Re: [boost] [guidelines] why template errors suck

Subject: Re: [boost] [guidelines] why template errors suck
From: David Abrahams (dave_at_[hidden])
Date: 2010-09-27 15:20:57

At Mon, 27 Sep 2010 11:18:47 -0800,
Robert Ramey wrote:
> 
> My experience was in trying to apply this to the serialization
> library.

That is, admittedly, a tough one.

> My example was
> 
> ar << t;
> 
> where t is of serializable type T  and ar is an instance of an Archive class
> 
> The "archive concept" would admit a wide variety of semantics.
> The archive classes included in the library include code for deep
> copy/recover of pointers to polymorphic classes.  But in another
> context, such as one which only displays data for debugging,
> one might implement this functionality as just a top level display
> so that he can make a header only version.  

Well, there's license to break almost any rule when debugging :-) Try
to come up with a non-debugging example.

> So here one has two useful widely varying implementation symantics
> with exactly the same syntax requirements.  Is it wrong to use
> concepts here?

No, but you'll need decide what the abstraction's (abstract) semantics
are if you're going to document/understand the code as anything more
than an English-language translation of its exact implementation
(which would be pretty useless as documentation).  As soon as your
documentation is something more than that, there are abstract
semantics involved.  You can build and use concepts that have no
attached semantics, but you can't say anything interesting about what
the code using those concepts "does."

> Is it wrong to make such a library?  Is is wrong, to "hijack" a
> types concepts for unintended uses?

I don't have a moral judgement about this.  The only right/wrong
position I'm taking here is about the interpretation of the term
Concept.  I happen to know what it was intended to mean, so I can
speak with authority on that.

> I concluded that it not really possible to specify symantics and any
> formal or rigorous way without actually writing the code itself.  

I recall a conversation on this list where someone (named Sebastian?)
successfully, and impressively, described a semantic relationship
required between the load/save operations of a read/write archive
pair.  

The fact that it's hard to do doesn't make it a pointless affair.
Until you've done it, you have at best an "intuitive feel" for the
problem your library is solving.

> It doesn't make the idea of validation of template parameters
> useless, it just puts a limit of one can realistically expect from
> it.

That's unrelated to the rest of the conversation, because as I have
stated many times, validation of template parameters by computer can
never check semantic constraints.  It's equivalent to solving the
halting problem.

> >> Of course the nomenclature doesn't help any either.
> >
> > Sorry, I don't know what you're talking about here.
> 
> I was referring to the usage of the word "concept" in this context.  I
> found this to be very misleading and not at all descriptive of the
> the function of "checking template parameters at compile time"

Because that's not what a concept is.  That's just one of the most
important things you can do with a concept.  Concepts are about
mathematical abstraction.

-- 
Dave Abrahams
BoostPro Computing
http://www.boostpro.com