Subject: Re: [boost] [c++11]
From: Niall Douglas (ndouglas_at_[hidden])
Date: 2013-06-17 17:54:37
> I wasn't suggesting making the library strictly C++03-compatible. My main
> point was that the library has to be compatible with _todays_ and even
better
> - _yesterdays_ compilers to be actually useful. My choice of "reasonable"
time
> frame is 3-5 years, as I mentioned. That includes VS 2008 and VS 2010,
which
> already had some C++11 features, including rvalue references. Making VS
2013
> absolute minimum is a no-go, IMHO.
I don't think anyone is making VS2013 an absolute minimum - rather they're
making variadic templates an absolute minimum. Something which any other
compiler than VS has had for years, and something which unlike many, if not
most other C++11 language features is something which lets you write a ton
of stuff you simply can't write without them.
Now, nothing in proposed Boost.AFIO uses variadic templates in a way not
replaceable with preprocessor programming. However it very /nearly/ did: I
had this neato bit of evil variadic instantiation which let you convert an
unknown set of trailing args into a std::tuple<> for storage, then /unpack/
said std::tuple<> back into an unknown set of trailing args, all with
perfect forwarding. Effectively it was a compile-time std::function<>
implementation.
As proposed Boost.AFIO is also a closure execution engine where that closure
can take arbitrary args and return arbitrary returns, that evil variadic
template machinery was really useful because it lets you skip using
std::function<> with its implied virtual function call. As proposed
Boost.AFIO currently nests, in the worst case scenario, a fully bound
std::function<> calling a partially bound std::function<> calling a
partially bound std::function<>, that's three nested virtual indirect
function calls. Recent Intel CPUs can prefetch and branch predict two
nesting levels, but they can't cope with three, so performance takes a real
nose dive and hence one of the reasons we currently see a peak 80k IOPS.
That evil variadic template machinery gets rid of that problem. And its only
replacement is std::function<> - you *cannot* implement that functionality
(pure compile time std::function<>) without variadic templates because you
can't properly unpack them without.
None of this matters for proposed Boost.AFIO. But it certainly matters for
the forthcoming C++11 only libraries coming soon to Boost, because they make
/very/ heavy usage of functionality only possible with variadic templates.
That means that they will require, as an absolute minimum for VS, VS2013 or
Nov 2012 CTP. Other C++11 features one can make do without without too much
cost, but losing variadic templates means a ton of extra source code
complexity, much longer compile times and chopping out a lot of what C++11
enables which simply can't be done in C++03.
I suspect that those libraries are probably going to force a new "variadic
template support is mandatory" category of Boost libraries, probably sooner
rather than later. I agree it's going to be frustrating for those stuck with
older compilers: I pity those whose employers force MSVC6 on them to this
very day.
Niall
--- Opinions expressed here are my own and do not necessarily represent those of BlackBerry Inc.