BLD: Switch order of test for lapack_mkl and openblas_lapack

[rmalouf]

The check for openblas_lapack comes before the check for lapack_mkl in lapack_opt_info. This means that if openblas is installed on the system in one of the usual places, distutils will find it and use it even when mkl is specific in site.cfg. Reverse the order, so it checks mkl's lapack first and then openblas, lets it use mkl when requested. (See #6669)

[pv]

And what if you have mkl installed, but want to use openblas?

[rmalouf]

mkl wouldn't ordinarily be installed in /usr/lib, /usr/local/lib, or any of the other places the config code searches by default, so unless you specify the mkl install path in site.cfg, it's not going to find it. But yeah, I get your point, the fact that system_info tries to "do the right thing" when site.cfg is incomplete means that it'll probably do the wrong thing under some circumstances.

[rgommers]

unless you specify the mkl install path in site.cfg, it's not going to find it.

Fixing this more completely by always preferring things specified in site.cfg over anything in the standard locations would require fairly significant surgery in system_info. Which is painful and very hard to test. So I'm tempted to just agree that switching the order here is an improvement.

[charris]

system_info really needs a rewrite, but... painful.

[rgommers]

system_info really needs a rewrite, but... painful.

I will send a follow-up PR with some doc improvements and cleanups, because I'm tired of re-discovering the wheel.

A full rewrite isn't actually that hard and could make it much simpler, but the big problem is that there are no tests. And there's a lot of state, so just some unit tests really aren't enough.

[charris]

@rgommers So merge? I have no objections, but ISTM that one problem just gets exchanged for a less likely problem.

[rgommers]

Yes, in it goes. Thanks @rmalouf

[rgommers]

ISTM that one problem just gets exchanged for a less likely problem.

Indeed. The logic still needs a makeover at some point; it should respect site.cfg first of all and it doesn't do that now if multiple libs are installed in standard locations. But well, there's more that needs a makeover.....

[matthew-brett]

I just ran into this one - I wanted to build scipy against ATLAS on a machine with openblas installed. I believe I would have to edit systeminfo.py to do this - is that right?

[rgommers]

correct. really system_info should be reworked to always respect site.cfg first, but it doesn't do that. I have a partial cleanup of system_info somewhere, but didn't finish it (yet)