3 files changed, 65 insertions, 1 deletions

diff --git a/tools/test/avx_sig/Makefile b/tools/test/avx_sig/Makefile
index 695cdc0a4a94..13a8d3546a83 100644
--- a/tools/test/avx_sig/Makefile
+++ b/tools/test/avx_sig/Makefile
@@ -10,5 +10,8 @@ LIBADD=		pthread
 .if ${MACHINE_CPUARCH} == "amd64"
 SRCS+=		c2x2c_amd64.S
 .endif
+.if ${MACHINE_CPUARCH} == "aarch64"
+SRCS+=		c2x2c_aarch64.S
+.endif
 
 .include <bsd.prog.mk>
diff --git a/tools/test/avx_sig/avx_sig.c b/tools/test/avx_sig/avx_sig.c
index 6ead5d8ce312..19fb85f2e803 100644
--- a/tools/test/avx_sig/avx_sig.c
+++ b/tools/test/avx_sig/avx_sig.c
@@ -48,14 +48,20 @@
 #define	nitems(x)	(sizeof((x)) / sizeof((x)[0]))
 #endif
 
+#if defined(__amd64__)
 #define	SIMDRNAM	"xmm"
+#define	NREGS		16
+#elif defined(__aarch64__)
+#define	SIMDRNAM	"q"
+#define	NREGS		32
+#endif
 
 struct simdreg {
 	uint8_t simd_bytes[16];
 };
 
 struct simd {
-	struct simdreg simdreg[16];
+	struct simdreg simdreg[NREGS];
 };
 
 void cpu_to_simd(struct simd *simd);
diff --git a/tools/test/avx_sig/c2x2c_aarch64.S b/tools/test/avx_sig/c2x2c_aarch64.S
new file mode 100644
index 000000000000..eebe1f334546
--- /dev/null
+++ b/tools/test/avx_sig/c2x2c_aarch64.S
@@ -0,0 +1,55 @@
+/*
+ * This file is in public domain.
+ * Written by Dmitry Chagin <dchagin@FreeBSD.org>
+ *
+ * $FreeBSD$
+ */
+
+	.global cpu_to_simd
+	.type cpu_to_simd, @function
+cpu_to_simd:
+	stp  q0,  q1, [x0, #( 0 * 16)]
+	stp  q2,  q3, [x0, #( 2 * 16)]
+	stp  q4,  q5, [x0, #( 4 * 16)]
+	stp  q6,  q7, [x0, #( 6 * 16)]
+	stp  q8,  q9, [x0, #( 8 * 16)]
+	stp q10, q11, [x0, #(10 * 16)]
+	stp q12, q13, [x0, #(12 * 16)]
+	stp q14, q15, [x0, #(14 * 16)]
+	stp q16, q17, [x0, #(16 * 16)]
+	stp q18, q19, [x0, #(18 * 16)]
+	stp q20, q21, [x0, #(20 * 16)]
+	stp q22, q23, [x0, #(22 * 16)]
+	stp q24, q25, [x0, #(24 * 16)]
+	stp q26, q27, [x0, #(26 * 16)]
+	stp q28, q29, [x0, #(28 * 16)]
+	stp q30, q31, [x0, #(30 * 16)]
+	ret
+
+	.size cpu_to_simd, . - cpu_to_simd
+
+
+	.global simd_to_cpu
+	.type simd_to_cpu, @function
+simd_to_cpu:
+	ldp  q0,  q1, [x0, #( 0 * 16)]
+	ldp  q2,  q3, [x0, #( 2 * 16)]
+	ldp  q4,  q5, [x0, #( 4 * 16)]
+	ldp  q6,  q7, [x0, #( 6 * 16)]
+	ldp  q8,  q9, [x0, #( 8 * 16)]
+	ldp q10, q11, [x0, #(10 * 16)]
+	ldp q12, q13, [x0, #(12 * 16)]
+	ldp q14, q15, [x0, #(14 * 16)]
+	ldp q16, q17, [x0, #(16 * 16)]
+	ldp q18, q19, [x0, #(18 * 16)]
+	ldp q20, q21, [x0, #(20 * 16)]
+	ldp q22, q23, [x0, #(22 * 16)]
+	ldp q24, q25, [x0, #(24 * 16)]
+	ldp q26, q27, [x0, #(26 * 16)]
+	ldp q28, q29, [x0, #(28 * 16)]
+	ldp q30, q31, [x0, #(30 * 16)]
+	ret
+
+	.size simd_to_cpu, . - simd_to_cpu
+
+	.section        .note.GNU-stack,"",@progbits