Template Haskell determinism

[Michael Sloan]

+1 to solving this.  Not sure about the approach, but assuming the
following concerns are addressed, I'm (+1) on it too:

This solution is clever!  However, I think there is some difficulty to
determining this ordering key.  Namely, what happens when I construct the
(Set Name) using results from multiple reifies?

One solution is to have the ordering key be a consecutive supply that's
initialized on a per-module basis.  There is still an issue there, though,
which is that you might store one of these names in a global IORef that's
used by a later TH splice.  Or, similarly, serialize the names to a file
and later load them.  At least in those cases you need to use 'runIO' to
break determinism.

If names get different ordering keys when reified from different modules
(seems like they'd have to, particularly given ghc's "-j"), then we end up
with an unpleasant circumstance where these do not compare as equal.  How
about having the Eq instance ignore the ordering key?  I think that mostly
resolves this concern.  This implies that the Ord instance should also
yield EQ and ignore the ordering key, when the unique key matches.

One issue with this is that switching the order of reify could unexpectedly
vary the behavior.

Does the map in TcGblEnv imply that a reify from a later module will get
the same ordering key?  So does this mean that the keys used in a given
reify depend on which things have already been reified?  In that case, then
this is also an issue with your solution.  Now, it's not a big problem at
all, just surprising to the user.


If the internal API for Name does change, may as well address
https://ghc.haskell.org/trac/ghc/ticket/10311 too.  I agree with SPJ's
suggested solution of having both the traditional package identifier and
package keys in 'Name'.

-Michael

On Tue, May 31, 2016 at 6:54 AM, Bartosz Nitka <niteria at gmail.com> wrote:

> Template Haskell with its ability to do arbitrary IO is non-deterministic
> by
> design. You could for example embed the current date in a file. There is
> however one kind of non-deterministic behavior that you can trigger
> accidentally. It has to do with how Names are reified. If you take a look
> at
> the definition of reifyName you can see that it puts the assigned Unique
> in a
> NameU:
>
>   reifyName :: NamedThing n => n -> TH.Name
>   reifyName thing
>     | isExternalName name = mk_varg pkg_str mod_str occ_str
>     | otherwise           = TH.mkNameU occ_str (getKey (getUnique name))
>     ...
> NameFlavour which NameU is a constructor of has a default Ord instance,
> meaning
> that it ends up comparing the Uniques. The relative ordering of Uniques is
> not
> guaranteed to be stable across recompilations [1], so this can lead to
> ABI-incompatible binaries.
>
> This isn't an abstract problem and it actually happens in practice. The
> microlens package keeps Names in a Set and later turns that set into a
> list.
> The results have different orders of TyVars resulting in different ABI
> hashes
> and can potentially be optimized differently.
>
> I believe it's worth to handle this case in a deterministic way and I have
> a
> solution in mind. The idea is to extend NameU (and potentially NameL) with
> an
> ordering key. To be more concrete:
>
> -   | NameU !Int
> +   | NameU !Int !Int
>
> This way the Ord instance can use a stable key and the problem reduces to
> ensuring the keys are stable. To generate stable keys we can use the fact
> that
> reify traverses the expressions in the same order every time and
> sequentially
> allocate new keys based on traversal order. The way I have it implemented
> now
> is to add a new field in TcGblEnv which maps Uniques to allocated keys:
>
> +        tcg_th_names :: TcRef (UniqFM Int, Int),
>
> Then the reifyName and qNewName do the necessary bookkeeping and translate
> the
> Uniques on the fly.
>
> This is a breaking change and it doesn't fix the problem that NameFlavour
> is
> not abstract and leaks the Uniques. It would break at least:
>
> - singletons
> - th-lift
> - haskell-src-meta
> - shakespeare
> - distributed-closure
>
> I'd like to get feedback if this is an acceptable solution and if the
> problem
> is worth solving.
>
> Cheers,
> Bartosz
>
> [1]
> https://ghc.haskell.org/trac/ghc/wiki/DeterministicBuilds#NondeterministicUniques
>
> _______________________________________________
> ghc-devs mailing list
> ghc-devs at haskell.org
> http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-devs/attachments/20160601/a3776778/attachment-0001.html>