fix bugged IMPORTED_LOCATION

[jpfeuffer]

conda-forge/boost-feedstock#250

maybe distantly related: #32 #73

[jpfeuffer]

@pdimov FYI. Would be great if we can get this in sooner than later.
I tested it on my own CI: https://github.com/jpfeuffer/boost_install/actions/runs/24077063215/job/70228055696#step:4:585

[pdimov]

I think I see what you're trying to do, but as proposed, this doesn't seem to actually do anything, because stage-bindir always returns the same directory as state-libdir.

What's the intended use?

[jpfeuffer]

The staging is not the problem. I just made it more explicit that libdir will equal bindir in this case.

This fixes a problem for the install case.

[jpfeuffer]

See the outputs of my linked action run, then you can see what it changes.
It adds the bindir to the cmake config and correctly points the imported_locstion property to the location of the DLL (which is in bindir NOT libdir) during installation.

I am 99.9% sure that this really changes and fixes things.

[pdimov]

I'm not sure it does. It seems to me that this changes the cmake configs to reference bindir, but doesn't actually install the DLLs into bindir; they remain installed in libdir.

It probably works for you because you move the DLLs manually from libdir to bindir after the install, but that doesn't seem correct in general.

Either way, changing things to install the DLLs into bindir is probably the right thing to do (#32) but might be more disruptive of a change than what we need.

I think I'll add support for --dlldir instead (#73) which would be backward compatible.

[jpfeuffer]

Aah ok sorry! Yes the thread in #32 was a bit confusing, and yes we do copy it to bin during Conda packaging.
Ok then we also need a flag to control this.

[jpfeuffer]

@pdimov Done. What do you think?

[pdimov]

I applied something along those lines to the develop branch, 41163c1 and be91f92.

In addition to the --dlldir support, I also changed the default under non-Cygwin Windows to also be --bindir, instead of --libdir as it used to be. That's because the current code couldn't really handle distinct locations for those two <target-os>-es. A check for [ os.name ] doesn't work, because os.name gives the OS used to build b2, not the target-os for which the libraries are currently being built.

To make matters even more interesting, b2 actually supports building and installing for both Windows and Cygwin simutaneoulsy (e.g. b2 install toolset=msvc,gcc.)

Either way, this takes care of #32 as well as #73.

This isn't going into the next release though, sorry. 1.91 is currently in beta, closes for release in two days, and this change is too disruptive for the post-beta period. It will be in 1.92.

[jpfeuffer]

Thank you! Ah being the default sounds nice, I really think that bindir is the standard for non-Cygwin Windows.
Bummer that it will take another release but it is what it is. Maybe we can patch it in conda until then!