Ruby

Date#ajd (astronomical Julian Day number) returns incorrect positive values instead of negative values for certain large negative years due to signed integer overflow in the RB_INT2FIX function.

require 'date'

# Works correctly
Date.civil(-110823815392979, 2, 19).ajd
#=> (-80956797141128945/2)

# BUG - returns positive instead of negative  
Date.civil(-11082381539297990, 2, 19).ajd  
#=> (1127692322401036327/2)  # Should be negative!

# Works correctly (takes a different code path)
Date.civil(-111111111082381539297990, 2, 19).ajd
#=> (-81166666645679714453739481/2)

In ext/date/date_core.c, the ir value can overflow

if (FIXNUM_P(r) && FIX2LONG(r) <= (FIXNUM_MAX / 2)) {
    long ir = FIX2LONG(r);
    ir = ir * 2 - 1;  
    // On Date.civil(-11082381539297990, 2, 19).ajd
    //
    // ir = -8095679714453739481
    return rb_rational_new2(LONG2FIX(ir), INT2FIX(2));  // Overflow in LONG2FIX
}

Then in include/ruby/internal/arithmetic/long.h (LONG2FIX is an alias for RB_INT2FIX)

static inline VALUE
RB_INT2FIX(long i)
{
    RBIMPL_ASSERT_OR_ASSUME(RB_FIXABLE(i));

    /* :NOTE: VALUE can be wider than long.  As j being unsigned, 2j+1 is fully
     * defined. Also it can be compiled into a single LEA instruction. */
    const unsigned long j = RBIMPL_CAST((unsigned long)i);
    const unsigned long k = (j << 1) + RUBY_FIXNUM_FLAG;
    const long          l = RBIMPL_CAST((long)k);
    const SIGNED_VALUE  m = l; /* Sign extend */
    const VALUE         n = RBIMPL_CAST((VALUE)m);

    RBIMPL_ASSERT_OR_ASSUME(RB_FIXNUM_P(n));
    return n;
}

It effectively goes:

j = (unsigned long)-8095679714453739481  // 10351064359255812135
k = (10351064359255812135 << 1) + RUBY_FIXNUM_FLAG // 2255384644802072655
...

So, 2255384644802072655 encodes 1127692322401036327, which is the numerator for the observed ajd return value in the reproducer above.

It also affects Date comparisons (since ajd is used in cmp_gen)