python 32-bit and 64-bit integer math with intentional overflow

Question

What's the best way to do integer math in 32- and 64-bit, so that overflow happens like it does in C?

e.g. (65536*65536+1)*(65536*65536+1) should be 0x0000000200000001 in 64-bit math, and not its exact value (non-overflowing) 0x10000000200000001.

Signed overflow causes undefined behaviour in C, so strictly speaking the question is meaningless. — Cairnarvon, May 25 '13 at 01:22
^ yet another reason C isn't helpful for certain types of numerical computation. :-( — Jason S, May 25 '13 at 04:32
Similar, but not a duplicate. The 32-bit question (or 16-bit, which I didn't ask about) is different from the int -> long int behavior. — Jason S, Jul 27 '15 at 13:42

score 33 · Accepted Answer · answered May 25 '13 at 00:56

33

Just & the result with the appropriate 32- or 64-bit mask (0xffffffff or 0xffffffffffffffff).

answered May 25 '13 at 00:56

martineau

With signed int you have to use modulo e.g. `% 0x100000000` or `% 0x10000000000000000`. Unfortunately this is not really efficient. Neither is the bitmask when compared to any language that supports native processor sized integers. NumPy exposes them so its probably the best solution still... – Gregory Morse Mar 16 '20 at 12:16

score 16 · Answer 2 · edited Apr 16 '16 at 05:18

16

Use NumPy with the appropriate integer size and the overflow is more C like:

32 bit:

>>> np.uint32(2**32-3) + np.uint32(5)
__main__:1: RuntimeWarning: overflow encountered in uint_scalars
2

64 bit:

>>> i64=np.uint64(65536*65536+1)
>>> hex(i64*i64)
'0x200000001L'

Compare with Python's native int:

>>> hex((65536*65536+1)*(65536*65536+1))
'0x10000000200000001L'

You can see that NumPy is doing as you desire.

edited Apr 16 '16 at 05:18

Jens

answered May 25 '13 at 00:57

dawg

Can NumPy omit the warnings though? This is probably the best solution as it would be the fastest and deals with signed/unsigned rather than hacking it with bit masks or modulo over in the Python native bigint side of things. – Gregory Morse Mar 16 '20 at 12:18
Yes, you can catch or suppress the errors and warnings. You can do that with the [warnings module](https://docs.python.org/2/library/warnings.html#warnings.filterwarnings) or with multiple numpy options. – dawg Mar 16 '20 at 15:16

2 Answers2