// RUN: fir-opt %s --acc-optimize-firstprivate-map -split-input-file | FileCheck %s // Test: Integer variable - should optimize func.func private @use_i32(i32) // CHECK-LABEL: func.func @test_trivial_scalar_hoist func.func @test_trivial_scalar_hoist() { %scalar = fir.alloca i32 {bindc_name = "scalar_var"} %decl = fir.declare %scalar {uniq_name = "_QFtest_trivial_scalarEscalar_var"} : (!fir.ref) -> !fir.ref %fpmap = acc.firstprivate_map varPtr(%decl : !fir.ref) -> !fir.ref // CHECK: %[[DECL:.*]] = fir.declare // CHECK: %[[LOAD:.*]] = fir.load %[[DECL]] : !fir.ref // CHECK-NOT: acc.firstprivate_map // CHECK: acc.parallel acc.parallel { %load = fir.load %fpmap : !fir.ref // CHECK: fir.call @use_i32(%[[LOAD]]) fir.call @use_i32(%load) : (i32) -> () acc.yield } return } // ----- // Test: Inside offload region - should optimize by getting rid of firstprivate_map func.func private @use_i32(i32) // CHECK-LABEL: func.func @test_inside_offload_region func.func @test_inside_offload_region() { %scalar = fir.alloca i32 {bindc_name = "scalar_var"} %decl = fir.declare %scalar {uniq_name = "_QFtest_inside_offloadEscalar_var"} : (!fir.ref) -> !fir.ref %copyin = acc.copyin varPtr(%decl : !fir.ref) -> !fir.ref // CHECK: %[[DECL:.*]] = fir.declare // CHECK: %[[COPYIN:.*]] = acc.copyin varPtr(%[[DECL]] : !fir.ref) -> !fir.ref acc.parallel dataOperands(%copyin : !fir.ref) { %fpmap = acc.firstprivate_map varPtr(%copyin : !fir.ref) -> !fir.ref %load = fir.load %fpmap : !fir.ref // CHECK: acc.parallel dataOperands(%[[COPYIN]] : !fir.ref) // CHECK-NOT: acc.firstprivate_map // CHECK: %[[LOAD:.*]] = fir.load %[[COPYIN]] : !fir.ref fir.call @use_i32(%load) : (i32) -> () acc.yield } return } // ----- // Test: Local alloca inside offload region - should optimize (erase) func.func private @use_i32(i32) // CHECK-LABEL: func.func @test_local_alloca_inside_offload func.func @test_local_alloca_inside_offload() { acc.parallel { %local = fir.alloca i32 {bindc_name = "local_var"} %decl = fir.declare %local {uniq_name = "_QFtest_local_allocaElocal_var"} : (!fir.ref) -> !fir.ref %fpmap = acc.firstprivate_map varPtr(%decl : !fir.ref) -> !fir.ref %load = fir.load %fpmap : !fir.ref // CHECK: acc.parallel // CHECK: %[[LOCAL:.*]] = fir.alloca i32 // CHECK: %[[DECL:.*]] = fir.declare %[[LOCAL]] // CHECK-NOT: acc.firstprivate_map // CHECK: %[[LOAD:.*]] = fir.load %[[DECL]] : !fir.ref fir.call @use_i32(%load) : (i32) -> () acc.yield } return } // ----- // Test: firstprivate_map with acc.private input - should optimize (erase) func.func private @use_i32(i32) // CHECK-LABEL: func.func @test_private_input func.func @test_private_input() { %scalar = fir.alloca i32 {bindc_name = "scalar_var"} %decl = fir.declare %scalar {uniq_name = "_QFtest_private_inputEscalar_var"} : (!fir.ref) -> !fir.ref %private = acc.private varPtr(%decl : !fir.ref) -> !fir.ref // CHECK: %[[DECL:.*]] = fir.declare // CHECK: %[[PRIVATE:.*]] = acc.private varPtr(%[[DECL]] : !fir.ref) -> !fir.ref acc.parallel private(%private : !fir.ref) { %fpmap = acc.firstprivate_map varPtr(%private : !fir.ref) -> !fir.ref %load = fir.load %fpmap : !fir.ref // CHECK: acc.parallel private(%[[PRIVATE]] : !fir.ref) // CHECK-NOT: acc.firstprivate_map // CHECK: %[[LOAD:.*]] = fir.load %[[PRIVATE]] : !fir.ref fir.call @use_i32(%load) : (i32) -> () acc.yield } return } // ----- // Test: Variable defined outside offload region but firstprivate_map inside - // should hoist out and then optimize func.func private @use_i32(i32) // CHECK-LABEL: func.func @test_hoist_from_offload_region func.func @test_hoist_from_offload_region() { %scalar = fir.alloca i32 {bindc_name = "scalar_var"} %decl = fir.declare %scalar {uniq_name = "_QFtest_hoistEscalar_var"} : (!fir.ref) -> !fir.ref // CHECK: %[[DECL:.*]] = fir.declare // CHECK: %[[LOAD:.*]] = fir.load %[[DECL]] : !fir.ref // CHECK-NOT: acc.firstprivate_map // CHECK: acc.parallel acc.parallel { %fpmap = acc.firstprivate_map varPtr(%decl : !fir.ref) -> !fir.ref %load = fir.load %fpmap : !fir.ref // CHECK: fir.call @use_i32(%[[LOAD]]) fir.call @use_i32(%load) : (i32) -> () acc.yield } return } // ----- // Test: Optional variable - should NOT optimize func.func private @use_i32(i32) // CHECK-LABEL: func.func @test_optional_no_hoist func.func @test_optional_no_hoist(%arg0: !fir.ref) { %decl = fir.declare %arg0 {fortran_attrs = #fir.var_attrs, uniq_name = "_QFtest_optionalEopt_var"} : (!fir.ref) -> !fir.ref %fpmap = acc.firstprivate_map varPtr(%decl : !fir.ref) -> !fir.ref // CHECK: %[[DECL:.*]] = fir.declare {{.*}} {fortran_attrs = #fir.var_attrs // CHECK: %[[FPMAP:.*]] = acc.firstprivate_map varPtr(%[[DECL]] : !fir.ref) -> !fir.ref // CHECK: acc.parallel acc.parallel { %load = fir.load %fpmap : !fir.ref // CHECK: %[[LOAD:.*]] = fir.load %[[FPMAP]] : !fir.ref fir.call @use_i32(%load) : (i32) -> () acc.yield } return } // ----- // Test: Array type (non-trivial) - should NOT optimize func.func private @use_array(!fir.array<10xi32>) // CHECK-LABEL: func.func @test_array_no_hoist func.func @test_array_no_hoist() { %c10 = arith.constant 10 : index %array = fir.alloca !fir.array<10xi32> {bindc_name = "array_var"} %shape = fir.shape %c10 : (index) -> !fir.shape<1> %decl = fir.declare %array(%shape) {uniq_name = "_QFtest_arrayEarray_var"} : (!fir.ref>, !fir.shape<1>) -> !fir.ref> %fpmap = acc.firstprivate_map varPtr(%decl : !fir.ref>) -> !fir.ref> // CHECK: %[[DECL:.*]] = fir.declare // CHECK: %[[FPMAP:.*]] = acc.firstprivate_map varPtr(%[[DECL]] : !fir.ref>) -> !fir.ref> // CHECK: acc.parallel acc.parallel { %load = fir.load %fpmap : !fir.ref> // CHECK: %[[LOAD:.*]] = fir.load %[[FPMAP]] : !fir.ref> fir.call @use_array(%load) : (!fir.array<10xi32>) -> () acc.yield } return } // ----- // Test: Multiple loads from same firstprivate_map - should optimize and hoist all func.func private @use_i32_i32(i32, i32) // CHECK-LABEL: func.func @test_multiple_loads_hoist func.func @test_multiple_loads_hoist() { %scalar = fir.alloca i32 {bindc_name = "scalar_var"} %decl = fir.declare %scalar {uniq_name = "_QFtest_multiple_loadsEscalar_var"} : (!fir.ref) -> !fir.ref %fpmap = acc.firstprivate_map varPtr(%decl : !fir.ref) -> !fir.ref // CHECK: %[[DECL:.*]] = fir.declare // CHECK-DAG: %[[LOAD1:.*]] = fir.load %[[DECL]] : !fir.ref // CHECK-DAG: %[[LOAD2:.*]] = fir.load %[[DECL]] : !fir.ref // CHECK-NOT: acc.firstprivate_map // CHECK: acc.parallel acc.parallel { %load1 = fir.load %fpmap : !fir.ref %load2 = fir.load %fpmap : !fir.ref fir.call @use_i32_i32(%load1, %load2) : (i32, i32) -> () acc.yield } return } // ----- // Test: Variable through fir.convert - should optimize func.func private @use_i32(i32) // CHECK-LABEL: func.func @test_through_convert func.func @test_through_convert() { %scalar = fir.alloca i32 {bindc_name = "scalar_var"} %decl = fir.declare %scalar {uniq_name = "_QFtest_convertEscalar_var"} : (!fir.ref) -> !fir.ref %convert = fir.convert %decl : (!fir.ref) -> !fir.ref %fpmap = acc.firstprivate_map varPtr(%convert : !fir.ref) -> !fir.ref // CHECK: %[[DECL:.*]] = fir.declare // CHECK: %[[CONVERT:.*]] = fir.convert %[[DECL]] // CHECK: %[[LOAD:.*]] = fir.load %[[CONVERT]] : !fir.ref // CHECK-NOT: acc.firstprivate_map // CHECK: acc.parallel acc.parallel { %load = fir.load %fpmap : !fir.ref // CHECK: fir.call @use_i32(%[[LOAD]]) fir.call @use_i32(%load) : (i32) -> () acc.yield } return } // ----- // Test: Block argument (unknown origin) - should NOT optimize func.func private @use_i32(i32) // CHECK-LABEL: func.func @test_block_arg_no_hoist func.func @test_block_arg_no_hoist(%arg0: !fir.ref) { // No declare op, so we can't determine if it's optional - conservative no-op %fpmap = acc.firstprivate_map varPtr(%arg0 : !fir.ref) -> !fir.ref // CHECK: %[[FPMAP:.*]] = acc.firstprivate_map varPtr(%arg0 : !fir.ref) -> !fir.ref // CHECK: acc.parallel acc.parallel { %load = fir.load %fpmap : !fir.ref // CHECK: %[[LOAD:.*]] = fir.load %[[FPMAP]] : !fir.ref fir.call @use_i32(%load) : (i32) -> () acc.yield } return }