Monday, December 19, 2016

C++ "MetaProgramming" and Why C++ Should Die

I saw a pretty awesome accomplishment on Twitter today - this fellow write a little raytracer that does all its calculations at compile time.
 
 
Pretty impressive, even if it takes a while. It's all done with templates and "metaprogramming", which is a fancy term used to excuse the complexity of programming both the computer AND the compiler.
 
No, I don't like it. And it's a great example of why C++ is done and needs to die.
 
I've loved C++ for a long time. I've encouraged friends learning in University without realizing the true pain they were experiencing. You see, I've been blissfully ignorant of how far things had gone for a long time. For the most part I ignored C++11 and C++14, until a recent project forced me into the deep end, and I got the O'Reilly book out and started reading.
 
I was pretty horrified, in general, but we'll focus on this particular aspect.
 
So the article above is about this fellow learning these new concepts with an ambitious and fairly impressive task, inspired more or less by this example:
 
Take this example:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
template<int base,int power>
struct Pow
{
const static int result = Pow<base,power-1>::result*base;
};
 
template<int base>
struct Pow<base,0>
{
const static int result = 1;
};
 
int main(int argc,char *argv[])
{
return Pow<5,2>::result;
}
 
If we look into the assembly file produced we just see the constant 25 being written to a register.
 
 
 
mov eax, 25
 
Your formatting is crap, Blogger.
 
Anyway, so what happens there is that main() invokes a template (Pow<5,2>), which generates a recursive chain of structures, each containing the next power, until the power is zero and the final template is invoked. The compiler runs through all this, and the final result is that single assembly instruction that generates a single const value "25" (5^2 = 25).
 
Fans of this style of programming point at the amazing efficiency of this resulting code as a major win. It's so much faster, they will tell you, than running the code the old fashioned way. But I call bullshit. Because in "the old way" we wouldn't have done that anyway. Not if performance mattered. Do you know what'd we do? Hell, this goes all the way back to "C", not even one plus!
 
const static int result = 25;
 
"Oh! Oh! Oh!" cries the peanut gallery. "But the computer didn't calculate that for you!!"
 
Of course it did. We did it offline. Or we did it in a separate program. Or we used a calculator. Or we did it at startup and cached the value. Or in the worst case, maybe we used a code generator. (Deliberately ignoring the fact that this example is very simple and didn't need code at all).
 
But! Isn't using a code generator for the most complex cases exactly what we did here? It's just built into the language now, isn't it?
 
Well, yes and no.
 
Yes, you essentially used a code generator to calculate the problem and reduce the code to the important single constant. But this is about the most complicated, difficult to debug way of doing it that I could have imagined.
 
First of all, you just littered your namespace with three different Pow structures. You didn't need the other two, but the compiler did, and they exist. It was a lot more expensive for the compiler to calculate all other structures and then decide what was really needed than just about any other technique would have been, which means your compile time is increased (substantially, in fact, depending on how many Pow's you need and how deep they have to go!) And suppose you typoed in the base,0 template? Well, then your error code is going to have to reflect the entire chain. In this case, it's a short chain of just three entities, and the error is a single line per entity, since it's very simple.
 
$ gcc test.cpp -otest
test.cpp:10: error: `into' does not name a type
test.cpp: In instantiation of `Pow<5, 1>':
test.cpp:4:   instantiated from `Pow<5, 2>'
test.cpp:15:   instantiated from here
test.cpp:4: error: `result' is not a member of `Pow<5, 0>'
 
But real life templates tend not to be so simple. And because of the nature of the templates, you can trigger errors simply by specifying the wrong type of argument to a parameter (for instance, forgetting to std::move can break some templates). The result can be pages of template chain errors, making troubleshooting difficult. And indeed, that is what our experimenter found:
 
This was the first time I had tinkered with metaprogramming so it was pretty hard at first, basically once one template evaluation fails you get a huge chain of thousands of failures.
 
The entire direction of the language's development seems to have shifted towards programming the compiler to generate constants for you at compile time. This is a good thing and we've often done it in the past, but it's not ALWAYS the right answer, and today it's being taken to ridiculous extremes. Some of the things I've seen mean that the compile-time ray tracer isn't even that outrageous to me, I've seen things attempted that feel on the same level of complexity. And I don't believe that we should be doing that.
 
Why not?
 
Well, this impacts you in several ways:
 
-compile time is longer. How many of those complicated template chains result in the single constant that the example above shows? You've built the code hundreds of times and never changed that value, have you? Make it a damned constant and save the time.
-typos in the code are MUCH harder to understand. If you've done std or boost template programming, you already know what I mean. If you haven't yet, you will. If you're a god who never makes mistakes, go back to cartoon land. This costs time - a simple typo that may be as simple as a missing modifier goes from a 10 second change to a minute or more, just to determine what line the error actually occurred on. I know people who switch to a different, non-production compiler for testing, just because the error messages are less verbose (meaning an entire compile phase is wasted). This time adds up substantially.
-learning time is longer. If you're using your own complex template chains (in addition to std, boost, or other common ones), then you have a larger and deeper codebase for a new developer to have to come to terms with -- and the two issues above are not going to help with that. Since most developers on most projects are thrown in with little more than an incomplete wiki and a promise to get around to guidance, you'd think that simple, easy to follow code would have some value.
 
I'm reminded of an old quote by DadHacker (http://www.dadhacker.com/blog/):
 
The future of computing is its own past, mashed-up and remixed by young'uns who have yet to fear the dark corners, the places where us old farts went in with similar bushy-tailed attitudes and came out with ashen-faced, eyes barn-door wide and with fifty new words for "pucker." Heed us. The stove is hot if you touch it. The stove is not only hot, it will incinerate your soul. At some point you will want to make pancakes or wash dishes for a living rather than run another build or merge another check-in or fix another bug...
-Dadhacker