[GHC] #12368: Demand Analyzer: Cunnig plan not adhered to with aborting fixpoint interation

[GHC]

#12368: Demand Analyzer: Cunnig plan not adhered to with aborting fixpoint
interation
-------------------------------------+-------------------------------------
        Reporter:  nomeata           |                Owner:
            Type:  bug               |               Status:  patch
        Priority:  low               |            Milestone:
       Component:  Compiler          |              Version:  8.0.1
      Resolution:                    |             Keywords:
Operating System:  Unknown/Multiple  |         Architecture:
                                     |  Unknown/Multiple
 Type of failure:  None/Unknown      |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:                    |  Differential Rev(s):  Phab:D2392
       Wiki Page:                    |
-------------------------------------+-------------------------------------

Comment (by nomeata):

 > So we return an un-decorated binding orig_pairs

 It is not necessarily un-decorated, but may have information from a
 previous pass of the Demand Analyzer. I was under the impression that we
 generally want to make sure that running the demand analyzer updates _all_
 strictness signatures and demand information, and does not leave any
 around from a previous iteration.

 I am never fully confident that I can overlook all the assumptions in the
 code and the consequences, that’s why I am much more confident with simple
 invariants such as “all code has been processed and annotated by the
 demand analyzer at least once.”

 > We should just return the un-extended environment! (If a variable isn't
 in the environment it's treated as having topSig.) Simple.

 Is that safe? `topSig` (actually `nopSig`, since there is no top
 strictness signature) has an empty environment, so if the RHS of the
 definition makes use of a free variable, `nopSig` effectively has an
 absent demand on that.

 I guess this is the reason for the `lazy_fv` fuss. I think in some other
 discussion you wondered whether that is really required. By having the
 analysis set to `topSig` it definitely has to.

 > Conclusion: if the fixpoint limit is reached:

 >  * return an environment that simply does not mention the new binders

 Agreed, if `lazy_fv` works as advertised.

 >  * set all the binders to no-strictness-at-all

 What about demand and strictness signatures attached to binders somewhere
 nested in the RHS of one of the equation. Should we zap them as well?

 > Finally, the loop function in dmdFix would be much better if it took a
 SigEnv rather than an AnalEnv, wouldn't it?

 Or maybe even just a list of `StrSig`s, to make it clear that only the
 strictness signatures of the recursive binders vary.


 All in all I wonder if avoiding an extra iteration (after we already did
 10) in the corner case is worth the extra complication of having to think
 about how to properly abort the analysis from in an invalid state. Simply
 jumping to a definitely sound state, and being able to guarantee that a
 pass of the demand analyzer processes all code appears to be simpler and
 more reliable to me.

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12368#comment:9>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler