Move taskwait inside if where it should be

Author: sterrettm2

src/xss-common-argsort.h

@@ -640,6 +640,7 @@ X86_SIMD_SORT_INLINE void xss_argsort(T *arr,
                                        arrsize - 1,
                                        2 * (arrsize_t)log2(arrsize),
                                        task_threshold);
+#pragma omp taskwait
         }
         else {
             argsort_<vectype, argtype>(arr,
@@ -649,7 +650,6 @@ X86_SIMD_SORT_INLINE void xss_argsort(T *arr,
                                        2 * (arrsize_t)log2(arrsize),
                                        std::numeric_limits<arrsize_t>::max());
         }
-#pragma omp taskwait
 #else
         argsort_<vectype, argtype>(
                 arr, arg, 0, arrsize - 1, 2 * (arrsize_t)log2(arrsize), 0);

src/xss-common-keyvaluesort.hpp

@@ -627,6 +627,7 @@ X86_SIMD_SORT_INLINE void xss_qsort_kv(
                                       index_last_elem,
                                       maxiters,
                                       task_threshold);
+#pragma omp taskwait
         }
         else {
             kvsort_<keytype, valtype>(keys,
@@ -636,7 +637,6 @@ X86_SIMD_SORT_INLINE void xss_qsort_kv(
                                       maxiters,
                                       std::numeric_limits<arrsize_t>::max());
         }
-#pragma omp taskwait
 #else
         kvsort_<keytype, valtype>(
                 keys, indexes, 0, index_last_elem, maxiters, 0);

src/xss-common-qsort.h

@@ -682,6 +682,7 @@ X86_SIMD_SORT_INLINE void xss_qsort(T *arr, arrsize_t arrsize, bool hasnan)
                                          arrsize - 1,
                                          2 * (arrsize_t)log2(arrsize),
                                          task_threshold);
+#pragma omp taskwait
         }
         else {
             qsort_<vtype, comparator, T>(arr,
@@ -690,7 +691,6 @@ X86_SIMD_SORT_INLINE void xss_qsort(T *arr, arrsize_t arrsize, bool hasnan)
                                          2 * (arrsize_t)log2(arrsize),
                                          std::numeric_limits<arrsize_t>::max());
         }
-#pragma omp taskwait
 #else
         qsort_<vtype, comparator, T>(
                 arr, 0, arrsize - 1, 2 * (arrsize_t)log2(arrsize), 0);