Strictness problem

[nilsanders]

As requested, this is a resubmission (and update) of a
previously reported bug, to get it into the bug tracker.

The following program should output something involving
Correct:

> module Main where

> f x = case x of
>   x@True  -> \y -> x && y
>   x@False -> \y -> x && y

> main = putStrLn $ f (error "Correct") `seq` "Error"

However, whether it does so is a non-trivial function
of the GHC version and optimisation settings:

GHC version     -O2?  Correct?
------------------------------
4.08.1          No    Yes
4.08.1          Yes   No
5.04.2          No    No
5.04.2          Yes   Yes
6.0.1           _     No
6.2.2           _     No
6.4             _     No
6.4.1.20050820  _     No

All tests were run on a Solaris system.

Different fs give different behaviour, at least for
6.0.1. Try e.g.

> f x = case x of
>   True  -> id
>   False -> id

[simonpj]

Related to #317.

Caused by the fact that we eta-expand around the case expression. Not eta-expanding hurts performance in important cases (we think). Still thinking what to do about this; it's technically wrong but seldom seems to bite in practice.

Simon

[simonmar]

See also #826

[igloo]

This is still wrong in 6.6 and the HEAD.

[igloo]

Still happens in 6.12.

[michalt]

Replying to igloo:

Still happens in 6.12.

Current HEAD (7.1.20101015):

• without optimisations - correct
• with -O or -O2 - error