[nem-en] patch for matching compiler (second take)

[Marcin Grzeskowiak]

Hi

This is the second patch with new matching compiler. This
time it is for revision 5869 of Nemerle repository.

Changes since the first patch:

1) Span for all enum types (with or without System.Flags
   attribute) is now -1 ("infinite") and as a result
   the decision tree always contains a Decision.Failure
   node for matching on enums (lack of this node was
   responsible for problems mentioned in the earlier
   post). I also changed Con.FindValueExcept ()
   to mask some counter examples with enums (this
   is similar to the way counter examples with nulls
   are masked). The semantics changed a bit:
   
   enum E { | R | G | B }

   foo (e : E) : void
   {
     | R => {}
     | G => {}
     | B => {}
     | _ => {}
   }

   Now the last clause ('_') is not reported as unused
   (and it indeed matches foo (E.R | E.B)). This is
   similar to the way semantics of null matching changed
   in the previous patch ('_' also matches null).
   I added a testcase positive/new-matching-enums.n 
   and modified frommcs/test-234.n

2) To solve the problem with wrong line numbers in 
   warnings (a test case where matching was embedded in
   a try-with block) I changed function ImplicitCast () in 
   typing/Typer.n so it sets location of TExpr.TypeConversion
   node to that of casted expression. 
   
3) This patch includes TupleType.IsTupleMember ()
   function (from Kamil with my corrections).

4) In test driver (testsuite/test.n) there is now additional
   test before any process is killed, e.g.:

   unless (nem_compile.WaitForExit (20000) || nem_compile.HasExited)
     nem_compile.Kill ();

   This solves the problems I had with InvalidOperationException
   on Mono (but again, this just works and I didn't put much effort
   into investigating the source of the problem).

5) The code to recognize shared nodes in decision tree changed
   a bit. As a result all nodes marked as shared have in_deg > 1
   also in the decision DAG and the generated TExpr is not
   cluttered with unused labels.
   I added new-matching-shared.n test case.

My list of TODO items is still not empty but I decided to not
work on any of them at least until I complete other parts of
my thesis. Here are some of the issues (maybe it is a good
idea to add them to the bug database):

- modify Con.FindValueExcept()/DecisionTreeBuilder.CheckMatching()
  functions to not use exceptions to block unwanted warnings 
  (this was suggested by Kamil as throwing can be slow)

- I think that counter examples with 'null's are OK when the
  null value was explicit in the matching, e.g.:

  variant V { | R | G | B }
  match (vv : V * V) {
    | (R, _) => {}
    | (G, _) => {}
    | (B, _) => {}
    | (null, R) => {}
  }

  this code could be reported with (null, B) or (null, G) 
  counter example, but currently all counter examples that
  contain 'null' are blocked

- add optimization that uses GetVariantCode () for variants.

- matching compiler sometimes generates deeply nested if-else
  instructions which then are translated to IL with 'br'
  sequences similar to that one:
  
  IL_0140: br IL_014a
  IL_0145: br IL_0136
  IL_014a: br IL_0154
  IL_014f: br IL_0136
  IL_0154: br IL_015e
  IL_0159: br IL_0136
  IL_015e: br IL_0168
  IL_0163: br IL_0136
  IL_0168: br IL_0172
  IL_016d: br IL_00b2
  IL_0172: br IL_017c
  IL_0177: br IL_00b2
  IL_017c: br IL_0186
  IL_0181: br IL_00b2
  IL_0186: br IL_0190
  IL_018b: br IL_00b2
  IL_0190: ret

  (this is from:
     match (x) {
       | ([1,2], _) => {}
       | (_, [1,2]) => {}
       | _ => {}
     }
  )

  I think that some of flow analysis optimizations could
  help here (if I remember correctly there was something
  like "elimination of jumps to jumps" doable on 
  basic-block DAGs).

- remove castclass instructions from the generated IL
  code (one suggestion was to use stloc/ldoc pairs or
  dup/pop)

- add some message to MatchFailureException whenever
  the failure is due to unmatched null value



Marcin


*************************************************
    Śląski Pasaż Handlowy - wpadnij na zakupy
            http://pasaz.silesianet.pl
*************************************************
-------------- next part --------------
A non-text attachment was scrubbed...
Name: new-matching.patch
Type: application/octet-stream
Size: 59638 bytes
Desc: not available
Url : /mailman/pipermail/devel-en/attachments/20051031/0ae2f5f5/new-matching-0001.obj