Use SIMD instructions to update data pointers.

[Shark64]

Hi, I've noticed that even if the hashes code uses vector instructions for the computation the update of the data pointers at the end is still done using scalar code. We can save some time and code size by loading the pointers as a vector, update them all together and store them back.
I've run ``make test'' and it worked without errors on Linux_x86-64 with RocketLake CPU.

[Bulat-Ziganshin]

Do you know about latencies of [failed] store-forwarding? Read e.g. https://stackoverflow.com/questions/46135766/can-modern-x86-implementations-store-forward-from-more-than-one-prior-store :

A read that is bigger than the write, or a read that covers both written and unwritten bytes, takes approximately 11 clock cycles extra.

OTOH, I don't know the rest of this code. The better strategy may be storing pointers in a SIMD register permanently and using e.g. PEXTRD to copy them into scalar registers. Or even using VPGATHER where it available: https://stackoverflow.com/questions/21774454/how-are-the-gather-instructions-in-avx2-implemented

[Shark64]

I'm familiar with store-forwarding, but it doesn't look like it should be a problem here. The pointers are updated only once at the end of the function, outside the hot loop. My patch it's mainly to reduce code size, not for performace. Also, as your quote says the store-fowarding fails if your read back a bigger chuck than the older write, so writing them as a single vector should help by having all the data in one or two store buffers.

[Bulat-Ziganshin]

you are right, sorry for the noise. i didn't check the code before commenting :(

[Shark64]

you are right, sorry for the noise. i didn't check the code before commenting :(

No worries ;)

[pablodelara]

Thanks for your PR. I left a comment.

[pablodelara]

One more change needed, thanks

[pablodelara]

Thanks for the changes. I will squash the third commit into the first commit once this is merged.

[pablodelara]

After giving an extra thought, I don't think making this change is worth it, for two reasons:
1 - It's replacing 4 instructions with 5 instructions
2 - We are removing AVX code in the next few months (already mentioned it).
So could you drop the changes in the AVX implementations and leave just the other ones?
Many thanks for the work!

[Shark64]

Sure, I've just reverted the _avx functions to the old version.

[pablodelara]

Thanks for the work @Shark64. This is now merged in mainline.