Provide and use various quad word primitives.

This includes bn_qwaddqw(), bn_qwsubqw(), bn_qwmulw_addw() and
bn_qwmulw_addqw_addw(). These can typically be optimised on architectures
that have a reasonable number of general purpose registers.

ok tb@

1 files changed, 6 insertions, 13 deletions

diff --git a/src/lib/libcrypto/bn/bn_mul.c b/src/lib/libcrypto/bn/bn_mul.c
index 118e8cddc5..65088cc5c4 100644
--- a/src/lib/libcrypto/bn/bn_mul.c
+++ b/src/lib/libcrypto/bn/bn_mul.c
@@ -1,4 +1,4 @@
-/* $OpenBSD: bn_mul.c,v 1.37 2023/04/19 10:51:22 jsing Exp $ */
+/* $OpenBSD: bn_mul.c,v 1.38 2023/06/12 16:17:24 jsing Exp $ */
 /* Copyright (C) 1995-1998 Eric Young (eay@cryptsoft.com)
  * All rights reserved.
  *
@@ -210,17 +210,13 @@ bn_mul_words(BN_ULONG *r, const BN_ULONG *a, int num, BN_ULONG w)
         if (num <= 0)
                 return 0;
 
-#ifndef OPENSSL_SMALL_FOOTPRINT
         while (num & ~3) {
-                bn_mulw_addw(a[0], w, carry, &carry, &r[0]);
-                bn_mulw_addw(a[1], w, carry, &carry, &r[1]);
-                bn_mulw_addw(a[2], w, carry, &carry, &r[2]);
-                bn_mulw_addw(a[3], w, carry, &carry, &r[3]);
+                bn_qwmulw_addw(a[3], a[2], a[1], a[0], w, carry, &carry,
+                    &r[3], &r[2], &r[1], &r[0]);
                 a += 4;
                 r += 4;
                 num -= 4;
         }
-#endif
         while (num) {
                 bn_mulw_addw(a[0], w, carry, &carry, &r[0]);
                 a++;
@@ -247,17 +243,14 @@ bn_mul_add_words(BN_ULONG *r, const BN_ULONG *a, int num, BN_ULONG w)
         if (num <= 0)
                 return 0;
 
-#ifndef OPENSSL_SMALL_FOOTPRINT
         while (num & ~3) {
-                bn_mulw_addw_addw(a[0], w, r[0], carry, &carry, &r[0]);
-                bn_mulw_addw_addw(a[1], w, r[1], carry, &carry, &r[1]);
-                bn_mulw_addw_addw(a[2], w, r[2], carry, &carry, &r[2]);
-                bn_mulw_addw_addw(a[3], w, r[3], carry, &carry, &r[3]);
+                bn_qwmulw_addqw_addw(a[3], a[2], a[1], a[0], w,
+                    r[3], r[2], r[1], r[0], carry, &carry,
+                    &r[3], &r[2], &r[1], &r[0]);
                 a += 4;
                 r += 4;
                 num -= 4;
         }
-#endif
         while (num) {
                 bn_mulw_addw_addw(a[0], w, r[0], carry, &carry, &r[0]);
                 a++;