Compiling with double-precision for macOS <10.14 fails about joltphysics HOT 2 CLOSED

So if I understand the issue correctly, the problem occurs on x86_64 only and only on macOS 10.12. I'm not sure what the stack alignment on macOS is but it is indeed most likely 16. The warning is then probably triggered because the compiler cannot guarantee that the struct will be aligned to 32 bytes. If this is a problem or not really depends on the instructions that the compiler generates, in the old days compilers generated 'load aligned' (e.g. _mm256_load_pd) instructions which would break if the object wasn't 32 byte aligned. I think modern compilers don't do this anymore (they emit e.g. _mm256_loadu_pd), so I would expect that ignoring the warning would not really cause problems.

B.t.w. it doesn't help for this particular case because they only support Apple Silicon macs, but I don't own a mac either and I use scaleway to rent a virtual mac for day if I need one.

from joltphysics.

and only on macOS 10.12

Technically when targeting any version prior to 10.14 I guess, but yes, 10.12 and 10.13 in this case.

The warning is then probably triggered because the compiler cannot guarantee that the struct will be aligned to 32 bytes

I'm out of my depth here, but I don't get the impression that this is what the problem is. Judging by the error I get the impression that the compiler has some generated code in the implicit destructor (although I can't fathom what exactly) that makes use of the global delete that takes a std::align_val_t, which from what I've gathered it would have to use for deleting anything with an alignment larger than sizeof(max_align_t), as is the case here. macOS/OSX 10.12 doesn't seem to provide these in its C++ runtime, presumably for the same reason it didn't provide aligned_alloc prior to 10.15, as mentioned in #738, which I guess makes some kind of sense since they're both relatively recent additions (C++17).

Either way, given that this would probably require testing with whatever C++ runtime 10.12 shipped with, I don't think either one of us will be able to definitively know if any fix is truly viable, but I'd probably be fine with adding -faligned-allocation either on my end or yours and call it a day.

B.t.w. it doesn't help for this particular case because they only support Apple Silicon macs, but I don't own a mac either and I use scaleway to rent a virtual mac for day if I need one.

Ah, good to know, thank you!

from joltphysics.