[GHC] #13032: Redundant forcing of Given dictionaries

[GHC]

#13032: Redundant forcing of Given dictionaries
-------------------------------------+-------------------------------------
        Reporter:  simonpj           |                Owner:
            Type:  bug               |               Status:  new
        Priority:  normal            |            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):
       Wiki Page:                    |
-------------------------------------+-------------------------------------

Comment (by simonpj):

 Let's leave aside the deferred-type-error case for now.

 Other than that, I claim that '''every evidence variable should be non-
 bottom''', and so

 * Can be evaluated strictly, and/or
 * Its evaluation can be discarded if its payload is not required

 This is true even of recursively-defined dictionaries; see `Note
 [Recursive superclasses]` in `TcInstDcls`. (A recursively-defined
 dictionary may be infinite, so you can't evaluate it hyper-strictly, but
 it should not be bottom.)

 But dictionaries still ''can'' be defined recursively (unlike say unboxed
 values).

 Now, I believe all this is true of the programs generated by source
 Haskell, but nothing in Core checks this invariant; and in Core you could
 certainly declare a bottom dictionary `(let d = d in ...)`.   I don't know
 how to make it a statically-checkable invariant of Core, but I expect that
 it'd be possible to make a decent conservative approximation that was
 enough to capture the programs generated from source Haskell.

 There's even an old flag `-fdicts-strict` that I would like to get back
 to, which flirted with using call-by-value for evidence variables;  c.f.
 #2436

 That leaves deferred type errors.  But let's not allow the tail to wag the
 dog.

--
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/13032#comment:2>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler