Re: [boost] Re: static sized strings

From: Rob Stewart (stewart_at_[hidden])
Date: 2004-05-13 13:14:56

From: "Reece Dunn" <msclrhd_at_[hidden]>
> Rob Stewart wrote:
> >From: David Abrahams > John Nagle writes:
> > >
> > > >     If you're willing to accept a call to "new", you can just
> > > > use std::string and reserve some initial size.
> > >
> > > It's not the same thing.  You might only be willing to accept a call
> > > to "new" if things go beyond a certain length.  That's *usually* the
> > > only reason I'd ever have a fixed-size buffer.  Very occasionally
> > > I'll be in code that where I don't want to throw an exception, but
> > > then a string that responds to overflow by throwing is no better.
> >
> >char_string can throw a more meaningful exception than
> >std::bad_alloc.
> >
> >In the small string optimization or std::string with reserve
> >approaches, the buffer grows when the input is too large, and
> >other, downstream code may overflow or behave badly.  Stopping
> >the overflow early may be better and having an automated
> >mechanism for preventing overflow is wise.
> 
> There is not meant to be any buffer allocation. char_string is a 
> fixed-capacity string that will not write beyond the length of that buffer.

I know that's what you meant, but Dave was suggesting using the
small string optimization as a means to get the best of both
worlds.  It would give the no-allocation behavior of your
proposal plus the ability to grow via the free store when that
initial allocation was exceeded.  There could even be a template
parameter to control whether to allow the buffer to grow beyond
its initial, stack-based allocation.  One way would fit your
idea, another would use the small string optimization.

> > > > So you don't need that capabiilty in char_string.  char_string is
> > > > for situations when you don't want to invoke dynamic allocation at
> > > > all.  In desktop applications, that's not too common, but in real
> > > > time work, inside operating systems, and in embedded applications
> > > > it's not unusual.
> > >
> > > I'm saying, it's almost always better to degrade performance
> > > gracefully as the program's input grows than it is to introduce an
> > > execution discontinuity like an exception (if possible).  In many of
> > > those scenarios you've named, exceptions, like dynamic allocation have
> > > been banned anyway for similar real and/or imaginary reasons.
> >
> >There's probably room and need for both types of strings.
> 
> I don't intend the char_string buffer to be increased. I intended that the 
> operations on the buffer would not go beyond the buffer (which would lead to 
> overflow) and also that going over the buffer would not generate an 
> exception.

I understand that's your intention, but Dave's suggesting that it
would be better if instead of dropping information on the floor
silently or throwing an exception when there's no room for a null
terminator, a less efficient, but correct class might be
appropriate.  (Actually, I think I'm putting words in Dave's
mouth.  IIRC, he was only addressing the "exception on
insufficient space for a terminator" matter.)

Perhaps what you meant to say was that you think a critical
design criteria of char_string is that it should *never* grow.  I
didn't read that from the above.

> It may be beneficial to have an error policy to determin the behaviour in 
> the event of an overflow. I am not advocating this approach, just putting it 
> up for discussion.

That's why I suggested there's probably room for both.  A policy
might be a good way to handle it.  A separate class could do it,
too.

It's a matter of use cases.  Are there cases in which the string
should never exceed its initial capacity?  What will happen if it
does (and it is designed to work correctly if it does)?  Are
there uses cases in which the string should have an initial
capacity for efficiency reasons but should grow if the input
exceeds its initial capacity?

I can imagine uses for both, but I'm not certain that there are.

> The only exception I can see keeping is the one in the constructor to test 
> for a buffer of 0, but this would be better re-implemented as a compile-time 
> check using mpl code, e.g.:
> 
> template< int n, ... >
> class char_string
> {
>    class ok{ typedef char value; };
>    class zero_buffer_error{};
> 
>    typedef mpl::if_< mpl::equal< n, 0 >, zero_buffer_error, ok >::value 
> value;
> };

Yes, that's a good idea.

-- 
Rob Stewart                           stewart_at_[hidden]
Software Engineer                     http://www.sig.com
Susquehanna International Group, LLP  using std::disclaimer;