Re: [RFC PATCH v2 07/30] target/loongarch: Add MMU support for LoongArch CPU.

[Richard Henderson]

On 11/17/21 7:37 AM, yangxiaojuan wrote:
> > > +        *physical = (tlb_ppn1 << 12) | (address & ((1 << tlb_ps) - 1));
> >
> > TARGET_PAGE_SIZE.
>
> Maybe TARGET_PAGE_SIZE is not fit for a huge page. MAKE_64BIT_MASK(0, tlb_ps) 
> is ok?

I meant the first << 12.  But, yes, MAKE_64BIT_MASK is a good improvement.

> > Bit of a shame to have a linear search.  I guess it's ok for a start, but 
> > you'll find that this function is critical to the emulation speed of qemu, so 
> > you should think about other ways to organize the data.
>
> The stlb search by the set, the mtlb uses the linear search, I have no other 
> idea, do you have some advice?

Well, stlb are all the same page size, and duplicate matches are undefined.

A hash table of the valid entries might be workable, with the virtual page and asid taken 
into account.  I'd imagine that would get you down from 2048 comparisons to just a few.

The mtlb entries may all have different page size, so I don't immediately know how to 
search them more efficiently.  But there are only 64 of them.

It might be worth having a couple of bitsets that summarize the valid entries.  This could 
be used to search less than 64 mtlb entries.  It could also be used to influence the 
"random" selection of a new tlb entry: always select an empty tlb entry first, then evict 
a random entry.


r~