Re: [RFC PATCH] target/s390x: don't double ld_code() when reading instructions

[Richard Henderson]

On 10/12/21 7:52 AM, Alex Bennée wrote:
> > I think the plugin stuff could be more clever, knowing where the read
> > occurs within the sequence.  Otherwise, we should simplify the
> > interface so that it is not possible to make this mistake.
>
> It's plugin_insn_append which is doing the tracking here so we could
> extend the interface to include the current pc of the load and make the
> appropriate adjustments. That said it's a bunch hoops to jump every
> instruction when we could just as easily add an assert and fix up any
> cases where we do. I guess it comes down to how prevalent double dipping
> in the instruction stream is when constructing a translation?

Yes, which is why I suggested simplifying the interface to translate_ld*.  It currently 
takes the DisasContextBase; it could potentially read from pc_next, and increment it.  It 
would completely eliminate the problem you're encountering.

> What happens if the protection of the code area changes half way through
> a translation? Could a mapping change in flight?

No, we hold mmap_lock.

r~