Add support for multiple microarchitectures in single binary

 * compiler -march= parameter is changed from native to corei7
   so code is always genereted with instructions which are available
   on the Nehalem microarchitecture (up to SSE4.2)

 * compiler -mtune= parameter is added so code is optimized for
   corei7-avx which equals to Sandy Bridge microarchitecture

 * set of macros is added which allows run-time detection of available
   cpu instructions (e.g. clib_cpu_supports_avx())

 * set of macros is added which allows us to clone graph node funcitons
   where cloned function is optmized for different microarchitecture
   Those macros are using following attributes:
     __attribute__((flatten))
     __attribute__((target("arch=core-avx2)))

   I.e. If applied to foo_node_fn() macro will generate cloned
   functions foo_node_fn_avx2() and foo_node_fn_avx512() (future)
   It will also generate function void * foo_node_fn_multiarch_select()
   which detects available instruction set and returns pointer to the
   best matching function clone.

Change-Id: I2dce0ac92a5ede95fcb56f47f3d1f3c4c040bac0
Signed-off-by: Damjan Marion <damarion@cisco.com>

2 files changed, 95 insertions, 6 deletions

diff --git a/vppinfra/vppinfra/cpu.c b/vppinfra/vppinfra/cpu.c
index f2dbaf1f7d8..9008ee3dd89 100644
--- a/vppinfra/vppinfra/cpu.c
+++ b/vppinfra/vppinfra/cpu.c
@@ -16,10 +16,6 @@
 #include <vppinfra/format.h>
 #include <vppinfra/cpu.h>
 
-#if __x86_64__
-#include <cpuid.h>
-#endif
-
 #define foreach_x86_cpu_uarch \
  _(0x06, 0x4f, "Broadwell", "Broadwell-EP/EX") \
  _(0x06, 0x3d, "Broadwell", "Broadwell") \
@@ -108,4 +104,21 @@ format_cpu_model_name (u8 * s, va_list * args)
 #else /* ! __x86_64__ */
   return format (s, "unknown");
 #endif
-}
\ No newline at end of file
+}
+
+u8 *
+format_cpu_flags (u8 * s, va_list * args)
+{
+#if __x86_64__
+#define _(flag, func, reg, bit) \
+  if (clib_cpu_supports_ ## flag()) \
+    s = format (s, #flag " ");
+  foreach_x86_64_flags
+  return s;
+#undef _
+#else /* ! __x86_64__ */
+  return format (s, "unknown");
+#endif
+}
+
+
diff --git a/vppinfra/vppinfra/cpu.h b/vppinfra/vppinfra/cpu.h
index 79cdf74f84b..961af709a63 100644
--- a/vppinfra/vppinfra/cpu.h
+++ b/vppinfra/vppinfra/cpu.h
@@ -16,7 +16,83 @@
 #ifndef included_clib_cpu_h
 #define included_clib_cpu_h
 
+#include <vppinfra/format.h>
+
+/*
+ * multiarchitecture support. Adding new entry will produce
+ * new graph node function variant optimized for specific cpu
+ * microarchitecture.
+ * Order is important for runtime selection, as 1st match wins...
+ */
+
+#if __x86_64__ && CLIB_DEBUG == 0
+#define foreach_march_variant(macro, x) \
+  macro(avx2,  x, "arch=core-avx2")
+#else
+#define foreach_march_variant(macro, x)
+#endif
+
+
+#if __GNUC__ > 4  && !__clang__
+#define CLIB_CPU_OPTIMIZED __attribute__ ((optimize ("tree-vectorize")))
+#else
+#define CLIB_CPU_OPTIMIZED
+#endif
+
+
+#define CLIB_MULTIARCH_ARCH_CHECK(arch, fn, tgt)			\
+  if (clib_cpu_supports_ ## arch())					\
+    return & fn ## _ ##arch;
+
+#define CLIB_MULTIARCH_SELECT_FN(fn,...)                               \
+  __VA_ARGS__ void * fn ## _multiarch_select(void)                     \
+{                                                                      \
+  foreach_march_variant(CLIB_MULTIARCH_ARCH_CHECK, fn)                 \
+  return & fn;                                                         \
+}
+
+#if __x86_64__
+#include "cpuid.h"
+
+#define foreach_x86_64_flags \
+_ (sse3,     1, ecx, 0)   \
+_ (ssse3,    1, ecx, 9)   \
+_ (sse41,    1, ecx, 19)  \
+_ (sse42,    1, ecx, 20)  \
+_ (avx,      1, ecx, 28)  \
+_ (avx2,     7, ebx, 5)   \
+_ (avx512f,  7, ebx, 16)  \
+_ (aes,      1, ecx, 25)  \
+_ (sha,      7, ebx, 29)
+
+static inline int
+clib_get_cpuid(const u32 lev, u32 * eax, u32 *ebx, u32 * ecx, u32 * edx)
+{
+  if ((u32) __get_cpuid_max (0x80000000 & lev, 0) < lev)
+    return 0;
+  if (lev == 7)
+    __cpuid_count(lev, 0, *eax, *ebx, *ecx, *edx);
+  else
+    __cpuid(lev, *eax, *ebx, *ecx, *edx);
+  return 1;
+}
+
+
+#define _(flag, func, reg, bit) \
+static inline int							\
+clib_cpu_supports_ ## flag()						\
+{									\
+  u32 __attribute__((unused)) eax, ebx = 0, ecx = 0, edx  = 0;		\
+  clib_get_cpuid (func, &eax, &ebx, &ecx, &edx);			\
+									\
+  return ((reg & (1 << bit)) != 0);					\
+}
+  foreach_x86_64_flags
+#undef _
+#endif
+
 format_function_t format_cpu_uarch;
 format_function_t format_cpu_model_name;
+format_function_t format_cpu_flags;
 
-#endif
\ No newline at end of file
+#endif