Confusing flags for rewrite rules

[dons]

Functions exported via rewrite rules are incorrectly flagged as "defined but not used" when -Wall is enabled:

module M ({- rules -}) where

eq :: Eq a => a -> a -> Bool
eq = (==)

{-# RULES
    "rule 1" forall x. x == y = y `eq` x
  #-}

Will emit a bogus:

  M.hs:4:0: Warning: Defined but not used: `eq'

[dons]

Ah, worked this one out.

-fglasgow-exts is needed to enable rules, otherwise they're considered comments. (Hence the type error in this rule is silently ignored).

Note however, strangely, -frewrite-rules *doesn't* fix this.

[igloo]

This is currently rather odd. You need one of ScopedTypeVariables, PolymorphicComponents, ExistentialQuantification, Rank2Types and RankNTypes enabled, as these all turn on lexing of forall. Really, there ought to be a flag for rewrite rules instead (which also turns on lexing of forall).

[simonpj]

I agree this is totally weird, and we must fix it for 6.10. The current situation is:

• RULES pragmas are discarded by the lexer unless explicitForallEnabled is on
• explicitForAllEnabled is set only by the flags Ian mentions above
• Those flags are set by -fglasgow-exts

There is an optimisation flag -frewrite-rules, which controls whether rewrite rules are applied during optimisation. It's not meant to be a language-extension flag.

We could have a language-extension flag -XRewriteRules, which switches on RULES in source code. (Even without it, optimisation could still use imported rules, as controlled by the optimisation flag -frewrite-rules.) It could be switched on by -fglasgow-exts. It would switch on lexing of forall.

Presumably RULES should be ignored without -XRewriteRules. (Would we want a warning? And yet another warning-supression flag?) I'm a bit concerned that this would break old programs that don't use -fglasgow-exts, but do have -XScopedTypeVariables (say), by silently discarding their rewrite rules. This seems like a strong argument for adding a warning for discarded RULES. This warning would also have warned Don in the program that started this ticket.

Should the optimisation flag still be called -frewrite-rules. I think we should re-title it -frun-rewrite-rules (implied by -O of course).

Nothing very hard here, but better to get the design right first. Comments?

Meanwhile I'll re-title the ticket.

[claus]

i ran into -frewrite-rules here:

 http://www.haskell.org/pipermail/glasgow-haskell-users/2008-June/014912.html

the (slightly incorrectly) observed behaviour is *very* confusing:

* -frewrite-rules on its own is not sufficient to enable RULES, presumably because they are discarded before they could be enabled

* -fglasgow-exts on its own is sufficient, presumably because -O implied -frewrite-rules anyway

imho, wanting RULES to be applied implies wanting them to be parsed, not discarded, so -frewrite-rules should imply all relevant language flags. until that is the case, the odd dependency should be documented in the flag reference.

btw, i changed the odd cc, moving rules, warnings to keywords. or do these exist as ccs?

[simonpj]

See the discussion and fix in #2497.

Simon