Improve AVX* detection.

Check for all the necessary capabilities before allowing the build
system to generate the AVX512 code.

Signed-off-by: George Bosilca <bosilca@icl.utk.edu>

Author: bosilca

ompi/mca/op/avx/configure.m4

@@ -38,7 +38,13 @@ AC_DEFUN([MCA_ompi_op_avx_CONFIG],[
 
     OPAL_VAR_SCOPE_PUSH([op_avx_cflags_save])
 
-    AS_IF([test "$opal_cv_asm_arch" = "X86_64"],
+    case "${host}" in
+        x86_64-*x32|i?86-*|x86_64*|amd64*)
+            check_avx="yes";;
+        *)
+            check_avx="no";;
+    esac
+    AS_IF([test "$check_avx" = "yes"],
           [AC_LANG_PUSH([C])
 
            #
@@ -123,6 +129,28 @@ AC_DEFUN([MCA_ompi_op_avx_CONFIG],[
                               MCA_BUILD_OP_AVX512_FLAGS=""
                               AC_MSG_RESULT([no])])
                          CFLAGS="$op_avx_cflags_save"
+                        ])
+                  #
+                  # Check for combination of AVX512F + AVX512VL
+                  #
+                  AS_IF([test $op_avx512_support -eq 1],
+                        [AC_MSG_CHECKING([if _mm_max_epi64 generates code that can be compiled])
+                         op_avx_cflags_save="$CFLAGS"
+                         CFLAGS="$CFLAGS_WITHOUT_OPTFLAGS -O0 $MCA_BUILD_OP_AVX512_FLAGS"
+                         AC_LINK_IFELSE(
+                             [AC_LANG_PROGRAM([[#include <immintrin.h>]],
+                                      [[
+#if !defined(__AVX512F__) || !defined(__AVX512VL__) || !defined(__AVX512BW__)
+#error "icc needs the -m flags to provide the AVX* detection macros"
+#endif
+    __m128i vA, vB;
+    _mm_max_epi64(vA, vB)
+                                      ]])],
+                             [AC_MSG_RESULT([yes])],
+                             [op_avx512_support=0
+                              MCA_BUILD_OP_AVX512_FLAGS=""
+                              AC_MSG_RESULT([no])])
+                         CFLAGS="$op_avx_cflags_save"
                         ])])
            #
            # Check support for AVX2

ompi/mca/op/avx/op_avx_functions.c

@@ -32,16 +32,18 @@
  * to a lesser support (AVX512 -> AVX2, AVX2 -> AVX, AVX -> error out).
  */
 #if defined(GENERATE_AVX512_CODE)
+#  define PREPEND _avx512
 #  if defined(__AVX512BW__) && defined(__AVX512F__) && defined(__AVX512VL__)
-#    define PREPEND _avx512
+/* all good */
 #  else
 #    undef GENERATE_AVX512_CODE
 #  endif  /* defined(__AVX512BW__) && defined(__AVX512F__) && defined(__AVX512VL__) */
 #endif  /* defined(GENERATE_AVX512_CODE) */
 
 #if !defined(PREPEND) && defined(GENERATE_AVX2_CODE)
+#  define PREPEND _avx2
 #  if defined(__AVX2__)
-#    define PREPEND _avx2
+/* all good */
 #  else
 #    undef GENERATE_AVX2_CODE
 #  endif  /* defined(__AVX2__) */