Proposal: Add atomic IORef operations to Data.IORef

[David Feuer]

The error handling situation for modification operations is somewhat
deceptive. Unlike MVar operations, where the modifying thread can be fully
responsible for exceptions, a thread can *try* to force the modified IORef
value, but it may *actually* be forced by another thread if that thread
gets there first. Someone using atomicModifyIORef2' should be very careful
not to produce bottom as the new IORef value. Similarly, they must not use
unsafePerformIO to produce the result if the IO in question must be
performed in a particular thread.

On Fri, Jan 15, 2021, 4:13 PM Alexey Kuleshevich <alexey at kuleshevi.ch>
wrote:

> Despite that this one is easy `atomicSwapIORef' ref !a = atomicSwapIORef
> ref a` I think it is important to include it. I also think
> `atomicWriteIORef'` should be added as well. Point is that most of the time
> in a concurrent setup it is the thread that does the IORef modification
> should be responsible for computing the value that is being written. This
> is important not only for performance but also for error handling.
>
> Alexey.
>
>
> ‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐
> On Saturday, January 16, 2021 12:01 AM, David Feuer <david.feuer at gmail.com>
> wrote:
>
> > Cool. Do you also want a strict atomicSwapIORef'?
> >
> > On Fri, Jan 15, 2021 at 3:38 PM Alexey Kuleshevich alexey at kuleshevi.ch
> wrote:
> >
> > > I totally agree with everything that you just said. Sorry, I should
> have said originally "-1 for adding JUST these lazy versions". As long as
> there are strict versions (whatever the names they might get, IMHO Lazy
> suffix is nice) and a warning on lazy versions about leaking memory, then I
> am 100% on board as well. Also +1 on not forcing the result.
> > > Alexey.
> > > ‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐
> > > On Friday, January 15, 2021 11:29 PM, David Feuer
> david.feuer at gmail.com wrote:
> > >
> > > > I disagree. I believe we should offer both atomicModifyIORef2 and an
> > > > atomicModifyIORef2'. The latter should force the new value in the
> > > > IORef but, unlike atomicModifyIORef', should not force the returned
> > > > value. Or if you and others prefer, we could offer a strict
> > > > atomicModifyIORef and a lazy atomicModifyIORef2Lazy. As a Haskell
> > > > programmer, I really don't want totally polymorphic values getting
> > > > forced behind my back. It's actually possible to write an even lazier
> > > > version that doesn't even force the function result pair, but I don't
> > > > think that's a very useful idea.
> > > > One other thing: the natural approach to atomicModifyIORef2' uses a
> > > > "half-strict pair" type
> > > > data HSPair a b = HSPair !a b
> > > > atomicModifyIORef2' :: IORef a -> (a -> HSPair a b) -> IO (a, HSPair
> a b)
> > > > I don't know to what extent users are willing to tolerate such an
> > > > extra datatype.
> > > > Side note: the peculiarly strict behavior of atomicModifyIORef' is a
> > > > result of the poor design of the old atomicModifyMutVar# primop.
> > > > On Fri, Jan 15, 2021 at 3:19 PM Alexey Kuleshevich
> alexey at kuleshevi.ch wrote:
> > > >
> > > > > Both of these functions are lazy with respect to the new value
> being written into the IORef, which is a horrible default for atomic
> operations. That is why atomicModifyIORef is a source of memory leaks and
> terrible performance, which is also why atomicModifyIORef' is almost always
> used instead.
> > > > > So +1 from me on adding strict versions of these functions that
> force new value to whnf, but -1 for adding these lazy versions as they are
> currently defined in GHC.IORef
> > > > > Alexey.
> > > > > ‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐
> > > > > On Friday, January 15, 2021 11:05 PM, David Feuer
> david.feuer at gmail.com wrote:
> > > > >
> > > > > > GHC.IORef currently exports
> > > > > > atomicModifyIORef2 :: IORef a -> (a -> (a,b)) -> IO (a, (a, b))
> > > > > > atomicSwapIORef :: IORef a -> a -> IO a
> > > > > > atomicModifyIORef2 is a lot like atomicModifyIORef, but it
> returns
> > > > > > both the old value in the IORef and the full result of applying
> the
> > > > > > user-supplied function to that. This is a pretty thin wrapper
> around
> > > > > > the newish atomicModifyMutVar2# primop, which has replaced the
> less
> > > > > > powerful atomicModifyMutVar#.
> > > > > > atomicSwapIORef atomically installs a user-supplied value in an
> IORef
> > > > > > and returns the old value. It is currently implemented using
> > > > > > atomicModifyIORef2, but it can and should be reimplemented using
> its
> > > > > > own, more efficient primop.
> > > > > > I propose to add both of these functions to Data.IORef.
> > > > > > David
> > > > > > Libraries mailing list
> > > > > > Libraries at haskell.org
> > > > > > http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/libraries/attachments/20210115/ea51b770/attachment.html>