vendor/llvm-project/llvmorg-11-init-20887-g2e10b7a39b9vendor/llvm-project/master

1 files changed, 35 insertions, 1 deletions

diff --git a/openmp/runtime/src/kmp_csupport.cpp b/openmp/runtime/src/kmp_csupport.cpp
index ac9a93590ad0..9cfa64d6ff9e 100644
--- a/openmp/runtime/src/kmp_csupport.cpp
+++ b/openmp/runtime/src/kmp_csupport.cpp
@@ -92,7 +92,7 @@ construct, since the master thread is necessarily thread zero).
 
 If multiple non-OpenMP threads all enter an OpenMP construct then this
 will be a unique thread identifier among all the threads created by
-the OpenMP runtime (but the value cannote be defined in terms of
+the OpenMP runtime (but the value cannot be defined in terms of
 OpenMP thread ids returned by omp_get_thread_num()).
 */
 kmp_int32 __kmpc_global_thread_num(ident_t *loc) {
@@ -4023,6 +4023,9 @@ void __kmpc_doacross_wait(ident_t *loc, int gtid, const kmp_int64 *vec) {
   lo = pr_buf->th_doacross_info[2];
   up = pr_buf->th_doacross_info[3];
   st = pr_buf->th_doacross_info[4];
+#if OMPT_SUPPORT && OMPT_OPTIONAL
+  ompt_dependence_t deps[num_dims];
+#endif
   if (st == 1) { // most common case
     if (vec[0] < lo || vec[0] > up) {
       KA_TRACE(20, ("__kmpc_doacross_wait() exit: T#%d iter %lld is out of "
@@ -4048,6 +4051,10 @@ void __kmpc_doacross_wait(ident_t *loc, int gtid, const kmp_int64 *vec) {
     }
     iter_number = (kmp_uint64)(lo - vec[0]) / (-st);
   }
+#if OMPT_SUPPORT && OMPT_OPTIONAL
+  deps[0].variable.value = iter_number;
+  deps[0].dependence_type = ompt_dependence_type_sink;
+#endif
   for (i = 1; i < num_dims; ++i) {
     kmp_int64 iter, ln;
     kmp_int32 j = i * 4;
@@ -4081,6 +4088,10 @@ void __kmpc_doacross_wait(ident_t *loc, int gtid, const kmp_int64 *vec) {
       iter = (kmp_uint64)(lo - vec[i]) / (-st);
     }
     iter_number = iter + ln * iter_number;
+#if OMPT_SUPPORT && OMPT_OPTIONAL
+    deps[i].variable.value = iter;
+    deps[i].dependence_type = ompt_dependence_type_sink;
+#endif
   }
   shft = iter_number % 32; // use 32-bit granularity
   iter_number >>= 5; // divided by 32
@@ -4089,6 +4100,12 @@ void __kmpc_doacross_wait(ident_t *loc, int gtid, const kmp_int64 *vec) {
     KMP_YIELD(TRUE);
   }
   KMP_MB();
+#if OMPT_SUPPORT && OMPT_OPTIONAL
+  if (ompt_enabled.ompt_callback_dependences) {
+    ompt_callbacks.ompt_callback(ompt_callback_dependences)(
+        &(OMPT_CUR_TASK_INFO(th)->task_data), deps, num_dims);
+  }
+#endif
   KA_TRACE(20,
            ("__kmpc_doacross_wait() exit: T#%d wait for iter %lld completed\n",
             gtid, (iter_number << 5) + shft));
@@ -4116,6 +4133,9 @@ void __kmpc_doacross_post(ident_t *loc, int gtid, const kmp_int64 *vec) {
   num_dims = pr_buf->th_doacross_info[0];
   lo = pr_buf->th_doacross_info[2];
   st = pr_buf->th_doacross_info[4];
+#if OMPT_SUPPORT && OMPT_OPTIONAL
+  ompt_dependence_t deps[num_dims];
+#endif
   if (st == 1) { // most common case
     iter_number = vec[0] - lo;
   } else if (st > 0) {
@@ -4123,6 +4143,10 @@ void __kmpc_doacross_post(ident_t *loc, int gtid, const kmp_int64 *vec) {
   } else { // negative increment
     iter_number = (kmp_uint64)(lo - vec[0]) / (-st);
   }
+#if OMPT_SUPPORT && OMPT_OPTIONAL
+  deps[0].variable.value = iter_number;
+  deps[0].dependence_type = ompt_dependence_type_source;
+#endif
   for (i = 1; i < num_dims; ++i) {
     kmp_int64 iter, ln;
     kmp_int32 j = i * 4;
@@ -4137,7 +4161,17 @@ void __kmpc_doacross_post(ident_t *loc, int gtid, const kmp_int64 *vec) {
       iter = (kmp_uint64)(lo - vec[i]) / (-st);
     }
     iter_number = iter + ln * iter_number;
+#if OMPT_SUPPORT && OMPT_OPTIONAL
+    deps[i].variable.value = iter;
+    deps[i].dependence_type = ompt_dependence_type_source;
+#endif
+  }
+#if OMPT_SUPPORT && OMPT_OPTIONAL
+  if (ompt_enabled.ompt_callback_dependences) {
+    ompt_callbacks.ompt_callback(ompt_callback_dependences)(
+        &(OMPT_CUR_TASK_INFO(th)->task_data), deps, num_dims);
   }
+#endif
   shft = iter_number % 32; // use 32-bit granularity
   iter_number >>= 5; // divided by 32
   flag = 1 << shft;