We see currently two miscompiles in LNT with -polly-parallel: http://lab.llvm.org:8011/builders/perf-x86_64-penryn-O3-polly-parallel-fast/builds/15119
I’m not able to reproduce them on my target machine. I’ve run the LNT with the following options: ~/mysandbox/bin/lnt runtest nt --sandbox SANDBOX --cc "/home/roman/Documents/polly/build/llvm/bin/clang" --cxx "/home/roman/Documents/polly/build/llvm/bin/clang++" --cflags=" -lgomp -Xclang -load -Xclang /tmp_home/compiled/llvm/lib/LLVMPolly.so -O3 -mllvm -polly -mllvm -polly-parallel" --test-suite ~/llvm-test-suite Have I missed something?
I had issues reproducing them as well. So no, this does not seem to be an easy/simple to debug issue.
This was likely not a miscompile, but rather too many threads running in parallel as we run LNT with -j16 and then allowed each program to create itself threads without limitation (which resulted in one thread per core and process). This issue was resolved by hardcoding the number of parallel threads per process with OMP_NUM_THREADS=2.