Add optimized versions of ChaCha20 by @floodyberry

1 files changed, 185 insertions, 0 deletions

diff --git a/src/libcryptobox/chacha20/macro.S b/src/libcryptobox/chacha20/macro.S
new file mode 100644
index 000000000..f4b4d0940
--- /dev/null
+++ b/src/libcryptobox/chacha20/macro.S
@@ -0,0 +1,185 @@
+#include "platform_config.h"
+
+#define IS_MACH (defined(__MACH__))
+
+#if !defined(HAVE_SLASHMACRO) && !defined(HAVE_DOLLARMACRO)
+	#error Unknown gnu as macro parameter convention! Run ./configure
+#endif
+
+#if (IS_MACH)
+	.macro FN name
+	#if defined(HAVE_SLASHMACRO)
+		\name:
+		_\name:
+	#elif defined(HAVE_DOLLARMACRO)
+		$0:
+		_$0:
+	#endif
+	.endm
+
+	.macro FN_EXT name, args, xmmused
+	#if defined(HAVE_SLASHMACRO)
+		FN \name
+	#elif defined(HAVE_DOLLARMACRO)
+		FN $0
+	#endif
+	.endm
+
+	.macro FN_END name
+	.endm
+
+	.macro HIDDEN name
+	#if defined(HAVE_AS_PRIVATE_EXTERN)
+	#if defined(HAVE_SLASHMACRO)
+		.private_extern \name
+		.private_extern _\name
+	#elif defined(HAVE_DOLLARMACRO)
+		.private_extern $0
+		.private_extern _$0
+	#endif
+	#endif
+	.endm
+#else
+	.macro FN name
+		\name:
+		_\name:
+	.endm
+
+	.macro FN_EXT name, args, xmmused
+		FN \name
+	.endm
+
+	.macro FN_END name
+		.size \name, .-\name
+		.size _\name, .-_\name
+		.type \name, @function
+	 	.type _\name, @function
+	.endm
+
+	.macro HIDDEN name
+	#if defined(HAVE_AS_HIDDEN)
+		.hidden \name
+		.hidden _\name
+	#endif
+	.endm
+
+	/* set NX for stack */
+	.section .note.GNU-stack,"",@progbits
+#endif
+#if (IS_MACH)
+	.macro SECTION_TEXT
+		.section __TEXT,__text,regular
+	.endm
+
+	.macro SECTION_RODATA
+		.section __TEXT,__text,regular
+	.endm
+#else
+	/* put everything in the code segment to simplify things */
+	.macro SECTION_TEXT
+		.text
+	.endm
+
+	.macro SECTION_RODATA
+		.text
+	.endm
+#endif
+
+/* declare a global function */
+.macro GLOBAL name
+#if defined(HAVE_SLASHMACRO)
+	.globl \name
+	.globl _\name
+#elif defined(HAVE_DOLLARMACRO)
+	.globl $0
+	.globl _$0
+#endif
+.endm
+
+.macro FN_LOCAL_PREFIX name
+#if defined(HAVE_SLASHMACRO)
+	FN LOCAL_PREFIX(\name)
+#elif defined(HAVE_DOLLARMACRO)
+	FN LOCAL_PREFIX($0)
+#endif
+.endm
+
+.macro FN_EXT_LOCAL_PREFIX name, args, xmmused
+#if defined(HAVE_SLASHMACRO)
+	FN_EXT LOCAL_PREFIX(\name), \args, \xmmused
+#elif defined(HAVE_DOLLARMACRO)
+	FN_EXT LOCAL_PREFIX($0), $1, $2
+#endif
+.endm
+
+.macro FN_END_LOCAL_PREFIX name
+#if defined(HAVE_SLASHMACRO)
+	FN_END LOCAL_PREFIX(\name)
+#elif defined(HAVE_DOLLARMACRO)
+	FN_END LOCAL_PREFIX($0)
+#endif
+.endm
+
+.macro GLOBAL_LOCAL_PREFIX name
+#if defined(HAVE_SLASHMACRO)
+	GLOBAL LOCAL_PREFIX(\name)
+	HIDDEN LOCAL_PREFIX(\name)
+#elif defined(HAVE_DOLLARMACRO)
+	GLOBAL LOCAL_PREFIX($0)
+	HIDDEN LOCAL_PREFIX($0)
+#endif
+.endm
+
+.macro GLOBAL_HIDDEN_FN name
+#if defined(HAVE_SLASHMACRO)
+	GLOBAL \name
+	HIDDEN \name
+	FN \name
+#elif defined(HAVE_DOLLARMACRO)
+	GLOBAL $0
+	HIDDEN $0
+	FN $0
+#endif
+.endm
+
+.macro GLOBAL_HIDDEN_FN_EXT name, args, xmmused
+#if defined(HAVE_SLASHMACRO)
+	GLOBAL \name
+	HIDDEN \name
+	FN_EXT \name, \args, \xmmused
+#elif defined(HAVE_DOLLARMACRO)
+	GLOBAL $0
+	HIDDEN $0
+	FN_EXT $0, $1, $2
+#endif
+.endm
+
+/* pic support */
+.macro LOAD_VAR_PIC var, reg
+#if (IS_X86_32)
+	#if defined(HAVE_SLASHMACRO)
+		call 1f
+		1:
+		popl \reg
+		leal \var - 1b(\reg), \reg
+	#elif defined(HAVE_DOLLARMACRO)
+		call 1f
+		1:
+		popl $1
+		leal $0 - 1b($1), $1
+	#endif
+#else
+	#if defined(HAVE_SLASHMACRO)
+		leaq \var(%rip), \reg
+	#elif defined(HAVE_DOLLARMACRO)
+		leaq $0(%rip), $1
+	#endif
+#endif
+.endm
+
+SECTION_RODATA
+.p2align 4,,15
+chacha_constants:
+.long 0x61707865,0x3320646e,0x79622d32,0x6b206574 /* "expand 32-byte k" */
+.byte 2,3,0,1,6,7,4,5,10,11,8,9,14,15,12,13 /* pshufb rotate by 16 */
+.byte 3,0,1,2,7,4,5,6,11,8,9,10,15,12,13,14 /* pshufb rotate by 8 */