Add cmake option to enable avx512

cmake/ConfigureCompiler.cmake

@@ -209,6 +209,25 @@ else()
   #   -mavx
   #   -msse4.2)
 
+  # Tentatively re-enabling vector instructions
+  set(USE_AVX512F OFF CACHE BOOL "Enable AVX 512F instructions")
+  if (USE_AVX512F)
+    add_compile_options(-mavx512f)
+  endif()
+  set(USE_AVX ON CACHE BOOL "Enable AVX instructions")
+  if (USE_AVX)
+    add_compile_options(-mavx)
+  endif()
+
+  # Intentionally using builtin memcpy.  G++ does a good job on small memcpy's when the size is known at runtime.
+  # If the size is not known, then it falls back on the memcpy that's available at runtime (rte_memcpy, as of this
+  # writing; see flow.cpp).
+  #
+  # The downside of the builtin memcpy is that it's slower at large copies, so if we spend a lot of time on large
+  # copies of sizes that are known at compile time, this might not be a win.  See the output of performance/memcpy
+  # for more information.
+  #add_compile_options(-fno-builtin-memcpy)
+
   if (USE_VALGRIND)
     add_compile_options(-DVALGRIND -DUSE_VALGRIND)
   endif()
@@ -243,12 +262,6 @@ else()
       -Wno-tautological-pointer-compare
       -Wno-format
       -Woverloaded-virtual)
-    set(USE_AVX ON CACHE BOOL "Enable AVX instructions")
-    if (USE_AVX)
-      add_compile_options(-mavx)
-    else()
-      add_compile_options(-msse4)
-    endif()
     if (USE_CCACHE)
       add_compile_options(
         -Wno-register
@@ -260,20 +273,6 @@ else()
   endif()
   if (GCC)
     add_compile_options(-Wno-pragmas)
-    set(USE_AVX ON CACHE BOOL "Enable AVX instructions")
-    if (USE_AVX)
-      add_compile_options(-mavx)
-    else()
-      add_compile_options(-msse4)
-    endif()
-    # Intentionally using builtin memcpy.  G++ does a good job on small memcpy's when the size is known at runtime.
-    # If the size is not known, then it falls back on the memcpy that's available at runtime (rte_memcpy, as of this
-    # writing; see flow.cpp).
-    #
-    # The downside of the builtin memcpy is that it's slower at large copies, so if we spend a lot of time on large
-    # copies of sizes that are known at compile time, this might not be a win.  See the output of performance/memcpy
-    # for more information.
-    #add_compile_options(-fno-builtin-memcpy)
     # Otherwise `state [[maybe_unused]] int x;` will issue a warning.
     # https://stackoverflow.com/questions/50646334/maybe-unused-on-member-variable-gcc-warns-incorrectly-that-attribute-is
     add_compile_options(-Wno-attributes)

flow/flow.cpp

@@ -26,6 +26,7 @@
 #include <stdarg.h>
 #include <cinttypes>
 
+#if defined (__linux__) || defined (__FreeBSD__)
 // For benchmarking; need a version of rte_memcpy that doesn't live in the same compilation unit as the test.
 void * rte_memcpy_noinline(void *__restrict __dest, const void *__restrict __src, size_t __n) {
 	return rte_memcpy(__dest, __src, __n);
@@ -36,6 +37,7 @@ __attribute__((visibility ("default"))) void *memcpy (void *__restrict __dest, c
 	// folly_memcpy is faster for small copies, but rte seems to win out in most other circumstances
 	return rte_memcpy(__dest, __src, __n);
 }
+#endif // defined (__linux__) || defined (__FreeBSD__)
 
 INetwork *g_network = 0;
 

flow/rte_memcpy.h

@@ -28,6 +28,8 @@ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
 
 #include <flow/Platform.h>
 
+#if defined (__linux__) || defined (__FreeBSD__)
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -50,8 +52,9 @@ extern "C" {
 static force_inline void *
 rte_memcpy(void *dst, const void *src, size_t n);
 
-#ifdef __AVX__
-//#define RTE_MACHINE_CPUFLAG_AVX512F  -- our g++ is too old for this
+#ifdef __AVX512F__
+#define RTE_MACHINE_CPUFLAG_AVX512F
+#elif defined(__AVX__)
 #define RTE_MACHINE_CPUFLAG_AVX2
 #endif
 
@@ -905,4 +908,6 @@ rte_rdtsc(void)
 }
 #endif
 
+#endif /* defined (__linux__) || defined (__FreeBSD__) */
+
 #endif /* _RTE_MEMCPY_X86_64_H_ */

flow/test_memcpy_perf.cpp

@@ -6,13 +6,13 @@
 #include <stdio.h>
 #include <string.h>
 #include <stdlib.h>
-#include <sys/time.h>
 
 #include "flow/rte_memcpy.h"
 #include "flow/IRandom.h"
 #include "flow/UnitTest.h"
 #include "flow/flow.h"
 
+#if defined (__linux__) || defined (__FreeBSD__)
 extern "C" {
 	void* folly_memcpy(void* dst, const void* src, uint32_t length);
 }
@@ -352,4 +352,6 @@ TEST_CASE("performance/memcpy/rte") {
 	return Void();
 }
 
-void forceLinkMemcpyPerfTests() {}
+#endif // defined (__linux__) || defined (__FreeBSD__)
+
+void forceLinkMemcpyPerfTests() {}