Proposal: Add newUniqueSTM to Data.Unique

[Bas van Dijk]

On 19 January 2011 20:50, Edward Z. Yang <ezyang at mit.edu> wrote:
> +1 for reverting to using an IORef.

Here's a hastily constructed benchmark (based on the one from #3838)
that compares the two implementations:

{-# LANGUAGE BangPatterns #-}

module Main where

import Control.Concurrent.STM.TVar
import Control.Monad.STM
import Control.Concurrent.MVar
import Control.Concurrent
import Control.Monad
import Foreign
import Data.IORef

newtype Unique = Unique Integer deriving (Eq, Ord)


-- Current implementation of Unique using TVars

uniqSource :: TVar Integer
uniqSource = unsafePerformIO (newTVarIO 0)
{-# NOINLINE uniqSource #-}

newUnique :: IO Unique
newUnique = atomically $ do
  val <- readTVar uniqSource
  let next = val+1
  writeTVar uniqSource $! next
  return (Unique next)


-- New implementation of Unique using IORefs

uniqSource' :: IORef Integer
uniqSource' = unsafePerformIO (newIORef 0)
{-# NOINLINE uniqSource' #-}

newUnique' :: IO Unique
newUnique' = do !next <- atomicModifyIORef uniqSource' $ \val ->
                           let !next = val+1 in (next, next)
                return (Unique next)


-- Benchmark

numThreads :: Integer
numThreads = 1000000

bench doNewUnique = do
  done <- newEmptyMVar
  let loop :: Integer -> IO ()
      loop i = do
          when (i < numThreads) $ do
              forkIO $ do threadDelay 1000
                          Unique u <- doNewUnique
                          when (u == numThreads) $ putMVar done ()
              loop (i + 1)
  loop 0
  takeMVar done

main = bench newUnique
main = bench newUnique'

(Build with -O2)

$ time ./benchUniqueTVar

real	0m16.575s
user	0m16.080s
sys	0m0.430s

$ time ./benchUniqueIORef

real	0m16.378s
user	0m15.840s
sys	0m0.480s

So using an IORef is a tiny bit faster.

Does newUniqueSTM give you a performance advantage in a STM
transaction? Or is it just a convenience that you don't need to use
unsafeIOToSTM?

Because of the slightly simpler implementation and slightly better
performance I'm for reverting to IORefs.

Regards,

Bas