brintos

brintos / llvm-project-archived public Read only

0
0
Text · 22.6 KiB · f66cf7a Raw
522 lines · plain
1// RUN: mlir-opt %s -one-shot-bufferize="bufferize-function-boundaries" -drop-equivalent-buffer-results -split-input-file | FileCheck %s2 3// Run fuzzer with different seeds.4// RUN: mlir-opt %s -one-shot-bufferize="test-analysis-only analysis-heuristic=fuzzer analysis-fuzzer-seed=23 bufferize-function-boundaries" -split-input-file -o /dev/null5// RUN: mlir-opt %s -one-shot-bufferize="test-analysis-only analysis-heuristic=fuzzer analysis-fuzzer-seed=59 bufferize-function-boundaries" -split-input-file -o /dev/null6// RUN: mlir-opt %s -one-shot-bufferize="test-analysis-only analysis-heuristic=fuzzer analysis-fuzzer-seed=91 bufferize-function-boundaries" -split-input-file -o /dev/null7 8// Test bufferization using memref types that have no layout map.9// RUN: mlir-opt %s -one-shot-bufferize="unknown-type-conversion=identity-layout-map bufferize-function-boundaries" -split-input-file -o /dev/null10 11// CHECK-LABEL: func private @insert_slice_fun12//  CHECK-SAME:   %[[A0:[a-zA-Z0-9]*]]: memref<?xf32, strided<[?], offset: ?>>,13//  CHECK-SAME:   %[[A1:[a-zA-Z0-9]*]]: memref<?xf32, strided<[?], offset: ?>>,14//  CHECK-SAME:   %[[t0:[a-zA-Z0-9]*]]: memref<4xf32, strided<[?], offset: ?>>,15//  CHECK-SAME:   %[[t1:[a-zA-Z0-9]*]]: memref<4xf32, strided<[?], offset: ?>>16func.func private @insert_slice_fun(17    %A0 : tensor<?xf32> {bufferization.writable = false},18    %A1 : tensor<?xf32> {bufferization.writable = true},19    %t0 : tensor<4xf32> {bufferization.writable = false},20    %t1 : tensor<4xf32> {bufferization.writable = true})21  ->  (tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>)22{23  // Alloc and copy the whole result tensor. Copy the tensor.extract_slice.24  //      CHECK: %[[REALLOC3:.*]] = memref.alloc25  //      CHECK: memref.copy %[[A0]], %[[REALLOC3]]26  //      CHECK: %[[SV_A0:.*]] = memref.subview %[[REALLOC3]]27  //      CHECK: memref.copy %[[t0]], %[[SV_A0]]28  %r0 = tensor.insert_slice %t0 into %A0[0][4][1] : tensor<4xf32> into tensor<?xf32>29 30  // Alloc and copy the whole result tensor. Copy the tensor.extract_slice.31  //      CHECK: %[[REALLOC2:.*]] = memref.alloc32  //      CHECK: memref.copy %[[A0]]33  //      CHECK: %[[SV_A0_2:.*]] = memref.subview %[[REALLOC2]]34  //      CHECK: memref.copy %[[t1]], %[[SV_A0_2]]35  %r1 = tensor.insert_slice %t1 into %A0[0][4][1] : tensor<4xf32> into tensor<?xf32>36 37  //  Still alloc the large tensor because %A1 is read after. Copy the tensor.extract_slice.38  //      CHECK: %[[REALLOC1:.*]] = memref.alloc39  //      CHECK: memref.copy %[[A1]]40  //      CHECK: %[[SV_A1:.*]] = memref.subview %[[REALLOC1]]41  //      CHECK: memref.copy %[[t0]], %[[SV_A1]]42  %r2 = tensor.insert_slice %t0 into %A1[0][4][1] : tensor<4xf32> into tensor<?xf32>43 44  //  Do not realloc the large tensor. Copy the tensor.extract_slice.45  //  CHECK-NOT: alloc46  //      CHECK: %[[SV_A1_2:.*]] = memref.subview %[[A1]]47  //      CHECK: memref.copy %[[t1]], %[[SV_A1_2]]48  %r3 = tensor.insert_slice %t1 into %A1[0][4][1] : tensor<4xf32> into tensor<?xf32>49 50  //      CHECK: return %[[REALLOC3]], %[[REALLOC2]], %[[REALLOC1]] :51  // CHECK-SAME:   memref<?xf32>, memref<?xf32>, memref<?xf32>52  return %r0, %r1, %r2, %r3: tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>53}54 55// -----56 57// CHECK-LABEL: func @insert_slice_fun58//  CHECK-SAME:   %[[A:[a-zA-Z0-9]*]]: memref<?xf32, strided<[?], offset: ?>>59//  CHECK-SAME:   %[[t:[a-zA-Z0-9]*]]: memref<4xf32, strided<[?], offset: ?>>60func.func @insert_slice_fun(61    %A : tensor<?xf32> {bufferization.writable = true},62    %t : tensor<4xf32> {bufferization.writable = false})63  -> tensor<?xf32>64{65  %f0 = arith.constant 0.0 : f3266 67  //  CHECK-NOT: alloc68  //      CHECK: %[[SV_A:.*]] = memref.subview %[[A]]69  //      CHECK: memref.copy %[[t]], %[[SV_A]]70  %r0 = tensor.insert_slice %t into %A[0][4][1] : tensor<4xf32> into tensor<?xf32>71 72  /// Overwrite A inplace.73  //      CHECK: linalg.fill ins({{.*}}{{.*}}outs(%[[A]]74  %r1 = linalg.fill ins(%f0 : f32) outs(%r0 : tensor<?xf32>) -> tensor<?xf32>75 76  //     CHECK: return77  // CHECK-NOT: tensor78  return %r1: tensor<?xf32>79}80 81// -----82 83// CHECK-LABEL: func @insert_slice_fun84//  CHECK-SAME:   %[[A:[a-zA-Z0-9]*]]: memref<?xf32, strided<[?], offset: ?>>85//  CHECK-SAME:   %[[t:[a-zA-Z0-9]*]]: memref<4xf32, strided<[?], offset: ?>>86func.func @insert_slice_fun(87    %A : tensor<?xf32> {bufferization.writable = true},88    %t : tensor<4xf32> {bufferization.writable = false})89  -> tensor<?xf32>90{91  %f0 = arith.constant 0.0 : f3292 93  //      CHECK: linalg.fill ins({{.*}}{{.*}}outs(%[[A]]94  %r0 = linalg.fill ins(%f0 : f32) outs(%A : tensor<?xf32>) -> tensor<?xf32>95 96  //  CHECK-NOT: alloc97  //      CHECK: %[[SV_A:.*]] = memref.subview %[[A]]98  /// Overwrite A inplace by copying into the subview.99  //      CHECK: memref.copy %[[t]], %[[SV_A]]100  %r1 = tensor.insert_slice %t into %r0[0][4][1] : tensor<4xf32> into tensor<?xf32>101 102  //     CHECK: return103  // CHECK-NOT: tensor104  return %r1: tensor<?xf32>105}106 107// -----108 109// CHECK-LABEL: func @insert_slice_fun_not_inplace110//  CHECK-SAME:   %[[A:[a-zA-Z0-9]*]]: memref<?xf32, strided<[?], offset: ?>>111//  CHECK-SAME:   %[[t:[a-zA-Z0-9]*]]: memref<4xf32, strided<[?], offset: ?>>112func.func @insert_slice_fun_not_inplace(113    %A : tensor<?xf32> {bufferization.writable = false},114    %t : tensor<4xf32> {bufferization.writable = false})115  -> tensor<?xf32>116{117  //      CHECK: %[[ALLOC:.*]] = memref.alloc(%{{.*}}) {alignment = 64 : i64} : memref<?xf32>118  //      CHECK: memref.copy %[[A]], %[[ALLOC]] : memref<?xf32{{.*}} to memref<?xf32>119  //      CHECK: %[[SV:.*]] = memref.subview %[[ALLOC]][0] [4] [1] : memref<?xf32> to memref<4xf32, strided<[1]>>120  //      CHECK: memref.copy %[[t]], %[[SV]] : memref<4xf32, strided{{.*}}> to memref<4xf32, strided<[1]>>121  %r0 = tensor.insert_slice %t into %A[0][4][1] : tensor<4xf32> into tensor<?xf32>122 123  //     CHECK: return %{{.*}} : memref<?xf32>124  return %r0: tensor<?xf32>125}126 127// -----128 129// This test case could bufferize in-place with a better analysis. However, it130// is simpler to let the canonicalizer fold away the tensor.insert_slice.131 132// CHECK-LABEL: func @tensor_cast_not_in_place(133//  CHECK-SAME:     %[[A:.*]]: memref<?xf32{{.*}}>, %[[B:.*]]: memref<?xf32{{.*}}>134//       CHECK:   %[[casted:.*]] = memref.cast %[[A]] : memref<?xf32, strided<[?], offset: ?>> to memref<4xf32, strided<[?], offset: ?>>135//       CHECK:   %[[alloc:.*]] = memref.alloc136//       CHECK:   memref.copy %[[casted]], %[[alloc]]137//       CHECK:   %[[subview:.*]] = memref.subview %[[A]][{{.*}}] [4] [1] : {{.*}} to memref<4xf32138//       CHECK:   memref.copy %[[alloc]], %[[subview]]139func.func @tensor_cast_not_in_place(140    %A : tensor<?xf32> {bufferization.writable = true},141    %B : tensor<?xf32> {bufferization.writable = false}, %idx: index)142  -> (tensor<?xf32>)143{144  %r0 = tensor.cast %A : tensor<?xf32> to tensor<4xf32>145  %r1 = tensor.insert_slice %r0 into %A[%idx][4][1] : tensor<4xf32> into tensor<?xf32>146  return %r1 : tensor<?xf32>147}148 149// -----150 151// CHECK-LABEL: func @insert_op152//  CHECK-SAME:     %[[t1:.*]]: memref<?xf32, {{.*}}>, %[[s:.*]]: f32, %[[i:.*]]: index153func.func @insert_op(%t1 : tensor<?xf32> {bufferization.writable = true},154                     %s : f32, %i : index) -> tensor<?xf32> {155  // CHECK: memref.store %[[s]], %[[t1]][%[[i]]]156  %0 = tensor.insert %s into %t1[%i] : tensor<?xf32>157  // CHECK: return158  return %0 : tensor<?xf32>159}160 161// -----162 163// A regression test to make sure that we handle rank-reducing extract_slice164// correctly.165 166// CHECK-LABEL: func @rank_reducing167func.func @rank_reducing(168    %i: index, %j: index,169    %arg0: tensor<8x18x32xf32>)170      -> tensor<?x1x6x8xf32> {171  %c1 = arith.constant 1 : index172  %c6 = arith.constant 6 : index173  %c8 = arith.constant 8 : index174  %c32 = arith.constant 32 : index175  %c0 = arith.constant 0 : index176  %0 = bufferization.alloc_tensor() : tensor<4x1x6x8xf32>177  %1 = tensor.cast %0 : tensor<4x1x6x8xf32> to tensor<?x1x6x8xf32>178  %2 = bufferization.alloc_tensor() : tensor<1x6x8xf32>179  %5 = scf.for %arg7 = %c0 to %c32 step %c8 iter_args(%arg8 = %1) -> (tensor<?x1x6x8xf32>) {180    %7 = affine.apply affine_map<(d0) -> (d0 ceildiv 8)>(%arg7)181    %8 = tensor.extract_slice %arg0[%i, %j, %arg7] [1, 6, 8] [1, 1, 1] : tensor<8x18x32xf32> to tensor<1x6x8xf32>182    %9 = scf.for %arg9 = %c0 to %c6 step %c1 iter_args(%arg10 = %2) -> (tensor<1x6x8xf32>) {183      %11 = tensor.extract_slice %8[0, %arg9, 0] [1, 1, 8] [1, 1, 1] : tensor<1x6x8xf32> to tensor<1x1x8xf32>184      %12 = tensor.insert_slice %11 into %arg10[0, %arg9, 0] [1, 1, 8] [1, 1, 1] : tensor<1x1x8xf32> into tensor<1x6x8xf32>185      scf.yield %12 : tensor<1x6x8xf32>186    }187    %10 = tensor.insert_slice %9 into %arg8[%7, 0, 0, 0] [1, 1, 6, 8] [1, 1, 1, 1] : tensor<1x6x8xf32> into tensor<?x1x6x8xf32>188    scf.yield %10 : tensor<?x1x6x8xf32>189  }190  return %5: tensor<?x1x6x8xf32>191}192 193// -----194 195// CHECK-LABEL: func.func @rank_reducing_parallel_insert_slice196func.func @rank_reducing_parallel_insert_slice(%in: tensor<100xf32>, %out: tensor<200x100xf32>) {197  %c1 = arith.constant 1 : index198  %num_threads = arith.constant 100 : index199 200  // CHECK: scf.forall {{.*}} {201  %result = scf.forall (%thread_idx) in (%num_threads) shared_outs (%o = %out) -> tensor<200x100xf32> {202      %1 = tensor.extract_slice %in[%thread_idx][1][1] : tensor<100xf32> to tensor<1xf32>203      scf.forall.in_parallel {204        // CHECK: memref.subview %{{.*}}[%{{.*}}] [1] [1] : memref<100xf32, strided<[?], offset: ?>> to memref<1xf32, strided<[?], offset: ?>>205        // CHECK: memref.subview %{{.*}}[1, %{{.*}}] [1, 1] [1, 1] : memref<200x100xf32, strided<[?, ?], offset: ?>> to memref<1xf32, strided<[?], offset: ?>>206        tensor.parallel_insert_slice %1 into %o[1, %thread_idx][1, 1][1, 1] :207          tensor<1xf32> into tensor<200x100xf32>208      }209  }210  // CHECK: }211  return212}213 214// -----215 216// CHECK-LABEL: func.func @parallel_insert_full_slice_in_place217// CHECK-NOT:     memref.alloc()218func.func @parallel_insert_full_slice_in_place(%2: tensor<2xf32>) -> tensor<2xf32> {219  %cst = arith.constant 0.000000e+00 : f32220  %3 = scf.forall (%arg0) in (1) shared_outs(%arg2 = %2) -> (tensor<2xf32>) {221    %fill = linalg.fill ins(%cst : f32) outs(%arg2 : tensor<2xf32>) -> tensor<2xf32>222    scf.forall.in_parallel {223      tensor.parallel_insert_slice %fill into %arg2[0] [2] [1] : tensor<2xf32> into tensor<2xf32>224    }225  } {mapping = [#gpu.thread<linear_dim_0>]}226  return %3 : tensor<2xf32>227}228 229// -----230 231// This test case could bufferize in-place with a better analysis. However, it232// is simpler to let the canonicalizer fold away the tensor.insert_slice.233 234// CHECK-LABEL: func @insert_equivalent_tensor235func.func @insert_equivalent_tensor(%t: tensor<10xf32>) -> tensor<10xf32> {236  // CHECK: memref.alloc237  %cst = arith.constant 4.200000e+01 : f32238  // CHECK: linalg.fill239  %0 = linalg.fill ins(%cst : f32) outs(%t : tensor<10xf32>) -> tensor<10xf32>240  // CHECK: memref.copy241  %1 = tensor.insert_slice %0 into %t[0][10][1] : tensor<10xf32> into tensor<10xf32>242  return %1 : tensor<10xf32>243}244 245// -----246 247// CHECK-LABEL: func @pad_memory_space(248//  CHECK-SAME:     %[[t:.*]]: memref<?xf32, strided<[?], offset: ?>>249func.func @pad_memory_space(%t: tensor<?xf32>, %h1: index, %f: f32, %pos: index) -> f32250{251  // CHECK: %[[alloc_tensor:.*]] = memref.alloc{{.*}} : memref<?xf32, 3>252  // CHECK: memref.copy %[[t]], %[[alloc_tensor]]253  %0 = bufferization.alloc_tensor() copy(%t)254      {memory_space = 3 : i64} : tensor<?xf32>255  // CHECK: %[[padded_alloc:.*]] = memref.alloc() {{.*}} : memref<15xf32, 3>256  // CHECK: linalg.map257  // CHECK:     outs(%[[padded_alloc]] : memref<15xf32, 3>)258  // CHECK:   linalg.yield %{{.*}}259  // CHECK: }260  // CHECK: %[[subview:.*]] = memref.subview {{.*}} : memref<15xf32, 3> to memref<?xf32, strided<[1], offset: 2>, 3>261  // CHECK: memref.copy %[[alloc_tensor]], %[[subview]]262  %1 = tensor.pad %0 low[2] high[%h1] {263  ^bb0(%arg0: index):264    tensor.yield %f : f32265  } : tensor<?xf32> to tensor<15xf32>266  // CHECK: memref.load {{.*}} : memref<15xf32, 3>267  %2 = tensor.extract %1[%pos] : tensor<15xf32>268  return %2 : f32269}270 271// -----272 273// CHECK-LABEL: func @insert_slice_regression(274//  CHECK-SAME:   %[[t:.*]]: memref<10xf32,{{.*}}>, %[[b:.*]]: memref<5xf32275func.func @insert_slice_regression(%t: tensor<10xf32>, %b: tensor<5xf32>) -> tensor<10xf32> {276  %cst = arith.constant 0.0 : f32277  %c0 = arith.constant 0 : index278  // CHECK: %[[alloc:.*]] = memref.alloc() {{.*}} : memref<10xf32>279  // CHECK: linalg.fill {{.*}} outs(%[[alloc]] : memref<10xf32>)280  %1 = linalg.fill ins(%cst : f32) outs(%t : tensor<10xf32>) -> tensor<10xf32>281 282  // Read %1 so that it does not DCE away.283  %vec = vector.transfer_read %1[%c0], %cst : tensor<10xf32>, vector<10xf32>284  vector.print %vec : vector<10xf32>285 286  // Write back a different value (not %1).287  // CHECK: %[[subview:.*]] = memref.subview %[[t]][0] [5] [1]288  // CHECK: memref.copy %[[b]], %[[subview]]289  %2 = tensor.insert_slice %b into %t[0][5][1] : tensor<5xf32> into tensor<10xf32>290  return %2 : tensor<10xf32>291}292 293// -----294 295// CHECK-LABEL: func @insert_slice_full_overwrite(296//  CHECK-SAME:   %[[t:.*]]: memref<10xf32,{{.*}}>, %[[b:.*]]: memref<10xf32,{{.*}}>297func.func @insert_slice_full_overwrite(%t: tensor<10xf32>, %b: tensor<10xf32>) -> tensor<10xf32> {298  %cst = arith.constant 0.0 : f32299  %c0 = arith.constant 0 : index300  // CHECK: linalg.fill {{.*}} outs(%[[t]] : memref<10xf32,{{.*}}>)301  %1 = linalg.fill ins(%cst : f32) outs(%t : tensor<10xf32>) -> tensor<10xf32>302 303  // Read %1 so that it does not DCE away.304  %vec = vector.transfer_read %1[%c0], %cst : tensor<10xf32>, vector<10xf32>305  vector.print %vec : vector<10xf32>306 307  // Write back a different value (not %1).308  // CHECK: %[[subview:.*]] = memref.subview %[[t]][0] [10] [1]309  // CHECK: memref.copy %[[b]], %[[subview]]310  %2 = tensor.insert_slice %b into %t[0][10][1] : tensor<10xf32> into tensor<10xf32>311  return %2 : tensor<10xf32>312}313 314// -----315 316// CHECK-LABEL: func @dim_not_reading(317//  CHECK-SAME:     %[[t:.*]]: memref<?xf32318func.func @dim_not_reading(%t: tensor<?xf32>, %f: f32, %pos: index)319    -> (tensor<?xf32>, index)320{321  %c0 = arith.constant 0 : index322  // CHECK-NOT: memref.alloc323  // CHECK-NOT: memref.copy324  //     CHECK: memref.store %{{.*}}, %[[t]]325  %0 = tensor.insert %f into %t[%pos] : tensor<?xf32>326  //     CHECK: memref.dim %[[t]]327  %1 = tensor.dim %t, %c0 : tensor<?xf32>328  return %0, %1 : tensor<?xf32>, index329}330 331// -----332 333//       CHECK: #[[$map:.*]] = affine_map<(d0) -> (d0 + 5)>334// CHECK-LABEL: func.func private @cast_retains_buffer_layout(335//  CHECK-SAME:     %[[t:.*]]: memref<?xf32, #[[$map]]>, %[[sz:.*]]: index) -> memref<?xf32, strided<[1], offset: 7>> {336//       CHECK:   %[[casted:.*]] = memref.cast %[[t]] : memref<?xf32, #[[$map]]> to memref<10xf32, #[[$map]]>337//       CHECK:   %[[slice:.*]] = memref.subview %[[casted]][2] [%[[sz]]] [1] : memref<10xf32, #[[$map]]> to memref<?xf32, strided<[1], offset: 7>>338//       CHECK:   return %[[slice]]339func.func private @cast_retains_buffer_layout(340    %t: tensor<?xf32>341        {bufferization.buffer_layout = affine_map<(d0) -> (d0 + 5)>},342    %sz: index)343  -> (tensor<10xf32>, tensor<?xf32>)344{345  %casted = tensor.cast %t : tensor<?xf32> to tensor<10xf32>346  %slice = tensor.extract_slice %casted[2][%sz][1] : tensor<10xf32> to tensor<?xf32>347 348  // Note: The %casted return type is folded away because both buffers are349  // equivalent. Therefore, we currently loose some static type information350  // in the caller.351  return %casted, %slice : tensor<10xf32>, tensor<?xf32>352}353 354// -----355 356// CHECK-LABEL: func private @cast_retains_buffer_layout_strided(357//  CHECK-SAME:     %[[t:.*]]: memref<?xf32, strided<[1], offset: 5>>, %[[sz:.*]]: index) -> memref<?xf32, strided<[1], offset: 7>> {358//       CHECK:   %[[casted:.*]] = memref.cast %[[t]] : memref<?xf32, strided<[1], offset: 5>> to memref<10xf32, strided<[1], offset: 5>>359//       CHECK:   %[[slice:.*]] = memref.subview %[[casted]][2] [%[[sz]]] [1] : memref<10xf32, strided<[1], offset: 5>> to memref<?xf32, strided<[1], offset: 7>>360//       CHECK:   return %[[slice]]361func.func private @cast_retains_buffer_layout_strided(362    %t: tensor<?xf32>363        {bufferization.buffer_layout = strided<[1], offset: 5>},364    %sz: index)365  -> (tensor<10xf32>, tensor<?xf32>)366{367  %casted = tensor.cast %t : tensor<?xf32> to tensor<10xf32>368  %slice = tensor.extract_slice %casted[2][%sz][1] : tensor<10xf32> to tensor<?xf32>369 370  // Note: The %casted return type is folded away because both buffers are371  // equivalent. Therefore, we currently loose some static type information372  // in the caller.373  return %casted, %slice : tensor<10xf32>, tensor<?xf32>374}375 376// -----377 378// CHECK-LABEL: func.func @parallel_insert_slice_source_out_of_place379func.func @parallel_insert_slice_source_out_of_place(%in: tensor<1xf32>, %out: tensor<100xf32>, %f: f32) {380  %c0 = arith.constant 0 : index381  %c1 = arith.constant 1 : index382  %num_threads = arith.constant 50 : index383 384  // CHECK: scf.forall {{.*}} {385  %result = scf.forall (%thread_idx) in (%num_threads) shared_outs (%o = %out) -> tensor<100xf32> {386      // The tensor.insert must bufferize out-of-place.387      // CHECK: memref.alloc388      // CHECK: memref.store389      %insert = tensor.insert %f into %in[%c0] : tensor<1xf32>390      %r = tensor.extract %in[%c0] : tensor<1xf32>391      vector.print %r : f32392 393      // CHECK: memref.copy394      scf.forall.in_parallel {395        tensor.parallel_insert_slice %insert into %o[%thread_idx][1][1] :396          tensor<1xf32> into tensor<100xf32>397      }398  }399  // CHECK: }400  return401}402 403// -----404 405// CHECK-LABEL: func @tensor.reshape(406func.func @tensor.reshape() -> tensor<2x2x5xf32> {407  // CHECK-DAG: %[[M1:.*]] = memref.cast %{{.*}} : memref<2x10xf32> to memref<?x10xf32>408  %t1_static = arith.constant dense<0.> : tensor<2x10xf32>409  %t1 = tensor.cast %t1_static : tensor<2x10xf32> to tensor<?x10xf32>410 411  // CHECK: %[[SHAPE:.*]] = memref.get_global @{{.*}} : memref<3xi64>412  %shape = arith.constant dense<[2, 2, 5]> : tensor<3xi64>413 414  // CHECK: %[[RESHAPED:.*]] = memref.reshape %[[M1]](%[[SHAPE]]) : (memref<?x10xf32>, memref<3xi64>) -> memref<2x2x5xf32>415  %reshaped = tensor.reshape %t1(%shape) : (tensor<?x10xf32>, tensor<3xi64>) -> tensor<2x2x5xf32>416 417  // CHECK: return %[[RESHAPED]]418  return %reshaped : tensor<2x2x5xf32>419}420 421// -----422 423// CHECK-LABEL: func @tensor_reshape_aliasing424//  CHECK-SAME:  (%[[ARG0:.+]]: index, %[[ARG1:.+]]: index)425func.func @tensor_reshape_aliasing(%arg0: index, %arg1: index) -> tensor<?x?xf32> {426  %t1_static = arith.constant dense<0.> : tensor<10xf32>427  // CHECK-DAG: %[[T1:.+]] = memref.cast428  %t1 = tensor.cast %t1_static : tensor<10xf32> to tensor<?xf32>429 430  // CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index431  %c0 = arith.constant 0 : index432  // CHECK-DAG: %[[C1:.+]] = arith.constant 1 : index433  %c1 = arith.constant 1 : index434 435  // CHECK-DAG: %[[SHAPE:.+]] = memref.alloc() {{.*}} : memref<2xindex>436  %shape = bufferization.alloc_tensor() : tensor<2xindex>437  // CHECK: memref.store %[[ARG0]], %[[SHAPE]][%[[C0]]]438  %shape.0 = tensor.insert %arg0 into %shape[%c0] : tensor<2xindex>439  // CHECK: memref.store %[[ARG1]], %[[SHAPE]][%[[C1]]]440  %shape.1 = tensor.insert %arg1 into %shape.0[%c1] : tensor<2xindex>441 442  // CHECK: %[[RESHAPED:.+]] = memref.reshape %[[T1]](%[[SHAPE]])443  %reshaped = tensor.reshape %t1(%shape.1) : (tensor<?xf32>, tensor<2xindex>) -> tensor<?x?xf32>444  // CHECK: return %[[RESHAPED]]445  return %reshaped : tensor<?x?xf32>446}447 448// -----449 450// CHECK-LABEL: @reshape_with_non_identity_layout(451// CHECK-SAME:    %[[INPUT:[a-zA-Z0-9]*]]: memref<2x2xf32, strided<[?, ?], offset: ?>, 3>,452// CHECK-SAME:    %[[LAYOUT:[a-zA-Z0-9]*]]: memref<2xi32, strided<[?], offset: ?>>,453func.func @reshape_with_non_identity_layout(%arg0: memref<2x2xf32, strided<[?, ?], offset: ?>, 3>, %arg1: tensor<2xi32>, %idx: index) -> f32 {454  %t = bufferization.to_tensor %arg0 restrict : memref<2x2xf32, strided<[?, ?], offset: ?>, 3> to tensor<2x2xf32>455 456  // CHECK: %[[SUBVIEW:.+]] = memref.subview %[[INPUT]][1, 0] [1, 2] [1, 1] : memref<2x2xf32, strided<[?, ?], offset: ?>, 3> to memref<2xf32, strided<[?], offset: ?>, 3>457  %extracted_slice = tensor.extract_slice %t[1, 0] [1, 2] [1, 1] : tensor<2x2xf32> to tensor<2xf32>458 459  // To satisify the constraints of memref.reshape, the subview must be460  // reallocated a buffer with an identity layout.461  // CHECK: %[[ALLOC:.+]] = memref.alloc() {{.*}} : memref<2xf32, 3>462  // CHECK: memref.copy %[[SUBVIEW]], %[[ALLOC]]463  // CHECK: %[[RESHAPED:.+]] = memref.reshape %[[ALLOC]](%[[LAYOUT]]) : (memref<2xf32, 3>, memref<2xi32, strided<[?], offset: ?>>) -> memref<1x2xf32, 3>464  %reshape = tensor.reshape %extracted_slice(%arg1) : (tensor<2xf32>, tensor<2xi32>) -> tensor<1x2xf32>465 466  %r = tensor.extract %reshape[%idx, %idx] : tensor<1x2xf32>467  return %r : f32468}469 470// -----471 472// CHECK-LABEL: func @collapse_shape_regression(473//  CHECK-SAME:     %[[t:.*]]: memref<10x20xf32,474func.func @collapse_shape_regression(475    %t: tensor<10x20xf32>, %f: f32, %idx: index) {476  // CHECK: %[[subview:.*]] = memref.subview %[[t]]477  %0 = tensor.extract_slice %t [2, 3] [5, 6] [1, 1]478      : tensor<10x20xf32> to tensor<5x6xf32>479 480  // Insert a copy because the original %0 is read later.481  // CHECK: %[[alloc1:.*]] = memref.alloc() {{.*}} : memref<5x6xf32>482  // CHECK: memref.copy %[[subview]], %[[alloc1]]483  // CHECK: memref.store {{.*}}, %[[alloc1]]484  tensor.insert %f into %0[%idx, %idx] : tensor<5x6xf32>485 486  // Insert a copy because the shape cannot be collapsed.487  // CHECK: %[[alloc2:.*]] = memref.alloc() {{.*}} : memref<5x6xf32>488  // CHECK: memref.copy %[[subview]], %[[alloc2]]489  // CHECK: memref.collapse_shape %[[alloc2]]490  tensor.collapse_shape %0[[0, 1]] : tensor<5x6xf32> into tensor<30xf32>491  return492}493 494// -----495 496// CHECK-LABEL: func private @mult_return_callee(497//  CHECK-SAME:   %[[T:.*]]: memref<?xf32, strided<[?], offset: ?>>, %[[COND:.*]]: i1,498//  CHECK-SAME:   %[[A:.*]]: index, %[[B:.*]]: index) -> index {499//       CHECK:   cf.cond_br %[[COND]], ^bb1, ^bb2500//       CHECK: ^bb1:501//       CHECK:   return %[[A]] : index502//       CHECK: ^bb2:503//       CHECK:   return %[[B]] : index504func.func private @mult_return_callee(%t: tensor<?xf32>,  %cond:i1, %a: index, %b: index) -> (tensor<10xf32>, index) {505  %casted = tensor.cast %t : tensor<?xf32> to tensor<10xf32>506  cf.cond_br %cond,^a, ^b507^a:508  return %casted, %a : tensor<10xf32>, index509^b:510  return %casted, %b : tensor<10xf32>, index511}512 513// CHECK-LABEL: func @mult_return(514//  CHECK-SAME:   %[[T:.*]]: memref<?xf32, strided<[?], offset: ?>>, %[[COND:.*]]: i1,515//  CHECK-SAME:   %[[A:.*]]: index, %[[B:.*]]: index) -> (memref<?xf32, strided<[?], offset: ?>>, index) {516func.func @mult_return(%t: tensor<?xf32>,  %cond:i1, %a: index, %b: index) -> (tensor<10xf32>, index) {517  // CHECK: %[[RET:.*]] = call @mult_return_callee(%[[T]], %[[COND]], %[[A]], %[[B]]) : (memref<?xf32, strided<[?], offset: ?>>, i1, index, index) -> index518  // CHECK: return %[[T]], %[[RET]] : memref<?xf32, strided<[?], offset: ?>>, index519  %t_res, %v = func.call @mult_return_callee(%t, %cond, %a, %b) : (tensor<?xf32>, i1, index, index) -> (tensor<10xf32>, index) 520  return %t_res, %v : tensor<10xf32>, index521}522