Simplify Visual C configuration for -m64/-m32mscoff builds

[CyberShadow]

Ping @rainers

[CyberShadow]

Crap, why is this failing? What make variant does the autotester use?

[CyberShadow]

Why is this failing?!? I can't reproduce this with any make variant I have on this computer. @braddr?

[CyberShadow]

Searching for the error message shows that it is emitted by DigitalMars make, but obviously it doesn't happen here. Does the autotester patch the make files in some way?

[CyberShadow]

Aha: https://github.com/braddr/d-tester/blob/master/client/src/diff-druntime-win64.diff

[rainers]

Sorry for being late at commenting, I could have warned about the makefile being patched.

I'm not sure making the sub-directory a dependent option is the best approach. How about specifying the bin path as patched by the build-server? BTW: link/lib are not used, so specifying CC on the command line is good enough. That is what I am doing.

Regarding errno.c: I'd rather like to see the object files in different directories (that will need more patches, though). I don't think building works across platforms without cleaning up yet. At least, you'll also have to patch the zlib makefile in phobos aswell to make that work.

[braddr]

The two changes (druntime and phobos) should be essentially the same rather than subtly different in how VCBINDIR is defined. I'll deal with cleaning up the auto-tester end of things. This ought to allow me to remove one of the last reasons there's an external diff applied at all. Something I should have done ages ago.

[CyberShadow]

I'm not sure making the sub-directory a dependent option is the best approach. How about specifying the bin path as patched by the build-server?

That's another way to do it. I was thinking that splitting up the path parts would allow e.g. compiling for 32 and 64 bits in the same makefile in the future, or easily choosing a VS version (e.g. VC=11).

The two changes (druntime and phobos) should be essentially the same rather than subtly different in how VCBINDIR is defined.

Yeah, I was going back and forth trying different things while trying to figure out the autotester failure. I considered that the leading backslash was triggering some special syntax.

This ought to allow me to remove one of the last reasons there's an external diff applied at all.

One effect of the autotester patch is that it removes the definition from the makefiles completely, causing make to read the variables from the environment. Since we still want a default value for backwards compatibility, without patching the makefiles the bin path will now need to be passed as a parameter to make. Is that OK?

[braddr]

I really don't want to have to have make invoked differently based on OS. Adding VCBINDIR to all make invocations wouldn't be the worst thing in the world, but it'd be really dumb to include for non-win64 environments. The best we can do with the digitalmars make, given its macro rule logic and inability to do this sort of default setting, is to also set MAKEFLAGS to "VDBINDIR=path" in the environment. I've never tried that, but it ought to work.

[CyberShadow]

BTW: link/lib are not used, so specifying CC on the command line is good enough.

lib appears to be used when building zlib, no? Or it doesn't matter which bitness the librarian it is?

I really don't want to have to have make invoked differently based on OS.

I noticed the build scripts are already customizing Make's command line (EXTRA_ARGS), so I guess one option would be something like:

if [ -n "$VCBINDIR" ] then; EXTRA_ARGS="VCBINDIR=$VCBINDIR"; fi

But MAKEFLAGS sounds better.

---

Anyway, updated. Now there is VC (to easily switch VC versions if desired), VCDIR (backwards compatibility), VCBIN_SUBDIR (for easy m32mscoff or cross-compile builds), and VCBIN_DIR a la autotester. Note that VCDIR also dictates the C include directory.

[braddr]

Move at least the 'bin' part ouf of VCBIN_DIR into VCBIN_SUBDIR. Maybe the \ as well. If you're going to make it overridable, don't go just half way.

[CyberShadow]

Why? I don't think it's even possible to install VS in a way that this would ever be useful, and for those cases you can just override VCBIN_DIR instead.. It would just make the shortest command line for m32mscoff or cross-compilation longer than necessary.

[rainers]

Why? I don't think it's even possible to install VS in a way that this would ever be useful

Actually I have an extracted a few files from VS2010 to get a small reference version that works building and debugging phobos (65 MB), and the folders are actually named differently ;-)
I think using VCBIN_SUBDIR=bin is even a bit more obvious than specifying an empty argument.

[CyberShadow]

OK, updated.

[rainers]

LGTM, though I'd still recommend to put "/.obj" into .gitignore instead of "/errno_c.obj" to exclude unittest.obj. Other refactorings can be added later.

@braddr I guess this should now be good enough to run the test builds without patching the makefile. If you can adjust the tester, please pull.

[rainers]

ping @braddr

[quickfur]

ping

[MartinNowak]

You should use MODEL in the make file to set the VCBIN_SUBDIR.
Not sure if it can be done though, maybe double expansion works.
VCBIN_SUBDIR_64=bin\x86_64
VCBIN_SUBDIR=$(VCBIN_SUBDIR_$(MODEL))

[rainers]

Unfortunately, DM make cannot do that.

[MartinNowak]

I know little to nothing about VC integration. What problem does this pull solve?

[CyberShadow]

I know little to nothing about VC integration. What problem does this pull solve?

Makes it easier to build -m32mscoff Phobos/Druntime, or cross-compile 32-bit to 64-bit.

[MartinNowak]

Waiting for @braddr to do something with the auto-tester. We really need more people that can maintain the auto-tester, because Brad is a busy person and we too often get stuck.

[wilzbach]

Waiting for @braddr to do something with the auto-tester. We really need more people that can maintain the auto-tester, because Brad is a busy person and we too often get stuck.

At mir we combined AppVeyor and Travis to test for Win32,Win64,Linux32,Linux64 and OS X 64 (OS X 32 has been deprecated for a couple of years).
At ldc they started to experiment with buildbot. Jenkins also supports distributed building...

ping @braddr

[wilzbach]

ping @braddr - do you have time to update the auto-tester?

[Geod24]

Druntime have been merged into DMD. Please re-submit your PR to dlang/dmd repository.