Using unsafePerformIO safely?

[atravers]

@nomeata

The question “when do we know that unsafePerformIO is safe” can be specialized to “why do we believe that runST is safe?” (which is one possible use of unsafePerformIO).

Correct.

I am not sure how easy it is to extend to other uses of unsafePerformIO […] or to safety in the presence of concurrency.

The problem is observable effects:

• ST actions like readSTRef and writeSTRef only manipulate the contents of the program’s memory - usually only made visible with a debugger.

• however, IO actions like getChar and putChar work with I/O devices which can potentially manipulate anything, including the program itself.

As for safety with concurrency, everyone’s favourite “type-smith” has already shown how unsafeInterleaveST (and consequently unsafeInterleaveIO) breaks regular Haskell semantics, even though their results are monadic:

unsafeInterleaveST :: ST s a -> ST s a
unsafeInterleaveIO :: IO a -> IO a

as opposed to:

unsafePerformST :: ST s a -> a
unsafePerformIO :: IO a -> a

If only there was a denotative way to write system software…