We are the windows package maintainers of PostGIS. We build these packages using MingW chain of tools. For other packages we fancy that do not come packaged with the windows VC++ builds, we also build these under mingw. We've described some of these already in File FDW family. For windows 32 builds we build
with the commonly known MSys/Mingw32 chain (but an older version 1.0.11)
because of issues we have building with the newer msys/mingw 32.
For windows 64-bit installs, we build with the mingw-w64 chain and in fact
we like the ming-w64 chain so much that we plan to migrate our Mingw32 to mingw64. We have PostgreSQL 9.2 and PostgreSQL 9.3 successfully installing under the mingw-w64 for windows 32 just fine (older PostgreSQL we experience a winsock.h something or other error which we are working on troubleshooting. For 64-bit we use ming-w64 for building extensions for PostgreSQL 9.0-9.2 and soon 9.3 with some minor issues. Some people have asked us, why put yourself thru this torture? Why not just build on MS VC++ for everything? Originally we had started on mingw because PostGIS needed a Unix like environment to compile and run thru the battery of tests. This is still the case, though PostGIS is planning a CMake move with help from Mateusz Lostkot which hopefuly will provide a better cross-platform experience and allows us to integrate pgRouting (which already is on CMake). Paul Ramsey rewrote many of the regression test scripts to be completely Perl based and not require sh. The other reality is we just prefer mingw and can't really stomach having to work with VC++. I'll describe why and why not build with mingw and deploy on VC++ compiled PostgreSQL.
MingW (at least for us is just more portable), perhaps it's because we are more familar with it. I have my workstation with a ming32/ming64 folder setup as we've described here. When I get a new workstation, I just copy the folders to the new pc and I'm ready to go. I'm builidng a buildbot for windows to auto build the PostGIS windows packages using jenkins and to get that compiling -- again I copied the ming32 and ming64 folders to the server and again it just works. The only trick launching from a batch script needed by the build bot, is I had to create a script that looks like this:
%WD%sh --login -i -c '"C:\path\to\myscript.sh"'
Depending on whether we build for 32 or 64 I just toggle 32/64. The login seemed to be required to get libtools to work properly.
- VC++ we just don't care for. That is because we are web developers and only use VS for ASP.NET development. Having to then add in the VC++ is just another hassle.
- Each version of PostgreSQL (at least the EDB ones that most people use), are compiled with different versions of VC++ each requiring their own VC++ runtime distributed. So for example 9.1 was built with VS 2008. 9.2 will be built with VS 2010. PostGIS has a couple of dependency libraries which we build once for each PostGIS release and share across all the PostgreSQL versions and there is only one or two extra helper dlls to package for these (the stdc++ ones) which get copied in the PostgreSQL bin. If we were to build under VS we would have two choices: 1) build using the VC++ EDB builds with, yet another clutter of our workspace or 2) package whatever runtime libraries we are using which would make our setup exe much more complicated.
- Many of the extensions people build for PostgreSQL, PostGIS included, assume a pgxs capable install. VC builds don't support pgxs. So if you are trying to compile an extension built by a Unix/Mac User on windows with VC++, you are screwed or have to rollup your sleeves to figure out how to hack their build script to work for you. Our VC++ skills are nil. The story is even worse for 64-bit PostgreSQL. Under mingw 32/64, it generally just works even for building 64-bit libraries. I like being special but I don't like being that special. I want to be able to compile things much the same way as anyone else working on the projects I work with. I don't want to ask other developers to make special concessions to get it to work for me and continually baby sit these changes. As it stands windows VC++ is probably the only environment that requires a non-Unix like environment to work. Sure others have idiosyncracies, but they don't require you to rearchitect your whole build system to work for them.
- I find it a lot easier to get a Unix developer to stab themselves with a pitch fork than to get a Windows developer to lift a finger for anything that is not a pure windows project. I'll explain in the why not section. I think part of the reason for that is that it's more common to find Unix developers that started life as windows developers. So they are familiar with both environments and the underlying codebase. The other reason is a lot of large corporations still run on windows and it makes good business sense to support people who have money assuming you can get them to shoot some of that your way.
- Sadly, some extensions built under MingW will not work under VC++. One example is PL/R. When PostgreSQL 8.3 came out and the packaged PostgreSQL windows installers switched to VC++ builds from MingW, it left many PL/R windows users very unhappy. For some reason the PL/R windows users were compiling under MingW wouldn't load on these new VC++ creatures. After much crying from windows folks and several of us trying to figure out what was wrong, we all concluded it was the VC++ switch. Joe Conway stepped up to the plate and proverbially stabbed himself with a pitchfork to revise his build scripts to work on VC++. This meant for him setting up a windows VM VC++ Express, running into unchartered scary territory I imagine for him.
Andrew Dunstan is another fine example. Andrew, BTW your changes to FDW Fixed Length still don't work for me. I'm going to verify to make sure it's not something wrong with my environment before I bug you.
- Mixing build systems is really not supported and some are shocked it works. If you ask some PostgreSQL key contributors, their jaws sometimes drop when you tell them what you are doing. Enough said.
- Memory issues even for PostGIS the 64-bit mingW compiled works fine against a mingW PostgreSQL but crashes on the input KML/GM tests when run from sh. The odd thing is it works fine when you manually run the test from PostgreSQL VC++. So it's hard to tell where this issue is coming from and if it would exist in a pure VC++ environment. I suspect it would.
General Tips for using MingW for compiling
We build on windows, so all the cross-compiling features of MingW are a bit lost on us. I don't have much of a clue how it works on Unix systems, though Andrew described a bit how to get going with it on Unix in Cross-compiling PostgreSQL for windows
PATH variable on your scripts. Most issues we've had with compiling not working the same on one pc from another is all the path junk brought in from windows from other installs you've got going. Suddenly your code has dependencies on dlls distributed by Tortoise etc. and you're hunting around what stray cat came thru.
For example the header of our sh scripts looks something like this:
This is especially true when building for 64-bit.
Don't install non-standard dependencies in the default location. Just add them to your path variable as needed. Case in point, each version of PostGIS uses a different version of GEOS and sometimes we have to test many. We also build for many versions of PostgreSQL. Keeping these separate helps immensely.
- Use dependency walker to check what dlls you are missing. Most often, a mingw library doesn't load on VC++ build because you are missing libstdc++ or some other c helper library. Often times you need others (PostGIS has several like libgdal, geos, proj, libxml-xml2 etc.). Dependency Walker will help you sniff out these issues.