[Haskell-cafe] TVars & throw

[Chris Kuklewicz]

Thomas Conway wrote:
> On 3/8/07, Chris Kuklewicz <haskell at list.mightyreason.com> wrote:
>> http://haskell.org/haskellwiki/New_monads/MonadAdvSTM
> 
> Ah. Yes, I've independently discovered an equivalent technique. I'm
> building external btree/hash table code, and need to handle the case
> where you go to use a node that hasn't been paged in. My solution was
> simply to throw an exception, catch it, do the IO, then rerun the
> transaction.

Which sounds *almost* equivalent.  My onRetry is different in that it lets one
create a sequence of actions that are contingent on retrying the whole
transaction.  If the transaction commits then these actions are not performed;
the actions queued by onCommit will be performed instead.

What happens in your throw/catch case if I have
  stm1 = do some_stm_code_that_throws_your_exception
  stm2 = return Foo
and I run "atomically (stm1 `orElse` stm2)" ?
Answer: The exception will prevent running stm2.

So I think a mixture of the two would be interesting.  Perhaps I will add a
"forceRetry" command to the MonadAdvSTM code that uses such an exception to
prevent any orElse code from running.

> Um, is
>    unsafeIOToSTM $ atomically trans
> going to run you into problems?

YES!

My code does not use "unsafeIOToSTM $ atomically" nor does it use
"unsafePerformIO $ atomically".  Both of those will kill the runtime system and
thus the whole program.  That fatal error is why there is the "*IO" functions
like "newTVarIO" were added.

My code uses "unsafeIOToSTM $ (IO code using MVar)" which is safe.

-- 
Chris