|
| 1 | +// Test affine promotion pass |
| 2 | + |
| 3 | +// RUN: fir-opt --split-input-file --promote-to-affine --affine-loop-invariant-code-motion --cse %s | FileCheck %s |
| 4 | + |
| 5 | +!arr_d1 = type !fir.ref<!fir.array<?xf32>> |
| 6 | +#arr_len = affine_map<()[j1,k1] -> (k1 - j1 + 1)> |
| 7 | + |
| 8 | +func @loop_with_load_and_store(%a1: !arr_d1, %a2: !arr_d1, %a3: !arr_d1) { |
| 9 | + %c1 = constant 1 : index |
| 10 | + %c0 = constant 0 : index |
| 11 | + %len = constant 100 : index |
| 12 | + %dims = fir.shape %len : (index) -> !fir.shape<1> |
| 13 | + %siz = affine.apply #arr_len()[%c1,%len] |
| 14 | + %t1 = fir.alloca !fir.array<?xf32>, %siz |
| 15 | + |
| 16 | + fir.do_loop %i = %c1 to %len step %c1 { |
| 17 | + %a1_idx = fir.array_coor %a1(%dims) %i |
| 18 | + : (!arr_d1, !fir.shape<1>, index) -> !fir.ref<f32> |
| 19 | + %a1_v = fir.load %a1_idx : !fir.ref<f32> |
| 20 | + |
| 21 | + %a2_idx = fir.array_coor %a2(%dims) %i |
| 22 | + : (!arr_d1, !fir.shape<1>, index) -> !fir.ref<f32> |
| 23 | + %a2_v = fir.load %a2_idx : !fir.ref<f32> |
| 24 | + |
| 25 | + %v = addf %a1_v, %a2_v : f32 |
| 26 | + %t1_idx = fir.array_coor %t1(%dims) %i |
| 27 | + : (!arr_d1, !fir.shape<1>, index) -> !fir.ref<f32> |
| 28 | + |
| 29 | + fir.store %v to %t1_idx : !fir.ref<f32> |
| 30 | + } |
| 31 | + fir.do_loop %i = %c1 to %len step %c1 { |
| 32 | + %t1_idx = fir.array_coor %t1(%dims) %i |
| 33 | + : (!arr_d1, !fir.shape<1>, index) -> !fir.ref<f32> |
| 34 | + %t1_v = fir.load %t1_idx : !fir.ref<f32> |
| 35 | + |
| 36 | + %a2_idx = fir.array_coor %a2(%dims) %i |
| 37 | + : (!arr_d1, !fir.shape<1>, index) -> !fir.ref<f32> |
| 38 | + %a2_v = fir.load %a2_idx : !fir.ref<f32> |
| 39 | + |
| 40 | + %v = mulf %t1_v, %a2_v : f32 |
| 41 | + %a3_idx = fir.array_coor %a3(%dims) %i |
| 42 | + : (!arr_d1, !fir.shape<1>, index) -> !fir.ref<f32> |
| 43 | + |
| 44 | + fir.store %v to %a3_idx : !fir.ref<f32> |
| 45 | + } |
| 46 | + return |
| 47 | +} |
| 48 | + |
| 49 | +// CHECK: func @loop_with_load_and_store(%[[VAL_0:.*]]: !fir.ref<!fir.array<?xf32>>, %[[VAL_1:.*]]: !fir.ref<!fir.array<?xf32>>, %[[VAL_2:.*]]: !fir.ref<!fir.array<?xf32>>) { |
| 50 | +// CHECK: %[[VAL_3:.*]] = constant 1 : index |
| 51 | +// CHECK: %[[VAL_4:.*]] = constant 100 : index |
| 52 | +// CHECK: %[[VAL_5:.*]] = fir.shape %[[VAL_4]] : (index) -> !fir.shape<1> |
| 53 | +// CHECK: %[[VAL_6:.*]] = affine.apply #map0(){{\[}}%[[VAL_3]], %[[VAL_4]]] |
| 54 | +// CHECK: %[[VAL_7:.*]] = fir.alloca !fir.array<?xf32>, %[[VAL_6]] |
| 55 | +// CHECK: %[[VAL_8:.*]] = fir.convert %[[VAL_0]] : (!fir.ref<!fir.array<?xf32>>) -> memref<?xf32> |
| 56 | +// CHECK: %[[VAL_9:.*]] = fir.convert %[[VAL_1]] : (!fir.ref<!fir.array<?xf32>>) -> memref<?xf32> |
| 57 | +// CHECK: %[[VAL_10:.*]] = fir.convert %[[VAL_7]] : (!fir.ref<!fir.array<?xf32>>) -> memref<?xf32> |
| 58 | +// CHECK: affine.for %[[VAL_11:.*]] = %[[VAL_3]] to #map1(){{\[}}%[[VAL_4]]] { |
| 59 | +// CHECK: %[[VAL_12:.*]] = affine.apply #map2(%[[VAL_11]]){{\[}}%[[VAL_3]], %[[VAL_4]], %[[VAL_3]]] |
| 60 | +// CHECK: %[[VAL_13:.*]] = affine.load %[[VAL_8]]{{\[}}%[[VAL_12]]] : memref<?xf32> |
| 61 | +// CHECK: %[[VAL_14:.*]] = affine.load %[[VAL_9]]{{\[}}%[[VAL_12]]] : memref<?xf32> |
| 62 | +// CHECK: %[[VAL_15:.*]] = addf %[[VAL_13]], %[[VAL_14]] : f32 |
| 63 | +// CHECK: affine.store %[[VAL_15]], %[[VAL_10]]{{\[}}%[[VAL_12]]] : memref<?xf32> |
| 64 | +// CHECK: } |
| 65 | +// CHECK: %[[VAL_16:.*]] = fir.convert %[[VAL_2]] : (!fir.ref<!fir.array<?xf32>>) -> memref<?xf32> |
| 66 | +// CHECK: affine.for %[[VAL_17:.*]] = %[[VAL_3]] to #map1(){{\[}}%[[VAL_4]]] { |
| 67 | +// CHECK: %[[VAL_18:.*]] = affine.apply #map2(%[[VAL_17]]){{\[}}%[[VAL_3]], %[[VAL_4]], %[[VAL_3]]] |
| 68 | +// CHECK: %[[VAL_19:.*]] = affine.load %[[VAL_10]]{{\[}}%[[VAL_18]]] : memref<?xf32> |
| 69 | +// CHECK: %[[VAL_20:.*]] = affine.load %[[VAL_9]]{{\[}}%[[VAL_18]]] : memref<?xf32> |
| 70 | +// CHECK: %[[VAL_21:.*]] = mulf %[[VAL_19]], %[[VAL_20]] : f32 |
| 71 | +// CHECK: affine.store %[[VAL_21]], %[[VAL_16]]{{\[}}%[[VAL_18]]] : memref<?xf32> |
| 72 | +// CHECK: } |
| 73 | +// CHECK: return |
| 74 | +// CHECK: } |
| 75 | + |
| 76 | +// ----- |
| 77 | + |
| 78 | +!arr_d1 = type !fir.ref<!fir.array<?xf32>> |
| 79 | +#arr_len = affine_map<()[j1,k1] -> (k1 - j1 + 1)> |
| 80 | + |
| 81 | +func @loop_with_if(%a: !arr_d1, %v: f32) { |
| 82 | + %c0 = constant 0 : index |
| 83 | + %c1 = constant 1 : index |
| 84 | + %c2 = constant 2 : index |
| 85 | + %len = constant 100 : index |
| 86 | + %dims = fir.shape %len : (index) -> !fir.shape<1> |
| 87 | + |
| 88 | + fir.do_loop %i = %c1 to %len step %c1 { |
| 89 | + fir.do_loop %j = %c1 to %len step %c1 { |
| 90 | + fir.do_loop %k = %c1 to %len step %c1 { |
| 91 | + %im2 = subi %i, %c2 : index |
| 92 | + %cond = cmpi "sgt", %im2, %c0 : index |
| 93 | + fir.if %cond { |
| 94 | + %a_idx = fir.array_coor %a(%dims) %i |
| 95 | + : (!arr_d1, !fir.shape<1>, index) -> !fir.ref<f32> |
| 96 | + fir.store %v to %a_idx : !fir.ref<f32> |
| 97 | + } |
| 98 | + %aj_idx = fir.array_coor %a(%dims) %j |
| 99 | + : (!arr_d1, !fir.shape<1>, index) -> !fir.ref<f32> |
| 100 | + fir.store %v to %aj_idx : !fir.ref<f32> |
| 101 | + %ak_idx = fir.array_coor %a(%dims) %k |
| 102 | + : (!arr_d1, !fir.shape<1>, index) -> !fir.ref<f32> |
| 103 | + fir.store %v to %ak_idx : !fir.ref<f32> |
| 104 | + } |
| 105 | + } |
| 106 | + } |
| 107 | + return |
| 108 | +} |
| 109 | + |
| 110 | +// CHECK: func @loop_with_if(%[[VAL_0:.*]]: !fir.ref<!fir.array<?xf32>>, %[[VAL_1:.*]]: f32) { |
| 111 | +// CHECK: %[[VAL_2:.*]] = constant 0 : index |
| 112 | +// CHECK: %[[VAL_3:.*]] = constant 1 : index |
| 113 | +// CHECK: %[[VAL_4:.*]] = constant 2 : index |
| 114 | +// CHECK: %[[VAL_5:.*]] = constant 100 : index |
| 115 | +// CHECK: %[[VAL_6:.*]] = fir.shape %[[VAL_5]] : (index) -> !fir.shape<1> |
| 116 | +// CHECK: %[[VAL_7:.*]] = fir.convert %[[VAL_0]] : (!fir.ref<!fir.array<?xf32>>) -> memref<?xf32> |
| 117 | +// CHECK: affine.for %[[VAL_8:.*]] = %[[VAL_3]] to #map0(){{\[}}%[[VAL_5]]] { |
| 118 | +// CHECK: %[[VAL_9:.*]] = affine.apply #map1(%[[VAL_8]]){{\[}}%[[VAL_3]], %[[VAL_5]], %[[VAL_3]]] |
| 119 | +// CHECK: affine.store %[[VAL_1]], %[[VAL_7]]{{\[}}%[[VAL_9]]] : memref<?xf32> |
| 120 | +// CHECK: } |
| 121 | +// CHECK: affine.for %[[VAL_10:.*]] = %[[VAL_3]] to #map0(){{\[}}%[[VAL_5]]] { |
| 122 | +// CHECK: %[[VAL_11:.*]] = affine.apply #map1(%[[VAL_10]]){{\[}}%[[VAL_3]], %[[VAL_5]], %[[VAL_3]]] |
| 123 | +// CHECK: affine.store %[[VAL_1]], %[[VAL_7]]{{\[}}%[[VAL_11]]] : memref<?xf32> |
| 124 | +// CHECK: } |
| 125 | +// CHECK: affine.for %[[VAL_12:.*]] = %[[VAL_3]] to #map0(){{\[}}%[[VAL_5]]] { |
| 126 | +// CHECK: %[[VAL_13:.*]] = subi %[[VAL_12]], %[[VAL_4]] : index |
| 127 | +// CHECK: affine.if #set(%[[VAL_12]]) { |
| 128 | +// CHECK: %[[VAL_14:.*]] = affine.apply #map1(%[[VAL_12]]){{\[}}%[[VAL_3]], %[[VAL_5]], %[[VAL_3]]] |
| 129 | +// CHECK: affine.store %[[VAL_1]], %[[VAL_7]]{{\[}}%[[VAL_14]]] : memref<?xf32> |
| 130 | +// CHECK: } |
| 131 | +// CHECK: } |
| 132 | +// CHECK: return |
| 133 | +// CHECK: } |
0 commit comments