MFILE removal

[Nicolas Alvarez]

There are only two things left using MFILE: the GUI RPC client library
(which I intend to completely replace some day), and the client_state.xml
saving code.

All the code that writes XML to make the client_state.xml file was already
converted to iostreams. Only the "top-level" step, creating the file
(write_state_file), is still using a MFILE/MIOFILE. Then that MIOFILE is
wrapped in an ostream when passed to write_state().

write_state_file creates an MFILE because that way all the data is buffered
in memory first, and only written to disk when closed. But I don't see why
this is needed. It's writing to client_state_next.xml, closing, and then
renaming it to client_state.xml. I don't see any potential problem in
writing the _next file "incrementally" instead of first buffering it all.

I attach a proposed patch to make write_state_file use a std::ofstream.
I'd like some pre-commit reviewing...

The diff may be a bit hard to follow. It may be better to apply the patch
locally and review the resulting code instead of looking at the patch.

[Der Meister]

Nicolas Alvarez wrote:
> write_state_file creates an MFILE because that way all the data is buffered
> in memory first, and only written to disk when closed. But I don't see why
> this is needed. It's writing to client_state_next.xml, closing, and then
> renaming it to client_state.xml. I don't see any potential problem in
> writing the _next file "incrementally" instead of first buffering it all.

This might have some implications on the performance, mainly increase
disc-load, if the code writing to the ofstream uses std::endl or
std::flush more than once. As long as this isn't the case it shouldn't
be a problem as std::ofstream and the underlying implementation should
already use some buffering which may even be better then the one
provided by MFILE (regarding to memory usage).

[Nicolas Alvarez]

An ofstream is of course buffered, but that buffer will probably not fit the
entire client_state.xml file.

Synecdoche isn't using std::flush or std::endl anywhere, except when writing
to std::cerr.

[David Barnard]

On Thu, Oct 22, 2009 at 5:37 PM, Nicolas Alvarez
<nicolas...@gmail.com> wrote:
> An ofstream is of course buffered, but that buffer will probably not fit the
> entire client_state.xml file.

I think the size of the buffer can be changed using the filebuf.

[Der Meister]

I don't think that adjusting the buffer size is necessary but using
std::endl or std::flush would flush the buffer and doing so after
(almost) each line would be really terrible. That's why I've asked.