| ! Tests LOCAL_INIT + REDUCE together through the device OpenMP conversion | |
| ! pipeline. Verifies that LOCAL_INIT's per-iteration copy is inlined at the | |
| ! top of the omp.loop_nest body and the reduction clause is on omp.wsloop, | |
| ! within the full target teams distribute parallel do structure. | |
| ! RUN: %flang_fc1 -emit-hlfir -fopenmp -fdo-concurrent-to-openmp=device %s -o - \ | |
| ! RUN: | FileCheck %s | |
| subroutine test_local_init_reduce_device(n) | |
| implicit none | |
| integer, intent(in) :: n | |
| real :: tmp, total | |
| integer :: i | |
| tmp = 1.0 | |
| total = 0.0 | |
| do concurrent (i = 1:n) local_init(tmp) reduce(+:total) | |
| tmp = tmp + real(i) | |
| total = total + tmp | |
| end do | |
| end subroutine | |
| ! Verify: omp.private for LOCAL_INIT (no copy_region) | |
| ! CHECK: omp.private {type = private} @[[PRIVATIZER:.*tmp.*\.omp]] : f32 | |
| ! Verify: full device nesting with reduction on teams+wsloop, private on parallel | |
| ! CHECK: omp.target | |
| ! CHECK: omp.teams reduction(@{{.*}} %{{.*}} -> %{{.*}} : !fir.ref<f32>) { | |
| ! CHECK: omp.parallel private(@[[PRIVATIZER]] %[[ORIG:.*]] -> %[[PRIV:.*]] : !fir.ref<f32>) { | |
| ! CHECK: omp.distribute { | |
| ! CHECK: omp.wsloop reduction(@{{.*}} %{{.*}} -> %{{.*}} : !fir.ref<f32>) { | |
| ! CHECK: omp.loop_nest | |
| ! Verify: per-iteration copy at top of loop_nest body | |
| ! CHECK: %[[VAL:.*]] = fir.load %[[ORIG]] : !fir.ref<f32> | |
| ! CHECK-NEXT: hlfir.assign %[[VAL]] to %[[PRIV]] : f32, !fir.ref<f32> |