| // Tests that LOCAL_INIT locality specifier is correctly converted to OpenMP | |
| // for the device path. The key verification points are: | |
| // | |
| // 1. The omp.private op has type=private with NO copy_region (same as LOCAL). | |
| // 2. The private variable is attached to omp.parallel (not omp.wsloop) on the | |
| // device path. | |
| // 3. The copy_region operations from fir.local are inlined at the top of | |
| // the omp.loop_nest body, giving per-iteration initialization. | |
| // | |
| // This ensures LOCAL_INIT's per-iteration semantics (each iteration sees the | |
| // original value) rather than OpenMP firstprivate's per-thread semantics. | |
| // RUN: fir-opt --omp-do-concurrent-conversion="map-to=device" %s | FileCheck %s | |
| // A LOCAL_INIT localizer for f32. The copy_region loads from the original | |
| // variable and stores to the private. | |
| fir.local {type = local_init} @_QFtestEtmp_firstprivate_f32 : f32 copy { | |
| ^bb0(%arg0: !fir.ref<f32>, %arg1: !fir.ref<f32>): | |
| %0 = fir.load %arg0 : !fir.ref<f32> | |
| fir.store %0 to %arg1 : !fir.ref<f32> | |
| fir.yield(%arg1 : !fir.ref<f32>) | |
| } | |
| func.func @_QPtest() { | |
| %c1 = arith.constant 1 : index | |
| %c10 = arith.constant 10 : index | |
| %cst = arith.constant 3.140000e+00 : f32 | |
| %0 = fir.alloca f32 {bindc_name = "tmp", uniq_name = "_QFtestEtmp"} | |
| %1 = fir.declare %0 uniq_name("_QFtestEtmp") : (!fir.ref<f32>) -> !fir.ref<f32> | |
| fir.store %cst to %1 : !fir.ref<f32> | |
| %2 = fir.address_of(@_QFEa) : !fir.ref<!fir.array<10xf32>> | |
| %3 = fir.shape %c10 : (index) -> !fir.shape<1> | |
| %4:2 = hlfir.declare %2(%3) uniq_name("_QFtestEa") : (!fir.ref<!fir.array<10xf32>>, !fir.shape<1>) -> (!fir.ref<!fir.array<10xf32>>, !fir.ref<!fir.array<10xf32>>) | |
| fir.do_concurrent { | |
| %5 = fir.alloca i32 {bindc_name = "i"} | |
| %6:2 = hlfir.declare %5 uniq_name("_QFtestEi") : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>) | |
| fir.do_concurrent.loop (%arg0) = (%c1) to (%c10) step (%c1) | |
| local(@_QFtestEtmp_firstprivate_f32 %1 -> %arg1 : !fir.ref<f32>) { | |
| %7 = fir.convert %arg0 : (index) -> i32 | |
| fir.store %7 to %6#0 : !fir.ref<i32> | |
| %8 = fir.declare %arg1 uniq_name("_QFtestEtmp") : (!fir.ref<f32>) -> !fir.ref<f32> | |
| %9 = fir.load %8 : !fir.ref<f32> | |
| %10 = fir.load %6#0 : !fir.ref<i32> | |
| %11 = fir.convert %10 : (i32) -> f32 | |
| %12 = arith.addf %9, %11 fastmath<contract> : f32 | |
| hlfir.assign %12 to %8 : f32, !fir.ref<f32> | |
| } | |
| } | |
| return | |
| } | |
| fir.global @_QFEa : !fir.array<10xf32> | |
| // Verify: omp.private has type=private and NO copy_region | |
| // CHECK: omp.private {type = private} @[[PRIVATIZER:.*_QFtestEtmp.*\.omp]] : f32 | |
| // Verify: full device nesting structure with private on omp.parallel | |
| // CHECK: omp.target | |
| // CHECK: omp.teams { | |
| // CHECK: omp.parallel private(@[[PRIVATIZER]] %[[ORIG:.*]] -> %[[PRIV:.*]] : !fir.ref<f32>) { | |
| // CHECK: omp.distribute { | |
| // CHECK: omp.wsloop { | |
| // CHECK: omp.loop_nest | |
| // Verify: per-iteration copy is inlined at the top of the loop_nest body | |
| // CHECK: %[[LOADED:.*]] = fir.load %[[ORIG]] : !fir.ref<f32> | |
| // CHECK-NEXT: fir.store %[[LOADED]] to %[[PRIV]] : !fir.ref<f32> |