opt out of accursedUnutterablePerformIO

[Ben Gamari]

Ben Franksen <ben.franksen at online.de> writes:

> Am 04.03.2018 um 15:58 schrieb Ben Gamari:
>> I'm afraid it's not possible to provide the interfaces exposed by
>> bytestring without some form of unsafety. Lazy IO alone requires
>> unsafeInterleaveIO and the bytestring indexing operations require at
>> very least unsafePerformIO since GHC treats access to foreign memory
>> as an effect.
>
> I am well aware of that.
>
>> accursedUnutterablePerformIO is an optimized form of unsafePerformIO
>> which likely won't cause any issues that wouldn't otherwise manifest
>> with plain unsafePerformIO. Consequently I am not sure it's worth
>> providing a means to disable its usage.
>
> I envy your confidence. The documentation mentions two bytestring
> commits that fix bugs by reverting accursedUnutterablePerformIO to
> unsafePerformIO. I guess this is plain evidence that there can in fact
> be "issues that wouldn't otherwise manifest with plain unsafePerformIO".
>
Ahh yes, I had forgotten that accursedUnutterablePerformIO is
implemented directly in terms of realWorld#. This implementation is
terribly bug-prone since floating can result in inappropriate sharing,
as seen in the tickets referred to in the documentation.

However, GHC 8.0 and later offer a much safer primitive, runRW#, which
can be used to implement things like accursedUnutterablePerformIO
without fear of over-zealous simplification. I believe the issues
pointed out in documentation could not have happened with an
implementation built on runRW#. Moreover, I highly doubt that the
switching to runRW# would incur any measurable performance penalty.


> Anyway. Would you perchance have any idea what could possibly make a
> program work fine with the first version of linesPS and crash with the
> second one? I find this pretty scary and would like to understand it.
>
Indeed that is quite scary. The cause is not at all obvious. Do you have
an isolated reproducer that you could share?

Cheers,

- Ben
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 487 bytes
Desc: not available
URL: <http://mail.haskell.org/pipermail/libraries/attachments/20180304/3e5b2f88/attachment.sig>