• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# 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
9from test import test_support
10
11if getattr(sys, 'float_repr_style', '') != 'short':
12    raise unittest.SkipTest('correctly-rounded string->float conversions '
13                            'not available on this system')
14
15# Correctly rounded str -> float in pure Python, for comparison.
16
17strtod_parser = re.compile(r"""    # A numeric string consists of:
18    (?P<sign>[-+])?          # an optional sign, followed by
19    (?=\d|\.\d)              # a number with at least one digit
20    (?P<int>\d*)             # having a (possibly empty) integer part
21    (?:\.(?P<frac>\d*))?     # followed by an optional fractional part
22    (?:E(?P<exp>[-+]?\d+))?  # and an optional exponent
23    \Z
24""", re.VERBOSE | re.IGNORECASE).match
25
26# Pure Python version of correctly rounded string->float conversion.
27# Avoids any use of floating-point by returning the result as a hex string.
28def strtod(s, mant_dig=53, min_exp = -1021, max_exp = 1024):
29    """Convert a finite decimal string to a hex string representing an
30    IEEE 754 binary64 float.  Return 'inf' or '-inf' on overflow.
31    This function makes no use of floating-point arithmetic at any
32    stage."""
33
34    # parse string into a pair of integers 'a' and 'b' such that
35    # abs(decimal value) = a/b, along with a boolean 'negative'.
36    m = strtod_parser(s)
37    if m is None:
38        raise ValueError('invalid numeric string')
39    fraction = m.group('frac') or ''
40    intpart = int(m.group('int') + fraction)
41    exp = int(m.group('exp') or '0') - len(fraction)
42    negative = m.group('sign') == '-'
43    a, b = intpart*10**max(exp, 0), 10**max(0, -exp)
44
45    # quick return for zeros
46    if not a:
47        return '-0x0.0p+0' if negative else '0x0.0p+0'
48
49    # compute exponent e for result; may be one too small in the case
50    # that the rounded value of a/b lies in a different binade from a/b
51    d = a.bit_length() - b.bit_length()
52    d += (a >> d if d >= 0 else a << -d) >= b
53    e = max(d, min_exp) - mant_dig
54
55    # approximate a/b by number of the form q * 2**e; adjust e if necessary
56    a, b = a << max(-e, 0), b << max(e, 0)
57    q, r = divmod(a, b)
58    if 2*r > b or 2*r == b and q & 1:
59        q += 1
60        if q.bit_length() == mant_dig+1:
61            q //= 2
62            e += 1
63
64    # double check that (q, e) has the right form
65    assert q.bit_length() <= mant_dig and e >= min_exp - mant_dig
66    assert q.bit_length() == mant_dig or e == min_exp - mant_dig
67
68    # check for overflow and underflow
69    if e + q.bit_length() > max_exp:
70        return '-inf' if negative else 'inf'
71    if not q:
72        return '-0x0.0p+0' if negative else '0x0.0p+0'
73
74    # for hex representation, shift so # bits after point is a multiple of 4
75    hexdigs = 1 + (mant_dig-2)//4
76    shift = 3 - (mant_dig-2)%4
77    q, e = q << shift, e - shift
78    return '{}0x{:x}.{:0{}x}p{:+d}'.format(
79        '-' if negative else '',
80        q // 16**hexdigs,
81        q % 16**hexdigs,
82        hexdigs,
83        e + 4*hexdigs)
84
85TEST_SIZE = 10
86
87class StrtodTests(unittest.TestCase):
88    def check_strtod(self, s):
89        """Compare the result of Python's builtin correctly rounded
90        string->float conversion (using float) to a pure Python
91        correctly rounded string->float implementation.  Fail if the
92        two methods give different results."""
93
94        try:
95            fs = float(s)
96        except OverflowError:
97            got = '-inf' if s[0] == '-' else 'inf'
98        except MemoryError:
99            got = 'memory error'
100        else:
101            got = fs.hex()
102        expected = strtod(s)
103        self.assertEqual(expected, got,
104                         "Incorrectly rounded str->float conversion for {}: "
105                         "expected {}, got {}".format(s, expected, got))
106
107    def test_short_halfway_cases(self):
108        # exact halfway cases with a small number of significant digits
109        for k in 0, 5, 10, 15, 20:
110            # upper = smallest integer >= 2**54/5**k
111            upper = -(-2**54//5**k)
112            # lower = smallest odd number >= 2**53/5**k
113            lower = -(-2**53//5**k)
114            if lower % 2 == 0:
115                lower += 1
116            for i in xrange(TEST_SIZE):
117                # Select a random odd n in [2**53/5**k,
118                # 2**54/5**k). Then n * 10**k gives a halfway case
119                # with small number of significant digits.
120                n, e = random.randrange(lower, upper, 2), k
121
122                # Remove any additional powers of 5.
123                while n % 5 == 0:
124                    n, e = n // 5, e + 1
125                assert n % 10 in (1, 3, 7, 9)
126
127                # Try numbers of the form n * 2**p2 * 10**e, p2 >= 0,
128                # until n * 2**p2 has more than 20 significant digits.
129                digits, exponent = n, e
130                while digits < 10**20:
131                    s = '{}e{}'.format(digits, exponent)
132                    self.check_strtod(s)
133                    # Same again, but with extra trailing zeros.
134                    s = '{}e{}'.format(digits * 10**40, exponent - 40)
135                    self.check_strtod(s)
136                    digits *= 2
137
138                # Try numbers of the form n * 5**p2 * 10**(e - p5), p5
139                # >= 0, with n * 5**p5 < 10**20.
140                digits, exponent = n, e
141                while digits < 10**20:
142                    s = '{}e{}'.format(digits, exponent)
143                    self.check_strtod(s)
144                    # Same again, but with extra trailing zeros.
145                    s = '{}e{}'.format(digits * 10**40, exponent - 40)
146                    self.check_strtod(s)
147                    digits *= 5
148                    exponent -= 1
149
150    def test_halfway_cases(self):
151        # test halfway cases for the round-half-to-even rule
152        for i in xrange(100 * TEST_SIZE):
153            # bit pattern for a random finite positive (or +0.0) float
154            bits = random.randrange(2047*2**52)
155
156            # convert bit pattern to a number of the form m * 2**e
157            e, m = divmod(bits, 2**52)
158            if e:
159                m, e = m + 2**52, e - 1
160            e -= 1074
161
162            # add 0.5 ulps
163            m, e = 2*m + 1, e - 1
164
165            # convert to a decimal string
166            if e >= 0:
167                digits = m << e
168                exponent = 0
169            else:
170                # m * 2**e = (m * 5**-e) * 10**e
171                digits = m * 5**-e
172                exponent = e
173            s = '{}e{}'.format(digits, exponent)
174            self.check_strtod(s)
175
176    def test_boundaries(self):
177        # boundaries expressed as triples (n, e, u), where
178        # n*10**e is an approximation to the boundary value and
179        # u*10**e is 1ulp
180        boundaries = [
181            (10000000000000000000, -19, 1110),   # a power of 2 boundary (1.0)
182            (17976931348623159077, 289, 1995),   # overflow boundary (2.**1024)
183            (22250738585072013831, -327, 4941),  # normal/subnormal (2.**-1022)
184            (0, -327, 4941),                     # zero
185            ]
186        for n, e, u in boundaries:
187            for j in xrange(1000):
188                digits = n + random.randrange(-3*u, 3*u)
189                exponent = e
190                s = '{}e{}'.format(digits, exponent)
191                self.check_strtod(s)
192                n *= 10
193                u *= 10
194                e -= 1
195
196    def test_underflow_boundary(self):
197        # test values close to 2**-1075, the underflow boundary; similar
198        # to boundary_tests, except that the random error doesn't scale
199        # with n
200        for exponent in xrange(-400, -320):
201            base = 10**-exponent // 2**1075
202            for j in xrange(TEST_SIZE):
203                digits = base + random.randrange(-1000, 1000)
204                s = '{}e{}'.format(digits, exponent)
205                self.check_strtod(s)
206
207    def test_bigcomp(self):
208        for ndigs in 5, 10, 14, 15, 16, 17, 18, 19, 20, 40, 41, 50:
209            dig10 = 10**ndigs
210            for i in xrange(10 * TEST_SIZE):
211                digits = random.randrange(dig10)
212                exponent = random.randrange(-400, 400)
213                s = '{}e{}'.format(digits, exponent)
214                self.check_strtod(s)
215
216    def test_parsing(self):
217        # make '0' more likely to be chosen than other digits
218        digits = '000000123456789'
219        signs = ('+', '-', '')
220
221        # put together random short valid strings
222        # \d*[.\d*]?e
223        for i in xrange(1000):
224            for j in xrange(TEST_SIZE):
225                s = random.choice(signs)
226                intpart_len = random.randrange(5)
227                s += ''.join(random.choice(digits) for _ in xrange(intpart_len))
228                if random.choice([True, False]):
229                    s += '.'
230                    fracpart_len = random.randrange(5)
231                    s += ''.join(random.choice(digits)
232                                 for _ in xrange(fracpart_len))
233                else:
234                    fracpart_len = 0
235                if random.choice([True, False]):
236                    s += random.choice(['e', 'E'])
237                    s += random.choice(signs)
238                    exponent_len = random.randrange(1, 4)
239                    s += ''.join(random.choice(digits)
240                                 for _ in xrange(exponent_len))
241
242                if intpart_len + fracpart_len:
243                    self.check_strtod(s)
244                else:
245                    try:
246                        float(s)
247                    except ValueError:
248                        pass
249                    else:
250                        assert False, "expected ValueError"
251
252    @test_support.precisionbigmemtest(size=test_support._2G, memuse=3,
253                                      dry_run=False)
254    def test_oversized_digit_strings(self, maxsize):
255        # Input string whose length doesn't fit in an INT.
256        s = "1." + "1" * int(2.2e9)
257        with self.assertRaises(ValueError):
258            float(s)
259        del s
260
261        s = "0." + "0" * int(2.2e9) + "1"
262        with self.assertRaises(ValueError):
263            float(s)
264        del s
265
266    def test_large_exponents(self):
267        # Verify that the clipping of the exponent in strtod doesn't affect the
268        # output values.
269        def positive_exp(n):
270            """ Long string with value 1.0 and exponent n"""
271            return '0.{}1e+{}'.format('0'*(n-1), n)
272
273        def negative_exp(n):
274            """ Long string with value 1.0 and exponent -n"""
275            return '1{}e-{}'.format('0'*n, n)
276
277        self.assertEqual(float(positive_exp(10000)), 1.0)
278        self.assertEqual(float(positive_exp(20000)), 1.0)
279        self.assertEqual(float(positive_exp(30000)), 1.0)
280        self.assertEqual(float(negative_exp(10000)), 1.0)
281        self.assertEqual(float(negative_exp(20000)), 1.0)
282        self.assertEqual(float(negative_exp(30000)), 1.0)
283
284    def test_particular(self):
285        # inputs that produced crashes or incorrectly rounded results with
286        # previous versions of dtoa.c, for various reasons
287        test_strings = [
288            # issue 7632 bug 1, originally reported failing case
289            '2183167012312112312312.23538020374420446192e-370',
290            # 5 instances of issue 7632 bug 2
291            '12579816049008305546974391768996369464963024663104e-357',
292            '17489628565202117263145367596028389348922981857013e-357',
293            '18487398785991994634182916638542680759613590482273e-357',
294            '32002864200581033134358724675198044527469366773928e-358',
295            '94393431193180696942841837085033647913224148539854e-358',
296            '73608278998966969345824653500136787876436005957953e-358',
297            '64774478836417299491718435234611299336288082136054e-358',
298            '13704940134126574534878641876947980878824688451169e-357',
299            '46697445774047060960624497964425416610480524760471e-358',
300            # failing case for bug introduced by METD in r77451 (attempted
301            # fix for issue 7632, bug 2), and fixed in r77482.
302            '28639097178261763178489759107321392745108491825303e-311',
303            # two numbers demonstrating a flaw in the bigcomp 'dig == 0'
304            # correction block (issue 7632, bug 3)
305            '1.00000000000000001e44',
306            '1.0000000000000000100000000000000000000001e44',
307            # dtoa.c bug for numbers just smaller than a power of 2 (issue
308            # 7632, bug 4)
309            '99999999999999994487665465554760717039532578546e-47',
310            # failing case for off-by-one error introduced by METD in
311            # r77483 (dtoa.c cleanup), fixed in r77490
312            '965437176333654931799035513671997118345570045914469' #...
313            '6213413350821416312194420007991306908470147322020121018368e0',
314            # incorrect lsb detection for round-half-to-even when
315            # bc->scale != 0 (issue 7632, bug 6).
316            '104308485241983990666713401708072175773165034278685' #...
317            '682646111762292409330928739751702404658197872319129' #...
318            '036519947435319418387839758990478549477777586673075' #...
319            '945844895981012024387992135617064532141489278815239' #...
320            '849108105951619997829153633535314849999674266169258' #...
321            '928940692239684771590065027025835804863585454872499' #...
322            '320500023126142553932654370362024104462255244034053' #...
323            '203998964360882487378334860197725139151265590832887' #...
324            '433736189468858614521708567646743455601905935595381' #...
325            '852723723645799866672558576993978025033590728687206' #...
326            '296379801363024094048327273913079612469982585674824' #...
327            '156000783167963081616214710691759864332339239688734' #...
328            '656548790656486646106983450809073750535624894296242' #...
329            '072010195710276073042036425579852459556183541199012' #...
330            '652571123898996574563824424330960027873516082763671875e-1075',
331            # demonstration that original fix for issue 7632 bug 1 was
332            # buggy; the exit condition was too strong
333            '247032822920623295e-341',
334            # demonstrate similar problem to issue 7632 bug1: crash
335            # with 'oversized quotient in quorem' message.
336            '99037485700245683102805043437346965248029601286431e-373',
337            '99617639833743863161109961162881027406769510558457e-373',
338            '98852915025769345295749278351563179840130565591462e-372',
339            '99059944827693569659153042769690930905148015876788e-373',
340            '98914979205069368270421829889078356254059760327101e-372',
341            # issue 7632 bug 5: the following 2 strings convert differently
342            '1000000000000000000000000000000000000000e-16',
343            '10000000000000000000000000000000000000000e-17',
344            # issue 7632 bug 7
345            '991633793189150720000000000000000000000000000000000000000e-33',
346            # And another, similar, failing halfway case
347            '4106250198039490000000000000000000000000000000000000000e-38',
348            # issue 7632 bug 8:  the following produced 10.0
349            '10.900000000000000012345678912345678912345',
350
351            # two humongous values from issue 7743
352            '116512874940594195638617907092569881519034793229385' #...
353            '228569165191541890846564669771714896916084883987920' #...
354            '473321268100296857636200926065340769682863349205363' #...
355            '349247637660671783209907949273683040397979984107806' #...
356            '461822693332712828397617946036239581632976585100633' #...
357            '520260770761060725403904123144384571612073732754774' #...
358            '588211944406465572591022081973828448927338602556287' #...
359            '851831745419397433012491884869454462440536895047499' #...
360            '436551974649731917170099387762871020403582994193439' #...
361            '761933412166821484015883631622539314203799034497982' #...
362            '130038741741727907429575673302461380386596501187482' #...
363            '006257527709842179336488381672818798450229339123527' #...
364            '858844448336815912020452294624916993546388956561522' #...
365            '161875352572590420823607478788399460162228308693742' #...
366            '05287663441403533948204085390898399055004119873046875e-1075',
367
368            '525440653352955266109661060358202819561258984964913' #...
369            '892256527849758956045218257059713765874251436193619' #...
370            '443248205998870001633865657517447355992225852945912' #...
371            '016668660000210283807209850662224417504752264995360' #...
372            '631512007753855801075373057632157738752800840302596' #...
373            '237050247910530538250008682272783660778181628040733' #...
374            '653121492436408812668023478001208529190359254322340' #...
375            '397575185248844788515410722958784640926528544043090' #...
376            '115352513640884988017342469275006999104519620946430' #...
377            '818767147966495485406577703972687838176778993472989' #...
378            '561959000047036638938396333146685137903018376496408' #...
379            '319705333868476925297317136513970189073693314710318' #...
380            '991252811050501448326875232850600451776091303043715' #...
381            '157191292827614046876950225714743118291034780466325' #...
382            '085141343734564915193426994587206432697337118211527' #...
383            '278968731294639353354774788602467795167875117481660' #...
384            '4738791256853675690543663283782215866825e-1180',
385
386            # exercise exit conditions in bigcomp comparison loop
387            '2602129298404963083833853479113577253105939995688e2',
388            '260212929840496308383385347911357725310593999568896e0',
389            '26021292984049630838338534791135772531059399956889601e-2',
390            '260212929840496308383385347911357725310593999568895e0',
391            '260212929840496308383385347911357725310593999568897e0',
392            '260212929840496308383385347911357725310593999568996e0',
393            '260212929840496308383385347911357725310593999568866e0',
394            # 2**53
395            '9007199254740992.00',
396            # 2**1024 - 2**970:  exact overflow boundary.  All values
397            # smaller than this should round to something finite;  any value
398            # greater than or equal to this one overflows.
399            '179769313486231580793728971405303415079934132710037' #...
400            '826936173778980444968292764750946649017977587207096' #...
401            '330286416692887910946555547851940402630657488671505' #...
402            '820681908902000708383676273854845817711531764475730' #...
403            '270069855571366959622842914819860834936475292719074' #...
404            '168444365510704342711559699508093042880177904174497792',
405            # 2**1024 - 2**970 - tiny
406            '179769313486231580793728971405303415079934132710037' #...
407            '826936173778980444968292764750946649017977587207096' #...
408            '330286416692887910946555547851940402630657488671505' #...
409            '820681908902000708383676273854845817711531764475730' #...
410            '270069855571366959622842914819860834936475292719074' #...
411            '168444365510704342711559699508093042880177904174497791.999',
412            # 2**1024 - 2**970 + tiny
413            '179769313486231580793728971405303415079934132710037' #...
414            '826936173778980444968292764750946649017977587207096' #...
415            '330286416692887910946555547851940402630657488671505' #...
416            '820681908902000708383676273854845817711531764475730' #...
417            '270069855571366959622842914819860834936475292719074' #...
418            '168444365510704342711559699508093042880177904174497792.001',
419            # 1 - 2**-54, +-tiny
420            '999999999999999944488848768742172978818416595458984375e-54',
421            '9999999999999999444888487687421729788184165954589843749999999e-54',
422            '9999999999999999444888487687421729788184165954589843750000001e-54',
423            ]
424        for s in test_strings:
425            self.check_strtod(s)
426
427def test_main():
428    test_support.run_unittest(StrtodTests)
429
430if __name__ == "__main__":
431    test_main()
432