This ticket is a continuation of #11978 (closed) and #12009 (closed). After fixing a couple of issues in those two tickets I found that the profiling run time is not thread safe.

Have a trivial test program (written as one of the tests for #11978 (closed)):

import Control.Concurrent
import Control.Concurrent.MVar
import Control.Exception
import Control.Monad

main :: IO ()
main = do
    putStrLn "Start ..."
    mvar <- newMVar (0 :: Int)

    let count = 50

    forM_ [ 1 .. count ] $ const $ forkIO $ do
            threadDelay 100
            i <- takeMVar mvar
            putMVar mvar $! i + 1

    threadDelay 1000000
    end <- takeMVar mvar
    putStrLn $ "Final result " ++ show end
    assert (end == count) $ return ()

Compiling that with a compiler that has bug fixes arising from #11978 (closed) and #12009 (closed) as:

inplace/bin/ghc-stage2 testsuite/tests/profiling/should_run/T11978b.hs \
    -fforce-recomp -rtsopts -fno-warn-tabs -O -prof -static -auto-all \
    -threaded -debug -o T11978b

and run as:

./T11978b +RTS -p -hb -N10 

crashes in a number of different ways. I've seen at least 3 different assertion failures and numerous segfaults (in different stg_ap_* functions).

Replace -hb with other profiling options like -hr etc do not seem to crash.

Looking at code, one example of lack of thread safetly is the function LDV_recordDead which mutates global variable censuses which does not have any locking around it. Only figured this out because the following assert (in LDV_recordDead) was being triggered occasionally.

    ASSERT(censuses[t].void_total < censuses[t].not_used);