Copyright | (c) 2013-2016 Galois Inc. |
---|---|
License | BSD3 |
Maintainer | cryptol@galois.com |
Stability | provisional |
Portability | portable |
Safe Haskell | Safe |
Language | Haskell2010 |
This module implements a transformation, which tries to avoid exponential slow down in some cases. What's the problem? Consider the following (common) patterns:
fibs = [0,1] # [ x + y | x <- fibs, y <- drop`{1} fibs ]
The type of fibs
is:
{a} (a >= 1, fin a) => [inf][a]
Here a
is the number of bits to be used in the values computed by fibs
.
When we evaluate fibs
, a
becomes a parameter to fibs
, which works
except that now fibs
is a function, and we don't get any of the memoization
we might expect! What looked like an efficient implementation has all
of a sudden become exponential!
Note that this is only a problem for polymorphic values: if fibs
was
already a function, it would not be that surprising that it does not
get cached.
So, to avoid this, we try to spot recursive polymorphic values,
where the recursive occurrences have the exact same type parameters
as the definition. For example, this is the case in fibs
: each
recursive call to fibs
is instantiated with exactly the same
type parameter (i.e., a
). The rewrite we do is as follows:
fibs : {a} (a >= 1, fin a) => [inf][a] fibs = \{a} (a >= 1, fin a) -> fibs' where fibs' : [inf][a] fibs' = [0,1] # [ x + y | x <- fibs', y <- drop`{1} fibs' ]
After the rewrite, the recursion is monomorphic (i.e., we are always using
the same type). As a result, fibs'
is an ordinary recursive value,
where we get the benefit of caching.
The rewrite is a bit more complex, when there are multiple mutually recursive functions. Here is an example:
zig : {a} (a >= 2, fin a) => [inf][a] zig = [1] # zag zag : {a} (a >= 2, fin a) => [inf][a] zag = [2] # zig
This gets rewritten to:
newName : {a} (a >= 2, fin a) => ([inf][a], [inf][a]) newName = \{a} (a >= 2, fin a) -> (zig', zag') where zig' : [inf][a] zig' = [1] # zag' zag' : [inf][a] zag' = [2] # zig' zig : {a} (a >= 2, fin a) => [inf][a] zig = \{a} (a >= 2, fin a) -> (newName a <> <> ).1 zag : {a} (a >= 2, fin a) => [inf][a] zag = \{a} (a >= 2, fin a) -> (newName a <> <> ).2
NOTE: We are assuming that no capture would occur with binders. For values, this is because we replaces things with freshly chosen variables. For types, this should be because there should be no shadowing in the types. XXX: Make sure that this really is the case for types!!