[GHC] #14626: No need to enter a scrutinised value

[GHC]

#14626: No need to enter a scrutinised value
-------------------------------------+-------------------------------------
        Reporter:  heisenbug         |                Owner:  heisenbug
            Type:  bug               |               Status:  new
        Priority:  normal            |            Milestone:
       Component:  Compiler          |              Version:  8.2.2
      Resolution:                    |             Keywords:  performance
Operating System:  Unknown/Multiple  |         Architecture:
                                     |  Unknown/Multiple
 Type of failure:  None/Unknown      |            Test Case:
      Blocked By:                    |             Blocking:
 Related Tickets:  #13861            |  Differential Rev(s):
       Wiki Page:                    |
-------------------------------------+-------------------------------------

Comment (by heisenbug):

 Replying to [comment:5 simonpj]:
 > > So something is definitely fishy here.
 >
 > I think that there's something different about '''top-level''' binders.
 Consider
 > {{{#!hs
 > data T = MkT !Bool
 > top_x = True
 >
 > f True  (MkT local_y) = local_y
 > f False _             = top_x
 > }}}
 > Here I think that `local_y` gets bound to a correctly-tagged pointer,
 fetched out of the `MkT` constructor.
 >
 > But, in contrast, I think that `top_x` is bound to the label for the
 top-level closure for `top_x`, which is 8-byte aligned.  So the label
 isn't tagged; instead, the code generator has to tag it.  Is that
 happening in your "return it instead of eantering it" path?


 No, I am not returning top-level bindings, I would be okay with entering
 those. What I want is to avoid entering local bindings from the `StgCase`:
 {{{#!hs
 f = \inp -> case inp of wildBind { True -> ...;  False -> wildBind }
 }}}
 `wildBind` is known to be evaled and properly tagged. I want  to
 `ReturnIt`.

 >
 > Other than that I have no idea why it crashes.  Things you might try:
 >
 > * Switch off the new optimisation when building stage2 and the
 libraries.  Use it only for the test suite: these are small programs and
 easier to debug.
 >
 > * Maybe add an assertion in the Cmm: the claim is that on the "return
 it" path, the thing beign returned is a correctly-tagged pointer.  So, the
 assertion can follow the pointer and check that the thing pointed to is a
 value, with the right tag, etc.  There must be some code in the RTS (or
 somewhere) that checks that.

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/14626#comment:6>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler