I have a program with this data structure: data Element = Element { elementOrigin :: V, elementSubs :: [Element] } and this important bit of code that operates on it: transform :: T -> Element -> Element transform t e = Element { elementOrigin = tmulv t (elementOrigin e), elementSubs = map (transform t) (elementSubs e) } Where V is a vertex and T is a matrix transformation. The following is true: transform a (transform b e) == transform (a * b) e I have read about rewrite rules, such would efficiently rewrite the obvious transformations. However, given that the program will be applying *many* transformations to very deep and wide Element trees, and applying those transformations at many levels, I don't see how the optimizer would be able to handle all cases, for instance a thunk getting left over from perhaps evaluation prior to some IO function. FWIW here's an example "tree-builder" I'm using to test with: mkElems 0 _ = Element { elementOrigin = V 0 0, elementSubs = [] } mkElems depth width = Element { elementOrigin = V 0 0, elementSubs = replicate width $ transform (rotation (pi / (fromIntegral width))) $ transform (translation $ V 1 1) $ mkElems (depth - 1) width } with depth = width = 5 If I could somehow arrange for the transform function to detect when it is being applied to a unevaluated thunk, and then modify the thunk in place, that would basically be the behavior I need. Any suggestions? -- http://petertodd.org 'peter'[:-1]@petertodd.org