While benchmarking the performance of GHC's SMP parallelism, I ran into an issue where (rarely) a test would fail with <<loop>>. I tried hard to reduce this to a single file, which I am attaching to this report.

To reproduce, compile the program with --make -threaded -feager-blackholing -rtsopts -O1 and run multiple times with -Nm with m>=2. In my case, using N8 (on a machine with 8 cores), I once got the error after 35 tries, but sometimes it takes thousands of attempts. I'm also attaching a bash script to compile and run the test repeatedly.

Tested with current HEAD, built on a Intel Xeon E5320, "perf" settings without dynamic linking. But I think I also encountered this behavior in GHC 7.0.3.