This shouldn't really be a fundamental limitation of the changes we did to auto-save, it's more like something that was overlooked.
I did notice that there is this line: `const int fileflags =…
Would be good to make sure that the path is actually absolute. From what I know, it may still be relative here.
Yeah, I optimized it a bit more to avoid performance regressions in a few cases, but I think there are more potential optimizations. For example, one could always process 64 bools at once: convert…
Yeah, that would be good. Two notes:
BitVectoris a bit special in that sense, because it initializes the bits always currently, even if no explicit initializer is given. That's necessary to…
Yeah, thought so, will leave this to someone else. The code here should not run in practice ideally.
It doesn't really matter if this buffer is oversized. Everything that's not used of it is never touched.
That said, I did notice that a buffer larger than max_segment_size/2 is probably never…
The u in _mm_loadu_si128 stands for unaligned. So that should be fine.
Not sure what you mean specifically. This seems somewhat unrelated to offset indices.
Not sure how popcount could be faster than just comparing with zero.