Simplest way to for loop

kshegunov

@SimonSchroeder said in Simplest way to for loop:

Personally, with a all the new features of C++20, like ranges and generators, I would really like to see someone implement something like this:
for(int i : "[0..10)"_rng) // numbers 0 through 9
...
for(int i : "[1..10]"_rng) // numbers 1 through 10
...

Ah, the C++ code is JavaScript argument. Well, I really wish someone wouldn't. auto is terrible enough on its own, I personally don't need even more vague headaches.

Chris Kawa

@SimonSchroeder said:

Personally, with a all the new features of C++20, like ranges and generators, I would really like to see someone implement something like this:

I was bored :P

auto operator"" _rng(const char* str, size_t len)
{
    int start = atoi(str+1);
    int end = atoi(strrchr(str, '.') + 1);

    if (*str == '(') ++start;
    if (*(str+len-1) == ']') ++end;

    return ranges::views::iota(start, end);
}

int main()
{
    for (auto i : "[1..10)"_rng)
    {
        std::cout << i << '\n';
    }
}

Please don't ever use this anywhere. No, seriously. It's a horrible waste.

Q139

#define el(t) errorLine(t)
#define il(t) infoLine(t)
#define vo void
#define re return
#define ei   else if
#define el   else
#define w(c)             while(c)
//for loops zero index based
#define f0(v,max)        for(int v=0; v<max; v++)
#define f(v,min,max)     for(int v=min; v<max; v++)
#define fRev(v,min,max)  for(int v=max-1; v>=min; v--)
#define fa(v,c)          for(auto v:c)
//data types
#define D double
#define F float
//int types
#define I int
#define UI unsigned int
#define L long int
#define UL unsigned long int
#define I64 qint64
#define I32 qint32
#define I16 qint16
#define U64 quint64
#define U32 quint32
#define U16 quint16

#define C char
#define UC unsigned char
#define B bool

I tryed coding with those , in some places i think it can actually increase readability ,if familiar with the macros.
Also single letter capital variable types make code look cleaner.
If i continue adding stuff some point 1-2 letter combos will run out and will have to start using more keystrokes.

JonB

@Q139
I don't mean to be rude, and of course your code is up to you, but I think you will be the only person to find these macros "clearer". I would simply class them as "dangerous" --- goodness knows what you might "break" by defining single- or even two-character sequences as macros. And you are overloading, say, el versus el(t) to do completely different things. On top of everything, have you looked whether you will break, say, Qt Creator's auto-completion or folding etc. with these definitions?

This is just not the C++ way to do things. Up to you.

kshegunov

@Q139 said in Simplest way to for loop:

I tryed coding with those , in some places i think it can actually increase readability ,if familiar with the macros.

@JonB said in Simplest way to for loop:

This is just not the C++ way to do things.

This is not even the C way to do things, that's simply trying to reinvent the language by means of the preprocessor, what could possibly go wrong ...

Chris Kawa

Reminds me of a guy that transitioned form Pascal to C++ and first thing he did was define bunch of macros like

#define begin {
#define end }
#define procedure void

If you decided to code in C++ code in C++. I completely agree with @kshegunov . Using preprocessor to reinvent the language is just straight up horrible.

artwaw

@Chris-Kawa I used to code a lot in Borland Pascal + TASM back in '90 but I would not even consider such... THING as the above. That's truly WAT moment for me.

Q139

@JonB said in Simplest way to for loop:

On top of everything, have you looked whether you will break, say, Qt Creator's auto-completion or folding etc. with these definitions?
Project 35k lines ,no problems so far.

If I continue adding macros at some point I would need a converter to detect the macros and convert to clean code.

It is sad that precompiler fails if adding other preprocessor macros like #define ompFor #pragma omp parrallel for schedule(dynamic)

mrjj

@Q139
Hi
Creator has a very good refactor function.
https://doc.qt.io/qtcreator/creator-editor-refactoring.html

Also, if you tried this due to being tired of typing something always, please notice
that Creator can help with its auto text.
Pressing ctrl+space allows you to insert for loop , all ready to name it
and so on.

alt text

Trust me when i say you dont really want to use macros to change the core language. Been there, done that and it was not a good idea. One year later when i looked at the project, then suddenly it was not so clear anymore and that is actually how most people see it right away. :)

Q139

@mrjj Is it possible to set refactor to start at 2 letters instead of 3?
I know it is bad idea due to exponentially increasing results etc..
Never saw for refactor , never pressed the ctrl+space on purpose.

mrjj

@Q139
Hi
Well the auto code text thing can trigger on any keyword and with length 1 i think
alt text

For refactor, its a super search and replace basically but it does know the real type so its smarter then
plain text replace. (to show only the right ones)

fcarney

For most programmers I have met, macro programming is considered something you should avoid unless you have a specific need. Learn the language and learn the libraries. You will find that you don't need macros. You will eventually run into a coder reviewer and they will not be happy with excessive macro usage. Macros have logic bombs that appear to be evaluated differently than you might expect. I would suggest researching why people avoid macros in the first place. You would do better to learn about templating to customize things. For instance you could use templating to make your own range class that allows iteration using for(auto i: items(0,10,1)). Where items is a templated class.

Q139

I was reading on c++ performance optimization and saw usage of for(int i=10; i--; ) supposedly is is faster for cpu to compare ==0 and also easyer to write.

sierdzio

@Q139 said in Simplest way to for loop:

I was reading on c++ performance optimization and saw usage of for(int i=10; i--; ) supposedly is is faster for cpu to compare ==0 and also easyer to write.

I'd say this: what matters in coding is not whether something is easier to write but what is easier to read. You will read the code far more often than you will write it. Even when writing code, we all, constantly, refer to (==read) a lot of already existing code.

And when it comes to "faster" - well, ranged for is slower than traditional one. What you wrote is even faster. But "faster" is very deceptive. Are you running a small loop once in a while? You won't notice a difference between 100 nanoseconds and 60 nanoseconds for the faster solution. If you have a big, tight loop, running in some performance-critical section - then sure, optimizing it beyond readability makes sense. But even then, usually much bigger speed gains can be achieved by other means.

JonB

@Q139
Apart from what @sierdzio has (correctly) observed.

It probably is faster, though really marginal. Unless you are doing "nothing" inside the loop, I would bet whatever you do do there (far?) outweighs the loop counter test itself.

However, I would hazard that a fair number --- possibly the majority --- of for loop cases do not do what is needed/you want if you count downward instead of upward, so it's a dubious generic optimization.

Meanwhile you should read through, say https://stackify.com/premature-optimization-evil/ or https://medium.com/@okaleniuk/premature-optimization-is-the-root-of-all-evil-is-the-root-of-evil-a8ab8056c6b.

Q139

@sierdzio I have not counted but it is many for loops running machine learning task from minutes-days depending on parameters , outer for loop caching lz4 compressed ram to disk, other inner loops iterate and access different data, outside a while loop that tweaks parameters.

Main data that gets acessed is is a long single dimensional vector that gets used as multidimensional, each dimension can have different length.

But access locations are scattered quite randomly for most parts of code as it uses 5-10+ dimensions.
One thing that improved performance was moving for loops that access memory in similar regions to be most inner but it probably only brings 1 dimension close.

@JonB For loop may be faster backwards but idk if reading data backwards would get better optimization in data acess also.

sierdzio

OK, that does look like a special case where it might matter, indeed.

JonB

@Q139
And OOI you've measured the performance difference between upward & downward loops with your code, compiled for release with whatever optimizations, and you're seeing a noticeable change?

Q139

@JonB Ddownward loops have not tested and inner loop where it could benefit most has do-while that generates permutations.

From compiler optimizations -O3 -march=nativeatm , -ffast-math didn't change anything.

Also im not sure O2 or 03 do much.
I read that compilers have a feature that allows collecting runtime stats and optimizing to fewer cache misses on next compile. Will test that at some point.

Best gains were from shifting for loops around, and comparing releases.
Also separating less accessed variables from structs in vector that gets many acesses to separate vectors helped performance 25+%.

Also in structs setting 4 bools in row let's compiler optimize to single byte, if put other variables between ,each bool will take 8 bits ,not sure why compiler won't optimize over 4bools into single byte if could fit 8, maybe bottleneck for cpu to access.

I now put variables in structs same types together if it performance intensive , it makes readability worse if same purpose variables scattered but compiler does not optimize or shift those around and vectors with structs grow in size.
Probably because it could break code if acessing struct components via memory address but during compile it could be checked and still optimized to reduce memory.
Idk how icc compiler performs, tryed it but didn't get running with Qt windows build.

If can get openAcc working with Qt project could try on various for loops to auto parralelize into gpu.

Chris Kawa

@Q139 said:

I was reading on c++ performance optimization and saw usage of for(int i=10; i--; ) supposedly is is faster for cpu to compare ==0 and also easyer to write.

Those are "tricks" from the 90s. Compilers can do that (and a lot more) themselves now: https://godbolt.org/z/6_3_qx
Heck, they will unroll the loop and not do any counting in some cases.

each bool will take 8 bits ,not sure why compiler won't optimize over 4bools into single byte if could fit 8, maybe bottleneck for cpu to access.

Each platform and data type has what is called a natural alignment - it's a size that is a best fit for given cpu architecture so that no internal shifts and masking is needed to access a variable. When you declare a struct each member is aligned to that size and some padding between variables might occur. For example:

struct S {
   bool b;
   int a;
};

could occupy 8 bytes. 4 for the int and 4 for the bool.

struct S {
    bool b1;
    bool b2;
    bool b3;
    bool b4;
    int a;
};

This will still occupy 8 bytes, but if you add 5th bool it will jump to 12 because of this padding. If you don't need every bit in a type (a bool only needs 1bit really) you can pack them like this:

struct S {
    int b1 : 1;
    int b2 : 1;
    int b3 : 1;
    int b4 : 1;
    int a : 28;
};

and now it's just 4 bytes.
Compilers also have custom means to force different packing. For example MSVC has #pragma pack(x):

#pragma pack(1)
struct S {
    bool b;
    int a;
};

This will occupy 5 bytes instead of the original 8.

Compilers have tools to show you the resulting layout of your structs and classes. For example MSVC has /d1reportSingleClassLayoutXXX switch, where XXX is the name of your struct. It will output a detailed information about sizes, alignment and packing of the struct.

As for loops - it's important to know your hardware and access data in a hardware friendly manner. In case of long loops it means knowing a size of your cache line, sizing your struct so you don't waste any space in them. For example if your cache line is 64 bytes in size and your struct is 65 you're gonna need two cache lines and potentially waste 63 bytes in the other one. When you have your structs aligned with cache lines next step is access them in a way CPU can optimize. This means access things as close to each other and in as predictable pattern as possible - linear in one direction is optimal, but there are a lot others - again, know your hardware.