Re: [PATCH v3 27/30] target/ppc: Move D/DS/X-form integer loads to decodetree

[Matheus K. Ferst]

On 30/04/2021 21:50, Richard Henderson wrote:
> On 4/30/21 4:54 PM, Matheus K. Ferst wrote:
> > The only difference between those two is 
> > tcg_gen_addi_tl/tcg_gen_movi_tl and tcg_gen_add_tl/tcg_gen_mov_tl. We 
> > could do this in a single method if we tcg_const_tl(a->si) in 
> > do_ldst_D. I'm not sure about the costs involved, and we'd need to 
> > tcg_temp_free it. If you want to give it a look, I did some tests on 
> > https://github.com/ppc64/qemu/tree/ferst-tcg-decodetree64 .
>
> I guess it works, but it feels ugly with the comparison vs cpu_gpr[0].  
> Maybe pass in rt and ra as integers, and only pass in rb/si as TCGv addend?

I tried that first and it looked just as weird, but if you prefer I'll 
change it back. We could also hide the ugliness in a helper e.g. bool 
gpr_is_zero(TCGv).

> The upper-case argument names clash with docs/devel/style.rst.

Good call, I'll fix this before push this branch again;

> You can use tcg_constant_tl() to grab a hashed constant that need not be 
> freed.

Then we could

> return do_ldst(ctx, cpu_gpr[a->rt], 
a->ra?cpu->gpr[a->ra]:tcg_const_tl(0), ...

At the cost of always calling tcg_gen_add_tl and repeating the RA=0 
logic in do_ldst_X and do_ldst_D.

> BTW, I thought I had a comment about PLS being used for both MLS and 
> 8LS.  I guess that must have gotten lost at some point.  If you go with 
> your renaming to MLS, you'll want a different comment about PLD et al.
>
>
> r~

Don't mind that, it was a bad idea, I gave up on this change while 
rebasing v3.

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