Блог пользователя dendi239

Micro- and macro-management for your next cpp source: General tips'n'tricks
Автор dendi239, история, 4 года назад, По-английски

Hello, Codeforces!

In this series of articles, I would like to tell you about tips'n'tricks I use in C++ sources. Some of them require a strong C++ background, some not, but all of them are aimed to reduce the time you spend on basic things like reading input/debugging/etc, so it might be useful for your next template.

This part is dedicated to some basic tips (and probably should've been the first part but it's not ¯\_(ツ)_/¯).

TL;DR

bits/stdc++.h

Just add this header to your source file and you won't have to add any other STL headers. Probably the only reason why you still haven't done it (if you haven't) is that it might not work on your machine: that happens because bits/stdc++.h is a feature of GCC, not the entire cpp standard. However, there're a few ways to work around it.

How to fix bits/stdc++.h if it's not working

Content

Well, before we start putting stdc++.h to the proper location, there's one thing to determine: content to put in it. You either can determine what default headers you might need or just google "GCC bits/stdc++.h source" and copy/paste it [if you've decided to do so, you might need to remove some gcc-specific headers: just try to compile it and remove one by one non-compiling includes].

Option 1: Add it to default location

Well, the easiest one is to find one of the include paths, like /usr/local/include (might add to /usr/include, but it's a system directory, do it at your own risk) on Mac, or somewhere else for Windows. All you need to do after this is to create folder bits and file stdc++.h inside it.

Option 2: Add to user-defined header search paths

If you compile/run you code manually, there's the option of adding proper location as compiler flag isystem directly:

$ g++ -std=c++2077 a.cpp -o run_a -isystem /path/to/your/custom/includes

Note that inside the folder /path/to/your/custom/includes must be the folder bits with file stdc++.h.

Option 3: Add to you build system

If you're using CMake based build system (which CLion is), then you could edit your CMakeLists.txt directly: 

include_subdirectory(/path/to/your/custom/includes)

Typealiases

In C++ there are a lot of strange things that came from ancient times, like signed long long int constructions. However, if you don't like such long type names, you could write using NewType = ExistingType or typedef ExistingType NewType.

A lot of programmers here, on CodeForces, use ll instead of long long and a bit less use ld instead of long double. Since C++11, we have int64_t, which is like long long but not that long. I could also suggest using the following (inspired by Rust):

using i64 = int64_t;
using f80 = long double;

using has an advantage over typedef: it supports templates, so you could shorten annoyingly long type names such as vector or unordered_set(map):

using Str = string;
template <class T> using Vec = vector<T>;
template <class K, class V> using UM = unordered_map<K, V>;

By the way, if you've noticed that sometimes map works faster than unordered_map even on big inputs, it happens because of expensive and frequent rehashing hidden in unordered_set(map). To work around it, I suggest calling um.reserve(4 * MAX_ELEMENTS_EXPECTED). It will work much faster. You can also read this article to learn more about unordered_map.

Prewritten library code

According to CodeForces' official rules, you might use some prewritten code. When it comes to some library-like code for things like fft, flows, or smth else you can find it annoying to search over your sources to copy/paste once you need it. One of the ways is to put all functionality you might need to separate files like z.hpp and simply include it [keep in mind the section about modifying header search paths].

If you don't use the include, just remove it before submission, and if you need it, you can simply jump to definition (most of IDEs can do it using ctrl + click to source and copy/paste content before submitting. For example with this file you could write smth like:

#include <bits/stdc++.h>
#include <z.hpp>

using namespace std;

constexpr int mod = 1000 * 1000 * 1000 + 7;
using Mint = Z<mod>;

template <class T>
T bin_pow(T base, int64_t exponent) {
  if (exponent == 0) return T(1);
  if (exponent % 2 == 1)
    return base * bin_pow(base * base, exponent / 2);
  return bin_pow(base * base, exponent / 2);
}

int main() {
  Mint a, b;
  cin >> a >> b;

  // note that the following addition is performed by modulo
  cout << a + b << '\n';

  // note that because of the type of `a` bin_pow uses modular arithmetics
  cout << bin_pow(a, b) << '\n';
}

Don't forget to copy/paste the actual source (if you're using it) to your file before submitting: there's no z.hpp accessible by codeforces's compiler. It may be enhanced by custom script doing it instead of you, you can find one such here.

stdin, stdout, stderr

Your program is using at least three streams to communicate with OS: stdin for input, stdout for output, and stderr for errors. Moreover, by default, there's sync for cin and cout which may slow down your solution on huge inputs. You could discover more here. If you don't like ios::sync_with_stdio(false); cin.tie(nullptr); appearing in main every time, you could use the following trick: if you need some side effects to happen even before main starts, you can do it when global variables instantiate:

int main() {
  // some code
}

// empty namespace to not conflict with any other possible names
namespace {

// global variable is assigned to the result of calling the newly created lambda function.
// since your variable cannot be void, put `int` as a result:
auto fast_io = [] {
  ios::sync_with_stdio(false);
  cin.tie(nullptr);

  return 0;
}();

}  // namespace

Since any checker/grader reads only stdout, I strongly recommend using stderr for debug information (cerr instead of cout) — that way once you find a bug and solve it, you won't need to remove all the debugging stuff before submitting. However, if you output to stderr a lot, it can affect the performance.

Interactive, custom files and other

Sometimes you don't need cin.tie for interactive problems or debugging in your IDE. Moreover, sometimes you need to work with custom files. To resolve it, I can suggest using defines INTERACTIVE and FILES:

// example of usage, don't use both in real life
#define INTERACTIVE
#define FILES "k3"

auto fast_io = [] {
#ifndef DEBUG
#  ifndef INTERACTIVE
    ios::sync_with_stdio(false);
    cin.tie(nullptr);
#  endif // INTERACTIVE
#  ifdef FILES
    freopen(FILES".in", "r", stdin);
    freopen(FILES".out", "w", stdout);
#  endif // FILES
#endif // DEBUG
  return 0;
};

Use some tooling

Since you're writing some code and test it against a few test cases it may be annoying to write all of them on your own every time you run the program. To get rid of copypasting them from a webpage, or trying to copy with spaces from pdf, you could've put it into files at once and get input from these files. There's an amazing tool from xalanq that does it for you. It ignores all your debug output while checking, but still prints it above the testing verdicts. Take a look at this submission:

➜  b cf test  
g++ -std=c++17 b.cpp -o b.exe -O2 -DLOCAL -DDEBUG
[ n ] = 4
[ n ] = 3
[ n ] = 1
[ n ] = 7
Passed #1 ... 0.001s 0B
[ n ] = 4
[ n ] = 3
[ n ] = 1
[ n ] = 7
Passed #2 ... 0.001s 0B
rm b.exe
➜  b cf submit
Submit CONTEST 1367, problem b
Current user: dendi239
Submitted
      #: 85413433
   when: 2020-06-29 03:31
   prob: B - Even Array
   lang: GNU C++17
 status: Accepted                   
   time: 61 ms
 memory: 3.71 MB

Run debug code only in debug mode

You may need to run some extra code for debugging purposes only. For example, printing a graph, performing additional checks, or comparing with the naive solution. It may affect the resulting performance and move you from the AC zone to the TL one.

There's the straightforward solution: use #ifdef preprocessor directive: 

int main() {
  int a, b;
  cin >> a >> b;
#ifdef DEBUG
  cerr << "a = " << a << " b = " << b << '\n'
#endif // DEBUG
  cout << a + b << '\n';

To run this code in debug mode, you need to pass -DDEBUG or -D DEBUG option to your compiler. If you don't know how to do it, or simply don't wanna to do it, codeforces defines ONLINE_JUDGE for you:

int main() {
  int a, b;
  cin >> a >> b;
#ifndef ONLINE_JUDGE
  cerr << "a = " << a << " b = " << b << '\n'
#endif // ONLINE_JUDGE
  cout << a + b << '\n';

Plenty of code just to output two variables, isn't it? Let's solve two problems:

  • print all variables given to function (macro)
  • write code (block, or just function) that runs in debug mode and doesn't evaluate arguments in release mode

Printing all variables

Ok, let's provide macro LOG(x, y, z) that works like cerr << "x = " << x << " y = " << y << " z = " << z << "\n", but works with plenty of arguments. Since we can't take x as string from function argument, let's start from simple scratch:

#define LOG(...)  cerr << "[ "#__VA_ARGS__" ] = "; log(__VA_ARGS__) << '\n'

Where log function prints all arguments given to it. You could implement it the way I mentioned in my previous article: 

template <class ...Args>
auto &log(const Args &...args) {
  return ((cerr << " " << as), ...) << '\n';
}

Let's put it all together:

template <class ...Args>
auto &log(const Args &...args) {
  return ((cerr << " " << as), ...) << '\n';
}

#define LOG(...)  cerr << "[ "#__VA_ARGS__" ] = "; log(__VA_ARGS__) << '\n'

int main() {
  int a, b;
  cin >> a >> b;

  LOG(a, b, a + b);
  // prints: [ a, b, a + b ] = 2 2 4

  cout << a + b << '\n';
}

Well, there are plenty of things you could've improved, as mentioned in this discussion. However, it still doesn't work fine for things with a comma in subexpressions like:

LOG(a, b, some_func(a, b))

Running code in debug mode only

How to run code in debug and don't evaluate arguments in release mode? Well, here's a trick you might have seen here involving if:

#ifndef DEBUG

#define CERR if (false) cerr

#endif

// somewhere in code

CERR << "a = " << a << '\n';

// expands to

if (false) cerr <<  << "a = " << a << '\n';

But there's bug here:

if (a == 0)
  CERR << "b = " << b << '\n';
else {
  CERR << "c = " << c << '\n';
  // dance like no one's watching
  // with a != 0
}

// expands to

if (a == 0)
  if (false)
    cerr << "b = " << b << '\n';
  else {
    if (false)
      cerr << "c = " << c << '\n';
    // a is still zero here :(
  }

Well, there's an alternative with while (false) which doesn't have such bugs. But what to do instead of while (false) to run any code? We may try to put nothing [don't know if there's bug], but I'd prefer a one time running for. Let's put it all together:

template <class ...Args>
auto &log(const Args &...args) 
{ return ((cerr << " " << as), ...) << '\n'; }

#ifdef DEBUG
#  define LOG(...)  cerr << "[ "#__VA_ARGS__" ] = "; log(__VA_ARGS__) << '\n'
#  define RUN       for (bool _flag = true; _flag; _flag = !_flag)
#else  // DEBUG
#  define LOG(...)  while (false) cerr
#  define RUN       while (false)
#endif // DEBUG

int main() {
  int n, m;
  cin >> n >> m;

  LOG(n, m);
  
  vector<vector<int>> g(n);
  for (int e = 0; e < m; ++e) {
    int u, v;
    cin >> u >> v;

    g[--u].push_back(--v);
    g[v].push_back(u);
  }
   
  RUN {
    for (int u = 0; u < n; ++i) {
      cerr << u << ":";
      for (int v : g[u]) 
        cerr << " " << v;
      cerr << "\n";
    }
  }
}

auto

There's the keyword to replace all the boring types you don't need to write explicitly:

for (auto it = xs.begin(); it != xs.end(); ++it)
for (auto foo : foos)

But there's a hidden problem: unnecessary copying: 

for (auto row : table) {
  // here row has a type vector<Cell> 
  // and it will be copied even if you don't modify it
}

To avoid it, you could've used:

for (auto &row : table)

But, it might be a temporary object, that's kind of annoying. To resolve it, you could've used universal reference everywhere:

// works just fine: no copies, 
//   reference if regular object
//   value if temporary one
for (auto &&row : table)

Ok, where to use it?

Predicates

If you're using lambdas, you might still not know about generic lambdas, let me show the use case:

vector<pair<int, int>> some_events = ...;
sort(some_events.begin(), some_events.end(), [&](auto &&lhs, auto &&rhs) {
  // compare two pair<int, int> here
});

You might have noticed that you often write code like return foo(lhs) < foo(rhs);. Here's the way to get rid of it: 

#define BY_FUNC(func) [&](auto &&lhs, auto &&rhs) { \
  return func(lhs) < func(rhs);                     \
}

sort(some_events.begin(), some_events.end(), BY_FUNC(some_func));

Much better (you can also come up with a shorter name for macro), but if you want to only compare by .second, it's not suitable (you'll still have to define some_func which takes a few extra lines) :(. However, we could generate all necessary code on our own (thanks to viskonsin for pitching me this idea):

#define ALL(xs) begin(xs), end(xs)

#define BY(...) [&](auto &&lhs, auto &&rhs) { \
  auto predicate = [&](auto &&x) {            \
    return __VA_ARGS__;                       \
  };                                          \
  return predicate(lhs) < predicate(rhs);     \
}

sort(ALL(pairs), BY(x.second));

// expands to:

sort(pairs.begin(), pairs.end(), [&](auto &&lhs, auto &&rhs) {
  auto predicate = [&](auto &&x) {
    return x.second;
  };
  return predicate(lhs) < predicate(rhs);
});

// such plenty of code out of one short macro!

dbg macro

Thanks to ramchandra for his amazing dbg macro, really nice job. There's also the dbg! macro in Rust. Let's develop the same behavior: dbg(expr) works as a simple expr in release mode and additionally prints it in debug mode. Well, the implementation for release mode is pretty straightforward:

#define dbg(...) (__VA_ARGS__)

And what about the debug one? I'd like to work with any kind of references here, to write smth like: 

auto it = dbg(order).find(dbg(x + 1));

Okay, it must be a result of expression with necessary debug output. Let's use a simple lambda on the fly (note that forward is used since we need to keep lvalue if it was passed and fallback to moving value if not: for dbg(i + j) we don't need to return a reference to local object):

int recursion_depth = 0;
#define dbg(...) [&]() -> auto && {       \
  ++recursion_depth;                      \
  auto&& value = __VA_ARGS__;             \
  --recursion_depth;                      \
  cerr << string(recursion_depth, '\t')   \
       << #__VA_ARGS__" = " << value      \
       << endl;                           \
  return forward<decltype(value)>(value); \
}()

Conclusion

You might have noticed that there's a lot of complicated things involved. It took me way more than 21 days to come up with all of these tips'n'tricks. Some of them contained bugs in the first version, others not. However, anytime you use macros it's unclear what they expand to and pretty easy to make a mistake. Don't use any single macro you're not feeling ready to debug a bit if necessary. If you don't feel comfortable in understanding some of these macros, prefer not to use them.

Here's a simple way to make a mistake when dealing with READ macro:

// does not compile
for (int READ(t); t--;)

// expands to
for (int t; io::read(t); t--;)

// why there're three ";" inside for?

It took over half an hour to come up with dbg macro that worked as I wanted it to. I tried a lot of things, even stack overflow didn't give me an answer, but months of exploring STL sources gave me a simple hint (actually it was a question: "what I need to use std::forward for?").

There's plenty of code, isn't it? Note that it may be overcomplicating to use every single of macros: just choose which ones you like and add them to your template. However, you could add all of them:

#include <bits/stdc++.h>

using namespace std;

using i64 = int64_t;
using f80 = long double;

using Str = string;
template <class T> using Vec = vector<T>;
template <class K, class V> using UM = unordered_map<K, V>;

template <class T> auto &operator>>(istream &is, vector<T> &xs)
{ for (auto &x : xs) is >> x; return is; }

namespace io {
template <class ...As> auto &read(As &...as) { return (cin >> ... >> as); }
template <class ...As> auto &log(const As &...as) { return ((cerr << " " << as), ...); }
}  // namespace io

#define ALL(xs) begin(xs), end(xs)
#define BY(...) [&](auto &&lhs, auto &&rhs) { \
  auto predicate = [&](auto &&x) {            \
    return __VA_ARGS__;                       \
  };                                          \
  return predicate(lhs) < predicate(rhs);     \
}  
#define READ(...)             __VA_ARGS__; io::read(__VA_ARGS__)
#define READ_CTOR(name, ...) name(__VA_ARGS__); io::read(name)

#ifndef ONLINE_JUDGE
  int recursion_depth = 0;
#  define RUN       for (bool _flag = true; _flag; _flag = !_flag)
#  define LOG(...)  cerr << string(recursion_depth, '\t') \
                         << "[ "#__VA_ARGS__" ] =";       \
                    io::log(__VA_ARGS__) << '\n'
#  define dbg(...) [&]() -> auto && {         \
      ++recursion_depth;                      \
      auto&& value = __VA_ARGS__;             \
      --recursion_depth;                      \
      cerr << string(recursion_depth, '\t')   \
           << #__VA_ARGS__" = " << value      \
           << endl;                           \
      return forward<decltype(value)>(value); \
  }()
#else
#  define LOG(...) while (false) cerr
#  define RUN      while (false)
#  define dbg(...) (__VA_ARGS__)
#endif

auto main() -> int {
  // write code here
}

namespace {
auto fast_io = [] {
#ifndef DEBUG
#  ifndef INTERACTIVE
    ios::sync_with_stdio(false);
    cin.tie(nullptr);
#  endif // INTERACTIVE
#  ifdef FILES
    freopen(FILES".in", "r", stdin);
    freopen(FILES".out", "w", stdout);
#  endif // FILES
#endif // DEBUG
  return 0;
};
} // namespace

What's next?

There're amazing articles from Golovanov399 here, and from HosseinYousefi here — check them, they cover a lot of things (like all(xs) macro) I omitted. There's also another my article about reading and declaring variables all at once. And more will come, stay tuned!

I would like to thank:

See ya!

  • Проголосовать: нравится
  • +137
  • Проголосовать: не нравится

»
4 года назад, # |
  Проголосовать: нравится +5 Проголосовать: не нравится

Auto comment: topic has been updated by dendi239 (previous revision, new revision, compare).

→ Ответить
»
4 года назад, # |
← Rev. 2   Проголосовать: нравится 0 Проголосовать: не нравится

Don't forget to copy/paste the actual source (if you're using it) to your file before submitting: there's no z.hpp accessible by codeforces's compiler. It may be enhanced by custom script doing it instead of you, but there's no such script yet ¯\_(ツ)_/¯.

Did you check this post? (I am not sure, but seems like it can help).

→ Ответить
»
4 года назад, # |
  Проголосовать: нравится 0 Проголосовать: не нравится

Auto comment: topic has been updated by dendi239 (previous revision, new revision, compare).

→ Ответить
»
4 года назад, # |
  Проголосовать: нравится 0 Проголосовать: не нравится

Auto comment: topic has been updated by dendi239 (previous revision, new revision, compare).

→ Ответить
»
4 года назад, # |
  Проголосовать: нравится 0 Проголосовать: не нравится

Nice blog! Although, a very important point that I couldn't find mentioned, is that even a very innocent looking cerr can cause your code to TLE. Here is AC and TLE submission.

The only difference? In TLE code, I'm printing two integers to cerr.

→ Ответить
  • »
    »
    4 года назад, # ^ |
      Проголосовать: нравится 0 Проголосовать: не нравится

    However, if you output to stderr a lot, it can affect the performance

    Yeah, I mentioned it but didn't elaborate importance of it. However, you still could use smth like LOG macro to get rid of these problems.

    That happens mostly because of endl, not cerr itself: flushing stderr is expensive enough.

    → Ответить