Conan: [workspaces] Support for the cmake_paths generator

[AlessandroA]

Conan version: 1.6.1

Hi, it would be nice if workspaces had a specialized behaviour for the
cmake_paths generator.

The issue with find_package

cmake_paths is useful to keep build and packaging systems separate because it
allows us to use find_package commands as we would do without conan.

The issue is that find_package does not work well with add_subdirectory. The
recommended solution is to re-map find_package from the top-level CMakeLists
file so that it becomes a no-op when the package has been added as a sub-project
(see Daniel Pfeifer's talk at 50m30s).

Feature request

When using workspaces with cmake_paths, create a
top-level CMakeLists that defines a list of sub-projects and re-defines
find_package for them.

The top-level CMakeLists file could look something like this:

cmake_minimum_required(VERSION 3.3)
project(MyProject CXX)

# Add this #######################
set(AS_SUBPROJECT
    HelloA
    HelloB
    HelloC
)

macro(find_package)
    if(NOT "${ARGV0}" IN_LIST AS_SUBPROJECT)
        _find_package(${ARGV})
    endif()
endmacro()
# ###############################

add_subdirectory(HelloA "/home/user/code/build/helloa")
add_subdirectory(HelloB "/home/user/code/build/hellob")
add_subdirectory(HelloC "/home/user/code/build/helloc")

---

One could maintain their own top-level CMakeLists file, but conan helps a lot
because the order of the add_subdirectory commands is important.

Not sure if this would be consistent with your idea of workspaces and the use of
cmake_paths!

[lasote]

Hi @AlessandroA, this is very interesting, thanks. We will have a look.

[Lawrencemm]

I believe this would break my project.

My solution to this problem is to always link against the interface target generated by a FindXXX.cmake script whether my executable is in a workspace or not.
See here:
https://github.com/Lawrencemm/workspace

lmeditor consumes lmlib as a package requirement.

Making find_package a no-op would break my project I think.

[Lawrencemm]

I have a different idea for supporting cmake_paths workspace generator.

When specifying cmake_paths as the workspace generator, the generated conanworkspace.cmake would look like this:

set(PACKAGE_foo_SRC "...")
set(PACKAGE_foo_BUILD "...")
set(PACKAGE_bar_SRC "...")
set(PACKAGE_bar_BUILD "...")
macro(conan_workspace_subdirectories)
    set(CMAKE_PROJECT_foo_INCLUDE ${PACKAGE_foo_BUILD}/conan_paths.cmake)
    set(CMAKE_PROJECT_bar_INCLUDE ${PACKAGE_bar_BUILD}/conan_paths.cmake)
    add_subdirectory(${PACKAGE_foo_SRC} ${PACKAGE_foo_BUILD})
    add_subdirectory(${PACKAGE_bar_SRC} ${PACKAGE_bar_BUILD})
endmacro()

The only difference is the addition of the setting of CMAKE_PROJECT_foo_INCLUDE variables.

[lasote]

Sounds very good. Did you try?

[Lawrencemm]

Success! I tried it by manually hacking conanworkspace.cmake, FindXXX was failing before, after adding the project include variables manually the packages were found.

There are some things that are really important to note, however.

1. This will only work if project name in the CMakeLists.txt of each package has the same name as the package itself. (spent a bit of time debugging that one :sweat_smile:). This might be overcome with an override in the workspace yml file.

2. The include variables should only be defined for packages that have the cmake_paths generator set. Otherwise, the inclusion will obviously fail as the cmake_paths.cmake file won't be there.

[lasote]

Mhh, after taking a look at this: https://github.com/conan-io/conan/issues/4808
Maybe what you are expecting from a cmake_paths generator is only adjust the CMAKE_MODULE_PATH and CMAKE_PREFIX_PATH to the [buildirs] directories from the projects in the workspace. You could assume that there, CMake will find some FindXXX.cmake or ConfigXXX.cmake files. That would be useful when the projects export their own targets.
WDYT?

[lasote]

To consider: If you edit a file in a subproject, how well the root project is able to "detect" it and recompile if needed etc. It exceeds my knowledge. I would need to try, so I'm tagging it to take a look to get more knowledge to decide what/when to implement.

[mbodmer]

What about ExternalProject_Add() to create the superbuild project? ... as I suggested here: https://github.com/conan-io/conan/issues/3034#issuecomment-477997115

[jgsogo]

Really interesting thread 👍

Some of the alternatives proposed here are very related to some POC I did in #4879. In this POC the standard CMakeLists.txt for any library file looks like this:

conan_basic_setup(TARGETS)

add_library(mylib ...)
target_link_libraries(mylib PUBLIC CONAN_PKG::mydep)

Here CMake/Conan machinery is creating an _imported target_ called CONAN_PKG::mydep (it would be mydep::mydep with find package approach) that I'm adding as a dependency of my project.

But when I'm working in a workspace these CMakeLists.txt are being added to the _parent_ project using the add_subdirectory() and I no longer want to link with the _imported target_ but to the target itself, otherwise CMake won't build the correct graph and some issues could arise.

To bypass it and keep my project files the same we need to do something different, like some of you are proposing in this issue. In my POC I don't need the imported target anymore, CONAN_PKG::mydep should be the target itself, so I can create an ALIAS target (no need to modify my target_link_libraries line). The generated conanworkspace.cmake could look like the following:

set(PACKAGE_mydep_SRC "path/to/mydep/src")
set(PACKAGE_mylib_SRC "path/to/mylib/src")

macro(conan_workspace_subdirectories)
    set(CONAN_IS_WS TRUE)

    add_subdirectory(${PACKAGE_mydep_SRC} "mydep/build/folder")
    add_library(CONAN_PKG::mydep ALIAS mydep)  # for cmake generator

    add_subdirectory(${PACKAGE_mylib_SRC} "mylib/build/folder")    
    add_library(CONAN_PKG::mylib ALIAS mylib)  # for cmake generator
endmacro()

And also, it could make sense to no-op all the conan_basic_setup (or find_package) calls for the packages included in the workspace. With the find_package approach it is easy as @AlessandroA suggested, for the conan_basic_setup one it would require a conan_workspace_basic_setup call in the root CMakeLists.txt of the workspace to do the Conan magic.

---

The PR and this comment is just a draft, I haven't tried it with cmake_paths or cmake_find_package generators, but I think that the problem is almost always the same: changing the imported target by the actual one for the libraries in the workspace.

[lasote]

We have still some things to discuss about this so no time enough for 1.15. I will reconsider it for 1.16

[kenfred]

Here CMake/Conan machinery is creating an _imported target_ called CONAN_PKG::mydep (it would be mydep::mydep with find package approach) that I'm adding as a dependency of my project.

But when I'm working in a workspace these CMakeLists.txt are being added to the _parent_ project using the add_subdirectory() and I no longer want to link with the _imported target_ but to the target itself, otherwise CMake won't build the correct graph and some issues could arise.

As I mentioned in #5762, I think we should use externalproject_add and not add_subdirectory. I don't think you gain anything by avoiding the imported targets. I believe the concern about CMake not building the correct graph is mitigated by using the dependency mechanisms provided by externalproject_add.

Although I like that add_subdirectory will result in a nice makefile or Visual Studio project with all the source files there, you lose several things by doing this:

• You lose an opportunity to run conan build, package, or test from the super project. With externalproject_add you run these as commands. You could potentially avoid handling special, alternate layouts due to dependencies reaching into build directories instead of package directories.
• You're limiting the workspace packages to CMake projects only. With externalproject_add you could have a variety of build systems under one super build. A cmake workspace generator shouldn't have to require the projects all be cmake.
• Unifying the cmake configure across the packages could be bad! There could be conflicting targets, redefinition of CMake variables, etc. In other words, you risk affecting a package's build by joining its cmake configuration with other packages.
• You have to conditionally do the Conan setup or find_package because with add_subdirectory you're changing how targets are defined. I shouldn't need to modify CMakelists.txt to use workspaces. So the cmake workspace generator should do externalproject_add to isolate the builds, providing module paths, toolchains, etc for the individual builds.

[Hopobcn]

Using ExternalProject_Add offers also some challenges:

• cmake configure step is deferred until the superbuild target is called. This could break projects that have targets depending on other workspace targets provided by conan
• Conan should provide the dependency graph via cmake DEPENDS

but also some benefits:

• multiple toolchains could be mixed (ideal for using with CMakeToolchain)

Something like this: (didn't work in my case)

include(ExternalProject)

set(PACKAGE_libA_SRC "path/to/libA/.")
set(PACKAGE_libA_BUILD "path/to/libA/build/Release")
set(PACKAGE_libB_SRC "path/to/libB/.")
set(PACKAGE_libB_BUILD "path/to/libB/build/Release")
set(PACKAGE_libC_SRC "path/to/libC/.")
set(PACKAGE_libC_BUILD path/to/libC/build/Release")
set(PACKAGE_libD_SRC "path/to/libD/.")
set(PACKAGE_libD_BUILD "path/to/libD/build/Release")
set(PACKAGE_App1_SRC "path/to/App1/.")
set(PACKAGE_App1_BUILD "path/to/App1/build/Release")

macro (conan_add_package)
    set(oneValueArgs NAME SRC BUILD)
    set(multiValueArgs DEPENDS)
    cmake_parse_arguments(CONAN_ADD_PACKAGE "" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
    if (NOT CONAN_ADD_PACKAGE_NAME)
        message(FATAL_ERROR "You must provide a name")
    endif ()
    if (NOT CONAN_ADD_PACKAGE_SRC)
        message(FATAL_ERROR "You must provide a source dir")
    endif ()
    if (NOT CONAN_ADD_PACKAGE_BUILD)
        message(FATAL_ERROR "You must provide a build dir")
    endif ()
    if (NOT CONAN_ADD_PACKAGE_DEPENDS)
        message(WARNING "You must provide all dependencies")
    endif ()

    message("Adding ${CONAN_ADD_PACKAGE_NAME}")

    ExternalProject_Add(${CONAN_ADD_PACKAGE_NAME}
        PREFIX ${CONAN_ADD_PACKAGE_NAME}
        SOURCE_DIR ${CONAN_ADD_PACKAGE_SRC}
        BINARY_DIR ${CONAN_ADD_PACKAGE_BUILD}
        CMAKE_ARGS -DCMAKE_TOOLCHAIN_FILE=${CONAN_ADD_PACKAGE_BUILD}/conan_toolchain.cmake
        DEPENDS ${CONAN_ADD_PACKAGE_DEPENDS}
    )
endmacro ()

macro(conan_workspace_subdirectories)
    conan_add_package(NAME libA SRC ${PACKAGE_libA_SRC} BUILD ${PACKAGE_libA_BUILD} DEPENDS "")
    conan_add_package(NAME libB SRC ${PACKAGE_libB_SRC} BUILD ${PACKAGE_libB_BUILD} DEPENDS libA)
    conan_add_package(NAME libC SRC ${PACKAGE_libC_SRC} BUILD ${PACKAGE_libC_BUILD} DEPENDS libA)
    conan_add_package(NAME libD SRC ${PACKAGE_libD_SRC} BUILD ${PACKAGE_libD_BUILD} DEPENDS libB libC)
    conan_add_package(NAME App1 SRC ${PACKAGE_App1_SRC} BUILD ${PACKAGE_App1_BUILD} DEPENDS libD)
endmacro()

But I'm not sure ExternalProject_Add is a good idea. In my company, we are using it and it's a pain in the ass. Once you start using it there's no way back. And that is never good.

To be honest, I would prefer conan workspace to stick to add_subdirectory. But generating FindXXX.cmake and conan_toolchains.cmake in the workspace build directory and not in the build directories from each component.