In nearly every programming languages there are often many different possibilities to write code that in the end does exactly the same thing. In most cases we want to use the version that is the easiest to understand to a human reader.
Let's say we have a vector of integers and want another one with all elements being the square of the elements in the first one.
[4, 1, 7] -> [16, 1, 49]
Sure, one could use gotos for this, but I guess nobody in their right mind would do this voluntarily, unless for trolling:
vector<int> squareVec1(const vector<int>& v)
{
vector<int> result;
result.reserve(v.size());
auto it = begin(v);
loopBegin:
if (it == end(v))
goto loopEnd;
result.push_back(*it * *it);
++it;
goto loopBegin;
loopEnd:
return result;
}
On the first look you have no idea what this code is doing. You have to kind of interpret it in your head and follow the control flow manually to find out what's going on.
The next more readable step would be to use a loop as control structure.
vector<int> squareVec2(const vector<int>& v)
{
vector<int> result;
result.reserve(v.size());
auto it = begin(v);
while (it != end(v))
{
result.push_back(*it * *it);
++it;
}
return result;
}
This is bit better, at least you can immediatly see that there is some kind of loop, but most people would probably use a
vector<int> squareVec3(const vector<int>& v)
{
vector<int> result;
result.reserve(v.size());
for (auto it = begin(v); it != end(v); ++it)
{
result.push_back(*it * *it);
}
return result;
}
Here you can immediatly see that the algorithm iterates ofter the elements of v
but you still have to read the whole line for (auto it = begin(v); it != end(v); ++it)
until you know that every single element is used and not e.g. every second, since the increase could also be it += 2
or something else instead of ++it
.
vector<int> squareVec4(const vector<int>& v)
{
vector<int> result;
result.reserve(v.size());
for (int i : v)
{
result.push_back(i*i);
}
return result;
}
This time the for
line already tells the reader that probably every element of v
is used, but still only probably. One still has to look into the body of the for loop and look for if
, continue
or even break
statements to really know that result
is guaranteed to have the same size as v
in the end.
Many people stop here, but we can do better in terms of readability ease.
OK, how can we express more clearly without explicit comments what our code does, i.e. make it self explaining?
vector<int> squareVec5(const vector<int>& v)
{
return accumulate(begin(v), end(v), vector<int>(),
[](vector<int> acc, int i)
{
acc.push_back(i*i);
return acc;
});
}
We use std::accumulate
. Everybody reading this knows without thinking, that every element of v
will be iterated over and that these values probably will be used to generate one resulting value. But apart from the performance problem of this solution, the "loop header" still does not say something about the shape of the result.
Eventually this is our final form.
vector<int> squareVec6(const vector<int>& v)
{
vector<int> result;
result.reserve(v.size());
transform(begin(v), end(v), back_inserter(result), [](int i)
{
return i*i;
});
return result;
}
std::transform
tells the reader at one glance that v.size()
elements will be pushed into result
and that every single element from v
will be used to generate exactly one new element for result
.
Now one just has to look at return i*i
and he directly knows everything.
This is much easier than to decypher a for loop every time.
Range-based for vs. <algorithm>
A for loop also beginning with for (int i : v)
could do something totally unrelated to std::transform
. E.g. it could implement a filter:
for (int i : v)
{
if (i % 2 == 0)
result.push_back(i);
}
Here a more expressive version would be:
copy_if(begin(v), end(v), back_inserter(result), [](int i)
{
return i % 2 == 0;
});
transform
and copy_if
show the map filter difference more clearly than the two for loops with the same header and just a differing body.
"But the range-based for loop is shorter and thus more readable." you say? In this very small example, this may be the case, but if the loop body would be much longer, the character count difference dissolves and you will be happy that you do not have to look at the body at all in the transform
/find_if
version to figure out what it is doing.
Also passing along a strategy in form of a std::function
will become easier, since you can just plug it in.
If you just can not stand the manual usage of begin(v)
and end(v)
you are free to write a wrapper in case you have to use std::transform
often enough:
template <typename T>
vector<T> transformVec(const vector<T>& v, const function<T(T)>& op)
{
vector<T> result;
result.reserve(v.size());
transform(begin(v), end(v), back_inserter(result), [&op](int i)
{
return op(i);
});
return result;
}
vector<int> squareVec7(const vector<int>& v)
{
return transformVec<int>(v, [](int i)
{
return i*i;
});
}
"But I have to use the hand written for loop for better performance!" - Nope, you do not have to.
Even if the std::transform
version looks like much abstraction induced function call overhead, especially with the lambda function, there is none. It is all optimized away by the compiler.
For 100 million values the different implementations (source code) took the following cpu times on my machine:
goto - elapsed time: 0.906895s
while - elapsed time: 0.915255s
for - elapsed time: 0.910748s
range based for - elapsed time: 0.933982s
std::transform - elapsed time: 0.903244s
Sure, readability also has something to with taste or to be precise familiarity, for in my opinion you should avoid explicit loops and make use of the cool stuff in the <algorithm>
header for better maintainability of your C++ software. Once you get used to it you will enjoy every for loop you do not have to read. ;-)
With effective stl Scott Meyers has written a very nice book covering this and more in depths. Herb Sutter's talk about lambdas can also help to get more into this topic. Also you can discuss this article on reddit.