[CMake] correct dependencies of update-if-different generated files?`

[Matthew Woehlke]

On 2013-07-29 19:02, Matthew Woehlke wrote:
> Let's say¹ I have a code generation tool, 'gen', that I am naïvely using
> to build a library, like so:
>
> add_custom_command(
>    OUTPUT a.cpp b.cpp
>    COMMAND gen in.xml
>    DEPENDS in.xml
> )
> add_library(foo a.cpp b.cpp)
>
> Now let's say that 'gen' is clever and does not overwrite the output
> files with identical content (sort-of like configure_file).
>
> Now, if in.xml is changed, the 'gen' command will run every time because
> the mtime of in.xml is newer than that of a.cpp and b.cpp.
>
> BUT... this "only" affects gmake (at least, it does not affect ninja²
> and reportedly does not affect VS).
>
> Has anyone run into this before, and/or is there a way to express the
> build rules so that the build will completely converge (will not execute
> any commands) *without* unnecessarily recompiling the generated source
> files?
>
> And, since I don't believe that there is, how would folks feel about
> improving add_custom_command to specifically handle this case? I think
> in pure make it is possible to solve the problem by having the custom
> command also update a sentinel file on successful completion, and doing
> like:
>
> a.cpp b.cpp: sentinel
>    # No command here!
>
> sentinel: in.xml
>    gen in.xml
>    touch sentinel
>
> Such an enhancement would also be desirable as it would allow
> 'automatic' adjusting of the build rules for generators that have the
> problem with only a minor change to the CMake rules, and without
> penalizing generators that are more clever.
>
> (¹ Original/actual code where I ran into this:
> https://codereview.qt-project.org/61497)
>
> (² Ninja is not affected because it records the last time it updated the
> output files independent of their mtime.)

No comments?

-- 
Matthew