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

[Noah Goldstein]

On Wed, Apr 27, 2022 at 11:17 AM Alexander Monakov <amonakov@ispras.ru> wrote:
>
> On Wed, 27 Apr 2022, Noah Goldstein wrote:
>
> > > However, when you reroll the loop and overlap two iterations, multiplication
> > > by 33*33 no longer has this nice property and runs with two 4cyc paths
> > > overlapped (so effective critical path is the same as original).
> >
> > the 33 * c0 can still use `addl; sall; addl` so not sure what you mean by
> > two 4cyc paths overlapped. Its one 4c path.
> >
> > `imul; addl` and `addl; sall; addl`.
> >
> > But it's fair that either wait its 4c of computation for 2 iterations. The
> > difference is the 5c load latency being amortized over 2 iterations
> > or 1 iteration.
>
> Right, it's one 4c path, I was thinking about something else for a moment.
> I'm not sure it's correct to amortize load latency like that, I'd say the
> difference is just that the original loop cannot issue two loads at once
> because of the dependency in its address computation.
>
> I see you dropped libc-alpha@ from Cc:, was that intentional?
No misclick sorry. Adding it back.

I think it is the way you're doing your analysis as a loop-carried
dependency. I.e really 7c per iteration with no unroll (although
its fair the loads on address can speculate ahead so it will
indeed be faster) vs 9c per 2x iterations.

>
> Alexander