Provide inline assembly versions of bn_umul_hilo() for aarch64/amd64/i386.

ok tb@

3 files changed, 67 insertions, 3 deletions

diff --git a/src/lib/libcrypto/bn/arch/aarch64/bn_arch.h b/src/lib/libcrypto/bn/arch/aarch64/bn_arch.h
index 136adf0e97..5cf25adc48 100644
--- a/src/lib/libcrypto/bn/arch/aarch64/bn_arch.h
+++ b/src/lib/libcrypto/bn/arch/aarch64/bn_arch.h
@@ -1,4 +1,4 @@
-/*      $OpenBSD: bn_arch.h,v 1.1 2023/01/20 10:04:33 jsing Exp $ */
+/*      $OpenBSD: bn_arch.h,v 1.2 2023/01/31 05:53:49 jsing Exp $ */
 /*
  * Copyright (c) 2023 Joel Sing <jsing@openbsd.org>
  *
@@ -15,10 +15,30 @@
  * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
  */
 
+#include <openssl/bn.h>
+
 #ifndef HEADER_BN_ARCH_H
 #define HEADER_BN_ARCH_H
 
 #ifndef OPENSSL_NO_ASM
 
+#if defined(__GNUC__)
+#define HAVE_BN_UMUL_HILO
+
+static inline void
+bn_umul_hilo(BN_ULONG a, BN_ULONG b, BN_ULONG *out_h, BN_ULONG *out_l)
+{
+        BN_ULONG h, l;
+
+        /* Unsigned multiplication using a umulh/mul pair. */
+        __asm__ ("umulh %0, %2, %3; mul %1, %2, %3"
+            : "=r"(h), "=r"(l)
+            : "r"(a), "r"(b));
+
+        *out_h = h;
+        *out_l = l;
+}
+#endif /* __GNUC__ */
+
 #endif
 #endif
diff --git a/src/lib/libcrypto/bn/arch/amd64/bn_arch.h b/src/lib/libcrypto/bn/arch/amd64/bn_arch.h
index 6b7eaf5eee..9e4b6b9442 100644
--- a/src/lib/libcrypto/bn/arch/amd64/bn_arch.h
+++ b/src/lib/libcrypto/bn/arch/amd64/bn_arch.h
@@ -1,4 +1,4 @@
-/*      $OpenBSD: bn_arch.h,v 1.8 2023/01/28 16:33:34 jsing Exp $ */
+/*      $OpenBSD: bn_arch.h,v 1.9 2023/01/31 05:53:49 jsing Exp $ */
 /*
  * Copyright (c) 2023 Joel Sing <jsing@openbsd.org>
  *
@@ -61,5 +61,27 @@ bn_div_rem_words_inline(BN_ULONG h, BN_ULONG l, BN_ULONG d, BN_ULONG *out_q,
 }
 #endif /* __GNUC__ */
 
+#if defined(__GNUC__)
+#define HAVE_BN_UMUL_HILO
+
+static inline void
+bn_umul_hilo(BN_ULONG a, BN_ULONG b, BN_ULONG *out_h, BN_ULONG *out_l)
+{
+        BN_ULONG h, l;
+
+        /*
+         * Unsigned multiplication of %rax, with the double word result being
+         * stored in %rdx:%rax.
+         */
+        __asm__ ("mulq %3"
+            : "=d"(h), "=a"(l)
+            : "a"(a), "rm"(b)
+            : "cc");
+
+        *out_h = h;
+        *out_l = l;
+}
+#endif /* __GNUC__ */
+
 #endif
 #endif
diff --git a/src/lib/libcrypto/bn/arch/i386/bn_arch.h b/src/lib/libcrypto/bn/arch/i386/bn_arch.h
index e2b4957efc..268c51e41a 100644
--- a/src/lib/libcrypto/bn/arch/i386/bn_arch.h
+++ b/src/lib/libcrypto/bn/arch/i386/bn_arch.h
@@ -1,4 +1,4 @@
-/*      $OpenBSD: bn_arch.h,v 1.7 2023/01/28 16:33:34 jsing Exp $ */
+/*      $OpenBSD: bn_arch.h,v 1.8 2023/01/31 05:53:49 jsing Exp $ */
 /*
  * Copyright (c) 2023 Joel Sing <jsing@openbsd.org>
  *
@@ -60,5 +60,27 @@ bn_div_rem_words_inline(BN_ULONG h, BN_ULONG l, BN_ULONG d, BN_ULONG *out_q,
 }
 #endif /* __GNUC__ */
 
+#if defined(__GNUC__)
+#define HAVE_BN_UMUL_HILO
+
+static inline void
+bn_umul_hilo(BN_ULONG a, BN_ULONG b, BN_ULONG *out_h, BN_ULONG *out_l)
+{
+        BN_ULONG h, l;
+
+        /*
+         * Unsigned multiplication of %eax, with the double word result being
+         * stored in %edx:%eax.
+         */
+        __asm__ ("mull %3"
+            : "=d"(h), "=a"(l)
+            : "a"(a), "rm"(b)
+            : "cc");
+
+        *out_h = h;
+        *out_l = l;
+}
+#endif /* __GNUC__ */
+
 #endif
 #endif