Re: Haskell Digest, Vol 52, Issue 1
Simon Peyton-Jones said:
But since the read may block, it matters *when* you perform it. For example if you print "Hello" and then read the IVar, you'll block after printing; but if you read the IVar and then print, the print won't come out. If the operation was pure (no IO) then you'd have a lot less control over when it happened.
But this is true of any expression in a non-strict language. Why the special treatment of IVars? Nikhil
Rishiyur Nikhil wrote:
Simon Peyton-Jones said:
But since the read may block, it matters *when* you perform it. For example if you print "Hello" and then read the IVar, you'll block after printing; but if you read the IVar and then print, the print won't come out. If the operation was pure (no IO) then you'd have a lot less control over when it happened.
But this is true of any expression in a non-strict language.
Why the special treatment of IVars?
Consider this: do x <- newIVar let y = readIVar x writeIVar x 3 print y (I wrote the let to better illustrate the problem, of course you can inline y if you want). Now suppose the compiler decided to evaluate y before the writeIVar. What's to prevent it doing that? Nothing in the Haskell spec, only implementation convention. GHC will not transform the program in this way, because it can have some other undesirable effects. e.g. see http://hackage.haskell.org/trac/ghc/ticket/1592 A pure readIVar would be just like lazy I/O, with similar drawbacks. With readIVar, the effect that lets you observe the evaluation order is writeIVar; with hGetContents it is hClose. Conclusion: it's probably no worse than lazy I/O. Cheers, Simon
On Dec 4, 2007, at 8:53 AM, Simon Marlow wrote:
Rishiyur Nikhil wrote:
Simon Peyton-Jones said:
But since the read may block, it matters *when* you perform it. For example if you print "Hello" and then read the IVar, you'll block after printing; but if you read the IVar and then print, the print won't come out. If the operation was pure (no IO) then you'd have a lot less control over when it happened. But this is true of any expression in a non-strict language. Why the special treatment of IVars?
Consider this:
do x <- newIVar let y = readIVar x writeIVar x 3 print y
(I wrote the let to better illustrate the problem, of course you can inline y if you want). Now suppose the compiler decided to evaluate y before the writeIVar. What's to prevent it doing that? Nothing in the Haskell spec, only implementation convention.
Nope, semantics. If we have a cyclic dependency, we have to respect it---it's just like thunk evaluation order in that respect. If we decide to do the evaluation eagerly, we have to be prepared to bail out when the information we require is not available to us. The easiest expedient from GHC's perspective is to perform the evaluation lazily, just as we'd expect.
GHC will not transform the program in this way, because it can have some other undesirable effects. e.g. see
http://hackage.haskell.org/trac/ghc/ticket/1592
A pure readIVar would be just like lazy I/O, with similar drawbacks. With readIVar, the effect that lets you observe the evaluation order is writeIVar; with hGetContents it is hClose. Conclusion: it's probably no worse than lazy I/O.
Actually, it's considerably better. The only thing we can observe is that the implementation failed to respect the desired semantics (which state that computation blocks until the value is available). That causes problems with the strictness properties of I/O, but everything to do with the strictness properties of I/O is problematic (because we have to decide whether exit = _|_ or not as the above report notes). It doesn't, so far as I can see, cause problems with the strictness properties of pure computations except that a thread can block in the middle of a pure computation waiting for an IVar to fill. Possibly tricky from an RTS perspective, but there's no *semantic* problem. -Jan
Cheers, Simon
_______________________________________________ Haskell mailing list Haskell@haskell.org http://www.haskell.org/mailman/listinfo/haskell
On 12/4/07, Simon Marlow <simonmarhaskell@gmail.com> wrote:
do x <- newIVar let y = readIVar x writeIVar x 3 print y
(I wrote the let to better illustrate the problem, of course you can inline y if you want). Now suppose the compiler decided to evaluate y before the writeIVar. What's to prevent it doing that?
Look at the translation: newIVar >>= (\x -> let y = readIVar x in writeIVar x 3 >> print y) y can't be floated out because it depends on x. -- Taral <taralx@gmail.com> "Please let me know if there's any further trouble I can give you." -- Unknown
participants (4)
-
Jan-Willem Maessen -
Rishiyur Nikhil -
Simon Marlow -
Taral