| 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 = !fir.ref<!fir.array<?xf32>> |
| 6 |
#arr_len = affine_map<()[j1,k1] -> (k1 - j1 + 1)> |
| 7 |
|
| 8 |
func.func @loop_with_load_and_store(%a1: !arr_d1, %a2: !arr_d1, %a3: !arr_d1) { |
| 9 |
%c1 = arith.constant 1 : index |
| 10 |
%c0 = arith.constant 0 : index |
| 11 |
%len = arith.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 = arith.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 = arith.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:.*]] = arith.constant 1 : index |
| 51 |
// CHECK: %[[VAL_4:.*]] = arith.constant 100 : index |
| 52 |
// CHECK: %[[VAL_5:.*]] = fir.shape %[[VAL_4]] : (index) -> !fir.shape<1> |
| 53 |
// CHECK: %[[VAL_6:.*]] = affine.apply #{{.*}}(){{\[}}%[[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 #{{.*}}(){{\[}}%[[VAL_4]]] { |
| 59 |
// CHECK: %[[VAL_12:.*]] = affine.apply #{{.*}}(%[[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:.*]] = arith.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 #{{.*}}(){{\[}}%[[VAL_4]]] { |
| 67 |
// CHECK: %[[VAL_18:.*]] = affine.apply #{{.*}}(%[[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:.*]] = arith.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 = !fir.ref<!fir.array<?xf32>> |
| 79 |
#arr_len = affine_map<()[j1,k1] -> (k1 - j1 + 1)> |
| 80 |
|
| 81 |
func.func @loop_with_if(%a: !arr_d1, %v: f32) { |
| 82 |
%c0 = arith.constant 0 : index |
| 83 |
%c1 = arith.constant 1 : index |
| 84 |
%c2 = arith.constant 2 : index |
| 85 |
%len = arith.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 = arith.subi %i, %c2 : index |
| 92 |
%cond = arith.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:.*]] = arith.constant 0 : index |
| 112 |
// CHECK: %[[VAL_3:.*]] = arith.constant 1 : index |
| 113 |
// CHECK: %[[VAL_4:.*]] = arith.constant 2 : index |
| 114 |
// CHECK: %[[VAL_5:.*]] = arith.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 #{{.*}}(){{\[}}%[[VAL_5]]] { |
| 118 |
// CHECK: %[[VAL_9:.*]] = affine.apply #{{.*}}(%[[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 #{{.*}}(){{\[}}%[[VAL_5]]] { |
| 122 |
// CHECK: %[[VAL_11:.*]] = affine.apply #{{.*}}(%[[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 #{{.*}}(){{\[}}%[[VAL_5]]] { |
| 126 |
// CHECK: %[[VAL_13:.*]] = arith.subi %[[VAL_12]], %[[VAL_4]] : index |
| 127 |
// CHECK: affine.if #set(%[[VAL_12]]) { |
| 128 |
// CHECK: %[[VAL_14:.*]] = affine.apply #{{.*}}(%[[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: } |
| 134 |
|
| 135 |
func.func @loop_with_result(%arg0: !fir.ref<!fir.array<100xf32>>, %arg1: !fir.ref<!fir.array<100x100xf32>>, %arg2: !fir.ref<!fir.array<100xf32>>) -> f32 { |
| 136 |
%c1 = arith.constant 1 : index |
| 137 |
%cst = arith.constant 0.000000e+00 : f32 |
| 138 |
%c100 = arith.constant 100 : index |
| 139 |
%0 = fir.shape %c100 : (index) -> !fir.shape<1> |
| 140 |
%1 = fir.shape %c100, %c100 : (index, index) -> !fir.shape<2> |
| 141 |
%2 = fir.alloca i32 |
| 142 |
%3:2 = fir.do_loop %arg3 = %c1 to %c100 step %c1 iter_args(%arg4 = %cst) -> (index, f32) { |
| 143 |
%8 = fir.array_coor %arg0(%0) %arg3 : (!fir.ref<!fir.array<100xf32>>, !fir.shape<1>, index) -> !fir.ref<f32> |
| 144 |
%9 = fir.load %8 : !fir.ref<f32> |
| 145 |
%10 = arith.addf %arg4, %9 fastmath<contract> : f32 |
| 146 |
%11 = arith.addi %arg3, %c1 overflow<nsw> : index |
| 147 |
fir.result %11, %10 : index, f32 |
| 148 |
} |
| 149 |
%4:2 = fir.do_loop %arg3 = %c1 to %c100 step %c1 iter_args(%arg4 = %3#1) -> (index, f32) { |
| 150 |
%8 = fir.array_coor %arg1(%1) %c1, %arg3 : (!fir.ref<!fir.array<100x100xf32>>, !fir.shape<2>, index, index) -> !fir.ref<f32> |
| 151 |
%9 = fir.convert %8 : (!fir.ref<f32>) -> !fir.ref<!fir.array<100xf32>> |
| 152 |
%10 = fir.do_loop %arg5 = %c1 to %c100 step %c1 iter_args(%arg6 = %arg4) -> (f32) { |
| 153 |
%12 = fir.array_coor %9(%0) %arg5 : (!fir.ref<!fir.array<100xf32>>, !fir.shape<1>, index) -> !fir.ref<f32> |
| 154 |
%13 = fir.load %12 : !fir.ref<f32> |
| 155 |
%14 = arith.addf %arg6, %13 fastmath<contract> : f32 |
| 156 |
fir.result %14 : f32 |
| 157 |
} |
| 158 |
%11 = arith.addi %arg3, %c1 overflow<nsw> : index |
| 159 |
fir.result %11, %10 : index, f32 |
| 160 |
} |
| 161 |
%5:2 = fir.do_loop %arg3 = %c1 to %c100 step %c1 iter_args(%arg4 = %4#1, %arg5 = %cst) -> (f32, f32) { |
| 162 |
%8 = fir.array_coor %arg0(%0) %arg3 : (!fir.ref<!fir.array<100xf32>>, !fir.shape<1>, index) -> !fir.ref<f32> |
| 163 |
%9 = fir.load %8 : !fir.ref<f32> |
| 164 |
%10 = arith.addf %arg4, %9 fastmath<contract> : f32 |
| 165 |
%11 = fir.array_coor %arg2(%0) %arg3 : (!fir.ref<!fir.array<100xf32>>, !fir.shape<1>, index) -> !fir.ref<f32> |
| 166 |
%12 = fir.load %11 : !fir.ref<f32> |
| 167 |
%13 = arith.addf %arg5, %12 fastmath<contract> : f32 |
| 168 |
fir.result %10, %13 : f32, f32 |
| 169 |
} |
| 170 |
%6 = arith.addf %5#0, %5#1 fastmath<contract> : f32 |
| 171 |
%7 = fir.convert %4#0 : (index) -> i32 |
| 172 |
fir.store %7 to %2 : !fir.ref<i32> |
| 173 |
return %6 : f32 |
| 174 |
} |
| 175 |
|
| 176 |
// CHECK-LABEL: func.func @loop_with_result( |
| 177 |
// CHECK-SAME: %[[ARG0:.*]]: !fir.ref<!fir.array<100xf32>>, |
| 178 |
// CHECK-SAME: %[[ARG1:.*]]: !fir.ref<!fir.array<100x100xf32>>, |
| 179 |
// CHECK-SAME: %[[ARG2:.*]]: !fir.ref<!fir.array<100xf32>>) -> f32 { |
| 180 |
// CHECK: %[[VAL_0:.*]] = arith.constant 1 : index |
| 181 |
// CHECK: %[[VAL_1:.*]] = arith.constant 0.000000e+00 : f32 |
| 182 |
// CHECK: %[[VAL_2:.*]] = arith.constant 100 : index |
| 183 |
// CHECK: %[[VAL_3:.*]] = fir.shape %[[VAL_2]] : (index) -> !fir.shape<1> |
| 184 |
// CHECK: %[[VAL_4:.*]] = fir.shape %[[VAL_2]], %[[VAL_2]] : (index, index) -> !fir.shape<2> |
| 185 |
// CHECK: %[[VAL_5:.*]] = fir.alloca i32 |
| 186 |
// CHECK: %[[VAL_6:.*]] = fir.convert %[[ARG0]] : (!fir.ref<!fir.array<100xf32>>) -> memref<?xf32> |
| 187 |
// CHECK: %[[VAL_7:.*]] = affine.for %[[VAL_8:.*]] = %[[VAL_0]] to #{{.*}}(){{\[}}%[[VAL_2]]] iter_args(%[[VAL_9:.*]] = %[[VAL_1]]) -> (f32) { |
| 188 |
// CHECK: %[[VAL_10:.*]] = affine.apply #{{.*}}(%[[VAL_8]]){{\[}}%[[VAL_0]], %[[VAL_2]], %[[VAL_0]]] |
| 189 |
// CHECK: %[[VAL_11:.*]] = affine.load %[[VAL_6]]{{\[}}%[[VAL_10]]] : memref<?xf32> |
| 190 |
// CHECK: %[[VAL_12:.*]] = arith.addf %[[VAL_9]], %[[VAL_11]] fastmath<contract> : f32 |
| 191 |
// CHECK: affine.yield %[[VAL_12]] : f32 |
| 192 |
// CHECK: } |
| 193 |
// CHECK: %[[VAL_13:.*]]:2 = fir.do_loop %[[VAL_14:.*]] = %[[VAL_0]] to %[[VAL_2]] step %[[VAL_0]] iter_args(%[[VAL_15:.*]] = %[[VAL_7]]) -> (index, f32) { |
| 194 |
// CHECK: %[[VAL_16:.*]] = fir.array_coor %[[ARG1]](%[[VAL_4]]) %[[VAL_0]], %[[VAL_14]] : (!fir.ref<!fir.array<100x100xf32>>, !fir.shape<2>, index, index) -> !fir.ref<f32> |
| 195 |
// CHECK: %[[VAL_17:.*]] = fir.convert %[[VAL_16]] : (!fir.ref<f32>) -> !fir.ref<!fir.array<100xf32>> |
| 196 |
// CHECK: %[[VAL_18:.*]] = fir.convert %[[VAL_17]] : (!fir.ref<!fir.array<100xf32>>) -> memref<?xf32> |
| 197 |
// CHECK: %[[VAL_19:.*]] = affine.for %[[VAL_20:.*]] = %[[VAL_0]] to #{{.*}}(){{\[}}%[[VAL_2]]] iter_args(%[[VAL_21:.*]] = %[[VAL_15]]) -> (f32) { |
| 198 |
// CHECK: %[[VAL_22:.*]] = affine.apply #{{.*}}(%[[VAL_20]]){{\[}}%[[VAL_0]], %[[VAL_2]], %[[VAL_0]]] |
| 199 |
// CHECK: %[[VAL_23:.*]] = affine.load %[[VAL_18]]{{\[}}%[[VAL_22]]] : memref<?xf32> |
| 200 |
// CHECK: %[[VAL_24:.*]] = arith.addf %[[VAL_21]], %[[VAL_23]] fastmath<contract> : f32 |
| 201 |
// CHECK: affine.yield %[[VAL_24]] : f32 |
| 202 |
// CHECK: } |
| 203 |
// CHECK: %[[VAL_25:.*]] = arith.addi %[[VAL_14]], %[[VAL_0]] overflow<nsw> : index |
| 204 |
// CHECK: fir.result %[[VAL_25]], %[[VAL_19]] : index, f32 |
| 205 |
// CHECK: } |
| 206 |
// CHECK: %[[VAL_26:.*]] = fir.convert %[[ARG2]] : (!fir.ref<!fir.array<100xf32>>) -> memref<?xf32> |
| 207 |
// CHECK: %[[VAL_27:.*]]:2 = affine.for %[[VAL_28:.*]] = %[[VAL_0]] to #{{.*}}(){{\[}}%[[VAL_2]]] iter_args(%[[VAL_29:.*]] = %[[VAL_30:.*]]#1, %[[VAL_31:.*]] = %[[VAL_1]]) -> (f32, f32) { |
| 208 |
// CHECK: %[[VAL_32:.*]] = affine.apply #{{.*}}(%[[VAL_28]]){{\[}}%[[VAL_0]], %[[VAL_2]], %[[VAL_0]]] |
| 209 |
// CHECK: %[[VAL_33:.*]] = affine.load %[[VAL_6]]{{\[}}%[[VAL_32]]] : memref<?xf32> |
| 210 |
// CHECK: %[[VAL_34:.*]] = arith.addf %[[VAL_29]], %[[VAL_33]] fastmath<contract> : f32 |
| 211 |
// CHECK: %[[VAL_35:.*]] = affine.load %[[VAL_26]]{{\[}}%[[VAL_32]]] : memref<?xf32> |
| 212 |
// CHECK: %[[VAL_36:.*]] = arith.addf %[[VAL_31]], %[[VAL_35]] fastmath<contract> : f32 |
| 213 |
// CHECK: affine.yield %[[VAL_34]], %[[VAL_36]] : f32, f32 |
| 214 |
// CHECK: } |
| 215 |
// CHECK: %[[VAL_37:.*]] = arith.addf %[[VAL_38:.*]]#0, %[[VAL_38]]#1 fastmath<contract> : f32 |
| 216 |
// CHECK: %[[VAL_39:.*]] = fir.convert %[[VAL_40:.*]]#0 : (index) -> i32 |
| 217 |
// CHECK: fir.store %[[VAL_39]] to %[[VAL_5]] : !fir.ref<i32> |
| 218 |
// CHECK: return %[[VAL_37]] : f32 |
| 219 |
// CHECK: } |