From 588b959bae4a5b78fd706db29b9c009cf25b538e Mon Sep 17 00:00:00 2001
From: jsing <jsing@openbsd.org>
Date: Sun, 29 Jan 2023 14:00:41 +0000
Subject: [PATCH] Use s2n-bignum assembly implementations for libcrypto bignum
 on amd64.

This switches the core bignum assembly implementations from x86_64-gcc.c to
s2n-bignum for amd64.

ok miod@ tb@
---
 lib/libcrypto/arch/amd64/Makefile.inc | 13 ++++-
 lib/libcrypto/bn/arch/amd64/bn_arch.c | 80 ++++++++++++++++++++++++++-
 2 files changed, 90 insertions(+), 3 deletions(-)

diff --git a/lib/libcrypto/arch/amd64/Makefile.inc b/lib/libcrypto/arch/amd64/Makefile.inc
index 1fd9f689197..5e433b572d4 100644
--- a/lib/libcrypto/arch/amd64/Makefile.inc
+++ b/lib/libcrypto/arch/amd64/Makefile.inc
@@ -1,4 +1,4 @@
-# $OpenBSD: Makefile.inc,v 1.10 2023/01/21 17:29:56 jsing Exp $
+# $OpenBSD: Makefile.inc,v 1.11 2023/01/29 14:00:41 jsing Exp $
 
 # amd64-specific libcrypto build rules
 
@@ -29,7 +29,16 @@ SSLASM+= bn x86_64-gf2m
 
 # bn s2n-bignum
 SRCS += bn_arch.c
+SRCS += bignum_add.S
+SRCS += bignum_cmadd.S
+SRCS += bignum_cmul.S
+SRCS += bignum_mul.S
+SRCS += bignum_mul_4_8_alt.S
+SRCS += bignum_mul_8_16_alt.S
 SRCS += bignum_sqr.S
+SRCS += bignum_sqr_4_8_alt.S
+SRCS += bignum_sqr_8_16_alt.S
+SRCS += bignum_sub.S
 
 # camellia
 SRCS+=	cmll_misc.c
@@ -75,7 +84,7 @@ ${f}.S: ${LCRYPTO_SRC}/${dir}/asm/${f}.pl ${EXTRA_PL}
 .endfor
 
 CFLAGS+= -DOPENSSL_CPUID_OBJ
-SRCS+=	x86_64cpuid.S x86_64-gcc.c
+SRCS+=	x86_64cpuid.S
 GENERATED+=x86_64cpuid.S
 
 x86_64cpuid.S: ${LCRYPTO_SRC}/x86_64cpuid.pl ${EXTRA_PL}
diff --git a/lib/libcrypto/bn/arch/amd64/bn_arch.c b/lib/libcrypto/bn/arch/amd64/bn_arch.c
index 240575955ce..aedefc76e22 100644
--- a/lib/libcrypto/bn/arch/amd64/bn_arch.c
+++ b/lib/libcrypto/bn/arch/amd64/bn_arch.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: bn_arch.c,v 1.1 2023/01/21 17:29:56 jsing Exp $ */
+/*	$OpenBSD: bn_arch.c,v 1.2 2023/01/29 14:00:41 jsing Exp $ */
 /*
  * Copyright (c) 2023 Joel Sing <jsing@openbsd.org>
  *
@@ -21,6 +21,58 @@
 #include "bn_local.h"
 #include "s2n_bignum.h"
 
+#ifdef HAVE_BN_ADD_WORDS
+BN_ULONG
+bn_add_words(BN_ULONG *rd, const BN_ULONG *ad, const BN_ULONG *bd, int n)
+{
+	return bignum_add(n, (uint64_t *)rd, n, (uint64_t *)ad, n,
+	    (uint64_t *)bd);
+}
+#endif
+
+#ifdef HAVE_BN_SUB_WORDS
+BN_ULONG
+bn_sub_words(BN_ULONG *rd, const BN_ULONG *ad, const BN_ULONG *bd, int n)
+{
+	return bignum_sub(n, (uint64_t *)rd, n, (uint64_t *)ad, n,
+	    (uint64_t *)bd);
+}
+#endif
+
+#ifdef HAVE_BN_MUL_ADD_WORDS
+BN_ULONG
+bn_mul_add_words(BN_ULONG *rd, const BN_ULONG *ad, int num, BN_ULONG w)
+{
+	return bignum_cmadd(num, (uint64_t *)rd, w, num, (uint64_t *)ad);
+}
+#endif
+
+#ifdef HAVE_BN_MUL_WORDS
+BN_ULONG
+bn_mul_words(BN_ULONG *rd, const BN_ULONG *ad, int num, BN_ULONG w)
+{
+	return bignum_cmul(num, (uint64_t *)rd, w, num, (uint64_t *)ad);
+}
+#endif
+
+#ifdef HAVE_BN_MUL_COMBA4
+void
+bn_mul_comba4(BN_ULONG *rd, BN_ULONG *ad, BN_ULONG *bd)
+{
+	/* XXX - consider using non-alt on CPUs that have the ADX extension. */
+	bignum_mul_4_8_alt((uint64_t *)rd, (uint64_t *)ad, (uint64_t *)bd);
+}
+#endif
+
+#ifdef HAVE_BN_MUL_COMBA8
+void
+bn_mul_comba8(BN_ULONG *rd, BN_ULONG *ad, BN_ULONG *bd)
+{
+	/* XXX - consider using non-alt on CPUs that have the ADX extension. */
+	bignum_mul_8_16_alt((uint64_t *)rd, (uint64_t *)ad, (uint64_t *)bd);
+}
+#endif
+
 #ifdef HAVE_BN_SQR
 int
 bn_sqr(BIGNUM *r, const BIGNUM *a, int rn, BN_CTX *ctx)
@@ -30,3 +82,29 @@ bn_sqr(BIGNUM *r, const BIGNUM *a, int rn, BN_CTX *ctx)
 	return 1;
 }
 #endif
+
+#ifdef HAVE_BN_SQR_COMBA4
+void
+bn_sqr_comba4(BN_ULONG *rd, const BN_ULONG *ad)
+{
+	/* XXX - consider using non-alt on CPUs that have the ADX extension. */
+	bignum_sqr_4_8_alt((uint64_t *)rd, (uint64_t *)ad);
+}
+#endif
+
+#ifdef HAVE_BN_SQR_COMBA8
+void
+bn_sqr_comba8(BN_ULONG *rd, const BN_ULONG *ad)
+{
+	/* XXX - consider using non-alt on CPUs that have the ADX extension. */
+	bignum_sqr_8_16_alt((uint64_t *)rd, (uint64_t *)ad);
+}
+#endif
+
+#ifdef HAVE_BN_SQR_WORDS
+void
+bn_sqr_words(BN_ULONG *rd, const BN_ULONG *ad, int num)
+{
+	bignum_sqr(num, (uint64_t *)rd, num, (uint64_t *)ad);
+}
+#endif
-- 
2.20.1