Massive binary bloat in std::print

[vitaut]

Consider the following trivial program (std-print.cc):

#include <print>

int main() {
  std::print("{}\n", 42);
}

Compiling it with libc++ and Apple clang version 17.0.0 (clang-1700.3.19.1) on macOS results in a ~300kB object file:

% c++ -c -std=c++23 std-print.cc
% ls -l std-print.o
-rw-r--r--  1 viz  staff  297384 Oct 11 07:07 std-print.o

For comparison, an equivalent program using {fmt} (fmt-print.cc)

#include <fmt/base.h>

int main() {
  fmt::print("{}\n", 42);
}

compiles to less than 1kB.

% c++ -c -std=c++23 fmt-print.cc -I include
% ls -l fmt-print.o
-rw-r--r--  1 viz  staff  968 Oct 11 07:08 fmt-print.o

So the implementation of std::print in libc++ generates ~300 times (!) more binary code and is ~5.5 times slower to compile. This is likely due to unnecessary templatization and defining in the header of vprint_unicode:

llvm-project/libcxx/include/print

Lines 387 to 390 in 7eee672

	template <class = void> // TODO PRINT template or availability markup fires too eagerly (http://llvm.org/PR61563).
	_LIBCPP_HIDE_FROM_ABI inline void vprint_unicode(string_view __fmt, format_args __args) {
	std::vprint_unicode(stdout, __fmt, __args);
	}

std::print was explicitly designed such that vprint* and other similar functions were defined in library source files and not headers specifically to avoid such bloat and slow compile times.

[EugeneZelenko]

Could you please try 21 or main branch, LLVM, not Apple?

[vitaut]

The bloat is essentially the same on 21 and can even be demonstrated on godbolt: https://www.godbolt.org/z/x36f9nxas. It shows that vprint_unicode, that should be compiled separately, is emitted in the same TU and pulls in the rest of implementation details.