tinyformat.h

std::string weekday = "Wednesday";
const char* month = "July";
size_t day = 27;
long hour = 14;
int min = 44;

tfm::printf("%s, %s %d, %.2d:%.2d\n", weekday, month, day, hour, min);

tfm::format(std::cerr, "%s, %s %d, %.2d:%.2d\n",
 weekday, month, day, hour, min);

std::string date = tfm::format("%s, %s %d, %.2d:%.2d\n",
 weekday, month, day, hour, min);
std::cout << date;

template<typename T>
void myPrint(const T& value)
{
 tfm::printf("My value is '%s'\n", value);
}

template<typename... Args>
void format(std::ostream& stream, const char* formatString,
 const Args&... args);

template<typename... Args>
std::string format(const char* formatString, const Args&... args);

template<typename... Args>
void printf(const char* formatString, const Args&... args);

[1] Generating the code to support more arguments is quite easy using the in-source code generator based on the excellent code generation script cog.py (Format strings and type safety

Tinyformat parses C99 format strings to guide the formatting process --- please refer to any standard C99 printf documentation for format string syntax. In contrast to printf, tinyformat does not use the format string to decide on the type to be formatted so this does not compromise the type safety: you may use any format specifier with any C++ type. The author suggests standardising on the %s conversion unless formatting numeric types.

Let's look at what happens when you execute the function call:

tfm::format(outStream, "%+6.4f", yourType);

First, the library parses the format string, and uses it to modify the state of outStream:

1. The outStream formatting flags are cleared and the width, precision and fill reset to the default.
2. The flag '+' means to prefix positive numbers with a '+'; tinyformat executes outStream.setf(std::ios::showpos)
3. The number 6 gives the field width; execute outStream.width(6).
4. The number 4 gives the precision; execute outStream.precision(4).
5. The conversion specification character 'f' means that floats should be formatted with a fixed number of digits; this corresponds to executing outStream.setf(std::ios::fixed, std::ios::floatfield);

After all these steps, tinyformat executes:

outStream << yourType;

and finally restores the stream flags, precision and fill.

What happens if yourType isn't actually a floating point type? In this case the flags set above are probably irrelevant and will be ignored by the underlying std::ostream implementation. The field width of six may cause some padding in the output of yourType, but that's about it.

Special cases for "%p", "%c" and "%s"

Tinyformat normally uses operator<< to convert types to strings. However, the "%p" and "%c" conversions require special rules for robustness. Consider:

uint8_t* pixels = get_pixels(/* ... */);
tfm::printf("%p", pixels);

Clearly the intention here is to print a representation of the pointer to pixels, but since uint8_t is a character type the compiler would attempt to print it as a C string if we blindly fed it into operator<<. To counter this kind of madness, tinyformat tries to static_cast any type fed to the "%p" conversion into a const void* before printing. If this can't be done at compile time the library falls back to using operator<< as usual.

The "%c" conversion has a similar problem: it signifies that the given integral type should be converted into a char before printing. The solution is identical: attempt to convert the provided type into a char using static_cast if possible, and if not fall back to using operator<<.

The "%s" conversion sets the boolalpha flag on the formatting stream. This means that a bool variable printed with "%s" will come out as true or false rather than the 1 or 0 that you would otherwise get.

Incompatibilities with C99 printf

Not all features of printf can be simulated simply using standard iostreams. Here's a list of known incompatibilities:

• The C99 "%a" and "%A" hexadecimal floating point conversions are not supported since the iostreams don't have the necessary flags. Using either of these flags will result in a call to TINYFORMAT_ERROR.
• The precision for integer conversions cannot be supported by the iostreams state independently of the field width. (Note: this is only a problem for certain obscure integer conversions; float conversions like %6.4f work correctly.) In tinyformat the field width takes precedence, so the 4 in %6.4d will be ignored. However, if the field width is not specified, the width used internally is set equal to the precision and padded with zeros on the left. That is, a conversion like %.4d effectively becomes %04d internally. This isn't correct for every case (eg, negative numbers end up with one less digit than desired) but it's about the closest simple solution within the iostream model.
• The "%n" query specifier isn't supported to keep things simple and will result in a call to TINYFORMAT_ERROR.
• The "%ls" conversion is not supported, and attempting to format a wchar_t array will cause a compile time error to minimise unexpected surprises. If you know the encoding of your wchar_t strings, you could write your own std::ostream insertion operator for them, and disable the compile time check by defining the macro TINYFORMAT_ALLOW_WCHAR_STRINGS. If you want to print the address of a wide character with the "%p" conversion, you should cast it to a void* before passing it to one of the formatting functions.

Error handling

By default, tinyformat calls assert() if it encounters an error in the format string or number of arguments. This behaviour can be changed (for example, to throw an exception) by defining the TINYFORMAT_ERROR macro before including tinyformat.h, or editing the config section of the header.

Formatting user defined types

User defined types with a stream insertion operator will be formatted using operator<<(std::ostream&, T) by default. The "%s" format specifier is suggested for user defined types, unless the type is inherently numeric.

For further customization, the user can override the formatValue() function to specify formatting independently of the stream insertion operator. If you override this function, the library will have already parsed the format specification and set the stream flags accordingly - see the source for details.

Wrapping tfm::format() inside a user defined format function

Suppose you wanted to define your own function which wraps tfm::format. For example, consider an error function taking an error code, which in C++11 might be written simply as:

template<typename... Args>
void error(int code, const char* fmt, const Args&... args)
{
 std::cerr << "error (code " << code << ")";
 tfm::format(std::cerr, fmt, args...);
}

Simulating this functionality in C++98 is pretty painful since it requires writing out a version of error() for each desired number of arguments. To make this bearable tinyformat comes with a set of macros which are used internally to generate the API, but which may also be used in user code.

The three macros TINYFORMAT_ARGTYPES(n), TINYFORMAT_VARARGS(n) and TINYFORMAT_PASSARGS(n) will generate a list of n argument types, type/name pairs and argument names respectively when called with an integer n between 1 and 16. We can use these to define a macro which generates the desired user defined function with n arguments. This should be followed by a call to TINYFORMAT_FOREACH_ARGNUM to generate the set of functions for all supported n:

#define MAKE_ERROR_FUNC(n) \
template<TINYFORMAT_ARGTYPES(n)> \
void error(int code, const char* fmt, TINYFORMAT_VARARGS(n)) \
{ \
 std::cerr << "error (code " << code << ")"; \
 tfm::format(std::cerr, fmt, TINYFORMAT_PASSARGS(n)); \
}
TINYFORMAT_FOREACH_ARGNUM(MAKE_ERROR_FUNC)

Sometimes it's useful to be able to pass a list of format arguments through to a non-template function. The FormatList class is provided as a way to do this by storing the argument list in a type-opaque way. For example:

template<typename... Args>
void error(int code, const char* fmt, const Args&... args)
{
 tfm::FormatListRef formatList = tfm::makeFormatList(args...);
 errorImpl(code, fmt, formatList);
}

What's interesting here is that errorImpl() is a non-template function so it could be separately compiled if desired. The FormatList instance can be used via a call to the vformat() function (the name chosen for semantic similarity to vprintf()):

void errorImpl(int code, const char* fmt, tfm::FormatListRef formatList)
{
 std::cerr << "error (code " << code << ")";
 tfm::vformat(std::cout, fmt, formatList);
}

The construction of a FormatList instance is very lightweight - it defers all formatting and simply stores a couple of function pointers and a value pointer per argument. Since most of the actual work is done inside vformat(), any logic which causes an early exit of errorImpl() - filtering of verbose log messages based on error code for example - could be a useful optimization for programs using tinyformat. (A faster option would be to write any early bailout code inside error(), though this must be done in the header.)

Benchmarks

Compile time and code bloat

The script bloat_test.sh included in the repository tests whether tinyformat succeeds in avoiding compile time and code bloat for nontrivial projects. The idea is to include tinyformat.h into 100 translation units and use printf() five times in each to simulate a medium sized project. The resulting executable size and compile time (g++-4.8.2, linux ubuntu 14.04) is shown in the following tables, which can be regenerated using make bloat_test:

Non-optimized build

test name	compiler wall time	executable size (stripped)
libc printf	1.8s	48K (36K)
std::ostream	10.7s	96K (76K)
tinyformat, no inlines	18.9s	140K (104K)
tinyformat	21.1s	220K (180K)
tinyformat, c++0x mode	20.7s	220K (176K)
boost::format	70.1s	844K (736K)

Optimized build (-O3 -DNDEBUG)

test name	compiler wall time	executable size (stripped)
libc printf	2.3s	40K (28K)
std::ostream	11.8s	104K (80K)
tinyformat, no inlines	23.0s	128K (104K)
tinyformat	32.9s	128K (104K)
tinyformat, c++0x mode	34.0s	128K (104K)
boost::format	147.9s	644K (600K)

For large projects it's arguably worthwhile to do separate compilation of the non-templated parts of tinyformat, as shown in the rows labelled tinyformat, no inlines. These were generated by putting the implementation of vformat (detail::formatImpl() etc) it into a separate file, tinyformat.cpp. Note that the results above can vary considerably with different compilers. For example, the -fipa-cp-clone optimization pass in g++-4.6 resulted in excessively large binaries. On the other hand, the g++-4.8 results are quite similar to using clang++-3.4.

Speed tests

The following speed tests results were generated by building tinyformat_speed_test.cpp on an Intel core i7-2600K running Linux Ubuntu 14.04 with g++-4.8.2 using -O3 -DNDEBUG. In the test, the format string "%0.10f:%04d:%+g:%s:%p:%c:%%\n" is filled 2000000 times with output sent to /dev/null; for further details see the source and Makefile.

test name	run time
libc printf	1.20s
std::ostream	1.82s
tinyformat	2.08s
boost::format	9.04s

It's likely that tinyformat has an advantage over boost.format because it tries reasonably hard to avoid formatting into temporary strings, preferring instead to send the results directly to the stream buffer. Tinyformat cannot be faster than the iostreams because it uses them internally, but it comes acceptably close.

Rationale

Or, why did I reinvent this particularly well studied wheel?

Nearly every program needs text formatting in some form but in many cases such formatting is incidental to the main purpose of the program. In these cases, you really want a library which is simple to use but as lightweight as possible.

The ultimate in lightweight dependencies are the solutions provided by the C++ and C libraries. However, both the C++ iostreams and C's printf() have well known usability problems: iostreams are hopelessly verbose for complicated formatting and printf() lacks extensibility and type safety. For example:

// Verbose; hard to read, hard to type:
std::cout << std::setprecision(2) << std::fixed << 1.23456 << "\n";
// The alternative using a format string is much easier on the eyes
tfm::printf("%.2f\n", 1.23456);

// Type mismatch between "%s" and int: will cause a segfault at runtime!
printf("%s", 1);
// The following is perfectly fine, and will result in "1" being printed.
tfm::printf("%s", 1);

On the other hand, there are plenty of excellent and complete libraries which solve the formatting problem in great generality (boost.format and fastformat come to mind, but there are many others). Unfortunately these kind of libraries tend to be rather heavy dependencies, far too heavy for projects which need to do only a little formatting. Problems include

1. Having many large source files. This makes a heavy dependency unsuitable to bundle within other projects for convenience.
2. Slow build times for every file using any sort of formatting (this is very noticeable with g++ and boost/format.hpp. I'm not sure about the various other alternatives.)
3. Code bloat due to instantiating many templates

Tinyformat tries to solve these problems while providing formatting which is sufficiently general and fast for incidental day to day uses.

License

For minimum license-related fuss, tinyformat.h is distributed under the boost software license, version 1.0. (Summary: you must keep the license text on all source copies, but don't have to mention tinyformat when distributing binaries.)

Author and acknowledgements

Tinyformat was written by Chris Foster, with contributions from various people as recorded in the git repository. The implementation owes a lot to boost::format for showing that it's fairly easy to use stream based formatting to simulate most of the printf() syntax. Douglas Gregor's introduction to variadic templates --- see Bugs

Here's a list of known bugs which are probably cumbersome to fix:

• Field padding won't work correctly with complicated user defined types. For general types, the only way to do this correctly seems to be format to a temporary string stream, check the length, and finally send to the output stream with padding if necessary. Doing this for all types would be quite inelegant because it implies extra allocations to make the temporary stream. A workaround is to add logic to operator<<() for composite user defined types so they are aware of the stream field width.