712 lines · plain
1// RUN: mlir-opt -expand-strided-metadata -finalize-memref-to-llvm -lower-affine -convert-arith-to-llvm -cse %s -split-input-file | FileCheck %s2// RUN: mlir-opt -expand-strided-metadata -finalize-memref-to-llvm='index-bitwidth=32' -lower-affine -convert-arith-to-llvm='index-bitwidth=32' -cse %s -split-input-file | FileCheck %s --check-prefix=CHECK323//4// This test demonstrates a full "memref to llvm" pipeline where5// we first expand some of the memref operations (using affine,6// arith, and memref operations) and convert each individual7// dialect to llvm.8//9// Note: We run CSE in that test to get rid of duplicated10// unrealized_conversion_cast that are inserted with11// finalize-memref-to-llvm and then convert-arith-to-llvm.12// The final code is still not perfect, because we have13// noop unrealized_conversion_cast from i64 to index14// and back.15 16// -----17 18// Final offset = baseOffset + subOffset0 * baseStride0 + subOffset1 * baseStride119// = 0 + %arg0 * 4 + %arg1 * 120// = %arg0 * 4 + %arg121// Because of how subviews are lowered (i.e., using22// reintpret_cast(extract_strided_metadata(subSrc))),23// we end up with two casts of the base.24// More specifically, when we lower this sequence:25// ```26// base, ... = extract_strided_metadata subSrc27// dst = reinterpret_cast base, ...28// ```29//30// extract_strided_metadata gets lowered into:31// ```32// castedSrc = unrealized_conversion_cast subSrc33// base = extractvalue castedSrc[0]34// ```35//36// And reinterpret_cast gets lowered into:37// ```38// castedBase = unrealized_conversion_cast base39// dst = extractvalue %castedBase[0]40// ```41//42// Which give us:43// ```44// castedSrc = unrealized_conversion_cast src45// base = extractvalue castedSrc[0]46// castedBase = unrealized_conversion_cast base47// dst = extractvalue %castedBase[0] <-- dst and base are effectively equal.48// ```49// CHECK-LABEL: func @subview(50// CHECK: %[[MEM:.*]]: memref<{{.*}}>,51// CHECK: %[[ARG0f:[a-zA-Z0-9]*]]: index,52// CHECK: %[[ARG1f:[a-zA-Z0-9]*]]: index,53// CHECK: %[[ARG2f:.*]]: index)54func.func @subview(%0 : memref<64x4xf32, strided<[4, 1], offset: 0>>, %arg0 : index, %arg1 : index, %arg2 : index)55-> memref<?x?xf32, strided<[?, ?], offset: ?>> {56 // CHECK-DAG: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]57 // CHECK-DAG: %[[ARG0:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]58 // CHECK-DAG: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]59 60 // CHECK: %[[BASE:.*]] = llvm.extractvalue %[[MEMREF]][0] : !llvm.struct<(ptr, ptr, i6461 // CHECK: %[[BASE_ALIGNED:.*]] = llvm.extractvalue %[[MEMREF]][1] : !llvm.struct<(ptr, ptr, i6462 // CHECK: %[[STRIDE0:.*]] = llvm.mlir.constant(4 : index) : i6463 // CHECK: %[[DESCSTRIDE0:.*]] = llvm.mul %[[ARG0]], %[[STRIDE0]] overflow<nsw> : i6464 // CHECK: %[[TMP:.*]] = builtin.unrealized_conversion_cast %[[DESCSTRIDE0]] : i64 to index65 // CHECK: %[[DESCSTRIDE0_V2:.*]] = builtin.unrealized_conversion_cast %[[TMP]] : index to i6466 // CHECK: %[[OFF2:.*]] = llvm.add %[[DESCSTRIDE0]], %[[ARG1]] : i6467 // CHECK: %[[TMP:.*]] = builtin.unrealized_conversion_cast %[[OFF2]] : i64 to index68 // CHECK: %[[OFF2:.*]] = builtin.unrealized_conversion_cast %[[TMP]] : index to i6469 // CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>70 // CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BASE]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>71 // CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[BASE_ALIGNED]], %[[DESC0]][1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>72 // CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[OFF2]], %[[DESC1]][2] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>73 // CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[ARG0]], %[[DESC2]][3, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>74 // CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[DESCSTRIDE0_V2]], %[[DESC3]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>75 // CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[ARG1]], %[[DESC4]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>76 // CHECK: %[[DESC6:.*]] = llvm.insertvalue %[[ARG1]], %[[DESC5]][4, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>77 78 %1 = memref.subview %0[%arg0, %arg1][%arg0, %arg1][%arg0, %arg1] :79 memref<64x4xf32, strided<[4, 1], offset: 0>>80 to memref<?x?xf32, strided<[?, ?], offset: ?>>81 return %1 : memref<?x?xf32, strided<[?, ?], offset: ?>>82}83 84// -----85 86// CHECK-LABEL: func @subview_non_zero_addrspace(87// CHECK: %[[MEM:[a-zA-Z0-9]*]]: memref<{{[^%]*}}>,88// CHECK: %[[ARG0f:[a-zA-Z0-9]*]]: index,89// CHECK: %[[ARG1f:[a-zA-Z0-9]*]]: index,90// CHECK: %[[ARG2f:.*]]: index)91func.func @subview_non_zero_addrspace(%0 : memref<64x4xf32, strided<[4, 1], offset: 0>, 3>, %arg0 : index, %arg1 : index, %arg2 : index) -> memref<?x?xf32, strided<[?, ?], offset: ?>, 3> {92 // CHECK-DAG: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]93 // CHECK-DAG: %[[ARG0:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]94 // CHECK-DAG: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]95 96 // CHECK: %[[BASE:.*]] = llvm.extractvalue %[[MEMREF]][0] : !llvm.struct<(ptr<3>, ptr<3>, i64, array<2 x i64>, array<2 x i64>)>97 // CHECK: %[[BASE_ALIGNED:.*]] = llvm.extractvalue %[[MEMREF]][1] : !llvm.struct<(ptr<3>, ptr<3>, i64, array<2 x i64>, array<2 x i64>)>98 // CHECK: %[[STRIDE0:.*]] = llvm.mlir.constant(4 : index) : i6499 // CHECK: %[[DESCSTRIDE0:.*]] = llvm.mul %[[ARG0]], %[[STRIDE0]] overflow<nsw> : i64100 // CHECK: %[[TMP:.*]] = builtin.unrealized_conversion_cast %[[DESCSTRIDE0]] : i64 to index101 // CHECK: %[[DESCSTRIDE0_V2:.*]] = builtin.unrealized_conversion_cast %[[TMP]] : index to i64102 // CHECK: %[[OFF2:.*]] = llvm.add %[[DESCSTRIDE0]], %[[ARG1]] : i64103 // CHECK: %[[TMP:.*]] = builtin.unrealized_conversion_cast %[[OFF2]] : i64 to index104 // CHECK: %[[OFF2:.*]] = builtin.unrealized_conversion_cast %[[TMP]] : index to i64105 // CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr<3>, ptr<3>, i64, array<2 x i64>, array<2 x i64>)>106 // CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BASE]], %[[DESC]][0] : !llvm.struct<(ptr<3>, ptr<3>, i64, array<2 x i64>, array<2 x i64>)>107 // CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[BASE_ALIGNED]], %[[DESC0]][1] : !llvm.struct<(ptr<3>, ptr<3>, i64, array<2 x i64>, array<2 x i64>)>108 // CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[OFF2]], %[[DESC1]][2] : !llvm.struct<(ptr<3>, ptr<3>, i64, array<2 x i64>, array<2 x i64>)>109 // CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[ARG0]], %[[DESC2]][3, 0] : !llvm.struct<(ptr<3>, ptr<3>, i64, array<2 x i64>, array<2 x i64>)>110 // CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[DESCSTRIDE0_V2]], %[[DESC3]][4, 0] : !llvm.struct<(ptr<3>, ptr<3>, i64, array<2 x i64>, array<2 x i64>)>111 // CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[ARG1]], %[[DESC4]][3, 1] : !llvm.struct<(ptr<3>, ptr<3>, i64, array<2 x i64>, array<2 x i64>)>112 // CHECK: %[[DESC6:.*]] = llvm.insertvalue %[[ARG1]], %[[DESC5]][4, 1] : !llvm.struct<(ptr<3>, ptr<3>, i64, array<2 x i64>, array<2 x i64>)>113 114 %1 = memref.subview %0[%arg0, %arg1][%arg0, %arg1][%arg0, %arg1] :115 memref<64x4xf32, strided<[4, 1], offset: 0>, 3>116 to memref<?x?xf32, strided<[?, ?], offset: ?>, 3>117 return %1 : memref<?x?xf32, strided<[?, ?], offset: ?>, 3>118}119 120// -----121 122// CHECK-LABEL: func @subview_const_size(123// CHECK-SAME: %[[MEM:[a-zA-Z0-9]*]]: memref<{{[^%]*}}>,124// CHECK-SAME: %[[ARG0f:[a-zA-Z0-9]*]]: index125// CHECK-SAME: %[[ARG1f:[a-zA-Z0-9]*]]: index126// CHECK-SAME: %[[ARG2f:[a-zA-Z0-9]*]]: index127func.func @subview_const_size(%0 : memref<64x4xf32, strided<[4, 1], offset: 0>>, %arg0 : index, %arg1 : index, %arg2 : index) -> memref<4x2xf32, strided<[?, ?], offset: ?>> {128 // CHECK-DAG: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]129 // CHECK-DAG: %[[ARG0:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]130 // CHECK-DAG: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]131 132 // CHECK: %[[BASE:.*]] = llvm.extractvalue %[[MEMREF]][0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>133 // CHECK: %[[BASE_ALIGNED:.*]] = llvm.extractvalue %[[MEMREF]][1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>134 // CHECK: %[[C4:.*]] = llvm.mlir.constant(4 : index) : i64135 // CHECK: %[[DESCSTRIDE0:.*]] = llvm.mul %[[ARG0]], %[[C4]] overflow<nsw> : i64136 // CHECK: %[[TMP:.*]] = builtin.unrealized_conversion_cast %[[DESCSTRIDE0]] : i64 to index137 // CHECK: %[[DESCSTRIDE0_V2:.*]] = builtin.unrealized_conversion_cast %[[TMP]] : index to i64138 // CHECK: %[[OFF2:.*]] = llvm.add %[[DESCSTRIDE0]], %[[ARG1]] : i64139 // CHECK: %[[TMP:.*]] = builtin.unrealized_conversion_cast %[[OFF2]] : i64 to index140 // CHECK: %[[OFF2:.*]] = builtin.unrealized_conversion_cast %[[TMP]] : index to i64141 // CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>142 // CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BASE]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>143 // CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[BASE_ALIGNED]], %[[DESC0]][1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>144 // CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[OFF2]], %[[DESC1]][2] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>145 // CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[C4]], %[[DESC2]][3, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>146 // CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[DESCSTRIDE0_V2]], %[[DESC3]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>147 // CHECK: %[[CST_SIZE1:.*]] = llvm.mlir.constant(2 : index) : i64148 // CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[CST_SIZE1]], %[[DESC4]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>149 // CHECK: %[[DESC6:.*]] = llvm.insertvalue %[[ARG1]], %[[DESC5]][4, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>150 151 %1 = memref.subview %0[%arg0, %arg1][4, 2][%arg0, %arg1] :152 memref<64x4xf32, strided<[4, 1], offset: 0>>153 to memref<4x2xf32, strided<[?, ?], offset: ?>>154 return %1 : memref<4x2xf32, strided<[?, ?], offset: ?>>155}156 157// -----158 159// CHECK-LABEL: func @subview_const_stride(160// CHECK-SAME: %[[MEM:[a-zA-Z0-9]*]]: memref<{{[^%]*}}>,161// CHECK-SAME: %[[ARG0f:[a-zA-Z0-9]*]]: index162// CHECK-SAME: %[[ARG1f:[a-zA-Z0-9]*]]: index163// CHECK-SAME: %[[ARG2f:[a-zA-Z0-9]*]]: index164func.func @subview_const_stride(%0 : memref<64x4xf32, strided<[4, 1], offset: 0>>, %arg0 : index, %arg1 : index, %arg2 : index) -> memref<?x?xf32, strided<[4, 2], offset: ?>> {165 // CHECK-DAG: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]166 // CHECK-DAG: %[[ARG0:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]167 // CHECK-DAG: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]168 169 // CHECK: %[[BASE:.*]] = llvm.extractvalue %[[MEMREF]][0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>170 // CHECK: %[[BASE_ALIGNED:.*]] = llvm.extractvalue %[[MEMREF]][1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>171 // CHECK: %[[C4:.*]] = llvm.mlir.constant(4 : index) : i64172 // CHECK: %[[OFF0:.*]] = llvm.mul %[[ARG0]], %[[C4]] overflow<nsw> : i64173 // CHECK: %[[OFF2:.*]] = llvm.add %[[OFF0]], %[[ARG1]] : i64174 // CHECK: %[[TMP:.*]] = builtin.unrealized_conversion_cast %[[OFF2]] : i64 to index175 // CHECK: %[[OFF2:.*]] = builtin.unrealized_conversion_cast %[[TMP]] : index to i64176 // CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>177 // CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BASE]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>178 // CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[BASE_ALIGNED]], %[[DESC0]][1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>179 // CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[OFF2]], %[[DESC1]][2] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>180 // CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[ARG0]], %[[DESC2]][3, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>181 // CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[C4]], %[[DESC3]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>182 // CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[ARG1]], %[[DESC4]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>183 // CHECK: %[[CST_STRIDE1:.*]] = llvm.mlir.constant(2 : index) : i64184 // CHECK: %[[DESC6:.*]] = llvm.insertvalue %[[CST_STRIDE1]], %[[DESC5]][4, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>185 186 %1 = memref.subview %0[%arg0, %arg1][%arg0, %arg1][1, 2] :187 memref<64x4xf32, strided<[4, 1], offset: 0>>188 to memref<?x?xf32, strided<[4, 2], offset: ?>>189 return %1 : memref<?x?xf32, strided<[4, 2], offset: ?>>190}191 192// -----193 194// CHECK-LABEL: func @subview_const_stride_and_offset(195// CHECK-SAME: %[[MEM:.*]]: memref<{{.*}}>196func.func @subview_const_stride_and_offset(%0 : memref<64x8xf32, strided<[8, 1], offset: 0>>) -> memref<62x3xf32, strided<[8, 1], offset: 2>> {197 // The last "insertvalue" that populates the memref descriptor from the function arguments.198 // CHECK: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]199 200 // CHECK: %[[BASE:.*]] = llvm.extractvalue %[[MEMREF]][0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>201 // CHECK: %[[BASE_ALIGNED:.*]] = llvm.extractvalue %[[MEMREF]][1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>202 // CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>203 // CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BASE]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>204 // CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[BASE_ALIGNED]], %[[DESC0]][1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>205 // CHECK: %[[CST_OFF:.*]] = llvm.mlir.constant(2 : index) : i64206 // CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[CST_OFF]], %[[DESC1]][2] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>207 // CHECK: %[[CST_SIZE0:.*]] = llvm.mlir.constant(62 : index) : i64208 // CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[CST_SIZE0]], %[[DESC2]][3, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>209 // CHECK: %[[CST_STRIDE0:.*]] = llvm.mlir.constant(8 : index) : i64210 // CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[CST_STRIDE0]], %[[DESC3]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>211 // CHECK: %[[CST_SIZE1:.*]] = llvm.mlir.constant(3 : index) : i64212 // CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[CST_SIZE1]], %[[DESC4]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>213 // CHECK: %[[CST_STRIDE1:.*]] = llvm.mlir.constant(1 : index) : i64214 // CHECK: %[[DESC6:.*]] = llvm.insertvalue %[[CST_STRIDE1]], %[[DESC5]][4, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>215 216 %1 = memref.subview %0[0, 2][62, 3][1, 1] :217 memref<64x8xf32, strided<[8, 1], offset: 0>>218 to memref<62x3xf32, strided<[8, 1], offset: 2>>219 return %1 : memref<62x3xf32, strided<[8, 1], offset: 2>>220}221 222// -----223 224// CHECK-LABEL: func @subview_mixed_static_dynamic(225// CHECK: %[[MEM:[a-zA-Z0-9]*]]: memref<{{[^%]*}}>,226// CHECK: %[[ARG0f:[a-zA-Z0-9]*]]: index,227// CHECK: %[[ARG1f:[a-zA-Z0-9]*]]: index,228// CHECK: %[[ARG2f:.*]]: index)229func.func @subview_mixed_static_dynamic(%0 : memref<64x4xf32, strided<[4, 1], offset: 0>>, %arg0 : index, %arg1 : index, %arg2 : index) -> memref<62x?xf32, strided<[?, 1], offset: ?>> {230 // CHECK-DAG: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]231 // CHECK-DAG: %[[ARG0:.*]] = builtin.unrealized_conversion_cast %[[ARG0f]]232 // CHECK-DAG: %[[ARG1:.*]] = builtin.unrealized_conversion_cast %[[ARG1f]]233 // CHECK-DAG: %[[ARG2:.*]] = builtin.unrealized_conversion_cast %[[ARG2f]]234 235 // CHECK: %[[BASE:.*]] = llvm.extractvalue %[[MEMREF]][0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>236 // CHECK: %[[BASE_ALIGNED:.*]] = llvm.extractvalue %[[MEMREF]][1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>237 // CHECK: %[[STRIDE0:.*]] = llvm.mlir.constant(4 : index) : i64238 // CHECK: %[[DESCSTRIDE0:.*]] = llvm.mul %[[ARG0]], %[[STRIDE0]] overflow<nsw> : i64239 // CHECK: %[[TMP:.*]] = builtin.unrealized_conversion_cast %[[DESCSTRIDE0]] : i64 to index240 // CHECK: %[[DESCSTRIDE0_V2:.*]] = builtin.unrealized_conversion_cast %[[TMP]] : index to i64241 // CHECK: %[[OFF0:.*]] = llvm.mul %[[ARG1]], %[[STRIDE0]] overflow<nsw> : i64242 // CHECK: %[[BASE_OFF:.*]] = llvm.mlir.constant(2 : index) : i64243 // CHECK: %[[OFF2:.*]] = llvm.add %[[OFF0]], %[[BASE_OFF]] : i64244 // CHECK: %[[TMP:.*]] = builtin.unrealized_conversion_cast %[[OFF2]] : i64 to index245 // CHECK: %[[OFF2:.*]] = builtin.unrealized_conversion_cast %[[TMP]] : index to i64246 // CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>247 // CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BASE]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>248 // CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[BASE_ALIGNED]], %[[DESC0]][1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>249 // CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[OFF2]], %[[DESC1]][2] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>250 // CHECK: %[[CST_SIZE0:.*]] = llvm.mlir.constant(62 : index) : i64251 // CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[CST_SIZE0]], %[[DESC2]][3, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>252 // CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[DESCSTRIDE0_V2]], %[[DESC3]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>253 // CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[ARG2]], %[[DESC4]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>254 // CHECK: %[[CST_STRIDE1:.*]] = llvm.mlir.constant(1 : index) : i64255 // CHECK: %[[DESC6:.*]] = llvm.insertvalue %[[CST_STRIDE1]], %[[DESC5]][4, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>256 257 %1 = memref.subview %0[%arg1, 2][62, %arg2][%arg0, 1] :258 memref<64x4xf32, strided<[4, 1], offset: 0>>259 to memref<62x?xf32, strided<[?, 1], offset: ?>>260 return %1 : memref<62x?xf32, strided<[?, 1], offset: ?>>261}262 263// -----264 265// CHECK-LABEL: func @subview_leading_operands(266// CHECK: %[[MEM:.*]]: memref<{{.*}}>,267func.func @subview_leading_operands(%0 : memref<5x3xf32>, %1: memref<5x?xf32>) -> memref<3x3xf32, strided<[3, 1], offset: 6>> {268 // CHECK: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]269 // Alloc ptr270 // CHECK: %[[BASE:.*]] = llvm.extractvalue %[[MEMREF]][0] : !llvm.struct<(ptr, ptr, i64271 // Aligned ptr272 // CHECK: %[[BASE_ALIGNED:.*]] = llvm.extractvalue %[[MEMREF]][1] : !llvm.struct<(ptr, ptr, i64273 // CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>274 // CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BASE]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>275 // CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[BASE_ALIGNED]], %[[DESC0]][1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>276 // Offset277 // CHECK: %[[CST_OFF:.*]] = llvm.mlir.constant(6 : index) : i64278 // CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[CST_OFF]], %[[DESC1]][2] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>279 // Sizes and strides @rank 0: both static extracted from type.280 // CHECK: %[[C3:.*]] = llvm.mlir.constant(3 : index) : i64281 // CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[C3]], %[[DESC2]][3, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>282 // CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[C3]], %[[DESC3]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>283 // Sizes and strides @rank 1: both static.284 // CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[C3]], %[[DESC4]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>285 // CHECK: %[[CST_STRIDE1:.*]] = llvm.mlir.constant(1 : index) : i64286 // CHECK: %[[DESC6:.*]] = llvm.insertvalue %[[CST_STRIDE1]], %[[DESC5]][4, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>287 %2 = memref.subview %0[2, 0][3, 3][1, 1]: memref<5x3xf32> to memref<3x3xf32, strided<[3, 1], offset: 6>>288 289 return %2 : memref<3x3xf32, strided<[3, 1], offset: 6>>290}291 292// -----293 294// CHECK-LABEL: func @subview_leading_operands_dynamic(295// CHECK: %[[MEM:[a-zA-Z0-9]*]]: memref296func.func @subview_leading_operands_dynamic(%0 : memref<5x?xf32>) -> memref<3x?xf32, strided<[?, 1], offset: ?>> {297 // CHECK: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]298 // CHECK: %[[SIZE1:.*]] = llvm.extractvalue %[[MEMREF]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>299 // CHECK: %[[BASE:.*]] = llvm.extractvalue %[[MEMREF]][0] : !llvm.struct<(ptr, ptr, i64300 // CHECK: %[[BASE_ALIGNED:.*]] = llvm.extractvalue %[[MEMREF]][1] : !llvm.struct<(ptr, ptr, i64301 // Extract strides302 // CHECK: %[[STRIDE0:.*]] = llvm.extractvalue %[[MEMREF]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>303 // Compute and insert offset from 2 + dynamic value.304 // CHECK: %[[CST_OFF0:.*]] = llvm.mlir.constant(2 : index) : i64305 // CHECK: %[[OFF0:.*]] = llvm.mul %[[STRIDE0]], %[[CST_OFF0]] overflow<nsw> : i64306 // CHECK: %[[TMP:.*]] = builtin.unrealized_conversion_cast %[[OFF0]] : i64 to index307 // CHECK: %[[OFF0:.*]] = builtin.unrealized_conversion_cast %[[TMP]] : index to i64308 // CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>309 // Alloc ptr310 // CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BASE]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>311 // Aligned ptr312 // CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[BASE_ALIGNED]], %[[DESC0]][1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>313 // CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[OFF0]], %[[DESC1]][2] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>314 // Sizes and strides @rank 0: both static.315 // CHECK: %[[CST_SIZE0:.*]] = llvm.mlir.constant(3 : index) : i64316 // CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[CST_SIZE0]], %[[DESC2]][3, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>317 // CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[STRIDE0]], %[[DESC3]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>318 // Sizes and strides @rank 1: static stride 1, dynamic size unchanged from source memref.319 // CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[SIZE1]], %[[DESC4]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>320 // CHECK: %[[CST_STRIDE1:.*]] = llvm.mlir.constant(1 : index) : i64321 // CHECK: %[[DESC6:.*]] = llvm.insertvalue %[[CST_STRIDE1]], %[[DESC5]][4, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>322 323 %c0 = arith.constant 1 : index324 %d0 = memref.dim %0, %c0 : memref<5x?xf32>325 %1 = memref.subview %0[2, 0][3, %d0][1, 1]: memref<5x?xf32> to memref<3x?xf32, strided<[?, 1], offset: ?>>326 return %1 : memref<3x?xf32, strided<[?, 1], offset: ?>>327}328 329// -----330 331// CHECK-LABEL: func @subview_rank_reducing_leading_operands(332// CHECK: %[[MEM:.*]]: memref333func.func @subview_rank_reducing_leading_operands(%0 : memref<5x3xf32>) -> memref<3xf32, strided<[1], offset: 3>> {334 // CHECK: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]335 // CHECK: %[[BASE:.*]] = llvm.extractvalue %[[MEMREF]][0] : !llvm.struct<(ptr, ptr, i64336 // CHECK: %[[BASE_ALIGNED:.*]] = llvm.extractvalue %[[MEMREF]][1] : !llvm.struct<(ptr, ptr, i64337 // CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>338 // Alloc ptr339 // CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BASE]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>340 // Aligned ptr341 // CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[BASE_ALIGNED]], %[[DESC0]][1] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>342 // CHECK: %[[C3:.*]] = llvm.mlir.constant(3 : index) : i64343 // CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[C3]], %[[DESC1]][2] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>344 // Sizes and strides @rank 0: both static.345 // CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[C3]], %[[DESC2]][3, 0] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>346 // CHECK: %[[CST_STRIDE0:.*]] = llvm.mlir.constant(1 : index) : i64347 // CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[CST_STRIDE0]], %[[DESC3]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>348 349 %1 = memref.subview %0[1, 0][1, 3][1, 1]: memref<5x3xf32> to memref<3xf32, strided<[1], offset: 3>>350 351 return %1 : memref<3xf32, strided<[1], offset: 3>>352}353 354// -----355 356// CHECK-LABEL: func @subview_negative_stride357// CHECK-SAME: (%[[MEM:.*]]: memref<7xf32>)358func.func @subview_negative_stride(%arg0 : memref<7xf32>) -> memref<7xf32, strided<[-1], offset: 6>> {359 // CHECK: %[[MEMREF:.*]] = builtin.unrealized_conversion_cast %[[MEM]]360 // CHECK: %[[BASE:.*]] = llvm.extractvalue %[[MEMREF]][0] : !llvm.struct<(ptr, ptr, i64361 // CHECK: %[[BASE_ALIGNED:.*]] = llvm.extractvalue %[[MEMREF]][1] : !llvm.struct<(ptr, ptr, i64362 // CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>363 // CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BASE]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>364 // CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[BASE_ALIGNED]], %[[DESC0]][1] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>365 // CHECK: %[[CST_OFF0:.*]] = llvm.mlir.constant(6 : index) : i64366 // CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[CST_OFF0]], %[[DESC1]][2] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>367 // CHECK: %[[CST_SIZE0:.*]] = llvm.mlir.constant(7 : index) : i64368 // CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[CST_SIZE0]], %[[DESC2]][3, 0] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>369 // CHECK: %[[CST_STRIDE0:.*]] = llvm.mlir.constant(-1 : index) : i64370 // CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[CST_STRIDE0]], %[[DESC3]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>371 // CHECK: %[[RES:.*]] = builtin.unrealized_conversion_cast %[[DESC4]] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)> to memref<7xf32, strided<[-1], offset: 6>>372 // CHECK: return %[[RES]] : memref<7xf32, strided<[-1], offset: 6>>373 374 %0 = memref.subview %arg0[6] [7] [-1] : memref<7xf32> to memref<7xf32, strided<[-1], offset: 6>>375 return %0 : memref<7xf32, strided<[-1], offset: 6>>376}377 378// -----379 380func.func @collapse_shape_static(%arg0: memref<1x3x4x1x5xf32>) -> memref<3x4x5xf32> {381 %0 = memref.collapse_shape %arg0 [[0, 1], [2], [3, 4]] :382 memref<1x3x4x1x5xf32> into memref<3x4x5xf32>383 return %0 : memref<3x4x5xf32>384}385// CHECK-LABEL: func @collapse_shape_static386// CHECK-SAME: %[[ARG:.*]]: memref<1x3x4x1x5xf32>) -> memref<3x4x5xf32> {387// CHECK: %[[MEM:.*]] = builtin.unrealized_conversion_cast %[[ARG]] : memref<1x3x4x1x5xf32> to !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)>388// CHECK: %[[BASE_BUFFER:.*]] = llvm.extractvalue %[[MEM]][0] : !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)>389// CHECK: %[[ALIGNED_BUFFER:.*]] = llvm.extractvalue %[[MEM]][1] : !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)>390// CHECK: %[[C0:.*]] = llvm.mlir.constant(0 : index) : i64391// CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>392// CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BASE_BUFFER]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>393// CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[ALIGNED_BUFFER]], %[[DESC0]][1] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>394// CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[C0]], %[[DESC1]][2] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>395// CHECK: %[[C3:.*]] = llvm.mlir.constant(3 : index) : i64396// CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[C3]], %[[DESC2]][3, 0] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>397// CHECK: %[[C20:.*]] = llvm.mlir.constant(20 : index) : i64398// CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[C20]], %[[DESC3]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>399// CHECK: %[[C4:.*]] = llvm.mlir.constant(4 : index) : i64400// CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[C4]], %[[DESC4]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>401// CHECK: %[[C5:.*]] = llvm.mlir.constant(5 : index) : i64402// CHECK: %[[DESC6:.*]] = llvm.insertvalue %[[C5]], %[[DESC5]][4, 1] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>403// CHECK: %[[DESC7:.*]] = llvm.insertvalue %[[C5]], %[[DESC6]][3, 2] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>404// CHECK: %[[C1:.*]] = llvm.mlir.constant(1 : index) : i64405// CHECK: %[[DESC8:.*]] = llvm.insertvalue %[[C1]], %[[DESC7]][4, 2] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>406// CHECK: %[[RES:.*]] = builtin.unrealized_conversion_cast %[[DESC8]] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)> to memref<3x4x5xf32>407// CHECK: return %[[RES]] : memref<3x4x5xf32>408// CHECK: }409 410// -----411 412func.func @collapse_shape_dynamic_with_non_identity_layout(413 %arg0 : memref<4x?x?xf32, strided<[?, 4, 1], offset: ?>>) ->414 memref<4x?xf32, strided<[?, ?], offset: ?>> {415 %0 = memref.collapse_shape %arg0 [[0], [1, 2]]:416 memref<4x?x?xf32, strided<[?, 4, 1], offset: ?>> into417 memref<4x?xf32, strided<[?, ?], offset: ?>>418 return %0 : memref<4x?xf32, strided<[?, ?], offset: ?>>419}420// CHECK-LABEL: func.func @collapse_shape_dynamic_with_non_identity_layout(421// CHECK-SAME: %[[ARG:.*]]: memref<4x?x?xf32, strided<[?, 4, 1], offset: ?>>) -> memref<4x?xf32, strided<[?, ?], offset: ?>> {422// CHECK: %[[MEM:.*]] = builtin.unrealized_conversion_cast %[[ARG]] : memref<4x?x?xf32, strided<[?, 4, 1], offset: ?>> to !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>423// CHECK: %[[BASE_BUFFER:.*]] = llvm.extractvalue %[[MEM]][0] : !llvm.struct<(ptr, ptr, i64,424// CHECK: %[[ALIGNED_BUFFER:.*]] = llvm.extractvalue %[[MEM]][1] : !llvm.struct<(ptr, ptr, i64,425// CHECK: %[[OFFSET:.*]] = llvm.extractvalue %[[MEM]][2] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>426// CHECK: %[[SIZE1:.*]] = llvm.extractvalue %[[MEM]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>427// CHECK: %[[SIZE2:.*]] = llvm.extractvalue %[[MEM]][3, 2] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>428// CHECK: %[[STRIDE0:.*]] = llvm.extractvalue %[[MEM]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>429// CHECK: %[[FINAL_SIZE1:.*]] = llvm.mul %[[SIZE1]], %[[SIZE2]] overflow<nsw> : i64430// CHECK: %[[SIZE1_TO_IDX:.*]] = builtin.unrealized_conversion_cast %[[FINAL_SIZE1]] : i64 to index431// CHECK: %[[FINAL_SIZE1:.*]] = builtin.unrealized_conversion_cast %[[SIZE1_TO_IDX]] : index to i64432// CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>433// CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BASE_BUFFER]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>434// CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[ALIGNED_BUFFER]], %[[DESC0]][1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>435// CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[OFFSET]], %[[DESC1]][2] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>436// CHECK: %[[C4:.*]] = llvm.mlir.constant(4 : index) : i64437// CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[C4]], %[[DESC2]][3, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>438// CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[STRIDE0]], %[[DESC3]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>439// CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[FINAL_SIZE1]], %[[DESC4]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>440// CHECK: %[[C1:.*]] = llvm.mlir.constant(1 : index) : i64441// CHECK: %[[DESC6:.*]] = llvm.insertvalue %[[C1]], %[[DESC5]][4, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>442// CHECK: %[[RES:.*]] = builtin.unrealized_conversion_cast %[[DESC6]] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)> to memref<4x?xf32, strided<[?, ?], offset: ?>>443// CHECK: return %[[RES]] : memref<4x?xf32, strided<[?, ?], offset: ?>>444// CHECK: }445// CHECK32-LABEL: func.func @collapse_shape_dynamic_with_non_identity_layout(446// CHECK32-SAME: %[[ARG:.*]]: memref<4x?x?xf32, strided<[?, 4, 1], offset: ?>>) -> memref<4x?xf32, strided<[?, ?], offset: ?>> {447// CHECK32: %[[MEM:.*]] = builtin.unrealized_conversion_cast %[[ARG]] : memref<4x?x?xf32, strided<[?, 4, 1], offset: ?>> to !llvm.struct<(ptr, ptr, i32, array<3 x i32>, array<3 x i32>)>448// CHECK32: %[[BASE_BUFFER:.*]] = llvm.extractvalue %[[MEM]][0] : !llvm.struct<(ptr, ptr, i32,449// CHECK32: %[[ALIGNED_BUFFER:.*]] = llvm.extractvalue %[[MEM]][1] : !llvm.struct<(ptr, ptr, i32,450// CHECK32: %[[OFFSET:.*]] = llvm.extractvalue %[[MEM]][2] : !llvm.struct<(ptr, ptr, i32, array<3 x i32>, array<3 x i32>)>451// CHECK32: %[[SIZE1:.*]] = llvm.extractvalue %[[MEM]][3, 1] : !llvm.struct<(ptr, ptr, i32, array<3 x i32>, array<3 x i32>)>452// CHECK32: %[[SIZE2:.*]] = llvm.extractvalue %[[MEM]][3, 2] : !llvm.struct<(ptr, ptr, i32, array<3 x i32>, array<3 x i32>)>453// CHECK32: %[[STRIDE0:.*]] = llvm.extractvalue %[[MEM]][4, 0] : !llvm.struct<(ptr, ptr, i32, array<3 x i32>, array<3 x i32>)>454// CHECK32: %[[FINAL_SIZE1_I32:.*]] = llvm.mul %[[SIZE1]], %[[SIZE2]] overflow<nsw> : i32455// CHECK32: %[[SIZE1_TO_IDX:.*]] = builtin.unrealized_conversion_cast %[[FINAL_SIZE1_I32]] : i32 to index456// CHECK32: %[[FINAL_SIZE1_CAST:.*]] = builtin.unrealized_conversion_cast %[[SIZE1_TO_IDX]] : index to i32457// CHECK32: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i32, array<2 x i32>, array<2 x i32>)>458// CHECK32: %[[DESC0:.*]] = llvm.insertvalue %[[BASE_BUFFER]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i32, array<2 x i32>, array<2 x i32>)>459// CHECK32: %[[DESC1:.*]] = llvm.insertvalue %[[ALIGNED_BUFFER]], %[[DESC0]][1] : !llvm.struct<(ptr, ptr, i32, array<2 x i32>, array<2 x i32>)>460// CHECK32: %[[DESC2:.*]] = llvm.insertvalue %[[OFFSET]], %[[DESC1]][2] : !llvm.struct<(ptr, ptr, i32, array<2 x i32>, array<2 x i32>)>461// CHECK32: %[[C4_I32:.*]] = llvm.mlir.constant(4 : index) : i32462// CHECK32: %[[DESC3:.*]] = llvm.insertvalue %[[C4_I32]], %[[DESC2]][3, 0] : !llvm.struct<(ptr, ptr, i32, array<2 x i32>, array<2 x i32>)>463// CHECK32: %[[DESC4:.*]] = llvm.insertvalue %[[STRIDE0]], %[[DESC3]][4, 0] : !llvm.struct<(ptr, ptr, i32, array<2 x i32>, array<2 x i32>)>464// CHECK32: %[[DESC5:.*]] = llvm.insertvalue %[[FINAL_SIZE1_CAST]], %[[DESC4]][3, 1] : !llvm.struct<(ptr, ptr, i32, array<2 x i32>, array<2 x i32>)>465// CHECK32: %[[C1_I32:.*]] = llvm.mlir.constant(1 : index) : i32466// CHECK32: %[[DESC6:.*]] = llvm.insertvalue %[[C1_I32]], %[[DESC5]][4, 1] : !llvm.struct<(ptr, ptr, i32, array<2 x i32>, array<2 x i32>)>467// CHECK32: %[[RES:.*]] = builtin.unrealized_conversion_cast %[[DESC6]] : !llvm.struct<(ptr, ptr, i32, array<2 x i32>, array<2 x i32>)> to memref<4x?xf32, strided<[?, ?], offset: ?>>468// CHECK32: return %[[RES]] : memref<4x?xf32, strided<[?, ?], offset: ?>>469// CHECK32: }470 471// -----472 473 474func.func @expand_shape_static(%arg0: memref<3x4x5xf32>) -> memref<1x3x4x1x5xf32> {475 // Reshapes that expand a contiguous tensor with some 1's.476 %0 = memref.expand_shape %arg0 [[0, 1], [2], [3, 4]] output_shape [1, 3, 4, 1, 5]477 : memref<3x4x5xf32> into memref<1x3x4x1x5xf32>478 return %0 : memref<1x3x4x1x5xf32>479}480// CHECK-LABEL: func @expand_shape_static481// CHECK-SAME: %[[ARG:.*]]: memref<3x4x5xf32>) -> memref<1x3x4x1x5xf32> {482// CHECK: %[[MEM:.*]] = builtin.unrealized_conversion_cast %[[ARG]] : memref<3x4x5xf32> to !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>483// CHECK: %[[BASE_BUFFER:.*]] = llvm.extractvalue %[[MEM]][0] : !llvm.struct<(ptr, ptr, i64,484// CHECK: %[[ALIGNED_BUFFER:.*]] = llvm.extractvalue %[[MEM]][1] : !llvm.struct<(ptr, ptr, i64,485// CHECK: %[[C0:.*]] = llvm.mlir.constant(0 : index) : i64486// CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)>487// CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BASE_BUFFER]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)>488// CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[ALIGNED_BUFFER]], %[[DESC0]][1] : !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)>489// CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[C0]], %[[DESC1]][2] : !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)>490// CHECK: %[[C1:.*]] = llvm.mlir.constant(1 : index) : i64491// CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[C1]], %[[DESC2]][3, 0] : !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)>492// CHECK: %[[C60:.*]] = llvm.mlir.constant(60 : index) : i64493// CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[C60]], %[[DESC3]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)>494// CHECK: %[[C3:.*]] = llvm.mlir.constant(3 : index) : i64495// CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[C3]], %[[DESC4]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)>496// CHECK: %[[C20:.*]] = llvm.mlir.constant(20 : index) : i64497// CHECK: %[[DESC6:.*]] = llvm.insertvalue %[[C20]], %[[DESC5]][4, 1] : !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)>498// CHECK: %[[C4:.*]] = llvm.mlir.constant(4 : index) : i64499// CHECK: %[[DESC7:.*]] = llvm.insertvalue %[[C4]], %[[DESC6]][3, 2] : !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)>500// CHECK: %[[C5:.*]] = llvm.mlir.constant(5 : index) : i64501// CHECK: %[[DESC8:.*]] = llvm.insertvalue %[[C5]], %[[DESC7]][4, 2] : !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)>502// CHECK: %[[DESC9:.*]] = llvm.insertvalue %[[C1]], %[[DESC8]][3, 3] : !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)>503// CHECK: %[[DESC10:.*]] = llvm.insertvalue %[[C5]], %[[DESC9]][4, 3] : !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)>504// CHECK: %[[DESC11:.*]] = llvm.insertvalue %[[C5]], %[[DESC10]][3, 4] : !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)>505// CHECK: %[[DESC12:.*]] = llvm.insertvalue %[[C1]], %[[DESC11]][4, 4] : !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)>506// CHECK: %[[RES:.*]] = builtin.unrealized_conversion_cast %[[DESC12]] : !llvm.struct<(ptr, ptr, i64, array<5 x i64>, array<5 x i64>)> to memref<1x3x4x1x5xf32>507// CHECK: return %[[RES]] : memref<1x3x4x1x5xf32>508// CHECK: }509 510// -----511 512func.func @collapse_shape_fold_zero_dim(%arg0 : memref<1x1xf32>) -> memref<f32> {513 %0 = memref.collapse_shape %arg0 [] : memref<1x1xf32> into memref<f32>514 return %0 : memref<f32>515}516// CHECK-LABEL: func.func @collapse_shape_fold_zero_dim(517// CHECK-SAME: %[[ARG:.*]]: memref<1x1xf32>) -> memref<f32> {518// CHECK: %[[MEM:.*]] = builtin.unrealized_conversion_cast %[[ARG]] : memref<1x1xf32> to !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>519// CHECK: %[[BASE_BUFFER:.*]] = llvm.extractvalue %[[MEM]][0] : !llvm.struct<(ptr, ptr, i64,520// CHECK: %[[ALIGNED_BUFFER:.*]] = llvm.extractvalue %[[MEM]][1] : !llvm.struct<(ptr, ptr, i64,521// CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i64)>522// CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BASE_BUFFER]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i64)>523// CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[ALIGNED_BUFFER]], %[[DESC0]][1] : !llvm.struct<(ptr, ptr, i64)>524// CHECK: %[[C0:.*]] = llvm.mlir.constant(0 : index) : i64525// CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[C0]], %[[DESC1]][2] : !llvm.struct<(ptr, ptr, i64)>526// CHECK: %[[RES:.*]] = builtin.unrealized_conversion_cast %[[DESC2]] : !llvm.struct<(ptr, ptr, i64)> to memref<f32>527// CHECK: return %[[RES]] : memref<f32>528// CHECK: }529 530// -----531 532func.func @expand_shape_zero_dim(%arg0 : memref<f32>) -> memref<1x1xf32> {533 %0 = memref.expand_shape %arg0 [] output_shape [1, 1] : memref<f32> into memref<1x1xf32>534 return %0 : memref<1x1xf32>535}536 537// CHECK-LABEL: func.func @expand_shape_zero_dim(538// CHECK-SAME: %[[ARG:.*]]: memref<f32>) -> memref<1x1xf32> {539// CHECK: %[[MEM:.*]] = builtin.unrealized_conversion_cast %[[ARG]] : memref<f32> to !llvm.struct<(ptr, ptr, i64)>540// CHECK: %[[BASE_BUFFER:.*]] = llvm.extractvalue %[[MEM]][0] : !llvm.struct<(ptr, ptr, i64)>541// CHECK: %[[ALIGNED_BUFFER:.*]] = llvm.extractvalue %[[MEM]][1] : !llvm.struct<(ptr, ptr, i64)>542// CHECK: %[[C0:.*]] = llvm.mlir.constant(0 : index) : i64543// CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>544// CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BASE_BUFFER]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>545// CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[ALIGNED_BUFFER]], %[[DESC0]][1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>546// CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[C0]], %[[DESC1]][2] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>547// CHECK: %[[C1:.*]] = llvm.mlir.constant(1 : index) : i64548// CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[C1]], %[[DESC2]][3, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>549// CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[C1]], %[[DESC3]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>550// CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[C1]], %[[DESC4]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>551// CHECK: %[[DESC6:.*]] = llvm.insertvalue %[[C1]], %[[DESC5]][4, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>552// CHECK: %[[RES:.*]] = builtin.unrealized_conversion_cast %[[DESC6]] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)> to memref<1x1xf32>553// CHECK: return %[[RES]] : memref<1x1xf32>554// CHECK: }555 556// -----557 558func.func @collapse_shape_dynamic(%arg0 : memref<1x2x?xf32>) -> memref<1x?xf32> {559 %0 = memref.collapse_shape %arg0 [[0], [1, 2]]: memref<1x2x?xf32> into memref<1x?xf32>560 return %0 : memref<1x?xf32>561}562 563// CHECK-LABEL: func.func @collapse_shape_dynamic(564// CHECK-SAME: %[[ARG:.*]]: memref<1x2x?xf32>) -> memref<1x?xf32> {565// CHECK: %[[MEM:.*]] = builtin.unrealized_conversion_cast %[[ARG]] : memref<1x2x?xf32> to !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>566// CHECK: %[[BASE_BUFFER:.*]] = llvm.extractvalue %[[MEM]][0] : !llvm.struct<(ptr, ptr, i64,567// CHECK: %[[ALIGNED_BUFFER:.*]] = llvm.extractvalue %[[MEM]][1] : !llvm.struct<(ptr, ptr, i64,568// CHECK: %[[C0:.*]] = llvm.mlir.constant(0 : index) : i64569// CHECK: %[[SIZE2:.*]] = llvm.extractvalue %[[MEM]][3, 2] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>570// CHECK: %[[STRIDE0:.*]] = llvm.extractvalue %[[MEM]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>571// CHECK: %[[C2:.*]] = llvm.mlir.constant(2 : index) : i64572// CHECK: %[[FINAL_SIZE1:.*]] = llvm.mul %[[SIZE2]], %[[C2]] overflow<nsw> : i64573// CHECK: %[[SIZE1_TO_IDX:.*]] = builtin.unrealized_conversion_cast %[[FINAL_SIZE1]] : i64 to index574// CHECK: %[[FINAL_SIZE1:.*]] = builtin.unrealized_conversion_cast %[[SIZE1_TO_IDX]] : index to i64575// CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>576// CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BASE_BUFFER]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>577// CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[ALIGNED_BUFFER]], %[[DESC0]][1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>578// CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[C0]], %[[DESC1]][2] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>579// CHECK: %[[C1:.*]] = llvm.mlir.constant(1 : index) : i64580// CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[C1]], %[[DESC2]][3, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>581// CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[STRIDE0]], %[[DESC3]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>582// CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[FINAL_SIZE1]], %[[DESC4]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>583// CHECK: %[[DESC6:.*]] = llvm.insertvalue %[[C1]], %[[DESC5]][4, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>584// CHECK: %[[RES:.*]] = builtin.unrealized_conversion_cast %[[DESC6]] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)> to memref<1x?xf32>585// CHECK: return %[[RES]] : memref<1x?xf32>586// CHECK: }587 588// -----589 590func.func @expand_shape_dynamic(%arg0 : memref<1x?xf32>, %sz0: index) -> memref<1x2x?xf32> {591 %0 = memref.expand_shape %arg0 [[0], [1, 2]] output_shape [1, 2, %sz0]: memref<1x?xf32> into memref<1x2x?xf32>592 return %0 : memref<1x2x?xf32>593}594 595// CHECK-LABEL: func.func @expand_shape_dynamic(596// CHECK-SAME: %[[ARG:.*]]: memref<1x?xf32>, %[[SZ0:.*]]: index) -> memref<1x2x?xf32> {597// CHECK: %[[MEM:.*]] = builtin.unrealized_conversion_cast %[[ARG]] : memref<1x?xf32> to !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>598// CHECK: %[[BASE_BUFFER:.*]] = llvm.extractvalue %[[MEM]][0] : !llvm.struct<(ptr, ptr, i64,599// CHECK: %[[ALIGNED_BUFFER:.*]] = llvm.extractvalue %[[MEM]][1] : !llvm.struct<(ptr, ptr, i64,600// CHECK: %[[C0:.*]] = llvm.mlir.constant(0 : index) : i64601// CHECK: %[[SIZE1:.*]] = llvm.extractvalue %[[MEM]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>602// CHECK: %[[STRIDE0:.*]] = llvm.extractvalue %[[MEM]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>603// CHECK: %[[C2:.*]] = llvm.mlir.constant(2 : index) : i64604// CHECK: %[[CMINUS1:.*]] = llvm.mlir.constant(-1 : index) : i64605// CHECK: %[[IS_NEGATIVE_SIZE1:.*]] = llvm.icmp "slt" %[[SIZE1]], %[[C0]] : i64606// CHECK: %[[ABS_SIZE1_MINUS_1:.*]] = llvm.sub %[[CMINUS1]], %[[SIZE1]] : i64607// CHECK: %[[ADJ_SIZE1:.*]] = llvm.select %[[IS_NEGATIVE_SIZE1]], %[[ABS_SIZE1_MINUS_1]], %[[SIZE1]] : i1, i64608// CHECK: %[[SIZE2:.*]] = llvm.sdiv %[[ADJ_SIZE1]], %[[C2]] : i64609// CHECK: %[[NEGATIVE_SIZE2:.*]] = llvm.sub %[[CMINUS1]], %[[SIZE2]] : i64610// CHECK: %[[FINAL_SIZE2:.*]] = llvm.select %[[IS_NEGATIVE_SIZE1]], %[[NEGATIVE_SIZE2]], %[[SIZE2]] : i1, i64611// CHECK: %[[SIZE2_TO_IDX:.*]] = builtin.unrealized_conversion_cast %[[FINAL_SIZE2]] : i64 to index612// CHECK: %[[FINAL_SIZE2:.*]] = builtin.unrealized_conversion_cast %[[SIZE2_TO_IDX]] : index to i64613// CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>614// CHECK: %[[DESC0:.*]] = llvm.insertvalue %[[BASE_BUFFER]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>615// CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[ALIGNED_BUFFER]], %[[DESC0]][1] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>616// CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[C0]], %[[DESC1]][2] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>617// CHECK: %[[C1:.*]] = llvm.mlir.constant(1 : index) : i64618// CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[C1]], %[[DESC2]][3, 0] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>619// CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[STRIDE0]], %[[DESC3]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>620// CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[C2]], %[[DESC4]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>621// In this example stride1 and size2 are the same.622// Hence with CSE, we get the same SSA value.623// CHECK: %[[DESC6:.*]] = llvm.insertvalue %[[FINAL_SIZE2]], %[[DESC5]][4, 1] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>624// CHECK: %[[DESC7:.*]] = llvm.insertvalue %[[FINAL_SIZE2]], %[[DESC6]][3, 2] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>625// CHECK: %[[DESC8:.*]] = llvm.insertvalue %[[C1]], %[[DESC7]][4, 2] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>626// CHECK: %[[RES:.*]] = builtin.unrealized_conversion_cast %[[DESC8]] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)> to memref<1x2x?xf32>627// CHECK: return %[[RES]] : memref<1x2x?xf32>628// CHECK: }629 630// -----631 632func.func @expand_shape_dynamic_with_non_identity_layout(633 %arg0 : memref<1x?xf32, strided<[?, ?], offset: ?>>, %sz0: index) ->634 memref<1x2x?xf32, strided<[?, ?, ?], offset: ?>> {635 %0 = memref.expand_shape %arg0 [[0], [1, 2]] output_shape [1, 2, %sz0] :636 memref<1x?xf32, strided<[?, ?], offset: ?>> into637 memref<1x2x?xf32, strided<[?, ?, ?], offset: ?>>638 return %0 : memref<1x2x?xf32, strided<[?, ?, ?], offset: ?>>639}640// CHECK-LABEL: func.func @expand_shape_dynamic_with_non_identity_layout(641// CHECK-SAME: %[[ARG:.*]]: memref<1x?xf32, strided<[?, ?], offset: ?>>, %[[SZ0:.*]]: index) -> memref<1x2x?xf32, strided<[?, ?, ?], offset: ?>> {642// CHECK: %[[MEM:.*]] = builtin.unrealized_conversion_cast %[[ARG]] : memref<1x?xf32, strided<[?, ?], offset: ?>> to !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>643// CHECK: %[[BASE_BUFFER:.*]] = llvm.extractvalue %[[MEM]][0] : !llvm.struct<(ptr, ptr, i64,644// CHECK: %[[ALIGNED_BUFFER:.*]] = llvm.extractvalue %[[MEM]][1] : !llvm.struct<(ptr, ptr, i64,645// CHECK: %[[C0:.*]] = llvm.mlir.constant(0 : index) : i64646// CHECK: %[[OFFSET:.*]] = llvm.extractvalue %[[MEM]][2] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>647// CHECK: %[[SIZE1:.*]] = llvm.extractvalue %[[MEM]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>648// CHECK: %[[STRIDE0:.*]] = llvm.extractvalue %[[MEM]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>649// CHECK: %[[STRIDE1:.*]] = llvm.extractvalue %[[MEM]][4, 1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>650// CHECK: %[[C2:.*]] = llvm.mlir.constant(2 : index) : i64651// CHECK: %[[CMINUS1:.*]] = llvm.mlir.constant(-1 : index) : i64652// CHECK: %[[IS_NEGATIVE_SIZE1:.*]] = llvm.icmp "slt" %[[SIZE1]], %[[C0]] : i64653// CHECK: %[[ABS_SIZE1_MINUS_1:.*]] = llvm.sub %[[CMINUS1]], %[[SIZE1]] : i64654// CHECK: %[[ADJ_SIZE1:.*]] = llvm.select %[[IS_NEGATIVE_SIZE1]], %[[ABS_SIZE1_MINUS_1]], %[[SIZE1]] : i1, i64655// CHECK: %[[SIZE2:.*]] = llvm.sdiv %[[ADJ_SIZE1]], %[[C2]] : i64656// CHECK: %[[NEGATIVE_SIZE2:.*]] = llvm.sub %[[CMINUS1]], %[[SIZE2]] : i64657// CHECK: %[[TMP_SIZE2:.*]] = llvm.select %[[IS_NEGATIVE_SIZE1]], %[[NEGATIVE_SIZE2]], %[[SIZE2]] : i1, i64658// CHECK: %[[SIZE2_TO_IDX:.*]] = builtin.unrealized_conversion_cast %[[TMP_SIZE2]] : i64 to index659// CHECK: %[[FINAL_SIZE2:.*]] = builtin.unrealized_conversion_cast %[[SIZE2_TO_IDX]] : index to i64660// CHECK: %[[FINAL_STRIDE1:.*]] = llvm.mul %[[TMP_SIZE2]], %[[STRIDE1]]661// CHECK: %[[STRIDE1_TO_IDX:.*]] = builtin.unrealized_conversion_cast %[[FINAL_STRIDE1]] : i64 to index662// CHECK: %[[FINAL_STRIDE1:.*]] = builtin.unrealized_conversion_cast %[[STRIDE1_TO_IDX]] : index to i64663// CHECK: %[[DESC:.*]] = llvm.mlir.poison : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>664// CHECK: %[[DESC1:.*]] = llvm.insertvalue %[[BASE_BUFFER]], %[[DESC]][0] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>665// CHECK: %[[DESC2:.*]] = llvm.insertvalue %[[ALIGNED_BUFFER]], %[[DESC1]][1] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>666// CHECK: %[[DESC3:.*]] = llvm.insertvalue %[[OFFSET]], %[[DESC2]][2] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>667// CHECK: %[[C1:.*]] = llvm.mlir.constant(1 : index) : i64668// CHECK: %[[DESC4:.*]] = llvm.insertvalue %[[C1]], %[[DESC3]][3, 0] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>669// CHECK: %[[DESC5:.*]] = llvm.insertvalue %[[STRIDE0]], %[[DESC4]][4, 0] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>670// CHECK: %[[DESC6:.*]] = llvm.insertvalue %[[C2]], %[[DESC5]][3, 1] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>671// CHECK: %[[DESC7:.*]] = llvm.insertvalue %[[FINAL_STRIDE1]], %[[DESC6]][4, 1] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>672// CHECK: %[[DESC8:.*]] = llvm.insertvalue %[[FINAL_SIZE2]], %[[DESC7]][3, 2] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>673// CHECK: %[[DESC9:.*]] = llvm.insertvalue %[[STRIDE1]], %[[DESC8]][4, 2] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)>674// CHECK: %[[RES:.*]] = builtin.unrealized_conversion_cast %[[DESC9]] : !llvm.struct<(ptr, ptr, i64, array<3 x i64>, array<3 x i64>)> to memref<1x2x?xf32, strided<[?, ?, ?], offset: ?>>675// CHECK: return %[[RES]] : memref<1x2x?xf32, strided<[?, ?, ?], offset: ?>>676// CHECK: }677 678// -----679 680// CHECK-LABEL: func @collapse_static_shape_with_non_identity_layout681func.func @collapse_static_shape_with_non_identity_layout(%arg: memref<1x1x8x8xf32, strided<[64, 64, 8, 1], offset: ?>>) -> memref<64xf32, strided<[1], offset: ?>> {682// CHECK-NOT: memref.collapse_shape683 %1 = memref.collapse_shape %arg [[0, 1, 2, 3]] : memref<1x1x8x8xf32, strided<[64, 64, 8, 1], offset: ?>> into memref<64xf32, strided<[1], offset: ?>>684 return %1 : memref<64xf32, strided<[1], offset: ?>>685}686 687// -----688 689// Check that the address of %arg0 appears with the same value690// in both the llvm.assume and as base of the load.691// This is to make sure that later CSEs and alignment propagation692// will be able to do their job easily.693 694// CHECK-LABEL: func @load_and_assume(695// CHECK-SAME: %[[ARG0:.*]]: memref<?x?xf32, strided<[?, ?], offset: ?>>,696// CHECK: %[[DESC:.*]] = builtin.unrealized_conversion_cast %[[ARG0]] : memref<?x?xf32, strided<[?, ?], offset: ?>> to !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>697// CHECK: %[[ALIGNED_PTR:.*]] = llvm.extractvalue %[[DESC]][1] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>698// CHECK: %[[OFFSET:.*]] = llvm.extractvalue %[[DESC]][2] : !llvm.struct<(ptr, ptr, i64, array<2 x i64>, array<2 x i64>)>699// CHECK: %[[BUFF_ADDR:.*]] = llvm.getelementptr %[[ALIGNED_PTR]][%[[OFFSET]]] : (!llvm.ptr, i64) -> !llvm.ptr, f32700// CHECK: llvm.intr.assume %{{.*}} ["align"(%[[BUFF_ADDR]], %{{.*}} : !llvm.ptr, i64)] : i1701// CHECK: %[[LD_ADDR:.*]] = llvm.getelementptr inbounds|nuw %[[BUFF_ADDR]][%{{.*}}] : (!llvm.ptr, i64) -> !llvm.ptr, f32702// CHECK: %[[VAL:.*]] = llvm.load %[[LD_ADDR]] : !llvm.ptr -> f32703// CHECK: return %[[VAL]] : f32704func.func @load_and_assume(705 %arg0: memref<?x?xf32, strided<[?, ?], offset: ?>>,706 %i0: index, %i1: index)707 -> f32 {708 %arg0_align = memref.assume_alignment %arg0, 16 : memref<?x?xf32, strided<[?, ?], offset: ?>>709 %2 = memref.load %arg0_align[%i0, %i1] : memref<?x?xf32, strided<[?, ?], offset: ?>>710 func.return %2 : f32711}712