cmake failures due to relative include dirs in pkg-config files and caches

[meyerj]

With the newest catkin version introduced in hydro (currently 0.5.73) I am getting the following cmake errors whenever cmake is run the second time in the same catkin workspace (e.g. if a CMakeLists.txt file has been changed):

-- ==> add_subdirectory(msgs/fmp_msgs)
-- fmp_msgs: 2 messages, 0 services
CMake Error at /opt/ros/hydro/share/catkin/cmake/catkin_package.cmake:335 (message):
  catkin_package() include dir 'include' is neither an absolute directory nor
  exists relative to '/opt/uxvcos/hydro/msgs/fmp_msgs'
Call Stack (most recent call first):
  /opt/ros/hydro/share/catkin/cmake/catkin_package.cmake:98 (_catkin_package)
  msgs/fmp_msgs/CMakeLists.txt:51 (catkin_package)

The fmp_msgs is nothing special, just a standard message only package that has no include directories at all:

...
catkin_package(
#  INCLUDE_DIRS include
#  LIBRARIES fmp_msgs
  CATKIN_DEPENDS message_runtime std_msgs
#  DEPENDS system_lib
)
...

I guess the problem is that the cmake cache contains the fmp_msgs_INCLUDE_DIRS variable after the first invocation and catkin_package() prepends these paths to the INCLUDE_DIRS specified in its arguments in catkin_package.cmake:119 and then complains about the missing directory in catkin_package.cmake:335:

fmp_msgs_INCLUDE_DIRS:INTERNAL=/opt/uxvcos/hydro/devel/include;include

The relative include entry was pulled in from std_msgs.pc somehow:

prefix=/opt/ros/hydro

Name: std_msgs
Description: Description of std_msgs
Version: 0.5.8
Cflags: -Iinclude
Libs: -L/opt/ros/hydro/lib 
Requires: message_runtime

Is this problem related to #490, or is there anything else I might be doing wrong here?

[meyerj]

Hmm, I cannot reproduce the cmake error anymore in a clean workspace. Nevertheless, it might be wrong if pkg-config reports a relative include directory on a fresh installation from the ROS Ubuntu repositories:

$ pkg-config --cflags-only-I std_msgs
include
$ rospack cflags-only-I std_msgs
include

[dirk-thomas]

I could not reproduce your original issue either but the relative include path in the .pc file was definitely a regression of #490.