Poor accuracy in true division of integers #474

Activity

1. msimacek

   self-assigned this

   on Feb 12, 2025
2. graalvmbot

   closed this as completedin 9513ba0on Feb 13, 2025
3. timfel

   added a commit that references this issue on Mar 18, 2025

   Increase precision of integer truediv

   ...

   This issue will close once commit 9513ba0 is merged into the 'master' branch.

   9513ba0
4. skirpichev commented on Nov 5, 2025

   

   skirpichev

   on Nov 5, 2025

   Author

   More actions

   @msimacek is this included in the 25.0.1 or in 25.0.0?

   In CI I get following failure (building gp312_250-manylinux_x86_64 wheel):

   E           assert (mpz(2305843008955759739) / mpz(784637716923335094969050127462454050937374032249308213819)) == 2.9387358767269886e-39
   

   CPython's result (which is same for mpz):

   Python 3.14.0 (tags/v3.14.0:ebf955df7a8, Oct 15 2025, 06:24:48) [GCC 12.2.0] on linux
   Type "help", "copyright", "credits" or "license" for more information.
   >>> 2305843008955759739 / 784637716923335094969050127462454050937374032249308213819
   2.9387358767269883e-39

   I don't think this is fixed.
5. skirpichev commented on Nov 19, 2025

   

   skirpichev

   on Nov 19, 2025

   Author

   More actions

   @msimacek can this be reopened or should I open a new issue?

   I downloaded latest stable and run. It seems the fix was applied, but in my new example it's not enough:

   Python 3.12.8 (Thu Oct 09 17:26:36 UTC 2025)
   [Graal, Oracle GraalVM, Java 25.0.1 (amd64)] on linux
   Type "help", "copyright", "credits" or "license" for more information.
   >>> x, y = 15649, 2024092432744435552  # now ok
   >>> x/y
   7.731366288831862e-15
   >>> 2305843008955759739 / 784637716923335094969050127462454050937374032249308213819  # this is wrong
   2.9387358767269886e-39
   >>> _.hex()  # one bit off, correct answer is '0x1.ffffffff0a03cp-129'
   '0x1.ffffffff0a03dp-129'
6. msimacek commented on Nov 19, 2025

   

   msimacek

   on Nov 19, 2025

   Contributor

   More actions

   Hi, sorry, let's reopen this
7. msimacek

   reopened this on Nov 19, 2025
8. skirpichev commented on Dec 12, 2025

   

   skirpichev

   on Dec 12, 2025

   Author

   More actions

   BTW,

   The result expected to be correctly rounded with the round-half-to-even rule.

   it seems, this is not something documented. The Python doesn't specify behavior of true division at all, see discussion in python/cpython#142449. And with the current optimization in the algorithm, above property might be violated on some systems.