Slice Operations

Header: include/stratax/indexing/Slicing.hpp.

Shared storage and lifetime

stratax::indexing::slice returns ArrayView with copied metadata and shared elements. Mutable sources yield mutable views; const owning sources yield read-only elements. The caller must keep the source allocation alive. Destroying or assigning the owner can invalidate a view; moving or swapping storage changes which object must keep that allocation alive.

Vector takes one Slice, Matrix takes row and column Slices, and Tensor takes one Slice per axis either as a parameter pack or a vector of Slices. C++ also supports slicing an ArrayView with a vector of Slices. All axes are retained, including length-one and zero axes. No selected elements are copied.

Normalization and errors

Bounds are signed and half-open. Negative bounds count from the end and positive-step bounds clamp to [0, extent]. Directionally empty selections produce empty views with the fully selected shape. Slice construction rejects step zero. Public view-producing overloads reject negative steps with Exceptions::IndexError, because their strides are unsigned.

The internal normalization helper can describe reverse ranges, including its special stop=-1 sentinel. That does not enable reverse views in the public API. Extents that do not fit ptrdiff_t raise IndexError. Tensor/view slice-count mismatches raise RankError; checked offset/stride overflow raises DimensionError. Metadata allocation can fail. For empty results the base pointer is retained without shifting it beyond the source allocation; it need not be null.

Example

#include <stratax.h>
#include <cassert>
 
int main() {
    stratax::container::Matrix<int> source{{0, 1, 2}, {3, 4, 5}};
    auto view = stratax::indexing::slice(source,
        stratax::core::Slice{0, 2}, stratax::core::Slice{0, 3, 2});
    assert(view.shape() == stratax::core::Shape({2, 2}));
    assert(view.strides() == stratax::core::Shape({3, 2}));
    view.at(1) = 20;
    assert(source(0, 2) == 20);
    auto copy = stratax::conversion::to_matrix(view);
    copy(0, 0) = 99;
    assert(source(0, 0) == 0);
}

Construction takes O(r) time and metadata storage for rank r, independent of selected element count. Logical flat access takes O(r). Vector and Matrix slice construction have fixed rank and O(1) metadata cost. Python slice views keep their owner alive but do not yet support further slicing; see Python API Reference and ArrayView.