argon2: add workaround for big 64-byte aligned allocations?

[newpavlov]

It was previously discussed in #566.

The claim is that for big allocations we can get higher performance by allocating len + 64 bytes with 1-byte alignment (we then can manually construct a 64-byte aligned region in the allocated memory with size len), than by directly allocating len bytes with 64-byte alignment.

cc @jonasmalacofilho

[jonasmalacofilho]

than by directly allocating len bytes with 16-byte alignment.

Small correction: the issue is with allocations using alignment greater than 16 (or, more generally, more than the maximum alignment supported by calloc).

In particular, the case we care about is allocating Blocks, which are 64-byte aligned.

[newpavlov]

Oh, you are right. Fixed.

BTW do we really need the 64 byte alignment in the first place? IIUC this alignment is too strict for SIMD vectors and it looks like an optimization which accounts for cache line size.

[jonasmalacofilho]

Yes, (I think) it's more about cache line size and, specifically, preventing false sharing. It gives a ~5% improvement over 16-byte alignment, if I recall correctly.

I also tried 128-byte alignment, which in theory makes sense for modern 64-bit architectures (including x86-64), and is the value adopted by most general solutions for false sharing (e.g. crossbeam::utils::CachePadded) on these platforms, but any further improvements were offset by more instructions being generated, at least when I tested it (c0ce7f9).

It probably also matters that in Argon2 we can't false share any block, just blocks on the boundaries of the slices. False sharing here isn't as much of an issue as it can be in other cases.

[newpavlov]

Have you tried to directly mmap the memory?

[jonasmalacofilho]

No, I haven't.