No need to try anything else. I'll see if I can fix the problem. The option
is a new one I added this week to try and reduce testing hard drive space.
Any progress on this?
I can obviously disable things like pre-compiled-headers and fix things that way, but that feels like the wrong fix to me (longer compile times etc).
Thanks, John.
The --remove-test-targets bjam option seems to have shrunk things down a bit. I'm not hitting the size limit anymore. It looks like we're down to 30GB/compiler from 60GB.
This solved my problem, but if we could have less churn on keeping unnecessary object files, tests would run quicker (I believe that they are usually disk IO limited, especially on windows). I'm not sure what if any other solutions there are to this.
For the time being I've disabled PCH support in Boost.Math when --remove-test-targets is on the command line. Hopefully we'll actually get some regression test results now at the expense of even slower build times. This is still the wrong fix though: we shouldn't break a documented feature (PCH) with an undocumented one. John.