On Fri, Feb 21, 2025 at 09:40:45 -0500, Paul Smith wrote: > On Fri, 2025-02-21 at 12:26 +0100, Ben Boeckel wrote: > > CMake's Makefile implementation can support it; it supports Fortran > > modules after all. However, `ninja` support is far easier as it (in > > CMake's implementation) has a single graph for the entire build. The > > Makefiles generator is a 3-level recursive make in order to implement > > things like CMake's per-directory entry points. Implementing C++ > > modules and fixing the *ninja* build graph found a number of > > assumptions that this structure ends up relying on to provide CMake's > > semantics to projects. A rewrite of the Makefiles generator to not > > require this is something we'd like to do, but there's no real way to > > put it on the timeline at the moment. Basically it would look like > > the Ninja generator output where there's a single build graph instead > > of the static recursion level. > > I fully agree that the makefiles generated by cmake are... sub-optimal. > > Automake-generated makefiles use a single build graph (or can do so) > and they are POSIX-standard makefiles. So it should be possible to > create POSIX-standard makefiles which still have these features. This > is actually simpler for CMake than for "hand-rolled" makefile > environments since the generated makefiles can avoid tricky parts like > conditionalized content, implicit rules, etc. > > I think it should be possible to preserve the per-directory entry > points, now that POSIX make supports "include". Does it preserve the "defer actually including until its associated recipe is up-to-date" behavior? If my memory serves at least…I've been able to live happily in ninja-land for a long time :) . > Another alternative is to leave the UNIX Makefiles generator as-is, and > create a new generator for GNU Make makefiles, which take full > advantage of GNU Make features. > > I don't have any idea what the appetite is in CMake-land for doing this > work. Right now, CMake also makes POSIX-compatible Makefiles (and we now generate a complete suppression of builtin rules since I discovered the last batch when debugging the graph and had to skip over the builtin CVS rules cluttering the logs). I'm not sure the prevalence of non-GNU make usage (perhaps all of those users have already migrated to Ninja?), but a new `GNU Makefiles` generator could be a thing. Ideally it would share a lot of its implementation with the existing Ninja generators. But again, it's not on our (Kitware's) road map at the moment either (though it is on the wishlist). Thanks, --Ben
