And here is alternative patch which uses a hashtable.

Both patches have about the same performance for *.pyc files, but marshal_hashtable.patch is much faster for duplicated values. Marshalling [1000]*10**6, [1000.0]*10**6 and [1000.0j]*10**6 with version 3 an 4 is so fast as marshalling [1000]*10**6 with version 2 (i.e. 5 times faster than current implementation).

data              ver.  dumps(ms)  loads(ms)  size(KiB)

genData            2       99.9      188.9     4090.7
genData            3      148.2      189.1     4090.7
genData            4      121.4      177.4     3651.3

[1000]*10**6       2       97.7      131.6     4882.8
[1000]*10**6       3       95.1       63.1     4882.8
[1000]*10**6       4       95.1       64.4     4882.8

[1000.0]*10**6     2      172.9      153.5     8789.1
[1000.0]*10**6     3       97.4       61.9     4882.8
[1000.0]*10**6     4       95.7       61.6     4882.8

[1000.0j]*10**6    2      288.6      228.2    16601.6
[1000.0j]*10**6    3       94.9       61.6     4882.8
[1000.0j]*10**6    4       95.1       62.2     4882.8

20 pydecimals      2       88.0      111.4     3929.6
20 pydecimals      3       57.0       51.4     3368.5
20 pydecimals      4       46.6       39.9     3292.8