LLVM-6 build is much faster than LLVM-14 build about genmc HOT 5 CLOSED

Thanks for the report.

LLVM-14 is not supported yet, so I don't know whether something breaks or there's any regression. In principle, as long as the GenMC version is the same, the running time should be comparable.

I'll try to add support for LLVM-14 in the next version (along with some new features), though this might take some time as I'll be out of office for a while. In the meantime, if you upload the .ll files produced by GenMC (via -transform-output), I might be able to guess why this happens and/or push some workaround.

Michalis

from genmc.

Thanks for reply!

I attached an archive with two .ll files: genmc-llvm.zip.

Pavel

from genmc.

Sorry for the delay.

I assume that GenMC did not actually terminate for LLVM 14?

In any case, the code that clang produces for the two versions is different, hence the different behavior. In LLVM 14, GenMC fails to automatically bound some loops and gets stuck into some infinite loop.

I will try to fix this soon, but in the meantime using -unroll=N should make the tool terminate.

Michalis

from genmc.

Thank you very much!
The option -unroll=100 helps us.

Pavel

from genmc.

It turns out that there was a bug in GenMC that is not fixed. The behavior should now be consistent across LLVM versions.

BTW, I'd recommend making BackOff in fmutex.cpp a no-op when MC_ON; having it enabled only slows down verification.

from genmc.