From C++11, a std::string is always zero-terminated. Thus, of all the characters, exactly the last one is falsy. So, these loops are equivalent:

#include<string>
std::string s;

// indexed loop, yielding an index
// ungolfed
for(int i = 0; i < s.size(); i++)
{
    std::cout << s[i];
}

// golfed C++98
for(int i=0;i++<s.size();)std::cout<<s[i];

// golfed C++11
for(int i=0;s[i++];)std::cout<<s[i];

Note, that depending on the surrounding code, the following alternative looping techniques may be more efficient:

// C++11 iterator loop, yielding an iterator
for(auto i=s.begin();i++<s.end();)std::cout<<*i;

// C++11 ranged-for loop, yielding a character
for(auto c:s)std::cout<<c;

The latter work likewise for std::vector.

If you have some variables declared in a row :

int a,b,c,d;

You can access b by (&a)[1], or c by (&a)[2] ...

It can be interesting there :

(i?a:i-1?b:i-2?c:d)=k;
VS
(&a)[i]=k;

You could have declared an array a[4], but if you use a, b, c, d in other ways, it's not always interesting to replace b by a[1], c by a[2]

Moreover, you know that x[y] = *(x+y) = y[x], so :

(i?a:i-1?b:i-2?c:d)=k;
VS
(&a)[i]=k;
VS
i[&a]=k;

If you need a type alias, using is two bytes shorter than #define:

#define T ... // newline
using T=...;

When possible, change && and || to & and | respectively.

When using simple if statements:

if(<condition>)<stuff>;

can be changed to:

<condition>?<stuff>:<any single character variable or literal>;

which saves a character.

It's rather straightforward to cacluate when replacing something with a #define results in shorter code.

#define a blabla

Where a is a single letter name and blabla is the replacement of length \$x\$ has in total \$10+x\$ characters.

If you have \$n\$ times blabla in the code then the #define pays off when

$$10 + x + n < nx$$

With the define you have \$10 + x\$ to define it and \$n\$-times a single character. Without it you have \$n\$-times blabla of length \$x\$. The following table lists how often blabla has to appear given its length for the define to make the code shorter:

Because the inequality is symmetric, the same table applies when \$n\$ and \$x\$ are swapped, ie if blabla appears \$n = 2\$ times in the code then the define results in shorter code when blabla is of length \$x \ge 13\$. The formula is

$$n > \frac {10 + x} {x - 1}$$

or

$$x > \frac {10 + n} {n - 1}$$

PS: Sometines the single letter named preprocessor symbol has to be followed by a space. This shifts the matters a little towards not using the define, but just a little.

Some compilers (e.g. GCC) support multi-character constants. This can save a few characters when a large integer value is required. Example:

int n='  ';

The value is implementation-specific. Usually the value of 'ab' is 256*'a'+'b' or 'a'+256*'b'. You can specify up to 4 characters between the quotation marks.

Functions in <algorithm> often requires passing a.begin(),a.end() which is really long, instead you can use &a[0],&*end(a) to save a byte or two if a is vector or string.

std::sort(a.begin(),a.end());
std::sort(&a[0],&*a.end());

// with using namespace std;
sort(begin(a),end(a));
sort(&a[0],&*end(a));
```

Use auto instead of complex type names.

This can be done with generic lambdas and C++20 abbreviated function templates

int x(std::set<std::string>y){}
template<class T>int x(T&y){}
int x(auto&y){}
[](auto&x){}

This also allows you to avoid most includes.

Note that g++ on TIO requires -fconcepts.

G++ allows for variable-sized arrays, so there is no need to go through the lengthy pointer-malloc-new stuff.

int* array = new int[array_size];

becomes

int array[array_size];

Saves 9 bytes minimum with int.

When writing a full program you can use argc to initialize an integer variable to 1:

main(a){

is the same as:

main(){int a=1;

Use GCC builtins instead of importing

If you are using a GCC compiler, it sometimes helps to use their builtin functions, such as __builtin_puts or __builtin_clz. For example,

44 bytes:

int main(){__builtin_puts("Hello, world!");}`

50 bytes:

#import<cstdio>
int main(){puts("Hello, world!");}

If you're doing C++11 or newer (which should always be the case now), use auto for complex types, if possible.

Example: 54 Bytes instead of 66

#include<vector>
std::vector<int> f(std::vector<int> l){return l;}

#include<vector>
auto f(std::vector<int> l){return l;}

Also, as performance does not matter, for some challenges a std::list may just do the job for a few bytes less:

#include<list>
auto f(std::list<int> l){return l;}

If you want to swap two integer variables a and b then ,

a^=b^=a^=b;

can be used , saving 5 characters than the standard way

a+=b;
b=a-b;
a-=b;

Use generic lambdas as cheap templates

For types other than int, using them as function arguments can be expensive. However, generic lambdas were introduced (in C++14?) and allow any lambda to be a template - using auto for the argument types can save bytes. Compare:

double f(double x, double y)
[](auto x, auto y)

Generic lambdas are also very convenient for accepting iterators - probably the best way to accept array inputs in C++ is [](auto a, auto z), where a and z are passed as begin() and end() of the array/vector/list/etc.

Shorter header

This is GCC specific, it may be extensible to other compilers.

Precompiled header.

In G++ bits/stdc++.h is the precompiled header consists of all other headers. If you need to import 2 different ones you can just use this.

Shorter header.

This is all headers listed on http://en.cppreference.com/w/cpp/header:

$("#a").text(pako.inflate(atob('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'), {to: 'string'}));

<script src="https://cdn.rawgit.com/nodeca/pako/master/dist/pako.min.js"></script>
<script src="https://ajax.googleapis.com/ajax/libs/jquery/2.1.1/jquery.min.js"></script>
<pre id="a">
  
</pre>

sorted in increasing order of length.

Some of them are already longer than bits/stdc++.h, and some of them requires C++17 support. Some others are not supported by TIO G++ (for reasons I don't know of). Filter out them we have:

$("#a").text(pako.inflate(atob('eJx9VcuS3CAMvOc39p5Tak+p/MoUI2SbhIcDYrz++4BdW6WG1B7VYI36IebNRfLV8s/Ix69vb59VYVFVMLuqXCqq8q7oqyQusKql7l7XJmdzqtry36rPixj6o+p2CucUjGyqzkZc0ucLx5c+55U/NJUqUD+dIFfacoq6Y+CQ8gk/4ZMRkCC0LvoG5ww9iOTcgcZBm8kaqANQJLu4KuDFJCkD9WhTQGOWw+qmRSyo4xMZtENScKT92jIb3WOpUrP+pAv8XVvQHQGgRwCAPod3T0DcGo0HhOV32IevTNYSVHHeyTncsbxoZHajqxDBY1MKZ0E/RocmAyiFlmUdnlgDZ5CvLUPTT5uSJmSZkDIhLgUTHagxeUfJMr3ka507K/DiiiYgV5wBuWIDyJVOQI4JIVmH5SRX0vuP9y+RPqCLi06pE25rDVl/GT++HG5W9rYVhrrTgNAlJBK+sofQFdBRlpbHEWrxm8SLk57NlgHq6RoUncyiOXO3yHDr1nTauGdKfhLaQjNqk3Zcrwt/EO/tUY1I4Ia12H5N2ckWkNUc7gt4VljSmxaO/D+sS+6bEzhLZ4YRrpH6yPCifHKbPOwN8cFq1x6SDb4b5SzC4dEWBqK4pHyYbB/DH19p68D2cf+//APa+54V'), {to: 'string'}));

<script src="https://cdn.rawgit.com/nodeca/pako/master/dist/pako.min.js"></script>
<script src="https://ajax.googleapis.com/ajax/libs/jquery/2.1.1/jquery.min.js"></script>
<pre id="a">
  
</pre>

It may happens that some of them can be replaced by shorter ones. Just binary search whether the one you need can be replaced. In particular:

cstdio -> ios        (-3 bytes)
algorithm -> regex   (-4 bytes)
vector -> queue      (-1 byte)
string -> map        (-3 bytes)
bitset -> regex      (-1 byte)
numeric -> random    (-1 byte)

Kind of late to the party I guess...

If you want to turn an expression into -1 and 1 instead of 0 and 1, instead of this:

int x;
if (a * 10 > 5)
    x = 1;
else
    x = -1;

do this:

int x = (a * 10 > 5) * 2 - 1;

It can save some bytes depending on usage.

The ternary conditional operator ?: can often be used as a stand in for simple if--else statements at considerable savings.

It is of special value in that it can be used to select alternate lvalues as in

#include <iostream>
#include <cstdlib>
int main(int c, char**v){
  int o=0,e=0,u;
  while(--c) ((u=atoi(v[c]))%2?o:e)+=u;
  std::cout << "Sum of odds " << o <<std::endl
            << "Sum of evens " << e <<std::endl;
}

Since array elements are stored directly after one another in memory, instead of something like this:

for(int x = 0; x < 25; x++) {
    for(int y = 0; y < 25; y++)
        array[x][y] = whatever;
}

You can do something like this:

int* pointer = array;
for(int i = 0; i < 25*25; i++, pointer++)
    *pointer = whatever;

Obviously neither of the above are golfed, for readability, but explicitly using pointers can save you a lot of space.

Using the comma operator in lieu of open and close braces can save a few characters, if you have a situation where your clauses have more than one statement in them:

if(c){x=1;cout<<"Hi";y=2;}else{x=2;cout<<"Bye";y=3;}

vs.

if(c)x=1,cout<<"Hi",y=2;else x=2,cout<<"Bye",y=3;###

Two characters saved on a plain IF, or three total for an IF/ELSE.

As a point of distinction between C and C++, the result of a comma expression in C++ as a whole may be used as an lvalue...FWIW.

You may use the ternary operator ?: without any expressions in the true-block (it saves a byte)

#include <iostream>

int foo()
{
    std::cout << "Foo\n";
}

int main()
{
    1?foo():0;  // if (true) foo()
    0?:foo();   // if (!false) foo()
}

Check it here

In my first attempt at code golf for task "Subtract the next numbers" I have started from function (58 bytes)

int f(int N, int P){int F;for(F=N;P;F-=++N,P--);return F;}

then safe 5 bytes with shifting to lambda and moving initialization out of for (53)

[](int N,int P){int F=N;for(;P;F-=++N,P--);return F;}

and finally after switching from for to while I got 51 bytes:

[](int N,int P){int F=N;while(P--)F-=++N;return F;}

The ungolfed test code is something like:

#include <iostream>
int main(void)
{
    int N, P;
    std::cin >> N >> P;
    auto f = [](int N,int P)
    {
        int F = N;
        while (P--)
            F -= ++N;
        return F;
    };
    std::cout << f(N, P) << std::endl;
    return 0;
}

UPDATE:

Actually for can reach the same length as while:

[](int N,int P){int F=N;for(;P--;F-=++N);return F;}

One that I found handy:

Taking advantage of the fact that non-zero values evaluate to true in boolean expressions, and that x&&y evaluates to x*y when dealing with booleans

(x!=0 && y!=0)

evaluates to

(x*y)

You just have to be aware of overflows, as pointed out below.

#import instead of #include gives you one more byte.

Also, the space character between #import and header is not necessarily:

#include <map>
// vs
#import<map>

And if you need something from stdlib header, you may import any header with STL container (preferable set or map) instead of cstdlib.

Don't use string(""), use "". It saves 8 bytes.

Instead of writing big powers of 10, use e notation. For example, a=1000000000 is longer than a=1e9. This can be extended to other numbers like a=1e9+24 is better than a=1000000024.

It is useful to remember is that a[i] is the same as *(a+i).

Replace a[0] with *a for two character savings. Also, a[i][0] is equivalent to *a[i] and a[0][i] shrinks down to i[*a]. So if you are hard-coding a 0 index in your array, a better way probably exists.

Instead of using while(1), use for(;;), saving one character :)

Arithmetic operations on Booleans:

Although

a*=b>0?.5:-.5

is better than

if(b>0)a*=.5;else a*=-.5;

it is not as good as

a*=(b>0)-.5

Also, using #define on anything that is used a lot. It is often shorter than using functions, since type names are not necessary.

Combine things as much as possible:

a+=a--;

is the same as

a=2*a-1;

Sometimes you can save two characters by using the fact that static storage duration variables (that especially includes all global scope variables) are automatically zero-initialized at the beginning (unlike automatic variables where you have no such guarantee). So instead of

int main()
{
  int a=0;
  // ...
}

you can write

int a;
int main()
{
  // ...
}

Quite an obvious one, but it you are using a lot of the standard library, using namespace std; might save a few characters.

If you're willing to use C++0x, you can use new features like lambdas.

Use the following types:

u64, s64, u32, s32 (or int)

For repetitive words/types, use #defines:

#define a while

It's only worth it if you use while a lot to make up for the extra 10 characters. (About 4.)