Merge pull request #172 from sterrettm2/mmx_fix

Fix for MMX instructions being generated without emms

README.md

@@ -80,16 +80,6 @@ benchmark](https://github.com/google/benchmark) frameworks respectively. You
 can configure meson to build them both by using `-Dbuild_tests=true` and
 `-Dbuild_benchmarks=true`.
 
-### Note about building with avx512 by g++ v9 and v10
-
-There is a risk when compile with avx512 by g++ v9 and v10,
-as some `MMX Technology` instructions is used by g++ v9/v10 
-without clearing fpu state. 
-Check [issue 154](https://github.com/intel/x86-simd-sort/issues/154) 
-for more details.
-
-Adding `g++` option `-mno-mmx`, which disables `MMX Technology` instructions, is a possible workaround.
-
 ## Example usage
 
 #### Sort an array of floats

src/xss-common-argsort.h

@@ -575,6 +575,11 @@ X86_SIMD_SORT_INLINE void xss_argsort(T *arr,
 
         if (descending) { std::reverse(arg, arg + arrsize); }
     }
+
+#ifdef __MMX__
+    // Workaround for compiler bug generating MMX instructions without emms
+    _mm_empty();
+#endif
 }
 
 template <typename T>
@@ -632,6 +637,11 @@ X86_SIMD_SORT_INLINE void xss_argselect(T *arr,
         argselect_<vectype, argtype>(
                 arr, arg, k, 0, arrsize - 1, 2 * (arrsize_t)log2(arrsize));
     }
+
+#ifdef __MMX__
+    // Workaround for compiler bug generating MMX instructions without emms
+    _mm_empty();
+#endif
 }
 
 template <typename T>