Re: [PATCH v3 6/6] elf: Optimize _dl_new_hash in dl-new-hash.h

[Alexander Monakov <amonakov@ispras.ru>]

Hi,

I managed to run the new benchtest; I understand how you get a 2x speedup
on long inputs now.

Tigerlake belongs to Intel CPU generations that don't perform move
elimination for general registers (along with Icelake and Sandybridge).
So register copy in preparation to shift 'h' by 5 costs an extra cycle
on the critical path.

Furthermore, h*33 + *c gets evaluated as '((h + h*32) + *c) instead of
'((h + *c) + h*32)', which prevents interleaving the additions and thus
puts one extra add on the critical path.

(gcc-11 gets this right if you assist it by writing 'h = h + *c + h*32')

So due to the above issues, on Tigerlake you get 4 cycles for the original
loop, and also 4 cycles for the modified two-at-a-time loop.

(on top of that, gcc-10 fails to eliminate extra zero-extends and ends up
with two zero-extends for *c in the loop)

Alexander