How does (boost like) header only libraries influence compilation size?

Question

When I include boost library hpp header in two translation units, does the boost code compile twice (and the binary size is double, comparing to the traditional precompiled library?)?

Answer 1

You've asked two distinct questions:

does the boost code compile twice? Yes it does. The net effect is that compilation time is probably a little longer as the compiler has to digest all the headers for each compilation unit.

the binary size is double? No it probably does not, but that will come down to your choice of optimisation flags. A template instantiated in unit A will notionally share the same implementation code as one instantiated in unit B with exact same type parameters.

Whether it actually shares the same code will depend on whether your optimisation flags permit inlining of the template implementation. If you have allowed inlining and the compiler has chosen to do it then your binary size will increase as the compiler places the template implementation inline with your code in order to achieve your stated optimisation goals.

Inlining is never possible with a binary-only library and so this is one reason why your binary size could grow when using header-only libraries.

Here's an example to show how gcc shares template implementation when inlining is not enabled:

a.cpp

#include <vector>

void test1() {
  std::vector<int> myvec;
  myvec.push_back(1);
}

b.cpp

#include <vector>

void test2() {
  std::vector<int> myvec;
  myvec.push_back(1);
}

m.cpp

extern void test1(),test2();

main() {
  test1();
  test2();
}

compilation

g++ -g -O0 -fno-inline -c m.cpp
g++ -g -O0 -fno-inline -c a.cpp
g++ -g -O0 -fno-inline -c b.cpp
g++ -o a.out a.o b.o m.o
objdump -S a.out |less

analysis of objdump

void test1() {
  // [snip]
  myvec.push_back(1);
  // [snip]
  4008a0:       e8 d1 00 00 00          callq  400976 <_ZNSt6vectorIiSaIiEE9push_backERKi>

void test2() {
  // [snip]
  myvec.push_back(1);
  // [snip]
  401548:       e8 29 f4 ff ff          callq  400976 <_ZNSt6vectorIiSaIiEE9push_backERKi>

Note how the same implementation of push_back is used (location 400976) even though it was compiled into completely separate compilation units.

Answer 2

There are some boost header files which are particularly large(I'm looking at you boost/lexical_cast.hpp), which will cause the binary to be larger. However, compilers offer some options to help:

MSVC has an option /LTCG (Link-time Code Generation)

GCC has -flto(I believe is enabled with -O3)

These options generally permit the linker to discard unused components and reduce duplication across the compilation unit.