[Mlir-commits] [mlir] [mlir][ArmSVE] Add `-arm-sve-legalize-vector-storage` pass (PR #68794)

[Andrzej Warzyński]

================
@@ -0,0 +1,121 @@
+// RUN: mlir-opt %s -convert-vector-to-scf -arm-sve-legalize-vector-storage -convert-vector-to-llvm="enable-arm-sve" -test-lower-to-llvm | \
+// RUN: %mcr_aarch64_cmd -e=entry -entry-point-result=void --march=aarch64 --mattr="+sve" -shared-libs=%mlir_lib_dir/libmlir_c_runner_utils%shlibext | \
+// RUN: FileCheck %s
+
+/// This tests basic functionality of arrays of scalable vectors, which in MLIR
+/// are vectors with a single trailing scalable dimension. This test requires
+/// the -arm-sve-legalize-vector-storage pass to ensure the loads/stores done
+/// here are be legal for the LLVM backend.
+
+func.func @read_and_print_2d_vector(%memref: memref<3x?xf32>)  {
+  %cst = arith.constant 0.000000e+00 : f32
+  %c0 = arith.constant 0 : index
+  %c1 = arith.constant 1 : index
+  %c2 = arith.constant 2 : index
+  %dim = memref.dim %memref, %c1 : memref<3x?xf32>
+  %mask = vector.create_mask %c2, %dim : vector<3x[8]xi1>
+  %vector = vector.transfer_read %memref[%c0,%c0], %cst, %mask {in_bounds = [true, true]} : memref<3x?xf32>, vector<3x[8]xf32>
+
+  /// TODO: Support vector.print for arrays of scalable vectors.
+  %row0 = vector.extract %vector[0] : vector<[8]xf32> from vector<3x[8]xf32>
+  %row1 = vector.extract %vector[1] : vector<[8]xf32> from vector<3x[8]xf32>
+  %row2 = vector.extract %vector[2] : vector<[8]xf32> from vector<3x[8]xf32>
+
+  /// Print each of the vectors. This only checks the first eight elements (which
+  /// works for all vscale >= 1).
----------------
banach-space wrote:

[nit] I prefer to highlight the assumption first: "vscale is >= 1, hence at least 8 elements will be printed". 

https://github.com/llvm/llvm-project/pull/68794