[GHC] #10897: Incorrect ASSERT for buildPatSyn

18 Sep 2015

      #10897: Incorrect ASSERT for buildPatSyn
-------------------------------------+-------------------------------------
              Reporter:  ezyang      |             Owner:  cactus
                  Type:  bug         |            Status:  new
              Priority:  normal      |         Milestone:
             Component:  Compiler    |           Version:  7.11
  (Type checker)                     |
              Keywords:              |  Operating System:  Unknown/Multiple
          Architecture:              |   Type of failure:  Compile-time
  Unknown/Multiple                   |  crash
             Test Case:              |        Blocked By:
              Blocking:              |   Related Tickets:
Differential Revisions:              |
-------------------------------------+-------------------------------------
 Consider the following two files:

 {{{
 -- A.hs
 {-# LANGUAGE PatternSynonyms #-}
 module A where
 pattern Single :: a -> a
 pattern Single x = x

 -- B.hs
 module B where
 import A
 Single y = True
 }}}

 When I build these using one-shot compilation using a debugged GHC (i.e.
 with ASSERTs) I get the following error:

 {{{
 [ezyang@hs01 ghc-quick3]$ inplace/bin/ghc-stage2 -c A.hs -fforce-recomp
 [ezyang@hs01 ghc-quick3]$ inplace/bin/ghc-stage2 -c B.hs -fforce-recomp
 ghc-stage2: panic! (the 'impossible' happened)
   (GHC version 7.11.20150918 for x86_64-unknown-linux):
         ASSERT failed! file compiler/iface/BuildTyCl.hs, line 210

 Please report this as a GHC bug:  http://www.haskell.org/ghc/reportabug
 }}}

 I expanded the assert and discovered it was claiming the universal type
 variables of the matcher should be equal to the type variables of the
 pattern declaration. This assert cannot possibly be right: the matcher and
 the pattern declaration are typechecked separately and there's no reason
 that the local binders should actually be the same. The equality test here
 should be done up to alpha-renaming.

 The other thing I found a bit puzzling was whether or not it mattered
 whether or not we used the local type variables from the matcher or the
 freshly bound ones. I suppose if we are consistent it shouldn't matter, so
 I don't think the code is buggy, just a bad ASSERT.

 BTW: this assert problem doesn't show up with `--make` because the assert
 occurs during typechecking of interface files, and with `--make` we don't
 need to typecheck an interface file.

--
Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/10897
GHC http://www.haskell.org/ghc/
The Glasgow Haskell Compiler