Provide optimised bn_mulw_{addw,addw_addw,addtw}() for aarch64.

This results in bn_mul_comba4() and bn_mul_comba8() requiring ~30% less
instructions than they did previously.

diff --git a/lib/libcrypto/bn/arch/aarch64/bn_arch.h b/lib/libcrypto/bn/arch/aarch64/bn_arch.h
index 1b9358e..708083a 100644 (file)
--- a/lib/libcrypto/bn/arch/aarch64/bn_arch.h
+++ b/lib/libcrypto/bn/arch/aarch64/bn_arch.h
@@ -1,4 +1,4 @@
-/*     $OpenBSD: bn_arch.h,v 1.8 2023/05/28 17:22:04 jsing Exp $ */
+/*     $OpenBSD: bn_arch.h,v 1.9 2023/05/28 17:42:30 jsing Exp $ */
 /*
  * Copyright (c) 2023 Joel Sing <jsing@openbsd.org>
  *
@@ -81,6 +81,73 @@ bn_mulw(BN_ULONG a, BN_ULONG b, BN_ULONG *out_r1, BN_ULONG *out_r0)
        *out_r0 = r0;
 }
 
+#define HAVE_BN_MULW_ADDW
+
+static inline void
+bn_mulw_addw(BN_ULONG a, BN_ULONG b, BN_ULONG c, BN_ULONG *out_r1,
+    BN_ULONG *out_r0)
+{
+       BN_ULONG r1, r0;
+
+       __asm__ (
+           "umulh  %[r1], %[a], %[b] \n"
+           "mul    %[r0], %[a], %[b] \n"
+           "adds   %[r0], %[r0], %[c] \n"
+           "adc    %[r1], %[r1], xzr \n"
+           : [r1]"=&r"(r1), [r0]"=&r"(r0)
+           : [a]"r"(a), [b]"r"(b), [c]"r"(c)
+           : "cc");
+
+       *out_r1 = r1;
+       *out_r0 = r0;
+}
+
+#define HAVE_BN_MULW_ADDW_ADDW
+
+static inline void
+bn_mulw_addw_addw(BN_ULONG a, BN_ULONG b, BN_ULONG c, BN_ULONG d,
+    BN_ULONG *out_r1, BN_ULONG *out_r0)
+{
+       BN_ULONG r1, r0;
+
+       __asm__ (
+           "umulh  %[r1], %[a], %[b] \n"
+           "mul    %[r0], %[a], %[b] \n"
+           "adds   %[r0], %[r0], %[c] \n"
+           "adc    %[r1], %[r1], xzr \n"
+           "adds   %[r0], %[r0], %[d] \n"
+           "adc    %[r1], %[r1], xzr \n"
+           : [r1]"=&r"(r1), [r0]"=&r"(r0)
+           : [a]"r"(a), [b]"r"(b), [c]"r"(c), [d]"r"(d)
+           : "cc");
+
+       *out_r1 = r1;
+       *out_r0 = r0;
+}
+
+#define HAVE_BN_MULW_ADDTW
+
+static inline void
+bn_mulw_addtw(BN_ULONG a, BN_ULONG b, BN_ULONG c2, BN_ULONG c1, BN_ULONG c0,
+    BN_ULONG *out_r2, BN_ULONG *out_r1, BN_ULONG *out_r0)
+{
+       BN_ULONG r2, r1, r0;
+
+       __asm__ (
+           "umulh  %[r1], %[a], %[b] \n"
+           "mul    %[r0], %[a], %[b] \n"
+           "adds   %[r0], %[r0], %[c0] \n"
+           "adcs   %[r1], %[r1], %[c1] \n"
+           "adc    %[r2], xzr, %[c2] \n"
+           : [r2]"=&r"(r2), [r1]"=&r"(r1), [r0]"=&r"(r0)
+           : [a]"r"(a), [b]"r"(b), [c2]"r"(c2), [c1]"r"(c1), [c0]"r"(c0)
+           : "cc");
+
+       *out_r2 = r2;
+       *out_r1 = r1;
+       *out_r0 = r0;
+}
+
 #define HAVE_BN_SUBW
 
 static inline void