int M_get_stack_ptr(int sz)
{
int i, k;
UCHAR *cp;
k = ++M_mul_stack_ptr;
/* if size is 0, just need to malloc and return */
if (mul_stack_data_size[k] == 0)
{
if ((i = sz) < 16)
i = 16;
if ((cp = (UCHAR *)MAPM_MALLOC(i + 4)) == NULL)
{
/* fatal, this does not return */
M_apm_log_error_msg(M_APM_FATAL, M_stack_ptr_error_msg);
}
mul_stack_data[k] = cp;
mul_stack_data_size[k] = i;
}
else /* it has been malloc'ed, see if it's big enough */
{
if (sz > mul_stack_data_size[k])
{
cp = mul_stack_data[k];
if ((cp = (UCHAR *)MAPM_REALLOC(cp, (sz + 4))) == NULL)
{
/* fatal, this does not return */
M_apm_log_error_msg(M_APM_FATAL, M_stack_ptr_error_msg);
}
mul_stack_data[k] = cp;
mul_stack_data_size[k] = sz;
}
}
return(k);
}
void m_apm_multiply(M_APM r, M_APM a, M_APM b)
{
int ai, itmp, sign, nexp, ii, jj, indexa, indexb, index0, numdigits;
UCHAR *cp, *cpr, *cp_div, *cp_rem;
void *vp;
sign = a->m_apm_sign * b->m_apm_sign;
nexp = a->m_apm_exponent + b->m_apm_exponent;
if (sign == 0) /* one number is zero, result is zero */
{
M_set_to_zero(r);
return;
}
numdigits = a->m_apm_datalength + b->m_apm_datalength;
indexa = (a->m_apm_datalength + 1) >> 1;
indexb = (b->m_apm_datalength + 1) >> 1;
/*
* If we are multiplying 2 'big' numbers, use the fast algorithm.
*
* This is a **very** approx break even point between this algorithm
* and the FFT multiply. Note that different CPU's, operating systems,
* and compiler's may yield a different break even point. This point
* (~96 decimal digits) is how the test came out on the author's system.
*/
if (indexa >= 48 && indexb >= 48)
{
M_fast_multiply(r, a, b);
return;
}
ii = (numdigits + 1) >> 1; /* required size of result, in bytes */
if (ii > r->m_apm_malloclength)
{
if ((vp = MAPM_REALLOC(r->m_apm_data, (ii + 32))) == NULL)
{
/* fatal, this does not return */
M_apm_log_error_msg(M_APM_FATAL, "\'m_apm_multiply\', Out of memory");
}
r->m_apm_malloclength = ii + 28;
r->m_apm_data = (UCHAR *)vp;
}
M_get_div_rem_addr(&cp_div, &cp_rem);
index0 = indexa + indexb;
cp = r->m_apm_data;
memset(cp, 0, index0);
ii = indexa;
while (TRUE)
{
index0--;
cpr = cp + index0;
jj = indexb;
ai = (int)a->m_apm_data[--ii];
while (TRUE)
{
itmp = ai * b->m_apm_data[--jj];
*(cpr-1) += cp_div[itmp];
*cpr += cp_rem[itmp];
if (*cpr >= 100)
{
*cpr -= 100;
*(cpr-1) += 1;
}
cpr--;
if (*cpr >= 100)
{
*cpr -= 100;
*(cpr-1) += 1;
}
if (jj == 0)
break;
}
if (ii == 0)
break;
}
r->m_apm_sign = sign;
r->m_apm_exponent = nexp;
r->m_apm_datalength = numdigits;
M_apm_normalize(r);
}
void M_fast_multiply(M_APM rr, M_APM aa, M_APM bb)
{
void *vp;
int ii, k, nexp, sign;
if (M_firsttimef)
{
M_firsttimef = FALSE;
for (k=0; k < M_STACK_SIZE; k++)
mul_stack_data_size[k] = 0;
size_flag = M_get_sizeof_int();
bit_limit = 8 * size_flag + 1;
M_ain = m_apm_init();
M_bin = m_apm_init();
}
exp_stack_ptr = -1;
M_mul_stack_ptr = -1;
m_apm_copy(M_ain, aa);
m_apm_copy(M_bin, bb);
sign = M_ain->m_apm_sign * M_bin->m_apm_sign;
nexp = M_ain->m_apm_exponent + M_bin->m_apm_exponent;
if (M_ain->m_apm_datalength >= M_bin->m_apm_datalength)
ii = M_ain->m_apm_datalength;
else
ii = M_bin->m_apm_datalength;
ii = (ii + 1) >> 1;
ii = M_next_power_of_2(ii);
/* Note: 'ii' must be >= 4 here. this is guaranteed
by the caller: m_apm_multiply
*/
k = 2 * ii; /* required size of result, in bytes */
M_apm_pad(M_ain, k); /* fill out the data so the number of */
M_apm_pad(M_bin, k); /* bytes is an exact power of 2 */
if (k > rr->m_apm_malloclength)
{
if ((vp = MAPM_REALLOC(rr->m_apm_data, (k + 32))) == NULL)
{
/* fatal, this does not return */
M_apm_log_error_msg(M_APM_FATAL, "\'M_fast_multiply\', Out of memory");
}
rr->m_apm_malloclength = k + 28;
rr->m_apm_data = (UCHAR *)vp;
}
#ifdef NO_FFT_MULTIPLY
M_fmul_div_conq(rr->m_apm_data, M_ain->m_apm_data,
M_bin->m_apm_data, ii);
#else
/*
* if the numbers are *really* big, use the divide-and-conquer
* routine first until the numbers are small enough to be handled
* by the FFT algorithm. If the numbers are already small enough,
* call the FFT multiplication now.
*
* Note that 'ii' here is (and must be) an exact power of 2.
*/
if (size_flag == 2) /* if still using 16 bit compilers .... */
{
M_fast_mul_fft(rr->m_apm_data, M_ain->m_apm_data,
M_bin->m_apm_data, ii);
}
else /* >= 32 bit compilers */
{
if (ii > (MAX_FFT_BYTES + 2))
{
M_fmul_div_conq(rr->m_apm_data, M_ain->m_apm_data,
M_bin->m_apm_data, ii);
}
else
{
M_fast_mul_fft(rr->m_apm_data, M_ain->m_apm_data,
M_bin->m_apm_data, ii);
}
}
#endif
rr->m_apm_sign = sign;
rr->m_apm_exponent = nexp;
rr->m_apm_datalength = 4 * ii;
M_apm_normalize(rr);
}
void M_apm_sdivide(M_APM r, int places, M_APM a, M_APM b)
{
int j, k, m, b0, sign, nexp, indexr, icompare, iterations;
long trial_numer;
void *vp;
if (M_div_firsttime)
{
M_div_firsttime = FALSE;
M_div_worka = m_apm_init();
M_div_workb = m_apm_init();
M_div_tmp7 = m_apm_init();
M_div_tmp8 = m_apm_init();
M_div_tmp9 = m_apm_init();
}
sign = a->m_apm_sign * b->m_apm_sign;
if (sign == 0) /* one number is zero, result is zero */
{
if (b->m_apm_sign == 0)
{
M_apm_log_error_msg(M_APM_RETURN, "\'M_apm_sdivide\', Divide by 0");
}
M_set_to_zero(r);
return;
}
/*
* Knuth step D1. Since base = 100, base / 2 = 50.
* (also make the working copies positive)
*/
if (b->m_apm_data[0] >= 50)
{
m_apm_absolute_value(M_div_worka, a);
m_apm_absolute_value(M_div_workb, b);
}
else /* 'normal' step D1 */
{
k = 100 / (b->m_apm_data[0] + 1);
m_apm_set_long(M_div_tmp9, (long)k);
m_apm_multiply(M_div_worka, M_div_tmp9, a);
m_apm_multiply(M_div_workb, M_div_tmp9, b);
M_div_worka->m_apm_sign = 1;
M_div_workb->m_apm_sign = 1;
}
/* setup trial denominator for step D3 */
b0 = 100 * (int)M_div_workb->m_apm_data[0];
if (M_div_workb->m_apm_datalength >= 3)
b0 += M_div_workb->m_apm_data[1];
nexp = M_div_worka->m_apm_exponent - M_div_workb->m_apm_exponent;
if (nexp > 0)
iterations = nexp + places + 1;
else
iterations = places + 1;
k = (iterations + 1) >> 1; /* required size of result, in bytes */
if (k > r->m_apm_malloclength)
{
if ((vp = MAPM_REALLOC(r->m_apm_data, (k + 32))) == NULL)
{
/* fatal, this does not return */
M_apm_log_error_msg(M_APM_FATAL, "\'M_apm_sdivide\', Out of memory");
}
r->m_apm_malloclength = k + 28;
r->m_apm_data = (UCHAR *)vp;
}
/* clear the exponent in the working copies */
M_div_worka->m_apm_exponent = 0;
M_div_workb->m_apm_exponent = 0;
/* if numbers are equal, ratio == 1.00000... */
if ((icompare = m_apm_compare(M_div_worka, M_div_workb)) == 0)
{
iterations = 1;
r->m_apm_data[0] = 10;
nexp++;
}
else /* ratio not 1, do the real division */
{
if (icompare == 1) /* numerator > denominator */
{
nexp++; /* to adjust the final exponent */
M_div_worka->m_apm_exponent += 1; /* multiply numerator by 10 */
}
else /* numerator < denominator */
{
M_div_worka->m_apm_exponent += 2; /* multiply numerator by 100 */
}
indexr = 0;
m = 0;
while (TRUE)
{
/*
* Knuth step D3. Only use the 3rd -> 6th digits if the number
* actually has that many digits.
*/
trial_numer = 10000L * (long)M_div_worka->m_apm_data[0];
if (M_div_worka->m_apm_datalength >= 5)
{
trial_numer += 100 * M_div_worka->m_apm_data[1]
+ M_div_worka->m_apm_data[2];
}
else
{
if (M_div_worka->m_apm_datalength >= 3)
trial_numer += 100 * M_div_worka->m_apm_data[1];
}
j = (int)(trial_numer / b0);
/*
* Since the library 'normalizes' all the results, we need
* to look at the exponent of the number to decide if we
* have a lead in 0n or 00.
*/
if ((k = 2 - M_div_worka->m_apm_exponent) > 0)
{
while (TRUE)
{
j /= 10;
if (--k == 0)
break;
}
}
if (j == 100) /* qhat == base ?? */
j = 99; /* if so, decrease by 1 */
m_apm_set_long(M_div_tmp8, (long)j);
m_apm_multiply(M_div_tmp7, M_div_tmp8, M_div_workb);
/*
* Compare our q-hat (j) against the desired number.
* j is either correct, 1 too large, or 2 too large
* per Theorem B on pg 272 of Art of Compter Programming,
* Volume 2, 3rd Edition.
*
* The above statement is only true if using the 2 leading
* digits of the numerator and the leading digit of the
* denominator. Since we are using the (3) leading digits
* of the numerator and the (2) leading digits of the
* denominator, we eliminate the case where our q-hat is
* 2 too large, (and q-hat being 1 too large is quite remote).
*/
if (m_apm_compare(M_div_tmp7, M_div_worka) == 1)
{
j--;
m_apm_subtract(M_div_tmp8, M_div_tmp7, M_div_workb);
m_apm_copy(M_div_tmp7, M_div_tmp8);
}
/*
* Since we know q-hat is correct, step D6 is unnecessary.
*
* Store q-hat, step D5. Since D6 is unnecessary, we can
* do D5 before D4 and decide if we are done.
*/
r->m_apm_data[indexr++] = (UCHAR)j; /* j == 'qhat' */
m += 2;
if (m >= iterations)
break;
/* step D4 */
m_apm_subtract(M_div_tmp9, M_div_worka, M_div_tmp7);
/*
* if the subtraction yields zero, the division is exact
* and we are done early.
*/
if (M_div_tmp9->m_apm_sign == 0)
{
iterations = m;
break;
}
/* multiply by 100 and re-save */
M_div_tmp9->m_apm_exponent += 2;
m_apm_copy(M_div_worka, M_div_tmp9);
}
}
r->m_apm_sign = sign;
r->m_apm_exponent = nexp;
r->m_apm_datalength = iterations;
M_apm_normalize(r);
}
