[CMake] Correct way to specify multiple build configurations for single config generators?

[Craig Scott]

On Mon, Dec 10, 2018 at 1:37 PM Zaak Beekman <zbeekman at gmail.com> wrote:

> I have been reading the *excellent* book "Professional CMake". The author,
> Craig Scott, recommends the following best practices:
>
>    - check the existence of `CMAKE_CONFIGURATION_TYPES` and only adding
>    or pruning configurations if it's preset, *AFTER* your call to `project()`
>    - do not set `CMAKE_CONFIGURATION_TYPES` and use the `STRINGS`
>    property on `CMAKE_BUILD_TYPE` for single config generators
>    - forcibly set the `CMAKE_BUILD_TYPE` cache variable if it is unset,
>    otherwise check the passed `CMAKE_BUILD_TYPE` against allowable configs and
>    throw an error if it differs
>    - set `CMAKE_<LANG>_FLAGS_<CONFIG>` and
>    `CMAKE_<TARGETTYPE>_LINKER_FLAGS_<CONFIG>` variables as cache variables for
>    newly defined configuration
>
> The problem I encounter is that, if I set
> `-DCMAKE_BUILD_TYPE:STRING=<MY_CONFIG>` to CMake on the command line, then
> CMake seems to initialize all of the `<MY_CONFIG>` cache variables to the
> empty string *BEFORE* I can set them. All other custom configurations get
> set to my specified default cache variable value.
>
> What is the best practice for setting the new configuration's default
> flags?
>

This is only going to be an issue for single-configuration generators. For
multi-config generators, you don't set CMAKE_BUILD_TYPE and the flags you
set as cache variables after the first project() call should be honoured.
For single-config generators, you have a few choices:

   - Run CMake with CMAKE_BUILD_TYPE set to some other config first, then
   change it to the one you want after the first run. If the custom config is
   one that is only occasionally used and typically just one you switch to
   temporarily from time-to-time, then this may be okay for your situation.
   It's also easy enough to handle in a CI build, albeit slightly less
   efficient since you have to run CMake twice. It's a bit annoying, but it is
   at least easy.
   - Rather than setting the custom config's flags in the project, do it in
   a toolchain file instead (with the various ..._<CONFIG>_INIT variables).
   The flags are going to be toolchain-specific anyway, so you don't lose
   generality. You also gain the advantage that you can take that same
   toolchain file and apply it to other projects, so you get better
   reusability. This would be what I would use personally, but I'm comfortable
   with toolchain files. The main downside to this approach is that those new
   to CMake can feel a bit intimidated by toolchain files or see them as an
   unnecessary complication for a build that should just pick up the default
   compilers and work out of the box. It's a tradeoff, so I erred on the side
   of simplicity in the book and didn't make toolchains a recommendation for
   defining custom configs. If you and your users are happy with toolchain
   files though, I'd go this route.
   - You could set your custom config's cache variables (again I'd
   recommend setting the ..._<CONFIG>_INIT variables) before the first
   project() command, but you won't have the compiler ID or any other compiler
   information at that point. You also won't have any of the default configs'
   flags available to you, so you can't make your custom config an extension
   of an existing one (e.g. you can't start with flags from, say,
   RelWithDebInfo and add a few more to enable profiling like in the book's
   example). If your project only needs to consider one compiler and the flags
   are simple and well-defined, this is going to be the easiest for your
   developers, since it will work out-of-the-box with no extra effort on their
   part.
   - Instead of setting cache variables, you could append to regular
   non-cache ...<CONFIG> variables after the project() command. This has the
   advantage of being simple, but your config-specific flags won't show in the
   cache and developers can't override them without editing the project.

So as you can see from the above, there's no one perfect solution. The best
choice really depends on your project and your users.



>
> I am aware of `CMAKE_USER_MAKE_RULES_OVERRIDE` and all of the
> `CMAKE_<LANG|TARGET_TYPE>_FLAGS_<CONFIG>_INIT` variable, but as far as I
> can tell, there is no access to knowing what compiler is being used, since
> the compilers have yet to be probed. Is it possible to use the
> `CMAKE_..._INIT` variables to set per-compiler flags some how?
>

Use a toolchain file where you are in control of which compiler is
selected. Use different toolchain files for different compilers.



> Any answers or pointers here would be most appreciated.
>
> FYI, Before being enlightened to some of the dangers or non-standard ways
> of what I was doing, I would typically set `CMAKE_CONFIGURATION_TYPES`
> *before* my call to `project()`. Something like this:
>
> ```cmake
> set ( CMAKE_CONFIGURATION_TYPES "Debug" "Release" "MinSizeRel"
> "RelWithDebInfo" "CodeCoverage" )
> set ( CMAKE_BUILD_TYPE "Release"
>   CACHE STRING "Select which configuration to build." )
> set_property ( CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS
> ${CMAKE_CONFIGURATION_TYPES} )
> ```
>
> I would then call `project()` so that the compilers would be probed and I
> could make specific flag choices based on the compiler being used, and set
> those as Cache variables. Something like this:
>
> ```cmake
> if ("${CMAKE_Fortran_COMPILER_ID}" MATCHES "GNU" )
>   set(gfortran_compiler true)
>   set ( CMAKE_C_FLAGS_CODECOVERAGE "-fprofile-arcs -ftest-coverage -O0"
>     CACHE STRING "Code coverage C compiler flags")
>   set ( CMAKE_Fortran_FLAGS_CODECOVERAGE "-fprofile-arcs -ftest-coverage
> -O0"
>     CACHE STRING "Code coverage Fortran compiler flags")
> ```
>
> This seemed to work fine and specifying, e.g.,
> `-DCMAKE_BUILD_TYPE:STRING=CodeCoverage`, the
> `CMAKE_<LANG>_FLAGS_CODECOVERAGE` cache variable would be set to the
> appropriate value.
>

This didn't work for me, it showed the same problem you described in your
original issue. Perhaps you forgot to clear out a previous build first? It
would only work after an initial run of CMake where you didn't set
CMAKE_BUILD_TYPE or you set it to something else.


-- 
Craig Scott
Melbourne, Australia
https://crascit.com

New book released: Professional CMake: A Practical Guide
<https://crascit.com/professional-cmake/>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://cmake.org/pipermail/cmake/attachments/20181210/a8666278/attachment.html>