https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106635
--- Comment #11 from Richard Earnshaw <rearnsha at gcc dot gnu.org> --- (In reply to Xiaoguang from comment #9) > Yeah, I also find such description, my memory type is uncachable normal > memory, but not device memory > I use mmap to get the virtual address with an O_SYNC in fd Yes, I think you're correct: you can have non-cacheable normal memory and that supports unaligned accesses. However, I don't know how you are supposed to set that up with Linux mmap calls. It's beyond my knowledge of how the kernel works with the system. You'd be best asking about that in a different forum where some kernel experts hang out. I don't know what other constraints you might have on your system, but have you considered using cacheable memory and then inserting cache synchronization calls at the points where you really need to synchronize data? That may well have significantly better performance than trying to map the data as non-cacheable. Note that if you have simultaneous readers of the data, you'll need to mark the pointers to it as volatile anyway, because otherwise the compiler might re-order the stores to separate locations.
