! RUN: bbc -emit-hlfir -fcuda %s -o - | FileCheck %s
! RUN: bbc -emit-hlfir -fcuda %s -o - | fir-opt | FileCheck %s

! Test lowering of CUDA procedure calls.

module test_call
  use, intrinsic :: __fortran_builtins, only: __builtin_dim3
contains
  attributes(global) subroutine dev_kernel0()
  end

  attributes(global) subroutine dev_kernel1(a)
    real :: a
  end

  subroutine host()
    real, device :: a
    integer(8) :: stream
    integer(4) :: nbytes

! CHECK-LABEL: func.func @_QMtest_callPhost()
! CHECK: %[[A:.*]]:2 = hlfir.declare %{{.*}} {data_attr = #cuf.cuda<device>, uniq_name = "_QMtest_callFhostEa"} : (!fir.ref<f32>) -> (!fir.ref<f32>, !fir.ref<f32>)

    call dev_kernel0<<<10, 20>>>()
! CHECK: cuf.kernel_launch @_QMtest_callPdev_kernel0<<<%c10{{.*}}, %c1{{.*}}, %c1{{.*}}, %c20{{.*}}, %c1{{.*}}, %c1{{.*}}>>>()

    call dev_kernel0<<< __builtin_dim3(1,1,4), __builtin_dim3(32,1,1) >>>
! CHECK: %[[ADDR_DIM3_GRID:.*]] = fir.address_of(@_QQro._QM__fortran_builtinsT__builtin_dim3.{{.*}}) : !fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>
! CHECK: %[[DIM3_GRID:.*]]:2 = hlfir.declare %[[ADDR_DIM3_GRID]] {fortran_attrs = #fir.var_attrs<parameter>, uniq_name = "_QQro._QM__fortran_builtinsT__builtin_dim3.0"} : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>, !fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>)
! CHECK: %[[GRID_X:.*]] = hlfir.designate %[[DIM3_GRID]]#0{"x"}   : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<i32>
! CHECK: %[[GRID_X_LOAD:.*]] = fir.load %[[GRID_X]] : !fir.ref<i32>
! CHECK: %[[GRID_Y:.*]] = hlfir.designate %[[DIM3_GRID]]#0{"y"}   : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<i32>
! CHECK: %[[GRID_Y_LOAD:.*]] = fir.load %[[GRID_Y]] : !fir.ref<i32>
! CHECK: %[[GRID_Z:.*]] = hlfir.designate %[[DIM3_GRID]]#0{"z"}   : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<i32>
! CHECK: %[[GRID_Z_LOAD:.*]] = fir.load %[[GRID_Z]] : !fir.ref<i32>
! CHECK: %[[ADDR_DIM3_BLOCK:.*]] = fir.address_of(@_QQro._QM__fortran_builtinsT__builtin_dim3.{{.*}}) : !fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>
! CHECK: %[[DIM3_BLOCK:.*]]:2 = hlfir.declare %[[ADDR_DIM3_BLOCK]] {fortran_attrs = #fir.var_attrs<parameter>, uniq_name = "_QQro._QM__fortran_builtinsT__builtin_dim3.1"} : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>, !fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>)
! CHECK: %[[BLOCK_X:.*]] = hlfir.designate %[[DIM3_BLOCK]]#0{"x"}   : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<i32>
! CHECK: %[[BLOCK_X_LOAD:.*]] = fir.load %[[BLOCK_X]] : !fir.ref<i32>
! CHECK: %[[BLOCK_Y:.*]] = hlfir.designate %[[DIM3_BLOCK]]#0{"y"}   : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<i32>
! CHECK: %[[BLOCK_Y_LOAD:.*]] = fir.load %[[BLOCK_Y]] : !fir.ref<i32>
! CHECK: %[[BLOCK_Z:.*]] = hlfir.designate %[[DIM3_BLOCK]]#0{"z"}   : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<i32>
! CHECK: %[[BLOCK_Z_LOAD:.*]] = fir.load %[[BLOCK_Z]] : !fir.ref<i32>
! CHECK: cuf.kernel_launch @_QMtest_callPdev_kernel0<<<%[[GRID_X_LOAD]], %[[GRID_Y_LOAD]], %[[GRID_Z_LOAD]], %[[BLOCK_X_LOAD]], %[[BLOCK_Y_LOAD]], %[[BLOCK_Z_LOAD]]>>>()

    call dev_kernel0<<<10, 20, 2>>>()
! CHECK: cuf.kernel_launch @_QMtest_callPdev_kernel0<<<%c10{{.*}}, %c1{{.*}}, %c1{{.*}}, %c20{{.*}}, %c1{{.*}}, %c1{{.*}}, %c2{{.*}}>>>() 

    call dev_kernel0<<<10, 20, 2, 0_8>>>()
! CHECK: cuf.kernel_launch @_QMtest_callPdev_kernel0<<<%c10{{.*}}, %c1{{.*}}, %c1{{.*}}, %c20{{.*}}, %c1{{.*}}, %c1{{.*}}, %c2{{.*}}, %{{.*}} : !fir.ref<i64>>>>()

    call dev_kernel1<<<1, 32>>>(a)
! CHECK: cuf.kernel_launch @_QMtest_callPdev_kernel1<<<%c1{{.*}}, %c1{{.*}}, %c1{{.*}}, %c32{{.*}}, %c1{{.*}}, %c1{{.*}}>>>(%{{.*}}) : (!fir.ref<f32>)

    call dev_kernel1<<<*, 32>>>(a)
! CHECK: cuf.kernel_launch @_QMtest_callPdev_kernel1<<<%c-1{{.*}}, %c1{{.*}}, %c1{{.*}}, %c32{{.*}}, %c1{{.*}}, %c1{{.*}}>>>(%{{.*}})

    call dev_kernel1<<<*,32,0,stream>>>(a)
! CHECK: cuf.kernel_launch @_QMtest_callPdev_kernel1<<<%c-1{{.*}}, %c1{{.*}}, %c1{{.*}}, %c32{{.*}}, %c1{{.*}}, %c1{{.*}}, %c0{{.*}}, %{{.*}} : !fir.ref<i64>>>>(%{{.*}}) : (!fir.ref<f32>)

    call dev_kernel1<<<*, 32, 0.8 * nbytes>>>(a)
! CHECK: %[[MUL:.*]] = arith.mulf %{{.*}}, %{{.*}} fastmath<contract> : f32
! CHECK: %[[BYTES:.*]] = fir.convert %[[MUL]] : (f32) -> i32
! CHECK: cuf.kernel_launch @_QMtest_callPdev_kernel1<<<%{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %[[BYTES]]>>>(%{{.*}}) : (!fir.ref<f32>)
  end

end