Add VLIB_NODE_FN() macro to simplify multiversioning of node functions

Change-Id: Ibab5e27277f618ceb2d543b9d6a1a5f191e7d1db
Signed-off-by: Damjan Marion <damarion@cisco.com>

19 files changed, 160 insertions, 284 deletions

diff --git a/src/plugins/acl.am b/src/plugins/acl.am
index 7d723402615..8508a1dc025 100644
--- a/src/plugins/acl.am
+++ b/src/plugins/acl.am
@@ -34,7 +34,7 @@ acl_plugin_la_SOURCES =				\
 API_FILES += acl/acl.api
 
 if CPU_X86_64
-acl_multiversioning_files =                                    \
+acl_multiversioning_sources =                                    \
         acl/dataplane_node.c
 
 
@@ -42,10 +42,10 @@ if CC_SUPPORTS_AVX2
 ###############################################################
 # AVX2
 ###############################################################
-libacl_plugin_avx2_la_SOURCES = $(acl_multiversioning_files)
+libacl_plugin_avx2_la_SOURCES = $(acl_multiversioning_sources)
 libacl_plugin_avx2_la_CFLAGS =                                 \
         $(AM_CFLAGS)  @CPU_AVX2_FLAGS@                          \
-        -DCLIB_MULTIARCH_VARIANT=avx2
+        -DCLIB_MARCH_VARIANT=avx2
 noinst_LTLIBRARIES += libacl_plugin_avx2.la
 acl_plugin_la_LIBADD += libacl_plugin_avx2.la
 endif
@@ -54,10 +54,10 @@ if CC_SUPPORTS_AVX512
 ###############################################################
 # AVX512
 ###############################################################
-libacl_plugin_avx512_la_SOURCES = $(acl_multiversioning_files)
+libacl_plugin_avx512_la_SOURCES = $(acl_multiversioning_sources)
 libacl_plugin_avx512_la_CFLAGS =                               \
         $(AM_CFLAGS) @CPU_AVX512_FLAGS@                         \
-        -DCLIB_MULTIARCH_VARIANT=avx512
+        -DCLIB_MARCH_VARIANT=avx512
 noinst_LTLIBRARIES += libacl_plugin_avx512.la
 acl_plugin_la_LIBADD += libacl_plugin_avx512.la
 endif
diff --git a/src/plugins/acl/dataplane_node.c b/src/plugins/acl/dataplane_node.c
index 5393091f26f..351cbbd8fb2 100644
--- a/src/plugins/acl/dataplane_node.c
+++ b/src/plugins/acl/dataplane_node.c
@@ -327,36 +327,10 @@ acl_fa_node_fn (vlib_main_t * vm,
   return frame->n_vectors;
 }
 
-vlib_node_function_t __clib_weak acl_in_ip4_l2_node_fn_avx512;
-vlib_node_function_t __clib_weak acl_in_ip4_l2_node_fn_avx2;
-
-vlib_node_function_t __clib_weak acl_out_ip4_l2_node_fn_avx512;
-vlib_node_function_t __clib_weak acl_out_ip4_l2_node_fn_avx2;
-
-vlib_node_function_t __clib_weak acl_in_ip6_l2_node_fn_avx512;
-vlib_node_function_t __clib_weak acl_in_ip6_l2_node_fn_avx2;
-
-vlib_node_function_t __clib_weak acl_out_ip6_l2_node_fn_avx512;
-vlib_node_function_t __clib_weak acl_out_ip6_l2_node_fn_avx2;
-
-vlib_node_function_t __clib_weak acl_in_ip4_fa_node_fn_avx512;
-vlib_node_function_t __clib_weak acl_in_ip4_fa_node_fn_avx2;
-
-vlib_node_function_t __clib_weak acl_out_ip4_fa_node_fn_avx512;
-vlib_node_function_t __clib_weak acl_out_ip4_fa_node_fn_avx2;
-
-vlib_node_function_t __clib_weak acl_in_ip6_fa_node_fn_avx512;
-vlib_node_function_t __clib_weak acl_in_ip6_fa_node_fn_avx2;
-
-vlib_node_function_t __clib_weak acl_out_ip6_fa_node_fn_avx512;
-vlib_node_function_t __clib_weak acl_out_ip6_fa_node_fn_avx2;
-
-
 vlib_node_registration_t acl_in_l2_ip6_node;
-uword CLIB_CPU_OPTIMIZED
-CLIB_MULTIARCH_FN (acl_in_ip6_l2_node_fn) (vlib_main_t * vm,
-					   vlib_node_runtime_t * node,
-					   vlib_frame_t * frame)
+VLIB_NODE_FN (acl_in_l2_ip6_node) (vlib_main_t * vm,
+				   vlib_node_runtime_t * node,
+				   vlib_frame_t * frame)
 {
   acl_main_t *am = &acl_main;
   return acl_fa_node_fn (vm, node, frame, 1, 1, 1,
@@ -365,10 +339,9 @@ CLIB_MULTIARCH_FN (acl_in_ip6_l2_node_fn) (vlib_main_t * vm,
 }
 
 vlib_node_registration_t acl_in_l2_ip4_node;
-uword CLIB_CPU_OPTIMIZED
-CLIB_MULTIARCH_FN (acl_in_ip4_l2_node_fn) (vlib_main_t * vm,
-					   vlib_node_runtime_t * node,
-					   vlib_frame_t * frame)
+VLIB_NODE_FN (acl_in_l2_ip4_node) (vlib_main_t * vm,
+				   vlib_node_runtime_t * node,
+				   vlib_frame_t * frame)
 {
   acl_main_t *am = &acl_main;
   return acl_fa_node_fn (vm, node, frame, 0, 1, 1,
@@ -377,10 +350,9 @@ CLIB_MULTIARCH_FN (acl_in_ip4_l2_node_fn) (vlib_main_t * vm,
 }
 
 vlib_node_registration_t acl_out_l2_ip6_node;
-uword CLIB_CPU_OPTIMIZED
-CLIB_MULTIARCH_FN (acl_out_ip6_l2_node_fn) (vlib_main_t * vm,
-					    vlib_node_runtime_t * node,
-					    vlib_frame_t * frame)
+VLIB_NODE_FN (acl_out_l2_ip6_node) (vlib_main_t * vm,
+				    vlib_node_runtime_t * node,
+				    vlib_frame_t * frame)
 {
   acl_main_t *am = &acl_main;
   return acl_fa_node_fn (vm, node, frame, 1, 0, 1,
@@ -389,10 +361,9 @@ CLIB_MULTIARCH_FN (acl_out_ip6_l2_node_fn) (vlib_main_t * vm,
 }
 
 vlib_node_registration_t acl_out_l2_ip4_node;
-uword CLIB_CPU_OPTIMIZED
-CLIB_MULTIARCH_FN (acl_out_ip4_l2_node_fn) (vlib_main_t * vm,
-					    vlib_node_runtime_t * node,
-					    vlib_frame_t * frame)
+VLIB_NODE_FN (acl_out_l2_ip4_node) (vlib_main_t * vm,
+				    vlib_node_runtime_t * node,
+				    vlib_frame_t * frame)
 {
   acl_main_t *am = &acl_main;
   return acl_fa_node_fn (vm, node, frame, 0, 0, 1,
@@ -403,93 +374,38 @@ CLIB_MULTIARCH_FN (acl_out_ip4_l2_node_fn) (vlib_main_t * vm,
 /**** L3 processing path nodes ****/
 
 vlib_node_registration_t acl_in_fa_ip6_node;
-uword CLIB_CPU_OPTIMIZED
-CLIB_MULTIARCH_FN (acl_in_ip6_fa_node_fn) (vlib_main_t * vm,
-					   vlib_node_runtime_t * node,
-					   vlib_frame_t * frame)
+VLIB_NODE_FN (acl_in_fa_ip6_node) (vlib_main_t * vm,
+				   vlib_node_runtime_t * node,
+				   vlib_frame_t * frame)
 {
   return acl_fa_node_fn (vm, node, frame, 1, 1, 0, 0, &acl_in_fa_ip6_node);
 }
 
 vlib_node_registration_t acl_in_fa_ip4_node;
-uword CLIB_CPU_OPTIMIZED
-CLIB_MULTIARCH_FN (acl_in_ip4_fa_node_fn) (vlib_main_t * vm,
-					   vlib_node_runtime_t * node,
-					   vlib_frame_t * frame)
+VLIB_NODE_FN (acl_in_fa_ip4_node) (vlib_main_t * vm,
+				   vlib_node_runtime_t * node,
+				   vlib_frame_t * frame)
 {
   return acl_fa_node_fn (vm, node, frame, 0, 1, 0, 0, &acl_in_fa_ip4_node);
 }
 
 vlib_node_registration_t acl_out_fa_ip6_node;
-uword CLIB_CPU_OPTIMIZED
-CLIB_MULTIARCH_FN (acl_out_ip6_fa_node_fn) (vlib_main_t * vm,
-					    vlib_node_runtime_t * node,
-					    vlib_frame_t * frame)
+VLIB_NODE_FN (acl_out_fa_ip6_node) (vlib_main_t * vm,
+				    vlib_node_runtime_t * node,
+				    vlib_frame_t * frame)
 {
   return acl_fa_node_fn (vm, node, frame, 1, 0, 0, 0, &acl_out_fa_ip6_node);
 }
 
 vlib_node_registration_t acl_out_fa_ip4_node;
-uword CLIB_CPU_OPTIMIZED
-CLIB_MULTIARCH_FN (acl_out_ip4_fa_node_fn) (vlib_main_t * vm,
-					    vlib_node_runtime_t * node,
-					    vlib_frame_t * frame)
+VLIB_NODE_FN (acl_out_fa_ip4_node) (vlib_main_t * vm,
+				    vlib_node_runtime_t * node,
+				    vlib_frame_t * frame)
 {
   return acl_fa_node_fn (vm, node, frame, 0, 0, 0, 0, &acl_out_fa_ip4_node);
 }
 
-
-
-#if __x86_64__
-static void __clib_constructor
-acl_plugin_multiarch_select (void)
-{
-  if (acl_in_ip4_l2_node_fn_avx512 && clib_cpu_supports_avx512f ())
-    acl_in_l2_ip4_node.function = acl_in_ip4_l2_node_fn_avx512;
-  else if (acl_in_ip4_l2_node_fn_avx2 && clib_cpu_supports_avx2 ())
-    acl_in_l2_ip4_node.function = acl_in_ip4_l2_node_fn_avx2;
-
-  if (acl_out_ip4_l2_node_fn_avx512 && clib_cpu_supports_avx512f ())
-    acl_out_l2_ip4_node.function = acl_out_ip4_l2_node_fn_avx512;
-  else if (acl_out_ip4_l2_node_fn_avx2 && clib_cpu_supports_avx2 ())
-    acl_out_l2_ip4_node.function = acl_out_ip4_l2_node_fn_avx2;
-
-  if (acl_in_ip6_l2_node_fn_avx512 && clib_cpu_supports_avx512f ())
-    acl_in_l2_ip6_node.function = acl_in_ip6_l2_node_fn_avx512;
-  else if (acl_in_ip6_l2_node_fn_avx2 && clib_cpu_supports_avx2 ())
-    acl_in_l2_ip6_node.function = acl_in_ip6_l2_node_fn_avx2;
-
-  if (acl_out_ip6_l2_node_fn_avx512 && clib_cpu_supports_avx512f ())
-    acl_out_l2_ip6_node.function = acl_out_ip6_l2_node_fn_avx512;
-  else if (acl_out_ip6_l2_node_fn_avx2 && clib_cpu_supports_avx2 ())
-    acl_out_l2_ip6_node.function = acl_out_ip6_l2_node_fn_avx2;
-
-  if (acl_in_ip4_fa_node_fn_avx512 && clib_cpu_supports_avx512f ())
-    acl_in_fa_ip4_node.function = acl_in_ip4_fa_node_fn_avx512;
-  else if (acl_in_ip4_fa_node_fn_avx2 && clib_cpu_supports_avx2 ())
-    acl_in_fa_ip4_node.function = acl_in_ip4_fa_node_fn_avx2;
-
-  if (acl_out_ip4_fa_node_fn_avx512 && clib_cpu_supports_avx512f ())
-    acl_out_fa_ip4_node.function = acl_out_ip4_fa_node_fn_avx512;
-  else if (acl_out_ip4_fa_node_fn_avx2 && clib_cpu_supports_avx2 ())
-    acl_out_fa_ip4_node.function = acl_out_ip4_fa_node_fn_avx2;
-
-  if (acl_in_ip6_fa_node_fn_avx512 && clib_cpu_supports_avx512f ())
-    acl_in_fa_ip6_node.function = acl_in_ip6_fa_node_fn_avx512;
-  else if (acl_in_ip6_fa_node_fn_avx2 && clib_cpu_supports_avx2 ())
-    acl_in_fa_ip6_node.function = acl_in_ip6_fa_node_fn_avx2;
-
-  if (acl_out_ip6_fa_node_fn_avx512 && clib_cpu_supports_avx512f ())
-    acl_out_fa_ip6_node.function = acl_out_ip6_fa_node_fn_avx512;
-  else if (acl_out_ip6_fa_node_fn_avx2 && clib_cpu_supports_avx2 ())
-    acl_out_fa_ip6_node.function = acl_out_ip6_fa_node_fn_avx2;
-
-}
-#endif
-
-
-
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 static u8 *
 format_fa_5tuple (u8 * s, va_list * args)
 {
@@ -549,7 +465,6 @@ static char *acl_fa_error_strings[] = {
 
 VLIB_REGISTER_NODE (acl_in_l2_ip6_node) =
 {
-  .function = acl_in_ip6_l2_node_fn,
   .name = "acl-plugin-in-ip6-l2",
   .vector_size = sizeof (u32),
   .format_trace = format_acl_plugin_trace,
@@ -565,7 +480,6 @@ VLIB_REGISTER_NODE (acl_in_l2_ip6_node) =
 
 VLIB_REGISTER_NODE (acl_in_l2_ip4_node) =
 {
-  .function = acl_in_ip4_l2_node_fn,
   .name = "acl-plugin-in-ip4-l2",
   .vector_size = sizeof (u32),
   .format_trace = format_acl_plugin_trace,
@@ -581,7 +495,6 @@ VLIB_REGISTER_NODE (acl_in_l2_ip4_node) =
 
 VLIB_REGISTER_NODE (acl_out_l2_ip6_node) =
 {
-  .function = acl_out_ip6_l2_node_fn,
   .name = "acl-plugin-out-ip6-l2",
   .vector_size = sizeof (u32),
   .format_trace = format_acl_plugin_trace,
@@ -597,7 +510,6 @@ VLIB_REGISTER_NODE (acl_out_l2_ip6_node) =
 
 VLIB_REGISTER_NODE (acl_out_l2_ip4_node) =
 {
-  .function = acl_out_ip4_l2_node_fn,
   .name = "acl-plugin-out-ip4-l2",
   .vector_size = sizeof (u32),
   .format_trace = format_acl_plugin_trace,
@@ -614,7 +526,6 @@ VLIB_REGISTER_NODE (acl_out_l2_ip4_node) =
 
 VLIB_REGISTER_NODE (acl_in_fa_ip6_node) =
 {
-  .function = acl_in_ip6_fa_node_fn,
   .name = "acl-plugin-in-ip6-fa",
   .vector_size = sizeof (u32),
   .format_trace = format_acl_plugin_trace,
@@ -637,7 +548,6 @@ VNET_FEATURE_INIT (acl_in_ip6_fa_feature, static) =
 
 VLIB_REGISTER_NODE (acl_in_fa_ip4_node) =
 {
-  .function = acl_in_ip4_fa_node_fn,
   .name = "acl-plugin-in-ip4-fa",
   .vector_size = sizeof (u32),
   .format_trace = format_acl_plugin_trace,
@@ -661,7 +571,6 @@ VNET_FEATURE_INIT (acl_in_ip4_fa_feature, static) =
 
 VLIB_REGISTER_NODE (acl_out_fa_ip6_node) =
 {
-  .function = acl_out_ip6_fa_node_fn,
   .name = "acl-plugin-out-ip6-fa",
   .vector_size = sizeof (u32),
   .format_trace = format_acl_plugin_trace,
@@ -684,7 +593,6 @@ VNET_FEATURE_INIT (acl_out_ip6_fa_feature, static) =
 
 VLIB_REGISTER_NODE (acl_out_fa_ip4_node) =
 {
-  .function = acl_out_ip4_fa_node_fn,
   .name = "acl-plugin-out-ip4-fa",
   .vector_size = sizeof (u32),
   .format_trace = format_acl_plugin_trace,
diff --git a/src/plugins/avf.am b/src/plugins/avf.am
index 76231a16bcb..1c7885fdc44 100644
--- a/src/plugins/avf.am
+++ b/src/plugins/avf.am
@@ -25,7 +25,7 @@ avf_plugin_la_SOURCES = \
 noinst_HEADERS += avf/avf.h
 
 if CPU_X86_64
-avf_multiversioning_files = 					\
+avf_multiversioning_sources = 					\
 	avf/input.c						\
 	avf/output.c
 
@@ -33,10 +33,10 @@ if CC_SUPPORTS_AVX2
 ###############################################################
 # AVX2
 ###############################################################
-libavf_plugin_avx2_la_SOURCES = $(avf_multiversioning_files)
+libavf_plugin_avx2_la_SOURCES = $(avf_multiversioning_sources)
 libavf_plugin_avx2_la_CFLAGS =				\
 	$(AM_CFLAGS)  @CPU_AVX2_FLAGS@				\
-	-DCLIB_MULTIARCH_VARIANT=avx2
+	-DCLIB_MARCH_VARIANT=avx2
 noinst_LTLIBRARIES += libavf_plugin_avx2.la
 avf_plugin_la_LIBADD += libavf_plugin_avx2.la
 endif
@@ -45,10 +45,10 @@ if CC_SUPPORTS_AVX512
 ###############################################################
 # AVX512
 ###############################################################
-libavf_plugin_avx512_la_SOURCES = $(avf_multiversioning_files)
+libavf_plugin_avx512_la_SOURCES = $(avf_multiversioning_sources)
 libavf_plugin_avx512_la_CFLAGS =				\
 	$(AM_CFLAGS) @CPU_AVX512_FLAGS@				\
-	-DCLIB_MULTIARCH_VARIANT=avx512
+	-DCLIB_MARCH_VARIANT=avx512
 noinst_LTLIBRARIES += libavf_plugin_avx512.la
 avf_plugin_la_LIBADD += libavf_plugin_avx512.la
 endif
diff --git a/src/plugins/avf/input.c b/src/plugins/avf/input.c
index 931bfa39de9..5b7f48a3ed7 100644
--- a/src/plugins/avf/input.c
+++ b/src/plugins/avf/input.c
@@ -375,8 +375,7 @@ avf_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
   return n_rx_packets;
 }
 
-uword
-CLIB_MULTIARCH_FN (avf_input) (vlib_main_t * vm, vlib_node_runtime_t * node,
+VLIB_NODE_FN (avf_input_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
 			       vlib_frame_t * frame)
 {
   u32 n_rx = 0;
@@ -399,10 +398,9 @@ CLIB_MULTIARCH_FN (avf_input) (vlib_main_t * vm, vlib_node_runtime_t * node,
   return n_rx;
 }
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 /* *INDENT-OFF* */
 VLIB_REGISTER_NODE (avf_input_node) = {
-  .function = avf_input,
   .name = "avf-input",
   .sibling_of = "device-input",
   .format_trace = format_avf_input_trace,
@@ -411,20 +409,6 @@ VLIB_REGISTER_NODE (avf_input_node) = {
   .n_errors = AVF_INPUT_N_ERROR,
   .error_strings = avf_input_error_strings,
 };
-
-#if __x86_64__
-vlib_node_function_t __clib_weak avf_input_avx512;
-vlib_node_function_t __clib_weak avf_input_avx2;
-static void __clib_constructor
-avf_input_multiarch_select (void)
-{
-  if (avf_input_avx512 && clib_cpu_supports_avx512f ())
-    avf_input_node.function = avf_input_avx512;
-  else if (avf_input_avx2 && clib_cpu_supports_avx2 ())
-    avf_input_node.function = avf_input_avx2;
-}
-
-#endif
 #endif
 
 /* *INDENT-ON* */
diff --git a/src/plugins/avf/output.c b/src/plugins/avf/output.c
index 8e5fa6a43f8..ec4d7f6c0a2 100644
--- a/src/plugins/avf/output.c
+++ b/src/plugins/avf/output.c
@@ -159,7 +159,7 @@ CLIB_MULTIARCH_FN (avf_interface_tx) (vlib_main_t * vm,
   return frame->n_vectors - n_left;
 }
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 #if __x86_64__
 vlib_node_function_t __clib_weak avf_interface_tx_avx512;
 vlib_node_function_t __clib_weak avf_interface_tx_avx2;
diff --git a/src/plugins/dpdk.am b/src/plugins/dpdk.am
index af7d96d7c17..6c15d0abd8e 100644
--- a/src/plugins/dpdk.am
+++ b/src/plugins/dpdk.am
@@ -57,7 +57,7 @@ dpdk_plugin_la_SOURCES =					\
 API_FILES += dpdk/api/dpdk.api
 
 if CPU_X86_64
-dpdk_multiversioning_files = 					\
+dpdk_multiversioning_sources = 					\
 	dpdk/buffer.c						\
 	dpdk/device/node.c					\
 	dpdk/device/device.c
@@ -66,10 +66,10 @@ if CC_SUPPORTS_AVX2
 ###############################################################
 # AVX2
 ###############################################################
-libdpdk_plugin_avx2_la_SOURCES = $(dpdk_multiversioning_files)
+libdpdk_plugin_avx2_la_SOURCES = $(dpdk_multiversioning_sources)
 libdpdk_plugin_avx2_la_CFLAGS =					\
 	$(AM_CFLAGS)  @CPU_AVX2_FLAGS@				\
-	-DCLIB_MULTIARCH_VARIANT=avx2
+	-DCLIB_MARCH_VARIANT=avx2
 noinst_LTLIBRARIES += libdpdk_plugin_avx2.la
 dpdk_plugin_la_LIBADD += libdpdk_plugin_avx2.la
 endif
@@ -78,10 +78,10 @@ if CC_SUPPORTS_AVX512
 ###############################################################
 # AVX512
 ###############################################################
-libdpdk_plugin_avx512_la_SOURCES = $(dpdk_multiversioning_files)
+libdpdk_plugin_avx512_la_SOURCES = $(dpdk_multiversioning_sources)
 libdpdk_plugin_avx512_la_CFLAGS =				\
 	$(AM_CFLAGS) @CPU_AVX512_FLAGS@				\
-	-DCLIB_MULTIARCH_VARIANT=avx512
+	-DCLIB_MARCH_VARIANT=avx512
 noinst_LTLIBRARIES += libdpdk_plugin_avx512.la
 dpdk_plugin_la_LIBADD += libdpdk_plugin_avx512.la
 endif
diff --git a/src/plugins/dpdk/buffer.c b/src/plugins/dpdk/buffer.c
index 452c47931ca..3324c049511 100644
--- a/src/plugins/dpdk/buffer.c
+++ b/src/plugins/dpdk/buffer.c
@@ -146,7 +146,7 @@ next:
     }
 }
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 static void
 del_free_list (vlib_main_t * vm, vlib_buffer_free_list_t * f)
 {
@@ -436,7 +436,7 @@ CLIB_MULTIARCH_FN (dpdk_buffer_free_no_next) (vlib_main_t * vm, u32 * buffers,
 			   0);
 }
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 static void
 dpdk_packet_template_init (vlib_main_t * vm,
 			   void *vt,
diff --git a/src/plugins/dpdk/device/device.c b/src/plugins/dpdk/device/device.c
index 5409fe4f50e..fe659fbeaf1 100644
--- a/src/plugins/dpdk/device/device.c
+++ b/src/plugins/dpdk/device/device.c
@@ -37,7 +37,7 @@ typedef enum
     DPDK_TX_FUNC_N_ERROR,
 } dpdk_tx_func_error_t;
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 static char *dpdk_tx_func_error_strings[] = {
 #define _(n,s) s,
   foreach_dpdk_tx_func_error
@@ -517,7 +517,7 @@ CLIB_MULTIARCH_FN (dpdk_interface_tx) (vlib_main_t * vm,
   return tx_pkts;
 }
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 static void
 dpdk_clear_hw_interface_counters (u32 instance)
 {
@@ -695,7 +695,7 @@ dpdk_interface_tx_multiarch_select (void)
 
 #define UP_DOWN_FLAG_EVENT 1
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 uword
 admin_up_down_process (vlib_main_t * vm,
 		       vlib_node_runtime_t * rt, vlib_frame_t * f)
diff --git a/src/plugins/dpdk/device/node.c b/src/plugins/dpdk/device/node.c
index daccf733372..7f4b2cb8145 100644
--- a/src/plugins/dpdk/device/node.c
+++ b/src/plugins/dpdk/device/node.c
@@ -28,7 +28,7 @@
 
 #include <dpdk/device/dpdk_priv.h>
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 static char *dpdk_error_strings[] = {
 #define _(n,s) s,
   foreach_dpdk_error
@@ -636,8 +636,7 @@ dpdk_device_input (vlib_main_t * vm, dpdk_main_t * dm, dpdk_device_t * xd,
   return n_rx_packets;
 }
 
-uword CLIB_CPU_OPTIMIZED
-CLIB_MULTIARCH_FN (dpdk_input) (vlib_main_t * vm, vlib_node_runtime_t * node,
+VLIB_NODE_FN (dpdk_input_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
 				vlib_frame_t * f)
 {
   dpdk_main_t *dm = &dpdk_main;
@@ -666,10 +665,9 @@ CLIB_MULTIARCH_FN (dpdk_input) (vlib_main_t * vm, vlib_node_runtime_t * node,
   return n_rx_packets;
 }
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 /* *INDENT-OFF* */
 VLIB_REGISTER_NODE (dpdk_input_node) = {
-  .function = dpdk_input,
   .type = VLIB_NODE_TYPE_INPUT,
   .name = "dpdk-input",
   .sibling_of = "device-input",
@@ -684,20 +682,6 @@ VLIB_REGISTER_NODE (dpdk_input_node) = {
   .error_strings = dpdk_error_strings,
 };
 /* *INDENT-ON* */
-
-vlib_node_function_t __clib_weak dpdk_input_avx512;
-vlib_node_function_t __clib_weak dpdk_input_avx2;
-
-#if __x86_64__
-static void __clib_constructor
-dpdk_input_multiarch_select (void)
-{
-  if (dpdk_input_avx512 && clib_cpu_supports_avx512f ())
-    dpdk_input_node.function = dpdk_input_avx512;
-  else if (dpdk_input_avx2 && clib_cpu_supports_avx2 ())
-    dpdk_input_node.function = dpdk_input_avx2;
-}
-#endif
 #endif
 
 /*
diff --git a/src/plugins/memif.am b/src/plugins/memif.am
index 6e9aa779155..33a378f4a3f 100644
--- a/src/plugins/memif.am
+++ b/src/plugins/memif.am
@@ -36,7 +36,7 @@ nobase_apiinclude_HEADERS +=              \
 API_FILES += memif/memif.api
 
 if CPU_X86_64
-memif_multiversioning_files = 					\
+memif_multiversioning_sources = 					\
 	memif/node.c						\
 	memif/device.c
 
@@ -44,10 +44,10 @@ if CC_SUPPORTS_AVX2
 ###############################################################
 # AVX2
 ###############################################################
-libmemif_plugin_avx2_la_SOURCES = $(memif_multiversioning_files)
+libmemif_plugin_avx2_la_SOURCES = $(memif_multiversioning_sources)
 libmemif_plugin_avx2_la_CFLAGS =				\
 	$(AM_CFLAGS)  @CPU_AVX2_FLAGS@				\
-	-DCLIB_MULTIARCH_VARIANT=avx2
+	-DCLIB_MARCH_VARIANT=avx2
 noinst_LTLIBRARIES += libmemif_plugin_avx2.la
 memif_plugin_la_LIBADD += libmemif_plugin_avx2.la
 endif
@@ -56,10 +56,10 @@ if CC_SUPPORTS_AVX512
 ###############################################################
 # AVX512
 ###############################################################
-libmemif_plugin_avx512_la_SOURCES = $(memif_multiversioning_files)
+libmemif_plugin_avx512_la_SOURCES = $(memif_multiversioning_sources)
 libmemif_plugin_avx512_la_CFLAGS =				\
 	$(AM_CFLAGS) @CPU_AVX512_FLAGS@				\
-	-DCLIB_MULTIARCH_VARIANT=avx512
+	-DCLIB_MARCH_VARIANT=avx512
 noinst_LTLIBRARIES += libmemif_plugin_avx512.la
 memif_plugin_la_LIBADD += libmemif_plugin_avx512.la
 endif
diff --git a/src/plugins/memif/device.c b/src/plugins/memif/device.c
index c70a4ea3549..14b9d547234 100644
--- a/src/plugins/memif/device.c
+++ b/src/plugins/memif/device.c
@@ -46,7 +46,7 @@ static __clib_unused char *memif_tx_func_error_strings[] = {
 #undef _
 };
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 u8 *
 format_memif_device_name (u8 * s, va_list * args)
 {
@@ -500,7 +500,7 @@ memif_subif_add_del_function (vnet_main_t * vnm,
   return 0;
 }
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 /* *INDENT-OFF* */
 VNET_DEVICE_CLASS (memif_device_class) = {
   .name = "memif",
diff --git a/src/plugins/memif/node.c b/src/plugins/memif/node.c
index 029e25d814a..6e2807a2054 100644
--- a/src/plugins/memif/node.c
+++ b/src/plugins/memif/node.c
@@ -854,10 +854,10 @@ done:
   return n_rx_packets;
 }
 
-uword
-CLIB_MULTIARCH_FN (memif_input_fn) (vlib_main_t * vm,
-				    vlib_node_runtime_t * node,
-				    vlib_frame_t * frame)
+
+VLIB_NODE_FN (memif_input_node) (vlib_main_t * vm,
+				 vlib_node_runtime_t * node,
+				 vlib_frame_t * frame)
 {
   u32 n_rx = 0;
   memif_main_t *mm = &memif_main;
@@ -910,10 +910,9 @@ CLIB_MULTIARCH_FN (memif_input_fn) (vlib_main_t * vm,
   return n_rx;
 }
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 /* *INDENT-OFF* */
 VLIB_REGISTER_NODE (memif_input_node) = {
-  .function = memif_input_fn,
   .name = "memif-input",
   .sibling_of = "device-input",
   .format_trace = format_memif_input_trace,
@@ -922,20 +921,6 @@ VLIB_REGISTER_NODE (memif_input_node) = {
   .n_errors = MEMIF_INPUT_N_ERROR,
   .error_strings = memif_input_error_strings,
 };
-
-vlib_node_function_t __clib_weak memif_input_fn_avx512;
-vlib_node_function_t __clib_weak memif_input_fn_avx2;
-
-#if __x86_64__
-static void __clib_constructor
-memif_input_multiarch_select (void)
-{
-  if (memif_input_fn_avx512 && clib_cpu_supports_avx512f ())
-    memif_input_node.function = memif_input_fn_avx512;
-  else if (memif_input_fn_avx2 && clib_cpu_supports_avx2 ())
-    memif_input_node.function = memif_input_fn_avx2;
-}
-#endif
 #endif
 
 /* *INDENT-ON* */
diff --git a/src/vlib/node.c b/src/vlib/node.c
index 17dd2ea87d4..cc1732bb90d 100644
--- a/src/vlib/node.c
+++ b/src/vlib/node.c
@@ -324,6 +324,26 @@ register_node (vlib_main_t * vm, vlib_node_registration_t * r)
       ASSERT (VLIB_NODE_TYPE_INTERNAL == zero.type);
     }
 
+  if (r->node_fn_registrations)
+    {
+      vlib_node_fn_registration_t *fnr = r->node_fn_registrations;
+      int priority = -1;
+
+      /* to avoid confusion, please remove ".function " statiement from
+         CLIB_NODE_REGISTRATION() if using function function candidates */
+      ASSERT (r->function == 0);
+
+      while (fnr)
+	{
+	  if (fnr->priority > priority)
+	    {
+	      priority = fnr->priority;
+	      r->function = fnr->function;
+	    }
+	  fnr = fnr->next_registration;
+	}
+    }
+
   ASSERT (r->function != 0);
 
   n = clib_mem_alloc_no_fail (sizeof (n[0]));
diff --git a/src/vlib/node.h b/src/vlib/node.h
index 7177bebeffb..67eaea3f0e8 100644
--- a/src/vlib/node.h
+++ b/src/vlib/node.h
@@ -75,11 +75,21 @@ typedef enum
   VLIB_N_NODE_TYPE,
 } vlib_node_type_t;
 
+typedef struct _vlib_node_fn_registration
+{
+  vlib_node_function_t *function;
+  int priority;
+  struct _vlib_node_fn_registration *next_registration;
+} vlib_node_fn_registration_t;
+
 typedef struct _vlib_node_registration
 {
   /* Vector processing function for this node. */
   vlib_node_function_t *function;
 
+  /* Node function candidate registration with priority */
+  vlib_node_fn_registration_t *node_fn_registrations;
+
   /* Node name. */
   char *name;
 
@@ -160,6 +170,24 @@ static void __vlib_rm_node_registration_##x (void)                      \
 }                                                                       \
 __VA_ARGS__ vlib_node_registration_t x
 
+#define VLIB_NODE_FN(node)						\
+uword CLIB_MARCH_SFX (node##_fn)();					\
+static vlib_node_fn_registration_t					\
+  CLIB_MARCH_SFX(node##_fn_registration) =				\
+  { .function = &CLIB_MARCH_SFX (node##_fn), };				\
+									\
+static void __clib_constructor						\
+CLIB_MARCH_SFX (node##_multiarch_register) (void)			\
+{									\
+  extern vlib_node_registration_t node;					\
+  vlib_node_fn_registration_t *r;					\
+  r = & CLIB_MARCH_SFX (node##_fn_registration);			\
+  r->priority = CLIB_MARCH_FN_PRIORITY();				\
+  r->next_registration = node.node_fn_registrations;			\
+  node.node_fn_registrations = r;					\
+}									\
+uword CLIB_CPU_OPTIMIZED CLIB_MARCH_SFX (node##_fn)
+
 #if CLIB_DEBUG > 0
 #define VLIB_NODE_FUNCTION_CLONE_TEMPLATE(arch, fn)
 #define VLIB_NODE_FUNCTION_MULTIARCH_CLONE(fn)
diff --git a/src/vnet.am b/src/vnet.am
index 47159fc4591..86f3ad00f33 100644
--- a/src/vnet.am
+++ b/src/vnet.am
@@ -21,7 +21,7 @@ libvnet_la_DEPENDENCIES = \
 	libsvmdb.la		\
 	libsvm.la 		\
 	libvlibmemory.la
-libvnet_multiversioning_files =
+libvnet_multiversioning_sources =
 
 libvnet_la_LIBADD = $(libvnet_la_DEPENDENCIES) -lm -lpthread -ldl -lrt
 
@@ -128,7 +128,7 @@ libvnet_la_SOURCES +=				\
  vnet/ethernet/p2p_ethernet_input.c		\
  vnet/ethernet/p2p_ethernet_api.c
 
-libvnet_multiversioning_files +=		\
+libvnet_multiversioning_sources +=		\
  vnet/l2/l2_output.c
 
 nobase_include_HEADERS +=			\
@@ -294,7 +294,7 @@ nobase_include_HEADERS +=			\
   vnet/bonding/node.h                           \
   vnet/bonding/bond.api.h
 
-libvnet_multiversioning_files = 		\
+libvnet_multiversioning_sources +=		\
   vnet/bonding/node.c
 
 API_FILES += vnet/bonding/bond.api
@@ -416,7 +416,7 @@ API_FILES += 					\
  vnet/ip/rd_cp.api                              \
  vnet/ip/punt.api
 
-libvnet_multiversioning_files = 		\
+libvnet_multiversioning_sources +=		\
  vnet/ip/ip4_input.c
 
 ########################################
@@ -1286,10 +1286,10 @@ if CC_SUPPORTS_AVX2
 ###############################################################
 # AVX2
 ###############################################################
-libvnet_avx2_la_SOURCES = $(libvnet_multiversioning_files)
+libvnet_avx2_la_SOURCES = $(libvnet_multiversioning_sources)
 libvnet_avx2_la_CFLAGS =			\
 	$(AM_CFLAGS)  @CPU_AVX2_FLAGS@		\
-	-DCLIB_MULTIARCH_VARIANT=avx2
+	-DCLIB_MARCH_VARIANT=avx2
 noinst_LTLIBRARIES += libvnet_avx2.la
 libvnet_la_LIBADD += libvnet_avx2.la
 endif
@@ -1298,10 +1298,10 @@ if CC_SUPPORTS_AVX512
 ###############################################################
 # AVX512
 ###############################################################
-libvnet_avx512_la_SOURCES = $(libvnet_multiversioning_files)
+libvnet_avx512_la_SOURCES = $(libvnet_multiversioning_sources)
 libvnet_avx512_la_CFLAGS =			\
 	$(AM_CFLAGS) @CPU_AVX512_FLAGS@		\
-	-DCLIB_MULTIARCH_VARIANT=avx512
+	-DCLIB_MARCH_VARIANT=avx512
 noinst_LTLIBRARIES += libvnet_avx512.la
 libvnet_la_LIBADD += libvnet_avx512.la
 endif
diff --git a/src/vnet/bonding/node.c b/src/vnet/bonding/node.c
index df2da5c66fc..7d2acf7d0f7 100644
--- a/src/vnet/bonding/node.c
+++ b/src/vnet/bonding/node.c
@@ -36,7 +36,7 @@ typedef enum
     BOND_INPUT_N_ERROR,
 } bond_input_error_t;
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 static char *bond_input_error_strings[] = {
 #define _(n,s) s,
   foreach_bond_input_error
@@ -172,10 +172,9 @@ bond_update_next (vlib_main_t * vm, vlib_node_runtime_t * node,
   vnet_feature_next ( /* not used */ 0, next_index, b);
 }
 
-uword CLIB_CPU_OPTIMIZED
-CLIB_MULTIARCH_FN (bond_input_fn) (vlib_main_t * vm,
-				   vlib_node_runtime_t * node,
-				   vlib_frame_t * frame)
+VLIB_NODE_FN (bond_input_node) (vlib_main_t * vm,
+				vlib_node_runtime_t * node,
+				vlib_frame_t * frame)
 {
   u16 thread_index = vlib_get_thread_index ();
   u32 *from, n_left;
@@ -367,7 +366,7 @@ CLIB_MULTIARCH_FN (bond_input_fn) (vlib_main_t * vm,
   return frame->n_vectors;
 }
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 static clib_error_t *
 bond_input_init (vlib_main_t * vm)
 {
@@ -376,7 +375,6 @@ bond_input_init (vlib_main_t * vm)
 
 /* *INDENT-OFF* */
 VLIB_REGISTER_NODE (bond_input_node) = {
-  .function = bond_input_fn,
   .name = "bond-input",
   .vector_size = sizeof (u32),
   .format_buffer = format_ethernet_header_with_length,
@@ -391,20 +389,6 @@ VLIB_REGISTER_NODE (bond_input_node) = {
   }
 };
 
-#if __x86_64__
-vlib_node_function_t __clib_weak bond_input_fn_avx512;
-vlib_node_function_t __clib_weak bond_input_fn_avx2;
-static void __clib_constructor
-bond_input_multiarch_select (void)
-{
-  if (bond_input_fn_avx512 && clib_cpu_supports_avx512f ())
-    bond_input_node.function = bond_input_fn_avx512;
-  else if (bond_input_fn_avx2 && clib_cpu_supports_avx2 ())
-    bond_input_node.function = bond_input_fn_avx2;
-}
-#endif
-
-
 VLIB_INIT_FUNCTION (bond_input_init);
 
 VNET_FEATURE_INIT (bond_input, static) =
diff --git a/src/vnet/ip/ip4_input.c b/src/vnet/ip/ip4_input.c
index 7af2bee187a..b476f95ab8b 100644
--- a/src/vnet/ip/ip4_input.c
+++ b/src/vnet/ip/ip4_input.c
@@ -47,7 +47,7 @@ typedef struct
   u8 packet_data[64];
 } ip4_input_trace_t;
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 static u8 *
 format_ip4_input_trace (u8 * s, va_list * va)
 {
@@ -292,22 +292,20 @@ ip4_input_inline (vlib_main_t * vm,
       <code> vnet_get_config_data (... &next0 ...); </code>
       or @c error-drop
 */
-uword CLIB_CPU_OPTIMIZED
-CLIB_MULTIARCH_FN (ip4_input) (vlib_main_t * vm, vlib_node_runtime_t * node,
+VLIB_NODE_FN (ip4_input_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
 			       vlib_frame_t * frame)
 {
   return ip4_input_inline (vm, node, frame, /* verify_checksum */ 1);
 }
 
-uword CLIB_CPU_OPTIMIZED
-CLIB_MULTIARCH_FN (ip4_input_no_checksum) (vlib_main_t * vm,
+VLIB_NODE_FN (ip4_input_no_checksum_node) (vlib_main_t * vm,
 					   vlib_node_runtime_t * node,
 					   vlib_frame_t * frame)
 {
   return ip4_input_inline (vm, node, frame, /* verify_checksum */ 0);
 }
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 char *ip4_error_strings[] = {
 #define _(sym,string) string,
   foreach_ip4_error
@@ -316,7 +314,6 @@ char *ip4_error_strings[] = {
 
 /* *INDENT-OFF* */
 VLIB_REGISTER_NODE (ip4_input_node) = {
-  .function = ip4_input,
   .name = "ip4-input",
   .vector_size = sizeof (u32),
 
@@ -337,8 +334,7 @@ VLIB_REGISTER_NODE (ip4_input_node) = {
   .format_trace = format_ip4_input_trace,
 };
 
-VLIB_REGISTER_NODE (ip4_input_no_checksum_node,static) = {
-  .function = ip4_input_no_checksum,
+VLIB_REGISTER_NODE (ip4_input_no_checksum_node) = {
   .name = "ip4-input-no-checksum",
   .vector_size = sizeof (u32),
 
@@ -357,26 +353,6 @@ VLIB_REGISTER_NODE (ip4_input_no_checksum_node,static) = {
 };
 /* *INDENT-ON* */
 
-#if __x86_64__
-vlib_node_function_t __clib_weak ip4_input_avx512;
-vlib_node_function_t __clib_weak ip4_input_avx2;
-vlib_node_function_t __clib_weak ip4_input_no_checksum_avx512;
-vlib_node_function_t __clib_weak ip4_input_no_checksum_avx2;
-static void __clib_constructor
-ip4_input_multiarch_select (void)
-{
-  if (ip4_input_no_checksum_avx512 && clib_cpu_supports_avx512f ())
-    ip4_input_no_checksum_node.function = ip4_input_no_checksum_avx512;
-  else if (ip4_input_no_checksum_avx2 && clib_cpu_supports_avx2 ())
-    ip4_input_no_checksum_node.function = ip4_input_no_checksum_avx2;
-
-  if (ip4_input_avx512 && clib_cpu_supports_avx512f ())
-    ip4_input_node.function = ip4_input_avx512;
-  else if (ip4_input_avx2 && clib_cpu_supports_avx2 ())
-    ip4_input_node.function = ip4_input_avx2;
-}
-#endif
-
 static clib_error_t *
 ip4_init (vlib_main_t * vm)
 {
diff --git a/src/vnet/l2/l2_output.c b/src/vnet/l2/l2_output.c
index 306049bd0db..1fa042ffd54 100644
--- a/src/vnet/l2/l2_output.c
+++ b/src/vnet/l2/l2_output.c
@@ -27,7 +27,7 @@
 #include <vnet/l2/l2_output.h>
 
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 /* Feature graph node names */
 static char *l2output_feat_names[] = {
 #define _(sym,name) name,
@@ -76,7 +76,7 @@ typedef struct
   u8 raw[12];			/* raw data */
 } l2output_trace_t;
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 /* packet trace format function */
 static u8 *
 format_l2output_trace (u8 * s, va_list * args)
@@ -275,10 +275,9 @@ l2output_process_batch (vlib_main_t * vm, vlib_node_runtime_t * node,
 				   l2_efp, l2_vtr, l2_pbb, 1, 1);
 }
 
-uword CLIB_CPU_OPTIMIZED
-CLIB_MULTIARCH_FN (l2output_node_fn) (vlib_main_t * vm,
-				      vlib_node_runtime_t * node,
-				      vlib_frame_t * frame)
+VLIB_NODE_FN (l2output_node) (vlib_main_t * vm,
+			      vlib_node_runtime_t * node,
+			      vlib_frame_t * frame)
 {
   u32 n_left, *from;
   l2output_main_t *msm = &l2output_main;
@@ -431,10 +430,9 @@ CLIB_MULTIARCH_FN (l2output_node_fn) (vlib_main_t * vm,
   return frame->n_vectors;
 }
 
-#ifndef CLIB_MULTIARCH_VARIANT
+#ifndef CLIB_MARCH_VARIANT
 /* *INDENT-OFF* */
 VLIB_REGISTER_NODE (l2output_node) = {
-  .function = l2output_node_fn,
   .name = "l2-output",
   .vector_size = sizeof (u32),
   .format_trace = format_l2output_trace,
@@ -451,20 +449,6 @@ VLIB_REGISTER_NODE (l2output_node) = {
         [L2OUTPUT_NEXT_BAD_INTF] = "l2-output-bad-intf",
   },
 };
-
-#if __x86_64__
-vlib_node_function_t __clib_weak l2output_node_fn_avx512;
-vlib_node_function_t __clib_weak l2output_node_fn_avx2;
-static void __clib_constructor
-l2output_multiarch_select (void)
-{
-  if (l2output_node_fn_avx512 && clib_cpu_supports_avx512f ())
-    l2output_node.function = l2output_node_fn_avx512;
-  else if (l2output_node_fn_avx2 && clib_cpu_supports_avx2 ())
-    l2output_node.function = l2output_node_fn_avx2;
-}
-#endif
-
 /* *INDENT-ON* */
 
 
diff --git a/src/vppinfra/cpu.h b/src/vppinfra/cpu.h
index f670f3e964b..110815c9f87 100644
--- a/src/vppinfra/cpu.h
+++ b/src/vppinfra/cpu.h
@@ -51,14 +51,16 @@
   return & fn;                                                         \
 }
 
-#ifdef CLIB_MULTIARCH_VARIANT
+#ifdef CLIB_MARCH_VARIANT
 #define __CLIB_MULTIARCH_FN(a,b) a##_##b
 #define _CLIB_MULTIARCH_FN(a,b) __CLIB_MULTIARCH_FN(a,b)
-#define CLIB_MULTIARCH_FN(fn) _CLIB_MULTIARCH_FN(fn,CLIB_MULTIARCH_VARIANT)
+#define CLIB_MULTIARCH_FN(fn) _CLIB_MULTIARCH_FN(fn,CLIB_MARCH_VARIANT)
 #else
 #define CLIB_MULTIARCH_FN(fn) fn
 #endif
 
+#define CLIB_MARCH_SFX CLIB_MULTIARCH_FN
+
 #define foreach_x86_64_flags \
 _ (sse3,     1, ecx, 0)   \
 _ (ssse3,    1, ecx, 9)   \
@@ -165,6 +167,27 @@ clib_cpu_supports_aes ()
 #endif
 }
 
+static inline int
+clib_cpu_march_priority_avx512 ()
+{
+  if (clib_cpu_supports_avx512f ())
+    return 20;
+  return -1;
+}
+
+static inline int
+clib_cpu_march_priority_avx2 ()
+{
+  if (clib_cpu_supports_avx2 ())
+    return 10;
+  return -1;
+}
+
+#ifdef CLIB_MARCH_VARIANT
+#define CLIB_MARCH_FN_PRIORITY() CLIB_MARCH_SFX(clib_cpu_march_priority)()
+#else
+#define CLIB_MARCH_FN_PRIORITY() 0
+#endif
 #endif /* included_clib_cpu_h */
 
 format_function_t format_cpu_uarch;