Re: Library dependency files

[Erik Rull]

Todd Showalter wrote:
>     Easier solution: rerun makedepend on all files every build, and
> eat the half second it costs,  Seriously, I've timed it, and it's
> ~500ms on our project on a decent machine and getting faster every
> time the machines get better  My i860 takes an average of ~110ms
> (real) for a nothing-to-do build.  Mind you, the way we've laid our
> project out probably helps us on this; every file in the project
> depends on the whole header set (which is pretty much inescapable in
> console game development), so with the exception of private header
> files the whole dependency tree is parsed on the first file and
> cached.  Subsequent files just get copies of the same tree.

Very interesting - how does your make rule look like for this case?

>     Doing a complete dependency sweep every build means there are no
> special hacks in the makefile, no chance we missed a corner case or a
> variable, and the execution time is nearly equivalent, even on
> nothing-to-be-done builds.  Saving some fraction of a half second in
> exchange for making the build process more complex doesn't strike me
> as a win.

That's true, you're context free.

>     Now, if your dependencies were taking on the order of tens of
> seconds or more to generate, that would be another question entirely.
> And if you aren't using environment variables to drive your build
> process, the paper we're discussing works fine.  The idea the paper
> describes is good, and I don't mean to say it isn't.  My only point
> here is that use of environment variables to drive the build breaks
> assumptions the paper makes (ie: that make will detect all changes
> that require dependency recalculation), and force you to add
> additional complexity to the system.

My "doing nothing" takes currently ~ 4 seconds - how large is your project?

>     I don't see this as a fault in make, or in the paper.  It's just a
> consequence of the particular way our build is set up.  We've got
> multiple platforms and multiple configurations supported in a single
> source tree, and we want to be able to access those without having to
> run a "configure" step.  Our build supports multiple game consoles and
> operating systems, enables or disables features on those platforms
> (networking, movie playback, debug printing, storage access
> mechanisms, graphics APIs...), and allows variant compilation for
> different language and legal regions.  To make it all work without a
> "configure" step, we have environment variables that control various
> flags.

I'm compiling also for different cpu targets - a short taste of your 
makefile would be interesting.

>     We could probably can them all up in config files and add them to
> the dependency graph; that would get rid of the "dependency change
> sneaks past make" problem, but once again it would add complexity
> (more files to maintain, a more complex build process, would need some
> way of telling make which file to use...) and buy us nearly nothing in
> return over the system we already have.
>
>                                                      Todd.

Best regards,

Erik