I am looking for a portable way to implement lazy evaluation in C++ for logging class.
Let's say that I have a simple logging function like
void syslog(int priority, const char *format, ...);
then in syslog() function we can do:
if (priority < current_priority)
return;
so we never actually call the formatting function (sprintf).
On the other hand, if we use logging stream like
log << LOG_NOTICE << "test " << 123;
all the formating is always executed, which may take a lot of time.
Is there any possibility to actually use all the goodies of ostream (like custom << operator for classes, type safety, elegant syntax...) in a way that the formating is executed AFTER the logging level is checked ?
This looks like something that could be handled with expression templates. Beware, however, that expression templates can be decidedly non-trivial to implement.
The general idea of how they work is that the operators just build up a temporary object, and you pass that temporary object to your logging object. The logging object would look at the logging level and decide whether to carry out the actions embodied in the temporary object, or just discard it.
What I've done in our apps is to return a boost::iostreams::null_stream in the case where the logging level filters that statement. That works reasonably well, but will still call all << operators.
If the log level is set at compile time, you could switch to an object with a null << operator.
Otherwise, it's expression templates as Jerry said.
The easiest and most straight-forward way is to simply move the check outside of the formatting:
MyLogger log; // Probably a global variable or similar.
if (log.notice())
log << "notified!\n" << some_function("which takes forever to compute"
" and which it is impossible to elide if the check is inside log's"
" op<< or similar");
if (log.warn()) {
log << "warned!\n";
T x;
longer_code_computing(value_for, x); // easily separate out this logic
log << x;
}
If you really wanted to shorten the common case, you could use a macro:
#define LOG_NOTICE(logger) if (logger.notice()) logger <<
LOG_NOTICE(log) << "foo\n";
// instead of:
if (log.notice()) log << "foo\n";
But the savings is marginal.
One possible MyLogger:
struct MyLogger {
int priority_threshold;
bool notice() const { return notice_priority < current_priority; }
bool warn() const { return warn_priority < current_priority; }
bool etc() const { return etc_priority < current_priority; }
template<class T>
MyLogger& operator<<(T const &x) {
do_something_with(x);
return *this;
}
};
The problem here is mixing iostream-style operator overloading with a printf-like logging function – specifically translating manipulators and formatting flags/fields from iostreams into a format string. You could write to a stringstream and then chunk that to your syslog function, or try something fancier. The above MyLogger works easiest if it also contains an ostream reference to which it can forward, but you'll need a few more op<< overloads for iomanips (e.g. endl) if you do that.
For mine I made a debug_ostream class which has templated << operators. These operators check the debug level before calling the real operator.
You will need to define non-template overrides for const char* and std::ostream& (*x)(std::ostream&) because otherwise those don't work. I'm not sure why.
With inlining and high enough optimization levels the compiler will turn the whole output line into a single check of the debug level instead of one per output item.
I should add to this that this doesn't solve the original problem. For example if part of the debug line is to call an expensive function to get a value for output, that function will still be called. My solution only skips the formatting overhead.
Related
When you stream variables to an output stream such as cout, type conversion is automatic. What I'm trying to figure out is how to do this via a function call, for example:
inline void DEBUG(ostream& s) // Don't know if this prototype is appropriate
{
cout << s;
}
main()
{
int i = 5;
DEBUG("The value is: " << i << endl); // This doesn't compile
DEBUG("The value is: " + i + endl); // Neither does this
}
I found similar questions on here, but they all involve passing the stream object as a parameter, whereas I'm trying to pass the "streamed data" to a function that already has the stream object, so it's the other way round. Is this even possible? I don't want to resort to explicit type conversions. I also found this question, but I really don't want to write a whole logger class if I can avoid it.
At the moment I'm implementing it as a macro, which works, but I'd rather use an inline function if possible.
#define DEBUG(s) (cout << s)
Of course it does not compile. There are many reasons for that.
First, Operator << is not defined for standard streams, and you are trying to do exactly that: stream stream into stream in your DEBUG(). (Pun intended).
Second, operator << is not defined for string literals, and you are trying to invoke it here:
"The value is: " << i
+ is not defined for literals either, by the way.
To achieve the semantic you want to see, you will have to start with the stream. String literal need to be converted to stream first, and than you can apply << to it. This is ONLY way to achieve what you want.
Edit:
Now since I understand the rationale, I can give a better answer. There are many ways how people are trying to segregate different levels of debugging uniformely, and there are several libraries aiming for that (log4cpp, boost.log to name just few). Before you start implementing your own logging, I would definitely suggest looking into those. There is much more to the good logging than just debug levels.
If, for any reason, you want to use your own homebrew, here are the couple of recepies you might explore:
Use your own logger class (one of the very rare examples, close to
the single one! where Singleton is appropriate). You can than set the
logging level in the beggining of your application, and than just
call Logger::debug() << ...
Enrich above solution with macros. The problem with functions is that, unlike macros, they loose context. So if you want to log file and line number of the logging invocation (and you usually do!), you might want to do LOG_DEBUG << ...; here LOG_DEBUG would expand into something like Logger::debug() << __FILE__ << ":" << __LINE__ << ....
Once you've done this, you will see that sometimes you call other functions inside the << chain. At this point you might realize that those functions would be called regardless of your debug level, and might think you do not want to call them when debugging is not enabled (something along the lines LOG_DEBUG << " Object now is " << object.serialize(); So you will want to enrich the LOG_DEBUG macro to not execute anything when debug level does not match.
And the saga continues... Ready to use the library?
Well, what (at least some) logging libraries would do is create a temporary proxy object that would act as a stream:
#include <iostream>
struct LoggerProxy {
LoggerProxy(const char* file, int line)
{
std::cout << "File " << file << ", line " << line << ": ";
}
template<typename T>
LoggerProxy& operator<<(T&& t)
{
std::cout << t;
return *this;
}
};
#define LOG_DEBUG LoggerProxy{__FILE__, __LINE__}
int main()
{
LOG_DEBUG << "Value is: " << 4;
}
You can do a lot of fancy stuff with this, such as debug level checks, output to different streams or multiple backends (such as simultaneous output to std::cout/cerr and log file) and many more.
What does the following code do in C/C++?
if (blah(), 5) {
//do something
}
Comma operator is applied and the value 5 is used to determine the conditional's true/false.
It will execute blah() and get something back (presumably), then the comma operator is employed and 5 will be the only thing that is used to determine the true/false value for the expression.
Note that the , operator could be overloaded for the return type of the blah() function (which wasn't specified), making the result non-obvious.
If the comma operator is not overloaded, the code is similar to this:
blah();
if (5) {
// do something
}
If the comma operator is overloaded, the result will be based on that function.
#include <iostream>
#include <string>
using namespace std;
string blah()
{
return "blah";
}
bool operator,(const string& key, const int& val) {
return false;
}
int main (int argc, char * const argv[]) {
if (blah(), 5) {
cout << "if block";
} else {
cout << "else block";
}
return 0;
}
(edited to show comma operator overloading scenario. thanks to David Pierre for commenting on this)
I know one thing that this kind of code should do: it should get the coder fired. I would be quite a bit afraid to work next to someone who writes like this.
In the pathological case, it depends on what the comma operator does...
class PlaceHolder
{
};
PlaceHolder Blah() { return PlaceHolder(); }
bool operator,(PlaceHolder, int) { return false; }
if (Blah(), 5)
{
cout << "This will never run.";
}
I would say that depends on blah().
On a more broad answer. The comma operator (non overloaded) resolves as in, execute the first part and return the second part.
So if you have (foo(),bar()) Both functions will be executed, but the value of the expression evaluates to bar() (and the type of the expression as well).
While I won't say there are fair usages for that, is usually considered a bit hard to read code. Mainly because not many languages shares such constructs. So As a personal rule of thumb I avoid it unless I am adding code to a preexistent expression and don't want to change completely its format.
Example: I have a Macro (not discussing if you should use macros or not, sometimes its not even you that wrote it)
FIND_SOMETHING(X) (x>2) ? find_fruits(x) : find_houses(x)
And I usually use it in assignments like my_possession = FIND_SOMETHING(34);
Now I want to add log to it for debuggin purposes but I cannot change the find functions,. I could do :
FIND_SOMETHING(X) (x>2)? (LOG("looking for fruits"),find_fruits(x)):(LOG("looking for houses"),find_houses(x))
I use sometimes constructs like this for debugging purposes. When I force the if close to be true regardless of the return value of blah.
It's obvious that it should never appear in production code.
The following was written assuming it is C code, either in a C file or within a C block of a C++ file:
It is a pointless if. It will call blah(), however the result of blah() is not considered by if at all. The only thing being considered is 5, thus the if will always evaluate to true. IOW you could write this code as
blah();
// do something
without any if at all.
I am trying to do a simple logging library only for my own. I know there exists several once, but I have not found any header-only, small and very "c++" like logging library which fits into my application.
Currently I have the following syntax:
logger << debug << "A debug message" << end; //debug and end is my custom manipulators
I have implemented all necessary operator<< and it works great, specially when it have backward compatibility with std::ostream. But I wonder, just for efficiency if it is a why to stop evaluate anything if some message should not be logged (after debug in the example)? Making everything after the severity manipulator "disappear"?
Just now do I have the following code in short:
template <typename Type>
Logger & Logger::operator<<(const Type & message)
{
if(this->get_message_level() <= this->get_severity())
{
BOOST_FOREACH(std::ostream* stream, this->_sinks)
{
*stream << message;
}
}
return *this;
}
Logger & Logger::operator<< (Logger & (*pf)(Logger&))
{
return pf(*this);
}
Logger & debug(Logger& logger)
{
logger.lock();
logger.set_severity(7);
//...
return logger;
}
Logger & end(Logger& logger)
{
logger << std::endl;
logger.unlock();
return logger;
}
Thanks in advance.
You could do something as simple as
extern "C" bool want_log;
#define LOG(Arg) do { if (want_log) \
cout << __FILE__ << ":" << __LINE__ ": " << Arg << endl; } while(0)
and use it as LOG("x=" << x)
It can be a bit tricky, depending on what compromizes you're willing to
accept in the syntax. If you really want to support everything that
outputting to an ostream does, then the best you can do (as far as I
know) is a wrapper class, along the lines of:
class Logger
{
std::ostream* myDest;
public:
// Appropriate constructors...
template<typename T>
Logger& operator<<( T const& obj )
{
if ( myDest != NULL )
(*myDest) << obj;
return *this;
}
};
If logging is turned off (Logger::myDest == NULL), none of the
conversion code will execute, but you'll still evaluate each of the
arguments. In my experience, this is not usually an issue, since most
of the arguments are either string literals or a simple variable, but
it's not a total 0 cost. It also has the potential disadvantage that
the propagated type is not std::ostream& (although I've never found
this to be a problem in practice).
A somewhat tricker solution would be to use a macro along the lines of:
#define logger loggerInstance != NULL && (*loggerInstance)
This will still allow most of the actual uses with the same syntax
(because the && operator has very low precedence), but could fail in
cases where someone has tried to be too clever, and embedded the logging
output in a more complicated expression. In addition to not doing the
conversions, the arguments are not even evaluated if logging is turned
off.
Finally, if you accept a different syntax, you can write something like:
#ifndef NDEBUG
#define LOG(argList) logger << argList
#else
#define LOG(argList)
#endif
This requires the user to write LOG("x = " << x), instead of log <<
"x = " << x, and requires recompiling if you want to turn logging on,
but it is the only solution I know which has absolute 0 cost if logging
is turned off.
In my experience, most applications can support the first solution; in a
very few cases, you might want to use the second; and I've never seen an
application where performance required the third.
Note that even with the first, you'll probably want to use a macro to
get the logger instance, in order to automatically insert __FILE__ and
__LINE__, and that in the second, you'll probably still want to use a
wrapper class, in order to ensure a flush in the destructor; if the
application is multithreaded, you'll want the wrapper in all cases, in
order to make the entire sequence of output atomic.
You could check for the severity in the Logger & Logger::operator<< (Logger & (*pf)(Logger&)) operator and just return an "empty" logger that doesn't print anything in that case:
class EmptyLogger : public Logger {
template <typename Type>
Logger & Logger::operator<<(const Type & message) { return *this; }
};
EmptyLogger empty; // Can be global/static, it won't cause any race conditions as it does nothing
Logger & Logger::operator<< (Logger & (*pf)(Logger&))
{
if (logger.get_severity() < 5) // Replace with a proper condition
return empty;
return pf(*this); // What does pf() do? Aren't you returning reference to a temporary object?
}
Take a look at this article in dr dobbs about logging in c++:
http://drdobbs.com/cpp/201804215?pgno=2
This page addresses your particular concern but I'd recommend reading the whole article.
I implemented a logging system based on this article and was really pleased with it.
I am in progress of refactoring a C++ application of mine. Before I used a macro like
LOG("something interesting") // 1
LOG("something ended") // 2
LOG("foo: " << bar) // 3
My idea now was to write a Log class like so:
Log(std::string _init_message):
init_message(_init_message)
{ PrintLogLine(init_message + " ...");}
~Log()
{ PrintLogLine(init_message + " done.");}
to get "automatic" logging of specific actions (i.e. when they start, stop + additionally timings etc.) when I use it like
void ActionXYZ() {
Log log("xyz");
// do stuff
}
Where I am struggling is in defining a way to make it work for case 3). In Java I could use a method which takes one String argument since the compiler takes care of automatically string building. What possibilities do I have in C++?
Can I make it work so that I can use it like either one option?
// in "do stuff"
log("foo:" + bar); // OR
log << "foo:" << bar;
As I've mentioned in the comments, you could use Boost.Format. It also helps with the problem of int-to-string conversions, etc. It might get a bit verbose, though — to avoid calling .str() to call std::string constructor, you could make one that accepts boost::format directly.
Log log(boost::format("foo %1% bar") % 42); // with Log(boost::format)
Log log((boost::format("foo %1% bar") % 42).str()); // with only Log(std::string)
Refer to Boost.Format documentation for details.
Two immediate possibilities come to mind. The first is to take advantage of std::string appending:
Log log(std::string("foo:") + bar);
The second is to make more log constructors that take additional parameters:
Log log("foo:", bar);
You really should consider using Boost.Log for your logging. Logging can be a complex thing; it's useful to get a fully formed implementation.
You can create a logging class that inherits from std::strstream.
class Mstream : public std::strstream
{
public:
Mstream() : std::strstream(Buffer = new char[BUFLEN], BUFLEN, ios::out);
ostream& endf(ostream& s);
void Write();
private:
char* Buffer;
};
Now you can log output as,
Mstream m;
m <<"Foo"<<s<<endf;
In endf(ostream& s) you can cast ostream to Mstream and call the Write(). In Write(), you format the output and print it to console or to a file.
Is it possible to define a second insertion operator to have two modes of outputting a class? Say e.g. one that outputs all members and one that just outputs some basic unique identifier that is grep-able in a log? If so, is there an operator that is usually chosen? I would guess as analogy to << one might use <<< if that is legal?
Thanks
If you want to output only the id, then the best idea is probably to provide a method to get the id in a type that's streamable (e.g. std::string id() const;). That's much more intuitive to other people working on the code than some strange operator use.
Your suggestion of <<< (it's not possible to create new operators in C++, but ignoring that for a moment) reveals that you're happy for there to be different code at the point of call. Therefore, the only benefit you'd get would be the saving of a few character's source code; it isn't worth the obfuscation.
By way of contrast, there are situations where you want the same streaming notation to invoke different behaviours, such as switching between id-only and full data, or different representations such as tag/value, CSV, XML, and binary. These alternatives are usually best communicated by either:
using different stream types (e.g. XMLStream rather than std::ostream), and defining XMLStream& operator<<(XMLStream&, const My_Type&) etc, and/or
using stream manipulators - you can create your own - random Google result: http://www.informit.com/articles/article.aspx?p=171014&seqNum=2
There's no such thing already defined or in use by convention.
Also, you cannot define your own operators in C++, you have to use one of the ones already in the language and overloadable, and <<< isn't an operator in C++, so it is out anyway.
I'd strongly recommend you don't use some other operator for this. (See rule #1 here for a more thorough explanation.) If you have subtle differences between output operations, well-chosen functions names go a long way for making better code than unclear operators arbitrarily picked.
No. You can't define your own operators (<<< doesn't exist in C++). But you can define a id() method returning a string and output this.
There is no such operator as <<< in C++.
You are, however, free to implement, for example operator <(ostream&,Object&), which would do what you want. The problem is, code may get unreadable when you try to chain < and << together.
you can use operator | for instance. Another way of doing this is to define small tag classes for which the operator is overloaded; example (pretty simplistic but you get the point):
template< class T >
struct GrepTag
{
GrepTag( const T& );
T value;
}
template< class T >
Greptag< T > MakeGrepTag( const T& x )
{
return GrepTag< T >( x );
}
template< class T >
MyClass& MyClass::operator << ( const GrepTag< T >& g )
{
//output g.value here
}
MyClass() << MakeGrepTag( "text" );
Yet another way, more like the standard streams, is to use a tag as well but keep some state internally:
struct GrepTag
{
}
MyClass& MyClass::operator << ( const GrepTag& g )
{
grepState = true;
}
template< class T >
MyClass& MyClass::operator << ( const T& )
{
if( grepState )
{
//output special
grepState = false;
}
else
{
//output normal
}
}
MyClass() << GrepTag() << "text";
You cannot define your own operators in C++. You can only overload those that exist.
So I recomend not using an operator for outputting basic unique identifier grep-able in a log. This doesn't correspond to any existing operator role. Use a method instead, such as exportToLog().