[OpenMP][libomptarget] Initialize global memory stack only once.

[gtbercea]

Summary

The global stack initialization function may be called multiple times. The initialization of the shared memory slots should only happen when the function is called for the first time for a given warp master thread.

[grokos]

Looks good.

libomptarget/deviceRTLs/nvptx/src/data_sharing.cu

EXTERN void __kmpc_data_sharing_init_stack() {
   // Initialize the data sharing structures. This section should only be
   // executed by the warp active master threads.
   if (IsWarpMasterActiveThread()) {
     unsigned WID = getWarpId();
     omptarget_nvptx_TeamDescr *teamDescr =
         &omptarget_nvptx_threadPrivateContext->TeamContext();
     __kmpc_data_sharing_slot *RootS = teamDescr->RootS(WID, IsMasterThread());
 
+    // If a valid address has been returned then proceed with the initalization.
+    // Otherwise the initialization of the slot has already happened in a
+    // previous call to this function.
+    if (RootS) {
       DataSharingState.SlotPtr[WID] = RootS;
       DataSharingState.TailPtr[WID] = RootS;
       DataSharingState.StackPtr[WID] = (void *)&RootS->Data[0];
     }
+  }
 
   // Currently we only support the sharing of variables between master and
   // workers. The list of references to shared variables exists only for
   // the master thread.
   if (IsMasterThread()) {
     // Initialize the list of references to arguments.
     omptarget_nvptx_globalArgs.Init();
   }

libomptarget/deviceRTLs/nvptx/src/omptarget-nvptx.h

public:
 
   // init
   INLINE void InitTeamDescr();
 
   INLINE __kmpc_data_sharing_slot *RootS(int wid, bool IsMasterThread) {
     // If this is invoked by the master thread of the master warp then intialize
     // it with a smaller slot.
     if (IsMasterThread) {
+      // Do not initalize this slot again if it has already been initalized.
+      if (master_rootS[0].DataEnd == &master_rootS[0].Data[0] + DS_Slot_Size)
+        return 0;
       // Initialize the pointer to the end of the slot given the size of the
       // data section. DataEnd is non-inclusive.
       master_rootS[0].DataEnd = &master_rootS[0].Data[0] + DS_Slot_Size;
       // We currently do not have a next slot.
       master_rootS[0].Next = 0;
       master_rootS[0].Prev = 0;
       master_rootS[0].PrevSlotStackPtr = 0;
       return (__kmpc_data_sharing_slot *)&master_rootS[0];
     }
+    // Do not initalize this slot again if it has already been initalized.
+    if (worker_rootS[wid].DataEnd ==
+        &worker_rootS[wid].Data[0] + DS_Worker_Warp_Slot_Size)
+      return 0;
     // Initialize the pointer to the end of the slot given the size of the data
     // section. DataEnd is non-inclusive.
     worker_rootS[wid].DataEnd =
         &worker_rootS[wid].Data[0] + DS_Worker_Warp_Slot_Size;
     // We currently do not have a next slot.
     worker_rootS[wid].Next = 0;
     worker_rootS[wid].Prev = 0;
     worker_rootS[wid].PrevSlotStackPtr = 0;