Merge pull request #183 from sterrettm2/defensive-emms

Raghuveer Devulapalli · web-flow · commit 7604f9e35ec4 · 2025-02-24T20:40:17.000-08:00
Add defensive EMMS instructions to each SIMD sort function
diff --git a/src/avx512-16bit-qsort.hpp b/src/avx512-16bit-qsort.hpp
@@ -572,6 +572,11 @@ avx512_qsort_fp16(uint16_t *arr,
         }
         replace_inf_with_nan(arr, arrsize, nan_count, descending);
     }
+
+#ifdef __MMX__
+    // Workaround for compiler bug generating MMX instructions without emms
+    _mm_empty();
+#endif
 }
 
 [[maybe_unused]] X86_SIMD_SORT_INLINE void
@@ -605,6 +610,11 @@ avx512_qselect_fp16(uint16_t *arr,
                     2 * (arrsize_t)log2(indx_last_elem));
         }
     }
+
+#ifdef __MMX__
+    // Workaround for compiler bug generating MMX instructions without emms
+    _mm_empty();
+#endif
 }
 
 [[maybe_unused]] X86_SIMD_SORT_INLINE void
diff --git a/src/xss-common-keyvaluesort.hpp b/src/xss-common-keyvaluesort.hpp
@@ -652,6 +652,11 @@ X86_SIMD_SORT_INLINE void xss_qsort_kv(
             std::reverse(indexes, indexes + arrsize);
         }
     }
+
+#ifdef __MMX__
+    // Workaround for compiler bug generating MMX instructions without emms
+    _mm_empty();
+#endif
 }
 
 template <typename T1,
@@ -712,6 +717,11 @@ X86_SIMD_SORT_INLINE void xss_select_kv(T1 *keys,
             std::reverse(indexes, indexes + arrsize);
         }
     }
+
+#ifdef __MMX__
+    // Workaround for compiler bug generating MMX instructions without emms
+    _mm_empty();
+#endif
 }
 
 template <typename T1,
diff --git a/src/xss-common-qsort.h b/src/xss-common-qsort.h
@@ -632,6 +632,11 @@ X86_SIMD_SORT_INLINE void xss_qsort(T *arr, arrsize_t arrsize, bool hasnan)
 
         replace_inf_with_nan(arr, arrsize, nan_count, descending);
     }
+
+#ifdef __MMX__
+    // Workaround for compiler bug generating MMX instructions without emms
+    _mm_empty();
+#endif
 }
 
 // Quick select methods
@@ -666,6 +671,11 @@ xss_qselect(T *arr, arrsize_t k, arrsize_t arrsize, bool hasnan)
                                        index_last_elem,
                                        2 * (arrsize_t)log2(arrsize));
     }
+
+#ifdef __MMX__
+    // Workaround for compiler bug generating MMX instructions without emms
+    _mm_empty();
+#endif
 }
 
 // Partial sort methods:

Original file line number	Diff line number	Diff line change
`@@ -572,6 +572,11 @@ avx512_qsort_fp16(uint16_t *arr,`
`572`	`572`	`}`
`573`	`573`	`replace_inf_with_nan(arr, arrsize, nan_count, descending);`
`574`	`574`	`}`
	`575`	`+`
	`576`	`+#ifdef __MMX__`
	`577`	`+ // Workaround for compiler bug generating MMX instructions without emms`
	`578`	`+ _mm_empty();`
	`579`	`+#endif`
`575`	`580`	`}`
`576`	`581`
`577`	`582`	`[[maybe_unused]] X86_SIMD_SORT_INLINE void`
`@@ -605,6 +610,11 @@ avx512_qselect_fp16(uint16_t *arr,`
`605`	`610`	`2 * (arrsize_t)log2(indx_last_elem));`
`606`	`611`	`}`
`607`	`612`	`}`
	`613`	`+`
	`614`	`+#ifdef __MMX__`
	`615`	`+ // Workaround for compiler bug generating MMX instructions without emms`
	`616`	`+ _mm_empty();`
	`617`	`+#endif`
`608`	`618`	`}`
`609`	`619`
`610`	`620`	`[[maybe_unused]] X86_SIMD_SORT_INLINE void`
Original file line number	Diff line number	Diff line change
`@@ -652,6 +652,11 @@ X86_SIMD_SORT_INLINE void xss_qsort_kv(`
`652`	`652`	`std::reverse(indexes, indexes + arrsize);`
`653`	`653`	`}`
`654`	`654`	`}`
	`655`	`+`
	`656`	`+#ifdef __MMX__`
	`657`	`+ // Workaround for compiler bug generating MMX instructions without emms`
	`658`	`+ _mm_empty();`
	`659`	`+#endif`
`655`	`660`	`}`
`656`	`661`
`657`	`662`	`template <typename T1,`
`@@ -712,6 +717,11 @@ X86_SIMD_SORT_INLINE void xss_select_kv(T1 *keys,`
`712`	`717`	`std::reverse(indexes, indexes + arrsize);`
`713`	`718`	`}`
`714`	`719`	`}`
	`720`	`+`
	`721`	`+#ifdef __MMX__`
	`722`	`+ // Workaround for compiler bug generating MMX instructions without emms`
	`723`	`+ _mm_empty();`
	`724`	`+#endif`
`715`	`725`	`}`
`716`	`726`
`717`	`727`	`template <typename T1,`
Original file line number	Diff line number	Diff line change
`@@ -632,6 +632,11 @@ X86_SIMD_SORT_INLINE void xss_qsort(T *arr, arrsize_t arrsize, bool hasnan)`
`632`	`632`
`633`	`633`	`replace_inf_with_nan(arr, arrsize, nan_count, descending);`
`634`	`634`	`}`
	`635`	`+`
	`636`	`+#ifdef __MMX__`
	`637`	`+ // Workaround for compiler bug generating MMX instructions without emms`
	`638`	`+ _mm_empty();`
	`639`	`+#endif`
`635`	`640`	`}`
`636`	`641`
`637`	`642`	`// Quick select methods`
`@@ -666,6 +671,11 @@ xss_qselect(T *arr, arrsize_t k, arrsize_t arrsize, bool hasnan)`
`666`	`671`	`index_last_elem,`
`667`	`672`	`2 * (arrsize_t)log2(arrsize));`
`668`	`673`	`}`
	`674`	`+`
	`675`	`+#ifdef __MMX__`
	`676`	`+ // Workaround for compiler bug generating MMX instructions without emms`
	`677`	`+ _mm_empty();`
	`678`	`+#endif`
`669`	`679`	`}`
`670`	`680`
`671`	`681`	`// Partial sort methods:`