3 files changed, 34 insertions, 0 deletions
diff --git a/test cases/common/139 simd/meson.build b/test cases/common/139 simd/meson.build
index 1d23ffb..6031218 100644
--- a/test cases/common/139 simd/meson.build
+++ b/test cases/common/139 simd/meson.build
@@ -30,6 +30,7 @@ simdarr = [['-mmmx', 'HAVE_MMX', 'simd_mmx', 'simd_mmx.c'],
  ['-msse4.1', 'HAVE_SSE41', 'simd_sse41', 'simd_sse41.c'],
  ['-msse4.2', 'HAVE_SSE42', 'simd_sse42', 'simd_sse42.c'],
  ['-mavx', 'HAVE_AVX', 'simd_avx', 'simd_avx.c'],
+ ['-mavx2', 'HAVE_AVX2', 'simd_avx2', 'simd_avx2.c'],
 ]
 
 foreach ia : simdarr
diff --git a/test cases/common/139 simd/simd_avx2.c b/test cases/common/139 simd/simd_avx2.c
new file mode 100644
index 0000000..bca8c99
--- /dev/null
+++ b/test cases/common/139 simd/simd_avx2.c
@@ -0,0 +1,27 @@
+#include<simdconfig.h>
+#include<simdfuncs.h>
+
+#include<immintrin.h>
+#include<cpuid.h>
+#include<stdint.h>
+
+int avx2_available() {
+    return __builtin_cpu_supports("avx2");
+}
+
+void increment_avx2(float arr[4]) {
+    double darr[4];
+    darr[0] = arr[0];
+    darr[1] = arr[1];
+    darr[2] = arr[2];
+    darr[3] = arr[3];
+    __m256d val = _mm256_loadu_pd(darr);
+    __m256d one = _mm256_set1_pd(1.0);
+    __m256d result = _mm256_add_pd(val, one);
+    _mm256_storeu_pd(darr, result);
+    one = _mm256_permute4x64_pd(one, 66); /* A no-op, just here to use AVX2. */
+    arr[0] = (float)darr[0];
+    arr[1] = (float)darr[1];
+    arr[2] = (float)darr[2];
+    arr[3] = (float)darr[3];
+}
diff --git a/test cases/common/139 simd/simdchecker.c b/test cases/common/139 simd/simdchecker.c
index da53f0a..1502ae7 100644
--- a/test cases/common/139 simd/simdchecker.c
+++ b/test cases/common/139 simd/simdchecker.c
@@ -17,6 +17,12 @@ int main(int argc, char **argv) {
 /* Add here. The first matched one is used so put "better" instruction
  * sets at the top.
  */
+#if HAVE_AVX2
+    if(fptr == NULL && avx2_available()) {
+        fptr = increment_avx2;
+        type = "AVX2";
+    }
+#endif
 #if HAVE_AVX
     if(fptr == NULL && avx_available()) {
         fptr = increment_avx;