Get rid of the superstitious "~" in dict hashing's "i = (~hash) & mask".

The comment following used to say:
	/* We use ~hash instead of hash, as degenerate hash functions, such
	   as for ints <sigh>, can have lots of leading zeros. It's not
	   really a performance risk, but better safe than sorry.
	   12-Dec-00 tim:  so ~hash produces lots of leading ones instead --
	   what's the gain? */
That is, there was never a good reason for doing it.  And to the contrary,
as explained on Python-Dev last December, it tended to make the *sum*
(i + incr) & mask (which is the first table index examined in case of
collison) the same "too often" across distinct hashes.

Changing to the simpler "i = hash & mask" reduced the number of string-dict
collisions (== # number of times we go around the lookup for-loop) from about
6 million to 5 million during a full run of the test suite (these are
approximate because the test suite does some random stuff from run to run).
The number of collisions in non-string dicts also decreased, but not as
dramatically.

Note that this may, for a given dict, change the order (wrt previous
releases) of entries exposed by .keys(), .values() and .items().  A number
of std tests suffered bogus failures as a result.  For dicts keyed by
small ints, or (less so) by characters, the order is much more likely to be
in increasing order of key now; e.g.,

>>> d = {}
>>> for i in range(10):
...    d[i] = i
...
>>> d
{0: 0, 1: 1, 2: 2, 3: 3, 4: 4, 5: 5, 6: 6, 7: 7, 8: 8, 9: 9}
>>>

Unfortunately. people may latch on to that in small examples and draw a
bogus conclusion.

test_support.py
    Moved test_extcall's sortdict() into test_support, made it stronger,
    and imported sortdict into other std tests that needed it.
test_unicode.py
    Excluced cp875 from the "roundtrip over range(128)" test, because
    cp875 doesn't have a well-defined inverse for unicode("?", "cp875").
    See Python-Dev for excruciating details.
Cookie.py
    Chaged various output functions to sort dicts before building
    strings from them.
test_extcall
    Fiddled the expected-result file.  This remains sensitive to native
    dict ordering, because, e.g., if there are multiple errors in a
    keyword-arg dict (and test_extcall sets up many cases like that), the
    specific error Python complains about first depends on native dict
    ordering.

1 files changed, 5 insertions, 2 deletions

diff --git a/Lib/test/test_unicode.py b/Lib/test/test_unicode.py
index 0df8217..c82ac69 100644
--- a/Lib/test/test_unicode.py
+++ b/Lib/test/test_unicode.py
@@ -5,7 +5,7 @@ Written by Marc-Andre Lemburg (mal@lemburg.com).
 (c) Copyright CNRI, All Rights Reserved. NO WARRANTY.
 
 """#"
-from test_support import verify, verbose
+from test_support import verify, verbose, TestFailed
 import sys
 
 def test(method, input, output, *args):
@@ -493,11 +493,14 @@ for encoding in (
     'cp856', 'cp857', 'cp864', 'cp869', 'cp874',
 
     'mac_greek', 'mac_iceland','mac_roman', 'mac_turkish',
-    'cp1006', 'cp875', 'iso8859_8',
+    'cp1006', 'iso8859_8',
 
     ### These have undefined mappings:
     #'cp424',
 
+    ### These fail the round-trip:
+    #'cp875'
+
     ):
     try:
         verify(unicode(s,encoding).encode(encoding) == s)