Re: TCG performance on PPC64

[Matheus K. Ferst]

On 18/05/2022 11:11, Mark Cave-Ayland wrote:
> Perhaps related to your comment above about "something else", Talospace 
> reports that
> running TCG on a Talos II (Power 9) is actually quicker for running a 
> MacOS install
> under qemu-system-ppc compared with KVM-PR [1]. That's certainly odd, 
> and at odds
> with running KVM-PR on my G4 Mac Mini which gives a huge performance 
> boost over TCG
> here. That might suggest IO has a noticeable effect on 
> qemu-system-ppc(64) here.

That's an interesting result. I guess we can mitigate the possible IO 
impact by working from an initramfs in future tests.

> Slightly more back on topic, TCG can itself emulate certain instructions 
> that aren't
> implemented directly in the backends: e.g. compare tcg/i386/tcg-target.h 
> with
> tcg/ppc/tcg-target.h. Perhaps there is a "hot instruction" which is used 
> a lot during
> kernel boot which is being emulated rather than implemented directly in 
> the TCG PPC
> backend compared with i386?

Comparing the two files, PPC has zeros for 6 TCG_HAS_*_vec, 12 
TCG_HAS_*_i32, and 8 TCG_HAS_*_i64 macros that i386 implements. It might 
be possible to identify how frequent they are used with uprobes or 
something like that.

> Finally another comment from Richard about vector instruction use from 
> [2]: "As an
> aside, this does suggest to me that target/ppc might be well served in 
> moving the
> ppc_vsr_t members of CPUPPCState earlier, so that this offset is 
> smaller". Presumably
> this is because calculating smaller offsets can be done using fewer 
> instructions?
> However I suppose this would only have an effect on vector-heavy workloads.

It may be worth a try, ppc uses some vector insns for string 
manipulation that some services at early boot might use.

Thanks for your suggestions Mark,
Matheus K. Ferst
Instituto de Pesquisas ELDORADO <http://www.eldorado.org.br/>
Analista de Software
Aviso Legal - Disclaimer <https://www.eldorado.org.br/disclaimer.html>