[mlir][linalg] Fix empty outer dim case for packing reshape op (llvm#96732)

This PR fixes the issue reported in
[comment](llvm#93529 (comment)).

Author: yifeizh2

mlir/lib/Dialect/Linalg/Transforms/DataLayoutPropagation.cpp

@@ -605,7 +605,8 @@ static bool isDimsDivisibleByTileSizes(ArrayRef<int64_t> dimsPos,
 static int64_t applyPermutationAndReindexReassoc(
     SmallVector<ReassociationIndices> &reassocIndices,
     ArrayRef<int64_t> permutation) {
-  applyPermutationToVector<ReassociationIndices>(reassocIndices, permutation);
+  if (!permutation.empty())
+    applyPermutationToVector<ReassociationIndices>(reassocIndices, permutation);
   int64_t nextPos = 0;
   for (ReassociationIndices &indices : reassocIndices) {
     for (auto &index : indices) {

mlir/test/Dialect/Linalg/data-layout-propagation.mlir

@@ -926,6 +926,24 @@ func.func @bubble_up_pack_through_collapse(%1: tensor<?x16x4xf32>, %dim : index)
 
 // -----
 
+func.func @bubble_up_pack_through_collapse_empty_outer_dims_perm(%1: tensor<?x16x4xf32>, %dim : index) -> tensor<?x4x8x1xf32> {
+  %collapsed = tensor.collapse_shape %1 [[0, 1], [2]] : tensor<?x16x4xf32> into tensor<?x4xf32>
+  %2 = tensor.empty(%dim) : tensor<?x4x8x1xf32>
+  %pack = tensor.pack %collapsed inner_dims_pos = [0, 1] inner_tiles = [8, 1] into %2 : tensor<?x4xf32> -> tensor<?x4x8x1xf32>
+  func.return %pack : tensor<?x4x8x1xf32>
+}
+// CHECK-LABEL: func.func @bubble_up_pack_through_collapse_empty_outer_dims_perm
+// CHECK-SAME:      %[[ARG0:[a-zA-Z0-9]+]]
+// CHECK-SAME:      %[[ARG1:[a-zA-Z0-9]+]]
+// CHECK:         %[[C0:.+]] = arith.constant 0 : index
+// CHECK:         %[[DIM:.+]] = tensor.dim %[[ARG0]], %[[C0]] : tensor<?x16x4xf32>
+// CHECK:         %[[EMPTY:.+]] = tensor.empty(%[[DIM]]) : tensor<?x2x4x8x1xf32>
+// CHECK:         %[[PACK:.+]] = tensor.pack %[[ARG0]] inner_dims_pos = [1, 2] inner_tiles = [8, 1] into %[[EMPTY]] : tensor<?x16x4xf32> -> tensor<?x2x4x8x1xf32>
+// CHECK:         %[[COLLAPSED:.+]] = tensor.collapse_shape %[[PACK]] {{\[}}[0, 1], [2], [3], [4]] : tensor<?x2x4x8x1xf32> into tensor<?x4x8x1xf32>
+// CHECK:         return %[[COLLAPSED]] : tensor<?x4x8x1xf32>
+
+// -----
+
 func.func @bubble_up_permuted_pack_through_collapse(%1: tensor<4x192x16x256xf32>) -> tensor<4x32x3072x8x1xf32> {
   %collapsed = tensor.collapse_shape %1 [[0], [1, 2], [3]] : tensor<4x192x16x256xf32> into tensor<4x3072x256xf32>
   %2 = tensor.empty() : tensor<4x32x3072x8x1xf32>
@@ -1269,6 +1287,27 @@ func.func @push_down_unpack_through_expand(%5: tensor<?x32x8x8xf32>, %dim: index
 
 // -----
 
+func.func @push_down_unpack_through_expand_empty_outer_dims_perm(%5: tensor<?x32x8x8xf32>, %dim: index, %sz0: index) -> tensor<?x256x256xf32> {
+  %6 = tensor.empty(%dim) : tensor<?x256xf32>
+  %unpack = tensor.unpack %5 inner_dims_pos = [0, 1] inner_tiles = [8, 8] into %6 : tensor<?x32x8x8xf32> -> tensor<?x256xf32>
+  %expanded = tensor.expand_shape %unpack [[0, 1], [2]] output_shape [%sz0, 256, 256] : tensor<?x256xf32> into tensor<?x256x256xf32>
+  func.return %expanded : tensor<?x256x256xf32>
+}
+// CHECK-LABEL: func.func @push_down_unpack_through_expand_empty_outer_dims_perm
+// CHECK-SAME:      %[[ARG0:[a-zA-Z0-9]+]]
+// CHECK-SAME:      %[[ARG1:[a-zA-Z0-9]+]]
+// CHECK:         %[[C32:.+]] = arith.constant 32 : index
+// CHECK:         %[[C0:.+]] = arith.constant 0 : index
+// CHECK:         %[[DIM0:.+]] = tensor.dim %[[ARG0]], %[[C0]] : tensor<?x32x8x8xf32>
+// CHECK:         %[[SZ0:.+]] = arith.divui %[[DIM0]], %[[C32]] : index
+// CHECK:         %[[EXPANDED:.+]] = tensor.expand_shape %[[ARG0]] {{\[}}[0, 1], [2], [3], [4]] output_shape [%[[SZ0]], 32, 32, 8, 8] : tensor<?x32x8x8xf32> into tensor<?x32x32x8x8xf32>
+// CHECK:         %[[DIM:.+]] = tensor.dim %[[EXPANDED]], %[[C0]] : tensor<?x32x32x8x8xf32>
+// CHECK:         %[[EMPTY:.+]] = tensor.empty(%[[DIM]]) : tensor<?x256x256xf32>
+// CHECK:         %[[UNPACK:.+]] = tensor.unpack %[[EXPANDED:.+]] inner_dims_pos = [1, 2] inner_tiles = [8, 8] into %[[EMPTY]] : tensor<?x32x32x8x8xf32> -> tensor<?x256x256xf32>
+// CHECK:         return %[[UNPACK]] : tensor<?x256x256xf32>
+
+// -----
+
 func.func @push_down_permuted_unpack_through_expand(%5: tensor<4x32x384x8x8xf32>) -> tensor<4x12x256x256xf32> {
   %6 = tensor.empty() : tensor<4x3072x256xf32>
   %unpack = tensor.unpack %5 outer_dims_perm = [0, 2, 1] inner_dims_pos = [2, 1] inner_tiles = [8, 8] into %6 : tensor<4x32x384x8x8xf32> -> tensor<4x3072x256xf32>