97 lines · plain
1// RUN: mlir-translate -mlir-to-llvmir -split-input-file %s | FileCheck %s2 3// test a parallel reduction with a cleanup region4 5 omp.declare_reduction @add_reduction_i_32 : !llvm.ptr init {6 ^bb0(%arg0: !llvm.ptr):7 %0 = llvm.mlir.constant(0 : i32) : i328 %c4 = llvm.mlir.constant(4 : i64) : i649 %2 = llvm.call @malloc(%c4) : (i64) -> !llvm.ptr10 llvm.store %0, %2 : i32, !llvm.ptr11 omp.yield(%2 : !llvm.ptr)12 } combiner {13 ^bb0(%arg0: !llvm.ptr, %arg1: !llvm.ptr):14 %0 = llvm.load %arg0 : !llvm.ptr -> i3215 %1 = llvm.load %arg1 : !llvm.ptr -> i3216 %2 = llvm.add %0, %1 : i3217 llvm.store %2, %arg0 : i32, !llvm.ptr18 omp.yield(%arg0 : !llvm.ptr)19 } cleanup {20 ^bb0(%arg0: !llvm.ptr):21 llvm.call @free(%arg0) : (!llvm.ptr) -> ()22 omp.yield23 }24 25 // CHECK-LABEL: @main26 llvm.func @main() {27 %0 = llvm.mlir.constant(-1 : i32) : i3228 %1 = llvm.mlir.addressof @i : !llvm.ptr29 %2 = llvm.mlir.addressof @j : !llvm.ptr30 omp.parallel reduction(byref @add_reduction_i_32 %1 -> %arg0, byref @add_reduction_i_32 %2 -> %arg1 : !llvm.ptr, !llvm.ptr) {31 llvm.store %0, %arg0 : i32, !llvm.ptr32 llvm.store %0, %arg1 : i32, !llvm.ptr33 omp.terminator34 }35 llvm.return36 }37 llvm.mlir.global internal @i() {addr_space = 0 : i32} : i32 {38 %0 = llvm.mlir.constant(0 : i32) : i3239 llvm.return %0 : i3240 }41 llvm.mlir.global internal @j() {addr_space = 0 : i32} : i32 {42 %0 = llvm.mlir.constant(0 : i32) : i3243 llvm.return %0 : i3244 }45 llvm.func @malloc(%arg0 : i64) -> !llvm.ptr46 llvm.func @free(%arg0 : !llvm.ptr) -> ()47 48// CHECK: %{{.+}} = 49// Call to the outlined function.50// CHECK: call void {{.*}} @__kmpc_fork_call51// CHECK-SAME: @[[OUTLINED:[A-Za-z_.][A-Za-z0-9_.]*]]52 53// Outlined function.54// CHECK: define internal void @[[OUTLINED]]55 56// Private reduction variable and its initialization.57// CHECK: %tid.addr.local = alloca i3258// CHECK: %[[PRIV_PTR_I:.+]] = alloca ptr59// CHECK: %[[PRIV_PTR_J:.+]] = alloca ptr60// CHECK: %[[MALLOC_I:.+]] = call ptr @malloc(i64 4)61// CHECK: store ptr %[[MALLOC_I]], ptr %[[PRIV_PTR_I]]62// CHECK: %[[MALLOC_J:.+]] = call ptr @malloc(i64 4)63// CHECK: store ptr %[[MALLOC_J]], ptr %[[PRIV_PTR_J]]64 65// Call to the reduction function.66// CHECK: call i32 @__kmpc_reduce67// CHECK-SAME: @[[REDFUNC:[A-Za-z_.][A-Za-z0-9_.]*]]68 69 70// Non-atomic reduction:71// CHECK: %[[PRIV_VAL_PTR_I:.+]] = load ptr, ptr %[[PRIV_PTR_I]]72// CHECK: %[[LOAD_I:.+]] = load i32, ptr @i73// CHECK: %[[PRIV_VAL_I:.+]] = load i32, ptr %[[PRIV_VAL_PTR_I]]74// CHECK: %[[SUM_I:.+]] = add i32 %[[LOAD_I]], %[[PRIV_VAL_I]]75// CHECK: store i32 %[[SUM_I]], ptr @i76// CHECK: %[[PRIV_VAL_PTR_J:.+]] = load ptr, ptr %[[PRIV_PTR_J]]77// CHECK: %[[LOAD_J:.+]] = load i32, ptr @j78// CHECK: %[[PRIV_VAL_J:.+]] = load i32, ptr %[[PRIV_VAL_PTR_J]]79// CHECK: %[[SUM_J:.+]] = add i32 %[[LOAD_J]], %[[PRIV_VAL_J]]80// CHECK: store i32 %[[SUM_J]], ptr @j81// CHECK: call void @__kmpc_end_reduce82// CHECK: br label %[[FINALIZE:.+]]83 84// CHECK: [[FINALIZE]]:85// CHECK: br label %[[OMP_FINALIZE:.+]]86 87// Cleanup region:88// CHECK: [[OMP_FINALIZE]]:89// CHECK: %[[PRIV_I:.+]] = load ptr, ptr %[[PRIV_PTR_I]], align 890// CHECK: call void @free(ptr %[[PRIV_I]])91// CHECK: %[[PRIV_J:.+]] = load ptr, ptr %[[PRIV_PTR_J]], align 892// CHECK: call void @free(ptr %[[PRIV_J]])93 94// Reduction function.95// CHECK: define internal void @[[REDFUNC]]96// CHECK: add i3297