TheRock/compiler/amd-llvm/flang/test/Lower/CUDA/cuda-doconc.cuf

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

! Check if do concurrent works inside cuf kernel directive

subroutine doconc1
  integer :: i, n
  integer, managed :: a(3)
  a(:) = -1
  n = 3
  n = n - 1
  !$cuf kernel do
  do concurrent(i=1:n)
    a(i) = 1
  end do
end

! CHECK-LABEL: func.func @_QPdoconc1()
! CHECK: %[[IMEM:.*]] = fir.alloca i32 {bindc_name = "i"}
! CHECK: %[[IV:.*]]:2 = hlfir.declare %[[IMEM]] {uniq_name = "_QFdoconc1Ei"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
! CHECK: cuf.kernel<<<*, *>>> (%[[ARG0:[^ ]*]] : index) =
! CHECK: %[[CONV:.*]] = fir.convert %[[ARG0]] : (index) -> i32
! CHECK: fir.store %[[CONV]] to %[[IV]]#0 : !fir.ref<i32>
! CHECK: fir.load %[[IV]]#0 : !fir.ref<i32>

subroutine doconc2
  integer :: i, j, m, n
  integer, managed :: a(2, 4)
  m = 2
  n = 4
  a(:,:) = -1
  !$cuf kernel do
  do concurrent(i=1:m,j=1:n)
    a(i,j) = i+j
  end do
end

! CHECK-LABEL: func.func @_QPdoconc2()
! CHECK: %[[JMEM:.*]] = fir.alloca i32 {bindc_name = "j"}
! CHECK: %[[IMEM:.*]] = fir.alloca i32 {bindc_name = "i"}
! CHECK: %[[IDECL:.*]]:2 = hlfir.declare %[[IMEM]] {uniq_name = "_QFdoconc2Ei"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
! CHECK: %[[JDECL:.*]]:2 = hlfir.declare %[[JMEM]] {uniq_name = "_QFdoconc2Ej"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
! CHECK: cuf.kernel<<<*, *>>> (%[[ARG0:[^ ]*]] : index, %[[ARG1:[^ ]*]] : index) =
! CHECK: %[[CONVI:.*]] = fir.convert %[[ARG0]] : (index) -> i32
! CHECK: fir.store %[[CONVI]] to %[[IDECL]]#0 : !fir.ref<i32>
! CHECK: %[[CONVJ:.*]] = fir.convert %[[ARG1]] : (index) -> i32
! CHECK: fir.store %[[CONVJ]] to %[[JDECL]]#0 : !fir.ref<i32>
! CHECK: fir.load %[[IDECL]]#0 : !fir.ref<i32>
! CHECK: fir.load %[[JDECL]]#0 : !fir.ref<i32>

! Test that do concurrent index variables in a cuf.kernel are allocated with
! their declared Fortran integer type (i32), not the loop index type (index).
! Previously the allocation used index, which prevented type conversions such
! as real(i) from compiling correctly.
subroutine doconc_real
  integer :: i, n
  real, managed :: a(4)
  n = 4
  !$cuf kernel do
  do concurrent(i=1:n)
    a(i) = real(i)
  end do
end

! CHECK-LABEL: func.func @_QPdoconc_real()
! CHECK: %[[IMEM:.*]] = fir.alloca i32 {bindc_name = "i"}
! CHECK: %[[IV:.*]]:2 = hlfir.declare %[[IMEM]] {uniq_name = "_QFdoconc_realEi"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
! CHECK: cuf.kernel<<<*, *>>> (%[[ARG0:[^ ]*]] : index) =
! CHECK: %[[CONV:.*]] = fir.convert %[[ARG0]] : (index) -> i32
! CHECK: fir.store %[[CONV]] to %[[IV]]#0 : !fir.ref<i32>
! CHECK: %[[LOADED:.*]] = fir.load %[[IV]]#0 : !fir.ref<i32>
! CHECK: fir.convert %[[LOADED]] : (i32) -> f32

! Test that do concurrent with multiple index variables and an explicit loop
! count (do(2)) generates correct IR.  Previously the PFT navigation loop
! assumed nested regular do-loops and asserted when it encountered the flat
! do concurrent structure.
subroutine doconc_do2
  integer :: i, j, k, n
  real, managed :: a(4, 2)
  k = 2
  n = 4
  !$cuf kernel do(2) <<<k, n>>>
  do concurrent(i=1:k, j=1:n)
    a(j,i) = 1.0
  end do
end

! CHECK-LABEL: func.func @_QPdoconc_do2()
! j is allocated before i in the alloca block (loop control order).
! CHECK: %[[JMEM:.*]] = fir.alloca i32 {bindc_name = "j"}
! CHECK: %[[IMEM:.*]] = fir.alloca i32 {bindc_name = "i"}
! CHECK: %[[IDECL:.*]]:2 = hlfir.declare %[[IMEM]] {uniq_name = "_QFdoconc_do2Ei"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
! CHECK: %[[JDECL:.*]]:2 = hlfir.declare %[[JMEM]] {uniq_name = "_QFdoconc_do2Ej"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
! CHECK: cuf.kernel{{.*}} (%[[ARG0:[^ ]*]] : index, %[[ARG1:[^ ]*]] : index) =
! CHECK: %[[CONVI:.*]] = fir.convert %[[ARG0]] : (index) -> i32
! CHECK: fir.store %[[CONVI]] to %[[IDECL]]#0 : !fir.ref<i32>
! CHECK: %[[CONVJ:.*]] = fir.convert %[[ARG1]] : (index) -> i32
! CHECK: fir.store %[[CONVJ]] to %[[JDECL]]#0 : !fir.ref<i32>