C++ Weekly - Ep 154 - One Simple Trick For Reducing Code Bloat
A friend sent me the link to this c++ weekly episode and I followed that clickbait. Just recently I met the topic of compiler-generated functions at work so the episode caught some special interest.
So what is it about already? In a nutshell, the episode is about the different amount of compiler-generated assembly output depending on the (non-)definition of the destructor. (I just wonder. Probably this will apply to other non-generated functions as well.)
See this
struct S {
std::string s;
// ~S();
};
int main()
{
std::vector<S> vec;
vec.emplace_back();
vec.emplace_back();
}
vs. changed S
destructor here
struct S {
std::string s;
~S();
};
S::~S() {}
Essentially, the episode noted the following:
- The second case generated more assembly code, especially it contained an explicit calls to the destructor function.
- This did not always happen, e.g. if
S
contained a native type the destructor call was optimized away (with-O3
). - The rule of zero was advertised.
I refrain from getting into more details of what happened when, since that would not be TL;DR-ish and also because the results have changed since the episode was published almost two years ago.
For me, the episode was a starter to dive into the trivially-desctructible topic which crossed my way before, I believe also in one of Jason’s episodes. Was that about multiple destructor’s feature in C++20?
My personal takeaways:
- It’s a difference if a class is cleaned up by excecuting a function (i.e. destructor) and then freeing the memory of just freeing the memory, saving a possibly empty function call. Reading it that way, that appears obvious now. The insight for me is apparently that it is possible to free an object without calling any destrcutor.
Just freeing the memory, as I understand it now, this is what trivially destructible means.I’m not sure about this any more, I read up on the topic a bit and by definition, e.g. thestruct S
with astd::string
is never trivially destructible sincestd::string
is not trivially destructible.- Never underestimate the power of compiler optimization. Even what’s not optimized today, might get optimized tomorrow.
Maybe it’s worth to comment on the last point. In code reviews I’ve seen people argue relentlessly for optimizing code in favor of assumed perfomance (and possible C++ developers are (in)famous for it). While I would agree that in this example (or not ever if not needed, for that matter), you should define an empty destrutor, don’t design your code around performance. Clarity for example is often favorable and it’s not clear that even a performance hit that exists with proof won’t get an optimization in the future.
Hope that was TL;DR enough and not too much of a full blog post ;) I also hope it was useful to you.
References:
- C++ Weekly episode #154: https://youtu.be/D8eCPl2zit4
- https://en.cppreference.com/w/cpp/language/destructor#Trivial_destructor
- C++17 standard, §6.8.5 Object Lifetime
Note to self: I’d really like to follow up on §6.8.5 and what it means exactly.