Re: [RFC PATCH 0/3] translation performance improvements

["Nicholas Piggin" <npiggin@gmail.com>]

On Tue Apr 1, 2025 at 5:40 AM AEST, Richard Henderson wrote:
> On 3/31/25 10:54, Nicholas Piggin wrote:
>> I've been struggling with these couple of performance issues with
>> TB coherency. I almost thought deferring flush to icbi would be
>> workable, buta note in the docs says that exceptions require TB
>> to be coherent... I don't know what requires that, maybe it could
>> be worked around?
>
> Which note?  Anyway, qemu implements accurate tb invalidation for x86 and s390x, which 
> means we don't really need to do anything special for other targets.

In docs/devel/tcg.rst

On RISC targets, correctly written software uses memory barriers and
cache flushes, so some of the protection above would not be
necessary. However, QEMU still requires that the generated code always
matches the target instructions in memory in order to handle
exceptions correctly.

>
> Compare aarch64 "IC_IVAU" which (at least for system mode) is implemented as a nop.

I'll take a look at it.


>> Another thing is PowerVM runtime firmware runs with MMU disabled
>> for ifetch. This means a fixed linear map with no memory protection.
>> Is it possible we can enable goto tb across TARGET_PAGE_SIZE for
>> ifetches in this mode?
>
> No, there are several things that assume nothing jumps across TARGET_PAGE_SIZE, including 
> breakpoints.

I see. It did actually work and run fine, so I wonder how much effort
it would take to cater for these issues. I guess for this rather niche
"real mode" it may not be worth bending over backward.

Thanks,
Nick