Doesn't build when using netbsd curses on Linux

[ericonr]

fish 3.3.0

I wanted to build fish locally, but had forgotten I had https://github.com/sabotage-linux/netbsd-curses installed instead of ncurses. It failed to build because

/// Under curses, tputs expects an int (*func)(char) as its last parameter, but in ncurses, tputs
/// expects a int (*func)(int) as its last parameter. tputs_arg_t is defined to always be what tputs
/// expects. Hopefully.
#if !defined(NCURSES_VERSION) || defined(__NetBSD__)
typedef int tputs_arg_t;
#else
typedef char tputs_arg_t;
#endif

doesn't cover this specific use case. Ideally it would be possible to run a proper test in CMakeLists and use the result of that? Not urgent, it's easy to fix locally, but still a build bug.

[ericonr]

Fixing this is not enough for it to build; this setup also confuses FindPackage(Curses) and fish's pkg_config fallback isn't 100% correct.

[faho]

The FindPackage(Curses) bit would of course have to be improved in cmake (or whatever else provides that).

Tbh this isn't something I would want to work on as it's not a use case I have. Patches welcome!

[ericonr]

No worries, this is mostly a documentation issue. If I can come up with a clean fix, I will open a PR; in the meantime I just ripped some stuff out.

[ericonr]

The FindPackage(Curses) bit would of course have to be improved in cmake (or whatever else provides that).

Using pkgconfig first instead of as a fallback would be almost enough, it's CMake's weird logic in FindCurses that gets things wrong.

[mqudsi]

CMake's weird FindCurses implementation has caused us lots of grief. We don't run on any platforms where pkgconfig isn't a thing; personally, if you want to open a PR that fixes the netbsd curses problem and happens to change FindCurses to a fallback for pkgconfig rather than the other way around, I would happy to review it.

[alarixnia]

NetBSD doesn't provide pkgconfig files for libcurses. That appears to be something the Linux port added... It should be easy to detect ncurses because it defines NCURSES_VERSION_MAJOR etc.

[faho]

It should be easy to detect ncurses because it defines NCURSES_VERSION_MAJOR etc.

@alarixnia I'm assuming you mean we should replace the curses-related __NetBSD__ checks with NCURSES_VERSION_MAJOR checks?

I'm not sure if that would work, given we do

#if !defined(NCURSES_VERSION) || defined(__NetBSD__)

here. So unless my brain is refusing to logic again (not an impossibility, it does that sometimes), that should already be taken on NetBSD and on systems without ncurses.

Or is this specifically about NCURSES_VERSION_MAJOR?

[alarixnia]

@faho Hm, honestly I just scanned ncurses.h to look for something that would uniquely identify it as ncurses. But now I realize the problem is more complicated...

The problematic function is clearly a terminfo function, and I guess terminfo is usually provided by ncurses on non-NetBSD systems, but NetBSD libterminfo is still used on NetBSD if you're using ncurses (edit: actually, no, probably not?). Is the Linux port of NetBSD libcurses not using NetBSD libterminfo (that seems like a mistake), or is NCURSES_VERSION being defined somehow (also a mistake)?

[faho]

This appears ŧo have its own tputs and I can't find an NCURSES_VERSION define anywhere.

There is sabotage-linux/netbsd-curses#43, which explains that libterminfo is still separate, so it appears our curses-finding magic finds ncurses terminfo and NetBSD curses.

Which, uh, yeah, that's... not gonna work, is it?

Using pkg-config if possible and falling back to FindCurses would probably fix even that, tho.

[ericonr]

There is sabotage-linux/netbsd-curses#43, which explains that libterminfo is still separate, so it appears our curses-finding magic finds ncurses terminfo and NetBSD curses.

That isn't the case here. I have only netbsd-curses headers.

[faho]

That isn't the case here. I have only netbsd-curses headers.

Then where/why is NCURSES_VERSION defined?

[faho]

Okay, I don't see how this fails to build.

Either you have ncurses, and NCURSES_VERSION is defined, so it picks the datatype that's correct for ncurses,

or you don't have ncurses, and NCURSES_VERSION is not defined, so it picks the datatype that's appropriate for NetBSD-curses - including the linux port.

If NCURSES_VERSION is defined, then you probably also have ncurses installed and cmake picks it up partially (or you defined it yourself and that's the issue).

Either way, I would quite like to have at the very least the actual compiler error, and ideally some reproduction steps.