rdkit/Code/RDBoost/python_streambuf.h

//
//  This file is part of the CCTBX distribution:
//     http://cctbx.sourceforge.net/
//  Downloaded from here:
//    http://cctbx.svn.sourceforge.net/viewvc/cctbx/trunk/boost_adaptbx/python_streambuf.h?revision=13619
//
//  Copyright (c) 2006, The Regents of the University of
//  California, through Lawrence Berkeley National Laboratory (subject to
//  receipt of any required approvals from the U.S. Dept. of Energy).  All
//  rights reserved.
//
//  The license is here:
//    http://cctbx.svn.sourceforge.net/viewvc/cctbx/trunk/boost_adaptbx/LICENSE_2_0.txt?revision=5148
//
#include <RDGeneral/export.h>
#ifndef BOOST_ADAPTBX_PYTHON_STREAMBUF_H
#define BOOST_ADAPTBX_PYTHON_STREAMBUF_H
#include <RDGeneral/BoostStartInclude.h>
#include <boost/python/object.hpp>
#include <boost/python/str.hpp>
#include <boost/python/extract.hpp>

#include <boost/optional.hpp>
#include <boost/utility/typed_in_place_factory.hpp>
#include <RDGeneral/BoostEndInclude.h>

#include <RDGeneral/Invariant.h>
#include <RDGeneral/Exceptions.h>

#include <streambuf>
#include <iostream>

namespace boost_adaptbx {
namespace python {

namespace bp = boost::python;

/// A stream buffer getting data from and putting data into a Python file object
/** The aims are as follow:

    - Given a C++ function acting on a standard stream, e.g.

      \code
      void read_inputs(std::istream& input) {
        ...
        input >> something >> something_else;
      }
      \endcode

      and given a piece of Python code which creates a file-like object,
      to be able to pass this file object to that C++ function, e.g.

      \code
      import gzip
      gzip_file_obj = gzip.GzipFile(...)
      read_inputs(gzip_file_obj)
      \endcode

      and have the standard stream pull data from and put data into the Python
      file object.

    - When Python \c read_inputs() returns, the Python object is able to
      continue reading or writing where the C++ code left off.

    - Operations in C++ on mere files should be competitively fast compared
      to the direct use of \c std::fstream.


    \b Motivation

      - the standard Python library offer of file-like objects (files,
        compressed files and archives, network, ...) is far superior to the
        offer of streams in the C++ standard library and Boost C++ libraries.

      - i/o code involves a fair amount of text processing which is more
        efficiently prototyped in Python but then one may need to rewrite
        a time-critical part in C++, in as seamless a manner as possible.

    \b Usage

    This is 2-step:

      - a trivial wrapper function

        \code
          using boost_adaptbx::python::streambuf;
          void read_inputs_wrapper(streambuf& input)
          {
            streambuf::istream is(input);
            read_inputs(is);
          }

          def("read_inputs", read_inputs_wrapper);
        \endcode

        which has to be written every time one wants a Python binding for
        such a C++ function.

      - the Python side

        \code
          from boost.python import streambuf
          read_inputs(streambuf(python_file_obj=obj, buffer_size=1024))
        \endcode

        \c buffer_size is optional. See also: \c default_buffer_size

  Note: references are to the C++ standard (the numbers between parentheses
  at the end of references are margin markers).
*/
class streambuf : public std::basic_streambuf<char> {
 private:
  typedef std::basic_streambuf<char> base_t;

 public:
  /* The syntax
      using base_t::char_type;
     would be nicer but Visual Studio C++ 8 chokes on it
  */
  typedef base_t::char_type char_type;
  typedef base_t::int_type int_type;
  typedef base_t::pos_type pos_type;
  typedef base_t::off_type off_type;
  typedef base_t::traits_type traits_type;

  // work around Visual C++ 7.1 problem
  inline static int traits_type_eof() { return traits_type::eof(); }

  /// The default size of the read and write buffer.
  /** They are respectively used to buffer data read from and data written to
      the Python file object. It can be modified from Python.
  */
  const static std::size_t default_buffer_size = 1024;

  /// Construct from a Python file object
  /** if buffer_size is 0 the current default_buffer_size is used.
   */
  streambuf(bp::object &python_file_obj, std::size_t buffer_size_ = 0)
      : py_read(getattr(python_file_obj, "read", bp::object())),
        py_write(getattr(python_file_obj, "write", bp::object())),
        py_seek(getattr(python_file_obj, "seek", bp::object())),
        py_tell(getattr(python_file_obj, "tell", bp::object())),
        buffer_size(buffer_size_ != 0 ? buffer_size_ : default_buffer_size),
        write_buffer(nullptr),
        pos_of_read_buffer_end_in_py_file(0),
        pos_of_write_buffer_end_in_py_file(buffer_size),
        farthest_pptr(nullptr) {
    TEST_ASSERT(buffer_size != 0);
    /* Some Python file objects (e.g. sys.stdout and sys.stdin)
       have non-functional seek and tell. If so, assign None to
       py_tell and py_seek.
     */
    if (py_tell != bp::object()) {
      try {
        off_type py_pos = bp::extract<off_type>(py_tell());
        if (py_seek != bp::object()) {
          /* Make sure we can actually seek.
             bzip2 readers from python have a seek method, but it fails
             when they are in write mode.
           */
          py_seek(py_pos);
        }
      } catch (bp::error_already_set &) {
        py_tell = bp::object();
        py_seek = bp::object();
        /* Boost.Python does not do any Python exception handling whatsoever
           So we need to catch it by hand like so.
         */
        PyErr_Clear();
      }
    }

    if (py_write != bp::object()) {
      // C-like string to make debugging easier
      write_buffer = new char[buffer_size + 1];
      write_buffer[buffer_size] = '\0';
      setp(write_buffer, write_buffer + buffer_size);  // 27.5.2.4.5 (5)
      farthest_pptr = pptr();
    } else {
      // The first attempt at output will result in a call to overflow
      setp(nullptr, nullptr);
    }

    if (py_tell != bp::object()) {
      off_type py_pos = bp::extract<off_type>(py_tell());
      pos_of_read_buffer_end_in_py_file = py_pos;
      pos_of_write_buffer_end_in_py_file = py_pos;
    }
  }

  /// constructor to enforce a mode (binary or text)
  streambuf(bp::object &python_file_obj, char mode,
            std::size_t buffer_size_ = 0)
      : streambuf(python_file_obj, buffer_size_) {
#if 1
    bp::object io_mod = bp::import("io");
    CHECK_INVARIANT(io_mod, "module not found");
    bp::object iobase = io_mod.attr("TextIOBase");
    CHECK_INVARIANT(iobase, "base class not found");
#else
    // using statics to save an undetermined amount of time results in
    // alarming seg faults on windows. so we don't do it. Keep this here
    // for the moment though in case someone manages to figure that out in
    // the future
    static bp::object io_mod = bp::object();
    static bp::object iobase = bp::object();
    if (!io_mod) io_mod = bp::import("io");
    if (io_mod && !iobase) iobase = io_mod.attr("TextIOBase");
    CHECK_INVARIANT(io_mod, "module not found");
    CHECK_INVARIANT(iobase, "base class not found");
#endif

    df_isTextMode = PyObject_IsInstance(python_file_obj.ptr(), iobase.ptr());
    switch (mode) {
      case 's':  /// yeah, is redundant, but it is somehow natural to do "s"
      case 't':
        if (!df_isTextMode) {
          throw ValueErrorException(
              "Need a text mode file object like StringIO or a file opened "
              "with mode 't'");
        }
        break;
      case 'b':
        if (df_isTextMode) {
          throw ValueErrorException(
              "Need a binary mode file object like BytesIO or a file opened "
              "with mode 'b'");
        }
        break;
      default:
        throw std::invalid_argument("bad mode character");
    }
  }

  /// Mundane destructor freeing the allocated resources
  ~streambuf() override {
    if (write_buffer) {
      delete[] write_buffer;
    }
  }

  /// C.f. C++ standard section 27.5.2.4.3
  /** It is essential to override this virtual function for the stream
      member function readsome to work correctly (c.f. 27.6.1.3, alinea 30)
   */
  std::streamsize showmanyc() override {
    int_type const failure = traits_type::eof();
    int_type status = underflow();
    if (status == failure) {
      return -1;
    }
    return egptr() - gptr();
  }

  /// C.f. C++ standard section 27.5.2.4.3
  int_type underflow() override {
    int_type const failure = traits_type::eof();
    if (py_read == bp::object()) {
      throw std::invalid_argument(
          "That Python file object has no 'read' attribute");
    }
    read_buffer = py_read(buffer_size);
    char *read_buffer_data;
    bp::ssize_t py_n_read;
    if (PyBytes_AsStringAndSize(read_buffer.ptr(), &read_buffer_data,
                                &py_n_read) == -1) {
      setg(nullptr, nullptr, nullptr);
      throw std::invalid_argument(
          "The method 'read' of the Python file object "
          "did not return a string.");
    }
    off_type n_read = (off_type)py_n_read;
    pos_of_read_buffer_end_in_py_file += n_read;
    setg(read_buffer_data, read_buffer_data, read_buffer_data + n_read);
    // ^^^27.5.2.3.1 (4)
    if (n_read == 0) {
      return failure;
    }
    return traits_type::to_int_type(read_buffer_data[0]);
  }

  /// C.f. C++ standard section 27.5.2.4.5
  int_type overflow(int_type c = traits_type_eof()) override {
    if (py_write == bp::object()) {
      throw std::invalid_argument(
          "That Python file object has no 'write' attribute");
    }
    farthest_pptr = std::max(farthest_pptr, pptr());
    off_type n_written = (off_type)(farthest_pptr - pbase());
    off_type orig_n_written = n_written;
    const unsigned int STD_ASCII = 0x7F;
    if (df_isTextMode && static_cast<unsigned int>(c) > STD_ASCII) {
      // we're somewhere in the middle of a utf8 block. If we
      // only write part of it we'll end up with an exception,
      // so push everything that could be utf8 into the next block
      while (n_written > 0 && static_cast<unsigned int>(
                                  write_buffer[n_written - 1]) > STD_ASCII) {
        --n_written;
      }
    }
    bp::str chunk(pbase(), pbase() + n_written);
    py_write(chunk);

    if ((!df_isTextMode || static_cast<unsigned int>(c) <= STD_ASCII) &&
        !traits_type::eq_int_type(c, traits_type::eof())) {
      py_write(traits_type::to_char_type(c));
      n_written++;
    }

    setp(pbase(), epptr());
    // ^^^ 27.5.2.4.5 (5)
    farthest_pptr = pptr();
    if (n_written) {
      pos_of_write_buffer_end_in_py_file += n_written;
      if (df_isTextMode && static_cast<unsigned int>(c) > STD_ASCII &&
          !traits_type::eq_int_type(c, traits_type::eof())) {
        size_t n_to_copy = orig_n_written - n_written;

        for (size_t i = 0; i < n_to_copy; ++i) {
          sputc(write_buffer[n_written + i]);
          ++farthest_pptr;
        }
        sputc(c);
        ++farthest_pptr;
      }
    }
    return traits_type::eq_int_type(c, traits_type::eof())
               ? traits_type::not_eof(c)
               : c;
  }

  /// Update the python file to reflect the state of this stream buffer
  /** Empty the write buffer into the Python file object and set the seek
      position of the latter accordingly (C++ standard section 27.5.2.4.2).
      If there is no write buffer or it is empty, but there is a non-empty
      read buffer, set the Python file object seek position to the
      seek position in that read buffer.
  */
  int sync() override {
    int result = 0;
    farthest_pptr = std::max(farthest_pptr, pptr());
    if (farthest_pptr && farthest_pptr > pbase()) {
      off_type delta = pptr() - farthest_pptr;
      int_type status = overflow();
      if (traits_type::eq_int_type(status, traits_type::eof())) {
        result = -1;
      }
      if (py_seek != bp::object()) {
        py_seek(delta, 1);
      }
    } else if (gptr() && gptr() < egptr()) {
      if (py_seek != bp::object()) {
        py_seek(gptr() - egptr(), 1);
      }
    }
    return result;
  }

  /// C.f. C++ standard section 27.5.2.4.2
  /** This implementation is optimised to look whether the position is within
      the buffers, so as to avoid calling Python seek or tell. It is
      important for many applications that the overhead of calling into Python
      is avoided as much as possible (e.g. parsers which may do a lot of
      backtracking)
  */
  pos_type seekoff(off_type off, std::ios_base::seekdir way,
                   std::ios_base::openmode which =
                       std::ios_base::in | std::ios_base::out) override {
    /* In practice, "which" is either std::ios_base::in or out
       since we end up here because either seekp or seekg was called
       on the stream using this buffer. That simplifies the code
       in a few places.
    */
    int const failure = off_type(-1);

    if (py_seek == bp::object()) {
      throw std::invalid_argument(
          "That Python file object has no 'seek' attribute");
    }

    // we need the read buffer to contain something!
    if (which == std::ios_base::in && !gptr()) {
      if (traits_type::eq_int_type(underflow(), traits_type::eof())) {
        return failure;
      }
    }

    // compute the whence parameter for Python seek
    int whence;
    switch (way) {
      case std::ios_base::beg:
        whence = 0;
        break;
      case std::ios_base::cur:
        whence = 1;
        break;
      case std::ios_base::end:
        whence = 2;
        break;
      default:
        return failure;
    }

    // Let's have a go
    boost::optional<off_type> result =
        seekoff_without_calling_python(off, way, which);
    if (!result) {
      // we need to call Python
      if (which == std::ios_base::out) {
        overflow();
      }
      if (way == std::ios_base::cur) {
        if (which == std::ios_base::in) {
          off -= egptr() - gptr();
        } else if (which == std::ios_base::out) {
          off += pptr() - pbase();
        }
      }
      py_seek(off, whence);
      result = off_type(bp::extract<off_type>(py_tell()));
      if (which == std::ios_base::in) {
        underflow();
      }
    }
    return *result;
  }

  /// C.f. C++ standard section 27.5.2.4.2
  pos_type seekpos(pos_type sp,
                   std::ios_base::openmode which =
                       std::ios_base::in | std::ios_base::out) override {
    return streambuf::seekoff(sp, std::ios_base::beg, which);
  }

 private:
  bp::object py_read, py_write, py_seek, py_tell;

  std::size_t buffer_size;

  /* This is actually a Python string and the actual read buffer is
     its internal data, i.e. an array of characters. We use a Boost.Python
     object so as to hold on it: as a result, the actual buffer can't
     go away.
  */
  bp::object read_buffer;

  /* A mere array of char's allocated on the heap at construction time and
     de-allocated only at destruction time.
  */
  char *write_buffer;
  bool df_isTextMode;

  off_type pos_of_read_buffer_end_in_py_file,
      pos_of_write_buffer_end_in_py_file;

  // the farthest place the buffer has been written into
  char *farthest_pptr;

  boost::optional<off_type> seekoff_without_calling_python(
      off_type off, std::ios_base::seekdir way, std::ios_base::openmode which) {
    boost::optional<off_type> const failure = off_type(-1);

    // Buffer range and current position
    off_type buf_begin, buf_end, buf_cur, upper_bound;
    off_type pos_of_buffer_end_in_py_file;
    if (which == std::ios_base::in) {
      pos_of_buffer_end_in_py_file = pos_of_read_buffer_end_in_py_file;
      buf_begin = reinterpret_cast<std::streamsize>(eback());
      buf_cur = reinterpret_cast<std::streamsize>(gptr());
      buf_end = reinterpret_cast<std::streamsize>(egptr());
      upper_bound = buf_end;
    } else if (which == std::ios_base::out) {
      pos_of_buffer_end_in_py_file = pos_of_write_buffer_end_in_py_file;
      buf_begin = reinterpret_cast<std::streamsize>(pbase());
      buf_cur = reinterpret_cast<std::streamsize>(pptr());
      buf_end = reinterpret_cast<std::streamsize>(epptr());
      farthest_pptr = std::max(farthest_pptr, pptr());
      upper_bound = reinterpret_cast<std::streamsize>(farthest_pptr) + 1;
    } else {
      CHECK_INVARIANT(0, "unreachable code");
    }

    // Sought position in "buffer coordinate"
    off_type buf_sought;
    if (way == std::ios_base::cur) {
      buf_sought = buf_cur + off;
    } else if (way == std::ios_base::beg) {
      buf_sought = buf_end + (off - pos_of_buffer_end_in_py_file);
    } else if (way == std::ios_base::end) {
      return failure;
    } else {
      CHECK_INVARIANT(0, "unreachable code");
    }

    // if the sought position is not in the buffer, give up
    if (buf_sought < buf_begin || buf_sought >= upper_bound) {
      return failure;
    }

    // we are in wonderland
    if (which == std::ios_base::in) {
      gbump(buf_sought - buf_cur);
    } else if (which == std::ios_base::out) {
      pbump(buf_sought - buf_cur);
    }
    return pos_of_buffer_end_in_py_file + (buf_sought - buf_end);
  }

 public:
  class istream : public std::istream {
   public:
    istream(streambuf &buf) : std::istream(&buf) {
      exceptions(std::ios_base::badbit);
    }

    ~istream() override {
      // do nothing.
      // This used to do:
      // if (this->good()) this->sync();
      // but that caused problems if the underlying file had been closed
      // (see github #579) and really doesn't seem necessary for what we're
      // doing.
    }
  };

  class ostream : public std::ostream {
   public:
    ostream(streambuf &buf) : std::ostream(&buf) {
      exceptions(std::ios_base::badbit);
    }

    // overload that takes ownership of the streambuf ptr
    ostream(streambuf *buf) : std::ostream(buf), m_buf(buf) {
      exceptions(std::ios_base::badbit);
    }

    ~ostream() override {
      if (this->good()) {
        this->flush();
      }
      delete m_buf;
    }

   private:
    streambuf *m_buf = nullptr;
  };
};

// std::size_t streambuf::default_buffer_size = 1024;

struct streambuf_capsule {
  streambuf python_streambuf;

  streambuf_capsule(bp::object &python_file_obj, std::size_t buffer_size = 0)
      : python_streambuf(python_file_obj, buffer_size) {}
};

struct ostream : private streambuf_capsule, streambuf::ostream {
  ostream(bp::object &python_file_obj, std::size_t buffer_size = 0)
      : streambuf_capsule(python_file_obj, buffer_size),
        streambuf::ostream(python_streambuf) {}

  ~ostream() noexcept override {
    if (this->good()) {
      this->flush();
    }
  }
};
}  // namespace python
}  // namespace boost_adaptbx

#endif  // GUARD