Provide bn_umul_hilo().

The bignum code needs to be able to multiply two words, producing a
double word result. Some architectures do not have native support for
this, hence a pure C version is required. bn_umul_hilo() provides this
functionality.

There are currently two implementations, both of which are branch free.
The first uses bitwise operations for the carry, while the second uses
accumulators. The accumulator version uses fewer instructions, however
requires more variables/registers and seems to be slower, at least on
amd64/i386. The accumulator version may be faster on architectures that
have more registers available. Further testing can be performed and one
of the two implementations can be removed at a later date.

ok tb@

1 files changed, 159 insertions, 0 deletions

diff --git a/src/lib/libcrypto/bn/bn_internal.h b/src/lib/libcrypto/bn/bn_internal.h
new file mode 100644
index 0000000000..de5cd22424
--- /dev/null
+++ b/src/lib/libcrypto/bn/bn_internal.h
@@ -0,0 +1,159 @@
+/*      $OpenBSD: bn_internal.h,v 1.1 2023/01/31 05:48:39 jsing Exp $ */
+/*
+ * Copyright (c) 2023 Joel Sing <jsing@openbsd.org>
+ *
+ * Permission to use, copy, modify, and distribute this software for any
+ * purpose with or without fee is hereby granted, provided that the above
+ * copyright notice and this permission notice appear in all copies.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+ * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+ * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+ * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+ * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+ * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+ */
+
+#include <openssl/bn.h>
+
+#include "bn_arch.h"
+
+#ifndef HEADER_BN_INTERNAL_H
+#define HEADER_BN_INTERNAL_H
+
+#ifndef HAVE_BN_UMUL_HILO
+#ifdef BN_LLONG
+static inline void
+bn_umul_hilo(BN_ULONG a, BN_ULONG b, BN_ULONG *out_h, BN_ULONG *out_l)
+{
+        BN_ULLONG r;
+
+        r = (BN_ULLONG)a * (BN_ULLONG)b;
+
+        *out_h = r >> BN_BITS2;
+        *out_l = r & BN_MASK2;
+}
+
+#else /* !BN_LLONG */
+/*
+ * Multiply two words (a * b) producing a double word result (h:l).
+ *
+ * This can be rewritten as:
+ *
+ *  a * b = (hi32(a) * 2^32 + lo32(a)) * (hi32(b) * 2^32 + lo32(b))
+ *        = hi32(a) * hi32(b) * 2^64 +
+ *          hi32(a) * lo32(b) * 2^32 +
+ *          hi32(b) * lo32(a) * 2^32 +
+ *          lo32(a) * lo32(b)
+ *
+ * The multiplication for each part of a and b can be calculated for each of
+ * these four terms without overflowing a BN_ULONG, as the maximum value of a
+ * 32 bit x 32 bit multiplication is 32 + 32 = 64 bits. Once these
+ * multiplications have been performed the result can be partitioned and summed
+ * into a double word (h:l). The same applies on a 32 bit system, substituting
+ * 16 for 32 and 32 for 64.
+ */
+#if 1
+static inline void
+bn_umul_hilo(BN_ULONG a, BN_ULONG b, BN_ULONG *out_h, BN_ULONG *out_l)
+{
+        BN_ULONG ah, al, bh, bl, h, l, x, c1, c2;
+
+        ah = a >> BN_BITS4;
+        al = a & BN_MASK2l;
+        bh = b >> BN_BITS4;
+        bl = b & BN_MASK2l;
+
+        h = ah * bh;
+        l = al * bl;
+
+        /* (ah * bl) << BN_BITS4, partition the result across h:l with carry. */
+        x = ah * bl;
+        h += x >> BN_BITS4;
+        x <<= BN_BITS4;
+        c1 = l | x;
+        c2 = l & x;
+        l += x;
+        h += ((c1 & ~l) | c2) >> (BN_BITS2 - 1); /* carry */
+        
+        /* (bh * al) << BN_BITS4, partition the result across h:l with carry. */
+        x = bh * al;
+        h += x >> BN_BITS4;
+        x <<= BN_BITS4;
+        c1 = l | x;
+        c2 = l & x;
+        l += x;
+        h += ((c1 & ~l) | c2) >> (BN_BITS2 - 1); /* carry */
+
+        *out_h = h;
+        *out_l = l;
+}
+#else
+
+/*
+ * XXX - this accumulator based version uses fewer instructions, however
+ * requires more variables/registers. It seems to be slower on at least amd64
+ * and i386, however may be faster on other architectures that have more
+ * registers available. Further testing is required and one of the two
+ * implementations should eventually be removed.
+ */
+static inline void
+bn_umul_hilo(BN_ULONG a, BN_ULONG b, BN_ULONG *out_h, BN_ULONG *out_l)
+{
+        BN_ULONG ah, bh, al, bl, x, h, l;
+        BN_ULONG acc0, acc1, acc2, acc3;
+
+        ah = a >> BN_BITS4;
+        bh = b >> BN_BITS4;
+        al = a & BN_MASK2l;
+        bl = b & BN_MASK2l;
+
+        h = ah * bh;
+        l = al * bl;
+
+        acc0 = l & BN_MASK2l;
+        acc1 = l >> BN_BITS4;
+        acc2 = h & BN_MASK2l;
+        acc3 = h >> BN_BITS4;
+
+        /* (ah * bl) << BN_BITS4, partition the result across h:l. */
+        x = ah * bl;
+        acc1 += x & BN_MASK2l;
+        acc2 += (acc1 >> BN_BITS4) + (x >> BN_BITS4);
+        acc3 += acc2 >> BN_BITS4;
+
+        /* (bh * al) << BN_BITS4, partition the result across h:l. */
+        x = bh * al;
+        acc1 += x & BN_MASK2l;
+        acc2 += (acc1 >> BN_BITS4) + (x >> BN_BITS4);
+        acc3 += acc2 >> BN_BITS4;
+
+        *out_h = (acc3 << BN_BITS4) | acc2;
+        *out_l = (acc1 << BN_BITS4) | acc0;
+}
+#endif
+#endif /* !BN_LLONG */
+#endif
+
+#ifndef HAVE_BN_UMUL_LO
+static inline BN_ULONG
+bn_umul_lo(BN_ULONG a, BN_ULONG b)
+{
+        return a * b;
+}
+#endif
+
+#ifndef HAVE_BN_UMUL_HI
+static inline BN_ULONG
+bn_umul_hi(BN_ULONG a, BN_ULONG b)
+{
+        BN_ULONG h, l;
+
+        bn_umul_hilo(a, b, &h, &l);
+
+        return h;
+}
+#endif
+
+#endif