Fixed GH-17398: Decrease BCG's refcount before error

[SakiTakamachi]

No description provided.

[Girgias]

This makes sense to me.

I wonder if for master it makes sense to add a custom ZPP specifier like I did for GMP to do the error handling up front.

[nielsdos]

Doesn't make sense to me.

The fix is incomplete: you only guard the allocation bailout for bc_num2str_ex, but there are many other ways to get an allocation bailout that leaves the refcount of the global numbers incremented, resulting in a leak.

It also increases the complexity quite a bit for an exceptional case, which I don't like.

I think you should fix this by creating a dedicated bc_destroy_num function that destroys a bcmath number regardless of the refcount. This can then be used in GSHUTDOWN. This is safe because we're so far into shutdown that no PHP code could run anyway.

[SakiTakamachi]

@nielsdos

Indeed.
First of all, do you think a refcount is even necessary here?

edit:
I guess I was sleepy. ↑

Understood. I will try to discard the BCG regardless of the refcount value.