No, it cannot be done before CMake runs. It is a target like any other build target. It depends on other files or targets, and should be rerun if any of them change, and if it is rebuilt, there are other targets that depend on it and should be rebuilt, and so on.
The entire point of the build system is to concisely and accurate specify the build graph, and only rebuild what is necessary.
For the big majority of projects out there, it seems to me that it is perfectly fine to handle that manually, because projects are not that big. Do we agree on that?
Then of course, if you are a GAFAM with a huge monorepo, probably you need more advanced tools (they don't build their own tools for nothing). My feeling, though, is that most people complaining that CMake is crap actually work on small projects and can't be bothered to run protoc manually "because their philosophy is that it should all work with one button".
Now it's not a pain for me, and apparently you're on the side of those who find cmake painful. Not sure how unreasonable my way is :)
Let's also agree that every time you change a .cpp file you compile it, and every time you change a .h file you compile all the affected .cpp files.
No we don't agree on that. That is the entire point of having a build system.
Even proto files you gave as an example, may import other proto files. When you change one of the imported files you have to recompile all proto files that import that file, transitively.
I have, unfortunately, worked with build systems that do not capture all dependencies and systems that capture way too many dependencies. The former resulted in many hours spent debugging non existent problems, and the latter resulted in many hours spent waiting for compilation to finish.
Most people either work on programs small enough where having a conservative approximation of the dependency graph is "fine" and excessive rebuilds are OK, or structure the project with a little repetition between the between the build time dependencies and the code. Or, they just let the module system in their language of language choice figure it out. These are all completely OK. But it's very expensive to do conservative recompilation and very error prone to do manual repetition in larger projects, and gets very complex in multi-language setups.
Where does the presupposition of just some build tools, but not the others (if I am reading it correctly), resulting in excessive rebuilds come from?
Apart from building a dependencies DAG for a project, the next fundamental responsibility of a build tool is to watch for / detect changed graph nodes and apply relevant build rules (expressed as graph edges) to rebuild only the necessary parts of the graph by traversing the path upward (if the apex node is up at the top). Naturally, if it is a dependency at the very bottom of a very leafy graph that has changed and upstream nodes have a dependency on it, it will trigger a large (of a very large) rebuild of the entire dependency path. But that is true of all build systems, not just select ones, and that is what we need buid tools for.
The problem is that many people either can't or don't express dependency rules correctly or dependencies generate build byproducts that are to be dynamically injected into the DAG, and they not properly accounted for in the build rules, and that results in an incorrectly built DAG at the build system run time resulting in excessive rebuilds.
In the ideal world, dependency management is relegated to a module system of a programming language. But since long serving incumbents cough cough C, C++ and Java (not until Java 9 anyway) do not have a module system, people have created a kludge (make) and later multiple kludges (autoconf/automake, cmake, maven/ivy etc) to bandaid the problem.
However a generic build tool can't account for a permutation of all possible scenarios, all possible programming languages, linkers, pre- and post-processors and whatever else needs to be done to produce a build artefact and excel at it in every possible use case. Some do it better than the others in very specific scenarios but do poorly in some other or fringe use cases. Then there are personal preferences. Therefore, the discourse over which build system is better than the other(s) will never end.
The module system is tasked with figuring out dependencies (i.e. files comprising a module), which is true for all programming languages starting from Standard ML through to Rust and anything in between.
Now it really feels that most people complaining about CMake and using Meson instead are not in this situation. And if they were, it would seem more pragmatic to me to say "CMake is fine for small projects, but at some size we needed something else" than "Meson is obviously, definitely orders of magnitude better because it took me less time learning how to compile my 20-files project".
And this isn't something that ultra-massive billion line projects singularly suffer from. You can easily suffer from all these issues and more in small and big software projects alike. I've suffered from these problems plenty in FOSS and commercial software that were small and large. "Computer program that generates another computer program" is actually a pretty normal and fine way to build software. But when compilation is expensive, these things matter, and you can't keep making excuses about it.
A build system should accurately track dependencies and rebuild things. It's that simple. You can say that running two commands is OK or whatever, but don't try to imply that people who want that criteria -- accurate deps and rebuilding -- aren't trying to solve problems they actually have, and then ditch the example when it becomes inconvenient to you that they are.
I don't try to have cmake do too much, e.g. I don't run the generation (or python scripts or custom functions) from cmake, because it quickly becomes a mess and running the steps manually / in a script is usually easy.
I haven't seen a project yet where the cmake config is not a mess (because let's be honest, most CMakeLists out there are pretty bad) and where this is clearly a limitation of cmake (and not of meson).
Happy to get an actual open source example where that's clearly the case, though.
and to me, build procedures that aren't completely implemented in cmake but require to install who knows what shell / platform / language / etc... in addition are the bane of my life as they make everything harder and more brittle and less cross-platform.
I don't get it. Also shell scripts are pretty good at running a couple commands, if really that's important.
I actually have maintained FLOSS projects, and in my experience, throwing everything in one cmake project makes it harder for contributors to understand. Because they only ever use one command, and the day something fails they have to dig into it to understand what that custom command does. On the contrary, if they need to run the generation script first, then they know that there is a generation step involved, and suddenly the whole thing feels less like magic.
So that is my opinion: if your FLOSS project uses a big custom build system (because you customized your cmake to do all sorts of custom things), then probably I will not contribute. Because I have other things to do than learning your custom integration of standards steps I actually would already know.
What's better is what is easier to understand and maintain. If your whole project is made of one file and builds with one simple call to a compiler, IMO it's fine to just have this line in the README, no need for a custom_target in cmake.
Now the thing is that often, I see that people put a lot of importance into the "it should be easy to get started for beginners who don't know the tools", and everything is glued together using some custom functions. The goal being that a beginner can essentially run "make" (yep, I've seen Makefiles wrapping CMakeLists for that) and the project will build (sometimes it will even install dependencies on the system, that's crazy). But then if the beginner wants to understand what's happening under the hood, well... it's a big pile of custom glue.
Now if, instead of that, you document the tools you are using in your README, and use the tools in a standard way, then people who know the tools will get started quickly ("oh, it's protoc then cmake, I know exactly how to use that"). And yeah, beginners will need to learn how to run those tools. But that's fine IMO, because if you use the tools the standard way (without custom functions), then the beginner can just look for the official docs.
Standard is better than custom.
One thing that I find make CMake configurations completely unreadable is to try to have it do everything, coffee included. Where often a script could be run separately first (e.g. to generate those files), and then CMake could run normally afterwards.