ec_cvt.c: ARM comparison results were wrong, clarify the background.

1 files changed, 12 insertions, 6 deletions

diff --git a/crypto/ec/ec_cvt.c b/crypto/ec/ec_cvt.c
index a99d762d3..dffd70521 100644
--- a/crypto/ec/ec_cvt.c
+++ b/crypto/ec/ec_cvt.c
@@ -85,15 +85,21 @@ EC_GROUP *EC_GROUP_new_curve_GFp(const BIGNUM *p, const BIGNUM *a, const BIGNUM
 	 * 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%
+	 * improvement on IA-64, ~25% 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...
+	 * bit keys respectively. Choice of these boundary values is
+	 * arguable, because the dependency of improvement coefficient
+	 * from key length is not a "monotone" curve. For example while
+	 * 571-bit result is 23% on ARM, 384-bit one is -1%. But it's
+	 * generally faster, sometimes "respectfully" faster, or
+	 * "tolerably" slower... What effectively happens is that loop
+	 * with bn_mul_add_words is put against bn_mul_mont, and the
+	 * 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();