[OpenMP][DeviceRTL] Fix an issue that thread array might be corrupted

[tianshilei1992]

Summary

The shared memory stack in the device runtime assumes no intervined uses.
D135037 breaks the assumption, potentially causing the shared stack corruption.
This patch moves the thread array to heap memory. Since it is already the slow
path, it doesn't matter that much anyway.

[jhuber6]

openmp/libomptarget/DeviceRTL/src/State.cpp

266

[tianshilei1992]

fix build issue

[jhuber6]

Allocating these from global memory will definitely slow down the slow path further, but I believe the better performance on the fast path is worth it. This won't fix the problem on AMDGPU until we support a malloc implementation upstream (@ronlieb).

openmp/libomptarget/DeviceRTL/src/State.cpp

void state::enterDataEnvironment(IdentTy *Ident) {
 
   unsigned TId = mapping::getThreadIdInBlock();
   ThreadStateTy *NewThreadState =
       static_cast<ThreadStateTy *>(__kmpc_alloc_shared(sizeof(ThreadStateTy)));
   uintptr_t *ThreadStatesBitsPtr = reinterpret_cast<uintptr_t *>(&ThreadStates);
   if (!atomic::load(ThreadStatesBitsPtr, atomic::seq_cst)) {
     uint32_t Bytes = sizeof(ThreadStates[0]) * mapping::getBlockSize();
     void *ThreadStatesPtr =
-        memory::allocShared(Bytes, "Thread state array allocation");
+        memory::allocGlogal(Bytes, "Thread state array allocation");
     if (!atomic::cas(ThreadStatesBitsPtr, uintptr_t(0),
                      reinterpret_cast<uintptr_t>(ThreadStatesPtr),
                      atomic::seq_cst, atomic::seq_cst))
-      memory::freeShared(ThreadStatesPtr, Bytes,
+      memory::freeGlobal(ThreadStatesPtr, Bytes,
                          "Thread state array allocated multiple times");
     ASSERT(atomic::load(ThreadStatesBitsPtr, atomic::seq_cst) &&
            "Expected valid thread states bit!");
   }
   NewThreadState->init(ThreadStates[TId]);
   TeamState.HasThreadState = true;
   ThreadStates[TId] = NewThreadState;
 }