ec_cvt.c: avoid EC_GFp_nist_method on platforms with bn_mul_mont [see

commentary for details].

diff --git a/crypto/ec/ec_cvt.c b/crypto/ec/ec_cvt.c
index 0c3b93a43bb3635d7159f53288838c7ada341fd0..a99d762d3b93ad67febb6ea9a3b531f035599860 100644 (file)
--- a/crypto/ec/ec_cvt.c
+++ b/crypto/ec/ec_cvt.c
@@ -80,10 +80,29 @@ EC_GROUP *EC_GROUP_new_curve_GFp(const BIGNUM *p, const BIGNUM *a, const BIGNUM
        const EC_METHOD *meth;
        EC_GROUP *ret;
 
+#if defined(OPENSSL_BN_ASM_MONT) && !defined(__sparc)
+       /*
+        * This might appear controversial, but the fact is that generic
+        * prime method was observed to deliver better performance even
+        * for NIST primes on a range of platforms, e.g.: 60%-15%
+        * improvement on IA-64, 50%-20% on ARM, 30%-90% on P4, 20%-25%
+        * in 32-bit build and 35%--12% in 64-bit build on Core2...
+        * Coefficients are relative to optimized bn_nist.c for most
+        * intensive ECDSA verify and ECDH operations for 192- and 521-
+        * bit keys respectively. What effectively happens is that loop
+        * with bn_mul_add_words is put against bn_mul_mont, and latter
+        * wins on short vectors. Correct solution should be implementing
+        * dedicated NxN multiplication subroutines for small N. But till
+        * it materializes, let's stick to generic prime method...
+        *                                              <appro>
+        */
+       meth = EC_GFp_mont_method();
+#else
        if (BN_nist_mod_func(p))
                meth = EC_GFp_nist_method();
        else
                meth = EC_GFp_mont_method();
+#endif
        
        ret = EC_GROUP_new(meth);
        if (ret == NULL)