0,0 +1,59 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#ifndef IO_H +#define IO_H + + +#include <errno.h> +#include "mpdecimal.h" + + +#if SIZE_MAX == MPD_SIZE_MAX + #define mpd_strtossize _mpd_strtossize +#else +static inline mpd_ssize_t +mpd_strtossize(const char *s, char **end, int base) +{ + int64_t retval; + + errno = 0; + retval = _mpd_strtossize(s, end, base); + if (errno == 0 && (retval > MPD_SSIZE_MAX || retval < MPD_SSIZE_MIN)) { + errno = ERANGE; + } + if (errno == ERANGE) { + return (retval < 0) ? MPD_SSIZE_MIN : MPD_SSIZE_MAX; + } + + return (mpd_ssize_t)retval; +} +#endif + + +#endif diff --git a/Modules/_decimal/libmpdec/literature/REFERENCES.txt b/Modules/_decimal/libmpdec/literature/REFERENCES.txt new file mode 100644 index 0000000000..9ed5782656 --- /dev/null +++ b/Modules/_decimal/libmpdec/literature/REFERENCES.txt @@ -0,0 +1,51 @@ + + +This document contains links to the literature used in the process of +creating the library. The list is probably not complete. + + +Mike Cowlishaw: General Decimal Arithmetic Specification +http://speleotrove.com/decimal/decarith.html + + +Jean-Michel Muller: On the definition of ulp (x) +lara.inist.fr/bitstream/2332/518/1/LIP-RR2005-09.pdf + + +T. E. Hull, A. Abrham: Properly rounded variable precision square root +http://portal.acm.org/citation.cfm?id=214413 + + +T. E. Hull, A. Abrham: Variable precision exponential function +http://portal.acm.org/citation.cfm?id=6498 + + +Roman E. Maeder: Storage allocation for the Karatsuba integer multiplication +algorithm. http://www.springerlink.com/content/w15058mj6v59t565/ + + +J. M. Pollard: The fast Fourier transform in a finite field +http://www.ams.org/journals/mcom/1971-25-114/S0025-5718-1971-0301966-0/home.html + + +David H. Bailey: FFTs in External or Hierarchical Memory +http://crd.lbl.gov/~dhbailey/dhbpapers/ + + +W. Morven Gentleman: Matrix Multiplication and Fast Fourier Transforms +http://www.alcatel-lucent.com/bstj/vol47-1968/articles/bstj47-6-1099.pdf + + +Mikko Tommila: Apfloat documentation +http://www.apfloat.org/apfloat/2.41/apfloat.pdf + + +Joerg Arndt: "Matters Computational" +http://www.jjj.de/fxt/ + + +Karl Hasselstrom: Fast Division of Large Integers +www.treskal.com/kalle/exjobb/original-report.pdf + + + diff --git a/Modules/_decimal/libmpdec/literature/bignum.txt b/Modules/_decimal/libmpdec/literature/bignum.txt new file mode 100644 index 0000000000..8a8731da9b --- /dev/null +++ b/Modules/_decimal/libmpdec/literature/bignum.txt @@ -0,0 +1,83 @@ + + +Bignum support (Fast Number Theoretic Transform or FNT): +======================================================== + +Bignum arithmetic in libmpdec uses the scheme for fast convolution +of integer sequences from: + +J. M. Pollard: The fast Fourier transform in a finite field +http://www.ams.org/journals/mcom/1971-25-114/S0025-5718-1971-0301966-0/home.html + + +The transform in a finite field can be used for convolution in the same +way as the Fourier Transform. The main advantages of the Number Theoretic +Transform are that it is both exact and very memory efficient. + + +Convolution in pseudo-code: +~~~~~~~~~~~~~~~~~~~~~~~~~~~ + + fnt_convolute(a, b): + x = fnt(a) # forward transform of a + y = fnt(b) # forward transform of b + z = pairwise multiply x[i] and y[i] + result = inv_fnt(z) # backward transform of z. + + +Extending the maximum transform length (Chinese Remainder Theorem): +------------------------------------------------------------------- + +The maximum transform length is quite limited when using a single +prime field. However, it is possible to use multiple primes and +recover the result using the Chinese Remainder Theorem. + + +Multiplication in pseudo-code: +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + + _mpd_fntmul(u, v): + c1 = fnt_convolute(u, v, P1) # convolute modulo prime1 + c2 = fnt_convolute(u, v, P2) # convolute modulo prime2 + c3 = fnt_convolute(u, v, P3) # convolute modulo prime3 + result = crt3(c1, c2, c3) # Chinese Remainder Theorem + + +Optimized transform functions: +------------------------------ + +There are three different fnt() functions: + + std_fnt: "standard" decimation in frequency transform for array lengths + of 2**n. Performs well up to 1024 words. + + sixstep: Cache-friendly algorithm for array lengths of 2**n. Outperforms + std_fnt for large arrays. + + fourstep: Algorithm for array lengths of 3 * 2**n. Also cache friendly + in large parts. + + +List of bignum-only files: +-------------------------- + +Functions from these files are only used in _mpd_fntmul(). + + umodarith.h -> fast low level routines for unsigned modular arithmetic + numbertheory.c -> routines for setting up the FNT + difradix2.c -> decimation in frequency transform, used as the + "base case" by the following three files: + + fnt.c -> standard transform for smaller arrays + sixstep.c -> transform large arrays of length 2**n + fourstep.c -> transform arrays of length 3 * 2**n + + convolute.c -> do the actual fast convolution, using one of + the three transform functions. + transpose.c -> transpositions needed for the sixstep algorithm. + crt.c -> Chinese Remainder Theorem: use information from three + transforms modulo three different primes to get the + final result. + + + diff --git a/Modules/_decimal/libmpdec/literature/fnt.py b/Modules/_decimal/libmpdec/literature/fnt.py new file mode 100644 index 0000000000..bf937459f5 --- /dev/null +++ b/Modules/_decimal/libmpdec/literature/fnt.py @@ -0,0 +1,208 @@ +# +# Copyright (c) 2008-2012 Stefan Krah. All rights reserved. +# +# Redistribution and use in source and binary forms, with or without +# modification, are permitted provided that the following conditions +# are met: +# +# 1. Redistributions of source code must retain the above copyright +# notice, this list of conditions and the following disclaimer. +# +# 2. Redistributions in binary form must reproduce the above copyright +# notice, this list of conditions and the following disclaimer in the +# documentation and/or other materials provided with the distribution. +# +# THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND +# ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +# ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE +# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL +# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS +# OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) +# HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +# LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY +# OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF +# SUCH DAMAGE. +# + + +###################################################################### +# This file lists and checks some of the constants and limits used # +# in libmpdec's Number Theoretic Transform. At the end of the file # +# there is an example function for the plain DFT transform. # +###################################################################### + + +# +# Number theoretic transforms are done in subfields of F(p). P[i] +# are the primes, D[i] = P[i] - 1 are highly composite and w[i] +# are the respective primitive roots of F(p). +# +# The strategy is to convolute two coefficients modulo all three +# primes, then use the Chinese Remainder Theorem on the three +# result arrays to recover the result in the usual base RADIX +# form. +# + +# ====================================================================== +# Primitive roots +# ====================================================================== + +# +# Verify primitive roots: +# +# For a prime field, r is a primitive root if and only if for all prime +# factors f of p-1, r**((p-1)/f) =/= 1 (mod p). +# +def prod(F, E): + """Check that the factorization of P-1 is correct. F is the list of + factors of P-1, E lists the number of occurrences of each factor.""" + x = 1 + for y, z in zip(F, E): + x *= y**z + return x + +def is_primitive_root(r, p, factors, exponents): + """Check if r is a primitive root of F(p).""" + if p != prod(factors, exponents) + 1: + return False + for f in factors: + q, control = divmod(p-1, f) + if control != 0: + return False + if pow(r, q, p) == 1: + return False + return True + + +# ================================================================= +# Constants and limits for the 64-bit version +# ================================================================= + +RADIX = 10**19 + +# Primes P1, P2 and P3: +P = [2**64-2**32+1, 2**64-2**34+1, 2**64-2**40+1] + +# P-1, highly composite. The transform length d is variable and +# must divide D = P-1. Since all D are divisible by 3 * 2**32, +# transform lengths can be 2**n or 3 * 2**n (where n <= 32). +D = [2**32 * 3 * (5 * 17 * 257 * 65537), + 2**34 * 3**2 * (7 * 11 * 31 * 151 * 331), + 2**40 * 3**2 * (5 * 7 * 13 * 17 * 241)] + +# Prime factors of P-1 and their exponents: +F = [(2,3,5,17,257,65537), (2,3,7,11,31,151,331), (2,3,5,7,13,17,241)] +E = [(32,1,1,1,1,1), (34,2,1,1,1,1,1), (40,2,1,1,1,1,1)] + +# Maximum transform length for 2**n. Above that only 3 * 2**31 +# or 3 * 2**32 are possible. +MPD_MAXTRANSFORM_2N = 2**32 + + +# Limits in the terminology of Pollard's paper: +m2 = (MPD_MAXTRANSFORM_2N * 3) // 2 # Maximum length of the smaller array. +M1 = M2 = RADIX-1 # Maximum value per single word. +L = m2 * M1 * M2 +P[0] * P[1] * P[2] > 2 * L + + +# Primitive roots of F(P1), F(P2) and F(P3): +w = [7, 10, 19] + +# The primitive roots are correct: +for i in range(3): + if not is_primitive_root(w[i], P[i], F[i], E[i]): + print("FAIL") + + +# ================================================================= +# Constants and limits for the 32-bit version +# ================================================================= + +RADIX = 10**9 + +# Primes P1, P2 and P3: +P = [2113929217, 2013265921, 1811939329] + +# P-1, highly composite. All D = P-1 are divisible by 3 * 2**25, +# allowing for transform lengths up to 3 * 2**25 words. +D = [2**25 * 3**2 * 7, + 2**27 * 3 * 5, + 2**26 * 3**3] + +# Prime factors of P-1 and their exponents: +F = [(2,3,7), (2,3,5), (2,3)] +E = [(25,2,1), (27,1,1), (26,3)] + +# Maximum transform length for 2**n. Above that only 3 * 2**24 or +# 3 * 2**25 are possible. +MPD_MAXTRANSFORM_2N = 2**25 + + +# Limits in the terminology of Pollard's paper: +m2 = (MPD_MAXTRANSFORM_2N * 3) // 2 # Maximum length of the smaller array. +M1 = M2 = RADIX-1 # Maximum value per single word. +L = m2 * M1 * M2 +P[0] * P[1] * P[2] > 2 * L + + +# Primitive roots of F(P1), F(P2) and F(P3): +w = [5, 31, 13] + +# The primitive roots are correct: +for i in range(3): + if not is_primitive_root(w[i], P[i], F[i], E[i]): + print("FAIL") + + +# ====================================================================== +# Example transform using a single prime +# ====================================================================== + +def ntt(lst, dir): + """Perform a transform on the elements of lst. len(lst) must + be 2**n or 3 * 2**n, where n <= 25. This is the slow DFT.""" + p = 2113929217 # prime + d = len(lst) # transform length + d_prime = pow(d, (p-2), p) # inverse of d + xi = (p-1)//d + w = 5 # primitive root of F(p) + r = pow(w, xi, p) # primitive root of the subfield + r_prime = pow(w, (p-1-xi), p) # inverse of r + if dir == 1: # forward transform + a = lst # input array + A = [0] * d # transformed values + for i in range(d): + s = 0 + for j in range(d): + s += a[j] * pow(r, i*j, p) + A[i] = s % p + return A + elif dir == -1: # backward transform + A = lst # input array + a = [0] * d # transformed values + for j in range(d): + s = 0 + for i in range(d): + s += A[i] * pow(r_prime, i*j, p) + a[j] = (d_prime * s) % p + return a + +def ntt_convolute(a, b): + """convolute arrays a and b.""" + assert(len(a) == len(b)) + x = ntt(a, 1) + y = ntt(b, 1) + for i in range(len(a)): + y[i] = y[i] * x[i] + r = ntt(y, -1) + return r + + +# Example: Two arrays representing 21 and 81 in little-endian: +a = [1, 2, 0, 0] +b = [1, 8, 0, 0] + +assert(ntt_convolute(a, b) == [1, 10, 16, 0]) +assert(21 * 81 == (1*10**0 + 10*10**1 + 16*10**2 + 0*10**3)) diff --git a/Modules/_decimal/libmpdec/literature/matrix-transform.txt b/Modules/_decimal/libmpdec/literature/matrix-transform.txt new file mode 100644 index 0000000000..ff6219810e --- /dev/null +++ b/Modules/_decimal/libmpdec/literature/matrix-transform.txt @@ -0,0 +1,256 @@ + + +(* Copyright (c) 2011 Stefan Krah. All rights reserved. *) + + +The Matrix Fourier Transform: +============================= + +In libmpdec, the Matrix Fourier Transform [1] is called four-step transform +after a variant that appears in [2]. The algorithm requires that the input +array can be viewed as an R*C matrix. + +All operations are done modulo p. For readability, the proofs drop all +instances of (mod p). + + +Algorithm four-step (forward transform): +---------------------------------------- + + a := input array + d := len(a) = R * C + p := prime + w := primitive root of unity of the prime field + r := w**((p-1)/d) + A := output array + + 1) Apply a length R FNT to each column. + + 2) Multiply each matrix element (addressed by j*C+m) by r**(j*m). + + 3) Apply a length C FNT to each row. + + 4) Transpose the matrix. + + +Proof (forward transform): +-------------------------- + + The algorithm can be derived starting from the regular definition of + the finite-field transform of length d: + + d-1 + ,---- + \ + A[k] = | a[l] * r**(k * l) + / + `---- + l = 0 + + + The sum can be rearranged into the sum of the sums of columns: + + C-1 R-1 + ,---- ,---- + \ \ + = | | a[i * C + j] * r**(k * (i * C + j)) + / / + `---- `---- + j = 0 i = 0 + + + Extracting a constant from the inner sum: + + C-1 R-1 + ,---- ,---- + \ \ + = | r**k*j * | a[i * C + j] * r**(k * i * C) + / / + `---- `---- + j = 0 i = 0 + + + Without any loss of generality, let k = n * R + m, + where n < C and m < R: + + C-1 R-1 + ,---- ,---- + \ \ + A[n*R+m] = | r**(R*n*j) * r**(m*j) * | a[i*C+j] * r**(R*C*n*i) * r**(C*m*i) + / / + `---- `---- + j = 0 i = 0 + + + Since r = w ** ((p-1) / (R*C)): + + a) r**(R*C*n*i) = w**((p-1)*n*i) = 1 + + b) r**(C*m*i) = w**((p-1) / R) ** (m*i) = r_R ** (m*i) + + c) r**(R*n*j) = w**((p-1) / C) ** (n*j) = r_C ** (n*j) + + r_R := root of the subfield of length R. + r_C := root of the subfield of length C. + + + C-1 R-1 + ,---- ,---- + \ \ + A[n*R+m] = | r_C**(n*j) * [ r**(m*j) * | a[i*C+j] * r_R**(m*i) ] + / ^ / + `---- | `---- 1) transform the columns + j = 0 | i = 0 + ^ | + | `-- 2) multiply + | + `-- 3) transform the rows + + + Note that the entire RHS is a function of n and m and that the results + for each pair (n, m) are stored in Fortran order. + + Let the term in square brackets be f(m, j). Step 1) and 2) precalculate + the term for all (m, j). After that, the original matrix is now a lookup + table with the mth element in the jth column at location m * C + j. + + Let the complete RHS be g(m, n). Step 3) does an in-place transform of + length n on all rows. After that, the original matrix is now a lookup + table with the mth element in the nth column at location m * C + n. + + But each (m, n) pair should be written to location n * R + m. Therefore, + step 4) transposes the result of step 3). + + + +Algorithm four-step (inverse transform): +---------------------------------------- + + A := input array + d := len(A) = R * C + p := prime + d' := d**(p-2) # inverse of d + w := primitive root of unity of the prime field + r := w**((p-1)/d) # root of the subfield + r' := w**((p-1) - (p-1)/d) # inverse of r + a := output array + + 0) View the matrix as a C*R matrix. + + 1) Transpose the matrix, producing an R*C matrix. + + 2) Apply a length C FNT to each row. + + 3) Multiply each matrix element (addressed by i*C+n) by r**(i*n). + + 4) Apply a length R FNT to each column. + + +Proof (inverse transform): +-------------------------- + + The algorithm can be derived starting from the regular definition of + the finite-field inverse transform of length d: + + d-1 + ,---- + \ + a[k] = d' * | A[l] * r' ** (k * l) + / + `---- + l = 0 + + + The sum can be rearranged into the sum of the sums of columns. Note + that at this stage we still have a C*R matrix, so C denotes the number + of rows: + + R-1 C-1 + ,---- ,---- + \ \ + = d' * | | a[j * R + i] * r' ** (k * (j * R + i)) + / / + `---- `---- + i = 0 j = 0 + + + Extracting a constant from the inner sum: + + R-1 C-1 + ,---- ,---- + \ \ + = d' * | r' ** (k*i) * | a[j * R + i] * r' ** (k * j * R) + / / + `---- `---- + i = 0 j = 0 + + + Without any loss of generality, let k = m * C + n, + where m < R and n < C: + + R-1 C-1 + ,---- ,---- + \ \ + A[m*C+n] = d' * | r' ** (C*m*i) * r' ** (n*i) * | a[j*R+i] * r' ** (R*C*m*j) * r' ** (R*n*j) + / / + `---- `---- + i = 0 j = 0 + + + Since r' = w**((p-1) - (p-1)/d) and d = R*C: + + a) r' ** (R*C*m*j) = w**((p-1)*R*C*m*j - (p-1)*m*j) = 1 + + b) r' ** (C*m*i) = w**((p-1)*C - (p-1)/R) ** (m*i) = r_R' ** (m*i) + + c) r' ** (R*n*j) = r_C' ** (n*j) + + d) d' = d**(p-2) = (R*C) ** (p-2) = R**(p-2) * C**(p-2) = R' * C' + + r_R' := inverse of the root of the subfield of length R. + r_C' := inverse of the root of the subfield of length C. + R' := inverse of R + C' := inverse of C + + + R-1 C-1 + ,---- ,---- 2) transform the rows of a^T + \ \ + A[m*C+n] = R' * | r_R' ** (m*i) * [ r' ** (n*i) * C' * | a[j*R+i] * r_C' ** (n*j) ] + / ^ / ^ + `---- | `---- | + i = 0 | j = 0 | + ^ | `-- 1) Transpose input matrix + | `-- 3) multiply to address elements by + | i * C + j + `-- 3) transform the columns + + + + Note that the entire RHS is a function of m and n and that the results + for each pair (m, n) are stored in C order. + + Let the term in square brackets be f(n, i). Without step 1), the sum + would perform a length C transform on the columns of the input matrix. + This is a) inefficient and b) the results are needed in C order, so + step 1) exchanges rows and columns. + + Step 2) and 3) precalculate f(n, i) for all (n, i). After that, the + original matrix is now a lookup table with the ith element in the nth + column at location i * C + n. + + Let the complete RHS be g(m, n). Step 4) does an in-place transform of + length m on all columns. After that, the original matrix is now a lookup + table with the mth element in the nth column at location m * C + n, + which means that all A[k] = A[m * C + n] are in the correct order. + + +-- + + [1] Joerg Arndt: "Matters Computational" + http://www.jjj.de/fxt/ + [2] David H. Bailey: FFTs in External or Hierarchical Memory + http://crd.lbl.gov/~dhbailey/dhbpapers/ + + + diff --git a/Modules/_decimal/libmpdec/literature/mulmod-64.txt b/Modules/_decimal/libmpdec/literature/mulmod-64.txt new file mode 100644 index 0000000000..93bf22e9fe --- /dev/null +++ b/Modules/_decimal/libmpdec/literature/mulmod-64.txt @@ -0,0 +1,127 @@ + + +(* Copyright (c) 2011 Stefan Krah. All rights reserved. *) + + +========================================================================== + Calculate (a * b) % p using special primes +========================================================================== + +A description of the algorithm can be found in the apfloat manual by +Tommila [1]. + + +Definitions: +------------ + +In the whole document, "==" stands for "is congruent with". + +Result of a * b in terms of high/low words: + + (1) hi * 2**64 + lo = a * b + +Special primes: + + (2) p = 2**64 - z + 1, where z = 2**n + +Single step modular reduction: + + (3) R(hi, lo) = hi * z - hi + lo + + +Strategy: +--------- + + a) Set (hi, lo) to the result of a * b. + + b) Set (hi', lo') to the result of R(hi, lo). + + c) Repeat step b) until 0 <= hi' * 2**64 + lo' < 2*p. + + d) If the result is less than p, return lo'. Otherwise return lo' - p. + + +The reduction step b) preserves congruence: +------------------------------------------- + + hi * 2**64 + lo == hi * z - hi + lo (mod p) + + Proof: + ~~~~~~ + + hi * 2**64 + lo = (2**64 - z + 1) * hi + z * hi - hi + lo + + = p * hi + z * hi - hi + lo + + == z * hi - hi + lo (mod p) + + +Maximum numbers of step b): +--------------------------- + +# To avoid unneccessary formalism, define: + +def R(hi, lo, z): + return divmod(hi * z - hi + lo, 2**64) + +# For simplicity, assume hi=2**64-1, lo=2**64-1 after the +# initial multiplication a * b. This is of course impossible +# but certainly covers all cases. + +# Then, for p1: +hi=2**64-1; lo=2**64-1; z=2**32 +p1 = 2**64 - z + 1 + +hi, lo = R(hi, lo, z) # First reduction +hi, lo = R(hi, lo, z) # Second reduction +hi * 2**64 + lo < 2 * p1 # True + +# For p2: +hi=2**64-1; lo=2**64-1; z=2**34 +p2 = 2**64 - z + 1 + +hi, lo = R(hi, lo, z) # First reduction +hi, lo = R(hi, lo, z) # Second reduction +hi, lo = R(hi, lo, z) # Third reduction +hi * 2**64 + lo < 2 * p2 # True + +# For p3: +hi=2**64-1; lo=2**64-1; z=2**40 +p3 = 2**64 - z + 1 + +hi, lo = R(hi, lo, z) # First reduction +hi, lo = R(hi, lo, z) # Second reduction +hi, lo = R(hi, lo, z) # Third reduction +hi * 2**64 + lo < 2 * p3 # True + + +Step d) preserves congruence and yields a result < p: +----------------------------------------------------- + + Case hi = 0: + + Case lo < p: trivial. + + Case lo >= p: + + lo == lo - p (mod p) # result is congruent + + p <= lo < 2*p -> 0 <= lo - p < p # result is in the correct range + + Case hi = 1: + + p < 2**64 /\ 2**64 + lo < 2*p -> lo < p # lo is always less than p + + 2**64 + lo == 2**64 + (lo - p) (mod p) # result is congruent + + = lo - p # exactly the same value as the previous RHS + # in uint64_t arithmetic. + + p < 2**64 + lo < 2*p -> 0 < 2**64 + (lo - p) < p # correct range + + + +[1] http://www.apfloat.org/apfloat/2.40/apfloat.pdf + + + diff --git a/Modules/_decimal/libmpdec/literature/mulmod-ppro.txt b/Modules/_decimal/libmpdec/literature/mulmod-ppro.txt new file mode 100644 index 0000000000..43e6e4e4ba --- /dev/null +++ b/Modules/_decimal/libmpdec/literature/mulmod-ppro.txt @@ -0,0 +1,269 @@ + + +(* Copyright (c) 2011 Stefan Krah. All rights reserved. *) + + +======================================================================== + Calculate (a * b) % p using the 80-bit x87 FPU +======================================================================== + +A description of the algorithm can be found in the apfloat manual by +Tommila [1]. + +The proof follows an argument made by Granlund/Montgomery in [2]. + + +Definitions and assumptions: +---------------------------- + +The 80-bit extended precision format uses 64 bits for the significand: + + (1) F = 64 + +The modulus is prime and less than 2**31: + + (2) 2 <= p < 2**31 + +The factors are less than p: + + (3) 0 <= a < p + (4) 0 <= b < p + +The product a * b is less than 2**62 and is thus exact in 64 bits: + + (5) n = a * b + +The product can be represented in terms of quotient and remainder: + + (6) n = q * p + r + +Using (3), (4) and the fact that p is prime, the remainder is always +greater than zero: + + (7) 0 <= q < p /\ 1 <= r < p + + +Strategy: +--------- + +Precalculate the 80-bit long double inverse of p, with a maximum +relative error of 2**(1-F): + + (8) pinv = (long double)1.0 / p + +Calculate an estimate for q = floor(n/p). The multiplication has another +maximum relative error of 2**(1-F): + + (9) qest = n * pinv + +If we can show that q < qest < q+1, then trunc(qest) = q. It is then +easy to recover the remainder r. The complete algorithm is: + + a) Set the control word to 64-bit precision and truncation mode. + + b) n = a * b # Calculate exact product. + + c) qest = n * pinv # Calculate estimate for the quotient. + + d) q = (qest+2**63)-2**63 # Truncate qest to the exact quotient. + + f) r = n - q * p # Calculate remainder. + + +Proof for q < qest < q+1: +------------------------- + +Using the cumulative error, the error bounds for qest are: + + n n * (1 + 2**(1-F))**2 + (9) --------------------- <= qest <= --------------------- + p * (1 + 2**(1-F))**2 p + + + Lemma 1: + -------- + n q * p + r + (10) q < --------------------- = --------------------- + p * (1 + 2**(1-F))**2 p * (1 + 2**(1-F))**2 + + + Proof: + ~~~~~~ + + (I) q * p * (1 + 2**(1-F))**2 < q * p + r + + (II) q * p * 2**(2-F) + q * p * 2**(2-2*F) < r + + Using (1) and (7), it is sufficient to show that: + + (III) q * p * 2**(-62) + q * p * 2**(-126) < 1 <= r + + (III) can easily be verified by substituting the largest possible + values p = 2**31-1 and q = 2**31-2. + + The critical cases occur when r = 1, n = m * p + 1. These cases + can be exhaustively verified with a test program. + + + Lemma 2: + -------- + + n * (1 + 2**(1-F))**2 (q * p + r) * (1 + 2**(1-F))**2 + (11) --------------------- = ------------------------------- < q + 1 + p p + + Proof: + ~~~~~~ + + (I) (q * p + r) + (q * p + r) * 2**(2-F) + (q * p + r) * 2**(2-2*F) < q * p + p + + (II) (q * p + r) * 2**(2-F) + (q * p + r) * 2**(2-2*F) < p - r + + Using (1) and (7), it is sufficient to show that: + + (III) (q * p + r) * 2**(-62) + (q * p + r) * 2**(-126) < 1 <= p - r + + (III) can easily be verified by substituting the largest possible + values p = 2**31-1, q = 2**31-2 and r = 2**31-2. + + The critical cases occur when r = (p - 1), n = m * p - 1. These cases + can be exhaustively verified with a test program. + + +[1] http://www.apfloat.org/apfloat/2.40/apfloat.pdf +[2] http://gmplib.org/~tege/divcnst-pldi94.pdf + [Section 7: "Use of floating point"] + + + +(* Coq proof for (10) and (11) *) + +Require Import ZArith. +Require Import QArith. +Require Import Qpower. +Require Import Qabs. +Require Import Psatz. + +Open Scope Q_scope. + + +Ltac qreduce T := + rewrite <- (Qred_correct (T)); simpl (Qred (T)). + +Theorem Qlt_move_right : + forall x y z:Q, x + z < y <-> x < y - z. +Proof. + intros. + split. + intros. + psatzl Q. + intros. + psatzl Q. +Qed. + +Theorem Qlt_mult_by_z : + forall x y z:Q, 0 < z -> (x < y <-> x * z < y * z). +Proof. + intros. + split. + intros. + apply Qmult_lt_compat_r. trivial. trivial. + intros. + rewrite <- (Qdiv_mult_l x z). rewrite <- (Qdiv_mult_l y z). + apply Qmult_lt_compat_r. + apply Qlt_shift_inv_l. + trivial. psatzl Q. trivial. psatzl Q. psatzl Q. +Qed. + +Theorem Qle_mult_quad : + forall (a b c d:Q), + 0 <= a -> a <= c -> + 0 <= b -> b <= d -> + a * b <= c * d. + intros. + psatz Q. +Qed. + + +Theorem q_lt_qest: + forall (p q r:Q), + (0 < p) -> (p <= (2#1)^31 - 1) -> + (0 <= q) -> (q <= p - 1) -> + (1 <= r) -> (r <= p - 1) -> + q < (q * p + r) / (p * (1 + (2#1)^(-63))^2). +Proof. + intros. + rewrite Qlt_mult_by_z with (z := (p * (1 + (2#1)^(-63))^2)). + + unfold Qdiv. + rewrite <- Qmult_assoc. + rewrite (Qmult_comm (/ (p * (1 + (2 # 1) ^ (-63)) ^ 2)) (p * (1 + (2 # 1) ^ (-63)) ^ 2)). + rewrite Qmult_inv_r. + rewrite Qmult_1_r. + + assert (q * (p * (1 + (2 # 1) ^ (-63)) ^ 2) == q * p + (q * p) * ((2 # 1) ^ (-62) + (2 # 1) ^ (-126))). + qreduce ((1 + (2 # 1) ^ (-63)) ^ 2). + qreduce ((2 # 1) ^ (-62) + (2 # 1) ^ (-126)). + ring_simplify. + reflexivity. + rewrite H5. + + rewrite Qplus_comm. + rewrite Qlt_move_right. + ring_simplify (q * p + r - q * p). + qreduce ((2 # 1) ^ (-62) + (2 # 1) ^ (-126)). + + apply Qlt_le_trans with (y := 1). + rewrite Qlt_mult_by_z with (z := 85070591730234615865843651857942052864 # 18446744073709551617). + ring_simplify. + + apply Qle_lt_trans with (y := ((2 # 1) ^ 31 - (2#1)) * ((2 # 1) ^ 31 - 1)). + apply Qle_mult_quad. + assumption. psatzl Q. psatzl Q. psatzl Q. psatzl Q. psatzl Q. assumption. psatzl Q. psatzl Q. +Qed. + +Theorem qest_lt_qplus1: + forall (p q r:Q), + (0 < p) -> (p <= (2#1)^31 - 1) -> + (0 <= q) -> (q <= p - 1) -> + (1 <= r) -> (r <= p - 1) -> + ((q * p + r) * (1 + (2#1)^(-63))^2) / p < q + 1. +Proof. + intros. + rewrite Qlt_mult_by_z with (z := p). + + unfold Qdiv. + rewrite <- Qmult_assoc. + rewrite (Qmult_comm (/ p) p). + rewrite Qmult_inv_r. + rewrite Qmult_1_r. + + assert ((q * p + r) * (1 + (2 # 1) ^ (-63)) ^ 2 == q * p + r + (q * p + r) * ((2 # 1) ^ (-62) + (2 # 1) ^ (-126))). + qreduce ((1 + (2 # 1) ^ (-63)) ^ 2). + qreduce ((2 # 1) ^ (-62) + (2 # 1) ^ (-126)). + ring_simplify. reflexivity. + rewrite H5. + + rewrite <- Qplus_assoc. rewrite <- Qplus_comm. rewrite Qlt_move_right. + ring_simplify ((q + 1) * p - q * p). + + rewrite <- Qplus_comm. rewrite Qlt_move_right. + + apply Qlt_le_trans with (y := 1). + qreduce ((2 # 1) ^ (-62) + (2 # 1) ^ (-126)). + + rewrite Qlt_mult_by_z with (z := 85070591730234615865843651857942052864 # 18446744073709551617). + ring_simplify. + + ring_simplify in H0. + apply Qle_lt_trans with (y := (2147483646 # 1) * (2147483647 # 1) + (2147483646 # 1)). + + apply Qplus_le_compat. + apply Qle_mult_quad. + assumption. psatzl Q. auto with qarith. assumption. psatzl Q. + auto with qarith. auto with qarith. + psatzl Q. psatzl Q. assumption. +Qed. + + + diff --git a/Modules/_decimal/libmpdec/literature/six-step.txt b/Modules/_decimal/libmpdec/literature/six-step.txt new file mode 100644 index 0000000000..759147ff1c --- /dev/null +++ b/Modules/_decimal/libmpdec/literature/six-step.txt @@ -0,0 +1,63 @@ + + +(* Copyright (c) 2011 Stefan Krah. All rights reserved. *) + + +The Six Step Transform: +======================= + +In libmpdec, the six-step transform is the Matrix Fourier Transform (See +matrix-transform.txt) in disguise. It is called six-step transform after +a variant that appears in [1]. The algorithm requires that the input +array can be viewed as an R*C matrix. + + +Algorithm six-step (forward transform): +--------------------------------------- + + 1a) Transpose the matrix. + + 1b) Apply a length R FNT to each row. + + 1c) Transpose the matrix. + + 2) Multiply each matrix element (addressed by j*C+m) by r**(j*m). + + 3) Apply a length C FNT to each row. + + 4) Transpose the matrix. + +Note that steps 1a) - 1c) are exactly equivalent to step 1) of the Matrix +Fourier Transform. For large R, it is faster to transpose twice and do +a transform on the rows than to perform a column transpose directly. + + + +Algorithm six-step (inverse transform): +--------------------------------------- + + 0) View the matrix as a C*R matrix. + + 1) Transpose the matrix, producing an R*C matrix. + + 2) Apply a length C FNT to each row. + + 3) Multiply each matrix element (addressed by i*C+n) by r**(i*n). + + 4a) Transpose the matrix. + + 4b) Apply a length R FNT to each row. + + 4c) Transpose the matrix. + +Again, steps 4a) - 4c) are equivalent to step 4) of the Matrix Fourier +Transform. + + + +-- + + [1] David H. Bailey: FFTs in External or Hierarchical Memory + http://crd.lbl.gov/~dhbailey/dhbpapers/ + + diff --git a/Modules/_decimal/libmpdec/literature/umodarith.lisp b/Modules/_decimal/libmpdec/literature/umodarith.lisp new file mode 100644 index 0000000000..60a14a4e56 --- /dev/null +++ b/Modules/_decimal/libmpdec/literature/umodarith.lisp @@ -0,0 +1,692 @@ +; +; Copyright (c) 2008-2012 Stefan Krah. All rights reserved. +; +; Redistribution and use in source and binary forms, with or without +; modification, are permitted provided that the following conditions +; are met: +; +; 1. Redistributions of source code must retain the above copyright +; notice, this list of conditions and the following disclaimer. +; +; 2. Redistributions in binary form must reproduce the above copyright +; notice, this list of conditions and the following disclaimer in the +; documentation and/or other materials provided with the distribution. +; +; THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND +; ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +; IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +; ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE +; FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL +; DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS +; OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) +; HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +; LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY +; OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF +; SUCH DAMAGE. +; + + +(in-package "ACL2") + +(include-book "arithmetic/top-with-meta" :dir :system) +(include-book "arithmetic-2/floor-mod/floor-mod" :dir :system) + + +;; ===================================================================== +;; Proofs for several functions in umodarith.h +;; ===================================================================== + + + +;; ===================================================================== +;; Helper theorems +;; ===================================================================== + +(defthm elim-mod-m<x<2*m + (implies (and (<= m x) + (< x (* 2 m)) + (rationalp x) (rationalp m)) + (equal (mod x m) + (+ x (- m))))) + +(defthm modaux-1a + (implies (and (< x m) (< 0 x) (< 0 m) + (rationalp x) (rationalp m)) + (equal (mod (- x) m) + (+ (- x) m)))) + +(defthm modaux-1b + (implies (and (< (- x) m) (< x 0) (< 0 m) + (rationalp x) (rationalp m)) + (equal (mod x m) + (+ x m))) + :hints (("Goal" :use ((:instance modaux-1a + (x (- x))))))) + +(defthm modaux-1c + (implies (and (< x m) (< 0 x) (< 0 m) + (rationalp x) (rationalp m)) + (equal (mod x m) + x))) + +(defthm modaux-2a + (implies (and (< 0 b) (< b m) + (natp x) (natp b) (natp m) + (< (mod (+ b x) m) b)) + (equal (mod (+ (- m) b x) m) + (+ (- m) b (mod x m))))) + +(defthm modaux-2b + (implies (and (< 0 b) (< b m) + (natp x) (natp b) (natp m) + (< (mod (+ b x) m) b)) + (equal (mod (+ b x) m) + (+ (- m) b (mod x m)))) + :hints (("Goal" :use (modaux-2a)))) + +(defthm linear-mod-1 + (implies (and (< x m) (< b m) + (natp x) (natp b) + (rationalp m)) + (equal (< x (mod (+ (- b) x) m)) + (< x b))) + :hints (("Goal" :use ((:instance modaux-1a + (x (+ b (- x)))))))) + +(defthm linear-mod-2 + (implies (and (< 0 b) (< b m) + (natp x) (natp b) + (natp m)) + (equal (< (mod x m) + (mod (+ (- b) x) m)) + (< (mod x m) b)))) + +(defthm linear-mod-3 + (implies (and (< x m) (< b m) + (natp x) (natp b) + (rationalp m)) + (equal (<= b (mod (+ b x) m)) + (< (+ b x) m))) + :hints (("Goal" :use ((:instance elim-mod-m<x<2*m + (x (+ b x))))))) + +(defthm modaux-2c + (implies (and (< 0 b) (< b m) + (natp x) (natp b) (natp m) + (<= b (mod (+ b x) m))) + (equal (mod (+ b x) m) + (+ b (mod x m)))) + :hints (("Subgoal *1/8''" :use (linear-mod-3)))) + +(defthmd modaux-2d + (implies (and (< x m) (< 0 x) (< 0 m) + (< (- m) b) (< b 0) (rationalp m) + (<= x (mod (+ b x) m)) + (rationalp x) (rationalp b)) + (equal (+ (- m) (mod (+ b x) m)) + (+ b x))) + :hints (("Goal" :cases ((<= 0 (+ b x)))) + ("Subgoal 2'" :use ((:instance modaux-1b + (x (+ b x))))))) + +(defthm mod-m-b + (implies (and (< 0 x) (< 0 b) (< 0 m) + (< x b) (< b m) + (natp x) (natp b) (natp m)) + (equal (mod (+ (mod (- x) m) b) m) + (mod (- x) b)))) + + +;; ===================================================================== +;; addmod, submod +;; ===================================================================== + +(defun addmod (a b m base) + (let* ((s (mod (+ a b) base)) + (s (if (< s a) (mod (- s m) base) s)) + (s (if (>= s m) (mod (- s m) base) s))) + s)) + +(defthmd addmod-correct + (implies (and (< 0 m) (< m base) + (< a m) (<= b m) + (natp m) (natp base) + (natp a) (natp b)) + (equal (addmod a b m base) + (mod (+ a b) m))) + :hints (("Goal" :cases ((<= base (+ a b)))) + ("Subgoal 2.1'" :use ((:instance elim-mod-m<x<2*m + (x (+ a b))))))) + +(defun submod (a b m base) + (let* ((d (mod (- a b) base)) + (d (if (< a d) (mod (+ d m) base) d))) + d)) + +(defthmd submod-aux1 + (implies (and (< a (mod (+ a (- b)) base)) + (< 0 base) (< a base) (<= b base) + (natp base) (natp a) (natp b)) + (< a b)) + :rule-classes :forward-chaining) + +(defthmd submod-aux2 + (implies (and (<= (mod (+ a (- b)) base) a) + (< 0 base) (< a base) (< b base) + (natp base) (natp a) (natp b)) + (<= b a)) + :rule-classes :forward-chaining) + +(defthmd submod-correct + (implies (and (< 0 m) (< m base) + (< a m) (<= b m) + (natp m) (natp base) + (natp a) (natp b)) + (equal (submod a b m base) + (mod (- a b) m))) + :hints (("Goal" :cases ((<= base (+ a b)))) + ("Subgoal 2.2" :use ((:instance submod-aux1))) + ("Subgoal 2.2'''" :cases ((and (< 0 (+ a (- b) m)) + (< (+ a (- b) m) m)))) + ("Subgoal 2.1" :use ((:instance submod-aux2))) + ("Subgoal 1.2" :use ((:instance submod-aux1))) + ("Subgoal 1.1" :use ((:instance submod-aux2))))) + + +(defun submod-2 (a b m base) + (let* ((d (mod (- a b) base)) + (d (if (< a b) (mod (+ d m) base) d))) + d)) + +(defthm submod-2-correct + (implies (and (< 0 m) (< m base) + (< a m) (<= b m) + (natp m) (natp base) + (natp a) (natp b)) + (equal (submod-2 a b m base) + (mod (- a b) m))) + :hints (("Subgoal 2'" :cases ((and (< 0 (+ a (- b) m)) + (< (+ a (- b) m) m)))))) + + +;; ========================================================================= +;; ext-submod is correct +;; ========================================================================= + +; a < 2*m, b < 2*m +(defun ext-submod (a b m base) + (let* ((a (if (>= a m) (- a m) a)) + (b (if (>= b m) (- b m) b)) + (d (mod (- a b) base)) + (d (if (< a b) (mod (+ d m) base) d))) + d)) + +; a < 2*m, b < 2*m +(defun ext-submod-2 (a b m base) + (let* ((a (mod a m)) + (b (mod b m)) + (d (mod (- a b) base)) + (d (if (< a b) (mod (+ d m) base) d))) + d)) + +(defthmd ext-submod-ext-submod-2-equal + (implies (and (< 0 m) (< m base) + (< a (* 2 m)) (< b (* 2 m)) + (natp m) (natp base) + (natp a) (natp b)) + (equal (ext-submod a b m base) + (ext-submod-2 a b m base)))) + +(defthmd ext-submod-2-correct + (implies (and (< 0 m) (< m base) + (< a (* 2 m)) (< b (* 2 m)) + (natp m) (natp base) + (natp a) (natp b)) + (equal (ext-submod-2 a b m base) + (mod (- a b) m)))) + + +;; ========================================================================= +;; dw-reduce is correct +;; ========================================================================= + +(defun dw-reduce (hi lo m base) + (let* ((r1 (mod hi m)) + (r2 (mod (+ (* r1 base) lo) m))) + r2)) + +(defthmd dw-reduce-correct + (implies (and (< 0 m) (< m base) + (< hi base) (< lo base) + (natp m) (natp base) + (natp hi) (natp lo)) + (equal (dw-reduce hi lo m base) + (mod (+ (* hi base) lo) m)))) + +(defthmd <=-multiply-both-sides-by-z + (implies (and (rationalp x) (rationalp y) + (< 0 z) (rationalp z)) + (equal (<= x y) + (<= (* z x) (* z y))))) + +(defthmd dw-reduce-aux1 + (implies (and (< 0 m) (< m base) + (natp m) (natp base) + (< lo base) (natp lo) + (< x m) (natp x)) + (< (+ lo (* base x)) (* base m))) + :hints (("Goal" :cases ((<= (+ x 1) m))) + ("Subgoal 1''" :cases ((<= (* base (+ x 1)) (* base m)))) + ("subgoal 1.2" :use ((:instance <=-multiply-both-sides-by-z + (x (+ 1 x)) + (y m) + (z base)))))) + +(defthm dw-reduce-aux2 + (implies (and (< x (* base m)) + (< 0 m) (< m base) + (natp m) (natp base) (natp x)) + (< (floor x m) base))) + +;; This is the necessary condition for using _mpd_div_words(). +(defthmd dw-reduce-second-quotient-fits-in-single-word + (implies (and (< 0 m) (< m base) + (< hi base) (< lo base) + (natp m) (natp base) + (natp hi) (natp lo) + (equal r1 (mod hi m))) + (< (floor (+ (* r1 base) lo) m) + base)) + :hints (("Goal" :cases ((< r1 m))) + ("Subgoal 1''" :cases ((< (+ lo (* base (mod hi m))) (* base m)))) + ("Subgoal 1.2" :use ((:instance dw-reduce-aux1 + (x (mod hi m))))))) + + +;; ========================================================================= +;; dw-submod is correct +;; ========================================================================= + +(defun dw-submod (a hi lo m base) + (let* ((r (dw-reduce hi lo m base)) + (d (mod (- a r) base)) + (d (if (< a r) (mod (+ d m) base) d))) + d)) + +(defthmd dw-submod-aux1 + (implies (and (natp a) (< 0 m) (natp m) + (natp x) (equal r (mod x m))) + (equal (mod (- a x) m) + (mod (- a r) m)))) + +(defthmd dw-submod-correct + (implies (and (< 0 m) (< m base) + (natp a) (< a m) + (< hi base) (< lo base) + (natp m) (natp base) + (natp hi) (natp lo)) + (equal (dw-submod a hi lo m base) + (mod (- a (+ (* base hi) lo)) m))) + :hints (("Goal" :in-theory (disable dw-reduce) + :use ((:instance dw-submod-aux1 + (x (+ lo (* base hi))) + (r (dw-reduce hi lo m base))) + (:instance dw-reduce-correct))))) + + +;; ========================================================================= +;; ANSI C arithmetic for uint64_t +;; ========================================================================= + +(defun add (a b) + (mod (+ a b) + (expt 2 64))) + +(defun sub (a b) + (mod (- a b) + (expt 2 64))) + +(defun << (w n) + (mod (* w (expt 2 n)) + (expt 2 64))) + +(defun >> (w n) + (floor w (expt 2 n))) + +;; join upper and lower half of a double word, yielding a 128 bit number +(defun join (hi lo) + (+ (* (expt 2 64) hi) lo)) + + +;; ============================================================================= +;; Fast modular reduction +;; ============================================================================= + +;; These are the three primes used in the Number Theoretic Transform. +;; A fast modular reduction scheme exists for all of them. +(defmacro p1 () + (+ (expt 2 64) (- (expt 2 32)) 1)) + +(defmacro p2 () + (+ (expt 2 64) (- (expt 2 34)) 1)) + +(defmacro p3 () + (+ (expt 2 64) (- (expt 2 40)) 1)) + + +;; reduce the double word number hi*2**64 + lo (mod p1) +(defun simple-mod-reduce-p1 (hi lo) + (+ (* (expt 2 32) hi) (- hi) lo)) + +;; reduce the double word number hi*2**64 + lo (mod p2) +(defun simple-mod-reduce-p2 (hi lo) + (+ (* (expt 2 34) hi) (- hi) lo)) + +;; reduce the double word number hi*2**64 + lo (mod p3) +(defun simple-mod-reduce-p3 (hi lo) + (+ (* (expt 2 40) hi) (- hi) lo)) + + +; ---------------------------------------------------------- +; The modular reductions given above are correct +; ---------------------------------------------------------- + +(defthmd congruence-p1-aux + (equal (* (expt 2 64) hi) + (+ (* (p1) hi) + (* (expt 2 32) hi) + (- hi)))) + +(defthmd congruence-p2-aux + (equal (* (expt 2 64) hi) + (+ (* (p2) hi) + (* (expt 2 34) hi) + (- hi)))) + +(defthmd congruence-p3-aux + (equal (* (expt 2 64) hi) + (+ (* (p3) hi) + (* (expt 2 40) hi) + (- hi)))) + +(defthmd mod-augment + (implies (and (rationalp x) + (rationalp y) + (rationalp m)) + (equal (mod (+ x y) m) + (mod (+ x (mod y m)) m)))) + +(defthmd simple-mod-reduce-p1-congruent + (implies (and (integerp hi) + (integerp lo)) + (equal (mod (simple-mod-reduce-p1 hi lo) (p1)) + (mod (join hi lo) (p1)))) + :hints (("Goal''" :use ((:instance congruence-p1-aux) + (:instance mod-augment + (m (p1)) + (x (+ (- hi) lo (* (expt 2 32) hi))) + (y (* (p1) hi))))))) + +(defthmd simple-mod-reduce-p2-congruent + (implies (and (integerp hi) + (integerp lo)) + (equal (mod (simple-mod-reduce-p2 hi lo) (p2)) + (mod (join hi lo) (p2)))) + :hints (("Goal''" :use ((:instance congruence-p2-aux) + (:instance mod-augment + (m (p2)) + (x (+ (- hi) lo (* (expt 2 34) hi))) + (y (* (p2) hi))))))) + +(defthmd simple-mod-reduce-p3-congruent + (implies (and (integerp hi) + (integerp lo)) + (equal (mod (simple-mod-reduce-p3 hi lo) (p3)) + (mod (join hi lo) (p3)))) + :hints (("Goal''" :use ((:instance congruence-p3-aux) + (:instance mod-augment + (m (p3)) + (x (+ (- hi) lo (* (expt 2 40) hi))) + (y (* (p3) hi))))))) + + +; --------------------------------------------------------------------- +; We need a number less than 2*p, so that we can use the trick from +; elim-mod-m<x<2*m for the final reduction. +; For p1, two modular reductions are sufficient, for p2 and p3 three. +; --------------------------------------------------------------------- + +;; p1: the first reduction is less than 2**96 +(defthmd simple-mod-reduce-p1-<-2**96 + (implies (and (< hi (expt 2 64)) + (< lo (expt 2 64)) + (natp hi) (natp lo)) + (< (simple-mod-reduce-p1 hi lo) + (expt 2 96)))) + +;; p1: the second reduction is less than 2*p1 +(defthmd simple-mod-reduce-p1-<-2*p1 + (implies (and (< hi (expt 2 64)) + (< lo (expt 2 64)) + (< (join hi lo) (expt 2 96)) + (natp hi) (natp lo)) + (< (simple-mod-reduce-p1 hi lo) + (* 2 (p1))))) + + +;; p2: the first reduction is less than 2**98 +(defthmd simple-mod-reduce-p2-<-2**98 + (implies (and (< hi (expt 2 64)) + (< lo (expt 2 64)) + (natp hi) (natp lo)) + (< (simple-mod-reduce-p2 hi lo) + (expt 2 98)))) + +;; p2: the second reduction is less than 2**69 +(defthmd simple-mod-reduce-p2-<-2*69 + (implies (and (< hi (expt 2 64)) + (< lo (expt 2 64)) + (< (join hi lo) (expt 2 98)) + (natp hi) (natp lo)) + (< (simple-mod-reduce-p2 hi lo) + (expt 2 69)))) + +;; p3: the third reduction is less than 2*p2 +(defthmd simple-mod-reduce-p2-<-2*p2 + (implies (and (< hi (expt 2 64)) + (< lo (expt 2 64)) + (< (join hi lo) (expt 2 69)) + (natp hi) (natp lo)) + (< (simple-mod-reduce-p2 hi lo) + (* 2 (p2))))) + + +;; p3: the first reduction is less than 2**104 +(defthmd simple-mod-reduce-p3-<-2**104 + (implies (and (< hi (expt 2 64)) + (< lo (expt 2 64)) + (natp hi) (natp lo)) + (< (simple-mod-reduce-p3 hi lo) + (expt 2 104)))) + +;; p3: the second reduction is less than 2**81 +(defthmd simple-mod-reduce-p3-<-2**81 + (implies (and (< hi (expt 2 64)) + (< lo (expt 2 64)) + (< (join hi lo) (expt 2 104)) + (natp hi) (natp lo)) + (< (simple-mod-reduce-p3 hi lo) + (expt 2 81)))) + +;; p3: the third reduction is less than 2*p3 +(defthmd simple-mod-reduce-p3-<-2*p3 + (implies (and (< hi (expt 2 64)) + (< lo (expt 2 64)) + (< (join hi lo) (expt 2 81)) + (natp hi) (natp lo)) + (< (simple-mod-reduce-p3 hi lo) + (* 2 (p3))))) + + +; ------------------------------------------------------------------------- +; The simple modular reductions, adapted for compiler friendly C +; ------------------------------------------------------------------------- + +(defun mod-reduce-p1 (hi lo) + (let* ((y hi) + (x y) + (hi (>> hi 32)) + (x (sub lo x)) + (hi (if (> x lo) (+ hi -1) hi)) + (y (<< y 32)) + (lo (add y x)) + (hi (if (< lo y) (+ hi 1) hi))) + (+ (* hi (expt 2 64)) lo))) + +(defun mod-reduce-p2 (hi lo) + (let* ((y hi) + (x y) + (hi (>> hi 30)) + (x (sub lo x)) + (hi (if (> x lo) (+ hi -1) hi)) + (y (<< y 34)) + (lo (add y x)) + (hi (if (< lo y) (+ hi 1) hi))) + (+ (* hi (expt 2 64)) lo))) + +(defun mod-reduce-p3 (hi lo) + (let* ((y hi) + (x y) + (hi (>> hi 24)) + (x (sub lo x)) + (hi (if (> x lo) (+ hi -1) hi)) + (y (<< y 40)) + (lo (add y x)) + (hi (if (< lo y) (+ hi 1) hi))) + (+ (* hi (expt 2 64)) lo))) + + +; ------------------------------------------------------------------------- +; The compiler friendly versions are equal to the simple versions +; ------------------------------------------------------------------------- + +(defthm mod-reduce-aux1 + (implies (and (<= 0 a) (natp a) (natp m) + (< (- m) b) (<= b 0) + (integerp b) + (< (mod (+ b a) m) + (mod a m))) + (equal (mod (+ b a) m) + (+ b (mod a m)))) + :hints (("Subgoal 2" :use ((:instance modaux-1b + (x (+ a b))))))) + +(defthm mod-reduce-aux2 + (implies (and (<= 0 a) (natp a) (natp m) + (< b m) (natp b) + (< (mod (+ b a) m) + (mod a m))) + (equal (+ m (mod (+ b a) m)) + (+ b (mod a m))))) + + +(defthm mod-reduce-aux3 + (implies (and (< 0 a) (natp a) (natp m) + (< (- m) b) (< b 0) + (integerp b) + (<= (mod a m) + (mod (+ b a) m))) + (equal (+ (- m) (mod (+ b a) m)) + (+ b (mod a m)))) + :hints (("Subgoal 1.2'" :use ((:instance modaux-1b + (x b)))) + ("Subgoal 1''" :use ((:instance modaux-2d + (x I)))))) + + +(defthm mod-reduce-aux4 + (implies (and (< 0 a) (natp a) (natp m) + (< b m) (natp b) + (<= (mod a m) + (mod (+ b a) m))) + (equal (mod (+ b a) m) + (+ b (mod a m))))) + + +(defthm mod-reduce-p1==simple-mod-reduce-p1 + (implies (and (< hi (expt 2 64)) + (< lo (expt 2 64)) + (natp hi) (natp lo)) + (equal (mod-reduce-p1 hi lo) + (simple-mod-reduce-p1 hi lo))) + :hints (("Goal" :in-theory (disable expt) + :cases ((< 0 hi))) + ("Subgoal 1.2.2'" :use ((:instance mod-reduce-aux1 + (m (expt 2 64)) + (b (+ (- HI) LO)) + (a (* (expt 2 32) hi))))) + ("Subgoal 1.2.1'" :use ((:instance mod-reduce-aux3 + (m (expt 2 64)) + (b (+ (- HI) LO)) + (a (* (expt 2 32) hi))))) + ("Subgoal 1.1.2'" :use ((:instance mod-reduce-aux2 + (m (expt 2 64)) + (b (+ (- HI) LO)) + (a (* (expt 2 32) hi))))) + ("Subgoal 1.1.1'" :use ((:instance mod-reduce-aux4 + (m (expt 2 64)) + (b (+ (- HI) LO)) + (a (* (expt 2 32) hi))))))) + + +(defthm mod-reduce-p2==simple-mod-reduce-p2 + (implies (and (< hi (expt 2 64)) + (< lo (expt 2 64)) + (natp hi) (natp lo)) + (equal (mod-reduce-p2 hi lo) + (simple-mod-reduce-p2 hi lo))) + :hints (("Goal" :cases ((< 0 hi))) + ("Subgoal 1.2.2'" :use ((:instance mod-reduce-aux1 + (m (expt 2 64)) + (b (+ (- HI) LO)) + (a (* (expt 2 34) hi))))) + ("Subgoal 1.2.1'" :use ((:instance mod-reduce-aux3 + (m (expt 2 64)) + (b (+ (- HI) LO)) + (a (* (expt 2 34) hi))))) + ("Subgoal 1.1.2'" :use ((:instance mod-reduce-aux2 + (m (expt 2 64)) + (b (+ (- HI) LO)) + (a (* (expt 2 34) hi))))) + ("Subgoal 1.1.1'" :use ((:instance mod-reduce-aux4 + (m (expt 2 64)) + (b (+ (- HI) LO)) + (a (* (expt 2 34) hi))))))) + + +(defthm mod-reduce-p3==simple-mod-reduce-p3 + (implies (and (< hi (expt 2 64)) + (< lo (expt 2 64)) + (natp hi) (natp lo)) + (equal (mod-reduce-p3 hi lo) + (simple-mod-reduce-p3 hi lo))) + :hints (("Goal" :cases ((< 0 hi))) + ("Subgoal 1.2.2'" :use ((:instance mod-reduce-aux1 + (m (expt 2 64)) + (b (+ (- HI) LO)) + (a (* (expt 2 40) hi))))) + ("Subgoal 1.2.1'" :use ((:instance mod-reduce-aux3 + (m (expt 2 64)) + (b (+ (- HI) LO)) + (a (* (expt 2 40) hi))))) + ("Subgoal 1.1.2'" :use ((:instance mod-reduce-aux2 + (m (expt 2 64)) + (b (+ (- HI) LO)) + (a (* (expt 2 40) hi))))) + ("Subgoal 1.1.1'" :use ((:instance mod-reduce-aux4 + (m (expt 2 64)) + (b (+ (- HI) LO)) + (a (* (expt 2 40) hi))))))) + + + diff --git a/Modules/_decimal/libmpdec/memory.c b/Modules/_decimal/libmpdec/memory.c new file mode 100644 index 0000000000..bf6350f904 --- /dev/null +++ b/Modules/_decimal/libmpdec/memory.c @@ -0,0 +1,292 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#include "mpdecimal.h" +#include <stdio.h> +#include <stdlib.h> +#include "typearith.h" +#include "memory.h" + + +/* Guaranteed minimum allocation for a coefficient. May be changed once + at program start using mpd_setminalloc(). */ +mpd_ssize_t MPD_MINALLOC = MPD_MINALLOC_MIN; + +/* Custom allocation and free functions */ +void *(* mpd_mallocfunc)(size_t size) = malloc; +void *(* mpd_reallocfunc)(void *ptr, size_t size) = realloc; +void *(* mpd_callocfunc)(size_t nmemb, size_t size) = calloc; +void (* mpd_free)(void *ptr) = free; + + +/* emulate calloc if it is not available */ +void * +mpd_callocfunc_em(size_t nmemb, size_t size) +{ + void *ptr; + size_t req; + mpd_size_t overflow; + +#if MPD_SIZE_MAX < SIZE_MAX + /* full_coverage test only */ + if (nmemb > MPD_SIZE_MAX || size > MPD_SIZE_MAX) { + return NULL; + } +#endif + + req = mul_size_t_overflow((mpd_size_t)nmemb, (mpd_size_t)size, + &overflow); + if (overflow) { + return NULL; + } + + ptr = mpd_mallocfunc(req); + if (ptr == NULL) { + return NULL; + } + /* used on uint32_t or uint64_t */ + memset(ptr, 0, req); + + return ptr; +} + + +/* malloc with overflow checking */ +void * +mpd_alloc(mpd_size_t nmemb, mpd_size_t size) +{ + mpd_size_t req, overflow; + + req = mul_size_t_overflow(nmemb, size, &overflow); + if (overflow) { + return NULL; + } + + return mpd_mallocfunc(req); +} + +/* calloc with overflow checking */ +void * +mpd_calloc(mpd_size_t nmemb, mpd_size_t size) +{ + mpd_size_t overflow; + + (void)mul_size_t_overflow(nmemb, size, &overflow); + if (overflow) { + return NULL; + } + + return mpd_callocfunc(nmemb, size); +} + +/* realloc with overflow checking */ +void * +mpd_realloc(void *ptr, mpd_size_t nmemb, mpd_size_t size, uint8_t *err) +{ + void *new; + mpd_size_t req, overflow; + + req = mul_size_t_overflow(nmemb, size, &overflow); + if (overflow) { + *err = 1; + return ptr; + } + + new = mpd_reallocfunc(ptr, req); + if (new == NULL) { + *err = 1; + return ptr; + } + + return new; +} + +/* struct hack malloc with overflow checking */ +void * +mpd_sh_alloc(mpd_size_t struct_size, mpd_size_t nmemb, mpd_size_t size) +{ + mpd_size_t req, overflow; + + req = mul_size_t_overflow(nmemb, size, &overflow); + if (overflow) { + return NULL; + } + + req = add_size_t_overflow(req, struct_size, &overflow); + if (overflow) { + return NULL; + } + + return mpd_mallocfunc(req); +} + + +/* Allocate a new decimal with a coefficient of length 'nwords'. In case + of an error the return value is NULL. */ +mpd_t * +mpd_qnew_size(mpd_ssize_t nwords) +{ + mpd_t *result; + + nwords = (nwords < MPD_MINALLOC) ? MPD_MINALLOC : nwords; + + result = mpd_alloc(1, sizeof *result); + if (result == NULL) { + return NULL; + } + + result->data = mpd_alloc(nwords, sizeof *result->data); + if (result->data == NULL) { + mpd_free(result); + return NULL; + } + + result->flags = 0; + result->exp = 0; + result->digits = 0; + result->len = 0; + result->alloc = nwords; + + return result; +} + +/* Allocate a new decimal with a coefficient of length MPD_MINALLOC. + In case of an error the return value is NULL. */ +mpd_t * +mpd_qnew(void) +{ + return mpd_qnew_size(MPD_MINALLOC); +} + +/* Allocate new decimal. Caller can check for NULL or MPD_Malloc_error. + Raises on error. */ +mpd_t * +mpd_new(mpd_context_t *ctx) +{ + mpd_t *result; + + result = mpd_qnew(); + if (result == NULL) { + mpd_addstatus_raise(ctx, MPD_Malloc_error); + } + return result; +} + +/* + * Input: 'result' is a static mpd_t with a static coefficient. + * Assumption: 'nwords' >= result->alloc. + * + * Resize the static coefficient to a larger dynamic one and copy the + * existing data. If successful, the value of 'result' is unchanged. + * Otherwise, set 'result' to NaN and update 'status' with MPD_Malloc_error. + */ +int +mpd_switch_to_dyn(mpd_t *result, mpd_ssize_t nwords, uint32_t *status) +{ + mpd_uint_t *p = result->data; + + assert(nwords >= result->alloc); + + result->data = mpd_alloc(nwords, sizeof *result->data); + if (result->data == NULL) { + result->data = p; + mpd_set_qnan(result); + mpd_set_positive(result); + result->exp = result->digits = result->len = 0; + *status |= MPD_Malloc_error; + return 0; + } + + memcpy(result->data, p, result->alloc * (sizeof *result->data)); + result->alloc = nwords; + mpd_set_dynamic_data(result); + return 1; +} + +/* + * Input: 'result' is a static mpd_t with a static coefficient. + * + * Convert the coefficient to a dynamic one that is initialized to zero. If + * malloc fails, set 'result' to NaN and update 'status' with MPD_Malloc_error. + */ +int +mpd_switch_to_dyn_zero(mpd_t *result, mpd_ssize_t nwords, uint32_t *status) +{ + mpd_uint_t *p = result->data; + + result->data = mpd_calloc(nwords, sizeof *result->data); + if (result->data == NULL) { + result->data = p; + mpd_set_qnan(result); + mpd_set_positive(result); + result->exp = result->digits = result->len = 0; + *status |= MPD_Malloc_error; + return 0; + } + + result->alloc = nwords; + mpd_set_dynamic_data(result); + + return 1; +} + +/* + * Input: 'result' is a static or a dynamic mpd_t with a dynamic coefficient. + * Resize the coefficient to length 'nwords': + * Case nwords > result->alloc: + * If realloc is successful: + * 'result' has a larger coefficient but the same value. Return 1. + * Otherwise: + * Set 'result' to NaN, update status with MPD_Malloc_error and return 0. + * Case nwords < result->alloc: + * If realloc is successful: + * 'result' has a smaller coefficient. result->len is undefined. Return 1. + * Otherwise (unlikely): + * 'result' is unchanged. Reuse the now oversized coefficient. Return 1. + */ +int +mpd_realloc_dyn(mpd_t *result, mpd_ssize_t nwords, uint32_t *status) +{ + uint8_t err = 0; + + result->data = mpd_realloc(result->data, nwords, sizeof *result->data, &err); + if (!err) { + result->alloc = nwords; + } + else if (nwords > result->alloc) { + mpd_set_qnan(result); + mpd_set_positive(result); + result->exp = result->digits = result->len = 0; + *status |= MPD_Malloc_error; + return 0; + } + + return 1; +} + + diff --git a/Modules/_decimal/libmpdec/memory.h b/Modules/_decimal/libmpdec/memory.h new file mode 100644 index 0000000000..b3a4a56e96 --- /dev/null +++ b/Modules/_decimal/libmpdec/memory.h @@ -0,0 +1,44 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#ifndef MEMORY_H +#define MEMORY_H + + +#include "mpdecimal.h" + + +int mpd_switch_to_dyn(mpd_t *result, mpd_ssize_t size, uint32_t *status); +int mpd_switch_to_dyn_zero(mpd_t *result, mpd_ssize_t size, uint32_t *status); +int mpd_realloc_dyn(mpd_t *result, mpd_ssize_t size, uint32_t *status); + + +#endif + + + diff --git a/Modules/_decimal/libmpdec/mpdecimal.c b/Modules/_decimal/libmpdec/mpdecimal.c new file mode 100644 index 0000000000..89be5355aa --- /dev/null +++ b/Modules/_decimal/libmpdec/mpdecimal.c @@ -0,0 +1,7623 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#include "mpdecimal.h" +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <limits.h> +#include <math.h> +#include "basearith.h" +#include "bits.h" +#include "convolute.h" +#include "crt.h" +#include "errno.h" +#include "memory.h" +#include "typearith.h" +#include "umodarith.h" + +#ifdef PPRO + #if defined(_MSC_VER) + #include <float.h> + #pragma fenv_access(on) + #elif !defined(__OpenBSD__) && !defined(__NetBSD__) + /* C99 */ + #include <fenv.h> + #pragma STDC FENV_ACCESS ON + #endif +#endif + +#if defined(__x86_64__) && defined(__GLIBC__) && !defined(__INTEL_COMPILER) + #define USE_80BIT_LONG_DOUBLE +#endif + +#if defined(_MSC_VER) + #define ALWAYS_INLINE __forceinline +#elif defined(LEGACY_COMPILER) + #define ALWAYS_INLINE + #undef inline + #define inline +#else + #ifdef TEST_COVERAGE + #define ALWAYS_INLINE + #else + #define ALWAYS_INLINE inline __attribute__ ((always_inline)) + #endif +#endif + + +#define MPD_NEWTONDIV_CUTOFF 1024L + +#define MPD_NEW_STATIC(name, flags, exp, digits, len) \ + mpd_uint_t name##_data[MPD_MINALLOC_MAX]; \ + mpd_t name = {flags|MPD_STATIC|MPD_STATIC_DATA, exp, digits, \ + len, MPD_MINALLOC_MAX, name##_data} + +#define MPD_NEW_CONST(name, flags, exp, digits, len, alloc, initval) \ + mpd_uint_t name##_data[alloc] = {initval}; \ + mpd_t name = {flags|MPD_STATIC|MPD_CONST_DATA, exp, digits, \ + len, alloc, name##_data} + +#define MPD_NEW_SHARED(name, a) \ + mpd_t name = {(a->flags&~MPD_DATAFLAGS)|MPD_STATIC|MPD_SHARED_DATA, \ + a->exp, a->digits, a->len, a->alloc, a->data} + + +static mpd_uint_t data_one[1] = {1}; +static mpd_uint_t data_zero[1] = {0}; +static const mpd_t one = {MPD_STATIC|MPD_CONST_DATA, 0, 1, 1, 1, data_one}; +static const mpd_t minus_one = {MPD_NEG|MPD_STATIC|MPD_CONST_DATA, 0, 1, 1, 1, + data_one}; +static const mpd_t zero = {MPD_STATIC|MPD_CONST_DATA, 0, 1, 1, 1, data_zero}; + +static inline void _mpd_check_exp(mpd_t *dec, const mpd_context_t *ctx, + uint32_t *status); +static void _settriple(mpd_t *result, uint8_t sign, mpd_uint_t a, + mpd_ssize_t exp); +static inline mpd_ssize_t _mpd_real_size(mpd_uint_t *data, mpd_ssize_t size); + +static void _mpd_qadd(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status); +static inline void _mpd_qmul(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status); +static void _mpd_qbarrett_divmod(mpd_t *q, mpd_t *r, const mpd_t *a, + const mpd_t *b, uint32_t *status); +static inline void _mpd_qpow_uint(mpd_t *result, mpd_t *base, mpd_uint_t exp, + uint8_t resultsign, const mpd_context_t *ctx, uint32_t *status); + +mpd_uint_t mpd_qsshiftr(mpd_t *result, const mpd_t *a, mpd_ssize_t n); + + +/******************************************************************************/ +/* Performance critical inline functions */ +/******************************************************************************/ + +#ifdef CONFIG_64 +/* Digits in a word, primarily useful for the most significant word. */ +ALWAYS_INLINE int +mpd_word_digits(mpd_uint_t word) +{ + if (word < mpd_pow10[9]) { + if (word < mpd_pow10[4]) { + if (word < mpd_pow10[2]) { + return (word < mpd_pow10[1]) ? 1 : 2; + } + return (word < mpd_pow10[3]) ? 3 : 4; + } + if (word < mpd_pow10[6]) { + return (word < mpd_pow10[5]) ? 5 : 6; + } + if (word < mpd_pow10[8]) { + return (word < mpd_pow10[7]) ? 7 : 8; + } + return 9; + } + if (word < mpd_pow10[14]) { + if (word < mpd_pow10[11]) { + return (word < mpd_pow10[10]) ? 10 : 11; + } + if (word < mpd_pow10[13]) { + return (word < mpd_pow10[12]) ? 12 : 13; + } + return 14; + } + if (word < mpd_pow10[18]) { + if (word < mpd_pow10[16]) { + return (word < mpd_pow10[15]) ? 15 : 16; + } + return (word < mpd_pow10[17]) ? 17 : 18; + } + + return (word < mpd_pow10[19]) ? 19 : 20; +} +#else +ALWAYS_INLINE int +mpd_word_digits(mpd_uint_t word) +{ + if (word < mpd_pow10[4]) { + if (word < mpd_pow10[2]) { + return (word < mpd_pow10[1]) ? 1 : 2; + } + return (word < mpd_pow10[3]) ? 3 : 4; + } + if (word < mpd_pow10[6]) { + return (word < mpd_pow10[5]) ? 5 : 6; + } + if (word < mpd_pow10[8]) { + return (word < mpd_pow10[7]) ? 7 : 8; + } + + return (word < mpd_pow10[9]) ? 9 : 10; +} +#endif + + +/* Adjusted exponent */ +ALWAYS_INLINE mpd_ssize_t +mpd_adjexp(const mpd_t *dec) +{ + return (dec->exp + dec->digits) - 1; +} + +/* Etiny */ +ALWAYS_INLINE mpd_ssize_t +mpd_etiny(const mpd_context_t *ctx) +{ + return ctx->emin - (ctx->prec - 1); +} + +/* Etop: used for folding down in IEEE clamping */ +ALWAYS_INLINE mpd_ssize_t +mpd_etop(const mpd_context_t *ctx) +{ + return ctx->emax - (ctx->prec - 1); +} + +/* Most significant word */ +ALWAYS_INLINE mpd_uint_t +mpd_msword(const mpd_t *dec) +{ + assert(dec->len > 0); + return dec->data[dec->len-1]; +} + +/* Most significant digit of a word */ +inline mpd_uint_t +mpd_msd(mpd_uint_t word) +{ + int n; + + n = mpd_word_digits(word); + return word / mpd_pow10[n-1]; +} + +/* Least significant digit of a word */ +ALWAYS_INLINE mpd_uint_t +mpd_lsd(mpd_uint_t word) +{ + return word % 10; +} + +/* Coefficient size needed to store 'digits' */ +ALWAYS_INLINE mpd_ssize_t +mpd_digits_to_size(mpd_ssize_t digits) +{ + mpd_ssize_t q, r; + + _mpd_idiv_word(&q, &r, digits, MPD_RDIGITS); + return (r == 0) ? q : q+1; +} + +/* Number of digits in the exponent. Not defined for MPD_SSIZE_MIN. */ +inline int +mpd_exp_digits(mpd_ssize_t exp) +{ + exp = (exp < 0) ? -exp : exp; + return mpd_word_digits(exp); +} + +/* Canonical */ +ALWAYS_INLINE int +mpd_iscanonical(const mpd_t *dec UNUSED) +{ + return 1; +} + +/* Finite */ +ALWAYS_INLINE int +mpd_isfinite(const mpd_t *dec) +{ + return !(dec->flags & MPD_SPECIAL); +} + +/* Infinite */ +ALWAYS_INLINE int +mpd_isinfinite(const mpd_t *dec) +{ + return dec->flags & MPD_INF; +} + +/* NaN */ +ALWAYS_INLINE int +mpd_isnan(const mpd_t *dec) +{ + return dec->flags & (MPD_NAN|MPD_SNAN); +} + +/* Negative */ +ALWAYS_INLINE int +mpd_isnegative(const mpd_t *dec) +{ + return dec->flags & MPD_NEG; +} + +/* Positive */ +ALWAYS_INLINE int +mpd_ispositive(const mpd_t *dec) +{ + return !(dec->flags & MPD_NEG); +} + +/* qNaN */ +ALWAYS_INLINE int +mpd_isqnan(const mpd_t *dec) +{ + return dec->flags & MPD_NAN; +} + +/* Signed */ +ALWAYS_INLINE int +mpd_issigned(const mpd_t *dec) +{ + return dec->flags & MPD_NEG; +} + +/* sNaN */ +ALWAYS_INLINE int +mpd_issnan(const mpd_t *dec) +{ + return dec->flags & MPD_SNAN; +} + +/* Special */ +ALWAYS_INLINE int +mpd_isspecial(const mpd_t *dec) +{ + return dec->flags & MPD_SPECIAL; +} + +/* Zero */ +ALWAYS_INLINE int +mpd_iszero(const mpd_t *dec) +{ + return !mpd_isspecial(dec) && mpd_msword(dec) == 0; +} + +/* Test for zero when specials have been ruled out already */ +ALWAYS_INLINE int +mpd_iszerocoeff(const mpd_t *dec) +{ + return mpd_msword(dec) == 0; +} + +/* Normal */ +inline int +mpd_isnormal(const mpd_t *dec, const mpd_context_t *ctx) +{ + if (mpd_isspecial(dec)) return 0; + if (mpd_iszerocoeff(dec)) return 0; + + return mpd_adjexp(dec) >= ctx->emin; +} + +/* Subnormal */ +inline int +mpd_issubnormal(const mpd_t *dec, const mpd_context_t *ctx) +{ + if (mpd_isspecial(dec)) return 0; + if (mpd_iszerocoeff(dec)) return 0; + + return mpd_adjexp(dec) < ctx->emin; +} + +/* Odd word */ +ALWAYS_INLINE int +mpd_isoddword(mpd_uint_t word) +{ + return word & 1; +} + +/* Odd coefficient */ +ALWAYS_INLINE int +mpd_isoddcoeff(const mpd_t *dec) +{ + return mpd_isoddword(dec->data[0]); +} + +/* 0 if dec is positive, 1 if dec is negative */ +ALWAYS_INLINE uint8_t +mpd_sign(const mpd_t *dec) +{ + return dec->flags & MPD_NEG; +} + +/* 1 if dec is positive, -1 if dec is negative */ +ALWAYS_INLINE int +mpd_arith_sign(const mpd_t *dec) +{ + return 1 - 2 * mpd_isnegative(dec); +} + +/* Radix */ +ALWAYS_INLINE long +mpd_radix(void) +{ + return 10; +} + +/* Dynamic decimal */ +ALWAYS_INLINE int +mpd_isdynamic(mpd_t *dec) +{ + return !(dec->flags & MPD_STATIC); +} + +/* Static decimal */ +ALWAYS_INLINE int +mpd_isstatic(mpd_t *dec) +{ + return dec->flags & MPD_STATIC; +} + +/* Data of decimal is dynamic */ +ALWAYS_INLINE int +mpd_isdynamic_data(mpd_t *dec) +{ + return !(dec->flags & MPD_DATAFLAGS); +} + +/* Data of decimal is static */ +ALWAYS_INLINE int +mpd_isstatic_data(mpd_t *dec) +{ + return dec->flags & MPD_STATIC_DATA; +} + +/* Data of decimal is shared */ +ALWAYS_INLINE int +mpd_isshared_data(mpd_t *dec) +{ + return dec->flags & MPD_SHARED_DATA; +} + +/* Data of decimal is const */ +ALWAYS_INLINE int +mpd_isconst_data(mpd_t *dec) +{ + return dec->flags & MPD_CONST_DATA; +} + + +/******************************************************************************/ +/* Inline memory handling */ +/******************************************************************************/ + +/* Fill destination with zeros */ +ALWAYS_INLINE void +mpd_uint_zero(mpd_uint_t *dest, mpd_size_t len) +{ + mpd_size_t i; + + for (i = 0; i < len; i++) { + dest[i] = 0; + } +} + +/* Free a decimal */ +ALWAYS_INLINE void +mpd_del(mpd_t *dec) +{ + if (mpd_isdynamic_data(dec)) { + mpd_free(dec->data); + } + if (mpd_isdynamic(dec)) { + mpd_free(dec); + } +} + +/* + * Resize the coefficient. Existing data up to 'nwords' is left untouched. + * Return 1 on success, 0 otherwise. + * + * Input invariant: MPD_MINALLOC <= result->alloc. + * + * Case nwords == result->alloc: + * 'result' is unchanged. Return 1. + * + * Case nwords > result->alloc: + * Case realloc success: + * The value of 'result' does not change. Return 1. + * Case realloc failure: + * 'result' is NaN, status is updated with MPD_Malloc_error. Return 0. + * + * Case nwords < result->alloc: + * Case is_static_data or realloc failure [1]: + * 'result' is unchanged. Return 1. + * Case realloc success: + * The value of result is undefined (expected). Return 1. + * + * + * [1] In that case the old (now oversized) area is still valid. + */ +ALWAYS_INLINE int +mpd_qresize(mpd_t *result, mpd_ssize_t nwords, uint32_t *status) +{ + assert(!mpd_isconst_data(result)); /* illegal operation for a const */ + assert(!mpd_isshared_data(result)); /* illegal operation for a shared */ + assert(MPD_MINALLOC <= result->alloc); + + nwords = (nwords <= MPD_MINALLOC) ? MPD_MINALLOC : nwords; + if (nwords == result->alloc) { + return 1; + } + if (mpd_isstatic_data(result)) { + if (nwords > result->alloc) { + return mpd_switch_to_dyn(result, nwords, status); + } + return 1; + } + + return mpd_realloc_dyn(result, nwords, status); +} + +/* Same as mpd_qresize, but the complete coefficient (including the old + * memory area!) is initialized to zero. */ +ALWAYS_INLINE int +mpd_qresize_zero(mpd_t *result, mpd_ssize_t nwords, uint32_t *status) +{ + assert(!mpd_isconst_data(result)); /* illegal operation for a const */ + assert(!mpd_isshared_data(result)); /* illegal operation for a shared */ + assert(MPD_MINALLOC <= result->alloc); + + nwords = (nwords <= MPD_MINALLOC) ? MPD_MINALLOC : nwords; + if (nwords != result->alloc) { + if (mpd_isstatic_data(result)) { + if (nwords > result->alloc) { + return mpd_switch_to_dyn_zero(result, nwords, status); + } + } + else if (!mpd_realloc_dyn(result, nwords, status)) { + return 0; + } + } + + mpd_uint_zero(result->data, nwords); + return 1; +} + +/* + * Reduce memory size for the coefficient to MPD_MINALLOC. In theory, + * realloc may fail even when reducing the memory size. But in that case + * the old memory area is always big enough, so checking for MPD_Malloc_error + * is not imperative. + */ +ALWAYS_INLINE void +mpd_minalloc(mpd_t *result) +{ + assert(!mpd_isconst_data(result)); /* illegal operation for a const */ + assert(!mpd_isshared_data(result)); /* illegal operation for a shared */ + + if (!mpd_isstatic_data(result) && result->alloc > MPD_MINALLOC) { + uint8_t err = 0; + result->data = mpd_realloc(result->data, MPD_MINALLOC, + sizeof *result->data, &err); + if (!err) { + result->alloc = MPD_MINALLOC; + } + } +} + +int +mpd_resize(mpd_t *result, mpd_ssize_t nwords, mpd_context_t *ctx) +{ + uint32_t status = 0; + if (!mpd_qresize(result, nwords, &status)) { + mpd_addstatus_raise(ctx, status); + return 0; + } + return 1; +} + +int +mpd_resize_zero(mpd_t *result, mpd_ssize_t nwords, mpd_context_t *ctx) +{ + uint32_t status = 0; + if (!mpd_qresize_zero(result, nwords, &status)) { + mpd_addstatus_raise(ctx, status); + return 0; + } + return 1; +} + + +/******************************************************************************/ +/* Set attributes of a decimal */ +/******************************************************************************/ + +/* Set digits. Assumption: result->len is initialized and > 0. */ +inline void +mpd_setdigits(mpd_t *result) +{ + mpd_ssize_t wdigits = mpd_word_digits(mpd_msword(result)); + result->digits = wdigits + (result->len-1) * MPD_RDIGITS; +} + +/* Set sign */ +ALWAYS_INLINE void +mpd_set_sign(mpd_t *result, uint8_t sign) +{ + result->flags &= ~MPD_NEG; + result->flags |= sign; +} + +/* Copy sign from another decimal */ +ALWAYS_INLINE void +mpd_signcpy(mpd_t *result, mpd_t *a) +{ + uint8_t sign = a->flags&MPD_NEG; + + result->flags &= ~MPD_NEG; + result->flags |= sign; +} + +/* Set infinity */ +ALWAYS_INLINE void +mpd_set_infinity(mpd_t *result) +{ + result->flags &= ~MPD_SPECIAL; + result->flags |= MPD_INF; +} + +/* Set qNaN */ +ALWAYS_INLINE void +mpd_set_qnan(mpd_t *result) +{ + result->flags &= ~MPD_SPECIAL; + result->flags |= MPD_NAN; +} + +/* Set sNaN */ +ALWAYS_INLINE void +mpd_set_snan(mpd_t *result) +{ + result->flags &= ~MPD_SPECIAL; + result->flags |= MPD_SNAN; +} + +/* Set to negative */ +ALWAYS_INLINE void +mpd_set_negative(mpd_t *result) +{ + result->flags |= MPD_NEG; +} + +/* Set to positive */ +ALWAYS_INLINE void +mpd_set_positive(mpd_t *result) +{ + result->flags &= ~MPD_NEG; +} + +/* Set to dynamic */ +ALWAYS_INLINE void +mpd_set_dynamic(mpd_t *result) +{ + result->flags &= ~MPD_STATIC; +} + +/* Set to static */ +ALWAYS_INLINE void +mpd_set_static(mpd_t *result) +{ + result->flags |= MPD_STATIC; +} + +/* Set data to dynamic */ +ALWAYS_INLINE void +mpd_set_dynamic_data(mpd_t *result) +{ + result->flags &= ~MPD_DATAFLAGS; +} + +/* Set data to static */ +ALWAYS_INLINE void +mpd_set_static_data(mpd_t *result) +{ + result->flags &= ~MPD_DATAFLAGS; + result->flags |= MPD_STATIC_DATA; +} + +/* Set data to shared */ +ALWAYS_INLINE void +mpd_set_shared_data(mpd_t *result) +{ + result->flags &= ~MPD_DATAFLAGS; + result->flags |= MPD_SHARED_DATA; +} + +/* Set data to const */ +ALWAYS_INLINE void +mpd_set_const_data(mpd_t *result) +{ + result->flags &= ~MPD_DATAFLAGS; + result->flags |= MPD_CONST_DATA; +} + +/* Clear flags, preserving memory attributes. */ +ALWAYS_INLINE void +mpd_clear_flags(mpd_t *result) +{ + result->flags &= (MPD_STATIC|MPD_DATAFLAGS); +} + +/* Set flags, preserving memory attributes. */ +ALWAYS_INLINE void +mpd_set_flags(mpd_t *result, uint8_t flags) +{ + result->flags &= (MPD_STATIC|MPD_DATAFLAGS); + result->flags |= flags; +} + +/* Copy flags, preserving memory attributes of result. */ +ALWAYS_INLINE void +mpd_copy_flags(mpd_t *result, const mpd_t *a) +{ + uint8_t aflags = a->flags; + result->flags &= (MPD_STATIC|MPD_DATAFLAGS); + result->flags |= (aflags & ~(MPD_STATIC|MPD_DATAFLAGS)); +} + +/* Initialize a workcontext from ctx. Set traps, flags and newtrap to 0. */ +static inline void +mpd_workcontext(mpd_context_t *workctx, const mpd_context_t *ctx) +{ + workctx->prec = ctx->prec; + workctx->emax = ctx->emax; + workctx->emin = ctx->emin; + workctx->round = ctx->round; + workctx->traps = 0; + workctx->status = 0; + workctx->newtrap = 0; + workctx->clamp = ctx->clamp; + workctx->allcr = ctx->allcr; +} + + +/******************************************************************************/ +/* Getting and setting parts of decimals */ +/******************************************************************************/ + +/* Flip the sign of a decimal */ +static inline void +_mpd_negate(mpd_t *dec) +{ + dec->flags ^= MPD_NEG; +} + +/* Set coefficient to zero */ +void +mpd_zerocoeff(mpd_t *result) +{ + mpd_minalloc(result); + result->digits = 1; + result->len = 1; + result->data[0] = 0; +} + +/* Set the coefficient to all nines. */ +void +mpd_qmaxcoeff(mpd_t *result, const mpd_context_t *ctx, uint32_t *status) +{ + mpd_ssize_t len, r; + + _mpd_idiv_word(&len, &r, ctx->prec, MPD_RDIGITS); + len = (r == 0) ? len : len+1; + + if (!mpd_qresize(result, len, status)) { + return; + } + + result->len = len; + result->digits = ctx->prec; + + --len; + if (r > 0) { + result->data[len--] = mpd_pow10[r]-1; + } + for (; len >= 0; --len) { + result->data[len] = MPD_RADIX-1; + } +} + +/* + * Cut off the most significant digits so that the rest fits in ctx->prec. + * Cannot fail. + */ +static void +_mpd_cap(mpd_t *result, const mpd_context_t *ctx) +{ + uint32_t dummy; + mpd_ssize_t len, r; + + if (result->len > 0 && result->digits > ctx->prec) { + _mpd_idiv_word(&len, &r, ctx->prec, MPD_RDIGITS); + len = (r == 0) ? len : len+1; + + if (r != 0) { + result->data[len-1] %= mpd_pow10[r]; + } + + len = _mpd_real_size(result->data, len); + /* resize to fewer words cannot fail */ + mpd_qresize(result, len, &dummy); + result->len = len; + mpd_setdigits(result); + } + if (mpd_iszero(result)) { + _settriple(result, mpd_sign(result), 0, result->exp); + } +} + +/* + * Cut off the most significant digits of a NaN payload so that the rest + * fits in ctx->prec - ctx->clamp. Cannot fail. + */ +static void +_mpd_fix_nan(mpd_t *result, const mpd_context_t *ctx) +{ + uint32_t dummy; + mpd_ssize_t prec; + mpd_ssize_t len, r; + + prec = ctx->prec - ctx->clamp; + if (result->len > 0 && result->digits > prec) { + if (prec == 0) { + mpd_minalloc(result); + result->len = result->digits = 0; + } + else { + _mpd_idiv_word(&len, &r, prec, MPD_RDIGITS); + len = (r == 0) ? len : len+1; + + if (r != 0) { + result->data[len-1] %= mpd_pow10[r]; + } + + len = _mpd_real_size(result->data, len); + /* resize to fewer words cannot fail */ + mpd_qresize(result, len, &dummy); + result->len = len; + mpd_setdigits(result); + if (mpd_iszerocoeff(result)) { + /* NaN0 is not a valid representation */ + result->len = result->digits = 0; + } + } + } +} + +/* + * Get n most significant digits from a decimal, where 0 < n <= MPD_UINT_DIGITS. + * Assumes MPD_UINT_DIGITS == MPD_RDIGITS+1, which is true for 32 and 64 bit + * machines. + * + * The result of the operation will be in lo. If the operation is impossible, + * hi will be nonzero. This is used to indicate an error. + */ +static inline void +_mpd_get_msdigits(mpd_uint_t *hi, mpd_uint_t *lo, const mpd_t *dec, + unsigned int n) +{ + mpd_uint_t r, tmp; + + assert(0 < n && n <= MPD_RDIGITS+1); + + _mpd_div_word(&tmp, &r, dec->digits, MPD_RDIGITS); + r = (r == 0) ? MPD_RDIGITS : r; /* digits in the most significant word */ + + *hi = 0; + *lo = dec->data[dec->len-1]; + if (n <= r) { + *lo /= mpd_pow10[r-n]; + } + else if (dec->len > 1) { + /* at this point 1 <= r < n <= MPD_RDIGITS+1 */ + _mpd_mul_words(hi, lo, *lo, mpd_pow10[n-r]); + tmp = dec->data[dec->len-2] / mpd_pow10[MPD_RDIGITS-(n-r)]; + *lo = *lo + tmp; + if (*lo < tmp) (*hi)++; + } +} + + +/******************************************************************************/ +/* Gathering information about a decimal */ +/******************************************************************************/ + +/* The real size of the coefficient without leading zero words. */ +static inline mpd_ssize_t +_mpd_real_size(mpd_uint_t *data, mpd_ssize_t size) +{ + while (size > 1 && data[size-1] == 0) { + size--; + } + + return size; +} + +/* Return number of trailing zeros. No errors are possible. */ +mpd_ssize_t +mpd_trail_zeros(const mpd_t *dec) +{ + mpd_uint_t word; + mpd_ssize_t i, tz = 0; + + for (i=0; i < dec->len; ++i) { + if (dec->data[i] != 0) { + word = dec->data[i]; + tz = i * MPD_RDIGITS; + while (word % 10 == 0) { + word /= 10; + tz++; + } + break; + } + } + + return tz; +} + +/* Integer: Undefined for specials */ +static int +_mpd_isint(const mpd_t *dec) +{ + mpd_ssize_t tz; + + if (mpd_iszerocoeff(dec)) { + return 1; + } + + tz = mpd_trail_zeros(dec); + return (dec->exp + tz >= 0); +} + +/* Integer */ +int +mpd_isinteger(const mpd_t *dec) +{ + if (mpd_isspecial(dec)) { + return 0; + } + return _mpd_isint(dec); +} + +/* Word is a power of 10 */ +static int +mpd_word_ispow10(mpd_uint_t word) +{ + int n; + + n = mpd_word_digits(word); + if (word == mpd_pow10[n-1]) { + return 1; + } + + return 0; +} + +/* Coefficient is a power of 10 */ +static int +mpd_coeff_ispow10(const mpd_t *dec) +{ + if (mpd_word_ispow10(mpd_msword(dec))) { + if (_mpd_isallzero(dec->data, dec->len-1)) { + return 1; + } + } + + return 0; +} + +/* All digits of a word are nines */ +static int +mpd_word_isallnine(mpd_uint_t word) +{ + int n; + + n = mpd_word_digits(word); + if (word == mpd_pow10[n]-1) { + return 1; + } + + return 0; +} + +/* All digits of the coefficient are nines */ +static int +mpd_coeff_isallnine(const mpd_t *dec) +{ + if (mpd_word_isallnine(mpd_msword(dec))) { + if (_mpd_isallnine(dec->data, dec->len-1)) { + return 1; + } + } + + return 0; +} + +/* Odd decimal: Undefined for non-integers! */ +int +mpd_isodd(const mpd_t *dec) +{ + mpd_uint_t q, r; + assert(mpd_isinteger(dec)); + if (mpd_iszerocoeff(dec)) return 0; + if (dec->exp < 0) { + _mpd_div_word(&q, &r, -dec->exp, MPD_RDIGITS); + q = dec->data[q] / mpd_pow10[r]; + return mpd_isoddword(q); + } + return dec->exp == 0 && mpd_isoddword(dec->data[0]); +} + +/* Even: Undefined for non-integers! */ +int +mpd_iseven(const mpd_t *dec) +{ + return !mpd_isodd(dec); +} + +/******************************************************************************/ +/* Getting and setting decimals */ +/******************************************************************************/ + +/* Internal function: Set a static decimal from a triple, no error checking. */ +static void +_ssettriple(mpd_t *result, uint8_t sign, mpd_uint_t a, mpd_ssize_t exp) +{ + mpd_set_flags(result, sign); + result->exp = exp; + _mpd_div_word(&result->data[1], &result->data[0], a, MPD_RADIX); + result->len = (result->data[1] == 0) ? 1 : 2; + mpd_setdigits(result); +} + +/* Internal function: Set a decimal from a triple, no error checking. */ +static void +_settriple(mpd_t *result, uint8_t sign, mpd_uint_t a, mpd_ssize_t exp) +{ + mpd_minalloc(result); + mpd_set_flags(result, sign); + result->exp = exp; + _mpd_div_word(&result->data[1], &result->data[0], a, MPD_RADIX); + result->len = (result->data[1] == 0) ? 1 : 2; + mpd_setdigits(result); +} + +/* Set a special number from a triple */ +void +mpd_setspecial(mpd_t *result, uint8_t sign, uint8_t type) +{ + mpd_minalloc(result); + result->flags &= ~(MPD_NEG|MPD_SPECIAL); + result->flags |= (sign|type); + result->exp = result->digits = result->len = 0; +} + +/* Set result of NaN with an error status */ +void +mpd_seterror(mpd_t *result, uint32_t flags, uint32_t *status) +{ + mpd_minalloc(result); + mpd_set_qnan(result); + mpd_set_positive(result); + result->exp = result->digits = result->len = 0; + *status |= flags; +} + +/* quietly set a static decimal from an mpd_ssize_t */ +void +mpd_qsset_ssize(mpd_t *result, mpd_ssize_t a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_uint_t u; + uint8_t sign = MPD_POS; + + if (a < 0) { + if (a == MPD_SSIZE_MIN) { + u = (mpd_uint_t)MPD_SSIZE_MAX + + (-(MPD_SSIZE_MIN+MPD_SSIZE_MAX)); + } + else { + u = -a; + } + sign = MPD_NEG; + } + else { + u = a; + } + _ssettriple(result, sign, u, 0); + mpd_qfinalize(result, ctx, status); +} + +/* quietly set a static decimal from an mpd_uint_t */ +void +mpd_qsset_uint(mpd_t *result, mpd_uint_t a, const mpd_context_t *ctx, + uint32_t *status) +{ + _ssettriple(result, MPD_POS, a, 0); + mpd_qfinalize(result, ctx, status); +} + +/* quietly set a static decimal from an int32_t */ +void +mpd_qsset_i32(mpd_t *result, int32_t a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_qsset_ssize(result, a, ctx, status); +} + +/* quietly set a static decimal from a uint32_t */ +void +mpd_qsset_u32(mpd_t *result, uint32_t a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_qsset_uint(result, a, ctx, status); +} + +#ifdef CONFIG_64 +/* quietly set a static decimal from an int64_t */ +void +mpd_qsset_i64(mpd_t *result, int64_t a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_qsset_ssize(result, a, ctx, status); +} + +/* quietly set a static decimal from a uint64_t */ +void +mpd_qsset_u64(mpd_t *result, uint64_t a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_qsset_uint(result, a, ctx, status); +} +#endif + +/* quietly set a decimal from an mpd_ssize_t */ +void +mpd_qset_ssize(mpd_t *result, mpd_ssize_t a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_minalloc(result); + mpd_qsset_ssize(result, a, ctx, status); +} + +/* quietly set a decimal from an mpd_uint_t */ +void +mpd_qset_uint(mpd_t *result, mpd_uint_t a, const mpd_context_t *ctx, + uint32_t *status) +{ + _settriple(result, MPD_POS, a, 0); + mpd_qfinalize(result, ctx, status); +} + +/* quietly set a decimal from an int32_t */ +void +mpd_qset_i32(mpd_t *result, int32_t a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_qset_ssize(result, a, ctx, status); +} + +/* quietly set a decimal from a uint32_t */ +void +mpd_qset_u32(mpd_t *result, uint32_t a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_qset_uint(result, a, ctx, status); +} + +#if defined(CONFIG_32) && !defined(LEGACY_COMPILER) +/* set a decimal from a uint64_t */ +static void +_c32setu64(mpd_t *result, uint64_t u, uint8_t sign, uint32_t *status) +{ + mpd_uint_t w[3]; + uint64_t q; + int i, len; + + len = 0; + do { + q = u / MPD_RADIX; + w[len] = (mpd_uint_t)(u - q * MPD_RADIX); + u = q; len++; + } while (u != 0); + + if (!mpd_qresize(result, len, status)) { + return; + } + for (i = 0; i < len; i++) { + result->data[i] = w[i]; + } + + mpd_set_sign(result, sign); + result->exp = 0; + result->len = len; + mpd_setdigits(result); +} + +static void +_c32_qset_u64(mpd_t *result, uint64_t a, const mpd_context_t *ctx, + uint32_t *status) +{ + _c32setu64(result, a, MPD_POS, status); + mpd_qfinalize(result, ctx, status); +} + +/* set a decimal from an int64_t */ +static void +_c32_qset_i64(mpd_t *result, int64_t a, const mpd_context_t *ctx, + uint32_t *status) +{ + uint64_t u; + uint8_t sign = MPD_POS; + + if (a < 0) { + if (a == INT64_MIN) { + u = (uint64_t)INT64_MAX + (-(INT64_MIN+INT64_MAX)); + } + else { + u = -a; + } + sign = MPD_NEG; + } + else { + u = a; + } + _c32setu64(result, u, sign, status); + mpd_qfinalize(result, ctx, status); +} +#endif /* CONFIG_32 && !LEGACY_COMPILER */ + +#ifndef LEGACY_COMPILER +/* quietly set a decimal from an int64_t */ +void +mpd_qset_i64(mpd_t *result, int64_t a, const mpd_context_t *ctx, + uint32_t *status) +{ +#ifdef CONFIG_64 + mpd_qset_ssize(result, a, ctx, status); +#else + _c32_qset_i64(result, a, ctx, status); +#endif +} + +/* quietly set a decimal from a uint64_t */ +void +mpd_qset_u64(mpd_t *result, uint64_t a, const mpd_context_t *ctx, + uint32_t *status) +{ +#ifdef CONFIG_64 + mpd_qset_uint(result, a, ctx, status); +#else + _c32_qset_u64(result, a, ctx, status); +#endif +} +#endif /* !LEGACY_COMPILER */ + + +/* + * Quietly get an mpd_uint_t from a decimal. Assumes + * MPD_UINT_DIGITS == MPD_RDIGITS+1, which is true for + * 32 and 64 bit machines. + * + * If the operation is impossible, MPD_Invalid_operation is set. + */ +static mpd_uint_t +_mpd_qget_uint(int use_sign, const mpd_t *a, uint32_t *status) +{ + mpd_t tmp; + mpd_uint_t tmp_data[2]; + mpd_uint_t lo, hi; + + if (mpd_isspecial(a)) { + *status |= MPD_Invalid_operation; + return MPD_UINT_MAX; + } + if (mpd_iszero(a)) { + return 0; + } + if (use_sign && mpd_isnegative(a)) { + *status |= MPD_Invalid_operation; + return MPD_UINT_MAX; + } + + if (a->digits+a->exp > MPD_RDIGITS+1) { + *status |= MPD_Invalid_operation; + return MPD_UINT_MAX; + } + + if (a->exp < 0) { + if (!_mpd_isint(a)) { + *status |= MPD_Invalid_operation; + return MPD_UINT_MAX; + } + /* At this point a->digits+a->exp <= MPD_RDIGITS+1, + * so the shift fits. */ + tmp.data = tmp_data; + tmp.flags = MPD_STATIC|MPD_STATIC_DATA; + tmp.alloc = 2; + mpd_qsshiftr(&tmp, a, -a->exp); + tmp.exp = 0; + a = &tmp; + } + + _mpd_get_msdigits(&hi, &lo, a, MPD_RDIGITS+1); + if (hi) { + *status |= MPD_Invalid_operation; + return MPD_UINT_MAX; + } + + if (a->exp > 0) { + _mpd_mul_words(&hi, &lo, lo, mpd_pow10[a->exp]); + if (hi) { + *status |= MPD_Invalid_operation; + return MPD_UINT_MAX; + } + } + + return lo; +} + +/* + * Sets Invalid_operation for: + * - specials + * - negative numbers (except negative zero) + * - non-integers + * - overflow + */ +mpd_uint_t +mpd_qget_uint(const mpd_t *a, uint32_t *status) +{ + return _mpd_qget_uint(1, a, status); +} + +/* Same as above, but gets the absolute value, i.e. the sign is ignored. */ +mpd_uint_t +mpd_qabs_uint(const mpd_t *a, uint32_t *status) +{ + return _mpd_qget_uint(0, a, status); +} + +/* quietly get an mpd_ssize_t from a decimal */ +mpd_ssize_t +mpd_qget_ssize(const mpd_t *a, uint32_t *status) +{ + mpd_uint_t u; + int isneg; + + u = mpd_qabs_uint(a, status); + if (*status&MPD_Invalid_operation) { + return MPD_SSIZE_MAX; + } + + isneg = mpd_isnegative(a); + if (u <= MPD_SSIZE_MAX) { + return isneg ? -((mpd_ssize_t)u) : (mpd_ssize_t)u; + } + else if (isneg && u+(MPD_SSIZE_MIN+MPD_SSIZE_MAX) == MPD_SSIZE_MAX) { + return MPD_SSIZE_MIN; + } + + *status |= MPD_Invalid_operation; + return MPD_SSIZE_MAX; +} + +#ifdef CONFIG_64 +/* quietly get a uint64_t from a decimal */ +uint64_t +mpd_qget_u64(const mpd_t *a, uint32_t *status) +{ + return mpd_qget_uint(a, status); +} + +/* quietly get an int64_t from a decimal */ +int64_t +mpd_qget_i64(const mpd_t *a, uint32_t *status) +{ + return mpd_qget_ssize(a, status); +} +#else +/* quietly get a uint32_t from a decimal */ +uint32_t +mpd_qget_u32(const mpd_t *a, uint32_t *status) +{ + return mpd_qget_uint(a, status); +} + +/* quietly get an int32_t from a decimal */ +int32_t +mpd_qget_i32(const mpd_t *a, uint32_t *status) +{ + return mpd_qget_ssize(a, status); +} +#endif + + +/******************************************************************************/ +/* Filtering input of functions, finalizing output of functions */ +/******************************************************************************/ + +/* + * Check if the operand is NaN, copy to result and return 1 if this is + * the case. Copying can fail since NaNs are allowed to have a payload that + * does not fit in MPD_MINALLOC. + */ +int +mpd_qcheck_nan(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + if (mpd_isnan(a)) { + *status |= mpd_issnan(a) ? MPD_Invalid_operation : 0; + mpd_qcopy(result, a, status); + mpd_set_qnan(result); + _mpd_fix_nan(result, ctx); + return 1; + } + return 0; +} + +/* + * Check if either operand is NaN, copy to result and return 1 if this + * is the case. Copying can fail since NaNs are allowed to have a payload + * that does not fit in MPD_MINALLOC. + */ +int +mpd_qcheck_nans(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + if ((a->flags|b->flags)&(MPD_NAN|MPD_SNAN)) { + const mpd_t *choice = b; + if (mpd_issnan(a)) { + choice = a; + *status |= MPD_Invalid_operation; + } + else if (mpd_issnan(b)) { + *status |= MPD_Invalid_operation; + } + else if (mpd_isqnan(a)) { + choice = a; + } + mpd_qcopy(result, choice, status); + mpd_set_qnan(result); + _mpd_fix_nan(result, ctx); + return 1; + } + return 0; +} + +/* + * Check if one of the operands is NaN, copy to result and return 1 if this + * is the case. Copying can fail since NaNs are allowed to have a payload + * that does not fit in MPD_MINALLOC. + */ +static int +mpd_qcheck_3nans(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_t *c, + const mpd_context_t *ctx, uint32_t *status) +{ + if ((a->flags|b->flags|c->flags)&(MPD_NAN|MPD_SNAN)) { + const mpd_t *choice = c; + if (mpd_issnan(a)) { + choice = a; + *status |= MPD_Invalid_operation; + } + else if (mpd_issnan(b)) { + choice = b; + *status |= MPD_Invalid_operation; + } + else if (mpd_issnan(c)) { + *status |= MPD_Invalid_operation; + } + else if (mpd_isqnan(a)) { + choice = a; + } + else if (mpd_isqnan(b)) { + choice = b; + } + mpd_qcopy(result, choice, status); + mpd_set_qnan(result); + _mpd_fix_nan(result, ctx); + return 1; + } + return 0; +} + +/* Check if rounding digit 'rnd' leads to an increment. */ +static inline int +_mpd_rnd_incr(const mpd_t *dec, mpd_uint_t rnd, const mpd_context_t *ctx) +{ + int ld; + + switch (ctx->round) { + case MPD_ROUND_DOWN: case MPD_ROUND_TRUNC: + return 0; + case MPD_ROUND_HALF_UP: + return (rnd >= 5); + case MPD_ROUND_HALF_EVEN: + return (rnd > 5) || ((rnd == 5) && mpd_isoddcoeff(dec)); + case MPD_ROUND_CEILING: + return !(rnd == 0 || mpd_isnegative(dec)); + case MPD_ROUND_FLOOR: + return !(rnd == 0 || mpd_ispositive(dec)); + case MPD_ROUND_HALF_DOWN: + return (rnd > 5); + case MPD_ROUND_UP: + return !(rnd == 0); + case MPD_ROUND_05UP: + ld = (int)mpd_lsd(dec->data[0]); + return (!(rnd == 0) && (ld == 0 || ld == 5)); + default: + /* Without a valid context, further results will be undefined. */ + return 0; /* GCOV_NOT_REACHED */ + } +} + +/* + * Apply rounding to a decimal that has been right-shifted into a full + * precision decimal. If an increment leads to an overflow of the precision, + * adjust the coefficient and the exponent and check the new exponent for + * overflow. + */ +static inline void +_mpd_apply_round(mpd_t *dec, mpd_uint_t rnd, const mpd_context_t *ctx, + uint32_t *status) +{ + if (_mpd_rnd_incr(dec, rnd, ctx)) { + /* We have a number with exactly ctx->prec digits. The increment + * can only lead to an overflow if the decimal is all nines. In + * that case, the result is a power of ten with prec+1 digits. + * + * If the precision is a multiple of MPD_RDIGITS, this situation is + * detected by _mpd_baseincr returning a carry. + * If the precision is not a multiple of MPD_RDIGITS, we have to + * check if the result has one digit too many. + */ + mpd_uint_t carry = _mpd_baseincr(dec->data, dec->len); + if (carry) { + dec->data[dec->len-1] = mpd_pow10[MPD_RDIGITS-1]; + dec->exp += 1; + _mpd_check_exp(dec, ctx, status); + return; + } + mpd_setdigits(dec); + if (dec->digits > ctx->prec) { + mpd_qshiftr_inplace(dec, 1); + dec->exp += 1; + dec->digits = ctx->prec; + _mpd_check_exp(dec, ctx, status); + } + } +} + +/* + * Apply rounding to a decimal. Allow overflow of the precision. + */ +static inline void +_mpd_apply_round_excess(mpd_t *dec, mpd_uint_t rnd, const mpd_context_t *ctx, + uint32_t *status) +{ + if (_mpd_rnd_incr(dec, rnd, ctx)) { + mpd_uint_t carry = _mpd_baseincr(dec->data, dec->len); + if (carry) { + if (!mpd_qresize(dec, dec->len+1, status)) { + return; + } + dec->data[dec->len] = 1; + dec->len += 1; + } + mpd_setdigits(dec); + } +} + +/* + * Apply rounding to a decimal that has been right-shifted into a decimal + * with full precision or less. Return failure if an increment would + * overflow the precision. + */ +static inline int +_mpd_apply_round_fit(mpd_t *dec, mpd_uint_t rnd, const mpd_context_t *ctx, + uint32_t *status) +{ + if (_mpd_rnd_incr(dec, rnd, ctx)) { + mpd_uint_t carry = _mpd_baseincr(dec->data, dec->len); + if (carry) { + if (!mpd_qresize(dec, dec->len+1, status)) { + return 0; + } + dec->data[dec->len] = 1; + dec->len += 1; + } + mpd_setdigits(dec); + if (dec->digits > ctx->prec) { + mpd_seterror(dec, MPD_Invalid_operation, status); + return 0; + } + } + return 1; +} + +/* Check a normal number for overflow, underflow, clamping. If the operand + is modified, it will be zero, special or (sub)normal with a coefficient + that fits into the current context precision. */ +static inline void +_mpd_check_exp(mpd_t *dec, const mpd_context_t *ctx, uint32_t *status) +{ + mpd_ssize_t adjexp, etiny, shift; + int rnd; + + adjexp = mpd_adjexp(dec); + if (adjexp > ctx->emax) { + + if (mpd_iszerocoeff(dec)) { + dec->exp = ctx->emax; + if (ctx->clamp) { + dec->exp -= (ctx->prec-1); + } + mpd_zerocoeff(dec); + *status |= MPD_Clamped; + return; + } + + switch (ctx->round) { + case MPD_ROUND_HALF_UP: case MPD_ROUND_HALF_EVEN: + case MPD_ROUND_HALF_DOWN: case MPD_ROUND_UP: + case MPD_ROUND_TRUNC: + mpd_setspecial(dec, mpd_sign(dec), MPD_INF); + break; + case MPD_ROUND_DOWN: case MPD_ROUND_05UP: + mpd_qmaxcoeff(dec, ctx, status); + dec->exp = ctx->emax - ctx->prec + 1; + break; + case MPD_ROUND_CEILING: + if (mpd_isnegative(dec)) { + mpd_qmaxcoeff(dec, ctx, status); + dec->exp = ctx->emax - ctx->prec + 1; + } + else { + mpd_setspecial(dec, MPD_POS, MPD_INF); + } + break; + case MPD_ROUND_FLOOR: + if (mpd_ispositive(dec)) { + mpd_qmaxcoeff(dec, ctx, status); + dec->exp = ctx->emax - ctx->prec + 1; + } + else { + mpd_setspecial(dec, MPD_NEG, MPD_INF); + } + break; + default: /* debug */ + abort(); /* GCOV_NOT_REACHED */ + } + + *status |= MPD_Overflow|MPD_Inexact|MPD_Rounded; + + } /* fold down */ + else if (ctx->clamp && dec->exp > mpd_etop(ctx)) { + /* At this point adjexp=exp+digits-1 <= emax and exp > etop=emax-prec+1: + * (1) shift = exp -emax+prec-1 > 0 + * (2) digits+shift = exp+digits-1 - emax + prec <= prec */ + shift = dec->exp - mpd_etop(ctx); + if (!mpd_qshiftl(dec, dec, shift, status)) { + return; + } + dec->exp -= shift; + *status |= MPD_Clamped; + if (!mpd_iszerocoeff(dec) && adjexp < ctx->emin) { + /* Underflow is impossible, since exp < etiny=emin-prec+1 + * and exp > etop=emax-prec+1 would imply emax < emin. */ + *status |= MPD_Subnormal; + } + } + else if (adjexp < ctx->emin) { + + etiny = mpd_etiny(ctx); + + if (mpd_iszerocoeff(dec)) { + if (dec->exp < etiny) { + dec->exp = etiny; + mpd_zerocoeff(dec); + *status |= MPD_Clamped; + } + return; + } + + *status |= MPD_Subnormal; + if (dec->exp < etiny) { + /* At this point adjexp=exp+digits-1 < emin and exp < etiny=emin-prec+1: + * (1) shift = emin-prec+1 - exp > 0 + * (2) digits-shift = exp+digits-1 - emin + prec < prec */ + shift = etiny - dec->exp; + rnd = (int)mpd_qshiftr_inplace(dec, shift); + dec->exp = etiny; + /* We always have a spare digit in case of an increment. */ + _mpd_apply_round_excess(dec, rnd, ctx, status); + *status |= MPD_Rounded; + if (rnd) { + *status |= (MPD_Inexact|MPD_Underflow); + if (mpd_iszerocoeff(dec)) { + mpd_zerocoeff(dec); + *status |= MPD_Clamped; + } + } + } + /* Case exp >= etiny=emin-prec+1: + * (1) adjexp=exp+digits-1 < emin + * (2) digits < emin-exp+1 <= prec */ + } +} + +/* Transcendental functions do not always set Underflow reliably, + * since they only use as much precision as is necessary for correct + * rounding. If a result like 1.0000000000e-101 is finalized, there + * is no rounding digit that would trigger Underflow. But we can + * assume Inexact, so a short check suffices. */ +static inline void +mpd_check_underflow(mpd_t *dec, const mpd_context_t *ctx, uint32_t *status) +{ + if (mpd_adjexp(dec) < ctx->emin && !mpd_iszero(dec) && + dec->exp < mpd_etiny(ctx)) { + *status |= MPD_Underflow; + } +} + +/* Check if a normal number must be rounded after the exponent has been checked. */ +static inline void +_mpd_check_round(mpd_t *dec, const mpd_context_t *ctx, uint32_t *status) +{ + mpd_uint_t rnd; + mpd_ssize_t shift; + + /* must handle specials: _mpd_check_exp() can produce infinities or NaNs */ + if (mpd_isspecial(dec)) { + return; + } + + if (dec->digits > ctx->prec) { + shift = dec->digits - ctx->prec; + rnd = mpd_qshiftr_inplace(dec, shift); + dec->exp += shift; + _mpd_apply_round(dec, rnd, ctx, status); + *status |= MPD_Rounded; + if (rnd) { + *status |= MPD_Inexact; + } + } +} + +/* Finalize all operations. */ +void +mpd_qfinalize(mpd_t *result, const mpd_context_t *ctx, uint32_t *status) +{ + if (mpd_isspecial(result)) { + if (mpd_isnan(result)) { + _mpd_fix_nan(result, ctx); + } + return; + } + + _mpd_check_exp(result, ctx, status); + _mpd_check_round(result, ctx, status); +} + + +/******************************************************************************/ +/* Copying */ +/******************************************************************************/ + +/* Internal function: Copy a decimal, share data with src: USE WITH CARE! */ +static inline void +_mpd_copy_shared(mpd_t *dest, const mpd_t *src) +{ + dest->flags = src->flags; + dest->exp = src->exp; + dest->digits = src->digits; + dest->len = src->len; + dest->alloc = src->alloc; + dest->data = src->data; + + mpd_set_shared_data(dest); +} + +/* + * Copy a decimal. In case of an error, status is set to MPD_Malloc_error. + */ +int +mpd_qcopy(mpd_t *result, const mpd_t *a, uint32_t *status) +{ + if (result == a) return 1; + + if (!mpd_qresize(result, a->len, status)) { + return 0; + } + + mpd_copy_flags(result, a); + result->exp = a->exp; + result->digits = a->digits; + result->len = a->len; + memcpy(result->data, a->data, a->len * (sizeof *result->data)); + + return 1; +} + +/* + * Copy to a decimal with a static buffer. The caller has to make sure that + * the buffer is big enough. Cannot fail. + */ +static void +mpd_qcopy_static(mpd_t *result, const mpd_t *a) +{ + if (result == a) return; + + memcpy(result->data, a->data, a->len * (sizeof *result->data)); + + mpd_copy_flags(result, a); + result->exp = a->exp; + result->digits = a->digits; + result->len = a->len; +} + +/* + * Return a newly allocated copy of the operand. In case of an error, + * status is set to MPD_Malloc_error and the return value is NULL. + */ +mpd_t * +mpd_qncopy(const mpd_t *a) +{ + mpd_t *result; + + if ((result = mpd_qnew_size(a->len)) == NULL) { + return NULL; + } + memcpy(result->data, a->data, a->len * (sizeof *result->data)); + mpd_copy_flags(result, a); + result->exp = a->exp; + result->digits = a->digits; + result->len = a->len; + + return result; +} + +/* + * Copy a decimal and set the sign to positive. In case of an error, the + * status is set to MPD_Malloc_error. + */ +int +mpd_qcopy_abs(mpd_t *result, const mpd_t *a, uint32_t *status) +{ + if (!mpd_qcopy(result, a, status)) { + return 0; + } + mpd_set_positive(result); + return 1; +} + +/* + * Copy a decimal and negate the sign. In case of an error, the + * status is set to MPD_Malloc_error. + */ +int +mpd_qcopy_negate(mpd_t *result, const mpd_t *a, uint32_t *status) +{ + if (!mpd_qcopy(result, a, status)) { + return 0; + } + _mpd_negate(result); + return 1; +} + +/* + * Copy a decimal, setting the sign of the first operand to the sign of the + * second operand. In case of an error, the status is set to MPD_Malloc_error. + */ +int +mpd_qcopy_sign(mpd_t *result, const mpd_t *a, const mpd_t *b, uint32_t *status) +{ + uint8_t sign_b = mpd_sign(b); /* result may equal b! */ + + if (!mpd_qcopy(result, a, status)) { + return 0; + } + mpd_set_sign(result, sign_b); + return 1; +} + + +/******************************************************************************/ +/* Comparisons */ +/******************************************************************************/ + +/* + * For all functions that compare two operands and return an int the usual + * convention applies to the return value: + * + * -1 if op1 < op2 + * 0 if op1 == op2 + * 1 if op1 > op2 + * + * INT_MAX for error + */ + + +/* Convenience macro. If a and b are not equal, return from the calling + * function with the correct comparison value. */ +#define CMP_EQUAL_OR_RETURN(a, b) \ + if (a != b) { \ + if (a < b) { \ + return -1; \ + } \ + return 1; \ + } + +/* + * Compare the data of big and small. This function does the equivalent + * of first shifting small to the left and then comparing the data of + * big and small, except that no allocation for the left shift is needed. + */ +static int +_mpd_basecmp(mpd_uint_t *big, mpd_uint_t *small, mpd_size_t n, mpd_size_t m, + mpd_size_t shift) +{ +#if defined(__GNUC__) && !defined(__INTEL_COMPILER) && !defined(__clang__) + /* spurious uninitialized warnings */ + mpd_uint_t l=l, lprev=lprev, h=h; +#else + mpd_uint_t l, lprev, h; +#endif + mpd_uint_t q, r; + mpd_uint_t ph, x; + + assert(m > 0 && n >= m && shift > 0); + + _mpd_div_word(&q, &r, (mpd_uint_t)shift, MPD_RDIGITS); + + if (r != 0) { + + ph = mpd_pow10[r]; + + --m; --n; + _mpd_divmod_pow10(&h, &lprev, small[m--], MPD_RDIGITS-r); + if (h != 0) { + CMP_EQUAL_OR_RETURN(big[n], h) + --n; + } + for (; m != MPD_SIZE_MAX; m--,n--) { + _mpd_divmod_pow10(&h, &l, small[m], MPD_RDIGITS-r); + x = ph * lprev + h; + CMP_EQUAL_OR_RETURN(big[n], x) + lprev = l; + } + x = ph * lprev; + CMP_EQUAL_OR_RETURN(big[q], x) + } + else { + while (--m != MPD_SIZE_MAX) { + CMP_EQUAL_OR_RETURN(big[m+q], small[m]) + } + } + + return !_mpd_isallzero(big, q); +} + +/* Compare two decimals with the same adjusted exponent. */ +static int +_mpd_cmp_same_adjexp(const mpd_t *a, const mpd_t *b) +{ + mpd_ssize_t shift, i; + + if (a->exp != b->exp) { + /* Cannot wrap: a->exp + a->digits = b->exp + b->digits, so + * a->exp - b->exp = b->digits - a->digits. */ + shift = a->exp - b->exp; + if (shift > 0) { + return -1 * _mpd_basecmp(b->data, a->data, b->len, a->len, shift); + } + else { + return _mpd_basecmp(a->data, b->data, a->len, b->len, -shift); + } + } + + /* + * At this point adjexp(a) == adjexp(b) and a->exp == b->exp, + * so a->digits == b->digits, therefore a->len == b->len. + */ + for (i = a->len-1; i >= 0; --i) { + CMP_EQUAL_OR_RETURN(a->data[i], b->data[i]) + } + + return 0; +} + +/* Compare two numerical values. */ +static int +_mpd_cmp(const mpd_t *a, const mpd_t *b) +{ + mpd_ssize_t adjexp_a, adjexp_b; + + /* equal pointers */ + if (a == b) { + return 0; + } + + /* infinities */ + if (mpd_isinfinite(a)) { + if (mpd_isinfinite(b)) { + return mpd_isnegative(b) - mpd_isnegative(a); + } + return mpd_arith_sign(a); + } + if (mpd_isinfinite(b)) { + return -mpd_arith_sign(b); + } + + /* zeros */ + if (mpd_iszerocoeff(a)) { + if (mpd_iszerocoeff(b)) { + return 0; + } + return -mpd_arith_sign(b); + } + if (mpd_iszerocoeff(b)) { + return mpd_arith_sign(a); + } + + /* different signs */ + if (mpd_sign(a) != mpd_sign(b)) { + return mpd_sign(b) - mpd_sign(a); + } + + /* different adjusted exponents */ + adjexp_a = mpd_adjexp(a); + adjexp_b = mpd_adjexp(b); + if (adjexp_a != adjexp_b) { + if (adjexp_a < adjexp_b) { + return -1 * mpd_arith_sign(a); + } + return mpd_arith_sign(a); + } + + /* same adjusted exponents */ + return _mpd_cmp_same_adjexp(a, b) * mpd_arith_sign(a); +} + +/* Compare the absolutes of two numerical values. */ +static int +_mpd_cmp_abs(const mpd_t *a, const mpd_t *b) +{ + mpd_ssize_t adjexp_a, adjexp_b; + + /* equal pointers */ + if (a == b) { + return 0; + } + + /* infinities */ + if (mpd_isinfinite(a)) { + if (mpd_isinfinite(b)) { + return 0; + } + return 1; + } + if (mpd_isinfinite(b)) { + return -1; + } + + /* zeros */ + if (mpd_iszerocoeff(a)) { + if (mpd_iszerocoeff(b)) { + return 0; + } + return -1; + } + if (mpd_iszerocoeff(b)) { + return 1; + } + + /* different adjusted exponents */ + adjexp_a = mpd_adjexp(a); + adjexp_b = mpd_adjexp(b); + if (adjexp_a != adjexp_b) { + if (adjexp_a < adjexp_b) { + return -1; + } + return 1; + } + + /* same adjusted exponents */ + return _mpd_cmp_same_adjexp(a, b); +} + +/* Compare two values and return an integer result. */ +int +mpd_qcmp(const mpd_t *a, const mpd_t *b, uint32_t *status) +{ + if (mpd_isspecial(a) || mpd_isspecial(b)) { + if (mpd_isnan(a) || mpd_isnan(b)) { + *status |= MPD_Invalid_operation; + return INT_MAX; + } + } + + return _mpd_cmp(a, b); +} + +/* + * Compare a and b, convert the the usual integer result to a decimal and + * store it in 'result'. For convenience, the integer result of the comparison + * is returned. Comparisons involving NaNs return NaN/INT_MAX. + */ +int +mpd_qcompare(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + int c; + + if (mpd_isspecial(a) || mpd_isspecial(b)) { + if (mpd_qcheck_nans(result, a, b, ctx, status)) { + return INT_MAX; + } + } + + c = _mpd_cmp(a, b); + _settriple(result, (c < 0), (c != 0), 0); + return c; +} + +/* Same as mpd_compare(), but signal for all NaNs, i.e. also for quiet NaNs. */ +int +mpd_qcompare_signal(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + int c; + + if (mpd_isspecial(a) || mpd_isspecial(b)) { + if (mpd_qcheck_nans(result, a, b, ctx, status)) { + *status |= MPD_Invalid_operation; + return INT_MAX; + } + } + + c = _mpd_cmp(a, b); + _settriple(result, (c < 0), (c != 0), 0); + return c; +} + +/* Compare the operands using a total order. */ +int +mpd_cmp_total(const mpd_t *a, const mpd_t *b) +{ + mpd_t aa, bb; + int nan_a, nan_b; + int c; + + if (mpd_sign(a) != mpd_sign(b)) { + return mpd_sign(b) - mpd_sign(a); + } + + + if (mpd_isnan(a)) { + c = 1; + if (mpd_isnan(b)) { + nan_a = (mpd_isqnan(a)) ? 1 : 0; + nan_b = (mpd_isqnan(b)) ? 1 : 0; + if (nan_b == nan_a) { + if (a->len > 0 && b->len > 0) { + _mpd_copy_shared(&aa, a); + _mpd_copy_shared(&bb, b); + aa.exp = bb.exp = 0; + /* compare payload */ + c = _mpd_cmp_abs(&aa, &bb); + } + else { + c = (a->len > 0) - (b->len > 0); + } + } + else { + c = nan_a - nan_b; + } + } + } + else if (mpd_isnan(b)) { + c = -1; + } + else { + c = _mpd_cmp_abs(a, b); + if (c == 0 && a->exp != b->exp) { + c = (a->exp < b->exp) ? -1 : 1; + } + } + + return c * mpd_arith_sign(a); +} + +/* + * Compare a and b according to a total order, convert the usual integer result + * to a decimal and store it in 'result'. For convenience, the integer result + * of the comparison is returned. + */ +int +mpd_compare_total(mpd_t *result, const mpd_t *a, const mpd_t *b) +{ + int c; + + c = mpd_cmp_total(a, b); + _settriple(result, (c < 0), (c != 0), 0); + return c; +} + +/* Compare the magnitude of the operands using a total order. */ +int +mpd_cmp_total_mag(const mpd_t *a, const mpd_t *b) +{ + mpd_t aa, bb; + + _mpd_copy_shared(&aa, a); + _mpd_copy_shared(&bb, b); + + mpd_set_positive(&aa); + mpd_set_positive(&bb); + + return mpd_cmp_total(&aa, &bb); +} + +/* + * Compare the magnitude of a and b according to a total order, convert the + * the usual integer result to a decimal and store it in 'result'. + * For convenience, the integer result of the comparison is returned. + */ +int +mpd_compare_total_mag(mpd_t *result, const mpd_t *a, const mpd_t *b) +{ + int c; + + c = mpd_cmp_total_mag(a, b); + _settriple(result, (c < 0), (c != 0), 0); + return c; +} + +/* Determine an ordering for operands that are numerically equal. */ +static inline int +_mpd_cmp_numequal(const mpd_t *a, const mpd_t *b) +{ + int sign_a, sign_b; + int c; + + sign_a = mpd_sign(a); + sign_b = mpd_sign(b); + if (sign_a != sign_b) { + c = sign_b - sign_a; + } + else { + c = (a->exp < b->exp) ? -1 : 1; + c *= mpd_arith_sign(a); + } + + return c; +} + + +/******************************************************************************/ +/* Shifting the coefficient */ +/******************************************************************************/ + +/* + * Shift the coefficient of the operand to the left, no check for specials. + * Both operands may be the same pointer. If the result length has to be + * increased, mpd_qresize() might fail with MPD_Malloc_error. + */ +int +mpd_qshiftl(mpd_t *result, const mpd_t *a, mpd_ssize_t n, uint32_t *status) +{ + mpd_ssize_t size; + + assert(!mpd_isspecial(a)); + assert(n >= 0); + + if (mpd_iszerocoeff(a) || n == 0) { + return mpd_qcopy(result, a, status); + } + + size = mpd_digits_to_size(a->digits+n); + if (!mpd_qresize(result, size, status)) { + return 0; /* result is NaN */ + } + + _mpd_baseshiftl(result->data, a->data, size, a->len, n); + + mpd_copy_flags(result, a); + result->exp = a->exp; + result->digits = a->digits+n; + result->len = size; + + return 1; +} + +/* Determine the rounding indicator if all digits of the coefficient are shifted + * out of the picture. */ +static mpd_uint_t +_mpd_get_rnd(const mpd_uint_t *data, mpd_ssize_t len, int use_msd) +{ + mpd_uint_t rnd = 0, rest = 0, word; + + word = data[len-1]; + /* special treatment for the most significant digit if shift == digits */ + if (use_msd) { + _mpd_divmod_pow10(&rnd, &rest, word, mpd_word_digits(word)-1); + if (len > 1 && rest == 0) { + rest = !_mpd_isallzero(data, len-1); + } + } + else { + rest = !_mpd_isallzero(data, len); + } + + return (rnd == 0 || rnd == 5) ? rnd + !!rest : rnd; +} + +/* + * Same as mpd_qshiftr(), but 'result' is an mpd_t with a static coefficient. + * It is the caller's responsibility to ensure that the coefficient is big + * enough. The function cannot fail. + */ +mpd_uint_t +mpd_qsshiftr(mpd_t *result, const mpd_t *a, mpd_ssize_t n) +{ + mpd_uint_t rnd; + mpd_ssize_t size; + + assert(!mpd_isspecial(a)); + assert(n >= 0); + + if (mpd_iszerocoeff(a) || n == 0) { + mpd_qcopy_static(result, a); + return 0; + } + + if (n >= a->digits) { + rnd = _mpd_get_rnd(a->data, a->len, (n==a->digits)); + mpd_zerocoeff(result); + } + else { + result->digits = a->digits-n; + size = mpd_digits_to_size(result->digits); + rnd = _mpd_baseshiftr(result->data, a->data, a->len, n); + result->len = size; + } + + mpd_copy_flags(result, a); + result->exp = a->exp; + + return rnd; +} + +/* + * Inplace shift of the coefficient to the right, no check for specials. + * Returns the rounding indicator for mpd_rnd_incr(). + * The function cannot fail. + */ +mpd_uint_t +mpd_qshiftr_inplace(mpd_t *result, mpd_ssize_t n) +{ + uint32_t dummy; + mpd_uint_t rnd; + mpd_ssize_t size; + + assert(!mpd_isspecial(result)); + assert(n >= 0); + + if (mpd_iszerocoeff(result) || n == 0) { + return 0; + } + + if (n >= result->digits) { + rnd = _mpd_get_rnd(result->data, result->len, (n==result->digits)); + mpd_zerocoeff(result); + } + else { + rnd = _mpd_baseshiftr(result->data, result->data, result->len, n); + result->digits -= n; + size = mpd_digits_to_size(result->digits); + /* reducing the size cannot fail */ + mpd_qresize(result, size, &dummy); + result->len = size; + } + + return rnd; +} + +/* + * Shift the coefficient of the operand to the right, no check for specials. + * Both operands may be the same pointer. Returns the rounding indicator to + * be used by mpd_rnd_incr(). If the result length has to be increased, + * mpd_qcopy() or mpd_qresize() might fail with MPD_Malloc_error. In those + * cases, MPD_UINT_MAX is returned. + */ +mpd_uint_t +mpd_qshiftr(mpd_t *result, const mpd_t *a, mpd_ssize_t n, uint32_t *status) +{ + mpd_uint_t rnd; + mpd_ssize_t size; + + assert(!mpd_isspecial(a)); + assert(n >= 0); + + if (mpd_iszerocoeff(a) || n == 0) { + if (!mpd_qcopy(result, a, status)) { + return MPD_UINT_MAX; + } + return 0; + } + + if (n >= a->digits) { + rnd = _mpd_get_rnd(a->data, a->len, (n==a->digits)); + mpd_zerocoeff(result); + } + else { + result->digits = a->digits-n; + size = mpd_digits_to_size(result->digits); + if (result == a) { + rnd = _mpd_baseshiftr(result->data, a->data, a->len, n); + /* reducing the size cannot fail */ + mpd_qresize(result, size, status); + } + else { + if (!mpd_qresize(result, size, status)) { + return MPD_UINT_MAX; + } + rnd = _mpd_baseshiftr(result->data, a->data, a->len, n); + } + result->len = size; + } + + mpd_copy_flags(result, a); + result->exp = a->exp; + + return rnd; +} + + +/******************************************************************************/ +/* Miscellaneous operations */ +/******************************************************************************/ + +/* Logical And */ +void +mpd_qand(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + const mpd_t *big = a, *small = b; + mpd_uint_t x, y, z, xbit, ybit; + int k, mswdigits; + mpd_ssize_t i; + + if (mpd_isspecial(a) || mpd_isspecial(b) || + mpd_isnegative(a) || mpd_isnegative(b) || + a->exp != 0 || b->exp != 0) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if (b->digits > a->digits) { + big = b; + small = a; + } + if (!mpd_qresize(result, big->len, status)) { + return; + } + + + /* full words */ + for (i = 0; i < small->len-1; i++) { + x = small->data[i]; + y = big->data[i]; + z = 0; + for (k = 0; k < MPD_RDIGITS; k++) { + xbit = x % 10; + x /= 10; + ybit = y % 10; + y /= 10; + if (xbit > 1 || ybit > 1) { + goto invalid_operation; + } + z += (xbit&ybit) ? mpd_pow10[k] : 0; + } + result->data[i] = z; + } + /* most significant word of small */ + x = small->data[i]; + y = big->data[i]; + z = 0; + mswdigits = mpd_word_digits(x); + for (k = 0; k < mswdigits; k++) { + xbit = x % 10; + x /= 10; + ybit = y % 10; + y /= 10; + if (xbit > 1 || ybit > 1) { + goto invalid_operation; + } + z += (xbit&ybit) ? mpd_pow10[k] : 0; + } + result->data[i++] = z; + + /* scan the rest of y for digits > 1 */ + for (; k < MPD_RDIGITS; k++) { + ybit = y % 10; + y /= 10; + if (ybit > 1) { + goto invalid_operation; + } + } + /* scan the rest of big for digits > 1 */ + for (; i < big->len; i++) { + y = big->data[i]; + for (k = 0; k < MPD_RDIGITS; k++) { + ybit = y % 10; + y /= 10; + if (ybit > 1) { + goto invalid_operation; + } + } + } + + mpd_clear_flags(result); + result->exp = 0; + result->len = _mpd_real_size(result->data, small->len); + mpd_qresize(result, result->len, status); + mpd_setdigits(result); + _mpd_cap(result, ctx); + return; + +invalid_operation: + mpd_seterror(result, MPD_Invalid_operation, status); +} + +/* Class of an operand. Returns a pointer to the constant name. */ +const char * +mpd_class(const mpd_t *a, const mpd_context_t *ctx) +{ + if (mpd_isnan(a)) { + if (mpd_isqnan(a)) + return "NaN"; + else + return "sNaN"; + } + else if (mpd_ispositive(a)) { + if (mpd_isinfinite(a)) + return "+Infinity"; + else if (mpd_iszero(a)) + return "+Zero"; + else if (mpd_isnormal(a, ctx)) + return "+Normal"; + else + return "+Subnormal"; + } + else { + if (mpd_isinfinite(a)) + return "-Infinity"; + else if (mpd_iszero(a)) + return "-Zero"; + else if (mpd_isnormal(a, ctx)) + return "-Normal"; + else + return "-Subnormal"; + } +} + +/* Logical Xor */ +void +mpd_qinvert(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_uint_t x, z, xbit; + mpd_ssize_t i, digits, len; + mpd_ssize_t q, r; + int k; + + if (mpd_isspecial(a) || mpd_isnegative(a) || a->exp != 0) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + + digits = (a->digits < ctx->prec) ? ctx->prec : a->digits; + _mpd_idiv_word(&q, &r, digits, MPD_RDIGITS); + len = (r == 0) ? q : q+1; + if (!mpd_qresize(result, len, status)) { + return; + } + + for (i = 0; i < len; i++) { + x = (i < a->len) ? a->data[i] : 0; + z = 0; + for (k = 0; k < MPD_RDIGITS; k++) { + xbit = x % 10; + x /= 10; + if (xbit > 1) { + goto invalid_operation; + } + z += !xbit ? mpd_pow10[k] : 0; + } + result->data[i] = z; + } + + mpd_clear_flags(result); + result->exp = 0; + result->len = _mpd_real_size(result->data, len); + mpd_qresize(result, result->len, status); + mpd_setdigits(result); + _mpd_cap(result, ctx); + return; + +invalid_operation: + mpd_seterror(result, MPD_Invalid_operation, status); +} + +/* Exponent of the magnitude of the most significant digit of the operand. */ +void +mpd_qlogb(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + if (mpd_isspecial(a)) { + if (mpd_qcheck_nan(result, a, ctx, status)) { + return; + } + mpd_setspecial(result, MPD_POS, MPD_INF); + } + else if (mpd_iszerocoeff(a)) { + mpd_setspecial(result, MPD_NEG, MPD_INF); + *status |= MPD_Division_by_zero; + } + else { + mpd_qset_ssize(result, mpd_adjexp(a), ctx, status); + } +} + +/* Logical Or */ +void +mpd_qor(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + const mpd_t *big = a, *small = b; + mpd_uint_t x, y, z, xbit, ybit; + int k, mswdigits; + mpd_ssize_t i; + + if (mpd_isspecial(a) || mpd_isspecial(b) || + mpd_isnegative(a) || mpd_isnegative(b) || + a->exp != 0 || b->exp != 0) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if (b->digits > a->digits) { + big = b; + small = a; + } + if (!mpd_qresize(result, big->len, status)) { + return; + } + + + /* full words */ + for (i = 0; i < small->len-1; i++) { + x = small->data[i]; + y = big->data[i]; + z = 0; + for (k = 0; k < MPD_RDIGITS; k++) { + xbit = x % 10; + x /= 10; + ybit = y % 10; + y /= 10; + if (xbit > 1 || ybit > 1) { + goto invalid_operation; + } + z += (xbit|ybit) ? mpd_pow10[k] : 0; + } + result->data[i] = z; + } + /* most significant word of small */ + x = small->data[i]; + y = big->data[i]; + z = 0; + mswdigits = mpd_word_digits(x); + for (k = 0; k < mswdigits; k++) { + xbit = x % 10; + x /= 10; + ybit = y % 10; + y /= 10; + if (xbit > 1 || ybit > 1) { + goto invalid_operation; + } + z += (xbit|ybit) ? mpd_pow10[k] : 0; + } + + /* scan for digits > 1 and copy the rest of y */ + for (; k < MPD_RDIGITS; k++) { + ybit = y % 10; + y /= 10; + if (ybit > 1) { + goto invalid_operation; + } + z += ybit*mpd_pow10[k]; + } + result->data[i++] = z; + /* scan for digits > 1 and copy the rest of big */ + for (; i < big->len; i++) { + y = big->data[i]; + for (k = 0; k < MPD_RDIGITS; k++) { + ybit = y % 10; + y /= 10; + if (ybit > 1) { + goto invalid_operation; + } + } + result->data[i] = big->data[i]; + } + + mpd_clear_flags(result); + result->exp = 0; + result->len = _mpd_real_size(result->data, big->len); + mpd_qresize(result, result->len, status); + mpd_setdigits(result); + _mpd_cap(result, ctx); + return; + +invalid_operation: + mpd_seterror(result, MPD_Invalid_operation, status); +} + +/* + * Rotate the coefficient of 'a' by 'b' digits. 'b' must be an integer with + * exponent 0. + */ +void +mpd_qrotate(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + uint32_t workstatus = 0; + MPD_NEW_STATIC(tmp,0,0,0,0); + MPD_NEW_STATIC(big,0,0,0,0); + MPD_NEW_STATIC(small,0,0,0,0); + mpd_ssize_t n, lshift, rshift; + + if (mpd_isspecial(a) || mpd_isspecial(b)) { + if (mpd_qcheck_nans(result, a, b, ctx, status)) { + return; + } + } + if (b->exp != 0 || mpd_isinfinite(b)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + + n = mpd_qget_ssize(b, &workstatus); + if (workstatus&MPD_Invalid_operation) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if (n > ctx->prec || n < -ctx->prec) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if (mpd_isinfinite(a)) { + mpd_qcopy(result, a, status); + return; + } + + if (n >= 0) { + lshift = n; + rshift = ctx->prec-n; + } + else { + lshift = ctx->prec+n; + rshift = -n; + } + + if (a->digits > ctx->prec) { + if (!mpd_qcopy(&tmp, a, status)) { + mpd_seterror(result, MPD_Malloc_error, status); + goto finish; + } + _mpd_cap(&tmp, ctx); + a = &tmp; + } + + if (!mpd_qshiftl(&big, a, lshift, status)) { + mpd_seterror(result, MPD_Malloc_error, status); + goto finish; + } + _mpd_cap(&big, ctx); + + if (mpd_qshiftr(&small, a, rshift, status) == MPD_UINT_MAX) { + mpd_seterror(result, MPD_Malloc_error, status); + goto finish; + } + _mpd_qadd(result, &big, &small, ctx, status); + + +finish: + mpd_del(&tmp); + mpd_del(&big); + mpd_del(&small); +} + +/* + * b must be an integer with exponent 0 and in the range +-2*(emax + prec). + * XXX: In my opinion +-(2*emax + prec) would be more sensible. + * The result is a with the value of b added to its exponent. + */ +void +mpd_qscaleb(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + uint32_t workstatus = 0; + mpd_uint_t n, maxjump; +#ifndef LEGACY_COMPILER + int64_t exp; +#else + mpd_uint_t x; + int x_sign, n_sign; + mpd_ssize_t exp; +#endif + + if (mpd_isspecial(a) || mpd_isspecial(b)) { + if (mpd_qcheck_nans(result, a, b, ctx, status)) { + return; + } + } + if (b->exp != 0 || mpd_isinfinite(b)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + + n = mpd_qabs_uint(b, &workstatus); + /* the spec demands this */ + maxjump = 2 * (mpd_uint_t)(ctx->emax + ctx->prec); + + if (n > maxjump || workstatus&MPD_Invalid_operation) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if (mpd_isinfinite(a)) { + mpd_qcopy(result, a, status); + return; + } + +#ifndef LEGACY_COMPILER + exp = a->exp + (int64_t)n * mpd_arith_sign(b); + exp = (exp > MPD_EXP_INF) ? MPD_EXP_INF : exp; + exp = (exp < MPD_EXP_CLAMP) ? MPD_EXP_CLAMP : exp; +#else + x = (a->exp < 0) ? -a->exp : a->exp; + x_sign = (a->exp < 0) ? 1 : 0; + n_sign = mpd_isnegative(b) ? 1 : 0; + + if (x_sign == n_sign) { + x = x + n; + if (x < n) x = MPD_UINT_MAX; + } + else { + x_sign = (x >= n) ? x_sign : n_sign; + x = (x >= n) ? x - n : n - x; + } + if (!x_sign && x > MPD_EXP_INF) x = MPD_EXP_INF; + if (x_sign && x > -MPD_EXP_CLAMP) x = -MPD_EXP_CLAMP; + exp = x_sign ? -((mpd_ssize_t)x) : (mpd_ssize_t)x; +#endif + + mpd_qcopy(result, a, status); + result->exp = (mpd_ssize_t)exp; + + mpd_qfinalize(result, ctx, status); +} + +/* + * Shift the coefficient by n digits, positive n is a left shift. In the case + * of a left shift, the result is decapitated to fit the context precision. If + * you don't want that, use mpd_shiftl(). + */ +void +mpd_qshiftn(mpd_t *result, const mpd_t *a, mpd_ssize_t n, const mpd_context_t *ctx, + uint32_t *status) +{ + if (mpd_isspecial(a)) { + if (mpd_qcheck_nan(result, a, ctx, status)) { + return; + } + mpd_qcopy(result, a, status); + return; + } + + if (n >= 0 && n <= ctx->prec) { + mpd_qshiftl(result, a, n, status); + _mpd_cap(result, ctx); + } + else if (n < 0 && n >= -ctx->prec) { + if (!mpd_qcopy(result, a, status)) { + return; + } + _mpd_cap(result, ctx); + mpd_qshiftr_inplace(result, -n); + } + else { + mpd_seterror(result, MPD_Invalid_operation, status); + } +} + +/* + * Same as mpd_shiftn(), but the shift is specified by the decimal b, which + * must be an integer with a zero exponent. Infinities remain infinities. + */ +void +mpd_qshift(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, + uint32_t *status) +{ + uint32_t workstatus = 0; + mpd_ssize_t n; + + if (mpd_isspecial(a) || mpd_isspecial(b)) { + if (mpd_qcheck_nans(result, a, b, ctx, status)) { + return; + } + } + if (b->exp != 0 || mpd_isinfinite(b)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + + n = mpd_qget_ssize(b, &workstatus); + if (workstatus&MPD_Invalid_operation) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if (n > ctx->prec || n < -ctx->prec) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if (mpd_isinfinite(a)) { + mpd_qcopy(result, a, status); + return; + } + + if (n >= 0) { + mpd_qshiftl(result, a, n, status); + _mpd_cap(result, ctx); + } + else { + if (!mpd_qcopy(result, a, status)) { + return; + } + _mpd_cap(result, ctx); + mpd_qshiftr_inplace(result, -n); + } +} + +/* Logical Xor */ +void +mpd_qxor(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + const mpd_t *big = a, *small = b; + mpd_uint_t x, y, z, xbit, ybit; + int k, mswdigits; + mpd_ssize_t i; + + if (mpd_isspecial(a) || mpd_isspecial(b) || + mpd_isnegative(a) || mpd_isnegative(b) || + a->exp != 0 || b->exp != 0) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if (b->digits > a->digits) { + big = b; + small = a; + } + if (!mpd_qresize(result, big->len, status)) { + return; + } + + + /* full words */ + for (i = 0; i < small->len-1; i++) { + x = small->data[i]; + y = big->data[i]; + z = 0; + for (k = 0; k < MPD_RDIGITS; k++) { + xbit = x % 10; + x /= 10; + ybit = y % 10; + y /= 10; + if (xbit > 1 || ybit > 1) { + goto invalid_operation; + } + z += (xbit^ybit) ? mpd_pow10[k] : 0; + } + result->data[i] = z; + } + /* most significant word of small */ + x = small->data[i]; + y = big->data[i]; + z = 0; + mswdigits = mpd_word_digits(x); + for (k = 0; k < mswdigits; k++) { + xbit = x % 10; + x /= 10; + ybit = y % 10; + y /= 10; + if (xbit > 1 || ybit > 1) { + goto invalid_operation; + } + z += (xbit^ybit) ? mpd_pow10[k] : 0; + } + + /* scan for digits > 1 and copy the rest of y */ + for (; k < MPD_RDIGITS; k++) { + ybit = y % 10; + y /= 10; + if (ybit > 1) { + goto invalid_operation; + } + z += ybit*mpd_pow10[k]; + } + result->data[i++] = z; + /* scan for digits > 1 and copy the rest of big */ + for (; i < big->len; i++) { + y = big->data[i]; + for (k = 0; k < MPD_RDIGITS; k++) { + ybit = y % 10; + y /= 10; + if (ybit > 1) { + goto invalid_operation; + } + } + result->data[i] = big->data[i]; + } + + mpd_clear_flags(result); + result->exp = 0; + result->len = _mpd_real_size(result->data, big->len); + mpd_qresize(result, result->len, status); + mpd_setdigits(result); + _mpd_cap(result, ctx); + return; + +invalid_operation: + mpd_seterror(result, MPD_Invalid_operation, status); +} + + +/******************************************************************************/ +/* Arithmetic operations */ +/******************************************************************************/ + +/* + * The absolute value of a. If a is negative, the result is the same + * as the result of the minus operation. Otherwise, the result is the + * result of the plus operation. + */ +void +mpd_qabs(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + if (mpd_isspecial(a)) { + if (mpd_qcheck_nan(result, a, ctx, status)) { + return; + } + } + + if (mpd_isnegative(a)) { + mpd_qminus(result, a, ctx, status); + } + else { + mpd_qplus(result, a, ctx, status); + } +} + +static inline void +_mpd_ptrswap(mpd_t **a, mpd_t **b) +{ + mpd_t *t = *a; + *a = *b; + *b = t; +} + +/* Add or subtract infinities. */ +static void +_mpd_qaddsub_inf(mpd_t *result, const mpd_t *a, const mpd_t *b, uint8_t sign_b, + uint32_t *status) +{ + if (mpd_isinfinite(a)) { + if (mpd_sign(a) != sign_b && mpd_isinfinite(b)) { + mpd_seterror(result, MPD_Invalid_operation, status); + } + else { + mpd_setspecial(result, mpd_sign(a), MPD_INF); + } + return; + } + assert(mpd_isinfinite(b)); + mpd_setspecial(result, sign_b, MPD_INF); +} + +/* Add or subtract non-special numbers. */ +static void +_mpd_qaddsub(mpd_t *result, const mpd_t *a, const mpd_t *b, uint8_t sign_b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_t *big, *small; + MPD_NEW_STATIC(big_aligned,0,0,0,0); + MPD_NEW_CONST(tiny,0,0,1,1,1,1); + mpd_uint_t carry; + mpd_ssize_t newsize, shift; + mpd_ssize_t exp, i; + int swap = 0; + + + /* compare exponents */ + big = (mpd_t *)a; small = (mpd_t *)b; + if (big->exp != small->exp) { + if (small->exp > big->exp) { + _mpd_ptrswap(&big, &small); + swap++; + } + /* align the coefficients */ + if (!mpd_iszerocoeff(big)) { + exp = big->exp - 1; + exp += (big->digits > ctx->prec) ? 0 : big->digits-ctx->prec-1; + if (mpd_adjexp(small) < exp) { + /* + * Avoid huge shifts by substituting a value for small that is + * guaranteed to produce the same results. + * + * adjexp(small) < exp if and only if: + * + * bdigits <= prec AND + * bdigits+shift >= prec+2+sdigits AND + * exp = bexp+bdigits-prec-2 + * + * 1234567000000000 -> bdigits + shift + * ----------XX1234 -> sdigits + * ----------X1 -> tiny-digits + * |- prec -| + * + * OR + * + * bdigits > prec AND + * shift > sdigits AND + * exp = bexp-1 + * + * 1234567892100000 -> bdigits + shift + * ----------XX1234 -> sdigits + * ----------X1 -> tiny-digits + * |- prec -| + * + * If tiny is zero, adding or subtracting is a no-op. + * Otherwise, adding tiny generates a non-zero digit either + * below the rounding digit or the least significant digit + * of big. When subtracting, tiny is in the same position as + * the carry that would be generated by subtracting sdigits. + */ + mpd_copy_flags(&tiny, small); + tiny.exp = exp; + tiny.digits = 1; + tiny.len = 1; + tiny.data[0] = mpd_iszerocoeff(small) ? 0 : 1; + small = &tiny; + } + /* This cannot wrap: the difference is positive and <= maxprec */ + shift = big->exp - small->exp; + if (!mpd_qshiftl(&big_aligned, big, shift, status)) { + mpd_seterror(result, MPD_Malloc_error, status); + goto finish; + } + big = &big_aligned; + } + } + result->exp = small->exp; + + + /* compare length of coefficients */ + if (big->len < small->len) { + _mpd_ptrswap(&big, &small); + swap++; + } + + newsize = big->len; + if (!mpd_qresize(result, newsize, status)) { + goto finish; + } + + if (mpd_sign(a) == sign_b) { + + carry = _mpd_baseadd(result->data, big->data, small->data, + big->len, small->len); + + if (carry) { + newsize = big->len + 1; + if (!mpd_qresize(result, newsize, status)) { + goto finish; + } + result->data[newsize-1] = carry; + } + + result->len = newsize; + mpd_set_flags(result, sign_b); + } + else { + if (big->len == small->len) { + for (i=big->len-1; i >= 0; --i) { + if (big->data[i] != small->data[i]) { + if (big->data[i] < small->data[i]) { + _mpd_ptrswap(&big, &small); + swap++; + } + break; + } + } + } + + _mpd_basesub(result->data, big->data, small->data, + big->len, small->len); + newsize = _mpd_real_size(result->data, big->len); + /* resize to smaller cannot fail */ + (void)mpd_qresize(result, newsize, status); + + result->len = newsize; + sign_b = (swap & 1) ? sign_b : mpd_sign(a); + mpd_set_flags(result, sign_b); + + if (mpd_iszerocoeff(result)) { + mpd_set_positive(result); + if (ctx->round == MPD_ROUND_FLOOR) { + mpd_set_negative(result); + } + } + } + + mpd_setdigits(result); + +finish: + mpd_del(&big_aligned); +} + +/* Add a and b. No specials, no finalizing. */ +static void +_mpd_qadd(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + _mpd_qaddsub(result, a, b, mpd_sign(b), ctx, status); +} + +/* Subtract b from a. No specials, no finalizing. */ +static void +_mpd_qsub(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + _mpd_qaddsub(result, a, b, !mpd_sign(b), ctx, status); +} + +/* Add a and b. */ +void +mpd_qadd(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + if (mpd_isspecial(a) || mpd_isspecial(b)) { + if (mpd_qcheck_nans(result, a, b, ctx, status)) { + return; + } + _mpd_qaddsub_inf(result, a, b, mpd_sign(b), status); + return; + } + + _mpd_qaddsub(result, a, b, mpd_sign(b), ctx, status); + mpd_qfinalize(result, ctx, status); +} + +/* Subtract b from a. */ +void +mpd_qsub(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + if (mpd_isspecial(a) || mpd_isspecial(b)) { + if (mpd_qcheck_nans(result, a, b, ctx, status)) { + return; + } + _mpd_qaddsub_inf(result, a, b, !mpd_sign(b), status); + return; + } + + _mpd_qaddsub(result, a, b, !mpd_sign(b), ctx, status); + mpd_qfinalize(result, ctx, status); +} + +/* Add decimal and mpd_ssize_t. */ +void +mpd_qadd_ssize(mpd_t *result, const mpd_t *a, mpd_ssize_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_context_t maxcontext; + MPD_NEW_STATIC(bb,0,0,0,0); + + mpd_maxcontext(&maxcontext); + mpd_qsset_ssize(&bb, b, &maxcontext, status); + mpd_qadd(result, a, &bb, ctx, status); + mpd_del(&bb); +} + +/* Add decimal and mpd_uint_t. */ +void +mpd_qadd_uint(mpd_t *result, const mpd_t *a, mpd_uint_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_context_t maxcontext; + MPD_NEW_STATIC(bb,0,0,0,0); + + mpd_maxcontext(&maxcontext); + mpd_qsset_uint(&bb, b, &maxcontext, status); + mpd_qadd(result, a, &bb, ctx, status); + mpd_del(&bb); +} + +/* Subtract mpd_ssize_t from decimal. */ +void +mpd_qsub_ssize(mpd_t *result, const mpd_t *a, mpd_ssize_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_context_t maxcontext; + MPD_NEW_STATIC(bb,0,0,0,0); + + mpd_maxcontext(&maxcontext); + mpd_qsset_ssize(&bb, b, &maxcontext, status); + mpd_qsub(result, a, &bb, ctx, status); + mpd_del(&bb); +} + +/* Subtract mpd_uint_t from decimal. */ +void +mpd_qsub_uint(mpd_t *result, const mpd_t *a, mpd_uint_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_context_t maxcontext; + MPD_NEW_STATIC(bb,0,0,0,0); + + mpd_maxcontext(&maxcontext); + mpd_qsset_uint(&bb, b, &maxcontext, status); + mpd_qsub(result, a, &bb, ctx, status); + mpd_del(&bb); +} + +/* Add decimal and int32_t. */ +void +mpd_qadd_i32(mpd_t *result, const mpd_t *a, int32_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_qadd_ssize(result, a, b, ctx, status); +} + +/* Add decimal and uint32_t. */ +void +mpd_qadd_u32(mpd_t *result, const mpd_t *a, uint32_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_qadd_uint(result, a, b, ctx, status); +} + +#ifdef CONFIG_64 +/* Add decimal and int64_t. */ +void +mpd_qadd_i64(mpd_t *result, const mpd_t *a, int64_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_qadd_ssize(result, a, b, ctx, status); +} + +/* Add decimal and uint64_t. */ +void +mpd_qadd_u64(mpd_t *result, const mpd_t *a, uint64_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_qadd_uint(result, a, b, ctx, status); +} +#endif + +/* Subtract int32_t from decimal. */ +void +mpd_qsub_i32(mpd_t *result, const mpd_t *a, int32_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_qsub_ssize(result, a, b, ctx, status); +} + +/* Subtract uint32_t from decimal. */ +void +mpd_qsub_u32(mpd_t *result, const mpd_t *a, uint32_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_qsub_uint(result, a, b, ctx, status); +} + +#ifdef CONFIG_64 +/* Subtract int64_t from decimal. */ +void +mpd_qsub_i64(mpd_t *result, const mpd_t *a, int64_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_qsub_ssize(result, a, b, ctx, status); +} + +/* Subtract uint64_t from decimal. */ +void +mpd_qsub_u64(mpd_t *result, const mpd_t *a, uint64_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_qsub_uint(result, a, b, ctx, status); +} +#endif + + +/* Divide infinities. */ +static void +_mpd_qdiv_inf(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + if (mpd_isinfinite(a)) { + if (mpd_isinfinite(b)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + mpd_setspecial(result, mpd_sign(a)^mpd_sign(b), MPD_INF); + return; + } + assert(mpd_isinfinite(b)); + _settriple(result, mpd_sign(a)^mpd_sign(b), 0, mpd_etiny(ctx)); + *status |= MPD_Clamped; +} + +enum {NO_IDEAL_EXP, SET_IDEAL_EXP}; +/* Divide a by b. */ +static void +_mpd_qdiv(int action, mpd_t *q, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + MPD_NEW_STATIC(aligned,0,0,0,0); + mpd_uint_t ld; + mpd_ssize_t shift, exp, tz; + mpd_ssize_t newsize; + mpd_ssize_t ideal_exp; + mpd_uint_t rem; + uint8_t sign_a = mpd_sign(a); + uint8_t sign_b = mpd_sign(b); + + + if (mpd_isspecial(a) || mpd_isspecial(b)) { + if (mpd_qcheck_nans(q, a, b, ctx, status)) { + return; + } + _mpd_qdiv_inf(q, a, b, ctx, status); + return; + } + if (mpd_iszerocoeff(b)) { + if (mpd_iszerocoeff(a)) { + mpd_seterror(q, MPD_Division_undefined, status); + } + else { + mpd_setspecial(q, sign_a^sign_b, MPD_INF); + *status |= MPD_Division_by_zero; + } + return; + } + if (mpd_iszerocoeff(a)) { + exp = a->exp - b->exp; + _settriple(q, sign_a^sign_b, 0, exp); + mpd_qfinalize(q, ctx, status); + return; + } + + shift = (b->digits - a->digits) + ctx->prec + 1; + ideal_exp = a->exp - b->exp; + exp = ideal_exp - shift; + if (shift > 0) { + if (!mpd_qshiftl(&aligned, a, shift, status)) { + mpd_seterror(q, MPD_Malloc_error, status); + goto finish; + } + a = &aligned; + } + else if (shift < 0) { + shift = -shift; + if (!mpd_qshiftl(&aligned, b, shift, status)) { + mpd_seterror(q, MPD_Malloc_error, status); + goto finish; + } + b = &aligned; + } + + + newsize = a->len - b->len + 1; + if ((q != b && q != a) || (q == b && newsize > b->len)) { + if (!mpd_qresize(q, newsize, status)) { + mpd_seterror(q, MPD_Malloc_error, status); + goto finish; + } + } + + + if (b->len == 1) { + rem = _mpd_shortdiv(q->data, a->data, a->len, b->data[0]); + } + else if (a->len < 2*MPD_NEWTONDIV_CUTOFF && + b->len < MPD_NEWTONDIV_CUTOFF) { + int ret = _mpd_basedivmod(q->data, NULL, a->data, b->data, + a->len, b->len); + if (ret < 0) { + mpd_seterror(q, MPD_Malloc_error, status); + goto finish; + } + rem = ret; + } + else { + MPD_NEW_STATIC(r,0,0,0,0); + _mpd_qbarrett_divmod(q, &r, a, b, status); + if (mpd_isspecial(q) || mpd_isspecial(&r)) { + mpd_del(&r); + goto finish; + } + rem = !mpd_iszerocoeff(&r); + mpd_del(&r); + newsize = q->len; + } + + newsize = _mpd_real_size(q->data, newsize); + /* resize to smaller cannot fail */ + mpd_qresize(q, newsize, status); + mpd_set_flags(q, sign_a^sign_b); + q->len = newsize; + mpd_setdigits(q); + + shift = ideal_exp - exp; + if (rem) { + ld = mpd_lsd(q->data[0]); + if (ld == 0 || ld == 5) { + q->data[0] += 1; + } + } + else if (action == SET_IDEAL_EXP && shift > 0) { + tz = mpd_trail_zeros(q); + shift = (tz > shift) ? shift : tz; + mpd_qshiftr_inplace(q, shift); + exp += shift; + } + + q->exp = exp; + + +finish: + mpd_del(&aligned); + mpd_qfinalize(q, ctx, status); +} + +/* Divide a by b. */ +void +mpd_qdiv(mpd_t *q, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + _mpd_qdiv(SET_IDEAL_EXP, q, a, b, ctx, status); +} + +/* Internal function. */ +static void +_mpd_qdivmod(mpd_t *q, mpd_t *r, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + MPD_NEW_STATIC(aligned,0,0,0,0); + mpd_ssize_t qsize, rsize; + mpd_ssize_t ideal_exp, expdiff, shift; + uint8_t sign_a = mpd_sign(a); + uint8_t sign_ab = mpd_sign(a)^mpd_sign(b); + + + ideal_exp = (a->exp > b->exp) ? b->exp : a->exp; + if (mpd_iszerocoeff(a)) { + if (!mpd_qcopy(r, a, status)) { + goto nanresult; /* GCOV_NOT_REACHED */ + } + r->exp = ideal_exp; + _settriple(q, sign_ab, 0, 0); + return; + } + + expdiff = mpd_adjexp(a) - mpd_adjexp(b); + if (expdiff < 0) { + if (a->exp > b->exp) { + /* positive and less than b->digits - a->digits */ + shift = a->exp - b->exp; + if (!mpd_qshiftl(r, a, shift, status)) { + goto nanresult; + } + r->exp = ideal_exp; + } + else { + if (!mpd_qcopy(r, a, status)) { + goto nanresult; + } + } + _settriple(q, sign_ab, 0, 0); + return; + } + if (expdiff > ctx->prec) { + *status |= MPD_Division_impossible; + goto nanresult; + } + + + /* + * At this point we have: + * (1) 0 <= a->exp + a->digits - b->exp - b->digits <= prec + * (2) a->exp - b->exp >= b->digits - a->digits + * (3) a->exp - b->exp <= prec + b->digits - a->digits + */ + if (a->exp != b->exp) { + shift = a->exp - b->exp; + if (shift > 0) { + /* by (3), after the shift a->digits <= prec + b->digits */ + if (!mpd_qshiftl(&aligned, a, shift, status)) { + goto nanresult; + } + a = &aligned; + } + else { + shift = -shift; + /* by (2), after the shift b->digits <= a->digits */ + if (!mpd_qshiftl(&aligned, b, shift, status)) { + goto nanresult; + } + b = &aligned; + } + } + + + qsize = a->len - b->len + 1; + if (!(q == a && qsize < a->len) && !(q == b && qsize < b->len)) { + if (!mpd_qresize(q, qsize, status)) { + goto nanresult; + } + } + + rsize = b->len; + if (!(r == a && rsize < a->len)) { + if (!mpd_qresize(r, rsize, status)) { + goto nanresult; + } + } + + if (b->len == 1) { + if (a->len == 1) { + _mpd_div_word(&q->data[0], &r->data[0], a->data[0], b->data[0]); + } + else { + r->data[0] = _mpd_shortdiv(q->data, a->data, a->len, b->data[0]); + } + } + else if (a->len < 2*MPD_NEWTONDIV_CUTOFF && + b->len < MPD_NEWTONDIV_CUTOFF) { + int ret; + ret = _mpd_basedivmod(q->data, r->data, a->data, b->data, + a->len, b->len); + if (ret == -1) { + *status |= MPD_Malloc_error; + goto nanresult; + } + } + else { + _mpd_qbarrett_divmod(q, r, a, b, status); + if (mpd_isspecial(q) || mpd_isspecial(r)) { + goto nanresult; + } + if (mpd_isinfinite(q) || q->digits > ctx->prec) { + *status |= MPD_Division_impossible; + goto nanresult; + } + qsize = q->len; + rsize = r->len; + } + + qsize = _mpd_real_size(q->data, qsize); + /* resize to smaller cannot fail */ + mpd_qresize(q, qsize, status); + q->len = qsize; + mpd_setdigits(q); + mpd_set_flags(q, sign_ab); + q->exp = 0; + if (q->digits > ctx->prec) { + *status |= MPD_Division_impossible; + goto nanresult; + } + + rsize = _mpd_real_size(r->data, rsize); + /* resize to smaller cannot fail */ + mpd_qresize(r, rsize, status); + r->len = rsize; + mpd_setdigits(r); + mpd_set_flags(r, sign_a); + r->exp = ideal_exp; + +out: + mpd_del(&aligned); + return; + +nanresult: + mpd_setspecial(q, MPD_POS, MPD_NAN); + mpd_setspecial(r, MPD_POS, MPD_NAN); + goto out; +} + +/* Integer division with remainder. */ +void +mpd_qdivmod(mpd_t *q, mpd_t *r, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + uint8_t sign = mpd_sign(a)^mpd_sign(b); + + if (mpd_isspecial(a) || mpd_isspecial(b)) { + if (mpd_qcheck_nans(q, a, b, ctx, status)) { + mpd_qcopy(r, q, status); + return; + } + if (mpd_isinfinite(a)) { + if (mpd_isinfinite(b)) { + mpd_setspecial(q, MPD_POS, MPD_NAN); + } + else { + mpd_setspecial(q, sign, MPD_INF); + } + mpd_setspecial(r, MPD_POS, MPD_NAN); + *status |= MPD_Invalid_operation; + return; + } + if (mpd_isinfinite(b)) { + if (!mpd_qcopy(r, a, status)) { + mpd_seterror(q, MPD_Malloc_error, status); + return; + } + mpd_qfinalize(r, ctx, status); + _settriple(q, sign, 0, 0); + return; + } + /* debug */ + abort(); /* GCOV_NOT_REACHED */ + } + if (mpd_iszerocoeff(b)) { + if (mpd_iszerocoeff(a)) { + mpd_setspecial(q, MPD_POS, MPD_NAN); + mpd_setspecial(r, MPD_POS, MPD_NAN); + *status |= MPD_Division_undefined; + } + else { + mpd_setspecial(q, sign, MPD_INF); + mpd_setspecial(r, MPD_POS, MPD_NAN); + *status |= (MPD_Division_by_zero|MPD_Invalid_operation); + } + return; + } + + _mpd_qdivmod(q, r, a, b, ctx, status); + mpd_qfinalize(q, ctx, status); + mpd_qfinalize(r, ctx, status); +} + +void +mpd_qdivint(mpd_t *q, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + MPD_NEW_STATIC(r,0,0,0,0); + uint8_t sign = mpd_sign(a)^mpd_sign(b); + + if (mpd_isspecial(a) || mpd_isspecial(b)) { + if (mpd_qcheck_nans(q, a, b, ctx, status)) { + return; + } + if (mpd_isinfinite(a) && mpd_isinfinite(b)) { + mpd_seterror(q, MPD_Invalid_operation, status); + return; + } + if (mpd_isinfinite(a)) { + mpd_setspecial(q, sign, MPD_INF); + return; + } + if (mpd_isinfinite(b)) { + _settriple(q, sign, 0, 0); + return; + } + /* debug */ + abort(); /* GCOV_NOT_REACHED */ + } + if (mpd_iszerocoeff(b)) { + if (mpd_iszerocoeff(a)) { + mpd_seterror(q, MPD_Division_undefined, status); + } + else { + mpd_setspecial(q, sign, MPD_INF); + *status |= MPD_Division_by_zero; + } + return; + } + + + _mpd_qdivmod(q, &r, a, b, ctx, status); + mpd_del(&r); + mpd_qfinalize(q, ctx, status); +} + +/* Divide decimal by mpd_ssize_t. */ +void +mpd_qdiv_ssize(mpd_t *result, const mpd_t *a, mpd_ssize_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_context_t maxcontext; + MPD_NEW_STATIC(bb,0,0,0,0); + + mpd_maxcontext(&maxcontext); + mpd_qsset_ssize(&bb, b, &maxcontext, status); + mpd_qdiv(result, a, &bb, ctx, status); + mpd_del(&bb); +} + +/* Divide decimal by mpd_uint_t. */ +void +mpd_qdiv_uint(mpd_t *result, const mpd_t *a, mpd_uint_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_context_t maxcontext; + MPD_NEW_STATIC(bb,0,0,0,0); + + mpd_maxcontext(&maxcontext); + mpd_qsset_uint(&bb, b, &maxcontext, status); + mpd_qdiv(result, a, &bb, ctx, status); + mpd_del(&bb); +} + +/* Divide decimal by int32_t. */ +void +mpd_qdiv_i32(mpd_t *result, const mpd_t *a, int32_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_qdiv_ssize(result, a, b, ctx, status); +} + +/* Divide decimal by uint32_t. */ +void +mpd_qdiv_u32(mpd_t *result, const mpd_t *a, uint32_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_qdiv_uint(result, a, b, ctx, status); +} + +#ifdef CONFIG_64 +/* Divide decimal by int64_t. */ +void +mpd_qdiv_i64(mpd_t *result, const mpd_t *a, int64_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_qdiv_ssize(result, a, b, ctx, status); +} + +/* Divide decimal by uint64_t. */ +void +mpd_qdiv_u64(mpd_t *result, const mpd_t *a, uint64_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_qdiv_uint(result, a, b, ctx, status); +} +#endif + +#if defined(_MSC_VER) + /* conversion from 'double' to 'mpd_ssize_t', possible loss of data */ + #pragma warning(disable:4244) +#endif +/* + * Get the number of iterations for the Horner scheme in _mpd_qexp(). + */ +static inline mpd_ssize_t +_mpd_get_exp_iterations(const mpd_t *a, mpd_ssize_t prec) +{ + mpd_uint_t dummy; + mpd_uint_t msdigits; + double f; + + /* 9 is MPD_RDIGITS for 32 bit platforms */ + _mpd_get_msdigits(&dummy, &msdigits, a, 9); + f = ((double)msdigits + 1) / mpd_pow10[mpd_word_digits(msdigits)]; + +#ifdef CONFIG_64 + #ifdef USE_80BIT_LONG_DOUBLE + return ceill((1.435*(long double)prec - 1.182) + / log10l((long double)prec/f)); + #else + /* prec > floor((1ULL<<53) / 1.435) */ + if (prec > 6276793905742851LL) { + return MPD_SSIZE_MAX; + } + return ceil((1.435*(double)prec - 1.182) / log10((double)prec/f)); + #endif +#else /* CONFIG_32 */ + return ceil((1.435*(double)prec - 1.182) / log10((double)prec/f)); + #if defined(_MSC_VER) + #pragma warning(default:4244) + #endif +#endif +} + +/* + * Internal function, specials have been dealt with. + * + * The algorithm is from Hull&Abrham, Variable Precision Exponential Function, + * ACM Transactions on Mathematical Software, Vol. 12, No. 2, June 1986. + * + * Main differences: + * + * - The number of iterations for the Horner scheme is calculated using the + * C log10() function. + * + * - The analysis for early abortion has been adapted for the mpd_t + * ranges. + */ +static void +_mpd_qexp(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_context_t workctx; + MPD_NEW_STATIC(tmp,0,0,0,0); + MPD_NEW_STATIC(sum,0,0,0,0); + MPD_NEW_CONST(word,0,0,0,1,1,1); + mpd_ssize_t j, n, t; + + assert(!mpd_isspecial(a)); + + /* + * We are calculating e^x = e^(r*10^t) = (e^r)^(10^t), where r < 1 and t >= 0. + * + * If t > 0, we have: + * + * (1) 0.1 <= r < 1, so e^r >= e^0.1. Overflow in the final power operation + * will occur when (e^0.1)^(10^t) > 10^(emax+1). If we consider MAX_EMAX, + * this will happen for t > 10 (32 bit) or (t > 19) (64 bit). + * + * (2) -1 < r <= -0.1, so e^r > e^-1. Underflow in the final power operation + * will occur when (e^-1)^(10^t) < 10^(etiny-1). If we consider MIN_ETINY, + * this will also happen for t > 10 (32 bit) or (t > 19) (64 bit). + */ +#if defined(CONFIG_64) + #define MPD_EXP_MAX_T 19 +#elif defined(CONFIG_32) + #define MPD_EXP_MAX_T 10 +#endif + t = a->digits + a->exp; + t = (t > 0) ? t : 0; + if (t > MPD_EXP_MAX_T) { + if (mpd_ispositive(a)) { + mpd_setspecial(result, MPD_POS, MPD_INF); + *status |= MPD_Overflow|MPD_Inexact|MPD_Rounded; + } + else { + _settriple(result, MPD_POS, 0, mpd_etiny(ctx)); + *status |= (MPD_Inexact|MPD_Rounded|MPD_Subnormal| + MPD_Underflow|MPD_Clamped); + } + return; + } + + mpd_maxcontext(&workctx); + workctx.prec = ctx->prec + t + 2; + workctx.prec = (workctx.prec < 9) ? 9 : workctx.prec; + workctx.round = MPD_ROUND_HALF_EVEN; + + if ((n = _mpd_get_exp_iterations(a, workctx.prec)) == MPD_SSIZE_MAX) { + mpd_seterror(result, MPD_Invalid_operation, status); /* GCOV_UNLIKELY */ + goto finish; /* GCOV_UNLIKELY */ + } + + if (!mpd_qcopy(result, a, status)) { + goto finish; + } + result->exp -= t; + + _settriple(&sum, MPD_POS, 1, 0); + + for (j = n-1; j >= 1; j--) { + word.data[0] = j; + mpd_setdigits(&word); + mpd_qdiv(&tmp, result, &word, &workctx, &workctx.status); + mpd_qmul(&sum, &sum, &tmp, &workctx, &workctx.status); + mpd_qadd(&sum, &sum, &one, &workctx, &workctx.status); + } + +#ifdef CONFIG_64 + _mpd_qpow_uint(result, &sum, mpd_pow10[t], MPD_POS, &workctx, status); +#else + if (t <= MPD_MAX_POW10) { + _mpd_qpow_uint(result, &sum, mpd_pow10[t], MPD_POS, &workctx, status); + } + else { + t -= MPD_MAX_POW10; + _mpd_qpow_uint(&tmp, &sum, mpd_pow10[MPD_MAX_POW10], MPD_POS, + &workctx, status); + _mpd_qpow_uint(result, &tmp, mpd_pow10[t], MPD_POS, &workctx, status); + } +#endif + + +finish: + mpd_del(&tmp); + mpd_del(&sum); + *status |= (workctx.status&MPD_Errors); + *status |= (MPD_Inexact|MPD_Rounded); +} + +/* exp(a) */ +void +mpd_qexp(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_context_t workctx; + + if (mpd_isspecial(a)) { + if (mpd_qcheck_nan(result, a, ctx, status)) { + return; + } + if (mpd_isnegative(a)) { + _settriple(result, MPD_POS, 0, 0); + } + else { + mpd_setspecial(result, MPD_POS, MPD_INF); + } + return; + } + if (mpd_iszerocoeff(a)) { + _settriple(result, MPD_POS, 1, 0); + return; + } + + workctx = *ctx; + workctx.round = MPD_ROUND_HALF_EVEN; + + if (ctx->allcr) { + MPD_NEW_STATIC(t1, 0,0,0,0); + MPD_NEW_STATIC(t2, 0,0,0,0); + MPD_NEW_STATIC(ulp, 0,0,0,0); + MPD_NEW_STATIC(aa, 0,0,0,0); + mpd_ssize_t prec; + + if (result == a) { + if (!mpd_qcopy(&aa, a, status)) { + mpd_seterror(result, MPD_Malloc_error, status); + return; + } + a = &aa; + } + + workctx.clamp = 0; + prec = ctx->prec + 3; + while (1) { + workctx.prec = prec; + _mpd_qexp(result, a, &workctx, status); + _ssettriple(&ulp, MPD_POS, 1, + result->exp + result->digits-workctx.prec-1); + + workctx.prec = ctx->prec; + mpd_qadd(&t1, result, &ulp, &workctx, &workctx.status); + mpd_qsub(&t2, result, &ulp, &workctx, &workctx.status); + if (mpd_isspecial(result) || mpd_iszerocoeff(result) || + mpd_qcmp(&t1, &t2, status) == 0) { + workctx.clamp = ctx->clamp; + mpd_check_underflow(result, &workctx, status); + mpd_qfinalize(result, &workctx, status); + break; + } + prec += MPD_RDIGITS; + } + mpd_del(&t1); + mpd_del(&t2); + mpd_del(&ulp); + mpd_del(&aa); + } + else { + _mpd_qexp(result, a, &workctx, status); + mpd_check_underflow(result, &workctx, status); + mpd_qfinalize(result, &workctx, status); + } +} + +/* Fused multiply-add: (a * b) + c, with a single final rounding. */ +void +mpd_qfma(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_t *c, + const mpd_context_t *ctx, uint32_t *status) +{ + uint32_t workstatus = 0; + mpd_t *cc = (mpd_t *)c; + + if (result == c) { + if ((cc = mpd_qncopy(c)) == NULL) { + mpd_seterror(result, MPD_Malloc_error, status); + return; + } + } + + _mpd_qmul(result, a, b, ctx, &workstatus); + if (!(workstatus&MPD_Invalid_operation)) { + mpd_qadd(result, result, cc, ctx, &workstatus); + } + + if (cc != c) mpd_del(cc); + *status |= workstatus; +} + +static inline int +ln_schedule_prec(mpd_ssize_t klist[MPD_MAX_PREC_LOG2], mpd_ssize_t maxprec, + mpd_ssize_t initprec) +{ + mpd_ssize_t k; + int i; + + assert(maxprec >= 2 && initprec >= 2); + if (maxprec <= initprec) return -1; + + i = 0; k = maxprec; + do { + k = (k+2) / 2; + klist[i++] = k; + } while (k > initprec); + + return i-1; +} + +#ifdef CONFIG_64 +#if MPD_RDIGITS != 19 + #error "mpdecimal.c: MPD_RDIGITS must be 19." +#endif +static const mpd_uint_t mpd_ln10_data[MPD_MINALLOC_MAX] = { + 6983716328982174407ULL, 9089704281976336583ULL, 1515961135648465461ULL, + 4416816335727555703ULL, 2900988039194170265ULL, 2307925037472986509ULL, + 107598438319191292ULL, 3466624107184669231ULL, 4450099781311469159ULL, + 9807828059751193854ULL, 7713456862091670584ULL, 1492198849978748873ULL, + 6528728696511086257ULL, 2385392051446341972ULL, 8692180205189339507ULL, + 6518769751037497088ULL, 2375253577097505395ULL, 9095610299291824318ULL, + 982748238504564801ULL, 5438635917781170543ULL, 7547331541421808427ULL, + 752371033310119785ULL, 3171643095059950878ULL, 9785265383207606726ULL, + 2932258279850258550ULL, 5497347726624257094ULL, 2976979522110718264ULL, + 9221477656763693866ULL, 1979650047149510504ULL, 6674183485704422507ULL, + 9702766860595249671ULL, 9278096762712757753ULL, 9314848524948644871ULL, + 6826928280848118428ULL, 754403708474699401ULL, 230105703089634572ULL, + 1929203337658714166ULL, 7589402567763113569ULL, 4208241314695689016ULL, + 2922455440575892572ULL, 9356734206705811364ULL, 2684916746550586856ULL, + 644507064800027750ULL, 9476834636167921018ULL, 5659121373450747856ULL, + 2835522011480466371ULL, 6470806855677432162ULL, 7141748003688084012ULL, + 9619404400222105101ULL, 5504893431493939147ULL, 6674744042432743651ULL, + 2287698219886746543ULL, 7773262884616336622ULL, 1985283935053089653ULL, + 4680843799894826233ULL, 8168948290720832555ULL, 8067566662873690987ULL, + 6248633409525465082ULL, 9829834196778404228ULL, 3524802359972050895ULL, + 3327900967572609677ULL, 110148862877297603ULL, 179914546843642076ULL, + 2302585092994045684ULL +}; +#else +#if MPD_RDIGITS != 9 + #error "mpdecimal.c: MPD_RDIGITS must be 9." +#endif +static const mpd_uint_t mpd_ln10_data[MPD_MINALLOC_MAX] = { + 401682692UL, 708474699UL, 720754403UL, 30896345UL, 602301057UL, 765871416UL, + 192920333UL, 763113569UL, 589402567UL, 956890167UL, 82413146UL, 589257242UL, + 245544057UL, 811364292UL, 734206705UL, 868569356UL, 167465505UL, 775026849UL, + 706480002UL, 18064450UL, 636167921UL, 569476834UL, 734507478UL, 156591213UL, + 148046637UL, 283552201UL, 677432162UL, 470806855UL, 880840126UL, 417480036UL, + 210510171UL, 940440022UL, 939147961UL, 893431493UL, 436515504UL, 440424327UL, + 654366747UL, 821988674UL, 622228769UL, 884616336UL, 537773262UL, 350530896UL, + 319852839UL, 989482623UL, 468084379UL, 720832555UL, 168948290UL, 736909878UL, + 675666628UL, 546508280UL, 863340952UL, 404228624UL, 834196778UL, 508959829UL, + 23599720UL, 967735248UL, 96757260UL, 603332790UL, 862877297UL, 760110148UL, + 468436420UL, 401799145UL, 299404568UL, 230258509UL +}; +#endif +/* _mpd_ln10 is used directly for precisions smaller than MINALLOC_MAX*RDIGITS. + Otherwise, it serves as the initial approximation for calculating ln(10). */ +static const mpd_t _mpd_ln10 = { + MPD_STATIC|MPD_CONST_DATA, -(MPD_MINALLOC_MAX*MPD_RDIGITS-1), + MPD_MINALLOC_MAX*MPD_RDIGITS, MPD_MINALLOC_MAX, MPD_MINALLOC_MAX, + (mpd_uint_t *)mpd_ln10_data +}; + +/* Set 'result' to ln(10), with 'prec' digits, using ROUND_HALF_EVEN. */ +void +mpd_qln10(mpd_t *result, mpd_ssize_t prec, uint32_t *status) +{ + mpd_context_t varcontext, maxcontext; + MPD_NEW_STATIC(tmp, 0,0,0,0); + MPD_NEW_CONST(static10, 0,0,2,1,1,10); + mpd_ssize_t klist[MPD_MAX_PREC_LOG2]; + mpd_uint_t rnd; + mpd_ssize_t shift; + int i; + + assert(prec >= 1); + + shift = MPD_MINALLOC_MAX*MPD_RDIGITS-prec; + shift = shift < 0 ? 0 : shift; + + rnd = mpd_qshiftr(result, &_mpd_ln10, shift, status); + if (rnd == MPD_UINT_MAX) { + mpd_seterror(result, MPD_Malloc_error, status); + return; + } + result->exp = -(result->digits-1); + + mpd_maxcontext(&maxcontext); + if (prec < MPD_MINALLOC_MAX*MPD_RDIGITS) { + maxcontext.prec = prec; + _mpd_apply_round_excess(result, rnd, &maxcontext, status); + *status |= (MPD_Inexact|MPD_Rounded); + return; + } + + mpd_maxcontext(&varcontext); + varcontext.round = MPD_ROUND_TRUNC; + + i = ln_schedule_prec(klist, prec+2, result->digits); + for (; i >= 0; i--) { + varcontext.prec = 2*klist[i]+3; + result->flags ^= MPD_NEG; + _mpd_qexp(&tmp, result, &varcontext, status); + result->flags ^= MPD_NEG; + mpd_qmul(&tmp, &static10, &tmp, &varcontext, status); + mpd_qsub(&tmp, &tmp, &one, &maxcontext, status); + mpd_qadd(result, result, &tmp, &maxcontext, status); + if (mpd_isspecial(result)) { + break; + } + } + + mpd_del(&tmp); + maxcontext.prec = prec; + mpd_qfinalize(result, &maxcontext, status); +} + +/* Initial approximations for the ln() iteration */ +static const uint16_t lnapprox[900] = { + /* index 0 - 400: log((i+100)/100) * 1000 */ + 0, 10, 20, 30, 39, 49, 58, 68, 77, 86, 95, 104, 113, 122, 131, 140, 148, 157, + 166, 174, 182, 191, 199, 207, 215, 223, 231, 239, 247, 255, 262, 270, 278, + 285, 293, 300, 308, 315, 322, 329, 336, 344, 351, 358, 365, 372, 378, 385, + 392, 399, 406, 412, 419, 425, 432, 438, 445, 451, 457, 464, 470, 476, 482, + 489, 495, 501, 507, 513, 519, 525, 531, 536, 542, 548, 554, 560, 565, 571, + 577, 582, 588, 593, 599, 604, 610, 615, 621, 626, 631, 637, 642, 647, 652, + 658, 663, 668, 673, 678, 683, 688, 693, 698, 703, 708, 713, 718, 723, 728, + 732, 737, 742, 747, 751, 756, 761, 766, 770, 775, 779, 784, 788, 793, 798, + 802, 806, 811, 815, 820, 824, 829, 833, 837, 842, 846, 850, 854, 859, 863, + 867, 871, 876, 880, 884, 888, 892, 896, 900, 904, 908, 912, 916, 920, 924, + 928, 932, 936, 940, 944, 948, 952, 956, 959, 963, 967, 971, 975, 978, 982, + 986, 990, 993, 997, 1001, 1004, 1008, 1012, 1015, 1019, 1022, 1026, 1030, + 1033, 1037, 1040, 1044, 1047, 1051, 1054, 1058, 1061, 1065, 1068, 1072, 1075, + 1078, 1082, 1085, 1089, 1092, 1095, 1099, 1102, 1105, 1109, 1112, 1115, 1118, + 1122, 1125, 1128, 1131, 1135, 1138, 1141, 1144, 1147, 1151, 1154, 1157, 1160, + 1163, 1166, 1169, 1172, 1176, 1179, 1182, 1185, 1188, 1191, 1194, 1197, 1200, + 1203, 1206, 1209, 1212, 1215, 1218, 1221, 1224, 1227, 1230, 1233, 1235, 1238, + 1241, 1244, 1247, 1250, 1253, 1256, 1258, 1261, 1264, 1267, 1270, 1273, 1275, + 1278, 1281, 1284, 1286, 1289, 1292, 1295, 1297, 1300, 1303, 1306, 1308, 1311, + 1314, 1316, 1319, 1322, 1324, 1327, 1330, 1332, 1335, 1338, 1340, 1343, 1345, + 1348, 1351, 1353, 1356, 1358, 1361, 1364, 1366, 1369, 1371, 1374, 1376, 1379, + 1381, 1384, 1386, 1389, 1391, 1394, 1396, 1399, 1401, 1404, 1406, 1409, 1411, + 1413, 1416, 1418, 1421, 1423, 1426, 1428, 1430, 1433, 1435, 1437, 1440, 1442, + 1445, 1447, 1449, 1452, 1454, 1456, 1459, 1461, 1463, 1466, 1468, 1470, 1472, + 1475, 1477, 1479, 1482, 1484, 1486, 1488, 1491, 1493, 1495, 1497, 1500, 1502, + 1504, 1506, 1509, 1511, 1513, 1515, 1517, 1520, 1522, 1524, 1526, 1528, 1530, + 1533, 1535, 1537, 1539, 1541, 1543, 1545, 1548, 1550, 1552, 1554, 1556, 1558, + 1560, 1562, 1564, 1567, 1569, 1571, 1573, 1575, 1577, 1579, 1581, 1583, 1585, + 1587, 1589, 1591, 1593, 1595, 1597, 1599, 1601, 1603, 1605, 1607, 1609, + /* index 401 - 899: -log((i+100)/1000) * 1000 */ + 691, 689, 687, 685, 683, 681, 679, 677, 675, 673, 671, 669, 668, 666, 664, + 662, 660, 658, 656, 654, 652, 650, 648, 646, 644, 642, 641, 639, 637, 635, + 633, 631, 629, 627, 626, 624, 622, 620, 618, 616, 614, 612, 611, 609, 607, + 605, 603, 602, 600, 598, 596, 594, 592, 591, 589, 587, 585, 583, 582, 580, + 578, 576, 574, 573, 571, 569, 567, 566, 564, 562, 560, 559, 557, 555, 553, + 552, 550, 548, 546, 545, 543, 541, 540, 538, 536, 534, 533, 531, 529, 528, + 526, 524, 523, 521, 519, 518, 516, 514, 512, 511, 509, 508, 506, 504, 502, + 501, 499, 498, 496, 494, 493, 491, 489, 488, 486, 484, 483, 481, 480, 478, + 476, 475, 473, 472, 470, 468, 467, 465, 464, 462, 460, 459, 457, 456, 454, + 453, 451, 449, 448, 446, 445, 443, 442, 440, 438, 437, 435, 434, 432, 431, + 429, 428, 426, 425, 423, 422, 420, 419, 417, 416, 414, 412, 411, 410, 408, + 406, 405, 404, 402, 400, 399, 398, 396, 394, 393, 392, 390, 389, 387, 386, + 384, 383, 381, 380, 378, 377, 375, 374, 372, 371, 370, 368, 367, 365, 364, + 362, 361, 360, 358, 357, 355, 354, 352, 351, 350, 348, 347, 345, 344, 342, + 341, 340, 338, 337, 336, 334, 333, 331, 330, 328, 327, 326, 324, 323, 322, + 320, 319, 318, 316, 315, 313, 312, 311, 309, 308, 306, 305, 304, 302, 301, + 300, 298, 297, 296, 294, 293, 292, 290, 289, 288, 286, 285, 284, 282, 281, + 280, 278, 277, 276, 274, 273, 272, 270, 269, 268, 267, 265, 264, 263, 261, + 260, 259, 258, 256, 255, 254, 252, 251, 250, 248, 247, 246, 245, 243, 242, + 241, 240, 238, 237, 236, 234, 233, 232, 231, 229, 228, 227, 226, 224, 223, + 222, 221, 219, 218, 217, 216, 214, 213, 212, 211, 210, 208, 207, 206, 205, + 203, 202, 201, 200, 198, 197, 196, 195, 194, 192, 191, 190, 189, 188, 186, + 185, 184, 183, 182, 180, 179, 178, 177, 176, 174, 173, 172, 171, 170, 168, + 167, 166, 165, 164, 162, 161, 160, 159, 158, 157, 156, 154, 153, 152, 151, + 150, 148, 147, 146, 145, 144, 143, 142, 140, 139, 138, 137, 136, 135, 134, + 132, 131, 130, 129, 128, 127, 126, 124, 123, 122, 121, 120, 119, 118, 116, + 115, 114, 113, 112, 111, 110, 109, 108, 106, 105, 104, 103, 102, 101, 100, + 99, 98, 97, 95, 94, 93, 92, 91, 90, 89, 88, 87, 86, 84, 83, 82, 81, 80, 79, + 78, 77, 76, 75, 74, 73, 72, 70, 69, 68, 67, 66, 65, 64, 63, 62, 61, 60, 59, + 58, 57, 56, 54, 53, 52, 51, 50, 49, 48, 47, 46, 45, 44, 43, 42, 41, 40, 39, + 38, 37, 36, 35, 34, 33, 31, 30, 29, 28, 27, 26, 25, 24, 23, 22, 21, 20, 19, + 18, 17, 16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1 +}; + +/* Internal ln() function that does not check for specials, zero or one. */ +static void +_mpd_qln(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_context_t varcontext, maxcontext; + mpd_t *z = (mpd_t *) result; + MPD_NEW_STATIC(v,0,0,0,0); + MPD_NEW_STATIC(vtmp,0,0,0,0); + MPD_NEW_STATIC(tmp,0,0,0,0); + mpd_ssize_t klist[MPD_MAX_PREC_LOG2]; + mpd_ssize_t maxprec, shift, t; + mpd_ssize_t a_digits, a_exp; + mpd_uint_t dummy, x; + int i; + + assert(!mpd_isspecial(a) && !mpd_iszerocoeff(a)); + + /* + * We are calculating ln(a) = ln(v * 10^t) = ln(v) + t*ln(10), + * where 0.5 < v <= 5. + */ + if (!mpd_qcopy(&v, a, status)) { + mpd_seterror(result, MPD_Malloc_error, status); + goto finish; + } + + /* Initial approximation: we have at least one non-zero digit */ + _mpd_get_msdigits(&dummy, &x, &v, 3); + if (x < 10) x *= 10; + if (x < 100) x *= 10; + x -= 100; + + /* a may equal z */ + a_digits = a->digits; + a_exp = a->exp; + + mpd_minalloc(z); + mpd_clear_flags(z); + z->data[0] = lnapprox[x]; + z->len = 1; + z->exp = -3; + mpd_setdigits(z); + + if (x <= 400) { + v.exp = -(a_digits - 1); + t = a_exp + a_digits - 1; + } + else { + v.exp = -a_digits; + t = a_exp + a_digits; + mpd_set_negative(z); + } + + mpd_maxcontext(&maxcontext); + mpd_maxcontext(&varcontext); + varcontext.round = MPD_ROUND_TRUNC; + + maxprec = ctx->prec + 2; + if (x <= 10 || x >= 805) { + /* v is close to 1: Estimate the magnitude of the logarithm. + * If v = 1 or ln(v) will underflow, skip the loop. Otherwise, + * adjust the precision upwards in order to obtain a sufficient + * number of significant digits. + * + * 1) x/(1+x) < ln(1+x) < x, for x > -1, x != 0 + * + * 2) (v-1)/v < ln(v) < v-1 + */ + mpd_t *lower = &tmp; + mpd_t *upper = &vtmp; + int cmp = _mpd_cmp(&v, &one); + + varcontext.round = MPD_ROUND_CEILING; + varcontext.prec = maxprec; + mpd_qsub(upper, &v, &one, &varcontext, &varcontext.status); + varcontext.round = MPD_ROUND_FLOOR; + mpd_qdiv(lower, upper, &v, &varcontext, &varcontext.status); + varcontext.round = MPD_ROUND_TRUNC; + + if (cmp < 0) { + _mpd_ptrswap(&upper, &lower); + } + if (mpd_adjexp(upper) < mpd_etiny(ctx)) { + _settriple(z, (cmp<0), 1, mpd_etiny(ctx)-1); + goto postloop; + } + /* XXX optimization: t == 0 && mpd_adjexp(lower) < 0 */ + if (mpd_adjexp(lower) < 0) { + maxprec = maxprec - mpd_adjexp(lower); + } + } + + i = ln_schedule_prec(klist, maxprec, 2); + for (; i >= 0; i--) { + varcontext.prec = 2*klist[i]+3; + z->flags ^= MPD_NEG; + _mpd_qexp(&tmp, z, &varcontext, status); + z->flags ^= MPD_NEG; + + if (v.digits > varcontext.prec) { + shift = v.digits - varcontext.prec; + mpd_qshiftr(&vtmp, &v, shift, status); + vtmp.exp += shift; + mpd_qmul(&tmp, &vtmp, &tmp, &varcontext, status); + } + else { + mpd_qmul(&tmp, &v, &tmp, &varcontext, status); + } + + mpd_qsub(&tmp, &tmp, &one, &maxcontext, status); + mpd_qadd(z, z, &tmp, &maxcontext, status); + if (mpd_isspecial(z)) { + break; + } + } + +postloop: + mpd_qln10(&v, maxprec+2, status); + mpd_qmul_ssize(&tmp, &v, t, &maxcontext, status); + varcontext.prec = maxprec+2; + mpd_qadd(result, &tmp, z, &varcontext, status); + + +finish: + mpd_del(&v); + mpd_del(&vtmp); + mpd_del(&tmp); +} + +/* ln(a) */ +void +mpd_qln(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_context_t workctx; + mpd_ssize_t adjexp, t; + + if (mpd_isspecial(a)) { + if (mpd_qcheck_nan(result, a, ctx, status)) { + return; + } + if (mpd_isnegative(a)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + mpd_setspecial(result, MPD_POS, MPD_INF); + return; + } + if (mpd_iszerocoeff(a)) { + mpd_setspecial(result, MPD_NEG, MPD_INF); + return; + } + if (mpd_isnegative(a)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if (_mpd_cmp(a, &one) == 0) { + _settriple(result, MPD_POS, 0, 0); + return; + } + /* Check if the result will overflow. + * + * 1) adjexp(a) + 1 > log10(a) >= adjexp(a) + * + * 2) |log10(a)| >= adjexp(a), if adjexp(a) >= 0 + * |log10(a)| > -adjexp(a)-1, if adjexp(a) < 0 + * + * 3) |log(a)| > 2*|log10(a)| + */ + adjexp = mpd_adjexp(a); + t = (adjexp < 0) ? -adjexp-1 : adjexp; + t *= 2; + if (mpd_exp_digits(t)-1 > ctx->emax) { + *status |= MPD_Overflow|MPD_Inexact|MPD_Rounded; + mpd_setspecial(result, (adjexp<0), MPD_INF); + return; + } + + workctx = *ctx; + workctx.round = MPD_ROUND_HALF_EVEN; + + if (ctx->allcr) { + MPD_NEW_STATIC(t1, 0,0,0,0); + MPD_NEW_STATIC(t2, 0,0,0,0); + MPD_NEW_STATIC(ulp, 0,0,0,0); + MPD_NEW_STATIC(aa, 0,0,0,0); + mpd_ssize_t prec; + + if (result == a) { + if (!mpd_qcopy(&aa, a, status)) { + mpd_seterror(result, MPD_Malloc_error, status); + return; + } + a = &aa; + } + + workctx.clamp = 0; + prec = ctx->prec + 3; + while (1) { + workctx.prec = prec; + _mpd_qln(result, a, &workctx, status); + _ssettriple(&ulp, MPD_POS, 1, + result->exp + result->digits-workctx.prec-1); + + workctx.prec = ctx->prec; + mpd_qadd(&t1, result, &ulp, &workctx, &workctx.status); + mpd_qsub(&t2, result, &ulp, &workctx, &workctx.status); + if (mpd_isspecial(result) || mpd_iszerocoeff(result) || + mpd_qcmp(&t1, &t2, status) == 0) { + workctx.clamp = ctx->clamp; + mpd_check_underflow(result, &workctx, status); + mpd_qfinalize(result, &workctx, status); + break; + } + prec += MPD_RDIGITS; + } + mpd_del(&t1); + mpd_del(&t2); + mpd_del(&ulp); + mpd_del(&aa); + } + else { + _mpd_qln(result, a, &workctx, status); + mpd_check_underflow(result, &workctx, status); + mpd_qfinalize(result, &workctx, status); + } +} + +/* Internal log10() function that does not check for specials, zero, ... */ +static void +_mpd_qlog10(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_context_t workctx; + MPD_NEW_STATIC(ln10,0,0,0,0); + + mpd_maxcontext(&workctx); + workctx.prec = ctx->prec + 3; + _mpd_qln(result, a, &workctx, status); + mpd_qln10(&ln10, workctx.prec, status); + + workctx = *ctx; + workctx.round = MPD_ROUND_HALF_EVEN; + _mpd_qdiv(NO_IDEAL_EXP, result, result, &ln10, &workctx, status); + + mpd_del(&ln10); +} + +/* log10(a) */ +void +mpd_qlog10(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_context_t workctx; + mpd_ssize_t adjexp, t; + + workctx = *ctx; + workctx.round = MPD_ROUND_HALF_EVEN; + + if (mpd_isspecial(a)) { + if (mpd_qcheck_nan(result, a, ctx, status)) { + return; + } + if (mpd_isnegative(a)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + mpd_setspecial(result, MPD_POS, MPD_INF); + return; + } + if (mpd_iszerocoeff(a)) { + mpd_setspecial(result, MPD_NEG, MPD_INF); + return; + } + if (mpd_isnegative(a)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if (mpd_coeff_ispow10(a)) { + uint8_t sign = 0; + adjexp = mpd_adjexp(a); + if (adjexp < 0) { + sign = 1; + adjexp = -adjexp; + } + _settriple(result, sign, adjexp, 0); + mpd_qfinalize(result, &workctx, status); + return; + } + /* Check if the result will overflow. + * + * 1) adjexp(a) + 1 > log10(a) >= adjexp(a) + * + * 2) |log10(a)| >= adjexp(a), if adjexp(a) >= 0 + * |log10(a)| > -adjexp(a)-1, if adjexp(a) < 0 + */ + adjexp = mpd_adjexp(a); + t = (adjexp < 0) ? -adjexp-1 : adjexp; + if (mpd_exp_digits(t)-1 > ctx->emax) { + *status |= MPD_Overflow|MPD_Inexact|MPD_Rounded; + mpd_setspecial(result, (adjexp<0), MPD_INF); + return; + } + + if (ctx->allcr) { + MPD_NEW_STATIC(t1, 0,0,0,0); + MPD_NEW_STATIC(t2, 0,0,0,0); + MPD_NEW_STATIC(ulp, 0,0,0,0); + MPD_NEW_STATIC(aa, 0,0,0,0); + mpd_ssize_t prec; + + if (result == a) { + if (!mpd_qcopy(&aa, a, status)) { + mpd_seterror(result, MPD_Malloc_error, status); + return; + } + a = &aa; + } + + workctx.clamp = 0; + prec = ctx->prec + 3; + while (1) { + workctx.prec = prec; + _mpd_qlog10(result, a, &workctx, status); + _ssettriple(&ulp, MPD_POS, 1, + result->exp + result->digits-workctx.prec-1); + + workctx.prec = ctx->prec; + mpd_qadd(&t1, result, &ulp, &workctx, &workctx.status); + mpd_qsub(&t2, result, &ulp, &workctx, &workctx.status); + if (mpd_isspecial(result) || mpd_iszerocoeff(result) || + mpd_qcmp(&t1, &t2, status) == 0) { + workctx.clamp = ctx->clamp; + mpd_check_underflow(result, &workctx, status); + mpd_qfinalize(result, &workctx, status); + break; + } + prec += MPD_RDIGITS; + } + mpd_del(&t1); + mpd_del(&t2); + mpd_del(&ulp); + mpd_del(&aa); + } + else { + _mpd_qlog10(result, a, &workctx, status); + mpd_check_underflow(result, &workctx, status); + } +} + +/* + * Maximum of the two operands. Attention: If one operand is a quiet NaN and the + * other is numeric, the numeric operand is returned. This may not be what one + * expects. + */ +void +mpd_qmax(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + int c; + + if (mpd_isqnan(a) && !mpd_isnan(b)) { + mpd_qcopy(result, b, status); + } + else if (mpd_isqnan(b) && !mpd_isnan(a)) { + mpd_qcopy(result, a, status); + } + else if (mpd_qcheck_nans(result, a, b, ctx, status)) { + return; + } + else { + c = _mpd_cmp(a, b); + if (c == 0) { + c = _mpd_cmp_numequal(a, b); + } + + if (c < 0) { + mpd_qcopy(result, b, status); + } + else { + mpd_qcopy(result, a, status); + } + } + + mpd_qfinalize(result, ctx, status); +} + +/* + * Maximum magnitude: Same as mpd_max(), but compares the operands with their + * sign ignored. + */ +void +mpd_qmax_mag(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + int c; + + if (mpd_isqnan(a) && !mpd_isnan(b)) { + mpd_qcopy(result, b, status); + } + else if (mpd_isqnan(b) && !mpd_isnan(a)) { + mpd_qcopy(result, a, status); + } + else if (mpd_qcheck_nans(result, a, b, ctx, status)) { + return; + } + else { + c = _mpd_cmp_abs(a, b); + if (c == 0) { + c = _mpd_cmp_numequal(a, b); + } + + if (c < 0) { + mpd_qcopy(result, b, status); + } + else { + mpd_qcopy(result, a, status); + } + } + + mpd_qfinalize(result, ctx, status); +} + +/* + * Minimum of the two operands. Attention: If one operand is a quiet NaN and the + * other is numeric, the numeric operand is returned. This may not be what one + * expects. + */ +void +mpd_qmin(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + int c; + + if (mpd_isqnan(a) && !mpd_isnan(b)) { + mpd_qcopy(result, b, status); + } + else if (mpd_isqnan(b) && !mpd_isnan(a)) { + mpd_qcopy(result, a, status); + } + else if (mpd_qcheck_nans(result, a, b, ctx, status)) { + return; + } + else { + c = _mpd_cmp(a, b); + if (c == 0) { + c = _mpd_cmp_numequal(a, b); + } + + if (c < 0) { + mpd_qcopy(result, a, status); + } + else { + mpd_qcopy(result, b, status); + } + } + + mpd_qfinalize(result, ctx, status); +} + +/* + * Minimum magnitude: Same as mpd_min(), but compares the operands with their + * sign ignored. + */ +void +mpd_qmin_mag(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + int c; + + if (mpd_isqnan(a) && !mpd_isnan(b)) { + mpd_qcopy(result, b, status); + } + else if (mpd_isqnan(b) && !mpd_isnan(a)) { + mpd_qcopy(result, a, status); + } + else if (mpd_qcheck_nans(result, a, b, ctx, status)) { + return; + } + else { + c = _mpd_cmp_abs(a, b); + if (c == 0) { + c = _mpd_cmp_numequal(a, b); + } + + if (c < 0) { + mpd_qcopy(result, a, status); + } + else { + mpd_qcopy(result, b, status); + } + } + + mpd_qfinalize(result, ctx, status); +} + +/* Minimum space needed for the result array in _karatsuba_rec(). */ +static inline mpd_size_t +_kmul_resultsize(mpd_size_t la, mpd_size_t lb) +{ + mpd_size_t n, m; + + n = add_size_t(la, lb); + n = add_size_t(n, 1); + + m = (la+1)/2 + 1; + m = mul_size_t(m, 3); + + return (m > n) ? m : n; +} + +/* Work space needed in _karatsuba_rec(). lim >= 4 */ +static inline mpd_size_t +_kmul_worksize(mpd_size_t n, mpd_size_t lim) +{ + mpd_size_t m; + + if (n <= lim) { + return 0; + } + + m = (n+1)/2 + 1; + + return add_size_t(mul_size_t(m, 2), _kmul_worksize(m, lim)); +} + + +#define MPD_KARATSUBA_BASECASE 16 /* must be >= 4 */ + +/* + * Add the product of a and b to c. + * c must be _kmul_resultsize(la, lb) in size. + * w is used as a work array and must be _kmul_worksize(a, lim) in size. + * Roman E. Maeder, Storage Allocation for the Karatsuba Integer Multiplication + * Algorithm. In "Design and implementation of symbolic computation systems", + * Springer, 1993, ISBN 354057235X, 9783540572350. + */ +static void +_karatsuba_rec(mpd_uint_t *c, const mpd_uint_t *a, const mpd_uint_t *b, + mpd_uint_t *w, mpd_size_t la, mpd_size_t lb) +{ + mpd_size_t m, lt; + + assert(la >= lb && lb > 0); + assert(la <= MPD_KARATSUBA_BASECASE || w != NULL); + + if (la <= MPD_KARATSUBA_BASECASE) { + _mpd_basemul(c, a, b, la, lb); + return; + } + + m = (la+1)/2; // ceil(la/2) + + /* lb <= m < la */ + if (lb <= m) { + + /* lb can now be larger than la-m */ + if (lb > la-m) { + lt = lb + lb + 1; // space needed for result array + mpd_uint_zero(w, lt); // clear result array + _karatsuba_rec(w, b, a+m, w+lt, lb, la-m); // b*ah + } + else { + lt = (la-m) + (la-m) + 1; // space needed for result array + mpd_uint_zero(w, lt); // clear result array + _karatsuba_rec(w, a+m, b, w+lt, la-m, lb); // ah*b + } + _mpd_baseaddto(c+m, w, (la-m)+lb); // add ah*b*B**m + + lt = m + m + 1; // space needed for the result array + mpd_uint_zero(w, lt); // clear result array + _karatsuba_rec(w, a, b, w+lt, m, lb); // al*b + _mpd_baseaddto(c, w, m+lb); // add al*b + + return; + } + + /* la >= lb > m */ + memcpy(w, a, m * sizeof *w); + w[m] = 0; + _mpd_baseaddto(w, a+m, la-m); + + memcpy(w+(m+1), b, m * sizeof *w); + w[m+1+m] = 0; + _mpd_baseaddto(w+(m+1), b+m, lb-m); + + _karatsuba_rec(c+m, w, w+(m+1), w+2*(m+1), m+1, m+1); + + lt = (la-m) + (la-m) + 1; + mpd_uint_zero(w, lt); + + _karatsuba_rec(w, a+m, b+m, w+lt, la-m, lb-m); + + _mpd_baseaddto(c+2*m, w, (la-m) + (lb-m)); + _mpd_basesubfrom(c+m, w, (la-m) + (lb-m)); + + lt = m + m + 1; + mpd_uint_zero(w, lt); + + _karatsuba_rec(w, a, b, w+lt, m, m); + _mpd_baseaddto(c, w, m+m); + _mpd_basesubfrom(c+m, w, m+m); + + return; +} + +/* + * Multiply u and v, using Karatsuba multiplication. Returns a pointer + * to the result or NULL in case of failure (malloc error). + * Conditions: ulen >= vlen, ulen >= 4 + */ +mpd_uint_t * +_mpd_kmul(const mpd_uint_t *u, const mpd_uint_t *v, + mpd_size_t ulen, mpd_size_t vlen, + mpd_size_t *rsize) +{ + mpd_uint_t *result = NULL, *w = NULL; + mpd_size_t m; + + assert(ulen >= 4); + assert(ulen >= vlen); + + *rsize = _kmul_resultsize(ulen, vlen); + if ((result = mpd_calloc(*rsize, sizeof *result)) == NULL) { + return NULL; + } + + m = _kmul_worksize(ulen, MPD_KARATSUBA_BASECASE); + if (m && ((w = mpd_calloc(m, sizeof *w)) == NULL)) { + mpd_free(result); + return NULL; + } + + _karatsuba_rec(result, u, v, w, ulen, vlen); + + + if (w) mpd_free(w); + return result; +} + + +/* Determine the minimum length for the number theoretic transform. */ +static inline mpd_size_t +_mpd_get_transform_len(mpd_size_t rsize) +{ + mpd_size_t log2rsize; + mpd_size_t x, step; + + assert(rsize >= 4); + log2rsize = mpd_bsr(rsize); + + if (rsize <= 1024) { + x = ((mpd_size_t)1)<<log2rsize; + return (rsize == x) ? x : x<<1; + } + else if (rsize <= MPD_MAXTRANSFORM_2N) { + x = ((mpd_size_t)1)<<log2rsize; + if (rsize == x) return x; + step = x>>1; + x += step; + return (rsize <= x) ? x : x + step; + } + else if (rsize <= MPD_MAXTRANSFORM_2N+MPD_MAXTRANSFORM_2N/2) { + return MPD_MAXTRANSFORM_2N+MPD_MAXTRANSFORM_2N/2; + } + else if (rsize <= 3*MPD_MAXTRANSFORM_2N) { + return 3*MPD_MAXTRANSFORM_2N; + } + else { + return MPD_SIZE_MAX; + } +} + +#ifdef PPRO +#ifndef _MSC_VER +static inline unsigned short +_mpd_get_control87(void) +{ + unsigned short cw; + + __asm__ __volatile__ ("fnstcw %0" : "=m" (cw)); + return cw; +} + +static inline void +_mpd_set_control87(unsigned short cw) +{ + __asm__ __volatile__ ("fldcw %0" : : "m" (cw)); +} +#endif + +unsigned int +mpd_set_fenv(void) +{ + unsigned int cw; +#ifdef _MSC_VER + unsigned int flags = + _EM_INVALID|_EM_DENORMAL|_EM_ZERODIVIDE|_EM_OVERFLOW| + _EM_UNDERFLOW|_EM_INEXACT|_RC_CHOP|_PC_64; + unsigned int mask = _MCW_EM|_MCW_RC|_MCW_PC; + unsigned int dummy; + + __control87_2(0, 0, &cw, NULL); + __control87_2(flags, mask, &dummy, NULL); +#else + cw = _mpd_get_control87(); + _mpd_set_control87(cw|0xF3F); +#endif + return cw; +} + +void +mpd_restore_fenv(unsigned int cw) +{ +#ifdef _MSC_VER + unsigned int mask = _MCW_EM|_MCW_RC|_MCW_PC; + unsigned int dummy; + + __control87_2(cw, mask, &dummy, NULL); +#else + _mpd_set_control87((unsigned short)cw); +#endif +} +#endif /* PPRO */ + +/* + * Multiply u and v, using the fast number theoretic transform. Returns + * a pointer to the result or NULL in case of failure (malloc error). + */ +mpd_uint_t * +_mpd_fntmul(const mpd_uint_t *u, const mpd_uint_t *v, + mpd_size_t ulen, mpd_size_t vlen, + mpd_size_t *rsize) +{ + mpd_uint_t *c1 = NULL, *c2 = NULL, *c3 = NULL, *vtmp = NULL; + mpd_size_t n; + +#ifdef PPRO + unsigned int cw; + cw = mpd_set_fenv(); +#endif + + *rsize = add_size_t(ulen, vlen); + if ((n = _mpd_get_transform_len(*rsize)) == MPD_SIZE_MAX) { + goto malloc_error; + } + + if ((c1 = mpd_calloc(sizeof *c1, n)) == NULL) { + goto malloc_error; + } + if ((c2 = mpd_calloc(sizeof *c2, n)) == NULL) { + goto malloc_error; + } + if ((c3 = mpd_calloc(sizeof *c3, n)) == NULL) { + goto malloc_error; + } + + memcpy(c1, u, ulen * (sizeof *c1)); + memcpy(c2, u, ulen * (sizeof *c2)); + memcpy(c3, u, ulen * (sizeof *c3)); + + if (u == v) { + if (!fnt_autoconvolute(c1, n, P1) || + !fnt_autoconvolute(c2, n, P2) || + !fnt_autoconvolute(c3, n, P3)) { + goto malloc_error; + } + } + else { + if ((vtmp = mpd_calloc(sizeof *vtmp, n)) == NULL) { + goto malloc_error; + } + + memcpy(vtmp, v, vlen * (sizeof *vtmp)); + if (!fnt_convolute(c1, vtmp, n, P1)) { + mpd_free(vtmp); + goto malloc_error; + } + + memcpy(vtmp, v, vlen * (sizeof *vtmp)); + mpd_uint_zero(vtmp+vlen, n-vlen); + if (!fnt_convolute(c2, vtmp, n, P2)) { + mpd_free(vtmp); + goto malloc_error; + } + + memcpy(vtmp, v, vlen * (sizeof *vtmp)); + mpd_uint_zero(vtmp+vlen, n-vlen); + if (!fnt_convolute(c3, vtmp, n, P3)) { + mpd_free(vtmp); + goto malloc_error; + } + + mpd_free(vtmp); + } + + crt3(c1, c2, c3, *rsize); + +out: +#ifdef PPRO + mpd_restore_fenv(cw); +#endif + if (c2) mpd_free(c2); + if (c3) mpd_free(c3); + return c1; + +malloc_error: + if (c1) mpd_free(c1); + c1 = NULL; + goto out; +} + + +/* + * Karatsuba multiplication with FNT/basemul as the base case. + */ +static int +_karatsuba_rec_fnt(mpd_uint_t *c, const mpd_uint_t *a, const mpd_uint_t *b, + mpd_uint_t *w, mpd_size_t la, mpd_size_t lb) +{ + mpd_size_t m, lt; + + assert(la >= lb && lb > 0); + assert(la <= 3*(MPD_MAXTRANSFORM_2N/2) || w != NULL); + + if (la <= 3*(MPD_MAXTRANSFORM_2N/2)) { + + if (lb <= 192) { + _mpd_basemul(c, b, a, lb, la); + } + else { + mpd_uint_t *result; + mpd_size_t dummy; + + if ((result = _mpd_fntmul(a, b, la, lb, &dummy)) == NULL) { + return 0; + } + memcpy(c, result, (la+lb) * (sizeof *result)); + mpd_free(result); + } + return 1; + } + + m = (la+1)/2; // ceil(la/2) + + /* lb <= m < la */ + if (lb <= m) { + + /* lb can now be larger than la-m */ + if (lb > la-m) { + lt = lb + lb + 1; // space needed for result array + mpd_uint_zero(w, lt); // clear result array + if (!_karatsuba_rec_fnt(w, b, a+m, w+lt, lb, la-m)) { // b*ah + return 0; /* GCOV_UNLIKELY */ + } + } + else { + lt = (la-m) + (la-m) + 1; // space needed for result array + mpd_uint_zero(w, lt); // clear result array + if (!_karatsuba_rec_fnt(w, a+m, b, w+lt, la-m, lb)) { // ah*b + return 0; /* GCOV_UNLIKELY */ + } + } + _mpd_baseaddto(c+m, w, (la-m)+lb); // add ah*b*B**m + + lt = m + m + 1; // space needed for the result array + mpd_uint_zero(w, lt); // clear result array + if (!_karatsuba_rec_fnt(w, a, b, w+lt, m, lb)) { // al*b + return 0; /* GCOV_UNLIKELY */ + } + _mpd_baseaddto(c, w, m+lb); // add al*b + + return 1; + } + + /* la >= lb > m */ + memcpy(w, a, m * sizeof *w); + w[m] = 0; + _mpd_baseaddto(w, a+m, la-m); + + memcpy(w+(m+1), b, m * sizeof *w); + w[m+1+m] = 0; + _mpd_baseaddto(w+(m+1), b+m, lb-m); + + if (!_karatsuba_rec_fnt(c+m, w, w+(m+1), w+2*(m+1), m+1, m+1)) { + return 0; /* GCOV_UNLIKELY */ + } + + lt = (la-m) + (la-m) + 1; + mpd_uint_zero(w, lt); + + if (!_karatsuba_rec_fnt(w, a+m, b+m, w+lt, la-m, lb-m)) { + return 0; /* GCOV_UNLIKELY */ + } + + _mpd_baseaddto(c+2*m, w, (la-m) + (lb-m)); + _mpd_basesubfrom(c+m, w, (la-m) + (lb-m)); + + lt = m + m + 1; + mpd_uint_zero(w, lt); + + if (!_karatsuba_rec_fnt(w, a, b, w+lt, m, m)) { + return 0; /* GCOV_UNLIKELY */ + } + _mpd_baseaddto(c, w, m+m); + _mpd_basesubfrom(c+m, w, m+m); + + return 1; +} + +/* + * Multiply u and v, using Karatsuba multiplication with the FNT as the + * base case. Returns a pointer to the result or NULL in case of failure + * (malloc error). Conditions: ulen >= vlen, ulen >= 4. + */ +mpd_uint_t * +_mpd_kmul_fnt(const mpd_uint_t *u, const mpd_uint_t *v, + mpd_size_t ulen, mpd_size_t vlen, + mpd_size_t *rsize) +{ + mpd_uint_t *result = NULL, *w = NULL; + mpd_size_t m; + + assert(ulen >= 4); + assert(ulen >= vlen); + + *rsize = _kmul_resultsize(ulen, vlen); + if ((result = mpd_calloc(*rsize, sizeof *result)) == NULL) { + return NULL; + } + + m = _kmul_worksize(ulen, 3*(MPD_MAXTRANSFORM_2N/2)); + if (m && ((w = mpd_calloc(m, sizeof *w)) == NULL)) { + mpd_free(result); /* GCOV_UNLIKELY */ + return NULL; /* GCOV_UNLIKELY */ + } + + if (!_karatsuba_rec_fnt(result, u, v, w, ulen, vlen)) { + mpd_free(result); + result = NULL; + } + + + if (w) mpd_free(w); + return result; +} + + +/* Deal with the special cases of multiplying infinities. */ +static void +_mpd_qmul_inf(mpd_t *result, const mpd_t *a, const mpd_t *b, uint32_t *status) +{ + if (mpd_isinfinite(a)) { + if (mpd_iszero(b)) { + mpd_seterror(result, MPD_Invalid_operation, status); + } + else { + mpd_setspecial(result, mpd_sign(a)^mpd_sign(b), MPD_INF); + } + return; + } + assert(mpd_isinfinite(b)); + if (mpd_iszero(a)) { + mpd_seterror(result, MPD_Invalid_operation, status); + } + else { + mpd_setspecial(result, mpd_sign(a)^mpd_sign(b), MPD_INF); + } +} + +/* + * Internal function: Multiply a and b. _mpd_qmul deals with specials but + * does NOT finalize the result. This is for use in mpd_fma(). + */ +static inline void +_mpd_qmul(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_t *big = (mpd_t *)a, *small = (mpd_t *)b; + mpd_uint_t *rdata = NULL; + mpd_uint_t rbuf[MPD_MINALLOC_MAX]; + mpd_size_t rsize, i; + + + if (mpd_isspecial(a) || mpd_isspecial(b)) { + if (mpd_qcheck_nans(result, a, b, ctx, status)) { + return; + } + _mpd_qmul_inf(result, a, b, status); + return; + } + + if (small->len > big->len) { + _mpd_ptrswap(&big, &small); + } + + rsize = big->len + small->len; + + if (big->len == 1) { + _mpd_singlemul(result->data, big->data[0], small->data[0]); + goto finish; + } + if (rsize <= (mpd_size_t)MPD_MINALLOC_MAX) { + if (big->len == 2) { + _mpd_mul_2_le2(rbuf, big->data, small->data, small->len); + } + else { + mpd_uint_zero(rbuf, rsize); + if (small->len == 1) { + _mpd_shortmul(rbuf, big->data, big->len, small->data[0]); + } + else { + _mpd_basemul(rbuf, small->data, big->data, small->len, big->len); + } + } + if (!mpd_qresize(result, rsize, status)) { + return; + } + for(i = 0; i < rsize; i++) { + result->data[i] = rbuf[i]; + } + goto finish; + } + + + if (small->len == 1) { + if ((rdata = mpd_calloc(rsize, sizeof *rdata)) == NULL) { + mpd_seterror(result, MPD_Malloc_error, status); + return; + } + _mpd_shortmul(rdata, big->data, big->len, small->data[0]); + } + else if (rsize <= 1024) { + rdata = _mpd_kmul(big->data, small->data, big->len, small->len, &rsize); + if (rdata == NULL) { + mpd_seterror(result, MPD_Malloc_error, status); + return; + } + } + else if (rsize <= 3*MPD_MAXTRANSFORM_2N) { + rdata = _mpd_fntmul(big->data, small->data, big->len, small->len, &rsize); + if (rdata == NULL) { + mpd_seterror(result, MPD_Malloc_error, status); + return; + } + } + else { + rdata = _mpd_kmul_fnt(big->data, small->data, big->len, small->len, &rsize); + if (rdata == NULL) { + mpd_seterror(result, MPD_Malloc_error, status); /* GCOV_UNLIKELY */ + return; /* GCOV_UNLIKELY */ + } + } + + if (mpd_isdynamic_data(result)) { + mpd_free(result->data); + } + result->data = rdata; + result->alloc = rsize; + mpd_set_dynamic_data(result); + + +finish: + mpd_set_flags(result, mpd_sign(a)^mpd_sign(b)); + result->exp = big->exp + small->exp; + result->len = _mpd_real_size(result->data, rsize); + /* resize to smaller cannot fail */ + mpd_qresize(result, result->len, status); + mpd_setdigits(result); +} + +/* Multiply a and b. */ +void +mpd_qmul(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + _mpd_qmul(result, a, b, ctx, status); + mpd_qfinalize(result, ctx, status); +} + +/* Multiply decimal and mpd_ssize_t. */ +void +mpd_qmul_ssize(mpd_t *result, const mpd_t *a, mpd_ssize_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_context_t maxcontext; + MPD_NEW_STATIC(bb,0,0,0,0); + + mpd_maxcontext(&maxcontext); + mpd_qsset_ssize(&bb, b, &maxcontext, status); + mpd_qmul(result, a, &bb, ctx, status); + mpd_del(&bb); +} + +/* Multiply decimal and mpd_uint_t. */ +void +mpd_qmul_uint(mpd_t *result, const mpd_t *a, mpd_uint_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_context_t maxcontext; + MPD_NEW_STATIC(bb,0,0,0,0); + + mpd_maxcontext(&maxcontext); + mpd_qsset_uint(&bb, b, &maxcontext, status); + mpd_qmul(result, a, &bb, ctx, status); + mpd_del(&bb); +} + +void +mpd_qmul_i32(mpd_t *result, const mpd_t *a, int32_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_qmul_ssize(result, a, b, ctx, status); +} + +void +mpd_qmul_u32(mpd_t *result, const mpd_t *a, uint32_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_qmul_uint(result, a, b, ctx, status); +} + +#ifdef CONFIG_64 +void +mpd_qmul_i64(mpd_t *result, const mpd_t *a, int64_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_qmul_ssize(result, a, b, ctx, status); +} + +void +mpd_qmul_u64(mpd_t *result, const mpd_t *a, uint64_t b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_qmul_uint(result, a, b, ctx, status); +} +#endif + +/* Like the minus operator. */ +void +mpd_qminus(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + if (mpd_isspecial(a)) { + if (mpd_qcheck_nan(result, a, ctx, status)) { + return; + } + } + + if (mpd_iszero(a) && ctx->round != MPD_ROUND_FLOOR) { + mpd_qcopy_abs(result, a, status); + } + else { + mpd_qcopy_negate(result, a, status); + } + + mpd_qfinalize(result, ctx, status); +} + +/* Like the plus operator. */ +void +mpd_qplus(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + if (mpd_isspecial(a)) { + if (mpd_qcheck_nan(result, a, ctx, status)) { + return; + } + } + + if (mpd_iszero(a) && ctx->round != MPD_ROUND_FLOOR) { + mpd_qcopy_abs(result, a, status); + } + else { + mpd_qcopy(result, a, status); + } + + mpd_qfinalize(result, ctx, status); +} + +/* The largest representable number that is smaller than the operand. */ +void +mpd_qnext_minus(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_context_t workctx; /* function context */ + MPD_NEW_CONST(tiny,MPD_POS,mpd_etiny(ctx)-1,1,1,1,1); + + if (mpd_isspecial(a)) { + if (mpd_qcheck_nan(result, a, ctx, status)) { + return; + } + if (mpd_isinfinite(a)) { + if (mpd_isnegative(a)) { + mpd_qcopy(result, a, status); + return; + } + else { + mpd_clear_flags(result); + mpd_qmaxcoeff(result, ctx, status); + if (mpd_isnan(result)) { + return; + } + result->exp = ctx->emax - ctx->prec + 1; + return; + } + } + /* debug */ + abort(); /* GCOV_NOT_REACHED */ + } + + mpd_workcontext(&workctx, ctx); + workctx.round = MPD_ROUND_FLOOR; + + if (!mpd_qcopy(result, a, status)) { + return; + } + + mpd_qfinalize(result, &workctx, &workctx.status); + if (workctx.status&(MPD_Inexact|MPD_Errors)) { + *status |= (workctx.status&MPD_Errors); + return; + } + + workctx.status = 0; + mpd_qsub(result, a, &tiny, &workctx, &workctx.status); + *status |= (workctx.status&MPD_Errors); +} + +/* The smallest representable number that is larger than the operand. */ +void +mpd_qnext_plus(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_context_t workctx; + MPD_NEW_CONST(tiny,MPD_POS,mpd_etiny(ctx)-1,1,1,1,1); + + if (mpd_isspecial(a)) { + if (mpd_qcheck_nan(result, a, ctx, status)) { + return; + } + if (mpd_isinfinite(a)) { + if (mpd_ispositive(a)) { + mpd_qcopy(result, a, status); + } + else { + mpd_clear_flags(result); + mpd_qmaxcoeff(result, ctx, status); + if (mpd_isnan(result)) { + return; + } + mpd_set_flags(result, MPD_NEG); + result->exp = mpd_etop(ctx); + } + return; + } + } + + mpd_workcontext(&workctx, ctx); + workctx.round = MPD_ROUND_CEILING; + + if (!mpd_qcopy(result, a, status)) { + return; + } + + mpd_qfinalize(result, &workctx, &workctx.status); + if (workctx.status & (MPD_Inexact|MPD_Errors)) { + *status |= (workctx.status&MPD_Errors); + return; + } + + workctx.status = 0; + mpd_qadd(result, a, &tiny, &workctx, &workctx.status); + *status |= (workctx.status&MPD_Errors); +} + +/* + * The number closest to the first operand that is in the direction towards + * the second operand. + */ +void +mpd_qnext_toward(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + int c; + + if (mpd_isnan(a) || mpd_isnan(b)) { + if (mpd_qcheck_nans(result, a, b, ctx, status)) + return; + } + + c = _mpd_cmp(a, b); + if (c == 0) { + mpd_qcopy_sign(result, a, b, status); + return; + } + + if (c < 0) { + mpd_qnext_plus(result, a, ctx, status); + } + else { + mpd_qnext_minus(result, a, ctx, status); + } + + if (mpd_isinfinite(result)) { + *status |= (MPD_Overflow|MPD_Rounded|MPD_Inexact); + } + else if (mpd_adjexp(result) < ctx->emin) { + *status |= (MPD_Underflow|MPD_Subnormal|MPD_Rounded|MPD_Inexact); + if (mpd_iszero(result)) { + *status |= MPD_Clamped; + } + } +} + +/* + * Internal function: Integer power with mpd_uint_t exponent, base is modified! + * Function can fail with MPD_Malloc_error. + */ +static inline void +_mpd_qpow_uint(mpd_t *result, mpd_t *base, mpd_uint_t exp, uint8_t resultsign, + const mpd_context_t *ctx, uint32_t *status) +{ + uint32_t workstatus = 0; + mpd_uint_t n; + + if (exp == 0) { + _settriple(result, resultsign, 1, 0); /* GCOV_NOT_REACHED */ + return; /* GCOV_NOT_REACHED */ + } + + if (!mpd_qcopy(result, base, status)) { + return; + } + + n = mpd_bits[mpd_bsr(exp)]; + while (n >>= 1) { + mpd_qmul(result, result, result, ctx, &workstatus); + if (exp & n) { + mpd_qmul(result, result, base, ctx, &workstatus); + } + if (workstatus & (MPD_Overflow|MPD_Clamped)) { + break; + } + } + + *status |= workstatus; + mpd_set_sign(result, resultsign); +} + +/* + * Internal function: Integer power with mpd_t exponent, tbase and texp + * are modified!! Function can fail with MPD_Malloc_error. + */ +static inline void +_mpd_qpow_mpd(mpd_t *result, mpd_t *tbase, mpd_t *texp, uint8_t resultsign, + const mpd_context_t *ctx, uint32_t *status) +{ + uint32_t workstatus = 0; + mpd_context_t maxctx; + MPD_NEW_CONST(two,0,0,1,1,1,2); + + + mpd_maxcontext(&maxctx); + + /* resize to smaller cannot fail */ + mpd_qcopy(result, &one, status); + + while (!mpd_iszero(texp)) { + if (mpd_isodd(texp)) { + mpd_qmul(result, result, tbase, ctx, &workstatus); + *status |= workstatus; + if (workstatus & (MPD_Overflow|MPD_Clamped)) { + break; + } + } + mpd_qmul(tbase, tbase, tbase, ctx, &workstatus); + mpd_qdivint(texp, texp, &two, &maxctx, &workstatus); + if (mpd_isnan(tbase) || mpd_isnan(texp)) { + mpd_seterror(result, workstatus&MPD_Errors, status); + return; + } + } + mpd_set_sign(result, resultsign); +} + +/* + * The power function for integer exponents. + */ +static void +_mpd_qpow_int(mpd_t *result, const mpd_t *base, const mpd_t *exp, + uint8_t resultsign, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_context_t workctx; + MPD_NEW_STATIC(tbase,0,0,0,0); + MPD_NEW_STATIC(texp,0,0,0,0); + mpd_ssize_t n; + + + mpd_workcontext(&workctx, ctx); + workctx.prec += (exp->digits + exp->exp + 2); + workctx.round = MPD_ROUND_HALF_EVEN; + workctx.clamp = 0; + if (mpd_isnegative(exp)) { + mpd_qdiv(&tbase, &one, base, &workctx, status); + if (*status&MPD_Errors) { + mpd_setspecial(result, MPD_POS, MPD_NAN); + goto finish; + } + } + else { + if (!mpd_qcopy(&tbase, base, status)) { + mpd_setspecial(result, MPD_POS, MPD_NAN); + goto finish; + } + } + + n = mpd_qabs_uint(exp, &workctx.status); + if (workctx.status&MPD_Invalid_operation) { + if (!mpd_qcopy(&texp, exp, status)) { + mpd_setspecial(result, MPD_POS, MPD_NAN); /* GCOV_UNLIKELY */ + goto finish; /* GCOV_UNLIKELY */ + } + _mpd_qpow_mpd(result, &tbase, &texp, resultsign, &workctx, status); + } + else { + _mpd_qpow_uint(result, &tbase, n, resultsign, &workctx, status); + } + + if (mpd_isinfinite(result)) { + /* for ROUND_DOWN, ROUND_FLOOR, etc. */ + _settriple(result, resultsign, 1, MPD_EXP_INF); + } + +finish: + mpd_del(&tbase); + mpd_del(&texp); + mpd_qfinalize(result, ctx, status); +} + +/* + * This is an internal function that does not check for NaNs. + */ +static int +_qcheck_pow_one_inf(mpd_t *result, const mpd_t *base, uint8_t resultsign, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_ssize_t shift; + int cmp; + + if ((cmp = _mpd_cmp(base, &one)) == 0) { + shift = ctx->prec-1; + mpd_qshiftl(result, &one, shift, status); + result->exp = -shift; + mpd_set_flags(result, resultsign); + *status |= (MPD_Inexact|MPD_Rounded); + } + + return cmp; +} + +/* + * If base equals one, calculate the correct power of one result. + * Otherwise, result is undefined. Return the value of the comparison + * against 1. + * + * This is an internal function that does not check for specials. + */ +static int +_qcheck_pow_one(mpd_t *result, const mpd_t *base, const mpd_t *exp, + uint8_t resultsign, + const mpd_context_t *ctx, uint32_t *status) +{ + uint32_t workstatus = 0; + mpd_ssize_t shift; + int cmp; + + if ((cmp = _mpd_cmp_abs(base, &one)) == 0) { + if (_mpd_isint(exp)) { + if (mpd_isnegative(exp)) { + _settriple(result, resultsign, 1, 0); + return 0; + } + /* 1.000**3 = 1.000000000 */ + mpd_qmul_ssize(result, exp, -base->exp, ctx, &workstatus); + if (workstatus&MPD_Errors) { + *status |= (workstatus&MPD_Errors); + return 0; + } + /* digits-1 after exponentiation */ + shift = mpd_qget_ssize(result, &workstatus); + /* shift is MPD_SSIZE_MAX if result is too large */ + if (shift > ctx->prec-1) { + shift = ctx->prec-1; + *status |= MPD_Rounded; + } + } + else if (mpd_ispositive(base)) { + shift = ctx->prec-1; + *status |= (MPD_Inexact|MPD_Rounded); + } + else { + return -2; /* GCOV_NOT_REACHED */ + } + if (!mpd_qshiftl(result, &one, shift, status)) { + return 0; + } + result->exp = -shift; + mpd_set_flags(result, resultsign); + } + + return cmp; +} + +/* + * Detect certain over/underflow of x**y. + * ACL2 proof: pow_bounds.lisp. + * + * Symbols: + * + * e: EXP_INF or EXP_CLAMP + * x: base + * y: exponent + * + * omega(e) = log10(abs(e)) + * zeta(x) = log10(abs(log10(x))) + * theta(y) = log10(abs(y)) + * + * Upper and lower bounds: + * + * ub_omega(e) = ceil(log10(abs(e))) + * lb_theta(y) = floor(log10(abs(y))) + * + * | floor(log10(floor(abs(log10(x))))) if x < 1/10 or x >= 10 + * lb_zeta(x) = | floor(log10(abs(x-1)/10)) if 1/10 <= x < 1 + * | floor(log10(abs((x-1)/100))) if 1 < x < 10 + * + * ub_omega(e) and lb_theta(y) are obviously upper and lower bounds + * for omega(e) and theta(y). + * + * lb_zeta is a lower bound for zeta(x): + * + * x < 1/10 or x >= 10: + * + * abs(log10(x)) >= 1, so the outer log10 is well defined. Since log10 + * is strictly increasing, the end result is a lower bound. + * + * 1/10 <= x < 1: + * + * We use: log10(x) <= (x-1)/log(10) + * abs(log10(x)) >= abs(x-1)/log(10) + * abs(log10(x)) >= abs(x-1)/10 + * + * 1 < x < 10: + * + * We use: (x-1)/(x*log(10)) < log10(x) + * abs((x-1)/100) < abs(log10(x)) + * + * XXX: abs((x-1)/10) would work, need ACL2 proof. + * + * + * Let (0 < x < 1 and y < 0) or (x > 1 and y > 0). (H1) + * Let ub_omega(exp_inf) < lb_zeta(x) + lb_theta(y) (H2) + * + * Then: + * log10(abs(exp_inf)) < log10(abs(log10(x))) + log10(abs(y)). (1) + * exp_inf < log10(x) * y (2) + * 10**exp_inf < x**y (3) + * + * Let (0 < x < 1 and y > 0) or (x > 1 and y < 0). (H3) + * Let ub_omega(exp_clamp) < lb_zeta(x) + lb_theta(y) (H4) + * + * Then: + * log10(abs(exp_clamp)) < log10(abs(log10(x))) + log10(abs(y)). (4) + * log10(x) * y < exp_clamp (5) + * x**y < 10**exp_clamp (6) + * + */ +static mpd_ssize_t +_lower_bound_zeta(const mpd_t *x, uint32_t *status) +{ + mpd_context_t maxctx; + MPD_NEW_STATIC(scratch,0,0,0,0); + mpd_ssize_t t, u; + + t = mpd_adjexp(x); + if (t > 0) { + /* x >= 10 -> floor(log10(floor(abs(log10(x))))) */ + return mpd_exp_digits(t) - 1; + } + else if (t < -1) { + /* x < 1/10 -> floor(log10(floor(abs(log10(x))))) */ + return mpd_exp_digits(t+1) - 1; + } + else { + mpd_maxcontext(&maxctx); + mpd_qsub(&scratch, x, &one, &maxctx, status); + if (mpd_isspecial(&scratch)) { + mpd_del(&scratch); + return MPD_SSIZE_MAX; + } + u = mpd_adjexp(&scratch); + mpd_del(&scratch); + + /* t == -1, 1/10 <= x < 1 -> floor(log10(abs(x-1)/10)) + * t == 0, 1 < x < 10 -> floor(log10(abs(x-1)/100)) */ + return (t == 0) ? u-2 : u-1; + } +} + +/* + * Detect cases of certain overflow/underflow in the power function. + * Assumptions: x != 1, y != 0. The proof above is for positive x. + * If x is negative and y is an odd integer, x**y == -(abs(x)**y), + * so the analysis does not change. + */ +static int +_qcheck_pow_bounds(mpd_t *result, const mpd_t *x, const mpd_t *y, + uint8_t resultsign, + const mpd_context_t *ctx, uint32_t *status) +{ + MPD_NEW_SHARED(abs_x, x); + mpd_ssize_t ub_omega, lb_zeta, lb_theta; + uint8_t sign; + + mpd_set_positive(&abs_x); + + lb_theta = mpd_adjexp(y); + lb_zeta = _lower_bound_zeta(&abs_x, status); + if (lb_zeta == MPD_SSIZE_MAX) { + mpd_seterror(result, MPD_Malloc_error, status); + return 1; + } + + sign = (mpd_adjexp(&abs_x) < 0) ^ mpd_sign(y); + if (sign == 0) { + /* (0 < |x| < 1 and y < 0) or (|x| > 1 and y > 0) */ + ub_omega = mpd_exp_digits(ctx->emax); + if (ub_omega < lb_zeta + lb_theta) { + _settriple(result, resultsign, 1, MPD_EXP_INF); + mpd_qfinalize(result, ctx, status); + return 1; + } + } + else { + /* (0 < |x| < 1 and y > 0) or (|x| > 1 and y < 0). */ + ub_omega = mpd_exp_digits(mpd_etiny(ctx)); + if (ub_omega < lb_zeta + lb_theta) { + _settriple(result, resultsign, 1, mpd_etiny(ctx)-1); + mpd_qfinalize(result, ctx, status); + return 1; + } + } + + return 0; +} + +/* + * TODO: Implement algorithm for computing exact powers from decimal.py. + * In order to prevent infinite loops, this has to be called before + * using Ziv's strategy for correct rounding. + */ +/* +static int +_mpd_qpow_exact(mpd_t *result, const mpd_t *base, const mpd_t *exp, + const mpd_context_t *ctx, uint32_t *status) +{ + return 0; +} +*/ + +/* The power function for real exponents */ +static void +_mpd_qpow_real(mpd_t *result, const mpd_t *base, const mpd_t *exp, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_context_t workctx; + MPD_NEW_STATIC(texp,0,0,0,0); + + if (!mpd_qcopy(&texp, exp, status)) { + mpd_seterror(result, MPD_Malloc_error, status); + return; + } + + mpd_maxcontext(&workctx); + workctx.prec = (base->digits > ctx->prec) ? base->digits : ctx->prec; + workctx.prec += (4 + MPD_EXPDIGITS); + workctx.round = MPD_ROUND_HALF_EVEN; + workctx.allcr = ctx->allcr; + + mpd_qln(result, base, &workctx, &workctx.status); + mpd_qmul(result, result, &texp, &workctx, &workctx.status); + mpd_qexp(result, result, &workctx, status); + + mpd_del(&texp); + *status |= (workctx.status&MPD_Errors); + *status |= (MPD_Inexact|MPD_Rounded); +} + +/* The power function: base**exp */ +void +mpd_qpow(mpd_t *result, const mpd_t *base, const mpd_t *exp, + const mpd_context_t *ctx, uint32_t *status) +{ + uint8_t resultsign = 0; + int intexp = 0; + int cmp; + + if (mpd_isspecial(base) || mpd_isspecial(exp)) { + if (mpd_qcheck_nans(result, base, exp, ctx, status)) { + return; + } + } + if (mpd_isinteger(exp)) { + intexp = 1; + resultsign = mpd_isnegative(base) && mpd_isodd(exp); + } + + if (mpd_iszero(base)) { + if (mpd_iszero(exp)) { + mpd_seterror(result, MPD_Invalid_operation, status); + } + else if (mpd_isnegative(exp)) { + mpd_setspecial(result, resultsign, MPD_INF); + } + else { + _settriple(result, resultsign, 0, 0); + } + return; + } + if (mpd_isnegative(base)) { + if (!intexp || mpd_isinfinite(exp)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + } + if (mpd_isinfinite(exp)) { + /* power of one */ + cmp = _qcheck_pow_one_inf(result, base, resultsign, ctx, status); + if (cmp == 0) { + return; + } + else { + cmp *= mpd_arith_sign(exp); + if (cmp < 0) { + _settriple(result, resultsign, 0, 0); + } + else { + mpd_setspecial(result, resultsign, MPD_INF); + } + } + return; + } + if (mpd_isinfinite(base)) { + if (mpd_iszero(exp)) { + _settriple(result, resultsign, 1, 0); + } + else if (mpd_isnegative(exp)) { + _settriple(result, resultsign, 0, 0); + } + else { + mpd_setspecial(result, resultsign, MPD_INF); + } + return; + } + if (mpd_iszero(exp)) { + _settriple(result, resultsign, 1, 0); + return; + } + if (_qcheck_pow_one(result, base, exp, resultsign, ctx, status) == 0) { + return; + } + if (_qcheck_pow_bounds(result, base, exp, resultsign, ctx, status)) { + return; + } + + if (intexp) { + _mpd_qpow_int(result, base, exp, resultsign, ctx, status); + } + else { + _mpd_qpow_real(result, base, exp, ctx, status); + if (!mpd_isspecial(result) && _mpd_cmp(result, &one) == 0) { + mpd_ssize_t shift = ctx->prec-1; + mpd_qshiftl(result, &one, shift, status); + result->exp = -shift; + } + if (mpd_isinfinite(result)) { + /* for ROUND_DOWN, ROUND_FLOOR, etc. */ + _settriple(result, MPD_POS, 1, MPD_EXP_INF); + } + mpd_qfinalize(result, ctx, status); + } +} + +/* + * Internal function: Integer powmod with mpd_uint_t exponent, base is modified! + * Function can fail with MPD_Malloc_error. + */ +static inline void +_mpd_qpowmod_uint(mpd_t *result, mpd_t *base, mpd_uint_t exp, + mpd_t *mod, uint32_t *status) +{ + mpd_context_t maxcontext; + + mpd_maxcontext(&maxcontext); + + /* resize to smaller cannot fail */ + mpd_qcopy(result, &one, status); + + while (exp > 0) { + if (exp & 1) { + mpd_qmul(result, result, base, &maxcontext, status); + mpd_qrem(result, result, mod, &maxcontext, status); + } + mpd_qmul(base, base, base, &maxcontext, status); + mpd_qrem(base, base, mod, &maxcontext, status); + exp >>= 1; + } +} + +/* The powmod function: (base**exp) % mod */ +void +mpd_qpowmod(mpd_t *result, const mpd_t *base, const mpd_t *exp, + const mpd_t *mod, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_context_t maxcontext; + MPD_NEW_STATIC(tbase,0,0,0,0); + MPD_NEW_STATIC(texp,0,0,0,0); + MPD_NEW_STATIC(tmod,0,0,0,0); + MPD_NEW_STATIC(tmp,0,0,0,0); + MPD_NEW_CONST(two,0,0,1,1,1,2); + mpd_ssize_t tbase_exp, texp_exp; + mpd_ssize_t i; + mpd_t t; + mpd_uint_t r; + uint8_t sign; + + + if (mpd_isspecial(base) || mpd_isspecial(exp) || mpd_isspecial(mod)) { + if (mpd_qcheck_3nans(result, base, exp, mod, ctx, status)) { + return; + } + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + + + if (!_mpd_isint(base) || !_mpd_isint(exp) || !_mpd_isint(mod)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if (mpd_iszerocoeff(mod)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if (mod->digits+mod->exp > ctx->prec) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + + sign = (mpd_isnegative(base)) && (mpd_isodd(exp)); + if (mpd_iszerocoeff(exp)) { + if (mpd_iszerocoeff(base)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + r = (_mpd_cmp_abs(mod, &one)==0) ? 0 : 1; + _settriple(result, sign, r, 0); + return; + } + if (mpd_isnegative(exp)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if (mpd_iszerocoeff(base)) { + _settriple(result, sign, 0, 0); + return; + } + + if (!mpd_qcopy(&tmod, mod, status)) { + goto mpd_errors; + } + mpd_set_positive(&tmod); + + mpd_maxcontext(&maxcontext); + + mpd_qround_to_int(&tbase, base, &maxcontext, status); + mpd_qround_to_int(&texp, exp, &maxcontext, status); + mpd_qround_to_int(&tmod, &tmod, &maxcontext, status); + + tbase_exp = tbase.exp; + tbase.exp = 0; + texp_exp = texp.exp; + texp.exp = 0; + + /* base = (base.int % modulo * pow(10, base.exp, modulo)) % modulo */ + mpd_qrem(&tbase, &tbase, &tmod, &maxcontext, status); + _settriple(result, MPD_POS, 1, tbase_exp); + mpd_qrem(result, result, &tmod, &maxcontext, status); + mpd_qmul(&tbase, &tbase, result, &maxcontext, status); + mpd_qrem(&tbase, &tbase, &tmod, &maxcontext, status); + if (mpd_isspecial(&tbase) || + mpd_isspecial(&texp) || + mpd_isspecial(&tmod)) { + goto mpd_errors; + } + + for (i = 0; i < texp_exp; i++) { + _mpd_qpowmod_uint(&tmp, &tbase, 10, &tmod, status); + t = tmp; + tmp = tbase; + tbase = t; + } + if (mpd_isspecial(&tbase)) { + goto mpd_errors; /* GCOV_UNLIKELY */ + } + + /* resize to smaller cannot fail */ + mpd_qcopy(result, &one, status); + while (mpd_isfinite(&texp) && !mpd_iszero(&texp)) { + if (mpd_isodd(&texp)) { + mpd_qmul(result, result, &tbase, &maxcontext, status); + mpd_qrem(result, result, &tmod, &maxcontext, status); + } + mpd_qmul(&tbase, &tbase, &tbase, &maxcontext, status); + mpd_qrem(&tbase, &tbase, &tmod, &maxcontext, status); + mpd_qdivint(&texp, &texp, &two, &maxcontext, status); + } + if (mpd_isspecial(&texp) || mpd_isspecial(&tbase) || + mpd_isspecial(&tmod) || mpd_isspecial(result)) { + /* MPD_Malloc_error */ + goto mpd_errors; + } + else { + mpd_set_sign(result, sign); + } + +out: + mpd_del(&tbase); + mpd_del(&texp); + mpd_del(&tmod); + mpd_del(&tmp); + mpd_qfinalize(result, ctx, status); + return; + +mpd_errors: + mpd_setspecial(result, MPD_POS, MPD_NAN); + goto out; +} + +void +mpd_qquantize(mpd_t *result, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + uint32_t workstatus = 0; + mpd_ssize_t b_exp = b->exp; + mpd_ssize_t expdiff, shift; + mpd_uint_t rnd; + + if (mpd_isspecial(a) || mpd_isspecial(b)) { + if (mpd_qcheck_nans(result, a, b, ctx, status)) { + return; + } + if (mpd_isinfinite(a) && mpd_isinfinite(b)) { + mpd_qcopy(result, a, status); + return; + } + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + + if (b->exp > ctx->emax || b->exp < mpd_etiny(ctx)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + + if (mpd_iszero(a)) { + _settriple(result, mpd_sign(a), 0, b->exp); + mpd_qfinalize(result, ctx, status); + return; + } + + + expdiff = a->exp - b->exp; + if (a->digits + expdiff > ctx->prec) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + + if (expdiff >= 0) { + shift = expdiff; + if (!mpd_qshiftl(result, a, shift, status)) { + return; + } + result->exp = b_exp; + } + else { + /* At this point expdiff < 0 and a->digits+expdiff <= prec, + * so the shift before an increment will fit in prec. */ + shift = -expdiff; + rnd = mpd_qshiftr(result, a, shift, status); + if (rnd == MPD_UINT_MAX) { + return; + } + result->exp = b_exp; + if (!_mpd_apply_round_fit(result, rnd, ctx, status)) { + return; + } + workstatus |= MPD_Rounded; + if (rnd) { + workstatus |= MPD_Inexact; + } + } + + if (mpd_adjexp(result) > ctx->emax || + mpd_adjexp(result) < mpd_etiny(ctx)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + + *status |= workstatus; + mpd_qfinalize(result, ctx, status); +} + +void +mpd_qreduce(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_ssize_t shift, maxexp, maxshift; + uint8_t sign_a = mpd_sign(a); + + if (mpd_isspecial(a)) { + if (mpd_qcheck_nan(result, a, ctx, status)) { + return; + } + mpd_qcopy(result, a, status); + return; + } + + if (!mpd_qcopy(result, a, status)) { + return; + } + mpd_qfinalize(result, ctx, status); + if (mpd_isspecial(result)) { + return; + } + if (mpd_iszero(result)) { + _settriple(result, sign_a, 0, 0); + return; + } + + shift = mpd_trail_zeros(result); + maxexp = (ctx->clamp) ? mpd_etop(ctx) : ctx->emax; + /* After the finalizing above result->exp <= maxexp. */ + maxshift = maxexp - result->exp; + shift = (shift > maxshift) ? maxshift : shift; + + mpd_qshiftr_inplace(result, shift); + result->exp += shift; +} + +void +mpd_qrem(mpd_t *r, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, + uint32_t *status) +{ + MPD_NEW_STATIC(q,0,0,0,0); + + if (mpd_isspecial(a) || mpd_isspecial(b)) { + if (mpd_qcheck_nans(r, a, b, ctx, status)) { + return; + } + if (mpd_isinfinite(a)) { + mpd_seterror(r, MPD_Invalid_operation, status); + return; + } + if (mpd_isinfinite(b)) { + mpd_qcopy(r, a, status); + mpd_qfinalize(r, ctx, status); + return; + } + /* debug */ + abort(); /* GCOV_NOT_REACHED */ + } + if (mpd_iszerocoeff(b)) { + if (mpd_iszerocoeff(a)) { + mpd_seterror(r, MPD_Division_undefined, status); + } + else { + mpd_seterror(r, MPD_Invalid_operation, status); + } + return; + } + + _mpd_qdivmod(&q, r, a, b, ctx, status); + mpd_del(&q); + mpd_qfinalize(r, ctx, status); +} + +void +mpd_qrem_near(mpd_t *r, const mpd_t *a, const mpd_t *b, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_context_t workctx; + MPD_NEW_STATIC(btmp,0,0,0,0); + MPD_NEW_STATIC(q,0,0,0,0); + mpd_ssize_t expdiff, floordigits; + int cmp, isodd, allnine; + + if (mpd_isspecial(a) || mpd_isspecial(b)) { + if (mpd_qcheck_nans(r, a, b, ctx, status)) { + return; + } + if (mpd_isinfinite(a)) { + mpd_seterror(r, MPD_Invalid_operation, status); + return; + } + if (mpd_isinfinite(b)) { + mpd_qcopy(r, a, status); + mpd_qfinalize(r, ctx, status); + return; + } + /* debug */ + abort(); /* GCOV_NOT_REACHED */ + } + if (mpd_iszerocoeff(b)) { + if (mpd_iszerocoeff(a)) { + mpd_seterror(r, MPD_Division_undefined, status); + } + else { + mpd_seterror(r, MPD_Invalid_operation, status); + } + return; + } + + if (r == b) { + if (!mpd_qcopy(&btmp, b, status)) { + mpd_seterror(r, MPD_Malloc_error, status); + return; + } + b = &btmp; + } + + workctx = *ctx; + workctx.prec = a->digits; + workctx.prec = (workctx.prec > ctx->prec) ? workctx.prec : ctx->prec; + + _mpd_qdivmod(&q, r, a, b, &workctx, status); + if (mpd_isnan(&q) || mpd_isnan(r) || q.digits > ctx->prec) { + mpd_seterror(r, MPD_Division_impossible, status); + goto finish; + } + if (mpd_iszerocoeff(r)) { + goto finish; + } + + /* Deal with cases like rmnx078: + * remaindernear 999999999.5 1 -> NaN Division_impossible */ + expdiff = mpd_adjexp(b) - mpd_adjexp(r); + if (-1 <= expdiff && expdiff <= 1) { + + mpd_qtrunc(&q, &q, &workctx, &workctx.status); + allnine = mpd_coeff_isallnine(&q); + floordigits = q.digits; + isodd = mpd_isodd(&q); + + mpd_maxcontext(&workctx); + if (mpd_sign(a) == mpd_sign(b)) { + _mpd_qsub(&q, r, b, &workctx, &workctx.status); + if (workctx.status&MPD_Errors) { + mpd_seterror(r, workctx.status&MPD_Errors, status); + goto finish; + } + } + else { + _mpd_qadd(&q, r, b, &workctx, &workctx.status); + if (workctx.status&MPD_Errors) { + mpd_seterror(r, workctx.status&MPD_Errors, status); + goto finish; + } + } + + cmp = mpd_cmp_total_mag(&q, r); + if (cmp < 0 || (cmp == 0 && isodd)) { + if (allnine && floordigits == ctx->prec) { + mpd_seterror(r, MPD_Division_impossible, status); + goto finish; + } + mpd_qcopy(r, &q, status); + *status &= ~MPD_Rounded; + } + } + + +finish: + mpd_del(&btmp); + mpd_del(&q); + mpd_qfinalize(r, ctx, status); +} + +static void +_mpd_qrescale(mpd_t *result, const mpd_t *a, mpd_ssize_t exp, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_ssize_t expdiff, shift; + mpd_uint_t rnd; + + if (mpd_isspecial(a)) { + mpd_qcopy(result, a, status); + return; + } + + if (mpd_iszero(a)) { + _settriple(result, mpd_sign(a), 0, exp); + return; + } + + expdiff = a->exp - exp; + if (expdiff >= 0) { + shift = expdiff; + if (a->digits + shift > MPD_MAX_PREC+1) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if (!mpd_qshiftl(result, a, shift, status)) { + return; + } + result->exp = exp; + } + else { + shift = -expdiff; + rnd = mpd_qshiftr(result, a, shift, status); + if (rnd == MPD_UINT_MAX) { + return; + } + result->exp = exp; + _mpd_apply_round_excess(result, rnd, ctx, status); + *status |= MPD_Rounded; + if (rnd) { + *status |= MPD_Inexact; + } + } + + if (mpd_issubnormal(result, ctx)) { + *status |= MPD_Subnormal; + } +} + +/* + * Rescale a number so that it has exponent 'exp'. Does not regard context + * precision, emax, emin, but uses the rounding mode. Special numbers are + * quietly copied. Restrictions: + * + * MPD_MIN_ETINY <= exp <= MPD_MAX_EMAX+1 + * result->digits <= MPD_MAX_PREC+1 + */ +void +mpd_qrescale(mpd_t *result, const mpd_t *a, mpd_ssize_t exp, + const mpd_context_t *ctx, uint32_t *status) +{ + if (exp > MPD_MAX_EMAX+1 || exp < MPD_MIN_ETINY) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + + _mpd_qrescale(result, a, exp, ctx, status); +} + +/* + * Same as mpd_qrescale, but with relaxed restrictions. The result of this + * function should only be used for formatting a number and never as input + * for other operations. + * + * MPD_MIN_ETINY-MPD_MAX_PREC <= exp <= MPD_MAX_EMAX+1 + * result->digits <= MPD_MAX_PREC+1 + */ +void +mpd_qrescale_fmt(mpd_t *result, const mpd_t *a, mpd_ssize_t exp, + const mpd_context_t *ctx, uint32_t *status) +{ + if (exp > MPD_MAX_EMAX+1 || exp < MPD_MIN_ETINY-MPD_MAX_PREC) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + + _mpd_qrescale(result, a, exp, ctx, status); +} + +/* Round to an integer according to 'action' and ctx->round. */ +enum {TO_INT_EXACT, TO_INT_SILENT, TO_INT_TRUNC}; +static void +_mpd_qround_to_integral(int action, mpd_t *result, const mpd_t *a, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_uint_t rnd; + + if (mpd_isspecial(a)) { + if (mpd_qcheck_nan(result, a, ctx, status)) { + return; + } + mpd_qcopy(result, a, status); + return; + } + if (a->exp >= 0) { + mpd_qcopy(result, a, status); + return; + } + if (mpd_iszerocoeff(a)) { + _settriple(result, mpd_sign(a), 0, 0); + return; + } + + rnd = mpd_qshiftr(result, a, -a->exp, status); + if (rnd == MPD_UINT_MAX) { + return; + } + result->exp = 0; + + if (action == TO_INT_EXACT || action == TO_INT_SILENT) { + _mpd_apply_round_excess(result, rnd, ctx, status); + if (action == TO_INT_EXACT) { + *status |= MPD_Rounded; + if (rnd) { + *status |= MPD_Inexact; + } + } + } +} + +void +mpd_qround_to_intx(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + (void)_mpd_qround_to_integral(TO_INT_EXACT, result, a, ctx, status); +} + +void +mpd_qround_to_int(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + (void)_mpd_qround_to_integral(TO_INT_SILENT, result, a, ctx, status); +} + +void +mpd_qtrunc(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + (void)_mpd_qround_to_integral(TO_INT_TRUNC, result, a, ctx, status); +} + +void +mpd_qfloor(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_context_t workctx = *ctx; + workctx.round = MPD_ROUND_FLOOR; + (void)_mpd_qround_to_integral(TO_INT_SILENT, result, a, + &workctx, status); +} + +void +mpd_qceil(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_context_t workctx = *ctx; + workctx.round = MPD_ROUND_CEILING; + (void)_mpd_qround_to_integral(TO_INT_SILENT, result, a, + &workctx, status); +} + +int +mpd_same_quantum(const mpd_t *a, const mpd_t *b) +{ + if (mpd_isspecial(a) || mpd_isspecial(b)) { + return ((mpd_isnan(a) && mpd_isnan(b)) || + (mpd_isinfinite(a) && mpd_isinfinite(b))); + } + + return a->exp == b->exp; +} + +/* Schedule the increase in precision for the Newton iteration. */ +static inline int +recpr_schedule_prec(mpd_ssize_t klist[MPD_MAX_PREC_LOG2], + mpd_ssize_t maxprec, mpd_ssize_t initprec) +{ + mpd_ssize_t k; + int i; + + assert(maxprec > 0 && initprec > 0); + if (maxprec <= initprec) return -1; + + i = 0; k = maxprec; + do { + k = (k+1) / 2; + klist[i++] = k; + } while (k > initprec); + + return i-1; +} + +/* Initial approximation for the reciprocal. */ +static void +_mpd_qreciprocal_approx(mpd_t *z, const mpd_t *v, uint32_t *status) +{ + mpd_uint_t p10data[2] = {0, mpd_pow10[MPD_RDIGITS-2]}; /* 10**(2*MPD_RDIGITS-2) */ + mpd_uint_t dummy, word; + int n; + + assert(v->exp == -v->digits); + + _mpd_get_msdigits(&dummy, &word, v, MPD_RDIGITS); + n = mpd_word_digits(word); + word *= mpd_pow10[MPD_RDIGITS-n]; + + mpd_qresize(z, 2, status); + (void)_mpd_shortdiv(z->data, p10data, 2, word); + + mpd_clear_flags(z); + z->exp = -(MPD_RDIGITS-2); + z->len = (z->data[1] == 0) ? 1 : 2; + mpd_setdigits(z); +} + +/* Reciprocal, calculated with Newton's Method. Assumption: result != a. */ +static void +_mpd_qreciprocal(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + mpd_context_t varcontext, maxcontext; + mpd_t *z = result; /* current approximation */ + mpd_t *v; /* a, normalized to a number between 0.1 and 1 */ + MPD_NEW_SHARED(vtmp, a); /* v shares data with a */ + MPD_NEW_STATIC(s,0,0,0,0); /* temporary variable */ + MPD_NEW_STATIC(t,0,0,0,0); /* temporary variable */ + MPD_NEW_CONST(two,0,0,1,1,1,2); /* const 2 */ + mpd_ssize_t klist[MPD_MAX_PREC_LOG2]; + mpd_ssize_t adj, maxprec, initprec; + uint8_t sign = mpd_sign(a); + int i; + + assert(result != a); + + v = &vtmp; + mpd_clear_flags(v); + adj = v->digits + v->exp; + v->exp = -v->digits; + + /* initial approximation */ + _mpd_qreciprocal_approx(z, v, status); + + mpd_maxcontext(&varcontext); + mpd_maxcontext(&maxcontext); + varcontext.round = MPD_ROUND_TRUNC; + maxcontext.round = MPD_ROUND_TRUNC; + + maxprec = (v->digits > ctx->prec) ? v->digits : ctx->prec; + maxprec += 2; + initprec = MPD_RDIGITS-3; + + i = recpr_schedule_prec(klist, maxprec, initprec); + for (; i >= 0; i--) { + mpd_qmul(&s, z, z, &maxcontext, status); + varcontext.prec = 2*klist[i] + 5; + if (v->digits > varcontext.prec) { + mpd_qshiftr(&t, v, v->digits-varcontext.prec, status); + t.exp = -varcontext.prec; + mpd_qmul(&t, &t, &s, &varcontext, status); + } + else { + mpd_qmul(&t, v, &s, &varcontext, status); + } + mpd_qmul(&s, z, &two, &maxcontext, status); + mpd_qsub(z, &s, &t, &maxcontext, status); + } + + if (!mpd_isspecial(z)) { + z->exp -= adj; + mpd_set_flags(z, sign); + } + + mpd_del(&s); + mpd_del(&t); + mpd_qfinalize(z, ctx, status); +} + +/* + * Integer division with remainder of the coefficients: coeff(a) / coeff(b). + * This function is for large numbers where it is faster to divide by + * multiplying the dividend by the reciprocal of the divisor. + * The inexact result is fixed by a small loop, which should not take + * more than 2 iterations. + */ +static void +_mpd_qbarrett_divmod(mpd_t *q, mpd_t *r, const mpd_t *a, const mpd_t *b, + uint32_t *status) +{ + mpd_context_t workctx; + mpd_t *qq = q, *rr = r; + mpd_t aa, bb; + int k; + + mpd_maxcontext(&workctx); + _mpd_copy_shared(&aa, a); + _mpd_copy_shared(&bb, b); + + mpd_set_positive(&aa); + mpd_set_positive(&bb); + aa.exp = 0; + bb.exp = 0; + + if (q == a || q == b) { + if ((qq = mpd_qnew()) == NULL) { + *status |= MPD_Malloc_error; + goto nanresult; + } + } + if (r == a || r == b) { + if ((rr = mpd_qnew()) == NULL) { + *status |= MPD_Malloc_error; + goto nanresult; + } + } + + /* maximum length of q + 3 digits */ + workctx.prec = aa.digits - bb.digits + 1 + 3; + /* we get the reciprocal with precision maxlen(q) + 3 */ + _mpd_qreciprocal(rr, &bb, &workctx, &workctx.status); + + mpd_qmul(qq, &aa, rr, &workctx, &workctx.status); + mpd_qtrunc(qq, qq, &workctx, &workctx.status); + + workctx.prec = aa.digits + 3; + /* get the remainder */ + mpd_qmul(rr, &bb, qq, &workctx, &workctx.status); + mpd_qsub(rr, &aa, rr, &workctx, &workctx.status); + + /* Fix the result. Algorithm from: Karl Hasselstrom, Fast Division of Large Integers */ + for (k = 0;; k++) { + if (mpd_isspecial(rr)) { + *status |= (workctx.status&MPD_Errors); + goto nanresult; + } + if (k > 2) { + mpd_err_warn("libmpdec: internal error in " /* GCOV_NOT_REACHED */ + "_mpd_qbarrett_divmod: please report"); /* GCOV_NOT_REACHED */ + *status |= MPD_Invalid_operation; /* GCOV_NOT_REACHED */ + goto nanresult; /* GCOV_NOT_REACHED */ + } + else if (_mpd_cmp(&zero, rr) == 1) { + mpd_qadd(rr, rr, &bb, &workctx, &workctx.status); + mpd_qadd(qq, qq, &minus_one, &workctx, &workctx.status); + } + else if (_mpd_cmp(rr, &bb) == -1) { + break; + } + else { + mpd_qsub(rr, rr, &bb, &workctx, &workctx.status); + mpd_qadd(qq, qq, &one, &workctx, &workctx.status); + } + } + + if (qq != q) { + if (!mpd_qcopy(q, qq, status)) { + goto nanresult; /* GCOV_UNLIKELY */ + } + mpd_del(qq); + } + if (rr != r) { + if (!mpd_qcopy(r, rr, status)) { + goto nanresult; /* GCOV_UNLIKELY */ + } + mpd_del(rr); + } + + *status |= (workctx.status&MPD_Errors); + return; + + +nanresult: + if (qq && qq != q) mpd_del(qq); + if (rr && rr != r) mpd_del(rr); + mpd_setspecial(q, MPD_POS, MPD_NAN); + mpd_setspecial(r, MPD_POS, MPD_NAN); +} + +static inline int +invroot_schedule_prec(mpd_ssize_t klist[MPD_MAX_PREC_LOG2], + mpd_ssize_t maxprec, mpd_ssize_t initprec) +{ + mpd_ssize_t k; + int i; + + assert(maxprec >= 3 && initprec >= 3); + if (maxprec <= initprec) return -1; + + i = 0; k = maxprec; + do { + k = (k+3) / 2; + klist[i++] = k; + } while (k > initprec); + + return i-1; +} + +/* + * Initial approximation for the inverse square root. + * + * Input: + * v := 7 or 8 decimal digits with an implicit exponent of 10**-6, + * representing a number 1 <= x < 100. + * + * Output: + * An approximation to 1/sqrt(v) + */ +static inline void +_invroot_init_approx(mpd_t *z, mpd_uint_t v) +{ + mpd_uint_t lo = 1000; + mpd_uint_t hi = 10000; + mpd_uint_t a, sq; + + assert(v >= lo*lo && v < (hi+1)*(hi+1)); + + for(;;) { + a = (lo + hi) / 2; + sq = a * a; + if (v >= sq) { + if (v < sq + 2*a + 1) { + break; + } + lo = a + 1; + } + else { + hi = a - 1; + } + } + + /* At this point a/1000 is an approximation to sqrt(v). */ + mpd_minalloc(z); + mpd_clear_flags(z); + z->data[0] = 1000000000UL / a; + z->len = 1; + z->exp = -6; + mpd_setdigits(z); +} + +static void +_mpd_qinvroot(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + uint32_t workstatus = 0; + mpd_context_t varcontext, maxcontext; + mpd_t *z = result; /* current approximation */ + mpd_t *v; /* a, normalized to a number between 1 and 100 */ + MPD_NEW_SHARED(vtmp, a); /* by default v will share data with a */ + MPD_NEW_STATIC(s,0,0,0,0); /* temporary variable */ + MPD_NEW_STATIC(t,0,0,0,0); /* temporary variable */ + MPD_NEW_CONST(one_half,0,-1,1,1,1,5); + MPD_NEW_CONST(three,0,0,1,1,1,3); + mpd_ssize_t klist[MPD_MAX_PREC_LOG2]; + mpd_ssize_t ideal_exp, shift; + mpd_ssize_t adj, tz; + mpd_ssize_t maxprec, fracdigits; + mpd_uint_t x, dummy; + int i, n; + + + ideal_exp = -(a->exp - (a->exp & 1)) / 2; + + v = &vtmp; + if (result == a) { + if ((v = mpd_qncopy(a)) == NULL) { + mpd_seterror(result, MPD_Malloc_error, status); + return; + } + } + + /* normalize a to 1 <= v < 100 */ + if ((v->digits+v->exp) & 1) { + fracdigits = v->digits - 1; + v->exp = -fracdigits; + n = (v->digits > 7) ? 7 : (int)v->digits; + _mpd_get_msdigits(&dummy, &x, v, n); + if (n < 7) { + x *= mpd_pow10[7-n]; + } + } + else { + fracdigits = v->digits - 2; + v->exp = -fracdigits; + n = (v->digits > 8) ? 8 : (int)v->digits; + _mpd_get_msdigits(&dummy, &x, v, n); + if (n < 8) { + x *= mpd_pow10[8-n]; + } + } + adj = (a->exp-v->exp) / 2; + + /* initial approximation */ + _invroot_init_approx(z, x); + + mpd_maxcontext(&maxcontext); + mpd_maxcontext(&varcontext); + varcontext.round = MPD_ROUND_TRUNC; + maxprec = ctx->prec + 2; + + i = invroot_schedule_prec(klist, maxprec, 3); + for (; i >= 0; i--) { + varcontext.prec = 2*klist[i]+2; + mpd_qmul(&s, z, z, &maxcontext, &workstatus); + if (v->digits > varcontext.prec) { + shift = v->digits - varcontext.prec; + mpd_qshiftr(&t, v, shift, &workstatus); + t.exp += shift; + mpd_qmul(&t, &t, &s, &varcontext, &workstatus); + } + else { + mpd_qmul(&t, v, &s, &varcontext, &workstatus); + } + mpd_qsub(&t, &three, &t, &maxcontext, &workstatus); + mpd_qmul(z, z, &t, &varcontext, &workstatus); + mpd_qmul(z, z, &one_half, &maxcontext, &workstatus); + } + + z->exp -= adj; + + tz = mpd_trail_zeros(result); + shift = ideal_exp - result->exp; + shift = (tz > shift) ? shift : tz; + if (shift > 0) { + mpd_qshiftr_inplace(result, shift); + result->exp += shift; + } + + + mpd_del(&s); + mpd_del(&t); + if (v != &vtmp) mpd_del(v); + *status |= (workstatus&MPD_Errors); + varcontext = *ctx; + varcontext.round = MPD_ROUND_HALF_EVEN; + mpd_qfinalize(result, &varcontext, status); +} + +void +mpd_qinvroot(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + + if (mpd_isspecial(a)) { + if (mpd_qcheck_nan(result, a, ctx, status)) { + return; + } + if (mpd_isnegative(a)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + /* positive infinity */ + _settriple(result, MPD_POS, 0, mpd_etiny(ctx)); + *status |= MPD_Clamped; + return; + } + if (mpd_iszero(a)) { + mpd_setspecial(result, mpd_sign(a), MPD_INF); + *status |= MPD_Division_by_zero; + return; + } + if (mpd_isnegative(a)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + + _mpd_qinvroot(result, a, ctx, status); +} + +/* + * Ensure correct rounding. Algorithm after Hull & Abrham, "Properly Rounded + * Variable Precision Square Root", ACM Transactions on Mathematical Software, + * Vol. 11, No. 3. + */ +static void +_mpd_fix_sqrt(mpd_t *result, const mpd_t *a, mpd_t *tmp, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_context_t maxctx; + MPD_NEW_CONST(u,0,0,1,1,1,5); + + mpd_maxcontext(&maxctx); + u.exp = u.digits - ctx->prec + result->exp - 1; + + _mpd_qsub(tmp, result, &u, &maxctx, status); + if (*status&MPD_Errors) goto nanresult; + + _mpd_qmul(tmp, tmp, tmp, &maxctx, status); + if (*status&MPD_Errors) goto nanresult; + + if (_mpd_cmp(tmp, a) == 1) { + u.exp += 1; + u.data[0] = 1; + _mpd_qsub(result, result, &u, &maxctx, status); + } + else { + _mpd_qadd(tmp, result, &u, &maxctx, status); + if (*status&MPD_Errors) goto nanresult; + + _mpd_qmul(tmp, tmp, tmp, &maxctx, status); + if (*status&MPD_Errors) goto nanresult; + + if (_mpd_cmp(tmp, a) == -1) { + u.exp += 1; + u.data[0] = 1; + _mpd_qadd(result, result, &u, &maxctx, status); + } + } + + return; + +nanresult: + mpd_setspecial(result, MPD_POS, MPD_NAN); +} + +void +mpd_qsqrt(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, + uint32_t *status) +{ + uint32_t workstatus = 0; + mpd_context_t varcontext; + mpd_t *z = result; /* current approximation */ + MPD_NEW_STATIC(v,0,0,0,0); /* a, normalized to a number between 1 and 10 */ + MPD_NEW_STATIC(vtmp,0,0,0,0); + MPD_NEW_STATIC(tmp,0,0,0,0); + mpd_ssize_t ideal_exp, shift; + mpd_ssize_t target_prec, fracdigits; + mpd_ssize_t a_exp, a_digits; + mpd_ssize_t adj, tz; + mpd_uint_t dummy, t; + int exact = 0; + + + varcontext = *ctx; + varcontext.round = MPD_ROUND_HALF_EVEN; + ideal_exp = (a->exp - (a->exp & 1)) / 2; + + if (mpd_isspecial(a)) { + if (mpd_qcheck_nan(result, a, ctx, status)) { + return; + } + if (mpd_isnegative(a)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + mpd_setspecial(result, MPD_POS, MPD_INF); + return; + } + if (mpd_iszero(a)) { + _settriple(result, mpd_sign(a), 0, ideal_exp); + mpd_qfinalize(result, ctx, status); + return; + } + if (mpd_isnegative(a)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + + if (!mpd_qcopy(&v, a, status)) { + mpd_seterror(result, MPD_Malloc_error, status); + goto finish; + } + + a_exp = a->exp; + a_digits = a->digits; + + /* normalize a to 1 <= v < 100 */ + if ((v.digits+v.exp) & 1) { + fracdigits = v.digits - 1; + v.exp = -fracdigits; + _mpd_get_msdigits(&dummy, &t, &v, 3); + t = t < 100 ? t*10 : t; + t = t < 100 ? t*10 : t; + } + else { + fracdigits = v.digits - 2; + v.exp = -fracdigits; + _mpd_get_msdigits(&dummy, &t, &v, 4); + t = t < 1000 ? t*10 : t; + t = t < 1000 ? t*10 : t; + t = t < 1000 ? t*10 : t; + } + adj = (a_exp-v.exp) / 2; + + + /* use excess digits */ + target_prec = (a_digits > ctx->prec) ? a_digits : ctx->prec; + target_prec += 2; + varcontext.prec = target_prec + 3; + + /* invroot is much faster for large numbers */ + _mpd_qinvroot(&tmp, &v, &varcontext, &workstatus); + + varcontext.prec = target_prec; + _mpd_qdiv(NO_IDEAL_EXP, z, &one, &tmp, &varcontext, &workstatus); + + + tz = mpd_trail_zeros(result); + if ((result->digits-tz)*2-1 <= v.digits) { + _mpd_qmul(&tmp, result, result, &varcontext, &workstatus); + if (workstatus&MPD_Errors) { + mpd_seterror(result, workstatus&MPD_Errors, status); + goto finish; + } + exact = (_mpd_cmp(&tmp, &v) == 0); + } + *status |= (workstatus&MPD_Errors); + + if (!exact && !mpd_isspecial(result) && !mpd_iszero(result)) { + _mpd_fix_sqrt(result, &v, &tmp, &varcontext, status); + if (mpd_isspecial(result)) goto finish; + *status |= (MPD_Rounded|MPD_Inexact); + } + + result->exp += adj; + if (exact) { + shift = ideal_exp - result->exp; + shift = (tz > shift) ? shift : tz; + if (shift > 0) { + mpd_qshiftr_inplace(result, shift); + result->exp += shift; + } + } + + +finish: + mpd_del(&v); + mpd_del(&vtmp); + mpd_del(&tmp); + varcontext.prec = ctx->prec; + mpd_qfinalize(result, &varcontext, status); +} + + +/******************************************************************************/ +/* Base conversions */ +/******************************************************************************/ + +/* + * Returns the space needed to represent an integer mpd_t in base 'base'. + * The result is undefined for non-integers. + * + * Max space needed: + * + * base^n >= 10^(digits+exp) + * n >= log10(10^(digits+exp))/log10(base) = (digits+exp) / log10(base) + */ +size_t +mpd_sizeinbase(mpd_t *a, uint32_t base) +{ + size_t x; + + assert(mpd_isinteger(a)); + if (mpd_iszero(a)) { + return 1; + } + + x = a->digits+a->exp; + +#ifdef CONFIG_64 + #ifdef USE_80BIT_LONG_DOUBLE + return (long double)x / log10(base) + 3; + #else + /* x > floor(((1ULL<<53)-3) * log10(2)) */ + if (x > 2711437152599294ULL) { + return SIZE_MAX; + } + return (double)x / log10(base) + 3; + #endif +#else /* CONFIG_32 */ +{ + double y = x / log10(base) + 3; + return (y > SIZE_MAX) ? SIZE_MAX : (size_t)y; +} +#endif +} + +/* + * Returns the space needed to import a base 'base' integer of length 'srclen'. + */ +static inline mpd_ssize_t +_mpd_importsize(size_t srclen, uint32_t base) +{ +#if SIZE_MAX == UINT64_MAX + #ifdef USE_80BIT_LONG_DOUBLE + long double x = (long double)srclen * (log10(base)/MPD_RDIGITS) + 3; + #else + double x; + if (srclen > (1ULL<<53)) { + return MPD_SSIZE_MAX; + } + x = (double)srclen * (log10(base)/MPD_RDIGITS) + 3; + #endif +#else + double x = srclen * (log10(base)/MPD_RDIGITS) + 3; +#endif + return (x > MPD_MAXIMPORT) ? MPD_SSIZE_MAX : (mpd_ssize_t)x; +} + + +static inline size_t +_to_base_u16(uint16_t *w, size_t wlen, mpd_uint_t wbase, + mpd_uint_t *u, mpd_ssize_t ulen) +{ + size_t n = 0; + + assert(wlen > 0 && ulen > 0); + + do { + w[n++] = (uint16_t)_mpd_shortdiv(u, u, ulen, wbase); + /* ulen will be at least 1. u[ulen-1] can only be zero if ulen == 1 */ + ulen = _mpd_real_size(u, ulen); + + } while (u[ulen-1] != 0 && n < wlen); + + /* proper termination condition */ + assert(u[ulen-1] == 0); + + return n; +} + +static inline void +_from_base_u16(mpd_uint_t *w, mpd_ssize_t wlen, + const mpd_uint_t *u, size_t ulen, uint32_t ubase) +{ + mpd_ssize_t m = 1; + mpd_uint_t carry; + + assert(wlen > 0 && ulen > 0); + + w[0] = u[--ulen]; + while (--ulen != SIZE_MAX && m < wlen) { + _mpd_shortmul(w, w, m, ubase); + m = _mpd_real_size(w, m+1); + carry = _mpd_shortadd(w, m, u[ulen]); + if (carry) w[m++] = carry; + } + + /* proper termination condition */ + assert(ulen == SIZE_MAX); +} + +/* target base wbase <= source base ubase */ +static inline size_t +_baseconv_to_smaller(uint32_t *w, size_t wlen, mpd_uint_t wbase, + mpd_uint_t *u, mpd_ssize_t ulen, mpd_uint_t ubase) +{ + size_t n = 0; + + assert(wlen > 0 && ulen > 0); + + do { + w[n++] = (uint32_t)_mpd_shortdiv_b(u, u, ulen, wbase, ubase); + /* ulen will be at least 1. u[ulen-1] can only be zero if ulen == 1 */ + ulen = _mpd_real_size(u, ulen); + + } while (u[ulen-1] != 0 && n < wlen); + + /* proper termination condition */ + assert(u[ulen-1] == 0); + + return n; +} + +/* target base wbase >= source base ubase */ +static inline void +_baseconv_to_larger(mpd_uint_t *w, mpd_ssize_t wlen, mpd_uint_t wbase, + const mpd_uint_t *u, size_t ulen, mpd_uint_t ubase) +{ + mpd_ssize_t m = 1; + mpd_uint_t carry; + + assert(wlen > 0 && ulen > 0); + + w[0] = u[--ulen]; + while (--ulen != SIZE_MAX && m < wlen) { + _mpd_shortmul_b(w, w, m, ubase, wbase); + m = _mpd_real_size(w, m+1); + carry = _mpd_shortadd_b(w, m, u[ulen], wbase); + if (carry) w[m++] = carry; + } + + /* proper termination condition */ + assert(ulen == SIZE_MAX); +} + + +/* + * Converts an integer mpd_t to a multiprecision integer with + * base <= UINT16_MAX+1. The least significant word of the result + * is rdata[0]. + */ +size_t +mpd_qexport_u16(uint16_t *rdata, size_t rlen, uint32_t rbase, + const mpd_t *src, uint32_t *status) +{ + mpd_t *tsrc; + size_t n; + + assert(rbase <= (1U<<16)); + assert(rlen <= SIZE_MAX/(sizeof *rdata)); + + if (mpd_isspecial(src) || !_mpd_isint(src)) { + *status |= MPD_Invalid_operation; + return SIZE_MAX; + } + + memset(rdata, 0, rlen * (sizeof *rdata)); + + if (mpd_iszero(src)) { + return 1; + } + + if ((tsrc = mpd_qnew()) == NULL) { + *status |= MPD_Malloc_error; + return SIZE_MAX; + } + + if (src->exp >= 0) { + if (!mpd_qshiftl(tsrc, src, src->exp, status)) { + mpd_del(tsrc); + return SIZE_MAX; + } + } + else { + if (mpd_qshiftr(tsrc, src, -src->exp, status) == MPD_UINT_MAX) { + mpd_del(tsrc); + return SIZE_MAX; + } + } + + n = _to_base_u16(rdata, rlen, rbase, tsrc->data, tsrc->len); + + mpd_del(tsrc); + return n; +} + +/* + * Converts an integer mpd_t to a multiprecision integer with + * base <= UINT32_MAX. The least significant word of the result + * is rdata[0]. + */ +size_t +mpd_qexport_u32(uint32_t *rdata, size_t rlen, uint32_t rbase, + const mpd_t *src, uint32_t *status) +{ + mpd_t *tsrc; + size_t n; + + if (mpd_isspecial(src) || !_mpd_isint(src)) { + *status |= MPD_Invalid_operation; + return SIZE_MAX; + } +#if MPD_SIZE_MAX < SIZE_MAX + if (rlen > MPD_SSIZE_MAX) { + *status |= MPD_Invalid_operation; + return SIZE_MAX; + } +#endif + + assert(rlen <= SIZE_MAX/(sizeof *rdata)); + memset(rdata, 0, rlen * (sizeof *rdata)); + + if (mpd_iszero(src)) { + return 1; + } + + if ((tsrc = mpd_qnew()) == NULL) { + *status |= MPD_Malloc_error; + return SIZE_MAX; + } + + if (src->exp >= 0) { + if (!mpd_qshiftl(tsrc, src, src->exp, status)) { + mpd_del(tsrc); + return SIZE_MAX; + } + } + else { + if (mpd_qshiftr(tsrc, src, -src->exp, status) == MPD_UINT_MAX) { + mpd_del(tsrc); + return SIZE_MAX; + } + } + +#ifdef CONFIG_64 + n = _baseconv_to_smaller(rdata, rlen, rbase, + tsrc->data, tsrc->len, MPD_RADIX); +#else + if (rbase <= MPD_RADIX) { + n = _baseconv_to_smaller(rdata, rlen, rbase, + tsrc->data, tsrc->len, MPD_RADIX); + } + else { + _baseconv_to_larger(rdata, (mpd_ssize_t)rlen, rbase, + tsrc->data, tsrc->len, MPD_RADIX); + n = _mpd_real_size(rdata, (mpd_ssize_t)rlen); + } +#endif + + mpd_del(tsrc); + return n; +} + + +/* + * Converts a multiprecision integer with base <= UINT16_MAX+1 to an mpd_t. + * The least significant word of the source is srcdata[0]. + */ +void +mpd_qimport_u16(mpd_t *result, + const uint16_t *srcdata, size_t srclen, + uint8_t srcsign, uint32_t srcbase, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_uint_t *usrc; /* uint16_t src copied to an mpd_uint_t array */ + mpd_ssize_t rlen; /* length of the result */ + size_t n = 0; + + assert(srclen > 0); + assert(srcbase <= (1U<<16)); + + if ((rlen = _mpd_importsize(srclen, srcbase)) == MPD_SSIZE_MAX) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if (srclen > MPD_SIZE_MAX/(sizeof *usrc)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if ((usrc = mpd_alloc((mpd_size_t)srclen, sizeof *usrc)) == NULL) { + mpd_seterror(result, MPD_Malloc_error, status); + return; + } + for (n = 0; n < srclen; n++) { + usrc[n] = srcdata[n]; + } + + /* result->data is initialized to zero */ + if (!mpd_qresize_zero(result, rlen, status)) { + goto finish; + } + + _from_base_u16(result->data, rlen, usrc, srclen, srcbase); + + mpd_set_flags(result, srcsign); + result->exp = 0; + result->len = _mpd_real_size(result->data, rlen); + mpd_setdigits(result); + + mpd_qresize(result, result->len, status); + mpd_qfinalize(result, ctx, status); + + +finish: + mpd_free(usrc); +} + +/* + * Converts a multiprecision integer with base <= UINT32_MAX to an mpd_t. + * The least significant word of the source is srcdata[0]. + */ +void +mpd_qimport_u32(mpd_t *result, + const uint32_t *srcdata, size_t srclen, + uint8_t srcsign, uint32_t srcbase, + const mpd_context_t *ctx, uint32_t *status) +{ + mpd_uint_t *usrc; /* uint32_t src copied to an mpd_uint_t array */ + mpd_ssize_t rlen; /* length of the result */ + size_t n = 0; + + assert(srclen > 0); + + if ((rlen = _mpd_importsize(srclen, srcbase)) == MPD_SSIZE_MAX) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if (srclen > MPD_SIZE_MAX/(sizeof *usrc)) { + mpd_seterror(result, MPD_Invalid_operation, status); + return; + } + if ((usrc = mpd_alloc((mpd_size_t)srclen, sizeof *usrc)) == NULL) { + mpd_seterror(result, MPD_Malloc_error, status); + return; + } + for (n = 0; n < srclen; n++) { + usrc[n] = srcdata[n]; + } + + /* result->data is initialized to zero */ + if (!mpd_qresize_zero(result, rlen, status)) { + goto finish; + } + +#ifdef CONFIG_64 + _baseconv_to_larger(result->data, rlen, MPD_RADIX, + usrc, srclen, srcbase); +#else + if (srcbase <= MPD_RADIX) { + _baseconv_to_larger(result->data, rlen, MPD_RADIX, + usrc, srclen, srcbase); + } + else { + _baseconv_to_smaller(result->data, rlen, MPD_RADIX, + usrc, (mpd_ssize_t)srclen, srcbase); + } +#endif + + mpd_set_flags(result, srcsign); + result->exp = 0; + result->len = _mpd_real_size(result->data, rlen); + mpd_setdigits(result); + + mpd_qresize(result, result->len, status); + mpd_qfinalize(result, ctx, status); + + +finish: + mpd_free(usrc); +} + + + diff --git a/Modules/_decimal/libmpdec/mpdecimal.h b/Modules/_decimal/libmpdec/mpdecimal.h new file mode 100644 index 0000000000..f3f32ac601 --- /dev/null +++ b/Modules/_decimal/libmpdec/mpdecimal.h @@ -0,0 +1,800 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#ifndef MPDECIMAL_H +#define MPDECIMAL_H + + +#ifdef __cplusplus +extern "C" { +#define __STDC_LIMIT_MACROS +#endif + + +#ifndef _MSC_VER + #include "pyconfig.h" +#endif + +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <limits.h> +#include <assert.h> + +#ifdef _MSC_VER + #include "vccompat.h" + #ifndef UNUSED + #define UNUSED + #endif + #define EXTINLINE extern inline +#else + #ifdef HAVE_STDINT_H + #include <stdint.h> + #endif + #ifdef HAVE_INTTYPES_H + #include <inttypes.h> + #endif + #ifndef __GNUC_STDC_INLINE__ + #define __GNUC_STDC_INLINE__ + #endif + #if defined(__GNUC__) && !defined(__INTEL_COMPILER) + #define UNUSED __attribute__((unused)) + #else + #define UNUSED + #endif + #define EXTINLINE +#endif + + +#if !defined(LEGACY_COMPILER) + #if !defined(UINT64_MAX) + /* The following #error is just a warning. If the compiler indeed does + * not have uint64_t, it is perfectly safe to comment out the #error. */ + #error "Warning: Compiler without uint64_t. Comment out this line." + #define LEGACY_COMPILER + #endif +#endif + + +/******************************************************************************/ +/* Configuration */ +/******************************************************************************/ + +#if defined(UNIVERSAL) + #if defined(CONFIG_64) || defined(CONFIG_32) + #error "cannot use CONFIG_64 or CONFIG_32 with UNIVERSAL." + #endif + #if defined(__ppc__) + #define CONFIG_32 + #define ANSI + #elif defined(__ppc64__) + #define CONFIG_64 + #define ANSI + #elif defined(__i386__) + #define CONFIG_32 + #define ANSI + #elif defined(__x86_64__) + #define CONFIG_64 + #define ASM + #else + #error "unknown architecture for universal build." + #endif +#endif + + +/* BEGIN CONFIG_64 */ +#if defined(CONFIG_64) +/* types for modular and base arithmetic */ +#define MPD_UINT_MAX UINT64_MAX +#define MPD_BITS_PER_UINT 64 +typedef uint64_t mpd_uint_t; /* unsigned mod type */ + +#define MPD_SIZE_MAX SIZE_MAX +typedef size_t mpd_size_t; /* unsigned size type */ + +/* type for exp, digits, len, prec */ +#define MPD_SSIZE_MAX INT64_MAX +#define MPD_SSIZE_MIN INT64_MIN +typedef int64_t mpd_ssize_t; +#define _mpd_strtossize strtoll + +/* decimal arithmetic */ +#define MPD_RADIX 10000000000000000000ULL /* 10**19 */ +#define MPD_RDIGITS 19 +#define MPD_MAX_POW10 19 +#define MPD_EXPDIGITS 19 /* MPD_EXPDIGITS <= MPD_RDIGITS+1 */ + +#define MPD_MAXTRANSFORM_2N 4294967296ULL /* 2**32 */ +#define MPD_MAX_PREC 999999999999999999LL +#define MPD_MAX_PREC_LOG2 64 +#define MPD_ELIMIT 1000000000000000000LL +#define MPD_MAX_EMAX 999999999999999999LL /* ELIMIT-1 */ +#define MPD_MIN_EMIN (-999999999999999999LL) /* -EMAX */ +#define MPD_MIN_ETINY (MPD_MIN_EMIN-(MPD_MAX_PREC-1)) +#define MPD_EXP_INF 2000000000000000001LL +#define MPD_EXP_CLAMP (-4000000000000000001LL) +#define MPD_MAXIMPORT 105263157894736842L /* ceil((2*MPD_MAX_PREC)/MPD_RDIGITS) */ + +/* conversion specifiers */ +#define PRI_mpd_uint_t PRIu64 +#define PRI_mpd_ssize_t PRIi64 +/* END CONFIG_64 */ + + +/* BEGIN CONFIG_32 */ +#elif defined(CONFIG_32) +/* types for modular and base arithmetic */ +#define MPD_UINT_MAX UINT32_MAX +#define MPD_BITS_PER_UINT 32 +typedef uint32_t mpd_uint_t; /* unsigned mod type */ + +#ifndef LEGACY_COMPILER +#define MPD_UUINT_MAX UINT64_MAX +typedef uint64_t mpd_uuint_t; /* double width unsigned mod type */ +#endif + +#define MPD_SIZE_MAX SIZE_MAX +typedef size_t mpd_size_t; /* unsigned size type */ + +/* type for dec->len, dec->exp, ctx->prec */ +#define MPD_SSIZE_MAX INT32_MAX +#define MPD_SSIZE_MIN INT32_MIN +typedef int32_t mpd_ssize_t; +#define _mpd_strtossize strtol + +/* decimal arithmetic */ +#define MPD_RADIX 1000000000UL /* 10**9 */ +#define MPD_RDIGITS 9 +#define MPD_MAX_POW10 9 +#define MPD_EXPDIGITS 10 /* MPD_EXPDIGITS <= MPD_RDIGITS+1 */ + +#define MPD_MAXTRANSFORM_2N 33554432UL /* 2**25 */ +#define MPD_MAX_PREC 425000000L +#define MPD_MAX_PREC_LOG2 32 +#define MPD_ELIMIT 425000001L +#define MPD_MAX_EMAX 425000000L /* ELIMIT-1 */ +#define MPD_MIN_EMIN (-425000000L) /* -EMAX */ +#define MPD_MIN_ETINY (MPD_MIN_EMIN-(MPD_MAX_PREC-1)) +#define MPD_EXP_INF 1000000001L /* allows for emax=999999999 in the tests */ +#define MPD_EXP_CLAMP (-2000000001L) /* allows for emin=-999999999 in the tests */ +#define MPD_MAXIMPORT 94444445L /* ceil((2*MPD_MAX_PREC)/MPD_RDIGITS) */ + +/* conversion specifiers */ +#define PRI_mpd_uint_t PRIu32 +#define PRI_mpd_ssize_t PRIi32 +/* END CONFIG_32 */ + +#else + #error "define CONFIG_64 or CONFIG_32" +#endif +/* END CONFIG */ + + +#if MPD_SIZE_MAX != MPD_UINT_MAX + #error "unsupported platform: need mpd_size_t == mpd_uint_t" +#endif + + +/******************************************************************************/ +/* Context */ +/******************************************************************************/ + +enum { + MPD_ROUND_UP, /* round away from 0 */ + MPD_ROUND_DOWN, /* round toward 0 (truncate) */ + MPD_ROUND_CEILING, /* round toward +infinity */ + MPD_ROUND_FLOOR, /* round toward -infinity */ + MPD_ROUND_HALF_UP, /* 0.5 is rounded up */ + MPD_ROUND_HALF_DOWN, /* 0.5 is rounded down */ + MPD_ROUND_HALF_EVEN, /* 0.5 is rounded to even */ + MPD_ROUND_05UP, /* round zero or five away from 0 */ + MPD_ROUND_TRUNC, /* truncate, but set infinity */ + MPD_ROUND_GUARD +}; + +enum { MPD_CLAMP_DEFAULT, MPD_CLAMP_IEEE_754, MPD_CLAMP_GUARD }; + +extern const char *mpd_round_string[MPD_ROUND_GUARD]; +extern const char *mpd_clamp_string[MPD_CLAMP_GUARD]; + + +typedef struct { + mpd_ssize_t prec; /* precision */ + mpd_ssize_t emax; /* max positive exp */ + mpd_ssize_t emin; /* min negative exp */ + uint32_t traps; /* status events that should be trapped */ + uint32_t status; /* status flags */ + uint32_t newtrap; /* set by mpd_addstatus_raise() */ + int round; /* rounding mode */ + int clamp; /* clamp mode */ + int allcr; /* all functions correctly rounded */ +} mpd_context_t; + + +/* Status flags */ +#define MPD_Clamped 0x00000001U +#define MPD_Conversion_syntax 0x00000002U +#define MPD_Division_by_zero 0x00000004U +#define MPD_Division_impossible 0x00000008U +#define MPD_Division_undefined 0x00000010U +#define MPD_Fpu_error 0x00000020U +#define MPD_Inexact 0x00000040U +#define MPD_Invalid_context 0x00000080U +#define MPD_Invalid_operation 0x00000100U +#define MPD_Malloc_error 0x00000200U +#define MPD_Not_implemented 0x00000400U +#define MPD_Overflow 0x00000800U +#define MPD_Rounded 0x00001000U +#define MPD_Subnormal 0x00002000U +#define MPD_Underflow 0x00004000U +#define MPD_Max_status (0x00008000U-1U) + +/* Conditions that result in an IEEE 754 exception */ +#define MPD_IEEE_Invalid_operation (MPD_Conversion_syntax | \ + MPD_Division_impossible | \ + MPD_Division_undefined | \ + MPD_Fpu_error | \ + MPD_Invalid_context | \ + MPD_Invalid_operation | \ + MPD_Malloc_error) \ + +/* Errors that require the result of an operation to be set to NaN */ +#define MPD_Errors (MPD_IEEE_Invalid_operation | \ + MPD_Division_by_zero) + +/* Default traps */ +#define MPD_Traps (MPD_IEEE_Invalid_operation | \ + MPD_Division_by_zero | \ + MPD_Overflow | \ + MPD_Underflow) + +/* Official name */ +#define MPD_Insufficient_storage MPD_Malloc_error + +/* IEEE 754 interchange format contexts */ +#define MPD_IEEE_CONTEXT_MAX_BITS 512 /* 16*(log2(MPD_MAX_EMAX / 3)-3) */ +#define MPD_DECIMAL32 32 +#define MPD_DECIMAL64 64 +#define MPD_DECIMAL128 128 + + +#define MPD_MINALLOC_MIN 2 +#define MPD_MINALLOC_MAX 64 +extern mpd_ssize_t MPD_MINALLOC; +extern void (* mpd_traphandler)(mpd_context_t *); +void mpd_dflt_traphandler(mpd_context_t *); + +void mpd_setminalloc(mpd_ssize_t n); +void mpd_init(mpd_context_t *ctx, mpd_ssize_t prec); + +void mpd_maxcontext(mpd_context_t *ctx); +void mpd_defaultcontext(mpd_context_t *ctx); +void mpd_basiccontext(mpd_context_t *ctx); +int mpd_ieee_context(mpd_context_t *ctx, int bits); + +mpd_ssize_t mpd_getprec(const mpd_context_t *ctx); +mpd_ssize_t mpd_getemax(const mpd_context_t *ctx); +mpd_ssize_t mpd_getemin(const mpd_context_t *ctx); +int mpd_getround(const mpd_context_t *ctx); +uint32_t mpd_gettraps(const mpd_context_t *ctx); +uint32_t mpd_getstatus(const mpd_context_t *ctx); +int mpd_getclamp(const mpd_context_t *ctx); +int mpd_getcr(const mpd_context_t *ctx); + +int mpd_qsetprec(mpd_context_t *ctx, mpd_ssize_t prec); +int mpd_qsetemax(mpd_context_t *ctx, mpd_ssize_t emax); +int mpd_qsetemin(mpd_context_t *ctx, mpd_ssize_t emin); +int mpd_qsetround(mpd_context_t *ctx, int newround); +int mpd_qsettraps(mpd_context_t *ctx, uint32_t flags); +int mpd_qsetstatus(mpd_context_t *ctx, uint32_t flags); +int mpd_qsetclamp(mpd_context_t *ctx, int c); +int mpd_qsetcr(mpd_context_t *ctx, int c); +void mpd_addstatus_raise(mpd_context_t *ctx, uint32_t flags); + + +/******************************************************************************/ +/* Decimal Arithmetic */ +/******************************************************************************/ + +/* mpd_t flags */ +#define MPD_POS ((uint8_t)0) +#define MPD_NEG ((uint8_t)1) +#define MPD_INF ((uint8_t)2) +#define MPD_NAN ((uint8_t)4) +#define MPD_SNAN ((uint8_t)8) +#define MPD_SPECIAL (MPD_INF|MPD_NAN|MPD_SNAN) +#define MPD_STATIC ((uint8_t)16) +#define MPD_STATIC_DATA ((uint8_t)32) +#define MPD_SHARED_DATA ((uint8_t)64) +#define MPD_CONST_DATA ((uint8_t)128) +#define MPD_DATAFLAGS (MPD_STATIC_DATA|MPD_SHARED_DATA|MPD_CONST_DATA) + +/* mpd_t */ +typedef struct { + uint8_t flags; + mpd_ssize_t exp; + mpd_ssize_t digits; + mpd_ssize_t len; + mpd_ssize_t alloc; + mpd_uint_t *data; +} mpd_t; + + +typedef unsigned char uchar; + + +/******************************************************************************/ +/* Quiet, thread-safe functions */ +/******************************************************************************/ + +/* format specification */ +typedef struct { + mpd_ssize_t min_width; /* minimum field width */ + mpd_ssize_t prec; /* fraction digits or significant digits */ + char type; /* conversion specifier */ + char align; /* alignment */ + char sign; /* sign printing/alignment */ + char fill[5]; /* fill character */ + const char *dot; /* decimal point */ + const char *sep; /* thousands separator */ + const char *grouping; /* grouping of digits */ +} mpd_spec_t; + +/* output to a string */ +char *mpd_to_sci(const mpd_t *dec, int fmt); +char *mpd_to_eng(const mpd_t *dec, int fmt); +mpd_ssize_t mpd_to_sci_size(char **res, const mpd_t *dec, int fmt); +mpd_ssize_t mpd_to_eng_size(char **res, const mpd_t *dec, int fmt); +int mpd_validate_lconv(mpd_spec_t *spec); +int mpd_parse_fmt_str(mpd_spec_t *spec, const char *fmt, int caps); +char * mpd_qformat_spec(const mpd_t *dec, const mpd_spec_t *spec, const mpd_context_t *ctx, uint32_t *status); +char *mpd_qformat(const mpd_t *dec, const char *fmt, const mpd_context_t *ctx, uint32_t *status); + +#define MPD_NUM_FLAGS 15 +#define MPD_MAX_FLAG_STRING 208 +#define MPD_MAX_FLAG_LIST (MPD_MAX_FLAG_STRING+18) +#define MPD_MAX_SIGNAL_LIST 121 +int mpd_snprint_flags(char *dest, int nmemb, uint32_t flags); +int mpd_lsnprint_flags(char *dest, int nmemb, uint32_t flags, const char *flag_string[]); +int mpd_lsnprint_signals(char *dest, int nmemb, uint32_t flags, const char *signal_string[]); + +/* output to a file */ +void mpd_fprint(FILE *file, const mpd_t *dec); +void mpd_print(const mpd_t *dec); + +/* assignment from a string */ +void mpd_qset_string(mpd_t *dec, const char *s, const mpd_context_t *ctx, uint32_t *status); + +/* set to NaN with error flags */ +void mpd_seterror(mpd_t *result, uint32_t flags, uint32_t *status); +/* set a special with sign and type */ +void mpd_setspecial(mpd_t *dec, uint8_t sign, uint8_t type); +/* set coefficient to zero or all nines */ +void mpd_zerocoeff(mpd_t *result); +void mpd_qmaxcoeff(mpd_t *result, const mpd_context_t *ctx, uint32_t *status); + +/* quietly assign a C integer type to an mpd_t */ +void mpd_qset_ssize(mpd_t *result, mpd_ssize_t a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qset_i32(mpd_t *result, int32_t a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qset_uint(mpd_t *result, mpd_uint_t a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qset_u32(mpd_t *result, uint32_t a, const mpd_context_t *ctx, uint32_t *status); +#ifndef LEGACY_COMPILER +void mpd_qset_i64(mpd_t *result, int64_t a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qset_u64(mpd_t *result, uint64_t a, const mpd_context_t *ctx, uint32_t *status); +#endif + +/* quietly assign a C integer type to an mpd_t with a static coefficient */ +void mpd_qsset_ssize(mpd_t *result, mpd_ssize_t a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qsset_i32(mpd_t *result, int32_t a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qsset_uint(mpd_t *result, mpd_uint_t a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qsset_u32(mpd_t *result, uint32_t a, const mpd_context_t *ctx, uint32_t *status); + +/* quietly get a C integer type from an mpd_t */ +mpd_ssize_t mpd_qget_ssize(const mpd_t *dec, uint32_t *status); +mpd_uint_t mpd_qget_uint(const mpd_t *dec, uint32_t *status); +mpd_uint_t mpd_qabs_uint(const mpd_t *dec, uint32_t *status); + + +/* quiet functions */ +int mpd_qcheck_nan(mpd_t *nanresult, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); +int mpd_qcheck_nans(mpd_t *nanresult, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qfinalize(mpd_t *result, const mpd_context_t *ctx, uint32_t *status); + +const char * mpd_class(const mpd_t *a, const mpd_context_t *ctx); + +int mpd_qcopy(mpd_t *result, const mpd_t *a, uint32_t *status); +mpd_t *mpd_qncopy(const mpd_t *a); +int mpd_qcopy_abs(mpd_t *result, const mpd_t *a, uint32_t *status); +int mpd_qcopy_negate(mpd_t *result, const mpd_t *a, uint32_t *status); +int mpd_qcopy_sign(mpd_t *result, const mpd_t *a, const mpd_t *b, uint32_t *status); + +void mpd_qand(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qinvert(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qlogb(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qor(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qscaleb(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qxor(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +int mpd_same_quantum(const mpd_t *a, const mpd_t *b); + +void mpd_qrotate(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +int mpd_qshiftl(mpd_t *result, const mpd_t *a, mpd_ssize_t n, uint32_t *status); +mpd_uint_t mpd_qshiftr(mpd_t *result, const mpd_t *a, mpd_ssize_t n, uint32_t *status); +mpd_uint_t mpd_qshiftr_inplace(mpd_t *result, mpd_ssize_t n); +void mpd_qshift(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qshiftn(mpd_t *result, const mpd_t *a, mpd_ssize_t n, const mpd_context_t *ctx, uint32_t *status); + +int mpd_qcmp(const mpd_t *a, const mpd_t *b, uint32_t *status); +int mpd_qcompare(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +int mpd_qcompare_signal(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +int mpd_cmp_total(const mpd_t *a, const mpd_t *b); +int mpd_cmp_total_mag(const mpd_t *a, const mpd_t *b); +int mpd_compare_total(mpd_t *result, const mpd_t *a, const mpd_t *b); +int mpd_compare_total_mag(mpd_t *result, const mpd_t *a, const mpd_t *b); + +void mpd_qround_to_intx(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qround_to_int(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qtrunc(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qfloor(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qceil(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); + +void mpd_qabs(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qmax(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qmax_mag(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qmin(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qmin_mag(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qminus(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qplus(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qnext_minus(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qnext_plus(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qnext_toward(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qquantize(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qrescale(mpd_t *result, const mpd_t *a, mpd_ssize_t exp, const mpd_context_t *ctx, uint32_t *status); +void mpd_qrescale_fmt(mpd_t *result, const mpd_t *a, mpd_ssize_t exp, const mpd_context_t *ctx, uint32_t *status); +void mpd_qreduce(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qadd(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qadd_ssize(mpd_t *result, const mpd_t *a, mpd_ssize_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qadd_i32(mpd_t *result, const mpd_t *a, int32_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qadd_uint(mpd_t *result, const mpd_t *a, mpd_uint_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qadd_u32(mpd_t *result, const mpd_t *a, uint32_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qsub(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qsub_ssize(mpd_t *result, const mpd_t *a, mpd_ssize_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qsub_i32(mpd_t *result, const mpd_t *a, int32_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qsub_uint(mpd_t *result, const mpd_t *a, mpd_uint_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qsub_u32(mpd_t *result, const mpd_t *a, uint32_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qmul(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qmul_ssize(mpd_t *result, const mpd_t *a, mpd_ssize_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qmul_i32(mpd_t *result, const mpd_t *a, int32_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qmul_uint(mpd_t *result, const mpd_t *a, mpd_uint_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qmul_u32(mpd_t *result, const mpd_t *a, uint32_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qfma(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_t *c, const mpd_context_t *ctx, uint32_t *status); +void mpd_qdiv(mpd_t *q, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qdiv_ssize(mpd_t *result, const mpd_t *a, mpd_ssize_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qdiv_i32(mpd_t *result, const mpd_t *a, int32_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qdiv_uint(mpd_t *result, const mpd_t *a, mpd_uint_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qdiv_u32(mpd_t *result, const mpd_t *a, uint32_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qdivint(mpd_t *q, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qrem(mpd_t *r, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qrem_near(mpd_t *r, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qdivmod(mpd_t *q, mpd_t *r, const mpd_t *a, const mpd_t *b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qpow(mpd_t *result, const mpd_t *base, const mpd_t *exp, const mpd_context_t *ctx, uint32_t *status); +void mpd_qpowmod(mpd_t *result, const mpd_t *base, const mpd_t *exp, const mpd_t *mod, const mpd_context_t *ctx, uint32_t *status); +void mpd_qexp(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qln10(mpd_t *result, mpd_ssize_t prec, uint32_t *status); +void mpd_qln(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qlog10(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qsqrt(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qinvroot(mpd_t *result, const mpd_t *a, const mpd_context_t *ctx, uint32_t *status); + + +size_t mpd_sizeinbase(mpd_t *a, uint32_t base); +void mpd_qimport_u16(mpd_t *result, const uint16_t *srcdata, size_t srclen, + uint8_t srcsign, uint32_t srcbase, + const mpd_context_t *ctx, uint32_t *status); +void mpd_qimport_u32(mpd_t *result, const uint32_t *srcdata, size_t srclen, + uint8_t srcsign, uint32_t srcbase, + const mpd_context_t *ctx, uint32_t *status); +size_t mpd_qexport_u16(uint16_t *rdata, size_t rlen, uint32_t base, + const mpd_t *src, uint32_t *status); +size_t mpd_qexport_u32(uint32_t *rdata, size_t rlen, uint32_t base, + const mpd_t *src, uint32_t *status); + + +/******************************************************************************/ +/* Signalling functions */ +/******************************************************************************/ + +char * mpd_format(const mpd_t *dec, const char *fmt, mpd_context_t *ctx); +void mpd_import_u16(mpd_t *result, const uint16_t *srcdata, size_t srclen, uint8_t srcsign, uint32_t base, mpd_context_t *ctx); +void mpd_import_u32(mpd_t *result, const uint32_t *srcdata, size_t srclen, uint8_t srcsign, uint32_t base, mpd_context_t *ctx); +size_t mpd_export_u16(uint16_t *rdata, size_t rlen, uint32_t base, const mpd_t *src, mpd_context_t *ctx); +size_t mpd_export_u32(uint32_t *rdata, size_t rlen, uint32_t base, const mpd_t *src, mpd_context_t *ctx); +void mpd_finalize(mpd_t *result, mpd_context_t *ctx); +int mpd_check_nan(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +int mpd_check_nans(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_set_string(mpd_t *result, const char *s, mpd_context_t *ctx); +void mpd_maxcoeff(mpd_t *result, mpd_context_t *ctx); +void mpd_sset_ssize(mpd_t *result, mpd_ssize_t a, mpd_context_t *ctx); +void mpd_sset_i32(mpd_t *result, int32_t a, mpd_context_t *ctx); +void mpd_sset_uint(mpd_t *result, mpd_uint_t a, mpd_context_t *ctx); +void mpd_sset_u32(mpd_t *result, uint32_t a, mpd_context_t *ctx); +void mpd_set_ssize(mpd_t *result, mpd_ssize_t a, mpd_context_t *ctx); +void mpd_set_i32(mpd_t *result, int32_t a, mpd_context_t *ctx); +void mpd_set_uint(mpd_t *result, mpd_uint_t a, mpd_context_t *ctx); +void mpd_set_u32(mpd_t *result, uint32_t a, mpd_context_t *ctx); +#ifndef LEGACY_COMPILER +void mpd_set_i64(mpd_t *result, int64_t a, mpd_context_t *ctx); +void mpd_set_u64(mpd_t *result, uint64_t a, mpd_context_t *ctx); +#endif +mpd_ssize_t mpd_get_ssize(const mpd_t *a, mpd_context_t *ctx); +mpd_uint_t mpd_get_uint(const mpd_t *a, mpd_context_t *ctx); +mpd_uint_t mpd_abs_uint(const mpd_t *a, mpd_context_t *ctx); +void mpd_and(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_copy(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_canonical(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_copy_abs(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_copy_negate(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_copy_sign(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_invert(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_logb(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_or(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_rotate(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_scaleb(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_shiftl(mpd_t *result, const mpd_t *a, mpd_ssize_t n, mpd_context_t *ctx); +mpd_uint_t mpd_shiftr(mpd_t *result, const mpd_t *a, mpd_ssize_t n, mpd_context_t *ctx); +void mpd_shiftn(mpd_t *result, const mpd_t *a, mpd_ssize_t n, mpd_context_t *ctx); +void mpd_shift(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_xor(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_abs(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +int mpd_cmp(const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +int mpd_compare(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +int mpd_compare_signal(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_add(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_add_ssize(mpd_t *result, const mpd_t *a, mpd_ssize_t b, mpd_context_t *ctx); +void mpd_add_i32(mpd_t *result, const mpd_t *a, int32_t b, mpd_context_t *ctx); +void mpd_add_uint(mpd_t *result, const mpd_t *a, mpd_uint_t b, mpd_context_t *ctx); +void mpd_add_u32(mpd_t *result, const mpd_t *a, uint32_t b, mpd_context_t *ctx); +void mpd_sub(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_sub_ssize(mpd_t *result, const mpd_t *a, mpd_ssize_t b, mpd_context_t *ctx); +void mpd_sub_i32(mpd_t *result, const mpd_t *a, int32_t b, mpd_context_t *ctx); +void mpd_sub_uint(mpd_t *result, const mpd_t *a, mpd_uint_t b, mpd_context_t *ctx); +void mpd_sub_u32(mpd_t *result, const mpd_t *a, uint32_t b, mpd_context_t *ctx); +void mpd_div(mpd_t *q, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_div_ssize(mpd_t *result, const mpd_t *a, mpd_ssize_t b, mpd_context_t *ctx); +void mpd_div_i32(mpd_t *result, const mpd_t *a, int32_t b, mpd_context_t *ctx); +void mpd_div_uint(mpd_t *result, const mpd_t *a, mpd_uint_t b, mpd_context_t *ctx); +void mpd_div_u32(mpd_t *result, const mpd_t *a, uint32_t b, mpd_context_t *ctx); +void mpd_divmod(mpd_t *q, mpd_t *r, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_divint(mpd_t *q, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_exp(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_fma(mpd_t *result, const mpd_t *a, const mpd_t *b, const mpd_t *c, mpd_context_t *ctx); +void mpd_ln(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_log10(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_max(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_max_mag(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_min(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_min_mag(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_minus(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_mul(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_mul_ssize(mpd_t *result, const mpd_t *a, mpd_ssize_t b, mpd_context_t *ctx); +void mpd_mul_i32(mpd_t *result, const mpd_t *a, int32_t b, mpd_context_t *ctx); +void mpd_mul_uint(mpd_t *result, const mpd_t *a, mpd_uint_t b, mpd_context_t *ctx); +void mpd_mul_u32(mpd_t *result, const mpd_t *a, uint32_t b, mpd_context_t *ctx); +void mpd_next_minus(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_next_plus(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_next_toward(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_plus(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_pow(mpd_t *result, const mpd_t *base, const mpd_t *exp, mpd_context_t *ctx); +void mpd_powmod(mpd_t *result, const mpd_t *base, const mpd_t *exp, const mpd_t *mod, mpd_context_t *ctx); +void mpd_quantize(mpd_t *result, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_rescale(mpd_t *result, const mpd_t *a, mpd_ssize_t exp, mpd_context_t *ctx); +void mpd_reduce(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_rem(mpd_t *r, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_rem_near(mpd_t *r, const mpd_t *a, const mpd_t *b, mpd_context_t *ctx); +void mpd_round_to_intx(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_round_to_int(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_trunc(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_floor(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_ceil(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_sqrt(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); +void mpd_invroot(mpd_t *result, const mpd_t *a, mpd_context_t *ctx); + + +/******************************************************************************/ +/* Configuration specific */ +/******************************************************************************/ + +#ifdef CONFIG_64 +void mpd_qsset_i64(mpd_t *result, int64_t a, const mpd_context_t *ctx, uint32_t *status); +void mpd_qsset_u64(mpd_t *result, uint64_t a, const mpd_context_t *ctx, uint32_t *status); +int64_t mpd_qget_i64(const mpd_t *dec, uint32_t *status); +uint64_t mpd_qget_u64(const mpd_t *dec, uint32_t *status); + +void mpd_qadd_i64(mpd_t *result, const mpd_t *a, int64_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qadd_u64(mpd_t *result, const mpd_t *a, uint64_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qsub_i64(mpd_t *result, const mpd_t *a, int64_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qsub_u64(mpd_t *result, const mpd_t *a, uint64_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qmul_i64(mpd_t *result, const mpd_t *a, int64_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qmul_u64(mpd_t *result, const mpd_t *a, uint64_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qdiv_i64(mpd_t *result, const mpd_t *a, int64_t b, const mpd_context_t *ctx, uint32_t *status); +void mpd_qdiv_u64(mpd_t *result, const mpd_t *a, uint64_t b, const mpd_context_t *ctx, uint32_t *status); + +void mpd_sset_i64(mpd_t *result, int64_t a, mpd_context_t *ctx); +void mpd_sset_u64(mpd_t *result, uint64_t a, mpd_context_t *ctx); +int64_t mpd_get_i64(const mpd_t *a, mpd_context_t *ctx); +uint64_t mpd_get_u64(const mpd_t *a, mpd_context_t *ctx); + +void mpd_add_i64(mpd_t *result, const mpd_t *a, int64_t b, mpd_context_t *ctx); +void mpd_add_u64(mpd_t *result, const mpd_t *a, uint64_t b, mpd_context_t *ctx); +void mpd_sub_i64(mpd_t *result, const mpd_t *a, int64_t b, mpd_context_t *ctx); +void mpd_sub_u64(mpd_t *result, const mpd_t *a, uint64_t b, mpd_context_t *ctx); +void mpd_div_i64(mpd_t *result, const mpd_t *a, int64_t b, mpd_context_t *ctx); +void mpd_div_u64(mpd_t *result, const mpd_t *a, uint64_t b, mpd_context_t *ctx); +void mpd_mul_i64(mpd_t *result, const mpd_t *a, int64_t b, mpd_context_t *ctx); +void mpd_mul_u64(mpd_t *result, const mpd_t *a, uint64_t b, mpd_context_t *ctx); +#else +int32_t mpd_qget_i32(const mpd_t *dec, uint32_t *status); +uint32_t mpd_qget_u32(const mpd_t *dec, uint32_t *status); +int32_t mpd_get_i32(const mpd_t *a, mpd_context_t *ctx); +uint32_t mpd_get_u32(const mpd_t *a, mpd_context_t *ctx); +#endif + + +/******************************************************************************/ +/* Get attributes of a decimal */ +/******************************************************************************/ + +EXTINLINE mpd_ssize_t mpd_adjexp(const mpd_t *dec); +EXTINLINE mpd_ssize_t mpd_etiny(const mpd_context_t *ctx); +EXTINLINE mpd_ssize_t mpd_etop(const mpd_context_t *ctx); +EXTINLINE mpd_uint_t mpd_msword(const mpd_t *dec); +EXTINLINE int mpd_word_digits(mpd_uint_t word); +/* most significant digit of a word */ +EXTINLINE mpd_uint_t mpd_msd(mpd_uint_t word); +/* least significant digit of a word */ +EXTINLINE mpd_uint_t mpd_lsd(mpd_uint_t word); +/* coefficient size needed to store 'digits' */ +EXTINLINE mpd_ssize_t mpd_digits_to_size(mpd_ssize_t digits); +/* number of digits in the exponent, undefined for MPD_SSIZE_MIN */ +EXTINLINE int mpd_exp_digits(mpd_ssize_t exp); +EXTINLINE int mpd_iscanonical(const mpd_t *dec UNUSED); +EXTINLINE int mpd_isfinite(const mpd_t *dec); +EXTINLINE int mpd_isinfinite(const mpd_t *dec); +EXTINLINE int mpd_isinteger(const mpd_t *dec); +EXTINLINE int mpd_isnan(const mpd_t *dec); +EXTINLINE int mpd_isnegative(const mpd_t *dec); +EXTINLINE int mpd_ispositive(const mpd_t *dec); +EXTINLINE int mpd_isqnan(const mpd_t *dec); +EXTINLINE int mpd_issigned(const mpd_t *dec); +EXTINLINE int mpd_issnan(const mpd_t *dec); +EXTINLINE int mpd_isspecial(const mpd_t *dec); +EXTINLINE int mpd_iszero(const mpd_t *dec); +/* undefined for special numbers */ +EXTINLINE int mpd_iszerocoeff(const mpd_t *dec); +EXTINLINE int mpd_isnormal(const mpd_t *dec, const mpd_context_t *ctx); +EXTINLINE int mpd_issubnormal(const mpd_t *dec, const mpd_context_t *ctx); +/* odd word */ +EXTINLINE int mpd_isoddword(mpd_uint_t word); +/* odd coefficient */ +EXTINLINE int mpd_isoddcoeff(const mpd_t *dec); +/* odd decimal, only defined for integers */ +int mpd_isodd(const mpd_t *dec); +/* even decimal, only defined for integers */ +int mpd_iseven(const mpd_t *dec); +/* 0 if dec is positive, 1 if dec is negative */ +EXTINLINE uint8_t mpd_sign(const mpd_t *dec); +/* 1 if dec is positive, -1 if dec is negative */ +EXTINLINE int mpd_arith_sign(const mpd_t *dec); +EXTINLINE long mpd_radix(void); +EXTINLINE int mpd_isdynamic(mpd_t *dec); +EXTINLINE int mpd_isstatic(mpd_t *dec); +EXTINLINE int mpd_isdynamic_data(mpd_t *dec); +EXTINLINE int mpd_isstatic_data(mpd_t *dec); +EXTINLINE int mpd_isshared_data(mpd_t *dec); +EXTINLINE int mpd_isconst_data(mpd_t *dec); +EXTINLINE mpd_ssize_t mpd_trail_zeros(const mpd_t *dec); + + +/******************************************************************************/ +/* Set attributes of a decimal */ +/******************************************************************************/ + +/* set number of decimal digits in the coefficient */ +EXTINLINE void mpd_setdigits(mpd_t *result); +EXTINLINE void mpd_set_sign(mpd_t *result, uint8_t sign); +/* copy sign from another decimal */ +EXTINLINE void mpd_signcpy(mpd_t *result, mpd_t *a); +EXTINLINE void mpd_set_infinity(mpd_t *result); +EXTINLINE void mpd_set_qnan(mpd_t *result); +EXTINLINE void mpd_set_snan(mpd_t *result); +EXTINLINE void mpd_set_negative(mpd_t *result); +EXTINLINE void mpd_set_positive(mpd_t *result); +EXTINLINE void mpd_set_dynamic(mpd_t *result); +EXTINLINE void mpd_set_static(mpd_t *result); +EXTINLINE void mpd_set_dynamic_data(mpd_t *result); +EXTINLINE void mpd_set_static_data(mpd_t *result); +EXTINLINE void mpd_set_shared_data(mpd_t *result); +EXTINLINE void mpd_set_const_data(mpd_t *result); +EXTINLINE void mpd_clear_flags(mpd_t *result); +EXTINLINE void mpd_set_flags(mpd_t *result, uint8_t flags); +EXTINLINE void mpd_copy_flags(mpd_t *result, const mpd_t *a); + + +/******************************************************************************/ +/* Error Macros */ +/******************************************************************************/ + +#define mpd_err_fatal(...) \ + do {fprintf(stderr, "%s:%d: error: ", __FILE__, __LINE__); \ + fprintf(stderr, __VA_ARGS__); fputc('\n', stderr); \ + abort(); \ + } while (0) +#define mpd_err_warn(...) \ + do {fprintf(stderr, "%s:%d: warning: ", __FILE__, __LINE__); \ + fprintf(stderr, __VA_ARGS__); fputc('\n', stderr); \ + } while (0) + + +/******************************************************************************/ +/* Memory handling */ +/******************************************************************************/ + +extern void *(* mpd_mallocfunc)(size_t size); +extern void *(* mpd_callocfunc)(size_t nmemb, size_t size); +extern void *(* mpd_reallocfunc)(void *ptr, size_t size); +extern void (* mpd_free)(void *ptr); + +void *mpd_callocfunc_em(size_t nmemb, size_t size); + +void *mpd_alloc(mpd_size_t nmemb, mpd_size_t size); +void *mpd_calloc(mpd_size_t nmemb, mpd_size_t size); +void *mpd_realloc(void *ptr, mpd_size_t nmemb, mpd_size_t size, uint8_t *err); +void *mpd_sh_alloc(mpd_size_t struct_size, mpd_size_t nmemb, mpd_size_t size); + +mpd_t *mpd_qnew(void); +mpd_t *mpd_new(mpd_context_t *ctx); +mpd_t *mpd_qnew_size(mpd_ssize_t size); +void mpd_del(mpd_t *dec); + +void mpd_uint_zero(mpd_uint_t *dest, mpd_size_t len); +int mpd_qresize(mpd_t *result, mpd_ssize_t size, uint32_t *status); +int mpd_qresize_zero(mpd_t *result, mpd_ssize_t size, uint32_t *status); +void mpd_minalloc(mpd_t *result); + +int mpd_resize(mpd_t *result, mpd_ssize_t size, mpd_context_t *ctx); +int mpd_resize_zero(mpd_t *result, mpd_ssize_t size, mpd_context_t *ctx); + + +#ifdef __cplusplus +} /* END extern "C" */ +#endif + + +#endif /* MPDECIMAL_H */ + + + diff --git a/Modules/_decimal/libmpdec/numbertheory.c b/Modules/_decimal/libmpdec/numbertheory.c new file mode 100644 index 0000000000..10ce6dc146 --- /dev/null +++ b/Modules/_decimal/libmpdec/numbertheory.c @@ -0,0 +1,132 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#include "mpdecimal.h" +#include <stdlib.h> +#include <assert.h> +#include "bits.h" +#include "umodarith.h" +#include "numbertheory.h" + + +/* Bignum: Initialize the Number Theoretic Transform. */ + + +/* + * Return the nth root of unity in F(p). This corresponds to e**((2*pi*i)/n) + * in the Fourier transform. We have w**n == 1 (mod p). + * n := transform length. + * sign := -1 for forward transform, 1 for backward transform. + * modnum := one of {P1, P2, P3}. + */ +mpd_uint_t +_mpd_getkernel(mpd_uint_t n, int sign, int modnum) +{ + mpd_uint_t umod, p, r, xi; +#ifdef PPRO + double dmod; + uint32_t dinvmod[3]; +#endif + + SETMODULUS(modnum); + r = mpd_roots[modnum]; /* primitive root of F(p) */ + p = umod; + xi = (p-1) / n; + + if (sign == -1) + return POWMOD(r, (p-1-xi)); + else + return POWMOD(r, xi); +} + +/* + * Initialize and return transform parameters. + * n := transform length. + * sign := -1 for forward transform, 1 for backward transform. + * modnum := one of {P1, P2, P3}. + */ +struct fnt_params * +_mpd_init_fnt_params(mpd_size_t n, int sign, int modnum) +{ + struct fnt_params *tparams; + mpd_uint_t umod; +#ifdef PPRO + double dmod; + uint32_t dinvmod[3]; +#endif + mpd_uint_t kernel, w; + mpd_uint_t i; + mpd_size_t nhalf; + + assert(ispower2(n)); + assert(sign == -1 || sign == 1); + assert(P1 <= modnum && modnum <= P3); + + nhalf = n/2; + tparams = mpd_sh_alloc(sizeof *tparams, nhalf, sizeof (mpd_uint_t)); + if (tparams == NULL) { + return NULL; + } + + SETMODULUS(modnum); + kernel = _mpd_getkernel(n, sign, modnum); + + tparams->modnum = modnum; + tparams->modulus = umod; + tparams->kernel = kernel; + + /* wtable[] := w**0, w**1, ..., w**(nhalf-1) */ + w = 1; + for (i = 0; i < nhalf; i++) { + tparams->wtable[i] = w; + w = MULMOD(w, kernel); + } + + return tparams; +} + +/* Initialize wtable of size three. */ +void +_mpd_init_w3table(mpd_uint_t w3table[3], int sign, int modnum) +{ + mpd_uint_t umod; +#ifdef PPRO + double dmod; + uint32_t dinvmod[3]; +#endif + mpd_uint_t kernel; + + SETMODULUS(modnum); + kernel = _mpd_getkernel(3, sign, modnum); + + w3table[0] = 1; + w3table[1] = kernel; + w3table[2] = POWMOD(kernel, 2); +} + + diff --git a/Modules/_decimal/libmpdec/numbertheory.h b/Modules/_decimal/libmpdec/numbertheory.h new file mode 100644 index 0000000000..f54d11dc38 --- /dev/null +++ b/Modules/_decimal/libmpdec/numbertheory.h @@ -0,0 +1,71 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#ifndef NUMBER_THEORY_H +#define NUMBER_THEORY_H + + +#include "constants.h" +#include "mpdecimal.h" + + +/* transform parameters */ +struct fnt_params { + int modnum; + mpd_uint_t modulus; + mpd_uint_t kernel; + mpd_uint_t wtable[]; +}; + + +mpd_uint_t _mpd_getkernel(mpd_uint_t n, int sign, int modnum); +struct fnt_params *_mpd_init_fnt_params(mpd_size_t n, int sign, int modnum); +void _mpd_init_w3table(mpd_uint_t w3table[3], int sign, int modnum); + + +#ifdef PPRO +static inline void +ppro_setmodulus(int modnum, mpd_uint_t *umod, double *dmod, uint32_t dinvmod[3]) +{ + *dmod = *umod = mpd_moduli[modnum]; + dinvmod[0] = mpd_invmoduli[modnum][0]; + dinvmod[1] = mpd_invmoduli[modnum][1]; + dinvmod[2] = mpd_invmoduli[modnum][2]; +} +#else +static inline void +std_setmodulus(int modnum, mpd_uint_t *umod) +{ + *umod = mpd_moduli[modnum]; +} +#endif + + +#endif + + diff --git a/Modules/_decimal/libmpdec/sixstep.c b/Modules/_decimal/libmpdec/sixstep.c new file mode 100644 index 0000000000..7d0542d641 --- /dev/null +++ b/Modules/_decimal/libmpdec/sixstep.c @@ -0,0 +1,214 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#include "mpdecimal.h" +#include <stdio.h> +#include <stdlib.h> +#include <assert.h> +#include "bits.h" +#include "difradix2.h" +#include "numbertheory.h" +#include "transpose.h" +#include "umodarith.h" +#include "sixstep.h" + + +/* Bignum: Cache efficient Matrix Fourier Transform for arrays of the + form 2**n (See literature/six-step.txt). */ + + +/* forward transform with sign = -1 */ +int +six_step_fnt(mpd_uint_t *a, mpd_size_t n, int modnum) +{ + struct fnt_params *tparams; + mpd_size_t log2n, C, R; + mpd_uint_t kernel; + mpd_uint_t umod; +#ifdef PPRO + double dmod; + uint32_t dinvmod[3]; +#endif + mpd_uint_t *x, w0, w1, wstep; + mpd_size_t i, k; + + + assert(ispower2(n)); + assert(n >= 16); + assert(n <= MPD_MAXTRANSFORM_2N); + + log2n = mpd_bsr(n); + C = ((mpd_size_t)1) << (log2n / 2); /* number of columns */ + R = ((mpd_size_t)1) << (log2n - (log2n / 2)); /* number of rows */ + + + /* Transpose the matrix. */ + if (!transpose_pow2(a, R, C)) { + return 0; + } + + /* Length R transform on the rows. */ + if ((tparams = _mpd_init_fnt_params(R, -1, modnum)) == NULL) { + return 0; + } + for (x = a; x < a+n; x += R) { + fnt_dif2(x, R, tparams); + } + + /* Transpose the matrix. */ + if (!transpose_pow2(a, C, R)) { + mpd_free(tparams); + return 0; + } + + /* Multiply each matrix element (addressed by i*C+k) by r**(i*k). */ + SETMODULUS(modnum); + kernel = _mpd_getkernel(n, -1, modnum); + for (i = 1; i < R; i++) { + w0 = 1; /* r**(i*0): initial value for k=0 */ + w1 = POWMOD(kernel, i); /* r**(i*1): initial value for k=1 */ + wstep = MULMOD(w1, w1); /* r**(2*i) */ + for (k = 0; k < C; k += 2) { + mpd_uint_t x0 = a[i*C+k]; + mpd_uint_t x1 = a[i*C+k+1]; + MULMOD2(&x0, w0, &x1, w1); + MULMOD2C(&w0, &w1, wstep); /* r**(i*(k+2)) = r**(i*k) * r**(2*i) */ + a[i*C+k] = x0; + a[i*C+k+1] = x1; + } + } + + /* Length C transform on the rows. */ + if (C != R) { + mpd_free(tparams); + if ((tparams = _mpd_init_fnt_params(C, -1, modnum)) == NULL) { + return 0; + } + } + for (x = a; x < a+n; x += C) { + fnt_dif2(x, C, tparams); + } + mpd_free(tparams); + +#if 0 + /* An unordered transform is sufficient for convolution. */ + /* Transpose the matrix. */ + if (!transpose_pow2(a, R, C)) { + return 0; + } +#endif + + return 1; +} + + +/* reverse transform, sign = 1 */ +int +inv_six_step_fnt(mpd_uint_t *a, mpd_size_t n, int modnum) +{ + struct fnt_params *tparams; + mpd_size_t log2n, C, R; + mpd_uint_t kernel; + mpd_uint_t umod; +#ifdef PPRO + double dmod; + uint32_t dinvmod[3]; +#endif + mpd_uint_t *x, w0, w1, wstep; + mpd_size_t i, k; + + + assert(ispower2(n)); + assert(n >= 16); + assert(n <= MPD_MAXTRANSFORM_2N); + + log2n = mpd_bsr(n); + C = ((mpd_size_t)1) << (log2n / 2); /* number of columns */ + R = ((mpd_size_t)1) << (log2n - (log2n / 2)); /* number of rows */ + + +#if 0 + /* An unordered transform is sufficient for convolution. */ + /* Transpose the matrix, producing an R*C matrix. */ + if (!transpose_pow2(a, C, R)) { + return 0; + } +#endif + + /* Length C transform on the rows. */ + if ((tparams = _mpd_init_fnt_params(C, 1, modnum)) == NULL) { + return 0; + } + for (x = a; x < a+n; x += C) { + fnt_dif2(x, C, tparams); + } + + /* Multiply each matrix element (addressed by i*C+k) by r**(i*k). */ + SETMODULUS(modnum); + kernel = _mpd_getkernel(n, 1, modnum); + for (i = 1; i < R; i++) { + w0 = 1; + w1 = POWMOD(kernel, i); + wstep = MULMOD(w1, w1); + for (k = 0; k < C; k += 2) { + mpd_uint_t x0 = a[i*C+k]; + mpd_uint_t x1 = a[i*C+k+1]; + MULMOD2(&x0, w0, &x1, w1); + MULMOD2C(&w0, &w1, wstep); + a[i*C+k] = x0; + a[i*C+k+1] = x1; + } + } + + /* Transpose the matrix. */ + if (!transpose_pow2(a, R, C)) { + mpd_free(tparams); + return 0; + } + + /* Length R transform on the rows. */ + if (R != C) { + mpd_free(tparams); + if ((tparams = _mpd_init_fnt_params(R, 1, modnum)) == NULL) { + return 0; + } + } + for (x = a; x < a+n; x += R) { + fnt_dif2(x, R, tparams); + } + mpd_free(tparams); + + /* Transpose the matrix. */ + if (!transpose_pow2(a, C, R)) { + return 0; + } + + return 1; +} + + diff --git a/Modules/_decimal/libmpdec/sixstep.h b/Modules/_decimal/libmpdec/sixstep.h new file mode 100644 index 0000000000..4d251df059 --- /dev/null +++ b/Modules/_decimal/libmpdec/sixstep.h @@ -0,0 +1,41 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#ifndef SIX_STEP_H +#define SIX_STEP_H + + +#include "mpdecimal.h" +#include <stdio.h> + + +int six_step_fnt(mpd_uint_t *a, mpd_size_t n, int modnum); +int inv_six_step_fnt(mpd_uint_t *a, mpd_size_t n, int modnum); + + +#endif diff --git a/Modules/_decimal/libmpdec/transpose.c b/Modules/_decimal/libmpdec/transpose.c new file mode 100644 index 0000000000..5e5d4b6625 --- /dev/null +++ b/Modules/_decimal/libmpdec/transpose.c @@ -0,0 +1,276 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#include "mpdecimal.h" +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <limits.h> +#include <assert.h> +#include "bits.h" +#include "constants.h" +#include "typearith.h" +#include "transpose.h" + + +#define BUFSIZE 4096 +#define SIDE 128 + + +/* Bignum: The transpose functions are used for very large transforms + in sixstep.c and fourstep.c. */ + + +/* Definition of the matrix transpose */ +void +std_trans(mpd_uint_t dest[], mpd_uint_t src[], mpd_size_t rows, mpd_size_t cols) +{ + mpd_size_t idest, isrc; + mpd_size_t r, c; + + for (r = 0; r < rows; r++) { + isrc = r * cols; + idest = r; + for (c = 0; c < cols; c++) { + dest[idest] = src[isrc]; + isrc += 1; + idest += rows; + } + } +} + +/* + * Swap half-rows of 2^n * (2*2^n) matrix. + * FORWARD_CYCLE: even/odd permutation of the halfrows. + * BACKWARD_CYCLE: reverse the even/odd permutation. + */ +static int +swap_halfrows_pow2(mpd_uint_t *matrix, mpd_size_t rows, mpd_size_t cols, int dir) +{ + mpd_uint_t buf1[BUFSIZE]; + mpd_uint_t buf2[BUFSIZE]; + mpd_uint_t *readbuf, *writebuf, *hp; + mpd_size_t *done, dbits; + mpd_size_t b = BUFSIZE, stride; + mpd_size_t hn, hmax; /* halfrow number */ + mpd_size_t m, r=0; + mpd_size_t offset; + mpd_size_t next; + + + assert(cols == mul_size_t(2, rows)); + + if (dir == FORWARD_CYCLE) { + r = rows; + } + else if (dir == BACKWARD_CYCLE) { + r = 2; + } + else { + abort(); /* GCOV_NOT_REACHED */ + } + + m = cols - 1; + hmax = rows; /* cycles start at odd halfrows */ + dbits = 8 * sizeof *done; + if ((done = mpd_calloc(hmax/(sizeof *done) + 1, sizeof *done)) == NULL) { + return 0; + } + + for (hn = 1; hn <= hmax; hn += 2) { + + if (done[hn/dbits] & mpd_bits[hn%dbits]) { + continue; + } + + readbuf = buf1; writebuf = buf2; + + for (offset = 0; offset < cols/2; offset += b) { + + stride = (offset + b < cols/2) ? b : cols/2-offset; + + hp = matrix + hn*cols/2; + memcpy(readbuf, hp+offset, stride*(sizeof *readbuf)); + pointerswap(&readbuf, &writebuf); + + next = mulmod_size_t(hn, r, m); + hp = matrix + next*cols/2; + + while (next != hn) { + + memcpy(readbuf, hp+offset, stride*(sizeof *readbuf)); + memcpy(hp+offset, writebuf, stride*(sizeof *writebuf)); + pointerswap(&readbuf, &writebuf); + + done[next/dbits] |= mpd_bits[next%dbits]; + + next = mulmod_size_t(next, r, m); + hp = matrix + next*cols/2; + + } + + memcpy(hp+offset, writebuf, stride*(sizeof *writebuf)); + + done[hn/dbits] |= mpd_bits[hn%dbits]; + } + } + + mpd_free(done); + return 1; +} + +/* In-place transpose of a square matrix */ +static inline void +squaretrans(mpd_uint_t *buf, mpd_size_t cols) +{ + mpd_uint_t tmp; + mpd_size_t idest, isrc; + mpd_size_t r, c; + + for (r = 0; r < cols; r++) { + c = r+1; + isrc = r*cols + c; + idest = c*cols + r; + for (c = r+1; c < cols; c++) { + tmp = buf[isrc]; + buf[isrc] = buf[idest]; + buf[idest] = tmp; + isrc += 1; + idest += cols; + } + } +} + +/* + * Transpose 2^n * 2^n matrix. For cache efficiency, the matrix is split into + * square blocks with side length 'SIDE'. First, the blocks are transposed, + * then a square tranposition is done on each individual block. + */ +static void +squaretrans_pow2(mpd_uint_t *matrix, mpd_size_t size) +{ + mpd_uint_t buf1[SIDE*SIDE]; + mpd_uint_t buf2[SIDE*SIDE]; + mpd_uint_t *to, *from; + mpd_size_t b = size; + mpd_size_t r, c; + mpd_size_t i; + + while (b > SIDE) b >>= 1; + + for (r = 0; r < size; r += b) { + + for (c = r; c < size; c += b) { + + from = matrix + r*size + c; + to = buf1; + for (i = 0; i < b; i++) { + memcpy(to, from, b*(sizeof *to)); + from += size; + to += b; + } + squaretrans(buf1, b); + + if (r == c) { + to = matrix + r*size + c; + from = buf1; + for (i = 0; i < b; i++) { + memcpy(to, from, b*(sizeof *to)); + from += b; + to += size; + } + continue; + } + else { + from = matrix + c*size + r; + to = buf2; + for (i = 0; i < b; i++) { + memcpy(to, from, b*(sizeof *to)); + from += size; + to += b; + } + squaretrans(buf2, b); + + to = matrix + c*size + r; + from = buf1; + for (i = 0; i < b; i++) { + memcpy(to, from, b*(sizeof *to)); + from += b; + to += size; + } + + to = matrix + r*size + c; + from = buf2; + for (i = 0; i < b; i++) { + memcpy(to, from, b*(sizeof *to)); + from += b; + to += size; + } + } + } + } + +} + +/* + * In-place transposition of a 2^n x 2^n or a 2^n x (2*2^n) + * or a (2*2^n) x 2^n matrix. + */ +int +transpose_pow2(mpd_uint_t *matrix, mpd_size_t rows, mpd_size_t cols) +{ + mpd_size_t size = mul_size_t(rows, cols); + + assert(ispower2(rows)); + assert(ispower2(cols)); + + if (cols == rows) { + squaretrans_pow2(matrix, rows); + } + else if (cols == mul_size_t(2, rows)) { + if (!swap_halfrows_pow2(matrix, rows, cols, FORWARD_CYCLE)) { + return 0; + } + squaretrans_pow2(matrix, rows); + squaretrans_pow2(matrix+(size/2), rows); + } + else if (rows == mul_size_t(2, cols)) { + squaretrans_pow2(matrix, cols); + squaretrans_pow2(matrix+(size/2), cols); + if (!swap_halfrows_pow2(matrix, cols, rows, BACKWARD_CYCLE)) { + return 0; + } + } + else { + abort(); /* GCOV_NOT_REACHED */ + } + + return 1; +} + + diff --git a/Modules/_decimal/libmpdec/transpose.h b/Modules/_decimal/libmpdec/transpose.h new file mode 100644 index 0000000000..dd0aec6600 --- /dev/null +++ b/Modules/_decimal/libmpdec/transpose.h @@ -0,0 +1,55 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#ifndef TRANSPOSE_H +#define TRANSPOSE_H + + +#include "mpdecimal.h" +#include <stdio.h> + + +enum {FORWARD_CYCLE, BACKWARD_CYCLE}; + + +void std_trans(mpd_uint_t dest[], mpd_uint_t src[], mpd_size_t rows, mpd_size_t cols); +int transpose_pow2(mpd_uint_t *matrix, mpd_size_t rows, mpd_size_t cols); +void transpose_3xpow2(mpd_uint_t *matrix, mpd_size_t rows, mpd_size_t cols); + + +static inline void pointerswap(mpd_uint_t **a, mpd_uint_t **b) +{ + mpd_uint_t *tmp; + + tmp = *b; + *b = *a; + *a = tmp; +} + + +#endif diff --git a/Modules/_decimal/libmpdec/typearith.h b/Modules/_decimal/libmpdec/typearith.h new file mode 100644 index 0000000000..eeba8dd5ba --- /dev/null +++ b/Modules/_decimal/libmpdec/typearith.h @@ -0,0 +1,669 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#ifndef TYPEARITH_H +#define TYPEARITH_H + + +#include "mpdecimal.h" + + +/*****************************************************************************/ +/* Low level native arithmetic on basic types */ +/*****************************************************************************/ + + +/** ------------------------------------------------------------ + ** Double width multiplication and division + ** ------------------------------------------------------------ + */ + +#if defined(CONFIG_64) +#if defined(ANSI) +#if defined(HAVE_UINT128_T) +static inline void +_mpd_mul_words(mpd_uint_t *hi, mpd_uint_t *lo, mpd_uint_t a, mpd_uint_t b) +{ + __uint128_t hl; + + hl = (__uint128_t)a * b; + + *hi = hl >> 64; + *lo = (mpd_uint_t)hl; +} + +static inline void +_mpd_div_words(mpd_uint_t *q, mpd_uint_t *r, mpd_uint_t hi, mpd_uint_t lo, + mpd_uint_t d) +{ + __uint128_t hl; + + hl = ((__uint128_t)hi<<64) + lo; + *q = (mpd_uint_t)(hl / d); /* quotient is known to fit */ + *r = (mpd_uint_t)(hl - (__uint128_t)(*q) * d); +} +#else +static inline void +_mpd_mul_words(mpd_uint_t *hi, mpd_uint_t *lo, mpd_uint_t a, mpd_uint_t b) +{ + uint32_t w[4], carry; + uint32_t ah, al, bh, bl; + uint64_t hl; + + ah = (uint32_t)(a>>32); al = (uint32_t)a; + bh = (uint32_t)(b>>32); bl = (uint32_t)b; + + hl = (uint64_t)al * bl; + w[0] = (uint32_t)hl; + carry = (uint32_t)(hl>>32); + + hl = (uint64_t)ah * bl + carry; + w[1] = (uint32_t)hl; + w[2] = (uint32_t)(hl>>32); + + hl = (uint64_t)al * bh + w[1]; + w[1] = (uint32_t)hl; + carry = (uint32_t)(hl>>32); + + hl = ((uint64_t)ah * bh + w[2]) + carry; + w[2] = (uint32_t)hl; + w[3] = (uint32_t)(hl>>32); + + *hi = ((uint64_t)w[3]<<32) + w[2]; + *lo = ((uint64_t)w[1]<<32) + w[0]; +} + +/* + * By Henry S. Warren: http://www.hackersdelight.org/HDcode/divlu.c.txt + * http://www.hackersdelight.org/permissions.htm: + * "You are free to use, copy, and distribute any of the code on this web + * site, whether modified by you or not. You need not give attribution." + * + * Slightly modified, comments are mine. + */ +static inline int +nlz(uint64_t x) +{ + int n; + + if (x == 0) return(64); + + n = 0; + if (x <= 0x00000000FFFFFFFF) {n = n +32; x = x <<32;} + if (x <= 0x0000FFFFFFFFFFFF) {n = n +16; x = x <<16;} + if (x <= 0x00FFFFFFFFFFFFFF) {n = n + 8; x = x << 8;} + if (x <= 0x0FFFFFFFFFFFFFFF) {n = n + 4; x = x << 4;} + if (x <= 0x3FFFFFFFFFFFFFFF) {n = n + 2; x = x << 2;} + if (x <= 0x7FFFFFFFFFFFFFFF) {n = n + 1;} + + return n; +} + +static inline void +_mpd_div_words(mpd_uint_t *q, mpd_uint_t *r, mpd_uint_t u1, mpd_uint_t u0, + mpd_uint_t v) +{ + const mpd_uint_t b = 4294967296; + mpd_uint_t un1, un0, + vn1, vn0, + q1, q0, + un32, un21, un10, + rhat, t; + int s; + + assert(u1 < v); + + s = nlz(v); + v = v << s; + vn1 = v >> 32; + vn0 = v & 0xFFFFFFFF; + + t = (s == 0) ? 0 : u0 >> (64 - s); + un32 = (u1 << s) | t; + un10 = u0 << s; + + un1 = un10 >> 32; + un0 = un10 & 0xFFFFFFFF; + + q1 = un32 / vn1; + rhat = un32 - q1*vn1; +again1: + if (q1 >= b || q1*vn0 > b*rhat + un1) { + q1 = q1 - 1; + rhat = rhat + vn1; + if (rhat < b) goto again1; + } + + /* + * Before again1 we had: + * (1) q1*vn1 + rhat = un32 + * (2) q1*vn1*b + rhat*b + un1 = un32*b + un1 + * + * The statements inside the if-clause do not change the value + * of the left-hand side of (2), and the loop is only exited + * if q1*vn0 <= rhat*b + un1, so: + * + * (3) q1*vn1*b + q1*vn0 <= un32*b + un1 + * (4) q1*v <= un32*b + un1 + * (5) 0 <= un32*b + un1 - q1*v + * + * By (5) we are certain that the possible add-back step from + * Knuth's algorithm D is never required. + * + * Since the final quotient is less than 2**64, the following + * must be true: + * + * (6) un32*b + un1 - q1*v <= UINT64_MAX + * + * This means that in the following line, the high words + * of un32*b and q1*v can be discarded without any effect + * on the result. + */ + un21 = un32*b + un1 - q1*v; + + q0 = un21 / vn1; + rhat = un21 - q0*vn1; +again2: + if (q0 >= b || q0*vn0 > b*rhat + un0) { + q0 = q0 - 1; + rhat = rhat + vn1; + if (rhat < b) goto again2; + } + + *q = q1*b + q0; + *r = (un21*b + un0 - q0*v) >> s; +} +#endif + +/* END ANSI */ +#elif defined(ASM) +static inline void +_mpd_mul_words(mpd_uint_t *hi, mpd_uint_t *lo, mpd_uint_t a, mpd_uint_t b) +{ + mpd_uint_t h, l; + + asm ( "mulq %3\n\t" + : "=d" (h), "=a" (l) + : "%a" (a), "rm" (b) + : "cc" + ); + + *hi = h; + *lo = l; +} + +static inline void +_mpd_div_words(mpd_uint_t *q, mpd_uint_t *r, mpd_uint_t hi, mpd_uint_t lo, + mpd_uint_t d) +{ + mpd_uint_t qq, rr; + + asm ( "divq %4\n\t" + : "=a" (qq), "=d" (rr) + : "a" (lo), "d" (hi), "rm" (d) + : "cc" + ); + + *q = qq; + *r = rr; +} +/* END GCC ASM */ +#elif defined(MASM) +#include <intrin.h> +#pragma intrinsic(_umul128) + +static inline void +_mpd_mul_words(mpd_uint_t *hi, mpd_uint_t *lo, mpd_uint_t a, mpd_uint_t b) +{ + *lo = _umul128(a, b, hi); +} + +void _mpd_div_words(mpd_uint_t *q, mpd_uint_t *r, mpd_uint_t hi, mpd_uint_t lo, + mpd_uint_t d); + +/* END MASM (_MSC_VER) */ +#else + #error "need platform specific 128 bit multiplication and division" +#endif + +#define DIVMOD(q, r, v, d) *q = v / d; *r = v - *q * d +static inline void +_mpd_divmod_pow10(mpd_uint_t *q, mpd_uint_t *r, mpd_uint_t v, mpd_uint_t exp) +{ + assert(exp <= 19); + + if (exp <= 9) { + if (exp <= 4) { + switch (exp) { + case 0: *q = v; *r = 0; break; + case 1: DIVMOD(q, r, v, 10UL); break; + case 2: DIVMOD(q, r, v, 100UL); break; + case 3: DIVMOD(q, r, v, 1000UL); break; + case 4: DIVMOD(q, r, v, 10000UL); break; + } + } + else { + switch (exp) { + case 5: DIVMOD(q, r, v, 100000UL); break; + case 6: DIVMOD(q, r, v, 1000000UL); break; + case 7: DIVMOD(q, r, v, 10000000UL); break; + case 8: DIVMOD(q, r, v, 100000000UL); break; + case 9: DIVMOD(q, r, v, 1000000000UL); break; + } + } + } + else { + if (exp <= 14) { + switch (exp) { + case 10: DIVMOD(q, r, v, 10000000000ULL); break; + case 11: DIVMOD(q, r, v, 100000000000ULL); break; + case 12: DIVMOD(q, r, v, 1000000000000ULL); break; + case 13: DIVMOD(q, r, v, 10000000000000ULL); break; + case 14: DIVMOD(q, r, v, 100000000000000ULL); break; + } + } + else { + switch (exp) { + case 15: DIVMOD(q, r, v, 1000000000000000ULL); break; + case 16: DIVMOD(q, r, v, 10000000000000000ULL); break; + case 17: DIVMOD(q, r, v, 100000000000000000ULL); break; + case 18: DIVMOD(q, r, v, 1000000000000000000ULL); break; + case 19: DIVMOD(q, r, v, 10000000000000000000ULL); break; /* GCOV_NOT_REACHED */ + } + } + } +} + +/* END CONFIG_64 */ +#elif defined(CONFIG_32) +#if defined(ANSI) +#if !defined(LEGACY_COMPILER) +static inline void +_mpd_mul_words(mpd_uint_t *hi, mpd_uint_t *lo, mpd_uint_t a, mpd_uint_t b) +{ + mpd_uuint_t hl; + + hl = (mpd_uuint_t)a * b; + + *hi = hl >> 32; + *lo = (mpd_uint_t)hl; +} + +static inline void +_mpd_div_words(mpd_uint_t *q, mpd_uint_t *r, mpd_uint_t hi, mpd_uint_t lo, + mpd_uint_t d) +{ + mpd_uuint_t hl; + + hl = ((mpd_uuint_t)hi<<32) + lo; + *q = (mpd_uint_t)(hl / d); /* quotient is known to fit */ + *r = (mpd_uint_t)(hl - (mpd_uuint_t)(*q) * d); +} +/* END ANSI + uint64_t */ +#else +static inline void +_mpd_mul_words(mpd_uint_t *hi, mpd_uint_t *lo, mpd_uint_t a, mpd_uint_t b) +{ + uint16_t w[4], carry; + uint16_t ah, al, bh, bl; + uint32_t hl; + + ah = (uint16_t)(a>>16); al = (uint16_t)a; + bh = (uint16_t)(b>>16); bl = (uint16_t)b; + + hl = (uint32_t)al * bl; + w[0] = (uint16_t)hl; + carry = (uint16_t)(hl>>16); + + hl = (uint32_t)ah * bl + carry; + w[1] = (uint16_t)hl; + w[2] = (uint16_t)(hl>>16); + + hl = (uint32_t)al * bh + w[1]; + w[1] = (uint16_t)hl; + carry = (uint16_t)(hl>>16); + + hl = ((uint32_t)ah * bh + w[2]) + carry; + w[2] = (uint16_t)hl; + w[3] = (uint16_t)(hl>>16); + + *hi = ((uint32_t)w[3]<<16) + w[2]; + *lo = ((uint32_t)w[1]<<16) + w[0]; +} + +/* + * By Henry S. Warren: http://www.hackersdelight.org/HDcode/divlu.c.txt + * http://www.hackersdelight.org/permissions.htm: + * "You are free to use, copy, and distribute any of the code on this web + * site, whether modified by you or not. You need not give attribution." + * + * Slightly modified, comments are mine. + */ +static inline int +nlz(uint32_t x) +{ + int n; + + if (x == 0) return(32); + + n = 0; + if (x <= 0x0000FFFF) {n = n +16; x = x <<16;} + if (x <= 0x00FFFFFF) {n = n + 8; x = x << 8;} + if (x <= 0x0FFFFFFF) {n = n + 4; x = x << 4;} + if (x <= 0x3FFFFFFF) {n = n + 2; x = x << 2;} + if (x <= 0x7FFFFFFF) {n = n + 1;} + + return n; +} + +static inline void +_mpd_div_words(mpd_uint_t *q, mpd_uint_t *r, mpd_uint_t u1, mpd_uint_t u0, + mpd_uint_t v) +{ + const mpd_uint_t b = 65536; + mpd_uint_t un1, un0, + vn1, vn0, + q1, q0, + un32, un21, un10, + rhat, t; + int s; + + assert(u1 < v); + + s = nlz(v); + v = v << s; + vn1 = v >> 16; + vn0 = v & 0xFFFF; + + t = (s == 0) ? 0 : u0 >> (32 - s); + un32 = (u1 << s) | t; + un10 = u0 << s; + + un1 = un10 >> 16; + un0 = un10 & 0xFFFF; + + q1 = un32 / vn1; + rhat = un32 - q1*vn1; +again1: + if (q1 >= b || q1*vn0 > b*rhat + un1) { + q1 = q1 - 1; + rhat = rhat + vn1; + if (rhat < b) goto again1; + } + + /* + * Before again1 we had: + * (1) q1*vn1 + rhat = un32 + * (2) q1*vn1*b + rhat*b + un1 = un32*b + un1 + * + * The statements inside the if-clause do not change the value + * of the left-hand side of (2), and the loop is only exited + * if q1*vn0 <= rhat*b + un1, so: + * + * (3) q1*vn1*b + q1*vn0 <= un32*b + un1 + * (4) q1*v <= un32*b + un1 + * (5) 0 <= un32*b + un1 - q1*v + * + * By (5) we are certain that the possible add-back step from + * Knuth's algorithm D is never required. + * + * Since the final quotient is less than 2**32, the following + * must be true: + * + * (6) un32*b + un1 - q1*v <= UINT32_MAX + * + * This means that in the following line, the high words + * of un32*b and q1*v can be discarded without any effect + * on the result. + */ + un21 = un32*b + un1 - q1*v; + + q0 = un21 / vn1; + rhat = un21 - q0*vn1; +again2: + if (q0 >= b || q0*vn0 > b*rhat + un0) { + q0 = q0 - 1; + rhat = rhat + vn1; + if (rhat < b) goto again2; + } + + *q = q1*b + q0; + *r = (un21*b + un0 - q0*v) >> s; +} +#endif /* END ANSI + LEGACY_COMPILER */ + +/* END ANSI */ +#elif defined(ASM) +static inline void +_mpd_mul_words(mpd_uint_t *hi, mpd_uint_t *lo, mpd_uint_t a, mpd_uint_t b) +{ + mpd_uint_t h, l; + + asm ( "mull %3\n\t" + : "=d" (h), "=a" (l) + : "%a" (a), "rm" (b) + : "cc" + ); + + *hi = h; + *lo = l; +} + +static inline void +_mpd_div_words(mpd_uint_t *q, mpd_uint_t *r, mpd_uint_t hi, mpd_uint_t lo, + mpd_uint_t d) +{ + mpd_uint_t qq, rr; + + asm ( "divl %4\n\t" + : "=a" (qq), "=d" (rr) + : "a" (lo), "d" (hi), "rm" (d) + : "cc" + ); + + *q = qq; + *r = rr; +} +/* END GCC ASM */ +#elif defined(MASM) +static inline void __cdecl +_mpd_mul_words(mpd_uint_t *hi, mpd_uint_t *lo, mpd_uint_t a, mpd_uint_t b) +{ + mpd_uint_t h, l; + + __asm { + mov eax, a + mul b + mov h, edx + mov l, eax + } + + *hi = h; + *lo = l; +} + +static inline void __cdecl +_mpd_div_words(mpd_uint_t *q, mpd_uint_t *r, mpd_uint_t hi, mpd_uint_t lo, + mpd_uint_t d) +{ + mpd_uint_t qq, rr; + + __asm { + mov eax, lo + mov edx, hi + div d + mov qq, eax + mov rr, edx + } + + *q = qq; + *r = rr; +} +/* END MASM (_MSC_VER) */ +#else + #error "need platform specific 64 bit multiplication and division" +#endif + +#define DIVMOD(q, r, v, d) *q = v / d; *r = v - *q * d +static inline void +_mpd_divmod_pow10(mpd_uint_t *q, mpd_uint_t *r, mpd_uint_t v, mpd_uint_t exp) +{ + assert(exp <= 9); + + if (exp <= 4) { + switch (exp) { + case 0: *q = v; *r = 0; break; + case 1: DIVMOD(q, r, v, 10UL); break; + case 2: DIVMOD(q, r, v, 100UL); break; + case 3: DIVMOD(q, r, v, 1000UL); break; + case 4: DIVMOD(q, r, v, 10000UL); break; + } + } + else { + switch (exp) { + case 5: DIVMOD(q, r, v, 100000UL); break; + case 6: DIVMOD(q, r, v, 1000000UL); break; + case 7: DIVMOD(q, r, v, 10000000UL); break; + case 8: DIVMOD(q, r, v, 100000000UL); break; + case 9: DIVMOD(q, r, v, 1000000000UL); break; /* GCOV_NOT_REACHED */ + } + } +} +/* END CONFIG_32 */ + +/* NO CONFIG */ +#else + #error "define CONFIG_64 or CONFIG_32" +#endif /* CONFIG */ + + +static inline void +_mpd_div_word(mpd_uint_t *q, mpd_uint_t *r, mpd_uint_t v, mpd_uint_t d) +{ + *q = v / d; + *r = v - *q * d; +} + +static inline void +_mpd_idiv_word(mpd_ssize_t *q, mpd_ssize_t *r, mpd_ssize_t v, mpd_ssize_t d) +{ + *q = v / d; + *r = v - *q * d; +} + + +/** ------------------------------------------------------------ + ** Arithmetic with overflow checking + ** ------------------------------------------------------------ + */ + +/* The following macros do call exit() in case of an overflow. + If the library is used correctly (i.e. with valid context + parameters), such overflows cannot occur. The macros are used + as sanity checks in a couple of strategic places and should + be viewed as a handwritten version of gcc's -ftrapv option. */ + +static inline mpd_size_t +add_size_t(mpd_size_t a, mpd_size_t b) +{ + if (a > MPD_SIZE_MAX - b) { + mpd_err_fatal("add_size_t(): overflow: check the context"); /* GCOV_NOT_REACHED */ + } + return a + b; +} + +static inline mpd_size_t +sub_size_t(mpd_size_t a, mpd_size_t b) +{ + if (b > a) { + mpd_err_fatal("sub_size_t(): overflow: check the context"); /* GCOV_NOT_REACHED */ + } + return a - b; +} + +#if MPD_SIZE_MAX != MPD_UINT_MAX + #error "adapt mul_size_t() and mulmod_size_t()" +#endif + +static inline mpd_size_t +mul_size_t(mpd_size_t a, mpd_size_t b) +{ + mpd_uint_t hi, lo; + + _mpd_mul_words(&hi, &lo, (mpd_uint_t)a, (mpd_uint_t)b); + if (hi) { + mpd_err_fatal("mul_size_t(): overflow: check the context"); /* GCOV_NOT_REACHED */ + } + return lo; +} + +static inline mpd_size_t +add_size_t_overflow(mpd_size_t a, mpd_size_t b, mpd_size_t *overflow) +{ + mpd_size_t ret; + + *overflow = 0; + ret = a + b; + if (ret < a) *overflow = 1; + return ret; +} + +static inline mpd_size_t +mul_size_t_overflow(mpd_size_t a, mpd_size_t b, mpd_size_t *overflow) +{ + mpd_uint_t lo; + + _mpd_mul_words((mpd_uint_t *)overflow, &lo, (mpd_uint_t)a, + (mpd_uint_t)b); + return lo; +} + +static inline mpd_ssize_t +mod_mpd_ssize_t(mpd_ssize_t a, mpd_ssize_t m) +{ + mpd_ssize_t r = a % m; + return (r < 0) ? r + m : r; +} + +static inline mpd_size_t +mulmod_size_t(mpd_size_t a, mpd_size_t b, mpd_size_t m) +{ + mpd_uint_t hi, lo; + mpd_uint_t q, r; + + _mpd_mul_words(&hi, &lo, (mpd_uint_t)a, (mpd_uint_t)b); + _mpd_div_words(&q, &r, hi, lo, (mpd_uint_t)m); + + return r; +} + + +#endif /* TYPEARITH_H */ + + + diff --git a/Modules/_decimal/libmpdec/umodarith.h b/Modules/_decimal/libmpdec/umodarith.h new file mode 100644 index 0000000000..06cde0a7ad --- /dev/null +++ b/Modules/_decimal/libmpdec/umodarith.h @@ -0,0 +1,650 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#ifndef UMODARITH_H +#define UMODARITH_H + + +#include "constants.h" +#include "mpdecimal.h" +#include "typearith.h" + + +/* Bignum: Low level routines for unsigned modular arithmetic. These are + used in the fast convolution functions for very large coefficients. */ + + +/**************************************************************************/ +/* ANSI modular arithmetic */ +/**************************************************************************/ + + +/* + * Restrictions: a < m and b < m + * ACL2 proof: umodarith.lisp: addmod-correct + */ +static inline mpd_uint_t +addmod(mpd_uint_t a, mpd_uint_t b, mpd_uint_t m) +{ + mpd_uint_t s; + + s = a + b; + s = (s < a) ? s - m : s; + s = (s >= m) ? s - m : s; + + return s; +} + +/* + * Restrictions: a < m and b < m + * ACL2 proof: umodarith.lisp: submod-2-correct + */ +static inline mpd_uint_t +submod(mpd_uint_t a, mpd_uint_t b, mpd_uint_t m) +{ + mpd_uint_t d; + + d = a - b; + d = (a < b) ? d + m : d; + + return d; +} + +/* + * Restrictions: a < 2m and b < 2m + * ACL2 proof: umodarith.lisp: section ext-submod + */ +static inline mpd_uint_t +ext_submod(mpd_uint_t a, mpd_uint_t b, mpd_uint_t m) +{ + mpd_uint_t d; + + a = (a >= m) ? a - m : a; + b = (b >= m) ? b - m : b; + + d = a - b; + d = (a < b) ? d + m : d; + + return d; +} + +/* + * Reduce double word modulo m. + * Restrictions: m != 0 + * ACL2 proof: umodarith.lisp: section dw-reduce + */ +static inline mpd_uint_t +dw_reduce(mpd_uint_t hi, mpd_uint_t lo, mpd_uint_t m) +{ + mpd_uint_t r1, r2, w; + + _mpd_div_word(&w, &r1, hi, m); + _mpd_div_words(&w, &r2, r1, lo, m); + + return r2; +} + +/* + * Subtract double word from a. + * Restrictions: a < m + * ACL2 proof: umodarith.lisp: section dw-submod + */ +static inline mpd_uint_t +dw_submod(mpd_uint_t a, mpd_uint_t hi, mpd_uint_t lo, mpd_uint_t m) +{ + mpd_uint_t d, r; + + r = dw_reduce(hi, lo, m); + d = a - r; + d = (a < r) ? d + m : d; + + return d; +} + +#ifdef CONFIG_64 + +/**************************************************************************/ +/* 64-bit modular arithmetic */ +/**************************************************************************/ + +/* + * A proof of the algorithm is in literature/mulmod-64.txt. An ACL2 + * proof is in umodarith.lisp: section "Fast modular reduction". + * + * Algorithm: calculate (a * b) % p: + * + * a) hi, lo <- a * b # Calculate a * b. + * + * b) hi, lo <- R(hi, lo) # Reduce modulo p. + * + * c) Repeat step b) until 0 <= hi * 2**64 + lo < 2*p. + * + * d) If the result is less than p, return lo. Otherwise return lo - p. + */ + +static inline mpd_uint_t +x64_mulmod(mpd_uint_t a, mpd_uint_t b, mpd_uint_t m) +{ + mpd_uint_t hi, lo, x, y; + + + _mpd_mul_words(&hi, &lo, a, b); + + if (m & (1ULL<<32)) { /* P1 */ + + /* first reduction */ + x = y = hi; + hi >>= 32; + + x = lo - x; + if (x > lo) hi--; + + y <<= 32; + lo = y + x; + if (lo < y) hi++; + + /* second reduction */ + x = y = hi; + hi >>= 32; + + x = lo - x; + if (x > lo) hi--; + + y <<= 32; + lo = y + x; + if (lo < y) hi++; + + return (hi || lo >= m ? lo - m : lo); + } + else if (m & (1ULL<<34)) { /* P2 */ + + /* first reduction */ + x = y = hi; + hi >>= 30; + + x = lo - x; + if (x > lo) hi--; + + y <<= 34; + lo = y + x; + if (lo < y) hi++; + + /* second reduction */ + x = y = hi; + hi >>= 30; + + x = lo - x; + if (x > lo) hi--; + + y <<= 34; + lo = y + x; + if (lo < y) hi++; + + /* third reduction */ + x = y = hi; + hi >>= 30; + + x = lo - x; + if (x > lo) hi--; + + y <<= 34; + lo = y + x; + if (lo < y) hi++; + + return (hi || lo >= m ? lo - m : lo); + } + else { /* P3 */ + + /* first reduction */ + x = y = hi; + hi >>= 24; + + x = lo - x; + if (x > lo) hi--; + + y <<= 40; + lo = y + x; + if (lo < y) hi++; + + /* second reduction */ + x = y = hi; + hi >>= 24; + + x = lo - x; + if (x > lo) hi--; + + y <<= 40; + lo = y + x; + if (lo < y) hi++; + + /* third reduction */ + x = y = hi; + hi >>= 24; + + x = lo - x; + if (x > lo) hi--; + + y <<= 40; + lo = y + x; + if (lo < y) hi++; + + return (hi || lo >= m ? lo - m : lo); + } +} + +static inline void +x64_mulmod2c(mpd_uint_t *a, mpd_uint_t *b, mpd_uint_t w, mpd_uint_t m) +{ + *a = x64_mulmod(*a, w, m); + *b = x64_mulmod(*b, w, m); +} + +static inline void +x64_mulmod2(mpd_uint_t *a0, mpd_uint_t b0, mpd_uint_t *a1, mpd_uint_t b1, + mpd_uint_t m) +{ + *a0 = x64_mulmod(*a0, b0, m); + *a1 = x64_mulmod(*a1, b1, m); +} + +static inline mpd_uint_t +x64_powmod(mpd_uint_t base, mpd_uint_t exp, mpd_uint_t umod) +{ + mpd_uint_t r = 1; + + while (exp > 0) { + if (exp & 1) + r = x64_mulmod(r, base, umod); + base = x64_mulmod(base, base, umod); + exp >>= 1; + } + + return r; +} + +/* END CONFIG_64 */ +#else /* CONFIG_32 */ + + +/**************************************************************************/ +/* 32-bit modular arithmetic */ +/**************************************************************************/ + +#if defined(ANSI) +#if !defined(LEGACY_COMPILER) +/* HAVE_UINT64_T */ +static inline mpd_uint_t +std_mulmod(mpd_uint_t a, mpd_uint_t b, mpd_uint_t m) +{ + return ((mpd_uuint_t) a * b) % m; +} + +static inline void +std_mulmod2c(mpd_uint_t *a, mpd_uint_t *b, mpd_uint_t w, mpd_uint_t m) +{ + *a = ((mpd_uuint_t) *a * w) % m; + *b = ((mpd_uuint_t) *b * w) % m; +} + +static inline void +std_mulmod2(mpd_uint_t *a0, mpd_uint_t b0, mpd_uint_t *a1, mpd_uint_t b1, + mpd_uint_t m) +{ + *a0 = ((mpd_uuint_t) *a0 * b0) % m; + *a1 = ((mpd_uuint_t) *a1 * b1) % m; +} +/* END HAVE_UINT64_T */ +#else +/* LEGACY_COMPILER */ +static inline mpd_uint_t +std_mulmod(mpd_uint_t a, mpd_uint_t b, mpd_uint_t m) +{ + mpd_uint_t hi, lo, q, r; + _mpd_mul_words(&hi, &lo, a, b); + _mpd_div_words(&q, &r, hi, lo, m); + return r; +} + +static inline void +std_mulmod2c(mpd_uint_t *a, mpd_uint_t *b, mpd_uint_t w, mpd_uint_t m) +{ + *a = std_mulmod(*a, w, m); + *b = std_mulmod(*b, w, m); +} + +static inline void +std_mulmod2(mpd_uint_t *a0, mpd_uint_t b0, mpd_uint_t *a1, mpd_uint_t b1, + mpd_uint_t m) +{ + *a0 = std_mulmod(*a0, b0, m); + *a1 = std_mulmod(*a1, b1, m); +} +/* END LEGACY_COMPILER */ +#endif + +static inline mpd_uint_t +std_powmod(mpd_uint_t base, mpd_uint_t exp, mpd_uint_t umod) +{ + mpd_uint_t r = 1; + + while (exp > 0) { + if (exp & 1) + r = std_mulmod(r, base, umod); + base = std_mulmod(base, base, umod); + exp >>= 1; + } + + return r; +} +#endif /* ANSI CONFIG_32 */ + + +/**************************************************************************/ +/* Pentium Pro modular arithmetic */ +/**************************************************************************/ + +/* + * A proof of the algorithm is in literature/mulmod-ppro.txt. The FPU + * control word must be set to 64-bit precision and truncation mode + * prior to using these functions. + * + * Algorithm: calculate (a * b) % p: + * + * p := prime < 2**31 + * pinv := (long double)1.0 / p (precalculated) + * + * a) n = a * b # Calculate exact product. + * b) qest = n * pinv # Calculate estimate for q = n / p. + * c) q = (qest+2**63)-2**63 # Truncate qest to the exact quotient. + * d) r = n - q * p # Calculate remainder. + * + * Remarks: + * + * - p = dmod and pinv = dinvmod. + * - dinvmod points to an array of three uint32_t, which is interpreted + * as an 80 bit long double by fldt. + * - Intel compilers prior to version 11 do not seem to handle the + * __GNUC__ inline assembly correctly. + * - random tests are provided in tests/extended/ppro_mulmod.c + */ + +#if defined(PPRO) +#if defined(ASM) + +/* Return (a * b) % dmod */ +static inline mpd_uint_t +ppro_mulmod(mpd_uint_t a, mpd_uint_t b, double *dmod, uint32_t *dinvmod) +{ + mpd_uint_t retval; + + asm ( + "fildl %2\n\t" + "fildl %1\n\t" + "fmulp %%st, %%st(1)\n\t" + "fldt (%4)\n\t" + "fmul %%st(1), %%st\n\t" + "flds %5\n\t" + "fadd %%st, %%st(1)\n\t" + "fsubrp %%st, %%st(1)\n\t" + "fldl (%3)\n\t" + "fmulp %%st, %%st(1)\n\t" + "fsubrp %%st, %%st(1)\n\t" + "fistpl %0\n\t" + : "=m" (retval) + : "m" (a), "m" (b), "r" (dmod), "r" (dinvmod), "m" (MPD_TWO63) + : "st", "memory" + ); + + return retval; +} + +/* + * Two modular multiplications in parallel: + * *a0 = (*a0 * w) % dmod + * *a1 = (*a1 * w) % dmod + */ +static inline void +ppro_mulmod2c(mpd_uint_t *a0, mpd_uint_t *a1, mpd_uint_t w, + double *dmod, uint32_t *dinvmod) +{ + asm ( + "fildl %2\n\t" + "fildl (%1)\n\t" + "fmul %%st(1), %%st\n\t" + "fxch %%st(1)\n\t" + "fildl (%0)\n\t" + "fmulp %%st, %%st(1) \n\t" + "fldt (%4)\n\t" + "flds %5\n\t" + "fld %%st(2)\n\t" + "fmul %%st(2)\n\t" + "fadd %%st(1)\n\t" + "fsub %%st(1)\n\t" + "fmull (%3)\n\t" + "fsubrp %%st, %%st(3)\n\t" + "fxch %%st(2)\n\t" + "fistpl (%0)\n\t" + "fmul %%st(2)\n\t" + "fadd %%st(1)\n\t" + "fsubp %%st, %%st(1)\n\t" + "fmull (%3)\n\t" + "fsubrp %%st, %%st(1)\n\t" + "fistpl (%1)\n\t" + : : "r" (a0), "r" (a1), "m" (w), + "r" (dmod), "r" (dinvmod), + "m" (MPD_TWO63) + : "st", "memory" + ); +} + +/* + * Two modular multiplications in parallel: + * *a0 = (*a0 * b0) % dmod + * *a1 = (*a1 * b1) % dmod + */ +static inline void +ppro_mulmod2(mpd_uint_t *a0, mpd_uint_t b0, mpd_uint_t *a1, mpd_uint_t b1, + double *dmod, uint32_t *dinvmod) +{ + asm ( + "fildl %3\n\t" + "fildl (%2)\n\t" + "fmulp %%st, %%st(1)\n\t" + "fildl %1\n\t" + "fildl (%0)\n\t" + "fmulp %%st, %%st(1)\n\t" + "fldt (%5)\n\t" + "fld %%st(2)\n\t" + "fmul %%st(1), %%st\n\t" + "fxch %%st(1)\n\t" + "fmul %%st(2), %%st\n\t" + "flds %6\n\t" + "fldl (%4)\n\t" + "fxch %%st(3)\n\t" + "fadd %%st(1), %%st\n\t" + "fxch %%st(2)\n\t" + "fadd %%st(1), %%st\n\t" + "fxch %%st(2)\n\t" + "fsub %%st(1), %%st\n\t" + "fxch %%st(2)\n\t" + "fsubp %%st, %%st(1)\n\t" + "fxch %%st(1)\n\t" + "fmul %%st(2), %%st\n\t" + "fxch %%st(1)\n\t" + "fmulp %%st, %%st(2)\n\t" + "fsubrp %%st, %%st(3)\n\t" + "fsubrp %%st, %%st(1)\n\t" + "fxch %%st(1)\n\t" + "fistpl (%2)\n\t" + "fistpl (%0)\n\t" + : : "r" (a0), "m" (b0), "r" (a1), "m" (b1), + "r" (dmod), "r" (dinvmod), + "m" (MPD_TWO63) + : "st", "memory" + ); +} +/* END PPRO GCC ASM */ +#elif defined(MASM) + +/* Return (a * b) % dmod */ +static inline mpd_uint_t __cdecl +ppro_mulmod(mpd_uint_t a, mpd_uint_t b, double *dmod, uint32_t *dinvmod) +{ + mpd_uint_t retval; + + __asm { + mov eax, dinvmod + mov edx, dmod + fild b + fild a + fmulp st(1), st + fld TBYTE PTR [eax] + fmul st, st(1) + fld MPD_TWO63 + fadd st(1), st + fsubp st(1), st + fld QWORD PTR [edx] + fmulp st(1), st + fsubp st(1), st + fistp retval + } + + return retval; +} + +/* + * Two modular multiplications in parallel: + * *a0 = (*a0 * w) % dmod + * *a1 = (*a1 * w) % dmod + */ +static inline mpd_uint_t __cdecl +ppro_mulmod2c(mpd_uint_t *a0, mpd_uint_t *a1, mpd_uint_t w, + double *dmod, uint32_t *dinvmod) +{ + __asm { + mov ecx, dmod + mov edx, a1 + mov ebx, dinvmod + mov eax, a0 + fild w + fild DWORD PTR [edx] + fmul st, st(1) + fxch st(1) + fild DWORD PTR [eax] + fmulp st(1), st + fld TBYTE PTR [ebx] + fld MPD_TWO63 + fld st(2) + fmul st, st(2) + fadd st, st(1) + fsub st, st(1) + fmul QWORD PTR [ecx] + fsubp st(3), st + fxch st(2) + fistp DWORD PTR [eax] + fmul st, st(2) + fadd st, st(1) + fsubrp st(1), st + fmul QWORD PTR [ecx] + fsubp st(1), st + fistp DWORD PTR [edx] + } +} + +/* + * Two modular multiplications in parallel: + * *a0 = (*a0 * b0) % dmod + * *a1 = (*a1 * b1) % dmod + */ +static inline void __cdecl +ppro_mulmod2(mpd_uint_t *a0, mpd_uint_t b0, mpd_uint_t *a1, mpd_uint_t b1, + double *dmod, uint32_t *dinvmod) +{ + __asm { + mov ecx, dmod + mov edx, a1 + mov ebx, dinvmod + mov eax, a0 + fild b1 + fild DWORD PTR [edx] + fmulp st(1), st + fild b0 + fild DWORD PTR [eax] + fmulp st(1), st + fld TBYTE PTR [ebx] + fld st(2) + fmul st, st(1) + fxch st(1) + fmul st, st(2) + fld DWORD PTR MPD_TWO63 + fld QWORD PTR [ecx] + fxch st(3) + fadd st, st(1) + fxch st(2) + fadd st, st(1) + fxch st(2) + fsub st, st(1) + fxch st(2) + fsubrp st(1), st + fxch st(1) + fmul st, st(2) + fxch st(1) + fmulp st(2), st + fsubp st(3), st + fsubp st(1), st + fxch st(1) + fistp DWORD PTR [edx] + fistp DWORD PTR [eax] + } +} +#endif /* PPRO MASM (_MSC_VER) */ + + +/* Return (base ** exp) % dmod */ +static inline mpd_uint_t +ppro_powmod(mpd_uint_t base, mpd_uint_t exp, double *dmod, uint32_t *dinvmod) +{ + mpd_uint_t r = 1; + + while (exp > 0) { + if (exp & 1) + r = ppro_mulmod(r, base, dmod, dinvmod); + base = ppro_mulmod(base, base, dmod, dinvmod); + exp >>= 1; + } + + return r; +} +#endif /* PPRO */ +#endif /* CONFIG_32 */ + + +#endif /* UMODARITH_H */ + + + diff --git a/Modules/_decimal/libmpdec/vccompat.h b/Modules/_decimal/libmpdec/vccompat.h new file mode 100644 index 0000000000..276e0372cb --- /dev/null +++ b/Modules/_decimal/libmpdec/vccompat.h @@ -0,0 +1,62 @@ +/* + * Copyright (c) 2008-2012 Stefan Krah. All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + + +#ifndef VCCOMPAT_H +#define VCCOMPAT_H + + +/* Visual C fixes: no stdint.h, no snprintf ... */ +#ifdef _MSC_VER + #include "vcstdint.h" + #undef inline + #define inline __inline + #undef random + #define random rand + #undef srandom + #define srandom srand + #undef snprintf + #define snprintf sprintf_s + #define HAVE_SNPRINTF + #undef strncasecmp + #define strncasecmp _strnicmp + #undef strcasecmp + #define strcasecmp _stricmp + #undef strtoll + #define strtoll _strtoi64 + #define strdup _strdup + #define PRIi64 "I64i" + #define PRIu64 "I64u" + #define PRIi32 "I32i" + #define PRIu32 "I32u" +#endif + + +#endif /* VCCOMPAT_H */ + + + diff --git a/Modules/_decimal/libmpdec/vcdiv64.asm b/Modules/_decimal/libmpdec/vcdiv64.asm new file mode 100644 index 0000000000..31bba08b0c --- /dev/null +++ b/Modules/_decimal/libmpdec/vcdiv64.asm @@ -0,0 +1,48 @@ +; +; Copyright (c) 2008-2012 Stefan Krah. All rights reserved. +; +; Redistribution and use in source and binary forms, with or without +; modification, are permitted provided that the following conditions +; are met: +; +; 1. Redistributions of source code must retain the above copyright +; notice, this list of conditions and the following disclaimer. +; +; 2. Redistributions in binary form must reproduce the above copyright +; notice, this list of conditions and the following disclaimer in the +; documentation and/or other materials provided with the distribution. +; +; THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS "AS IS" AND +; ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +; IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +; ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE +; FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL +; DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS +; OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) +; HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT +; LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY +; OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF +; SUCH DAMAGE. +; + + +PUBLIC _mpd_div_words +_TEXT SEGMENT +q$ = 8 +r$ = 16 +hi$ = 24 +lo$ = 32 +d$ = 40 +_mpd_div_words PROC + mov r10, rdx + mov rdx, r8 + mov rax, r9 + div QWORD PTR d$[rsp] + mov QWORD PTR [r10], rdx + mov QWORD PTR [rcx], rax + ret 0 +_mpd_div_words ENDP +_TEXT ENDS +END + + diff --git a/Modules/_decimal/libmpdec/vcstdint.h b/Modules/_decimal/libmpdec/vcstdint.h new file mode 100644 index 0000000000..e032ff1605 --- /dev/null +++ b/Modules/_decimal/libmpdec/vcstdint.h @@ -0,0 +1,232 @@ +// ISO C9x compliant stdint.h for Microsoft Visual Studio +// Based on ISO/IEC 9899:TC2 Committee draft (May 6, 2005) WG14/N1124 +// +// Copyright (c) 2006-2008 Alexander Chemeris +// +// Redistribution and use in source and binary forms, with or without +// modification, are permitted provided that the following conditions are met: +// +// 1. Redistributions of source code must retain the above copyright notice, +// this list of conditions and the following disclaimer. +// +// 2. Redistributions in binary form must reproduce the above copyright +// notice, this list of conditions and the following disclaimer in the +// documentation and/or other materials provided with the distribution. +// +// 3. The name of the author may be used to endorse or promote products +// derived from this software without specific prior written permission. +// +// THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED +// WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF +// MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO +// EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, +// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; +// OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, +// WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR +// OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF +// ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +// +/////////////////////////////////////////////////////////////////////////////// + +#ifndef _MSC_VER // [ +#error "Use this header only with Microsoft Visual C++ compilers!" +#endif // _MSC_VER ] + +#ifndef _MSC_STDINT_H_ // [ +#define _MSC_STDINT_H_ + +#if _MSC_VER > 1000 +#pragma once +#endif + +#include <limits.h> + +// For Visual Studio 6 in C++ mode wrap <wchar.h> include with 'extern "C++" {}' +// or compiler give many errors like this: +// error C2733: second C linkage of overloaded function 'wmemchr' not allowed +#if (_MSC_VER < 1300) && defined(__cplusplus) + extern "C++" { +#endif +# include <wchar.h> +#if (_MSC_VER < 1300) && defined(__cplusplus) + } +#endif + +// Define _W64 macros to mark types changing their size, like intptr_t. +#ifndef _W64 +# if !defined(__midl) && (defined(_X86_) || defined(_M_IX86)) && _MSC_VER >= 1300 +# define _W64 __w64 +# else +# define _W64 +# endif +#endif + + +// 7.18.1 Integer types + +// 7.18.1.1 Exact-width integer types +typedef __int8 int8_t; +typedef __int16 int16_t; +typedef __int32 int32_t; +typedef __int64 int64_t; +typedef unsigned __int8 uint8_t; +typedef unsigned __int16 uint16_t; +typedef unsigned __int32 uint32_t; +typedef unsigned __int64 uint64_t; + +// 7.18.1.2 Minimum-width integer types +typedef int8_t int_least8_t; +typedef int16_t int_least16_t; +typedef int32_t int_least32_t; +typedef int64_t int_least64_t; +typedef uint8_t uint_least8_t; +typedef uint16_t uint_least16_t; +typedef uint32_t uint_least32_t; +typedef uint64_t uint_least64_t; + +// 7.18.1.3 Fastest minimum-width integer types +typedef int8_t int_fast8_t; +typedef int16_t int_fast16_t; +typedef int32_t int_fast32_t; +typedef int64_t int_fast64_t; +typedef uint8_t uint_fast8_t; +typedef uint16_t uint_fast16_t; +typedef uint32_t uint_fast32_t; +typedef uint64_t uint_fast64_t; + +// 7.18.1.4 Integer types capable of holding object pointers +#ifdef _WIN64 // [ + typedef __int64 intptr_t; + typedef unsigned __int64 uintptr_t; +#else // _WIN64 ][ + typedef _W64 int intptr_t; + typedef _W64 unsigned int uintptr_t; +#endif // _WIN64 ] + +// 7.18.1.5 Greatest-width integer types +typedef int64_t intmax_t; +typedef uint64_t uintmax_t; + + +// 7.18.2 Limits of specified-width integer types + +#if !defined(__cplusplus) || defined(__STDC_LIMIT_MACROS) // [ See footnote 220 at page 257 and footnote 221 at page 259 + +// 7.18.2.1 Limits of exact-width integer types +#define INT8_MIN ((int8_t)_I8_MIN) +#define INT8_MAX _I8_MAX +#define INT16_MIN ((int16_t)_I16_MIN) +#define INT16_MAX _I16_MAX +#define INT32_MIN ((int32_t)_I32_MIN) +#define INT32_MAX _I32_MAX +#define INT64_MIN ((int64_t)_I64_MIN) +#define INT64_MAX _I64_MAX +#define UINT8_MAX _UI8_MAX +#define UINT16_MAX _UI16_MAX +#define UINT32_MAX _UI32_MAX +#define UINT64_MAX _UI64_MAX + +// 7.18.2.2 Limits of minimum-width integer types +#define INT_LEAST8_MIN INT8_MIN +#define INT_LEAST8_MAX INT8_MAX +#define INT_LEAST16_MIN INT16_MIN +#define INT_LEAST16_MAX INT16_MAX +#define INT_LEAST32_MIN INT32_MIN +#define INT_LEAST32_MAX INT32_MAX +#define INT_LEAST64_MIN INT64_MIN +#define INT_LEAST64_MAX INT64_MAX +#define UINT_LEAST8_MAX UINT8_MAX +#define UINT_LEAST16_MAX UINT16_MAX +#define UINT_LEAST32_MAX UINT32_MAX +#define UINT_LEAST64_MAX UINT64_MAX + +// 7.18.2.3 Limits of fastest minimum-width integer types +#define INT_FAST8_MIN INT8_MIN +#define INT_FAST8_MAX INT8_MAX +#define INT_FAST16_MIN INT16_MIN +#define INT_FAST16_MAX INT16_MAX +#define INT_FAST32_MIN INT32_MIN +#define INT_FAST32_MAX INT32_MAX +#define INT_FAST64_MIN INT64_MIN +#define INT_FAST64_MAX INT64_MAX +#define UINT_FAST8_MAX UINT8_MAX +#define UINT_FAST16_MAX UINT16_MAX +#define UINT_FAST32_MAX UINT32_MAX +#define UINT_FAST64_MAX UINT64_MAX + +// 7.18.2.4 Limits of integer types capable of holding object pointers +#ifdef _WIN64 // [ +# define INTPTR_MIN INT64_MIN +# define INTPTR_MAX INT64_MAX +# define UINTPTR_MAX UINT64_MAX +#else // _WIN64 ][ +# define INTPTR_MIN INT32_MIN +# define INTPTR_MAX INT32_MAX +# define UINTPTR_MAX UINT32_MAX +#endif // _WIN64 ] + +// 7.18.2.5 Limits of greatest-width integer types +#define INTMAX_MIN INT64_MIN +#define INTMAX_MAX INT64_MAX +#define UINTMAX_MAX UINT64_MAX + +// 7.18.3 Limits of other integer types + +#ifdef _WIN64 // [ +# define PTRDIFF_MIN _I64_MIN +# define PTRDIFF_MAX _I64_MAX +#else // _WIN64 ][ +# define PTRDIFF_MIN _I32_MIN +# define PTRDIFF_MAX _I32_MAX +#endif // _WIN64 ] + +#define SIG_ATOMIC_MIN INT_MIN +#define SIG_ATOMIC_MAX INT_MAX + +#ifndef SIZE_MAX // [ +# ifdef _WIN64 // [ +# define SIZE_MAX _UI64_MAX +# else // _WIN64 ][ +# define SIZE_MAX _UI32_MAX +# endif // _WIN64 ] +#endif // SIZE_MAX ] + +// WCHAR_MIN and WCHAR_MAX are also defined in <wchar.h> +#ifndef WCHAR_MIN // [ +# define WCHAR_MIN 0 +#endif // WCHAR_MIN ] +#ifndef WCHAR_MAX // [ +# define WCHAR_MAX _UI16_MAX +#endif // WCHAR_MAX ] + +#define WINT_MIN 0 +#define WINT_MAX _UI16_MAX + +#endif // __STDC_LIMIT_MACROS ] + + +// 7.18.4 Limits of other integer types + +#if !defined(__cplusplus) || defined(__STDC_CONSTANT_MACROS) // [ See footnote 224 at page 260 + +// 7.18.4.1 Macros for minimum-width integer constants + +#define INT8_C(val) val##i8 +#define INT16_C(val) val##i16 +#define INT32_C(val) val##i32 +#define INT64_C(val) val##i64 + +#define UINT8_C(val) val##ui8 +#define UINT16_C(val) val##ui16 +#define UINT32_C(val) val##ui32 +#define UINT64_C(val) val##ui64 + +// 7.18.4.2 Macros for greatest-width integer constants +#define INTMAX_C INT64_C +#define UINTMAX_C UINT64_C + +#endif // __STDC_CONSTANT_MACROS ] + + +#endif // _MSC_STDINT_H_ ] |