-
simonpj authored
A regrettably-gigantic commit that puts in place what Simon PJ has been up to for the last month or so, on and off. The basic idea was to restore unfoldings to *occurrences* of variables without introducing a space leak. I wanted to make sure things improved relative to 4.04, and that proved depressingly hard. On the way I discovered several quite serious bugs in the simplifier. Here's a summary of what's gone on. ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ * No commas between for-alls in RULES. This makes the for-alls have the same syntax as in types. * Arrange that simplConArgs works in one less pass than before. This exposed a bug: a bogus call to completeBeta. * Add a top-level flag in CoreUnfolding, used in callSiteInline * Extend w/w to use etaExpandArity, so it does eta/coerce expansion * Implement inline phases. The meaning of the inline pragmas is described in CoreUnfold.lhs. You can say things like {#- INLINE 2 build #-} to mean "inline build in phase 2" * Don't float anything out of an INLINE. Don't float things to top level unless they also escape a value lambda. [see comments with SetLevels.lvlMFE Without at least one of these changes, I found that {-# INLINE concat #-} concat = __inline (/\a -> foldr (++) []) was getting floated to concat = __inline( /\a -> lvl a ) lvl = ...inlined version of foldr... Subsequently I found that not floating constants out of an INLINE gave really bad code like __inline (let x = e in \y -> ...) so I now let things float out of INLINE * Implement the "reverse-mapping" idea for CSE; actually it turned out to be easier to implement it in SetLevels, and may benefit full laziness too. * It's a good idea to inline inRange. Consider index (l,h) i = case inRange (l,h) i of True -> l+i False -> error inRange itself isn't strict in h, but if it't inlined then 'index' *does* become strict in h. Interesting! * Big change to the way unfoldings and occurrence info is propagated in the simplifier The plan is described in Subst.lhs with the Subst type Occurrence info is now in a separate IdInfo field than user pragmas * I found that (coerce T (coerce S (\x.e))) y didn't simplify in one round. First we get to (\x.e) y and only then do the beta. Solution: cancel the coerces in the continuation * Amazingly, CoreUnfold wasn't counting the cost of a function an application. * Disable rules in initial simplifier run. Otherwise full laziness doesn't get a chance to lift out a MFE before a rule (e.g. fusion) zaps it. queens is a case in point * Improve float-out stuff significantly. The big change is that if we have \x -> ... /\a -> ...let p = ..a.. in let q = ...p... where p's rhs doesn't x, we abstract a from p, so that we can get p past x. (We did that before.) But we also substitute (p a) for p in q, and then we can do the same thing for q. (We didn't do that, so q got stuck.) This is much better. It involves doing a substitution "as we go" in SetLevels, though.
30b5ebe4