Improve SpecConstr for free variables

This ticket records a suggestion for improving SpecConstr, so we don't lose it. The original SpecConstr transformation is described in "Call pattern specialisation for Haskell". Consider this program

   f x = let g y = ...case x of { z:zs -> e1; [] -> e2 } ...
          in
          ...case x of
                z:zs -> if ... then g 3 else g 4
                []   -> ...

Here 'x' is free in 'g', but x's value is known at g's call sites. It's not enough just to know "x is a cons" inside g; we must also have access to z,zs. So perhaps the thing to do is to imagine lambda-lifting 'g' to become 'gl' thus:

  gl x y = ...case x of { z:zs -> e1; [] -> e2 } ...

  f x = ...case x of
                z:zs -> if ... then gl x 3 else gl x 4
                []   -> ...

Now the SpecConstr transformation will apply nicely. And it's arguably a bad shortcoming that currently the mere act of lambda lifting can affect how effective SpecConstr is.

Of course, we only want to lambda-lift wrt the parameters that'll be specialised, so this transformation probably wants to be done at the same time as the rest of SpecConstr. I don't have much idea of how hard that'd be, but it's a nice idea.

Trac metadata

Trac field	Value
Version	6.8.2
Type	Bug
TypeOfFailure	OtherFailure
Priority	normal
Resolution	Unresolved
Component	Compiler
Test case
Differential revisions
BlockedBy
Related
Blocking
CC
Operating system	Unknown
Architecture	Unknown