blob: e39558a2f6048d71f3271c404c04e1c79103d7cb [file] [log] [blame]
Mark Dickinson9000c162010-01-14 15:43:57 +00001# Tests for the correctly-rounded string -> float conversions
2# introduced in Python 2.7 and 3.1.
3
4import random
5import struct
6import unittest
7import re
8import sys
9import test.support
10
11# Correctly rounded str -> float in pure Python, for comparison.
12
13strtod_parser = re.compile(r""" # A numeric string consists of:
14 (?P<sign>[-+])? # an optional sign, followed by
15 (?=\d|\.\d) # a number with at least one digit
16 (?P<int>\d*) # having a (possibly empty) integer part
17 (?:\.(?P<frac>\d*))? # followed by an optional fractional part
18 (?:E(?P<exp>[-+]?\d+))? # and an optional exponent
19 \Z
20""", re.VERBOSE | re.IGNORECASE).match
21
22def strtod(s, mant_dig=53, min_exp = -1021, max_exp = 1024):
23 """Convert a finite decimal string to a hex string representing an
24 IEEE 754 binary64 float. Return 'inf' or '-inf' on overflow.
25 This function makes no use of floating-point arithmetic at any
26 stage."""
27
28 # parse string into a pair of integers 'a' and 'b' such that
29 # abs(decimal value) = a/b, along with a boolean 'negative'.
30 m = strtod_parser(s)
31 if m is None:
32 raise ValueError('invalid numeric string')
33 fraction = m.group('frac') or ''
34 intpart = int(m.group('int') + fraction)
35 exp = int(m.group('exp') or '0') - len(fraction)
36 negative = m.group('sign') == '-'
37 a, b = intpart*10**max(exp, 0), 10**max(0, -exp)
38
39 # quick return for zeros
40 if not a:
41 return '-0x0.0p+0' if negative else '0x0.0p+0'
42
43 # compute exponent e for result; may be one too small in the case
44 # that the rounded value of a/b lies in a different binade from a/b
45 d = a.bit_length() - b.bit_length()
46 d += (a >> d if d >= 0 else a << -d) >= b
47 e = max(d, min_exp) - mant_dig
48
49 # approximate a/b by number of the form q * 2**e; adjust e if necessary
50 a, b = a << max(-e, 0), b << max(e, 0)
51 q, r = divmod(a, b)
52 if 2*r > b or 2*r == b and q & 1:
53 q += 1
54 if q.bit_length() == mant_dig+1:
55 q //= 2
56 e += 1
57
58 # double check that (q, e) has the right form
59 assert q.bit_length() <= mant_dig and e >= min_exp - mant_dig
60 assert q.bit_length() == mant_dig or e == min_exp - mant_dig
61
62 # check for overflow and underflow
63 if e + q.bit_length() > max_exp:
64 return '-inf' if negative else 'inf'
65 if not q:
66 return '-0x0.0p+0' if negative else '0x0.0p+0'
67
68 # for hex representation, shift so # bits after point is a multiple of 4
69 hexdigs = 1 + (mant_dig-2)//4
70 shift = 3 - (mant_dig-2)%4
71 q, e = q << shift, e - shift
72 return '{}0x{:x}.{:0{}x}p{:+d}'.format(
73 '-' if negative else '',
74 q // 16**hexdigs,
75 q % 16**hexdigs,
76 hexdigs,
77 e + 4*hexdigs)
78
Mark Dickinson6328f412010-02-21 14:51:02 +000079TEST_SIZE = 10
Mark Dickinson9000c162010-01-14 15:43:57 +000080
81@unittest.skipUnless(getattr(sys, 'float_repr_style', '') == 'short',
82 "applies only when using short float repr style")
83class StrtodTests(unittest.TestCase):
84 def check_strtod(self, s):
85 """Compare the result of Python's builtin correctly rounded
86 string->float conversion (using float) to a pure Python
87 correctly rounded string->float implementation. Fail if the
88 two methods give different results."""
89
90 try:
91 fs = float(s)
92 except OverflowError:
93 got = '-inf' if s[0] == '-' else 'inf'
Mark Dickinson863d61d2010-01-16 20:35:50 +000094 except MemoryError:
95 got = 'memory error'
Mark Dickinson9000c162010-01-14 15:43:57 +000096 else:
97 got = fs.hex()
98 expected = strtod(s)
99 self.assertEqual(expected, got,
100 "Incorrectly rounded str->float conversion for {}: "
101 "expected {}, got {}".format(s, expected, got))
102
Mark Dickinsone42ffae2010-01-21 19:57:43 +0000103 def test_short_halfway_cases(self):
104 # exact halfway cases with a small number of significant digits
105 for k in 0, 5, 10, 15, 20:
106 # upper = smallest integer >= 2**54/5**k
107 upper = -(-2**54//5**k)
108 # lower = smallest odd number >= 2**53/5**k
109 lower = -(-2**53//5**k)
110 if lower % 2 == 0:
111 lower += 1
Mark Dickinson6328f412010-02-21 14:51:02 +0000112 for i in range(TEST_SIZE):
Mark Dickinsone42ffae2010-01-21 19:57:43 +0000113 # Select a random odd n in [2**53/5**k,
114 # 2**54/5**k). Then n * 10**k gives a halfway case
115 # with small number of significant digits.
116 n, e = random.randrange(lower, upper, 2), k
117
118 # Remove any additional powers of 5.
119 while n % 5 == 0:
120 n, e = n // 5, e + 1
121 assert n % 10 in (1, 3, 7, 9)
122
123 # Try numbers of the form n * 2**p2 * 10**e, p2 >= 0,
124 # until n * 2**p2 has more than 20 significant digits.
125 digits, exponent = n, e
126 while digits < 10**20:
127 s = '{}e{}'.format(digits, exponent)
128 self.check_strtod(s)
129 # Same again, but with extra trailing zeros.
130 s = '{}e{}'.format(digits * 10**40, exponent - 40)
131 self.check_strtod(s)
132 digits *= 2
133
134 # Try numbers of the form n * 5**p2 * 10**(e - p5), p5
135 # >= 0, with n * 5**p5 < 10**20.
136 digits, exponent = n, e
137 while digits < 10**20:
138 s = '{}e{}'.format(digits, exponent)
139 self.check_strtod(s)
140 # Same again, but with extra trailing zeros.
141 s = '{}e{}'.format(digits * 10**40, exponent - 40)
142 self.check_strtod(s)
143 digits *= 5
144 exponent -= 1
145
Mark Dickinson9000c162010-01-14 15:43:57 +0000146 def test_halfway_cases(self):
147 # test halfway cases for the round-half-to-even rule
Mark Dickinson6328f412010-02-21 14:51:02 +0000148 for i in range(100 * TEST_SIZE):
149 # bit pattern for a random finite positive (or +0.0) float
150 bits = random.randrange(2047*2**52)
Mark Dickinson9000c162010-01-14 15:43:57 +0000151
Mark Dickinson6328f412010-02-21 14:51:02 +0000152 # convert bit pattern to a number of the form m * 2**e
153 e, m = divmod(bits, 2**52)
154 if e:
155 m, e = m + 2**52, e - 1
156 e -= 1074
Mark Dickinson9000c162010-01-14 15:43:57 +0000157
Mark Dickinson6328f412010-02-21 14:51:02 +0000158 # add 0.5 ulps
159 m, e = 2*m + 1, e - 1
Mark Dickinson9000c162010-01-14 15:43:57 +0000160
Mark Dickinson6328f412010-02-21 14:51:02 +0000161 # convert to a decimal string
162 if e >= 0:
163 digits = m << e
164 exponent = 0
165 else:
166 # m * 2**e = (m * 5**-e) * 10**e
167 digits = m * 5**-e
168 exponent = e
169 s = '{}e{}'.format(digits, exponent)
170 self.check_strtod(s)
Mark Dickinson9000c162010-01-14 15:43:57 +0000171
172 def test_boundaries(self):
173 # boundaries expressed as triples (n, e, u), where
174 # n*10**e is an approximation to the boundary value and
175 # u*10**e is 1ulp
176 boundaries = [
177 (10000000000000000000, -19, 1110), # a power of 2 boundary (1.0)
178 (17976931348623159077, 289, 1995), # overflow boundary (2.**1024)
179 (22250738585072013831, -327, 4941), # normal/subnormal (2.**-1022)
180 (0, -327, 4941), # zero
181 ]
182 for n, e, u in boundaries:
183 for j in range(1000):
Mark Dickinson6328f412010-02-21 14:51:02 +0000184 digits = n + random.randrange(-3*u, 3*u)
185 exponent = e
186 s = '{}e{}'.format(digits, exponent)
187 self.check_strtod(s)
Mark Dickinson9000c162010-01-14 15:43:57 +0000188 n *= 10
189 u *= 10
190 e -= 1
191
192 def test_underflow_boundary(self):
193 # test values close to 2**-1075, the underflow boundary; similar
194 # to boundary_tests, except that the random error doesn't scale
195 # with n
196 for exponent in range(-400, -320):
197 base = 10**-exponent // 2**1075
198 for j in range(TEST_SIZE):
199 digits = base + random.randrange(-1000, 1000)
200 s = '{}e{}'.format(digits, exponent)
201 self.check_strtod(s)
202
203 def test_bigcomp(self):
Mark Dickinsone42ffae2010-01-21 19:57:43 +0000204 for ndigs in 5, 10, 14, 15, 16, 17, 18, 19, 20, 40, 41, 50:
205 dig10 = 10**ndigs
Mark Dickinson6328f412010-02-21 14:51:02 +0000206 for i in range(10 * TEST_SIZE):
Mark Dickinsone42ffae2010-01-21 19:57:43 +0000207 digits = random.randrange(dig10)
Mark Dickinson9000c162010-01-14 15:43:57 +0000208 exponent = random.randrange(-400, 400)
209 s = '{}e{}'.format(digits, exponent)
210 self.check_strtod(s)
211
212 def test_parsing(self):
Mark Dickinson11f65782010-01-16 18:12:46 +0000213 # make '0' more likely to be chosen than other digits
214 digits = '000000123456789'
Mark Dickinson9000c162010-01-14 15:43:57 +0000215 signs = ('+', '-', '')
216
217 # put together random short valid strings
218 # \d*[.\d*]?e
219 for i in range(1000):
220 for j in range(TEST_SIZE):
221 s = random.choice(signs)
222 intpart_len = random.randrange(5)
223 s += ''.join(random.choice(digits) for _ in range(intpart_len))
224 if random.choice([True, False]):
225 s += '.'
226 fracpart_len = random.randrange(5)
227 s += ''.join(random.choice(digits)
228 for _ in range(fracpart_len))
229 else:
230 fracpart_len = 0
231 if random.choice([True, False]):
232 s += random.choice(['e', 'E'])
233 s += random.choice(signs)
234 exponent_len = random.randrange(1, 4)
235 s += ''.join(random.choice(digits)
236 for _ in range(exponent_len))
237
238 if intpart_len + fracpart_len:
239 self.check_strtod(s)
240 else:
241 try:
242 float(s)
243 except ValueError:
244 pass
245 else:
246 assert False, "expected ValueError"
247
248 def test_particular(self):
249 # inputs that produced crashes or incorrectly rounded results with
250 # previous versions of dtoa.c, for various reasons
251 test_strings = [
252 # issue 7632 bug 1, originally reported failing case
253 '2183167012312112312312.23538020374420446192e-370',
254 # 5 instances of issue 7632 bug 2
255 '12579816049008305546974391768996369464963024663104e-357',
256 '17489628565202117263145367596028389348922981857013e-357',
257 '18487398785991994634182916638542680759613590482273e-357',
258 '32002864200581033134358724675198044527469366773928e-358',
259 '94393431193180696942841837085033647913224148539854e-358',
260 # failing case for bug introduced by METD in r77451 (attempted
261 # fix for issue 7632, bug 2), and fixed in r77482.
262 '28639097178261763178489759107321392745108491825303e-311',
263 # two numbers demonstrating a flaw in the bigcomp 'dig == 0'
264 # correction block (issue 7632, bug 3)
265 '1.00000000000000001e44',
266 '1.0000000000000000100000000000000000000001e44',
267 # dtoa.c bug for numbers just smaller than a power of 2 (issue
268 # 7632, bug 4)
269 '99999999999999994487665465554760717039532578546e-47',
270 # failing case for off-by-one error introduced by METD in
271 # r77483 (dtoa.c cleanup), fixed in r77490
272 '965437176333654931799035513671997118345570045914469' #...
273 '6213413350821416312194420007991306908470147322020121018368e0',
274 # incorrect lsb detection for round-half-to-even when
275 # bc->scale != 0 (issue 7632, bug 6).
276 '104308485241983990666713401708072175773165034278685' #...
277 '682646111762292409330928739751702404658197872319129' #...
278 '036519947435319418387839758990478549477777586673075' #...
279 '945844895981012024387992135617064532141489278815239' #...
280 '849108105951619997829153633535314849999674266169258' #...
281 '928940692239684771590065027025835804863585454872499' #...
282 '320500023126142553932654370362024104462255244034053' #...
283 '203998964360882487378334860197725139151265590832887' #...
284 '433736189468858614521708567646743455601905935595381' #...
285 '852723723645799866672558576993978025033590728687206' #...
286 '296379801363024094048327273913079612469982585674824' #...
287 '156000783167963081616214710691759864332339239688734' #...
288 '656548790656486646106983450809073750535624894296242' #...
289 '072010195710276073042036425579852459556183541199012' #...
290 '652571123898996574563824424330960027873516082763671875e-1075',
291 # demonstration that original fix for issue 7632 bug 1 was
292 # buggy; the exit condition was too strong
293 '247032822920623295e-341',
Mark Dickinsone42ffae2010-01-21 19:57:43 +0000294 # demonstrate similar problem to issue 7632 bug1: crash
295 # with 'oversized quotient in quorem' message.
296 '99037485700245683102805043437346965248029601286431e-373',
297 '99617639833743863161109961162881027406769510558457e-373',
298 '98852915025769345295749278351563179840130565591462e-372',
299 '99059944827693569659153042769690930905148015876788e-373',
300 '98914979205069368270421829889078356254059760327101e-372',
Mark Dickinson9000c162010-01-14 15:43:57 +0000301 # issue 7632 bug 5: the following 2 strings convert differently
302 '1000000000000000000000000000000000000000e-16',
Mark Dickinson11f65782010-01-16 18:12:46 +0000303 '10000000000000000000000000000000000000000e-17',
Mark Dickinsone42ffae2010-01-21 19:57:43 +0000304 # issue 7632 bug 7
305 '991633793189150720000000000000000000000000000000000000000e-33',
306 # And another, similar, failing halfway case
307 '4106250198039490000000000000000000000000000000000000000e-38',
Mark Dickinson11f65782010-01-16 18:12:46 +0000308 # issue 7632 bug 8: the following produced 10.0
309 '10.900000000000000012345678912345678912345',
Benjamin Peterson23b9ef72010-02-03 02:43:37 +0000310
311 # two humongous values from issue 7743
312 '116512874940594195638617907092569881519034793229385' #...
313 '228569165191541890846564669771714896916084883987920' #...
314 '473321268100296857636200926065340769682863349205363' #...
315 '349247637660671783209907949273683040397979984107806' #...
316 '461822693332712828397617946036239581632976585100633' #...
317 '520260770761060725403904123144384571612073732754774' #...
318 '588211944406465572591022081973828448927338602556287' #...
319 '851831745419397433012491884869454462440536895047499' #...
320 '436551974649731917170099387762871020403582994193439' #...
321 '761933412166821484015883631622539314203799034497982' #...
322 '130038741741727907429575673302461380386596501187482' #...
323 '006257527709842179336488381672818798450229339123527' #...
324 '858844448336815912020452294624916993546388956561522' #...
325 '161875352572590420823607478788399460162228308693742' #...
326 '05287663441403533948204085390898399055004119873046875e-1075',
327
328 '525440653352955266109661060358202819561258984964913' #...
329 '892256527849758956045218257059713765874251436193619' #...
330 '443248205998870001633865657517447355992225852945912' #...
331 '016668660000210283807209850662224417504752264995360' #...
332 '631512007753855801075373057632157738752800840302596' #...
333 '237050247910530538250008682272783660778181628040733' #...
334 '653121492436408812668023478001208529190359254322340' #...
335 '397575185248844788515410722958784640926528544043090' #...
336 '115352513640884988017342469275006999104519620946430' #...
337 '818767147966495485406577703972687838176778993472989' #...
338 '561959000047036638938396333146685137903018376496408' #...
339 '319705333868476925297317136513970189073693314710318' #...
340 '991252811050501448326875232850600451776091303043715' #...
341 '157191292827614046876950225714743118291034780466325' #...
342 '085141343734564915193426994587206432697337118211527' #...
343 '278968731294639353354774788602467795167875117481660' #...
344 '4738791256853675690543663283782215866825e-1180',
345
Mark Dickinsone42ffae2010-01-21 19:57:43 +0000346 # exercise exit conditions in bigcomp comparison loop
347 '2602129298404963083833853479113577253105939995688e2',
348 '260212929840496308383385347911357725310593999568896e0',
349 '26021292984049630838338534791135772531059399956889601e-2',
350 '260212929840496308383385347911357725310593999568895e0',
351 '260212929840496308383385347911357725310593999568897e0',
352 '260212929840496308383385347911357725310593999568996e0',
353 '260212929840496308383385347911357725310593999568866e0',
354 # 2**53
355 '9007199254740992.00',
356 # 2**1024 - 2**970: exact overflow boundary. All values
357 # smaller than this should round to something finite; any value
358 # greater than or equal to this one overflows.
359 '179769313486231580793728971405303415079934132710037' #...
360 '826936173778980444968292764750946649017977587207096' #...
361 '330286416692887910946555547851940402630657488671505' #...
362 '820681908902000708383676273854845817711531764475730' #...
363 '270069855571366959622842914819860834936475292719074' #...
364 '168444365510704342711559699508093042880177904174497792',
365 # 2**1024 - 2**970 - tiny
366 '179769313486231580793728971405303415079934132710037' #...
367 '826936173778980444968292764750946649017977587207096' #...
368 '330286416692887910946555547851940402630657488671505' #...
369 '820681908902000708383676273854845817711531764475730' #...
370 '270069855571366959622842914819860834936475292719074' #...
371 '168444365510704342711559699508093042880177904174497791.999',
372 # 2**1024 - 2**970 + tiny
373 '179769313486231580793728971405303415079934132710037' #...
374 '826936173778980444968292764750946649017977587207096' #...
375 '330286416692887910946555547851940402630657488671505' #...
376 '820681908902000708383676273854845817711531764475730' #...
377 '270069855571366959622842914819860834936475292719074' #...
378 '168444365510704342711559699508093042880177904174497792.001',
379 # 1 - 2**-54, +-tiny
380 '999999999999999944488848768742172978818416595458984375e-54',
381 '9999999999999999444888487687421729788184165954589843749999999e-54',
382 '9999999999999999444888487687421729788184165954589843750000001e-54',
Mark Dickinson9000c162010-01-14 15:43:57 +0000383 ]
384 for s in test_strings:
385 self.check_strtod(s)
386
387def test_main():
388 test.support.run_unittest(StrtodTests)
389
390if __name__ == "__main__":
391 test_main()