Perldelta 5.8.2 says:
The hash randomisation introduced with 5.8.1 has been amended. It
transpired that although the implementation introduced in 5.8.1 was source
compatible with 5.8.0, it was not binary compatible in certain cases. 5.8.2
contains an improved implementation which is both source and binary
compatible with both 5.8.0 and 5.8.1, and remains robust against the form of
attack which prompted the change for 5.8.1.
What it doesn't say, is that an adaptive algorithm has been implemented that will cause a re-hashing of keys if any list (of identical hash keys) becomes too long. At least, that's what I remember from following the discussion from a distance. You might want to check out the p5p archives for specifics.