The Ix instance for (,) leaks space in range

[nomeata]

(at least if you take leak to mean unexpected space behaviour).

This was brought to my attention via  http://stackoverflow.com/questions/12780497/puzzling-memory-behavior-in-haskell where someone created a 6×10^6-array and was surprised that showing the array caused an additional (3+2)*8*10^6 bytes to be used. The reason (as far as I could tell) was this code:

instance (Ix a, Ix b) => Ix (a, b) where
    range ((l1,l2),(u1,u2)) = [ (i1,i2) | i1 <- range (l1,u1), i2 <- range (l2,u2) ]

in Arr.GHC. The result of range (l2,u2) is shared between every step in the first component of the index.

I guess it is reasonable to expect that the result of range is never worth sharing (is it?). I am not entirely sure what the best, cleanest, alternative implementation is, though. (At least not without a working dontshare annotation ;-))

[simonpj]

Aha. See #7206, still on my stack, but would be delighted if someone took this on.

[nomeata]

Would #7206 really help here? After all, the problematic code is polymophic, so there is not a concrete range that could be inlined to a cheapBuild.

In code where the tuple-range is inlined and the concrete range functions of the element types are known (as in the second example in the stack exchange question), things already work out nicely as the foldr/builder fusion actually kicks in.

[igloo]

Milestone _|_ to match #7206.