[flang][cuda] Fix GPULaunchKernelConversion to generate correct kernel launch parameters

Author: khaki3

flang/lib/Optimizer/Transforms/CUFGPUToLLVMConversion.cpp

@@ -58,7 +58,7 @@ static mlir::Value createKernelArgArray(mlir::Location loc,
         loc, ptrTy, structTy, argStruct, mlir::ArrayRef<mlir::Value>({indice}));
     rewriter.create<LLVM::StoreOp>(loc, arg, structMember);
     mlir::Value arrayMember = rewriter.create<LLVM::GEPOp>(
-        loc, ptrTy, structTy, argArray, mlir::ArrayRef<mlir::Value>({indice}));
+        loc, ptrTy, ptrTy, argArray, mlir::ArrayRef<mlir::Value>({indice}));
     rewriter.create<LLVM::StoreOp>(loc, structMember, arrayMember);
   }
   return argArray;

flang/test/Fir/CUDA/cuda-gpu-launch-func.mlir

@@ -99,9 +99,16 @@ module attributes {dlti.dl_spec = #dlti.dl_spec<#dlti.dl_entry<i1, dense<8> : ve
 }
 
 // CHECK-LABEL: _QMmod1Phost_sub
-
+// CHECK: %[[STRUCT:.*]] = llvm.alloca %{{.*}} x !llvm.struct<(ptr)> : (i32) -> !llvm.ptr
+// CHECK: %[[PARAMS:.*]] = llvm.alloca %{{.*}} x !llvm.ptr : (i32) -> !llvm.ptr
+// CHECK: %[[ZERO:.*]] = llvm.mlir.constant(0 : i32) : i32
+// CHECK: %[[STRUCT_PTR:.*]] = llvm.getelementptr %[[STRUCT]][%[[ZERO]]] : (!llvm.ptr, i32) -> !llvm.ptr, !llvm.struct<(ptr)>
+// CHECK: llvm.store %{{.*}}, %[[STRUCT_PTR]] : !llvm.ptr, !llvm.ptr
+// CHECK: %[[PARAM_PTR:.*]] = llvm.getelementptr %[[PARAMS]][%[[ZERO]]] : (!llvm.ptr, i32) -> !llvm.ptr, !llvm.ptr
+// CHECK: llvm.store %[[STRUCT_PTR]], %[[PARAM_PTR]] : !llvm.ptr, !llvm.ptr
 // CHECK: %[[KERNEL_PTR:.*]] = llvm.mlir.addressof @_QMmod1Psub1 : !llvm.ptr
-// CHECK: llvm.call @_FortranACUFLaunchKernel(%[[KERNEL_PTR]], {{.*}})
+// CHECK: %[[NULL:.*]] = llvm.mlir.zero : !llvm.ptr
+// CHECK: llvm.call @_FortranACUFLaunchKernel(%[[KERNEL_PTR]], {{.*}}, %[[PARAMS]], %[[NULL]])
 
 // -----