91 lines · plain
1// RUN: mlir-translate -mlir-to-llvmir -split-input-file %s | FileCheck %s2 3// test a wsloop 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 %loop_ub = llvm.mlir.constant(9 : i32) : i3231 %loop_lb = llvm.mlir.constant(0 : i32) : i3232 %loop_step = llvm.mlir.constant(1 : i32) : i32 33 omp.wsloop reduction(byref @add_reduction_i_32 %1 -> %arg0, byref @add_reduction_i_32 %2 -> %arg1 : !llvm.ptr, !llvm.ptr) {34 omp.loop_nest (%loop_cnt) : i32 = (%loop_lb) to (%loop_ub) inclusive step (%loop_step) {35 llvm.store %0, %arg0 : i32, !llvm.ptr36 llvm.store %0, %arg1 : i32, !llvm.ptr37 omp.yield38 }39 }40 llvm.return41 }42 llvm.mlir.global internal @i() {addr_space = 0 : i32} : i32 {43 %0 = llvm.mlir.constant(0 : i32) : i3244 llvm.return %0 : i3245 }46 llvm.mlir.global internal @j() {addr_space = 0 : i32} : i32 {47 %0 = llvm.mlir.constant(0 : i32) : i3248 llvm.return %0 : i3249 }50 llvm.func @malloc(%arg0 : i64) -> !llvm.ptr51 llvm.func @free(%arg0 : !llvm.ptr) -> ()52 53// Private reduction variable and its initialization.54// CHECK: %[[PRIV_PTR_I:.+]] = alloca ptr55// CHECK: %[[PRIV_PTR_J:.+]] = alloca ptr56// CHECK: %[[MALLOC_I:.+]] = call ptr @malloc(i64 4)57// CHECK: store ptr %[[MALLOC_I]], ptr %[[PRIV_PTR_I]]58// CHECK: %[[MALLOC_J:.+]] = call ptr @malloc(i64 4)59// CHECK: store ptr %[[MALLOC_J]], ptr %[[PRIV_PTR_J]]60 61// Call to the reduction function.62// CHECK: call i32 @__kmpc_reduce63// CHECK-SAME: @[[REDFUNC:[A-Za-z_.][A-Za-z0-9_.]*]]64 65// Weirdly the finalization block is generated before the reduction blocks:66// CHECK: [[FINALIZE:.+]]:67// CHECK: call void @__kmpc_barrier68// CHECK: %[[PRIV_I:.+]] = load ptr, ptr %[[PRIV_PTR_I]], align 869// CHECK: call void @free(ptr %[[PRIV_I]])70// CHECK: %[[PRIV_J:.+]] = load ptr, ptr %[[PRIV_PTR_J]], align 871// CHECK: call void @free(ptr %[[PRIV_J]])72// CHECK: ret void73 74// Non-atomic reduction:75// CHECK: %[[PRIV_VAL_PTR_I:.+]] = load ptr, ptr %[[PRIV_PTR_I]]76// CHECK: %[[LOAD_I:.+]] = load i32, ptr @i77// CHECK: %[[PRIV_VAL_I:.+]] = load i32, ptr %[[PRIV_VAL_PTR_I]]78// CHECK: %[[SUM_I:.+]] = add i32 %[[LOAD_I]], %[[PRIV_VAL_I]]79// CHECK: store i32 %[[SUM_I]], ptr @i80// CHECK: %[[PRIV_VAL_PTR_J:.+]] = load ptr, ptr %[[PRIV_PTR_J]]81// CHECK: %[[LOAD_J:.+]] = load i32, ptr @j82// CHECK: %[[PRIV_VAL_J:.+]] = load i32, ptr %[[PRIV_VAL_PTR_J]]83// CHECK: %[[SUM_J:.+]] = add i32 %[[LOAD_J]], %[[PRIV_VAL_J]]84// CHECK: store i32 %[[SUM_J]], ptr @j85// CHECK: call void @__kmpc_end_reduce86// CHECK: br label %[[FINALIZE]]87 88// Reduction function.89// CHECK: define internal void @[[REDFUNC]]90// CHECK: add i3291