Repository navigation
Arithmetic comparison errors involving sys.maxsize #1448
Description
Micro Python v1.4.4-251-g21ffa7c on 2015-08-30; PYBv1.0 with STM32F405RG
Type "help()" for more information.
>>> import sys
>>> sys.maxsize == sys.maxsize - 1 + 1 # expect True
False
>>> sys.maxsize > sys.maxsize - 1 + 1 # expect False
TrueThe error occurs within sys.maxsize and 2 * sys.maxsize + 1:
>>> sys.maxsize - 1 == sys.maxsize - 2 + 1 # expect True
True
>>> 2 * sys.maxsize + 1 == 2 * sys.maxsize - 1 + 1 + 1 # expect True
False
>>> 2 * sys.maxsize + 2 == 2 * sys.maxsize - 1 + 1 + 2 # expect True
TrueYet everywhere equivalence is expected, the string representations are equal:
>>> str(sys.maxsize) == str(sys.maxsize - 1 + 1) # expect True
True
>>> str(2 * sys.maxsize + 1) == str(2 * sys.maxsize - 1 + 1 + 1) # expect True
TrueThe unix build behaves the same.
Note that both stmhal/mpconfigport.h and unix/mpconfigport.h have:
#define MICROPY_LONGINT_IMPL (MICROPY_LONGINT_IMPL_MPZ)
...
#define MICROPY_PY_SYS_MAXSIZE (1)which puts us in objint_mpz.c. I had no luck finding odd behavior around bit crossings using:
def test(n):
x = 1
for i in range(n):
x += x+1
if x == sys.maxsize:
print("arithmetic == at", i+1, x, hex(x))
if hex(x) == hex(sys.maxsize):
print("hex() == at", i+1, x, hex(x))
yield x-1 == x-1 - 1 + 1
yield x == x - 1 + 1
yield x+1 == x+1 - 1 + 1On pyboard:
>>> all(test(2050))
hex() == at 30 2147483647 0x7fffffff
TrueAnd unix:
>>> all(test(2050))
hex() == at 62 9223372036854775807 0x7fffffffffffffff
TrueCpython of course gives:
>>> all(test(2050))
arithmetic == at 62 9223372036854775807 0x7fffffffffffffff
hex() == at 62 9223372036854775807 0x7fffffffffffffff
True
Activity
dpgeorge commentedon Aug 31, 2015
Nice find! The problem is in objint_mpz.c, in the definition of mp_maxsize_obj. I think the .len field is wrong (line 61). Try making .len=2 on 64-bit unix.
pramasoul commentedon Aug 31, 2015
Making .len=2 on 64-bit unix does indeed give correct behavior. What preprocessor expression should I use in place of
MPZ_NUM_DIG_FOR_INTto give the correct value across platforms?dpgeorge commentedon Sep 1, 2015
I don't know, it has to be worked out. Maybe the definition of MPZ_NUM_DIG_FOR_INT can be fixed.
hosaka commentedon Sep 15, 2015
I think the problem with the MPZ_NUM_DIG_FOR_INT macro was the +1 in it's definition. Removing it makes the number of digits stay at 2 for 32/64 bit unix. Let me know if it's okay to open a pull req.
py/mpz: Fix calculation of max digit storage for mpz; fix sys.maxsize.
dpgeorge commentedon Sep 15, 2015
Fixed in 8b4fb4f.
@Alex-March I just beat you to it! In this case the calculation was trying to round the division up, so the correct fix is (+ MPZ_DIG_SIZE-1) before the division. Also I fixed the definition of sys.maxsize to make it more robust.
Yes, you are always welcome to open a PR :)
hosaka commentedon Sep 15, 2015
@dpgeorge good timing, thanks!
py/mpz: Fix calculation of max digit storage for mpz; fix sys.maxsize.
Merge pull request micropython#1448 from dhalbert/sphinx-fixes