[Mlir-commits] [mlir] [MLIR] Folding unpack and pack sequence in data layout propagation from padded domain (PR #138332)

[Adam Siemieniuk]

================
@@ -455,13 +455,10 @@ func.func @unpack_on_output(%arg0: tensor<12x2x56x56x32xf32>) -> tensor<12x56x56
 // CHECK:         %[[UNPACKED_ARG0:.+]] = linalg.unpack %[[ARG0]]
 // CHECK-SAME:      outer_dims_perm = [0, 3, 1, 2] inner_dims_pos = [3] inner_tiles = [32]
 // CHECK-SAME:      into %[[ARG0_EMPTY_UNPACK]]
-// CHECK:         %[[ARG0_EMPTY_PACK:.+]] = tensor.empty() : tensor<12x2x56x56x32xf32>
-// CHECK:         %[[PACKED_ARG0:.+]] = linalg.pack %[[UNPACKED_ARG0]]
-// CHECK-SAME:      outer_dims_perm = [0, 3, 1, 2] inner_dims_pos = [3] inner_tiles = [32]
-// CHECK-SAME:      into %[[ARG0_EMPTY_PACK]]
+// CHECK:         %[[EMPTY:.+]] = tensor.empty() : tensor<12x2x56x56x32xf32>
 // CHECK:         %[[RES:.+]] = linalg.generic
 // CHECK-SAME:      indexing_maps = [#[[$MAP]]]
-// CHECK-SAME:      outs(%[[PACKED_ARG0]]
+// CHECK-SAME:      outs(%[[EMPTY]]
----------------
adam-smnk wrote:

This change invalidates IR.

In this example, the `linalg.generic` outs value is not only used as a destination but also used for its initial value. Before propagation unpacked `%1` would contain data from `%arg0` which was preserved by `%[[PACKED._ARG0]]`.
After this PR, the op uses empty uninitialized tensor effectively reading garbage data.

https://github.com/llvm/llvm-project/pull/138332