view liboctave/array/Array.h @ 29949:f254c302bb9c

remove JIT compiler from Octave sources As stated in the NEWS file entry added with this changeset, no one has ever seriously taken on further development of the JIT compiler in Octave since it was first added as part of a Google Summer of Code project in 2012 and it still does nothing significant. It is out of date with the default interpreter that walks the parse tree. Even though we have fixed the configure script to disable it by default, people still ask questions about how to build it, but it doesn’t seem that they are doing that to work on it but because they think it will make Octave code run faster (it never did, except for some extremely simple bits of code as examples for demonstration purposes only). * NEWS: Note change. * configure.ac, acinclude.m4: Eliminate checks and macros related to the JIT compiler and LLVM. * basics.txi, install.txi, octave.texi, vectorize.txi: Remove mention of JIT compiler and LLVM. * jit-ir.cc, jit-ir.h, jit-typeinfo.cc, jit-typeinfo.h, jit-util.cc, jit-util.h, pt-jit.cc, pt-jit.h: Delete. * libinterp/parse-tree/module.mk: Update. * Array-jit.cc: Delete. * libinterp/template-inst/module.mk: Update. * test/jit.tst: Delete. * test/module.mk: Update. * interpreter.cc (interpreter::interpreter): Don't check options for debug_jit or jit_compiler. * toplev.cc (F__octave_config_info__): Remove JIT compiler and LLVM info from struct. * ov-base.h (octave_base_value::grab, octave_base_value::release): Delete. * ov-builtin.h, ov-builtin.cc (octave_builtin::to_jit, octave_builtin::stash_jit): Delete. (octave_builtin::m_jtype): Delete data member and all uses. * ov-usr-fcn.h, ov-usr-fcn.cc (octave_user_function::m_jit_info): Delete data member and all uses. (octave_user_function::get_info, octave_user_function::stash_info): Delete. * options.h (DEBUG_JIT_OPTION, JIT_COMPILER_OPTION): Delete macro definitions and all uses. * octave.h, octave.cc (cmdline_options::cmdline_options): Don't handle DEBUG_JIT_OPTION, JIT_COMPILER_OPTION): Delete. (cmdline_options::debug_jit, cmdline_options::jit_compiler): Delete functions and all uses. (cmdline_options::m_debug_jit, cmdline_options::m_jit_compiler): Delete data members and all uses. (octave_getopt_options long_opts): Remove "debug-jit" and "jit-compiler" from the list. * pt-eval.cc (tree_evaluator::visit_simple_for_command, tree_evaluator::visit_complex_for_command, tree_evaluator::visit_while_command, tree_evaluator::execute_user_function): Eliminate JIT compiler code. * pt-loop.h, pt-loop.cc (tree_while_command::get_info, tree_while_command::stash_info, tree_simple_for_command::get_info, tree_simple_for_command::stash_info): Delete functions and all uses. (tree_while_command::m_compiled, tree_simple_for_command::m_compiled): Delete member variable and all uses. * usage.h (usage_string, octave_print_verbose_usage_and_exit): Remove [--debug-jit] and [--jit-compiler] from the message. * Array.h (Array<T>::Array): Remove constructor that was only intended to be used by the JIT compiler. (Array<T>::jit_ref_count, Array<T>::jit_slice_data, Array<T>::jit_dimensions, Array<T>::jit_array_rep): Delete. * Marray.h (MArray<T>::MArray): Remove constructor that was only intended to be used by the JIT compiler. * NDArray.h (NDArray::NDarray): Remove constructor that was only intended to be used by the JIT compiler. * dim-vector.h (dim_vector::to_jit): Delete. (dim_vector::dim_vector): Remove constructor that was only intended to be used by the JIT compiler. * codeql-analysis.yaml, make.yaml: Don't require llvm-dev. * subst-config-vals.in.sh, subst-cross-config-vals.in.sh: Don't substitute OCTAVE_CONF_LLVM_CPPFLAGS, OCTAVE_CONF_LLVM_LDFLAGS, or OCTAVE_CONF_LLVM_LIBS. * Doxyfile.in: Don't define HAVE_LLVM. * aspell-octave.en.pws: Eliminate jit, JIT, and LLVM from the list of spelling exceptions. * build-env.h, build-env.in.cc (LLVM_CPPFLAGS, LLVM_LDFLAGS, LLVM_LIBS): Delete variables and all uses. * libinterp/corefcn/module.mk (%canon_reldir%_libcorefcn_la_CPPFLAGS): Remove $(LLVM_CPPFLAGS) from the list. * libinterp/parse-tree/module.mk (%canon_reldir%_libparse_tree_la_CPPFLAGS): Remove $(LLVM_CPPFLAGS) from the list.
author John W. Eaton <jwe@octave.org>
date Tue, 10 Aug 2021 16:42:29 -0400
parents 3e5e88d9c85f
children 4c88a452519c
line wrap: on
line source

////////////////////////////////////////////////////////////////////////
//
// Copyright (C) 1993-2021 The Octave Project Developers
//
// See the file COPYRIGHT.md in the top-level directory of this
// distribution or <https://octave.org/copyright/>.
//
// This file is part of Octave.
//
// Octave is free software: you can redistribute it and/or modify it
// under the terms of the GNU General Public License as published by
// the Free Software Foundation, either version 3 of the License, or
// (at your option) any later version.
//
// Octave is distributed in the hope that it will be useful, but
// WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU General Public License
// along with Octave; see the file COPYING.  If not, see
// <https://www.gnu.org/licenses/>.
//
////////////////////////////////////////////////////////////////////////

#if ! defined (octave_Array_h)
#define octave_Array_h 1

#include "octave-config.h"

#include <cassert>
#include <cstddef>

#include <algorithm>
#include <iosfwd>
#include <string>

#include "dim-vector.h"
#include "idx-vector.h"
#include "lo-error.h"
#include "lo-traits.h"
#include "lo-utils.h"
#include "oct-refcount.h"
#include "oct-sort.h"
#include "quit.h"

//! N Dimensional Array with copy-on-write semantics.
//!
//! The Array class is at the root of Octave.  It provides a container
//! with an arbitrary number of dimensions.  The operator () provides
//! access to individual elements via subscript and linear indexing.
//! Indexing starts at 0.  Arrays are column-major order as in Fortran.
//!
//! @code{.cc}
//! // 3 D Array with 10 rows, 20 columns, and 5 pages, filled with 7.0
//! Array<double> A Array<double (dim_vector (10, 20, 5), 7.0);
//!
//! // set value for row 0, column 10, and page 3
//! A(0, 10, 3) = 2.5;
//!
//! // get value for row 1, column 2, and page 0
//! double v = A(1, 2, 0);
//!
//! // get value for 25th element (row 4, column 3, page 1)
//! double v = A(24);
//! @endcode
//!
//! ## Notes on STL compatibility
//!
//! ### size() and length()
//!
//! To access the total number of elements in an Array, use numel()
//! which is short for number of elements and is equivalent to the
//! Octave function with same name.
//!
//! @code{.cc}
//! Array<int> A (dim_vector (10, 20, 4), 1);
//!
//! octave_idx_type n = A.numel (); // returns 800 (10x20x4)
//!
//! octave_idx_type nr = A.size (0); // returns 10 (number of rows/dimension 0)
//! octave_idx_type nc = A.size (1); // returns 20 (number of columns)
//! octave_idx_type nc = A.size (2); // returns 4 (size of dimension 3)
//! octave_idx_type l6 = A.size (6); // returns 1 (implicit singleton dimension)
//!
//! // Alternatively, get a dim_vector which represents the dimensions.
//! dim_vector dims = A.dims ();
//! @endcode
//!
//! The methods size() and length() as they exist in the STL cause
//! confusion in the context of a N dimensional array.
//!
//! The size() of an array is the length of all dimensions.  In Octave,
//! the size() function returns a row vector with the length of each
//! dimension, or the size of a specific dimension.  Only the latter is
//! present in liboctave.
//!
//! Since there is more than 1 dimension, length() would not make sense
//! without expliciting which dimension.  If the function existed, which
//! length should it return?  Octave length() function returns the length
//! of the longest dimension which is an odd definition, only useful for
//! vectors and square matrices.  The alternatives numel(), rows(),
//! columns(), and size(d) are more explicit and recommended.
//!
//! ### size_type
//!
//! Array::size_type is 'octave_idx_type' which is a typedef for 'int'
//! or 'long int', depending whether Octave was configured for 64-bit
//! indexing.
//!
//! This is a signed integer which may cause problems when mixed with
//! STL containers.  The reason is that Octave interacts with Fortran
//! routines, providing an interface many Fortran numeric libraries.
//!
//! ## Subclasses
//!
//! The following subclasses specializations, will be of most use:
//!   - Matrix: Array<double> with only 2 dimensions
//!   - ComplexMatrix: Array<std::complex<double>> with only 2 dimensions
//!   - boolNDArray: N dimensional Array<bool>
//!   - ColumnVector: Array<double> with 1 column
//!   - string_vector: Array<std::string> with 1 column
//!   - Cell: Array<octave_value>, equivalent to an Octave cell.

// forward declare template with visibility attribute
template <typename T> class OCTARRAY_API Array;

template <typename T>
class
Array
{
protected:

  //! The real representation of all arrays.
  class ArrayRep
  {
  public:

    T *m_data;
    octave_idx_type m_len;
    octave::refcount<octave_idx_type> m_count;

    ArrayRep (T *d, octave_idx_type l)
      : m_data (new T [l]), m_len (l), m_count (1)
    {
      std::copy_n (d, l, m_data);
    }

    template <typename U>
    ArrayRep (U *d, octave_idx_type l)
      : m_data (new T [l]), m_len (l), m_count (1)
    {
      std::copy_n (d, l, m_data);
    }

    // Use new instead of setting data to 0 so that fortran_vec and
    // data always return valid addresses, even for zero-size arrays.

    ArrayRep (void) : m_data (new T [0]), m_len (0), m_count (1) { }

    explicit ArrayRep (octave_idx_type n)
      : m_data (new T [n]), m_len (n), m_count (1) { }

    explicit ArrayRep (octave_idx_type n, const T& val)
      : m_data (new T [n]), m_len (n), m_count (1)
    {
      std::fill_n (m_data, n, val);
    }

    ArrayRep (const ArrayRep& a)
      : m_data (new T [a.m_len]), m_len (a.m_len), m_count (1)
    {
      std::copy_n (a.m_data, a.m_len, m_data);
    }

    ~ArrayRep (void) { delete [] m_data; }

    octave_idx_type numel (void) const { return m_len; }

    // No assignment!

    ArrayRep& operator = (const ArrayRep&) = delete;
  };

  //--------------------------------------------------------------------

public:

  void make_unique (void)
  {
    if (m_rep->m_count > 1)
      {
        ArrayRep *r = new ArrayRep (m_slice_data, m_slice_len);

        if (--m_rep->m_count == 0)
          delete m_rep;

        m_rep = r;
        m_slice_data = m_rep->m_data;
      }
  }

  typedef T element_type;

  typedef T value_type;

  //! Used for operator(), and returned by numel() and size()
  //! (beware: signed integer)
  typedef octave_idx_type size_type;

  typedef typename ref_param<T>::type crefT;

  typedef bool (*compare_fcn_type) (typename ref_param<T>::type,
                                    typename ref_param<T>::type);

protected:

  dim_vector m_dimensions;

  typename Array<T>::ArrayRep *m_rep;

  // Rationale:
  // m_slice_data is a pointer to m_rep->m_data, denoting together with m_slice_len the
  // actual portion of the data referenced by this Array<T> object.  This
  // allows to make shallow copies not only of a whole array, but also of
  // contiguous subranges.  Every time m_rep is directly manipulated, m_slice_data
  // and m_slice_len need to be properly updated.

  T *m_slice_data;
  octave_idx_type m_slice_len;

  //! slice constructor
  Array (const Array<T>& a, const dim_vector& dv,
         octave_idx_type l, octave_idx_type u)
    : m_dimensions (dv), m_rep(a.m_rep), m_slice_data (a.m_slice_data+l), m_slice_len (u-l)
  {
    m_rep->m_count++;
    m_dimensions.chop_trailing_singletons ();
  }

private:

  static OCTARRAY_API typename Array<T>::ArrayRep *nil_rep (void);

public:

  //! Empty ctor (0 by 0).
  Array (void)
    : m_dimensions (), m_rep (nil_rep ()), m_slice_data (m_rep->m_data),
      m_slice_len (m_rep->m_len)
  {
    m_rep->m_count++;
  }

  //! nD uninitialized ctor.
  explicit Array (const dim_vector& dv)
    : m_dimensions (dv),
      m_rep (new typename Array<T>::ArrayRep (dv.safe_numel ())),
      m_slice_data (m_rep->m_data), m_slice_len (m_rep->m_len)
  {
    m_dimensions.chop_trailing_singletons ();
  }

  //! nD initialized ctor.
  explicit Array (const dim_vector& dv, const T& val)
    : m_dimensions (dv),
      m_rep (new typename Array<T>::ArrayRep (dv.safe_numel ())),
      m_slice_data (m_rep->m_data), m_slice_len (m_rep->m_len)
  {
    fill (val);
    m_dimensions.chop_trailing_singletons ();
  }

  //! Reshape constructor.
  OCTARRAY_API Array (const Array<T>& a, const dim_vector& dv);

  //! Constructor from standard library sequence containers.
  template<template <typename...> class Container>
  Array (const Container<T>& a, const dim_vector& dv);

  //! Type conversion case.
  template <typename U>
  Array (const Array<U>& a)
    : m_dimensions (a.dims ()),
      m_rep (new typename Array<T>::ArrayRep (a.data (), a.numel ())),
      m_slice_data (m_rep->m_data), m_slice_len (m_rep->m_len)
  { }

  //! No type conversion case.
  Array (const Array<T>& a)
    : m_dimensions (a.m_dimensions), m_rep (a.m_rep), m_slice_data (a.m_slice_data),
      m_slice_len (a.m_slice_len)
  {
    m_rep->m_count++;
  }

  Array (Array<T>&& a)
    : m_dimensions (std::move (a.m_dimensions)), m_rep (a.m_rep),
      m_slice_data (a.m_slice_data), m_slice_len (a.m_slice_len)
  {
    a.m_rep = nullptr;
    a.m_slice_data = nullptr;
    a.m_slice_len = 0;
  }

public:

  virtual ~Array (void)
  {
    // Because we define a move constructor and a move assignment
    // operator, m_rep may be a nullptr here.  We should only need to
    // protect the move assignment operator in a similar way.

    if (m_rep && --m_rep->m_count == 0)
      delete m_rep;
  }

  Array<T>& operator = (const Array<T>& a)
  {
    if (this != &a)
      {
        if (--m_rep->m_count == 0)
          delete m_rep;

        m_rep = a.m_rep;
        m_rep->m_count++;

        m_dimensions = a.m_dimensions;
        m_slice_data = a.m_slice_data;
        m_slice_len = a.m_slice_len;
      }

    return *this;
  }

  Array<T>& operator = (Array<T>&& a)
  {
    if (this != &a)
      {
        m_dimensions = std::move (a.m_dimensions);

        // Because we define a move constructor and a move assignment
        // operator, m_rep may be a nullptr here.  We should only need to
        // protect the destructor in a similar way.

        if (m_rep && --m_rep->m_count == 0)
          delete m_rep;

        m_rep = a.m_rep;
        m_slice_data = a.m_slice_data;
        m_slice_len = a.m_slice_len;

        a.m_rep = nullptr;
        a.m_slice_data = nullptr;
        a.m_slice_len = 0;
      }

    return *this;
  }

  OCTARRAY_API void fill (const T& val);

  OCTARRAY_API void clear (void);
  OCTARRAY_API void clear (const dim_vector& dv);

  void clear (octave_idx_type r, octave_idx_type c)
  { clear (dim_vector (r, c)); }

  //! Number of elements in the array.
  octave_idx_type numel (void) const { return m_slice_len; }
  //@}

  //! Return the array as a column vector.
  Array<T> as_column (void) const
  {
    Array<T> retval (*this);
    if (m_dimensions.ndims () != 2 || m_dimensions(1) != 1)
      retval.m_dimensions = dim_vector (numel (), 1);

    return retval;
  }

  //! Return the array as a row vector.
  Array<T> as_row (void) const
  {
    Array<T> retval (*this);
    if (m_dimensions.ndims () != 2 || m_dimensions(0) != 1)
      retval.m_dimensions = dim_vector (1, numel ());

    return retval;
  }

  //! Return the array as a matrix.
  Array<T> as_matrix (void) const
  {
    Array<T> retval (*this);
    if (m_dimensions.ndims () != 2)
      retval.m_dimensions = m_dimensions.redim (2);

    return retval;
  }

  //! @name First dimension
  //!
  //! Get the first dimension of the array (number of rows)
  //@{
  octave_idx_type dim1 (void) const { return m_dimensions(0); }
  octave_idx_type rows (void) const { return dim1 (); }
  //@}

  //! @name Second dimension
  //!
  //! Get the second dimension of the array (number of columns)
  //@{
  octave_idx_type dim2 (void) const { return m_dimensions(1); }
  octave_idx_type cols (void) const { return dim2 (); }
  octave_idx_type columns (void) const { return dim2 (); }
  //@}

  //! @name Third dimension
  //!
  //! Get the third dimension of the array (number of pages)
  //@{
  octave_idx_type dim3 (void) const { return m_dimensions(2); }
  octave_idx_type pages (void) const { return dim3 (); }
  //@}

  //! Size of the specified dimension.
  //!
  //! Dimensions beyond the Array number of dimensions return 1 as
  //! those are implicit singleton dimensions.
  //!
  //! Equivalent to Octave's 'size (A, DIM)'

  size_type size (const size_type d) const
  {
    // Should we throw for negative values?
    // Should >= ndims () be handled by dim_vector operator() instead ?
    return d >= ndims () ? 1 : m_dimensions(d);
  }

  std::size_t byte_size (void) const
  { return static_cast<std::size_t> (numel ()) * sizeof (T); }

  //! Return a const-reference so that dims ()(i) works efficiently.
  const dim_vector& dims (void) const { return m_dimensions; }

  //! Chop off leading singleton dimensions
  OCTARRAY_API Array<T> squeeze (void) const;

  OCTARRAY_API octave_idx_type compute_index (octave_idx_type i, octave_idx_type j) const;
  OCTARRAY_API octave_idx_type compute_index (octave_idx_type i, octave_idx_type j,
                                 octave_idx_type k) const;
  OCTARRAY_API octave_idx_type compute_index (const Array<octave_idx_type>& ra_idx) const;

  octave_idx_type compute_index_unchecked (const Array<octave_idx_type>& ra_idx)
  const
  { return m_dimensions.compute_index (ra_idx.data (), ra_idx.numel ()); }

  // No checking, even for multiple references, ever.

  T& xelem (octave_idx_type n) { return m_slice_data[n]; }
  crefT xelem (octave_idx_type n) const { return m_slice_data[n]; }

  T& xelem (octave_idx_type i, octave_idx_type j)
  { return xelem (dim1 ()*j+i); }
  crefT xelem (octave_idx_type i, octave_idx_type j) const
  { return xelem (dim1 ()*j+i); }

  T& xelem (octave_idx_type i, octave_idx_type j, octave_idx_type k)
  { return xelem (i, dim2 ()*k+j); }
  crefT xelem (octave_idx_type i, octave_idx_type j, octave_idx_type k) const
  { return xelem (i, dim2 ()*k+j); }

  T& xelem (const Array<octave_idx_type>& ra_idx)
  { return xelem (compute_index_unchecked (ra_idx)); }

  crefT xelem (const Array<octave_idx_type>& ra_idx) const
  { return xelem (compute_index_unchecked (ra_idx)); }

  // FIXME: would be nice to fix this so that we don't unnecessarily force
  //        a copy, but that is not so easy, and I see no clean way to do it.

  OCTARRAY_API T& checkelem (octave_idx_type n);

  OCTARRAY_API T& checkelem (octave_idx_type i, octave_idx_type j);

  OCTARRAY_API T& checkelem (octave_idx_type i, octave_idx_type j, octave_idx_type k);

  OCTARRAY_API T& checkelem (const Array<octave_idx_type>& ra_idx);

  T& elem (octave_idx_type n)
  {
    make_unique ();
    return xelem (n);
  }

  T& elem (octave_idx_type i, octave_idx_type j) { return elem (dim1 ()*j+i); }

  T& elem (octave_idx_type i, octave_idx_type j, octave_idx_type k)
  { return elem (i, dim2 ()*k+j); }

  T& elem (const Array<octave_idx_type>& ra_idx)
  { return Array<T>::elem (compute_index_unchecked (ra_idx)); }

  T& operator () (octave_idx_type n) { return elem (n); }
  T& operator () (octave_idx_type i, octave_idx_type j) { return elem (i, j); }
  T& operator () (octave_idx_type i, octave_idx_type j, octave_idx_type k)
  { return elem (i, j, k); }
  T& operator () (const Array<octave_idx_type>& ra_idx)
  { return elem (ra_idx); }

  OCTARRAY_API crefT checkelem (octave_idx_type n) const;

  OCTARRAY_API crefT checkelem (octave_idx_type i, octave_idx_type j) const;

  OCTARRAY_API crefT checkelem (octave_idx_type i, octave_idx_type j,
                   octave_idx_type k) const;

  OCTARRAY_API crefT checkelem (const Array<octave_idx_type>& ra_idx) const;

  crefT elem (octave_idx_type n) const { return xelem (n); }

  crefT elem (octave_idx_type i, octave_idx_type j) const
  { return xelem (i, j); }

  crefT elem (octave_idx_type i, octave_idx_type j, octave_idx_type k) const
  { return xelem (i, j, k); }

  crefT elem (const Array<octave_idx_type>& ra_idx) const
  { return Array<T>::xelem (compute_index_unchecked (ra_idx)); }

  crefT operator () (octave_idx_type n) const { return elem (n); }
  crefT operator () (octave_idx_type i, octave_idx_type j) const
  { return elem (i, j); }
  crefT operator () (octave_idx_type i, octave_idx_type j,
                     octave_idx_type k) const
  { return elem (i, j, k); }
  crefT operator () (const Array<octave_idx_type>& ra_idx) const
  { return elem (ra_idx); }

  // Fast extractors.  All of these produce shallow copies.

  //! Extract column: A(:,k+1).
  OCTARRAY_API Array<T> column (octave_idx_type k) const;
  //! Extract page: A(:,:,k+1).
  OCTARRAY_API Array<T> page (octave_idx_type k) const;

  //! Extract a slice from this array as a column vector: A(:)(lo+1:up).
  //! Must be 0 <= lo && up <= numel.  May be up < lo.
  OCTARRAY_API Array<T> linear_slice (octave_idx_type lo, octave_idx_type up) const;

  Array<T> reshape (octave_idx_type nr, octave_idx_type nc) const
  { return Array<T> (*this, dim_vector (nr, nc)); }

  Array<T> reshape (const dim_vector& new_dims) const
  { return Array<T> (*this, new_dims); }

  OCTARRAY_API Array<T> permute (const Array<octave_idx_type>& vec, bool inv = false) const;
  Array<T> ipermute (const Array<octave_idx_type>& vec) const
  { return permute (vec, true); }

  bool issquare (void) const { return (dim1 () == dim2 ()); }

  bool isempty (void) const { return numel () == 0; }

  bool isvector (void) const { return m_dimensions.isvector (); }

  bool is_nd_vector (void) const { return m_dimensions.is_nd_vector (); }

  OCTARRAY_API Array<T> transpose (void) const;
  OCTARRAY_API Array<T> hermitian (T (*fcn) (const T&) = nullptr) const;

  const T * data (void) const { return m_slice_data; }

  OCTAVE_DEPRECATED (7, "for read-only access, use 'data' method instead")
  const T * fortran_vec (void) const { return data (); }

  OCTARRAY_API T * fortran_vec (void);

  bool is_shared (void) { return m_rep->m_count > 1; }

  int ndims (void) const { return m_dimensions.ndims (); }

  //@{
  //! Indexing without resizing.
  OCTARRAY_API Array<T> index (const octave::idx_vector& i) const;

  OCTARRAY_API Array<T> index (const octave::idx_vector& i, const octave::idx_vector& j) const;

  OCTARRAY_API Array<T> index (const Array<octave::idx_vector>& ia) const;
  //@}

  virtual OCTARRAY_API T resize_fill_value (void) const;

  //@{
  //! Resizing (with fill).
  OCTARRAY_API void resize2 (octave_idx_type nr, octave_idx_type nc, const T& rfv);
  void resize2 (octave_idx_type nr, octave_idx_type nc)
  {
    resize2 (nr, nc, resize_fill_value ());
  }

  OCTARRAY_API void resize1 (octave_idx_type n, const T& rfv);
  void resize1 (octave_idx_type n) { resize1 (n, resize_fill_value ()); }

  OCTARRAY_API void resize (const dim_vector& dv, const T& rfv);
  void resize (const dim_vector& dv) { resize (dv, resize_fill_value ()); }
  //@}

  //@{
  //! Indexing with possible resizing and fill

  // FIXME: this is really a corner case, that should better be
  // handled directly in liboctinterp.

  OCTARRAY_API Array<T> index (const octave::idx_vector& i, bool resize_ok, const T& rfv) const;
  Array<T> index (const octave::idx_vector& i, bool resize_ok) const
  {
    return index (i, resize_ok, resize_fill_value ());
  }

  OCTARRAY_API Array<T> index (const octave::idx_vector& i, const octave::idx_vector& j, bool resize_ok,
                  const T& rfv) const;
  Array<T> index (const octave::idx_vector& i, const octave::idx_vector& j,
                  bool resize_ok) const
  {
    return index (i, j, resize_ok, resize_fill_value ());
  }

  OCTARRAY_API Array<T> index (const Array<octave::idx_vector>& ia, bool resize_ok,
                  const T& rfv) const;
  Array<T> index (const Array<octave::idx_vector>& ia, bool resize_ok) const
  {
    return index (ia, resize_ok, resize_fill_value ());
  }
  //@}

  //@{
  //! Indexed assignment (always with resize & fill).
  OCTARRAY_API void assign (const octave::idx_vector& i, const Array<T>& rhs, const T& rfv);
  void assign (const octave::idx_vector& i, const Array<T>& rhs)
  {
    assign (i, rhs, resize_fill_value ());
  }

  OCTARRAY_API void assign (const octave::idx_vector& i, const octave::idx_vector& j, const Array<T>& rhs,
               const T& rfv);
  void assign (const octave::idx_vector& i, const octave::idx_vector& j, const Array<T>& rhs)
  {
    assign (i, j, rhs, resize_fill_value ());
  }

  OCTARRAY_API void assign (const Array<octave::idx_vector>& ia, const Array<T>& rhs, const T& rfv);
  void assign (const Array<octave::idx_vector>& ia, const Array<T>& rhs)
  {
    assign (ia, rhs, resize_fill_value ());
  }
  //@}

  //@{
  //! Deleting elements.

  //! A(I) = [] (with a single subscript)
  OCTARRAY_API void delete_elements (const octave::idx_vector& i);

  //! A(:,...,I,...,:) = [] (>= 2 subscripts, one of them is non-colon)
  OCTARRAY_API void delete_elements (int dim, const octave::idx_vector& i);

  //! Dispatcher to the above two.
  OCTARRAY_API void delete_elements (const Array<octave::idx_vector>& ia);
  //@}

  //! Insert an array into another at a specified position.  If
  //! size (a) is [d1 d2 ... dN] and idx is [i1 i2 ... iN], this
  //! method is equivalent to x(i1:i1+d1-1, i2:i2+d2-1, ... ,
  //! iN:iN+dN-1) = a.
  OCTARRAY_API Array<T>& insert (const Array<T>& a, const Array<octave_idx_type>& idx);

  //! This is just a special case for idx = [r c 0 ...]
  OCTARRAY_API Array<T>& insert (const Array<T>& a, octave_idx_type r, octave_idx_type c);

  void maybe_economize (void)
  {
    if (m_rep->m_count == 1 && m_slice_len != m_rep->m_len)
      {
        ArrayRep *new_rep = new ArrayRep (m_slice_data, m_slice_len);
        delete m_rep;
        m_rep = new_rep;
        m_slice_data = m_rep->m_data;
      }
  }

  OCTARRAY_API void print_info (std::ostream& os, const std::string& prefix) const;

  //! Give a pointer to the data in mex format.  Unsafe.  This function
  //! exists to support the MEX interface.  You should not use it
  //! anywhere else.
  void * mex_get_data (void) const { return const_cast<T *> (data ()); }

  OCTARRAY_API Array<T> sort (int dim = 0, sortmode mode = ASCENDING) const;
  OCTARRAY_API Array<T> sort (Array<octave_idx_type> &sidx, int dim = 0,
                 sortmode mode = ASCENDING) const;

  //! Ordering is auto-detected or can be specified.
  OCTARRAY_API sortmode issorted (sortmode mode = UNSORTED) const;

  //! Sort by rows returns only indices.
  OCTARRAY_API Array<octave_idx_type> sort_rows_idx (sortmode mode = ASCENDING) const;

  //! Ordering is auto-detected or can be specified.
  OCTARRAY_API sortmode is_sorted_rows (sortmode mode = UNSORTED) const;

  //! Do a binary lookup in a sorted array.  Must not contain NaNs.
  //! Mode can be specified or is auto-detected by comparing 1st and last element.
  OCTARRAY_API octave_idx_type lookup (const T& value, sortmode mode = UNSORTED) const;

  //! Ditto, but for an array of values, specializing on the case when values
  //! are sorted.  NaNs get the value N.
  OCTARRAY_API Array<octave_idx_type> lookup (const Array<T>& values,
                                 sortmode mode = UNSORTED) const;

  //! Count nonzero elements.
  OCTARRAY_API octave_idx_type nnz (void) const;

  //! Find indices of (at most n) nonzero elements.  If n is specified,
  //! backward specifies search from backward.
  OCTARRAY_API Array<octave_idx_type> find (octave_idx_type n = -1,
                               bool backward = false) const;

  //! Returns the n-th element in increasing order, using the same
  //! ordering as used for sort.  n can either be a scalar index or a
  //! contiguous range.
  OCTARRAY_API Array<T> nth_element (const octave::idx_vector& n, int dim = 0) const;

  //! Get the kth super or subdiagonal.  The zeroth diagonal is the
  //! ordinary diagonal.
  OCTARRAY_API Array<T> diag (octave_idx_type k = 0) const;

  OCTARRAY_API Array<T> diag (octave_idx_type m, octave_idx_type n) const;

  //! Concatenation along a specified (0-based) dimension, equivalent
  //! to cat().  dim = -1 corresponds to dim = 0 and dim = -2
  //! corresponds to dim = 1, but apply the looser matching rules of
  //! vertcat/horzcat.
  static OCTARRAY_API Array<T>
  cat (int dim, octave_idx_type n, const Array<T> *array_list);

  //! Apply function fcn to each element of the Array<T>.  This function
  //! is optimized with a manually unrolled loop.
  template <typename U, typename F>
  Array<U>
  map (F fcn) const
  {
    octave_idx_type len = numel ();

    const T *m = data ();

    Array<U> result (dims ());
    U *p = result.fortran_vec ();

    octave_idx_type i;
    for (i = 0; i < len - 3; i += 4)
      {
        octave_quit ();

        p[i] = fcn (m[i]);
        p[i+1] = fcn (m[i+1]);
        p[i+2] = fcn (m[i+2]);
        p[i+3] = fcn (m[i+3]);
      }

    octave_quit ();

    for (; i < len; i++)
      p[i] = fcn (m[i]);

    return result;
  }

  //@{
  //! Overloads for function references.
  template <typename U>
  Array<U>
  map (U (&fcn) (T)) const
  { return map<U, U (&) (T)> (fcn); }

  template <typename U>
  Array<U>
  map (U (&fcn) (const T&)) const
  { return map<U, U (&) (const T&)> (fcn); }
  //@}

  //! Generic any/all test functionality with arbitrary predicate.
  template <typename F, bool zero>
  bool test (F fcn) const
  {
    return octave::any_all_test<F, T, zero> (fcn, data (), numel ());
  }

  //@{
  //! Simpler calls.
  template <typename F>
  bool test_any (F fcn) const
  { return test<F, false> (fcn); }

  template <typename F>
  bool test_all (F fcn) const
  { return test<F, true> (fcn); }
  //@}

  //@{
  //! Overloads for function references.
  bool test_any (bool (&fcn) (T)) const
  { return test<bool (&) (T), false> (fcn); }

  bool test_any (bool (&fcn) (const T&)) const
  { return test<bool (&) (const T&), false> (fcn); }

  bool test_all (bool (&fcn) (T)) const
  { return test<bool (&) (T), true> (fcn); }

  bool test_all (bool (&fcn) (const T&)) const
  { return test<bool (&) (const T&), true> (fcn); }
  //@}

  template <typename U> friend class Array;

  //! Returns true if this->dims () == dv, and if so, replaces this->m_dimensions
  //! by a shallow copy of dv.  This is useful for maintaining several arrays
  //! with supposedly equal dimensions (e.g. structs in the interpreter).
  OCTARRAY_API bool optimize_dimensions (const dim_vector& dv);

private:
  OCTARRAY_API static void instantiation_guard ();
};

// We use a variadic template for template template parameter so that
// we don't have to specify all the template parameters and limit this
// to Container<T>. http://stackoverflow.com/a/20499809/1609556
template<typename T>
template<template <typename...> class Container>
Array<T>::Array (const Container<T>& a, const dim_vector& dv)
  : m_dimensions (dv), m_rep (new typename Array<T>::ArrayRep (dv.safe_numel ())),
    m_slice_data (m_rep->m_data), m_slice_len (m_rep->m_len)
{
  if (m_dimensions.safe_numel () != octave_idx_type (a.size ()))
    {
      std::string new_dims_str = m_dimensions.str ();

      (*current_liboctave_error_handler)
        ("reshape: can't reshape %zi elements into %s array",
         a.size (), new_dims_str.c_str ());
    }

  octave_idx_type i = 0;
  for (const T& x : a)
    m_slice_data[i++] = x;

  m_dimensions.chop_trailing_singletons ();
}

template <typename T>
OCTARRAY_API std::ostream&
operator << (std::ostream& os, const Array<T>& a);

#endif