| 1 |
// RUN: fir-opt --stack-arrays %s | FileCheck %s |
| 2 |
|
| 3 |
// Simplest transformation |
| 4 |
func.func @simple() { |
| 5 |
%0 = fir.allocmem !fir.array<42xi32> |
| 6 |
fir.freemem %0 : !fir.heap<!fir.array<42xi32>> |
| 7 |
return |
| 8 |
} |
| 9 |
// CHECK: func.func @simple() { |
| 10 |
// CHECK-NEXT: fir.alloca !fir.array<42xi32> |
| 11 |
// CHECK-NEXT: return |
| 12 |
// CHECK-NEXT: } |
| 13 |
|
| 14 |
// Check fir.must_be_heap allocations are not moved |
| 15 |
func.func @must_be_heap() { |
| 16 |
%0 = fir.allocmem !fir.array<42xi32> {fir.must_be_heap = true} |
| 17 |
fir.freemem %0 : !fir.heap<!fir.array<42xi32>> |
| 18 |
return |
| 19 |
} |
| 20 |
// CHECK: func.func @must_be_heap() { |
| 21 |
// CHECK-NEXT: %[[ALLOC:.*]] = fir.allocmem !fir.array<42xi32> {fir.must_be_heap = true} |
| 22 |
// CHECK-NEXT: fir.freemem %[[ALLOC]] : !fir.heap<!fir.array<42xi32>> |
| 23 |
// CHECK-NEXT: return |
| 24 |
// CHECK-NEXT: } |
| 25 |
|
| 26 |
// Check the data-flow-analysis can detect cases where we aren't sure if memory |
| 27 |
// is freed by the end of the function |
| 28 |
func.func @dfa1(%arg0: !fir.ref<!fir.logical<4>> {fir.bindc_name = "cond"}) { |
| 29 |
%7 = arith.constant 42 : index |
| 30 |
%8 = fir.allocmem !fir.array<?xi32>, %7 {uniq_name = "_QFdfa1Earr.alloc"} |
| 31 |
%9 = fir.load %arg0 : !fir.ref<!fir.logical<4>> |
| 32 |
%10 = fir.convert %9 : (!fir.logical<4>) -> i1 |
| 33 |
fir.if %10 { |
| 34 |
fir.freemem %8 : !fir.heap<!fir.array<?xi32>> |
| 35 |
} else { |
| 36 |
} |
| 37 |
return |
| 38 |
} |
| 39 |
// CHECK: func.func @dfa1(%arg0: !fir.ref<!fir.logical<4>> {fir.bindc_name = "cond"}) { |
| 40 |
// CHECK-NEXT: %[[C42:.*]] = arith.constant 42 : index |
| 41 |
// CHECK-NEXT: %[[MEM:.*]] = fir.allocmem !fir.array<?xi32>, %[[C42]] {uniq_name = "_QFdfa1Earr.alloc"} |
| 42 |
// CHECK-NEXT: %[[LOGICAL:.*]] = fir.load %arg0 : !fir.ref<!fir.logical<4>> |
| 43 |
// CHECK-NEXT: %[[BOOL:.*]] = fir.convert %[[LOGICAL]] : (!fir.logical<4>) -> i1 |
| 44 |
// CHECK-NEXT: fir.if %[[BOOL]] { |
| 45 |
// CHECK-NEXT: fir.freemem %[[MEM]] : !fir.heap<!fir.array<?xi32>> |
| 46 |
// CHECK-NEXT: } else { |
| 47 |
// CHECK-NEXT: } |
| 48 |
// CHECK-NEXT: return |
| 49 |
// CHECK-NEXT: } |
| 50 |
|
| 51 |
// Check scf.if |
| 52 |
func.func @dfa2(%arg0: i1) { |
| 53 |
%a = fir.allocmem !fir.array<1xi8> |
| 54 |
scf.if %arg0 { |
| 55 |
fir.freemem %a : !fir.heap<!fir.array<1xi8>> |
| 56 |
} else { |
| 57 |
} |
| 58 |
return |
| 59 |
} |
| 60 |
// CHECK: func.func @dfa2(%arg0: i1) { |
| 61 |
// CHECK-NEXT: %[[MEM:.*]] = fir.allocmem !fir.array<1xi8> |
| 62 |
// CHECK-NEXT: scf.if %arg0 { |
| 63 |
// CHECK-NEXT: fir.freemem %[[MEM]] : !fir.heap<!fir.array<1xi8>> |
| 64 |
// CHECK-NEXT: } else { |
| 65 |
// CHECK-NEXT: } |
| 66 |
// CHECK-NEXT: return |
| 67 |
// CHECK-NEXT: } |
| 68 |
|
| 69 |
// Check freemem in both regions |
| 70 |
func.func @dfa3(%arg0: i1) { |
| 71 |
%a = fir.allocmem !fir.array<1xi8> |
| 72 |
fir.if %arg0 { |
| 73 |
fir.freemem %a : !fir.heap<!fir.array<1xi8>> |
| 74 |
} else { |
| 75 |
fir.freemem %a : !fir.heap<!fir.array<1xi8>> |
| 76 |
} |
| 77 |
return |
| 78 |
} |
| 79 |
// CHECK: func.func @dfa3(%arg0: i1) { |
| 80 |
// CHECK-NEXT: %[[MEM:.*]] = fir.alloca !fir.array<1xi8> |
| 81 |
// CHECK-NEXT: fir.if %arg0 { |
| 82 |
// CHECK-NEXT: } else { |
| 83 |
// CHECK-NEXT: } |
| 84 |
// CHECK-NEXT: return |
| 85 |
// CHECK-NEXT: } |
| 86 |
|
| 87 |
func.func private @dfa3a_foo(!fir.ref<!fir.array<1xi8>>) -> () |
| 88 |
func.func private @dfa3a_bar(!fir.ref<!fir.array<1xi8>>) -> () |
| 89 |
|
| 90 |
// Check freemem in both regions, with other uses |
| 91 |
func.func @dfa3a(%arg0: i1) { |
| 92 |
%a = fir.allocmem !fir.array<1xi8> |
| 93 |
fir.if %arg0 { |
| 94 |
%ref = fir.convert %a : (!fir.heap<!fir.array<1xi8>>) -> !fir.ref<!fir.array<1xi8>> |
| 95 |
func.call @dfa3a_foo(%ref) : (!fir.ref<!fir.array<1xi8>>) -> () |
| 96 |
fir.freemem %a : !fir.heap<!fir.array<1xi8>> |
| 97 |
} else { |
| 98 |
%ref = fir.convert %a : (!fir.heap<!fir.array<1xi8>>) -> !fir.ref<!fir.array<1xi8>> |
| 99 |
func.call @dfa3a_bar(%ref) : (!fir.ref<!fir.array<1xi8>>) -> () |
| 100 |
fir.freemem %a : !fir.heap<!fir.array<1xi8>> |
| 101 |
} |
| 102 |
return |
| 103 |
} |
| 104 |
// CHECK: func.func @dfa3a(%arg0: i1) { |
| 105 |
// CHECK-NEXT: %[[MEM:.*]] = fir.alloca !fir.array<1xi8> |
| 106 |
// CHECK-NEXT: %[[HEAP:.*]] = fir.convert %[[MEM]] : (!fir.ref<!fir.array<1xi8>>) -> !fir.heap<!fir.array<1xi8>> |
| 107 |
// CHECK-NEXT: fir.if %arg0 { |
| 108 |
// CHECK-NEXT: %[[REF:.*]] = fir.convert %[[HEAP]] : (!fir.heap<!fir.array<1xi8>>) -> !fir.ref<!fir.array<1xi8>> |
| 109 |
// CHECK-NEXT: func.call @dfa3a_foo(%[[REF]]) |
| 110 |
// CHECK-NEXT: } else { |
| 111 |
// CHECK-NEXT: %[[REF:.*]] = fir.convert %[[HEAP]] : (!fir.heap<!fir.array<1xi8>>) -> !fir.ref<!fir.array<1xi8>> |
| 112 |
// CHECK-NEXT: func.call @dfa3a_bar(%[[REF]]) |
| 113 |
// CHECK-NEXT: } |
| 114 |
// CHECK-NEXT: return |
| 115 |
// CHECK-NEXT: } |
| 116 |
|
| 117 |
// check the alloca is placed after all operands become available |
| 118 |
func.func @placement1() { |
| 119 |
// do some stuff with other ssa values |
| 120 |
%1 = arith.constant 1 : index |
| 121 |
%2 = arith.constant 2 : index |
| 122 |
%3 = arith.addi %1, %2 : index |
| 123 |
// operand is now available |
| 124 |
%4 = fir.allocmem !fir.array<?xi32>, %3 |
| 125 |
// ... |
| 126 |
fir.freemem %4 : !fir.heap<!fir.array<?xi32>> |
| 127 |
return |
| 128 |
} |
| 129 |
// CHECK: func.func @placement1() { |
| 130 |
// CHECK-NEXT: %[[ARG:.*]] = arith.constant 3 : index |
| 131 |
// CHECK-NEXT: %[[MEM:.*]] = fir.alloca !fir.array<?xi32>, %[[ARG]] |
| 132 |
// CHECK-NEXT: return |
| 133 |
// CHECK-NEXT: } |
| 134 |
|
| 135 |
// check that if there are no operands, then the alloca is placed early |
| 136 |
func.func @placement2() { |
| 137 |
// do some stuff with other ssa values |
| 138 |
%1 = arith.constant 1 : index |
| 139 |
%2 = arith.constant 2 : index |
| 140 |
%3 = arith.addi %1, %2 : index |
| 141 |
%4 = fir.allocmem !fir.array<42xi32> |
| 142 |
// ... |
| 143 |
fir.freemem %4 : !fir.heap<!fir.array<42xi32>> |
| 144 |
return |
| 145 |
} |
| 146 |
// CHECK: func.func @placement2() { |
| 147 |
// CHECK-NEXT: %[[MEM:.*]] = fir.alloca !fir.array<42xi32> |
| 148 |
// CHECK-NEXT: %[[ONE:.*]] = arith.constant 1 : index |
| 149 |
// CHECK-NEXT: %[[TWO:.*]] = arith.constant 2 : index |
| 150 |
// CHECK-NEXT: %[[SUM:.*]] = arith.addi %[[ONE]], %[[TWO]] : index |
| 151 |
// CHECK-NEXT: return |
| 152 |
// CHECK-NEXT: } |
| 153 |
|
| 154 |
// check that stack allocations which must be placed in loops use stacksave |
| 155 |
func.func @placement3() { |
| 156 |
%c1 = arith.constant 1 : index |
| 157 |
%c1_i32 = fir.convert %c1 : (index) -> i32 |
| 158 |
%c2 = arith.constant 2 : index |
| 159 |
%c10 = arith.constant 10 : index |
| 160 |
%0:2 = fir.do_loop %arg0 = %c1 to %c10 step %c1 iter_args(%arg1 = %c1_i32) -> (index, i32) { |
| 161 |
%3 = arith.addi %c1, %c2 : index |
| 162 |
// operand is now available |
| 163 |
%4 = fir.allocmem !fir.array<?xi32>, %3 |
| 164 |
// ... |
| 165 |
fir.freemem %4 : !fir.heap<!fir.array<?xi32>> |
| 166 |
fir.result %3, %c1_i32 : index, i32 |
| 167 |
} |
| 168 |
return |
| 169 |
} |
| 170 |
// CHECK: func.func @placement3() { |
| 171 |
// CHECK-NEXT: %[[C1:.*]] = arith.constant 1 : index |
| 172 |
// CHECK-NEXT: %[[C1_I32:.*]] = fir.convert %[[C1]] : (index) -> i32 |
| 173 |
// CHECK-NEXT: %[[C2:.*]] = arith.constant 2 : index |
| 174 |
// CHECK-NEXT: %[[C10:.*]] = arith.constant 10 : index |
| 175 |
// CHECK-NEXT: fir.do_loop |
| 176 |
// CHECK-NEXT: %[[SUM:.*]] = arith.addi %[[C1]], %[[C2]] : index |
| 177 |
// CHECK-NEXT: %[[SP:.*]] = llvm.intr.stacksave : !llvm.ptr |
| 178 |
// CHECK-NEXT: %[[MEM:.*]] = fir.alloca !fir.array<?xi32>, %[[SUM]] |
| 179 |
// CHECK-NEXT: llvm.intr.stackrestore %[[SP]] : !llvm.ptr |
| 180 |
// CHECK-NEXT: fir.result |
| 181 |
// CHECK-NEXT: } |
| 182 |
// CHECK-NEXT: return |
| 183 |
// CHECK-NEXT: } |
| 184 |
|
| 185 |
// check that stack save/restore are used in CFG loops |
| 186 |
func.func @placement4(%arg0 : i1) { |
| 187 |
%c1 = arith.constant 1 : index |
| 188 |
%c1_i32 = fir.convert %c1 : (index) -> i32 |
| 189 |
%c2 = arith.constant 2 : index |
| 190 |
%c10 = arith.constant 10 : index |
| 191 |
cf.br ^bb1 |
| 192 |
^bb1: |
| 193 |
%3 = arith.addi %c1, %c2 : index |
| 194 |
// operand is now available |
| 195 |
%4 = fir.allocmem !fir.array<?xi32>, %3 |
| 196 |
// ... |
| 197 |
fir.freemem %4 : !fir.heap<!fir.array<?xi32>> |
| 198 |
cf.cond_br %arg0, ^bb1, ^bb2 |
| 199 |
^bb2: |
| 200 |
return |
| 201 |
} |
| 202 |
// CHECK: func.func @placement4(%arg0: i1) { |
| 203 |
// CHECK-NEXT: %[[C1:.*]] = arith.constant 1 : index |
| 204 |
// CHECK-NEXT: %[[C1_I32:.*]] = fir.convert %[[C1]] : (index) -> i32 |
| 205 |
// CHECK-NEXT: %[[C10:.*]] = arith.constant 10 : index |
| 206 |
// CHECK-NEXT: cf.br ^bb1 |
| 207 |
// CHECK-NEXT: ^bb1: |
| 208 |
// CHECK-NEXT: %[[C3:.*]] = arith.constant 3 : index |
| 209 |
// CHECK-NEXT: %[[SP:.*]] = llvm.intr.stacksave : !llvm.ptr |
| 210 |
// CHECK-NEXT: %[[MEM:.*]] = fir.alloca !fir.array<?xi32>, %[[C3]] |
| 211 |
// CHECK-NEXT: llvm.intr.stackrestore %[[SP]] : !llvm.ptr |
| 212 |
// CHECK-NEXT: cf.cond_br %arg0, ^bb1, ^bb2 |
| 213 |
// CHECK-NEXT: ^bb2: |
| 214 |
// CHECK-NEXT: return |
| 215 |
// CHECK-NEXT: } |
| 216 |
|
| 217 |
// check that stacksave is not used when there is an intervening alloca |
| 218 |
func.func @placement5() { |
| 219 |
%c1 = arith.constant 1 : index |
| 220 |
%c1_i32 = fir.convert %c1 : (index) -> i32 |
| 221 |
%c2 = arith.constant 2 : index |
| 222 |
%c10 = arith.constant 10 : index |
| 223 |
%0:2 = fir.do_loop %arg0 = %c1 to %c10 step %c1 iter_args(%arg1 = %c1_i32) -> (index, i32) { |
| 224 |
%3 = arith.addi %c1, %c2 : index |
| 225 |
// operand is now available |
| 226 |
%4 = fir.allocmem !fir.array<?xi32>, %3 |
| 227 |
%5 = fir.alloca i32 |
| 228 |
fir.freemem %4 : !fir.heap<!fir.array<?xi32>> |
| 229 |
fir.result %3, %c1_i32 : index, i32 |
| 230 |
} |
| 231 |
return |
| 232 |
} |
| 233 |
// CHECK: func.func @placement5() { |
| 234 |
// CHECK-NEXT: %[[C1:.*]] = arith.constant 1 : index |
| 235 |
// CHECK-NEXT: %[[C1_I32:.*]] = fir.convert %[[C1]] : (index) -> i32 |
| 236 |
// CHECK-NEXT: %[[C2:.*]] = arith.constant 2 : index |
| 237 |
// CHECK-NEXT: %[[C10:.*]] = arith.constant 10 : index |
| 238 |
// CHECK-NEXT: fir.do_loop |
| 239 |
// CHECK-NEXT: %[[SUM:.*]] = arith.addi %[[C1]], %[[C2]] : index |
| 240 |
// CHECK-NEXT: %[[MEM:.*]] = fir.allocmem !fir.array<?xi32>, %[[SUM]] |
| 241 |
// CHECK-NEXT: %[[IDX:.*]] = fir.alloca i32 |
| 242 |
// CHECK-NEXT: fir.freemem %[[MEM]] : !fir.heap<!fir.array<?xi32>> |
| 243 |
// CHECK-NEXT: fir.result |
| 244 |
// CHECK-NEXT: } |
| 245 |
// CHECK-NEXT: return |
| 246 |
// CHECK-NEXT: } |
| 247 |
|
| 248 |
// check that stack save/restore are not used when the memalloc and freemem are |
| 249 |
// in different blocks |
| 250 |
func.func @placement6(%arg0: i1) { |
| 251 |
%c1 = arith.constant 1 : index |
| 252 |
%c1_i32 = fir.convert %c1 : (index) -> i32 |
| 253 |
%c2 = arith.constant 2 : index |
| 254 |
%c10 = arith.constant 10 : index |
| 255 |
cf.br ^bb1 |
| 256 |
^bb1: |
| 257 |
%3 = arith.addi %c1, %c2 : index |
| 258 |
// operand is now available |
| 259 |
%4 = fir.allocmem !fir.array<?xi32>, %3 |
| 260 |
// ... |
| 261 |
cf.cond_br %arg0, ^bb2, ^bb3 |
| 262 |
^bb2: |
| 263 |
// ... |
| 264 |
fir.freemem %4 : !fir.heap<!fir.array<?xi32>> |
| 265 |
cf.br ^bb1 |
| 266 |
^bb3: |
| 267 |
// ... |
| 268 |
fir.freemem %4 : !fir.heap<!fir.array<?xi32>> |
| 269 |
cf.br ^bb1 |
| 270 |
} |
| 271 |
// CHECK: func.func @placement6(%arg0: i1) { |
| 272 |
// CHECK-NEXT: %[[c1:.*]] = arith.constant 1 : index |
| 273 |
// CHECK-NEXT: %[[c1_i32:.*]] = fir.convert %[[c1]] : (index) -> i32 |
| 274 |
// CHECK-NEXT: %[[c2:.*]] = arith.constant 2 : index |
| 275 |
// CHECK-NEXT: %[[c10:.*]] = arith.constant 10 : index |
| 276 |
// CHECK-NEXT: cf.br ^bb1 |
| 277 |
// CHECK-NEXT: ^bb1: |
| 278 |
// CHECK-NEXT: %[[ADD:.*]] = arith.addi %[[c1]], %[[c2]] : index |
| 279 |
// CHECK-NEXT: %[[MEM:.*]] = fir.allocmem !fir.array<?xi32>, %[[ADD]] |
| 280 |
// CHECK-NEXT: cf.cond_br %arg0, ^bb2, ^bb3 |
| 281 |
// CHECK-NEXT: ^bb2: |
| 282 |
// CHECK-NEXT: fir.freemem %[[MEM]] : !fir.heap<!fir.array<?xi32>> |
| 283 |
// CHECK-NEXT: cf.br ^bb1 |
| 284 |
// CHECK-NEXT: ^bb3: |
| 285 |
// CHECK-NEXT: fir.freemem %[[MEM]] : !fir.heap<!fir.array<?xi32>> |
| 286 |
// CHECK-NEXT: cf.br ^bb1 |
| 287 |
// CHECK-NEXT: } |
| 288 |
|
| 289 |
// Check multiple returns, where the memory is always freed |
| 290 |
func.func @returns(%arg0: i1) { |
| 291 |
%0 = fir.allocmem !fir.array<42xi32> |
| 292 |
cf.cond_br %arg0, ^bb1, ^bb2 |
| 293 |
^bb1: |
| 294 |
fir.freemem %0 : !fir.heap<!fir.array<42xi32>> |
| 295 |
return |
| 296 |
^bb2: |
| 297 |
fir.freemem %0 : !fir.heap<!fir.array<42xi32>> |
| 298 |
return |
| 299 |
} |
| 300 |
// CHECK: func.func @returns(%[[COND:.*]]: i1) { |
| 301 |
// CHECK-NEXT: %[[ALLOC:.*]] = fir.alloca !fir.array<42xi32> |
| 302 |
// CHECK-NEXT: cf.cond_br %[[COND]], ^bb1, ^bb2 |
| 303 |
// CHECK-NEXT: ^bb1: |
| 304 |
// CHECK-NEXT: return |
| 305 |
// CHECK-NEXT: ^bb2: |
| 306 |
// CHECK-NEXT: return |
| 307 |
// CHECK-NEXT: } |
| 308 |
|
| 309 |
// Check multiple returns, where the memory is not freed on one branch |
| 310 |
func.func @returns2(%arg0: i1) { |
| 311 |
%0 = fir.allocmem !fir.array<42xi32> |
| 312 |
cf.cond_br %arg0, ^bb1, ^bb2 |
| 313 |
^bb1: |
| 314 |
fir.freemem %0 : !fir.heap<!fir.array<42xi32>> |
| 315 |
return |
| 316 |
^bb2: |
| 317 |
return |
| 318 |
} |
| 319 |
// CHECK: func.func @returns2(%[[COND:.*]]: i1) { |
| 320 |
// CHECK-NEXT: %[[ALLOC:.*]] = fir.allocmem !fir.array<42xi32> |
| 321 |
// CHECK-NEXT: cf.cond_br %[[COND]], ^bb1, ^bb2 |
| 322 |
// CHECK-NEXT: ^bb1: |
| 323 |
// CHECK-NEXT: fir.freemem %[[ALLOC]] : !fir.heap<!fir.array<42xi32>> |
| 324 |
// CHECK-NEXT: return |
| 325 |
// CHECK-NEXT: ^bb2: |
| 326 |
// CHECK-NEXT: return |
| 327 |
// CHECK-NEXT: } |
| 328 |
|
| 329 |
// Check allocations are not moved outside of an omp region |
| 330 |
func.func @omp_placement1() { |
| 331 |
omp.sections { |
| 332 |
omp.section { |
| 333 |
%mem = fir.allocmem !fir.array<42xi32> |
| 334 |
fir.freemem %mem : !fir.heap<!fir.array<42xi32>> |
| 335 |
omp.terminator |
| 336 |
} |
| 337 |
omp.terminator |
| 338 |
} |
| 339 |
return |
| 340 |
} |
| 341 |
// CHECK: func.func @omp_placement1() { |
| 342 |
// CHECK-NEXT: %[[MEM:.*]] = fir.alloca !fir.array<42xi32> |
| 343 |
// CHECK-NEXT: %[[MEM_CONV:.*]] = fir.convert %[[MEM]] : (!fir.ref<!fir.array<42xi32>>) -> !fir.heap<!fir.array<42xi32>> |
| 344 |
// CHECK-NEXT: omp.sections { |
| 345 |
// CHECK-NEXT: omp.section { |
| 346 |
// CHECK-NEXT: omp.terminator |
| 347 |
// CHECK-NEXT: } |
| 348 |
// CHECK-NEXT: omp.terminator |
| 349 |
// CHECK-NEXT: } |
| 350 |
// CHECK-NEXT: return |
| 351 |
// CHECK-NEXT: } |
| 352 |
|
| 353 |
// function terminated by stop statement |
| 354 |
func.func @stop_terminator() { |
| 355 |
%0 = fir.allocmem !fir.array<42xi32> |
| 356 |
fir.freemem %0 : !fir.heap<!fir.array<42xi32>> |
| 357 |
%c0_i32 = arith.constant 0 : i32 |
| 358 |
%false = arith.constant false |
| 359 |
fir.call @_FortranAStopStatement(%c0_i32, %false, %false) : (i32, i1, i1) -> () |
| 360 |
fir.unreachable |
| 361 |
} |
| 362 |
// CHECK: func.func @stop_terminator() { |
| 363 |
// CHECK-NEXT: fir.alloca !fir.array<42xi32> |
| 364 |
// CHECK-NEXT: %[[ZERO:.*]] = arith.constant 0 : i32 |
| 365 |
// CHECK-NEXT: %[[FALSE:.*]] = arith.constant false |
| 366 |
// CHECK-NEXT: fir.call @_FortranAStopStatement(%[[ZERO]], %[[FALSE]], %[[FALSE]]) : (i32, i1, i1) -> () |
| 367 |
// CHECK-NEXT: fir.unreachable |
| 368 |
// CHECK-NEXT: } |
| 369 |
|
| 370 |
|
| 371 |
// check that stack allocations that use fir.declare which must be placed in loops |
| 372 |
// use stacksave |
| 373 |
func.func @placement_loop_declare() { |
| 374 |
%c1 = arith.constant 1 : index |
| 375 |
%c1_i32 = fir.convert %c1 : (index) -> i32 |
| 376 |
%c2 = arith.constant 2 : index |
| 377 |
%c10 = arith.constant 10 : index |
| 378 |
%0:2 = fir.do_loop %arg0 = %c1 to %c10 step %c1 iter_args(%arg1 = %c1_i32) -> (index, i32) { |
| 379 |
%3 = arith.addi %c1, %c2 : index |
| 380 |
// operand is now available |
| 381 |
%4 = fir.allocmem !fir.array<?xi32>, %3 |
| 382 |
%shape = fir.shape %3 : (index) -> !fir.shape<1> |
| 383 |
%5 = fir.declare %4(%shape) {uniq_name = "temp"} : (!fir.heap<!fir.array<?xi32>>, !fir.shape<1>) -> !fir.heap<!fir.array<?xi32>> |
| 384 |
// ... |
| 385 |
fir.freemem %5 : !fir.heap<!fir.array<?xi32>> |
| 386 |
fir.result %3, %c1_i32 : index, i32 |
| 387 |
} |
| 388 |
return |
| 389 |
} |
| 390 |
// CHECK: func.func @placement_loop_declare() { |
| 391 |
// CHECK-NEXT: %[[C1:.*]] = arith.constant 1 : index |
| 392 |
// CHECK-NEXT: %[[C1_I32:.*]] = fir.convert %[[C1]] : (index) -> i32 |
| 393 |
// CHECK-NEXT: %[[C2:.*]] = arith.constant 2 : index |
| 394 |
// CHECK-NEXT: %[[C10:.*]] = arith.constant 10 : index |
| 395 |
// CHECK-NEXT: fir.do_loop |
| 396 |
// CHECK-NEXT: %[[SUM:.*]] = arith.addi %[[C1]], %[[C2]] : index |
| 397 |
// CHECK-NEXT: %[[SP:.*]] = llvm.intr.stacksave : !llvm.ptr |
| 398 |
// CHECK-NEXT: %[[MEM:.*]] = fir.alloca !fir.array<?xi32>, %[[SUM]] |
| 399 |
// CHECK: llvm.intr.stackrestore %[[SP]] : !llvm.ptr |
| 400 |
// CHECK-NEXT: fir.result |
| 401 |
// CHECK-NEXT: } |
| 402 |
// CHECK-NEXT: return |
| 403 |
// CHECK-NEXT: } |
| 404 |
|
| 405 |
// Can we look through fir.convert and fir.declare? |
| 406 |
func.func @lookthrough() { |
| 407 |
%0 = fir.allocmem !fir.array<42xi32> |
| 408 |
%c42 = arith.constant 42 : index |
| 409 |
%shape = fir.shape %c42 : (index) -> !fir.shape<1> |
| 410 |
%1 = fir.declare %0(%shape) {uniq_name = "name"} : (!fir.heap<!fir.array<42xi32>>, !fir.shape<1>) -> !fir.heap<!fir.array<42xi32>> |
| 411 |
%2 = fir.convert %1 : (!fir.heap<!fir.array<42xi32>>) -> !fir.ref<!fir.array<42xi32>> |
| 412 |
// use the ref so the converts aren't folded |
| 413 |
%3 = fir.load %2 : !fir.ref<!fir.array<42xi32>> |
| 414 |
%4 = fir.convert %2 : (!fir.ref<!fir.array<42xi32>>) -> !fir.heap<!fir.array<42xi32>> |
| 415 |
fir.freemem %4 : !fir.heap<!fir.array<42xi32>> |
| 416 |
return |
| 417 |
} |
| 418 |
// CHECK: func.func @lookthrough() { |
| 419 |
// CHECK: fir.alloca !fir.array<42xi32> |
| 420 |
// CHECK-NOT: fir.freemem |
| 421 |
|
| 422 |
// StackArrays is better to find fir.freemem ops corresponding to fir.allocmem |
| 423 |
// using the same look through mechanism as during the allocation analysis, |
| 424 |
// looking through fir.convert and fir.declare. |
| 425 |
func.func @finding_freemem_in_block() { |
| 426 |
%c0 = arith.constant 0 : index |
| 427 |
%c10_i32 = arith.constant 10 : i32 |
| 428 |
%c1_i32 = arith.constant 1 : i32 |
| 429 |
%0 = fir.alloca i32 {bindc_name = "k", uniq_name = "k"} |
| 430 |
%1 = fir.declare %0 {uniq_name = "k"} : (!fir.ref<i32>) -> !fir.ref<i32> |
| 431 |
fir.store %c1_i32 to %1 : !fir.ref<i32> |
| 432 |
cf.br ^bb1 |
| 433 |
^bb1: // 2 preds: ^bb0, ^bb2 |
| 434 |
%2 = fir.load %1 : !fir.ref<i32> |
| 435 |
%3 = arith.cmpi sle, %2, %c10_i32 : i32 |
| 436 |
cf.cond_br %3, ^bb2, ^bb3 |
| 437 |
^bb2: // pred: ^bb1 |
| 438 |
%4 = fir.declare %1 {fortran_attrs = #fir.var_attrs<intent_in>, uniq_name = "x"} : (!fir.ref<i32>) -> !fir.ref<i32> |
| 439 |
%5 = fir.load %4 : !fir.ref<i32> |
| 440 |
%6 = fir.convert %5 : (i32) -> index |
| 441 |
%7 = arith.cmpi sgt, %6, %c0 : index |
| 442 |
%8 = arith.select %7, %6, %c0 : index |
| 443 |
%9 = fir.shape %8 : (index) -> !fir.shape<1> |
| 444 |
%10 = fir.allocmem !fir.array<?xi32>, %8 {bindc_name = ".tmp.expr_result", uniq_name = ""} |
| 445 |
%11 = fir.convert %10 : (!fir.heap<!fir.array<?xi32>>) -> !fir.ref<!fir.array<?xi32>> |
| 446 |
%12 = fir.declare %11(%9) {uniq_name = ".tmp.expr_result"} : (!fir.ref<!fir.array<?xi32>>, !fir.shape<1>) -> !fir.ref<!fir.array<?xi32>> |
| 447 |
%13 = fir.embox %12(%9) : (!fir.ref<!fir.array<?xi32>>, !fir.shape<1>) -> !fir.box<!fir.array<?xi32>> |
| 448 |
%14 = fir.call @_QPfunc(%1) fastmath<fast> : (!fir.ref<i32>) -> !fir.array<?xi32> |
| 449 |
fir.save_result %14 to %12(%9) : !fir.array<?xi32>, !fir.ref<!fir.array<?xi32>>, !fir.shape<1> |
| 450 |
fir.call @_QPsub(%13) fastmath<fast> : (!fir.box<!fir.array<?xi32>>) -> () |
| 451 |
%15 = fir.convert %12 : (!fir.ref<!fir.array<?xi32>>) -> !fir.heap<!fir.array<?xi32>> |
| 452 |
fir.freemem %15 : !fir.heap<!fir.array<?xi32>> |
| 453 |
%16 = fir.load %1 : !fir.ref<i32> |
| 454 |
%17 = arith.addi %16, %c1_i32 : i32 |
| 455 |
fir.store %17 to %1 : !fir.ref<i32> |
| 456 |
cf.br ^bb1 |
| 457 |
^bb3: // pred: ^bb1 |
| 458 |
return |
| 459 |
} |
| 460 |
// CHECK: func.func @finding_freemem_in_block() { |
| 461 |
// CHECK: fir.alloca !fir.array<?xi32> |
| 462 |
// CHECK-NOT: fir.freemem |