[flang][cuda] Add fence after barrier_init

Author: clementval

flang/lib/Optimizer/Builder/IntrinsicCall.cpp

@@ -3236,6 +3236,11 @@ void IntrinsicLibrary::genBarrierInit(llvm::ArrayRef<fir::ExtendedValue> args) {
       convertBarrierToLLVM(builder, loc, fir::getBase(args[0]));
   mlir::NVVM::MBarrierInitSharedOp::create(builder, loc, barrier,
                                            fir::getBase(args[1]), {});
+  auto kind = mlir::NVVM::ProxyKindAttr::get(
+      builder.getContext(), mlir::NVVM::ProxyKind::async_shared);
+  auto space = mlir::NVVM::SharedSpaceAttr::get(
+      builder.getContext(), mlir::NVVM::SharedSpace::shared_cta);
+  mlir::NVVM::FenceProxyOp::create(builder, loc, kind, space);
 }
 
 // BESSEL_JN

flang/test/Lower/CUDA/cuda-device-proc.cuf

@@ -411,6 +411,7 @@ end subroutine
 ! CHECK: %[[LLVM_PTR:.*]] = fir.convert %[[DECL_SHARED]]#0 : (!fir.ref<i64>) -> !llvm.ptr
 ! CHECK: %[[SHARED_PTR:.*]] = llvm.addrspacecast %[[LLVM_PTR]] : !llvm.ptr to !llvm.ptr<3>
 ! CHECK: nvvm.mbarrier.init.shared %[[SHARED_PTR]], %[[COUNT]] : !llvm.ptr<3>, i32
+! CHECK: nvvm.fence.proxy {kind = #nvvm.proxy_kind<async.shared>, space = #nvvm.shared_space<cta>}
 
 ! CHECK: %[[LLVM_PTR:.*]] = fir.convert %[[DECL_SHARED]]#0 : (!fir.ref<i64>) -> !llvm.ptr
 ! CHECK: %[[SHARED_PTR:.*]] = llvm.addrspacecast %[[LLVM_PTR]] : !llvm.ptr to !llvm.ptr<3>