AcademySoftwareFoundation
Fix undefined aliasing issue when compiling with clang-cl
Fixes the inner ValueT alias throw an undefined error when compiling with clang/clang-cl on Windows by converting the convertValue into a template function.
The committers listed above are authorized under a signed CLA.
- :white_check_mark: login: bgottamd (776b6fd6beaa4805baf73f0fc731af08659087fd)
![linux-foundation-easycla[bot] avatar](https://avatars.githubusercontent.com/in/17893?v=4 "Published by a pub.dev verified publisher"){kind=small}
Adding a quick bump. This small code change is essential so that Blender can support both clang - merged PR here - and openvdb. Without it, blender fails to build due to the cited issue.
any progress here? We'd like to get this working out of the box for blender, and while i have no issue including selected upstream landed patches for our dependencies, upstream unreviewed patches are a bit of a different story.
I've hit the same clang bug before and reported it: llvm/llvm-project#64996. Using -fno-delayed-template-parsing, as was suggested there, makes OpenVDB compile successfully with clang-cl.
Thanks @jo-rn - this does fix the issue with the header file compilation in blender. However, there is some performance degradation when this flag is used:
| File | args | fastest | slowest | avg |
|---|---|---|---|---|
| classroom 3840x2160 256 samples | header template fix | 1209 | 1217 | 1213 |
| -fno-delayed-template-parsing | 1218 | 1228 | 1223 |
This equates to ~1% drop in performance just from this flag. Hoping this templating change or a modification of the alias can still be merged in.
Odd result, classroom is a blender 2.79 scene, openvdb wasn't integrated till blender 2.83, so the odds of it having an openvdb volume has to be near the low end of things. Did you put the -fno-delayed-template-parsing flag globally on the whole blender project? or isolated it to just the few openvdb related files that needed this fix?
I isolated the flag to projects that complained about the header issue. This ended up being only extern_mantaflow and cycles_scene
For those keeping track of this issue affecting Blender, I have created a PR to work around the issue using the approach described in this thread: https://projects.blender.org/blender/blender/pulls/126236
