Some very rough benchmarks: folding over these two implementations of a binary tree data Node1 a = Node1 a (Maybe (Node1 a)) (Maybe (Node1 a))
data Node2 a = Node2A a | Node2B a (Node2 a) | Node2C a (Node2 a) | Node2D a (Node2 a) (Node2 a)
with the latter of Simon's examples gives the following measurements when performing 100 traversals over randomly generated trees of size 3000: min mean +/-sd median max Packed: 12.001 16.001 4.000 16.001 24.002
Unpacked: 12.000 13.144 1.952 12.001 16.001
(I'm not entirely sure what I'm doing, as this is my first contribution to the compiler proper, and pointers as to how to obtain more convincing benchmarks would be appreciated.)
My motivation is more philosophical than anything else, though:
The behavior of unpacking several multi-constructor types can be highly useful to the programmer, and is also (precisely because of the exponential growth) difficult and time-consuming for the programmer to duplicate by hand.
Currently, UNPACK pragmas have *no effect* when used on multi-constructor types, and give no warnings or other hints that they are having no effect. There is currently no reason for a programmer to use an UNPACK pragma on a multi-constructor type, so I would not expect to see code bloat occurring in older programs. The primary exception to this would be in programmers that use the (already regarded as a sledgehammer) -funbox-strict-fields option.
As it stands, it is already the programmer's burden to know how many constructors an UNPACK'd type has, because the pragma will only have any effect if it is a single-constructor type. The effect of the proposal is to attach consequences to UNPACKing multi-constructor types, and placing the burden on the programmer to be sure that exponentially great code expansion does not get out of hand.
The first two bullets are really the ones that grate on me -- that use of the UNPACK pragma in this fashion on multi-constructor types could be seriously useful to a programmer, but currently has no effect, and as a result, I would expect that implementing this proposal wouldn't cause problems in old programs and could be useful in new ones.
I think I'm making sense. Would anyone else care to chime in?