[BUG] TypeDesc SEGFAULT from Copy Constructor

[Cthuflu]

Crash somewhat consistently occurs here in Blender 4.1.1 (OpenImageIO v2.5.12.0) on Fedora 40:

OpenImageIO/src/libutil/paramlist.cpp

Line 53 in 80541ad

memcpy(ptr, _value, size); //NOSONAR

When crash doesn't occur, programs like Blender (4.1.1 built with v2.5.6.0) continue running but with no icons, and the welcome image is also a black box. Tested with v2.5.6.0 up to v2.5.13.0, with v2.5.9.0 and later having the bug.

Cause of the issue seems to be the new default on Copy Constructor in PR #4162. Programs built with v2.5.6.0 but using a linked version later than v2.5.8.0 will SEGFAULT from the Copy Constructor change in distributed headers. In the Blender case it somehow changes the arraylen from 0 to 32727 for a TypeDesc from changing a property in ImageDesc.

Temporarily reverting these changes does fix the issue for those cases. With this being built into distributed system headers it seems like it would be a breaking change for a minor API version; the alternative is tracking down how the arraylen value changes without touching any other code. rebuilding applications with the newer headers.

[lgritz]

Sorry for the delay in looking at this.

I didn't expect that change to make an incompatibility, but I guess it did somehow. In retrospect, I wish I'd kept that fix strictly in master and not backported it.

I'm not sure that reverting is going to help much at this point -- it might make 2.5.>=15 be compatible again with 2.5.<=8, but we'll still (forever) be in the situation where mixing headers and libraries for 2.5.<=8 and 2.5.9-14 will be incompatible and so will be mixing 2.5.9-14 and 2.5.15+. That feels like it's just shifting the problem around and not eliminating it.

Given that 2.6/3.0 will be the main supported release family soon, is it worth trying to fix this in 2.5? Do you (or anyone else reading) even have a specific suggestion for how to fix 2.5 for real? Or is the best we can do to advise everybody to upgrade to 2.5.9+ and try not to mix headers and libraries from different sides of that boundary?

[Cthuflu]

I believe at this point it is best to mark that version as a boundary.