Fix GH-16890: array_sum() with GMP can loose precision (LLP64)

[cmb69]

We must use mpz_fits_si_p() instead of mpz_fits_slong_p() since the latter is not suitable for LLP64 data models.

---

Note that I've targeted PHP-8.3 since I haven't been able to find a suitable test case. Still, I think the actual fix should go into PHP-8.2.

[cmb69]

https://github.com/php/php-src/actions/runs/11957635359/job/33335259816?pr=16891 😢 mpz_fits_sint_p() is mpir only.

[devnexen]

Will we end up in another libreadline/libedit situation ? :)

[cmb69]

Will we end up in another libreadline/libedit situation ? :)

Not really. While libreadline and libedit have been developed independently, mpir is a fork of libgmp (5?), and as such has the same license. Either can be used on POSIX systems, but only mpir made fixes for LLP64 models; unfortunately, mpir is no longer maintained. Still using libgmp on Windows (if you can even built it with MSVC) doesn't make much sense, although in the long run we may need to stop using mpir on Windows. Also note that on Windows we're not using libedit, but rather WinEditLine, which is yet again somewhat different.

[Girgias]

Thank you, a lot of this code does seem a bit jank.... hopefully after I get round to merging the custom ZPP change for master we can improve it.

[cmb69]

Given that the fix required an ugly work-around for libgmp, that the reported issue was about array_sum(), that it has been self-reported, and PHP-8.2 is in security stage in a month, I've only applied to PHP-8.3 (and merged up).