[GHC] #15939: StgLint fails because Stg bindings are not dependency-ordered

[GHC]

#15939: StgLint fails because Stg bindings are not dependency-ordered
-------------------------------------+-------------------------------------
           Reporter:  osa1           |             Owner:  (none)
               Type:  bug            |            Status:  new
           Priority:  normal         |         Milestone:
          Component:  Compiler       |           Version:
           Keywords:                 |  Operating System:  Unknown/Multiple
       Architecture:                 |   Type of failure:  None/Unknown
  Unknown/Multiple                   |
          Test Case:                 |        Blocked By:
           Blocking:                 |   Related Tickets:  #9718
Differential Rev(s):                 |         Wiki Page:
-------------------------------------+-------------------------------------
 After Phab:D5370 if I enable `-dstg-lint` when running the test suite I
 get
 about a dozen failures. I didn't check all, but most of them are because
 Stg
 bindings are not dependency-ordered, so some bindings come later in the
 list of
 bindings than their uses. I added a few debug prints to figure out if we
 ever
 have the bindings ordered in the pipeline. As far as I can see, by the
 time we
 desugar we have dependency-ordered bindings, but simplifier breaks the
 ordering
 by introducing uses. There are also "implicit" bindings introduced after
 simplifications (e.g. in CorePrep). Because of this by the time we start
 Stg
 generation we already have incorrect ordering.

 I think we have these options:

 - Don't assume dependency ordering in StgLint. Bring all top-level binders
 into
   the scope before linting bindings.

 - Somehow maintain dependency ordering in all of the Core passes so that
 by the
   time we start generating Stg we have an ordered list of bindings. (seems
 hard)

 - Sort Stg bindings in `CoreToStg`.

 (1) and (3) are equally simple, but (3) requires an extra pass over
 bindings.
 However, to be able to do the `CafInfo` analysis for #9718 efficiently we
 need
 (3), so I suggest implementing (3). We can then enable `-dstg-lint` in the
 test
 suite for all tests (add it to `TEST_HC_OPTS` in `testsuite/mk/test/mk`).

-- 
Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/15939>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler