pjacklam / p5-math-bigint Goto Github PK

Hi,

as reported by Klaus Ethgen in https://bugs.debian.org/1012704 this hangs with recent versions of Math-BigInt:

% perl -Mbignum -e '1 | (1 >> 1)'
Deep recursion on subroutine "Math::BigInt::bior" at /usr/share/perl5/Math/BigFloat.pm line 3883.
Deep recursion on subroutine "Math::BigFloat::bior" at /usr/share/perl5/Math/BigInt.pm line 3513.

Also happens with other bitwise operations like & and ^ .

It seems to have regressed around 1.999832 (where it started to spit errors) and 1.999834 (where the errors became infinite recursion.)

In LedgerSMB, we have a class called LedgerSMB::PGNumber which ultimately ISA Math::BigFloat. Since this class adds a few methods, it's important that the result of each operation remains of the same class. While this used to work pretty well (mostly deployed on Debian versions), this is now broken (on Debian Bookworm). Going back to Bullseye and installing the latest Math::BigInt distribution from CPAN breaks the same way.

Reading CHANGES, I suspect 1.999830 and 1.999831: Bullseye has 1.999818 and Bookworm has 1.999838.

Could you please indicate whether the current state is permanent? If so, what is your advice as to how to deal with the situation?

Should you need to look at the code, the LedgerSMB::PGNumber code is here. It's inheriting from Math::BigFloat through here

Was previously running Math-BigInt-1.999818. Tried to upgrade and found unexpected behavior when calling numify after a bfround.

my $f = Math::BigFloat->new(123456.123456);
say $f->bfround(-2)->numify;
# 123000

Since this repo is now the official one, or at least the advertised one on MetaCPAN

Note that two PRs are pending from
https://github.com/rafl/Math-BigInt/pulls

I'm unsure about their state, if they are already merged here or not.

it would be good to close future PR to the parent repo, and point them here instead.

As seen here https://github.com/pjacklam/p5-Math-BigInt/tags
the last tag available is 1.999813, we are already at 1.999828

Having tags is convenient to use GitHub UI to perform some quick compare for example or even from the command line.

Unfortunately I do not have permission to create these tags, I will gladly help and provide them if needed.

Thanks for considering creating these tags

It looks like 46d1252 / 1.999836 broke numerical equivalence of Math::BigFloat subclass instances with Math::BigFloat instances. This affects at least the LedgerSMB test suite via the PGObject-Type-BigFloat CPAN module, which is how we noticed it in https://bugs.debian.org/1043234 .

A test case is

package MyFloat;
use base qw(Math::BigFloat);
1;

print MyFloat->new(9999.99) == Math::BigFloat->new(9999.99) ? "ok\n": "not ok\n";
print Math::BigFloat->new(9999.99) == MyFloat->new(9999.99) ? "ok\n": "not ok\n";

which fails the first test (but not the second one.)