It turns out passing libraries to link in target_link_libraries() as ${foo_pkg_LIBRARIES} introduces a lot of duplicate of the same library that needs to be linked. This the reason why the ld linker takes a REALLY long time to finish linking since it does not de-duplicate the duplicate libraries and links all them. The lld linker seems to be smarter about this.
Switching to ${foo_pkg_TARGETS} prevents the duplicate libraries that were previously included as transitives. This is also the pattern that ROS 2 follows. See.
With these changes, I can compile the workspace faster than ever with just ld.
I removed the need to install and use lld in the README so that users can follow regular ROS patterns to build the workspace. But kept it in CI so that i can be faster since at the end of the day lld with these changes is faster than lld without these changes.
It turns out passing libraries to link in
target_link_libraries()as${foo_pkg_LIBRARIES}introduces a lot of duplicate of the same library that needs to be linked. This the reason why theldlinker takes a REALLY long time to finish linking since it does not de-duplicate the duplicate libraries and links all them. Thelldlinker seems to be smarter about this.Switching to
${foo_pkg_TARGETS}prevents the duplicate libraries that were previously included as transitives. This is also the pattern that ROS 2 follows. See.With these changes, I can compile the workspace faster than ever with just
ld.I removed the need to install and use
lldin the README so that users can follow regular ROS patterns to build the workspace. But kept it in CI so that i can be faster since at the end of the daylldwith these changes is faster thanlldwithout these changes.