OpenSolaris

The following changes in big_mont_mul() Montogomery multiplication routine) may lead to better performance.

Currently (snv 111):
for (i = 0; i < nlen; i++) {
                digit = rr[i] * n0;
                //digit = digit * n0;

                c = BIG_MUL_ADD_VEC(rr + i, nn, nlen, digit);
                j = i + nlen;
                rr[j] += c;
                while (rr[j] < c) {
                        rr[j++ + 1] += 1;
                        //j++;
                        c = 1;
                }
        } 


Suggested:
        BIG_CHUNK_TYPE  c[BIGTMPSIZE];
        for (i = 0; i < nlen; i++) {
                //j = i + nlen;
                temp = rr+i;
                digit = *temp * n0;
                c[i] = BIG_MUL_ADD_VEC(temp, nn, nlen, digit);
        }

        for (i = 0; i < nlen; i++) {
                j = i + nlen;
                rr[j] += c[i];
                while (rr[j] < c[i]) {
                        rr[j++ + 1] += 1;
                        //j++;
                        c[i] = 1;
                }
        } 

This change reduces the dependency between the computation of c (with big_mul_add_vec) and adding the carryover bits, thus improving pipelining.

With the suggested code, the -fast compiler option in SS12 was found to give better performance.
null

N/A

N/A