build: rework x86 CPU variants

Type: improvement
Change-Id: Ief243f88e654e578ef9b8060fcf535b364aececb
Signed-off-by: Damjan Marion <damarion@cisco.com>

6 files changed, 31 insertions, 13 deletions

diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index 16ae7d5451d..1564fd0ad8c 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -13,7 +13,7 @@
 
 cmake_minimum_required(VERSION 3.5 FATAL_ERROR)
 
-set(CMAKE_C_COMPILER_NAMES clang-9 cc)
+set(CMAKE_C_COMPILER_NAMES clang-10 clang-9 gcc-9 cc)
 
 project(vpp C)
 
diff --git a/src/cmake/cpu.cmake b/src/cmake/cpu.cmake
index 0ce8bea31ec..0e47f3325cf 100644
--- a/src/cmake/cpu.cmake
+++ b/src/cmake/cpu.cmake
@@ -76,17 +76,21 @@ endif()
 ##############################################################################
 if(CMAKE_SYSTEM_PROCESSOR MATCHES "amd64.*|x86_64.*|AMD64.*")
   set(CMAKE_C_FLAGS "-march=corei7 -mtune=corei7-avx ${CMAKE_C_FLAGS}")
-  check_c_compiler_flag("-march=core-avx2" compiler_flag_march_core_avx2)
-  if(compiler_flag_march_core_avx2)
-    list(APPEND MARCH_VARIANTS "avx2\;-march=core-avx2 -mtune=core-avx2")
+  check_c_compiler_flag("-march=haswell" compiler_flag_march_haswell)
+  if(compiler_flag_march_haswell)
+    list(APPEND MARCH_VARIANTS "hsw\;-march=haswell -mtune=haswell")
   endif()
   if (GNU_ASSEMBLER_AVX512_BUG)
      message(WARNING "AVX-512 multiarch variant(s) disabled due to GNU Assembler bug")
   else()
+    check_c_compiler_flag("-mprefer-vector-width=256" compiler_flag_mprefer_vector_width)
     check_c_compiler_flag("-march=skylake-avx512" compiler_flag_march_skylake_avx512)
     check_c_compiler_flag("-march=icelake-client" compiler_flag_march_icelake_client)
-    if(compiler_flag_march_skylake_avx512)
-      list(APPEND MARCH_VARIANTS "avx512\;-march=skylake-avx512 -mtune=skylake-avx512")
+    if(compiler_flag_march_skylake_avx512 AND compiler_flag_mprefer_vector_width)
+      list(APPEND MARCH_VARIANTS "skx\;-march=skylake-avx512 -mtune=skylake-avx512 -mprefer-vector-width=256")
+    endif()
+    if(compiler_flag_march_icelake_client AND compiler_flag_mprefer_vector_width)
+      list(APPEND MARCH_VARIANTS "icl\;-march=icelake-client -mtune=icelake-client -mprefer-vector-width=512")
     endif()
   endif()
 elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64.*|AARCH64.*)")
diff --git a/src/vlib/node_cli.c b/src/vlib/node_cli.c
index 750f69e5d60..7cfc4a36f8c 100644
--- a/src/vlib/node_cli.c
+++ b/src/vlib/node_cli.c
@@ -623,7 +623,7 @@ show_node (vlib_main_t * vm, unformat_input_t * input,
 	  if (vec_len (s) == 0)
 	    s = format (s, "\n    %-15s  %=8s  %6s",
 			"Name", "Priority", "Active");
-	  s = format (s, "\n    %-15s  %=8u  %=6s", fnr->name, fnr->priority,
+	  s = format (s, "\n    %-15s  %8d  %=6s", fnr->name, fnr->priority,
 		      fnr->function == n->function ? "yes" : "");
 	  fnr = fnr->next_registration;
 	}
diff --git a/src/vppinfra/cpu.h b/src/vppinfra/cpu.h
index 7bd5dfe9722..dc73c90ca34 100644
--- a/src/vppinfra/cpu.h
+++ b/src/vppinfra/cpu.h
@@ -132,6 +132,9 @@ _ (x86_aes,  1, ecx, 25)  \
 _ (sha,      7, ebx, 29)  \
 _ (vaes,     7, ecx, 9)   \
 _ (vpclmulqdq, 7, ecx, 10)   \
+_ (avx512_vnni, 7, ecx, 11)   \
+_ (avx512_bitalg, 7, ecx, 12)   \
+_ (avx512_vpopcntdq, 7, ecx, 14)   \
 _ (invariant_tsc, 0x80000007, edx, 8)
 
 
@@ -245,15 +248,23 @@ clib_cpu_supports_aes ()
 }
 
 static inline int
-clib_cpu_march_priority_avx512 ()
+clib_cpu_march_priority_icl ()
+{
+  if (clib_cpu_supports_avx512_bitalg ())
+    return 200;
+  return -1;
+}
+
+static inline int
+clib_cpu_march_priority_skx ()
 {
   if (clib_cpu_supports_avx512f ())
-    return 20;
+    return 100;
   return -1;
 }
 
 static inline int
-clib_cpu_march_priority_avx2 ()
+clib_cpu_march_priority_hsw ()
 {
   if (clib_cpu_supports_avx2 ())
     return 50;
diff --git a/src/vppinfra/string.h b/src/vppinfra/string.h
index a3db9264cac..3dafe899a8f 100644
--- a/src/vppinfra/string.h
+++ b/src/vppinfra/string.h
@@ -71,7 +71,7 @@ void clib_memswap (void *_a, void *_b, uword bytes);
  * so don't let it anywhere near them.
  */
 #ifndef __COVERITY__
-#if __AVX512F__
+#if __AVX512BITALG__
 #include <vppinfra/memcpy_avx512.h>
 #elif __AVX2__
 #include <vppinfra/memcpy_avx2.h>
diff --git a/src/vppinfra/vector.h b/src/vppinfra/vector.h
index 906d8d8fbfd..8b08db22124 100644
--- a/src/vppinfra/vector.h
+++ b/src/vppinfra/vector.h
@@ -66,7 +66,7 @@
 #endif
 #endif
 
-#if defined (__AVX512F__)
+#if defined (__AVX512BITALG__)
 #define CLIB_HAVE_VEC512
 #endif
 
@@ -168,7 +168,10 @@ typedef u64 u64x _vector_size (8);
 #include <vppinfra/vector_avx2.h>
 #endif
 
-#if defined (__AVX512F__)
+#if defined (__AVX512BITALG__)
+/* Due to power level transition issues, we don't preffer AVX-512 on
+   Skylake X and CascadeLake CPUs, AVX512BITALG is introduced on
+   icelake CPUs  */
 #include <vppinfra/vector_avx512.h>
 #endif