| 1 |
from test import support, seq_tests |
| 2 |
import unittest |
| 3 |
|
| 4 |
import gc |
| 5 |
import pickle |
| 6 |
|
| 7 |
# For tuple hashes, we normally only run a test to ensure that we get |
| 8 |
# the same results across platforms in a handful of cases. If that's |
| 9 |
# so, there's no real point to running more. Set RUN_ALL_HASH_TESTS to |
| 10 |
# run more anyway. That's usually of real interest only when analyzing, |
| 11 |
# or changing, the hash algorithm. In which case it's usually also |
| 12 |
# most useful to set JUST_SHOW_HASH_RESULTS, to see all the results |
| 13 |
# instead of wrestling with test "failures". See the bottom of the |
| 14 |
# file for extensive notes on what we're testing here and why. |
| 15 |
RUN_ALL_HASH_TESTS = False |
| 16 |
JUST_SHOW_HASH_RESULTS = False # if RUN_ALL_HASH_TESTS, just display |
| 17 |
|
| 18 |
class TupleTest(seq_tests.CommonTest): |
| 19 |
type2test = tuple |
| 20 |
|
| 21 |
def test_getitem_error(self): |
| 22 |
t = () |
| 23 |
msg = "tuple indices must be integers or slices" |
| 24 |
with self.assertRaisesRegex(TypeError, msg): |
| 25 |
t['a'] |
| 26 |
|
| 27 |
def test_constructors(self): |
| 28 |
super().test_constructors() |
| 29 |
# calling built-in types without argument must return empty |
| 30 |
self.assertEqual(tuple(), ()) |
| 31 |
t0_3 = (0, 1, 2, 3) |
| 32 |
t0_3_bis = tuple(t0_3) |
| 33 |
self.assertTrue(t0_3 is t0_3_bis) |
| 34 |
self.assertEqual(tuple([]), ()) |
| 35 |
self.assertEqual(tuple([0, 1, 2, 3]), (0, 1, 2, 3)) |
| 36 |
self.assertEqual(tuple(''), ()) |
| 37 |
self.assertEqual(tuple('spam'), ('s', 'p', 'a', 'm')) |
| 38 |
self.assertEqual(tuple(x for x in range(10) if x % 2), |
| 39 |
(1, 3, 5, 7, 9)) |
| 40 |
|
| 41 |
def test_keyword_args(self): |
| 42 |
with self.assertRaisesRegex(TypeError, 'keyword argument'): |
| 43 |
tuple(sequence=()) |
| 44 |
|
| 45 |
def test_keywords_in_subclass(self): |
| 46 |
class subclass(tuple): |
| 47 |
pass |
| 48 |
u = subclass([1, 2]) |
| 49 |
self.assertIs(type(u), subclass) |
| 50 |
self.assertEqual(list(u), [1, 2]) |
| 51 |
with self.assertRaises(TypeError): |
| 52 |
subclass(sequence=()) |
| 53 |
|
| 54 |
class subclass_with_init(tuple): |
| 55 |
def __init__(self, arg, newarg=None): |
| 56 |
self.newarg = newarg |
| 57 |
u = subclass_with_init([1, 2], newarg=3) |
| 58 |
self.assertIs(type(u), subclass_with_init) |
| 59 |
self.assertEqual(list(u), [1, 2]) |
| 60 |
self.assertEqual(u.newarg, 3) |
| 61 |
|
| 62 |
class subclass_with_new(tuple): |
| 63 |
def __new__(cls, arg, newarg=None): |
| 64 |
self = super().__new__(cls, arg) |
| 65 |
self.newarg = newarg |
| 66 |
return self |
| 67 |
u = subclass_with_new([1, 2], newarg=3) |
| 68 |
self.assertIs(type(u), subclass_with_new) |
| 69 |
self.assertEqual(list(u), [1, 2]) |
| 70 |
self.assertEqual(u.newarg, 3) |
| 71 |
|
| 72 |
def test_truth(self): |
| 73 |
super().test_truth() |
| 74 |
self.assertTrue(not ()) |
| 75 |
self.assertTrue((42, )) |
| 76 |
|
| 77 |
def test_len(self): |
| 78 |
super().test_len() |
| 79 |
self.assertEqual(len(()), 0) |
| 80 |
self.assertEqual(len((0,)), 1) |
| 81 |
self.assertEqual(len((0, 1, 2)), 3) |
| 82 |
|
| 83 |
def test_iadd(self): |
| 84 |
super().test_iadd() |
| 85 |
u = (0, 1) |
| 86 |
u2 = u |
| 87 |
u += (2, 3) |
| 88 |
self.assertTrue(u is not u2) |
| 89 |
|
| 90 |
def test_imul(self): |
| 91 |
super().test_imul() |
| 92 |
u = (0, 1) |
| 93 |
u2 = u |
| 94 |
u *= 3 |
| 95 |
self.assertTrue(u is not u2) |
| 96 |
|
| 97 |
def test_tupleresizebug(self): |
| 98 |
# Check that a specific bug in _PyTuple_Resize() is squashed. |
| 99 |
def f(): |
| 100 |
for i in range(1000): |
| 101 |
yield i |
| 102 |
self.assertEqual(list(tuple(f())), list(range(1000))) |
| 103 |
|
| 104 |
# We expect tuples whose base components have deterministic hashes to |
| 105 |
# have deterministic hashes too - and, indeed, the same hashes across |
| 106 |
# platforms with hash codes of the same bit width. |
| 107 |
def test_hash_exact(self): |
| 108 |
def check_one_exact(t, e32, e64): |
| 109 |
got = hash(t) |
| 110 |
expected = e32 if support.NHASHBITS == 32 else e64 |
| 111 |
if got != expected: |
| 112 |
msg = f"FAIL hash({t!r}) == {got} != {expected}" |
| 113 |
self.fail(msg) |
| 114 |
|
| 115 |
check_one_exact((), 750394483, 5740354900026072187) |
| 116 |
check_one_exact((0,), 1214856301, -8753497827991233192) |
| 117 |
check_one_exact((0, 0), -168982784, -8458139203682520985) |
| 118 |
check_one_exact((0.5,), 2077348973, -408149959306781352) |
| 119 |
check_one_exact((0.5, (), (-2, 3, (4, 6))), 714642271, |
| 120 |
-1845940830829704396) |
| 121 |
|
| 122 |
# Various tests for hashing of tuples to check that we get few collisions. |
| 123 |
# Does something only if RUN_ALL_HASH_TESTS is true. |
| 124 |
# |
| 125 |
# Earlier versions of the tuple hash algorithm had massive collisions |
| 126 |
# reported at: |
| 127 |
# - https://bugs.python.org/issue942952 |
| 128 |
# - https://bugs.python.org/issue34751 |
| 129 |
def test_hash_optional(self): |
| 130 |
from itertools import product |
| 131 |
|
| 132 |
if not RUN_ALL_HASH_TESTS: |
| 133 |
return |
| 134 |
|
| 135 |
# If specified, `expected` is a 2-tuple of expected |
| 136 |
# (number_of_collisions, pileup) values, and the test fails if |
| 137 |
# those aren't the values we get. Also if specified, the test |
| 138 |
# fails if z > `zlimit`. |
| 139 |
def tryone_inner(tag, nbins, hashes, expected=None, zlimit=None): |
| 140 |
from collections import Counter |
| 141 |
|
| 142 |
nballs = len(hashes) |
| 143 |
mean, sdev = support.collision_stats(nbins, nballs) |
| 144 |
c = Counter(hashes) |
| 145 |
collisions = nballs - len(c) |
| 146 |
z = (collisions - mean) / sdev |
| 147 |
pileup = max(c.values()) - 1 |
| 148 |
del c |
| 149 |
got = (collisions, pileup) |
| 150 |
failed = False |
| 151 |
prefix = "" |
| 152 |
if zlimit is not None and z > zlimit: |
| 153 |
failed = True |
| 154 |
prefix = f"FAIL z > {zlimit}; " |
| 155 |
if expected is not None and got != expected: |
| 156 |
failed = True |
| 157 |
prefix += f"FAIL {got} != {expected}; " |
| 158 |
if failed or JUST_SHOW_HASH_RESULTS: |
| 159 |
msg = f"{prefix}{tag}; pileup {pileup:,} mean {mean:.1f} " |
| 160 |
msg += f"coll {collisions:,} z {z:+.1f}" |
| 161 |
if JUST_SHOW_HASH_RESULTS: |
| 162 |
import sys |
| 163 |
print(msg, file=sys.__stdout__) |
| 164 |
else: |
| 165 |
self.fail(msg) |
| 166 |
|
| 167 |
def tryone(tag, xs, |
| 168 |
native32=None, native64=None, hi32=None, lo32=None, |
| 169 |
zlimit=None): |
| 170 |
NHASHBITS = support.NHASHBITS |
| 171 |
hashes = list(map(hash, xs)) |
| 172 |
tryone_inner(tag + f"; {NHASHBITS}-bit hash codes", |
| 173 |
1 << NHASHBITS, |
| 174 |
hashes, |
| 175 |
native32 if NHASHBITS == 32 else native64, |
| 176 |
zlimit) |
| 177 |
|
| 178 |
if NHASHBITS > 32: |
| 179 |
shift = NHASHBITS - 32 |
| 180 |
tryone_inner(tag + "; 32-bit upper hash codes", |
| 181 |
1 << 32, |
| 182 |
[h >> shift for h in hashes], |
| 183 |
hi32, |
| 184 |
zlimit) |
| 185 |
|
| 186 |
mask = (1 << 32) - 1 |
| 187 |
tryone_inner(tag + "; 32-bit lower hash codes", |
| 188 |
1 << 32, |
| 189 |
[h & mask for h in hashes], |
| 190 |
lo32, |
| 191 |
zlimit) |
| 192 |
|
| 193 |
# Tuples of smallish positive integers are common - nice if we |
| 194 |
# get "better than random" for these. |
| 195 |
tryone("range(100) by 3", list(product(range(100), repeat=3)), |
| 196 |
(0, 0), (0, 0), (4, 1), (0, 0)) |
| 197 |
|
| 198 |
# A previous hash had systematic problems when mixing integers of |
| 199 |
# similar magnitude but opposite sign, obscurely related to that |
| 200 |
# j ^ -2 == -j when j is odd. |
| 201 |
cands = list(range(-10, -1)) + list(range(9)) |
| 202 |
|
| 203 |
# Note: -1 is omitted because hash(-1) == hash(-2) == -2, and |
| 204 |
# there's nothing the tuple hash can do to avoid collisions |
| 205 |
# inherited from collisions in the tuple components' hashes. |
| 206 |
tryone("-10 .. 8 by 4", list(product(cands, repeat=4)), |
| 207 |
(0, 0), (0, 0), (0, 0), (0, 0)) |
| 208 |
del cands |
| 209 |
|
| 210 |
# The hashes here are a weird mix of values where all the |
| 211 |
# variation is in the lowest bits and across a single high-order |
| 212 |
# bit - the middle bits are all zeroes. A decent hash has to |
| 213 |
# both propagate low bits to the left and high bits to the |
| 214 |
# right. This is also complicated a bit in that there are |
| 215 |
# collisions among the hashes of the integers in L alone. |
| 216 |
L = [n << 60 for n in range(100)] |
| 217 |
tryone("0..99 << 60 by 3", list(product(L, repeat=3)), |
| 218 |
(0, 0), (0, 0), (0, 0), (324, 1)) |
| 219 |
del L |
| 220 |
|
| 221 |
# Used to suffer a massive number of collisions. |
| 222 |
tryone("[-3, 3] by 18", list(product([-3, 3], repeat=18)), |
| 223 |
(7, 1), (0, 0), (7, 1), (6, 1)) |
| 224 |
|
| 225 |
# And even worse. hash(0.5) has only a single bit set, at the |
| 226 |
# high end. A decent hash needs to propagate high bits right. |
| 227 |
tryone("[0, 0.5] by 18", list(product([0, 0.5], repeat=18)), |
| 228 |
(5, 1), (0, 0), (9, 1), (12, 1)) |
| 229 |
|
| 230 |
# Hashes of ints and floats are the same across platforms. |
| 231 |
# String hashes vary even on a single platform across runs, due |
| 232 |
# to hash randomization for strings. So we can't say exactly |
| 233 |
# what this should do. Instead we insist that the # of |
| 234 |
# collisions is no more than 4 sdevs above the theoretically |
| 235 |
# random mean. Even if the tuple hash can't achieve that on its |
| 236 |
# own, the string hash is trying to be decently pseudo-random |
| 237 |
# (in all bit positions) on _its_ own. We can at least test |
| 238 |
# that the tuple hash doesn't systematically ruin that. |
| 239 |
tryone("4-char tuples", |
| 240 |
list(product("abcdefghijklmnopqrstuvwxyz", repeat=4)), |
| 241 |
zlimit=4.0) |
| 242 |
|
| 243 |
# The "old tuple test". See https://bugs.python.org/issue942952. |
| 244 |
# Ensures, for example, that the hash: |
| 245 |
# is non-commutative |
| 246 |
# spreads closely spaced values |
| 247 |
# doesn't exhibit cancellation in tuples like (x,(x,y)) |
| 248 |
N = 50 |
| 249 |
base = list(range(N)) |
| 250 |
xp = list(product(base, repeat=2)) |
| 251 |
inps = base + list(product(base, xp)) + \ |
| 252 |
list(product(xp, base)) + xp + list(zip(base)) |
| 253 |
tryone("old tuple test", inps, |
| 254 |
(2, 1), (0, 0), (52, 49), (7, 1)) |
| 255 |
del base, xp, inps |
| 256 |
|
| 257 |
# The "new tuple test". See https://bugs.python.org/issue34751. |
| 258 |
# Even more tortured nesting, and a mix of signed ints of very |
| 259 |
# small magnitude. |
| 260 |
n = 5 |
| 261 |
A = [x for x in range(-n, n+1) if x != -1] |
| 262 |
B = A + [(a,) for a in A] |
| 263 |
L2 = list(product(A, repeat=2)) |
| 264 |
L3 = L2 + list(product(A, repeat=3)) |
| 265 |
L4 = L3 + list(product(A, repeat=4)) |
| 266 |
# T = list of testcases. These consist of all (possibly nested |
| 267 |
# at most 2 levels deep) tuples containing at most 4 items from |
| 268 |
# the set A. |
| 269 |
T = A |
| 270 |
T += [(a,) for a in B + L4] |
| 271 |
T += product(L3, B) |
| 272 |
T += product(L2, repeat=2) |
| 273 |
T += product(B, L3) |
| 274 |
T += product(B, B, L2) |
| 275 |
T += product(B, L2, B) |
| 276 |
T += product(L2, B, B) |
| 277 |
T += product(B, repeat=4) |
| 278 |
assert len(T) == 345130 |
| 279 |
tryone("new tuple test", T, |
| 280 |
(9, 1), (0, 0), (21, 5), (6, 1)) |
| 281 |
|
| 282 |
def test_repr(self): |
| 283 |
l0 = tuple() |
| 284 |
l2 = (0, 1, 2) |
| 285 |
a0 = self.type2test(l0) |
| 286 |
a2 = self.type2test(l2) |
| 287 |
|
| 288 |
self.assertEqual(str(a0), repr(l0)) |
| 289 |
self.assertEqual(str(a2), repr(l2)) |
| 290 |
self.assertEqual(repr(a0), "()") |
| 291 |
self.assertEqual(repr(a2), "(0, 1, 2)") |
| 292 |
|
| 293 |
# Checks that t is not tracked without any GC collections. |
| 294 |
def _not_tracked_instantly(self, t): |
| 295 |
self.assertFalse(gc.is_tracked(t), t) |
| 296 |
|
| 297 |
# Checks that t is not tracked after GC collection. |
| 298 |
def _not_tracked(self, t): |
| 299 |
# Nested tuples can take several collections to untrack |
| 300 |
gc.collect() |
| 301 |
gc.collect() |
| 302 |
self.assertFalse(gc.is_tracked(t), t) |
| 303 |
|
| 304 |
# Checks that t continues to be tracked even after GC collection. |
| 305 |
def _tracked(self, t): |
| 306 |
self.assertTrue(gc.is_tracked(t), t) |
| 307 |
gc.collect() |
| 308 |
gc.collect() |
| 309 |
self.assertTrue(gc.is_tracked(t), t) |
| 310 |
|
| 311 |
@support.cpython_only |
| 312 |
def test_track_literals(self): |
| 313 |
# Test GC-optimization of tuple literals |
| 314 |
x, y, z = 1.5, "a", [] |
| 315 |
|
| 316 |
# We check that those objects aren't tracked at all. |
| 317 |
# It's essential for the GC performance, see gh-139951. |
| 318 |
self._not_tracked_instantly(()) |
| 319 |
self._not_tracked_instantly((1,)) |
| 320 |
self._not_tracked_instantly((1, 2)) |
| 321 |
self._not_tracked_instantly((1, 2, "a")) |
| 322 |
self._not_tracked_instantly((1, 2) * 5) |
| 323 |
self._not_tracked_instantly((12, 10**10, 'a_' * 100)) |
| 324 |
self._not_tracked_instantly((object(),)) |
| 325 |
|
| 326 |
self._not_tracked(((1, x), y, (2, 3))) |
| 327 |
self._not_tracked((1, 2, (None, True, False, ()), int)) |
| 328 |
self._not_tracked((object(), ())) |
| 329 |
|
| 330 |
# Tuples with mutable elements are always tracked, even if those |
| 331 |
# elements are not tracked right now. |
| 332 |
self._tracked(([],)) |
| 333 |
self._tracked(([1],)) |
| 334 |
self._tracked(({},)) |
| 335 |
self._tracked((set(),)) |
| 336 |
self._tracked((x, y, z)) |
| 337 |
|
| 338 |
def check_track_dynamic(self, tp, always_track): |
| 339 |
x, y, z = 1.5, "a", [] |
| 340 |
|
| 341 |
check = self._tracked if always_track else self._not_tracked |
| 342 |
check(tp()) |
| 343 |
check(tp([])) |
| 344 |
check(tp(set())) |
| 345 |
check(tp([1, x, y])) |
| 346 |
check(tp(obj for obj in [1, x, y])) |
| 347 |
check(tp(set([1, x, y]))) |
| 348 |
check(tp(tuple([obj]) for obj in [1, x, y])) |
| 349 |
check(tuple(tp([obj]) for obj in [1, x, y])) |
| 350 |
|
| 351 |
self._tracked(tp([z])) |
| 352 |
self._tracked(tp([[x, y]])) |
| 353 |
self._tracked(tp([{x: y}])) |
| 354 |
self._tracked(tp(obj for obj in [x, y, z])) |
| 355 |
self._tracked(tp(tuple([obj]) for obj in [x, y, z])) |
| 356 |
self._tracked(tuple(tp([obj]) for obj in [x, y, z])) |
| 357 |
|
| 358 |
t = tp([1, x, y, z]) |
| 359 |
self.assertEqual(type(t), tp) |
| 360 |
self._tracked(t) |
| 361 |
self.assertEqual(type(t[:]), tuple) |
| 362 |
self._tracked(t[:]) |
| 363 |
|
| 364 |
@support.cpython_only |
| 365 |
def test_track_dynamic(self): |
| 366 |
# Test GC-optimization of dynamically constructed tuples. |
| 367 |
self.check_track_dynamic(tuple, False) |
| 368 |
|
| 369 |
@support.cpython_only |
| 370 |
def test_track_subtypes(self): |
| 371 |
# Tuple subtypes must always be tracked |
| 372 |
class MyTuple(tuple): |
| 373 |
pass |
| 374 |
self.check_track_dynamic(MyTuple, True) |
| 375 |
|
| 376 |
@support.cpython_only |
| 377 |
def test_bug7466(self): |
| 378 |
# Trying to untrack an unfinished tuple could crash Python |
| 379 |
self._not_tracked(tuple(gc.collect() for i in range(101))) |
| 380 |
|
| 381 |
def test_repr_large(self): |
| 382 |
# Check the repr of large list objects |
| 383 |
def check(n): |
| 384 |
l = (0,) * n |
| 385 |
s = repr(l) |
| 386 |
self.assertEqual(s, |
| 387 |
'(' + ', '.join(['0'] * n) + ')') |
| 388 |
check(10) # check our checking code |
| 389 |
check(1000000) |
| 390 |
|
| 391 |
def test_iterator_pickle(self): |
| 392 |
# Userlist iterators don't support pickling yet since |
| 393 |
# they are based on generators. |
| 394 |
data = self.type2test([4, 5, 6, 7]) |
| 395 |
for proto in range(pickle.HIGHEST_PROTOCOL + 1): |
| 396 |
itorg = iter(data) |
| 397 |
d = pickle.dumps(itorg, proto) |
| 398 |
it = pickle.loads(d) |
| 399 |
self.assertEqual(type(itorg), type(it)) |
| 400 |
self.assertEqual(self.type2test(it), self.type2test(data)) |
| 401 |
|
| 402 |
it = pickle.loads(d) |
| 403 |
next(it) |
| 404 |
d = pickle.dumps(it, proto) |
| 405 |
self.assertEqual(self.type2test(it), self.type2test(data)[1:]) |
| 406 |
|
| 407 |
def test_reversed_pickle(self): |
| 408 |
data = self.type2test([4, 5, 6, 7]) |
| 409 |
for proto in range(pickle.HIGHEST_PROTOCOL + 1): |
| 410 |
itorg = reversed(data) |
| 411 |
d = pickle.dumps(itorg, proto) |
| 412 |
it = pickle.loads(d) |
| 413 |
self.assertEqual(type(itorg), type(it)) |
| 414 |
self.assertEqual(self.type2test(it), self.type2test(reversed(data))) |
| 415 |
|
| 416 |
it = pickle.loads(d) |
| 417 |
next(it) |
| 418 |
d = pickle.dumps(it, proto) |
| 419 |
self.assertEqual(self.type2test(it), self.type2test(reversed(data))[1:]) |
| 420 |
|
| 421 |
def test_no_comdat_folding(self): |
| 422 |
# Issue 8847: In the PGO build, the MSVC linker's COMDAT folding |
| 423 |
# optimization causes failures in code that relies on distinct |
| 424 |
# function addresses. |
| 425 |
class T(tuple): pass |
| 426 |
with self.assertRaises(TypeError): |
| 427 |
[3,] + T((1,2)) |
| 428 |
|
| 429 |
def test_lexicographic_ordering(self): |
| 430 |
# Issue 21100 |
| 431 |
a = self.type2test([1, 2]) |
| 432 |
b = self.type2test([1, 2, 0]) |
| 433 |
c = self.type2test([1, 3]) |
| 434 |
self.assertLess(a, b) |
| 435 |
self.assertLess(b, c) |
| 436 |
|
| 437 |
# Notes on testing hash codes. The primary thing is that Python doesn't |
| 438 |
# care about "random" hash codes. To the contrary, we like them to be |
| 439 |
# very regular when possible, so that the low-order bits are as evenly |
| 440 |
# distributed as possible. For integers this is easy: hash(i) == i for |
| 441 |
# all not-huge i except i==-1. |
| 442 |
# |
| 443 |
# For tuples of mixed type there's really no hope of that, so we want |
| 444 |
# "randomish" here instead. But getting close to pseudo-random in all |
| 445 |
# bit positions is more expensive than we've been willing to pay for. |
| 446 |
# |
| 447 |
# We can tolerate large deviations from random - what we don't want is |
| 448 |
# catastrophic pileups on a relative handful of hash codes. The dict |
| 449 |
# and set lookup routines remain effective provided that full-width hash |
| 450 |
# codes for not-equal objects are distinct. |
| 451 |
# |
| 452 |
# So we compute various statistics here based on what a "truly random" |
| 453 |
# hash would do, but don't automate "pass or fail" based on those |
| 454 |
# results. Instead those are viewed as inputs to human judgment, and the |
| 455 |
# automated tests merely ensure we get the _same_ results across |
| 456 |
# platforms. In fact, we normally don't bother to run them at all - |
| 457 |
# set RUN_ALL_HASH_TESTS to force it. |
| 458 |
# |
| 459 |
# When global JUST_SHOW_HASH_RESULTS is True, the tuple hash statistics |
| 460 |
# are just displayed to stdout. A typical output line looks like: |
| 461 |
# |
| 462 |
# old tuple test; 32-bit upper hash codes; \ |
| 463 |
# pileup 49 mean 7.4 coll 52 z +16.4 |
| 464 |
# |
| 465 |
# "old tuple test" is just a string name for the test being run. |
| 466 |
# |
| 467 |
# "32-bit upper hash codes" means this was run under a 64-bit build and |
| 468 |
# we've shifted away the lower 32 bits of the hash codes. |
| 469 |
# |
| 470 |
# "pileup" is 0 if there were no collisions across those hash codes. |
| 471 |
# It's 1 less than the maximum number of times any single hash code was |
| 472 |
# seen. So in this case, there was (at least) one hash code that was |
| 473 |
# seen 50 times: that hash code "piled up" 49 more times than ideal. |
| 474 |
# |
| 475 |
# "mean" is the number of collisions a perfectly random hash function |
| 476 |
# would have yielded, on average. |
| 477 |
# |
| 478 |
# "coll" is the number of collisions actually seen. |
| 479 |
# |
| 480 |
# "z" is "coll - mean" divided by the standard deviation of the number |
| 481 |
# of collisions a perfectly random hash function would suffer. A |
| 482 |
# positive value is "worse than random", and negative value "better than |
| 483 |
# random". Anything of magnitude greater than 3 would be highly suspect |
| 484 |
# for a hash function that claimed to be random. It's essentially |
| 485 |
# impossible that a truly random function would deliver a result 16.4 |
| 486 |
# sdevs "worse than random". |
| 487 |
# |
| 488 |
# But we don't care here! That's why the test isn't coded to fail. |
| 489 |
# Knowing something about how the high-order hash code bits behave |
| 490 |
# provides insight, but is irrelevant to how the dict and set lookup |
| 491 |
# code performs. The low-order bits are much more important to that, |
| 492 |
# and on the same test those did "just like random": |
| 493 |
# |
| 494 |
# old tuple test; 32-bit lower hash codes; \ |
| 495 |
# pileup 1 mean 7.4 coll 7 z -0.2 |
| 496 |
# |
| 497 |
# So there are always tradeoffs to consider. For another: |
| 498 |
# |
| 499 |
# 0..99 << 60 by 3; 32-bit hash codes; \ |
| 500 |
# pileup 0 mean 116.4 coll 0 z -10.8 |
| 501 |
# |
| 502 |
# That was run under a 32-bit build, and is spectacularly "better than |
| 503 |
# random". On a 64-bit build the wider hash codes are fine too: |
| 504 |
# |
| 505 |
# 0..99 << 60 by 3; 64-bit hash codes; \ |
| 506 |
# pileup 0 mean 0.0 coll 0 z -0.0 |
| 507 |
# |
| 508 |
# but their lower 32 bits are poor: |
| 509 |
# |
| 510 |
# 0..99 << 60 by 3; 32-bit lower hash codes; \ |
| 511 |
# pileup 1 mean 116.4 coll 324 z +19.2 |
| 512 |
# |
| 513 |
# In a statistical sense that's waaaaay too many collisions, but (a) 324 |
| 514 |
# collisions out of a million hash codes isn't anywhere near being a |
| 515 |
# real problem; and, (b) the worst pileup on a single hash code is a measly |
| 516 |
# 1 extra. It's a relatively poor case for the tuple hash, but still |
| 517 |
# fine for practical use. |
| 518 |
# |
| 519 |
# This isn't, which is what Python 3.7.1 produced for the hashes of |
| 520 |
# itertools.product([0, 0.5], repeat=18). Even with a fat 64-bit |
| 521 |
# hashcode, the highest pileup was over 16,000 - making a dict/set |
| 522 |
# lookup on one of the colliding values thousands of times slower (on |
| 523 |
# average) than we expect. |
| 524 |
# |
| 525 |
# [0, 0.5] by 18; 64-bit hash codes; \ |
| 526 |
# pileup 16,383 mean 0.0 coll 262,128 z +6073641856.9 |
| 527 |
# [0, 0.5] by 18; 32-bit lower hash codes; \ |
| 528 |
# pileup 262,143 mean 8.0 coll 262,143 z +92683.6 |
| 529 |
|
| 530 |
if __name__ == "__main__": |
| 531 |
unittest.main() |