diff --git a/build/cmake/cpu.cmake b/build/cmake/cpu.cmake
index acebe20..8c67d89 100644
--- a/build/cmake/cpu.cmake
+++ b/build/cmake/cpu.cmake
@@ -120,6 +120,19 @@ elseif("${AOM_TARGET_CPU}" MATCHES "^x86")
     set(RTCD_ARCH_X86_64 "yes")
   endif()
 
+  # AVX2 requires __m256i definition starting v3.9.0
+
+  if(ENABLE_AVX2)
+    aom_check_source_compiles("x86_64_avx2_m256i_available" "
+#include <emmintrin.h>
+#ifndef __m256i
+#error 1
+#endif" HAVE_AVX2_M256I)
+    if(HAVE_AVX2_M256I EQUAL 0)
+      set(ENABLE_AVX2 0)
+    endif()
+  endif()
+
   set(X86_FLAVORS "MMX;SSE;SSE2;SSE3;SSSE3;SSE4_1;SSE4_2;AVX;AVX2")
   foreach(flavor ${X86_FLAVORS})
     if(ENABLE_${flavor} AND NOT disable_remaining_flavors)
diff --git a/aom_dsp/x86/synonyms.h b/aom_dsp/x86/synonyms.h
index 0d51cdf..6744ec5 100644
--- a/aom_dsp/x86/synonyms.h
+++ b/aom_dsp/x86/synonyms.h
@@ -46,13 +46,6 @@ static INLINE __m128i xx_loadu_128(const void *a) {
   return _mm_loadu_si128((const __m128i *)a);
 }
 
-// Load 64 bits from each of hi and low, and pack into an SSE register
-// Since directly loading as `int64_t`s and using _mm_set_epi64 may violate
-// the strict aliasing rule, this takes a different approach
-static INLINE __m128i xx_loadu_2x64(const void *hi, const void *lo) {
-  return _mm_unpacklo_epi64(_mm_loadu_si64(lo), _mm_loadu_si64(hi));
-}
-
 static INLINE void xx_storel_32(void *const a, const __m128i v) {
   const int val = _mm_cvtsi128_si32(v);
   memcpy(a, &val, sizeof(val));
diff --git a/aom_dsp/x86/synonyms_avx2.h b/aom_dsp/x86/synonyms_avx2.h
index d4e8f69..45be17e 100644
--- a/aom_dsp/x86/synonyms_avx2.h
+++ b/aom_dsp/x86/synonyms_avx2.h
@@ -25,6 +25,13 @@
  * Intrinsics prefixed with yy_ operate on or return 256bit YMM registers.
  */
 
+// Load 64 bits from each of hi and low, and pack into an SSE register
+// Since directly loading as `int64_t`s and using _mm_set_epi64 may violate
+// the strict aliasing rule, this takes a different approach
+static INLINE __m128i xx_loadu_2x64(const void *hi, const void *lo) {
+  return _mm_unpacklo_epi64(_mm_loadu_si64(lo), _mm_loadu_si64(hi));
+}
+
 // Loads and stores to do away with the tedium of casting the address
 // to the right type.
 static INLINE __m256i yy_load_256(const void *a) {