Description

mlir-opt --canonicalize --inline crashes with std::bad_optional_access in getNewMixedTileSizes when canonicalizing a linalg.unpack op that has a tensor.cast as its source and a dynamic (non-constant) tile size. The FoldTensorCastUnPackOp pattern fires and calls getNewMixedTileSizes, which calls getConstantIntValue(tile).value() unconditionally. When the tile size is a dynamic index SSA value rather than a compile-time constant, getConstantIntValue returns std::nullopt and .value() aborts.

Reproducer

module {
  func.func @get_tile() -> index {
    %c256 = arith.constant 256 : index
    return %c256 : index
  }

  func.func @f(%A: tensor<1x3x8x1xi32>) -> tensor<7x3xi32> {
    %tile_size = func.call @get_tile() : () -> index
    %empty = tensor.empty() : tensor<7x3xi32>
    %cast = tensor.cast %A : tensor<1x3x8x1xi32> to tensor<?x3x?x1xi32>
    %unpack = linalg.unpack %cast
      inner_dims_pos = [0, 1]
      inner_tiles = [%tile_size, 1]
      into %empty : tensor<?x3x?x1xi32> -> tensor<7x3xi32>
    return %unpack : tensor<7x3xi32>
  }
}

Command

mlir-opt --canonicalize --inline reproduce.mlir

Expected behavior

FoldTensorCastUnPackOp should bail out (return failure) when a tile size is not a compile-time constant, rather than unconditionally calling .value() on the result of getConstantIntValue. Alternatively it should only fire when all tile sizes are statically known.

Actual behavior

mlir-opt: optional:1013: std::optional<long>::value() &&: Assertion `this->_M_is_engaged()' failed.
Aborted (core dumped)

The crash is in getNewMixedTileSizes at LinalgOps.cpp:5025, called from FoldTensorCastUnPackOp::matchAndRewrite at LinalgOps.cpp:6479.

Call chain

--canonicalize (runs after --inline propagates call)
  → FoldTensorCastUnPackOp::matchAndRewrite   (LinalgOps.cpp:6479)
    → getNewMixedTileSizes                    (LinalgOps.cpp:5025)
      → getConstantIntValue(tile)             → nullopt (tile is dynamic)
        → optional::value()                   ← abort: bad_optional_access