From: Paul Mensonides (pmenso57_at_[hidden])
Date: 2005-05-09 05:06:12
> -----Original Message-----
> From: boost-bounces_at_[hidden]
> [mailto:boost-bounces_at_[hidden]] On Behalf Of Andrei
> Alexandrescu (See Website For Email)
> > What high integration are you referring to here? I don't
> see much of
> > a high integration between template metaprogramming and the rest of
> > the language. In fact, there is very little integration. Instead,
> > it's more like they coexist without interferring with each
> other, and communication is entirely one way.
> > Even with the type system, the pattern-matching capabilities are
> > limited to (in
> > essence) syntactic matching rather than more general property
> > matching. You have to bend over backwards to do that
> (read: SFINAE).
> It's not syntactic any day of the week. Not at *all*.
There is a near direct correlation between syntax and its capabilities. I don't
mean that it literally is syntactic matching. It can only match those type
properties that can be formed syntactically--and it cannot even match all of
those. Furthermore, the primary utility of that pattern matching is with
templates themselves, not the "non-template part of C++." The template
mechanism alone (meaning without major hacks) cannot do even the simplest
manipulation on the non-template part. Even with major hacks, it can still only
do a very limited amount of such manipulation. Other than a few commonalities
(like name binding), there is virtually no integration between the two. The
non-template part can statically cause the instantiation of a template and
dynamically use it, and the template mechanism can statically introduce
non-template code, but the template mechanism cannot statically invoke the
non-template part, nor can the non-template part dynamically cause the
instantiation of a template. There are only two types of communication, and,
from the template side in particular, it is only one-way communication. That
isn't integration, that is a layered separation. For metaprogramming, the
template mechanism is usable (and a lot of interesting things can be done using
it), but it is not robust.
> Templates perform genuine recursive pattern matching. They
> feed on types and constants and offer types and constants
> back. I see good integration there.
The mechanism is systematic, but that doesn't say anything about its integration
with the non-template part of C++.
> If, on the other hand,
> you want to discuss the limited amount of introspection C++
> offers, that's a different subject.
No, it really isn't a different subject. We are talking about metaprogramming,
not simplistic type substitution ("regular" template usage). There are two
purposes of metaprogramming: introspection and generation (i.e. generative
metaprogramming). The template mechanism is only mediocre at both.
BTW, I'm not saying that the preprocessor does better overall. The preprocessor
is terrible at introspection, but (comparatively) excels at many types of
generation. There are serious strengths and weaknesses in both mechanisms.
> > What idioms are those? This, BTW, from the perspective of
> a user, not
> > library internals. What is really the difference,
> idiomwise, between
> > fold(list, state), fold<list, state>, and FOLD(list, state)?
> I am referring to idioms that ask you to #define certain
> things and then invoke macros that will use them (or worse,
> include files that will use
> them!) and then #undef them
Which is no different than limiting scope in any other context. How many
people, for example, complain about the internal linkage of local structures
which could otherwise be used for function objects (etc.)? When you write such
a #define (in the context of metaprogramming), you are defining a function,
which, in the case above, is passed to a higher-order algorithm. The #undef
merely explicitly controls the lifetime of the function. In some ways, the
preprocessor treats such 'functions' as first class objects better than the core
> , or that have you define files
> that rely on things like "I am iterating now through this
What is wrong with that? More specifically, what is wrong with viewing the
preprocessing pass as a sort of script execution? That is exactly what it is.
In fact, parsing non-preprocessor C++ is itself a sort of top-to-bottom script
execution (e.g. function prototypes, predeclarations, etc.). Thinking of it any
other way is, with C++ in its current form, fundamentally flawed.
> all that comma handling,
There is actually very little comma handling in real preprocessor code.
Granted, there is some, but with variadics, the difficulty in handling those
nearly disappears--as I showed you before. Having to write TYPELIST(int,
(std::pair<int, int>), double) is not that much of a burden.
As I've also said before, the preprocessor's primary strength is its
disassociation with the underlying language--for generation, the syntax and
semantics of the generated language interfere**. That is also its primary
weakness--it cannot do any realistic introspection into the syntax and semantics
of the underlying language. Commas are a mildly annoying case entirely because
a particular part of the syntax of the preprocessor is colliding with a
particular part of the syntax of the underlying language.
** This is both a good thing and a bad thing. It is a good thing because you
can directly express the generation. Syntactic (and semantic) integration leads
to roundabout generation. On the other hand, syntactic integration allows for
uniform syntax--which, in turn, allows for infinite "meta-levels" (i.e. metacode
that operates on metacode (etc.) which operates on runtime code).
> the necessarily long names,
For library facilities, yes. You need to apply a namespace. That isn't really
that different than normal C++, except when actually designing the library
facilities. In normal client C++, you can explicitly qualify names or manually
bring them into scope. You can do the same things with the preprocessor.
(There are also a variety of ways of locally shortening names without defining
> the dynamic binding (a historical mistake that keeps on
> coming again and again in PL design! yuck!) etc.
...which hardly matters when there is no lexical scoping or namespaces/packages.
Dynamic versus static binding is irrelevant when a symbol unambiguously refers
to a distinct entity.
> IMHO those idioms' place is in programming language history
> books as "awkward first attempts".
On of the most effective aspects of the preprocessor, BTW, is that there is no
fundamental distinction between input/output and argument/return value (i.e. A()
produces output, but the A() in B(A()) produces a return value). This lack of
distinction eliminates a great deal of cruft in the generation process. If you
do it in another language (i.e. external tool), you have to distinguish between
them--which is more verbose. The preprocessor is naturally more
domain-specialized for generation than existing external (general purpose)
tools. Granted, you could define such abstractions in another language (or
define another language), but that would eventually lead you nearly full
circle--you'd end up with a custom preprocessor that does similar things. Yes,
by throwing out other important considerations, you could make it better.
Now, improving the preprocessor (as you and Joel both mention) is a much better
alternative. However, I, for one, won't use non-standard extensions unless
there is 1) no other way or 2) using them is drastically better than the
alternative. There isn't much that you can do to the preprocessor to make it
better to that extent. What that means is that the improvements need to be
standardized--and the main thing preventing that is the common "wisdom" that the
use of macros is evil--for whatever reasons (or lack of reasons). In other
words, your approach leads to a circular non-solution. Denouncing the
preprocessor is not going to lead to a better preprocessor. Improving the
preprocessor requires first swaying the general opinion on preprocessor-like use
(the hard part--an upcliff battle) and then move to add or update features (the
Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk