x86_64 performance patch.

Configure

@@ -391,7 +391,7 @@ my %table=(
 "linux-s390",	"gcc:-DB_ENDIAN -DTERMIO -DNO_ASM -O3 -fomit-frame-pointer -Wall::-D_REENTRANT::-ldl:BN_LLONG::::::::::dlfcn:linux-shared:-fPIC::.so.\$(SHLIB_MAJOR).\$(SHLIB_MINOR)",
 "linux-s390x",	"gcc:-DB_ENDIAN -DTERMIO -DNO_ASM -O3 -fomit-frame-pointer -Wall::-D_REENTRANT::-ldl:SIXTY_FOUR_BIT_LONG::::::::::dlfcn:linux-shared:-fPIC::.so.\$(SHLIB_MAJOR).\$(SHLIB_MINOR)",
 "linux-ia64",   "gcc:-DL_ENDIAN -DTERMIO -O3 -fomit-frame-pointer -Wall::-D_REENTRANT::-ldl:SIXTY_FOUR_BIT_LONG RC4_CHUNK RC4_CHAR:asm/ia64.o:::::::::dlfcn:linux-shared:-fPIC::.so.\$(SHLIB_MAJOR).\$(SHLIB_MINOR)",
-"linux-x86_64", "gcc:-DL_ENDIAN -DNO_ASM ::-D_REENTRANT::-ldl:SIXTY_FOUR_BIT_LONG::::::::::dlfcn:linux-shared:-fPIC::.so.\$(SHLIB_MAJOR).\$(SHLIB_MINOR)",
+"linux-x86_64", "gcc:-m64 -DL_ENDIAN -DTERMIO -O3 -Wall -DMD32_REG_T=int::-D_REENTRANT::-ldl:SIXTY_FOUR_BIT_LONG RC4_CHUNK RC4_CHAR BF_PTR2 DES_INT DES_UNROLL:asm/x86_64-gcc.o:::::::::dlfcn:linux-shared:-fPIC:-m64:.so.\$(SHLIB_MAJOR).\$(SHLIB_MINOR)",
 "NetBSD-sparc",	"gcc:-DTERMIOS -O3 -fomit-frame-pointer -mv8 -Wall -DB_ENDIAN::(unknown):::BN_LLONG MD2_CHAR RC4_INDEX DES_UNROLL::::::::::dlfcn:bsd-gcc-shared:-fPIC::.so.\$(SHLIB_MAJOR).\$(SHLIB_MINOR)",
 "NetBSD-m68",	"gcc:-DTERMIOS -O3 -fomit-frame-pointer -Wall -DB_ENDIAN::(unknown):::BN_LLONG MD2_CHAR RC4_INDEX DES_UNROLL::::::::::dlfcn:bsd-gcc-shared:-fPIC::.so.\$(SHLIB_MAJOR).\$(SHLIB_MINOR)",
 "NetBSD-x86",	"gcc:-DTERMIOS -O3 -fomit-frame-pointer -m486 -Wall::(unknown):::BN_LLONG ${x86_gcc_des} ${x86_gcc_opts}::::::::::dlfcn:bsd-gcc-shared:-fPIC::.so.\$(SHLIB_MAJOR).\$(SHLIB_MINOR)",

Makefile.org

@@ -274,10 +274,7 @@ do_gnu-shared:
 	libs="$$libs -l$$i"; \
 	done
 
-DETECT_GNU_LD=${CC} -v 2>&1 | grep '^gcc' >/dev/null 2>&1 && \
+DETECT_GNU_LD=(${CC} -Wl,-V /dev/null 2>&1 | grep '^GNU ld' )>/dev/null
-	my_ld=`${CC} -print-prog-name=ld 2>&1` && \
-	[ -n "$$my_ld" ] && \
-	$$my_ld -v 2>&1 | grep 'GNU ld' >/dev/null 2>&1
 
 # For Darwin AKA Mac OS/X (dyld)
 do_darwin-shared: 

crypto/bn/Makefile.ssl

@@ -136,6 +136,8 @@ asm/ia64-cpp.o:	asm/ia64.S
 	$(CC) $(ASFLAGS) -c -o asm/ia64-cpp.o /tmp/ia64.$$$$.s;	\
 	rm -f /tmp/ia64.$$$$.s
 
+asm/x86_64-gcc.o: asm/x86_64-gcc.c
+
 files:
 	$(PERL) $(TOP)/util/files.pl Makefile.ssl >> $(TOP)/MINFO
 

crypto/bn/bn_div.c

@@ -150,6 +150,20 @@ int BN_div(BIGNUM *dv, BIGNUM *rem, const BIGNUM *m, const BIGNUM *d,
 	    q;					\
 	})
 #  define REMAINDER_IS_ALREADY_CALCULATED
+#  elif defined(__x86_64) && defined(SIXTY_FOUR_BIT_LONG)
+   /*
+    * Same story here, but it's 128-bit by 64-bit division. Wow!
+    *					<appro@fy.chalmers.se>
+    */
+#  define bn_div_words(n0,n1,d0)		\
+	({  asm volatile (			\
+		"divq	%4"			\
+		: "=a"(q), "=d"(rem)		\
+		: "a"(n1), "d"(n0), "g"(d0)	\
+		: "cc");			\
+	    q;					\
+	})
+#  define REMAINDER_IS_ALREADY_CALCULATED
 #  endif /* __<cpu> */
 # endif /* __GNUC__ */
 #endif /* OPENSSL_NO_ASM */
@@ -296,7 +310,9 @@ int BN_div(BIGNUM *dv, BIGNUM *rm, const BIGNUM *num, const BIGNUM *divisor,
 			rem=(n1-q*d0)&BN_MASK2;
 #endif
 
-#ifdef BN_UMULT_HIGH
+#if defined(BN_UMULT_LOHI)
+			BN_UMULT_LOHI(t2l,t2h,d1,q);
+#elif defined(BN_UMULT_HIGH)
 			t2l = d1 * q;
 			t2h = BN_UMULT_HIGH(d1,q);
 #else

crypto/bn/bn_lcl.h

@@ -230,6 +230,21 @@ struct bignum_ctx
 	     : "r"(a), "r"(b));		\
 	ret;			})
 #  endif	/* compiler */
+# elif defined(__x86_64) && defined(SIXTY_FOUR_BIT_LONG)
+#  if defined(__GNUC__)
+#   define BN_UMULT_HIGH(a,b)	({	\
+	register BN_ULONG ret,discard;	\
+	asm ("mulq	%3"		\
+	     : "=a"(discard),"=d"(ret)	\
+	     : "a"(a), "g"(b)		\
+	     : "cc");			\
+	ret;			})
+#   define BN_UMULT_LOHI(low,high,a,b)	\
+	asm ("mulq	%3"		\
+		: "=a"(low),"=d"(high)	\
+		: "a"(a),"g"(b)		\
+		: "cc");
+#  endif
 # endif		/* cpu */
 #endif		/* OPENSSL_NO_ASM */
 
@@ -337,7 +352,7 @@ struct bignum_ctx
 
 #define LBITS(a)	((a)&BN_MASK2l)
 #define HBITS(a)	(((a)>>BN_BITS4)&BN_MASK2l)
-#define	L2HBITS(a)	((BN_ULONG)((a)&BN_MASK2l)<<BN_BITS4)
+#define	L2HBITS(a)	(((a)<<BN_BITS4)&BN_MASK2)
 
 #define LLBITS(a)	((a)&BN_MASKl)
 #define LHBITS(a)	(((a)>>BN_BITS2)&BN_MASKl)

crypto/des/des_locl.h

@@ -162,6 +162,16 @@
 
 #if defined(OPENSSL_SYS_WIN32) && defined(_MSC_VER)
 #define	ROTATE(a,n)	(_lrotr(a,n))
+#elif defined(__GNUC__) && __GNUC__>=2 && !defined(NO_ASM) && !defined(NO_INLINE_ASM)
+# if defined(__i386) || defined(__i386__) || defined(__x86_64) || defined(__x86_64__)
+#  define ROTATE(a,n)	({ register unsigned int ret;	\
+				asm ("rorl %1,%0"	\
+					: "=r"(ret)	\
+					: "I"(n),"0"(a)	\
+					: "cc");	\
+			   ret;				\
+			})
+# endif
 #else
 #define	ROTATE(a,n)	(((a)>>(n))+((a)<<(32-(n))))
 #endif

crypto/md32_common.h

@@ -198,7 +198,7 @@
    *
    * 					<appro@fy.chalmers.se>
    */
-#  if defined(__i386) || defined(__i386__)
+#  if defined(__i386) || defined(__i386__) || defined(__x86_64) || defined(__x86_64__)
 #   define ROTATE(a,n)	({ register unsigned int ret;	\
 				asm (			\
 				"roll %1,%0"		\
@@ -224,7 +224,7 @@
  */
 # if defined(__GNUC__) && __GNUC__>=2 && !defined(OPENSSL_NO_ASM) && !defined(OPENSSL_NO_INLINE_ASM)
   /* some GNU C inline assembler templates by <appro@fy.chalmers.se> */
-#  if (defined(__i386) || defined(__i386__)) && !defined(I386_ONLY)
+#  if (defined(__i386) || defined(__i386__) || defined(__x86_64) || defined(__x86_64__)) && !defined(I386_ONLY)
 #   define BE_FETCH32(a)	({ register unsigned int l=(a);\
 				asm (			\
 				"bswapl %0"		\
@@ -610,3 +610,28 @@ int HASH_FINAL (unsigned char *md, HASH_CTX *c)
 	 */
 	return 1;
 	}
+
+#ifndef MD32_REG_T
+#define MD32_REG_T long
+/*
+ * This comment was originaly written for MD5, which is why it
+ * discusses A-D. But it basically applies to all 32-bit digests,
+ * which is why it was moved to common header file.
+ *
+ * In case you wonder why A-D are declared as long and not
+ * as MD5_LONG. Doing so results in slight performance
+ * boost on LP64 architectures. The catch is we don't
+ * really care if 32 MSBs of a 64-bit register get polluted
+ * with eventual overflows as we *save* only 32 LSBs in
+ * *either* case. Now declaring 'em long excuses the compiler
+ * from keeping 32 MSBs zeroed resulting in 13% performance
+ * improvement under SPARC Solaris7/64 and 5% under AlphaLinux.
+ * Well, to be honest it should say that this *prevents* 
+ * performance degradation.
+ *				<appro@fy.chalmers.se>
+ * Apparently there're LP64 compilers that generate better
+ * code if A-D are declared int. Most notably GCC-x86_64
+ * generates better code.
+ *				<appro@fy.chalmers.se>
+ */
+#endif

crypto/md4/md4_dgst.c

@@ -86,21 +86,7 @@ int MD4_Init(MD4_CTX *c)
 void md4_block_host_order (MD4_CTX *c, const void *data, int num)
 	{
 	const MD4_LONG *X=data;
-	register unsigned long A,B,C,D;
+	register unsigned MD32_REG_T A,B,C,D;
-	/*
-	 * In case you wonder why A-D are declared as long and not
-	 * as MD4_LONG. Doing so results in slight performance
-	 * boost on LP64 architectures. The catch is we don't
-	 * really care if 32 MSBs of a 64-bit register get polluted
-	 * with eventual overflows as we *save* only 32 LSBs in
-	 * *either* case. Now declaring 'em long excuses the compiler
-	 * from keeping 32 MSBs zeroed resulting in 13% performance
-	 * improvement under SPARC Solaris7/64 and 5% under AlphaLinux.
-	 * Well, to be honest it should say that this *prevents* 
-	 * performance degradation.
-	 *
-	 *				<appro@fy.chalmers.se>
-	 */
 
 	A=c->A;
 	B=c->B;
@@ -176,25 +162,11 @@ void md4_block_host_order (MD4_CTX *c, const void *data, int num)
 void md4_block_data_order (MD4_CTX *c, const void *data_, int num)
 	{
 	const unsigned char *data=data_;
-	register unsigned long A,B,C,D,l;
+	register unsigned MD32_REG_T A,B,C,D,l;
-	/*
-	 * In case you wonder why A-D are declared as long and not
-	 * as MD4_LONG. Doing so results in slight performance
-	 * boost on LP64 architectures. The catch is we don't
-	 * really care if 32 MSBs of a 64-bit register get polluted
-	 * with eventual overflows as we *save* only 32 LSBs in
-	 * *either* case. Now declaring 'em long excuses the compiler
-	 * from keeping 32 MSBs zeroed resulting in 13% performance
-	 * improvement under SPARC Solaris7/64 and 5% under AlphaLinux.
-	 * Well, to be honest it should say that this *prevents* 
-	 * performance degradation.
-	 *
-	 *				<appro@fy.chalmers.se>
-	 */
 #ifndef MD32_XARRAY
 	/* See comment in crypto/sha/sha_locl.h for details. */
-	unsigned long	XX0, XX1, XX2, XX3, XX4, XX5, XX6, XX7,
+	unsigned MD32_REG_T	XX0, XX1, XX2, XX3, XX4, XX5, XX6, XX7,
-			XX8, XX9,XX10,XX11,XX12,XX13,XX14,XX15;
+				XX8, XX9,XX10,XX11,XX12,XX13,XX14,XX15;
 # define X(i)	XX##i
 #else
 	MD4_LONG XX[MD4_LBLOCK];

crypto/md5/md5_dgst.c

@@ -86,21 +86,7 @@ int MD5_Init(MD5_CTX *c)
 void md5_block_host_order (MD5_CTX *c, const void *data, int num)
 	{
 	const MD5_LONG *X=data;
-	register unsigned long A,B,C,D;
+	register unsigned MD32_REG_T A,B,C,D;
-	/*
-	 * In case you wonder why A-D are declared as long and not
-	 * as MD5_LONG. Doing so results in slight performance
-	 * boost on LP64 architectures. The catch is we don't
-	 * really care if 32 MSBs of a 64-bit register get polluted
-	 * with eventual overflows as we *save* only 32 LSBs in
-	 * *either* case. Now declaring 'em long excuses the compiler
-	 * from keeping 32 MSBs zeroed resulting in 13% performance
-	 * improvement under SPARC Solaris7/64 and 5% under AlphaLinux.
-	 * Well, to be honest it should say that this *prevents* 
-	 * performance degradation.
-	 *
-	 *				<appro@fy.chalmers.se>
-	 */
 
 	A=c->A;
 	B=c->B;
@@ -193,25 +179,11 @@ void md5_block_host_order (MD5_CTX *c, const void *data, int num)
 void md5_block_data_order (MD5_CTX *c, const void *data_, int num)
 	{
 	const unsigned char *data=data_;
-	register unsigned long A,B,C,D,l;
+	register unsigned MD32_REG_T A,B,C,D,l;
-	/*
-	 * In case you wonder why A-D are declared as long and not
-	 * as MD5_LONG. Doing so results in slight performance
-	 * boost on LP64 architectures. The catch is we don't
-	 * really care if 32 MSBs of a 64-bit register get polluted
-	 * with eventual overflows as we *save* only 32 LSBs in
-	 * *either* case. Now declaring 'em long excuses the compiler
-	 * from keeping 32 MSBs zeroed resulting in 13% performance
-	 * improvement under SPARC Solaris7/64 and 5% under AlphaLinux.
-	 * Well, to be honest it should say that this *prevents* 
-	 * performance degradation.
-	 *
-	 *				<appro@fy.chalmers.se>
-	 */
 #ifndef MD32_XARRAY
 	/* See comment in crypto/sha/sha_locl.h for details. */
-	unsigned long	XX0, XX1, XX2, XX3, XX4, XX5, XX6, XX7,
+	unsigned MD32_REG_T	XX0, XX1, XX2, XX3, XX4, XX5, XX6, XX7,
-			XX8, XX9,XX10,XX11,XX12,XX13,XX14,XX15;
+				XX8, XX9,XX10,XX11,XX12,XX13,XX14,XX15;
 # define X(i)	XX##i
 #else
 	MD5_LONG XX[MD5_LBLOCK];

crypto/rc5/rc5_locl.h

@@ -149,6 +149,23 @@
 #if defined(OPENSSL_SYS_WIN32) && defined(_MSC_VER)
 #define ROTATE_l32(a,n)     _lrotl(a,n)
 #define ROTATE_r32(a,n)     _lrotr(a,n)
+#elif defined(__GNUC__) && __GNUC__>=2 && !defined(NO_ASM) && !defined(NO_INLINE_ASM)
+# if defined(__i386) || defined(__i386__) || defined(__x86_64) || defined(__x86_64__)
+#  define ROTATE_l32(a,n)	({ register unsigned int ret;	\
+					asm ("roll %%cl,%0"	\
+						: "=r"(ret)	\
+						: "c"(n),"0"(a)	\
+						: "cc");	\
+					ret;			\
+				})
+#  define ROTATE_r32(a,n)	({ register unsigned int ret;	\
+					asm ("rorl %%cl,%0"	\
+						: "=r"(ret)	\
+						: "c"(n),"0"(a)	\
+						: "cc");	\
+					ret;			\
+				})
+# endif
 #else
 #define ROTATE_l32(a,n)     (((a)<<(n&0x1f))|(((a)&0xffffffff)>>(32-(n&0x1f))))
 #define ROTATE_r32(a,n)     (((a)<<(32-(n&0x1f)))|(((a)&0xffffffff)>>(n&0x1f)))

crypto/ripemd/rmd_dgst.c

@@ -90,8 +90,8 @@ int RIPEMD160_Init(RIPEMD160_CTX *c)
 void ripemd160_block_host_order (RIPEMD160_CTX *ctx, const void *p, int num)
 	{
 	const RIPEMD160_LONG *XX=p;
-	register unsigned long A,B,C,D,E;
+	register unsigned MD32_REG_T A,B,C,D,E;
-	register unsigned long a,b,c,d,e;
+	register unsigned MD32_REG_T a,b,c,d,e;
 
 	for (;num--;XX+=HASH_LBLOCK)
 		{
@@ -290,12 +290,12 @@ void ripemd160_block_host_order (RIPEMD160_CTX *ctx, const void *p, int num)
 void ripemd160_block_data_order (RIPEMD160_CTX *ctx, const void *p, int num)
 	{
 	const unsigned char *data=p;
-	register unsigned long A,B,C,D,E;
+	register unsigned MD32_REG_T A,B,C,D,E;
-	unsigned long a,b,c,d,e,l;
+	unsigned MD32_REG_T a,b,c,d,e,l;
 #ifndef MD32_XARRAY
 	/* See comment in crypto/sha/sha_locl.h for details. */
-	unsigned long	XX0, XX1, XX2, XX3, XX4, XX5, XX6, XX7,
+	unsigned MD32_REG_T	XX0, XX1, XX2, XX3, XX4, XX5, XX6, XX7,
-			XX8, XX9,XX10,XX11,XX12,XX13,XX14,XX15;
+				XX8, XX9,XX10,XX11,XX12,XX13,XX14,XX15;
 # define X(i)	XX##i
 #else
 	RIPEMD160_LONG	XX[16];

crypto/sha/sha_locl.h

@@ -224,10 +224,10 @@ int HASH_INIT (SHA_CTX *c)
 void HASH_BLOCK_HOST_ORDER (SHA_CTX *c, const void *d, int num)
 	{
 	const SHA_LONG *W=d;
-	register unsigned long A,B,C,D,E,T;
+	register unsigned MD32_REG_T A,B,C,D,E,T;
 #ifndef MD32_XARRAY
-	unsigned long	XX0, XX1, XX2, XX3, XX4, XX5, XX6, XX7,
+	unsigned MD32_REG_T	XX0, XX1, XX2, XX3, XX4, XX5, XX6, XX7,
-			XX8, XX9,XX10,XX11,XX12,XX13,XX14,XX15;
+				XX8, XX9,XX10,XX11,XX12,XX13,XX14,XX15;
 #else
 	SHA_LONG	XX[16];
 #endif
@@ -349,10 +349,10 @@ void HASH_BLOCK_HOST_ORDER (SHA_CTX *c, const void *d, int num)
 void HASH_BLOCK_DATA_ORDER (SHA_CTX *c, const void *p, int num)
 	{
 	const unsigned char *data=p;
-	register unsigned long A,B,C,D,E,T,l;
+	register unsigned MD32_REG_T A,B,C,D,E,T,l;
 #ifndef MD32_XARRAY
-	unsigned long	XX0, XX1, XX2, XX3, XX4, XX5, XX6, XX7,
+	unsigned MD32_REG_T	XX0, XX1, XX2, XX3, XX4, XX5, XX6, XX7,
-			XX8, XX9,XX10,XX11,XX12,XX13,XX14,XX15;
+				XX8, XX9,XX10,XX11,XX12,XX13,XX14,XX15;
 #else
 	SHA_LONG	XX[16];
 #endif