/src/openssl34/crypto/bn/bn_mul.c

Source
/*
 * Copyright 1995-2018 The OpenSSL Project Authors. All Rights Reserved.
 *
 * Licensed under the Apache License 2.0 (the "License").  You may not use
 * this file except in compliance with the License.  You can obtain a copy
 * in the file LICENSE in the source distribution or at
 * https://www.openssl.org/source/license.html
 */

#include <assert.h>
#include "internal/cryptlib.h"
#include "bn_local.h"

#if defined(OPENSSL_NO_ASM) || !defined(OPENSSL_BN_ASM_PART_WORDS)
/*
 * Here follows specialised variants of bn_add_words() and bn_sub_words().
 * They have the property performing operations on arrays of different sizes.
 * The sizes of those arrays is expressed through cl, which is the common
 * length ( basically, min(len(a),len(b)) ), and dl, which is the delta
 * between the two lengths, calculated as len(a)-len(b). All lengths are the
 * number of BN_ULONGs...  For the operations that require a result array as
 * parameter, it must have the length cl+abs(dl). These functions should
 * probably end up in bn_asm.c as soon as there are assembler counterparts
 * for the systems that use assembler files.
 */

BN_ULONG bn_sub_part_words(BN_ULONG *r,
    const BN_ULONG *a, const BN_ULONG *b,
    int cl, int dl)
{
    BN_ULONG c, t;

    assert(cl >= 0);
    c = bn_sub_words(r, a, b, cl);

    if (dl == 0)
        return c;

    r += cl;
    a += cl;
    b += cl;

    if (dl < 0) {
        for (;;) {
            t = b[0];
            r[0] = (0 - t - c) & BN_MASK2;
            if (t != 0)
                c = 1;
            if (++dl >= 0)
                break;

            t = b[1];
            r[1] = (0 - t - c) & BN_MASK2;
            if (t != 0)
                c = 1;
            if (++dl >= 0)
                break;

            t = b[2];
            r[2] = (0 - t - c) & BN_MASK2;
            if (t != 0)
                c = 1;
            if (++dl >= 0)
                break;

            t = b[3];
            r[3] = (0 - t - c) & BN_MASK2;
            if (t != 0)
                c = 1;
            if (++dl >= 0)
                break;

            b += 4;
            r += 4;
        }
    } else {
        int save_dl = dl;
        while (c) {
            t = a[0];
            r[0] = (t - c) & BN_MASK2;
            if (t != 0)
                c = 0;
            if (--dl <= 0)
                break;

            t = a[1];
            r[1] = (t - c) & BN_MASK2;
            if (t != 0)
                c = 0;
            if (--dl <= 0)
                break;

            t = a[2];
            r[2] = (t - c) & BN_MASK2;
            if (t != 0)
                c = 0;
            if (--dl <= 0)
                break;

            t = a[3];
            r[3] = (t - c) & BN_MASK2;
            if (t != 0)
                c = 0;
            if (--dl <= 0)
                break;

            save_dl = dl;
            a += 4;
            r += 4;
        }
        if (dl > 0) {
            if (save_dl > dl) {
                switch (save_dl - dl) {
                case 1:
                    r[1] = a[1];
                    if (--dl <= 0)
                        break;
                    /* fall through */
                case 2:
                    r[2] = a[2];
                    if (--dl <= 0)
                        break;
                    /* fall through */
                case 3:
                    r[3] = a[3];
                    if (--dl <= 0)
                        break;
                }
                a += 4;
                r += 4;
            }
        }
        if (dl > 0) {
            for (;;) {
                r[0] = a[0];
                if (--dl <= 0)
                    break;
                r[1] = a[1];
                if (--dl <= 0)
                    break;
                r[2] = a[2];
                if (--dl <= 0)
                    break;
                r[3] = a[3];
                if (--dl <= 0)
                    break;

                a += 4;
                r += 4;
            }
        }
    }
    return c;
}
#endif

#ifdef BN_RECURSION
/*
 * Karatsuba recursive multiplication algorithm (cf. Knuth, The Art of
 * Computer Programming, Vol. 2)
 */

/*-
 * r is 2*n2 words in size,
 * a and b are both n2 words in size.
 * n2 must be a power of 2.
 * We multiply and return the result.
 * t must be 2*n2 words in size
 * We calculate
 * a[0]*b[0]
 * a[0]*b[0]+a[1]*b[1]+(a[0]-a[1])*(b[1]-b[0])
 * a[1]*b[1]
 */
/* dnX may not be positive, but n2/2+dnX has to be */
void bn_mul_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n2,
    int dna, int dnb, BN_ULONG *t)
{
    int n = n2 / 2, c1, c2;
    int tna = n + dna, tnb = n + dnb;
    unsigned int neg, zero;
    BN_ULONG ln, lo, *p;

#ifdef BN_MUL_COMBA
#if 0
    if (n2 == 4) {
        bn_mul_comba4(r, a, b);
        return;
    }
#endif
    /*
     * Only call bn_mul_comba 8 if n2 == 8 and the two arrays are complete
     * [steve]
     */
    if (n2 == 8 && dna == 0 && dnb == 0) {
        bn_mul_comba8(r, a, b);
        return;
    }
#endif /* BN_MUL_COMBA */
    /* Else do normal multiply */
    if (n2 < BN_MUL_RECURSIVE_SIZE_NORMAL) {
        bn_mul_normal(r, a, n2 + dna, b, n2 + dnb);
        if ((dna + dnb) < 0)
            memset(&r[2 * n2 + dna + dnb], 0,
                sizeof(BN_ULONG) * -(dna + dnb));
        return;
    }
    /* r=(a[0]-a[1])*(b[1]-b[0]) */
    c1 = bn_cmp_part_words(a, &(a[n]), tna, n - tna);
    c2 = bn_cmp_part_words(&(b[n]), b, tnb, tnb - n);
    zero = neg = 0;
    switch (c1 * 3 + c2) {
    case -4:
        bn_sub_part_words(t, &(a[n]), a, tna, tna - n); /* - */
        bn_sub_part_words(&(t[n]), b, &(b[n]), tnb, n - tnb); /* - */
        break;
    case -3:
        zero = 1;
        break;
    case -2:
        bn_sub_part_words(t, &(a[n]), a, tna, tna - n); /* - */
        bn_sub_part_words(&(t[n]), &(b[n]), b, tnb, tnb - n); /* + */
        neg = 1;
        break;
    case -1:
    case 0:
    case 1:
        zero = 1;
        break;
    case 2:
        bn_sub_part_words(t, a, &(a[n]), tna, n - tna); /* + */
        bn_sub_part_words(&(t[n]), b, &(b[n]), tnb, n - tnb); /* - */
        neg = 1;
        break;
    case 3:
        zero = 1;
        break;
    case 4:
        bn_sub_part_words(t, a, &(a[n]), tna, n - tna);
        bn_sub_part_words(&(t[n]), &(b[n]), b, tnb, tnb - n);
        break;
    }

#ifdef BN_MUL_COMBA
    if (n == 4 && dna == 0 && dnb == 0) { /* XXX: bn_mul_comba4 could take
                                           * extra args to do this well */
        if (!zero)
            bn_mul_comba4(&(t[n2]), t, &(t[n]));
        else
            memset(&t[n2], 0, sizeof(*t) * 8);

        bn_mul_comba4(r, a, b);
        bn_mul_comba4(&(r[n2]), &(a[n]), &(b[n]));
    } else if (n == 8 && dna == 0 && dnb == 0) { /* XXX: bn_mul_comba8 could
                                                  * take extra args to do
                                                  * this well */
        if (!zero)
            bn_mul_comba8(&(t[n2]), t, &(t[n]));
        else
            memset(&t[n2], 0, sizeof(*t) * 16);

        bn_mul_comba8(r, a, b);
        bn_mul_comba8(&(r[n2]), &(a[n]), &(b[n]));
    } else
#endif /* BN_MUL_COMBA */
    {
        p = &(t[n2 * 2]);
        if (!zero)
            bn_mul_recursive(&(t[n2]), t, &(t[n]), n, 0, 0, p);
        else
            memset(&t[n2], 0, sizeof(*t) * n2);
        bn_mul_recursive(r, a, b, n, 0, 0, p);
        bn_mul_recursive(&(r[n2]), &(a[n]), &(b[n]), n, dna, dnb, p);
    }

    /*-
     * t[32] holds (a[0]-a[1])*(b[1]-b[0]), c1 is the sign
     * r[10] holds (a[0]*b[0])
     * r[32] holds (b[1]*b[1])
     */

    c1 = (int)(bn_add_words(t, r, &(r[n2]), n2));

    if (neg) { /* if t[32] is negative */
        c1 -= (int)(bn_sub_words(&(t[n2]), t, &(t[n2]), n2));
    } else {
        /* Might have a carry */
        c1 += (int)(bn_add_words(&(t[n2]), &(t[n2]), t, n2));
    }

    /*-
     * t[32] holds (a[0]-a[1])*(b[1]-b[0])+(a[0]*b[0])+(a[1]*b[1])
     * r[10] holds (a[0]*b[0])
     * r[32] holds (b[1]*b[1])
     * c1 holds the carry bits
     */
    c1 += (int)(bn_add_words(&(r[n]), &(r[n]), &(t[n2]), n2));
    if (c1) {
        p = &(r[n + n2]);
        lo = *p;
        ln = (lo + c1) & BN_MASK2;
        *p = ln;

        /*
         * The overflow will stop before we over write words we should not
         * overwrite
         */
        if (ln < (BN_ULONG)c1) {
            do {
                p++;
                lo = *p;
                ln = (lo + 1) & BN_MASK2;
                *p = ln;
            } while (ln == 0);
        }
    }
}

/*
 * n+tn is the word length t needs to be n*4 is size, as does r
 */
/* tnX may not be negative but less than n */
void bn_mul_part_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n,
    int tna, int tnb, BN_ULONG *t)
{
    int i, j, n2 = n * 2;
    int c1, c2, neg;
    BN_ULONG ln, lo, *p;

    if (n < 8) {
        bn_mul_normal(r, a, n + tna, b, n + tnb);
        return;
    }

    /* r=(a[0]-a[1])*(b[1]-b[0]) */
    c1 = bn_cmp_part_words(a, &(a[n]), tna, n - tna);
    c2 = bn_cmp_part_words(&(b[n]), b, tnb, tnb - n);
    neg = 0;
    switch (c1 * 3 + c2) {
    case -4:
        bn_sub_part_words(t, &(a[n]), a, tna, tna - n); /* - */
        bn_sub_part_words(&(t[n]), b, &(b[n]), tnb, n - tnb); /* - */
        break;
    case -3:
    case -2:
        bn_sub_part_words(t, &(a[n]), a, tna, tna - n); /* - */
        bn_sub_part_words(&(t[n]), &(b[n]), b, tnb, tnb - n); /* + */
        neg = 1;
        break;
    case -1:
    case 0:
    case 1:
    case 2:
        bn_sub_part_words(t, a, &(a[n]), tna, n - tna); /* + */
        bn_sub_part_words(&(t[n]), b, &(b[n]), tnb, n - tnb); /* - */
        neg = 1;
        break;
    case 3:
    case 4:
        bn_sub_part_words(t, a, &(a[n]), tna, n - tna);
        bn_sub_part_words(&(t[n]), &(b[n]), b, tnb, tnb - n);
        break;
    }
    /*
     * The zero case isn't yet implemented here. The speedup would probably
     * be negligible.
     */
#if 0
    if (n == 4) {
        bn_mul_comba4(&(t[n2]), t, &(t[n]));
        bn_mul_comba4(r, a, b);
        bn_mul_normal(&(r[n2]), &(a[n]), tn, &(b[n]), tn);
        memset(&r[n2 + tn * 2], 0, sizeof(*r) * (n2 - tn * 2));
    } else
#endif
    if (n == 8) {
        bn_mul_comba8(&(t[n2]), t, &(t[n]));
        bn_mul_comba8(r, a, b);
        bn_mul_normal(&(r[n2]), &(a[n]), tna, &(b[n]), tnb);
        memset(&r[n2 + tna + tnb], 0, sizeof(*r) * (n2 - tna - tnb));
    } else {
        p = &(t[n2 * 2]);
        bn_mul_recursive(&(t[n2]), t, &(t[n]), n, 0, 0, p);
        bn_mul_recursive(r, a, b, n, 0, 0, p);
        i = n / 2;
        /*
         * If there is only a bottom half to the number, just do it
         */
        if (tna > tnb)
            j = tna - i;
        else
            j = tnb - i;
        if (j == 0) {
            bn_mul_recursive(&(r[n2]), &(a[n]), &(b[n]),
                i, tna - i, tnb - i, p);
            memset(&r[n2 + i * 2], 0, sizeof(*r) * (n2 - i * 2));
        } else if (j > 0) { /* eg, n == 16, i == 8 and tn == 11 */
            bn_mul_part_recursive(&(r[n2]), &(a[n]), &(b[n]),
                i, tna - i, tnb - i, p);
            memset(&(r[n2 + tna + tnb]), 0,
                sizeof(BN_ULONG) * (n2 - tna - tnb));
        } else { /* (j < 0) eg, n == 16, i == 8 and tn == 5 */

            memset(&r[n2], 0, sizeof(*r) * n2);
            if (tna < BN_MUL_RECURSIVE_SIZE_NORMAL
                && tnb < BN_MUL_RECURSIVE_SIZE_NORMAL) {
                bn_mul_normal(&(r[n2]), &(a[n]), tna, &(b[n]), tnb);
            } else {
                for (;;) {
                    i /= 2;
                    /*
                     * these simplified conditions work exclusively because
                     * difference between tna and tnb is 1 or 0
                     */
                    if (i < tna || i < tnb) {
                        bn_mul_part_recursive(&(r[n2]),
                            &(a[n]), &(b[n]),
                            i, tna - i, tnb - i, p);
                        break;
                    } else if (i == tna || i == tnb) {
                        bn_mul_recursive(&(r[n2]),
                            &(a[n]), &(b[n]),
                            i, tna - i, tnb - i, p);
                        break;
                    }
                }
            }
        }
    }

    /*-
     * t[32] holds (a[0]-a[1])*(b[1]-b[0]), c1 is the sign
     * r[10] holds (a[0]*b[0])
     * r[32] holds (b[1]*b[1])
     */

    c1 = (int)(bn_add_words(t, r, &(r[n2]), n2));

    if (neg) { /* if t[32] is negative */
        c1 -= (int)(bn_sub_words(&(t[n2]), t, &(t[n2]), n2));
    } else {
        /* Might have a carry */
        c1 += (int)(bn_add_words(&(t[n2]), &(t[n2]), t, n2));
    }

    /*-
     * t[32] holds (a[0]-a[1])*(b[1]-b[0])+(a[0]*b[0])+(a[1]*b[1])
     * r[10] holds (a[0]*b[0])
     * r[32] holds (b[1]*b[1])
     * c1 holds the carry bits
     */
    c1 += (int)(bn_add_words(&(r[n]), &(r[n]), &(t[n2]), n2));
    if (c1) {
        p = &(r[n + n2]);
        lo = *p;
        ln = (lo + c1) & BN_MASK2;
        *p = ln;

        /*
         * The overflow will stop before we over write words we should not
         * overwrite
         */
        if (ln < (BN_ULONG)c1) {
            do {
                p++;
                lo = *p;
                ln = (lo + 1) & BN_MASK2;
                *p = ln;
            } while (ln == 0);
        }
    }
}

/*-
 * a and b must be the same size, which is n2.
 * r needs to be n2 words and t needs to be n2*2
 */
void bn_mul_low_recursive(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n2,
    BN_ULONG *t)
{
    int n = n2 / 2;

    bn_mul_recursive(r, a, b, n, 0, 0, &(t[0]));
    if (n >= BN_MUL_LOW_RECURSIVE_SIZE_NORMAL) {
        bn_mul_low_recursive(&(t[0]), &(a[0]), &(b[n]), n, &(t[n2]));
        bn_add_words(&(r[n]), &(r[n]), &(t[0]), n);
        bn_mul_low_recursive(&(t[0]), &(a[n]), &(b[0]), n, &(t[n2]));
        bn_add_words(&(r[n]), &(r[n]), &(t[0]), n);
    } else {
        bn_mul_low_normal(&(t[0]), &(a[0]), &(b[n]), n);
        bn_mul_low_normal(&(t[n]), &(a[n]), &(b[0]), n);
        bn_add_words(&(r[n]), &(r[n]), &(t[0]), n);
        bn_add_words(&(r[n]), &(r[n]), &(t[n]), n);
    }
}
#endif /* BN_RECURSION */

int BN_mul(BIGNUM *r, const BIGNUM *a, const BIGNUM *b, BN_CTX *ctx)
{
    int ret = bn_mul_fixed_top(r, a, b, ctx);

    bn_correct_top(r);
    bn_check_top(r);

    return ret;
}

int bn_mul_fixed_top(BIGNUM *r, const BIGNUM *a, const BIGNUM *b, BN_CTX *ctx)
{
    int ret = 0;
    int top, al, bl;
    BIGNUM *rr;
#if defined(BN_MUL_COMBA) || defined(BN_RECURSION)
    int i;
#endif
#ifdef BN_RECURSION
    BIGNUM *t = NULL;
    int j = 0, k;
#endif

    bn_check_top(a);
    bn_check_top(b);
    bn_check_top(r);

    al = a->top;
    bl = b->top;

    if ((al == 0) || (bl == 0)) {
        BN_zero(r);
        return 1;
    }
    top = al + bl;

    BN_CTX_start(ctx);
    if ((r == a) || (r == b)) {
        if ((rr = BN_CTX_get(ctx)) == NULL)
            goto err;
    } else
        rr = r;

#if defined(BN_MUL_COMBA) || defined(BN_RECURSION)
    i = al - bl;
#endif
#ifdef BN_MUL_COMBA
    if (i == 0) {
#if 0
        if (al == 4) {
            if (bn_wexpand(rr, 8) == NULL)
                goto err;
            rr->top = 8;
            bn_mul_comba4(rr->d, a->d, b->d);
            goto end;
        }
#endif
        if (al == 8) {
            if (bn_wexpand(rr, 16) == NULL)
                goto err;
            rr->top = 16;
            bn_mul_comba8(rr->d, a->d, b->d);
            goto end;
        }
    }
#endif /* BN_MUL_COMBA */
#ifdef BN_RECURSION
    if ((al >= BN_MULL_SIZE_NORMAL) && (bl >= BN_MULL_SIZE_NORMAL)) {
        if (i >= -1 && i <= 1) {
            /*
             * Find out the power of two lower or equal to the longest of the
             * two numbers
             */
            if (i >= 0) {
                j = BN_num_bits_word((BN_ULONG)al);
            }
            if (i == -1) {
                j = BN_num_bits_word((BN_ULONG)bl);
            }
            j = 1 << (j - 1);
            assert(j <= al || j <= bl);
            k = j + j;
            t = BN_CTX_get(ctx);
            if (t == NULL)
                goto err;
            if (al > j || bl > j) {
                if (bn_wexpand(t, k * 4) == NULL)
                    goto err;
                if (bn_wexpand(rr, k * 4) == NULL)
                    goto err;
                bn_mul_part_recursive(rr->d, a->d, b->d,
                    j, al - j, bl - j, t->d);
            } else { /* al <= j || bl <= j */

                if (bn_wexpand(t, k * 2) == NULL)
                    goto err;
                if (bn_wexpand(rr, k * 2) == NULL)
                    goto err;
                bn_mul_recursive(rr->d, a->d, b->d, j, al - j, bl - j, t->d);
            }
            rr->top = top;
            goto end;
        }
    }
#endif /* BN_RECURSION */
    if (bn_wexpand(rr, top) == NULL)
        goto err;
    rr->top = top;
    bn_mul_normal(rr->d, a->d, al, b->d, bl);

#if defined(BN_MUL_COMBA) || defined(BN_RECURSION)
end:
#endif
    rr->neg = a->neg ^ b->neg;
    rr->flags |= BN_FLG_FIXED_TOP;
    if (r != rr && BN_copy(r, rr) == NULL)
        goto err;

    ret = 1;
err:
    bn_check_top(r);
    BN_CTX_end(ctx);
    return ret;
}

void bn_mul_normal(BN_ULONG *r, BN_ULONG *a, int na, BN_ULONG *b, int nb)
{
    BN_ULONG *rr;

    if (na < nb) {
        int itmp;
        BN_ULONG *ltmp;

        itmp = na;
        na = nb;
        nb = itmp;
        ltmp = a;
        a = b;
        b = ltmp;
    }
    rr = &(r[na]);
    if (nb <= 0) {
        (void)bn_mul_words(r, a, na, 0);
        return;
    } else
        rr[0] = bn_mul_words(r, a, na, b[0]);

    for (;;) {
        if (--nb <= 0)
            return;
        rr[1] = bn_mul_add_words(&(r[1]), a, na, b[1]);
        if (--nb <= 0)
            return;
        rr[2] = bn_mul_add_words(&(r[2]), a, na, b[2]);
        if (--nb <= 0)
            return;
        rr[3] = bn_mul_add_words(&(r[3]), a, na, b[3]);
        if (--nb <= 0)
            return;
        rr[4] = bn_mul_add_words(&(r[4]), a, na, b[4]);
        rr += 4;
        r += 4;
        b += 4;
    }
}

void bn_mul_low_normal(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n)
{
    bn_mul_words(r, a, n, b[0]);

    for (;;) {
        if (--n <= 0)
            return;
        bn_mul_add_words(&(r[1]), a, n, b[1]);
        if (--n <= 0)
            return;
        bn_mul_add_words(&(r[2]), a, n, b[2]);
        if (--n <= 0)
            return;
        bn_mul_add_words(&(r[3]), a, n, b[3]);
        if (--n <= 0)
            return;
        bn_mul_add_words(&(r[4]), a, n, b[4]);
        r += 4;
        b += 4;
    }
}

Coverage Report

Created: 2026-04-09 06:50

Line	Count	Source
1		/*
2		* Copyright 1995-2018 The OpenSSL Project Authors. All Rights Reserved.
3		*
4		* Licensed under the Apache License 2.0 (the "License"). You may not use
5		* this file except in compliance with the License. You can obtain a copy
6		* in the file LICENSE in the source distribution or at
7		* https://www.openssl.org/source/license.html
8		*/
9
10		#include <assert.h>
11		#include "internal/cryptlib.h"
12		#include "bn_local.h"
13
14		#if defined(OPENSSL_NO_ASM) \|\| !defined(OPENSSL_BN_ASM_PART_WORDS)
15		/*
16		* Here follows specialised variants of bn_add_words() and bn_sub_words().
17		* They have the property performing operations on arrays of different sizes.
18		* The sizes of those arrays is expressed through cl, which is the common
19		* length ( basically, min(len(a),len(b)) ), and dl, which is the delta
20		* between the two lengths, calculated as len(a)-len(b). All lengths are the
21		* number of BN_ULONGs... For the operations that require a result array as
22		* parameter, it must have the length cl+abs(dl). These functions should
23		* probably end up in bn_asm.c as soon as there are assembler counterparts
24		* for the systems that use assembler files.
25		*/
26
27		BN_ULONG bn_sub_part_words(BN_ULONG *r,
28		const BN_ULONG a, const BN_ULONG b,
29		int cl, int dl)
30	116M	{
31	116M	BN_ULONG c, t;
32
33	116M	assert(cl >= 0);
34	116M	c = bn_sub_words(r, a, b, cl);
35
36	116M	if (dl == 0)
37	113M	return c;
38
39	2.98M	r += cl;
40	2.98M	a += cl;
41	2.98M	b += cl;
42
43	2.98M	if (dl < 0) {
44	761k	for (;;) {
45	761k	t = b[0];
46	761k	r[0] = (0 - t - c) & BN_MASK2;
47	761k	if (t != 0)
48	0	c = 1;
49	761k	if (++dl >= 0)
50	53.3k	break;
51
52	707k	t = b[1];
53	707k	r[1] = (0 - t - c) & BN_MASK2;
54	707k	if (t != 0)
55	0	c = 1;
56	707k	if (++dl >= 0)
57	39.7k	break;
58
59	668k	t = b[2];
60	668k	r[2] = (0 - t - c) & BN_MASK2;
61	668k	if (t != 0)
62	0	c = 1;
63	668k	if (++dl >= 0)
64	157k	break;
65
66	510k	t = b[3];
67	510k	r[3] = (0 - t - c) & BN_MASK2;
68	510k	if (t != 0)
69	0	c = 1;
70	510k	if (++dl >= 0)
71	23.2k	break;
72
73	487k	b += 4;
74	487k	r += 4;
75	487k	}
76	2.71M	} else {
77	2.71M	int save_dl = dl;
78	3.26M	while (c) {
79	705k	t = a[0];
80	705k	r[0] = (t - c) & BN_MASK2;
81	705k	if (t != 0)
82	264k	c = 0;
83	705k	if (--dl <= 0)
84	52.2k	break;
85
86	653k	t = a[1];
87	653k	r[1] = (t - c) & BN_MASK2;
88	653k	if (t != 0)
89	215k	c = 0;
90	653k	if (--dl <= 0)
91	18.6k	break;
92
93	634k	t = a[2];
94	634k	r[2] = (t - c) & BN_MASK2;
95	634k	if (t != 0)
96	233k	c = 0;
97	634k	if (--dl <= 0)
98	66.6k	break;
99
100	568k	t = a[3];
101	568k	r[3] = (t - c) & BN_MASK2;
102	568k	if (t != 0)
103	198k	c = 0;
104	568k	if (--dl <= 0)
105	20.6k	break;
106
107	547k	save_dl = dl;
108	547k	a += 4;
109	547k	r += 4;
110	547k	}
111	2.71M	if (dl > 0) {
112	2.55M	if (save_dl > dl) {
113	0	switch (save_dl - dl) {
114	0	case 1:
115	0	r[1] = a[1];
116	0	if (--dl <= 0)
117	0	break;
118		/* fall through */
119	0	case 2:
120	0	r[2] = a[2];
121	0	if (--dl <= 0)
122	0	break;
123		/* fall through */
124	0	case 3:
125	0	r[3] = a[3];
126	0	if (--dl <= 0)
127	0	break;
128	0	}
129	0	a += 4;
130	0	r += 4;
131	0	}
132	2.55M	}
133	2.71M	if (dl > 0) {
134	18.4M	for (;;) {
135	18.4M	r[0] = a[0];
136	18.4M	if (--dl <= 0)
137	448k	break;
138	17.9M	r[1] = a[1];
139	17.9M	if (--dl <= 0)
140	307k	break;
141	17.6M	r[2] = a[2];
142	17.6M	if (--dl <= 0)
143	1.07M	break;
144	16.5M	r[3] = a[3];
145	16.5M	if (--dl <= 0)
146	730k	break;
147
148	15.8M	a += 4;
149	15.8M	r += 4;
150	15.8M	}
151	2.55M	}
152	2.71M	}
153	2.98M	return c;
154	2.98M	}
155		#endif
156
157		#ifdef BN_RECURSION
158		/*
159		* Karatsuba recursive multiplication algorithm (cf. Knuth, The Art of
160		* Computer Programming, Vol. 2)
161		*/
162
163		/*-
164		* r is 2*n2 words in size,
165		* a and b are both n2 words in size.
166		* n2 must be a power of 2.
167		* We multiply and return the result.
168		* t must be 2*n2 words in size
169		* We calculate
170		* a[0]*b[0]
171		* a[0]b[0]+a[1]b[1]+(a[0]-a[1])*(b[1]-b[0])
172		* a[1]*b[1]
173		*/
174		/* dnX may not be positive, but n2/2+dnX has to be */
175		void bn_mul_recursive(BN_ULONG r, BN_ULONG a, BN_ULONG *b, int n2,
176		int dna, int dnb, BN_ULONG *t)
177	60.8M	{
178	60.8M	int n = n2 / 2, c1, c2;
179	60.8M	int tna = n + dna, tnb = n + dnb;
180	60.8M	unsigned int neg, zero;
181	60.8M	BN_ULONG ln, lo, *p;
182
183	60.8M	#ifdef BN_MUL_COMBA
184		#if 0
185		if (n2 == 4) {
186		bn_mul_comba4(r, a, b);
187		return;
188		}
189		#endif
190		/*
191		* Only call bn_mul_comba 8 if n2 == 8 and the two arrays are complete
192		* [steve]
193		*/
194	60.8M	if (n2 == 8 && dna == 0 && dnb == 0) {
195	105k	bn_mul_comba8(r, a, b);
196	105k	return;
197	105k	}
198	60.7M	#endif /* BN_MUL_COMBA */
199		/* Else do normal multiply */
200	60.7M	if (n2 < BN_MUL_RECURSIVE_SIZE_NORMAL) {
201	54.8k	bn_mul_normal(r, a, n2 + dna, b, n2 + dnb);
202	54.8k	if ((dna + dnb) < 0)
203	54.8k	memset(&r[2 * n2 + dna + dnb], 0,
204	54.8k	sizeof(BN_ULONG) * -(dna + dnb));
205	54.8k	return;
206	54.8k	}
207		/* r=(a[0]-a[1])(b[1]-b[0]) /
208	60.6M	c1 = bn_cmp_part_words(a, &(a[n]), tna, n - tna);
209	60.6M	c2 = bn_cmp_part_words(&(b[n]), b, tnb, tnb - n);
210	60.6M	zero = neg = 0;
211	60.6M	switch (c1 * 3 + c2) {
212	11.7M	case -4:
213	11.7M	bn_sub_part_words(t, &(a[n]), a, tna, tna - n); /* - */
214	11.7M	bn_sub_part_words(&(t[n]), b, &(b[n]), tnb, n - tnb); /* - */
215	11.7M	break;
216	784k	case -3:
217	784k	zero = 1;
218	784k	break;
219	12.8M	case -2:
220	12.8M	bn_sub_part_words(t, &(a[n]), a, tna, tna - n); /* - */
221	12.8M	bn_sub_part_words(&(t[n]), &(b[n]), b, tnb, tnb - n); /* + */
222	12.8M	neg = 1;
223	12.8M	break;
224	205k	case -1:
225	1.88M	case 0:
226	2.08M	case 1:
227	2.08M	zero = 1;
228	2.08M	break;
229	20.5M	case 2:
230	20.5M	bn_sub_part_words(t, a, &(a[n]), tna, n - tna); /* + */
231	20.5M	bn_sub_part_words(&(t[n]), b, &(b[n]), tnb, n - tnb); /* - */
232	20.5M	neg = 1;
233	20.5M	break;
234	860k	case 3:
235	860k	zero = 1;
236	860k	break;
237	11.9M	case 4:
238	11.9M	bn_sub_part_words(t, a, &(a[n]), tna, n - tna);
239	11.9M	bn_sub_part_words(&(t[n]), &(b[n]), b, tnb, tnb - n);
240	11.9M	break;
241	60.6M	}
242
243	60.6M	#ifdef BN_MUL_COMBA
244	60.6M	if (n == 4 && dna == 0 && dnb == 0) { /* XXX: bn_mul_comba4 could take
245		* extra args to do this well */
246	0	if (!zero)
247	0	bn_mul_comba4(&(t[n2]), t, &(t[n]));
248	0	else
249	0	memset(&t[n2], 0, sizeof(t) 8);
250
251	0	bn_mul_comba4(r, a, b);
252	0	bn_mul_comba4(&(r[n2]), &(a[n]), &(b[n]));
253	60.6M	} else if (n == 8 && dna == 0 && dnb == 0) { /* XXX: bn_mul_comba8 could
254		* take extra args to do
255		* this well */
256	41.1M	if (!zero)
257	38.3M	bn_mul_comba8(&(t[n2]), t, &(t[n]));
258	2.73M	else
259	2.73M	memset(&t[n2], 0, sizeof(t) 16);
260
261	41.1M	bn_mul_comba8(r, a, b);
262	41.1M	bn_mul_comba8(&(r[n2]), &(a[n]), &(b[n]));
263	41.1M	} else
264	19.5M	#endif /* BN_MUL_COMBA */
265	19.5M	{
266	19.5M	p = &(t[n2 * 2]);
267	19.5M	if (!zero)
268	18.5M	bn_mul_recursive(&(t[n2]), t, &(t[n]), n, 0, 0, p);
269	989k	else
270	989k	memset(&t[n2], 0, sizeof(t) n2);
271	19.5M	bn_mul_recursive(r, a, b, n, 0, 0, p);
272	19.5M	bn_mul_recursive(&(r[n2]), &(a[n]), &(b[n]), n, dna, dnb, p);
273	19.5M	}
274
275		/*-
276		* t[32] holds (a[0]-a[1])*(b[1]-b[0]), c1 is the sign
277		* r[10] holds (a[0]*b[0])
278		* r[32] holds (b[1]*b[1])
279		*/
280
281	60.6M	c1 = (int)(bn_add_words(t, r, &(r[n2]), n2));
282
283	60.6M	if (neg) { /* if t[32] is negative */
284	33.3M	c1 -= (int)(bn_sub_words(&(t[n2]), t, &(t[n2]), n2));
285	33.3M	} else {
286		/* Might have a carry */
287	27.3M	c1 += (int)(bn_add_words(&(t[n2]), &(t[n2]), t, n2));
288	27.3M	}
289
290		/*-
291		* t[32] holds (a[0]-a[1])(b[1]-b[0])+(a[0]b[0])+(a[1]*b[1])
292		* r[10] holds (a[0]*b[0])
293		* r[32] holds (b[1]*b[1])
294		* c1 holds the carry bits
295		*/
296	60.6M	c1 += (int)(bn_add_words(&(r[n]), &(r[n]), &(t[n2]), n2));
297	60.6M	if (c1) {
298	23.3M	p = &(r[n + n2]);
299	23.3M	lo = *p;
300	23.3M	ln = (lo + c1) & BN_MASK2;
301	23.3M	*p = ln;
302
303		/*
304		* The overflow will stop before we over write words we should not
305		* overwrite
306		*/
307	23.3M	if (ln < (BN_ULONG)c1) {
308	866k	do {
309	866k	p++;
310	866k	lo = *p;
311	866k	ln = (lo + 1) & BN_MASK2;
312	866k	*p = ln;
313	866k	} while (ln == 0);
314	177k	}
315	23.3M	}
316	60.6M	}
317
318		/*
319		* n+tn is the word length t needs to be n*4 is size, as does r
320		*/
321		/* tnX may not be negative but less than n */
322		void bn_mul_part_recursive(BN_ULONG r, BN_ULONG a, BN_ULONG *b, int n,
323		int tna, int tnb, BN_ULONG *t)
324	1.45M	{
325	1.45M	int i, j, n2 = n * 2;
326	1.45M	int c1, c2, neg;
327	1.45M	BN_ULONG ln, lo, *p;
328
329	1.45M	if (n < 8) {
330	0	bn_mul_normal(r, a, n + tna, b, n + tnb);
331	0	return;
332	0	}
333
334		/* r=(a[0]-a[1])(b[1]-b[0]) /
335	1.45M	c1 = bn_cmp_part_words(a, &(a[n]), tna, n - tna);
336	1.45M	c2 = bn_cmp_part_words(&(b[n]), b, tnb, tnb - n);
337	1.45M	neg = 0;
338	1.45M	switch (c1 * 3 + c2) {
339	172k	case -4:
340	172k	bn_sub_part_words(t, &(a[n]), a, tna, tna - n); /* - */
341	172k	bn_sub_part_words(&(t[n]), b, &(b[n]), tnb, n - tnb); /* - */
342	172k	break;
343	5.23k	case -3:
344	24.3k	case -2:
345	24.3k	bn_sub_part_words(t, &(a[n]), a, tna, tna - n); /* - */
346	24.3k	bn_sub_part_words(&(t[n]), &(b[n]), b, tnb, tnb - n); /* + */
347	24.3k	neg = 1;
348	24.3k	break;
349	25.3k	case -1:
350	29.4k	case 0:
351	32.6k	case 1:
352	1.21M	case 2:
353	1.21M	bn_sub_part_words(t, a, &(a[n]), tna, n - tna); /* + */
354	1.21M	bn_sub_part_words(&(t[n]), b, &(b[n]), tnb, n - tnb); /* - */
355	1.21M	neg = 1;
356	1.21M	break;
357	19.8k	case 3:
358	46.4k	case 4:
359	46.4k	bn_sub_part_words(t, a, &(a[n]), tna, n - tna);
360	46.4k	bn_sub_part_words(&(t[n]), &(b[n]), b, tnb, tnb - n);
361	46.4k	break;
362	1.45M	}
363		/*
364		* The zero case isn't yet implemented here. The speedup would probably
365		* be negligible.
366		*/
367		#if 0
368		if (n == 4) {
369		bn_mul_comba4(&(t[n2]), t, &(t[n]));
370		bn_mul_comba4(r, a, b);
371		bn_mul_normal(&(r[n2]), &(a[n]), tn, &(b[n]), tn);
372		memset(&r[n2 + tn * 2], 0, sizeof(r) (n2 - tn * 2));
373		} else
374		#endif
375	1.45M	if (n == 8) {
376	179k	bn_mul_comba8(&(t[n2]), t, &(t[n]));
377	179k	bn_mul_comba8(r, a, b);
378	179k	bn_mul_normal(&(r[n2]), &(a[n]), tna, &(b[n]), tnb);
379	179k	memset(&r[n2 + tna + tnb], 0, sizeof(r) (n2 - tna - tnb));
380	1.27M	} else {
381	1.27M	p = &(t[n2 * 2]);
382	1.27M	bn_mul_recursive(&(t[n2]), t, &(t[n]), n, 0, 0, p);
383	1.27M	bn_mul_recursive(r, a, b, n, 0, 0, p);
384	1.27M	i = n / 2;
385		/*
386		* If there is only a bottom half to the number, just do it
387		*/
388	1.27M	if (tna > tnb)
389	95.4k	j = tna - i;
390	1.18M	else
391	1.18M	j = tnb - i;
392	1.27M	if (j == 0) {
393	47.7k	bn_mul_recursive(&(r[n2]), &(a[n]), &(b[n]),
394	47.7k	i, tna - i, tnb - i, p);
395	47.7k	memset(&r[n2 + i * 2], 0, sizeof(r) (n2 - i * 2));
396	1.22M	} else if (j > 0) { /* eg, n == 16, i == 8 and tn == 11 */
397	257k	bn_mul_part_recursive(&(r[n2]), &(a[n]), &(b[n]),
398	257k	i, tna - i, tnb - i, p);
399	257k	memset(&(r[n2 + tna + tnb]), 0,
400	257k	sizeof(BN_ULONG) * (n2 - tna - tnb));
401	972k	} else { /* (j < 0) eg, n == 16, i == 8 and tn == 5 */
402
403	972k	memset(&r[n2], 0, sizeof(r) n2);
404	972k	if (tna < BN_MUL_RECURSIVE_SIZE_NORMAL
405	835k	&& tnb < BN_MUL_RECURSIVE_SIZE_NORMAL) {
406	832k	bn_mul_normal(&(r[n2]), &(a[n]), tna, &(b[n]), tnb);
407	832k	} else {
408	158k	for (;;) {
409	158k	i /= 2;
410		/*
411		* these simplified conditions work exclusively because
412		* difference between tna and tnb is 1 or 0
413		*/
414	158k	if (i < tna \|\| i < tnb) {
415	36.0k	bn_mul_part_recursive(&(r[n2]),
416	36.0k	&(a[n]), &(b[n]),
417	36.0k	i, tna - i, tnb - i, p);
418	36.0k	break;
419	122k	} else if (i == tna \|\| i == tnb) {
420	103k	bn_mul_recursive(&(r[n2]),
421	103k	&(a[n]), &(b[n]),
422	103k	i, tna - i, tnb - i, p);
423	103k	break;
424	103k	}
425	158k	}
426	139k	}
427	972k	}
428	1.27M	}
429
430		/*-
431		* t[32] holds (a[0]-a[1])*(b[1]-b[0]), c1 is the sign
432		* r[10] holds (a[0]*b[0])
433		* r[32] holds (b[1]*b[1])
434		*/
435
436	1.45M	c1 = (int)(bn_add_words(t, r, &(r[n2]), n2));
437
438	1.45M	if (neg) { /* if t[32] is negative */
439	1.23M	c1 -= (int)(bn_sub_words(&(t[n2]), t, &(t[n2]), n2));
440	1.23M	} else {
441		/* Might have a carry */
442	218k	c1 += (int)(bn_add_words(&(t[n2]), &(t[n2]), t, n2));
443	218k	}
444
445		/*-
446		* t[32] holds (a[0]-a[1])(b[1]-b[0])+(a[0]b[0])+(a[1]*b[1])
447		* r[10] holds (a[0]*b[0])
448		* r[32] holds (b[1]*b[1])
449		* c1 holds the carry bits
450		*/
451	1.45M	c1 += (int)(bn_add_words(&(r[n]), &(r[n]), &(t[n2]), n2));
452	1.45M	if (c1) {
453	21.0k	p = &(r[n + n2]);
454	21.0k	lo = *p;
455	21.0k	ln = (lo + c1) & BN_MASK2;
456	21.0k	*p = ln;
457
458		/*
459		* The overflow will stop before we over write words we should not
460		* overwrite
461		*/
462	21.0k	if (ln < (BN_ULONG)c1) {
463	48.8k	do {
464	48.8k	p++;
465	48.8k	lo = *p;
466	48.8k	ln = (lo + 1) & BN_MASK2;
467	48.8k	*p = ln;
468	48.8k	} while (ln == 0);
469	10.0k	}
470	21.0k	}
471	1.45M	}
472
473		/*-
474		* a and b must be the same size, which is n2.
475		* r needs to be n2 words and t needs to be n2*2
476		*/
477		void bn_mul_low_recursive(BN_ULONG r, BN_ULONG a, BN_ULONG *b, int n2,
478		BN_ULONG *t)
479	0	{
480	0	int n = n2 / 2;
481
482	0	bn_mul_recursive(r, a, b, n, 0, 0, &(t[0]));
483	0	if (n >= BN_MUL_LOW_RECURSIVE_SIZE_NORMAL) {
484	0	bn_mul_low_recursive(&(t[0]), &(a[0]), &(b[n]), n, &(t[n2]));
485	0	bn_add_words(&(r[n]), &(r[n]), &(t[0]), n);
486	0	bn_mul_low_recursive(&(t[0]), &(a[n]), &(b[0]), n, &(t[n2]));
487	0	bn_add_words(&(r[n]), &(r[n]), &(t[0]), n);
488	0	} else {
489	0	bn_mul_low_normal(&(t[0]), &(a[0]), &(b[n]), n);
490	0	bn_mul_low_normal(&(t[n]), &(a[n]), &(b[0]), n);
491	0	bn_add_words(&(r[n]), &(r[n]), &(t[0]), n);
492	0	bn_add_words(&(r[n]), &(r[n]), &(t[n]), n);
493	0	}
494	0	}
495		#endif /* BN_RECURSION */
496
497		int BN_mul(BIGNUM r, const BIGNUM a, const BIGNUM b, BN_CTX ctx)
498	28.8M	{
499	28.8M	int ret = bn_mul_fixed_top(r, a, b, ctx);
500
501	28.8M	bn_correct_top(r);
502	28.8M	bn_check_top(r);
503
504	28.8M	return ret;
505	28.8M	}
506
507		int bn_mul_fixed_top(BIGNUM r, const BIGNUM a, const BIGNUM b, BN_CTX ctx)
508	33.7M	{
509	33.7M	int ret = 0;
510	33.7M	int top, al, bl;
511	33.7M	BIGNUM *rr;
512	33.7M	#if defined(BN_MUL_COMBA) \|\| defined(BN_RECURSION)
513	33.7M	int i;
514	33.7M	#endif
515	33.7M	#ifdef BN_RECURSION
516	33.7M	BIGNUM *t = NULL;
517	33.7M	int j = 0, k;
518	33.7M	#endif
519
520	33.7M	bn_check_top(a);
521	33.7M	bn_check_top(b);
522	33.7M	bn_check_top(r);
523
524	33.7M	al = a->top;
525	33.7M	bl = b->top;
526
527	33.7M	if ((al == 0) \|\| (bl == 0)) {
528	1.48M	BN_zero(r);
529	1.48M	return 1;
530	1.48M	}
531	32.2M	top = al + bl;
532
533	32.2M	BN_CTX_start(ctx);
534	32.2M	if ((r == a) \|\| (r == b)) {
535	181k	if ((rr = BN_CTX_get(ctx)) == NULL)
536	0	goto err;
537	181k	} else
538	32.0M	rr = r;
539
540	32.2M	#if defined(BN_MUL_COMBA) \|\| defined(BN_RECURSION)
541	32.2M	i = al - bl;
542	32.2M	#endif
543	32.2M	#ifdef BN_MUL_COMBA
544	32.2M	if (i == 0) {
545		#if 0
546		if (al == 4) {
547		if (bn_wexpand(rr, 8) == NULL)
548		goto err;
549		rr->top = 8;
550		bn_mul_comba4(rr->d, a->d, b->d);
551		goto end;
552		}
553		#endif
554	9.22M	if (al == 8) {
555	23.4k	if (bn_wexpand(rr, 16) == NULL)
556	0	goto err;
557	23.4k	rr->top = 16;
558	23.4k	bn_mul_comba8(rr->d, a->d, b->d);
559	23.4k	goto end;
560	23.4k	}
561	9.22M	}
562	32.2M	#endif /* BN_MUL_COMBA */
563	32.2M	#ifdef BN_RECURSION
564	32.2M	if ((al >= BN_MULL_SIZE_NORMAL) && (bl >= BN_MULL_SIZE_NORMAL)) {
565	1.64M	if (i >= -1 && i <= 1) {
566		/*
567		* Find out the power of two lower or equal to the longest of the
568		* two numbers
569		*/
570	1.58M	if (i >= 0) {
571	1.32M	j = BN_num_bits_word((BN_ULONG)al);
572	1.32M	}
573	1.58M	if (i == -1) {
574	256k	j = BN_num_bits_word((BN_ULONG)bl);
575	256k	}
576	1.58M	j = 1 << (j - 1);
577	1.58M	assert(j <= al \|\| j <= bl);
578	1.58M	k = j + j;
579	1.58M	t = BN_CTX_get(ctx);
580	1.58M	if (t == NULL)
581	0	goto err;
582	1.58M	if (al > j \|\| bl > j) {
583	1.16M	if (bn_wexpand(t, k * 4) == NULL)
584	0	goto err;
585	1.16M	if (bn_wexpand(rr, k * 4) == NULL)
586	0	goto err;
587	1.16M	bn_mul_part_recursive(rr->d, a->d, b->d,
588	1.16M	j, al - j, bl - j, t->d);
589	1.16M	} else { /* al <= j \|\| bl <= j */
590
591	418k	if (bn_wexpand(t, k * 2) == NULL)
592	0	goto err;
593	418k	if (bn_wexpand(rr, k * 2) == NULL)
594	0	goto err;
595	418k	bn_mul_recursive(rr->d, a->d, b->d, j, al - j, bl - j, t->d);
596	418k	}
597	1.58M	rr->top = top;
598	1.58M	goto end;
599	1.58M	}
600	1.64M	}
601	30.6M	#endif /* BN_RECURSION */
602	30.6M	if (bn_wexpand(rr, top) == NULL)
603	0	goto err;
604	30.6M	rr->top = top;
605	30.6M	bn_mul_normal(rr->d, a->d, al, b->d, bl);
606
607	30.6M	#if defined(BN_MUL_COMBA) \|\| defined(BN_RECURSION)
608	32.2M	end:
609	32.2M	#endif
610	32.2M	rr->neg = a->neg ^ b->neg;
611	32.2M	rr->flags \|= BN_FLG_FIXED_TOP;
612	32.2M	if (r != rr && BN_copy(r, rr) == NULL)
613	0	goto err;
614
615	32.2M	ret = 1;
616	32.2M	err:
617	32.2M	bn_check_top(r);
618	32.2M	BN_CTX_end(ctx);
619	32.2M	return ret;
620	32.2M	}
621
622		void bn_mul_normal(BN_ULONG r, BN_ULONG a, int na, BN_ULONG *b, int nb)
623	31.7M	{
624	31.7M	BN_ULONG *rr;
625
626	31.7M	if (na < nb) {
627	22.0M	int itmp;
628	22.0M	BN_ULONG *ltmp;
629
630	22.0M	itmp = na;
631	22.0M	na = nb;
632	22.0M	nb = itmp;
633	22.0M	ltmp = a;
634	22.0M	a = b;
635	22.0M	b = ltmp;
636	22.0M	}
637	31.7M	rr = &(r[na]);
638	31.7M	if (nb <= 0) {
639	238k	(void)bn_mul_words(r, a, na, 0);
640	238k	return;
641	238k	} else
642	31.4M	rr[0] = bn_mul_words(r, a, na, b[0]);
643
644	32.8M	for (;;) {
645	32.8M	if (--nb <= 0)
646	26.1M	return;
647	6.68M	rr[1] = bn_mul_add_words(&(r[1]), a, na, b[1]);
648	6.68M	if (--nb <= 0)
649	678k	return;
650	6.00M	rr[2] = bn_mul_add_words(&(r[2]), a, na, b[2]);
651	6.00M	if (--nb <= 0)
652	336k	return;
653	5.66M	rr[3] = bn_mul_add_words(&(r[3]), a, na, b[3]);
654	5.66M	if (--nb <= 0)
655	4.28M	return;
656	1.38M	rr[4] = bn_mul_add_words(&(r[4]), a, na, b[4]);
657	1.38M	rr += 4;
658	1.38M	r += 4;
659	1.38M	b += 4;
660	1.38M	}
661	31.4M	}
662
663		void bn_mul_low_normal(BN_ULONG r, BN_ULONG a, BN_ULONG *b, int n)
664	0	{
665	0	bn_mul_words(r, a, n, b[0]);
666
667	0	for (;;) {
668	0	if (--n <= 0)
669	0	return;
670	0	bn_mul_add_words(&(r[1]), a, n, b[1]);
671	0	if (--n <= 0)
672	0	return;
673	0	bn_mul_add_words(&(r[2]), a, n, b[2]);
674	0	if (--n <= 0)
675	0	return;
676	0	bn_mul_add_words(&(r[3]), a, n, b[3]);
677	0	if (--n <= 0)
678	0	return;
679	0	bn_mul_add_words(&(r[4]), a, n, b[4]);
680	0	r += 4;
681	0	b += 4;
682	0	}
683	0	}