Boost logo

Boost :

From: Jonathan Turkanis (technews_at_[hidden])
Date: 2004-09-06 12:07:34

"John Torjo" <john.lists_at_[hidden]> wrote in message:

Thanks for the review!

> tutorial.html:
> " The Iostreams Library allows users to create a standard input stream
> or stream buffer by defining a classes with a single member function
> read and several member types."
> Q: What do you mean, by 'several member types'?

Originally it was a char type a traits type and an i/o category tag. Now it's
just the char type and category. I was hoping to avoid listing the member types
in the tutorial, because by deriving from one of the convenience base classes
source, sink, etc.,

- unless you're writing a filter or resource templated on the char_type, you can
avoid dealing with char_type

- unless you need optional behavior beyond Closable and Localizable, you don't
need to worry about the i/o category tag.

Would you suggest a short description of these types in the tutorial, or just a
comment that beginners don't need to worry about them, plus a link to a detailed

> "Buffering and the ability to put back characters come for free."
> Q: Can I putback multiple chars, or just one?

The size of the putback buffer can be specified when the stream or stream buffer
is opened using open() or a constructor which specifies a resource -- That's if
you're using stream_facade or streambuf_facade.

If you're writing a filter, say:

       struct my_filter : input_filter {

            template<typename Source>

            int get(Source& src)


                 ... how many chars can I putback to src???



Here, it's only safe to put back one character. Currently source is a
streambuf_facade with a putback buffer size of 4 (configurable by a macro).
However, according to one of the optimizations I am considering, src could
sometimes be an arbitrary stream or streambuffer which a user has added to the
end of the filter chain.

> 3.4 Output Filters
> Here your examples use 'filtered_streambuf', while in the Reference
> section, you use 'filtering_streambuf'. Is this a typo?

Yes. Thanks. The current name is 'filtering_streambuf'; in the past I've used
'filtered_streambuf' and 'filter_streambuf'. What's your preference?

> examples.html
> I love the presidential filter ;)

I was hoping someone would notice that one ;-)

> policy_based_streams.html:
> I think you should visually show that streambuf_facade owns the policy
> (the source). I had to look through the code to realize that. I have
> later realized that you did say at the beginning " by delegating to a
> contained instance of a policy class". Maybe emphasize the "contained
> instance" (bold or something)?

Larry Evans's comments convinced me that I need a prominent section on lifetime
management issues. Here's what I suggested:

"Jonathan Turkanis" <technews_at_[hidden]> wrote:

> Lifetime Management of Filters and Resources. This material should either go
> under 'Concepts' in the User's Guide, or have its own section.
> A. By default, filters and resources are stored internally by value and
> be copy constructible. The reason for pass-by-value (really by const ref) is
> exception safety. This was accidentally omitted from the most recent rewrite
> the rationale.
> B. It is unspecified whether filters and resources which are copy
> constructible have deep copy semantics.
> C. Standard streams and stream buffers are models of Resource; they are
> always stored by reference.
> D. The library may make an arbitrary number of copies (usually just one)
> a filter or resource, but only one is stored, and no copies are made after i/o
> begins
> E. Filters and resources can be stored by reference using the function
> boost::ref (see This is useful in two types of
> 1. The filter or resource type is not copy-constructible
> 2. Cases like the one you present below, in which you keep an external
> instance of a filter, and want changes to this external instance to be
> directly in the filtered i/o.
> F. Filters and resources must free all associated resources (in the usual
> sense) either:
> 1. When the stored copy is destroyed, or
> 2. If the filter or resource type models Closable
> ( and i/o has commenced, when the function
> boost::io::close() is called.
> I believe these principles are fairly intuitive and easy to work with, but
> need to be spelled out in detail somewhere, and probably addressed in the
> examples.

> filtering_streambuf.html:
> Again, by looking at the code, it seems that you own the streambufs. Is
> that so? If it is, please state it in the docs (this is more of a
> general issue - it applies to other classes as well).
> I am worried about this, since I want to know what I need to destroy
> manually, and what is destroyed automatically for me.

All the internal streambufs are freed automatically by filtering_streambuf. If
you write:

     filtering_ostream out;

     out.push(base64_encoder()); // user-defined filter


You're responsible for freeing my_stream, if it was dynamically allocated.

The library went through a large number of conventions for passing filters and
resource (by pointer, by value,...). I forgot to discuss this in the latest
rewrite of the rationale. The main point was exception safety.

> General issues:
> I'm not sure if mixing streambuf and stream concepts is such a good idea.
> As I've seen in your examples, you only use streams. Tnus, you don't use
> streambufs directly.
> While having streambuf_facade is DEFINITELY a good idea, I think you
> should have an "Advanced" section, where you talk about the streambufs.
> The rest of us are quite happy with having and using stream_facade.

I think I need to add -- very early in the documentation -- a brief overview of
the standard iostreams library, in which I

1. indroduce the types streamsize and streamoff

2. mention that basic_streambuf is the component which represent a connection to
a data source/sink

3. explain that streams represent the formatting layer of standard i/o, and that
each stream contains a pointer to a streambuf to which its i/o functions and
operators delegate.

I can then explain that streambuf_facade is the principle component provided by
the library, that stream_facade and filtering streams are really only
convenience wrappers, but that the tutorial and examples will mostly use the
wrappers since these are more familiar to existing users of the standard library
and will probably be used most in practice.

How does that sound?

> -------------------------------------------
> * What is your evaluation of the potential usefulness of the library?
> Very useful
> -------------------------------------------
> * Did you try to use the library? With what compiler? Did you
> have any problems?
> VC7.1 Compiled some examples - they worked ok, after I slightly modified
> the jamfile.
> (which makes me not like bjam even more - it makes it so hard to build
> even trivial things)

I'm no expert on Boost.Build -- I got the Jamfile to work on my system by trial
and error.

Could you explain why you added the lines



> I think this library should be ACCEPTED.

Great! Thanks.

> Best,
> John


Boost list run by bdawes at, gregod at, cpdaniel at, john at