[OpenMP][Clang] Move device global stack init before master-workers split

[gtbercea]

Summary

This patch moves the call to the stack init data sharing function before the splitting of threads into master and workers. This ensures that the initialization happens for all active warp master threads. Test is modified appropriately.

[ABataev]

LG

[gtbercea]

Fix test.

[ABataev]

LG

[gtbercea]

This leads to usage of statically allocated shared data before their initialization in runtime structures by master thread in kernel_init() function. New patch available with worker and master-side initialization.

lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp

 }
 
 // Setup NVPTX threads for master-worker OpenMP scheme.
 void CGOpenMPRuntimeNVPTX::emitGenericEntryHeader(CodeGenFunction &CGF,
                                                   EntryFunctionState &EST,
                                                   WorkerFunctionState &WST) {
   CGBuilderTy &Bld = CGF.Builder;
 
+  // For data sharing, initialize the stack for master and workers.
+  CGF.EmitRuntimeCall(
+      createNVPTXRuntimeFunction(
+          OMPRTL_NVPTX__kmpc_data_sharing_init_stack));
+
   llvm::BasicBlock *WorkerBB = CGF.createBasicBlock(".worker");
   llvm::BasicBlock *MasterCheckBB = CGF.createBasicBlock(".mastercheck");
   llvm::BasicBlock *MasterBB = CGF.createBasicBlock(".master");
   EST.ExitBB = CGF.createBasicBlock(".exit");
 
   auto *IsWorker =
       Bld.CreateICmpULT(getNVPTXThreadID(CGF), getThreadLimit(CGF));
   Bld.CreateCondBr(IsWorker, WorkerBB, MasterCheckBB);
void CGOpenMPRuntimeNVPTX::emitGenericEntryHeader(CodeGenFunction &CGF,
   // First action in sequential region:
   // Initialize the state of the OpenMP runtime library on the GPU.
   // TODO: Optimize runtime initialization and pass in correct value.
   llvm::Value *Args[] = {getThreadLimit(CGF),
                          Bld.getInt16(/*RequiresOMPRuntime=*/1)};
   CGF.EmitRuntimeCall(
       createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_kernel_init), Args);
 
-  // For data sharing, we need to initialize the stack.
-  CGF.EmitRuntimeCall(
-      createNVPTXRuntimeFunction(
-          OMPRTL_NVPTX__kmpc_data_sharing_init_stack));
-
   emitGenericVarsProlog(CGF, WST.Loc);
 }
 
 void CGOpenMPRuntimeNVPTX::emitGenericEntryFooter(CodeGenFunction &CGF,
                                                   EntryFunctionState &EST) {
   emitGenericVarsEpilog(CGF);
   if (!CGF.HaveInsertPoint())
     return;

test/OpenMP/nvptx_data_sharing.cpp

void test_ds(){
   }
 }
 
 /// ========= In the kernel function ========= ///
 
 // CK1: {{.*}}define void @__omp_offloading{{.*}}test_ds{{.*}}()
 // CK1: [[SHAREDARGS1:%.+]] = alloca i8**
 // CK1: [[SHAREDARGS2:%.+]] = alloca i8**
-// CK1: call void @__kmpc_kernel_init
 // CK1: call void @__kmpc_data_sharing_init_stack
+// CK1: call void @__omp_offloading{{.*}}_worker()
+// CK1: call void @__kmpc_kernel_init
 // CK1: [[GLOBALSTACK:%.+]] = call i8* @__kmpc_data_sharing_push_stack(i64 8, i16 0)
 // CK1: [[GLOBALSTACK2:%.+]] = bitcast i8* [[GLOBALSTACK]] to %struct._globalized_locals_ty*
 // CK1: [[A:%.+]] = getelementptr inbounds %struct._globalized_locals_ty, %struct._globalized_locals_ty* [[GLOBALSTACK2]], i32 0, i32 0
 // CK1: [[B:%.+]] = getelementptr inbounds %struct._globalized_locals_ty, %struct._globalized_locals_ty* [[GLOBALSTACK2]], i32 0, i32 1
 // CK1: store i32 10, i32* [[A]]
 // CK1: call void @__kmpc_kernel_prepare_parallel({{.*}}, i16 1)
 // CK1: call void @__kmpc_begin_sharing_variables(i8*** [[SHAREDARGS1]], i64 1)
 // CK1: [[SHARGSTMP1:%.+]] = load i8**, i8*** [[SHAREDARGS1]]

test/OpenMP/nvptx_target_teams_distribute_parallel_for_generic_mode_codegen.cpp

 }
 
 // CHECK: define internal void @__omp_offloading_{{.*}}_main_l[[@LINE-6]]_worker()
 // CHECK: [[TID:%.+]] = call i32 @__kmpc_global_thread_num(%ident_t* @
 // CHECK: call void [[PARALLEL:@.+]]_wrapper(i16 0, i32 [[TID]])
 
 // CHECK: define void @__omp_offloading_{{.*}}_main_l[[@LINE-10]](i{{64|32}} %{{[^,].*}}, i32* dereferenceable{{[^,]*}}, i{{64|32}} %{{[^,)]*}})
 // CHECK: [[TID:%.+]] = call i32 @__kmpc_global_thread_num(%ident_t* @
-// CHECK: call void @__kmpc_kernel_init(
 // CHECK: call void @__kmpc_data_sharing_init_stack()
+// CHECK: call void @__kmpc_kernel_init(
 // CHECK: call void @__kmpc_for_static_init_4(
 // CHECK: call void @__kmpc_kernel_prepare_parallel(
 // CHECK: call void @__kmpc_begin_sharing_variables(i8*** [[BUF_PTR_PTR:%[^,]+]], i{{64|32}} 4)
 // CHECK: [[BUF_PTR:%.+]] = load i8**, i8*** [[BUF_PTR_PTR]],
 // CHECK: [[LB:%.+]] = inttoptr i{{64|32}} [[LB_:%.*]] to i8*
 // CHECK: store i8* [[LB]], i8** [[BUF_PTR]],
 // CHECK: [[BUF_PTR1:%.+]] = getelementptr inbounds i8*, i8** [[BUF_PTR]], i{{[0-9]+}} 1
 // CHECK: [[UB:%.+]] = inttoptr i{{64|32}} [[UB_:%.*]] to i8*